xref: /openbmc/linux/drivers/nvme/target/rdma.c (revision 99a15348)
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 struct nvmet_rdma_srq;
41 
42 struct nvmet_rdma_cmd {
43 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
44 	struct ib_cqe		cqe;
45 	struct ib_recv_wr	wr;
46 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
47 	struct nvme_command     *nvme_cmd;
48 	struct nvmet_rdma_queue	*queue;
49 	struct nvmet_rdma_srq   *nsrq;
50 };
51 
52 enum {
53 	NVMET_RDMA_REQ_INLINE_DATA	= (1 << 0),
54 	NVMET_RDMA_REQ_INVALIDATE_RKEY	= (1 << 1),
55 };
56 
57 struct nvmet_rdma_rsp {
58 	struct ib_sge		send_sge;
59 	struct ib_cqe		send_cqe;
60 	struct ib_send_wr	send_wr;
61 
62 	struct nvmet_rdma_cmd	*cmd;
63 	struct nvmet_rdma_queue	*queue;
64 
65 	struct ib_cqe		read_cqe;
66 	struct ib_cqe		write_cqe;
67 	struct rdma_rw_ctx	rw;
68 
69 	struct nvmet_req	req;
70 
71 	bool			allocated;
72 	u8			n_rdma;
73 	u32			flags;
74 	u32			invalidate_rkey;
75 
76 	struct list_head	wait_list;
77 	struct list_head	free_list;
78 };
79 
80 enum nvmet_rdma_queue_state {
81 	NVMET_RDMA_Q_CONNECTING,
82 	NVMET_RDMA_Q_LIVE,
83 	NVMET_RDMA_Q_DISCONNECTING,
84 };
85 
86 struct nvmet_rdma_queue {
87 	struct rdma_cm_id	*cm_id;
88 	struct ib_qp		*qp;
89 	struct nvmet_port	*port;
90 	struct ib_cq		*cq;
91 	atomic_t		sq_wr_avail;
92 	struct nvmet_rdma_device *dev;
93 	struct nvmet_rdma_srq   *nsrq;
94 	spinlock_t		state_lock;
95 	enum nvmet_rdma_queue_state state;
96 	struct nvmet_cq		nvme_cq;
97 	struct nvmet_sq		nvme_sq;
98 
99 	struct nvmet_rdma_rsp	*rsps;
100 	struct list_head	free_rsps;
101 	spinlock_t		rsps_lock;
102 	struct nvmet_rdma_cmd	*cmds;
103 
104 	struct work_struct	release_work;
105 	struct list_head	rsp_wait_list;
106 	struct list_head	rsp_wr_wait_list;
107 	spinlock_t		rsp_wr_wait_lock;
108 
109 	int			idx;
110 	int			host_qid;
111 	int			comp_vector;
112 	int			recv_queue_size;
113 	int			send_queue_size;
114 
115 	struct list_head	queue_list;
116 };
117 
118 struct nvmet_rdma_port {
119 	struct nvmet_port	*nport;
120 	struct sockaddr_storage addr;
121 	struct rdma_cm_id	*cm_id;
122 	struct delayed_work	repair_work;
123 };
124 
125 struct nvmet_rdma_srq {
126 	struct ib_srq            *srq;
127 	struct nvmet_rdma_cmd    *cmds;
128 	struct nvmet_rdma_device *ndev;
129 };
130 
131 struct nvmet_rdma_device {
132 	struct ib_device	*device;
133 	struct ib_pd		*pd;
134 	struct nvmet_rdma_srq	**srqs;
135 	int			srq_count;
136 	size_t			srq_size;
137 	struct kref		ref;
138 	struct list_head	entry;
139 	int			inline_data_size;
140 	int			inline_page_count;
141 };
142 
143 static bool nvmet_rdma_use_srq;
144 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
145 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
146 
147 static int srq_size_set(const char *val, const struct kernel_param *kp);
148 static const struct kernel_param_ops srq_size_ops = {
149 	.set = srq_size_set,
150 	.get = param_get_int,
151 };
152 
153 static int nvmet_rdma_srq_size = 1024;
154 module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
155 MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");
156 
157 static DEFINE_IDA(nvmet_rdma_queue_ida);
158 static LIST_HEAD(nvmet_rdma_queue_list);
159 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
160 
161 static LIST_HEAD(device_list);
162 static DEFINE_MUTEX(device_list_mutex);
163 
164 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
165 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
166 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
167 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
168 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
169 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
170 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
171 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
172 				struct nvmet_rdma_rsp *r);
173 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
174 				struct nvmet_rdma_rsp *r);
175 
176 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
177 
178 static int srq_size_set(const char *val, const struct kernel_param *kp)
179 {
180 	int n = 0, ret;
181 
182 	ret = kstrtoint(val, 10, &n);
183 	if (ret != 0 || n < 256)
184 		return -EINVAL;
185 
186 	return param_set_int(val, kp);
187 }
188 
189 static int num_pages(int len)
190 {
191 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
192 }
193 
194 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
195 {
196 	return nvme_is_write(rsp->req.cmd) &&
197 		rsp->req.transfer_len &&
198 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
199 }
200 
201 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
202 {
203 	return !nvme_is_write(rsp->req.cmd) &&
204 		rsp->req.transfer_len &&
205 		!rsp->req.cqe->status &&
206 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
207 }
208 
209 static inline struct nvmet_rdma_rsp *
210 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
211 {
212 	struct nvmet_rdma_rsp *rsp;
213 	unsigned long flags;
214 
215 	spin_lock_irqsave(&queue->rsps_lock, flags);
216 	rsp = list_first_entry_or_null(&queue->free_rsps,
217 				struct nvmet_rdma_rsp, free_list);
218 	if (likely(rsp))
219 		list_del(&rsp->free_list);
220 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
221 
222 	if (unlikely(!rsp)) {
223 		int ret;
224 
225 		rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
226 		if (unlikely(!rsp))
227 			return NULL;
228 		ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
229 		if (unlikely(ret)) {
230 			kfree(rsp);
231 			return NULL;
232 		}
233 
234 		rsp->allocated = true;
235 	}
236 
237 	return rsp;
238 }
239 
240 static inline void
241 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
242 {
243 	unsigned long flags;
244 
245 	if (unlikely(rsp->allocated)) {
246 		nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
247 		kfree(rsp);
248 		return;
249 	}
250 
251 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
252 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
253 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
254 }
255 
256 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
257 				struct nvmet_rdma_cmd *c)
258 {
259 	struct scatterlist *sg;
260 	struct ib_sge *sge;
261 	int i;
262 
263 	if (!ndev->inline_data_size)
264 		return;
265 
266 	sg = c->inline_sg;
267 	sge = &c->sge[1];
268 
269 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
270 		if (sge->length)
271 			ib_dma_unmap_page(ndev->device, sge->addr,
272 					sge->length, DMA_FROM_DEVICE);
273 		if (sg_page(sg))
274 			__free_page(sg_page(sg));
275 	}
276 }
277 
278 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
279 				struct nvmet_rdma_cmd *c)
280 {
281 	struct scatterlist *sg;
282 	struct ib_sge *sge;
283 	struct page *pg;
284 	int len;
285 	int i;
286 
287 	if (!ndev->inline_data_size)
288 		return 0;
289 
290 	sg = c->inline_sg;
291 	sg_init_table(sg, ndev->inline_page_count);
292 	sge = &c->sge[1];
293 	len = ndev->inline_data_size;
294 
295 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
296 		pg = alloc_page(GFP_KERNEL);
297 		if (!pg)
298 			goto out_err;
299 		sg_assign_page(sg, pg);
300 		sge->addr = ib_dma_map_page(ndev->device,
301 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
302 		if (ib_dma_mapping_error(ndev->device, sge->addr))
303 			goto out_err;
304 		sge->length = min_t(int, len, PAGE_SIZE);
305 		sge->lkey = ndev->pd->local_dma_lkey;
306 		len -= sge->length;
307 	}
308 
309 	return 0;
310 out_err:
311 	for (; i >= 0; i--, sg--, sge--) {
312 		if (sge->length)
313 			ib_dma_unmap_page(ndev->device, sge->addr,
314 					sge->length, DMA_FROM_DEVICE);
315 		if (sg_page(sg))
316 			__free_page(sg_page(sg));
317 	}
318 	return -ENOMEM;
319 }
320 
321 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
322 			struct nvmet_rdma_cmd *c, bool admin)
323 {
324 	/* NVMe command / RDMA RECV */
325 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
326 	if (!c->nvme_cmd)
327 		goto out;
328 
329 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
330 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
331 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
332 		goto out_free_cmd;
333 
334 	c->sge[0].length = sizeof(*c->nvme_cmd);
335 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
336 
337 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
338 		goto out_unmap_cmd;
339 
340 	c->cqe.done = nvmet_rdma_recv_done;
341 
342 	c->wr.wr_cqe = &c->cqe;
343 	c->wr.sg_list = c->sge;
344 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
345 
346 	return 0;
347 
348 out_unmap_cmd:
349 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
350 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
351 out_free_cmd:
352 	kfree(c->nvme_cmd);
353 
354 out:
355 	return -ENOMEM;
356 }
357 
358 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
359 		struct nvmet_rdma_cmd *c, bool admin)
360 {
361 	if (!admin)
362 		nvmet_rdma_free_inline_pages(ndev, c);
363 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
364 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
365 	kfree(c->nvme_cmd);
366 }
367 
368 static struct nvmet_rdma_cmd *
369 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
370 		int nr_cmds, bool admin)
371 {
372 	struct nvmet_rdma_cmd *cmds;
373 	int ret = -EINVAL, i;
374 
375 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
376 	if (!cmds)
377 		goto out;
378 
379 	for (i = 0; i < nr_cmds; i++) {
380 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
381 		if (ret)
382 			goto out_free;
383 	}
384 
385 	return cmds;
386 
387 out_free:
388 	while (--i >= 0)
389 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
390 	kfree(cmds);
391 out:
392 	return ERR_PTR(ret);
393 }
394 
395 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
396 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
397 {
398 	int i;
399 
400 	for (i = 0; i < nr_cmds; i++)
401 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
402 	kfree(cmds);
403 }
404 
405 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
406 		struct nvmet_rdma_rsp *r)
407 {
408 	/* NVMe CQE / RDMA SEND */
409 	r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
410 	if (!r->req.cqe)
411 		goto out;
412 
413 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
414 			sizeof(*r->req.cqe), DMA_TO_DEVICE);
415 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
416 		goto out_free_rsp;
417 
418 	if (!ib_uses_virt_dma(ndev->device))
419 		r->req.p2p_client = &ndev->device->dev;
420 	r->send_sge.length = sizeof(*r->req.cqe);
421 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
422 
423 	r->send_cqe.done = nvmet_rdma_send_done;
424 
425 	r->send_wr.wr_cqe = &r->send_cqe;
426 	r->send_wr.sg_list = &r->send_sge;
427 	r->send_wr.num_sge = 1;
428 	r->send_wr.send_flags = IB_SEND_SIGNALED;
429 
430 	/* Data In / RDMA READ */
431 	r->read_cqe.done = nvmet_rdma_read_data_done;
432 	/* Data Out / RDMA WRITE */
433 	r->write_cqe.done = nvmet_rdma_write_data_done;
434 
435 	return 0;
436 
437 out_free_rsp:
438 	kfree(r->req.cqe);
439 out:
440 	return -ENOMEM;
441 }
442 
443 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
444 		struct nvmet_rdma_rsp *r)
445 {
446 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
447 				sizeof(*r->req.cqe), DMA_TO_DEVICE);
448 	kfree(r->req.cqe);
449 }
450 
451 static int
452 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
453 {
454 	struct nvmet_rdma_device *ndev = queue->dev;
455 	int nr_rsps = queue->recv_queue_size * 2;
456 	int ret = -EINVAL, i;
457 
458 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
459 			GFP_KERNEL);
460 	if (!queue->rsps)
461 		goto out;
462 
463 	for (i = 0; i < nr_rsps; i++) {
464 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
465 
466 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
467 		if (ret)
468 			goto out_free;
469 
470 		list_add_tail(&rsp->free_list, &queue->free_rsps);
471 	}
472 
473 	return 0;
474 
475 out_free:
476 	while (--i >= 0) {
477 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
478 
479 		list_del(&rsp->free_list);
480 		nvmet_rdma_free_rsp(ndev, rsp);
481 	}
482 	kfree(queue->rsps);
483 out:
484 	return ret;
485 }
486 
487 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
488 {
489 	struct nvmet_rdma_device *ndev = queue->dev;
490 	int i, nr_rsps = queue->recv_queue_size * 2;
491 
492 	for (i = 0; i < nr_rsps; i++) {
493 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
494 
495 		list_del(&rsp->free_list);
496 		nvmet_rdma_free_rsp(ndev, rsp);
497 	}
498 	kfree(queue->rsps);
499 }
500 
501 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
502 		struct nvmet_rdma_cmd *cmd)
503 {
504 	int ret;
505 
506 	ib_dma_sync_single_for_device(ndev->device,
507 		cmd->sge[0].addr, cmd->sge[0].length,
508 		DMA_FROM_DEVICE);
509 
510 	if (cmd->nsrq)
511 		ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
512 	else
513 		ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);
514 
515 	if (unlikely(ret))
516 		pr_err("post_recv cmd failed\n");
517 
518 	return ret;
519 }
520 
521 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
522 {
523 	spin_lock(&queue->rsp_wr_wait_lock);
524 	while (!list_empty(&queue->rsp_wr_wait_list)) {
525 		struct nvmet_rdma_rsp *rsp;
526 		bool ret;
527 
528 		rsp = list_entry(queue->rsp_wr_wait_list.next,
529 				struct nvmet_rdma_rsp, wait_list);
530 		list_del(&rsp->wait_list);
531 
532 		spin_unlock(&queue->rsp_wr_wait_lock);
533 		ret = nvmet_rdma_execute_command(rsp);
534 		spin_lock(&queue->rsp_wr_wait_lock);
535 
536 		if (!ret) {
537 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
538 			break;
539 		}
540 	}
541 	spin_unlock(&queue->rsp_wr_wait_lock);
542 }
543 
544 static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
545 {
546 	struct ib_mr_status mr_status;
547 	int ret;
548 	u16 status = 0;
549 
550 	ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
551 	if (ret) {
552 		pr_err("ib_check_mr_status failed, ret %d\n", ret);
553 		return NVME_SC_INVALID_PI;
554 	}
555 
556 	if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
557 		switch (mr_status.sig_err.err_type) {
558 		case IB_SIG_BAD_GUARD:
559 			status = NVME_SC_GUARD_CHECK;
560 			break;
561 		case IB_SIG_BAD_REFTAG:
562 			status = NVME_SC_REFTAG_CHECK;
563 			break;
564 		case IB_SIG_BAD_APPTAG:
565 			status = NVME_SC_APPTAG_CHECK;
566 			break;
567 		}
568 		pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
569 		       mr_status.sig_err.err_type,
570 		       mr_status.sig_err.expected,
571 		       mr_status.sig_err.actual);
572 	}
573 
574 	return status;
575 }
576 
577 static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
578 		struct nvme_command *cmd, struct ib_sig_domain *domain,
579 		u16 control, u8 pi_type)
580 {
581 	domain->sig_type = IB_SIG_TYPE_T10_DIF;
582 	domain->sig.dif.bg_type = IB_T10DIF_CRC;
583 	domain->sig.dif.pi_interval = 1 << bi->interval_exp;
584 	domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
585 	if (control & NVME_RW_PRINFO_PRCHK_REF)
586 		domain->sig.dif.ref_remap = true;
587 
588 	domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
589 	domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
590 	domain->sig.dif.app_escape = true;
591 	if (pi_type == NVME_NS_DPS_PI_TYPE3)
592 		domain->sig.dif.ref_escape = true;
593 }
594 
595 static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
596 				     struct ib_sig_attrs *sig_attrs)
597 {
598 	struct nvme_command *cmd = req->cmd;
599 	u16 control = le16_to_cpu(cmd->rw.control);
600 	u8 pi_type = req->ns->pi_type;
601 	struct blk_integrity *bi;
602 
603 	bi = bdev_get_integrity(req->ns->bdev);
604 
605 	memset(sig_attrs, 0, sizeof(*sig_attrs));
606 
607 	if (control & NVME_RW_PRINFO_PRACT) {
608 		/* for WRITE_INSERT/READ_STRIP no wire domain */
609 		sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
610 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
611 					  pi_type);
612 		/* Clear the PRACT bit since HCA will generate/verify the PI */
613 		control &= ~NVME_RW_PRINFO_PRACT;
614 		cmd->rw.control = cpu_to_le16(control);
615 		/* PI is added by the HW */
616 		req->transfer_len += req->metadata_len;
617 	} else {
618 		/* for WRITE_PASS/READ_PASS both wire/memory domains exist */
619 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
620 					  pi_type);
621 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
622 					  pi_type);
623 	}
624 
625 	if (control & NVME_RW_PRINFO_PRCHK_REF)
626 		sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
627 	if (control & NVME_RW_PRINFO_PRCHK_GUARD)
628 		sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
629 	if (control & NVME_RW_PRINFO_PRCHK_APP)
630 		sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
631 }
632 
633 static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
634 				  struct ib_sig_attrs *sig_attrs)
635 {
636 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
637 	struct nvmet_req *req = &rsp->req;
638 	int ret;
639 
640 	if (req->metadata_len)
641 		ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
642 			cm_id->port_num, req->sg, req->sg_cnt,
643 			req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
644 			addr, key, nvmet_data_dir(req));
645 	else
646 		ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
647 				       req->sg, req->sg_cnt, 0, addr, key,
648 				       nvmet_data_dir(req));
649 
650 	return ret;
651 }
652 
653 static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
654 {
655 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
656 	struct nvmet_req *req = &rsp->req;
657 
658 	if (req->metadata_len)
659 		rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
660 			cm_id->port_num, req->sg, req->sg_cnt,
661 			req->metadata_sg, req->metadata_sg_cnt,
662 			nvmet_data_dir(req));
663 	else
664 		rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
665 				    req->sg, req->sg_cnt, nvmet_data_dir(req));
666 }
667 
668 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
669 {
670 	struct nvmet_rdma_queue *queue = rsp->queue;
671 
672 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
673 
674 	if (rsp->n_rdma)
675 		nvmet_rdma_rw_ctx_destroy(rsp);
676 
677 	if (rsp->req.sg != rsp->cmd->inline_sg)
678 		nvmet_req_free_sgls(&rsp->req);
679 
680 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
681 		nvmet_rdma_process_wr_wait_list(queue);
682 
683 	nvmet_rdma_put_rsp(rsp);
684 }
685 
686 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
687 {
688 	if (queue->nvme_sq.ctrl) {
689 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
690 	} else {
691 		/*
692 		 * we didn't setup the controller yet in case
693 		 * of admin connect error, just disconnect and
694 		 * cleanup the queue
695 		 */
696 		nvmet_rdma_queue_disconnect(queue);
697 	}
698 }
699 
700 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
701 {
702 	struct nvmet_rdma_rsp *rsp =
703 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
704 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
705 
706 	nvmet_rdma_release_rsp(rsp);
707 
708 	if (unlikely(wc->status != IB_WC_SUCCESS &&
709 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
710 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
711 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
712 		nvmet_rdma_error_comp(queue);
713 	}
714 }
715 
716 static void nvmet_rdma_queue_response(struct nvmet_req *req)
717 {
718 	struct nvmet_rdma_rsp *rsp =
719 		container_of(req, struct nvmet_rdma_rsp, req);
720 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
721 	struct ib_send_wr *first_wr;
722 
723 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
724 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
725 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
726 	} else {
727 		rsp->send_wr.opcode = IB_WR_SEND;
728 	}
729 
730 	if (nvmet_rdma_need_data_out(rsp)) {
731 		if (rsp->req.metadata_len)
732 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
733 					cm_id->port_num, &rsp->write_cqe, NULL);
734 		else
735 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
736 					cm_id->port_num, NULL, &rsp->send_wr);
737 	} else {
738 		first_wr = &rsp->send_wr;
739 	}
740 
741 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
742 
743 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
744 		rsp->send_sge.addr, rsp->send_sge.length,
745 		DMA_TO_DEVICE);
746 
747 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
748 		pr_err("sending cmd response failed\n");
749 		nvmet_rdma_release_rsp(rsp);
750 	}
751 }
752 
753 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
754 {
755 	struct nvmet_rdma_rsp *rsp =
756 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
757 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
758 	u16 status = 0;
759 
760 	WARN_ON(rsp->n_rdma <= 0);
761 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
762 	rsp->n_rdma = 0;
763 
764 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
765 		nvmet_rdma_rw_ctx_destroy(rsp);
766 		nvmet_req_uninit(&rsp->req);
767 		nvmet_rdma_release_rsp(rsp);
768 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
769 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
770 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
771 			nvmet_rdma_error_comp(queue);
772 		}
773 		return;
774 	}
775 
776 	if (rsp->req.metadata_len)
777 		status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
778 	nvmet_rdma_rw_ctx_destroy(rsp);
779 
780 	if (unlikely(status))
781 		nvmet_req_complete(&rsp->req, status);
782 	else
783 		rsp->req.execute(&rsp->req);
784 }
785 
786 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
787 {
788 	struct nvmet_rdma_rsp *rsp =
789 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
790 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
791 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
792 	u16 status;
793 
794 	if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
795 		return;
796 
797 	WARN_ON(rsp->n_rdma <= 0);
798 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
799 	rsp->n_rdma = 0;
800 
801 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
802 		nvmet_rdma_rw_ctx_destroy(rsp);
803 		nvmet_req_uninit(&rsp->req);
804 		nvmet_rdma_release_rsp(rsp);
805 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
806 			pr_info("RDMA WRITE for CQE failed with status %s (%d).\n",
807 				ib_wc_status_msg(wc->status), wc->status);
808 			nvmet_rdma_error_comp(queue);
809 		}
810 		return;
811 	}
812 
813 	/*
814 	 * Upon RDMA completion check the signature status
815 	 * - if succeeded send good NVMe response
816 	 * - if failed send bad NVMe response with appropriate error
817 	 */
818 	status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
819 	if (unlikely(status))
820 		rsp->req.cqe->status = cpu_to_le16(status << 1);
821 	nvmet_rdma_rw_ctx_destroy(rsp);
822 
823 	if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
824 		pr_err("sending cmd response failed\n");
825 		nvmet_rdma_release_rsp(rsp);
826 	}
827 }
828 
829 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
830 		u64 off)
831 {
832 	int sg_count = num_pages(len);
833 	struct scatterlist *sg;
834 	int i;
835 
836 	sg = rsp->cmd->inline_sg;
837 	for (i = 0; i < sg_count; i++, sg++) {
838 		if (i < sg_count - 1)
839 			sg_unmark_end(sg);
840 		else
841 			sg_mark_end(sg);
842 		sg->offset = off;
843 		sg->length = min_t(int, len, PAGE_SIZE - off);
844 		len -= sg->length;
845 		if (!i)
846 			off = 0;
847 	}
848 
849 	rsp->req.sg = rsp->cmd->inline_sg;
850 	rsp->req.sg_cnt = sg_count;
851 }
852 
853 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
854 {
855 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
856 	u64 off = le64_to_cpu(sgl->addr);
857 	u32 len = le32_to_cpu(sgl->length);
858 
859 	if (!nvme_is_write(rsp->req.cmd)) {
860 		rsp->req.error_loc =
861 			offsetof(struct nvme_common_command, opcode);
862 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
863 	}
864 
865 	if (off + len > rsp->queue->dev->inline_data_size) {
866 		pr_err("invalid inline data offset!\n");
867 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
868 	}
869 
870 	/* no data command? */
871 	if (!len)
872 		return 0;
873 
874 	nvmet_rdma_use_inline_sg(rsp, len, off);
875 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
876 	rsp->req.transfer_len += len;
877 	return 0;
878 }
879 
880 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
881 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
882 {
883 	u64 addr = le64_to_cpu(sgl->addr);
884 	u32 key = get_unaligned_le32(sgl->key);
885 	struct ib_sig_attrs sig_attrs;
886 	int ret;
887 
888 	rsp->req.transfer_len = get_unaligned_le24(sgl->length);
889 
890 	/* no data command? */
891 	if (!rsp->req.transfer_len)
892 		return 0;
893 
894 	if (rsp->req.metadata_len)
895 		nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);
896 
897 	ret = nvmet_req_alloc_sgls(&rsp->req);
898 	if (unlikely(ret < 0))
899 		goto error_out;
900 
901 	ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
902 	if (unlikely(ret < 0))
903 		goto error_out;
904 	rsp->n_rdma += ret;
905 
906 	if (invalidate) {
907 		rsp->invalidate_rkey = key;
908 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
909 	}
910 
911 	return 0;
912 
913 error_out:
914 	rsp->req.transfer_len = 0;
915 	return NVME_SC_INTERNAL;
916 }
917 
918 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
919 {
920 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
921 
922 	switch (sgl->type >> 4) {
923 	case NVME_SGL_FMT_DATA_DESC:
924 		switch (sgl->type & 0xf) {
925 		case NVME_SGL_FMT_OFFSET:
926 			return nvmet_rdma_map_sgl_inline(rsp);
927 		default:
928 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
929 			rsp->req.error_loc =
930 				offsetof(struct nvme_common_command, dptr);
931 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
932 		}
933 	case NVME_KEY_SGL_FMT_DATA_DESC:
934 		switch (sgl->type & 0xf) {
935 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
936 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
937 		case NVME_SGL_FMT_ADDRESS:
938 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
939 		default:
940 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
941 			rsp->req.error_loc =
942 				offsetof(struct nvme_common_command, dptr);
943 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
944 		}
945 	default:
946 		pr_err("invalid SGL type: %#x\n", sgl->type);
947 		rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
948 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
949 	}
950 }
951 
952 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
953 {
954 	struct nvmet_rdma_queue *queue = rsp->queue;
955 
956 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
957 			&queue->sq_wr_avail) < 0)) {
958 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
959 				1 + rsp->n_rdma, queue->idx,
960 				queue->nvme_sq.ctrl->cntlid);
961 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
962 		return false;
963 	}
964 
965 	if (nvmet_rdma_need_data_in(rsp)) {
966 		if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
967 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
968 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
969 	} else {
970 		rsp->req.execute(&rsp->req);
971 	}
972 
973 	return true;
974 }
975 
976 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
977 		struct nvmet_rdma_rsp *cmd)
978 {
979 	u16 status;
980 
981 	ib_dma_sync_single_for_cpu(queue->dev->device,
982 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
983 		DMA_FROM_DEVICE);
984 	ib_dma_sync_single_for_cpu(queue->dev->device,
985 		cmd->send_sge.addr, cmd->send_sge.length,
986 		DMA_TO_DEVICE);
987 
988 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
989 			&queue->nvme_sq, &nvmet_rdma_ops))
990 		return;
991 
992 	status = nvmet_rdma_map_sgl(cmd);
993 	if (status)
994 		goto out_err;
995 
996 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
997 		spin_lock(&queue->rsp_wr_wait_lock);
998 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
999 		spin_unlock(&queue->rsp_wr_wait_lock);
1000 	}
1001 
1002 	return;
1003 
1004 out_err:
1005 	nvmet_req_complete(&cmd->req, status);
1006 }
1007 
1008 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1009 {
1010 	struct nvmet_rdma_cmd *cmd =
1011 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
1012 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
1013 	struct nvmet_rdma_rsp *rsp;
1014 
1015 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1016 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
1017 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
1018 				wc->wr_cqe, ib_wc_status_msg(wc->status),
1019 				wc->status);
1020 			nvmet_rdma_error_comp(queue);
1021 		}
1022 		return;
1023 	}
1024 
1025 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
1026 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
1027 		nvmet_rdma_error_comp(queue);
1028 		return;
1029 	}
1030 
1031 	cmd->queue = queue;
1032 	rsp = nvmet_rdma_get_rsp(queue);
1033 	if (unlikely(!rsp)) {
1034 		/*
1035 		 * we get here only under memory pressure,
1036 		 * silently drop and have the host retry
1037 		 * as we can't even fail it.
1038 		 */
1039 		nvmet_rdma_post_recv(queue->dev, cmd);
1040 		return;
1041 	}
1042 	rsp->queue = queue;
1043 	rsp->cmd = cmd;
1044 	rsp->flags = 0;
1045 	rsp->req.cmd = cmd->nvme_cmd;
1046 	rsp->req.port = queue->port;
1047 	rsp->n_rdma = 0;
1048 
1049 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
1050 		unsigned long flags;
1051 
1052 		spin_lock_irqsave(&queue->state_lock, flags);
1053 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
1054 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
1055 		else
1056 			nvmet_rdma_put_rsp(rsp);
1057 		spin_unlock_irqrestore(&queue->state_lock, flags);
1058 		return;
1059 	}
1060 
1061 	nvmet_rdma_handle_command(queue, rsp);
1062 }
1063 
1064 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
1065 {
1066 	nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
1067 			     false);
1068 	ib_destroy_srq(nsrq->srq);
1069 
1070 	kfree(nsrq);
1071 }
1072 
1073 static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
1074 {
1075 	int i;
1076 
1077 	if (!ndev->srqs)
1078 		return;
1079 
1080 	for (i = 0; i < ndev->srq_count; i++)
1081 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1082 
1083 	kfree(ndev->srqs);
1084 }
1085 
1086 static struct nvmet_rdma_srq *
1087 nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
1088 {
1089 	struct ib_srq_init_attr srq_attr = { NULL, };
1090 	size_t srq_size = ndev->srq_size;
1091 	struct nvmet_rdma_srq *nsrq;
1092 	struct ib_srq *srq;
1093 	int ret, i;
1094 
1095 	nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
1096 	if (!nsrq)
1097 		return ERR_PTR(-ENOMEM);
1098 
1099 	srq_attr.attr.max_wr = srq_size;
1100 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
1101 	srq_attr.attr.srq_limit = 0;
1102 	srq_attr.srq_type = IB_SRQT_BASIC;
1103 	srq = ib_create_srq(ndev->pd, &srq_attr);
1104 	if (IS_ERR(srq)) {
1105 		ret = PTR_ERR(srq);
1106 		goto out_free;
1107 	}
1108 
1109 	nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
1110 	if (IS_ERR(nsrq->cmds)) {
1111 		ret = PTR_ERR(nsrq->cmds);
1112 		goto out_destroy_srq;
1113 	}
1114 
1115 	nsrq->srq = srq;
1116 	nsrq->ndev = ndev;
1117 
1118 	for (i = 0; i < srq_size; i++) {
1119 		nsrq->cmds[i].nsrq = nsrq;
1120 		ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
1121 		if (ret)
1122 			goto out_free_cmds;
1123 	}
1124 
1125 	return nsrq;
1126 
1127 out_free_cmds:
1128 	nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
1129 out_destroy_srq:
1130 	ib_destroy_srq(srq);
1131 out_free:
1132 	kfree(nsrq);
1133 	return ERR_PTR(ret);
1134 }
1135 
1136 static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
1137 {
1138 	int i, ret;
1139 
1140 	if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
1141 		/*
1142 		 * If SRQs aren't supported we just go ahead and use normal
1143 		 * non-shared receive queues.
1144 		 */
1145 		pr_info("SRQ requested but not supported.\n");
1146 		return 0;
1147 	}
1148 
1149 	ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
1150 			     nvmet_rdma_srq_size);
1151 	ndev->srq_count = min(ndev->device->num_comp_vectors,
1152 			      ndev->device->attrs.max_srq);
1153 
1154 	ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
1155 	if (!ndev->srqs)
1156 		return -ENOMEM;
1157 
1158 	for (i = 0; i < ndev->srq_count; i++) {
1159 		ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
1160 		if (IS_ERR(ndev->srqs[i])) {
1161 			ret = PTR_ERR(ndev->srqs[i]);
1162 			goto err_srq;
1163 		}
1164 	}
1165 
1166 	return 0;
1167 
1168 err_srq:
1169 	while (--i >= 0)
1170 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1171 	kfree(ndev->srqs);
1172 	return ret;
1173 }
1174 
1175 static void nvmet_rdma_free_dev(struct kref *ref)
1176 {
1177 	struct nvmet_rdma_device *ndev =
1178 		container_of(ref, struct nvmet_rdma_device, ref);
1179 
1180 	mutex_lock(&device_list_mutex);
1181 	list_del(&ndev->entry);
1182 	mutex_unlock(&device_list_mutex);
1183 
1184 	nvmet_rdma_destroy_srqs(ndev);
1185 	ib_dealloc_pd(ndev->pd);
1186 
1187 	kfree(ndev);
1188 }
1189 
1190 static struct nvmet_rdma_device *
1191 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
1192 {
1193 	struct nvmet_rdma_port *port = cm_id->context;
1194 	struct nvmet_port *nport = port->nport;
1195 	struct nvmet_rdma_device *ndev;
1196 	int inline_page_count;
1197 	int inline_sge_count;
1198 	int ret;
1199 
1200 	mutex_lock(&device_list_mutex);
1201 	list_for_each_entry(ndev, &device_list, entry) {
1202 		if (ndev->device->node_guid == cm_id->device->node_guid &&
1203 		    kref_get_unless_zero(&ndev->ref))
1204 			goto out_unlock;
1205 	}
1206 
1207 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
1208 	if (!ndev)
1209 		goto out_err;
1210 
1211 	inline_page_count = num_pages(nport->inline_data_size);
1212 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
1213 				cm_id->device->attrs.max_recv_sge) - 1;
1214 	if (inline_page_count > inline_sge_count) {
1215 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
1216 			nport->inline_data_size, cm_id->device->name,
1217 			inline_sge_count * PAGE_SIZE);
1218 		nport->inline_data_size = inline_sge_count * PAGE_SIZE;
1219 		inline_page_count = inline_sge_count;
1220 	}
1221 	ndev->inline_data_size = nport->inline_data_size;
1222 	ndev->inline_page_count = inline_page_count;
1223 
1224 	if (nport->pi_enable && !(cm_id->device->attrs.device_cap_flags &
1225 				  IB_DEVICE_INTEGRITY_HANDOVER)) {
1226 		pr_warn("T10-PI is not supported by device %s. Disabling it\n",
1227 			cm_id->device->name);
1228 		nport->pi_enable = false;
1229 	}
1230 
1231 	ndev->device = cm_id->device;
1232 	kref_init(&ndev->ref);
1233 
1234 	ndev->pd = ib_alloc_pd(ndev->device, 0);
1235 	if (IS_ERR(ndev->pd))
1236 		goto out_free_dev;
1237 
1238 	if (nvmet_rdma_use_srq) {
1239 		ret = nvmet_rdma_init_srqs(ndev);
1240 		if (ret)
1241 			goto out_free_pd;
1242 	}
1243 
1244 	list_add(&ndev->entry, &device_list);
1245 out_unlock:
1246 	mutex_unlock(&device_list_mutex);
1247 	pr_debug("added %s.\n", ndev->device->name);
1248 	return ndev;
1249 
1250 out_free_pd:
1251 	ib_dealloc_pd(ndev->pd);
1252 out_free_dev:
1253 	kfree(ndev);
1254 out_err:
1255 	mutex_unlock(&device_list_mutex);
1256 	return NULL;
1257 }
1258 
1259 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
1260 {
1261 	struct ib_qp_init_attr qp_attr = { };
1262 	struct nvmet_rdma_device *ndev = queue->dev;
1263 	int nr_cqe, ret, i, factor;
1264 
1265 	/*
1266 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
1267 	 */
1268 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1269 
1270 	queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
1271 				   queue->comp_vector, IB_POLL_WORKQUEUE);
1272 	if (IS_ERR(queue->cq)) {
1273 		ret = PTR_ERR(queue->cq);
1274 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1275 		       nr_cqe + 1, ret);
1276 		goto out;
1277 	}
1278 
1279 	qp_attr.qp_context = queue;
1280 	qp_attr.event_handler = nvmet_rdma_qp_event;
1281 	qp_attr.send_cq = queue->cq;
1282 	qp_attr.recv_cq = queue->cq;
1283 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1284 	qp_attr.qp_type = IB_QPT_RC;
1285 	/* +1 for drain */
1286 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1287 	factor = rdma_rw_mr_factor(ndev->device, queue->cm_id->port_num,
1288 				   1 << NVMET_RDMA_MAX_MDTS);
1289 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * factor;
1290 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1291 					ndev->device->attrs.max_send_sge);
1292 
1293 	if (queue->nsrq) {
1294 		qp_attr.srq = queue->nsrq->srq;
1295 	} else {
1296 		/* +1 for drain */
1297 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1298 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1299 	}
1300 
1301 	if (queue->port->pi_enable && queue->host_qid)
1302 		qp_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
1303 
1304 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1305 	if (ret) {
1306 		pr_err("failed to create_qp ret= %d\n", ret);
1307 		goto err_destroy_cq;
1308 	}
1309 	queue->qp = queue->cm_id->qp;
1310 
1311 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1312 
1313 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1314 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1315 		 qp_attr.cap.max_send_wr, queue->cm_id);
1316 
1317 	if (!queue->nsrq) {
1318 		for (i = 0; i < queue->recv_queue_size; i++) {
1319 			queue->cmds[i].queue = queue;
1320 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1321 			if (ret)
1322 				goto err_destroy_qp;
1323 		}
1324 	}
1325 
1326 out:
1327 	return ret;
1328 
1329 err_destroy_qp:
1330 	rdma_destroy_qp(queue->cm_id);
1331 err_destroy_cq:
1332 	ib_cq_pool_put(queue->cq, nr_cqe + 1);
1333 	goto out;
1334 }
1335 
1336 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1337 {
1338 	ib_drain_qp(queue->qp);
1339 	if (queue->cm_id)
1340 		rdma_destroy_id(queue->cm_id);
1341 	ib_destroy_qp(queue->qp);
1342 	ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
1343 		       queue->send_queue_size + 1);
1344 }
1345 
1346 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1347 {
1348 	pr_debug("freeing queue %d\n", queue->idx);
1349 
1350 	nvmet_sq_destroy(&queue->nvme_sq);
1351 
1352 	nvmet_rdma_destroy_queue_ib(queue);
1353 	if (!queue->nsrq) {
1354 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1355 				queue->recv_queue_size,
1356 				!queue->host_qid);
1357 	}
1358 	nvmet_rdma_free_rsps(queue);
1359 	ida_free(&nvmet_rdma_queue_ida, queue->idx);
1360 	kfree(queue);
1361 }
1362 
1363 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1364 {
1365 	struct nvmet_rdma_queue *queue =
1366 		container_of(w, struct nvmet_rdma_queue, release_work);
1367 	struct nvmet_rdma_device *dev = queue->dev;
1368 
1369 	nvmet_rdma_free_queue(queue);
1370 
1371 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1372 }
1373 
1374 static int
1375 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1376 				struct nvmet_rdma_queue *queue)
1377 {
1378 	struct nvme_rdma_cm_req *req;
1379 
1380 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1381 	if (!req || conn->private_data_len == 0)
1382 		return NVME_RDMA_CM_INVALID_LEN;
1383 
1384 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1385 		return NVME_RDMA_CM_INVALID_RECFMT;
1386 
1387 	queue->host_qid = le16_to_cpu(req->qid);
1388 
1389 	/*
1390 	 * req->hsqsize corresponds to our recv queue size plus 1
1391 	 * req->hrqsize corresponds to our send queue size
1392 	 */
1393 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1394 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1395 
1396 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1397 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1398 
1399 	/* XXX: Should we enforce some kind of max for IO queues? */
1400 
1401 	return 0;
1402 }
1403 
1404 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1405 				enum nvme_rdma_cm_status status)
1406 {
1407 	struct nvme_rdma_cm_rej rej;
1408 
1409 	pr_debug("rejecting connect request: status %d (%s)\n",
1410 		 status, nvme_rdma_cm_msg(status));
1411 
1412 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1413 	rej.sts = cpu_to_le16(status);
1414 
1415 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej),
1416 			   IB_CM_REJ_CONSUMER_DEFINED);
1417 }
1418 
1419 static struct nvmet_rdma_queue *
1420 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1421 		struct rdma_cm_id *cm_id,
1422 		struct rdma_cm_event *event)
1423 {
1424 	struct nvmet_rdma_port *port = cm_id->context;
1425 	struct nvmet_rdma_queue *queue;
1426 	int ret;
1427 
1428 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1429 	if (!queue) {
1430 		ret = NVME_RDMA_CM_NO_RSC;
1431 		goto out_reject;
1432 	}
1433 
1434 	ret = nvmet_sq_init(&queue->nvme_sq);
1435 	if (ret) {
1436 		ret = NVME_RDMA_CM_NO_RSC;
1437 		goto out_free_queue;
1438 	}
1439 
1440 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1441 	if (ret)
1442 		goto out_destroy_sq;
1443 
1444 	/*
1445 	 * Schedules the actual release because calling rdma_destroy_id from
1446 	 * inside a CM callback would trigger a deadlock. (great API design..)
1447 	 */
1448 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1449 	queue->dev = ndev;
1450 	queue->cm_id = cm_id;
1451 	queue->port = port->nport;
1452 
1453 	spin_lock_init(&queue->state_lock);
1454 	queue->state = NVMET_RDMA_Q_CONNECTING;
1455 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1456 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1457 	spin_lock_init(&queue->rsp_wr_wait_lock);
1458 	INIT_LIST_HEAD(&queue->free_rsps);
1459 	spin_lock_init(&queue->rsps_lock);
1460 	INIT_LIST_HEAD(&queue->queue_list);
1461 
1462 	queue->idx = ida_alloc(&nvmet_rdma_queue_ida, GFP_KERNEL);
1463 	if (queue->idx < 0) {
1464 		ret = NVME_RDMA_CM_NO_RSC;
1465 		goto out_destroy_sq;
1466 	}
1467 
1468 	/*
1469 	 * Spread the io queues across completion vectors,
1470 	 * but still keep all admin queues on vector 0.
1471 	 */
1472 	queue->comp_vector = !queue->host_qid ? 0 :
1473 		queue->idx % ndev->device->num_comp_vectors;
1474 
1475 
1476 	ret = nvmet_rdma_alloc_rsps(queue);
1477 	if (ret) {
1478 		ret = NVME_RDMA_CM_NO_RSC;
1479 		goto out_ida_remove;
1480 	}
1481 
1482 	if (ndev->srqs) {
1483 		queue->nsrq = ndev->srqs[queue->comp_vector % ndev->srq_count];
1484 	} else {
1485 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1486 				queue->recv_queue_size,
1487 				!queue->host_qid);
1488 		if (IS_ERR(queue->cmds)) {
1489 			ret = NVME_RDMA_CM_NO_RSC;
1490 			goto out_free_responses;
1491 		}
1492 	}
1493 
1494 	ret = nvmet_rdma_create_queue_ib(queue);
1495 	if (ret) {
1496 		pr_err("%s: creating RDMA queue failed (%d).\n",
1497 			__func__, ret);
1498 		ret = NVME_RDMA_CM_NO_RSC;
1499 		goto out_free_cmds;
1500 	}
1501 
1502 	return queue;
1503 
1504 out_free_cmds:
1505 	if (!queue->nsrq) {
1506 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1507 				queue->recv_queue_size,
1508 				!queue->host_qid);
1509 	}
1510 out_free_responses:
1511 	nvmet_rdma_free_rsps(queue);
1512 out_ida_remove:
1513 	ida_free(&nvmet_rdma_queue_ida, queue->idx);
1514 out_destroy_sq:
1515 	nvmet_sq_destroy(&queue->nvme_sq);
1516 out_free_queue:
1517 	kfree(queue);
1518 out_reject:
1519 	nvmet_rdma_cm_reject(cm_id, ret);
1520 	return NULL;
1521 }
1522 
1523 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1524 {
1525 	struct nvmet_rdma_queue *queue = priv;
1526 
1527 	switch (event->event) {
1528 	case IB_EVENT_COMM_EST:
1529 		rdma_notify(queue->cm_id, event->event);
1530 		break;
1531 	case IB_EVENT_QP_LAST_WQE_REACHED:
1532 		pr_debug("received last WQE reached event for queue=0x%p\n",
1533 			 queue);
1534 		break;
1535 	default:
1536 		pr_err("received IB QP event: %s (%d)\n",
1537 		       ib_event_msg(event->event), event->event);
1538 		break;
1539 	}
1540 }
1541 
1542 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1543 		struct nvmet_rdma_queue *queue,
1544 		struct rdma_conn_param *p)
1545 {
1546 	struct rdma_conn_param  param = { };
1547 	struct nvme_rdma_cm_rep priv = { };
1548 	int ret = -ENOMEM;
1549 
1550 	param.rnr_retry_count = 7;
1551 	param.flow_control = 1;
1552 	param.initiator_depth = min_t(u8, p->initiator_depth,
1553 		queue->dev->device->attrs.max_qp_init_rd_atom);
1554 	param.private_data = &priv;
1555 	param.private_data_len = sizeof(priv);
1556 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1557 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1558 
1559 	ret = rdma_accept(cm_id, &param);
1560 	if (ret)
1561 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1562 
1563 	return ret;
1564 }
1565 
1566 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1567 		struct rdma_cm_event *event)
1568 {
1569 	struct nvmet_rdma_device *ndev;
1570 	struct nvmet_rdma_queue *queue;
1571 	int ret = -EINVAL;
1572 
1573 	ndev = nvmet_rdma_find_get_device(cm_id);
1574 	if (!ndev) {
1575 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1576 		return -ECONNREFUSED;
1577 	}
1578 
1579 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1580 	if (!queue) {
1581 		ret = -ENOMEM;
1582 		goto put_device;
1583 	}
1584 
1585 	if (queue->host_qid == 0) {
1586 		/* Let inflight controller teardown complete */
1587 		flush_workqueue(nvmet_wq);
1588 	}
1589 
1590 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1591 	if (ret) {
1592 		/*
1593 		 * Don't destroy the cm_id in free path, as we implicitly
1594 		 * destroy the cm_id here with non-zero ret code.
1595 		 */
1596 		queue->cm_id = NULL;
1597 		goto free_queue;
1598 	}
1599 
1600 	mutex_lock(&nvmet_rdma_queue_mutex);
1601 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1602 	mutex_unlock(&nvmet_rdma_queue_mutex);
1603 
1604 	return 0;
1605 
1606 free_queue:
1607 	nvmet_rdma_free_queue(queue);
1608 put_device:
1609 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1610 
1611 	return ret;
1612 }
1613 
1614 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1615 {
1616 	unsigned long flags;
1617 
1618 	spin_lock_irqsave(&queue->state_lock, flags);
1619 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1620 		pr_warn("trying to establish a connected queue\n");
1621 		goto out_unlock;
1622 	}
1623 	queue->state = NVMET_RDMA_Q_LIVE;
1624 
1625 	while (!list_empty(&queue->rsp_wait_list)) {
1626 		struct nvmet_rdma_rsp *cmd;
1627 
1628 		cmd = list_first_entry(&queue->rsp_wait_list,
1629 					struct nvmet_rdma_rsp, wait_list);
1630 		list_del(&cmd->wait_list);
1631 
1632 		spin_unlock_irqrestore(&queue->state_lock, flags);
1633 		nvmet_rdma_handle_command(queue, cmd);
1634 		spin_lock_irqsave(&queue->state_lock, flags);
1635 	}
1636 
1637 out_unlock:
1638 	spin_unlock_irqrestore(&queue->state_lock, flags);
1639 }
1640 
1641 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1642 {
1643 	bool disconnect = false;
1644 	unsigned long flags;
1645 
1646 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1647 
1648 	spin_lock_irqsave(&queue->state_lock, flags);
1649 	switch (queue->state) {
1650 	case NVMET_RDMA_Q_CONNECTING:
1651 		while (!list_empty(&queue->rsp_wait_list)) {
1652 			struct nvmet_rdma_rsp *rsp;
1653 
1654 			rsp = list_first_entry(&queue->rsp_wait_list,
1655 					       struct nvmet_rdma_rsp,
1656 					       wait_list);
1657 			list_del(&rsp->wait_list);
1658 			nvmet_rdma_put_rsp(rsp);
1659 		}
1660 		fallthrough;
1661 	case NVMET_RDMA_Q_LIVE:
1662 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1663 		disconnect = true;
1664 		break;
1665 	case NVMET_RDMA_Q_DISCONNECTING:
1666 		break;
1667 	}
1668 	spin_unlock_irqrestore(&queue->state_lock, flags);
1669 
1670 	if (disconnect) {
1671 		rdma_disconnect(queue->cm_id);
1672 		queue_work(nvmet_wq, &queue->release_work);
1673 	}
1674 }
1675 
1676 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1677 {
1678 	bool disconnect = false;
1679 
1680 	mutex_lock(&nvmet_rdma_queue_mutex);
1681 	if (!list_empty(&queue->queue_list)) {
1682 		list_del_init(&queue->queue_list);
1683 		disconnect = true;
1684 	}
1685 	mutex_unlock(&nvmet_rdma_queue_mutex);
1686 
1687 	if (disconnect)
1688 		__nvmet_rdma_queue_disconnect(queue);
1689 }
1690 
1691 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1692 		struct nvmet_rdma_queue *queue)
1693 {
1694 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1695 
1696 	mutex_lock(&nvmet_rdma_queue_mutex);
1697 	if (!list_empty(&queue->queue_list))
1698 		list_del_init(&queue->queue_list);
1699 	mutex_unlock(&nvmet_rdma_queue_mutex);
1700 
1701 	pr_err("failed to connect queue %d\n", queue->idx);
1702 	queue_work(nvmet_wq, &queue->release_work);
1703 }
1704 
1705 /**
1706  * nvmet_rdma_device_removal() - Handle RDMA device removal
1707  * @cm_id:	rdma_cm id, used for nvmet port
1708  * @queue:      nvmet rdma queue (cm id qp_context)
1709  *
1710  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1711  * to unplug. Note that this event can be generated on a normal
1712  * queue cm_id and/or a device bound listener cm_id (where in this
1713  * case queue will be null).
1714  *
1715  * We registered an ib_client to handle device removal for queues,
1716  * so we only need to handle the listening port cm_ids. In this case
1717  * we nullify the priv to prevent double cm_id destruction and destroying
1718  * the cm_id implicitely by returning a non-zero rc to the callout.
1719  */
1720 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1721 		struct nvmet_rdma_queue *queue)
1722 {
1723 	struct nvmet_rdma_port *port;
1724 
1725 	if (queue) {
1726 		/*
1727 		 * This is a queue cm_id. we have registered
1728 		 * an ib_client to handle queues removal
1729 		 * so don't interfear and just return.
1730 		 */
1731 		return 0;
1732 	}
1733 
1734 	port = cm_id->context;
1735 
1736 	/*
1737 	 * This is a listener cm_id. Make sure that
1738 	 * future remove_port won't invoke a double
1739 	 * cm_id destroy. use atomic xchg to make sure
1740 	 * we don't compete with remove_port.
1741 	 */
1742 	if (xchg(&port->cm_id, NULL) != cm_id)
1743 		return 0;
1744 
1745 	/*
1746 	 * We need to return 1 so that the core will destroy
1747 	 * it's own ID.  What a great API design..
1748 	 */
1749 	return 1;
1750 }
1751 
1752 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1753 		struct rdma_cm_event *event)
1754 {
1755 	struct nvmet_rdma_queue *queue = NULL;
1756 	int ret = 0;
1757 
1758 	if (cm_id->qp)
1759 		queue = cm_id->qp->qp_context;
1760 
1761 	pr_debug("%s (%d): status %d id %p\n",
1762 		rdma_event_msg(event->event), event->event,
1763 		event->status, cm_id);
1764 
1765 	switch (event->event) {
1766 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1767 		ret = nvmet_rdma_queue_connect(cm_id, event);
1768 		break;
1769 	case RDMA_CM_EVENT_ESTABLISHED:
1770 		nvmet_rdma_queue_established(queue);
1771 		break;
1772 	case RDMA_CM_EVENT_ADDR_CHANGE:
1773 		if (!queue) {
1774 			struct nvmet_rdma_port *port = cm_id->context;
1775 
1776 			queue_delayed_work(nvmet_wq, &port->repair_work, 0);
1777 			break;
1778 		}
1779 		fallthrough;
1780 	case RDMA_CM_EVENT_DISCONNECTED:
1781 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1782 		nvmet_rdma_queue_disconnect(queue);
1783 		break;
1784 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1785 		ret = nvmet_rdma_device_removal(cm_id, queue);
1786 		break;
1787 	case RDMA_CM_EVENT_REJECTED:
1788 		pr_debug("Connection rejected: %s\n",
1789 			 rdma_reject_msg(cm_id, event->status));
1790 		fallthrough;
1791 	case RDMA_CM_EVENT_UNREACHABLE:
1792 	case RDMA_CM_EVENT_CONNECT_ERROR:
1793 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1794 		break;
1795 	default:
1796 		pr_err("received unrecognized RDMA CM event %d\n",
1797 			event->event);
1798 		break;
1799 	}
1800 
1801 	return ret;
1802 }
1803 
1804 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1805 {
1806 	struct nvmet_rdma_queue *queue;
1807 
1808 restart:
1809 	mutex_lock(&nvmet_rdma_queue_mutex);
1810 	list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1811 		if (queue->nvme_sq.ctrl == ctrl) {
1812 			list_del_init(&queue->queue_list);
1813 			mutex_unlock(&nvmet_rdma_queue_mutex);
1814 
1815 			__nvmet_rdma_queue_disconnect(queue);
1816 			goto restart;
1817 		}
1818 	}
1819 	mutex_unlock(&nvmet_rdma_queue_mutex);
1820 }
1821 
1822 static void nvmet_rdma_destroy_port_queues(struct nvmet_rdma_port *port)
1823 {
1824 	struct nvmet_rdma_queue *queue, *tmp;
1825 	struct nvmet_port *nport = port->nport;
1826 
1827 	mutex_lock(&nvmet_rdma_queue_mutex);
1828 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1829 				 queue_list) {
1830 		if (queue->port != nport)
1831 			continue;
1832 
1833 		list_del_init(&queue->queue_list);
1834 		__nvmet_rdma_queue_disconnect(queue);
1835 	}
1836 	mutex_unlock(&nvmet_rdma_queue_mutex);
1837 }
1838 
1839 static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
1840 {
1841 	struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
1842 
1843 	if (cm_id)
1844 		rdma_destroy_id(cm_id);
1845 
1846 	/*
1847 	 * Destroy the remaining queues, which are not belong to any
1848 	 * controller yet. Do it here after the RDMA-CM was destroyed
1849 	 * guarantees that no new queue will be created.
1850 	 */
1851 	nvmet_rdma_destroy_port_queues(port);
1852 }
1853 
1854 static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
1855 {
1856 	struct sockaddr *addr = (struct sockaddr *)&port->addr;
1857 	struct rdma_cm_id *cm_id;
1858 	int ret;
1859 
1860 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1861 			RDMA_PS_TCP, IB_QPT_RC);
1862 	if (IS_ERR(cm_id)) {
1863 		pr_err("CM ID creation failed\n");
1864 		return PTR_ERR(cm_id);
1865 	}
1866 
1867 	/*
1868 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1869 	 * at the same time.
1870 	 */
1871 	ret = rdma_set_afonly(cm_id, 1);
1872 	if (ret) {
1873 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1874 		goto out_destroy_id;
1875 	}
1876 
1877 	ret = rdma_bind_addr(cm_id, addr);
1878 	if (ret) {
1879 		pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
1880 		goto out_destroy_id;
1881 	}
1882 
1883 	ret = rdma_listen(cm_id, 128);
1884 	if (ret) {
1885 		pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
1886 		goto out_destroy_id;
1887 	}
1888 
1889 	port->cm_id = cm_id;
1890 	return 0;
1891 
1892 out_destroy_id:
1893 	rdma_destroy_id(cm_id);
1894 	return ret;
1895 }
1896 
1897 static void nvmet_rdma_repair_port_work(struct work_struct *w)
1898 {
1899 	struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
1900 			struct nvmet_rdma_port, repair_work);
1901 	int ret;
1902 
1903 	nvmet_rdma_disable_port(port);
1904 	ret = nvmet_rdma_enable_port(port);
1905 	if (ret)
1906 		queue_delayed_work(nvmet_wq, &port->repair_work, 5 * HZ);
1907 }
1908 
1909 static int nvmet_rdma_add_port(struct nvmet_port *nport)
1910 {
1911 	struct nvmet_rdma_port *port;
1912 	__kernel_sa_family_t af;
1913 	int ret;
1914 
1915 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1916 	if (!port)
1917 		return -ENOMEM;
1918 
1919 	nport->priv = port;
1920 	port->nport = nport;
1921 	INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
1922 
1923 	switch (nport->disc_addr.adrfam) {
1924 	case NVMF_ADDR_FAMILY_IP4:
1925 		af = AF_INET;
1926 		break;
1927 	case NVMF_ADDR_FAMILY_IP6:
1928 		af = AF_INET6;
1929 		break;
1930 	default:
1931 		pr_err("address family %d not supported\n",
1932 			nport->disc_addr.adrfam);
1933 		ret = -EINVAL;
1934 		goto out_free_port;
1935 	}
1936 
1937 	if (nport->inline_data_size < 0) {
1938 		nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1939 	} else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1940 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1941 			nport->inline_data_size,
1942 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1943 		nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1944 	}
1945 
1946 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1947 			nport->disc_addr.trsvcid, &port->addr);
1948 	if (ret) {
1949 		pr_err("malformed ip/port passed: %s:%s\n",
1950 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1951 		goto out_free_port;
1952 	}
1953 
1954 	ret = nvmet_rdma_enable_port(port);
1955 	if (ret)
1956 		goto out_free_port;
1957 
1958 	pr_info("enabling port %d (%pISpcs)\n",
1959 		le16_to_cpu(nport->disc_addr.portid),
1960 		(struct sockaddr *)&port->addr);
1961 
1962 	return 0;
1963 
1964 out_free_port:
1965 	kfree(port);
1966 	return ret;
1967 }
1968 
1969 static void nvmet_rdma_remove_port(struct nvmet_port *nport)
1970 {
1971 	struct nvmet_rdma_port *port = nport->priv;
1972 
1973 	cancel_delayed_work_sync(&port->repair_work);
1974 	nvmet_rdma_disable_port(port);
1975 	kfree(port);
1976 }
1977 
1978 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1979 		struct nvmet_port *nport, char *traddr)
1980 {
1981 	struct nvmet_rdma_port *port = nport->priv;
1982 	struct rdma_cm_id *cm_id = port->cm_id;
1983 
1984 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1985 		struct nvmet_rdma_rsp *rsp =
1986 			container_of(req, struct nvmet_rdma_rsp, req);
1987 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1988 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1989 
1990 		sprintf(traddr, "%pISc", addr);
1991 	} else {
1992 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1993 	}
1994 }
1995 
1996 static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
1997 {
1998 	if (ctrl->pi_support)
1999 		return NVMET_RDMA_MAX_METADATA_MDTS;
2000 	return NVMET_RDMA_MAX_MDTS;
2001 }
2002 
2003 static u16 nvmet_rdma_get_max_queue_size(const struct nvmet_ctrl *ctrl)
2004 {
2005 	return NVME_RDMA_MAX_QUEUE_SIZE;
2006 }
2007 
2008 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
2009 	.owner			= THIS_MODULE,
2010 	.type			= NVMF_TRTYPE_RDMA,
2011 	.msdbd			= 1,
2012 	.flags			= NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
2013 	.add_port		= nvmet_rdma_add_port,
2014 	.remove_port		= nvmet_rdma_remove_port,
2015 	.queue_response		= nvmet_rdma_queue_response,
2016 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
2017 	.disc_traddr		= nvmet_rdma_disc_port_addr,
2018 	.get_mdts		= nvmet_rdma_get_mdts,
2019 	.get_max_queue_size	= nvmet_rdma_get_max_queue_size,
2020 };
2021 
2022 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2023 {
2024 	struct nvmet_rdma_queue *queue, *tmp;
2025 	struct nvmet_rdma_device *ndev;
2026 	bool found = false;
2027 
2028 	mutex_lock(&device_list_mutex);
2029 	list_for_each_entry(ndev, &device_list, entry) {
2030 		if (ndev->device == ib_device) {
2031 			found = true;
2032 			break;
2033 		}
2034 	}
2035 	mutex_unlock(&device_list_mutex);
2036 
2037 	if (!found)
2038 		return;
2039 
2040 	/*
2041 	 * IB Device that is used by nvmet controllers is being removed,
2042 	 * delete all queues using this device.
2043 	 */
2044 	mutex_lock(&nvmet_rdma_queue_mutex);
2045 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
2046 				 queue_list) {
2047 		if (queue->dev->device != ib_device)
2048 			continue;
2049 
2050 		pr_info("Removing queue %d\n", queue->idx);
2051 		list_del_init(&queue->queue_list);
2052 		__nvmet_rdma_queue_disconnect(queue);
2053 	}
2054 	mutex_unlock(&nvmet_rdma_queue_mutex);
2055 
2056 	flush_workqueue(nvmet_wq);
2057 }
2058 
2059 static struct ib_client nvmet_rdma_ib_client = {
2060 	.name   = "nvmet_rdma",
2061 	.remove = nvmet_rdma_remove_one
2062 };
2063 
2064 static int __init nvmet_rdma_init(void)
2065 {
2066 	int ret;
2067 
2068 	ret = ib_register_client(&nvmet_rdma_ib_client);
2069 	if (ret)
2070 		return ret;
2071 
2072 	ret = nvmet_register_transport(&nvmet_rdma_ops);
2073 	if (ret)
2074 		goto err_ib_client;
2075 
2076 	return 0;
2077 
2078 err_ib_client:
2079 	ib_unregister_client(&nvmet_rdma_ib_client);
2080 	return ret;
2081 }
2082 
2083 static void __exit nvmet_rdma_exit(void)
2084 {
2085 	nvmet_unregister_transport(&nvmet_rdma_ops);
2086 	ib_unregister_client(&nvmet_rdma_ib_client);
2087 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
2088 	ida_destroy(&nvmet_rdma_queue_ida);
2089 }
2090 
2091 module_init(nvmet_rdma_init);
2092 module_exit(nvmet_rdma_exit);
2093 
2094 MODULE_LICENSE("GPL v2");
2095 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
2096