xref: /openbmc/linux/drivers/nvme/host/rdma.c (revision 8e774e02)
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 struct nvme_rdma_device {
38 	struct ib_device	*dev;
39 	struct ib_pd		*pd;
40 	struct kref		ref;
41 	struct list_head	entry;
42 	unsigned int		num_inline_segments;
43 };
44 
45 struct nvme_rdma_qe {
46 	struct ib_cqe		cqe;
47 	void			*data;
48 	u64			dma;
49 };
50 
51 struct nvme_rdma_queue;
52 struct nvme_rdma_request {
53 	struct nvme_request	req;
54 	struct ib_mr		*mr;
55 	struct nvme_rdma_qe	sqe;
56 	union nvme_result	result;
57 	__le16			status;
58 	refcount_t		ref;
59 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
60 	u32			num_sge;
61 	int			nents;
62 	struct ib_reg_wr	reg_wr;
63 	struct ib_cqe		reg_cqe;
64 	struct nvme_rdma_queue  *queue;
65 	struct sg_table		sg_table;
66 	struct scatterlist	first_sgl[];
67 };
68 
69 enum nvme_rdma_queue_flags {
70 	NVME_RDMA_Q_ALLOCATED		= 0,
71 	NVME_RDMA_Q_LIVE		= 1,
72 	NVME_RDMA_Q_TR_READY		= 2,
73 };
74 
75 struct nvme_rdma_queue {
76 	struct nvme_rdma_qe	*rsp_ring;
77 	int			queue_size;
78 	size_t			cmnd_capsule_len;
79 	struct nvme_rdma_ctrl	*ctrl;
80 	struct nvme_rdma_device	*device;
81 	struct ib_cq		*ib_cq;
82 	struct ib_qp		*qp;
83 
84 	unsigned long		flags;
85 	struct rdma_cm_id	*cm_id;
86 	int			cm_error;
87 	struct completion	cm_done;
88 };
89 
90 struct nvme_rdma_ctrl {
91 	/* read only in the hot path */
92 	struct nvme_rdma_queue	*queues;
93 
94 	/* other member variables */
95 	struct blk_mq_tag_set	tag_set;
96 	struct work_struct	err_work;
97 
98 	struct nvme_rdma_qe	async_event_sqe;
99 
100 	struct delayed_work	reconnect_work;
101 
102 	struct list_head	list;
103 
104 	struct blk_mq_tag_set	admin_tag_set;
105 	struct nvme_rdma_device	*device;
106 
107 	u32			max_fr_pages;
108 
109 	struct sockaddr_storage addr;
110 	struct sockaddr_storage src_addr;
111 
112 	struct nvme_ctrl	ctrl;
113 	bool			use_inline_data;
114 	u32			io_queues[HCTX_MAX_TYPES];
115 };
116 
117 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
118 {
119 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
120 }
121 
122 static LIST_HEAD(device_list);
123 static DEFINE_MUTEX(device_list_mutex);
124 
125 static LIST_HEAD(nvme_rdma_ctrl_list);
126 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
127 
128 /*
129  * Disabling this option makes small I/O goes faster, but is fundamentally
130  * unsafe.  With it turned off we will have to register a global rkey that
131  * allows read and write access to all physical memory.
132  */
133 static bool register_always = true;
134 module_param(register_always, bool, 0444);
135 MODULE_PARM_DESC(register_always,
136 	 "Use memory registration even for contiguous memory regions");
137 
138 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
139 		struct rdma_cm_event *event);
140 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
141 
142 static const struct blk_mq_ops nvme_rdma_mq_ops;
143 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
144 
145 /* XXX: really should move to a generic header sooner or later.. */
146 static inline void put_unaligned_le24(u32 val, u8 *p)
147 {
148 	*p++ = val;
149 	*p++ = val >> 8;
150 	*p++ = val >> 16;
151 }
152 
153 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
154 {
155 	return queue - queue->ctrl->queues;
156 }
157 
158 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
159 {
160 	return nvme_rdma_queue_idx(queue) >
161 		queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
162 		queue->ctrl->io_queues[HCTX_TYPE_READ];
163 }
164 
165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
166 {
167 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
168 }
169 
170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
171 		size_t capsule_size, enum dma_data_direction dir)
172 {
173 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
174 	kfree(qe->data);
175 }
176 
177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
178 		size_t capsule_size, enum dma_data_direction dir)
179 {
180 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
181 	if (!qe->data)
182 		return -ENOMEM;
183 
184 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
185 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
186 		kfree(qe->data);
187 		qe->data = NULL;
188 		return -ENOMEM;
189 	}
190 
191 	return 0;
192 }
193 
194 static void nvme_rdma_free_ring(struct ib_device *ibdev,
195 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
196 		size_t capsule_size, enum dma_data_direction dir)
197 {
198 	int i;
199 
200 	for (i = 0; i < ib_queue_size; i++)
201 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
202 	kfree(ring);
203 }
204 
205 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
206 		size_t ib_queue_size, size_t capsule_size,
207 		enum dma_data_direction dir)
208 {
209 	struct nvme_rdma_qe *ring;
210 	int i;
211 
212 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
213 	if (!ring)
214 		return NULL;
215 
216 	/*
217 	 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
218 	 * lifetime. It's safe, since any chage in the underlying RDMA device
219 	 * will issue error recovery and queue re-creation.
220 	 */
221 	for (i = 0; i < ib_queue_size; i++) {
222 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
223 			goto out_free_ring;
224 	}
225 
226 	return ring;
227 
228 out_free_ring:
229 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
230 	return NULL;
231 }
232 
233 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
234 {
235 	pr_debug("QP event %s (%d)\n",
236 		 ib_event_msg(event->event), event->event);
237 
238 }
239 
240 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
241 {
242 	int ret;
243 
244 	ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
245 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
246 	if (ret < 0)
247 		return ret;
248 	if (ret == 0)
249 		return -ETIMEDOUT;
250 	WARN_ON_ONCE(queue->cm_error > 0);
251 	return queue->cm_error;
252 }
253 
254 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
255 {
256 	struct nvme_rdma_device *dev = queue->device;
257 	struct ib_qp_init_attr init_attr;
258 	int ret;
259 
260 	memset(&init_attr, 0, sizeof(init_attr));
261 	init_attr.event_handler = nvme_rdma_qp_event;
262 	/* +1 for drain */
263 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
264 	/* +1 for drain */
265 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
266 	init_attr.cap.max_recv_sge = 1;
267 	init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
268 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
269 	init_attr.qp_type = IB_QPT_RC;
270 	init_attr.send_cq = queue->ib_cq;
271 	init_attr.recv_cq = queue->ib_cq;
272 
273 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
274 
275 	queue->qp = queue->cm_id->qp;
276 	return ret;
277 }
278 
279 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
280 		struct request *rq, unsigned int hctx_idx)
281 {
282 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
283 
284 	kfree(req->sqe.data);
285 }
286 
287 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
288 		struct request *rq, unsigned int hctx_idx,
289 		unsigned int numa_node)
290 {
291 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
292 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
293 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
294 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
295 
296 	nvme_req(rq)->ctrl = &ctrl->ctrl;
297 	req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
298 	if (!req->sqe.data)
299 		return -ENOMEM;
300 
301 	req->queue = queue;
302 
303 	return 0;
304 }
305 
306 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
307 		unsigned int hctx_idx)
308 {
309 	struct nvme_rdma_ctrl *ctrl = data;
310 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
311 
312 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
313 
314 	hctx->driver_data = queue;
315 	return 0;
316 }
317 
318 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
319 		unsigned int hctx_idx)
320 {
321 	struct nvme_rdma_ctrl *ctrl = data;
322 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
323 
324 	BUG_ON(hctx_idx != 0);
325 
326 	hctx->driver_data = queue;
327 	return 0;
328 }
329 
330 static void nvme_rdma_free_dev(struct kref *ref)
331 {
332 	struct nvme_rdma_device *ndev =
333 		container_of(ref, struct nvme_rdma_device, ref);
334 
335 	mutex_lock(&device_list_mutex);
336 	list_del(&ndev->entry);
337 	mutex_unlock(&device_list_mutex);
338 
339 	ib_dealloc_pd(ndev->pd);
340 	kfree(ndev);
341 }
342 
343 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
344 {
345 	kref_put(&dev->ref, nvme_rdma_free_dev);
346 }
347 
348 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
349 {
350 	return kref_get_unless_zero(&dev->ref);
351 }
352 
353 static struct nvme_rdma_device *
354 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
355 {
356 	struct nvme_rdma_device *ndev;
357 
358 	mutex_lock(&device_list_mutex);
359 	list_for_each_entry(ndev, &device_list, entry) {
360 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
361 		    nvme_rdma_dev_get(ndev))
362 			goto out_unlock;
363 	}
364 
365 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
366 	if (!ndev)
367 		goto out_err;
368 
369 	ndev->dev = cm_id->device;
370 	kref_init(&ndev->ref);
371 
372 	ndev->pd = ib_alloc_pd(ndev->dev,
373 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
374 	if (IS_ERR(ndev->pd))
375 		goto out_free_dev;
376 
377 	if (!(ndev->dev->attrs.device_cap_flags &
378 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
379 		dev_err(&ndev->dev->dev,
380 			"Memory registrations not supported.\n");
381 		goto out_free_pd;
382 	}
383 
384 	ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
385 					ndev->dev->attrs.max_send_sge - 1);
386 	list_add(&ndev->entry, &device_list);
387 out_unlock:
388 	mutex_unlock(&device_list_mutex);
389 	return ndev;
390 
391 out_free_pd:
392 	ib_dealloc_pd(ndev->pd);
393 out_free_dev:
394 	kfree(ndev);
395 out_err:
396 	mutex_unlock(&device_list_mutex);
397 	return NULL;
398 }
399 
400 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
401 {
402 	struct nvme_rdma_device *dev;
403 	struct ib_device *ibdev;
404 
405 	if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
406 		return;
407 
408 	dev = queue->device;
409 	ibdev = dev->dev;
410 
411 	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
412 
413 	/*
414 	 * The cm_id object might have been destroyed during RDMA connection
415 	 * establishment error flow to avoid getting other cma events, thus
416 	 * the destruction of the QP shouldn't use rdma_cm API.
417 	 */
418 	ib_destroy_qp(queue->qp);
419 	ib_free_cq(queue->ib_cq);
420 
421 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
422 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
423 
424 	nvme_rdma_dev_put(dev);
425 }
426 
427 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
428 {
429 	return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
430 		     ibdev->attrs.max_fast_reg_page_list_len);
431 }
432 
433 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
434 {
435 	struct ib_device *ibdev;
436 	const int send_wr_factor = 3;			/* MR, SEND, INV */
437 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
438 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
439 	enum ib_poll_context poll_ctx;
440 	int ret;
441 
442 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
443 	if (!queue->device) {
444 		dev_err(queue->cm_id->device->dev.parent,
445 			"no client data found!\n");
446 		return -ECONNREFUSED;
447 	}
448 	ibdev = queue->device->dev;
449 
450 	/*
451 	 * Spread I/O queues completion vectors according their queue index.
452 	 * Admin queues can always go on completion vector 0.
453 	 */
454 	comp_vector = idx == 0 ? idx : idx - 1;
455 
456 	/* Polling queues need direct cq polling context */
457 	if (nvme_rdma_poll_queue(queue))
458 		poll_ctx = IB_POLL_DIRECT;
459 	else
460 		poll_ctx = IB_POLL_SOFTIRQ;
461 
462 	/* +1 for ib_stop_cq */
463 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
464 				cq_factor * queue->queue_size + 1,
465 				comp_vector, poll_ctx);
466 	if (IS_ERR(queue->ib_cq)) {
467 		ret = PTR_ERR(queue->ib_cq);
468 		goto out_put_dev;
469 	}
470 
471 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
472 	if (ret)
473 		goto out_destroy_ib_cq;
474 
475 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
476 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
477 	if (!queue->rsp_ring) {
478 		ret = -ENOMEM;
479 		goto out_destroy_qp;
480 	}
481 
482 	ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
483 			      queue->queue_size,
484 			      IB_MR_TYPE_MEM_REG,
485 			      nvme_rdma_get_max_fr_pages(ibdev), 0);
486 	if (ret) {
487 		dev_err(queue->ctrl->ctrl.device,
488 			"failed to initialize MR pool sized %d for QID %d\n",
489 			queue->queue_size, idx);
490 		goto out_destroy_ring;
491 	}
492 
493 	set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
494 
495 	return 0;
496 
497 out_destroy_ring:
498 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
499 			    sizeof(struct nvme_completion), DMA_FROM_DEVICE);
500 out_destroy_qp:
501 	rdma_destroy_qp(queue->cm_id);
502 out_destroy_ib_cq:
503 	ib_free_cq(queue->ib_cq);
504 out_put_dev:
505 	nvme_rdma_dev_put(queue->device);
506 	return ret;
507 }
508 
509 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
510 		int idx, size_t queue_size)
511 {
512 	struct nvme_rdma_queue *queue;
513 	struct sockaddr *src_addr = NULL;
514 	int ret;
515 
516 	queue = &ctrl->queues[idx];
517 	queue->ctrl = ctrl;
518 	init_completion(&queue->cm_done);
519 
520 	if (idx > 0)
521 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
522 	else
523 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
524 
525 	queue->queue_size = queue_size;
526 
527 	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
528 			RDMA_PS_TCP, IB_QPT_RC);
529 	if (IS_ERR(queue->cm_id)) {
530 		dev_info(ctrl->ctrl.device,
531 			"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
532 		return PTR_ERR(queue->cm_id);
533 	}
534 
535 	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
536 		src_addr = (struct sockaddr *)&ctrl->src_addr;
537 
538 	queue->cm_error = -ETIMEDOUT;
539 	ret = rdma_resolve_addr(queue->cm_id, src_addr,
540 			(struct sockaddr *)&ctrl->addr,
541 			NVME_RDMA_CONNECT_TIMEOUT_MS);
542 	if (ret) {
543 		dev_info(ctrl->ctrl.device,
544 			"rdma_resolve_addr failed (%d).\n", ret);
545 		goto out_destroy_cm_id;
546 	}
547 
548 	ret = nvme_rdma_wait_for_cm(queue);
549 	if (ret) {
550 		dev_info(ctrl->ctrl.device,
551 			"rdma connection establishment failed (%d)\n", ret);
552 		goto out_destroy_cm_id;
553 	}
554 
555 	set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
556 
557 	return 0;
558 
559 out_destroy_cm_id:
560 	rdma_destroy_id(queue->cm_id);
561 	nvme_rdma_destroy_queue_ib(queue);
562 	return ret;
563 }
564 
565 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
566 {
567 	if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
568 		return;
569 
570 	rdma_disconnect(queue->cm_id);
571 	ib_drain_qp(queue->qp);
572 }
573 
574 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
575 {
576 	if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
577 		return;
578 
579 	nvme_rdma_destroy_queue_ib(queue);
580 	rdma_destroy_id(queue->cm_id);
581 }
582 
583 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
584 {
585 	int i;
586 
587 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
588 		nvme_rdma_free_queue(&ctrl->queues[i]);
589 }
590 
591 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
592 {
593 	int i;
594 
595 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
596 		nvme_rdma_stop_queue(&ctrl->queues[i]);
597 }
598 
599 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
600 {
601 	struct nvme_rdma_queue *queue = &ctrl->queues[idx];
602 	bool poll = nvme_rdma_poll_queue(queue);
603 	int ret;
604 
605 	if (idx)
606 		ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
607 	else
608 		ret = nvmf_connect_admin_queue(&ctrl->ctrl);
609 
610 	if (!ret)
611 		set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
612 	else
613 		dev_info(ctrl->ctrl.device,
614 			"failed to connect queue: %d ret=%d\n", idx, ret);
615 	return ret;
616 }
617 
618 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
619 {
620 	int i, ret = 0;
621 
622 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
623 		ret = nvme_rdma_start_queue(ctrl, i);
624 		if (ret)
625 			goto out_stop_queues;
626 	}
627 
628 	return 0;
629 
630 out_stop_queues:
631 	for (i--; i >= 1; i--)
632 		nvme_rdma_stop_queue(&ctrl->queues[i]);
633 	return ret;
634 }
635 
636 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
637 {
638 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
639 	struct ib_device *ibdev = ctrl->device->dev;
640 	unsigned int nr_io_queues, nr_default_queues;
641 	unsigned int nr_read_queues, nr_poll_queues;
642 	int i, ret;
643 
644 	nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
645 				min(opts->nr_io_queues, num_online_cpus()));
646 	nr_default_queues =  min_t(unsigned int, ibdev->num_comp_vectors,
647 				min(opts->nr_write_queues, num_online_cpus()));
648 	nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
649 	nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
650 
651 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
652 	if (ret)
653 		return ret;
654 
655 	ctrl->ctrl.queue_count = nr_io_queues + 1;
656 	if (ctrl->ctrl.queue_count < 2)
657 		return 0;
658 
659 	dev_info(ctrl->ctrl.device,
660 		"creating %d I/O queues.\n", nr_io_queues);
661 
662 	if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
663 		/*
664 		 * separate read/write queues
665 		 * hand out dedicated default queues only after we have
666 		 * sufficient read queues.
667 		 */
668 		ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
669 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
670 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
671 			min(nr_default_queues, nr_io_queues);
672 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
673 	} else {
674 		/*
675 		 * shared read/write queues
676 		 * either no write queues were requested, or we don't have
677 		 * sufficient queue count to have dedicated default queues.
678 		 */
679 		ctrl->io_queues[HCTX_TYPE_DEFAULT] =
680 			min(nr_read_queues, nr_io_queues);
681 		nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
682 	}
683 
684 	if (opts->nr_poll_queues && nr_io_queues) {
685 		/* map dedicated poll queues only if we have queues left */
686 		ctrl->io_queues[HCTX_TYPE_POLL] =
687 			min(nr_poll_queues, nr_io_queues);
688 	}
689 
690 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
691 		ret = nvme_rdma_alloc_queue(ctrl, i,
692 				ctrl->ctrl.sqsize + 1);
693 		if (ret)
694 			goto out_free_queues;
695 	}
696 
697 	return 0;
698 
699 out_free_queues:
700 	for (i--; i >= 1; i--)
701 		nvme_rdma_free_queue(&ctrl->queues[i]);
702 
703 	return ret;
704 }
705 
706 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
707 		bool admin)
708 {
709 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
710 	struct blk_mq_tag_set *set;
711 	int ret;
712 
713 	if (admin) {
714 		set = &ctrl->admin_tag_set;
715 		memset(set, 0, sizeof(*set));
716 		set->ops = &nvme_rdma_admin_mq_ops;
717 		set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
718 		set->reserved_tags = 2; /* connect + keep-alive */
719 		set->numa_node = nctrl->numa_node;
720 		set->cmd_size = sizeof(struct nvme_rdma_request) +
721 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
722 		set->driver_data = ctrl;
723 		set->nr_hw_queues = 1;
724 		set->timeout = ADMIN_TIMEOUT;
725 		set->flags = BLK_MQ_F_NO_SCHED;
726 	} else {
727 		set = &ctrl->tag_set;
728 		memset(set, 0, sizeof(*set));
729 		set->ops = &nvme_rdma_mq_ops;
730 		set->queue_depth = nctrl->sqsize + 1;
731 		set->reserved_tags = 1; /* fabric connect */
732 		set->numa_node = nctrl->numa_node;
733 		set->flags = BLK_MQ_F_SHOULD_MERGE;
734 		set->cmd_size = sizeof(struct nvme_rdma_request) +
735 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
736 		set->driver_data = ctrl;
737 		set->nr_hw_queues = nctrl->queue_count - 1;
738 		set->timeout = NVME_IO_TIMEOUT;
739 		set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
740 	}
741 
742 	ret = blk_mq_alloc_tag_set(set);
743 	if (ret)
744 		return ERR_PTR(ret);
745 
746 	return set;
747 }
748 
749 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
750 		bool remove)
751 {
752 	if (remove) {
753 		blk_cleanup_queue(ctrl->ctrl.admin_q);
754 		blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
755 	}
756 	if (ctrl->async_event_sqe.data) {
757 		nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
758 				sizeof(struct nvme_command), DMA_TO_DEVICE);
759 		ctrl->async_event_sqe.data = NULL;
760 	}
761 	nvme_rdma_free_queue(&ctrl->queues[0]);
762 }
763 
764 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
765 		bool new)
766 {
767 	int error;
768 
769 	error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
770 	if (error)
771 		return error;
772 
773 	ctrl->device = ctrl->queues[0].device;
774 	ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
775 
776 	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
777 
778 	/*
779 	 * Bind the async event SQE DMA mapping to the admin queue lifetime.
780 	 * It's safe, since any chage in the underlying RDMA device will issue
781 	 * error recovery and queue re-creation.
782 	 */
783 	error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
784 			sizeof(struct nvme_command), DMA_TO_DEVICE);
785 	if (error)
786 		goto out_free_queue;
787 
788 	if (new) {
789 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
790 		if (IS_ERR(ctrl->ctrl.admin_tagset)) {
791 			error = PTR_ERR(ctrl->ctrl.admin_tagset);
792 			goto out_free_async_qe;
793 		}
794 
795 		ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
796 		if (IS_ERR(ctrl->ctrl.admin_q)) {
797 			error = PTR_ERR(ctrl->ctrl.admin_q);
798 			goto out_free_tagset;
799 		}
800 	}
801 
802 	error = nvme_rdma_start_queue(ctrl, 0);
803 	if (error)
804 		goto out_cleanup_queue;
805 
806 	error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
807 			&ctrl->ctrl.cap);
808 	if (error) {
809 		dev_err(ctrl->ctrl.device,
810 			"prop_get NVME_REG_CAP failed\n");
811 		goto out_stop_queue;
812 	}
813 
814 	ctrl->ctrl.sqsize =
815 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
816 
817 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
818 	if (error)
819 		goto out_stop_queue;
820 
821 	ctrl->ctrl.max_hw_sectors =
822 		(ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
823 
824 	error = nvme_init_identify(&ctrl->ctrl);
825 	if (error)
826 		goto out_stop_queue;
827 
828 	return 0;
829 
830 out_stop_queue:
831 	nvme_rdma_stop_queue(&ctrl->queues[0]);
832 out_cleanup_queue:
833 	if (new)
834 		blk_cleanup_queue(ctrl->ctrl.admin_q);
835 out_free_tagset:
836 	if (new)
837 		blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
838 out_free_async_qe:
839 	nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
840 		sizeof(struct nvme_command), DMA_TO_DEVICE);
841 	ctrl->async_event_sqe.data = NULL;
842 out_free_queue:
843 	nvme_rdma_free_queue(&ctrl->queues[0]);
844 	return error;
845 }
846 
847 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
848 		bool remove)
849 {
850 	if (remove) {
851 		blk_cleanup_queue(ctrl->ctrl.connect_q);
852 		blk_mq_free_tag_set(ctrl->ctrl.tagset);
853 	}
854 	nvme_rdma_free_io_queues(ctrl);
855 }
856 
857 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
858 {
859 	int ret;
860 
861 	ret = nvme_rdma_alloc_io_queues(ctrl);
862 	if (ret)
863 		return ret;
864 
865 	if (new) {
866 		ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
867 		if (IS_ERR(ctrl->ctrl.tagset)) {
868 			ret = PTR_ERR(ctrl->ctrl.tagset);
869 			goto out_free_io_queues;
870 		}
871 
872 		ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
873 		if (IS_ERR(ctrl->ctrl.connect_q)) {
874 			ret = PTR_ERR(ctrl->ctrl.connect_q);
875 			goto out_free_tag_set;
876 		}
877 	} else {
878 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
879 			ctrl->ctrl.queue_count - 1);
880 	}
881 
882 	ret = nvme_rdma_start_io_queues(ctrl);
883 	if (ret)
884 		goto out_cleanup_connect_q;
885 
886 	return 0;
887 
888 out_cleanup_connect_q:
889 	if (new)
890 		blk_cleanup_queue(ctrl->ctrl.connect_q);
891 out_free_tag_set:
892 	if (new)
893 		blk_mq_free_tag_set(ctrl->ctrl.tagset);
894 out_free_io_queues:
895 	nvme_rdma_free_io_queues(ctrl);
896 	return ret;
897 }
898 
899 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
900 		bool remove)
901 {
902 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
903 	nvme_rdma_stop_queue(&ctrl->queues[0]);
904 	if (ctrl->ctrl.admin_tagset)
905 		blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
906 			nvme_cancel_request, &ctrl->ctrl);
907 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
908 	nvme_rdma_destroy_admin_queue(ctrl, remove);
909 }
910 
911 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
912 		bool remove)
913 {
914 	if (ctrl->ctrl.queue_count > 1) {
915 		nvme_stop_queues(&ctrl->ctrl);
916 		nvme_rdma_stop_io_queues(ctrl);
917 		if (ctrl->ctrl.tagset)
918 			blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
919 				nvme_cancel_request, &ctrl->ctrl);
920 		if (remove)
921 			nvme_start_queues(&ctrl->ctrl);
922 		nvme_rdma_destroy_io_queues(ctrl, remove);
923 	}
924 }
925 
926 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
927 {
928 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
929 
930 	if (list_empty(&ctrl->list))
931 		goto free_ctrl;
932 
933 	mutex_lock(&nvme_rdma_ctrl_mutex);
934 	list_del(&ctrl->list);
935 	mutex_unlock(&nvme_rdma_ctrl_mutex);
936 
937 	nvmf_free_options(nctrl->opts);
938 free_ctrl:
939 	kfree(ctrl->queues);
940 	kfree(ctrl);
941 }
942 
943 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
944 {
945 	/* If we are resetting/deleting then do nothing */
946 	if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
947 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
948 			ctrl->ctrl.state == NVME_CTRL_LIVE);
949 		return;
950 	}
951 
952 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
953 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
954 			ctrl->ctrl.opts->reconnect_delay);
955 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
956 				ctrl->ctrl.opts->reconnect_delay * HZ);
957 	} else {
958 		nvme_delete_ctrl(&ctrl->ctrl);
959 	}
960 }
961 
962 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
963 {
964 	int ret = -EINVAL;
965 	bool changed;
966 
967 	ret = nvme_rdma_configure_admin_queue(ctrl, new);
968 	if (ret)
969 		return ret;
970 
971 	if (ctrl->ctrl.icdoff) {
972 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
973 		goto destroy_admin;
974 	}
975 
976 	if (!(ctrl->ctrl.sgls & (1 << 2))) {
977 		dev_err(ctrl->ctrl.device,
978 			"Mandatory keyed sgls are not supported!\n");
979 		goto destroy_admin;
980 	}
981 
982 	if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
983 		dev_warn(ctrl->ctrl.device,
984 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
985 			ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
986 	}
987 
988 	if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
989 		dev_warn(ctrl->ctrl.device,
990 			"sqsize %u > ctrl maxcmd %u, clamping down\n",
991 			ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
992 		ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
993 	}
994 
995 	if (ctrl->ctrl.sgls & (1 << 20))
996 		ctrl->use_inline_data = true;
997 
998 	if (ctrl->ctrl.queue_count > 1) {
999 		ret = nvme_rdma_configure_io_queues(ctrl, new);
1000 		if (ret)
1001 			goto destroy_admin;
1002 	}
1003 
1004 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1005 	if (!changed) {
1006 		/* state change failure is ok if we're in DELETING state */
1007 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1008 		ret = -EINVAL;
1009 		goto destroy_io;
1010 	}
1011 
1012 	nvme_start_ctrl(&ctrl->ctrl);
1013 	return 0;
1014 
1015 destroy_io:
1016 	if (ctrl->ctrl.queue_count > 1)
1017 		nvme_rdma_destroy_io_queues(ctrl, new);
1018 destroy_admin:
1019 	nvme_rdma_stop_queue(&ctrl->queues[0]);
1020 	nvme_rdma_destroy_admin_queue(ctrl, new);
1021 	return ret;
1022 }
1023 
1024 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1025 {
1026 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1027 			struct nvme_rdma_ctrl, reconnect_work);
1028 
1029 	++ctrl->ctrl.nr_reconnects;
1030 
1031 	if (nvme_rdma_setup_ctrl(ctrl, false))
1032 		goto requeue;
1033 
1034 	dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1035 			ctrl->ctrl.nr_reconnects);
1036 
1037 	ctrl->ctrl.nr_reconnects = 0;
1038 
1039 	return;
1040 
1041 requeue:
1042 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1043 			ctrl->ctrl.nr_reconnects);
1044 	nvme_rdma_reconnect_or_remove(ctrl);
1045 }
1046 
1047 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1048 {
1049 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1050 			struct nvme_rdma_ctrl, err_work);
1051 
1052 	nvme_stop_keep_alive(&ctrl->ctrl);
1053 	nvme_rdma_teardown_io_queues(ctrl, false);
1054 	nvme_start_queues(&ctrl->ctrl);
1055 	nvme_rdma_teardown_admin_queue(ctrl, false);
1056 
1057 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1058 		/* state change failure is ok if we're in DELETING state */
1059 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1060 		return;
1061 	}
1062 
1063 	nvme_rdma_reconnect_or_remove(ctrl);
1064 }
1065 
1066 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1067 {
1068 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1069 		return;
1070 
1071 	queue_work(nvme_wq, &ctrl->err_work);
1072 }
1073 
1074 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1075 		const char *op)
1076 {
1077 	struct nvme_rdma_queue *queue = cq->cq_context;
1078 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1079 
1080 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1081 		dev_info(ctrl->ctrl.device,
1082 			     "%s for CQE 0x%p failed with status %s (%d)\n",
1083 			     op, wc->wr_cqe,
1084 			     ib_wc_status_msg(wc->status), wc->status);
1085 	nvme_rdma_error_recovery(ctrl);
1086 }
1087 
1088 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1089 {
1090 	if (unlikely(wc->status != IB_WC_SUCCESS))
1091 		nvme_rdma_wr_error(cq, wc, "MEMREG");
1092 }
1093 
1094 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1095 {
1096 	struct nvme_rdma_request *req =
1097 		container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1098 	struct request *rq = blk_mq_rq_from_pdu(req);
1099 
1100 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1101 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1102 		return;
1103 	}
1104 
1105 	if (refcount_dec_and_test(&req->ref))
1106 		nvme_end_request(rq, req->status, req->result);
1107 
1108 }
1109 
1110 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1111 		struct nvme_rdma_request *req)
1112 {
1113 	struct ib_send_wr wr = {
1114 		.opcode		    = IB_WR_LOCAL_INV,
1115 		.next		    = NULL,
1116 		.num_sge	    = 0,
1117 		.send_flags	    = IB_SEND_SIGNALED,
1118 		.ex.invalidate_rkey = req->mr->rkey,
1119 	};
1120 
1121 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1122 	wr.wr_cqe = &req->reg_cqe;
1123 
1124 	return ib_post_send(queue->qp, &wr, NULL);
1125 }
1126 
1127 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1128 		struct request *rq)
1129 {
1130 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1131 	struct nvme_rdma_device *dev = queue->device;
1132 	struct ib_device *ibdev = dev->dev;
1133 
1134 	if (!blk_rq_nr_phys_segments(rq))
1135 		return;
1136 
1137 	if (req->mr) {
1138 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1139 		req->mr = NULL;
1140 	}
1141 
1142 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1143 			req->nents, rq_data_dir(rq) ==
1144 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1145 
1146 	nvme_cleanup_cmd(rq);
1147 	sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
1148 }
1149 
1150 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1151 {
1152 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1153 
1154 	sg->addr = 0;
1155 	put_unaligned_le24(0, sg->length);
1156 	put_unaligned_le32(0, sg->key);
1157 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1158 	return 0;
1159 }
1160 
1161 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1162 		struct nvme_rdma_request *req, struct nvme_command *c,
1163 		int count)
1164 {
1165 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1166 	struct scatterlist *sgl = req->sg_table.sgl;
1167 	struct ib_sge *sge = &req->sge[1];
1168 	u32 len = 0;
1169 	int i;
1170 
1171 	for (i = 0; i < count; i++, sgl++, sge++) {
1172 		sge->addr = sg_dma_address(sgl);
1173 		sge->length = sg_dma_len(sgl);
1174 		sge->lkey = queue->device->pd->local_dma_lkey;
1175 		len += sge->length;
1176 	}
1177 
1178 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1179 	sg->length = cpu_to_le32(len);
1180 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1181 
1182 	req->num_sge += count;
1183 	return 0;
1184 }
1185 
1186 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1187 		struct nvme_rdma_request *req, struct nvme_command *c)
1188 {
1189 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1190 
1191 	sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1192 	put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1193 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1194 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1195 	return 0;
1196 }
1197 
1198 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1199 		struct nvme_rdma_request *req, struct nvme_command *c,
1200 		int count)
1201 {
1202 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1203 	int nr;
1204 
1205 	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1206 	if (WARN_ON_ONCE(!req->mr))
1207 		return -EAGAIN;
1208 
1209 	/*
1210 	 * Align the MR to a 4K page size to match the ctrl page size and
1211 	 * the block virtual boundary.
1212 	 */
1213 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1214 	if (unlikely(nr < count)) {
1215 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1216 		req->mr = NULL;
1217 		if (nr < 0)
1218 			return nr;
1219 		return -EINVAL;
1220 	}
1221 
1222 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1223 
1224 	req->reg_cqe.done = nvme_rdma_memreg_done;
1225 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1226 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1227 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1228 	req->reg_wr.wr.num_sge = 0;
1229 	req->reg_wr.mr = req->mr;
1230 	req->reg_wr.key = req->mr->rkey;
1231 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1232 			     IB_ACCESS_REMOTE_READ |
1233 			     IB_ACCESS_REMOTE_WRITE;
1234 
1235 	sg->addr = cpu_to_le64(req->mr->iova);
1236 	put_unaligned_le24(req->mr->length, sg->length);
1237 	put_unaligned_le32(req->mr->rkey, sg->key);
1238 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1239 			NVME_SGL_FMT_INVALIDATE;
1240 
1241 	return 0;
1242 }
1243 
1244 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1245 		struct request *rq, struct nvme_command *c)
1246 {
1247 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1248 	struct nvme_rdma_device *dev = queue->device;
1249 	struct ib_device *ibdev = dev->dev;
1250 	int count, ret;
1251 
1252 	req->num_sge = 1;
1253 	refcount_set(&req->ref, 2); /* send and recv completions */
1254 
1255 	c->common.flags |= NVME_CMD_SGL_METABUF;
1256 
1257 	if (!blk_rq_nr_phys_segments(rq))
1258 		return nvme_rdma_set_sg_null(c);
1259 
1260 	req->sg_table.sgl = req->first_sgl;
1261 	ret = sg_alloc_table_chained(&req->sg_table,
1262 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl,
1263 			SG_CHUNK_SIZE);
1264 	if (ret)
1265 		return -ENOMEM;
1266 
1267 	req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1268 
1269 	count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1270 		    rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1271 	if (unlikely(count <= 0)) {
1272 		ret = -EIO;
1273 		goto out_free_table;
1274 	}
1275 
1276 	if (count <= dev->num_inline_segments) {
1277 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1278 		    queue->ctrl->use_inline_data &&
1279 		    blk_rq_payload_bytes(rq) <=
1280 				nvme_rdma_inline_data_size(queue)) {
1281 			ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1282 			goto out;
1283 		}
1284 
1285 		if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1286 			ret = nvme_rdma_map_sg_single(queue, req, c);
1287 			goto out;
1288 		}
1289 	}
1290 
1291 	ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1292 out:
1293 	if (unlikely(ret))
1294 		goto out_unmap_sg;
1295 
1296 	return 0;
1297 
1298 out_unmap_sg:
1299 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1300 			req->nents, rq_data_dir(rq) ==
1301 			WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1302 out_free_table:
1303 	sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
1304 	return ret;
1305 }
1306 
1307 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1308 {
1309 	struct nvme_rdma_qe *qe =
1310 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1311 	struct nvme_rdma_request *req =
1312 		container_of(qe, struct nvme_rdma_request, sqe);
1313 	struct request *rq = blk_mq_rq_from_pdu(req);
1314 
1315 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1316 		nvme_rdma_wr_error(cq, wc, "SEND");
1317 		return;
1318 	}
1319 
1320 	if (refcount_dec_and_test(&req->ref))
1321 		nvme_end_request(rq, req->status, req->result);
1322 }
1323 
1324 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1325 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1326 		struct ib_send_wr *first)
1327 {
1328 	struct ib_send_wr wr;
1329 	int ret;
1330 
1331 	sge->addr   = qe->dma;
1332 	sge->length = sizeof(struct nvme_command),
1333 	sge->lkey   = queue->device->pd->local_dma_lkey;
1334 
1335 	wr.next       = NULL;
1336 	wr.wr_cqe     = &qe->cqe;
1337 	wr.sg_list    = sge;
1338 	wr.num_sge    = num_sge;
1339 	wr.opcode     = IB_WR_SEND;
1340 	wr.send_flags = IB_SEND_SIGNALED;
1341 
1342 	if (first)
1343 		first->next = &wr;
1344 	else
1345 		first = &wr;
1346 
1347 	ret = ib_post_send(queue->qp, first, NULL);
1348 	if (unlikely(ret)) {
1349 		dev_err(queue->ctrl->ctrl.device,
1350 			     "%s failed with error code %d\n", __func__, ret);
1351 	}
1352 	return ret;
1353 }
1354 
1355 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1356 		struct nvme_rdma_qe *qe)
1357 {
1358 	struct ib_recv_wr wr;
1359 	struct ib_sge list;
1360 	int ret;
1361 
1362 	list.addr   = qe->dma;
1363 	list.length = sizeof(struct nvme_completion);
1364 	list.lkey   = queue->device->pd->local_dma_lkey;
1365 
1366 	qe->cqe.done = nvme_rdma_recv_done;
1367 
1368 	wr.next     = NULL;
1369 	wr.wr_cqe   = &qe->cqe;
1370 	wr.sg_list  = &list;
1371 	wr.num_sge  = 1;
1372 
1373 	ret = ib_post_recv(queue->qp, &wr, NULL);
1374 	if (unlikely(ret)) {
1375 		dev_err(queue->ctrl->ctrl.device,
1376 			"%s failed with error code %d\n", __func__, ret);
1377 	}
1378 	return ret;
1379 }
1380 
1381 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1382 {
1383 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1384 
1385 	if (queue_idx == 0)
1386 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1387 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1388 }
1389 
1390 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1391 {
1392 	if (unlikely(wc->status != IB_WC_SUCCESS))
1393 		nvme_rdma_wr_error(cq, wc, "ASYNC");
1394 }
1395 
1396 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1397 {
1398 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1399 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1400 	struct ib_device *dev = queue->device->dev;
1401 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1402 	struct nvme_command *cmd = sqe->data;
1403 	struct ib_sge sge;
1404 	int ret;
1405 
1406 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1407 
1408 	memset(cmd, 0, sizeof(*cmd));
1409 	cmd->common.opcode = nvme_admin_async_event;
1410 	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1411 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1412 	nvme_rdma_set_sg_null(cmd);
1413 
1414 	sqe->cqe.done = nvme_rdma_async_done;
1415 
1416 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1417 			DMA_TO_DEVICE);
1418 
1419 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1420 	WARN_ON_ONCE(ret);
1421 }
1422 
1423 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1424 		struct nvme_completion *cqe, struct ib_wc *wc)
1425 {
1426 	struct request *rq;
1427 	struct nvme_rdma_request *req;
1428 
1429 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1430 	if (!rq) {
1431 		dev_err(queue->ctrl->ctrl.device,
1432 			"tag 0x%x on QP %#x not found\n",
1433 			cqe->command_id, queue->qp->qp_num);
1434 		nvme_rdma_error_recovery(queue->ctrl);
1435 		return;
1436 	}
1437 	req = blk_mq_rq_to_pdu(rq);
1438 
1439 	req->status = cqe->status;
1440 	req->result = cqe->result;
1441 
1442 	if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1443 		if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1444 			dev_err(queue->ctrl->ctrl.device,
1445 				"Bogus remote invalidation for rkey %#x\n",
1446 				req->mr->rkey);
1447 			nvme_rdma_error_recovery(queue->ctrl);
1448 		}
1449 	} else if (req->mr) {
1450 		int ret;
1451 
1452 		ret = nvme_rdma_inv_rkey(queue, req);
1453 		if (unlikely(ret < 0)) {
1454 			dev_err(queue->ctrl->ctrl.device,
1455 				"Queueing INV WR for rkey %#x failed (%d)\n",
1456 				req->mr->rkey, ret);
1457 			nvme_rdma_error_recovery(queue->ctrl);
1458 		}
1459 		/* the local invalidation completion will end the request */
1460 		return;
1461 	}
1462 
1463 	if (refcount_dec_and_test(&req->ref))
1464 		nvme_end_request(rq, req->status, req->result);
1465 }
1466 
1467 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1468 {
1469 	struct nvme_rdma_qe *qe =
1470 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1471 	struct nvme_rdma_queue *queue = cq->cq_context;
1472 	struct ib_device *ibdev = queue->device->dev;
1473 	struct nvme_completion *cqe = qe->data;
1474 	const size_t len = sizeof(struct nvme_completion);
1475 
1476 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1477 		nvme_rdma_wr_error(cq, wc, "RECV");
1478 		return;
1479 	}
1480 
1481 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1482 	/*
1483 	 * AEN requests are special as they don't time out and can
1484 	 * survive any kind of queue freeze and often don't respond to
1485 	 * aborts.  We don't even bother to allocate a struct request
1486 	 * for them but rather special case them here.
1487 	 */
1488 	if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1489 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1490 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1491 				&cqe->result);
1492 	else
1493 		nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1494 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1495 
1496 	nvme_rdma_post_recv(queue, qe);
1497 }
1498 
1499 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1500 {
1501 	int ret, i;
1502 
1503 	for (i = 0; i < queue->queue_size; i++) {
1504 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1505 		if (ret)
1506 			goto out_destroy_queue_ib;
1507 	}
1508 
1509 	return 0;
1510 
1511 out_destroy_queue_ib:
1512 	nvme_rdma_destroy_queue_ib(queue);
1513 	return ret;
1514 }
1515 
1516 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1517 		struct rdma_cm_event *ev)
1518 {
1519 	struct rdma_cm_id *cm_id = queue->cm_id;
1520 	int status = ev->status;
1521 	const char *rej_msg;
1522 	const struct nvme_rdma_cm_rej *rej_data;
1523 	u8 rej_data_len;
1524 
1525 	rej_msg = rdma_reject_msg(cm_id, status);
1526 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1527 
1528 	if (rej_data && rej_data_len >= sizeof(u16)) {
1529 		u16 sts = le16_to_cpu(rej_data->sts);
1530 
1531 		dev_err(queue->ctrl->ctrl.device,
1532 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1533 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1534 	} else {
1535 		dev_err(queue->ctrl->ctrl.device,
1536 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1537 	}
1538 
1539 	return -ECONNRESET;
1540 }
1541 
1542 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1543 {
1544 	int ret;
1545 
1546 	ret = nvme_rdma_create_queue_ib(queue);
1547 	if (ret)
1548 		return ret;
1549 
1550 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1551 	if (ret) {
1552 		dev_err(queue->ctrl->ctrl.device,
1553 			"rdma_resolve_route failed (%d).\n",
1554 			queue->cm_error);
1555 		goto out_destroy_queue;
1556 	}
1557 
1558 	return 0;
1559 
1560 out_destroy_queue:
1561 	nvme_rdma_destroy_queue_ib(queue);
1562 	return ret;
1563 }
1564 
1565 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1566 {
1567 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1568 	struct rdma_conn_param param = { };
1569 	struct nvme_rdma_cm_req priv = { };
1570 	int ret;
1571 
1572 	param.qp_num = queue->qp->qp_num;
1573 	param.flow_control = 1;
1574 
1575 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1576 	/* maximum retry count */
1577 	param.retry_count = 7;
1578 	param.rnr_retry_count = 7;
1579 	param.private_data = &priv;
1580 	param.private_data_len = sizeof(priv);
1581 
1582 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1583 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1584 	/*
1585 	 * set the admin queue depth to the minimum size
1586 	 * specified by the Fabrics standard.
1587 	 */
1588 	if (priv.qid == 0) {
1589 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1590 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1591 	} else {
1592 		/*
1593 		 * current interpretation of the fabrics spec
1594 		 * is at minimum you make hrqsize sqsize+1, or a
1595 		 * 1's based representation of sqsize.
1596 		 */
1597 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1598 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1599 	}
1600 
1601 	ret = rdma_connect(queue->cm_id, &param);
1602 	if (ret) {
1603 		dev_err(ctrl->ctrl.device,
1604 			"rdma_connect failed (%d).\n", ret);
1605 		goto out_destroy_queue_ib;
1606 	}
1607 
1608 	return 0;
1609 
1610 out_destroy_queue_ib:
1611 	nvme_rdma_destroy_queue_ib(queue);
1612 	return ret;
1613 }
1614 
1615 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1616 		struct rdma_cm_event *ev)
1617 {
1618 	struct nvme_rdma_queue *queue = cm_id->context;
1619 	int cm_error = 0;
1620 
1621 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1622 		rdma_event_msg(ev->event), ev->event,
1623 		ev->status, cm_id);
1624 
1625 	switch (ev->event) {
1626 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1627 		cm_error = nvme_rdma_addr_resolved(queue);
1628 		break;
1629 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1630 		cm_error = nvme_rdma_route_resolved(queue);
1631 		break;
1632 	case RDMA_CM_EVENT_ESTABLISHED:
1633 		queue->cm_error = nvme_rdma_conn_established(queue);
1634 		/* complete cm_done regardless of success/failure */
1635 		complete(&queue->cm_done);
1636 		return 0;
1637 	case RDMA_CM_EVENT_REJECTED:
1638 		nvme_rdma_destroy_queue_ib(queue);
1639 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1640 		break;
1641 	case RDMA_CM_EVENT_ROUTE_ERROR:
1642 	case RDMA_CM_EVENT_CONNECT_ERROR:
1643 	case RDMA_CM_EVENT_UNREACHABLE:
1644 		nvme_rdma_destroy_queue_ib(queue);
1645 		/* fall through */
1646 	case RDMA_CM_EVENT_ADDR_ERROR:
1647 		dev_dbg(queue->ctrl->ctrl.device,
1648 			"CM error event %d\n", ev->event);
1649 		cm_error = -ECONNRESET;
1650 		break;
1651 	case RDMA_CM_EVENT_DISCONNECTED:
1652 	case RDMA_CM_EVENT_ADDR_CHANGE:
1653 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1654 		dev_dbg(queue->ctrl->ctrl.device,
1655 			"disconnect received - connection closed\n");
1656 		nvme_rdma_error_recovery(queue->ctrl);
1657 		break;
1658 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1659 		/* device removal is handled via the ib_client API */
1660 		break;
1661 	default:
1662 		dev_err(queue->ctrl->ctrl.device,
1663 			"Unexpected RDMA CM event (%d)\n", ev->event);
1664 		nvme_rdma_error_recovery(queue->ctrl);
1665 		break;
1666 	}
1667 
1668 	if (cm_error) {
1669 		queue->cm_error = cm_error;
1670 		complete(&queue->cm_done);
1671 	}
1672 
1673 	return 0;
1674 }
1675 
1676 static enum blk_eh_timer_return
1677 nvme_rdma_timeout(struct request *rq, bool reserved)
1678 {
1679 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1680 	struct nvme_rdma_queue *queue = req->queue;
1681 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1682 
1683 	dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1684 		 rq->tag, nvme_rdma_queue_idx(queue));
1685 
1686 	if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1687 		/*
1688 		 * Teardown immediately if controller times out while starting
1689 		 * or we are already started error recovery. all outstanding
1690 		 * requests are completed on shutdown, so we return BLK_EH_DONE.
1691 		 */
1692 		flush_work(&ctrl->err_work);
1693 		nvme_rdma_teardown_io_queues(ctrl, false);
1694 		nvme_rdma_teardown_admin_queue(ctrl, false);
1695 		return BLK_EH_DONE;
1696 	}
1697 
1698 	dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1699 	nvme_rdma_error_recovery(ctrl);
1700 
1701 	return BLK_EH_RESET_TIMER;
1702 }
1703 
1704 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1705 		const struct blk_mq_queue_data *bd)
1706 {
1707 	struct nvme_ns *ns = hctx->queue->queuedata;
1708 	struct nvme_rdma_queue *queue = hctx->driver_data;
1709 	struct request *rq = bd->rq;
1710 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1711 	struct nvme_rdma_qe *sqe = &req->sqe;
1712 	struct nvme_command *c = sqe->data;
1713 	struct ib_device *dev;
1714 	bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1715 	blk_status_t ret;
1716 	int err;
1717 
1718 	WARN_ON_ONCE(rq->tag < 0);
1719 
1720 	if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1721 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1722 
1723 	dev = queue->device->dev;
1724 
1725 	req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
1726 					 sizeof(struct nvme_command),
1727 					 DMA_TO_DEVICE);
1728 	err = ib_dma_mapping_error(dev, req->sqe.dma);
1729 	if (unlikely(err))
1730 		return BLK_STS_RESOURCE;
1731 
1732 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1733 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1734 
1735 	ret = nvme_setup_cmd(ns, rq, c);
1736 	if (ret)
1737 		goto unmap_qe;
1738 
1739 	blk_mq_start_request(rq);
1740 
1741 	err = nvme_rdma_map_data(queue, rq, c);
1742 	if (unlikely(err < 0)) {
1743 		dev_err(queue->ctrl->ctrl.device,
1744 			     "Failed to map data (%d)\n", err);
1745 		nvme_cleanup_cmd(rq);
1746 		goto err;
1747 	}
1748 
1749 	sqe->cqe.done = nvme_rdma_send_done;
1750 
1751 	ib_dma_sync_single_for_device(dev, sqe->dma,
1752 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1753 
1754 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1755 			req->mr ? &req->reg_wr.wr : NULL);
1756 	if (unlikely(err)) {
1757 		nvme_rdma_unmap_data(queue, rq);
1758 		goto err;
1759 	}
1760 
1761 	return BLK_STS_OK;
1762 
1763 err:
1764 	if (err == -ENOMEM || err == -EAGAIN)
1765 		ret = BLK_STS_RESOURCE;
1766 	else
1767 		ret = BLK_STS_IOERR;
1768 unmap_qe:
1769 	ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
1770 			    DMA_TO_DEVICE);
1771 	return ret;
1772 }
1773 
1774 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1775 {
1776 	struct nvme_rdma_queue *queue = hctx->driver_data;
1777 
1778 	return ib_process_cq_direct(queue->ib_cq, -1);
1779 }
1780 
1781 static void nvme_rdma_complete_rq(struct request *rq)
1782 {
1783 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1784 	struct nvme_rdma_queue *queue = req->queue;
1785 	struct ib_device *ibdev = queue->device->dev;
1786 
1787 	nvme_rdma_unmap_data(queue, rq);
1788 	ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
1789 			    DMA_TO_DEVICE);
1790 	nvme_complete_rq(rq);
1791 }
1792 
1793 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1794 {
1795 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
1796 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1797 
1798 	if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
1799 		/* separate read/write queues */
1800 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
1801 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
1802 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1803 		set->map[HCTX_TYPE_READ].nr_queues =
1804 			ctrl->io_queues[HCTX_TYPE_READ];
1805 		set->map[HCTX_TYPE_READ].queue_offset =
1806 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
1807 	} else {
1808 		/* shared read/write queues */
1809 		set->map[HCTX_TYPE_DEFAULT].nr_queues =
1810 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
1811 		set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1812 		set->map[HCTX_TYPE_READ].nr_queues =
1813 			ctrl->io_queues[HCTX_TYPE_DEFAULT];
1814 		set->map[HCTX_TYPE_READ].queue_offset = 0;
1815 	}
1816 	blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1817 			ctrl->device->dev, 0);
1818 	blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1819 			ctrl->device->dev, 0);
1820 
1821 	if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
1822 		/* map dedicated poll queues only if we have queues left */
1823 		set->map[HCTX_TYPE_POLL].nr_queues =
1824 				ctrl->io_queues[HCTX_TYPE_POLL];
1825 		set->map[HCTX_TYPE_POLL].queue_offset =
1826 			ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1827 			ctrl->io_queues[HCTX_TYPE_READ];
1828 		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1829 	}
1830 
1831 	dev_info(ctrl->ctrl.device,
1832 		"mapped %d/%d/%d default/read/poll queues.\n",
1833 		ctrl->io_queues[HCTX_TYPE_DEFAULT],
1834 		ctrl->io_queues[HCTX_TYPE_READ],
1835 		ctrl->io_queues[HCTX_TYPE_POLL]);
1836 
1837 	return 0;
1838 }
1839 
1840 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1841 	.queue_rq	= nvme_rdma_queue_rq,
1842 	.complete	= nvme_rdma_complete_rq,
1843 	.init_request	= nvme_rdma_init_request,
1844 	.exit_request	= nvme_rdma_exit_request,
1845 	.init_hctx	= nvme_rdma_init_hctx,
1846 	.timeout	= nvme_rdma_timeout,
1847 	.map_queues	= nvme_rdma_map_queues,
1848 	.poll		= nvme_rdma_poll,
1849 };
1850 
1851 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1852 	.queue_rq	= nvme_rdma_queue_rq,
1853 	.complete	= nvme_rdma_complete_rq,
1854 	.init_request	= nvme_rdma_init_request,
1855 	.exit_request	= nvme_rdma_exit_request,
1856 	.init_hctx	= nvme_rdma_init_admin_hctx,
1857 	.timeout	= nvme_rdma_timeout,
1858 };
1859 
1860 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1861 {
1862 	cancel_work_sync(&ctrl->err_work);
1863 	cancel_delayed_work_sync(&ctrl->reconnect_work);
1864 
1865 	nvme_rdma_teardown_io_queues(ctrl, shutdown);
1866 	if (shutdown)
1867 		nvme_shutdown_ctrl(&ctrl->ctrl);
1868 	else
1869 		nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1870 	nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1871 }
1872 
1873 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1874 {
1875 	nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1876 }
1877 
1878 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1879 {
1880 	struct nvme_rdma_ctrl *ctrl =
1881 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1882 
1883 	nvme_stop_ctrl(&ctrl->ctrl);
1884 	nvme_rdma_shutdown_ctrl(ctrl, false);
1885 
1886 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1887 		/* state change failure should never happen */
1888 		WARN_ON_ONCE(1);
1889 		return;
1890 	}
1891 
1892 	if (nvme_rdma_setup_ctrl(ctrl, false))
1893 		goto out_fail;
1894 
1895 	return;
1896 
1897 out_fail:
1898 	++ctrl->ctrl.nr_reconnects;
1899 	nvme_rdma_reconnect_or_remove(ctrl);
1900 }
1901 
1902 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1903 	.name			= "rdma",
1904 	.module			= THIS_MODULE,
1905 	.flags			= NVME_F_FABRICS,
1906 	.reg_read32		= nvmf_reg_read32,
1907 	.reg_read64		= nvmf_reg_read64,
1908 	.reg_write32		= nvmf_reg_write32,
1909 	.free_ctrl		= nvme_rdma_free_ctrl,
1910 	.submit_async_event	= nvme_rdma_submit_async_event,
1911 	.delete_ctrl		= nvme_rdma_delete_ctrl,
1912 	.get_address		= nvmf_get_address,
1913 };
1914 
1915 /*
1916  * Fails a connection request if it matches an existing controller
1917  * (association) with the same tuple:
1918  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1919  *
1920  * if local address is not specified in the request, it will match an
1921  * existing controller with all the other parameters the same and no
1922  * local port address specified as well.
1923  *
1924  * The ports don't need to be compared as they are intrinsically
1925  * already matched by the port pointers supplied.
1926  */
1927 static bool
1928 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1929 {
1930 	struct nvme_rdma_ctrl *ctrl;
1931 	bool found = false;
1932 
1933 	mutex_lock(&nvme_rdma_ctrl_mutex);
1934 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1935 		found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1936 		if (found)
1937 			break;
1938 	}
1939 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1940 
1941 	return found;
1942 }
1943 
1944 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1945 		struct nvmf_ctrl_options *opts)
1946 {
1947 	struct nvme_rdma_ctrl *ctrl;
1948 	int ret;
1949 	bool changed;
1950 
1951 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1952 	if (!ctrl)
1953 		return ERR_PTR(-ENOMEM);
1954 	ctrl->ctrl.opts = opts;
1955 	INIT_LIST_HEAD(&ctrl->list);
1956 
1957 	if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1958 		opts->trsvcid =
1959 			kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1960 		if (!opts->trsvcid) {
1961 			ret = -ENOMEM;
1962 			goto out_free_ctrl;
1963 		}
1964 		opts->mask |= NVMF_OPT_TRSVCID;
1965 	}
1966 
1967 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1968 			opts->traddr, opts->trsvcid, &ctrl->addr);
1969 	if (ret) {
1970 		pr_err("malformed address passed: %s:%s\n",
1971 			opts->traddr, opts->trsvcid);
1972 		goto out_free_ctrl;
1973 	}
1974 
1975 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1976 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1977 			opts->host_traddr, NULL, &ctrl->src_addr);
1978 		if (ret) {
1979 			pr_err("malformed src address passed: %s\n",
1980 			       opts->host_traddr);
1981 			goto out_free_ctrl;
1982 		}
1983 	}
1984 
1985 	if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1986 		ret = -EALREADY;
1987 		goto out_free_ctrl;
1988 	}
1989 
1990 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
1991 			nvme_rdma_reconnect_ctrl_work);
1992 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1993 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1994 
1995 	ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
1996 				opts->nr_poll_queues + 1;
1997 	ctrl->ctrl.sqsize = opts->queue_size - 1;
1998 	ctrl->ctrl.kato = opts->kato;
1999 
2000 	ret = -ENOMEM;
2001 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2002 				GFP_KERNEL);
2003 	if (!ctrl->queues)
2004 		goto out_free_ctrl;
2005 
2006 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2007 				0 /* no quirks, we're perfect! */);
2008 	if (ret)
2009 		goto out_kfree_queues;
2010 
2011 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2012 	WARN_ON_ONCE(!changed);
2013 
2014 	ret = nvme_rdma_setup_ctrl(ctrl, true);
2015 	if (ret)
2016 		goto out_uninit_ctrl;
2017 
2018 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2019 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2020 
2021 	nvme_get_ctrl(&ctrl->ctrl);
2022 
2023 	mutex_lock(&nvme_rdma_ctrl_mutex);
2024 	list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2025 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2026 
2027 	return &ctrl->ctrl;
2028 
2029 out_uninit_ctrl:
2030 	nvme_uninit_ctrl(&ctrl->ctrl);
2031 	nvme_put_ctrl(&ctrl->ctrl);
2032 	if (ret > 0)
2033 		ret = -EIO;
2034 	return ERR_PTR(ret);
2035 out_kfree_queues:
2036 	kfree(ctrl->queues);
2037 out_free_ctrl:
2038 	kfree(ctrl);
2039 	return ERR_PTR(ret);
2040 }
2041 
2042 static struct nvmf_transport_ops nvme_rdma_transport = {
2043 	.name		= "rdma",
2044 	.module		= THIS_MODULE,
2045 	.required_opts	= NVMF_OPT_TRADDR,
2046 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2047 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2048 			  NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
2049 	.create_ctrl	= nvme_rdma_create_ctrl,
2050 };
2051 
2052 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2053 {
2054 	struct nvme_rdma_ctrl *ctrl;
2055 	struct nvme_rdma_device *ndev;
2056 	bool found = false;
2057 
2058 	mutex_lock(&device_list_mutex);
2059 	list_for_each_entry(ndev, &device_list, entry) {
2060 		if (ndev->dev == ib_device) {
2061 			found = true;
2062 			break;
2063 		}
2064 	}
2065 	mutex_unlock(&device_list_mutex);
2066 
2067 	if (!found)
2068 		return;
2069 
2070 	/* Delete all controllers using this device */
2071 	mutex_lock(&nvme_rdma_ctrl_mutex);
2072 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2073 		if (ctrl->device->dev != ib_device)
2074 			continue;
2075 		nvme_delete_ctrl(&ctrl->ctrl);
2076 	}
2077 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2078 
2079 	flush_workqueue(nvme_delete_wq);
2080 }
2081 
2082 static struct ib_client nvme_rdma_ib_client = {
2083 	.name   = "nvme_rdma",
2084 	.remove = nvme_rdma_remove_one
2085 };
2086 
2087 static int __init nvme_rdma_init_module(void)
2088 {
2089 	int ret;
2090 
2091 	ret = ib_register_client(&nvme_rdma_ib_client);
2092 	if (ret)
2093 		return ret;
2094 
2095 	ret = nvmf_register_transport(&nvme_rdma_transport);
2096 	if (ret)
2097 		goto err_unreg_client;
2098 
2099 	return 0;
2100 
2101 err_unreg_client:
2102 	ib_unregister_client(&nvme_rdma_ib_client);
2103 	return ret;
2104 }
2105 
2106 static void __exit nvme_rdma_cleanup_module(void)
2107 {
2108 	nvmf_unregister_transport(&nvme_rdma_transport);
2109 	ib_unregister_client(&nvme_rdma_ib_client);
2110 }
2111 
2112 module_init(nvme_rdma_init_module);
2113 module_exit(nvme_rdma_cleanup_module);
2114 
2115 MODULE_LICENSE("GPL v2");
2116