xref: /openbmc/linux/drivers/nvme/host/rdma.c (revision 61bf3293)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics RDMA host code.
4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <rdma/mr_pool.h>
11 #include <linux/err.h>
12 #include <linux/string.h>
13 #include <linux/atomic.h>
14 #include <linux/blk-mq.h>
15 #include <linux/blk-mq-rdma.h>
16 #include <linux/types.h>
17 #include <linux/list.h>
18 #include <linux/mutex.h>
19 #include <linux/scatterlist.h>
20 #include <linux/nvme.h>
21 #include <asm/unaligned.h>
22 
23 #include <rdma/ib_verbs.h>
24 #include <rdma/rdma_cm.h>
25 #include <linux/nvme-rdma.h>
26 
27 #include "nvme.h"
28 #include "fabrics.h"
29 
30 
31 #define NVME_RDMA_CONNECT_TIMEOUT_MS	3000		/* 3 second */
32 
33 #define NVME_RDMA_MAX_SEGMENTS		256
34 
35 #define NVME_RDMA_MAX_INLINE_SEGMENTS	4
36 
37 #define NVME_RDMA_DATA_SGL_SIZE \
38 	(sizeof(struct scatterlist) * NVME_INLINE_SG_CNT)
39 #define NVME_RDMA_METADATA_SGL_SIZE \
40 	(sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT)
41 
42 struct nvme_rdma_device {
43 	struct ib_device	*dev;
44 	struct ib_pd		*pd;
45 	struct kref		ref;
46 	struct list_head	entry;
47 	unsigned int		num_inline_segments;
48 };
49 
50 struct nvme_rdma_qe {
51 	struct ib_cqe		cqe;
52 	void			*data;
53 	u64			dma;
54 };
55 
56 struct nvme_rdma_sgl {
57 	int			nents;
58 	struct sg_table		sg_table;
59 };
60 
61 struct nvme_rdma_queue;
62 struct nvme_rdma_request {
63 	struct nvme_request	req;
64 	struct ib_mr		*mr;
65 	struct nvme_rdma_qe	sqe;
66 	union nvme_result	result;
67 	__le16			status;
68 	refcount_t		ref;
69 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
70 	u32			num_sge;
71 	struct ib_reg_wr	reg_wr;
72 	struct ib_cqe		reg_cqe;
73 	struct nvme_rdma_queue  *queue;
74 	struct nvme_rdma_sgl	data_sgl;
75 	struct nvme_rdma_sgl	*metadata_sgl;
76 	bool			use_sig_mr;
77 };
78 
79 enum nvme_rdma_queue_flags {
80 	NVME_RDMA_Q_ALLOCATED		= 0,
81 	NVME_RDMA_Q_LIVE		= 1,
82 	NVME_RDMA_Q_TR_READY		= 2,
83 };
84 
85 struct nvme_rdma_queue {
86 	struct nvme_rdma_qe	*rsp_ring;
87 	int			queue_size;
88 	size_t			cmnd_capsule_len;
89 	struct nvme_rdma_ctrl	*ctrl;
90 	struct nvme_rdma_device	*device;
91 	struct ib_cq		*ib_cq;
92 	struct ib_qp		*qp;
93 
94 	unsigned long		flags;
95 	struct rdma_cm_id	*cm_id;
96 	int			cm_error;
97 	struct completion	cm_done;
98 	bool			pi_support;
99 };
100 
101 struct nvme_rdma_ctrl {
102 	/* read only in the hot path */
103 	struct nvme_rdma_queue	*queues;
104 
105 	/* other member variables */
106 	struct blk_mq_tag_set	tag_set;
107 	struct work_struct	err_work;
108 
109 	struct nvme_rdma_qe	async_event_sqe;
110 
111 	struct delayed_work	reconnect_work;
112 
113 	struct list_head	list;
114 
115 	struct blk_mq_tag_set	admin_tag_set;
116 	struct nvme_rdma_device	*device;
117 
118 	u32			max_fr_pages;
119 
120 	struct sockaddr_storage addr;
121 	struct sockaddr_storage src_addr;
122 
123 	struct nvme_ctrl	ctrl;
124 	bool			use_inline_data;
125 	u32			io_queues[HCTX_MAX_TYPES];
126 };
127 
128 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
129 {
130 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
131 }
132 
133 static LIST_HEAD(device_list);
134 static DEFINE_MUTEX(device_list_mutex);
135 
136 static LIST_HEAD(nvme_rdma_ctrl_list);
137 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
138 
139 /*
140  * Disabling this option makes small I/O goes faster, but is fundamentally
141  * unsafe.  With it turned off we will have to register a global rkey that
142  * allows read and write access to all physical memory.
143  */
144 static bool register_always = true;
145 module_param(register_always, bool, 0444);
146 MODULE_PARM_DESC(register_always,
147 	 "Use memory registration even for contiguous memory regions");
148 
149 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
150 		struct rdma_cm_event *event);
151 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
152 
153 static const struct blk_mq_ops nvme_rdma_mq_ops;
154 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
155 
156 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
157 {
158 	return queue - queue->ctrl->queues;
159 }
160 
161 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
162 {
163 	return nvme_rdma_queue_idx(queue) >
164 		queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
165 		queue->ctrl->io_queues[HCTX_TYPE_READ];
166 }
167 
168 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
169 {
170 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
171 }
172 
173 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
174 		size_t capsule_size, enum dma_data_direction dir)
175 {
176 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
177 	kfree(qe->data);
178 }
179 
180 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
181 		size_t capsule_size, enum dma_data_direction dir)
182 {
183 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
184 	if (!qe->data)
185 		return -ENOMEM;
186 
187 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
188 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
189 		kfree(qe->data);
190 		qe->data = NULL;
191 		return -ENOMEM;
192 	}
193 
194 	return 0;
195 }
196 
197 static void nvme_rdma_free_ring(struct ib_device *ibdev,
198 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
199 		size_t capsule_size, enum dma_data_direction dir)
200 {
201 	int i;
202 
203 	for (i = 0; i < ib_queue_size; i++)
204 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
205 	kfree(ring);
206 }
207 
208 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
209 		size_t ib_queue_size, size_t capsule_size,
210 		enum dma_data_direction dir)
211 {
212 	struct nvme_rdma_qe *ring;
213 	int i;
214 
215 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
216 	if (!ring)
217 		return NULL;
218 
219 	/*
220 	 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
221 	 * lifetime. It's safe, since any chage in the underlying RDMA device
222 	 * will issue error recovery and queue re-creation.
223 	 */
224 	for (i = 0; i < ib_queue_size; i++) {
225 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
226 			goto out_free_ring;
227 	}
228 
229 	return ring;
230 
231 out_free_ring:
232 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
233 	return NULL;
234 }
235 
236 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
237 {
238 	pr_debug("QP event %s (%d)\n",
239 		 ib_event_msg(event->event), event->event);
240 
241 }
242 
243 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
244 {
245 	int ret;
246 
247 	ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
248 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
249 	if (ret < 0)
250 		return ret;
251 	if (ret == 0)
252 		return -ETIMEDOUT;
253 	WARN_ON_ONCE(queue->cm_error > 0);
254 	return queue->cm_error;
255 }
256 
257 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
258 {
259 	struct nvme_rdma_device *dev = queue->device;
260 	struct ib_qp_init_attr init_attr;
261 	int ret;
262 
263 	memset(&init_attr, 0, sizeof(init_attr));
264 	init_attr.event_handler = nvme_rdma_qp_event;
265 	/* +1 for drain */
266 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
267 	/* +1 for drain */
268 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
269 	init_attr.cap.max_recv_sge = 1;
270 	init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
271 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
272 	init_attr.qp_type = IB_QPT_RC;
273 	init_attr.send_cq = queue->ib_cq;
274 	init_attr.recv_cq = queue->ib_cq;
275 	if (queue->pi_support)
276 		init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
277 
278 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
279 
280 	queue->qp = queue->cm_id->qp;
281 	return ret;
282 }
283 
284 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
285 		struct request *rq, unsigned int hctx_idx)
286 {
287 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
288 
289 	kfree(req->sqe.data);
290 }
291 
292 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
293 		struct request *rq, unsigned int hctx_idx,
294 		unsigned int numa_node)
295 {
296 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
297 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
298 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
299 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
300 
301 	nvme_req(rq)->ctrl = &ctrl->ctrl;
302 	req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
303 	if (!req->sqe.data)
304 		return -ENOMEM;
305 
306 	/* metadata nvme_rdma_sgl struct is located after command's data SGL */
307 	if (queue->pi_support)
308 		req->metadata_sgl = (void *)nvme_req(rq) +
309 			sizeof(struct nvme_rdma_request) +
310 			NVME_RDMA_DATA_SGL_SIZE;
311 
312 	req->queue = queue;
313 
314 	return 0;
315 }
316 
317 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
318 		unsigned int hctx_idx)
319 {
320 	struct nvme_rdma_ctrl *ctrl = data;
321 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
322 
323 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
324 
325 	hctx->driver_data = queue;
326 	return 0;
327 }
328 
329 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
330 		unsigned int hctx_idx)
331 {
332 	struct nvme_rdma_ctrl *ctrl = data;
333 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
334 
335 	BUG_ON(hctx_idx != 0);
336 
337 	hctx->driver_data = queue;
338 	return 0;
339 }
340 
341 static void nvme_rdma_free_dev(struct kref *ref)
342 {
343 	struct nvme_rdma_device *ndev =
344 		container_of(ref, struct nvme_rdma_device, ref);
345 
346 	mutex_lock(&device_list_mutex);
347 	list_del(&ndev->entry);
348 	mutex_unlock(&device_list_mutex);
349 
350 	ib_dealloc_pd(ndev->pd);
351 	kfree(ndev);
352 }
353 
354 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
355 {
356 	kref_put(&dev->ref, nvme_rdma_free_dev);
357 }
358 
359 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
360 {
361 	return kref_get_unless_zero(&dev->ref);
362 }
363 
364 static struct nvme_rdma_device *
365 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
366 {
367 	struct nvme_rdma_device *ndev;
368 
369 	mutex_lock(&device_list_mutex);
370 	list_for_each_entry(ndev, &device_list, entry) {
371 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
372 		    nvme_rdma_dev_get(ndev))
373 			goto out_unlock;
374 	}
375 
376 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
377 	if (!ndev)
378 		goto out_err;
379 
380 	ndev->dev = cm_id->device;
381 	kref_init(&ndev->ref);
382 
383 	ndev->pd = ib_alloc_pd(ndev->dev,
384 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
385 	if (IS_ERR(ndev->pd))
386 		goto out_free_dev;
387 
388 	if (!(ndev->dev->attrs.device_cap_flags &
389 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
390 		dev_err(&ndev->dev->dev,
391 			"Memory registrations not supported.\n");
392 		goto out_free_pd;
393 	}
394 
395 	ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
396 					ndev->dev->attrs.max_send_sge - 1);
397 	list_add(&ndev->entry, &device_list);
398 out_unlock:
399 	mutex_unlock(&device_list_mutex);
400 	return ndev;
401 
402 out_free_pd:
403 	ib_dealloc_pd(ndev->pd);
404 out_free_dev:
405 	kfree(ndev);
406 out_err:
407 	mutex_unlock(&device_list_mutex);
408 	return NULL;
409 }
410 
411 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
412 {
413 	struct nvme_rdma_device *dev;
414 	struct ib_device *ibdev;
415 
416 	if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
417 		return;
418 
419 	dev = queue->device;
420 	ibdev = dev->dev;
421 
422 	if (queue->pi_support)
423 		ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
424 	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
425 
426 	/*
427 	 * The cm_id object might have been destroyed during RDMA connection
428 	 * establishment error flow to avoid getting other cma events, thus
429 	 * the destruction of the QP shouldn't use rdma_cm API.
430 	 */
431 	ib_destroy_qp(queue->qp);
432 	ib_free_cq(queue->ib_cq);
433 
434 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
435 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
436 
437 	nvme_rdma_dev_put(dev);
438 }
439 
440 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
441 {
442 	u32 max_page_list_len;
443 
444 	if (pi_support)
445 		max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
446 	else
447 		max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
448 
449 	return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
450 }
451 
452 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
453 {
454 	struct ib_device *ibdev;
455 	const int send_wr_factor = 3;			/* MR, SEND, INV */
456 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
457 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
458 	enum ib_poll_context poll_ctx;
459 	int ret, pages_per_mr;
460 
461 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
462 	if (!queue->device) {
463 		dev_err(queue->cm_id->device->dev.parent,
464 			"no client data found!\n");
465 		return -ECONNREFUSED;
466 	}
467 	ibdev = queue->device->dev;
468 
469 	/*
470 	 * Spread I/O queues completion vectors according their queue index.
471 	 * Admin queues can always go on completion vector 0.
472 	 */
473 	comp_vector = idx == 0 ? idx : idx - 1;
474 
475 	/* Polling queues need direct cq polling context */
476 	if (nvme_rdma_poll_queue(queue))
477 		poll_ctx = IB_POLL_DIRECT;
478 	else
479 		poll_ctx = IB_POLL_SOFTIRQ;
480 
481 	/* +1 for ib_stop_cq */
482 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
483 				cq_factor * queue->queue_size + 1,
484 				comp_vector, poll_ctx);
485 	if (IS_ERR(queue->ib_cq)) {
486 		ret = PTR_ERR(queue->ib_cq);
487 		goto out_put_dev;
488 	}
489 
490 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
491 	if (ret)
492 		goto out_destroy_ib_cq;
493 
494 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
495 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
496 	if (!queue->rsp_ring) {
497 		ret = -ENOMEM;
498 		goto out_destroy_qp;
499 	}
500 
501 	/*
502 	 * Currently we don't use SG_GAPS MR's so if the first entry is
503 	 * misaligned we'll end up using two entries for a single data page,
504 	 * so one additional entry is required.
505 	 */
506 	pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
507 	ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
508 			      queue->queue_size,
509 			      IB_MR_TYPE_MEM_REG,
510 			      pages_per_mr, 0);
511 	if (ret) {
512 		dev_err(queue->ctrl->ctrl.device,
513 			"failed to initialize MR pool sized %d for QID %d\n",
514 			queue->queue_size, idx);
515 		goto out_destroy_ring;
516 	}
517 
518 	if (queue->pi_support) {
519 		ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
520 				      queue->queue_size, IB_MR_TYPE_INTEGRITY,
521 				      pages_per_mr, pages_per_mr);
522 		if (ret) {
523 			dev_err(queue->ctrl->ctrl.device,
524 				"failed to initialize PI MR pool sized %d for QID %d\n",
525 				queue->queue_size, idx);
526 			goto out_destroy_mr_pool;
527 		}
528 	}
529 
530 	set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
531 
532 	return 0;
533 
534 out_destroy_mr_pool:
535 	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
536 out_destroy_ring:
537 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
538 			    sizeof(struct nvme_completion), DMA_FROM_DEVICE);
539 out_destroy_qp:
540 	rdma_destroy_qp(queue->cm_id);
541 out_destroy_ib_cq:
542 	ib_free_cq(queue->ib_cq);
543 out_put_dev:
544 	nvme_rdma_dev_put(queue->device);
545 	return ret;
546 }
547 
548 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
549 		int idx, size_t queue_size)
550 {
551 	struct nvme_rdma_queue *queue;
552 	struct sockaddr *src_addr = NULL;
553 	int ret;
554 
555 	queue = &ctrl->queues[idx];
556 	queue->ctrl = ctrl;
557 	if (idx && ctrl->ctrl.max_integrity_segments)
558 		queue->pi_support = true;
559 	else
560 		queue->pi_support = false;
561 	init_completion(&queue->cm_done);
562 
563 	if (idx > 0)
564 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
565 	else
566 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
567 
568 	queue->queue_size = queue_size;
569 
570 	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
571 			RDMA_PS_TCP, IB_QPT_RC);
572 	if (IS_ERR(queue->cm_id)) {
573 		dev_info(ctrl->ctrl.device,
574 			"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
575 		return PTR_ERR(queue->cm_id);
576 	}
577 
578 	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
579 		src_addr = (struct sockaddr *)&ctrl->src_addr;
580 
581 	queue->cm_error = -ETIMEDOUT;
582 	ret = rdma_resolve_addr(queue->cm_id, src_addr,
583 			(struct sockaddr *)&ctrl->addr,
584 			NVME_RDMA_CONNECT_TIMEOUT_MS);
585 	if (ret) {
586 		dev_info(ctrl->ctrl.device,
587 			"rdma_resolve_addr failed (%d).\n", ret);
588 		goto out_destroy_cm_id;
589 	}
590 
591 	ret = nvme_rdma_wait_for_cm(queue);
592 	if (ret) {
593 		dev_info(ctrl->ctrl.device,
594 			"rdma connection establishment failed (%d)\n", ret);
595 		goto out_destroy_cm_id;
596 	}
597 
598 	set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
599 
600 	return 0;
601 
602 out_destroy_cm_id:
603 	rdma_destroy_id(queue->cm_id);
604 	nvme_rdma_destroy_queue_ib(queue);
605 	return ret;
606 }
607 
608 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
609 {
610 	rdma_disconnect(queue->cm_id);
611 	ib_drain_qp(queue->qp);
612 }
613 
614 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
615 {
616 	if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
617 		return;
618 	__nvme_rdma_stop_queue(queue);
619 }
620 
621 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
622 {
623 	if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
624 		return;
625 
626 	nvme_rdma_destroy_queue_ib(queue);
627 	rdma_destroy_id(queue->cm_id);
628 }
629 
630 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
631 {
632 	int i;
633 
634 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
635 		nvme_rdma_free_queue(&ctrl->queues[i]);
636 }
637 
638 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
639 {
640 	int i;
641 
642 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
643 		nvme_rdma_stop_queue(&ctrl->queues[i]);
644 }
645 
646 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
647 {
648 	struct nvme_rdma_queue *queue = &ctrl->queues[idx];
649 	bool poll = nvme_rdma_poll_queue(queue);
650 	int ret;
651 
652 	if (idx)
653 		ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
654 	else
655 		ret = nvmf_connect_admin_queue(&ctrl->ctrl);
656 
657 	if (!ret) {
658 		set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
659 	} else {
660 		if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
661 			__nvme_rdma_stop_queue(queue);
662 		dev_info(ctrl->ctrl.device,
663 			"failed to connect queue: %d ret=%d\n", idx, ret);
664 	}
665 	return ret;
666 }
667 
668 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
669 {
670 	int i, ret = 0;
671 
672 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
673 		ret = nvme_rdma_start_queue(ctrl, i);
674 		if (ret)
675 			goto out_stop_queues;
676 	}
677 
678 	return 0;
679 
680 out_stop_queues:
681 	for (i--; i >= 1; i--)
682 		nvme_rdma_stop_queue(&ctrl->queues[i]);
683 	return ret;
684 }
685 
686 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
687 {
688 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
689 	struct ib_device *ibdev = ctrl->device->dev;
690 	unsigned int nr_io_queues, nr_default_queues;
691 	unsigned int nr_read_queues, nr_poll_queues;
692 	int i, ret;
693 
694 	nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
695 				min(opts->nr_io_queues, num_online_cpus()));
696 	nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
697 				min(opts->nr_write_queues, num_online_cpus()));
698 	nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
699 	nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
700 
701 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
702 	if (ret)
703 		return ret;
704 
705 	ctrl->ctrl.queue_count = nr_io_queues + 1;
706 	if (ctrl->ctrl.queue_count < 2)
707 		return 0;
708 
709 	dev_info(ctrl->ctrl.device,
710 		"creating %d I/O queues.\n", nr_io_queues);
711 
712 	if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
713 		/*
714 		 * separate read/write queues
715 		 * hand out dedicated default queues only after we have
716 		 * sufficient read queues.
717 		 */
718 		ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
719 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
720 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
721 			min(nr_default_queues, nr_io_queues);
722 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
723 	} else {
724 		/*
725 		 * shared read/write queues
726 		 * either no write queues were requested, or we don't have
727 		 * sufficient queue count to have dedicated default queues.
728 		 */
729 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
730 			min(nr_read_queues, nr_io_queues);
731 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
732 	}
733 
734 	if (opts->nr_poll_queues && nr_io_queues) {
735 		/* map dedicated poll queues only if we have queues left */
736 		ctrl->io_queues[HCTX_TYPE_POLL] =
737 			min(nr_poll_queues, nr_io_queues);
738 	}
739 
740 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
741 		ret = nvme_rdma_alloc_queue(ctrl, i,
742 				ctrl->ctrl.sqsize + 1);
743 		if (ret)
744 			goto out_free_queues;
745 	}
746 
747 	return 0;
748 
749 out_free_queues:
750 	for (i--; i >= 1; i--)
751 		nvme_rdma_free_queue(&ctrl->queues[i]);
752 
753 	return ret;
754 }
755 
756 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
757 		bool admin)
758 {
759 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
760 	struct blk_mq_tag_set *set;
761 	int ret;
762 
763 	if (admin) {
764 		set = &ctrl->admin_tag_set;
765 		memset(set, 0, sizeof(*set));
766 		set->ops = &nvme_rdma_admin_mq_ops;
767 		set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
768 		set->reserved_tags = 2; /* connect + keep-alive */
769 		set->numa_node = nctrl->numa_node;
770 		set->cmd_size = sizeof(struct nvme_rdma_request) +
771 				NVME_RDMA_DATA_SGL_SIZE;
772 		set->driver_data = ctrl;
773 		set->nr_hw_queues = 1;
774 		set->timeout = ADMIN_TIMEOUT;
775 		set->flags = BLK_MQ_F_NO_SCHED;
776 	} else {
777 		set = &ctrl->tag_set;
778 		memset(set, 0, sizeof(*set));
779 		set->ops = &nvme_rdma_mq_ops;
780 		set->queue_depth = nctrl->sqsize + 1;
781 		set->reserved_tags = 1; /* fabric connect */
782 		set->numa_node = nctrl->numa_node;
783 		set->flags = BLK_MQ_F_SHOULD_MERGE;
784 		set->cmd_size = sizeof(struct nvme_rdma_request) +
785 				NVME_RDMA_DATA_SGL_SIZE;
786 		if (nctrl->max_integrity_segments)
787 			set->cmd_size += sizeof(struct nvme_rdma_sgl) +
788 					 NVME_RDMA_METADATA_SGL_SIZE;
789 		set->driver_data = ctrl;
790 		set->nr_hw_queues = nctrl->queue_count - 1;
791 		set->timeout = NVME_IO_TIMEOUT;
792 		set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
793 	}
794 
795 	ret = blk_mq_alloc_tag_set(set);
796 	if (ret)
797 		return ERR_PTR(ret);
798 
799 	return set;
800 }
801 
802 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
803 		bool remove)
804 {
805 	if (remove) {
806 		blk_cleanup_queue(ctrl->ctrl.admin_q);
807 		blk_cleanup_queue(ctrl->ctrl.fabrics_q);
808 		blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
809 	}
810 	if (ctrl->async_event_sqe.data) {
811 		nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
812 				sizeof(struct nvme_command), DMA_TO_DEVICE);
813 		ctrl->async_event_sqe.data = NULL;
814 	}
815 	nvme_rdma_free_queue(&ctrl->queues[0]);
816 }
817 
818 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
819 		bool new)
820 {
821 	bool pi_capable = false;
822 	int error;
823 
824 	error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
825 	if (error)
826 		return error;
827 
828 	ctrl->device = ctrl->queues[0].device;
829 	ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
830 
831 	/* T10-PI support */
832 	if (ctrl->device->dev->attrs.device_cap_flags &
833 	    IB_DEVICE_INTEGRITY_HANDOVER)
834 		pi_capable = true;
835 
836 	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
837 							pi_capable);
838 
839 	/*
840 	 * Bind the async event SQE DMA mapping to the admin queue lifetime.
841 	 * It's safe, since any chage in the underlying RDMA device will issue
842 	 * error recovery and queue re-creation.
843 	 */
844 	error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
845 			sizeof(struct nvme_command), DMA_TO_DEVICE);
846 	if (error)
847 		goto out_free_queue;
848 
849 	if (new) {
850 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
851 		if (IS_ERR(ctrl->ctrl.admin_tagset)) {
852 			error = PTR_ERR(ctrl->ctrl.admin_tagset);
853 			goto out_free_async_qe;
854 		}
855 
856 		ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
857 		if (IS_ERR(ctrl->ctrl.fabrics_q)) {
858 			error = PTR_ERR(ctrl->ctrl.fabrics_q);
859 			goto out_free_tagset;
860 		}
861 
862 		ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
863 		if (IS_ERR(ctrl->ctrl.admin_q)) {
864 			error = PTR_ERR(ctrl->ctrl.admin_q);
865 			goto out_cleanup_fabrics_q;
866 		}
867 	}
868 
869 	error = nvme_rdma_start_queue(ctrl, 0);
870 	if (error)
871 		goto out_cleanup_queue;
872 
873 	error = nvme_enable_ctrl(&ctrl->ctrl);
874 	if (error)
875 		goto out_stop_queue;
876 
877 	ctrl->ctrl.max_segments = ctrl->max_fr_pages;
878 	ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
879 	if (pi_capable)
880 		ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
881 	else
882 		ctrl->ctrl.max_integrity_segments = 0;
883 
884 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
885 
886 	error = nvme_init_identify(&ctrl->ctrl);
887 	if (error)
888 		goto out_stop_queue;
889 
890 	return 0;
891 
892 out_stop_queue:
893 	nvme_rdma_stop_queue(&ctrl->queues[0]);
894 out_cleanup_queue:
895 	if (new)
896 		blk_cleanup_queue(ctrl->ctrl.admin_q);
897 out_cleanup_fabrics_q:
898 	if (new)
899 		blk_cleanup_queue(ctrl->ctrl.fabrics_q);
900 out_free_tagset:
901 	if (new)
902 		blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
903 out_free_async_qe:
904 	if (ctrl->async_event_sqe.data) {
905 		nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
906 			sizeof(struct nvme_command), DMA_TO_DEVICE);
907 		ctrl->async_event_sqe.data = NULL;
908 	}
909 out_free_queue:
910 	nvme_rdma_free_queue(&ctrl->queues[0]);
911 	return error;
912 }
913 
914 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
915 		bool remove)
916 {
917 	if (remove) {
918 		blk_cleanup_queue(ctrl->ctrl.connect_q);
919 		blk_mq_free_tag_set(ctrl->ctrl.tagset);
920 	}
921 	nvme_rdma_free_io_queues(ctrl);
922 }
923 
924 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
925 {
926 	int ret;
927 
928 	ret = nvme_rdma_alloc_io_queues(ctrl);
929 	if (ret)
930 		return ret;
931 
932 	if (new) {
933 		ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
934 		if (IS_ERR(ctrl->ctrl.tagset)) {
935 			ret = PTR_ERR(ctrl->ctrl.tagset);
936 			goto out_free_io_queues;
937 		}
938 
939 		ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
940 		if (IS_ERR(ctrl->ctrl.connect_q)) {
941 			ret = PTR_ERR(ctrl->ctrl.connect_q);
942 			goto out_free_tag_set;
943 		}
944 	} else {
945 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
946 			ctrl->ctrl.queue_count - 1);
947 	}
948 
949 	ret = nvme_rdma_start_io_queues(ctrl);
950 	if (ret)
951 		goto out_cleanup_connect_q;
952 
953 	return 0;
954 
955 out_cleanup_connect_q:
956 	if (new)
957 		blk_cleanup_queue(ctrl->ctrl.connect_q);
958 out_free_tag_set:
959 	if (new)
960 		blk_mq_free_tag_set(ctrl->ctrl.tagset);
961 out_free_io_queues:
962 	nvme_rdma_free_io_queues(ctrl);
963 	return ret;
964 }
965 
966 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
967 		bool remove)
968 {
969 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
970 	nvme_rdma_stop_queue(&ctrl->queues[0]);
971 	if (ctrl->ctrl.admin_tagset) {
972 		blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
973 			nvme_cancel_request, &ctrl->ctrl);
974 		blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
975 	}
976 	if (remove)
977 		blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
978 	nvme_rdma_destroy_admin_queue(ctrl, remove);
979 }
980 
981 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
982 		bool remove)
983 {
984 	if (ctrl->ctrl.queue_count > 1) {
985 		nvme_stop_queues(&ctrl->ctrl);
986 		nvme_rdma_stop_io_queues(ctrl);
987 		if (ctrl->ctrl.tagset) {
988 			blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
989 				nvme_cancel_request, &ctrl->ctrl);
990 			blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
991 		}
992 		if (remove)
993 			nvme_start_queues(&ctrl->ctrl);
994 		nvme_rdma_destroy_io_queues(ctrl, remove);
995 	}
996 }
997 
998 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
999 {
1000 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1001 
1002 	if (list_empty(&ctrl->list))
1003 		goto free_ctrl;
1004 
1005 	mutex_lock(&nvme_rdma_ctrl_mutex);
1006 	list_del(&ctrl->list);
1007 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1008 
1009 	nvmf_free_options(nctrl->opts);
1010 free_ctrl:
1011 	kfree(ctrl->queues);
1012 	kfree(ctrl);
1013 }
1014 
1015 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1016 {
1017 	/* If we are resetting/deleting then do nothing */
1018 	if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1019 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1020 			ctrl->ctrl.state == NVME_CTRL_LIVE);
1021 		return;
1022 	}
1023 
1024 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
1025 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1026 			ctrl->ctrl.opts->reconnect_delay);
1027 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1028 				ctrl->ctrl.opts->reconnect_delay * HZ);
1029 	} else {
1030 		nvme_delete_ctrl(&ctrl->ctrl);
1031 	}
1032 }
1033 
1034 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1035 {
1036 	int ret = -EINVAL;
1037 	bool changed;
1038 
1039 	ret = nvme_rdma_configure_admin_queue(ctrl, new);
1040 	if (ret)
1041 		return ret;
1042 
1043 	if (ctrl->ctrl.icdoff) {
1044 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1045 		goto destroy_admin;
1046 	}
1047 
1048 	if (!(ctrl->ctrl.sgls & (1 << 2))) {
1049 		dev_err(ctrl->ctrl.device,
1050 			"Mandatory keyed sgls are not supported!\n");
1051 		goto destroy_admin;
1052 	}
1053 
1054 	if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1055 		dev_warn(ctrl->ctrl.device,
1056 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
1057 			ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1058 	}
1059 
1060 	if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1061 		dev_warn(ctrl->ctrl.device,
1062 			"sqsize %u > ctrl maxcmd %u, clamping down\n",
1063 			ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1064 		ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1065 	}
1066 
1067 	if (ctrl->ctrl.sgls & (1 << 20))
1068 		ctrl->use_inline_data = true;
1069 
1070 	if (ctrl->ctrl.queue_count > 1) {
1071 		ret = nvme_rdma_configure_io_queues(ctrl, new);
1072 		if (ret)
1073 			goto destroy_admin;
1074 	}
1075 
1076 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1077 	if (!changed) {
1078 		/*
1079 		 * state change failure is ok if we're in DELETING state,
1080 		 * unless we're during creation of a new controller to
1081 		 * avoid races with teardown flow.
1082 		 */
1083 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1084 		WARN_ON_ONCE(new);
1085 		ret = -EINVAL;
1086 		goto destroy_io;
1087 	}
1088 
1089 	nvme_start_ctrl(&ctrl->ctrl);
1090 	return 0;
1091 
1092 destroy_io:
1093 	if (ctrl->ctrl.queue_count > 1)
1094 		nvme_rdma_destroy_io_queues(ctrl, new);
1095 destroy_admin:
1096 	nvme_rdma_stop_queue(&ctrl->queues[0]);
1097 	nvme_rdma_destroy_admin_queue(ctrl, new);
1098 	return ret;
1099 }
1100 
1101 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1102 {
1103 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1104 			struct nvme_rdma_ctrl, reconnect_work);
1105 
1106 	++ctrl->ctrl.nr_reconnects;
1107 
1108 	if (nvme_rdma_setup_ctrl(ctrl, false))
1109 		goto requeue;
1110 
1111 	dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1112 			ctrl->ctrl.nr_reconnects);
1113 
1114 	ctrl->ctrl.nr_reconnects = 0;
1115 
1116 	return;
1117 
1118 requeue:
1119 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1120 			ctrl->ctrl.nr_reconnects);
1121 	nvme_rdma_reconnect_or_remove(ctrl);
1122 }
1123 
1124 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1125 {
1126 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1127 			struct nvme_rdma_ctrl, err_work);
1128 
1129 	nvme_stop_keep_alive(&ctrl->ctrl);
1130 	nvme_rdma_teardown_io_queues(ctrl, false);
1131 	nvme_start_queues(&ctrl->ctrl);
1132 	nvme_rdma_teardown_admin_queue(ctrl, false);
1133 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1134 
1135 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1136 		/* state change failure is ok if we're in DELETING state */
1137 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1138 		return;
1139 	}
1140 
1141 	nvme_rdma_reconnect_or_remove(ctrl);
1142 }
1143 
1144 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1145 {
1146 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1147 		return;
1148 
1149 	queue_work(nvme_reset_wq, &ctrl->err_work);
1150 }
1151 
1152 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1153 		const char *op)
1154 {
1155 	struct nvme_rdma_queue *queue = cq->cq_context;
1156 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1157 
1158 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1159 		dev_info(ctrl->ctrl.device,
1160 			     "%s for CQE 0x%p failed with status %s (%d)\n",
1161 			     op, wc->wr_cqe,
1162 			     ib_wc_status_msg(wc->status), wc->status);
1163 	nvme_rdma_error_recovery(ctrl);
1164 }
1165 
1166 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1167 {
1168 	if (unlikely(wc->status != IB_WC_SUCCESS))
1169 		nvme_rdma_wr_error(cq, wc, "MEMREG");
1170 }
1171 
1172 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1173 {
1174 	struct nvme_rdma_request *req =
1175 		container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1176 	struct request *rq = blk_mq_rq_from_pdu(req);
1177 
1178 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1179 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1180 		return;
1181 	}
1182 
1183 	if (refcount_dec_and_test(&req->ref))
1184 		nvme_end_request(rq, req->status, req->result);
1185 
1186 }
1187 
1188 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1189 		struct nvme_rdma_request *req)
1190 {
1191 	struct ib_send_wr wr = {
1192 		.opcode		    = IB_WR_LOCAL_INV,
1193 		.next		    = NULL,
1194 		.num_sge	    = 0,
1195 		.send_flags	    = IB_SEND_SIGNALED,
1196 		.ex.invalidate_rkey = req->mr->rkey,
1197 	};
1198 
1199 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1200 	wr.wr_cqe = &req->reg_cqe;
1201 
1202 	return ib_post_send(queue->qp, &wr, NULL);
1203 }
1204 
1205 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1206 		struct request *rq)
1207 {
1208 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1209 	struct nvme_rdma_device *dev = queue->device;
1210 	struct ib_device *ibdev = dev->dev;
1211 	struct list_head *pool = &queue->qp->rdma_mrs;
1212 
1213 	if (!blk_rq_nr_phys_segments(rq))
1214 		return;
1215 
1216 	if (blk_integrity_rq(rq)) {
1217 		ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1218 				req->metadata_sgl->nents, rq_dma_dir(rq));
1219 		sg_free_table_chained(&req->metadata_sgl->sg_table,
1220 				      NVME_INLINE_METADATA_SG_CNT);
1221 	}
1222 
1223 	if (req->use_sig_mr)
1224 		pool = &queue->qp->sig_mrs;
1225 
1226 	if (req->mr) {
1227 		ib_mr_pool_put(queue->qp, pool, req->mr);
1228 		req->mr = NULL;
1229 	}
1230 
1231 	ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1232 			rq_dma_dir(rq));
1233 	sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1234 }
1235 
1236 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1237 {
1238 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1239 
1240 	sg->addr = 0;
1241 	put_unaligned_le24(0, sg->length);
1242 	put_unaligned_le32(0, sg->key);
1243 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1244 	return 0;
1245 }
1246 
1247 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1248 		struct nvme_rdma_request *req, struct nvme_command *c,
1249 		int count)
1250 {
1251 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1252 	struct scatterlist *sgl = req->data_sgl.sg_table.sgl;
1253 	struct ib_sge *sge = &req->sge[1];
1254 	u32 len = 0;
1255 	int i;
1256 
1257 	for (i = 0; i < count; i++, sgl++, sge++) {
1258 		sge->addr = sg_dma_address(sgl);
1259 		sge->length = sg_dma_len(sgl);
1260 		sge->lkey = queue->device->pd->local_dma_lkey;
1261 		len += sge->length;
1262 	}
1263 
1264 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1265 	sg->length = cpu_to_le32(len);
1266 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1267 
1268 	req->num_sge += count;
1269 	return 0;
1270 }
1271 
1272 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1273 		struct nvme_rdma_request *req, struct nvme_command *c)
1274 {
1275 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1276 
1277 	sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1278 	put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1279 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1280 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1281 	return 0;
1282 }
1283 
1284 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1285 		struct nvme_rdma_request *req, struct nvme_command *c,
1286 		int count)
1287 {
1288 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1289 	int nr;
1290 
1291 	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1292 	if (WARN_ON_ONCE(!req->mr))
1293 		return -EAGAIN;
1294 
1295 	/*
1296 	 * Align the MR to a 4K page size to match the ctrl page size and
1297 	 * the block virtual boundary.
1298 	 */
1299 	nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1300 			  SZ_4K);
1301 	if (unlikely(nr < count)) {
1302 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1303 		req->mr = NULL;
1304 		if (nr < 0)
1305 			return nr;
1306 		return -EINVAL;
1307 	}
1308 
1309 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1310 
1311 	req->reg_cqe.done = nvme_rdma_memreg_done;
1312 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1313 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1314 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1315 	req->reg_wr.wr.num_sge = 0;
1316 	req->reg_wr.mr = req->mr;
1317 	req->reg_wr.key = req->mr->rkey;
1318 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1319 			     IB_ACCESS_REMOTE_READ |
1320 			     IB_ACCESS_REMOTE_WRITE;
1321 
1322 	sg->addr = cpu_to_le64(req->mr->iova);
1323 	put_unaligned_le24(req->mr->length, sg->length);
1324 	put_unaligned_le32(req->mr->rkey, sg->key);
1325 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1326 			NVME_SGL_FMT_INVALIDATE;
1327 
1328 	return 0;
1329 }
1330 
1331 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1332 		struct nvme_command *cmd, struct ib_sig_domain *domain,
1333 		u16 control, u8 pi_type)
1334 {
1335 	domain->sig_type = IB_SIG_TYPE_T10_DIF;
1336 	domain->sig.dif.bg_type = IB_T10DIF_CRC;
1337 	domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1338 	domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1339 	if (control & NVME_RW_PRINFO_PRCHK_REF)
1340 		domain->sig.dif.ref_remap = true;
1341 
1342 	domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1343 	domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1344 	domain->sig.dif.app_escape = true;
1345 	if (pi_type == NVME_NS_DPS_PI_TYPE3)
1346 		domain->sig.dif.ref_escape = true;
1347 }
1348 
1349 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1350 		struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1351 		u8 pi_type)
1352 {
1353 	u16 control = le16_to_cpu(cmd->rw.control);
1354 
1355 	memset(sig_attrs, 0, sizeof(*sig_attrs));
1356 	if (control & NVME_RW_PRINFO_PRACT) {
1357 		/* for WRITE_INSERT/READ_STRIP no memory domain */
1358 		sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1359 		nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1360 					 pi_type);
1361 		/* Clear the PRACT bit since HCA will generate/verify the PI */
1362 		control &= ~NVME_RW_PRINFO_PRACT;
1363 		cmd->rw.control = cpu_to_le16(control);
1364 	} else {
1365 		/* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1366 		nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1367 					 pi_type);
1368 		nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1369 					 pi_type);
1370 	}
1371 }
1372 
1373 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1374 {
1375 	*mask = 0;
1376 	if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1377 		*mask |= IB_SIG_CHECK_REFTAG;
1378 	if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1379 		*mask |= IB_SIG_CHECK_GUARD;
1380 }
1381 
1382 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1383 {
1384 	if (unlikely(wc->status != IB_WC_SUCCESS))
1385 		nvme_rdma_wr_error(cq, wc, "SIG");
1386 }
1387 
1388 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1389 		struct nvme_rdma_request *req, struct nvme_command *c,
1390 		int count, int pi_count)
1391 {
1392 	struct nvme_rdma_sgl *sgl = &req->data_sgl;
1393 	struct ib_reg_wr *wr = &req->reg_wr;
1394 	struct request *rq = blk_mq_rq_from_pdu(req);
1395 	struct nvme_ns *ns = rq->q->queuedata;
1396 	struct bio *bio = rq->bio;
1397 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1398 	int nr;
1399 
1400 	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1401 	if (WARN_ON_ONCE(!req->mr))
1402 		return -EAGAIN;
1403 
1404 	nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1405 			     req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1406 			     SZ_4K);
1407 	if (unlikely(nr))
1408 		goto mr_put;
1409 
1410 	nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_disk), c,
1411 				req->mr->sig_attrs, ns->pi_type);
1412 	nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1413 
1414 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1415 
1416 	req->reg_cqe.done = nvme_rdma_sig_done;
1417 	memset(wr, 0, sizeof(*wr));
1418 	wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1419 	wr->wr.wr_cqe = &req->reg_cqe;
1420 	wr->wr.num_sge = 0;
1421 	wr->wr.send_flags = 0;
1422 	wr->mr = req->mr;
1423 	wr->key = req->mr->rkey;
1424 	wr->access = IB_ACCESS_LOCAL_WRITE |
1425 		     IB_ACCESS_REMOTE_READ |
1426 		     IB_ACCESS_REMOTE_WRITE;
1427 
1428 	sg->addr = cpu_to_le64(req->mr->iova);
1429 	put_unaligned_le24(req->mr->length, sg->length);
1430 	put_unaligned_le32(req->mr->rkey, sg->key);
1431 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1432 
1433 	return 0;
1434 
1435 mr_put:
1436 	ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1437 	req->mr = NULL;
1438 	if (nr < 0)
1439 		return nr;
1440 	return -EINVAL;
1441 }
1442 
1443 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1444 		struct request *rq, struct nvme_command *c)
1445 {
1446 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1447 	struct nvme_rdma_device *dev = queue->device;
1448 	struct ib_device *ibdev = dev->dev;
1449 	int pi_count = 0;
1450 	int count, ret;
1451 
1452 	req->num_sge = 1;
1453 	refcount_set(&req->ref, 2); /* send and recv completions */
1454 
1455 	c->common.flags |= NVME_CMD_SGL_METABUF;
1456 
1457 	if (!blk_rq_nr_phys_segments(rq))
1458 		return nvme_rdma_set_sg_null(c);
1459 
1460 	req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1461 	ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1462 			blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1463 			NVME_INLINE_SG_CNT);
1464 	if (ret)
1465 		return -ENOMEM;
1466 
1467 	req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1468 					    req->data_sgl.sg_table.sgl);
1469 
1470 	count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1471 			      req->data_sgl.nents, rq_dma_dir(rq));
1472 	if (unlikely(count <= 0)) {
1473 		ret = -EIO;
1474 		goto out_free_table;
1475 	}
1476 
1477 	if (blk_integrity_rq(rq)) {
1478 		req->metadata_sgl->sg_table.sgl =
1479 			(struct scatterlist *)(req->metadata_sgl + 1);
1480 		ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1481 				blk_rq_count_integrity_sg(rq->q, rq->bio),
1482 				req->metadata_sgl->sg_table.sgl,
1483 				NVME_INLINE_METADATA_SG_CNT);
1484 		if (unlikely(ret)) {
1485 			ret = -ENOMEM;
1486 			goto out_unmap_sg;
1487 		}
1488 
1489 		req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1490 				rq->bio, req->metadata_sgl->sg_table.sgl);
1491 		pi_count = ib_dma_map_sg(ibdev,
1492 					 req->metadata_sgl->sg_table.sgl,
1493 					 req->metadata_sgl->nents,
1494 					 rq_dma_dir(rq));
1495 		if (unlikely(pi_count <= 0)) {
1496 			ret = -EIO;
1497 			goto out_free_pi_table;
1498 		}
1499 	}
1500 
1501 	if (req->use_sig_mr) {
1502 		ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1503 		goto out;
1504 	}
1505 
1506 	if (count <= dev->num_inline_segments) {
1507 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1508 		    queue->ctrl->use_inline_data &&
1509 		    blk_rq_payload_bytes(rq) <=
1510 				nvme_rdma_inline_data_size(queue)) {
1511 			ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1512 			goto out;
1513 		}
1514 
1515 		if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1516 			ret = nvme_rdma_map_sg_single(queue, req, c);
1517 			goto out;
1518 		}
1519 	}
1520 
1521 	ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1522 out:
1523 	if (unlikely(ret))
1524 		goto out_unmap_pi_sg;
1525 
1526 	return 0;
1527 
1528 out_unmap_pi_sg:
1529 	if (blk_integrity_rq(rq))
1530 		ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1531 				req->metadata_sgl->nents, rq_dma_dir(rq));
1532 out_free_pi_table:
1533 	if (blk_integrity_rq(rq))
1534 		sg_free_table_chained(&req->metadata_sgl->sg_table,
1535 				      NVME_INLINE_METADATA_SG_CNT);
1536 out_unmap_sg:
1537 	ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1538 			rq_dma_dir(rq));
1539 out_free_table:
1540 	sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1541 	return ret;
1542 }
1543 
1544 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1545 {
1546 	struct nvme_rdma_qe *qe =
1547 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1548 	struct nvme_rdma_request *req =
1549 		container_of(qe, struct nvme_rdma_request, sqe);
1550 	struct request *rq = blk_mq_rq_from_pdu(req);
1551 
1552 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1553 		nvme_rdma_wr_error(cq, wc, "SEND");
1554 		return;
1555 	}
1556 
1557 	if (refcount_dec_and_test(&req->ref))
1558 		nvme_end_request(rq, req->status, req->result);
1559 }
1560 
1561 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1562 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1563 		struct ib_send_wr *first)
1564 {
1565 	struct ib_send_wr wr;
1566 	int ret;
1567 
1568 	sge->addr   = qe->dma;
1569 	sge->length = sizeof(struct nvme_command);
1570 	sge->lkey   = queue->device->pd->local_dma_lkey;
1571 
1572 	wr.next       = NULL;
1573 	wr.wr_cqe     = &qe->cqe;
1574 	wr.sg_list    = sge;
1575 	wr.num_sge    = num_sge;
1576 	wr.opcode     = IB_WR_SEND;
1577 	wr.send_flags = IB_SEND_SIGNALED;
1578 
1579 	if (first)
1580 		first->next = &wr;
1581 	else
1582 		first = &wr;
1583 
1584 	ret = ib_post_send(queue->qp, first, NULL);
1585 	if (unlikely(ret)) {
1586 		dev_err(queue->ctrl->ctrl.device,
1587 			     "%s failed with error code %d\n", __func__, ret);
1588 	}
1589 	return ret;
1590 }
1591 
1592 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1593 		struct nvme_rdma_qe *qe)
1594 {
1595 	struct ib_recv_wr wr;
1596 	struct ib_sge list;
1597 	int ret;
1598 
1599 	list.addr   = qe->dma;
1600 	list.length = sizeof(struct nvme_completion);
1601 	list.lkey   = queue->device->pd->local_dma_lkey;
1602 
1603 	qe->cqe.done = nvme_rdma_recv_done;
1604 
1605 	wr.next     = NULL;
1606 	wr.wr_cqe   = &qe->cqe;
1607 	wr.sg_list  = &list;
1608 	wr.num_sge  = 1;
1609 
1610 	ret = ib_post_recv(queue->qp, &wr, NULL);
1611 	if (unlikely(ret)) {
1612 		dev_err(queue->ctrl->ctrl.device,
1613 			"%s failed with error code %d\n", __func__, ret);
1614 	}
1615 	return ret;
1616 }
1617 
1618 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1619 {
1620 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1621 
1622 	if (queue_idx == 0)
1623 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1624 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1625 }
1626 
1627 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1628 {
1629 	if (unlikely(wc->status != IB_WC_SUCCESS))
1630 		nvme_rdma_wr_error(cq, wc, "ASYNC");
1631 }
1632 
1633 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1634 {
1635 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1636 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1637 	struct ib_device *dev = queue->device->dev;
1638 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1639 	struct nvme_command *cmd = sqe->data;
1640 	struct ib_sge sge;
1641 	int ret;
1642 
1643 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1644 
1645 	memset(cmd, 0, sizeof(*cmd));
1646 	cmd->common.opcode = nvme_admin_async_event;
1647 	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1648 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1649 	nvme_rdma_set_sg_null(cmd);
1650 
1651 	sqe->cqe.done = nvme_rdma_async_done;
1652 
1653 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1654 			DMA_TO_DEVICE);
1655 
1656 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1657 	WARN_ON_ONCE(ret);
1658 }
1659 
1660 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1661 		struct nvme_completion *cqe, struct ib_wc *wc)
1662 {
1663 	struct request *rq;
1664 	struct nvme_rdma_request *req;
1665 
1666 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1667 	if (!rq) {
1668 		dev_err(queue->ctrl->ctrl.device,
1669 			"tag 0x%x on QP %#x not found\n",
1670 			cqe->command_id, queue->qp->qp_num);
1671 		nvme_rdma_error_recovery(queue->ctrl);
1672 		return;
1673 	}
1674 	req = blk_mq_rq_to_pdu(rq);
1675 
1676 	req->status = cqe->status;
1677 	req->result = cqe->result;
1678 
1679 	if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1680 		if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1681 			dev_err(queue->ctrl->ctrl.device,
1682 				"Bogus remote invalidation for rkey %#x\n",
1683 				req->mr->rkey);
1684 			nvme_rdma_error_recovery(queue->ctrl);
1685 		}
1686 	} else if (req->mr) {
1687 		int ret;
1688 
1689 		ret = nvme_rdma_inv_rkey(queue, req);
1690 		if (unlikely(ret < 0)) {
1691 			dev_err(queue->ctrl->ctrl.device,
1692 				"Queueing INV WR for rkey %#x failed (%d)\n",
1693 				req->mr->rkey, ret);
1694 			nvme_rdma_error_recovery(queue->ctrl);
1695 		}
1696 		/* the local invalidation completion will end the request */
1697 		return;
1698 	}
1699 
1700 	if (refcount_dec_and_test(&req->ref))
1701 		nvme_end_request(rq, req->status, req->result);
1702 }
1703 
1704 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1705 {
1706 	struct nvme_rdma_qe *qe =
1707 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1708 	struct nvme_rdma_queue *queue = cq->cq_context;
1709 	struct ib_device *ibdev = queue->device->dev;
1710 	struct nvme_completion *cqe = qe->data;
1711 	const size_t len = sizeof(struct nvme_completion);
1712 
1713 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1714 		nvme_rdma_wr_error(cq, wc, "RECV");
1715 		return;
1716 	}
1717 
1718 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1719 	/*
1720 	 * AEN requests are special as they don't time out and can
1721 	 * survive any kind of queue freeze and often don't respond to
1722 	 * aborts.  We don't even bother to allocate a struct request
1723 	 * for them but rather special case them here.
1724 	 */
1725 	if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1726 				     cqe->command_id)))
1727 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1728 				&cqe->result);
1729 	else
1730 		nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1731 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1732 
1733 	nvme_rdma_post_recv(queue, qe);
1734 }
1735 
1736 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1737 {
1738 	int ret, i;
1739 
1740 	for (i = 0; i < queue->queue_size; i++) {
1741 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1742 		if (ret)
1743 			goto out_destroy_queue_ib;
1744 	}
1745 
1746 	return 0;
1747 
1748 out_destroy_queue_ib:
1749 	nvme_rdma_destroy_queue_ib(queue);
1750 	return ret;
1751 }
1752 
1753 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1754 		struct rdma_cm_event *ev)
1755 {
1756 	struct rdma_cm_id *cm_id = queue->cm_id;
1757 	int status = ev->status;
1758 	const char *rej_msg;
1759 	const struct nvme_rdma_cm_rej *rej_data;
1760 	u8 rej_data_len;
1761 
1762 	rej_msg = rdma_reject_msg(cm_id, status);
1763 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1764 
1765 	if (rej_data && rej_data_len >= sizeof(u16)) {
1766 		u16 sts = le16_to_cpu(rej_data->sts);
1767 
1768 		dev_err(queue->ctrl->ctrl.device,
1769 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1770 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1771 	} else {
1772 		dev_err(queue->ctrl->ctrl.device,
1773 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1774 	}
1775 
1776 	return -ECONNRESET;
1777 }
1778 
1779 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1780 {
1781 	struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1782 	int ret;
1783 
1784 	ret = nvme_rdma_create_queue_ib(queue);
1785 	if (ret)
1786 		return ret;
1787 
1788 	if (ctrl->opts->tos >= 0)
1789 		rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1790 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1791 	if (ret) {
1792 		dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1793 			queue->cm_error);
1794 		goto out_destroy_queue;
1795 	}
1796 
1797 	return 0;
1798 
1799 out_destroy_queue:
1800 	nvme_rdma_destroy_queue_ib(queue);
1801 	return ret;
1802 }
1803 
1804 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1805 {
1806 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1807 	struct rdma_conn_param param = { };
1808 	struct nvme_rdma_cm_req priv = { };
1809 	int ret;
1810 
1811 	param.qp_num = queue->qp->qp_num;
1812 	param.flow_control = 1;
1813 
1814 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1815 	/* maximum retry count */
1816 	param.retry_count = 7;
1817 	param.rnr_retry_count = 7;
1818 	param.private_data = &priv;
1819 	param.private_data_len = sizeof(priv);
1820 
1821 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1822 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1823 	/*
1824 	 * set the admin queue depth to the minimum size
1825 	 * specified by the Fabrics standard.
1826 	 */
1827 	if (priv.qid == 0) {
1828 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1829 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1830 	} else {
1831 		/*
1832 		 * current interpretation of the fabrics spec
1833 		 * is at minimum you make hrqsize sqsize+1, or a
1834 		 * 1's based representation of sqsize.
1835 		 */
1836 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1837 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1838 	}
1839 
1840 	ret = rdma_connect(queue->cm_id, &param);
1841 	if (ret) {
1842 		dev_err(ctrl->ctrl.device,
1843 			"rdma_connect failed (%d).\n", ret);
1844 		goto out_destroy_queue_ib;
1845 	}
1846 
1847 	return 0;
1848 
1849 out_destroy_queue_ib:
1850 	nvme_rdma_destroy_queue_ib(queue);
1851 	return ret;
1852 }
1853 
1854 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1855 		struct rdma_cm_event *ev)
1856 {
1857 	struct nvme_rdma_queue *queue = cm_id->context;
1858 	int cm_error = 0;
1859 
1860 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1861 		rdma_event_msg(ev->event), ev->event,
1862 		ev->status, cm_id);
1863 
1864 	switch (ev->event) {
1865 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1866 		cm_error = nvme_rdma_addr_resolved(queue);
1867 		break;
1868 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1869 		cm_error = nvme_rdma_route_resolved(queue);
1870 		break;
1871 	case RDMA_CM_EVENT_ESTABLISHED:
1872 		queue->cm_error = nvme_rdma_conn_established(queue);
1873 		/* complete cm_done regardless of success/failure */
1874 		complete(&queue->cm_done);
1875 		return 0;
1876 	case RDMA_CM_EVENT_REJECTED:
1877 		nvme_rdma_destroy_queue_ib(queue);
1878 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1879 		break;
1880 	case RDMA_CM_EVENT_ROUTE_ERROR:
1881 	case RDMA_CM_EVENT_CONNECT_ERROR:
1882 	case RDMA_CM_EVENT_UNREACHABLE:
1883 		nvme_rdma_destroy_queue_ib(queue);
1884 		/* fall through */
1885 	case RDMA_CM_EVENT_ADDR_ERROR:
1886 		dev_dbg(queue->ctrl->ctrl.device,
1887 			"CM error event %d\n", ev->event);
1888 		cm_error = -ECONNRESET;
1889 		break;
1890 	case RDMA_CM_EVENT_DISCONNECTED:
1891 	case RDMA_CM_EVENT_ADDR_CHANGE:
1892 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1893 		dev_dbg(queue->ctrl->ctrl.device,
1894 			"disconnect received - connection closed\n");
1895 		nvme_rdma_error_recovery(queue->ctrl);
1896 		break;
1897 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1898 		/* device removal is handled via the ib_client API */
1899 		break;
1900 	default:
1901 		dev_err(queue->ctrl->ctrl.device,
1902 			"Unexpected RDMA CM event (%d)\n", ev->event);
1903 		nvme_rdma_error_recovery(queue->ctrl);
1904 		break;
1905 	}
1906 
1907 	if (cm_error) {
1908 		queue->cm_error = cm_error;
1909 		complete(&queue->cm_done);
1910 	}
1911 
1912 	return 0;
1913 }
1914 
1915 static enum blk_eh_timer_return
1916 nvme_rdma_timeout(struct request *rq, bool reserved)
1917 {
1918 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1919 	struct nvme_rdma_queue *queue = req->queue;
1920 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1921 
1922 	dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1923 		 rq->tag, nvme_rdma_queue_idx(queue));
1924 
1925 	/*
1926 	 * Restart the timer if a controller reset is already scheduled. Any
1927 	 * timed out commands would be handled before entering the connecting
1928 	 * state.
1929 	 */
1930 	if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
1931 		return BLK_EH_RESET_TIMER;
1932 
1933 	if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1934 		/*
1935 		 * Teardown immediately if controller times out while starting
1936 		 * or we are already started error recovery. all outstanding
1937 		 * requests are completed on shutdown, so we return BLK_EH_DONE.
1938 		 */
1939 		flush_work(&ctrl->err_work);
1940 		nvme_rdma_teardown_io_queues(ctrl, false);
1941 		nvme_rdma_teardown_admin_queue(ctrl, false);
1942 		return BLK_EH_DONE;
1943 	}
1944 
1945 	dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1946 	nvme_rdma_error_recovery(ctrl);
1947 
1948 	return BLK_EH_RESET_TIMER;
1949 }
1950 
1951 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1952 		const struct blk_mq_queue_data *bd)
1953 {
1954 	struct nvme_ns *ns = hctx->queue->queuedata;
1955 	struct nvme_rdma_queue *queue = hctx->driver_data;
1956 	struct request *rq = bd->rq;
1957 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1958 	struct nvme_rdma_qe *sqe = &req->sqe;
1959 	struct nvme_command *c = sqe->data;
1960 	struct ib_device *dev;
1961 	bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1962 	blk_status_t ret;
1963 	int err;
1964 
1965 	WARN_ON_ONCE(rq->tag < 0);
1966 
1967 	if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1968 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1969 
1970 	dev = queue->device->dev;
1971 
1972 	req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
1973 					 sizeof(struct nvme_command),
1974 					 DMA_TO_DEVICE);
1975 	err = ib_dma_mapping_error(dev, req->sqe.dma);
1976 	if (unlikely(err))
1977 		return BLK_STS_RESOURCE;
1978 
1979 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1980 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1981 
1982 	ret = nvme_setup_cmd(ns, rq, c);
1983 	if (ret)
1984 		goto unmap_qe;
1985 
1986 	blk_mq_start_request(rq);
1987 
1988 	if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
1989 	    queue->pi_support &&
1990 	    (c->common.opcode == nvme_cmd_write ||
1991 	     c->common.opcode == nvme_cmd_read) &&
1992 	    nvme_ns_has_pi(ns))
1993 		req->use_sig_mr = true;
1994 	else
1995 		req->use_sig_mr = false;
1996 
1997 	err = nvme_rdma_map_data(queue, rq, c);
1998 	if (unlikely(err < 0)) {
1999 		dev_err(queue->ctrl->ctrl.device,
2000 			     "Failed to map data (%d)\n", err);
2001 		goto err;
2002 	}
2003 
2004 	sqe->cqe.done = nvme_rdma_send_done;
2005 
2006 	ib_dma_sync_single_for_device(dev, sqe->dma,
2007 			sizeof(struct nvme_command), DMA_TO_DEVICE);
2008 
2009 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2010 			req->mr ? &req->reg_wr.wr : NULL);
2011 	if (unlikely(err))
2012 		goto err_unmap;
2013 
2014 	return BLK_STS_OK;
2015 
2016 err_unmap:
2017 	nvme_rdma_unmap_data(queue, rq);
2018 err:
2019 	if (err == -ENOMEM || err == -EAGAIN)
2020 		ret = BLK_STS_RESOURCE;
2021 	else
2022 		ret = BLK_STS_IOERR;
2023 	nvme_cleanup_cmd(rq);
2024 unmap_qe:
2025 	ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2026 			    DMA_TO_DEVICE);
2027 	return ret;
2028 }
2029 
2030 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2031 {
2032 	struct nvme_rdma_queue *queue = hctx->driver_data;
2033 
2034 	return ib_process_cq_direct(queue->ib_cq, -1);
2035 }
2036 
2037 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2038 {
2039 	struct request *rq = blk_mq_rq_from_pdu(req);
2040 	struct ib_mr_status mr_status;
2041 	int ret;
2042 
2043 	ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2044 	if (ret) {
2045 		pr_err("ib_check_mr_status failed, ret %d\n", ret);
2046 		nvme_req(rq)->status = NVME_SC_INVALID_PI;
2047 		return;
2048 	}
2049 
2050 	if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2051 		switch (mr_status.sig_err.err_type) {
2052 		case IB_SIG_BAD_GUARD:
2053 			nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2054 			break;
2055 		case IB_SIG_BAD_REFTAG:
2056 			nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2057 			break;
2058 		case IB_SIG_BAD_APPTAG:
2059 			nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2060 			break;
2061 		}
2062 		pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2063 		       mr_status.sig_err.err_type, mr_status.sig_err.expected,
2064 		       mr_status.sig_err.actual);
2065 	}
2066 }
2067 
2068 static void nvme_rdma_complete_rq(struct request *rq)
2069 {
2070 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2071 	struct nvme_rdma_queue *queue = req->queue;
2072 	struct ib_device *ibdev = queue->device->dev;
2073 
2074 	if (req->use_sig_mr)
2075 		nvme_rdma_check_pi_status(req);
2076 
2077 	nvme_rdma_unmap_data(queue, rq);
2078 	ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2079 			    DMA_TO_DEVICE);
2080 	nvme_complete_rq(rq);
2081 }
2082 
2083 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2084 {
2085 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
2086 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2087 
2088 	if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2089 		/* separate read/write queues */
2090 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
2091 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2092 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2093 		set->map[HCTX_TYPE_READ].nr_queues =
2094 			ctrl->io_queues[HCTX_TYPE_READ];
2095 		set->map[HCTX_TYPE_READ].queue_offset =
2096 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2097 	} else {
2098 		/* shared read/write queues */
2099 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
2100 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2101 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2102 		set->map[HCTX_TYPE_READ].nr_queues =
2103 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
2104 		set->map[HCTX_TYPE_READ].queue_offset = 0;
2105 	}
2106 	blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2107 			ctrl->device->dev, 0);
2108 	blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2109 			ctrl->device->dev, 0);
2110 
2111 	if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2112 		/* map dedicated poll queues only if we have queues left */
2113 		set->map[HCTX_TYPE_POLL].nr_queues =
2114 				ctrl->io_queues[HCTX_TYPE_POLL];
2115 		set->map[HCTX_TYPE_POLL].queue_offset =
2116 			ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2117 			ctrl->io_queues[HCTX_TYPE_READ];
2118 		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2119 	}
2120 
2121 	dev_info(ctrl->ctrl.device,
2122 		"mapped %d/%d/%d default/read/poll queues.\n",
2123 		ctrl->io_queues[HCTX_TYPE_DEFAULT],
2124 		ctrl->io_queues[HCTX_TYPE_READ],
2125 		ctrl->io_queues[HCTX_TYPE_POLL]);
2126 
2127 	return 0;
2128 }
2129 
2130 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2131 	.queue_rq	= nvme_rdma_queue_rq,
2132 	.complete	= nvme_rdma_complete_rq,
2133 	.init_request	= nvme_rdma_init_request,
2134 	.exit_request	= nvme_rdma_exit_request,
2135 	.init_hctx	= nvme_rdma_init_hctx,
2136 	.timeout	= nvme_rdma_timeout,
2137 	.map_queues	= nvme_rdma_map_queues,
2138 	.poll		= nvme_rdma_poll,
2139 };
2140 
2141 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2142 	.queue_rq	= nvme_rdma_queue_rq,
2143 	.complete	= nvme_rdma_complete_rq,
2144 	.init_request	= nvme_rdma_init_request,
2145 	.exit_request	= nvme_rdma_exit_request,
2146 	.init_hctx	= nvme_rdma_init_admin_hctx,
2147 	.timeout	= nvme_rdma_timeout,
2148 };
2149 
2150 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2151 {
2152 	cancel_work_sync(&ctrl->err_work);
2153 	cancel_delayed_work_sync(&ctrl->reconnect_work);
2154 
2155 	nvme_rdma_teardown_io_queues(ctrl, shutdown);
2156 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2157 	if (shutdown)
2158 		nvme_shutdown_ctrl(&ctrl->ctrl);
2159 	else
2160 		nvme_disable_ctrl(&ctrl->ctrl);
2161 	nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2162 }
2163 
2164 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2165 {
2166 	nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2167 }
2168 
2169 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2170 {
2171 	struct nvme_rdma_ctrl *ctrl =
2172 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2173 
2174 	nvme_stop_ctrl(&ctrl->ctrl);
2175 	nvme_rdma_shutdown_ctrl(ctrl, false);
2176 
2177 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2178 		/* state change failure should never happen */
2179 		WARN_ON_ONCE(1);
2180 		return;
2181 	}
2182 
2183 	if (nvme_rdma_setup_ctrl(ctrl, false))
2184 		goto out_fail;
2185 
2186 	return;
2187 
2188 out_fail:
2189 	++ctrl->ctrl.nr_reconnects;
2190 	nvme_rdma_reconnect_or_remove(ctrl);
2191 }
2192 
2193 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2194 	.name			= "rdma",
2195 	.module			= THIS_MODULE,
2196 	.flags			= NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2197 	.reg_read32		= nvmf_reg_read32,
2198 	.reg_read64		= nvmf_reg_read64,
2199 	.reg_write32		= nvmf_reg_write32,
2200 	.free_ctrl		= nvme_rdma_free_ctrl,
2201 	.submit_async_event	= nvme_rdma_submit_async_event,
2202 	.delete_ctrl		= nvme_rdma_delete_ctrl,
2203 	.get_address		= nvmf_get_address,
2204 };
2205 
2206 /*
2207  * Fails a connection request if it matches an existing controller
2208  * (association) with the same tuple:
2209  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2210  *
2211  * if local address is not specified in the request, it will match an
2212  * existing controller with all the other parameters the same and no
2213  * local port address specified as well.
2214  *
2215  * The ports don't need to be compared as they are intrinsically
2216  * already matched by the port pointers supplied.
2217  */
2218 static bool
2219 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2220 {
2221 	struct nvme_rdma_ctrl *ctrl;
2222 	bool found = false;
2223 
2224 	mutex_lock(&nvme_rdma_ctrl_mutex);
2225 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2226 		found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2227 		if (found)
2228 			break;
2229 	}
2230 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2231 
2232 	return found;
2233 }
2234 
2235 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2236 		struct nvmf_ctrl_options *opts)
2237 {
2238 	struct nvme_rdma_ctrl *ctrl;
2239 	int ret;
2240 	bool changed;
2241 
2242 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2243 	if (!ctrl)
2244 		return ERR_PTR(-ENOMEM);
2245 	ctrl->ctrl.opts = opts;
2246 	INIT_LIST_HEAD(&ctrl->list);
2247 
2248 	if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2249 		opts->trsvcid =
2250 			kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2251 		if (!opts->trsvcid) {
2252 			ret = -ENOMEM;
2253 			goto out_free_ctrl;
2254 		}
2255 		opts->mask |= NVMF_OPT_TRSVCID;
2256 	}
2257 
2258 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2259 			opts->traddr, opts->trsvcid, &ctrl->addr);
2260 	if (ret) {
2261 		pr_err("malformed address passed: %s:%s\n",
2262 			opts->traddr, opts->trsvcid);
2263 		goto out_free_ctrl;
2264 	}
2265 
2266 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2267 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2268 			opts->host_traddr, NULL, &ctrl->src_addr);
2269 		if (ret) {
2270 			pr_err("malformed src address passed: %s\n",
2271 			       opts->host_traddr);
2272 			goto out_free_ctrl;
2273 		}
2274 	}
2275 
2276 	if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2277 		ret = -EALREADY;
2278 		goto out_free_ctrl;
2279 	}
2280 
2281 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
2282 			nvme_rdma_reconnect_ctrl_work);
2283 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2284 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2285 
2286 	ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2287 				opts->nr_poll_queues + 1;
2288 	ctrl->ctrl.sqsize = opts->queue_size - 1;
2289 	ctrl->ctrl.kato = opts->kato;
2290 
2291 	ret = -ENOMEM;
2292 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2293 				GFP_KERNEL);
2294 	if (!ctrl->queues)
2295 		goto out_free_ctrl;
2296 
2297 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2298 				0 /* no quirks, we're perfect! */);
2299 	if (ret)
2300 		goto out_kfree_queues;
2301 
2302 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2303 	WARN_ON_ONCE(!changed);
2304 
2305 	ret = nvme_rdma_setup_ctrl(ctrl, true);
2306 	if (ret)
2307 		goto out_uninit_ctrl;
2308 
2309 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2310 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2311 
2312 	mutex_lock(&nvme_rdma_ctrl_mutex);
2313 	list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2314 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2315 
2316 	return &ctrl->ctrl;
2317 
2318 out_uninit_ctrl:
2319 	nvme_uninit_ctrl(&ctrl->ctrl);
2320 	nvme_put_ctrl(&ctrl->ctrl);
2321 	if (ret > 0)
2322 		ret = -EIO;
2323 	return ERR_PTR(ret);
2324 out_kfree_queues:
2325 	kfree(ctrl->queues);
2326 out_free_ctrl:
2327 	kfree(ctrl);
2328 	return ERR_PTR(ret);
2329 }
2330 
2331 static struct nvmf_transport_ops nvme_rdma_transport = {
2332 	.name		= "rdma",
2333 	.module		= THIS_MODULE,
2334 	.required_opts	= NVMF_OPT_TRADDR,
2335 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2336 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2337 			  NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2338 			  NVMF_OPT_TOS,
2339 	.create_ctrl	= nvme_rdma_create_ctrl,
2340 };
2341 
2342 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2343 {
2344 	struct nvme_rdma_ctrl *ctrl;
2345 	struct nvme_rdma_device *ndev;
2346 	bool found = false;
2347 
2348 	mutex_lock(&device_list_mutex);
2349 	list_for_each_entry(ndev, &device_list, entry) {
2350 		if (ndev->dev == ib_device) {
2351 			found = true;
2352 			break;
2353 		}
2354 	}
2355 	mutex_unlock(&device_list_mutex);
2356 
2357 	if (!found)
2358 		return;
2359 
2360 	/* Delete all controllers using this device */
2361 	mutex_lock(&nvme_rdma_ctrl_mutex);
2362 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2363 		if (ctrl->device->dev != ib_device)
2364 			continue;
2365 		nvme_delete_ctrl(&ctrl->ctrl);
2366 	}
2367 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2368 
2369 	flush_workqueue(nvme_delete_wq);
2370 }
2371 
2372 static struct ib_client nvme_rdma_ib_client = {
2373 	.name   = "nvme_rdma",
2374 	.remove = nvme_rdma_remove_one
2375 };
2376 
2377 static int __init nvme_rdma_init_module(void)
2378 {
2379 	int ret;
2380 
2381 	ret = ib_register_client(&nvme_rdma_ib_client);
2382 	if (ret)
2383 		return ret;
2384 
2385 	ret = nvmf_register_transport(&nvme_rdma_transport);
2386 	if (ret)
2387 		goto err_unreg_client;
2388 
2389 	return 0;
2390 
2391 err_unreg_client:
2392 	ib_unregister_client(&nvme_rdma_ib_client);
2393 	return ret;
2394 }
2395 
2396 static void __exit nvme_rdma_cleanup_module(void)
2397 {
2398 	struct nvme_rdma_ctrl *ctrl;
2399 
2400 	nvmf_unregister_transport(&nvme_rdma_transport);
2401 	ib_unregister_client(&nvme_rdma_ib_client);
2402 
2403 	mutex_lock(&nvme_rdma_ctrl_mutex);
2404 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2405 		nvme_delete_ctrl(&ctrl->ctrl);
2406 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2407 	flush_workqueue(nvme_delete_wq);
2408 }
2409 
2410 module_init(nvme_rdma_init_module);
2411 module_exit(nvme_rdma_cleanup_module);
2412 
2413 MODULE_LICENSE("GPL v2");
2414