xref: /linux/drivers/nvme/host/rdma.c (revision bb1c928df78ee6e3665a0d013e74108cc9abf34b)
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/types.h>
23 #include <linux/list.h>
24 #include <linux/mutex.h>
25 #include <linux/scatterlist.h>
26 #include <linux/nvme.h>
27 #include <asm/unaligned.h>
28 
29 #include <rdma/ib_verbs.h>
30 #include <rdma/rdma_cm.h>
31 #include <linux/nvme-rdma.h>
32 
33 #include "nvme.h"
34 #include "fabrics.h"
35 
36 
37 #define NVME_RDMA_CONNECT_TIMEOUT_MS	3000		/* 3 second */
38 
39 #define NVME_RDMA_MAX_SEGMENT_SIZE	0xffffff	/* 24-bit SGL field */
40 
41 #define NVME_RDMA_MAX_SEGMENTS		256
42 
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS	1
44 
45 /*
46  * We handle AEN commands ourselves and don't even let the
47  * block layer know about them.
48  */
49 #define NVME_RDMA_NR_AEN_COMMANDS      1
50 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
51 	(NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
52 
53 struct nvme_rdma_device {
54 	struct ib_device       *dev;
55 	struct ib_pd	       *pd;
56 	struct kref		ref;
57 	struct list_head	entry;
58 };
59 
60 struct nvme_rdma_qe {
61 	struct ib_cqe		cqe;
62 	void			*data;
63 	u64			dma;
64 };
65 
66 struct nvme_rdma_queue;
67 struct nvme_rdma_request {
68 	struct nvme_request	req;
69 	struct ib_mr		*mr;
70 	struct nvme_rdma_qe	sqe;
71 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
72 	u32			num_sge;
73 	int			nents;
74 	bool			inline_data;
75 	struct ib_reg_wr	reg_wr;
76 	struct ib_cqe		reg_cqe;
77 	struct nvme_rdma_queue  *queue;
78 	struct sg_table		sg_table;
79 	struct scatterlist	first_sgl[];
80 };
81 
82 enum nvme_rdma_queue_flags {
83 	NVME_RDMA_Q_LIVE		= 0,
84 	NVME_RDMA_Q_DELETING		= 1,
85 };
86 
87 struct nvme_rdma_queue {
88 	struct nvme_rdma_qe	*rsp_ring;
89 	atomic_t		sig_count;
90 	int			queue_size;
91 	size_t			cmnd_capsule_len;
92 	struct nvme_rdma_ctrl	*ctrl;
93 	struct nvme_rdma_device	*device;
94 	struct ib_cq		*ib_cq;
95 	struct ib_qp		*qp;
96 
97 	unsigned long		flags;
98 	struct rdma_cm_id	*cm_id;
99 	int			cm_error;
100 	struct completion	cm_done;
101 };
102 
103 struct nvme_rdma_ctrl {
104 	/* read only in the hot path */
105 	struct nvme_rdma_queue	*queues;
106 
107 	/* other member variables */
108 	struct blk_mq_tag_set	tag_set;
109 	struct work_struct	delete_work;
110 	struct work_struct	err_work;
111 
112 	struct nvme_rdma_qe	async_event_sqe;
113 
114 	struct delayed_work	reconnect_work;
115 
116 	struct list_head	list;
117 
118 	struct blk_mq_tag_set	admin_tag_set;
119 	struct nvme_rdma_device	*device;
120 
121 	u32			max_fr_pages;
122 
123 	struct sockaddr_storage addr;
124 	struct sockaddr_storage src_addr;
125 
126 	struct nvme_ctrl	ctrl;
127 };
128 
129 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
130 {
131 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
132 }
133 
134 static LIST_HEAD(device_list);
135 static DEFINE_MUTEX(device_list_mutex);
136 
137 static LIST_HEAD(nvme_rdma_ctrl_list);
138 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
139 
140 /*
141  * Disabling this option makes small I/O goes faster, but is fundamentally
142  * unsafe.  With it turned off we will have to register a global rkey that
143  * allows read and write access to all physical memory.
144  */
145 static bool register_always = true;
146 module_param(register_always, bool, 0444);
147 MODULE_PARM_DESC(register_always,
148 	 "Use memory registration even for contiguous memory regions");
149 
150 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
151 		struct rdma_cm_event *event);
152 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
153 
154 /* XXX: really should move to a generic header sooner or later.. */
155 static inline void put_unaligned_le24(u32 val, u8 *p)
156 {
157 	*p++ = val;
158 	*p++ = val >> 8;
159 	*p++ = val >> 16;
160 }
161 
162 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
163 {
164 	return queue - queue->ctrl->queues;
165 }
166 
167 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
168 {
169 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
170 }
171 
172 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
173 		size_t capsule_size, enum dma_data_direction dir)
174 {
175 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
176 	kfree(qe->data);
177 }
178 
179 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
180 		size_t capsule_size, enum dma_data_direction dir)
181 {
182 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
183 	if (!qe->data)
184 		return -ENOMEM;
185 
186 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
187 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
188 		kfree(qe->data);
189 		return -ENOMEM;
190 	}
191 
192 	return 0;
193 }
194 
195 static void nvme_rdma_free_ring(struct ib_device *ibdev,
196 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
197 		size_t capsule_size, enum dma_data_direction dir)
198 {
199 	int i;
200 
201 	for (i = 0; i < ib_queue_size; i++)
202 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
203 	kfree(ring);
204 }
205 
206 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
207 		size_t ib_queue_size, size_t capsule_size,
208 		enum dma_data_direction dir)
209 {
210 	struct nvme_rdma_qe *ring;
211 	int i;
212 
213 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
214 	if (!ring)
215 		return NULL;
216 
217 	for (i = 0; i < ib_queue_size; i++) {
218 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
219 			goto out_free_ring;
220 	}
221 
222 	return ring;
223 
224 out_free_ring:
225 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
226 	return NULL;
227 }
228 
229 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
230 {
231 	pr_debug("QP event %s (%d)\n",
232 		 ib_event_msg(event->event), event->event);
233 
234 }
235 
236 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
237 {
238 	wait_for_completion_interruptible_timeout(&queue->cm_done,
239 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
240 	return queue->cm_error;
241 }
242 
243 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
244 {
245 	struct nvme_rdma_device *dev = queue->device;
246 	struct ib_qp_init_attr init_attr;
247 	int ret;
248 
249 	memset(&init_attr, 0, sizeof(init_attr));
250 	init_attr.event_handler = nvme_rdma_qp_event;
251 	/* +1 for drain */
252 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
253 	/* +1 for drain */
254 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
255 	init_attr.cap.max_recv_sge = 1;
256 	init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
257 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
258 	init_attr.qp_type = IB_QPT_RC;
259 	init_attr.send_cq = queue->ib_cq;
260 	init_attr.recv_cq = queue->ib_cq;
261 
262 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
263 
264 	queue->qp = queue->cm_id->qp;
265 	return ret;
266 }
267 
268 static int nvme_rdma_reinit_request(void *data, struct request *rq)
269 {
270 	struct nvme_rdma_ctrl *ctrl = data;
271 	struct nvme_rdma_device *dev = ctrl->device;
272 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
273 	int ret = 0;
274 
275 	ib_dereg_mr(req->mr);
276 
277 	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
278 			ctrl->max_fr_pages);
279 	if (IS_ERR(req->mr)) {
280 		ret = PTR_ERR(req->mr);
281 		req->mr = NULL;
282 		goto out;
283 	}
284 
285 	req->mr->need_inval = false;
286 
287 out:
288 	return ret;
289 }
290 
291 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
292 		struct request *rq, unsigned int hctx_idx)
293 {
294 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
295 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
296 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
297 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
298 	struct nvme_rdma_device *dev = queue->device;
299 
300 	if (req->mr)
301 		ib_dereg_mr(req->mr);
302 
303 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
304 			DMA_TO_DEVICE);
305 }
306 
307 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
308 		struct request *rq, unsigned int hctx_idx,
309 		unsigned int numa_node)
310 {
311 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
312 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
313 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
314 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
315 	struct nvme_rdma_device *dev = queue->device;
316 	struct ib_device *ibdev = dev->dev;
317 	int ret;
318 
319 	ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
320 			DMA_TO_DEVICE);
321 	if (ret)
322 		return ret;
323 
324 	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
325 			ctrl->max_fr_pages);
326 	if (IS_ERR(req->mr)) {
327 		ret = PTR_ERR(req->mr);
328 		goto out_free_qe;
329 	}
330 
331 	req->queue = queue;
332 
333 	return 0;
334 
335 out_free_qe:
336 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
337 			DMA_TO_DEVICE);
338 	return -ENOMEM;
339 }
340 
341 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
342 		unsigned int hctx_idx)
343 {
344 	struct nvme_rdma_ctrl *ctrl = data;
345 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
346 
347 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
348 
349 	hctx->driver_data = queue;
350 	return 0;
351 }
352 
353 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
354 		unsigned int hctx_idx)
355 {
356 	struct nvme_rdma_ctrl *ctrl = data;
357 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
358 
359 	BUG_ON(hctx_idx != 0);
360 
361 	hctx->driver_data = queue;
362 	return 0;
363 }
364 
365 static void nvme_rdma_free_dev(struct kref *ref)
366 {
367 	struct nvme_rdma_device *ndev =
368 		container_of(ref, struct nvme_rdma_device, ref);
369 
370 	mutex_lock(&device_list_mutex);
371 	list_del(&ndev->entry);
372 	mutex_unlock(&device_list_mutex);
373 
374 	ib_dealloc_pd(ndev->pd);
375 	kfree(ndev);
376 }
377 
378 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
379 {
380 	kref_put(&dev->ref, nvme_rdma_free_dev);
381 }
382 
383 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
384 {
385 	return kref_get_unless_zero(&dev->ref);
386 }
387 
388 static struct nvme_rdma_device *
389 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
390 {
391 	struct nvme_rdma_device *ndev;
392 
393 	mutex_lock(&device_list_mutex);
394 	list_for_each_entry(ndev, &device_list, entry) {
395 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
396 		    nvme_rdma_dev_get(ndev))
397 			goto out_unlock;
398 	}
399 
400 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
401 	if (!ndev)
402 		goto out_err;
403 
404 	ndev->dev = cm_id->device;
405 	kref_init(&ndev->ref);
406 
407 	ndev->pd = ib_alloc_pd(ndev->dev,
408 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
409 	if (IS_ERR(ndev->pd))
410 		goto out_free_dev;
411 
412 	if (!(ndev->dev->attrs.device_cap_flags &
413 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
414 		dev_err(&ndev->dev->dev,
415 			"Memory registrations not supported.\n");
416 		goto out_free_pd;
417 	}
418 
419 	list_add(&ndev->entry, &device_list);
420 out_unlock:
421 	mutex_unlock(&device_list_mutex);
422 	return ndev;
423 
424 out_free_pd:
425 	ib_dealloc_pd(ndev->pd);
426 out_free_dev:
427 	kfree(ndev);
428 out_err:
429 	mutex_unlock(&device_list_mutex);
430 	return NULL;
431 }
432 
433 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
434 {
435 	struct nvme_rdma_device *dev;
436 	struct ib_device *ibdev;
437 
438 	dev = queue->device;
439 	ibdev = dev->dev;
440 	rdma_destroy_qp(queue->cm_id);
441 	ib_free_cq(queue->ib_cq);
442 
443 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
444 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
445 
446 	nvme_rdma_dev_put(dev);
447 }
448 
449 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
450 {
451 	struct ib_device *ibdev;
452 	const int send_wr_factor = 3;			/* MR, SEND, INV */
453 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
454 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
455 	int ret;
456 
457 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
458 	if (!queue->device) {
459 		dev_err(queue->cm_id->device->dev.parent,
460 			"no client data found!\n");
461 		return -ECONNREFUSED;
462 	}
463 	ibdev = queue->device->dev;
464 
465 	/*
466 	 * The admin queue is barely used once the controller is live, so don't
467 	 * bother to spread it out.
468 	 */
469 	if (idx == 0)
470 		comp_vector = 0;
471 	else
472 		comp_vector = idx % ibdev->num_comp_vectors;
473 
474 
475 	/* +1 for ib_stop_cq */
476 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
477 				cq_factor * queue->queue_size + 1,
478 				comp_vector, IB_POLL_SOFTIRQ);
479 	if (IS_ERR(queue->ib_cq)) {
480 		ret = PTR_ERR(queue->ib_cq);
481 		goto out_put_dev;
482 	}
483 
484 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
485 	if (ret)
486 		goto out_destroy_ib_cq;
487 
488 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
489 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
490 	if (!queue->rsp_ring) {
491 		ret = -ENOMEM;
492 		goto out_destroy_qp;
493 	}
494 
495 	return 0;
496 
497 out_destroy_qp:
498 	ib_destroy_qp(queue->qp);
499 out_destroy_ib_cq:
500 	ib_free_cq(queue->ib_cq);
501 out_put_dev:
502 	nvme_rdma_dev_put(queue->device);
503 	return ret;
504 }
505 
506 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
507 		int idx, size_t queue_size)
508 {
509 	struct nvme_rdma_queue *queue;
510 	struct sockaddr *src_addr = NULL;
511 	int ret;
512 
513 	queue = &ctrl->queues[idx];
514 	queue->ctrl = ctrl;
515 	init_completion(&queue->cm_done);
516 
517 	if (idx > 0)
518 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
519 	else
520 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
521 
522 	queue->queue_size = queue_size;
523 	atomic_set(&queue->sig_count, 0);
524 
525 	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
526 			RDMA_PS_TCP, IB_QPT_RC);
527 	if (IS_ERR(queue->cm_id)) {
528 		dev_info(ctrl->ctrl.device,
529 			"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
530 		return PTR_ERR(queue->cm_id);
531 	}
532 
533 	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
534 		src_addr = (struct sockaddr *)&ctrl->src_addr;
535 
536 	queue->cm_error = -ETIMEDOUT;
537 	ret = rdma_resolve_addr(queue->cm_id, src_addr,
538 			(struct sockaddr *)&ctrl->addr,
539 			NVME_RDMA_CONNECT_TIMEOUT_MS);
540 	if (ret) {
541 		dev_info(ctrl->ctrl.device,
542 			"rdma_resolve_addr failed (%d).\n", ret);
543 		goto out_destroy_cm_id;
544 	}
545 
546 	ret = nvme_rdma_wait_for_cm(queue);
547 	if (ret) {
548 		dev_info(ctrl->ctrl.device,
549 			"rdma_resolve_addr wait failed (%d).\n", ret);
550 		goto out_destroy_cm_id;
551 	}
552 
553 	clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
554 
555 	return 0;
556 
557 out_destroy_cm_id:
558 	rdma_destroy_id(queue->cm_id);
559 	return ret;
560 }
561 
562 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
563 {
564 	rdma_disconnect(queue->cm_id);
565 	ib_drain_qp(queue->qp);
566 }
567 
568 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
569 {
570 	nvme_rdma_destroy_queue_ib(queue);
571 	rdma_destroy_id(queue->cm_id);
572 }
573 
574 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
575 {
576 	if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
577 		return;
578 	nvme_rdma_stop_queue(queue);
579 	nvme_rdma_free_queue(queue);
580 }
581 
582 static void nvme_rdma_free_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_and_free_queue(&ctrl->queues[i]);
588 }
589 
590 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
591 {
592 	int i, ret = 0;
593 
594 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
595 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
596 		if (ret) {
597 			dev_info(ctrl->ctrl.device,
598 				"failed to connect i/o queue: %d\n", ret);
599 			goto out_free_queues;
600 		}
601 		set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
602 	}
603 
604 	return 0;
605 
606 out_free_queues:
607 	nvme_rdma_free_io_queues(ctrl);
608 	return ret;
609 }
610 
611 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
612 {
613 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
614 	unsigned int nr_io_queues;
615 	int i, ret;
616 
617 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
618 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
619 	if (ret)
620 		return ret;
621 
622 	ctrl->ctrl.queue_count = nr_io_queues + 1;
623 	if (ctrl->ctrl.queue_count < 2)
624 		return 0;
625 
626 	dev_info(ctrl->ctrl.device,
627 		"creating %d I/O queues.\n", nr_io_queues);
628 
629 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
630 		ret = nvme_rdma_init_queue(ctrl, i,
631 					   ctrl->ctrl.opts->queue_size);
632 		if (ret) {
633 			dev_info(ctrl->ctrl.device,
634 				"failed to initialize i/o queue: %d\n", ret);
635 			goto out_free_queues;
636 		}
637 	}
638 
639 	return 0;
640 
641 out_free_queues:
642 	for (i--; i >= 1; i--)
643 		nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
644 
645 	return ret;
646 }
647 
648 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
649 {
650 	nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
651 			sizeof(struct nvme_command), DMA_TO_DEVICE);
652 	nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
653 	blk_cleanup_queue(ctrl->ctrl.admin_q);
654 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
655 	nvme_rdma_dev_put(ctrl->device);
656 }
657 
658 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
659 {
660 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
661 
662 	if (list_empty(&ctrl->list))
663 		goto free_ctrl;
664 
665 	mutex_lock(&nvme_rdma_ctrl_mutex);
666 	list_del(&ctrl->list);
667 	mutex_unlock(&nvme_rdma_ctrl_mutex);
668 
669 	kfree(ctrl->queues);
670 	nvmf_free_options(nctrl->opts);
671 free_ctrl:
672 	kfree(ctrl);
673 }
674 
675 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
676 {
677 	/* If we are resetting/deleting then do nothing */
678 	if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
679 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
680 			ctrl->ctrl.state == NVME_CTRL_LIVE);
681 		return;
682 	}
683 
684 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
685 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
686 			ctrl->ctrl.opts->reconnect_delay);
687 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
688 				ctrl->ctrl.opts->reconnect_delay * HZ);
689 	} else {
690 		dev_info(ctrl->ctrl.device, "Removing controller...\n");
691 		queue_work(nvme_wq, &ctrl->delete_work);
692 	}
693 }
694 
695 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
696 {
697 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
698 			struct nvme_rdma_ctrl, reconnect_work);
699 	bool changed;
700 	int ret;
701 
702 	++ctrl->ctrl.nr_reconnects;
703 
704 	if (ctrl->ctrl.queue_count > 1) {
705 		nvme_rdma_free_io_queues(ctrl);
706 
707 		ret = blk_mq_reinit_tagset(&ctrl->tag_set);
708 		if (ret)
709 			goto requeue;
710 	}
711 
712 	nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
713 
714 	ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
715 	if (ret)
716 		goto requeue;
717 
718 	ret = nvme_rdma_init_queue(ctrl, 0, NVME_AQ_DEPTH);
719 	if (ret)
720 		goto requeue;
721 
722 	ret = nvmf_connect_admin_queue(&ctrl->ctrl);
723 	if (ret)
724 		goto requeue;
725 
726 	set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
727 
728 	ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
729 	if (ret)
730 		goto requeue;
731 
732 	if (ctrl->ctrl.queue_count > 1) {
733 		ret = nvme_rdma_init_io_queues(ctrl);
734 		if (ret)
735 			goto requeue;
736 
737 		ret = nvme_rdma_connect_io_queues(ctrl);
738 		if (ret)
739 			goto requeue;
740 
741 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
742 				ctrl->ctrl.queue_count - 1);
743 	}
744 
745 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
746 	WARN_ON_ONCE(!changed);
747 	ctrl->ctrl.nr_reconnects = 0;
748 
749 	nvme_start_ctrl(&ctrl->ctrl);
750 
751 	dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
752 
753 	return;
754 
755 requeue:
756 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
757 			ctrl->ctrl.nr_reconnects);
758 	nvme_rdma_reconnect_or_remove(ctrl);
759 }
760 
761 static void nvme_rdma_error_recovery_work(struct work_struct *work)
762 {
763 	struct nvme_rdma_ctrl *ctrl = container_of(work,
764 			struct nvme_rdma_ctrl, err_work);
765 	int i;
766 
767 	nvme_stop_ctrl(&ctrl->ctrl);
768 
769 	for (i = 0; i < ctrl->ctrl.queue_count; i++)
770 		clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
771 
772 	if (ctrl->ctrl.queue_count > 1)
773 		nvme_stop_queues(&ctrl->ctrl);
774 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
775 
776 	/* We must take care of fastfail/requeue all our inflight requests */
777 	if (ctrl->ctrl.queue_count > 1)
778 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
779 					nvme_cancel_request, &ctrl->ctrl);
780 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
781 				nvme_cancel_request, &ctrl->ctrl);
782 
783 	/*
784 	 * queues are not a live anymore, so restart the queues to fail fast
785 	 * new IO
786 	 */
787 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
788 	nvme_start_queues(&ctrl->ctrl);
789 
790 	nvme_rdma_reconnect_or_remove(ctrl);
791 }
792 
793 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
794 {
795 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
796 		return;
797 
798 	queue_work(nvme_wq, &ctrl->err_work);
799 }
800 
801 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
802 		const char *op)
803 {
804 	struct nvme_rdma_queue *queue = cq->cq_context;
805 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
806 
807 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
808 		dev_info(ctrl->ctrl.device,
809 			     "%s for CQE 0x%p failed with status %s (%d)\n",
810 			     op, wc->wr_cqe,
811 			     ib_wc_status_msg(wc->status), wc->status);
812 	nvme_rdma_error_recovery(ctrl);
813 }
814 
815 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
816 {
817 	if (unlikely(wc->status != IB_WC_SUCCESS))
818 		nvme_rdma_wr_error(cq, wc, "MEMREG");
819 }
820 
821 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
822 {
823 	if (unlikely(wc->status != IB_WC_SUCCESS))
824 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
825 }
826 
827 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
828 		struct nvme_rdma_request *req)
829 {
830 	struct ib_send_wr *bad_wr;
831 	struct ib_send_wr wr = {
832 		.opcode		    = IB_WR_LOCAL_INV,
833 		.next		    = NULL,
834 		.num_sge	    = 0,
835 		.send_flags	    = 0,
836 		.ex.invalidate_rkey = req->mr->rkey,
837 	};
838 
839 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
840 	wr.wr_cqe = &req->reg_cqe;
841 
842 	return ib_post_send(queue->qp, &wr, &bad_wr);
843 }
844 
845 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
846 		struct request *rq)
847 {
848 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
849 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
850 	struct nvme_rdma_device *dev = queue->device;
851 	struct ib_device *ibdev = dev->dev;
852 	int res;
853 
854 	if (!blk_rq_bytes(rq))
855 		return;
856 
857 	if (req->mr->need_inval) {
858 		res = nvme_rdma_inv_rkey(queue, req);
859 		if (res < 0) {
860 			dev_err(ctrl->ctrl.device,
861 				"Queueing INV WR for rkey %#x failed (%d)\n",
862 				req->mr->rkey, res);
863 			nvme_rdma_error_recovery(queue->ctrl);
864 		}
865 	}
866 
867 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
868 			req->nents, rq_data_dir(rq) ==
869 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
870 
871 	nvme_cleanup_cmd(rq);
872 	sg_free_table_chained(&req->sg_table, true);
873 }
874 
875 static int nvme_rdma_set_sg_null(struct nvme_command *c)
876 {
877 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
878 
879 	sg->addr = 0;
880 	put_unaligned_le24(0, sg->length);
881 	put_unaligned_le32(0, sg->key);
882 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
883 	return 0;
884 }
885 
886 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
887 		struct nvme_rdma_request *req, struct nvme_command *c)
888 {
889 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
890 
891 	req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
892 	req->sge[1].length = sg_dma_len(req->sg_table.sgl);
893 	req->sge[1].lkey = queue->device->pd->local_dma_lkey;
894 
895 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
896 	sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
897 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
898 
899 	req->inline_data = true;
900 	req->num_sge++;
901 	return 0;
902 }
903 
904 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
905 		struct nvme_rdma_request *req, struct nvme_command *c)
906 {
907 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
908 
909 	sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
910 	put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
911 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
912 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
913 	return 0;
914 }
915 
916 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
917 		struct nvme_rdma_request *req, struct nvme_command *c,
918 		int count)
919 {
920 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
921 	int nr;
922 
923 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
924 	if (nr < count) {
925 		if (nr < 0)
926 			return nr;
927 		return -EINVAL;
928 	}
929 
930 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
931 
932 	req->reg_cqe.done = nvme_rdma_memreg_done;
933 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
934 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
935 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
936 	req->reg_wr.wr.num_sge = 0;
937 	req->reg_wr.mr = req->mr;
938 	req->reg_wr.key = req->mr->rkey;
939 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
940 			     IB_ACCESS_REMOTE_READ |
941 			     IB_ACCESS_REMOTE_WRITE;
942 
943 	req->mr->need_inval = true;
944 
945 	sg->addr = cpu_to_le64(req->mr->iova);
946 	put_unaligned_le24(req->mr->length, sg->length);
947 	put_unaligned_le32(req->mr->rkey, sg->key);
948 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
949 			NVME_SGL_FMT_INVALIDATE;
950 
951 	return 0;
952 }
953 
954 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
955 		struct request *rq, struct nvme_command *c)
956 {
957 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
958 	struct nvme_rdma_device *dev = queue->device;
959 	struct ib_device *ibdev = dev->dev;
960 	int count, ret;
961 
962 	req->num_sge = 1;
963 	req->inline_data = false;
964 	req->mr->need_inval = false;
965 
966 	c->common.flags |= NVME_CMD_SGL_METABUF;
967 
968 	if (!blk_rq_bytes(rq))
969 		return nvme_rdma_set_sg_null(c);
970 
971 	req->sg_table.sgl = req->first_sgl;
972 	ret = sg_alloc_table_chained(&req->sg_table,
973 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
974 	if (ret)
975 		return -ENOMEM;
976 
977 	req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
978 
979 	count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
980 		    rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
981 	if (unlikely(count <= 0)) {
982 		sg_free_table_chained(&req->sg_table, true);
983 		return -EIO;
984 	}
985 
986 	if (count == 1) {
987 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
988 		    blk_rq_payload_bytes(rq) <=
989 				nvme_rdma_inline_data_size(queue))
990 			return nvme_rdma_map_sg_inline(queue, req, c);
991 
992 		if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
993 			return nvme_rdma_map_sg_single(queue, req, c);
994 	}
995 
996 	return nvme_rdma_map_sg_fr(queue, req, c, count);
997 }
998 
999 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1000 {
1001 	if (unlikely(wc->status != IB_WC_SUCCESS))
1002 		nvme_rdma_wr_error(cq, wc, "SEND");
1003 }
1004 
1005 /*
1006  * We want to signal completion at least every queue depth/2.  This returns the
1007  * largest power of two that is not above half of (queue size + 1) to optimize
1008  * (avoid divisions).
1009  */
1010 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1011 {
1012 	int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1013 
1014 	return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1015 }
1016 
1017 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1018 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1019 		struct ib_send_wr *first, bool flush)
1020 {
1021 	struct ib_send_wr wr, *bad_wr;
1022 	int ret;
1023 
1024 	sge->addr   = qe->dma;
1025 	sge->length = sizeof(struct nvme_command),
1026 	sge->lkey   = queue->device->pd->local_dma_lkey;
1027 
1028 	qe->cqe.done = nvme_rdma_send_done;
1029 
1030 	wr.next       = NULL;
1031 	wr.wr_cqe     = &qe->cqe;
1032 	wr.sg_list    = sge;
1033 	wr.num_sge    = num_sge;
1034 	wr.opcode     = IB_WR_SEND;
1035 	wr.send_flags = 0;
1036 
1037 	/*
1038 	 * Unsignalled send completions are another giant desaster in the
1039 	 * IB Verbs spec:  If we don't regularly post signalled sends
1040 	 * the send queue will fill up and only a QP reset will rescue us.
1041 	 * Would have been way to obvious to handle this in hardware or
1042 	 * at least the RDMA stack..
1043 	 *
1044 	 * Always signal the flushes. The magic request used for the flush
1045 	 * sequencer is not allocated in our driver's tagset and it's
1046 	 * triggered to be freed by blk_cleanup_queue(). So we need to
1047 	 * always mark it as signaled to ensure that the "wr_cqe", which is
1048 	 * embedded in request's payload, is not freed when __ib_process_cq()
1049 	 * calls wr_cqe->done().
1050 	 */
1051 	if (nvme_rdma_queue_sig_limit(queue) || flush)
1052 		wr.send_flags |= IB_SEND_SIGNALED;
1053 
1054 	if (first)
1055 		first->next = &wr;
1056 	else
1057 		first = &wr;
1058 
1059 	ret = ib_post_send(queue->qp, first, &bad_wr);
1060 	if (ret) {
1061 		dev_err(queue->ctrl->ctrl.device,
1062 			     "%s failed with error code %d\n", __func__, ret);
1063 	}
1064 	return ret;
1065 }
1066 
1067 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1068 		struct nvme_rdma_qe *qe)
1069 {
1070 	struct ib_recv_wr wr, *bad_wr;
1071 	struct ib_sge list;
1072 	int ret;
1073 
1074 	list.addr   = qe->dma;
1075 	list.length = sizeof(struct nvme_completion);
1076 	list.lkey   = queue->device->pd->local_dma_lkey;
1077 
1078 	qe->cqe.done = nvme_rdma_recv_done;
1079 
1080 	wr.next     = NULL;
1081 	wr.wr_cqe   = &qe->cqe;
1082 	wr.sg_list  = &list;
1083 	wr.num_sge  = 1;
1084 
1085 	ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1086 	if (ret) {
1087 		dev_err(queue->ctrl->ctrl.device,
1088 			"%s failed with error code %d\n", __func__, ret);
1089 	}
1090 	return ret;
1091 }
1092 
1093 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1094 {
1095 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1096 
1097 	if (queue_idx == 0)
1098 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1099 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1100 }
1101 
1102 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1103 {
1104 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1105 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1106 	struct ib_device *dev = queue->device->dev;
1107 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1108 	struct nvme_command *cmd = sqe->data;
1109 	struct ib_sge sge;
1110 	int ret;
1111 
1112 	if (WARN_ON_ONCE(aer_idx != 0))
1113 		return;
1114 
1115 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1116 
1117 	memset(cmd, 0, sizeof(*cmd));
1118 	cmd->common.opcode = nvme_admin_async_event;
1119 	cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1120 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1121 	nvme_rdma_set_sg_null(cmd);
1122 
1123 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1124 			DMA_TO_DEVICE);
1125 
1126 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1127 	WARN_ON_ONCE(ret);
1128 }
1129 
1130 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1131 		struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1132 {
1133 	struct request *rq;
1134 	struct nvme_rdma_request *req;
1135 	int ret = 0;
1136 
1137 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1138 	if (!rq) {
1139 		dev_err(queue->ctrl->ctrl.device,
1140 			"tag 0x%x on QP %#x not found\n",
1141 			cqe->command_id, queue->qp->qp_num);
1142 		nvme_rdma_error_recovery(queue->ctrl);
1143 		return ret;
1144 	}
1145 	req = blk_mq_rq_to_pdu(rq);
1146 
1147 	if (rq->tag == tag)
1148 		ret = 1;
1149 
1150 	if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1151 	    wc->ex.invalidate_rkey == req->mr->rkey)
1152 		req->mr->need_inval = false;
1153 
1154 	nvme_end_request(rq, cqe->status, cqe->result);
1155 	return ret;
1156 }
1157 
1158 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1159 {
1160 	struct nvme_rdma_qe *qe =
1161 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1162 	struct nvme_rdma_queue *queue = cq->cq_context;
1163 	struct ib_device *ibdev = queue->device->dev;
1164 	struct nvme_completion *cqe = qe->data;
1165 	const size_t len = sizeof(struct nvme_completion);
1166 	int ret = 0;
1167 
1168 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1169 		nvme_rdma_wr_error(cq, wc, "RECV");
1170 		return 0;
1171 	}
1172 
1173 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1174 	/*
1175 	 * AEN requests are special as they don't time out and can
1176 	 * survive any kind of queue freeze and often don't respond to
1177 	 * aborts.  We don't even bother to allocate a struct request
1178 	 * for them but rather special case them here.
1179 	 */
1180 	if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1181 			cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1182 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1183 				&cqe->result);
1184 	else
1185 		ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1186 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1187 
1188 	nvme_rdma_post_recv(queue, qe);
1189 	return ret;
1190 }
1191 
1192 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1193 {
1194 	__nvme_rdma_recv_done(cq, wc, -1);
1195 }
1196 
1197 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1198 {
1199 	int ret, i;
1200 
1201 	for (i = 0; i < queue->queue_size; i++) {
1202 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1203 		if (ret)
1204 			goto out_destroy_queue_ib;
1205 	}
1206 
1207 	return 0;
1208 
1209 out_destroy_queue_ib:
1210 	nvme_rdma_destroy_queue_ib(queue);
1211 	return ret;
1212 }
1213 
1214 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1215 		struct rdma_cm_event *ev)
1216 {
1217 	struct rdma_cm_id *cm_id = queue->cm_id;
1218 	int status = ev->status;
1219 	const char *rej_msg;
1220 	const struct nvme_rdma_cm_rej *rej_data;
1221 	u8 rej_data_len;
1222 
1223 	rej_msg = rdma_reject_msg(cm_id, status);
1224 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1225 
1226 	if (rej_data && rej_data_len >= sizeof(u16)) {
1227 		u16 sts = le16_to_cpu(rej_data->sts);
1228 
1229 		dev_err(queue->ctrl->ctrl.device,
1230 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1231 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1232 	} else {
1233 		dev_err(queue->ctrl->ctrl.device,
1234 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1235 	}
1236 
1237 	return -ECONNRESET;
1238 }
1239 
1240 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1241 {
1242 	int ret;
1243 
1244 	ret = nvme_rdma_create_queue_ib(queue);
1245 	if (ret)
1246 		return ret;
1247 
1248 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1249 	if (ret) {
1250 		dev_err(queue->ctrl->ctrl.device,
1251 			"rdma_resolve_route failed (%d).\n",
1252 			queue->cm_error);
1253 		goto out_destroy_queue;
1254 	}
1255 
1256 	return 0;
1257 
1258 out_destroy_queue:
1259 	nvme_rdma_destroy_queue_ib(queue);
1260 	return ret;
1261 }
1262 
1263 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1264 {
1265 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1266 	struct rdma_conn_param param = { };
1267 	struct nvme_rdma_cm_req priv = { };
1268 	int ret;
1269 
1270 	param.qp_num = queue->qp->qp_num;
1271 	param.flow_control = 1;
1272 
1273 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1274 	/* maximum retry count */
1275 	param.retry_count = 7;
1276 	param.rnr_retry_count = 7;
1277 	param.private_data = &priv;
1278 	param.private_data_len = sizeof(priv);
1279 
1280 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1281 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1282 	/*
1283 	 * set the admin queue depth to the minimum size
1284 	 * specified by the Fabrics standard.
1285 	 */
1286 	if (priv.qid == 0) {
1287 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1288 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1289 	} else {
1290 		/*
1291 		 * current interpretation of the fabrics spec
1292 		 * is at minimum you make hrqsize sqsize+1, or a
1293 		 * 1's based representation of sqsize.
1294 		 */
1295 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1296 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1297 	}
1298 
1299 	ret = rdma_connect(queue->cm_id, &param);
1300 	if (ret) {
1301 		dev_err(ctrl->ctrl.device,
1302 			"rdma_connect failed (%d).\n", ret);
1303 		goto out_destroy_queue_ib;
1304 	}
1305 
1306 	return 0;
1307 
1308 out_destroy_queue_ib:
1309 	nvme_rdma_destroy_queue_ib(queue);
1310 	return ret;
1311 }
1312 
1313 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1314 		struct rdma_cm_event *ev)
1315 {
1316 	struct nvme_rdma_queue *queue = cm_id->context;
1317 	int cm_error = 0;
1318 
1319 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1320 		rdma_event_msg(ev->event), ev->event,
1321 		ev->status, cm_id);
1322 
1323 	switch (ev->event) {
1324 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1325 		cm_error = nvme_rdma_addr_resolved(queue);
1326 		break;
1327 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1328 		cm_error = nvme_rdma_route_resolved(queue);
1329 		break;
1330 	case RDMA_CM_EVENT_ESTABLISHED:
1331 		queue->cm_error = nvme_rdma_conn_established(queue);
1332 		/* complete cm_done regardless of success/failure */
1333 		complete(&queue->cm_done);
1334 		return 0;
1335 	case RDMA_CM_EVENT_REJECTED:
1336 		nvme_rdma_destroy_queue_ib(queue);
1337 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1338 		break;
1339 	case RDMA_CM_EVENT_ROUTE_ERROR:
1340 	case RDMA_CM_EVENT_CONNECT_ERROR:
1341 	case RDMA_CM_EVENT_UNREACHABLE:
1342 		nvme_rdma_destroy_queue_ib(queue);
1343 	case RDMA_CM_EVENT_ADDR_ERROR:
1344 		dev_dbg(queue->ctrl->ctrl.device,
1345 			"CM error event %d\n", ev->event);
1346 		cm_error = -ECONNRESET;
1347 		break;
1348 	case RDMA_CM_EVENT_DISCONNECTED:
1349 	case RDMA_CM_EVENT_ADDR_CHANGE:
1350 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1351 		dev_dbg(queue->ctrl->ctrl.device,
1352 			"disconnect received - connection closed\n");
1353 		nvme_rdma_error_recovery(queue->ctrl);
1354 		break;
1355 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1356 		/* device removal is handled via the ib_client API */
1357 		break;
1358 	default:
1359 		dev_err(queue->ctrl->ctrl.device,
1360 			"Unexpected RDMA CM event (%d)\n", ev->event);
1361 		nvme_rdma_error_recovery(queue->ctrl);
1362 		break;
1363 	}
1364 
1365 	if (cm_error) {
1366 		queue->cm_error = cm_error;
1367 		complete(&queue->cm_done);
1368 	}
1369 
1370 	return 0;
1371 }
1372 
1373 static enum blk_eh_timer_return
1374 nvme_rdma_timeout(struct request *rq, bool reserved)
1375 {
1376 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1377 
1378 	/* queue error recovery */
1379 	nvme_rdma_error_recovery(req->queue->ctrl);
1380 
1381 	/* fail with DNR on cmd timeout */
1382 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1383 
1384 	return BLK_EH_HANDLED;
1385 }
1386 
1387 /*
1388  * We cannot accept any other command until the Connect command has completed.
1389  */
1390 static inline blk_status_t
1391 nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1392 {
1393 	if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1394 		struct nvme_command *cmd = nvme_req(rq)->cmd;
1395 
1396 		if (!blk_rq_is_passthrough(rq) ||
1397 		    cmd->common.opcode != nvme_fabrics_command ||
1398 		    cmd->fabrics.fctype != nvme_fabrics_type_connect) {
1399 			/*
1400 			 * reconnecting state means transport disruption, which
1401 			 * can take a long time and even might fail permanently,
1402 			 * so we can't let incoming I/O be requeued forever.
1403 			 * fail it fast to allow upper layers a chance to
1404 			 * failover.
1405 			 */
1406 			if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING)
1407 				return BLK_STS_IOERR;
1408 			return BLK_STS_RESOURCE; /* try again later */
1409 		}
1410 	}
1411 
1412 	return 0;
1413 }
1414 
1415 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1416 		const struct blk_mq_queue_data *bd)
1417 {
1418 	struct nvme_ns *ns = hctx->queue->queuedata;
1419 	struct nvme_rdma_queue *queue = hctx->driver_data;
1420 	struct request *rq = bd->rq;
1421 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1422 	struct nvme_rdma_qe *sqe = &req->sqe;
1423 	struct nvme_command *c = sqe->data;
1424 	bool flush = false;
1425 	struct ib_device *dev;
1426 	blk_status_t ret;
1427 	int err;
1428 
1429 	WARN_ON_ONCE(rq->tag < 0);
1430 
1431 	ret = nvme_rdma_queue_is_ready(queue, rq);
1432 	if (unlikely(ret))
1433 		return ret;
1434 
1435 	dev = queue->device->dev;
1436 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1437 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1438 
1439 	ret = nvme_setup_cmd(ns, rq, c);
1440 	if (ret)
1441 		return ret;
1442 
1443 	blk_mq_start_request(rq);
1444 
1445 	err = nvme_rdma_map_data(queue, rq, c);
1446 	if (err < 0) {
1447 		dev_err(queue->ctrl->ctrl.device,
1448 			     "Failed to map data (%d)\n", err);
1449 		nvme_cleanup_cmd(rq);
1450 		goto err;
1451 	}
1452 
1453 	ib_dma_sync_single_for_device(dev, sqe->dma,
1454 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1455 
1456 	if (req_op(rq) == REQ_OP_FLUSH)
1457 		flush = true;
1458 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1459 			req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1460 	if (err) {
1461 		nvme_rdma_unmap_data(queue, rq);
1462 		goto err;
1463 	}
1464 
1465 	return BLK_STS_OK;
1466 err:
1467 	if (err == -ENOMEM || err == -EAGAIN)
1468 		return BLK_STS_RESOURCE;
1469 	return BLK_STS_IOERR;
1470 }
1471 
1472 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1473 {
1474 	struct nvme_rdma_queue *queue = hctx->driver_data;
1475 	struct ib_cq *cq = queue->ib_cq;
1476 	struct ib_wc wc;
1477 	int found = 0;
1478 
1479 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1480 		struct ib_cqe *cqe = wc.wr_cqe;
1481 
1482 		if (cqe) {
1483 			if (cqe->done == nvme_rdma_recv_done)
1484 				found |= __nvme_rdma_recv_done(cq, &wc, tag);
1485 			else
1486 				cqe->done(cq, &wc);
1487 		}
1488 	}
1489 
1490 	return found;
1491 }
1492 
1493 static void nvme_rdma_complete_rq(struct request *rq)
1494 {
1495 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1496 
1497 	nvme_rdma_unmap_data(req->queue, rq);
1498 	nvme_complete_rq(rq);
1499 }
1500 
1501 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1502 	.queue_rq	= nvme_rdma_queue_rq,
1503 	.complete	= nvme_rdma_complete_rq,
1504 	.init_request	= nvme_rdma_init_request,
1505 	.exit_request	= nvme_rdma_exit_request,
1506 	.reinit_request	= nvme_rdma_reinit_request,
1507 	.init_hctx	= nvme_rdma_init_hctx,
1508 	.poll		= nvme_rdma_poll,
1509 	.timeout	= nvme_rdma_timeout,
1510 };
1511 
1512 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1513 	.queue_rq	= nvme_rdma_queue_rq,
1514 	.complete	= nvme_rdma_complete_rq,
1515 	.init_request	= nvme_rdma_init_request,
1516 	.exit_request	= nvme_rdma_exit_request,
1517 	.reinit_request	= nvme_rdma_reinit_request,
1518 	.init_hctx	= nvme_rdma_init_admin_hctx,
1519 	.timeout	= nvme_rdma_timeout,
1520 };
1521 
1522 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1523 {
1524 	int error;
1525 
1526 	error = nvme_rdma_init_queue(ctrl, 0, NVME_AQ_DEPTH);
1527 	if (error)
1528 		return error;
1529 
1530 	ctrl->device = ctrl->queues[0].device;
1531 
1532 	/*
1533 	 * We need a reference on the device as long as the tag_set is alive,
1534 	 * as the MRs in the request structures need a valid ib_device.
1535 	 */
1536 	error = -EINVAL;
1537 	if (!nvme_rdma_dev_get(ctrl->device))
1538 		goto out_free_queue;
1539 
1540 	ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1541 		ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1542 
1543 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1544 	ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1545 	ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1546 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1547 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1548 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1549 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
1550 	ctrl->admin_tag_set.driver_data = ctrl;
1551 	ctrl->admin_tag_set.nr_hw_queues = 1;
1552 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1553 
1554 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1555 	if (error)
1556 		goto out_put_dev;
1557 
1558 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1559 	if (IS_ERR(ctrl->ctrl.admin_q)) {
1560 		error = PTR_ERR(ctrl->ctrl.admin_q);
1561 		goto out_free_tagset;
1562 	}
1563 
1564 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
1565 	if (error)
1566 		goto out_cleanup_queue;
1567 
1568 	set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1569 
1570 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP,
1571 			&ctrl->ctrl.cap);
1572 	if (error) {
1573 		dev_err(ctrl->ctrl.device,
1574 			"prop_get NVME_REG_CAP failed\n");
1575 		goto out_cleanup_queue;
1576 	}
1577 
1578 	ctrl->ctrl.sqsize =
1579 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
1580 
1581 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1582 	if (error)
1583 		goto out_cleanup_queue;
1584 
1585 	ctrl->ctrl.max_hw_sectors =
1586 		(ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1587 
1588 	error = nvme_init_identify(&ctrl->ctrl);
1589 	if (error)
1590 		goto out_cleanup_queue;
1591 
1592 	error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1593 			&ctrl->async_event_sqe, sizeof(struct nvme_command),
1594 			DMA_TO_DEVICE);
1595 	if (error)
1596 		goto out_cleanup_queue;
1597 
1598 	return 0;
1599 
1600 out_cleanup_queue:
1601 	blk_cleanup_queue(ctrl->ctrl.admin_q);
1602 out_free_tagset:
1603 	/* disconnect and drain the queue before freeing the tagset */
1604 	nvme_rdma_stop_queue(&ctrl->queues[0]);
1605 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
1606 out_put_dev:
1607 	nvme_rdma_dev_put(ctrl->device);
1608 out_free_queue:
1609 	nvme_rdma_free_queue(&ctrl->queues[0]);
1610 	return error;
1611 }
1612 
1613 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1614 {
1615 	cancel_work_sync(&ctrl->err_work);
1616 	cancel_delayed_work_sync(&ctrl->reconnect_work);
1617 
1618 	if (ctrl->ctrl.queue_count > 1) {
1619 		nvme_stop_queues(&ctrl->ctrl);
1620 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
1621 					nvme_cancel_request, &ctrl->ctrl);
1622 		nvme_rdma_free_io_queues(ctrl);
1623 	}
1624 
1625 	if (test_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags))
1626 		nvme_shutdown_ctrl(&ctrl->ctrl);
1627 
1628 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1629 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1630 				nvme_cancel_request, &ctrl->ctrl);
1631 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1632 	nvme_rdma_destroy_admin_queue(ctrl);
1633 }
1634 
1635 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1636 {
1637 	nvme_stop_ctrl(&ctrl->ctrl);
1638 	nvme_remove_namespaces(&ctrl->ctrl);
1639 	if (shutdown)
1640 		nvme_rdma_shutdown_ctrl(ctrl);
1641 
1642 	nvme_uninit_ctrl(&ctrl->ctrl);
1643 	if (ctrl->ctrl.tagset) {
1644 		blk_cleanup_queue(ctrl->ctrl.connect_q);
1645 		blk_mq_free_tag_set(&ctrl->tag_set);
1646 		nvme_rdma_dev_put(ctrl->device);
1647 	}
1648 
1649 	nvme_put_ctrl(&ctrl->ctrl);
1650 }
1651 
1652 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1653 {
1654 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1655 				struct nvme_rdma_ctrl, delete_work);
1656 
1657 	__nvme_rdma_remove_ctrl(ctrl, true);
1658 }
1659 
1660 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1661 {
1662 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1663 		return -EBUSY;
1664 
1665 	if (!queue_work(nvme_wq, &ctrl->delete_work))
1666 		return -EBUSY;
1667 
1668 	return 0;
1669 }
1670 
1671 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1672 {
1673 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1674 	int ret = 0;
1675 
1676 	/*
1677 	 * Keep a reference until all work is flushed since
1678 	 * __nvme_rdma_del_ctrl can free the ctrl mem
1679 	 */
1680 	if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1681 		return -EBUSY;
1682 	ret = __nvme_rdma_del_ctrl(ctrl);
1683 	if (!ret)
1684 		flush_work(&ctrl->delete_work);
1685 	nvme_put_ctrl(&ctrl->ctrl);
1686 	return ret;
1687 }
1688 
1689 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1690 {
1691 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1692 				struct nvme_rdma_ctrl, delete_work);
1693 
1694 	__nvme_rdma_remove_ctrl(ctrl, false);
1695 }
1696 
1697 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1698 {
1699 	struct nvme_rdma_ctrl *ctrl =
1700 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1701 	int ret;
1702 	bool changed;
1703 
1704 	nvme_stop_ctrl(&ctrl->ctrl);
1705 	nvme_rdma_shutdown_ctrl(ctrl);
1706 
1707 	ret = nvme_rdma_configure_admin_queue(ctrl);
1708 	if (ret) {
1709 		/* ctrl is already shutdown, just remove the ctrl */
1710 		INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1711 		goto del_dead_ctrl;
1712 	}
1713 
1714 	if (ctrl->ctrl.queue_count > 1) {
1715 		ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1716 		if (ret)
1717 			goto del_dead_ctrl;
1718 
1719 		ret = nvme_rdma_init_io_queues(ctrl);
1720 		if (ret)
1721 			goto del_dead_ctrl;
1722 
1723 		ret = nvme_rdma_connect_io_queues(ctrl);
1724 		if (ret)
1725 			goto del_dead_ctrl;
1726 
1727 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
1728 				ctrl->ctrl.queue_count - 1);
1729 	}
1730 
1731 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1732 	WARN_ON_ONCE(!changed);
1733 
1734 	nvme_start_ctrl(&ctrl->ctrl);
1735 
1736 	return;
1737 
1738 del_dead_ctrl:
1739 	/* Deleting this dead controller... */
1740 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1741 	WARN_ON(!queue_work(nvme_wq, &ctrl->delete_work));
1742 }
1743 
1744 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1745 	.name			= "rdma",
1746 	.module			= THIS_MODULE,
1747 	.flags			= NVME_F_FABRICS,
1748 	.reg_read32		= nvmf_reg_read32,
1749 	.reg_read64		= nvmf_reg_read64,
1750 	.reg_write32		= nvmf_reg_write32,
1751 	.free_ctrl		= nvme_rdma_free_ctrl,
1752 	.submit_async_event	= nvme_rdma_submit_async_event,
1753 	.delete_ctrl		= nvme_rdma_del_ctrl,
1754 	.get_address		= nvmf_get_address,
1755 };
1756 
1757 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1758 {
1759 	int ret;
1760 
1761 	ret = nvme_rdma_init_io_queues(ctrl);
1762 	if (ret)
1763 		return ret;
1764 
1765 	/*
1766 	 * We need a reference on the device as long as the tag_set is alive,
1767 	 * as the MRs in the request structures need a valid ib_device.
1768 	 */
1769 	ret = -EINVAL;
1770 	if (!nvme_rdma_dev_get(ctrl->device))
1771 		goto out_free_io_queues;
1772 
1773 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1774 	ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1775 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1776 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1777 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
1778 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1779 	ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1780 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
1781 	ctrl->tag_set.driver_data = ctrl;
1782 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
1783 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1784 
1785 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1786 	if (ret)
1787 		goto out_put_dev;
1788 	ctrl->ctrl.tagset = &ctrl->tag_set;
1789 
1790 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1791 	if (IS_ERR(ctrl->ctrl.connect_q)) {
1792 		ret = PTR_ERR(ctrl->ctrl.connect_q);
1793 		goto out_free_tag_set;
1794 	}
1795 
1796 	ret = nvme_rdma_connect_io_queues(ctrl);
1797 	if (ret)
1798 		goto out_cleanup_connect_q;
1799 
1800 	return 0;
1801 
1802 out_cleanup_connect_q:
1803 	blk_cleanup_queue(ctrl->ctrl.connect_q);
1804 out_free_tag_set:
1805 	blk_mq_free_tag_set(&ctrl->tag_set);
1806 out_put_dev:
1807 	nvme_rdma_dev_put(ctrl->device);
1808 out_free_io_queues:
1809 	nvme_rdma_free_io_queues(ctrl);
1810 	return ret;
1811 }
1812 
1813 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1814 		struct nvmf_ctrl_options *opts)
1815 {
1816 	struct nvme_rdma_ctrl *ctrl;
1817 	int ret;
1818 	bool changed;
1819 	char *port;
1820 
1821 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1822 	if (!ctrl)
1823 		return ERR_PTR(-ENOMEM);
1824 	ctrl->ctrl.opts = opts;
1825 	INIT_LIST_HEAD(&ctrl->list);
1826 
1827 	if (opts->mask & NVMF_OPT_TRSVCID)
1828 		port = opts->trsvcid;
1829 	else
1830 		port = __stringify(NVME_RDMA_IP_PORT);
1831 
1832 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1833 			opts->traddr, port, &ctrl->addr);
1834 	if (ret) {
1835 		pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1836 		goto out_free_ctrl;
1837 	}
1838 
1839 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1840 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1841 			opts->host_traddr, NULL, &ctrl->src_addr);
1842 		if (ret) {
1843 			pr_err("malformed src address passed: %s\n",
1844 			       opts->host_traddr);
1845 			goto out_free_ctrl;
1846 		}
1847 	}
1848 
1849 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1850 				0 /* no quirks, we're perfect! */);
1851 	if (ret)
1852 		goto out_free_ctrl;
1853 
1854 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
1855 			nvme_rdma_reconnect_ctrl_work);
1856 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1857 	INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1858 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1859 
1860 	ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1861 	ctrl->ctrl.sqsize = opts->queue_size - 1;
1862 	ctrl->ctrl.kato = opts->kato;
1863 
1864 	ret = -ENOMEM;
1865 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1866 				GFP_KERNEL);
1867 	if (!ctrl->queues)
1868 		goto out_uninit_ctrl;
1869 
1870 	ret = nvme_rdma_configure_admin_queue(ctrl);
1871 	if (ret)
1872 		goto out_kfree_queues;
1873 
1874 	/* sanity check icdoff */
1875 	if (ctrl->ctrl.icdoff) {
1876 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1877 		ret = -EINVAL;
1878 		goto out_remove_admin_queue;
1879 	}
1880 
1881 	/* sanity check keyed sgls */
1882 	if (!(ctrl->ctrl.sgls & (1 << 20))) {
1883 		dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1884 		ret = -EINVAL;
1885 		goto out_remove_admin_queue;
1886 	}
1887 
1888 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
1889 		/* warn if maxcmd is lower than queue_size */
1890 		dev_warn(ctrl->ctrl.device,
1891 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
1892 			opts->queue_size, ctrl->ctrl.maxcmd);
1893 		opts->queue_size = ctrl->ctrl.maxcmd;
1894 	}
1895 
1896 	if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1897 		/* warn if sqsize is lower than queue_size */
1898 		dev_warn(ctrl->ctrl.device,
1899 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
1900 			opts->queue_size, ctrl->ctrl.sqsize + 1);
1901 		opts->queue_size = ctrl->ctrl.sqsize + 1;
1902 	}
1903 
1904 	if (opts->nr_io_queues) {
1905 		ret = nvme_rdma_create_io_queues(ctrl);
1906 		if (ret)
1907 			goto out_remove_admin_queue;
1908 	}
1909 
1910 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1911 	WARN_ON_ONCE(!changed);
1912 
1913 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1914 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1915 
1916 	kref_get(&ctrl->ctrl.kref);
1917 
1918 	mutex_lock(&nvme_rdma_ctrl_mutex);
1919 	list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1920 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1921 
1922 	nvme_start_ctrl(&ctrl->ctrl);
1923 
1924 	return &ctrl->ctrl;
1925 
1926 out_remove_admin_queue:
1927 	nvme_rdma_destroy_admin_queue(ctrl);
1928 out_kfree_queues:
1929 	kfree(ctrl->queues);
1930 out_uninit_ctrl:
1931 	nvme_uninit_ctrl(&ctrl->ctrl);
1932 	nvme_put_ctrl(&ctrl->ctrl);
1933 	if (ret > 0)
1934 		ret = -EIO;
1935 	return ERR_PTR(ret);
1936 out_free_ctrl:
1937 	kfree(ctrl);
1938 	return ERR_PTR(ret);
1939 }
1940 
1941 static struct nvmf_transport_ops nvme_rdma_transport = {
1942 	.name		= "rdma",
1943 	.required_opts	= NVMF_OPT_TRADDR,
1944 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1945 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1946 	.create_ctrl	= nvme_rdma_create_ctrl,
1947 };
1948 
1949 static void nvme_rdma_add_one(struct ib_device *ib_device)
1950 {
1951 }
1952 
1953 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1954 {
1955 	struct nvme_rdma_ctrl *ctrl;
1956 
1957 	/* Delete all controllers using this device */
1958 	mutex_lock(&nvme_rdma_ctrl_mutex);
1959 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1960 		if (ctrl->device->dev != ib_device)
1961 			continue;
1962 		dev_info(ctrl->ctrl.device,
1963 			"Removing ctrl: NQN \"%s\", addr %pISp\n",
1964 			ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1965 		__nvme_rdma_del_ctrl(ctrl);
1966 	}
1967 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1968 
1969 	flush_workqueue(nvme_wq);
1970 }
1971 
1972 static struct ib_client nvme_rdma_ib_client = {
1973 	.name   = "nvme_rdma",
1974 	.add = nvme_rdma_add_one,
1975 	.remove = nvme_rdma_remove_one
1976 };
1977 
1978 static int __init nvme_rdma_init_module(void)
1979 {
1980 	int ret;
1981 
1982 	ret = ib_register_client(&nvme_rdma_ib_client);
1983 	if (ret)
1984 		return ret;
1985 
1986 	ret = nvmf_register_transport(&nvme_rdma_transport);
1987 	if (ret)
1988 		goto err_unreg_client;
1989 
1990 	return 0;
1991 
1992 err_unreg_client:
1993 	ib_unregister_client(&nvme_rdma_ib_client);
1994 	return ret;
1995 }
1996 
1997 static void __exit nvme_rdma_cleanup_module(void)
1998 {
1999 	nvmf_unregister_transport(&nvme_rdma_transport);
2000 	ib_unregister_client(&nvme_rdma_ib_client);
2001 }
2002 
2003 module_init(nvme_rdma_init_module);
2004 module_exit(nvme_rdma_cleanup_module);
2005 
2006 MODULE_LICENSE("GPL v2");
2007