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-integrity.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 <linux/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_CM_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	int cq_size;
100	struct mutex queue_lock;
101	};
102
103	struct nvme_rdma_ctrl {
104	/* read only in the hot path */
105	struct nvme_rdma_queue *queues;
106
107	/* other member variables */
108	struct blk_mq_tag_set tag_set;
109	struct work_struct err_work;
110
111	struct nvme_rdma_qe async_event_sqe;
112
113	struct delayed_work reconnect_work;
114
115	struct list_head list;
116
117	struct blk_mq_tag_set admin_tag_set;
118	struct nvme_rdma_device *device;
119
120	u32 max_fr_pages;
121
122	struct sockaddr_storage addr;
123	struct sockaddr_storage src_addr;
124
125	struct nvme_ctrl ctrl;
126	bool use_inline_data;
127	u32 io_queues[HCTX_MAX_TYPES];
128	};
129
130	static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
131	{
132	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
133	}
134
135	static LIST_HEAD(device_list);
136	static DEFINE_MUTEX(device_list_mutex);
137
138	static LIST_HEAD(nvme_rdma_ctrl_list);
139	static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
140
141	/*
142	* Disabling this option makes small I/O goes faster, but is fundamentally
143	* unsafe. With it turned off we will have to register a global rkey that
144	* allows read and write access to all physical memory.
145	*/
146	static bool register_always = true;
147	module_param(register_always, bool, 0444);
148	MODULE_PARM_DESC(register_always,
149	"Use memory registration even for contiguous memory regions");
150
151	static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
152	struct rdma_cm_event *event);
153	static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
154	static void nvme_rdma_complete_rq(struct request *rq);
155
156	static const struct blk_mq_ops nvme_rdma_mq_ops;
157	static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
158
159	static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
160	{
161	return queue - queue->ctrl->queues;
162	}
163
164	static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
165	{
166	return nvme_rdma_queue_idx(queue) >
167	queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
168	queue->ctrl->io_queues[HCTX_TYPE_READ];
169	}
170
171	static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
172	{
173	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
174	}
175
176	static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
177	size_t capsule_size, enum dma_data_direction dir)
178	{
179	ib_dma_unmap_single(dev: ibdev, addr: qe->dma, size: capsule_size, direction: dir);
180	kfree(objp: qe->data);
181	}
182
183	static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
184	size_t capsule_size, enum dma_data_direction dir)
185	{
186	qe->data = kzalloc(capsule_size, GFP_KERNEL);
187	if (!qe->data)
188	return -ENOMEM;
189
190	qe->dma = ib_dma_map_single(dev: ibdev, cpu_addr: qe->data, size: capsule_size, direction: dir);
191	if (ib_dma_mapping_error(dev: ibdev, dma_addr: qe->dma)) {
192	kfree(objp: qe->data);
193	qe->data = NULL;
194	return -ENOMEM;
195	}
196
197	return 0;
198	}
199
200	static void nvme_rdma_free_ring(struct ib_device *ibdev,
201	struct nvme_rdma_qe *ring, size_t ib_queue_size,
202	size_t capsule_size, enum dma_data_direction dir)
203	{
204	int i;
205
206	for (i = 0; i < ib_queue_size; i++)
207	nvme_rdma_free_qe(ibdev, qe: &ring[i], capsule_size, dir);
208	kfree(objp: ring);
209	}
210
211	static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
212	size_t ib_queue_size, size_t capsule_size,
213	enum dma_data_direction dir)
214	{
215	struct nvme_rdma_qe *ring;
216	int i;
217
218	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
219	if (!ring)
220	return NULL;
221
222	/*
223	* Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
224	* lifetime. It's safe, since any change in the underlying RDMA device
225	* will issue error recovery and queue re-creation.
226	*/
227	for (i = 0; i < ib_queue_size; i++) {
228	if (nvme_rdma_alloc_qe(ibdev, qe: &ring[i], capsule_size, dir))
229	goto out_free_ring;
230	}
231
232	return ring;
233
234	out_free_ring:
235	nvme_rdma_free_ring(ibdev, ring, ib_queue_size: i, capsule_size, dir);
236	return NULL;
237	}
238
239	static void nvme_rdma_qp_event(struct ib_event *event, void *context)
240	{
241	pr_debug("QP event %s (%d)\n",
242	ib_event_msg(event->event), event->event);
243
244	}
245
246	static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
247	{
248	int ret;
249
250	ret = wait_for_completion_interruptible(x: &queue->cm_done);
251	if (ret)
252	return ret;
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	init_attr.qp_context = queue;
278
279	ret = rdma_create_qp(id: queue->cm_id, pd: dev->pd, qp_init_attr: &init_attr);
280
281	queue->qp = queue->cm_id->qp;
282	return ret;
283	}
284
285	static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
286	struct request *rq, unsigned int hctx_idx)
287	{
288	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
289
290	kfree(objp: req->sqe.data);
291	}
292
293	static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
294	struct request *rq, unsigned int hctx_idx,
295	unsigned int numa_node)
296	{
297	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: set->driver_data);
298	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
299	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
300	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
301
302	nvme_req(req: rq)->ctrl = &ctrl->ctrl;
303	req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
304	if (!req->sqe.data)
305	return -ENOMEM;
306
307	/* metadata nvme_rdma_sgl struct is located after command's data SGL */
308	if (queue->pi_support)
309	req->metadata_sgl = (void *)nvme_req(req: rq) +
310	sizeof(struct nvme_rdma_request) +
311	NVME_RDMA_DATA_SGL_SIZE;
312
313	req->queue = queue;
314	nvme_req(req: rq)->cmd = req->sqe.data;
315
316	return 0;
317	}
318
319	static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
320	unsigned int hctx_idx)
321	{
322	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: data);
323	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
324
325	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
326
327	hctx->driver_data = queue;
328	return 0;
329	}
330
331	static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
332	unsigned int hctx_idx)
333	{
334	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: data);
335	struct nvme_rdma_queue *queue = &ctrl->queues[0];
336
337	BUG_ON(hctx_idx != 0);
338
339	hctx->driver_data = queue;
340	return 0;
341	}
342
343	static void nvme_rdma_free_dev(struct kref *ref)
344	{
345	struct nvme_rdma_device *ndev =
346	container_of(ref, struct nvme_rdma_device, ref);
347
348	mutex_lock(&device_list_mutex);
349	list_del(entry: &ndev->entry);
350	mutex_unlock(lock: &device_list_mutex);
351
352	ib_dealloc_pd(pd: ndev->pd);
353	kfree(objp: ndev);
354	}
355
356	static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
357	{
358	kref_put(kref: &dev->ref, release: nvme_rdma_free_dev);
359	}
360
361	static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
362	{
363	return kref_get_unless_zero(kref: &dev->ref);
364	}
365
366	static struct nvme_rdma_device *
367	nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
368	{
369	struct nvme_rdma_device *ndev;
370
371	mutex_lock(&device_list_mutex);
372	list_for_each_entry(ndev, &device_list, entry) {
373	if (ndev->dev->node_guid == cm_id->device->node_guid &&
374	nvme_rdma_dev_get(dev: ndev))
375	goto out_unlock;
376	}
377
378	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
379	if (!ndev)
380	goto out_err;
381
382	ndev->dev = cm_id->device;
383	kref_init(kref: &ndev->ref);
384
385	ndev->pd = ib_alloc_pd(ndev->dev,
386	register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
387	if (IS_ERR(ptr: ndev->pd))
388	goto out_free_dev;
389
390	if (!(ndev->dev->attrs.device_cap_flags &
391	IB_DEVICE_MEM_MGT_EXTENSIONS)) {
392	dev_err(&ndev->dev->dev,
393	"Memory registrations not supported.\n");
394	goto out_free_pd;
395	}
396
397	ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
398	ndev->dev->attrs.max_send_sge - 1);
399	list_add(new: &ndev->entry, head: &device_list);
400	out_unlock:
401	mutex_unlock(lock: &device_list_mutex);
402	return ndev;
403
404	out_free_pd:
405	ib_dealloc_pd(pd: ndev->pd);
406	out_free_dev:
407	kfree(objp: ndev);
408	out_err:
409	mutex_unlock(lock: &device_list_mutex);
410	return NULL;
411	}
412
413	static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
414	{
415	if (nvme_rdma_poll_queue(queue))
416	ib_free_cq(cq: queue->ib_cq);
417	else
418	ib_cq_pool_put(cq: queue->ib_cq, nr_cqe: queue->cq_size);
419	}
420
421	static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
422	{
423	struct nvme_rdma_device *dev;
424	struct ib_device *ibdev;
425
426	if (!test_and_clear_bit(nr: NVME_RDMA_Q_TR_READY, addr: &queue->flags))
427	return;
428
429	dev = queue->device;
430	ibdev = dev->dev;
431
432	if (queue->pi_support)
433	ib_mr_pool_destroy(qp: queue->qp, list: &queue->qp->sig_mrs);
434	ib_mr_pool_destroy(qp: queue->qp, list: &queue->qp->rdma_mrs);
435
436	/*
437	* The cm_id object might have been destroyed during RDMA connection
438	* establishment error flow to avoid getting other cma events, thus
439	* the destruction of the QP shouldn't use rdma_cm API.
440	*/
441	ib_destroy_qp(qp: queue->qp);
442	nvme_rdma_free_cq(queue);
443
444	nvme_rdma_free_ring(ibdev, ring: queue->rsp_ring, ib_queue_size: queue->queue_size,
445	capsule_size: sizeof(struct nvme_completion), dir: DMA_FROM_DEVICE);
446
447	nvme_rdma_dev_put(dev);
448	}
449
450	static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
451	{
452	u32 max_page_list_len;
453
454	if (pi_support)
455	max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
456	else
457	max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
458
459	return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
460	}
461
462	static int nvme_rdma_create_cq(struct ib_device *ibdev,
463	struct nvme_rdma_queue *queue)
464	{
465	int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
466
467	/*
468	* Spread I/O queues completion vectors according their queue index.
469	* Admin queues can always go on completion vector 0.
470	*/
471	comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
472
473	/* Polling queues need direct cq polling context */
474	if (nvme_rdma_poll_queue(queue))
475	queue->ib_cq = ib_alloc_cq(dev: ibdev, private: queue, nr_cqe: queue->cq_size,
476	comp_vector, poll_ctx: IB_POLL_DIRECT);
477	else
478	queue->ib_cq = ib_cq_pool_get(dev: ibdev, nr_cqe: queue->cq_size,
479	comp_vector_hint: comp_vector, poll_ctx: IB_POLL_SOFTIRQ);
480
481	if (IS_ERR(ptr: queue->ib_cq)) {
482	ret = PTR_ERR(ptr: queue->ib_cq);
483	return ret;
484	}
485
486	return 0;
487	}
488
489	static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
490	{
491	struct ib_device *ibdev;
492	const int send_wr_factor = 3; /* MR, SEND, INV */
493	const int cq_factor = send_wr_factor + 1; /* + RECV */
494	int ret, pages_per_mr;
495
496	queue->device = nvme_rdma_find_get_device(cm_id: queue->cm_id);
497	if (!queue->device) {
498	dev_err(queue->cm_id->device->dev.parent,
499	"no client data found!\n");
500	return -ECONNREFUSED;
501	}
502	ibdev = queue->device->dev;
503
504	/* +1 for ib_drain_qp */
505	queue->cq_size = cq_factor * queue->queue_size + 1;
506
507	ret = nvme_rdma_create_cq(ibdev, queue);
508	if (ret)
509	goto out_put_dev;
510
511	ret = nvme_rdma_create_qp(queue, factor: send_wr_factor);
512	if (ret)
513	goto out_destroy_ib_cq;
514
515	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, ib_queue_size: queue->queue_size,
516	capsule_size: sizeof(struct nvme_completion), dir: DMA_FROM_DEVICE);
517	if (!queue->rsp_ring) {
518	ret = -ENOMEM;
519	goto out_destroy_qp;
520	}
521
522	/*
523	* Currently we don't use SG_GAPS MR's so if the first entry is
524	* misaligned we'll end up using two entries for a single data page,
525	* so one additional entry is required.
526	*/
527	pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, pi_support: queue->pi_support) + 1;
528	ret = ib_mr_pool_init(qp: queue->qp, list: &queue->qp->rdma_mrs,
529	nr: queue->queue_size,
530	type: IB_MR_TYPE_MEM_REG,
531	max_num_sg: pages_per_mr, max_num_meta_sg: 0);
532	if (ret) {
533	dev_err(queue->ctrl->ctrl.device,
534	"failed to initialize MR pool sized %d for QID %d\n",
535	queue->queue_size, nvme_rdma_queue_idx(queue));
536	goto out_destroy_ring;
537	}
538
539	if (queue->pi_support) {
540	ret = ib_mr_pool_init(qp: queue->qp, list: &queue->qp->sig_mrs,
541	nr: queue->queue_size, type: IB_MR_TYPE_INTEGRITY,
542	max_num_sg: pages_per_mr, max_num_meta_sg: pages_per_mr);
543	if (ret) {
544	dev_err(queue->ctrl->ctrl.device,
545	"failed to initialize PI MR pool sized %d for QID %d\n",
546	queue->queue_size, nvme_rdma_queue_idx(queue));
547	goto out_destroy_mr_pool;
548	}
549	}
550
551	set_bit(nr: NVME_RDMA_Q_TR_READY, addr: &queue->flags);
552
553	return 0;
554
555	out_destroy_mr_pool:
556	ib_mr_pool_destroy(qp: queue->qp, list: &queue->qp->rdma_mrs);
557	out_destroy_ring:
558	nvme_rdma_free_ring(ibdev, ring: queue->rsp_ring, ib_queue_size: queue->queue_size,
559	capsule_size: sizeof(struct nvme_completion), dir: DMA_FROM_DEVICE);
560	out_destroy_qp:
561	rdma_destroy_qp(id: queue->cm_id);
562	out_destroy_ib_cq:
563	nvme_rdma_free_cq(queue);
564	out_put_dev:
565	nvme_rdma_dev_put(dev: queue->device);
566	return ret;
567	}
568
569	static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
570	int idx, size_t queue_size)
571	{
572	struct nvme_rdma_queue *queue;
573	struct sockaddr *src_addr = NULL;
574	int ret;
575
576	queue = &ctrl->queues[idx];
577	mutex_init(&queue->queue_lock);
578	queue->ctrl = ctrl;
579	if (idx && ctrl->ctrl.max_integrity_segments)
580	queue->pi_support = true;
581	else
582	queue->pi_support = false;
583	init_completion(x: &queue->cm_done);
584
585	if (idx > 0)
586	queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
587	else
588	queue->cmnd_capsule_len = sizeof(struct nvme_command);
589
590	queue->queue_size = queue_size;
591
592	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
593	RDMA_PS_TCP, IB_QPT_RC);
594	if (IS_ERR(ptr: queue->cm_id)) {
595	dev_info(ctrl->ctrl.device,
596	"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
597	ret = PTR_ERR(ptr: queue->cm_id);
598	goto out_destroy_mutex;
599	}
600
601	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
602	src_addr = (struct sockaddr *)&ctrl->src_addr;
603
604	queue->cm_error = -ETIMEDOUT;
605	ret = rdma_resolve_addr(id: queue->cm_id, src_addr,
606	dst_addr: (struct sockaddr *)&ctrl->addr,
607	NVME_RDMA_CM_TIMEOUT_MS);
608	if (ret) {
609	dev_info(ctrl->ctrl.device,
610	"rdma_resolve_addr failed (%d).\n", ret);
611	goto out_destroy_cm_id;
612	}
613
614	ret = nvme_rdma_wait_for_cm(queue);
615	if (ret) {
616	dev_info(ctrl->ctrl.device,
617	"rdma connection establishment failed (%d)\n", ret);
618	goto out_destroy_cm_id;
619	}
620
621	set_bit(nr: NVME_RDMA_Q_ALLOCATED, addr: &queue->flags);
622
623	return 0;
624
625	out_destroy_cm_id:
626	rdma_destroy_id(id: queue->cm_id);
627	nvme_rdma_destroy_queue_ib(queue);
628	out_destroy_mutex:
629	mutex_destroy(lock: &queue->queue_lock);
630	return ret;
631	}
632
633	static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
634	{
635	rdma_disconnect(id: queue->cm_id);
636	ib_drain_qp(qp: queue->qp);
637	}
638
639	static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
640	{
641	if (!test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
642	return;
643
644	mutex_lock(&queue->queue_lock);
645	if (test_and_clear_bit(nr: NVME_RDMA_Q_LIVE, addr: &queue->flags))
646	__nvme_rdma_stop_queue(queue);
647	mutex_unlock(lock: &queue->queue_lock);
648	}
649
650	static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
651	{
652	if (!test_and_clear_bit(nr: NVME_RDMA_Q_ALLOCATED, addr: &queue->flags))
653	return;
654
655	rdma_destroy_id(id: queue->cm_id);
656	nvme_rdma_destroy_queue_ib(queue);
657	mutex_destroy(lock: &queue->queue_lock);
658	}
659
660	static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
661	{
662	int i;
663
664	for (i = 1; i < ctrl->ctrl.queue_count; i++)
665	nvme_rdma_free_queue(queue: &ctrl->queues[i]);
666	}
667
668	static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
669	{
670	int i;
671
672	for (i = 1; i < ctrl->ctrl.queue_count; i++)
673	nvme_rdma_stop_queue(queue: &ctrl->queues[i]);
674	}
675
676	static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
677	{
678	struct nvme_rdma_queue *queue = &ctrl->queues[idx];
679	int ret;
680
681	if (idx)
682	ret = nvmf_connect_io_queue(ctrl: &ctrl->ctrl, qid: idx);
683	else
684	ret = nvmf_connect_admin_queue(ctrl: &ctrl->ctrl);
685
686	if (!ret) {
687	set_bit(nr: NVME_RDMA_Q_LIVE, addr: &queue->flags);
688	} else {
689	if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
690	__nvme_rdma_stop_queue(queue);
691	dev_info(ctrl->ctrl.device,
692	"failed to connect queue: %d ret=%d\n", idx, ret);
693	}
694	return ret;
695	}
696
697	static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl,
698	int first, int last)
699	{
700	int i, ret = 0;
701
702	for (i = first; i < last; i++) {
703	ret = nvme_rdma_start_queue(ctrl, idx: i);
704	if (ret)
705	goto out_stop_queues;
706	}
707
708	return 0;
709
710	out_stop_queues:
711	for (i--; i >= first; i--)
712	nvme_rdma_stop_queue(queue: &ctrl->queues[i]);
713	return ret;
714	}
715
716	static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
717	{
718	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
719	unsigned int nr_io_queues;
720	int i, ret;
721
722	nr_io_queues = nvmf_nr_io_queues(opts);
723	ret = nvme_set_queue_count(ctrl: &ctrl->ctrl, count: &nr_io_queues);
724	if (ret)
725	return ret;
726
727	if (nr_io_queues == 0) {
728	dev_err(ctrl->ctrl.device,
729	"unable to set any I/O queues\n");
730	return -ENOMEM;
731	}
732
733	ctrl->ctrl.queue_count = nr_io_queues + 1;
734	dev_info(ctrl->ctrl.device,
735	"creating %d I/O queues.\n", nr_io_queues);
736
737	nvmf_set_io_queues(opts, nr_io_queues, io_queues: ctrl->io_queues);
738	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
739	ret = nvme_rdma_alloc_queue(ctrl, idx: i,
740	queue_size: ctrl->ctrl.sqsize + 1);
741	if (ret)
742	goto out_free_queues;
743	}
744
745	return 0;
746
747	out_free_queues:
748	for (i--; i >= 1; i--)
749	nvme_rdma_free_queue(queue: &ctrl->queues[i]);
750
751	return ret;
752	}
753
754	static int nvme_rdma_alloc_tag_set(struct nvme_ctrl *ctrl)
755	{
756	unsigned int cmd_size = sizeof(struct nvme_rdma_request) +
757	NVME_RDMA_DATA_SGL_SIZE;
758
759	if (ctrl->max_integrity_segments)
760	cmd_size += sizeof(struct nvme_rdma_sgl) +
761	NVME_RDMA_METADATA_SGL_SIZE;
762
763	return nvme_alloc_io_tag_set(ctrl, set: &to_rdma_ctrl(ctrl)->tag_set,
764	ops: &nvme_rdma_mq_ops,
765	nr_maps: ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
766	cmd_size);
767	}
768
769	static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
770	{
771	if (ctrl->async_event_sqe.data) {
772	cancel_work_sync(work: &ctrl->ctrl.async_event_work);
773	nvme_rdma_free_qe(ibdev: ctrl->device->dev, qe: &ctrl->async_event_sqe,
774	capsule_size: sizeof(struct nvme_command), dir: DMA_TO_DEVICE);
775	ctrl->async_event_sqe.data = NULL;
776	}
777	nvme_rdma_free_queue(queue: &ctrl->queues[0]);
778	}
779
780	static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
781	bool new)
782	{
783	bool pi_capable = false;
784	int error;
785
786	error = nvme_rdma_alloc_queue(ctrl, idx: 0, NVME_AQ_DEPTH);
787	if (error)
788	return error;
789
790	ctrl->device = ctrl->queues[0].device;
791	ctrl->ctrl.numa_node = ibdev_to_node(ibdev: ctrl->device->dev);
792
793	/* T10-PI support */
794	if (ctrl->device->dev->attrs.kernel_cap_flags &
795	IBK_INTEGRITY_HANDOVER)
796	pi_capable = true;
797
798	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ibdev: ctrl->device->dev,
799	pi_support: pi_capable);
800
801	/*
802	* Bind the async event SQE DMA mapping to the admin queue lifetime.
803	* It's safe, since any change in the underlying RDMA device will issue
804	* error recovery and queue re-creation.
805	*/
806	error = nvme_rdma_alloc_qe(ibdev: ctrl->device->dev, qe: &ctrl->async_event_sqe,
807	capsule_size: sizeof(struct nvme_command), dir: DMA_TO_DEVICE);
808	if (error)
809	goto out_free_queue;
810
811	if (new) {
812	error = nvme_alloc_admin_tag_set(ctrl: &ctrl->ctrl,
813	set: &ctrl->admin_tag_set, ops: &nvme_rdma_admin_mq_ops,
814	cmd_size: sizeof(struct nvme_rdma_request) +
815	NVME_RDMA_DATA_SGL_SIZE);
816	if (error)
817	goto out_free_async_qe;
818
819	}
820
821	error = nvme_rdma_start_queue(ctrl, idx: 0);
822	if (error)
823	goto out_remove_admin_tag_set;
824
825	error = nvme_enable_ctrl(ctrl: &ctrl->ctrl);
826	if (error)
827	goto out_stop_queue;
828
829	ctrl->ctrl.max_segments = ctrl->max_fr_pages;
830	ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
831	if (pi_capable)
832	ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
833	else
834	ctrl->ctrl.max_integrity_segments = 0;
835
836	nvme_unquiesce_admin_queue(ctrl: &ctrl->ctrl);
837
838	error = nvme_init_ctrl_finish(ctrl: &ctrl->ctrl, was_suspended: false);
839	if (error)
840	goto out_quiesce_queue;
841
842	return 0;
843
844	out_quiesce_queue:
845	nvme_quiesce_admin_queue(ctrl: &ctrl->ctrl);
846	blk_sync_queue(q: ctrl->ctrl.admin_q);
847	out_stop_queue:
848	nvme_rdma_stop_queue(queue: &ctrl->queues[0]);
849	nvme_cancel_admin_tagset(ctrl: &ctrl->ctrl);
850	out_remove_admin_tag_set:
851	if (new)
852	nvme_remove_admin_tag_set(ctrl: &ctrl->ctrl);
853	out_free_async_qe:
854	if (ctrl->async_event_sqe.data) {
855	nvme_rdma_free_qe(ibdev: ctrl->device->dev, qe: &ctrl->async_event_sqe,
856	capsule_size: sizeof(struct nvme_command), dir: DMA_TO_DEVICE);
857	ctrl->async_event_sqe.data = NULL;
858	}
859	out_free_queue:
860	nvme_rdma_free_queue(queue: &ctrl->queues[0]);
861	return error;
862	}
863
864	static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
865	{
866	int ret, nr_queues;
867
868	ret = nvme_rdma_alloc_io_queues(ctrl);
869	if (ret)
870	return ret;
871
872	if (new) {
873	ret = nvme_rdma_alloc_tag_set(ctrl: &ctrl->ctrl);
874	if (ret)
875	goto out_free_io_queues;
876	}
877
878	/*
879	* Only start IO queues for which we have allocated the tagset
880	* and limitted it to the available queues. On reconnects, the
881	* queue number might have changed.
882	*/
883	nr_queues = min(ctrl->tag_set.nr_hw_queues + 1, ctrl->ctrl.queue_count);
884	ret = nvme_rdma_start_io_queues(ctrl, first: 1, last: nr_queues);
885	if (ret)
886	goto out_cleanup_tagset;
887
888	if (!new) {
889	nvme_start_freeze(ctrl: &ctrl->ctrl);
890	nvme_unquiesce_io_queues(ctrl: &ctrl->ctrl);
891	if (!nvme_wait_freeze_timeout(ctrl: &ctrl->ctrl, NVME_IO_TIMEOUT)) {
892	/*
893	* If we timed out waiting for freeze we are likely to
894	* be stuck. Fail the controller initialization just
895	* to be safe.
896	*/
897	ret = -ENODEV;
898	nvme_unfreeze(ctrl: &ctrl->ctrl);
899	goto out_wait_freeze_timed_out;
900	}
901	blk_mq_update_nr_hw_queues(set: ctrl->ctrl.tagset,
902	nr_hw_queues: ctrl->ctrl.queue_count - 1);
903	nvme_unfreeze(ctrl: &ctrl->ctrl);
904	}
905
906	/*
907	* If the number of queues has increased (reconnect case)
908	* start all new queues now.
909	*/
910	ret = nvme_rdma_start_io_queues(ctrl, first: nr_queues,
911	last: ctrl->tag_set.nr_hw_queues + 1);
912	if (ret)
913	goto out_wait_freeze_timed_out;
914
915	return 0;
916
917	out_wait_freeze_timed_out:
918	nvme_quiesce_io_queues(ctrl: &ctrl->ctrl);
919	nvme_sync_io_queues(ctrl: &ctrl->ctrl);
920	nvme_rdma_stop_io_queues(ctrl);
921	out_cleanup_tagset:
922	nvme_cancel_tagset(ctrl: &ctrl->ctrl);
923	if (new)
924	nvme_remove_io_tag_set(ctrl: &ctrl->ctrl);
925	out_free_io_queues:
926	nvme_rdma_free_io_queues(ctrl);
927	return ret;
928	}
929
930	static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
931	bool remove)
932	{
933	nvme_quiesce_admin_queue(ctrl: &ctrl->ctrl);
934	blk_sync_queue(q: ctrl->ctrl.admin_q);
935	nvme_rdma_stop_queue(queue: &ctrl->queues[0]);
936	nvme_cancel_admin_tagset(ctrl: &ctrl->ctrl);
937	if (remove) {
938	nvme_unquiesce_admin_queue(ctrl: &ctrl->ctrl);
939	nvme_remove_admin_tag_set(ctrl: &ctrl->ctrl);
940	}
941	nvme_rdma_destroy_admin_queue(ctrl);
942	}
943
944	static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
945	bool remove)
946	{
947	if (ctrl->ctrl.queue_count > 1) {
948	nvme_quiesce_io_queues(ctrl: &ctrl->ctrl);
949	nvme_sync_io_queues(ctrl: &ctrl->ctrl);
950	nvme_rdma_stop_io_queues(ctrl);
951	nvme_cancel_tagset(ctrl: &ctrl->ctrl);
952	if (remove) {
953	nvme_unquiesce_io_queues(ctrl: &ctrl->ctrl);
954	nvme_remove_io_tag_set(ctrl: &ctrl->ctrl);
955	}
956	nvme_rdma_free_io_queues(ctrl);
957	}
958	}
959
960	static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
961	{
962	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: nctrl);
963
964	flush_work(work: &ctrl->err_work);
965	cancel_delayed_work_sync(dwork: &ctrl->reconnect_work);
966	}
967
968	static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
969	{
970	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: nctrl);
971
972	if (list_empty(head: &ctrl->list))
973	goto free_ctrl;
974
975	mutex_lock(&nvme_rdma_ctrl_mutex);
976	list_del(entry: &ctrl->list);
977	mutex_unlock(lock: &nvme_rdma_ctrl_mutex);
978
979	nvmf_free_options(opts: nctrl->opts);
980	free_ctrl:
981	kfree(objp: ctrl->queues);
982	kfree(objp: ctrl);
983	}
984
985	static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl,
986	int status)
987	{
988	enum nvme_ctrl_state state = nvme_ctrl_state(ctrl: &ctrl->ctrl);
989
990	/* If we are resetting/deleting then do nothing */
991	if (state != NVME_CTRL_CONNECTING) {
992	WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE);
993	return;
994	}
995
996	if (nvmf_should_reconnect(ctrl: &ctrl->ctrl, status)) {
997	dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
998	ctrl->ctrl.opts->reconnect_delay);
999	queue_delayed_work(wq: nvme_wq, dwork: &ctrl->reconnect_work,
1000	delay: ctrl->ctrl.opts->reconnect_delay * HZ);
1001	} else {
1002	nvme_delete_ctrl(ctrl: &ctrl->ctrl);
1003	}
1004	}
1005
1006	static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1007	{
1008	int ret;
1009	bool changed;
1010	u16 max_queue_size;
1011
1012	ret = nvme_rdma_configure_admin_queue(ctrl, new);
1013	if (ret)
1014	return ret;
1015
1016	if (ctrl->ctrl.icdoff) {
1017	ret = -EOPNOTSUPP;
1018	dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1019	goto destroy_admin;
1020	}
1021
1022	if (!(ctrl->ctrl.sgls & NVME_CTRL_SGLS_KSDBDS)) {
1023	ret = -EOPNOTSUPP;
1024	dev_err(ctrl->ctrl.device,
1025	"Mandatory keyed sgls are not supported!\n");
1026	goto destroy_admin;
1027	}
1028
1029	if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1030	dev_warn(ctrl->ctrl.device,
1031	"queue_size %zu > ctrl sqsize %u, clamping down\n",
1032	ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1033	}
1034
1035	if (ctrl->ctrl.max_integrity_segments)
1036	max_queue_size = NVME_RDMA_MAX_METADATA_QUEUE_SIZE;
1037	else
1038	max_queue_size = NVME_RDMA_MAX_QUEUE_SIZE;
1039
1040	if (ctrl->ctrl.sqsize + 1 > max_queue_size) {
1041	dev_warn(ctrl->ctrl.device,
1042	"ctrl sqsize %u > max queue size %u, clamping down\n",
1043	ctrl->ctrl.sqsize + 1, max_queue_size);
1044	ctrl->ctrl.sqsize = max_queue_size - 1;
1045	}
1046
1047	if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1048	dev_warn(ctrl->ctrl.device,
1049	"sqsize %u > ctrl maxcmd %u, clamping down\n",
1050	ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1051	ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1052	}
1053
1054	if (ctrl->ctrl.sgls & NVME_CTRL_SGLS_SAOS)
1055	ctrl->use_inline_data = true;
1056
1057	if (ctrl->ctrl.queue_count > 1) {
1058	ret = nvme_rdma_configure_io_queues(ctrl, new);
1059	if (ret)
1060	goto destroy_admin;
1061	}
1062
1063	changed = nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_LIVE);
1064	if (!changed) {
1065	/*
1066	* state change failure is ok if we started ctrl delete,
1067	* unless we're during creation of a new controller to
1068	* avoid races with teardown flow.
1069	*/
1070	enum nvme_ctrl_state state = nvme_ctrl_state(ctrl: &ctrl->ctrl);
1071
1072	WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1073	state != NVME_CTRL_DELETING_NOIO);
1074	WARN_ON_ONCE(new);
1075	ret = -EINVAL;
1076	goto destroy_io;
1077	}
1078
1079	nvme_start_ctrl(ctrl: &ctrl->ctrl);
1080	return 0;
1081
1082	destroy_io:
1083	if (ctrl->ctrl.queue_count > 1) {
1084	nvme_quiesce_io_queues(ctrl: &ctrl->ctrl);
1085	nvme_sync_io_queues(ctrl: &ctrl->ctrl);
1086	nvme_rdma_stop_io_queues(ctrl);
1087	nvme_cancel_tagset(ctrl: &ctrl->ctrl);
1088	if (new)
1089	nvme_remove_io_tag_set(ctrl: &ctrl->ctrl);
1090	nvme_rdma_free_io_queues(ctrl);
1091	}
1092	destroy_admin:
1093	nvme_stop_keep_alive(ctrl: &ctrl->ctrl);
1094	nvme_rdma_teardown_admin_queue(ctrl, remove: new);
1095	return ret;
1096	}
1097
1098	static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1099	{
1100	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1101	struct nvme_rdma_ctrl, reconnect_work);
1102	int ret;
1103
1104	++ctrl->ctrl.nr_reconnects;
1105
1106	ret = nvme_rdma_setup_ctrl(ctrl, new: false);
1107	if (ret)
1108	goto requeue;
1109
1110	dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1111	ctrl->ctrl.nr_reconnects);
1112
1113	ctrl->ctrl.nr_reconnects = 0;
1114
1115	return;
1116
1117	requeue:
1118	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d/%d\n",
1119	ctrl->ctrl.nr_reconnects, ctrl->ctrl.opts->max_reconnects);
1120	nvme_rdma_reconnect_or_remove(ctrl, status: ret);
1121	}
1122
1123	static void nvme_rdma_error_recovery_work(struct work_struct *work)
1124	{
1125	struct nvme_rdma_ctrl *ctrl = container_of(work,
1126	struct nvme_rdma_ctrl, err_work);
1127
1128	nvme_stop_keep_alive(ctrl: &ctrl->ctrl);
1129	flush_work(work: &ctrl->ctrl.async_event_work);
1130	nvme_rdma_teardown_io_queues(ctrl, remove: false);
1131	nvme_unquiesce_io_queues(ctrl: &ctrl->ctrl);
1132	nvme_rdma_teardown_admin_queue(ctrl, remove: false);
1133	nvme_unquiesce_admin_queue(ctrl: &ctrl->ctrl);
1134	nvme_auth_stop(ctrl: &ctrl->ctrl);
1135
1136	if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_CONNECTING)) {
1137	/* state change failure is ok if we started ctrl delete */
1138	enum nvme_ctrl_state state = nvme_ctrl_state(ctrl: &ctrl->ctrl);
1139
1140	WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1141	state != NVME_CTRL_DELETING_NOIO);
1142	return;
1143	}
1144
1145	nvme_rdma_reconnect_or_remove(ctrl, status: 0);
1146	}
1147
1148	static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1149	{
1150	if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_RESETTING))
1151	return;
1152
1153	dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1154	queue_work(wq: nvme_reset_wq, work: &ctrl->err_work);
1155	}
1156
1157	static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1158	{
1159	struct request *rq = blk_mq_rq_from_pdu(pdu: req);
1160
1161	if (!refcount_dec_and_test(r: &req->ref))
1162	return;
1163	if (!nvme_try_complete_req(req: rq, status: req->status, result: req->result))
1164	nvme_rdma_complete_rq(rq);
1165	}
1166
1167	static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1168	const char *op)
1169	{
1170	struct nvme_rdma_queue *queue = wc->qp->qp_context;
1171	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1172
1173	if (nvme_ctrl_state(ctrl: &ctrl->ctrl) == NVME_CTRL_LIVE)
1174	dev_info(ctrl->ctrl.device,
1175	"%s for CQE 0x%p failed with status %s (%d)\n",
1176	op, wc->wr_cqe,
1177	ib_wc_status_msg(wc->status), wc->status);
1178	nvme_rdma_error_recovery(ctrl);
1179	}
1180
1181	static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1182	{
1183	if (unlikely(wc->status != IB_WC_SUCCESS))
1184	nvme_rdma_wr_error(cq, wc, op: "MEMREG");
1185	}
1186
1187	static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1188	{
1189	struct nvme_rdma_request *req =
1190	container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1191
1192	if (unlikely(wc->status != IB_WC_SUCCESS))
1193	nvme_rdma_wr_error(cq, wc, op: "LOCAL_INV");
1194	else
1195	nvme_rdma_end_request(req);
1196	}
1197
1198	static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1199	struct nvme_rdma_request *req)
1200	{
1201	struct ib_send_wr wr = {
1202	.opcode = IB_WR_LOCAL_INV,
1203	.next = NULL,
1204	.num_sge = 0,
1205	.send_flags = IB_SEND_SIGNALED,
1206	.ex.invalidate_rkey = req->mr->rkey,
1207	};
1208
1209	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1210	wr.wr_cqe = &req->reg_cqe;
1211
1212	return ib_post_send(qp: queue->qp, send_wr: &wr, NULL);
1213	}
1214
1215	static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1216	{
1217	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1218
1219	if (blk_integrity_rq(rq)) {
1220	ib_dma_unmap_sg(dev: ibdev, sg: req->metadata_sgl->sg_table.sgl,
1221	nents: req->metadata_sgl->nents, rq_dma_dir(rq));
1222	sg_free_table_chained(table: &req->metadata_sgl->sg_table,
1223	NVME_INLINE_METADATA_SG_CNT);
1224	}
1225
1226	ib_dma_unmap_sg(dev: ibdev, sg: req->data_sgl.sg_table.sgl, nents: req->data_sgl.nents,
1227	rq_dma_dir(rq));
1228	sg_free_table_chained(table: &req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1229	}
1230
1231	static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1232	struct request *rq)
1233	{
1234	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1235	struct nvme_rdma_device *dev = queue->device;
1236	struct ib_device *ibdev = dev->dev;
1237	struct list_head *pool = &queue->qp->rdma_mrs;
1238
1239	if (!blk_rq_nr_phys_segments(rq))
1240	return;
1241
1242	if (req->use_sig_mr)
1243	pool = &queue->qp->sig_mrs;
1244
1245	if (req->mr) {
1246	ib_mr_pool_put(qp: queue->qp, list: pool, mr: req->mr);
1247	req->mr = NULL;
1248	}
1249
1250	nvme_rdma_dma_unmap_req(ibdev, rq);
1251	}
1252
1253	static int nvme_rdma_set_sg_null(struct nvme_command *c)
1254	{
1255	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1256
1257	sg->addr = 0;
1258	put_unaligned_le24(val: 0, p: sg->length);
1259	put_unaligned_le32(val: 0, p: sg->key);
1260	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1261	return 0;
1262	}
1263
1264	static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1265	struct nvme_rdma_request *req, struct nvme_command *c,
1266	int count)
1267	{
1268	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1269	struct ib_sge *sge = &req->sge[1];
1270	struct scatterlist *sgl;
1271	u32 len = 0;
1272	int i;
1273
1274	for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1275	sge->addr = sg_dma_address(sgl);
1276	sge->length = sg_dma_len(sgl);
1277	sge->lkey = queue->device->pd->local_dma_lkey;
1278	len += sge->length;
1279	sge++;
1280	}
1281
1282	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1283	sg->length = cpu_to_le32(len);
1284	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1285
1286	req->num_sge += count;
1287	return 0;
1288	}
1289
1290	static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1291	struct nvme_rdma_request *req, struct nvme_command *c)
1292	{
1293	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1294
1295	sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1296	put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), p: sg->length);
1297	put_unaligned_le32(val: queue->device->pd->unsafe_global_rkey, p: sg->key);
1298	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1299	return 0;
1300	}
1301
1302	static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1303	struct nvme_rdma_request *req, struct nvme_command *c,
1304	int count)
1305	{
1306	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1307	int nr;
1308
1309	req->mr = ib_mr_pool_get(qp: queue->qp, list: &queue->qp->rdma_mrs);
1310	if (WARN_ON_ONCE(!req->mr))
1311	return -EAGAIN;
1312
1313	/*
1314	* Align the MR to a 4K page size to match the ctrl page size and
1315	* the block virtual boundary.
1316	*/
1317	nr = ib_map_mr_sg(mr: req->mr, sg: req->data_sgl.sg_table.sgl, sg_nents: count, NULL,
1318	SZ_4K);
1319	if (unlikely(nr < count)) {
1320	ib_mr_pool_put(qp: queue->qp, list: &queue->qp->rdma_mrs, mr: req->mr);
1321	req->mr = NULL;
1322	if (nr < 0)
1323	return nr;
1324	return -EINVAL;
1325	}
1326
1327	ib_update_fast_reg_key(mr: req->mr, newkey: ib_inc_rkey(rkey: req->mr->rkey));
1328
1329	req->reg_cqe.done = nvme_rdma_memreg_done;
1330	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1331	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1332	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1333	req->reg_wr.wr.num_sge = 0;
1334	req->reg_wr.mr = req->mr;
1335	req->reg_wr.key = req->mr->rkey;
1336	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1337	IB_ACCESS_REMOTE_READ |
1338	IB_ACCESS_REMOTE_WRITE;
1339
1340	sg->addr = cpu_to_le64(req->mr->iova);
1341	put_unaligned_le24(val: req->mr->length, p: sg->length);
1342	put_unaligned_le32(val: req->mr->rkey, p: sg->key);
1343	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1344	NVME_SGL_FMT_INVALIDATE;
1345
1346	return 0;
1347	}
1348
1349	static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1350	struct nvme_command *cmd, struct ib_sig_domain *domain,
1351	u16 control, u8 pi_type)
1352	{
1353	domain->sig_type = IB_SIG_TYPE_T10_DIF;
1354	domain->sig.dif.bg_type = IB_T10DIF_CRC;
1355	domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1356	domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1357	if (control & NVME_RW_PRINFO_PRCHK_REF)
1358	domain->sig.dif.ref_remap = true;
1359
1360	domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.lbat);
1361	domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.lbatm);
1362	domain->sig.dif.app_escape = true;
1363	if (pi_type == NVME_NS_DPS_PI_TYPE3)
1364	domain->sig.dif.ref_escape = true;
1365	}
1366
1367	static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1368	struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1369	u8 pi_type)
1370	{
1371	u16 control = le16_to_cpu(cmd->rw.control);
1372
1373	memset(sig_attrs, 0, sizeof(*sig_attrs));
1374	if (control & NVME_RW_PRINFO_PRACT) {
1375	/* for WRITE_INSERT/READ_STRIP no memory domain */
1376	sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1377	nvme_rdma_set_sig_domain(bi, cmd, domain: &sig_attrs->wire, control,
1378	pi_type);
1379	/* Clear the PRACT bit since HCA will generate/verify the PI */
1380	control &= ~NVME_RW_PRINFO_PRACT;
1381	cmd->rw.control = cpu_to_le16(control);
1382	} else {
1383	/* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1384	nvme_rdma_set_sig_domain(bi, cmd, domain: &sig_attrs->wire, control,
1385	pi_type);
1386	nvme_rdma_set_sig_domain(bi, cmd, domain: &sig_attrs->mem, control,
1387	pi_type);
1388	}
1389	}
1390
1391	static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1392	{
1393	*mask = 0;
1394	if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1395	*mask |= IB_SIG_CHECK_REFTAG;
1396	if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1397	*mask |= IB_SIG_CHECK_GUARD;
1398	}
1399
1400	static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1401	{
1402	if (unlikely(wc->status != IB_WC_SUCCESS))
1403	nvme_rdma_wr_error(cq, wc, op: "SIG");
1404	}
1405
1406	static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1407	struct nvme_rdma_request *req, struct nvme_command *c,
1408	int count, int pi_count)
1409	{
1410	struct nvme_rdma_sgl *sgl = &req->data_sgl;
1411	struct ib_reg_wr *wr = &req->reg_wr;
1412	struct request *rq = blk_mq_rq_from_pdu(pdu: req);
1413	struct nvme_ns *ns = rq->q->queuedata;
1414	struct bio *bio = rq->bio;
1415	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1416	struct blk_integrity *bi = blk_get_integrity(disk: bio->bi_bdev->bd_disk);
1417	u32 xfer_len;
1418	int nr;
1419
1420	req->mr = ib_mr_pool_get(qp: queue->qp, list: &queue->qp->sig_mrs);
1421	if (WARN_ON_ONCE(!req->mr))
1422	return -EAGAIN;
1423
1424	nr = ib_map_mr_sg_pi(mr: req->mr, data_sg: sgl->sg_table.sgl, data_sg_nents: count, NULL,
1425	meta_sg: req->metadata_sgl->sg_table.sgl, meta_sg_nents: pi_count, NULL,
1426	SZ_4K);
1427	if (unlikely(nr))
1428	goto mr_put;
1429
1430	nvme_rdma_set_sig_attrs(bi, cmd: c, sig_attrs: req->mr->sig_attrs, pi_type: ns->head->pi_type);
1431	nvme_rdma_set_prot_checks(cmd: c, mask: &req->mr->sig_attrs->check_mask);
1432
1433	ib_update_fast_reg_key(mr: req->mr, newkey: ib_inc_rkey(rkey: req->mr->rkey));
1434
1435	req->reg_cqe.done = nvme_rdma_sig_done;
1436	memset(wr, 0, sizeof(*wr));
1437	wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1438	wr->wr.wr_cqe = &req->reg_cqe;
1439	wr->wr.num_sge = 0;
1440	wr->wr.send_flags = 0;
1441	wr->mr = req->mr;
1442	wr->key = req->mr->rkey;
1443	wr->access = IB_ACCESS_LOCAL_WRITE |
1444	IB_ACCESS_REMOTE_READ |
1445	IB_ACCESS_REMOTE_WRITE;
1446
1447	sg->addr = cpu_to_le64(req->mr->iova);
1448	xfer_len = req->mr->length;
1449	/* Check if PI is added by the HW */
1450	if (!pi_count)
1451	xfer_len += (xfer_len >> bi->interval_exp) * ns->head->pi_size;
1452	put_unaligned_le24(val: xfer_len, p: sg->length);
1453	put_unaligned_le32(val: req->mr->rkey, p: sg->key);
1454	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1455
1456	return 0;
1457
1458	mr_put:
1459	ib_mr_pool_put(qp: queue->qp, list: &queue->qp->sig_mrs, mr: req->mr);
1460	req->mr = NULL;
1461	if (nr < 0)
1462	return nr;
1463	return -EINVAL;
1464	}
1465
1466	static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1467	int *count, int *pi_count)
1468	{
1469	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1470	int ret;
1471
1472	req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1473	ret = sg_alloc_table_chained(table: &req->data_sgl.sg_table,
1474	nents: blk_rq_nr_phys_segments(rq), first_chunk: req->data_sgl.sg_table.sgl,
1475	NVME_INLINE_SG_CNT);
1476	if (ret)
1477	return -ENOMEM;
1478
1479	req->data_sgl.nents = blk_rq_map_sg(rq, sglist: req->data_sgl.sg_table.sgl);
1480
1481	*count = ib_dma_map_sg(dev: ibdev, sg: req->data_sgl.sg_table.sgl,
1482	nents: req->data_sgl.nents, rq_dma_dir(rq));
1483	if (unlikely(*count <= 0)) {
1484	ret = -EIO;
1485	goto out_free_table;
1486	}
1487
1488	if (blk_integrity_rq(rq)) {
1489	req->metadata_sgl->sg_table.sgl =
1490	(struct scatterlist *)(req->metadata_sgl + 1);
1491	ret = sg_alloc_table_chained(table: &req->metadata_sgl->sg_table,
1492	nents: rq->nr_integrity_segments,
1493	first_chunk: req->metadata_sgl->sg_table.sgl,
1494	NVME_INLINE_METADATA_SG_CNT);
1495	if (unlikely(ret)) {
1496	ret = -ENOMEM;
1497	goto out_unmap_sg;
1498	}
1499
1500	req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq,
1501	req->metadata_sgl->sg_table.sgl);
1502	*pi_count = ib_dma_map_sg(dev: ibdev,
1503	sg: req->metadata_sgl->sg_table.sgl,
1504	nents: req->metadata_sgl->nents,
1505	rq_dma_dir(rq));
1506	if (unlikely(*pi_count <= 0)) {
1507	ret = -EIO;
1508	goto out_free_pi_table;
1509	}
1510	}
1511
1512	return 0;
1513
1514	out_free_pi_table:
1515	sg_free_table_chained(table: &req->metadata_sgl->sg_table,
1516	NVME_INLINE_METADATA_SG_CNT);
1517	out_unmap_sg:
1518	ib_dma_unmap_sg(dev: ibdev, sg: req->data_sgl.sg_table.sgl, nents: req->data_sgl.nents,
1519	rq_dma_dir(rq));
1520	out_free_table:
1521	sg_free_table_chained(table: &req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1522	return ret;
1523	}
1524
1525	static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1526	struct request *rq, struct nvme_command *c)
1527	{
1528	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1529	struct nvme_rdma_device *dev = queue->device;
1530	struct ib_device *ibdev = dev->dev;
1531	int pi_count = 0;
1532	int count, ret;
1533
1534	req->num_sge = 1;
1535	refcount_set(r: &req->ref, n: 2); /* send and recv completions */
1536
1537	c->common.flags |= NVME_CMD_SGL_METABUF;
1538
1539	if (!blk_rq_nr_phys_segments(rq))
1540	return nvme_rdma_set_sg_null(c);
1541
1542	ret = nvme_rdma_dma_map_req(ibdev, rq, count: &count, pi_count: &pi_count);
1543	if (unlikely(ret))
1544	return ret;
1545
1546	if (req->use_sig_mr) {
1547	ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1548	goto out;
1549	}
1550
1551	if (count <= dev->num_inline_segments) {
1552	if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1553	queue->ctrl->use_inline_data &&
1554	blk_rq_payload_bytes(rq) <=
1555	nvme_rdma_inline_data_size(queue)) {
1556	ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1557	goto out;
1558	}
1559
1560	if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1561	ret = nvme_rdma_map_sg_single(queue, req, c);
1562	goto out;
1563	}
1564	}
1565
1566	ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1567	out:
1568	if (unlikely(ret))
1569	goto out_dma_unmap_req;
1570
1571	return 0;
1572
1573	out_dma_unmap_req:
1574	nvme_rdma_dma_unmap_req(ibdev, rq);
1575	return ret;
1576	}
1577
1578	static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1579	{
1580	struct nvme_rdma_qe *qe =
1581	container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1582	struct nvme_rdma_request *req =
1583	container_of(qe, struct nvme_rdma_request, sqe);
1584
1585	if (unlikely(wc->status != IB_WC_SUCCESS))
1586	nvme_rdma_wr_error(cq, wc, op: "SEND");
1587	else
1588	nvme_rdma_end_request(req);
1589	}
1590
1591	static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1592	struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1593	struct ib_send_wr *first)
1594	{
1595	struct ib_send_wr wr;
1596	int ret;
1597
1598	sge->addr = qe->dma;
1599	sge->length = sizeof(struct nvme_command);
1600	sge->lkey = queue->device->pd->local_dma_lkey;
1601
1602	wr.next = NULL;
1603	wr.wr_cqe = &qe->cqe;
1604	wr.sg_list = sge;
1605	wr.num_sge = num_sge;
1606	wr.opcode = IB_WR_SEND;
1607	wr.send_flags = IB_SEND_SIGNALED;
1608
1609	if (first)
1610	first->next = &wr;
1611	else
1612	first = &wr;
1613
1614	ret = ib_post_send(qp: queue->qp, send_wr: first, NULL);
1615	if (unlikely(ret)) {
1616	dev_err(queue->ctrl->ctrl.device,
1617	"%s failed with error code %d\n", __func__, ret);
1618	}
1619	return ret;
1620	}
1621
1622	static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1623	struct nvme_rdma_qe *qe)
1624	{
1625	struct ib_recv_wr wr;
1626	struct ib_sge list;
1627	int ret;
1628
1629	list.addr = qe->dma;
1630	list.length = sizeof(struct nvme_completion);
1631	list.lkey = queue->device->pd->local_dma_lkey;
1632
1633	qe->cqe.done = nvme_rdma_recv_done;
1634
1635	wr.next = NULL;
1636	wr.wr_cqe = &qe->cqe;
1637	wr.sg_list = &list;
1638	wr.num_sge = 1;
1639
1640	ret = ib_post_recv(qp: queue->qp, recv_wr: &wr, NULL);
1641	if (unlikely(ret)) {
1642	dev_err(queue->ctrl->ctrl.device,
1643	"%s failed with error code %d\n", __func__, ret);
1644	}
1645	return ret;
1646	}
1647
1648	static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1649	{
1650	u32 queue_idx = nvme_rdma_queue_idx(queue);
1651
1652	if (queue_idx == 0)
1653	return queue->ctrl->admin_tag_set.tags[queue_idx];
1654	return queue->ctrl->tag_set.tags[queue_idx - 1];
1655	}
1656
1657	static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1658	{
1659	if (unlikely(wc->status != IB_WC_SUCCESS))
1660	nvme_rdma_wr_error(cq, wc, op: "ASYNC");
1661	}
1662
1663	static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1664	{
1665	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: arg);
1666	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1667	struct ib_device *dev = queue->device->dev;
1668	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1669	struct nvme_command *cmd = sqe->data;
1670	struct ib_sge sge;
1671	int ret;
1672
1673	ib_dma_sync_single_for_cpu(dev, addr: sqe->dma, size: sizeof(*cmd), dir: DMA_TO_DEVICE);
1674
1675	memset(cmd, 0, sizeof(*cmd));
1676	cmd->common.opcode = nvme_admin_async_event;
1677	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1678	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1679	nvme_rdma_set_sg_null(c: cmd);
1680
1681	sqe->cqe.done = nvme_rdma_async_done;
1682
1683	ib_dma_sync_single_for_device(dev, addr: sqe->dma, size: sizeof(*cmd),
1684	dir: DMA_TO_DEVICE);
1685
1686	ret = nvme_rdma_post_send(queue, qe: sqe, sge: &sge, num_sge: 1, NULL);
1687	WARN_ON_ONCE(ret);
1688	}
1689
1690	static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1691	struct nvme_completion *cqe, struct ib_wc *wc)
1692	{
1693	struct request *rq;
1694	struct nvme_rdma_request *req;
1695
1696	rq = nvme_find_rq(tags: nvme_rdma_tagset(queue), command_id: cqe->command_id);
1697	if (!rq) {
1698	dev_err(queue->ctrl->ctrl.device,
1699	"got bad command_id %#x on QP %#x\n",
1700	cqe->command_id, queue->qp->qp_num);
1701	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1702	return;
1703	}
1704	req = blk_mq_rq_to_pdu(rq);
1705
1706	req->status = cqe->status;
1707	req->result = cqe->result;
1708
1709	if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1710	if (unlikely(!req->mr ||
1711	wc->ex.invalidate_rkey != req->mr->rkey)) {
1712	dev_err(queue->ctrl->ctrl.device,
1713	"Bogus remote invalidation for rkey %#x\n",
1714	req->mr ? req->mr->rkey : 0);
1715	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1716	}
1717	} else if (req->mr) {
1718	int ret;
1719
1720	ret = nvme_rdma_inv_rkey(queue, req);
1721	if (unlikely(ret < 0)) {
1722	dev_err(queue->ctrl->ctrl.device,
1723	"Queueing INV WR for rkey %#x failed (%d)\n",
1724	req->mr->rkey, ret);
1725	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1726	}
1727	/* the local invalidation completion will end the request */
1728	return;
1729	}
1730
1731	nvme_rdma_end_request(req);
1732	}
1733
1734	static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1735	{
1736	struct nvme_rdma_qe *qe =
1737	container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1738	struct nvme_rdma_queue *queue = wc->qp->qp_context;
1739	struct ib_device *ibdev = queue->device->dev;
1740	struct nvme_completion *cqe = qe->data;
1741	const size_t len = sizeof(struct nvme_completion);
1742
1743	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1744	nvme_rdma_wr_error(cq, wc, op: "RECV");
1745	return;
1746	}
1747
1748	/* sanity checking for received data length */
1749	if (unlikely(wc->byte_len < len)) {
1750	dev_err(queue->ctrl->ctrl.device,
1751	"Unexpected nvme completion length(%d)\n", wc->byte_len);
1752	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1753	return;
1754	}
1755
1756	ib_dma_sync_single_for_cpu(dev: ibdev, addr: qe->dma, size: len, dir: DMA_FROM_DEVICE);
1757	/*
1758	* AEN requests are special as they don't time out and can
1759	* survive any kind of queue freeze and often don't respond to
1760	* aborts. We don't even bother to allocate a struct request
1761	* for them but rather special case them here.
1762	*/
1763	if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1764	cqe->command_id)))
1765	nvme_complete_async_event(ctrl: &queue->ctrl->ctrl, status: cqe->status,
1766	res: &cqe->result);
1767	else
1768	nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1769	ib_dma_sync_single_for_device(dev: ibdev, addr: qe->dma, size: len, dir: DMA_FROM_DEVICE);
1770
1771	nvme_rdma_post_recv(queue, qe);
1772	}
1773
1774	static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1775	{
1776	int ret, i;
1777
1778	for (i = 0; i < queue->queue_size; i++) {
1779	ret = nvme_rdma_post_recv(queue, qe: &queue->rsp_ring[i]);
1780	if (ret)
1781	return ret;
1782	}
1783
1784	return 0;
1785	}
1786
1787	static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1788	struct rdma_cm_event *ev)
1789	{
1790	struct rdma_cm_id *cm_id = queue->cm_id;
1791	int status = ev->status;
1792	const char *rej_msg;
1793	const struct nvme_rdma_cm_rej *rej_data;
1794	u8 rej_data_len;
1795
1796	rej_msg = rdma_reject_msg(id: cm_id, reason: status);
1797	rej_data = rdma_consumer_reject_data(id: cm_id, ev, data_len: &rej_data_len);
1798
1799	if (rej_data && rej_data_len >= sizeof(u16)) {
1800	u16 sts = le16_to_cpu(rej_data->sts);
1801
1802	dev_err(queue->ctrl->ctrl.device,
1803	"Connect rejected: status %d (%s) nvme status %d (%s).\n",
1804	status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1805	} else {
1806	dev_err(queue->ctrl->ctrl.device,
1807	"Connect rejected: status %d (%s).\n", status, rej_msg);
1808	}
1809
1810	return -ECONNRESET;
1811	}
1812
1813	static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1814	{
1815	struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1816	int ret;
1817
1818	ret = nvme_rdma_create_queue_ib(queue);
1819	if (ret)
1820	return ret;
1821
1822	if (ctrl->opts->tos >= 0)
1823	rdma_set_service_type(id: queue->cm_id, tos: ctrl->opts->tos);
1824	ret = rdma_resolve_route(id: queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS);
1825	if (ret) {
1826	dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1827	queue->cm_error);
1828	goto out_destroy_queue;
1829	}
1830
1831	return 0;
1832
1833	out_destroy_queue:
1834	nvme_rdma_destroy_queue_ib(queue);
1835	return ret;
1836	}
1837
1838	static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1839	{
1840	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1841	struct rdma_conn_param param = { };
1842	struct nvme_rdma_cm_req priv = { };
1843	int ret;
1844
1845	param.qp_num = queue->qp->qp_num;
1846	param.flow_control = 1;
1847
1848	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1849	/* maximum retry count */
1850	param.retry_count = 7;
1851	param.rnr_retry_count = 7;
1852	param.private_data = &priv;
1853	param.private_data_len = sizeof(priv);
1854
1855	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1856	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1857	/*
1858	* set the admin queue depth to the minimum size
1859	* specified by the Fabrics standard.
1860	*/
1861	if (priv.qid == 0) {
1862	priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1863	priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1864	} else {
1865	/*
1866	* current interpretation of the fabrics spec
1867	* is at minimum you make hrqsize sqsize+1, or a
1868	* 1's based representation of sqsize.
1869	*/
1870	priv.hrqsize = cpu_to_le16(queue->queue_size);
1871	priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1872	/* cntlid should only be set when creating an I/O queue */
1873	priv.cntlid = cpu_to_le16(ctrl->ctrl.cntlid);
1874	}
1875
1876	ret = rdma_connect_locked(id: queue->cm_id, conn_param: &param);
1877	if (ret) {
1878	dev_err(ctrl->ctrl.device,
1879	"rdma_connect_locked failed (%d).\n", ret);
1880	return ret;
1881	}
1882
1883	return 0;
1884	}
1885
1886	static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1887	struct rdma_cm_event *ev)
1888	{
1889	struct nvme_rdma_queue *queue = cm_id->context;
1890	int cm_error = 0;
1891
1892	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1893	rdma_event_msg(ev->event), ev->event,
1894	ev->status, cm_id);
1895
1896	switch (ev->event) {
1897	case RDMA_CM_EVENT_ADDR_RESOLVED:
1898	cm_error = nvme_rdma_addr_resolved(queue);
1899	break;
1900	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1901	cm_error = nvme_rdma_route_resolved(queue);
1902	break;
1903	case RDMA_CM_EVENT_ESTABLISHED:
1904	queue->cm_error = nvme_rdma_conn_established(queue);
1905	/* complete cm_done regardless of success/failure */
1906	complete(&queue->cm_done);
1907	return 0;
1908	case RDMA_CM_EVENT_REJECTED:
1909	cm_error = nvme_rdma_conn_rejected(queue, ev);
1910	break;
1911	case RDMA_CM_EVENT_ROUTE_ERROR:
1912	case RDMA_CM_EVENT_CONNECT_ERROR:
1913	case RDMA_CM_EVENT_UNREACHABLE:
1914	case RDMA_CM_EVENT_ADDR_ERROR:
1915	dev_dbg(queue->ctrl->ctrl.device,
1916	"CM error event %d\n", ev->event);
1917	cm_error = -ECONNRESET;
1918	break;
1919	case RDMA_CM_EVENT_DISCONNECTED:
1920	case RDMA_CM_EVENT_ADDR_CHANGE:
1921	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1922	dev_dbg(queue->ctrl->ctrl.device,
1923	"disconnect received - connection closed\n");
1924	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1925	break;
1926	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1927	/* device removal is handled via the ib_client API */
1928	break;
1929	default:
1930	dev_err(queue->ctrl->ctrl.device,
1931	"Unexpected RDMA CM event (%d)\n", ev->event);
1932	nvme_rdma_error_recovery(ctrl: queue->ctrl);
1933	break;
1934	}
1935
1936	if (cm_error) {
1937	queue->cm_error = cm_error;
1938	complete(&queue->cm_done);
1939	}
1940
1941	return 0;
1942	}
1943
1944	static void nvme_rdma_complete_timed_out(struct request *rq)
1945	{
1946	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1947	struct nvme_rdma_queue *queue = req->queue;
1948
1949	nvme_rdma_stop_queue(queue);
1950	nvmf_complete_timed_out_request(rq);
1951	}
1952
1953	static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq)
1954	{
1955	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1956	struct nvme_rdma_queue *queue = req->queue;
1957	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1958	struct nvme_command *cmd = req->req.cmd;
1959	int qid = nvme_rdma_queue_idx(queue);
1960
1961	dev_warn(ctrl->ctrl.device,
1962	"I/O tag %d (%04x) opcode %#x (%s) QID %d timeout\n",
1963	rq->tag, nvme_cid(rq), cmd->common.opcode,
1964	nvme_fabrics_opcode_str(qid, cmd), qid);
1965
1966	if (nvme_ctrl_state(ctrl: &ctrl->ctrl) != NVME_CTRL_LIVE) {
1967	/*
1968	* If we are resetting, connecting or deleting we should
1969	* complete immediately because we may block controller
1970	* teardown or setup sequence
1971	* - ctrl disable/shutdown fabrics requests
1972	* - connect requests
1973	* - initialization admin requests
1974	* - I/O requests that entered after unquiescing and
1975	* the controller stopped responding
1976	*
1977	* All other requests should be cancelled by the error
1978	* recovery work, so it's fine that we fail it here.
1979	*/
1980	nvme_rdma_complete_timed_out(rq);
1981	return BLK_EH_DONE;
1982	}
1983
1984	/*
1985	* LIVE state should trigger the normal error recovery which will
1986	* handle completing this request.
1987	*/
1988	nvme_rdma_error_recovery(ctrl);
1989	return BLK_EH_RESET_TIMER;
1990	}
1991
1992	static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1993	const struct blk_mq_queue_data *bd)
1994	{
1995	struct nvme_ns *ns = hctx->queue->queuedata;
1996	struct nvme_rdma_queue *queue = hctx->driver_data;
1997	struct request *rq = bd->rq;
1998	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1999	struct nvme_rdma_qe *sqe = &req->sqe;
2000	struct nvme_command *c = nvme_req(req: rq)->cmd;
2001	struct ib_device *dev;
2002	bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2003	blk_status_t ret;
2004	int err;
2005
2006	WARN_ON_ONCE(rq->tag < 0);
2007
2008	if (!nvme_check_ready(ctrl: &queue->ctrl->ctrl, rq, queue_live: queue_ready))
2009	return nvme_fail_nonready_command(ctrl: &queue->ctrl->ctrl, req: rq);
2010
2011	dev = queue->device->dev;
2012
2013	req->sqe.dma = ib_dma_map_single(dev, cpu_addr: req->sqe.data,
2014	size: sizeof(struct nvme_command),
2015	direction: DMA_TO_DEVICE);
2016	err = ib_dma_mapping_error(dev, dma_addr: req->sqe.dma);
2017	if (unlikely(err))
2018	return BLK_STS_RESOURCE;
2019
2020	ib_dma_sync_single_for_cpu(dev, addr: sqe->dma,
2021	size: sizeof(struct nvme_command), dir: DMA_TO_DEVICE);
2022
2023	ret = nvme_setup_cmd(ns, req: rq);
2024	if (ret)
2025	goto unmap_qe;
2026
2027	nvme_start_request(rq);
2028
2029	if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2030	queue->pi_support &&
2031	(c->common.opcode == nvme_cmd_write ||
2032	c->common.opcode == nvme_cmd_read) &&
2033	nvme_ns_has_pi(head: ns->head))
2034	req->use_sig_mr = true;
2035	else
2036	req->use_sig_mr = false;
2037
2038	err = nvme_rdma_map_data(queue, rq, c);
2039	if (unlikely(err < 0)) {
2040	dev_err(queue->ctrl->ctrl.device,
2041	"Failed to map data (%d)\n", err);
2042	goto err;
2043	}
2044
2045	sqe->cqe.done = nvme_rdma_send_done;
2046
2047	ib_dma_sync_single_for_device(dev, addr: sqe->dma,
2048	size: sizeof(struct nvme_command), dir: DMA_TO_DEVICE);
2049
2050	err = nvme_rdma_post_send(queue, qe: sqe, sge: req->sge, num_sge: req->num_sge,
2051	first: req->mr ? &req->reg_wr.wr : NULL);
2052	if (unlikely(err))
2053	goto err_unmap;
2054
2055	return BLK_STS_OK;
2056
2057	err_unmap:
2058	nvme_rdma_unmap_data(queue, rq);
2059	err:
2060	if (err == -EIO)
2061	ret = nvme_host_path_error(req: rq);
2062	else if (err == -ENOMEM || err == -EAGAIN)
2063	ret = BLK_STS_RESOURCE;
2064	else
2065	ret = BLK_STS_IOERR;
2066	nvme_cleanup_cmd(req: rq);
2067	unmap_qe:
2068	ib_dma_unmap_single(dev, addr: req->sqe.dma, size: sizeof(struct nvme_command),
2069	direction: DMA_TO_DEVICE);
2070	return ret;
2071	}
2072
2073	static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2074	{
2075	struct nvme_rdma_queue *queue = hctx->driver_data;
2076
2077	return ib_process_cq_direct(cq: queue->ib_cq, budget: -1);
2078	}
2079
2080	static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2081	{
2082	struct request *rq = blk_mq_rq_from_pdu(pdu: req);
2083	struct ib_mr_status mr_status;
2084	int ret;
2085
2086	ret = ib_check_mr_status(mr: req->mr, check_mask: IB_MR_CHECK_SIG_STATUS, mr_status: &mr_status);
2087	if (ret) {
2088	pr_err("ib_check_mr_status failed, ret %d\n", ret);
2089	nvme_req(req: rq)->status = NVME_SC_INVALID_PI;
2090	return;
2091	}
2092
2093	if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2094	switch (mr_status.sig_err.err_type) {
2095	case IB_SIG_BAD_GUARD:
2096	nvme_req(req: rq)->status = NVME_SC_GUARD_CHECK;
2097	break;
2098	case IB_SIG_BAD_REFTAG:
2099	nvme_req(req: rq)->status = NVME_SC_REFTAG_CHECK;
2100	break;
2101	case IB_SIG_BAD_APPTAG:
2102	nvme_req(req: rq)->status = NVME_SC_APPTAG_CHECK;
2103	break;
2104	}
2105	pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2106	mr_status.sig_err.err_type, mr_status.sig_err.expected,
2107	mr_status.sig_err.actual);
2108	}
2109	}
2110
2111	static void nvme_rdma_complete_rq(struct request *rq)
2112	{
2113	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2114	struct nvme_rdma_queue *queue = req->queue;
2115	struct ib_device *ibdev = queue->device->dev;
2116
2117	if (req->use_sig_mr)
2118	nvme_rdma_check_pi_status(req);
2119
2120	nvme_rdma_unmap_data(queue, rq);
2121	ib_dma_unmap_single(dev: ibdev, addr: req->sqe.dma, size: sizeof(struct nvme_command),
2122	direction: DMA_TO_DEVICE);
2123	nvme_complete_rq(req: rq);
2124	}
2125
2126	static void nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2127	{
2128	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(ctrl: set->driver_data);
2129
2130	nvmf_map_queues(set, ctrl: &ctrl->ctrl, io_queues: ctrl->io_queues);
2131	}
2132
2133	static const struct blk_mq_ops nvme_rdma_mq_ops = {
2134	.queue_rq = nvme_rdma_queue_rq,
2135	.complete = nvme_rdma_complete_rq,
2136	.init_request = nvme_rdma_init_request,
2137	.exit_request = nvme_rdma_exit_request,
2138	.init_hctx = nvme_rdma_init_hctx,
2139	.timeout = nvme_rdma_timeout,
2140	.map_queues = nvme_rdma_map_queues,
2141	.poll = nvme_rdma_poll,
2142	};
2143
2144	static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2145	.queue_rq = nvme_rdma_queue_rq,
2146	.complete = nvme_rdma_complete_rq,
2147	.init_request = nvme_rdma_init_request,
2148	.exit_request = nvme_rdma_exit_request,
2149	.init_hctx = nvme_rdma_init_admin_hctx,
2150	.timeout = nvme_rdma_timeout,
2151	};
2152
2153	static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2154	{
2155	nvme_rdma_teardown_io_queues(ctrl, remove: shutdown);
2156	nvme_quiesce_admin_queue(ctrl: &ctrl->ctrl);
2157	nvme_disable_ctrl(ctrl: &ctrl->ctrl, shutdown);
2158	nvme_rdma_teardown_admin_queue(ctrl, remove: shutdown);
2159	}
2160
2161	static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2162	{
2163	nvme_rdma_shutdown_ctrl(ctrl: to_rdma_ctrl(ctrl), shutdown: true);
2164	}
2165
2166	static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2167	{
2168	struct nvme_rdma_ctrl *ctrl =
2169	container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2170	int ret;
2171
2172	nvme_stop_ctrl(ctrl: &ctrl->ctrl);
2173	nvme_rdma_shutdown_ctrl(ctrl, shutdown: false);
2174
2175	if (!nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_CONNECTING)) {
2176	/* state change failure should never happen */
2177	WARN_ON_ONCE(1);
2178	return;
2179	}
2180
2181	ret = nvme_rdma_setup_ctrl(ctrl, new: false);
2182	if (ret)
2183	goto out_fail;
2184
2185	return;
2186
2187	out_fail:
2188	++ctrl->ctrl.nr_reconnects;
2189	nvme_rdma_reconnect_or_remove(ctrl, status: ret);
2190	}
2191
2192	static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2193	.name = "rdma",
2194	.module = THIS_MODULE,
2195	.flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2196	.reg_read32 = nvmf_reg_read32,
2197	.reg_read64 = nvmf_reg_read64,
2198	.reg_write32 = nvmf_reg_write32,
2199	.subsystem_reset = nvmf_subsystem_reset,
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	.stop_ctrl = nvme_rdma_stop_ctrl,
2205	};
2206
2207	/*
2208	* Fails a connection request if it matches an existing controller
2209	* (association) with the same tuple:
2210	* <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2211	*
2212	* if local address is not specified in the request, it will match an
2213	* existing controller with all the other parameters the same and no
2214	* local port address specified as well.
2215	*
2216	* The ports don't need to be compared as they are intrinsically
2217	* already matched by the port pointers supplied.
2218	*/
2219	static bool
2220	nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2221	{
2222	struct nvme_rdma_ctrl *ctrl;
2223	bool found = false;
2224
2225	mutex_lock(&nvme_rdma_ctrl_mutex);
2226	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2227	found = nvmf_ip_options_match(ctrl: &ctrl->ctrl, opts);
2228	if (found)
2229	break;
2230	}
2231	mutex_unlock(lock: &nvme_rdma_ctrl_mutex);
2232
2233	return found;
2234	}
2235
2236	static struct nvme_rdma_ctrl *nvme_rdma_alloc_ctrl(struct device *dev,
2237	struct nvmf_ctrl_options *opts)
2238	{
2239	struct nvme_rdma_ctrl *ctrl;
2240	int ret;
2241
2242	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2243	if (!ctrl)
2244	return ERR_PTR(error: -ENOMEM);
2245	ctrl->ctrl.opts = opts;
2246	INIT_LIST_HEAD(list: &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(net: &init_net, AF_UNSPEC,
2259	src: opts->traddr, port: opts->trsvcid, addr: &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(net: &init_net, AF_UNSPEC,
2268	src: opts->host_traddr, NULL, addr: &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->ctrl, dev, ops: &nvme_rdma_ctrl_ops,
2298	quirks: 0 /* no quirks, we're perfect! */);
2299	if (ret)
2300	goto out_kfree_queues;
2301
2302	return ctrl;
2303
2304	out_kfree_queues:
2305	kfree(objp: ctrl->queues);
2306	out_free_ctrl:
2307	kfree(objp: ctrl);
2308	return ERR_PTR(error: ret);
2309	}
2310
2311	static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2312	struct nvmf_ctrl_options *opts)
2313	{
2314	struct nvme_rdma_ctrl *ctrl;
2315	bool changed;
2316	int ret;
2317
2318	ctrl = nvme_rdma_alloc_ctrl(dev, opts);
2319	if (IS_ERR(ptr: ctrl))
2320	return ERR_CAST(ptr: ctrl);
2321
2322	ret = nvme_add_ctrl(ctrl: &ctrl->ctrl);
2323	if (ret)
2324	goto out_put_ctrl;
2325
2326	changed = nvme_change_ctrl_state(ctrl: &ctrl->ctrl, new_state: NVME_CTRL_CONNECTING);
2327	WARN_ON_ONCE(!changed);
2328
2329	ret = nvme_rdma_setup_ctrl(ctrl, new: true);
2330	if (ret)
2331	goto out_uninit_ctrl;
2332
2333	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs, hostnqn: %s\n",
2334	nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr, opts->host->nqn);
2335
2336	mutex_lock(&nvme_rdma_ctrl_mutex);
2337	list_add_tail(new: &ctrl->list, head: &nvme_rdma_ctrl_list);
2338	mutex_unlock(lock: &nvme_rdma_ctrl_mutex);
2339
2340	return &ctrl->ctrl;
2341
2342	out_uninit_ctrl:
2343	nvme_uninit_ctrl(ctrl: &ctrl->ctrl);
2344	out_put_ctrl:
2345	nvme_put_ctrl(ctrl: &ctrl->ctrl);
2346	if (ret > 0)
2347	ret = -EIO;
2348	return ERR_PTR(error: ret);
2349	}
2350
2351	static struct nvmf_transport_ops nvme_rdma_transport = {
2352	.name = "rdma",
2353	.module = THIS_MODULE,
2354	.required_opts = NVMF_OPT_TRADDR,
2355	.allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2356	NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2357	NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2358	NVMF_OPT_TOS,
2359	.create_ctrl = nvme_rdma_create_ctrl,
2360	};
2361
2362	static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2363	{
2364	struct nvme_rdma_ctrl *ctrl;
2365	struct nvme_rdma_device *ndev;
2366	bool found = false;
2367
2368	mutex_lock(&device_list_mutex);
2369	list_for_each_entry(ndev, &device_list, entry) {
2370	if (ndev->dev == ib_device) {
2371	found = true;
2372	break;
2373	}
2374	}
2375	mutex_unlock(lock: &device_list_mutex);
2376
2377	if (!found)
2378	return;
2379
2380	/* Delete all controllers using this device */
2381	mutex_lock(&nvme_rdma_ctrl_mutex);
2382	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2383	if (ctrl->device->dev != ib_device)
2384	continue;
2385	nvme_delete_ctrl(ctrl: &ctrl->ctrl);
2386	}
2387	mutex_unlock(lock: &nvme_rdma_ctrl_mutex);
2388
2389	flush_workqueue(nvme_delete_wq);
2390	}
2391
2392	static struct ib_client nvme_rdma_ib_client = {
2393	.name = "nvme_rdma",
2394	.remove = nvme_rdma_remove_one
2395	};
2396
2397	static int __init nvme_rdma_init_module(void)
2398	{
2399	int ret;
2400
2401	ret = ib_register_client(client: &nvme_rdma_ib_client);
2402	if (ret)
2403	return ret;
2404
2405	ret = nvmf_register_transport(ops: &nvme_rdma_transport);
2406	if (ret)
2407	goto err_unreg_client;
2408
2409	return 0;
2410
2411	err_unreg_client:
2412	ib_unregister_client(client: &nvme_rdma_ib_client);
2413	return ret;
2414	}
2415
2416	static void __exit nvme_rdma_cleanup_module(void)
2417	{
2418	struct nvme_rdma_ctrl *ctrl;
2419
2420	nvmf_unregister_transport(ops: &nvme_rdma_transport);
2421	ib_unregister_client(client: &nvme_rdma_ib_client);
2422
2423	mutex_lock(&nvme_rdma_ctrl_mutex);
2424	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2425	nvme_delete_ctrl(ctrl: &ctrl->ctrl);
2426	mutex_unlock(lock: &nvme_rdma_ctrl_mutex);
2427	flush_workqueue(nvme_delete_wq);
2428	}
2429
2430	module_init(nvme_rdma_init_module);
2431	module_exit(nvme_rdma_cleanup_module);
2432
2433	MODULE_DESCRIPTION("NVMe host RDMA transport driver");
2434	MODULE_LICENSE("GPL v2");
2435