xref: /linux/drivers/nvme/target/rdma.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics RDMA target.
4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/atomic.h>
8 #include <linux/blk-integrity.h>
9 #include <linux/ctype.h>
10 #include <linux/delay.h>
11 #include <linux/err.h>
12 #include <linux/init.h>
13 #include <linux/module.h>
14 #include <linux/nvme.h>
15 #include <linux/slab.h>
16 #include <linux/string.h>
17 #include <linux/wait.h>
18 #include <linux/inet.h>
19 #include <asm/unaligned.h>
20 
21 #include <rdma/ib_verbs.h>
22 #include <rdma/rdma_cm.h>
23 #include <rdma/rw.h>
24 #include <rdma/ib_cm.h>
25 
26 #include <linux/nvme-rdma.h>
27 #include "nvmet.h"
28 
29 /*
30  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
31  */
32 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE	PAGE_SIZE
33 #define NVMET_RDMA_MAX_INLINE_SGE		4
34 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE		max_t(int, SZ_16K, PAGE_SIZE)
35 
36 /* Assume mpsmin == device_page_size == 4KB */
37 #define NVMET_RDMA_MAX_MDTS			8
38 #define NVMET_RDMA_MAX_METADATA_MDTS		5
39 
40 #define NVMET_RDMA_BACKLOG 128
41 
42 struct nvmet_rdma_srq;
43 
44 struct nvmet_rdma_cmd {
45 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
46 	struct ib_cqe		cqe;
47 	struct ib_recv_wr	wr;
48 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
49 	struct nvme_command     *nvme_cmd;
50 	struct nvmet_rdma_queue	*queue;
51 	struct nvmet_rdma_srq   *nsrq;
52 };
53 
54 enum {
55 	NVMET_RDMA_REQ_INLINE_DATA	= (1 << 0),
56 };
57 
58 struct nvmet_rdma_rsp {
59 	struct ib_sge		send_sge;
60 	struct ib_cqe		send_cqe;
61 	struct ib_send_wr	send_wr;
62 
63 	struct nvmet_rdma_cmd	*cmd;
64 	struct nvmet_rdma_queue	*queue;
65 
66 	struct ib_cqe		read_cqe;
67 	struct ib_cqe		write_cqe;
68 	struct rdma_rw_ctx	rw;
69 
70 	struct nvmet_req	req;
71 
72 	bool			allocated;
73 	u8			n_rdma;
74 	u32			flags;
75 	u32			invalidate_rkey;
76 
77 	struct list_head	wait_list;
78 	struct list_head	free_list;
79 };
80 
81 enum nvmet_rdma_queue_state {
82 	NVMET_RDMA_Q_CONNECTING,
83 	NVMET_RDMA_Q_LIVE,
84 	NVMET_RDMA_Q_DISCONNECTING,
85 };
86 
87 struct nvmet_rdma_queue {
88 	struct rdma_cm_id	*cm_id;
89 	struct ib_qp		*qp;
90 	struct nvmet_port	*port;
91 	struct ib_cq		*cq;
92 	atomic_t		sq_wr_avail;
93 	struct nvmet_rdma_device *dev;
94 	struct nvmet_rdma_srq   *nsrq;
95 	spinlock_t		state_lock;
96 	enum nvmet_rdma_queue_state state;
97 	struct nvmet_cq		nvme_cq;
98 	struct nvmet_sq		nvme_sq;
99 
100 	struct nvmet_rdma_rsp	*rsps;
101 	struct list_head	free_rsps;
102 	spinlock_t		rsps_lock;
103 	struct nvmet_rdma_cmd	*cmds;
104 
105 	struct work_struct	release_work;
106 	struct list_head	rsp_wait_list;
107 	struct list_head	rsp_wr_wait_list;
108 	spinlock_t		rsp_wr_wait_lock;
109 
110 	int			idx;
111 	int			host_qid;
112 	int			comp_vector;
113 	int			recv_queue_size;
114 	int			send_queue_size;
115 
116 	struct list_head	queue_list;
117 };
118 
119 struct nvmet_rdma_port {
120 	struct nvmet_port	*nport;
121 	struct sockaddr_storage addr;
122 	struct rdma_cm_id	*cm_id;
123 	struct delayed_work	repair_work;
124 };
125 
126 struct nvmet_rdma_srq {
127 	struct ib_srq            *srq;
128 	struct nvmet_rdma_cmd    *cmds;
129 	struct nvmet_rdma_device *ndev;
130 };
131 
132 struct nvmet_rdma_device {
133 	struct ib_device	*device;
134 	struct ib_pd		*pd;
135 	struct nvmet_rdma_srq	**srqs;
136 	int			srq_count;
137 	size_t			srq_size;
138 	struct kref		ref;
139 	struct list_head	entry;
140 	int			inline_data_size;
141 	int			inline_page_count;
142 };
143 
144 static bool nvmet_rdma_use_srq;
145 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
146 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
147 
148 static int srq_size_set(const char *val, const struct kernel_param *kp);
149 static const struct kernel_param_ops srq_size_ops = {
150 	.set = srq_size_set,
151 	.get = param_get_int,
152 };
153 
154 static int nvmet_rdma_srq_size = 1024;
155 module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
156 MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");
157 
158 static DEFINE_IDA(nvmet_rdma_queue_ida);
159 static LIST_HEAD(nvmet_rdma_queue_list);
160 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
161 
162 static LIST_HEAD(device_list);
163 static DEFINE_MUTEX(device_list_mutex);
164 
165 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
166 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
167 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
168 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
169 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
170 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
171 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
172 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
173 				struct nvmet_rdma_rsp *r);
174 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
175 				struct nvmet_rdma_rsp *r);
176 
177 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
178 
179 static int srq_size_set(const char *val, const struct kernel_param *kp)
180 {
181 	int n = 0, ret;
182 
183 	ret = kstrtoint(val, 10, &n);
184 	if (ret != 0 || n < 256)
185 		return -EINVAL;
186 
187 	return param_set_int(val, kp);
188 }
189 
190 static int num_pages(int len)
191 {
192 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
193 }
194 
195 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
196 {
197 	return nvme_is_write(rsp->req.cmd) &&
198 		rsp->req.transfer_len &&
199 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
200 }
201 
202 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
203 {
204 	return !nvme_is_write(rsp->req.cmd) &&
205 		rsp->req.transfer_len &&
206 		!rsp->req.cqe->status &&
207 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
208 }
209 
210 static inline struct nvmet_rdma_rsp *
211 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
212 {
213 	struct nvmet_rdma_rsp *rsp;
214 	unsigned long flags;
215 
216 	spin_lock_irqsave(&queue->rsps_lock, flags);
217 	rsp = list_first_entry_or_null(&queue->free_rsps,
218 				struct nvmet_rdma_rsp, free_list);
219 	if (likely(rsp))
220 		list_del(&rsp->free_list);
221 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
222 
223 	if (unlikely(!rsp)) {
224 		int ret;
225 
226 		rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
227 		if (unlikely(!rsp))
228 			return NULL;
229 		ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
230 		if (unlikely(ret)) {
231 			kfree(rsp);
232 			return NULL;
233 		}
234 
235 		rsp->allocated = true;
236 	}
237 
238 	return rsp;
239 }
240 
241 static inline void
242 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
243 {
244 	unsigned long flags;
245 
246 	if (unlikely(rsp->allocated)) {
247 		nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
248 		kfree(rsp);
249 		return;
250 	}
251 
252 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
253 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
254 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
255 }
256 
257 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
258 				struct nvmet_rdma_cmd *c)
259 {
260 	struct scatterlist *sg;
261 	struct ib_sge *sge;
262 	int i;
263 
264 	if (!ndev->inline_data_size)
265 		return;
266 
267 	sg = c->inline_sg;
268 	sge = &c->sge[1];
269 
270 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
271 		if (sge->length)
272 			ib_dma_unmap_page(ndev->device, sge->addr,
273 					sge->length, DMA_FROM_DEVICE);
274 		if (sg_page(sg))
275 			__free_page(sg_page(sg));
276 	}
277 }
278 
279 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
280 				struct nvmet_rdma_cmd *c)
281 {
282 	struct scatterlist *sg;
283 	struct ib_sge *sge;
284 	struct page *pg;
285 	int len;
286 	int i;
287 
288 	if (!ndev->inline_data_size)
289 		return 0;
290 
291 	sg = c->inline_sg;
292 	sg_init_table(sg, ndev->inline_page_count);
293 	sge = &c->sge[1];
294 	len = ndev->inline_data_size;
295 
296 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
297 		pg = alloc_page(GFP_KERNEL);
298 		if (!pg)
299 			goto out_err;
300 		sg_assign_page(sg, pg);
301 		sge->addr = ib_dma_map_page(ndev->device,
302 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
303 		if (ib_dma_mapping_error(ndev->device, sge->addr))
304 			goto out_err;
305 		sge->length = min_t(int, len, PAGE_SIZE);
306 		sge->lkey = ndev->pd->local_dma_lkey;
307 		len -= sge->length;
308 	}
309 
310 	return 0;
311 out_err:
312 	for (; i >= 0; i--, sg--, sge--) {
313 		if (sge->length)
314 			ib_dma_unmap_page(ndev->device, sge->addr,
315 					sge->length, DMA_FROM_DEVICE);
316 		if (sg_page(sg))
317 			__free_page(sg_page(sg));
318 	}
319 	return -ENOMEM;
320 }
321 
322 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
323 			struct nvmet_rdma_cmd *c, bool admin)
324 {
325 	/* NVMe command / RDMA RECV */
326 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
327 	if (!c->nvme_cmd)
328 		goto out;
329 
330 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
331 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
332 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
333 		goto out_free_cmd;
334 
335 	c->sge[0].length = sizeof(*c->nvme_cmd);
336 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
337 
338 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
339 		goto out_unmap_cmd;
340 
341 	c->cqe.done = nvmet_rdma_recv_done;
342 
343 	c->wr.wr_cqe = &c->cqe;
344 	c->wr.sg_list = c->sge;
345 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
346 
347 	return 0;
348 
349 out_unmap_cmd:
350 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
351 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
352 out_free_cmd:
353 	kfree(c->nvme_cmd);
354 
355 out:
356 	return -ENOMEM;
357 }
358 
359 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
360 		struct nvmet_rdma_cmd *c, bool admin)
361 {
362 	if (!admin)
363 		nvmet_rdma_free_inline_pages(ndev, c);
364 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
365 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
366 	kfree(c->nvme_cmd);
367 }
368 
369 static struct nvmet_rdma_cmd *
370 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
371 		int nr_cmds, bool admin)
372 {
373 	struct nvmet_rdma_cmd *cmds;
374 	int ret = -EINVAL, i;
375 
376 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
377 	if (!cmds)
378 		goto out;
379 
380 	for (i = 0; i < nr_cmds; i++) {
381 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
382 		if (ret)
383 			goto out_free;
384 	}
385 
386 	return cmds;
387 
388 out_free:
389 	while (--i >= 0)
390 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
391 	kfree(cmds);
392 out:
393 	return ERR_PTR(ret);
394 }
395 
396 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
397 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
398 {
399 	int i;
400 
401 	for (i = 0; i < nr_cmds; i++)
402 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
403 	kfree(cmds);
404 }
405 
406 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
407 		struct nvmet_rdma_rsp *r)
408 {
409 	/* NVMe CQE / RDMA SEND */
410 	r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
411 	if (!r->req.cqe)
412 		goto out;
413 
414 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
415 			sizeof(*r->req.cqe), DMA_TO_DEVICE);
416 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
417 		goto out_free_rsp;
418 
419 	if (ib_dma_pci_p2p_dma_supported(ndev->device))
420 		r->req.p2p_client = &ndev->device->dev;
421 	r->send_sge.length = sizeof(*r->req.cqe);
422 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
423 
424 	r->send_cqe.done = nvmet_rdma_send_done;
425 
426 	r->send_wr.wr_cqe = &r->send_cqe;
427 	r->send_wr.sg_list = &r->send_sge;
428 	r->send_wr.num_sge = 1;
429 	r->send_wr.send_flags = IB_SEND_SIGNALED;
430 
431 	/* Data In / RDMA READ */
432 	r->read_cqe.done = nvmet_rdma_read_data_done;
433 	/* Data Out / RDMA WRITE */
434 	r->write_cqe.done = nvmet_rdma_write_data_done;
435 
436 	return 0;
437 
438 out_free_rsp:
439 	kfree(r->req.cqe);
440 out:
441 	return -ENOMEM;
442 }
443 
444 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
445 		struct nvmet_rdma_rsp *r)
446 {
447 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
448 				sizeof(*r->req.cqe), DMA_TO_DEVICE);
449 	kfree(r->req.cqe);
450 }
451 
452 static int
453 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
454 {
455 	struct nvmet_rdma_device *ndev = queue->dev;
456 	int nr_rsps = queue->recv_queue_size * 2;
457 	int ret = -EINVAL, i;
458 
459 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
460 			GFP_KERNEL);
461 	if (!queue->rsps)
462 		goto out;
463 
464 	for (i = 0; i < nr_rsps; i++) {
465 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
466 
467 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
468 		if (ret)
469 			goto out_free;
470 
471 		list_add_tail(&rsp->free_list, &queue->free_rsps);
472 	}
473 
474 	return 0;
475 
476 out_free:
477 	while (--i >= 0)
478 		nvmet_rdma_free_rsp(ndev, &queue->rsps[i]);
479 	kfree(queue->rsps);
480 out:
481 	return ret;
482 }
483 
484 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
485 {
486 	struct nvmet_rdma_device *ndev = queue->dev;
487 	int i, nr_rsps = queue->recv_queue_size * 2;
488 
489 	for (i = 0; i < nr_rsps; i++)
490 		nvmet_rdma_free_rsp(ndev, &queue->rsps[i]);
491 	kfree(queue->rsps);
492 }
493 
494 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
495 		struct nvmet_rdma_cmd *cmd)
496 {
497 	int ret;
498 
499 	ib_dma_sync_single_for_device(ndev->device,
500 		cmd->sge[0].addr, cmd->sge[0].length,
501 		DMA_FROM_DEVICE);
502 
503 	if (cmd->nsrq)
504 		ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
505 	else
506 		ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);
507 
508 	if (unlikely(ret))
509 		pr_err("post_recv cmd failed\n");
510 
511 	return ret;
512 }
513 
514 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
515 {
516 	spin_lock(&queue->rsp_wr_wait_lock);
517 	while (!list_empty(&queue->rsp_wr_wait_list)) {
518 		struct nvmet_rdma_rsp *rsp;
519 		bool ret;
520 
521 		rsp = list_entry(queue->rsp_wr_wait_list.next,
522 				struct nvmet_rdma_rsp, wait_list);
523 		list_del(&rsp->wait_list);
524 
525 		spin_unlock(&queue->rsp_wr_wait_lock);
526 		ret = nvmet_rdma_execute_command(rsp);
527 		spin_lock(&queue->rsp_wr_wait_lock);
528 
529 		if (!ret) {
530 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
531 			break;
532 		}
533 	}
534 	spin_unlock(&queue->rsp_wr_wait_lock);
535 }
536 
537 static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
538 {
539 	struct ib_mr_status mr_status;
540 	int ret;
541 	u16 status = 0;
542 
543 	ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
544 	if (ret) {
545 		pr_err("ib_check_mr_status failed, ret %d\n", ret);
546 		return NVME_SC_INVALID_PI;
547 	}
548 
549 	if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
550 		switch (mr_status.sig_err.err_type) {
551 		case IB_SIG_BAD_GUARD:
552 			status = NVME_SC_GUARD_CHECK;
553 			break;
554 		case IB_SIG_BAD_REFTAG:
555 			status = NVME_SC_REFTAG_CHECK;
556 			break;
557 		case IB_SIG_BAD_APPTAG:
558 			status = NVME_SC_APPTAG_CHECK;
559 			break;
560 		}
561 		pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
562 		       mr_status.sig_err.err_type,
563 		       mr_status.sig_err.expected,
564 		       mr_status.sig_err.actual);
565 	}
566 
567 	return status;
568 }
569 
570 static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
571 		struct nvme_command *cmd, struct ib_sig_domain *domain,
572 		u16 control, u8 pi_type)
573 {
574 	domain->sig_type = IB_SIG_TYPE_T10_DIF;
575 	domain->sig.dif.bg_type = IB_T10DIF_CRC;
576 	domain->sig.dif.pi_interval = 1 << bi->interval_exp;
577 	domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
578 	if (control & NVME_RW_PRINFO_PRCHK_REF)
579 		domain->sig.dif.ref_remap = true;
580 
581 	domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
582 	domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
583 	domain->sig.dif.app_escape = true;
584 	if (pi_type == NVME_NS_DPS_PI_TYPE3)
585 		domain->sig.dif.ref_escape = true;
586 }
587 
588 static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
589 				     struct ib_sig_attrs *sig_attrs)
590 {
591 	struct nvme_command *cmd = req->cmd;
592 	u16 control = le16_to_cpu(cmd->rw.control);
593 	u8 pi_type = req->ns->pi_type;
594 	struct blk_integrity *bi;
595 
596 	bi = bdev_get_integrity(req->ns->bdev);
597 
598 	memset(sig_attrs, 0, sizeof(*sig_attrs));
599 
600 	if (control & NVME_RW_PRINFO_PRACT) {
601 		/* for WRITE_INSERT/READ_STRIP no wire domain */
602 		sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
603 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
604 					  pi_type);
605 		/* Clear the PRACT bit since HCA will generate/verify the PI */
606 		control &= ~NVME_RW_PRINFO_PRACT;
607 		cmd->rw.control = cpu_to_le16(control);
608 		/* PI is added by the HW */
609 		req->transfer_len += req->metadata_len;
610 	} else {
611 		/* for WRITE_PASS/READ_PASS both wire/memory domains exist */
612 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
613 					  pi_type);
614 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
615 					  pi_type);
616 	}
617 
618 	if (control & NVME_RW_PRINFO_PRCHK_REF)
619 		sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
620 	if (control & NVME_RW_PRINFO_PRCHK_GUARD)
621 		sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
622 	if (control & NVME_RW_PRINFO_PRCHK_APP)
623 		sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
624 }
625 
626 static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
627 				  struct ib_sig_attrs *sig_attrs)
628 {
629 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
630 	struct nvmet_req *req = &rsp->req;
631 	int ret;
632 
633 	if (req->metadata_len)
634 		ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
635 			cm_id->port_num, req->sg, req->sg_cnt,
636 			req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
637 			addr, key, nvmet_data_dir(req));
638 	else
639 		ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
640 				       req->sg, req->sg_cnt, 0, addr, key,
641 				       nvmet_data_dir(req));
642 
643 	return ret;
644 }
645 
646 static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
647 {
648 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
649 	struct nvmet_req *req = &rsp->req;
650 
651 	if (req->metadata_len)
652 		rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
653 			cm_id->port_num, req->sg, req->sg_cnt,
654 			req->metadata_sg, req->metadata_sg_cnt,
655 			nvmet_data_dir(req));
656 	else
657 		rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
658 				    req->sg, req->sg_cnt, nvmet_data_dir(req));
659 }
660 
661 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
662 {
663 	struct nvmet_rdma_queue *queue = rsp->queue;
664 
665 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
666 
667 	if (rsp->n_rdma)
668 		nvmet_rdma_rw_ctx_destroy(rsp);
669 
670 	if (rsp->req.sg != rsp->cmd->inline_sg)
671 		nvmet_req_free_sgls(&rsp->req);
672 
673 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
674 		nvmet_rdma_process_wr_wait_list(queue);
675 
676 	nvmet_rdma_put_rsp(rsp);
677 }
678 
679 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
680 {
681 	if (queue->nvme_sq.ctrl) {
682 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
683 	} else {
684 		/*
685 		 * we didn't setup the controller yet in case
686 		 * of admin connect error, just disconnect and
687 		 * cleanup the queue
688 		 */
689 		nvmet_rdma_queue_disconnect(queue);
690 	}
691 }
692 
693 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
694 {
695 	struct nvmet_rdma_rsp *rsp =
696 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
697 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
698 
699 	nvmet_rdma_release_rsp(rsp);
700 
701 	if (unlikely(wc->status != IB_WC_SUCCESS &&
702 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
703 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
704 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
705 		nvmet_rdma_error_comp(queue);
706 	}
707 }
708 
709 static void nvmet_rdma_queue_response(struct nvmet_req *req)
710 {
711 	struct nvmet_rdma_rsp *rsp =
712 		container_of(req, struct nvmet_rdma_rsp, req);
713 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
714 	struct ib_send_wr *first_wr;
715 
716 	if (rsp->invalidate_rkey) {
717 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
718 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
719 	} else {
720 		rsp->send_wr.opcode = IB_WR_SEND;
721 	}
722 
723 	if (nvmet_rdma_need_data_out(rsp)) {
724 		if (rsp->req.metadata_len)
725 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
726 					cm_id->port_num, &rsp->write_cqe, NULL);
727 		else
728 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
729 					cm_id->port_num, NULL, &rsp->send_wr);
730 	} else {
731 		first_wr = &rsp->send_wr;
732 	}
733 
734 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
735 
736 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
737 		rsp->send_sge.addr, rsp->send_sge.length,
738 		DMA_TO_DEVICE);
739 
740 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
741 		pr_err("sending cmd response failed\n");
742 		nvmet_rdma_release_rsp(rsp);
743 	}
744 }
745 
746 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
747 {
748 	struct nvmet_rdma_rsp *rsp =
749 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
750 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
751 	u16 status = 0;
752 
753 	WARN_ON(rsp->n_rdma <= 0);
754 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
755 	rsp->n_rdma = 0;
756 
757 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
758 		nvmet_rdma_rw_ctx_destroy(rsp);
759 		nvmet_req_uninit(&rsp->req);
760 		nvmet_rdma_release_rsp(rsp);
761 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
762 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
763 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
764 			nvmet_rdma_error_comp(queue);
765 		}
766 		return;
767 	}
768 
769 	if (rsp->req.metadata_len)
770 		status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
771 	nvmet_rdma_rw_ctx_destroy(rsp);
772 
773 	if (unlikely(status))
774 		nvmet_req_complete(&rsp->req, status);
775 	else
776 		rsp->req.execute(&rsp->req);
777 }
778 
779 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
780 {
781 	struct nvmet_rdma_rsp *rsp =
782 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
783 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
784 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
785 	u16 status;
786 
787 	if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
788 		return;
789 
790 	WARN_ON(rsp->n_rdma <= 0);
791 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
792 	rsp->n_rdma = 0;
793 
794 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
795 		nvmet_rdma_rw_ctx_destroy(rsp);
796 		nvmet_req_uninit(&rsp->req);
797 		nvmet_rdma_release_rsp(rsp);
798 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
799 			pr_info("RDMA WRITE for CQE failed with status %s (%d).\n",
800 				ib_wc_status_msg(wc->status), wc->status);
801 			nvmet_rdma_error_comp(queue);
802 		}
803 		return;
804 	}
805 
806 	/*
807 	 * Upon RDMA completion check the signature status
808 	 * - if succeeded send good NVMe response
809 	 * - if failed send bad NVMe response with appropriate error
810 	 */
811 	status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
812 	if (unlikely(status))
813 		rsp->req.cqe->status = cpu_to_le16(status << 1);
814 	nvmet_rdma_rw_ctx_destroy(rsp);
815 
816 	if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
817 		pr_err("sending cmd response failed\n");
818 		nvmet_rdma_release_rsp(rsp);
819 	}
820 }
821 
822 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
823 		u64 off)
824 {
825 	int sg_count = num_pages(len);
826 	struct scatterlist *sg;
827 	int i;
828 
829 	sg = rsp->cmd->inline_sg;
830 	for (i = 0; i < sg_count; i++, sg++) {
831 		if (i < sg_count - 1)
832 			sg_unmark_end(sg);
833 		else
834 			sg_mark_end(sg);
835 		sg->offset = off;
836 		sg->length = min_t(int, len, PAGE_SIZE - off);
837 		len -= sg->length;
838 		if (!i)
839 			off = 0;
840 	}
841 
842 	rsp->req.sg = rsp->cmd->inline_sg;
843 	rsp->req.sg_cnt = sg_count;
844 }
845 
846 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
847 {
848 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
849 	u64 off = le64_to_cpu(sgl->addr);
850 	u32 len = le32_to_cpu(sgl->length);
851 
852 	if (!nvme_is_write(rsp->req.cmd)) {
853 		rsp->req.error_loc =
854 			offsetof(struct nvme_common_command, opcode);
855 		return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
856 	}
857 
858 	if (off + len > rsp->queue->dev->inline_data_size) {
859 		pr_err("invalid inline data offset!\n");
860 		return NVME_SC_SGL_INVALID_OFFSET | NVME_STATUS_DNR;
861 	}
862 
863 	/* no data command? */
864 	if (!len)
865 		return 0;
866 
867 	nvmet_rdma_use_inline_sg(rsp, len, off);
868 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
869 	rsp->req.transfer_len += len;
870 	return 0;
871 }
872 
873 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
874 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
875 {
876 	u64 addr = le64_to_cpu(sgl->addr);
877 	u32 key = get_unaligned_le32(sgl->key);
878 	struct ib_sig_attrs sig_attrs;
879 	int ret;
880 
881 	rsp->req.transfer_len = get_unaligned_le24(sgl->length);
882 
883 	/* no data command? */
884 	if (!rsp->req.transfer_len)
885 		return 0;
886 
887 	if (rsp->req.metadata_len)
888 		nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);
889 
890 	ret = nvmet_req_alloc_sgls(&rsp->req);
891 	if (unlikely(ret < 0))
892 		goto error_out;
893 
894 	ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
895 	if (unlikely(ret < 0))
896 		goto error_out;
897 	rsp->n_rdma += ret;
898 
899 	if (invalidate)
900 		rsp->invalidate_rkey = key;
901 
902 	return 0;
903 
904 error_out:
905 	rsp->req.transfer_len = 0;
906 	return NVME_SC_INTERNAL;
907 }
908 
909 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
910 {
911 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
912 
913 	switch (sgl->type >> 4) {
914 	case NVME_SGL_FMT_DATA_DESC:
915 		switch (sgl->type & 0xf) {
916 		case NVME_SGL_FMT_OFFSET:
917 			return nvmet_rdma_map_sgl_inline(rsp);
918 		default:
919 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
920 			rsp->req.error_loc =
921 				offsetof(struct nvme_common_command, dptr);
922 			return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
923 		}
924 	case NVME_KEY_SGL_FMT_DATA_DESC:
925 		switch (sgl->type & 0xf) {
926 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
927 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
928 		case NVME_SGL_FMT_ADDRESS:
929 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
930 		default:
931 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
932 			rsp->req.error_loc =
933 				offsetof(struct nvme_common_command, dptr);
934 			return NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
935 		}
936 	default:
937 		pr_err("invalid SGL type: %#x\n", sgl->type);
938 		rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
939 		return NVME_SC_SGL_INVALID_TYPE | NVME_STATUS_DNR;
940 	}
941 }
942 
943 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
944 {
945 	struct nvmet_rdma_queue *queue = rsp->queue;
946 
947 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
948 			&queue->sq_wr_avail) < 0)) {
949 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
950 				1 + rsp->n_rdma, queue->idx,
951 				queue->nvme_sq.ctrl->cntlid);
952 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
953 		return false;
954 	}
955 
956 	if (nvmet_rdma_need_data_in(rsp)) {
957 		if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
958 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
959 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
960 	} else {
961 		rsp->req.execute(&rsp->req);
962 	}
963 
964 	return true;
965 }
966 
967 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
968 		struct nvmet_rdma_rsp *cmd)
969 {
970 	u16 status;
971 
972 	ib_dma_sync_single_for_cpu(queue->dev->device,
973 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
974 		DMA_FROM_DEVICE);
975 	ib_dma_sync_single_for_cpu(queue->dev->device,
976 		cmd->send_sge.addr, cmd->send_sge.length,
977 		DMA_TO_DEVICE);
978 
979 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
980 			&queue->nvme_sq, &nvmet_rdma_ops))
981 		return;
982 
983 	status = nvmet_rdma_map_sgl(cmd);
984 	if (status)
985 		goto out_err;
986 
987 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
988 		spin_lock(&queue->rsp_wr_wait_lock);
989 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
990 		spin_unlock(&queue->rsp_wr_wait_lock);
991 	}
992 
993 	return;
994 
995 out_err:
996 	nvmet_req_complete(&cmd->req, status);
997 }
998 
999 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1000 {
1001 	struct nvmet_rdma_cmd *cmd =
1002 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
1003 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
1004 	struct nvmet_rdma_rsp *rsp;
1005 
1006 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1007 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
1008 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
1009 				wc->wr_cqe, ib_wc_status_msg(wc->status),
1010 				wc->status);
1011 			nvmet_rdma_error_comp(queue);
1012 		}
1013 		return;
1014 	}
1015 
1016 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
1017 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
1018 		nvmet_rdma_error_comp(queue);
1019 		return;
1020 	}
1021 
1022 	cmd->queue = queue;
1023 	rsp = nvmet_rdma_get_rsp(queue);
1024 	if (unlikely(!rsp)) {
1025 		/*
1026 		 * we get here only under memory pressure,
1027 		 * silently drop and have the host retry
1028 		 * as we can't even fail it.
1029 		 */
1030 		nvmet_rdma_post_recv(queue->dev, cmd);
1031 		return;
1032 	}
1033 	rsp->queue = queue;
1034 	rsp->cmd = cmd;
1035 	rsp->flags = 0;
1036 	rsp->req.cmd = cmd->nvme_cmd;
1037 	rsp->req.port = queue->port;
1038 	rsp->n_rdma = 0;
1039 	rsp->invalidate_rkey = 0;
1040 
1041 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
1042 		unsigned long flags;
1043 
1044 		spin_lock_irqsave(&queue->state_lock, flags);
1045 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
1046 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
1047 		else
1048 			nvmet_rdma_put_rsp(rsp);
1049 		spin_unlock_irqrestore(&queue->state_lock, flags);
1050 		return;
1051 	}
1052 
1053 	nvmet_rdma_handle_command(queue, rsp);
1054 }
1055 
1056 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
1057 {
1058 	nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
1059 			     false);
1060 	ib_destroy_srq(nsrq->srq);
1061 
1062 	kfree(nsrq);
1063 }
1064 
1065 static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
1066 {
1067 	int i;
1068 
1069 	if (!ndev->srqs)
1070 		return;
1071 
1072 	for (i = 0; i < ndev->srq_count; i++)
1073 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1074 
1075 	kfree(ndev->srqs);
1076 }
1077 
1078 static struct nvmet_rdma_srq *
1079 nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
1080 {
1081 	struct ib_srq_init_attr srq_attr = { NULL, };
1082 	size_t srq_size = ndev->srq_size;
1083 	struct nvmet_rdma_srq *nsrq;
1084 	struct ib_srq *srq;
1085 	int ret, i;
1086 
1087 	nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
1088 	if (!nsrq)
1089 		return ERR_PTR(-ENOMEM);
1090 
1091 	srq_attr.attr.max_wr = srq_size;
1092 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
1093 	srq_attr.attr.srq_limit = 0;
1094 	srq_attr.srq_type = IB_SRQT_BASIC;
1095 	srq = ib_create_srq(ndev->pd, &srq_attr);
1096 	if (IS_ERR(srq)) {
1097 		ret = PTR_ERR(srq);
1098 		goto out_free;
1099 	}
1100 
1101 	nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
1102 	if (IS_ERR(nsrq->cmds)) {
1103 		ret = PTR_ERR(nsrq->cmds);
1104 		goto out_destroy_srq;
1105 	}
1106 
1107 	nsrq->srq = srq;
1108 	nsrq->ndev = ndev;
1109 
1110 	for (i = 0; i < srq_size; i++) {
1111 		nsrq->cmds[i].nsrq = nsrq;
1112 		ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
1113 		if (ret)
1114 			goto out_free_cmds;
1115 	}
1116 
1117 	return nsrq;
1118 
1119 out_free_cmds:
1120 	nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
1121 out_destroy_srq:
1122 	ib_destroy_srq(srq);
1123 out_free:
1124 	kfree(nsrq);
1125 	return ERR_PTR(ret);
1126 }
1127 
1128 static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
1129 {
1130 	int i, ret;
1131 
1132 	if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
1133 		/*
1134 		 * If SRQs aren't supported we just go ahead and use normal
1135 		 * non-shared receive queues.
1136 		 */
1137 		pr_info("SRQ requested but not supported.\n");
1138 		return 0;
1139 	}
1140 
1141 	ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
1142 			     nvmet_rdma_srq_size);
1143 	ndev->srq_count = min(ndev->device->num_comp_vectors,
1144 			      ndev->device->attrs.max_srq);
1145 
1146 	ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
1147 	if (!ndev->srqs)
1148 		return -ENOMEM;
1149 
1150 	for (i = 0; i < ndev->srq_count; i++) {
1151 		ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
1152 		if (IS_ERR(ndev->srqs[i])) {
1153 			ret = PTR_ERR(ndev->srqs[i]);
1154 			goto err_srq;
1155 		}
1156 	}
1157 
1158 	return 0;
1159 
1160 err_srq:
1161 	while (--i >= 0)
1162 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1163 	kfree(ndev->srqs);
1164 	return ret;
1165 }
1166 
1167 static void nvmet_rdma_free_dev(struct kref *ref)
1168 {
1169 	struct nvmet_rdma_device *ndev =
1170 		container_of(ref, struct nvmet_rdma_device, ref);
1171 
1172 	mutex_lock(&device_list_mutex);
1173 	list_del(&ndev->entry);
1174 	mutex_unlock(&device_list_mutex);
1175 
1176 	nvmet_rdma_destroy_srqs(ndev);
1177 	ib_dealloc_pd(ndev->pd);
1178 
1179 	kfree(ndev);
1180 }
1181 
1182 static struct nvmet_rdma_device *
1183 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
1184 {
1185 	struct nvmet_rdma_port *port = cm_id->context;
1186 	struct nvmet_port *nport = port->nport;
1187 	struct nvmet_rdma_device *ndev;
1188 	int inline_page_count;
1189 	int inline_sge_count;
1190 	int ret;
1191 
1192 	mutex_lock(&device_list_mutex);
1193 	list_for_each_entry(ndev, &device_list, entry) {
1194 		if (ndev->device->node_guid == cm_id->device->node_guid &&
1195 		    kref_get_unless_zero(&ndev->ref))
1196 			goto out_unlock;
1197 	}
1198 
1199 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
1200 	if (!ndev)
1201 		goto out_err;
1202 
1203 	inline_page_count = num_pages(nport->inline_data_size);
1204 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
1205 				cm_id->device->attrs.max_recv_sge) - 1;
1206 	if (inline_page_count > inline_sge_count) {
1207 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
1208 			nport->inline_data_size, cm_id->device->name,
1209 			inline_sge_count * PAGE_SIZE);
1210 		nport->inline_data_size = inline_sge_count * PAGE_SIZE;
1211 		inline_page_count = inline_sge_count;
1212 	}
1213 	ndev->inline_data_size = nport->inline_data_size;
1214 	ndev->inline_page_count = inline_page_count;
1215 
1216 	if (nport->pi_enable && !(cm_id->device->attrs.kernel_cap_flags &
1217 				  IBK_INTEGRITY_HANDOVER)) {
1218 		pr_warn("T10-PI is not supported by device %s. Disabling it\n",
1219 			cm_id->device->name);
1220 		nport->pi_enable = false;
1221 	}
1222 
1223 	ndev->device = cm_id->device;
1224 	kref_init(&ndev->ref);
1225 
1226 	ndev->pd = ib_alloc_pd(ndev->device, 0);
1227 	if (IS_ERR(ndev->pd))
1228 		goto out_free_dev;
1229 
1230 	if (nvmet_rdma_use_srq) {
1231 		ret = nvmet_rdma_init_srqs(ndev);
1232 		if (ret)
1233 			goto out_free_pd;
1234 	}
1235 
1236 	list_add(&ndev->entry, &device_list);
1237 out_unlock:
1238 	mutex_unlock(&device_list_mutex);
1239 	pr_debug("added %s.\n", ndev->device->name);
1240 	return ndev;
1241 
1242 out_free_pd:
1243 	ib_dealloc_pd(ndev->pd);
1244 out_free_dev:
1245 	kfree(ndev);
1246 out_err:
1247 	mutex_unlock(&device_list_mutex);
1248 	return NULL;
1249 }
1250 
1251 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
1252 {
1253 	struct ib_qp_init_attr qp_attr = { };
1254 	struct nvmet_rdma_device *ndev = queue->dev;
1255 	int nr_cqe, ret, i, factor;
1256 
1257 	/*
1258 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
1259 	 */
1260 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1261 
1262 	queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
1263 				   queue->comp_vector, IB_POLL_WORKQUEUE);
1264 	if (IS_ERR(queue->cq)) {
1265 		ret = PTR_ERR(queue->cq);
1266 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1267 		       nr_cqe + 1, ret);
1268 		goto out;
1269 	}
1270 
1271 	qp_attr.qp_context = queue;
1272 	qp_attr.event_handler = nvmet_rdma_qp_event;
1273 	qp_attr.send_cq = queue->cq;
1274 	qp_attr.recv_cq = queue->cq;
1275 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1276 	qp_attr.qp_type = IB_QPT_RC;
1277 	/* +1 for drain */
1278 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1279 	factor = rdma_rw_mr_factor(ndev->device, queue->cm_id->port_num,
1280 				   1 << NVMET_RDMA_MAX_MDTS);
1281 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * factor;
1282 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1283 					ndev->device->attrs.max_send_sge);
1284 
1285 	if (queue->nsrq) {
1286 		qp_attr.srq = queue->nsrq->srq;
1287 	} else {
1288 		/* +1 for drain */
1289 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1290 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1291 	}
1292 
1293 	if (queue->port->pi_enable && queue->host_qid)
1294 		qp_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
1295 
1296 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1297 	if (ret) {
1298 		pr_err("failed to create_qp ret= %d\n", ret);
1299 		goto err_destroy_cq;
1300 	}
1301 	queue->qp = queue->cm_id->qp;
1302 
1303 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1304 
1305 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1306 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1307 		 qp_attr.cap.max_send_wr, queue->cm_id);
1308 
1309 	if (!queue->nsrq) {
1310 		for (i = 0; i < queue->recv_queue_size; i++) {
1311 			queue->cmds[i].queue = queue;
1312 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1313 			if (ret)
1314 				goto err_destroy_qp;
1315 		}
1316 	}
1317 
1318 out:
1319 	return ret;
1320 
1321 err_destroy_qp:
1322 	rdma_destroy_qp(queue->cm_id);
1323 err_destroy_cq:
1324 	ib_cq_pool_put(queue->cq, nr_cqe + 1);
1325 	goto out;
1326 }
1327 
1328 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1329 {
1330 	ib_drain_qp(queue->qp);
1331 	if (queue->cm_id)
1332 		rdma_destroy_id(queue->cm_id);
1333 	ib_destroy_qp(queue->qp);
1334 	ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
1335 		       queue->send_queue_size + 1);
1336 }
1337 
1338 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1339 {
1340 	pr_debug("freeing queue %d\n", queue->idx);
1341 
1342 	nvmet_sq_destroy(&queue->nvme_sq);
1343 
1344 	nvmet_rdma_destroy_queue_ib(queue);
1345 	if (!queue->nsrq) {
1346 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1347 				queue->recv_queue_size,
1348 				!queue->host_qid);
1349 	}
1350 	nvmet_rdma_free_rsps(queue);
1351 	ida_free(&nvmet_rdma_queue_ida, queue->idx);
1352 	kfree(queue);
1353 }
1354 
1355 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1356 {
1357 	struct nvmet_rdma_queue *queue =
1358 		container_of(w, struct nvmet_rdma_queue, release_work);
1359 	struct nvmet_rdma_device *dev = queue->dev;
1360 
1361 	nvmet_rdma_free_queue(queue);
1362 
1363 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1364 }
1365 
1366 static int
1367 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1368 				struct nvmet_rdma_queue *queue)
1369 {
1370 	struct nvme_rdma_cm_req *req;
1371 
1372 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1373 	if (!req || conn->private_data_len == 0)
1374 		return NVME_RDMA_CM_INVALID_LEN;
1375 
1376 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1377 		return NVME_RDMA_CM_INVALID_RECFMT;
1378 
1379 	queue->host_qid = le16_to_cpu(req->qid);
1380 
1381 	/*
1382 	 * req->hsqsize corresponds to our recv queue size plus 1
1383 	 * req->hrqsize corresponds to our send queue size
1384 	 */
1385 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1386 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1387 
1388 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1389 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1390 
1391 	/* XXX: Should we enforce some kind of max for IO queues? */
1392 
1393 	return 0;
1394 }
1395 
1396 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1397 				enum nvme_rdma_cm_status status)
1398 {
1399 	struct nvme_rdma_cm_rej rej;
1400 
1401 	pr_debug("rejecting connect request: status %d (%s)\n",
1402 		 status, nvme_rdma_cm_msg(status));
1403 
1404 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1405 	rej.sts = cpu_to_le16(status);
1406 
1407 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej),
1408 			   IB_CM_REJ_CONSUMER_DEFINED);
1409 }
1410 
1411 static struct nvmet_rdma_queue *
1412 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1413 		struct rdma_cm_id *cm_id,
1414 		struct rdma_cm_event *event)
1415 {
1416 	struct nvmet_rdma_port *port = cm_id->context;
1417 	struct nvmet_rdma_queue *queue;
1418 	int ret;
1419 
1420 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1421 	if (!queue) {
1422 		ret = NVME_RDMA_CM_NO_RSC;
1423 		goto out_reject;
1424 	}
1425 
1426 	ret = nvmet_sq_init(&queue->nvme_sq);
1427 	if (ret) {
1428 		ret = NVME_RDMA_CM_NO_RSC;
1429 		goto out_free_queue;
1430 	}
1431 
1432 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1433 	if (ret)
1434 		goto out_destroy_sq;
1435 
1436 	/*
1437 	 * Schedules the actual release because calling rdma_destroy_id from
1438 	 * inside a CM callback would trigger a deadlock. (great API design..)
1439 	 */
1440 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1441 	queue->dev = ndev;
1442 	queue->cm_id = cm_id;
1443 	queue->port = port->nport;
1444 
1445 	spin_lock_init(&queue->state_lock);
1446 	queue->state = NVMET_RDMA_Q_CONNECTING;
1447 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1448 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1449 	spin_lock_init(&queue->rsp_wr_wait_lock);
1450 	INIT_LIST_HEAD(&queue->free_rsps);
1451 	spin_lock_init(&queue->rsps_lock);
1452 	INIT_LIST_HEAD(&queue->queue_list);
1453 
1454 	queue->idx = ida_alloc(&nvmet_rdma_queue_ida, GFP_KERNEL);
1455 	if (queue->idx < 0) {
1456 		ret = NVME_RDMA_CM_NO_RSC;
1457 		goto out_destroy_sq;
1458 	}
1459 
1460 	/*
1461 	 * Spread the io queues across completion vectors,
1462 	 * but still keep all admin queues on vector 0.
1463 	 */
1464 	queue->comp_vector = !queue->host_qid ? 0 :
1465 		queue->idx % ndev->device->num_comp_vectors;
1466 
1467 
1468 	ret = nvmet_rdma_alloc_rsps(queue);
1469 	if (ret) {
1470 		ret = NVME_RDMA_CM_NO_RSC;
1471 		goto out_ida_remove;
1472 	}
1473 
1474 	if (ndev->srqs) {
1475 		queue->nsrq = ndev->srqs[queue->comp_vector % ndev->srq_count];
1476 	} else {
1477 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1478 				queue->recv_queue_size,
1479 				!queue->host_qid);
1480 		if (IS_ERR(queue->cmds)) {
1481 			ret = NVME_RDMA_CM_NO_RSC;
1482 			goto out_free_responses;
1483 		}
1484 	}
1485 
1486 	ret = nvmet_rdma_create_queue_ib(queue);
1487 	if (ret) {
1488 		pr_err("%s: creating RDMA queue failed (%d).\n",
1489 			__func__, ret);
1490 		ret = NVME_RDMA_CM_NO_RSC;
1491 		goto out_free_cmds;
1492 	}
1493 
1494 	return queue;
1495 
1496 out_free_cmds:
1497 	if (!queue->nsrq) {
1498 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1499 				queue->recv_queue_size,
1500 				!queue->host_qid);
1501 	}
1502 out_free_responses:
1503 	nvmet_rdma_free_rsps(queue);
1504 out_ida_remove:
1505 	ida_free(&nvmet_rdma_queue_ida, queue->idx);
1506 out_destroy_sq:
1507 	nvmet_sq_destroy(&queue->nvme_sq);
1508 out_free_queue:
1509 	kfree(queue);
1510 out_reject:
1511 	nvmet_rdma_cm_reject(cm_id, ret);
1512 	return NULL;
1513 }
1514 
1515 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1516 {
1517 	struct nvmet_rdma_queue *queue = priv;
1518 
1519 	switch (event->event) {
1520 	case IB_EVENT_COMM_EST:
1521 		rdma_notify(queue->cm_id, event->event);
1522 		break;
1523 	case IB_EVENT_QP_LAST_WQE_REACHED:
1524 		pr_debug("received last WQE reached event for queue=0x%p\n",
1525 			 queue);
1526 		break;
1527 	default:
1528 		pr_err("received IB QP event: %s (%d)\n",
1529 		       ib_event_msg(event->event), event->event);
1530 		break;
1531 	}
1532 }
1533 
1534 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1535 		struct nvmet_rdma_queue *queue,
1536 		struct rdma_conn_param *p)
1537 {
1538 	struct rdma_conn_param  param = { };
1539 	struct nvme_rdma_cm_rep priv = { };
1540 	int ret = -ENOMEM;
1541 
1542 	param.rnr_retry_count = 7;
1543 	param.flow_control = 1;
1544 	param.initiator_depth = min_t(u8, p->initiator_depth,
1545 		queue->dev->device->attrs.max_qp_init_rd_atom);
1546 	param.private_data = &priv;
1547 	param.private_data_len = sizeof(priv);
1548 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1549 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1550 
1551 	ret = rdma_accept(cm_id, &param);
1552 	if (ret)
1553 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1554 
1555 	return ret;
1556 }
1557 
1558 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1559 		struct rdma_cm_event *event)
1560 {
1561 	struct nvmet_rdma_device *ndev;
1562 	struct nvmet_rdma_queue *queue;
1563 	int ret = -EINVAL;
1564 
1565 	ndev = nvmet_rdma_find_get_device(cm_id);
1566 	if (!ndev) {
1567 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1568 		return -ECONNREFUSED;
1569 	}
1570 
1571 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1572 	if (!queue) {
1573 		ret = -ENOMEM;
1574 		goto put_device;
1575 	}
1576 
1577 	if (queue->host_qid == 0) {
1578 		struct nvmet_rdma_queue *q;
1579 		int pending = 0;
1580 
1581 		/* Check for pending controller teardown */
1582 		mutex_lock(&nvmet_rdma_queue_mutex);
1583 		list_for_each_entry(q, &nvmet_rdma_queue_list, queue_list) {
1584 			if (q->nvme_sq.ctrl == queue->nvme_sq.ctrl &&
1585 			    q->state == NVMET_RDMA_Q_DISCONNECTING)
1586 				pending++;
1587 		}
1588 		mutex_unlock(&nvmet_rdma_queue_mutex);
1589 		if (pending > NVMET_RDMA_BACKLOG)
1590 			return NVME_SC_CONNECT_CTRL_BUSY;
1591 	}
1592 
1593 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1594 	if (ret) {
1595 		/*
1596 		 * Don't destroy the cm_id in free path, as we implicitly
1597 		 * destroy the cm_id here with non-zero ret code.
1598 		 */
1599 		queue->cm_id = NULL;
1600 		goto free_queue;
1601 	}
1602 
1603 	mutex_lock(&nvmet_rdma_queue_mutex);
1604 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1605 	mutex_unlock(&nvmet_rdma_queue_mutex);
1606 
1607 	return 0;
1608 
1609 free_queue:
1610 	nvmet_rdma_free_queue(queue);
1611 put_device:
1612 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1613 
1614 	return ret;
1615 }
1616 
1617 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1618 {
1619 	unsigned long flags;
1620 
1621 	spin_lock_irqsave(&queue->state_lock, flags);
1622 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1623 		pr_warn("trying to establish a connected queue\n");
1624 		goto out_unlock;
1625 	}
1626 	queue->state = NVMET_RDMA_Q_LIVE;
1627 
1628 	while (!list_empty(&queue->rsp_wait_list)) {
1629 		struct nvmet_rdma_rsp *cmd;
1630 
1631 		cmd = list_first_entry(&queue->rsp_wait_list,
1632 					struct nvmet_rdma_rsp, wait_list);
1633 		list_del(&cmd->wait_list);
1634 
1635 		spin_unlock_irqrestore(&queue->state_lock, flags);
1636 		nvmet_rdma_handle_command(queue, cmd);
1637 		spin_lock_irqsave(&queue->state_lock, flags);
1638 	}
1639 
1640 out_unlock:
1641 	spin_unlock_irqrestore(&queue->state_lock, flags);
1642 }
1643 
1644 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1645 {
1646 	bool disconnect = false;
1647 	unsigned long flags;
1648 
1649 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1650 
1651 	spin_lock_irqsave(&queue->state_lock, flags);
1652 	switch (queue->state) {
1653 	case NVMET_RDMA_Q_CONNECTING:
1654 		while (!list_empty(&queue->rsp_wait_list)) {
1655 			struct nvmet_rdma_rsp *rsp;
1656 
1657 			rsp = list_first_entry(&queue->rsp_wait_list,
1658 					       struct nvmet_rdma_rsp,
1659 					       wait_list);
1660 			list_del(&rsp->wait_list);
1661 			nvmet_rdma_put_rsp(rsp);
1662 		}
1663 		fallthrough;
1664 	case NVMET_RDMA_Q_LIVE:
1665 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1666 		disconnect = true;
1667 		break;
1668 	case NVMET_RDMA_Q_DISCONNECTING:
1669 		break;
1670 	}
1671 	spin_unlock_irqrestore(&queue->state_lock, flags);
1672 
1673 	if (disconnect) {
1674 		rdma_disconnect(queue->cm_id);
1675 		queue_work(nvmet_wq, &queue->release_work);
1676 	}
1677 }
1678 
1679 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1680 {
1681 	bool disconnect = false;
1682 
1683 	mutex_lock(&nvmet_rdma_queue_mutex);
1684 	if (!list_empty(&queue->queue_list)) {
1685 		list_del_init(&queue->queue_list);
1686 		disconnect = true;
1687 	}
1688 	mutex_unlock(&nvmet_rdma_queue_mutex);
1689 
1690 	if (disconnect)
1691 		__nvmet_rdma_queue_disconnect(queue);
1692 }
1693 
1694 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1695 		struct nvmet_rdma_queue *queue)
1696 {
1697 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1698 
1699 	mutex_lock(&nvmet_rdma_queue_mutex);
1700 	if (!list_empty(&queue->queue_list))
1701 		list_del_init(&queue->queue_list);
1702 	mutex_unlock(&nvmet_rdma_queue_mutex);
1703 
1704 	pr_err("failed to connect queue %d\n", queue->idx);
1705 	queue_work(nvmet_wq, &queue->release_work);
1706 }
1707 
1708 /**
1709  * nvmet_rdma_device_removal() - Handle RDMA device removal
1710  * @cm_id:	rdma_cm id, used for nvmet port
1711  * @queue:      nvmet rdma queue (cm id qp_context)
1712  *
1713  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1714  * to unplug. Note that this event can be generated on a normal
1715  * queue cm_id and/or a device bound listener cm_id (where in this
1716  * case queue will be null).
1717  *
1718  * We registered an ib_client to handle device removal for queues,
1719  * so we only need to handle the listening port cm_ids. In this case
1720  * we nullify the priv to prevent double cm_id destruction and destroying
1721  * the cm_id implicitely by returning a non-zero rc to the callout.
1722  */
1723 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1724 		struct nvmet_rdma_queue *queue)
1725 {
1726 	struct nvmet_rdma_port *port;
1727 
1728 	if (queue) {
1729 		/*
1730 		 * This is a queue cm_id. we have registered
1731 		 * an ib_client to handle queues removal
1732 		 * so don't interfear and just return.
1733 		 */
1734 		return 0;
1735 	}
1736 
1737 	port = cm_id->context;
1738 
1739 	/*
1740 	 * This is a listener cm_id. Make sure that
1741 	 * future remove_port won't invoke a double
1742 	 * cm_id destroy. use atomic xchg to make sure
1743 	 * we don't compete with remove_port.
1744 	 */
1745 	if (xchg(&port->cm_id, NULL) != cm_id)
1746 		return 0;
1747 
1748 	/*
1749 	 * We need to return 1 so that the core will destroy
1750 	 * it's own ID.  What a great API design..
1751 	 */
1752 	return 1;
1753 }
1754 
1755 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1756 		struct rdma_cm_event *event)
1757 {
1758 	struct nvmet_rdma_queue *queue = NULL;
1759 	int ret = 0;
1760 
1761 	if (cm_id->qp)
1762 		queue = cm_id->qp->qp_context;
1763 
1764 	pr_debug("%s (%d): status %d id %p\n",
1765 		rdma_event_msg(event->event), event->event,
1766 		event->status, cm_id);
1767 
1768 	switch (event->event) {
1769 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1770 		ret = nvmet_rdma_queue_connect(cm_id, event);
1771 		break;
1772 	case RDMA_CM_EVENT_ESTABLISHED:
1773 		nvmet_rdma_queue_established(queue);
1774 		break;
1775 	case RDMA_CM_EVENT_ADDR_CHANGE:
1776 		if (!queue) {
1777 			struct nvmet_rdma_port *port = cm_id->context;
1778 
1779 			queue_delayed_work(nvmet_wq, &port->repair_work, 0);
1780 			break;
1781 		}
1782 		fallthrough;
1783 	case RDMA_CM_EVENT_DISCONNECTED:
1784 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1785 		nvmet_rdma_queue_disconnect(queue);
1786 		break;
1787 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1788 		ret = nvmet_rdma_device_removal(cm_id, queue);
1789 		break;
1790 	case RDMA_CM_EVENT_REJECTED:
1791 		pr_debug("Connection rejected: %s\n",
1792 			 rdma_reject_msg(cm_id, event->status));
1793 		fallthrough;
1794 	case RDMA_CM_EVENT_UNREACHABLE:
1795 	case RDMA_CM_EVENT_CONNECT_ERROR:
1796 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1797 		break;
1798 	default:
1799 		pr_err("received unrecognized RDMA CM event %d\n",
1800 			event->event);
1801 		break;
1802 	}
1803 
1804 	return ret;
1805 }
1806 
1807 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1808 {
1809 	struct nvmet_rdma_queue *queue, *n;
1810 
1811 	mutex_lock(&nvmet_rdma_queue_mutex);
1812 	list_for_each_entry_safe(queue, n, &nvmet_rdma_queue_list, queue_list) {
1813 		if (queue->nvme_sq.ctrl != ctrl)
1814 			continue;
1815 		list_del_init(&queue->queue_list);
1816 		__nvmet_rdma_queue_disconnect(queue);
1817 	}
1818 	mutex_unlock(&nvmet_rdma_queue_mutex);
1819 }
1820 
1821 static void nvmet_rdma_destroy_port_queues(struct nvmet_rdma_port *port)
1822 {
1823 	struct nvmet_rdma_queue *queue, *tmp;
1824 	struct nvmet_port *nport = port->nport;
1825 
1826 	mutex_lock(&nvmet_rdma_queue_mutex);
1827 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1828 				 queue_list) {
1829 		if (queue->port != nport)
1830 			continue;
1831 
1832 		list_del_init(&queue->queue_list);
1833 		__nvmet_rdma_queue_disconnect(queue);
1834 	}
1835 	mutex_unlock(&nvmet_rdma_queue_mutex);
1836 }
1837 
1838 static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
1839 {
1840 	struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
1841 
1842 	if (cm_id)
1843 		rdma_destroy_id(cm_id);
1844 
1845 	/*
1846 	 * Destroy the remaining queues, which are not belong to any
1847 	 * controller yet. Do it here after the RDMA-CM was destroyed
1848 	 * guarantees that no new queue will be created.
1849 	 */
1850 	nvmet_rdma_destroy_port_queues(port);
1851 }
1852 
1853 static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
1854 {
1855 	struct sockaddr *addr = (struct sockaddr *)&port->addr;
1856 	struct rdma_cm_id *cm_id;
1857 	int ret;
1858 
1859 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1860 			RDMA_PS_TCP, IB_QPT_RC);
1861 	if (IS_ERR(cm_id)) {
1862 		pr_err("CM ID creation failed\n");
1863 		return PTR_ERR(cm_id);
1864 	}
1865 
1866 	/*
1867 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1868 	 * at the same time.
1869 	 */
1870 	ret = rdma_set_afonly(cm_id, 1);
1871 	if (ret) {
1872 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1873 		goto out_destroy_id;
1874 	}
1875 
1876 	ret = rdma_bind_addr(cm_id, addr);
1877 	if (ret) {
1878 		pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
1879 		goto out_destroy_id;
1880 	}
1881 
1882 	ret = rdma_listen(cm_id, NVMET_RDMA_BACKLOG);
1883 	if (ret) {
1884 		pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
1885 		goto out_destroy_id;
1886 	}
1887 
1888 	port->cm_id = cm_id;
1889 	return 0;
1890 
1891 out_destroy_id:
1892 	rdma_destroy_id(cm_id);
1893 	return ret;
1894 }
1895 
1896 static void nvmet_rdma_repair_port_work(struct work_struct *w)
1897 {
1898 	struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
1899 			struct nvmet_rdma_port, repair_work);
1900 	int ret;
1901 
1902 	nvmet_rdma_disable_port(port);
1903 	ret = nvmet_rdma_enable_port(port);
1904 	if (ret)
1905 		queue_delayed_work(nvmet_wq, &port->repair_work, 5 * HZ);
1906 }
1907 
1908 static int nvmet_rdma_add_port(struct nvmet_port *nport)
1909 {
1910 	struct nvmet_rdma_port *port;
1911 	__kernel_sa_family_t af;
1912 	int ret;
1913 
1914 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1915 	if (!port)
1916 		return -ENOMEM;
1917 
1918 	nport->priv = port;
1919 	port->nport = nport;
1920 	INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
1921 
1922 	switch (nport->disc_addr.adrfam) {
1923 	case NVMF_ADDR_FAMILY_IP4:
1924 		af = AF_INET;
1925 		break;
1926 	case NVMF_ADDR_FAMILY_IP6:
1927 		af = AF_INET6;
1928 		break;
1929 	default:
1930 		pr_err("address family %d not supported\n",
1931 			nport->disc_addr.adrfam);
1932 		ret = -EINVAL;
1933 		goto out_free_port;
1934 	}
1935 
1936 	if (nport->inline_data_size < 0) {
1937 		nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1938 	} else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1939 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1940 			nport->inline_data_size,
1941 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1942 		nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1943 	}
1944 
1945 	if (nport->max_queue_size < 0) {
1946 		nport->max_queue_size = NVME_RDMA_DEFAULT_QUEUE_SIZE;
1947 	} else if (nport->max_queue_size > NVME_RDMA_MAX_QUEUE_SIZE) {
1948 		pr_warn("max_queue_size %u is too large, reducing to %u\n",
1949 			nport->max_queue_size, NVME_RDMA_MAX_QUEUE_SIZE);
1950 		nport->max_queue_size = NVME_RDMA_MAX_QUEUE_SIZE;
1951 	}
1952 
1953 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1954 			nport->disc_addr.trsvcid, &port->addr);
1955 	if (ret) {
1956 		pr_err("malformed ip/port passed: %s:%s\n",
1957 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1958 		goto out_free_port;
1959 	}
1960 
1961 	ret = nvmet_rdma_enable_port(port);
1962 	if (ret)
1963 		goto out_free_port;
1964 
1965 	pr_info("enabling port %d (%pISpcs)\n",
1966 		le16_to_cpu(nport->disc_addr.portid),
1967 		(struct sockaddr *)&port->addr);
1968 
1969 	return 0;
1970 
1971 out_free_port:
1972 	kfree(port);
1973 	return ret;
1974 }
1975 
1976 static void nvmet_rdma_remove_port(struct nvmet_port *nport)
1977 {
1978 	struct nvmet_rdma_port *port = nport->priv;
1979 
1980 	cancel_delayed_work_sync(&port->repair_work);
1981 	nvmet_rdma_disable_port(port);
1982 	kfree(port);
1983 }
1984 
1985 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1986 		struct nvmet_port *nport, char *traddr)
1987 {
1988 	struct nvmet_rdma_port *port = nport->priv;
1989 	struct rdma_cm_id *cm_id = port->cm_id;
1990 
1991 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1992 		struct nvmet_rdma_rsp *rsp =
1993 			container_of(req, struct nvmet_rdma_rsp, req);
1994 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1995 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1996 
1997 		sprintf(traddr, "%pISc", addr);
1998 	} else {
1999 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
2000 	}
2001 }
2002 
2003 static ssize_t nvmet_rdma_host_port_addr(struct nvmet_ctrl *ctrl,
2004 		char *traddr, size_t traddr_len)
2005 {
2006 	struct nvmet_sq *nvme_sq = ctrl->sqs[0];
2007 	struct nvmet_rdma_queue *queue =
2008 		container_of(nvme_sq, struct nvmet_rdma_queue, nvme_sq);
2009 
2010 	return snprintf(traddr, traddr_len, "%pISc",
2011 			(struct sockaddr *)&queue->cm_id->route.addr.dst_addr);
2012 }
2013 
2014 static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
2015 {
2016 	if (ctrl->pi_support)
2017 		return NVMET_RDMA_MAX_METADATA_MDTS;
2018 	return NVMET_RDMA_MAX_MDTS;
2019 }
2020 
2021 static u16 nvmet_rdma_get_max_queue_size(const struct nvmet_ctrl *ctrl)
2022 {
2023 	if (ctrl->pi_support)
2024 		return NVME_RDMA_MAX_METADATA_QUEUE_SIZE;
2025 	return NVME_RDMA_MAX_QUEUE_SIZE;
2026 }
2027 
2028 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
2029 	.owner			= THIS_MODULE,
2030 	.type			= NVMF_TRTYPE_RDMA,
2031 	.msdbd			= 1,
2032 	.flags			= NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
2033 	.add_port		= nvmet_rdma_add_port,
2034 	.remove_port		= nvmet_rdma_remove_port,
2035 	.queue_response		= nvmet_rdma_queue_response,
2036 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
2037 	.disc_traddr		= nvmet_rdma_disc_port_addr,
2038 	.host_traddr		= nvmet_rdma_host_port_addr,
2039 	.get_mdts		= nvmet_rdma_get_mdts,
2040 	.get_max_queue_size	= nvmet_rdma_get_max_queue_size,
2041 };
2042 
2043 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2044 {
2045 	struct nvmet_rdma_queue *queue, *tmp;
2046 	struct nvmet_rdma_device *ndev;
2047 	bool found = false;
2048 
2049 	mutex_lock(&device_list_mutex);
2050 	list_for_each_entry(ndev, &device_list, entry) {
2051 		if (ndev->device == ib_device) {
2052 			found = true;
2053 			break;
2054 		}
2055 	}
2056 	mutex_unlock(&device_list_mutex);
2057 
2058 	if (!found)
2059 		return;
2060 
2061 	/*
2062 	 * IB Device that is used by nvmet controllers is being removed,
2063 	 * delete all queues using this device.
2064 	 */
2065 	mutex_lock(&nvmet_rdma_queue_mutex);
2066 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
2067 				 queue_list) {
2068 		if (queue->dev->device != ib_device)
2069 			continue;
2070 
2071 		pr_info("Removing queue %d\n", queue->idx);
2072 		list_del_init(&queue->queue_list);
2073 		__nvmet_rdma_queue_disconnect(queue);
2074 	}
2075 	mutex_unlock(&nvmet_rdma_queue_mutex);
2076 
2077 	flush_workqueue(nvmet_wq);
2078 }
2079 
2080 static struct ib_client nvmet_rdma_ib_client = {
2081 	.name   = "nvmet_rdma",
2082 	.remove = nvmet_rdma_remove_one
2083 };
2084 
2085 static int __init nvmet_rdma_init(void)
2086 {
2087 	int ret;
2088 
2089 	ret = ib_register_client(&nvmet_rdma_ib_client);
2090 	if (ret)
2091 		return ret;
2092 
2093 	ret = nvmet_register_transport(&nvmet_rdma_ops);
2094 	if (ret)
2095 		goto err_ib_client;
2096 
2097 	return 0;
2098 
2099 err_ib_client:
2100 	ib_unregister_client(&nvmet_rdma_ib_client);
2101 	return ret;
2102 }
2103 
2104 static void __exit nvmet_rdma_exit(void)
2105 {
2106 	nvmet_unregister_transport(&nvmet_rdma_ops);
2107 	ib_unregister_client(&nvmet_rdma_ib_client);
2108 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
2109 	ida_destroy(&nvmet_rdma_queue_ida);
2110 }
2111 
2112 module_init(nvmet_rdma_init);
2113 module_exit(nvmet_rdma_exit);
2114 
2115 MODULE_DESCRIPTION("NVMe target RDMA transport driver");
2116 MODULE_LICENSE("GPL v2");
2117 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
2118