xref: /openbmc/linux/drivers/nvme/target/rdma.c (revision 2645d8d0)
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 
24 #include <linux/nvme-rdma.h>
25 #include "nvmet.h"
26 
27 /*
28  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
29  */
30 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE	PAGE_SIZE
31 #define NVMET_RDMA_MAX_INLINE_SGE		4
32 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE		max_t(int, SZ_16K, PAGE_SIZE)
33 
34 struct nvmet_rdma_cmd {
35 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
36 	struct ib_cqe		cqe;
37 	struct ib_recv_wr	wr;
38 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
39 	struct nvme_command     *nvme_cmd;
40 	struct nvmet_rdma_queue	*queue;
41 };
42 
43 enum {
44 	NVMET_RDMA_REQ_INLINE_DATA	= (1 << 0),
45 	NVMET_RDMA_REQ_INVALIDATE_RKEY	= (1 << 1),
46 };
47 
48 struct nvmet_rdma_rsp {
49 	struct ib_sge		send_sge;
50 	struct ib_cqe		send_cqe;
51 	struct ib_send_wr	send_wr;
52 
53 	struct nvmet_rdma_cmd	*cmd;
54 	struct nvmet_rdma_queue	*queue;
55 
56 	struct ib_cqe		read_cqe;
57 	struct rdma_rw_ctx	rw;
58 
59 	struct nvmet_req	req;
60 
61 	bool			allocated;
62 	u8			n_rdma;
63 	u32			flags;
64 	u32			invalidate_rkey;
65 
66 	struct list_head	wait_list;
67 	struct list_head	free_list;
68 };
69 
70 enum nvmet_rdma_queue_state {
71 	NVMET_RDMA_Q_CONNECTING,
72 	NVMET_RDMA_Q_LIVE,
73 	NVMET_RDMA_Q_DISCONNECTING,
74 };
75 
76 struct nvmet_rdma_queue {
77 	struct rdma_cm_id	*cm_id;
78 	struct nvmet_port	*port;
79 	struct ib_cq		*cq;
80 	atomic_t		sq_wr_avail;
81 	struct nvmet_rdma_device *dev;
82 	spinlock_t		state_lock;
83 	enum nvmet_rdma_queue_state state;
84 	struct nvmet_cq		nvme_cq;
85 	struct nvmet_sq		nvme_sq;
86 
87 	struct nvmet_rdma_rsp	*rsps;
88 	struct list_head	free_rsps;
89 	spinlock_t		rsps_lock;
90 	struct nvmet_rdma_cmd	*cmds;
91 
92 	struct work_struct	release_work;
93 	struct list_head	rsp_wait_list;
94 	struct list_head	rsp_wr_wait_list;
95 	spinlock_t		rsp_wr_wait_lock;
96 
97 	int			idx;
98 	int			host_qid;
99 	int			recv_queue_size;
100 	int			send_queue_size;
101 
102 	struct list_head	queue_list;
103 };
104 
105 struct nvmet_rdma_device {
106 	struct ib_device	*device;
107 	struct ib_pd		*pd;
108 	struct ib_srq		*srq;
109 	struct nvmet_rdma_cmd	*srq_cmds;
110 	size_t			srq_size;
111 	struct kref		ref;
112 	struct list_head	entry;
113 	int			inline_data_size;
114 	int			inline_page_count;
115 };
116 
117 static bool nvmet_rdma_use_srq;
118 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
119 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
120 
121 static DEFINE_IDA(nvmet_rdma_queue_ida);
122 static LIST_HEAD(nvmet_rdma_queue_list);
123 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
124 
125 static LIST_HEAD(device_list);
126 static DEFINE_MUTEX(device_list_mutex);
127 
128 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
129 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
130 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
131 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
132 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
133 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
134 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
135 				struct nvmet_rdma_rsp *r);
136 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
137 				struct nvmet_rdma_rsp *r);
138 
139 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
140 
141 static int num_pages(int len)
142 {
143 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
144 }
145 
146 /* XXX: really should move to a generic header sooner or later.. */
147 static inline u32 get_unaligned_le24(const u8 *p)
148 {
149 	return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
150 }
151 
152 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
153 {
154 	return nvme_is_write(rsp->req.cmd) &&
155 		rsp->req.transfer_len &&
156 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
157 }
158 
159 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
160 {
161 	return !nvme_is_write(rsp->req.cmd) &&
162 		rsp->req.transfer_len &&
163 		!rsp->req.rsp->status &&
164 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
165 }
166 
167 static inline struct nvmet_rdma_rsp *
168 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
169 {
170 	struct nvmet_rdma_rsp *rsp;
171 	unsigned long flags;
172 
173 	spin_lock_irqsave(&queue->rsps_lock, flags);
174 	rsp = list_first_entry_or_null(&queue->free_rsps,
175 				struct nvmet_rdma_rsp, free_list);
176 	if (likely(rsp))
177 		list_del(&rsp->free_list);
178 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
179 
180 	if (unlikely(!rsp)) {
181 		int ret;
182 
183 		rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
184 		if (unlikely(!rsp))
185 			return NULL;
186 		ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
187 		if (unlikely(ret)) {
188 			kfree(rsp);
189 			return NULL;
190 		}
191 
192 		rsp->allocated = true;
193 	}
194 
195 	return rsp;
196 }
197 
198 static inline void
199 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
200 {
201 	unsigned long flags;
202 
203 	if (unlikely(rsp->allocated)) {
204 		nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
205 		kfree(rsp);
206 		return;
207 	}
208 
209 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
210 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
211 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
212 }
213 
214 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
215 				struct nvmet_rdma_cmd *c)
216 {
217 	struct scatterlist *sg;
218 	struct ib_sge *sge;
219 	int i;
220 
221 	if (!ndev->inline_data_size)
222 		return;
223 
224 	sg = c->inline_sg;
225 	sge = &c->sge[1];
226 
227 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
228 		if (sge->length)
229 			ib_dma_unmap_page(ndev->device, sge->addr,
230 					sge->length, DMA_FROM_DEVICE);
231 		if (sg_page(sg))
232 			__free_page(sg_page(sg));
233 	}
234 }
235 
236 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
237 				struct nvmet_rdma_cmd *c)
238 {
239 	struct scatterlist *sg;
240 	struct ib_sge *sge;
241 	struct page *pg;
242 	int len;
243 	int i;
244 
245 	if (!ndev->inline_data_size)
246 		return 0;
247 
248 	sg = c->inline_sg;
249 	sg_init_table(sg, ndev->inline_page_count);
250 	sge = &c->sge[1];
251 	len = ndev->inline_data_size;
252 
253 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
254 		pg = alloc_page(GFP_KERNEL);
255 		if (!pg)
256 			goto out_err;
257 		sg_assign_page(sg, pg);
258 		sge->addr = ib_dma_map_page(ndev->device,
259 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
260 		if (ib_dma_mapping_error(ndev->device, sge->addr))
261 			goto out_err;
262 		sge->length = min_t(int, len, PAGE_SIZE);
263 		sge->lkey = ndev->pd->local_dma_lkey;
264 		len -= sge->length;
265 	}
266 
267 	return 0;
268 out_err:
269 	for (; i >= 0; 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 	return -ENOMEM;
277 }
278 
279 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
280 			struct nvmet_rdma_cmd *c, bool admin)
281 {
282 	/* NVMe command / RDMA RECV */
283 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
284 	if (!c->nvme_cmd)
285 		goto out;
286 
287 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
288 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
289 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
290 		goto out_free_cmd;
291 
292 	c->sge[0].length = sizeof(*c->nvme_cmd);
293 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
294 
295 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
296 		goto out_unmap_cmd;
297 
298 	c->cqe.done = nvmet_rdma_recv_done;
299 
300 	c->wr.wr_cqe = &c->cqe;
301 	c->wr.sg_list = c->sge;
302 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
303 
304 	return 0;
305 
306 out_unmap_cmd:
307 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
308 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
309 out_free_cmd:
310 	kfree(c->nvme_cmd);
311 
312 out:
313 	return -ENOMEM;
314 }
315 
316 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
317 		struct nvmet_rdma_cmd *c, bool admin)
318 {
319 	if (!admin)
320 		nvmet_rdma_free_inline_pages(ndev, c);
321 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
322 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
323 	kfree(c->nvme_cmd);
324 }
325 
326 static struct nvmet_rdma_cmd *
327 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
328 		int nr_cmds, bool admin)
329 {
330 	struct nvmet_rdma_cmd *cmds;
331 	int ret = -EINVAL, i;
332 
333 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
334 	if (!cmds)
335 		goto out;
336 
337 	for (i = 0; i < nr_cmds; i++) {
338 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
339 		if (ret)
340 			goto out_free;
341 	}
342 
343 	return cmds;
344 
345 out_free:
346 	while (--i >= 0)
347 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
348 	kfree(cmds);
349 out:
350 	return ERR_PTR(ret);
351 }
352 
353 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
354 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
355 {
356 	int i;
357 
358 	for (i = 0; i < nr_cmds; i++)
359 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
360 	kfree(cmds);
361 }
362 
363 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
364 		struct nvmet_rdma_rsp *r)
365 {
366 	/* NVMe CQE / RDMA SEND */
367 	r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
368 	if (!r->req.rsp)
369 		goto out;
370 
371 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
372 			sizeof(*r->req.rsp), DMA_TO_DEVICE);
373 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
374 		goto out_free_rsp;
375 
376 	r->send_sge.length = sizeof(*r->req.rsp);
377 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
378 
379 	r->send_cqe.done = nvmet_rdma_send_done;
380 
381 	r->send_wr.wr_cqe = &r->send_cqe;
382 	r->send_wr.sg_list = &r->send_sge;
383 	r->send_wr.num_sge = 1;
384 	r->send_wr.send_flags = IB_SEND_SIGNALED;
385 
386 	/* Data In / RDMA READ */
387 	r->read_cqe.done = nvmet_rdma_read_data_done;
388 	return 0;
389 
390 out_free_rsp:
391 	kfree(r->req.rsp);
392 out:
393 	return -ENOMEM;
394 }
395 
396 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
397 		struct nvmet_rdma_rsp *r)
398 {
399 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
400 				sizeof(*r->req.rsp), DMA_TO_DEVICE);
401 	kfree(r->req.rsp);
402 }
403 
404 static int
405 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
406 {
407 	struct nvmet_rdma_device *ndev = queue->dev;
408 	int nr_rsps = queue->recv_queue_size * 2;
409 	int ret = -EINVAL, i;
410 
411 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
412 			GFP_KERNEL);
413 	if (!queue->rsps)
414 		goto out;
415 
416 	for (i = 0; i < nr_rsps; i++) {
417 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
418 
419 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
420 		if (ret)
421 			goto out_free;
422 
423 		list_add_tail(&rsp->free_list, &queue->free_rsps);
424 	}
425 
426 	return 0;
427 
428 out_free:
429 	while (--i >= 0) {
430 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
431 
432 		list_del(&rsp->free_list);
433 		nvmet_rdma_free_rsp(ndev, rsp);
434 	}
435 	kfree(queue->rsps);
436 out:
437 	return ret;
438 }
439 
440 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
441 {
442 	struct nvmet_rdma_device *ndev = queue->dev;
443 	int i, nr_rsps = queue->recv_queue_size * 2;
444 
445 	for (i = 0; i < nr_rsps; i++) {
446 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
447 
448 		list_del(&rsp->free_list);
449 		nvmet_rdma_free_rsp(ndev, rsp);
450 	}
451 	kfree(queue->rsps);
452 }
453 
454 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
455 		struct nvmet_rdma_cmd *cmd)
456 {
457 	int ret;
458 
459 	ib_dma_sync_single_for_device(ndev->device,
460 		cmd->sge[0].addr, cmd->sge[0].length,
461 		DMA_FROM_DEVICE);
462 
463 	if (ndev->srq)
464 		ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
465 	else
466 		ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
467 
468 	if (unlikely(ret))
469 		pr_err("post_recv cmd failed\n");
470 
471 	return ret;
472 }
473 
474 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
475 {
476 	spin_lock(&queue->rsp_wr_wait_lock);
477 	while (!list_empty(&queue->rsp_wr_wait_list)) {
478 		struct nvmet_rdma_rsp *rsp;
479 		bool ret;
480 
481 		rsp = list_entry(queue->rsp_wr_wait_list.next,
482 				struct nvmet_rdma_rsp, wait_list);
483 		list_del(&rsp->wait_list);
484 
485 		spin_unlock(&queue->rsp_wr_wait_lock);
486 		ret = nvmet_rdma_execute_command(rsp);
487 		spin_lock(&queue->rsp_wr_wait_lock);
488 
489 		if (!ret) {
490 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
491 			break;
492 		}
493 	}
494 	spin_unlock(&queue->rsp_wr_wait_lock);
495 }
496 
497 
498 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
499 {
500 	struct nvmet_rdma_queue *queue = rsp->queue;
501 
502 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
503 
504 	if (rsp->n_rdma) {
505 		rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
506 				queue->cm_id->port_num, rsp->req.sg,
507 				rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
508 	}
509 
510 	if (rsp->req.sg != rsp->cmd->inline_sg)
511 		nvmet_req_free_sgl(&rsp->req);
512 
513 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
514 		nvmet_rdma_process_wr_wait_list(queue);
515 
516 	nvmet_rdma_put_rsp(rsp);
517 }
518 
519 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
520 {
521 	if (queue->nvme_sq.ctrl) {
522 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
523 	} else {
524 		/*
525 		 * we didn't setup the controller yet in case
526 		 * of admin connect error, just disconnect and
527 		 * cleanup the queue
528 		 */
529 		nvmet_rdma_queue_disconnect(queue);
530 	}
531 }
532 
533 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
534 {
535 	struct nvmet_rdma_rsp *rsp =
536 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
537 	struct nvmet_rdma_queue *queue = cq->cq_context;
538 
539 	nvmet_rdma_release_rsp(rsp);
540 
541 	if (unlikely(wc->status != IB_WC_SUCCESS &&
542 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
543 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
544 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
545 		nvmet_rdma_error_comp(queue);
546 	}
547 }
548 
549 static void nvmet_rdma_queue_response(struct nvmet_req *req)
550 {
551 	struct nvmet_rdma_rsp *rsp =
552 		container_of(req, struct nvmet_rdma_rsp, req);
553 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
554 	struct ib_send_wr *first_wr;
555 
556 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
557 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
558 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
559 	} else {
560 		rsp->send_wr.opcode = IB_WR_SEND;
561 	}
562 
563 	if (nvmet_rdma_need_data_out(rsp))
564 		first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
565 				cm_id->port_num, NULL, &rsp->send_wr);
566 	else
567 		first_wr = &rsp->send_wr;
568 
569 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
570 
571 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
572 		rsp->send_sge.addr, rsp->send_sge.length,
573 		DMA_TO_DEVICE);
574 
575 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
576 		pr_err("sending cmd response failed\n");
577 		nvmet_rdma_release_rsp(rsp);
578 	}
579 }
580 
581 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
582 {
583 	struct nvmet_rdma_rsp *rsp =
584 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
585 	struct nvmet_rdma_queue *queue = cq->cq_context;
586 
587 	WARN_ON(rsp->n_rdma <= 0);
588 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
589 	rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
590 			queue->cm_id->port_num, rsp->req.sg,
591 			rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
592 	rsp->n_rdma = 0;
593 
594 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
595 		nvmet_req_uninit(&rsp->req);
596 		nvmet_rdma_release_rsp(rsp);
597 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
598 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
599 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
600 			nvmet_rdma_error_comp(queue);
601 		}
602 		return;
603 	}
604 
605 	nvmet_req_execute(&rsp->req);
606 }
607 
608 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
609 		u64 off)
610 {
611 	int sg_count = num_pages(len);
612 	struct scatterlist *sg;
613 	int i;
614 
615 	sg = rsp->cmd->inline_sg;
616 	for (i = 0; i < sg_count; i++, sg++) {
617 		if (i < sg_count - 1)
618 			sg_unmark_end(sg);
619 		else
620 			sg_mark_end(sg);
621 		sg->offset = off;
622 		sg->length = min_t(int, len, PAGE_SIZE - off);
623 		len -= sg->length;
624 		if (!i)
625 			off = 0;
626 	}
627 
628 	rsp->req.sg = rsp->cmd->inline_sg;
629 	rsp->req.sg_cnt = sg_count;
630 }
631 
632 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
633 {
634 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
635 	u64 off = le64_to_cpu(sgl->addr);
636 	u32 len = le32_to_cpu(sgl->length);
637 
638 	if (!nvme_is_write(rsp->req.cmd)) {
639 		rsp->req.error_loc =
640 			offsetof(struct nvme_common_command, opcode);
641 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
642 	}
643 
644 	if (off + len > rsp->queue->dev->inline_data_size) {
645 		pr_err("invalid inline data offset!\n");
646 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
647 	}
648 
649 	/* no data command? */
650 	if (!len)
651 		return 0;
652 
653 	nvmet_rdma_use_inline_sg(rsp, len, off);
654 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
655 	rsp->req.transfer_len += len;
656 	return 0;
657 }
658 
659 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
660 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
661 {
662 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
663 	u64 addr = le64_to_cpu(sgl->addr);
664 	u32 key = get_unaligned_le32(sgl->key);
665 	int ret;
666 
667 	rsp->req.transfer_len = get_unaligned_le24(sgl->length);
668 
669 	/* no data command? */
670 	if (!rsp->req.transfer_len)
671 		return 0;
672 
673 	ret = nvmet_req_alloc_sgl(&rsp->req);
674 	if (ret < 0)
675 		goto error_out;
676 
677 	ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
678 			rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
679 			nvmet_data_dir(&rsp->req));
680 	if (ret < 0)
681 		goto error_out;
682 	rsp->n_rdma += ret;
683 
684 	if (invalidate) {
685 		rsp->invalidate_rkey = key;
686 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
687 	}
688 
689 	return 0;
690 
691 error_out:
692 	rsp->req.transfer_len = 0;
693 	return NVME_SC_INTERNAL;
694 }
695 
696 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
697 {
698 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
699 
700 	switch (sgl->type >> 4) {
701 	case NVME_SGL_FMT_DATA_DESC:
702 		switch (sgl->type & 0xf) {
703 		case NVME_SGL_FMT_OFFSET:
704 			return nvmet_rdma_map_sgl_inline(rsp);
705 		default:
706 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
707 			rsp->req.error_loc =
708 				offsetof(struct nvme_common_command, dptr);
709 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
710 		}
711 	case NVME_KEY_SGL_FMT_DATA_DESC:
712 		switch (sgl->type & 0xf) {
713 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
714 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
715 		case NVME_SGL_FMT_ADDRESS:
716 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
717 		default:
718 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
719 			rsp->req.error_loc =
720 				offsetof(struct nvme_common_command, dptr);
721 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
722 		}
723 	default:
724 		pr_err("invalid SGL type: %#x\n", sgl->type);
725 		rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
726 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
727 	}
728 }
729 
730 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
731 {
732 	struct nvmet_rdma_queue *queue = rsp->queue;
733 
734 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
735 			&queue->sq_wr_avail) < 0)) {
736 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
737 				1 + rsp->n_rdma, queue->idx,
738 				queue->nvme_sq.ctrl->cntlid);
739 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
740 		return false;
741 	}
742 
743 	if (nvmet_rdma_need_data_in(rsp)) {
744 		if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
745 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
746 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
747 	} else {
748 		nvmet_req_execute(&rsp->req);
749 	}
750 
751 	return true;
752 }
753 
754 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
755 		struct nvmet_rdma_rsp *cmd)
756 {
757 	u16 status;
758 
759 	ib_dma_sync_single_for_cpu(queue->dev->device,
760 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
761 		DMA_FROM_DEVICE);
762 	ib_dma_sync_single_for_cpu(queue->dev->device,
763 		cmd->send_sge.addr, cmd->send_sge.length,
764 		DMA_TO_DEVICE);
765 
766 	cmd->req.p2p_client = &queue->dev->device->dev;
767 
768 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
769 			&queue->nvme_sq, &nvmet_rdma_ops))
770 		return;
771 
772 	status = nvmet_rdma_map_sgl(cmd);
773 	if (status)
774 		goto out_err;
775 
776 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
777 		spin_lock(&queue->rsp_wr_wait_lock);
778 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
779 		spin_unlock(&queue->rsp_wr_wait_lock);
780 	}
781 
782 	return;
783 
784 out_err:
785 	nvmet_req_complete(&cmd->req, status);
786 }
787 
788 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
789 {
790 	struct nvmet_rdma_cmd *cmd =
791 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
792 	struct nvmet_rdma_queue *queue = cq->cq_context;
793 	struct nvmet_rdma_rsp *rsp;
794 
795 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
796 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
797 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
798 				wc->wr_cqe, ib_wc_status_msg(wc->status),
799 				wc->status);
800 			nvmet_rdma_error_comp(queue);
801 		}
802 		return;
803 	}
804 
805 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
806 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
807 		nvmet_rdma_error_comp(queue);
808 		return;
809 	}
810 
811 	cmd->queue = queue;
812 	rsp = nvmet_rdma_get_rsp(queue);
813 	if (unlikely(!rsp)) {
814 		/*
815 		 * we get here only under memory pressure,
816 		 * silently drop and have the host retry
817 		 * as we can't even fail it.
818 		 */
819 		nvmet_rdma_post_recv(queue->dev, cmd);
820 		return;
821 	}
822 	rsp->queue = queue;
823 	rsp->cmd = cmd;
824 	rsp->flags = 0;
825 	rsp->req.cmd = cmd->nvme_cmd;
826 	rsp->req.port = queue->port;
827 	rsp->n_rdma = 0;
828 
829 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
830 		unsigned long flags;
831 
832 		spin_lock_irqsave(&queue->state_lock, flags);
833 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
834 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
835 		else
836 			nvmet_rdma_put_rsp(rsp);
837 		spin_unlock_irqrestore(&queue->state_lock, flags);
838 		return;
839 	}
840 
841 	nvmet_rdma_handle_command(queue, rsp);
842 }
843 
844 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
845 {
846 	if (!ndev->srq)
847 		return;
848 
849 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
850 	ib_destroy_srq(ndev->srq);
851 }
852 
853 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
854 {
855 	struct ib_srq_init_attr srq_attr = { NULL, };
856 	struct ib_srq *srq;
857 	size_t srq_size;
858 	int ret, i;
859 
860 	srq_size = 4095;	/* XXX: tune */
861 
862 	srq_attr.attr.max_wr = srq_size;
863 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
864 	srq_attr.attr.srq_limit = 0;
865 	srq_attr.srq_type = IB_SRQT_BASIC;
866 	srq = ib_create_srq(ndev->pd, &srq_attr);
867 	if (IS_ERR(srq)) {
868 		/*
869 		 * If SRQs aren't supported we just go ahead and use normal
870 		 * non-shared receive queues.
871 		 */
872 		pr_info("SRQ requested but not supported.\n");
873 		return 0;
874 	}
875 
876 	ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
877 	if (IS_ERR(ndev->srq_cmds)) {
878 		ret = PTR_ERR(ndev->srq_cmds);
879 		goto out_destroy_srq;
880 	}
881 
882 	ndev->srq = srq;
883 	ndev->srq_size = srq_size;
884 
885 	for (i = 0; i < srq_size; i++) {
886 		ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
887 		if (ret)
888 			goto out_free_cmds;
889 	}
890 
891 	return 0;
892 
893 out_free_cmds:
894 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
895 out_destroy_srq:
896 	ib_destroy_srq(srq);
897 	return ret;
898 }
899 
900 static void nvmet_rdma_free_dev(struct kref *ref)
901 {
902 	struct nvmet_rdma_device *ndev =
903 		container_of(ref, struct nvmet_rdma_device, ref);
904 
905 	mutex_lock(&device_list_mutex);
906 	list_del(&ndev->entry);
907 	mutex_unlock(&device_list_mutex);
908 
909 	nvmet_rdma_destroy_srq(ndev);
910 	ib_dealloc_pd(ndev->pd);
911 
912 	kfree(ndev);
913 }
914 
915 static struct nvmet_rdma_device *
916 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
917 {
918 	struct nvmet_port *port = cm_id->context;
919 	struct nvmet_rdma_device *ndev;
920 	int inline_page_count;
921 	int inline_sge_count;
922 	int ret;
923 
924 	mutex_lock(&device_list_mutex);
925 	list_for_each_entry(ndev, &device_list, entry) {
926 		if (ndev->device->node_guid == cm_id->device->node_guid &&
927 		    kref_get_unless_zero(&ndev->ref))
928 			goto out_unlock;
929 	}
930 
931 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
932 	if (!ndev)
933 		goto out_err;
934 
935 	inline_page_count = num_pages(port->inline_data_size);
936 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
937 				cm_id->device->attrs.max_recv_sge) - 1;
938 	if (inline_page_count > inline_sge_count) {
939 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
940 			port->inline_data_size, cm_id->device->name,
941 			inline_sge_count * PAGE_SIZE);
942 		port->inline_data_size = inline_sge_count * PAGE_SIZE;
943 		inline_page_count = inline_sge_count;
944 	}
945 	ndev->inline_data_size = port->inline_data_size;
946 	ndev->inline_page_count = inline_page_count;
947 	ndev->device = cm_id->device;
948 	kref_init(&ndev->ref);
949 
950 	ndev->pd = ib_alloc_pd(ndev->device, 0);
951 	if (IS_ERR(ndev->pd))
952 		goto out_free_dev;
953 
954 	if (nvmet_rdma_use_srq) {
955 		ret = nvmet_rdma_init_srq(ndev);
956 		if (ret)
957 			goto out_free_pd;
958 	}
959 
960 	list_add(&ndev->entry, &device_list);
961 out_unlock:
962 	mutex_unlock(&device_list_mutex);
963 	pr_debug("added %s.\n", ndev->device->name);
964 	return ndev;
965 
966 out_free_pd:
967 	ib_dealloc_pd(ndev->pd);
968 out_free_dev:
969 	kfree(ndev);
970 out_err:
971 	mutex_unlock(&device_list_mutex);
972 	return NULL;
973 }
974 
975 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
976 {
977 	struct ib_qp_init_attr qp_attr;
978 	struct nvmet_rdma_device *ndev = queue->dev;
979 	int comp_vector, nr_cqe, ret, i;
980 
981 	/*
982 	 * Spread the io queues across completion vectors,
983 	 * but still keep all admin queues on vector 0.
984 	 */
985 	comp_vector = !queue->host_qid ? 0 :
986 		queue->idx % ndev->device->num_comp_vectors;
987 
988 	/*
989 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
990 	 */
991 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
992 
993 	queue->cq = ib_alloc_cq(ndev->device, queue,
994 			nr_cqe + 1, comp_vector,
995 			IB_POLL_WORKQUEUE);
996 	if (IS_ERR(queue->cq)) {
997 		ret = PTR_ERR(queue->cq);
998 		pr_err("failed to create CQ cqe= %d ret= %d\n",
999 		       nr_cqe + 1, ret);
1000 		goto out;
1001 	}
1002 
1003 	memset(&qp_attr, 0, sizeof(qp_attr));
1004 	qp_attr.qp_context = queue;
1005 	qp_attr.event_handler = nvmet_rdma_qp_event;
1006 	qp_attr.send_cq = queue->cq;
1007 	qp_attr.recv_cq = queue->cq;
1008 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1009 	qp_attr.qp_type = IB_QPT_RC;
1010 	/* +1 for drain */
1011 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1012 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1013 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1014 					ndev->device->attrs.max_send_sge);
1015 
1016 	if (ndev->srq) {
1017 		qp_attr.srq = ndev->srq;
1018 	} else {
1019 		/* +1 for drain */
1020 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1021 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1022 	}
1023 
1024 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1025 	if (ret) {
1026 		pr_err("failed to create_qp ret= %d\n", ret);
1027 		goto err_destroy_cq;
1028 	}
1029 
1030 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1031 
1032 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1033 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1034 		 qp_attr.cap.max_send_wr, queue->cm_id);
1035 
1036 	if (!ndev->srq) {
1037 		for (i = 0; i < queue->recv_queue_size; i++) {
1038 			queue->cmds[i].queue = queue;
1039 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1040 			if (ret)
1041 				goto err_destroy_qp;
1042 		}
1043 	}
1044 
1045 out:
1046 	return ret;
1047 
1048 err_destroy_qp:
1049 	rdma_destroy_qp(queue->cm_id);
1050 err_destroy_cq:
1051 	ib_free_cq(queue->cq);
1052 	goto out;
1053 }
1054 
1055 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1056 {
1057 	struct ib_qp *qp = queue->cm_id->qp;
1058 
1059 	ib_drain_qp(qp);
1060 	rdma_destroy_id(queue->cm_id);
1061 	ib_destroy_qp(qp);
1062 	ib_free_cq(queue->cq);
1063 }
1064 
1065 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1066 {
1067 	pr_debug("freeing queue %d\n", queue->idx);
1068 
1069 	nvmet_sq_destroy(&queue->nvme_sq);
1070 
1071 	nvmet_rdma_destroy_queue_ib(queue);
1072 	if (!queue->dev->srq) {
1073 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1074 				queue->recv_queue_size,
1075 				!queue->host_qid);
1076 	}
1077 	nvmet_rdma_free_rsps(queue);
1078 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1079 	kfree(queue);
1080 }
1081 
1082 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1083 {
1084 	struct nvmet_rdma_queue *queue =
1085 		container_of(w, struct nvmet_rdma_queue, release_work);
1086 	struct nvmet_rdma_device *dev = queue->dev;
1087 
1088 	nvmet_rdma_free_queue(queue);
1089 
1090 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1091 }
1092 
1093 static int
1094 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1095 				struct nvmet_rdma_queue *queue)
1096 {
1097 	struct nvme_rdma_cm_req *req;
1098 
1099 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1100 	if (!req || conn->private_data_len == 0)
1101 		return NVME_RDMA_CM_INVALID_LEN;
1102 
1103 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1104 		return NVME_RDMA_CM_INVALID_RECFMT;
1105 
1106 	queue->host_qid = le16_to_cpu(req->qid);
1107 
1108 	/*
1109 	 * req->hsqsize corresponds to our recv queue size plus 1
1110 	 * req->hrqsize corresponds to our send queue size
1111 	 */
1112 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1113 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1114 
1115 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1116 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1117 
1118 	/* XXX: Should we enforce some kind of max for IO queues? */
1119 
1120 	return 0;
1121 }
1122 
1123 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1124 				enum nvme_rdma_cm_status status)
1125 {
1126 	struct nvme_rdma_cm_rej rej;
1127 
1128 	pr_debug("rejecting connect request: status %d (%s)\n",
1129 		 status, nvme_rdma_cm_msg(status));
1130 
1131 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1132 	rej.sts = cpu_to_le16(status);
1133 
1134 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1135 }
1136 
1137 static struct nvmet_rdma_queue *
1138 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1139 		struct rdma_cm_id *cm_id,
1140 		struct rdma_cm_event *event)
1141 {
1142 	struct nvmet_rdma_queue *queue;
1143 	int ret;
1144 
1145 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1146 	if (!queue) {
1147 		ret = NVME_RDMA_CM_NO_RSC;
1148 		goto out_reject;
1149 	}
1150 
1151 	ret = nvmet_sq_init(&queue->nvme_sq);
1152 	if (ret) {
1153 		ret = NVME_RDMA_CM_NO_RSC;
1154 		goto out_free_queue;
1155 	}
1156 
1157 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1158 	if (ret)
1159 		goto out_destroy_sq;
1160 
1161 	/*
1162 	 * Schedules the actual release because calling rdma_destroy_id from
1163 	 * inside a CM callback would trigger a deadlock. (great API design..)
1164 	 */
1165 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1166 	queue->dev = ndev;
1167 	queue->cm_id = cm_id;
1168 
1169 	spin_lock_init(&queue->state_lock);
1170 	queue->state = NVMET_RDMA_Q_CONNECTING;
1171 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1172 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1173 	spin_lock_init(&queue->rsp_wr_wait_lock);
1174 	INIT_LIST_HEAD(&queue->free_rsps);
1175 	spin_lock_init(&queue->rsps_lock);
1176 	INIT_LIST_HEAD(&queue->queue_list);
1177 
1178 	queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1179 	if (queue->idx < 0) {
1180 		ret = NVME_RDMA_CM_NO_RSC;
1181 		goto out_destroy_sq;
1182 	}
1183 
1184 	ret = nvmet_rdma_alloc_rsps(queue);
1185 	if (ret) {
1186 		ret = NVME_RDMA_CM_NO_RSC;
1187 		goto out_ida_remove;
1188 	}
1189 
1190 	if (!ndev->srq) {
1191 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1192 				queue->recv_queue_size,
1193 				!queue->host_qid);
1194 		if (IS_ERR(queue->cmds)) {
1195 			ret = NVME_RDMA_CM_NO_RSC;
1196 			goto out_free_responses;
1197 		}
1198 	}
1199 
1200 	ret = nvmet_rdma_create_queue_ib(queue);
1201 	if (ret) {
1202 		pr_err("%s: creating RDMA queue failed (%d).\n",
1203 			__func__, ret);
1204 		ret = NVME_RDMA_CM_NO_RSC;
1205 		goto out_free_cmds;
1206 	}
1207 
1208 	return queue;
1209 
1210 out_free_cmds:
1211 	if (!ndev->srq) {
1212 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1213 				queue->recv_queue_size,
1214 				!queue->host_qid);
1215 	}
1216 out_free_responses:
1217 	nvmet_rdma_free_rsps(queue);
1218 out_ida_remove:
1219 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1220 out_destroy_sq:
1221 	nvmet_sq_destroy(&queue->nvme_sq);
1222 out_free_queue:
1223 	kfree(queue);
1224 out_reject:
1225 	nvmet_rdma_cm_reject(cm_id, ret);
1226 	return NULL;
1227 }
1228 
1229 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1230 {
1231 	struct nvmet_rdma_queue *queue = priv;
1232 
1233 	switch (event->event) {
1234 	case IB_EVENT_COMM_EST:
1235 		rdma_notify(queue->cm_id, event->event);
1236 		break;
1237 	default:
1238 		pr_err("received IB QP event: %s (%d)\n",
1239 		       ib_event_msg(event->event), event->event);
1240 		break;
1241 	}
1242 }
1243 
1244 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1245 		struct nvmet_rdma_queue *queue,
1246 		struct rdma_conn_param *p)
1247 {
1248 	struct rdma_conn_param  param = { };
1249 	struct nvme_rdma_cm_rep priv = { };
1250 	int ret = -ENOMEM;
1251 
1252 	param.rnr_retry_count = 7;
1253 	param.flow_control = 1;
1254 	param.initiator_depth = min_t(u8, p->initiator_depth,
1255 		queue->dev->device->attrs.max_qp_init_rd_atom);
1256 	param.private_data = &priv;
1257 	param.private_data_len = sizeof(priv);
1258 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1259 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1260 
1261 	ret = rdma_accept(cm_id, &param);
1262 	if (ret)
1263 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1264 
1265 	return ret;
1266 }
1267 
1268 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1269 		struct rdma_cm_event *event)
1270 {
1271 	struct nvmet_rdma_device *ndev;
1272 	struct nvmet_rdma_queue *queue;
1273 	int ret = -EINVAL;
1274 
1275 	ndev = nvmet_rdma_find_get_device(cm_id);
1276 	if (!ndev) {
1277 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1278 		return -ECONNREFUSED;
1279 	}
1280 
1281 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1282 	if (!queue) {
1283 		ret = -ENOMEM;
1284 		goto put_device;
1285 	}
1286 	queue->port = cm_id->context;
1287 
1288 	if (queue->host_qid == 0) {
1289 		/* Let inflight controller teardown complete */
1290 		flush_scheduled_work();
1291 	}
1292 
1293 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1294 	if (ret) {
1295 		schedule_work(&queue->release_work);
1296 		/* Destroying rdma_cm id is not needed here */
1297 		return 0;
1298 	}
1299 
1300 	mutex_lock(&nvmet_rdma_queue_mutex);
1301 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1302 	mutex_unlock(&nvmet_rdma_queue_mutex);
1303 
1304 	return 0;
1305 
1306 put_device:
1307 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1308 
1309 	return ret;
1310 }
1311 
1312 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1313 {
1314 	unsigned long flags;
1315 
1316 	spin_lock_irqsave(&queue->state_lock, flags);
1317 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1318 		pr_warn("trying to establish a connected queue\n");
1319 		goto out_unlock;
1320 	}
1321 	queue->state = NVMET_RDMA_Q_LIVE;
1322 
1323 	while (!list_empty(&queue->rsp_wait_list)) {
1324 		struct nvmet_rdma_rsp *cmd;
1325 
1326 		cmd = list_first_entry(&queue->rsp_wait_list,
1327 					struct nvmet_rdma_rsp, wait_list);
1328 		list_del(&cmd->wait_list);
1329 
1330 		spin_unlock_irqrestore(&queue->state_lock, flags);
1331 		nvmet_rdma_handle_command(queue, cmd);
1332 		spin_lock_irqsave(&queue->state_lock, flags);
1333 	}
1334 
1335 out_unlock:
1336 	spin_unlock_irqrestore(&queue->state_lock, flags);
1337 }
1338 
1339 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1340 {
1341 	bool disconnect = false;
1342 	unsigned long flags;
1343 
1344 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1345 
1346 	spin_lock_irqsave(&queue->state_lock, flags);
1347 	switch (queue->state) {
1348 	case NVMET_RDMA_Q_CONNECTING:
1349 	case NVMET_RDMA_Q_LIVE:
1350 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1351 		disconnect = true;
1352 		break;
1353 	case NVMET_RDMA_Q_DISCONNECTING:
1354 		break;
1355 	}
1356 	spin_unlock_irqrestore(&queue->state_lock, flags);
1357 
1358 	if (disconnect) {
1359 		rdma_disconnect(queue->cm_id);
1360 		schedule_work(&queue->release_work);
1361 	}
1362 }
1363 
1364 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1365 {
1366 	bool disconnect = false;
1367 
1368 	mutex_lock(&nvmet_rdma_queue_mutex);
1369 	if (!list_empty(&queue->queue_list)) {
1370 		list_del_init(&queue->queue_list);
1371 		disconnect = true;
1372 	}
1373 	mutex_unlock(&nvmet_rdma_queue_mutex);
1374 
1375 	if (disconnect)
1376 		__nvmet_rdma_queue_disconnect(queue);
1377 }
1378 
1379 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1380 		struct nvmet_rdma_queue *queue)
1381 {
1382 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1383 
1384 	mutex_lock(&nvmet_rdma_queue_mutex);
1385 	if (!list_empty(&queue->queue_list))
1386 		list_del_init(&queue->queue_list);
1387 	mutex_unlock(&nvmet_rdma_queue_mutex);
1388 
1389 	pr_err("failed to connect queue %d\n", queue->idx);
1390 	schedule_work(&queue->release_work);
1391 }
1392 
1393 /**
1394  * nvme_rdma_device_removal() - Handle RDMA device removal
1395  * @cm_id:	rdma_cm id, used for nvmet port
1396  * @queue:      nvmet rdma queue (cm id qp_context)
1397  *
1398  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1399  * to unplug. Note that this event can be generated on a normal
1400  * queue cm_id and/or a device bound listener cm_id (where in this
1401  * case queue will be null).
1402  *
1403  * We registered an ib_client to handle device removal for queues,
1404  * so we only need to handle the listening port cm_ids. In this case
1405  * we nullify the priv to prevent double cm_id destruction and destroying
1406  * the cm_id implicitely by returning a non-zero rc to the callout.
1407  */
1408 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1409 		struct nvmet_rdma_queue *queue)
1410 {
1411 	struct nvmet_port *port;
1412 
1413 	if (queue) {
1414 		/*
1415 		 * This is a queue cm_id. we have registered
1416 		 * an ib_client to handle queues removal
1417 		 * so don't interfear and just return.
1418 		 */
1419 		return 0;
1420 	}
1421 
1422 	port = cm_id->context;
1423 
1424 	/*
1425 	 * This is a listener cm_id. Make sure that
1426 	 * future remove_port won't invoke a double
1427 	 * cm_id destroy. use atomic xchg to make sure
1428 	 * we don't compete with remove_port.
1429 	 */
1430 	if (xchg(&port->priv, NULL) != cm_id)
1431 		return 0;
1432 
1433 	/*
1434 	 * We need to return 1 so that the core will destroy
1435 	 * it's own ID.  What a great API design..
1436 	 */
1437 	return 1;
1438 }
1439 
1440 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1441 		struct rdma_cm_event *event)
1442 {
1443 	struct nvmet_rdma_queue *queue = NULL;
1444 	int ret = 0;
1445 
1446 	if (cm_id->qp)
1447 		queue = cm_id->qp->qp_context;
1448 
1449 	pr_debug("%s (%d): status %d id %p\n",
1450 		rdma_event_msg(event->event), event->event,
1451 		event->status, cm_id);
1452 
1453 	switch (event->event) {
1454 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1455 		ret = nvmet_rdma_queue_connect(cm_id, event);
1456 		break;
1457 	case RDMA_CM_EVENT_ESTABLISHED:
1458 		nvmet_rdma_queue_established(queue);
1459 		break;
1460 	case RDMA_CM_EVENT_ADDR_CHANGE:
1461 	case RDMA_CM_EVENT_DISCONNECTED:
1462 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1463 		nvmet_rdma_queue_disconnect(queue);
1464 		break;
1465 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1466 		ret = nvmet_rdma_device_removal(cm_id, queue);
1467 		break;
1468 	case RDMA_CM_EVENT_REJECTED:
1469 		pr_debug("Connection rejected: %s\n",
1470 			 rdma_reject_msg(cm_id, event->status));
1471 		/* FALLTHROUGH */
1472 	case RDMA_CM_EVENT_UNREACHABLE:
1473 	case RDMA_CM_EVENT_CONNECT_ERROR:
1474 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1475 		break;
1476 	default:
1477 		pr_err("received unrecognized RDMA CM event %d\n",
1478 			event->event);
1479 		break;
1480 	}
1481 
1482 	return ret;
1483 }
1484 
1485 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1486 {
1487 	struct nvmet_rdma_queue *queue;
1488 
1489 restart:
1490 	mutex_lock(&nvmet_rdma_queue_mutex);
1491 	list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1492 		if (queue->nvme_sq.ctrl == ctrl) {
1493 			list_del_init(&queue->queue_list);
1494 			mutex_unlock(&nvmet_rdma_queue_mutex);
1495 
1496 			__nvmet_rdma_queue_disconnect(queue);
1497 			goto restart;
1498 		}
1499 	}
1500 	mutex_unlock(&nvmet_rdma_queue_mutex);
1501 }
1502 
1503 static int nvmet_rdma_add_port(struct nvmet_port *port)
1504 {
1505 	struct rdma_cm_id *cm_id;
1506 	struct sockaddr_storage addr = { };
1507 	__kernel_sa_family_t af;
1508 	int ret;
1509 
1510 	switch (port->disc_addr.adrfam) {
1511 	case NVMF_ADDR_FAMILY_IP4:
1512 		af = AF_INET;
1513 		break;
1514 	case NVMF_ADDR_FAMILY_IP6:
1515 		af = AF_INET6;
1516 		break;
1517 	default:
1518 		pr_err("address family %d not supported\n",
1519 				port->disc_addr.adrfam);
1520 		return -EINVAL;
1521 	}
1522 
1523 	if (port->inline_data_size < 0) {
1524 		port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1525 	} else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1526 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1527 			port->inline_data_size,
1528 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1529 		port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1530 	}
1531 
1532 	ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1533 			port->disc_addr.trsvcid, &addr);
1534 	if (ret) {
1535 		pr_err("malformed ip/port passed: %s:%s\n",
1536 			port->disc_addr.traddr, port->disc_addr.trsvcid);
1537 		return ret;
1538 	}
1539 
1540 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1541 			RDMA_PS_TCP, IB_QPT_RC);
1542 	if (IS_ERR(cm_id)) {
1543 		pr_err("CM ID creation failed\n");
1544 		return PTR_ERR(cm_id);
1545 	}
1546 
1547 	/*
1548 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1549 	 * at the same time.
1550 	 */
1551 	ret = rdma_set_afonly(cm_id, 1);
1552 	if (ret) {
1553 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1554 		goto out_destroy_id;
1555 	}
1556 
1557 	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1558 	if (ret) {
1559 		pr_err("binding CM ID to %pISpcs failed (%d)\n",
1560 			(struct sockaddr *)&addr, ret);
1561 		goto out_destroy_id;
1562 	}
1563 
1564 	ret = rdma_listen(cm_id, 128);
1565 	if (ret) {
1566 		pr_err("listening to %pISpcs failed (%d)\n",
1567 			(struct sockaddr *)&addr, ret);
1568 		goto out_destroy_id;
1569 	}
1570 
1571 	pr_info("enabling port %d (%pISpcs)\n",
1572 		le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1573 	port->priv = cm_id;
1574 	return 0;
1575 
1576 out_destroy_id:
1577 	rdma_destroy_id(cm_id);
1578 	return ret;
1579 }
1580 
1581 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1582 {
1583 	struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1584 
1585 	if (cm_id)
1586 		rdma_destroy_id(cm_id);
1587 }
1588 
1589 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1590 		struct nvmet_port *port, char *traddr)
1591 {
1592 	struct rdma_cm_id *cm_id = port->priv;
1593 
1594 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1595 		struct nvmet_rdma_rsp *rsp =
1596 			container_of(req, struct nvmet_rdma_rsp, req);
1597 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1598 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1599 
1600 		sprintf(traddr, "%pISc", addr);
1601 	} else {
1602 		memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1603 	}
1604 }
1605 
1606 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1607 	.owner			= THIS_MODULE,
1608 	.type			= NVMF_TRTYPE_RDMA,
1609 	.msdbd			= 1,
1610 	.has_keyed_sgls		= 1,
1611 	.add_port		= nvmet_rdma_add_port,
1612 	.remove_port		= nvmet_rdma_remove_port,
1613 	.queue_response		= nvmet_rdma_queue_response,
1614 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
1615 	.disc_traddr		= nvmet_rdma_disc_port_addr,
1616 };
1617 
1618 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1619 {
1620 	struct nvmet_rdma_queue *queue, *tmp;
1621 	struct nvmet_rdma_device *ndev;
1622 	bool found = false;
1623 
1624 	mutex_lock(&device_list_mutex);
1625 	list_for_each_entry(ndev, &device_list, entry) {
1626 		if (ndev->device == ib_device) {
1627 			found = true;
1628 			break;
1629 		}
1630 	}
1631 	mutex_unlock(&device_list_mutex);
1632 
1633 	if (!found)
1634 		return;
1635 
1636 	/*
1637 	 * IB Device that is used by nvmet controllers is being removed,
1638 	 * delete all queues using this device.
1639 	 */
1640 	mutex_lock(&nvmet_rdma_queue_mutex);
1641 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1642 				 queue_list) {
1643 		if (queue->dev->device != ib_device)
1644 			continue;
1645 
1646 		pr_info("Removing queue %d\n", queue->idx);
1647 		list_del_init(&queue->queue_list);
1648 		__nvmet_rdma_queue_disconnect(queue);
1649 	}
1650 	mutex_unlock(&nvmet_rdma_queue_mutex);
1651 
1652 	flush_scheduled_work();
1653 }
1654 
1655 static struct ib_client nvmet_rdma_ib_client = {
1656 	.name   = "nvmet_rdma",
1657 	.remove = nvmet_rdma_remove_one
1658 };
1659 
1660 static int __init nvmet_rdma_init(void)
1661 {
1662 	int ret;
1663 
1664 	ret = ib_register_client(&nvmet_rdma_ib_client);
1665 	if (ret)
1666 		return ret;
1667 
1668 	ret = nvmet_register_transport(&nvmet_rdma_ops);
1669 	if (ret)
1670 		goto err_ib_client;
1671 
1672 	return 0;
1673 
1674 err_ib_client:
1675 	ib_unregister_client(&nvmet_rdma_ib_client);
1676 	return ret;
1677 }
1678 
1679 static void __exit nvmet_rdma_exit(void)
1680 {
1681 	nvmet_unregister_transport(&nvmet_rdma_ops);
1682 	ib_unregister_client(&nvmet_rdma_ib_client);
1683 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1684 	ida_destroy(&nvmet_rdma_queue_ida);
1685 }
1686 
1687 module_init(nvmet_rdma_init);
1688 module_exit(nvmet_rdma_exit);
1689 
1690 MODULE_LICENSE("GPL v2");
1691 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
1692