xref: /openbmc/linux/drivers/nvme/target/rdma.c (revision 281e4684)
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/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/err.h>
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/nvme.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/wait.h>
17 #include <linux/inet.h>
18 #include <asm/unaligned.h>
19 
20 #include <rdma/ib_verbs.h>
21 #include <rdma/rdma_cm.h>
22 #include <rdma/rw.h>
23 #include <rdma/ib_cm.h>
24 
25 #include <linux/nvme-rdma.h>
26 #include "nvmet.h"
27 
28 /*
29  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
30  */
31 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE	PAGE_SIZE
32 #define NVMET_RDMA_MAX_INLINE_SGE		4
33 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE		max_t(int, SZ_16K, PAGE_SIZE)
34 
35 /* Assume mpsmin == device_page_size == 4KB */
36 #define NVMET_RDMA_MAX_MDTS			8
37 #define NVMET_RDMA_MAX_METADATA_MDTS		5
38 
39 struct nvmet_rdma_srq;
40 
41 struct nvmet_rdma_cmd {
42 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
43 	struct ib_cqe		cqe;
44 	struct ib_recv_wr	wr;
45 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
46 	struct nvme_command     *nvme_cmd;
47 	struct nvmet_rdma_queue	*queue;
48 	struct nvmet_rdma_srq   *nsrq;
49 };
50 
51 enum {
52 	NVMET_RDMA_REQ_INLINE_DATA	= (1 << 0),
53 	NVMET_RDMA_REQ_INVALIDATE_RKEY	= (1 << 1),
54 };
55 
56 struct nvmet_rdma_rsp {
57 	struct ib_sge		send_sge;
58 	struct ib_cqe		send_cqe;
59 	struct ib_send_wr	send_wr;
60 
61 	struct nvmet_rdma_cmd	*cmd;
62 	struct nvmet_rdma_queue	*queue;
63 
64 	struct ib_cqe		read_cqe;
65 	struct ib_cqe		write_cqe;
66 	struct rdma_rw_ctx	rw;
67 
68 	struct nvmet_req	req;
69 
70 	bool			allocated;
71 	u8			n_rdma;
72 	u32			flags;
73 	u32			invalidate_rkey;
74 
75 	struct list_head	wait_list;
76 	struct list_head	free_list;
77 };
78 
79 enum nvmet_rdma_queue_state {
80 	NVMET_RDMA_Q_CONNECTING,
81 	NVMET_RDMA_Q_LIVE,
82 	NVMET_RDMA_Q_DISCONNECTING,
83 };
84 
85 struct nvmet_rdma_queue {
86 	struct rdma_cm_id	*cm_id;
87 	struct ib_qp		*qp;
88 	struct nvmet_port	*port;
89 	struct ib_cq		*cq;
90 	atomic_t		sq_wr_avail;
91 	struct nvmet_rdma_device *dev;
92 	struct nvmet_rdma_srq   *nsrq;
93 	spinlock_t		state_lock;
94 	enum nvmet_rdma_queue_state state;
95 	struct nvmet_cq		nvme_cq;
96 	struct nvmet_sq		nvme_sq;
97 
98 	struct nvmet_rdma_rsp	*rsps;
99 	struct list_head	free_rsps;
100 	spinlock_t		rsps_lock;
101 	struct nvmet_rdma_cmd	*cmds;
102 
103 	struct work_struct	release_work;
104 	struct list_head	rsp_wait_list;
105 	struct list_head	rsp_wr_wait_list;
106 	spinlock_t		rsp_wr_wait_lock;
107 
108 	int			idx;
109 	int			host_qid;
110 	int			comp_vector;
111 	int			recv_queue_size;
112 	int			send_queue_size;
113 
114 	struct list_head	queue_list;
115 };
116 
117 struct nvmet_rdma_port {
118 	struct nvmet_port	*nport;
119 	struct sockaddr_storage addr;
120 	struct rdma_cm_id	*cm_id;
121 	struct delayed_work	repair_work;
122 };
123 
124 struct nvmet_rdma_srq {
125 	struct ib_srq            *srq;
126 	struct nvmet_rdma_cmd    *cmds;
127 	struct nvmet_rdma_device *ndev;
128 };
129 
130 struct nvmet_rdma_device {
131 	struct ib_device	*device;
132 	struct ib_pd		*pd;
133 	struct nvmet_rdma_srq	**srqs;
134 	int			srq_count;
135 	size_t			srq_size;
136 	struct kref		ref;
137 	struct list_head	entry;
138 	int			inline_data_size;
139 	int			inline_page_count;
140 };
141 
142 static bool nvmet_rdma_use_srq;
143 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
144 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
145 
146 static int srq_size_set(const char *val, const struct kernel_param *kp);
147 static const struct kernel_param_ops srq_size_ops = {
148 	.set = srq_size_set,
149 	.get = param_get_int,
150 };
151 
152 static int nvmet_rdma_srq_size = 1024;
153 module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
154 MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");
155 
156 static DEFINE_IDA(nvmet_rdma_queue_ida);
157 static LIST_HEAD(nvmet_rdma_queue_list);
158 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
159 
160 static LIST_HEAD(device_list);
161 static DEFINE_MUTEX(device_list_mutex);
162 
163 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
164 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
165 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
166 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
167 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
168 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
169 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
170 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
171 				struct nvmet_rdma_rsp *r);
172 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
173 				struct nvmet_rdma_rsp *r);
174 
175 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
176 
177 static int srq_size_set(const char *val, const struct kernel_param *kp)
178 {
179 	int n = 0, ret;
180 
181 	ret = kstrtoint(val, 10, &n);
182 	if (ret != 0 || n < 256)
183 		return -EINVAL;
184 
185 	return param_set_int(val, kp);
186 }
187 
188 static int num_pages(int len)
189 {
190 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
191 }
192 
193 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
194 {
195 	return nvme_is_write(rsp->req.cmd) &&
196 		rsp->req.transfer_len &&
197 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
198 }
199 
200 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
201 {
202 	return !nvme_is_write(rsp->req.cmd) &&
203 		rsp->req.transfer_len &&
204 		!rsp->req.cqe->status &&
205 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
206 }
207 
208 static inline struct nvmet_rdma_rsp *
209 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
210 {
211 	struct nvmet_rdma_rsp *rsp;
212 	unsigned long flags;
213 
214 	spin_lock_irqsave(&queue->rsps_lock, flags);
215 	rsp = list_first_entry_or_null(&queue->free_rsps,
216 				struct nvmet_rdma_rsp, free_list);
217 	if (likely(rsp))
218 		list_del(&rsp->free_list);
219 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
220 
221 	if (unlikely(!rsp)) {
222 		int ret;
223 
224 		rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
225 		if (unlikely(!rsp))
226 			return NULL;
227 		ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
228 		if (unlikely(ret)) {
229 			kfree(rsp);
230 			return NULL;
231 		}
232 
233 		rsp->allocated = true;
234 	}
235 
236 	return rsp;
237 }
238 
239 static inline void
240 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
241 {
242 	unsigned long flags;
243 
244 	if (unlikely(rsp->allocated)) {
245 		nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
246 		kfree(rsp);
247 		return;
248 	}
249 
250 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
251 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
252 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
253 }
254 
255 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
256 				struct nvmet_rdma_cmd *c)
257 {
258 	struct scatterlist *sg;
259 	struct ib_sge *sge;
260 	int i;
261 
262 	if (!ndev->inline_data_size)
263 		return;
264 
265 	sg = c->inline_sg;
266 	sge = &c->sge[1];
267 
268 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
269 		if (sge->length)
270 			ib_dma_unmap_page(ndev->device, sge->addr,
271 					sge->length, DMA_FROM_DEVICE);
272 		if (sg_page(sg))
273 			__free_page(sg_page(sg));
274 	}
275 }
276 
277 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
278 				struct nvmet_rdma_cmd *c)
279 {
280 	struct scatterlist *sg;
281 	struct ib_sge *sge;
282 	struct page *pg;
283 	int len;
284 	int i;
285 
286 	if (!ndev->inline_data_size)
287 		return 0;
288 
289 	sg = c->inline_sg;
290 	sg_init_table(sg, ndev->inline_page_count);
291 	sge = &c->sge[1];
292 	len = ndev->inline_data_size;
293 
294 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
295 		pg = alloc_page(GFP_KERNEL);
296 		if (!pg)
297 			goto out_err;
298 		sg_assign_page(sg, pg);
299 		sge->addr = ib_dma_map_page(ndev->device,
300 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
301 		if (ib_dma_mapping_error(ndev->device, sge->addr))
302 			goto out_err;
303 		sge->length = min_t(int, len, PAGE_SIZE);
304 		sge->lkey = ndev->pd->local_dma_lkey;
305 		len -= sge->length;
306 	}
307 
308 	return 0;
309 out_err:
310 	for (; i >= 0; i--, sg--, sge--) {
311 		if (sge->length)
312 			ib_dma_unmap_page(ndev->device, sge->addr,
313 					sge->length, DMA_FROM_DEVICE);
314 		if (sg_page(sg))
315 			__free_page(sg_page(sg));
316 	}
317 	return -ENOMEM;
318 }
319 
320 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
321 			struct nvmet_rdma_cmd *c, bool admin)
322 {
323 	/* NVMe command / RDMA RECV */
324 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
325 	if (!c->nvme_cmd)
326 		goto out;
327 
328 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
329 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
330 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
331 		goto out_free_cmd;
332 
333 	c->sge[0].length = sizeof(*c->nvme_cmd);
334 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
335 
336 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
337 		goto out_unmap_cmd;
338 
339 	c->cqe.done = nvmet_rdma_recv_done;
340 
341 	c->wr.wr_cqe = &c->cqe;
342 	c->wr.sg_list = c->sge;
343 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
344 
345 	return 0;
346 
347 out_unmap_cmd:
348 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
349 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
350 out_free_cmd:
351 	kfree(c->nvme_cmd);
352 
353 out:
354 	return -ENOMEM;
355 }
356 
357 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
358 		struct nvmet_rdma_cmd *c, bool admin)
359 {
360 	if (!admin)
361 		nvmet_rdma_free_inline_pages(ndev, c);
362 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
363 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
364 	kfree(c->nvme_cmd);
365 }
366 
367 static struct nvmet_rdma_cmd *
368 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
369 		int nr_cmds, bool admin)
370 {
371 	struct nvmet_rdma_cmd *cmds;
372 	int ret = -EINVAL, i;
373 
374 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
375 	if (!cmds)
376 		goto out;
377 
378 	for (i = 0; i < nr_cmds; i++) {
379 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
380 		if (ret)
381 			goto out_free;
382 	}
383 
384 	return cmds;
385 
386 out_free:
387 	while (--i >= 0)
388 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
389 	kfree(cmds);
390 out:
391 	return ERR_PTR(ret);
392 }
393 
394 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
395 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
396 {
397 	int i;
398 
399 	for (i = 0; i < nr_cmds; i++)
400 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
401 	kfree(cmds);
402 }
403 
404 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
405 		struct nvmet_rdma_rsp *r)
406 {
407 	/* NVMe CQE / RDMA SEND */
408 	r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
409 	if (!r->req.cqe)
410 		goto out;
411 
412 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
413 			sizeof(*r->req.cqe), DMA_TO_DEVICE);
414 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
415 		goto out_free_rsp;
416 
417 	if (!ib_uses_virt_dma(ndev->device))
418 		r->req.p2p_client = &ndev->device->dev;
419 	r->send_sge.length = sizeof(*r->req.cqe);
420 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
421 
422 	r->send_cqe.done = nvmet_rdma_send_done;
423 
424 	r->send_wr.wr_cqe = &r->send_cqe;
425 	r->send_wr.sg_list = &r->send_sge;
426 	r->send_wr.num_sge = 1;
427 	r->send_wr.send_flags = IB_SEND_SIGNALED;
428 
429 	/* Data In / RDMA READ */
430 	r->read_cqe.done = nvmet_rdma_read_data_done;
431 	/* Data Out / RDMA WRITE */
432 	r->write_cqe.done = nvmet_rdma_write_data_done;
433 
434 	return 0;
435 
436 out_free_rsp:
437 	kfree(r->req.cqe);
438 out:
439 	return -ENOMEM;
440 }
441 
442 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
443 		struct nvmet_rdma_rsp *r)
444 {
445 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
446 				sizeof(*r->req.cqe), DMA_TO_DEVICE);
447 	kfree(r->req.cqe);
448 }
449 
450 static int
451 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
452 {
453 	struct nvmet_rdma_device *ndev = queue->dev;
454 	int nr_rsps = queue->recv_queue_size * 2;
455 	int ret = -EINVAL, i;
456 
457 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
458 			GFP_KERNEL);
459 	if (!queue->rsps)
460 		goto out;
461 
462 	for (i = 0; i < nr_rsps; i++) {
463 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
464 
465 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
466 		if (ret)
467 			goto out_free;
468 
469 		list_add_tail(&rsp->free_list, &queue->free_rsps);
470 	}
471 
472 	return 0;
473 
474 out_free:
475 	while (--i >= 0) {
476 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
477 
478 		list_del(&rsp->free_list);
479 		nvmet_rdma_free_rsp(ndev, rsp);
480 	}
481 	kfree(queue->rsps);
482 out:
483 	return ret;
484 }
485 
486 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
487 {
488 	struct nvmet_rdma_device *ndev = queue->dev;
489 	int i, nr_rsps = queue->recv_queue_size * 2;
490 
491 	for (i = 0; i < nr_rsps; i++) {
492 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
493 
494 		list_del(&rsp->free_list);
495 		nvmet_rdma_free_rsp(ndev, rsp);
496 	}
497 	kfree(queue->rsps);
498 }
499 
500 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
501 		struct nvmet_rdma_cmd *cmd)
502 {
503 	int ret;
504 
505 	ib_dma_sync_single_for_device(ndev->device,
506 		cmd->sge[0].addr, cmd->sge[0].length,
507 		DMA_FROM_DEVICE);
508 
509 	if (cmd->nsrq)
510 		ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
511 	else
512 		ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);
513 
514 	if (unlikely(ret))
515 		pr_err("post_recv cmd failed\n");
516 
517 	return ret;
518 }
519 
520 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
521 {
522 	spin_lock(&queue->rsp_wr_wait_lock);
523 	while (!list_empty(&queue->rsp_wr_wait_list)) {
524 		struct nvmet_rdma_rsp *rsp;
525 		bool ret;
526 
527 		rsp = list_entry(queue->rsp_wr_wait_list.next,
528 				struct nvmet_rdma_rsp, wait_list);
529 		list_del(&rsp->wait_list);
530 
531 		spin_unlock(&queue->rsp_wr_wait_lock);
532 		ret = nvmet_rdma_execute_command(rsp);
533 		spin_lock(&queue->rsp_wr_wait_lock);
534 
535 		if (!ret) {
536 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
537 			break;
538 		}
539 	}
540 	spin_unlock(&queue->rsp_wr_wait_lock);
541 }
542 
543 static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
544 {
545 	struct ib_mr_status mr_status;
546 	int ret;
547 	u16 status = 0;
548 
549 	ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
550 	if (ret) {
551 		pr_err("ib_check_mr_status failed, ret %d\n", ret);
552 		return NVME_SC_INVALID_PI;
553 	}
554 
555 	if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
556 		switch (mr_status.sig_err.err_type) {
557 		case IB_SIG_BAD_GUARD:
558 			status = NVME_SC_GUARD_CHECK;
559 			break;
560 		case IB_SIG_BAD_REFTAG:
561 			status = NVME_SC_REFTAG_CHECK;
562 			break;
563 		case IB_SIG_BAD_APPTAG:
564 			status = NVME_SC_APPTAG_CHECK;
565 			break;
566 		}
567 		pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
568 		       mr_status.sig_err.err_type,
569 		       mr_status.sig_err.expected,
570 		       mr_status.sig_err.actual);
571 	}
572 
573 	return status;
574 }
575 
576 static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
577 		struct nvme_command *cmd, struct ib_sig_domain *domain,
578 		u16 control, u8 pi_type)
579 {
580 	domain->sig_type = IB_SIG_TYPE_T10_DIF;
581 	domain->sig.dif.bg_type = IB_T10DIF_CRC;
582 	domain->sig.dif.pi_interval = 1 << bi->interval_exp;
583 	domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
584 	if (control & NVME_RW_PRINFO_PRCHK_REF)
585 		domain->sig.dif.ref_remap = true;
586 
587 	domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
588 	domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
589 	domain->sig.dif.app_escape = true;
590 	if (pi_type == NVME_NS_DPS_PI_TYPE3)
591 		domain->sig.dif.ref_escape = true;
592 }
593 
594 static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
595 				     struct ib_sig_attrs *sig_attrs)
596 {
597 	struct nvme_command *cmd = req->cmd;
598 	u16 control = le16_to_cpu(cmd->rw.control);
599 	u8 pi_type = req->ns->pi_type;
600 	struct blk_integrity *bi;
601 
602 	bi = bdev_get_integrity(req->ns->bdev);
603 
604 	memset(sig_attrs, 0, sizeof(*sig_attrs));
605 
606 	if (control & NVME_RW_PRINFO_PRACT) {
607 		/* for WRITE_INSERT/READ_STRIP no wire domain */
608 		sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
609 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
610 					  pi_type);
611 		/* Clear the PRACT bit since HCA will generate/verify the PI */
612 		control &= ~NVME_RW_PRINFO_PRACT;
613 		cmd->rw.control = cpu_to_le16(control);
614 		/* PI is added by the HW */
615 		req->transfer_len += req->metadata_len;
616 	} else {
617 		/* for WRITE_PASS/READ_PASS both wire/memory domains exist */
618 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
619 					  pi_type);
620 		nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
621 					  pi_type);
622 	}
623 
624 	if (control & NVME_RW_PRINFO_PRCHK_REF)
625 		sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
626 	if (control & NVME_RW_PRINFO_PRCHK_GUARD)
627 		sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
628 	if (control & NVME_RW_PRINFO_PRCHK_APP)
629 		sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
630 }
631 
632 static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
633 				  struct ib_sig_attrs *sig_attrs)
634 {
635 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
636 	struct nvmet_req *req = &rsp->req;
637 	int ret;
638 
639 	if (req->metadata_len)
640 		ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
641 			cm_id->port_num, req->sg, req->sg_cnt,
642 			req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
643 			addr, key, nvmet_data_dir(req));
644 	else
645 		ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
646 				       req->sg, req->sg_cnt, 0, addr, key,
647 				       nvmet_data_dir(req));
648 
649 	return ret;
650 }
651 
652 static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
653 {
654 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
655 	struct nvmet_req *req = &rsp->req;
656 
657 	if (req->metadata_len)
658 		rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
659 			cm_id->port_num, req->sg, req->sg_cnt,
660 			req->metadata_sg, req->metadata_sg_cnt,
661 			nvmet_data_dir(req));
662 	else
663 		rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
664 				    req->sg, req->sg_cnt, nvmet_data_dir(req));
665 }
666 
667 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
668 {
669 	struct nvmet_rdma_queue *queue = rsp->queue;
670 
671 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
672 
673 	if (rsp->n_rdma)
674 		nvmet_rdma_rw_ctx_destroy(rsp);
675 
676 	if (rsp->req.sg != rsp->cmd->inline_sg)
677 		nvmet_req_free_sgls(&rsp->req);
678 
679 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
680 		nvmet_rdma_process_wr_wait_list(queue);
681 
682 	nvmet_rdma_put_rsp(rsp);
683 }
684 
685 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
686 {
687 	if (queue->nvme_sq.ctrl) {
688 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
689 	} else {
690 		/*
691 		 * we didn't setup the controller yet in case
692 		 * of admin connect error, just disconnect and
693 		 * cleanup the queue
694 		 */
695 		nvmet_rdma_queue_disconnect(queue);
696 	}
697 }
698 
699 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
700 {
701 	struct nvmet_rdma_rsp *rsp =
702 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
703 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
704 
705 	nvmet_rdma_release_rsp(rsp);
706 
707 	if (unlikely(wc->status != IB_WC_SUCCESS &&
708 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
709 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
710 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
711 		nvmet_rdma_error_comp(queue);
712 	}
713 }
714 
715 static void nvmet_rdma_queue_response(struct nvmet_req *req)
716 {
717 	struct nvmet_rdma_rsp *rsp =
718 		container_of(req, struct nvmet_rdma_rsp, req);
719 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
720 	struct ib_send_wr *first_wr;
721 
722 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
723 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
724 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
725 	} else {
726 		rsp->send_wr.opcode = IB_WR_SEND;
727 	}
728 
729 	if (nvmet_rdma_need_data_out(rsp)) {
730 		if (rsp->req.metadata_len)
731 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
732 					cm_id->port_num, &rsp->write_cqe, NULL);
733 		else
734 			first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
735 					cm_id->port_num, NULL, &rsp->send_wr);
736 	} else {
737 		first_wr = &rsp->send_wr;
738 	}
739 
740 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
741 
742 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
743 		rsp->send_sge.addr, rsp->send_sge.length,
744 		DMA_TO_DEVICE);
745 
746 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
747 		pr_err("sending cmd response failed\n");
748 		nvmet_rdma_release_rsp(rsp);
749 	}
750 }
751 
752 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
753 {
754 	struct nvmet_rdma_rsp *rsp =
755 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
756 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
757 	u16 status = 0;
758 
759 	WARN_ON(rsp->n_rdma <= 0);
760 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
761 	rsp->n_rdma = 0;
762 
763 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
764 		nvmet_rdma_rw_ctx_destroy(rsp);
765 		nvmet_req_uninit(&rsp->req);
766 		nvmet_rdma_release_rsp(rsp);
767 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
768 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
769 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
770 			nvmet_rdma_error_comp(queue);
771 		}
772 		return;
773 	}
774 
775 	if (rsp->req.metadata_len)
776 		status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
777 	nvmet_rdma_rw_ctx_destroy(rsp);
778 
779 	if (unlikely(status))
780 		nvmet_req_complete(&rsp->req, status);
781 	else
782 		rsp->req.execute(&rsp->req);
783 }
784 
785 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
786 {
787 	struct nvmet_rdma_rsp *rsp =
788 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
789 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
790 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
791 	u16 status;
792 
793 	if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
794 		return;
795 
796 	WARN_ON(rsp->n_rdma <= 0);
797 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
798 	rsp->n_rdma = 0;
799 
800 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
801 		nvmet_rdma_rw_ctx_destroy(rsp);
802 		nvmet_req_uninit(&rsp->req);
803 		nvmet_rdma_release_rsp(rsp);
804 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
805 			pr_info("RDMA WRITE for CQE failed with status %s (%d).\n",
806 				ib_wc_status_msg(wc->status), wc->status);
807 			nvmet_rdma_error_comp(queue);
808 		}
809 		return;
810 	}
811 
812 	/*
813 	 * Upon RDMA completion check the signature status
814 	 * - if succeeded send good NVMe response
815 	 * - if failed send bad NVMe response with appropriate error
816 	 */
817 	status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
818 	if (unlikely(status))
819 		rsp->req.cqe->status = cpu_to_le16(status << 1);
820 	nvmet_rdma_rw_ctx_destroy(rsp);
821 
822 	if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
823 		pr_err("sending cmd response failed\n");
824 		nvmet_rdma_release_rsp(rsp);
825 	}
826 }
827 
828 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
829 		u64 off)
830 {
831 	int sg_count = num_pages(len);
832 	struct scatterlist *sg;
833 	int i;
834 
835 	sg = rsp->cmd->inline_sg;
836 	for (i = 0; i < sg_count; i++, sg++) {
837 		if (i < sg_count - 1)
838 			sg_unmark_end(sg);
839 		else
840 			sg_mark_end(sg);
841 		sg->offset = off;
842 		sg->length = min_t(int, len, PAGE_SIZE - off);
843 		len -= sg->length;
844 		if (!i)
845 			off = 0;
846 	}
847 
848 	rsp->req.sg = rsp->cmd->inline_sg;
849 	rsp->req.sg_cnt = sg_count;
850 }
851 
852 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
853 {
854 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
855 	u64 off = le64_to_cpu(sgl->addr);
856 	u32 len = le32_to_cpu(sgl->length);
857 
858 	if (!nvme_is_write(rsp->req.cmd)) {
859 		rsp->req.error_loc =
860 			offsetof(struct nvme_common_command, opcode);
861 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
862 	}
863 
864 	if (off + len > rsp->queue->dev->inline_data_size) {
865 		pr_err("invalid inline data offset!\n");
866 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
867 	}
868 
869 	/* no data command? */
870 	if (!len)
871 		return 0;
872 
873 	nvmet_rdma_use_inline_sg(rsp, len, off);
874 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
875 	rsp->req.transfer_len += len;
876 	return 0;
877 }
878 
879 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
880 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
881 {
882 	u64 addr = le64_to_cpu(sgl->addr);
883 	u32 key = get_unaligned_le32(sgl->key);
884 	struct ib_sig_attrs sig_attrs;
885 	int ret;
886 
887 	rsp->req.transfer_len = get_unaligned_le24(sgl->length);
888 
889 	/* no data command? */
890 	if (!rsp->req.transfer_len)
891 		return 0;
892 
893 	if (rsp->req.metadata_len)
894 		nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);
895 
896 	ret = nvmet_req_alloc_sgls(&rsp->req);
897 	if (unlikely(ret < 0))
898 		goto error_out;
899 
900 	ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
901 	if (unlikely(ret < 0))
902 		goto error_out;
903 	rsp->n_rdma += ret;
904 
905 	if (invalidate) {
906 		rsp->invalidate_rkey = key;
907 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
908 	}
909 
910 	return 0;
911 
912 error_out:
913 	rsp->req.transfer_len = 0;
914 	return NVME_SC_INTERNAL;
915 }
916 
917 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
918 {
919 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
920 
921 	switch (sgl->type >> 4) {
922 	case NVME_SGL_FMT_DATA_DESC:
923 		switch (sgl->type & 0xf) {
924 		case NVME_SGL_FMT_OFFSET:
925 			return nvmet_rdma_map_sgl_inline(rsp);
926 		default:
927 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
928 			rsp->req.error_loc =
929 				offsetof(struct nvme_common_command, dptr);
930 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
931 		}
932 	case NVME_KEY_SGL_FMT_DATA_DESC:
933 		switch (sgl->type & 0xf) {
934 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
935 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
936 		case NVME_SGL_FMT_ADDRESS:
937 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
938 		default:
939 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
940 			rsp->req.error_loc =
941 				offsetof(struct nvme_common_command, dptr);
942 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
943 		}
944 	default:
945 		pr_err("invalid SGL type: %#x\n", sgl->type);
946 		rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
947 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
948 	}
949 }
950 
951 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
952 {
953 	struct nvmet_rdma_queue *queue = rsp->queue;
954 
955 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
956 			&queue->sq_wr_avail) < 0)) {
957 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
958 				1 + rsp->n_rdma, queue->idx,
959 				queue->nvme_sq.ctrl->cntlid);
960 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
961 		return false;
962 	}
963 
964 	if (nvmet_rdma_need_data_in(rsp)) {
965 		if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
966 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
967 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
968 	} else {
969 		rsp->req.execute(&rsp->req);
970 	}
971 
972 	return true;
973 }
974 
975 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
976 		struct nvmet_rdma_rsp *cmd)
977 {
978 	u16 status;
979 
980 	ib_dma_sync_single_for_cpu(queue->dev->device,
981 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
982 		DMA_FROM_DEVICE);
983 	ib_dma_sync_single_for_cpu(queue->dev->device,
984 		cmd->send_sge.addr, cmd->send_sge.length,
985 		DMA_TO_DEVICE);
986 
987 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
988 			&queue->nvme_sq, &nvmet_rdma_ops))
989 		return;
990 
991 	status = nvmet_rdma_map_sgl(cmd);
992 	if (status)
993 		goto out_err;
994 
995 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
996 		spin_lock(&queue->rsp_wr_wait_lock);
997 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
998 		spin_unlock(&queue->rsp_wr_wait_lock);
999 	}
1000 
1001 	return;
1002 
1003 out_err:
1004 	nvmet_req_complete(&cmd->req, status);
1005 }
1006 
1007 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1008 {
1009 	struct nvmet_rdma_cmd *cmd =
1010 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
1011 	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
1012 	struct nvmet_rdma_rsp *rsp;
1013 
1014 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1015 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
1016 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
1017 				wc->wr_cqe, ib_wc_status_msg(wc->status),
1018 				wc->status);
1019 			nvmet_rdma_error_comp(queue);
1020 		}
1021 		return;
1022 	}
1023 
1024 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
1025 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
1026 		nvmet_rdma_error_comp(queue);
1027 		return;
1028 	}
1029 
1030 	cmd->queue = queue;
1031 	rsp = nvmet_rdma_get_rsp(queue);
1032 	if (unlikely(!rsp)) {
1033 		/*
1034 		 * we get here only under memory pressure,
1035 		 * silently drop and have the host retry
1036 		 * as we can't even fail it.
1037 		 */
1038 		nvmet_rdma_post_recv(queue->dev, cmd);
1039 		return;
1040 	}
1041 	rsp->queue = queue;
1042 	rsp->cmd = cmd;
1043 	rsp->flags = 0;
1044 	rsp->req.cmd = cmd->nvme_cmd;
1045 	rsp->req.port = queue->port;
1046 	rsp->n_rdma = 0;
1047 
1048 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
1049 		unsigned long flags;
1050 
1051 		spin_lock_irqsave(&queue->state_lock, flags);
1052 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
1053 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
1054 		else
1055 			nvmet_rdma_put_rsp(rsp);
1056 		spin_unlock_irqrestore(&queue->state_lock, flags);
1057 		return;
1058 	}
1059 
1060 	nvmet_rdma_handle_command(queue, rsp);
1061 }
1062 
1063 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
1064 {
1065 	nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
1066 			     false);
1067 	ib_destroy_srq(nsrq->srq);
1068 
1069 	kfree(nsrq);
1070 }
1071 
1072 static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
1073 {
1074 	int i;
1075 
1076 	if (!ndev->srqs)
1077 		return;
1078 
1079 	for (i = 0; i < ndev->srq_count; i++)
1080 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1081 
1082 	kfree(ndev->srqs);
1083 }
1084 
1085 static struct nvmet_rdma_srq *
1086 nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
1087 {
1088 	struct ib_srq_init_attr srq_attr = { NULL, };
1089 	size_t srq_size = ndev->srq_size;
1090 	struct nvmet_rdma_srq *nsrq;
1091 	struct ib_srq *srq;
1092 	int ret, i;
1093 
1094 	nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
1095 	if (!nsrq)
1096 		return ERR_PTR(-ENOMEM);
1097 
1098 	srq_attr.attr.max_wr = srq_size;
1099 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
1100 	srq_attr.attr.srq_limit = 0;
1101 	srq_attr.srq_type = IB_SRQT_BASIC;
1102 	srq = ib_create_srq(ndev->pd, &srq_attr);
1103 	if (IS_ERR(srq)) {
1104 		ret = PTR_ERR(srq);
1105 		goto out_free;
1106 	}
1107 
1108 	nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
1109 	if (IS_ERR(nsrq->cmds)) {
1110 		ret = PTR_ERR(nsrq->cmds);
1111 		goto out_destroy_srq;
1112 	}
1113 
1114 	nsrq->srq = srq;
1115 	nsrq->ndev = ndev;
1116 
1117 	for (i = 0; i < srq_size; i++) {
1118 		nsrq->cmds[i].nsrq = nsrq;
1119 		ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
1120 		if (ret)
1121 			goto out_free_cmds;
1122 	}
1123 
1124 	return nsrq;
1125 
1126 out_free_cmds:
1127 	nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
1128 out_destroy_srq:
1129 	ib_destroy_srq(srq);
1130 out_free:
1131 	kfree(nsrq);
1132 	return ERR_PTR(ret);
1133 }
1134 
1135 static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
1136 {
1137 	int i, ret;
1138 
1139 	if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
1140 		/*
1141 		 * If SRQs aren't supported we just go ahead and use normal
1142 		 * non-shared receive queues.
1143 		 */
1144 		pr_info("SRQ requested but not supported.\n");
1145 		return 0;
1146 	}
1147 
1148 	ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
1149 			     nvmet_rdma_srq_size);
1150 	ndev->srq_count = min(ndev->device->num_comp_vectors,
1151 			      ndev->device->attrs.max_srq);
1152 
1153 	ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
1154 	if (!ndev->srqs)
1155 		return -ENOMEM;
1156 
1157 	for (i = 0; i < ndev->srq_count; i++) {
1158 		ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
1159 		if (IS_ERR(ndev->srqs[i])) {
1160 			ret = PTR_ERR(ndev->srqs[i]);
1161 			goto err_srq;
1162 		}
1163 	}
1164 
1165 	return 0;
1166 
1167 err_srq:
1168 	while (--i >= 0)
1169 		nvmet_rdma_destroy_srq(ndev->srqs[i]);
1170 	kfree(ndev->srqs);
1171 	return ret;
1172 }
1173 
1174 static void nvmet_rdma_free_dev(struct kref *ref)
1175 {
1176 	struct nvmet_rdma_device *ndev =
1177 		container_of(ref, struct nvmet_rdma_device, ref);
1178 
1179 	mutex_lock(&device_list_mutex);
1180 	list_del(&ndev->entry);
1181 	mutex_unlock(&device_list_mutex);
1182 
1183 	nvmet_rdma_destroy_srqs(ndev);
1184 	ib_dealloc_pd(ndev->pd);
1185 
1186 	kfree(ndev);
1187 }
1188 
1189 static struct nvmet_rdma_device *
1190 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
1191 {
1192 	struct nvmet_rdma_port *port = cm_id->context;
1193 	struct nvmet_port *nport = port->nport;
1194 	struct nvmet_rdma_device *ndev;
1195 	int inline_page_count;
1196 	int inline_sge_count;
1197 	int ret;
1198 
1199 	mutex_lock(&device_list_mutex);
1200 	list_for_each_entry(ndev, &device_list, entry) {
1201 		if (ndev->device->node_guid == cm_id->device->node_guid &&
1202 		    kref_get_unless_zero(&ndev->ref))
1203 			goto out_unlock;
1204 	}
1205 
1206 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
1207 	if (!ndev)
1208 		goto out_err;
1209 
1210 	inline_page_count = num_pages(nport->inline_data_size);
1211 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
1212 				cm_id->device->attrs.max_recv_sge) - 1;
1213 	if (inline_page_count > inline_sge_count) {
1214 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
1215 			nport->inline_data_size, cm_id->device->name,
1216 			inline_sge_count * PAGE_SIZE);
1217 		nport->inline_data_size = inline_sge_count * PAGE_SIZE;
1218 		inline_page_count = inline_sge_count;
1219 	}
1220 	ndev->inline_data_size = nport->inline_data_size;
1221 	ndev->inline_page_count = inline_page_count;
1222 
1223 	if (nport->pi_enable && !(cm_id->device->attrs.device_cap_flags &
1224 				  IB_DEVICE_INTEGRITY_HANDOVER)) {
1225 		pr_warn("T10-PI is not supported by device %s. Disabling it\n",
1226 			cm_id->device->name);
1227 		nport->pi_enable = false;
1228 	}
1229 
1230 	ndev->device = cm_id->device;
1231 	kref_init(&ndev->ref);
1232 
1233 	ndev->pd = ib_alloc_pd(ndev->device, 0);
1234 	if (IS_ERR(ndev->pd))
1235 		goto out_free_dev;
1236 
1237 	if (nvmet_rdma_use_srq) {
1238 		ret = nvmet_rdma_init_srqs(ndev);
1239 		if (ret)
1240 			goto out_free_pd;
1241 	}
1242 
1243 	list_add(&ndev->entry, &device_list);
1244 out_unlock:
1245 	mutex_unlock(&device_list_mutex);
1246 	pr_debug("added %s.\n", ndev->device->name);
1247 	return ndev;
1248 
1249 out_free_pd:
1250 	ib_dealloc_pd(ndev->pd);
1251 out_free_dev:
1252 	kfree(ndev);
1253 out_err:
1254 	mutex_unlock(&device_list_mutex);
1255 	return NULL;
1256 }
1257 
1258 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
1259 {
1260 	struct ib_qp_init_attr qp_attr;
1261 	struct nvmet_rdma_device *ndev = queue->dev;
1262 	int nr_cqe, ret, i, factor;
1263 
1264 	/*
1265 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
1266 	 */
1267 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1268 
1269 	queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
1270 				   queue->comp_vector, IB_POLL_WORKQUEUE);
1271 	if (IS_ERR(queue->cq)) {
1272 		ret = PTR_ERR(queue->cq);
1273 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1274 		       nr_cqe + 1, ret);
1275 		goto out;
1276 	}
1277 
1278 	memset(&qp_attr, 0, sizeof(qp_attr));
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_simple_remove(&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_simple_get(&nvmet_rdma_queue_ida, 0, 0, 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_simple_remove(&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_scheduled_work();
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 		schedule_work(&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 	schedule_work(&queue->release_work);
1703 }
1704 
1705 /**
1706  * nvme_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 			schedule_delayed_work(&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_disable_port(struct nvmet_rdma_port *port)
1823 {
1824 	struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
1825 
1826 	if (cm_id)
1827 		rdma_destroy_id(cm_id);
1828 }
1829 
1830 static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
1831 {
1832 	struct sockaddr *addr = (struct sockaddr *)&port->addr;
1833 	struct rdma_cm_id *cm_id;
1834 	int ret;
1835 
1836 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1837 			RDMA_PS_TCP, IB_QPT_RC);
1838 	if (IS_ERR(cm_id)) {
1839 		pr_err("CM ID creation failed\n");
1840 		return PTR_ERR(cm_id);
1841 	}
1842 
1843 	/*
1844 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1845 	 * at the same time.
1846 	 */
1847 	ret = rdma_set_afonly(cm_id, 1);
1848 	if (ret) {
1849 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1850 		goto out_destroy_id;
1851 	}
1852 
1853 	ret = rdma_bind_addr(cm_id, addr);
1854 	if (ret) {
1855 		pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
1856 		goto out_destroy_id;
1857 	}
1858 
1859 	ret = rdma_listen(cm_id, 128);
1860 	if (ret) {
1861 		pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
1862 		goto out_destroy_id;
1863 	}
1864 
1865 	port->cm_id = cm_id;
1866 	return 0;
1867 
1868 out_destroy_id:
1869 	rdma_destroy_id(cm_id);
1870 	return ret;
1871 }
1872 
1873 static void nvmet_rdma_repair_port_work(struct work_struct *w)
1874 {
1875 	struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
1876 			struct nvmet_rdma_port, repair_work);
1877 	int ret;
1878 
1879 	nvmet_rdma_disable_port(port);
1880 	ret = nvmet_rdma_enable_port(port);
1881 	if (ret)
1882 		schedule_delayed_work(&port->repair_work, 5 * HZ);
1883 }
1884 
1885 static int nvmet_rdma_add_port(struct nvmet_port *nport)
1886 {
1887 	struct nvmet_rdma_port *port;
1888 	__kernel_sa_family_t af;
1889 	int ret;
1890 
1891 	port = kzalloc(sizeof(*port), GFP_KERNEL);
1892 	if (!port)
1893 		return -ENOMEM;
1894 
1895 	nport->priv = port;
1896 	port->nport = nport;
1897 	INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
1898 
1899 	switch (nport->disc_addr.adrfam) {
1900 	case NVMF_ADDR_FAMILY_IP4:
1901 		af = AF_INET;
1902 		break;
1903 	case NVMF_ADDR_FAMILY_IP6:
1904 		af = AF_INET6;
1905 		break;
1906 	default:
1907 		pr_err("address family %d not supported\n",
1908 			nport->disc_addr.adrfam);
1909 		ret = -EINVAL;
1910 		goto out_free_port;
1911 	}
1912 
1913 	if (nport->inline_data_size < 0) {
1914 		nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1915 	} else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1916 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1917 			nport->inline_data_size,
1918 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1919 		nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1920 	}
1921 
1922 	ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1923 			nport->disc_addr.trsvcid, &port->addr);
1924 	if (ret) {
1925 		pr_err("malformed ip/port passed: %s:%s\n",
1926 			nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1927 		goto out_free_port;
1928 	}
1929 
1930 	ret = nvmet_rdma_enable_port(port);
1931 	if (ret)
1932 		goto out_free_port;
1933 
1934 	pr_info("enabling port %d (%pISpcs)\n",
1935 		le16_to_cpu(nport->disc_addr.portid),
1936 		(struct sockaddr *)&port->addr);
1937 
1938 	return 0;
1939 
1940 out_free_port:
1941 	kfree(port);
1942 	return ret;
1943 }
1944 
1945 static void nvmet_rdma_remove_port(struct nvmet_port *nport)
1946 {
1947 	struct nvmet_rdma_port *port = nport->priv;
1948 
1949 	cancel_delayed_work_sync(&port->repair_work);
1950 	nvmet_rdma_disable_port(port);
1951 	kfree(port);
1952 }
1953 
1954 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1955 		struct nvmet_port *nport, char *traddr)
1956 {
1957 	struct nvmet_rdma_port *port = nport->priv;
1958 	struct rdma_cm_id *cm_id = port->cm_id;
1959 
1960 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1961 		struct nvmet_rdma_rsp *rsp =
1962 			container_of(req, struct nvmet_rdma_rsp, req);
1963 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1964 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1965 
1966 		sprintf(traddr, "%pISc", addr);
1967 	} else {
1968 		memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1969 	}
1970 }
1971 
1972 static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
1973 {
1974 	if (ctrl->pi_support)
1975 		return NVMET_RDMA_MAX_METADATA_MDTS;
1976 	return NVMET_RDMA_MAX_MDTS;
1977 }
1978 
1979 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1980 	.owner			= THIS_MODULE,
1981 	.type			= NVMF_TRTYPE_RDMA,
1982 	.msdbd			= 1,
1983 	.flags			= NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
1984 	.add_port		= nvmet_rdma_add_port,
1985 	.remove_port		= nvmet_rdma_remove_port,
1986 	.queue_response		= nvmet_rdma_queue_response,
1987 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
1988 	.disc_traddr		= nvmet_rdma_disc_port_addr,
1989 	.get_mdts		= nvmet_rdma_get_mdts,
1990 };
1991 
1992 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1993 {
1994 	struct nvmet_rdma_queue *queue, *tmp;
1995 	struct nvmet_rdma_device *ndev;
1996 	bool found = false;
1997 
1998 	mutex_lock(&device_list_mutex);
1999 	list_for_each_entry(ndev, &device_list, entry) {
2000 		if (ndev->device == ib_device) {
2001 			found = true;
2002 			break;
2003 		}
2004 	}
2005 	mutex_unlock(&device_list_mutex);
2006 
2007 	if (!found)
2008 		return;
2009 
2010 	/*
2011 	 * IB Device that is used by nvmet controllers is being removed,
2012 	 * delete all queues using this device.
2013 	 */
2014 	mutex_lock(&nvmet_rdma_queue_mutex);
2015 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
2016 				 queue_list) {
2017 		if (queue->dev->device != ib_device)
2018 			continue;
2019 
2020 		pr_info("Removing queue %d\n", queue->idx);
2021 		list_del_init(&queue->queue_list);
2022 		__nvmet_rdma_queue_disconnect(queue);
2023 	}
2024 	mutex_unlock(&nvmet_rdma_queue_mutex);
2025 
2026 	flush_scheduled_work();
2027 }
2028 
2029 static struct ib_client nvmet_rdma_ib_client = {
2030 	.name   = "nvmet_rdma",
2031 	.remove = nvmet_rdma_remove_one
2032 };
2033 
2034 static int __init nvmet_rdma_init(void)
2035 {
2036 	int ret;
2037 
2038 	ret = ib_register_client(&nvmet_rdma_ib_client);
2039 	if (ret)
2040 		return ret;
2041 
2042 	ret = nvmet_register_transport(&nvmet_rdma_ops);
2043 	if (ret)
2044 		goto err_ib_client;
2045 
2046 	return 0;
2047 
2048 err_ib_client:
2049 	ib_unregister_client(&nvmet_rdma_ib_client);
2050 	return ret;
2051 }
2052 
2053 static void __exit nvmet_rdma_exit(void)
2054 {
2055 	nvmet_unregister_transport(&nvmet_rdma_ops);
2056 	ib_unregister_client(&nvmet_rdma_ib_client);
2057 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
2058 	ida_destroy(&nvmet_rdma_queue_ida);
2059 }
2060 
2061 module_init(nvmet_rdma_init);
2062 module_exit(nvmet_rdma_exit);
2063 
2064 MODULE_LICENSE("GPL v2");
2065 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
2066