xref: /linux/drivers/nvme/target/rdma.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 /*
2  * NVMe over Fabrics RDMA target.
3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  */
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/atomic.h>
16 #include <linux/ctype.h>
17 #include <linux/delay.h>
18 #include <linux/err.h>
19 #include <linux/init.h>
20 #include <linux/module.h>
21 #include <linux/nvme.h>
22 #include <linux/slab.h>
23 #include <linux/string.h>
24 #include <linux/wait.h>
25 #include <linux/inet.h>
26 #include <asm/unaligned.h>
27 
28 #include <rdma/ib_verbs.h>
29 #include <rdma/rdma_cm.h>
30 #include <rdma/rw.h>
31 
32 #include <linux/nvme-rdma.h>
33 #include "nvmet.h"
34 
35 /*
36  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
37  */
38 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE	PAGE_SIZE
39 #define NVMET_RDMA_MAX_INLINE_SGE		4
40 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE		max_t(int, SZ_16K, PAGE_SIZE)
41 
42 struct nvmet_rdma_cmd {
43 	struct ib_sge		sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
44 	struct ib_cqe		cqe;
45 	struct ib_recv_wr	wr;
46 	struct scatterlist	inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
47 	struct nvme_command     *nvme_cmd;
48 	struct nvmet_rdma_queue	*queue;
49 };
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 rdma_rw_ctx	rw;
66 
67 	struct nvmet_req	req;
68 
69 	u8			n_rdma;
70 	u32			flags;
71 	u32			invalidate_rkey;
72 
73 	struct list_head	wait_list;
74 	struct list_head	free_list;
75 };
76 
77 enum nvmet_rdma_queue_state {
78 	NVMET_RDMA_Q_CONNECTING,
79 	NVMET_RDMA_Q_LIVE,
80 	NVMET_RDMA_Q_DISCONNECTING,
81 };
82 
83 struct nvmet_rdma_queue {
84 	struct rdma_cm_id	*cm_id;
85 	struct nvmet_port	*port;
86 	struct ib_cq		*cq;
87 	atomic_t		sq_wr_avail;
88 	struct nvmet_rdma_device *dev;
89 	spinlock_t		state_lock;
90 	enum nvmet_rdma_queue_state state;
91 	struct nvmet_cq		nvme_cq;
92 	struct nvmet_sq		nvme_sq;
93 
94 	struct nvmet_rdma_rsp	*rsps;
95 	struct list_head	free_rsps;
96 	spinlock_t		rsps_lock;
97 	struct nvmet_rdma_cmd	*cmds;
98 
99 	struct work_struct	release_work;
100 	struct list_head	rsp_wait_list;
101 	struct list_head	rsp_wr_wait_list;
102 	spinlock_t		rsp_wr_wait_lock;
103 
104 	int			idx;
105 	int			host_qid;
106 	int			recv_queue_size;
107 	int			send_queue_size;
108 
109 	struct list_head	queue_list;
110 };
111 
112 struct nvmet_rdma_device {
113 	struct ib_device	*device;
114 	struct ib_pd		*pd;
115 	struct ib_srq		*srq;
116 	struct nvmet_rdma_cmd	*srq_cmds;
117 	size_t			srq_size;
118 	struct kref		ref;
119 	struct list_head	entry;
120 	int			inline_data_size;
121 	int			inline_page_count;
122 };
123 
124 static bool nvmet_rdma_use_srq;
125 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
126 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
127 
128 static DEFINE_IDA(nvmet_rdma_queue_ida);
129 static LIST_HEAD(nvmet_rdma_queue_list);
130 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
131 
132 static LIST_HEAD(device_list);
133 static DEFINE_MUTEX(device_list_mutex);
134 
135 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
136 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
137 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
138 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
139 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
140 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
141 
142 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
143 
144 static int num_pages(int len)
145 {
146 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
147 }
148 
149 /* XXX: really should move to a generic header sooner or later.. */
150 static inline u32 get_unaligned_le24(const u8 *p)
151 {
152 	return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
153 }
154 
155 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
156 {
157 	return nvme_is_write(rsp->req.cmd) &&
158 		rsp->req.transfer_len &&
159 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
160 }
161 
162 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
163 {
164 	return !nvme_is_write(rsp->req.cmd) &&
165 		rsp->req.transfer_len &&
166 		!rsp->req.rsp->status &&
167 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
168 }
169 
170 static inline struct nvmet_rdma_rsp *
171 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
172 {
173 	struct nvmet_rdma_rsp *rsp;
174 	unsigned long flags;
175 
176 	spin_lock_irqsave(&queue->rsps_lock, flags);
177 	rsp = list_first_entry(&queue->free_rsps,
178 				struct nvmet_rdma_rsp, free_list);
179 	list_del(&rsp->free_list);
180 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
181 
182 	return rsp;
183 }
184 
185 static inline void
186 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
187 {
188 	unsigned long flags;
189 
190 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
191 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
192 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
193 }
194 
195 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
196 				struct nvmet_rdma_cmd *c)
197 {
198 	struct scatterlist *sg;
199 	struct ib_sge *sge;
200 	int i;
201 
202 	if (!ndev->inline_data_size)
203 		return;
204 
205 	sg = c->inline_sg;
206 	sge = &c->sge[1];
207 
208 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
209 		if (sge->length)
210 			ib_dma_unmap_page(ndev->device, sge->addr,
211 					sge->length, DMA_FROM_DEVICE);
212 		if (sg_page(sg))
213 			__free_page(sg_page(sg));
214 	}
215 }
216 
217 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
218 				struct nvmet_rdma_cmd *c)
219 {
220 	struct scatterlist *sg;
221 	struct ib_sge *sge;
222 	struct page *pg;
223 	int len;
224 	int i;
225 
226 	if (!ndev->inline_data_size)
227 		return 0;
228 
229 	sg = c->inline_sg;
230 	sg_init_table(sg, ndev->inline_page_count);
231 	sge = &c->sge[1];
232 	len = ndev->inline_data_size;
233 
234 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
235 		pg = alloc_page(GFP_KERNEL);
236 		if (!pg)
237 			goto out_err;
238 		sg_assign_page(sg, pg);
239 		sge->addr = ib_dma_map_page(ndev->device,
240 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
241 		if (ib_dma_mapping_error(ndev->device, sge->addr))
242 			goto out_err;
243 		sge->length = min_t(int, len, PAGE_SIZE);
244 		sge->lkey = ndev->pd->local_dma_lkey;
245 		len -= sge->length;
246 	}
247 
248 	return 0;
249 out_err:
250 	for (; i >= 0; i--, sg--, sge--) {
251 		if (sge->length)
252 			ib_dma_unmap_page(ndev->device, sge->addr,
253 					sge->length, DMA_FROM_DEVICE);
254 		if (sg_page(sg))
255 			__free_page(sg_page(sg));
256 	}
257 	return -ENOMEM;
258 }
259 
260 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
261 			struct nvmet_rdma_cmd *c, bool admin)
262 {
263 	/* NVMe command / RDMA RECV */
264 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
265 	if (!c->nvme_cmd)
266 		goto out;
267 
268 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
269 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
270 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
271 		goto out_free_cmd;
272 
273 	c->sge[0].length = sizeof(*c->nvme_cmd);
274 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
275 
276 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
277 		goto out_unmap_cmd;
278 
279 	c->cqe.done = nvmet_rdma_recv_done;
280 
281 	c->wr.wr_cqe = &c->cqe;
282 	c->wr.sg_list = c->sge;
283 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
284 
285 	return 0;
286 
287 out_unmap_cmd:
288 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
289 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
290 out_free_cmd:
291 	kfree(c->nvme_cmd);
292 
293 out:
294 	return -ENOMEM;
295 }
296 
297 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
298 		struct nvmet_rdma_cmd *c, bool admin)
299 {
300 	if (!admin)
301 		nvmet_rdma_free_inline_pages(ndev, c);
302 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
303 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
304 	kfree(c->nvme_cmd);
305 }
306 
307 static struct nvmet_rdma_cmd *
308 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
309 		int nr_cmds, bool admin)
310 {
311 	struct nvmet_rdma_cmd *cmds;
312 	int ret = -EINVAL, i;
313 
314 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
315 	if (!cmds)
316 		goto out;
317 
318 	for (i = 0; i < nr_cmds; i++) {
319 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
320 		if (ret)
321 			goto out_free;
322 	}
323 
324 	return cmds;
325 
326 out_free:
327 	while (--i >= 0)
328 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
329 	kfree(cmds);
330 out:
331 	return ERR_PTR(ret);
332 }
333 
334 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
335 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
336 {
337 	int i;
338 
339 	for (i = 0; i < nr_cmds; i++)
340 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
341 	kfree(cmds);
342 }
343 
344 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
345 		struct nvmet_rdma_rsp *r)
346 {
347 	/* NVMe CQE / RDMA SEND */
348 	r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
349 	if (!r->req.rsp)
350 		goto out;
351 
352 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
353 			sizeof(*r->req.rsp), DMA_TO_DEVICE);
354 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
355 		goto out_free_rsp;
356 
357 	r->send_sge.length = sizeof(*r->req.rsp);
358 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
359 
360 	r->send_cqe.done = nvmet_rdma_send_done;
361 
362 	r->send_wr.wr_cqe = &r->send_cqe;
363 	r->send_wr.sg_list = &r->send_sge;
364 	r->send_wr.num_sge = 1;
365 	r->send_wr.send_flags = IB_SEND_SIGNALED;
366 
367 	/* Data In / RDMA READ */
368 	r->read_cqe.done = nvmet_rdma_read_data_done;
369 	return 0;
370 
371 out_free_rsp:
372 	kfree(r->req.rsp);
373 out:
374 	return -ENOMEM;
375 }
376 
377 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
378 		struct nvmet_rdma_rsp *r)
379 {
380 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
381 				sizeof(*r->req.rsp), DMA_TO_DEVICE);
382 	kfree(r->req.rsp);
383 }
384 
385 static int
386 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
387 {
388 	struct nvmet_rdma_device *ndev = queue->dev;
389 	int nr_rsps = queue->recv_queue_size * 2;
390 	int ret = -EINVAL, i;
391 
392 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
393 			GFP_KERNEL);
394 	if (!queue->rsps)
395 		goto out;
396 
397 	for (i = 0; i < nr_rsps; i++) {
398 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
399 
400 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
401 		if (ret)
402 			goto out_free;
403 
404 		list_add_tail(&rsp->free_list, &queue->free_rsps);
405 	}
406 
407 	return 0;
408 
409 out_free:
410 	while (--i >= 0) {
411 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
412 
413 		list_del(&rsp->free_list);
414 		nvmet_rdma_free_rsp(ndev, rsp);
415 	}
416 	kfree(queue->rsps);
417 out:
418 	return ret;
419 }
420 
421 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
422 {
423 	struct nvmet_rdma_device *ndev = queue->dev;
424 	int i, nr_rsps = queue->recv_queue_size * 2;
425 
426 	for (i = 0; i < nr_rsps; i++) {
427 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
428 
429 		list_del(&rsp->free_list);
430 		nvmet_rdma_free_rsp(ndev, rsp);
431 	}
432 	kfree(queue->rsps);
433 }
434 
435 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
436 		struct nvmet_rdma_cmd *cmd)
437 {
438 	int ret;
439 
440 	ib_dma_sync_single_for_device(ndev->device,
441 		cmd->sge[0].addr, cmd->sge[0].length,
442 		DMA_FROM_DEVICE);
443 
444 	if (ndev->srq)
445 		ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
446 	else
447 		ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
448 
449 	if (unlikely(ret))
450 		pr_err("post_recv cmd failed\n");
451 
452 	return ret;
453 }
454 
455 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
456 {
457 	spin_lock(&queue->rsp_wr_wait_lock);
458 	while (!list_empty(&queue->rsp_wr_wait_list)) {
459 		struct nvmet_rdma_rsp *rsp;
460 		bool ret;
461 
462 		rsp = list_entry(queue->rsp_wr_wait_list.next,
463 				struct nvmet_rdma_rsp, wait_list);
464 		list_del(&rsp->wait_list);
465 
466 		spin_unlock(&queue->rsp_wr_wait_lock);
467 		ret = nvmet_rdma_execute_command(rsp);
468 		spin_lock(&queue->rsp_wr_wait_lock);
469 
470 		if (!ret) {
471 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
472 			break;
473 		}
474 	}
475 	spin_unlock(&queue->rsp_wr_wait_lock);
476 }
477 
478 
479 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
480 {
481 	struct nvmet_rdma_queue *queue = rsp->queue;
482 
483 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
484 
485 	if (rsp->n_rdma) {
486 		rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
487 				queue->cm_id->port_num, rsp->req.sg,
488 				rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
489 	}
490 
491 	if (rsp->req.sg != rsp->cmd->inline_sg)
492 		sgl_free(rsp->req.sg);
493 
494 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
495 		nvmet_rdma_process_wr_wait_list(queue);
496 
497 	nvmet_rdma_put_rsp(rsp);
498 }
499 
500 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
501 {
502 	if (queue->nvme_sq.ctrl) {
503 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
504 	} else {
505 		/*
506 		 * we didn't setup the controller yet in case
507 		 * of admin connect error, just disconnect and
508 		 * cleanup the queue
509 		 */
510 		nvmet_rdma_queue_disconnect(queue);
511 	}
512 }
513 
514 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
515 {
516 	struct nvmet_rdma_rsp *rsp =
517 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
518 
519 	nvmet_rdma_release_rsp(rsp);
520 
521 	if (unlikely(wc->status != IB_WC_SUCCESS &&
522 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
523 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
524 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
525 		nvmet_rdma_error_comp(rsp->queue);
526 	}
527 }
528 
529 static void nvmet_rdma_queue_response(struct nvmet_req *req)
530 {
531 	struct nvmet_rdma_rsp *rsp =
532 		container_of(req, struct nvmet_rdma_rsp, req);
533 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
534 	struct ib_send_wr *first_wr;
535 
536 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
537 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
538 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
539 	} else {
540 		rsp->send_wr.opcode = IB_WR_SEND;
541 	}
542 
543 	if (nvmet_rdma_need_data_out(rsp))
544 		first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
545 				cm_id->port_num, NULL, &rsp->send_wr);
546 	else
547 		first_wr = &rsp->send_wr;
548 
549 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
550 
551 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
552 		rsp->send_sge.addr, rsp->send_sge.length,
553 		DMA_TO_DEVICE);
554 
555 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
556 		pr_err("sending cmd response failed\n");
557 		nvmet_rdma_release_rsp(rsp);
558 	}
559 }
560 
561 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
562 {
563 	struct nvmet_rdma_rsp *rsp =
564 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
565 	struct nvmet_rdma_queue *queue = cq->cq_context;
566 
567 	WARN_ON(rsp->n_rdma <= 0);
568 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
569 	rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
570 			queue->cm_id->port_num, rsp->req.sg,
571 			rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
572 	rsp->n_rdma = 0;
573 
574 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
575 		nvmet_req_uninit(&rsp->req);
576 		nvmet_rdma_release_rsp(rsp);
577 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
578 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
579 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
580 			nvmet_rdma_error_comp(queue);
581 		}
582 		return;
583 	}
584 
585 	nvmet_req_execute(&rsp->req);
586 }
587 
588 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
589 		u64 off)
590 {
591 	int sg_count = num_pages(len);
592 	struct scatterlist *sg;
593 	int i;
594 
595 	sg = rsp->cmd->inline_sg;
596 	for (i = 0; i < sg_count; i++, sg++) {
597 		if (i < sg_count - 1)
598 			sg_unmark_end(sg);
599 		else
600 			sg_mark_end(sg);
601 		sg->offset = off;
602 		sg->length = min_t(int, len, PAGE_SIZE - off);
603 		len -= sg->length;
604 		if (!i)
605 			off = 0;
606 	}
607 
608 	rsp->req.sg = rsp->cmd->inline_sg;
609 	rsp->req.sg_cnt = sg_count;
610 }
611 
612 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
613 {
614 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
615 	u64 off = le64_to_cpu(sgl->addr);
616 	u32 len = le32_to_cpu(sgl->length);
617 
618 	if (!nvme_is_write(rsp->req.cmd))
619 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
620 
621 	if (off + len > rsp->queue->dev->inline_data_size) {
622 		pr_err("invalid inline data offset!\n");
623 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
624 	}
625 
626 	/* no data command? */
627 	if (!len)
628 		return 0;
629 
630 	nvmet_rdma_use_inline_sg(rsp, len, off);
631 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
632 	rsp->req.transfer_len += len;
633 	return 0;
634 }
635 
636 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
637 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
638 {
639 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
640 	u64 addr = le64_to_cpu(sgl->addr);
641 	u32 len = get_unaligned_le24(sgl->length);
642 	u32 key = get_unaligned_le32(sgl->key);
643 	int ret;
644 
645 	/* no data command? */
646 	if (!len)
647 		return 0;
648 
649 	rsp->req.sg = sgl_alloc(len, GFP_KERNEL, &rsp->req.sg_cnt);
650 	if (!rsp->req.sg)
651 		return NVME_SC_INTERNAL;
652 
653 	ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
654 			rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
655 			nvmet_data_dir(&rsp->req));
656 	if (ret < 0)
657 		return NVME_SC_INTERNAL;
658 	rsp->req.transfer_len += len;
659 	rsp->n_rdma += ret;
660 
661 	if (invalidate) {
662 		rsp->invalidate_rkey = key;
663 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
664 	}
665 
666 	return 0;
667 }
668 
669 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
670 {
671 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
672 
673 	switch (sgl->type >> 4) {
674 	case NVME_SGL_FMT_DATA_DESC:
675 		switch (sgl->type & 0xf) {
676 		case NVME_SGL_FMT_OFFSET:
677 			return nvmet_rdma_map_sgl_inline(rsp);
678 		default:
679 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
680 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
681 		}
682 	case NVME_KEY_SGL_FMT_DATA_DESC:
683 		switch (sgl->type & 0xf) {
684 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
685 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
686 		case NVME_SGL_FMT_ADDRESS:
687 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
688 		default:
689 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
690 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
691 		}
692 	default:
693 		pr_err("invalid SGL type: %#x\n", sgl->type);
694 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
695 	}
696 }
697 
698 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
699 {
700 	struct nvmet_rdma_queue *queue = rsp->queue;
701 
702 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
703 			&queue->sq_wr_avail) < 0)) {
704 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
705 				1 + rsp->n_rdma, queue->idx,
706 				queue->nvme_sq.ctrl->cntlid);
707 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
708 		return false;
709 	}
710 
711 	if (nvmet_rdma_need_data_in(rsp)) {
712 		if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
713 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
714 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
715 	} else {
716 		nvmet_req_execute(&rsp->req);
717 	}
718 
719 	return true;
720 }
721 
722 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
723 		struct nvmet_rdma_rsp *cmd)
724 {
725 	u16 status;
726 
727 	ib_dma_sync_single_for_cpu(queue->dev->device,
728 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
729 		DMA_FROM_DEVICE);
730 	ib_dma_sync_single_for_cpu(queue->dev->device,
731 		cmd->send_sge.addr, cmd->send_sge.length,
732 		DMA_TO_DEVICE);
733 
734 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
735 			&queue->nvme_sq, &nvmet_rdma_ops))
736 		return;
737 
738 	status = nvmet_rdma_map_sgl(cmd);
739 	if (status)
740 		goto out_err;
741 
742 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
743 		spin_lock(&queue->rsp_wr_wait_lock);
744 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
745 		spin_unlock(&queue->rsp_wr_wait_lock);
746 	}
747 
748 	return;
749 
750 out_err:
751 	nvmet_req_complete(&cmd->req, status);
752 }
753 
754 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
755 {
756 	struct nvmet_rdma_cmd *cmd =
757 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
758 	struct nvmet_rdma_queue *queue = cq->cq_context;
759 	struct nvmet_rdma_rsp *rsp;
760 
761 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
762 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
763 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
764 				wc->wr_cqe, ib_wc_status_msg(wc->status),
765 				wc->status);
766 			nvmet_rdma_error_comp(queue);
767 		}
768 		return;
769 	}
770 
771 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
772 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
773 		nvmet_rdma_error_comp(queue);
774 		return;
775 	}
776 
777 	cmd->queue = queue;
778 	rsp = nvmet_rdma_get_rsp(queue);
779 	rsp->queue = queue;
780 	rsp->cmd = cmd;
781 	rsp->flags = 0;
782 	rsp->req.cmd = cmd->nvme_cmd;
783 	rsp->req.port = queue->port;
784 	rsp->n_rdma = 0;
785 
786 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
787 		unsigned long flags;
788 
789 		spin_lock_irqsave(&queue->state_lock, flags);
790 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
791 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
792 		else
793 			nvmet_rdma_put_rsp(rsp);
794 		spin_unlock_irqrestore(&queue->state_lock, flags);
795 		return;
796 	}
797 
798 	nvmet_rdma_handle_command(queue, rsp);
799 }
800 
801 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
802 {
803 	if (!ndev->srq)
804 		return;
805 
806 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
807 	ib_destroy_srq(ndev->srq);
808 }
809 
810 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
811 {
812 	struct ib_srq_init_attr srq_attr = { NULL, };
813 	struct ib_srq *srq;
814 	size_t srq_size;
815 	int ret, i;
816 
817 	srq_size = 4095;	/* XXX: tune */
818 
819 	srq_attr.attr.max_wr = srq_size;
820 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
821 	srq_attr.attr.srq_limit = 0;
822 	srq_attr.srq_type = IB_SRQT_BASIC;
823 	srq = ib_create_srq(ndev->pd, &srq_attr);
824 	if (IS_ERR(srq)) {
825 		/*
826 		 * If SRQs aren't supported we just go ahead and use normal
827 		 * non-shared receive queues.
828 		 */
829 		pr_info("SRQ requested but not supported.\n");
830 		return 0;
831 	}
832 
833 	ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
834 	if (IS_ERR(ndev->srq_cmds)) {
835 		ret = PTR_ERR(ndev->srq_cmds);
836 		goto out_destroy_srq;
837 	}
838 
839 	ndev->srq = srq;
840 	ndev->srq_size = srq_size;
841 
842 	for (i = 0; i < srq_size; i++) {
843 		ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
844 		if (ret)
845 			goto out_free_cmds;
846 	}
847 
848 	return 0;
849 
850 out_free_cmds:
851 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
852 out_destroy_srq:
853 	ib_destroy_srq(srq);
854 	return ret;
855 }
856 
857 static void nvmet_rdma_free_dev(struct kref *ref)
858 {
859 	struct nvmet_rdma_device *ndev =
860 		container_of(ref, struct nvmet_rdma_device, ref);
861 
862 	mutex_lock(&device_list_mutex);
863 	list_del(&ndev->entry);
864 	mutex_unlock(&device_list_mutex);
865 
866 	nvmet_rdma_destroy_srq(ndev);
867 	ib_dealloc_pd(ndev->pd);
868 
869 	kfree(ndev);
870 }
871 
872 static struct nvmet_rdma_device *
873 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
874 {
875 	struct nvmet_port *port = cm_id->context;
876 	struct nvmet_rdma_device *ndev;
877 	int inline_page_count;
878 	int inline_sge_count;
879 	int ret;
880 
881 	mutex_lock(&device_list_mutex);
882 	list_for_each_entry(ndev, &device_list, entry) {
883 		if (ndev->device->node_guid == cm_id->device->node_guid &&
884 		    kref_get_unless_zero(&ndev->ref))
885 			goto out_unlock;
886 	}
887 
888 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
889 	if (!ndev)
890 		goto out_err;
891 
892 	inline_page_count = num_pages(port->inline_data_size);
893 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
894 				cm_id->device->attrs.max_recv_sge) - 1;
895 	if (inline_page_count > inline_sge_count) {
896 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
897 			port->inline_data_size, cm_id->device->name,
898 			inline_sge_count * PAGE_SIZE);
899 		port->inline_data_size = inline_sge_count * PAGE_SIZE;
900 		inline_page_count = inline_sge_count;
901 	}
902 	ndev->inline_data_size = port->inline_data_size;
903 	ndev->inline_page_count = inline_page_count;
904 	ndev->device = cm_id->device;
905 	kref_init(&ndev->ref);
906 
907 	ndev->pd = ib_alloc_pd(ndev->device, 0);
908 	if (IS_ERR(ndev->pd))
909 		goto out_free_dev;
910 
911 	if (nvmet_rdma_use_srq) {
912 		ret = nvmet_rdma_init_srq(ndev);
913 		if (ret)
914 			goto out_free_pd;
915 	}
916 
917 	list_add(&ndev->entry, &device_list);
918 out_unlock:
919 	mutex_unlock(&device_list_mutex);
920 	pr_debug("added %s.\n", ndev->device->name);
921 	return ndev;
922 
923 out_free_pd:
924 	ib_dealloc_pd(ndev->pd);
925 out_free_dev:
926 	kfree(ndev);
927 out_err:
928 	mutex_unlock(&device_list_mutex);
929 	return NULL;
930 }
931 
932 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
933 {
934 	struct ib_qp_init_attr qp_attr;
935 	struct nvmet_rdma_device *ndev = queue->dev;
936 	int comp_vector, nr_cqe, ret, i;
937 
938 	/*
939 	 * Spread the io queues across completion vectors,
940 	 * but still keep all admin queues on vector 0.
941 	 */
942 	comp_vector = !queue->host_qid ? 0 :
943 		queue->idx % ndev->device->num_comp_vectors;
944 
945 	/*
946 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
947 	 */
948 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
949 
950 	queue->cq = ib_alloc_cq(ndev->device, queue,
951 			nr_cqe + 1, comp_vector,
952 			IB_POLL_WORKQUEUE);
953 	if (IS_ERR(queue->cq)) {
954 		ret = PTR_ERR(queue->cq);
955 		pr_err("failed to create CQ cqe= %d ret= %d\n",
956 		       nr_cqe + 1, ret);
957 		goto out;
958 	}
959 
960 	memset(&qp_attr, 0, sizeof(qp_attr));
961 	qp_attr.qp_context = queue;
962 	qp_attr.event_handler = nvmet_rdma_qp_event;
963 	qp_attr.send_cq = queue->cq;
964 	qp_attr.recv_cq = queue->cq;
965 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
966 	qp_attr.qp_type = IB_QPT_RC;
967 	/* +1 for drain */
968 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
969 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
970 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
971 					ndev->device->attrs.max_send_sge);
972 
973 	if (ndev->srq) {
974 		qp_attr.srq = ndev->srq;
975 	} else {
976 		/* +1 for drain */
977 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
978 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
979 	}
980 
981 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
982 	if (ret) {
983 		pr_err("failed to create_qp ret= %d\n", ret);
984 		goto err_destroy_cq;
985 	}
986 
987 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
988 
989 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
990 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
991 		 qp_attr.cap.max_send_wr, queue->cm_id);
992 
993 	if (!ndev->srq) {
994 		for (i = 0; i < queue->recv_queue_size; i++) {
995 			queue->cmds[i].queue = queue;
996 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
997 			if (ret)
998 				goto err_destroy_qp;
999 		}
1000 	}
1001 
1002 out:
1003 	return ret;
1004 
1005 err_destroy_qp:
1006 	rdma_destroy_qp(queue->cm_id);
1007 err_destroy_cq:
1008 	ib_free_cq(queue->cq);
1009 	goto out;
1010 }
1011 
1012 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1013 {
1014 	struct ib_qp *qp = queue->cm_id->qp;
1015 
1016 	ib_drain_qp(qp);
1017 	rdma_destroy_id(queue->cm_id);
1018 	ib_destroy_qp(qp);
1019 	ib_free_cq(queue->cq);
1020 }
1021 
1022 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1023 {
1024 	pr_debug("freeing queue %d\n", queue->idx);
1025 
1026 	nvmet_sq_destroy(&queue->nvme_sq);
1027 
1028 	nvmet_rdma_destroy_queue_ib(queue);
1029 	if (!queue->dev->srq) {
1030 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1031 				queue->recv_queue_size,
1032 				!queue->host_qid);
1033 	}
1034 	nvmet_rdma_free_rsps(queue);
1035 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1036 	kfree(queue);
1037 }
1038 
1039 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1040 {
1041 	struct nvmet_rdma_queue *queue =
1042 		container_of(w, struct nvmet_rdma_queue, release_work);
1043 	struct nvmet_rdma_device *dev = queue->dev;
1044 
1045 	nvmet_rdma_free_queue(queue);
1046 
1047 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1048 }
1049 
1050 static int
1051 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1052 				struct nvmet_rdma_queue *queue)
1053 {
1054 	struct nvme_rdma_cm_req *req;
1055 
1056 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1057 	if (!req || conn->private_data_len == 0)
1058 		return NVME_RDMA_CM_INVALID_LEN;
1059 
1060 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1061 		return NVME_RDMA_CM_INVALID_RECFMT;
1062 
1063 	queue->host_qid = le16_to_cpu(req->qid);
1064 
1065 	/*
1066 	 * req->hsqsize corresponds to our recv queue size plus 1
1067 	 * req->hrqsize corresponds to our send queue size
1068 	 */
1069 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1070 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1071 
1072 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1073 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1074 
1075 	/* XXX: Should we enforce some kind of max for IO queues? */
1076 
1077 	return 0;
1078 }
1079 
1080 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1081 				enum nvme_rdma_cm_status status)
1082 {
1083 	struct nvme_rdma_cm_rej rej;
1084 
1085 	pr_debug("rejecting connect request: status %d (%s)\n",
1086 		 status, nvme_rdma_cm_msg(status));
1087 
1088 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1089 	rej.sts = cpu_to_le16(status);
1090 
1091 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1092 }
1093 
1094 static struct nvmet_rdma_queue *
1095 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1096 		struct rdma_cm_id *cm_id,
1097 		struct rdma_cm_event *event)
1098 {
1099 	struct nvmet_rdma_queue *queue;
1100 	int ret;
1101 
1102 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1103 	if (!queue) {
1104 		ret = NVME_RDMA_CM_NO_RSC;
1105 		goto out_reject;
1106 	}
1107 
1108 	ret = nvmet_sq_init(&queue->nvme_sq);
1109 	if (ret) {
1110 		ret = NVME_RDMA_CM_NO_RSC;
1111 		goto out_free_queue;
1112 	}
1113 
1114 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1115 	if (ret)
1116 		goto out_destroy_sq;
1117 
1118 	/*
1119 	 * Schedules the actual release because calling rdma_destroy_id from
1120 	 * inside a CM callback would trigger a deadlock. (great API design..)
1121 	 */
1122 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1123 	queue->dev = ndev;
1124 	queue->cm_id = cm_id;
1125 
1126 	spin_lock_init(&queue->state_lock);
1127 	queue->state = NVMET_RDMA_Q_CONNECTING;
1128 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1129 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1130 	spin_lock_init(&queue->rsp_wr_wait_lock);
1131 	INIT_LIST_HEAD(&queue->free_rsps);
1132 	spin_lock_init(&queue->rsps_lock);
1133 	INIT_LIST_HEAD(&queue->queue_list);
1134 
1135 	queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1136 	if (queue->idx < 0) {
1137 		ret = NVME_RDMA_CM_NO_RSC;
1138 		goto out_destroy_sq;
1139 	}
1140 
1141 	ret = nvmet_rdma_alloc_rsps(queue);
1142 	if (ret) {
1143 		ret = NVME_RDMA_CM_NO_RSC;
1144 		goto out_ida_remove;
1145 	}
1146 
1147 	if (!ndev->srq) {
1148 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1149 				queue->recv_queue_size,
1150 				!queue->host_qid);
1151 		if (IS_ERR(queue->cmds)) {
1152 			ret = NVME_RDMA_CM_NO_RSC;
1153 			goto out_free_responses;
1154 		}
1155 	}
1156 
1157 	ret = nvmet_rdma_create_queue_ib(queue);
1158 	if (ret) {
1159 		pr_err("%s: creating RDMA queue failed (%d).\n",
1160 			__func__, ret);
1161 		ret = NVME_RDMA_CM_NO_RSC;
1162 		goto out_free_cmds;
1163 	}
1164 
1165 	return queue;
1166 
1167 out_free_cmds:
1168 	if (!ndev->srq) {
1169 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1170 				queue->recv_queue_size,
1171 				!queue->host_qid);
1172 	}
1173 out_free_responses:
1174 	nvmet_rdma_free_rsps(queue);
1175 out_ida_remove:
1176 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1177 out_destroy_sq:
1178 	nvmet_sq_destroy(&queue->nvme_sq);
1179 out_free_queue:
1180 	kfree(queue);
1181 out_reject:
1182 	nvmet_rdma_cm_reject(cm_id, ret);
1183 	return NULL;
1184 }
1185 
1186 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1187 {
1188 	struct nvmet_rdma_queue *queue = priv;
1189 
1190 	switch (event->event) {
1191 	case IB_EVENT_COMM_EST:
1192 		rdma_notify(queue->cm_id, event->event);
1193 		break;
1194 	default:
1195 		pr_err("received IB QP event: %s (%d)\n",
1196 		       ib_event_msg(event->event), event->event);
1197 		break;
1198 	}
1199 }
1200 
1201 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1202 		struct nvmet_rdma_queue *queue,
1203 		struct rdma_conn_param *p)
1204 {
1205 	struct rdma_conn_param  param = { };
1206 	struct nvme_rdma_cm_rep priv = { };
1207 	int ret = -ENOMEM;
1208 
1209 	param.rnr_retry_count = 7;
1210 	param.flow_control = 1;
1211 	param.initiator_depth = min_t(u8, p->initiator_depth,
1212 		queue->dev->device->attrs.max_qp_init_rd_atom);
1213 	param.private_data = &priv;
1214 	param.private_data_len = sizeof(priv);
1215 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1216 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1217 
1218 	ret = rdma_accept(cm_id, &param);
1219 	if (ret)
1220 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1221 
1222 	return ret;
1223 }
1224 
1225 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1226 		struct rdma_cm_event *event)
1227 {
1228 	struct nvmet_rdma_device *ndev;
1229 	struct nvmet_rdma_queue *queue;
1230 	int ret = -EINVAL;
1231 
1232 	ndev = nvmet_rdma_find_get_device(cm_id);
1233 	if (!ndev) {
1234 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1235 		return -ECONNREFUSED;
1236 	}
1237 
1238 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1239 	if (!queue) {
1240 		ret = -ENOMEM;
1241 		goto put_device;
1242 	}
1243 	queue->port = cm_id->context;
1244 
1245 	if (queue->host_qid == 0) {
1246 		/* Let inflight controller teardown complete */
1247 		flush_scheduled_work();
1248 	}
1249 
1250 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1251 	if (ret) {
1252 		schedule_work(&queue->release_work);
1253 		/* Destroying rdma_cm id is not needed here */
1254 		return 0;
1255 	}
1256 
1257 	mutex_lock(&nvmet_rdma_queue_mutex);
1258 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1259 	mutex_unlock(&nvmet_rdma_queue_mutex);
1260 
1261 	return 0;
1262 
1263 put_device:
1264 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1265 
1266 	return ret;
1267 }
1268 
1269 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1270 {
1271 	unsigned long flags;
1272 
1273 	spin_lock_irqsave(&queue->state_lock, flags);
1274 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1275 		pr_warn("trying to establish a connected queue\n");
1276 		goto out_unlock;
1277 	}
1278 	queue->state = NVMET_RDMA_Q_LIVE;
1279 
1280 	while (!list_empty(&queue->rsp_wait_list)) {
1281 		struct nvmet_rdma_rsp *cmd;
1282 
1283 		cmd = list_first_entry(&queue->rsp_wait_list,
1284 					struct nvmet_rdma_rsp, wait_list);
1285 		list_del(&cmd->wait_list);
1286 
1287 		spin_unlock_irqrestore(&queue->state_lock, flags);
1288 		nvmet_rdma_handle_command(queue, cmd);
1289 		spin_lock_irqsave(&queue->state_lock, flags);
1290 	}
1291 
1292 out_unlock:
1293 	spin_unlock_irqrestore(&queue->state_lock, flags);
1294 }
1295 
1296 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1297 {
1298 	bool disconnect = false;
1299 	unsigned long flags;
1300 
1301 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1302 
1303 	spin_lock_irqsave(&queue->state_lock, flags);
1304 	switch (queue->state) {
1305 	case NVMET_RDMA_Q_CONNECTING:
1306 	case NVMET_RDMA_Q_LIVE:
1307 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1308 		disconnect = true;
1309 		break;
1310 	case NVMET_RDMA_Q_DISCONNECTING:
1311 		break;
1312 	}
1313 	spin_unlock_irqrestore(&queue->state_lock, flags);
1314 
1315 	if (disconnect) {
1316 		rdma_disconnect(queue->cm_id);
1317 		schedule_work(&queue->release_work);
1318 	}
1319 }
1320 
1321 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1322 {
1323 	bool disconnect = false;
1324 
1325 	mutex_lock(&nvmet_rdma_queue_mutex);
1326 	if (!list_empty(&queue->queue_list)) {
1327 		list_del_init(&queue->queue_list);
1328 		disconnect = true;
1329 	}
1330 	mutex_unlock(&nvmet_rdma_queue_mutex);
1331 
1332 	if (disconnect)
1333 		__nvmet_rdma_queue_disconnect(queue);
1334 }
1335 
1336 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1337 		struct nvmet_rdma_queue *queue)
1338 {
1339 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1340 
1341 	mutex_lock(&nvmet_rdma_queue_mutex);
1342 	if (!list_empty(&queue->queue_list))
1343 		list_del_init(&queue->queue_list);
1344 	mutex_unlock(&nvmet_rdma_queue_mutex);
1345 
1346 	pr_err("failed to connect queue %d\n", queue->idx);
1347 	schedule_work(&queue->release_work);
1348 }
1349 
1350 /**
1351  * nvme_rdma_device_removal() - Handle RDMA device removal
1352  * @cm_id:	rdma_cm id, used for nvmet port
1353  * @queue:      nvmet rdma queue (cm id qp_context)
1354  *
1355  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1356  * to unplug. Note that this event can be generated on a normal
1357  * queue cm_id and/or a device bound listener cm_id (where in this
1358  * case queue will be null).
1359  *
1360  * We registered an ib_client to handle device removal for queues,
1361  * so we only need to handle the listening port cm_ids. In this case
1362  * we nullify the priv to prevent double cm_id destruction and destroying
1363  * the cm_id implicitely by returning a non-zero rc to the callout.
1364  */
1365 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1366 		struct nvmet_rdma_queue *queue)
1367 {
1368 	struct nvmet_port *port;
1369 
1370 	if (queue) {
1371 		/*
1372 		 * This is a queue cm_id. we have registered
1373 		 * an ib_client to handle queues removal
1374 		 * so don't interfear and just return.
1375 		 */
1376 		return 0;
1377 	}
1378 
1379 	port = cm_id->context;
1380 
1381 	/*
1382 	 * This is a listener cm_id. Make sure that
1383 	 * future remove_port won't invoke a double
1384 	 * cm_id destroy. use atomic xchg to make sure
1385 	 * we don't compete with remove_port.
1386 	 */
1387 	if (xchg(&port->priv, NULL) != cm_id)
1388 		return 0;
1389 
1390 	/*
1391 	 * We need to return 1 so that the core will destroy
1392 	 * it's own ID.  What a great API design..
1393 	 */
1394 	return 1;
1395 }
1396 
1397 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1398 		struct rdma_cm_event *event)
1399 {
1400 	struct nvmet_rdma_queue *queue = NULL;
1401 	int ret = 0;
1402 
1403 	if (cm_id->qp)
1404 		queue = cm_id->qp->qp_context;
1405 
1406 	pr_debug("%s (%d): status %d id %p\n",
1407 		rdma_event_msg(event->event), event->event,
1408 		event->status, cm_id);
1409 
1410 	switch (event->event) {
1411 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1412 		ret = nvmet_rdma_queue_connect(cm_id, event);
1413 		break;
1414 	case RDMA_CM_EVENT_ESTABLISHED:
1415 		nvmet_rdma_queue_established(queue);
1416 		break;
1417 	case RDMA_CM_EVENT_ADDR_CHANGE:
1418 	case RDMA_CM_EVENT_DISCONNECTED:
1419 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1420 		nvmet_rdma_queue_disconnect(queue);
1421 		break;
1422 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1423 		ret = nvmet_rdma_device_removal(cm_id, queue);
1424 		break;
1425 	case RDMA_CM_EVENT_REJECTED:
1426 		pr_debug("Connection rejected: %s\n",
1427 			 rdma_reject_msg(cm_id, event->status));
1428 		/* FALLTHROUGH */
1429 	case RDMA_CM_EVENT_UNREACHABLE:
1430 	case RDMA_CM_EVENT_CONNECT_ERROR:
1431 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1432 		break;
1433 	default:
1434 		pr_err("received unrecognized RDMA CM event %d\n",
1435 			event->event);
1436 		break;
1437 	}
1438 
1439 	return ret;
1440 }
1441 
1442 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1443 {
1444 	struct nvmet_rdma_queue *queue;
1445 
1446 restart:
1447 	mutex_lock(&nvmet_rdma_queue_mutex);
1448 	list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1449 		if (queue->nvme_sq.ctrl == ctrl) {
1450 			list_del_init(&queue->queue_list);
1451 			mutex_unlock(&nvmet_rdma_queue_mutex);
1452 
1453 			__nvmet_rdma_queue_disconnect(queue);
1454 			goto restart;
1455 		}
1456 	}
1457 	mutex_unlock(&nvmet_rdma_queue_mutex);
1458 }
1459 
1460 static int nvmet_rdma_add_port(struct nvmet_port *port)
1461 {
1462 	struct rdma_cm_id *cm_id;
1463 	struct sockaddr_storage addr = { };
1464 	__kernel_sa_family_t af;
1465 	int ret;
1466 
1467 	switch (port->disc_addr.adrfam) {
1468 	case NVMF_ADDR_FAMILY_IP4:
1469 		af = AF_INET;
1470 		break;
1471 	case NVMF_ADDR_FAMILY_IP6:
1472 		af = AF_INET6;
1473 		break;
1474 	default:
1475 		pr_err("address family %d not supported\n",
1476 				port->disc_addr.adrfam);
1477 		return -EINVAL;
1478 	}
1479 
1480 	if (port->inline_data_size < 0) {
1481 		port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1482 	} else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1483 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1484 			port->inline_data_size,
1485 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1486 		port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1487 	}
1488 
1489 	ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1490 			port->disc_addr.trsvcid, &addr);
1491 	if (ret) {
1492 		pr_err("malformed ip/port passed: %s:%s\n",
1493 			port->disc_addr.traddr, port->disc_addr.trsvcid);
1494 		return ret;
1495 	}
1496 
1497 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1498 			RDMA_PS_TCP, IB_QPT_RC);
1499 	if (IS_ERR(cm_id)) {
1500 		pr_err("CM ID creation failed\n");
1501 		return PTR_ERR(cm_id);
1502 	}
1503 
1504 	/*
1505 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1506 	 * at the same time.
1507 	 */
1508 	ret = rdma_set_afonly(cm_id, 1);
1509 	if (ret) {
1510 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1511 		goto out_destroy_id;
1512 	}
1513 
1514 	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1515 	if (ret) {
1516 		pr_err("binding CM ID to %pISpcs failed (%d)\n",
1517 			(struct sockaddr *)&addr, ret);
1518 		goto out_destroy_id;
1519 	}
1520 
1521 	ret = rdma_listen(cm_id, 128);
1522 	if (ret) {
1523 		pr_err("listening to %pISpcs failed (%d)\n",
1524 			(struct sockaddr *)&addr, ret);
1525 		goto out_destroy_id;
1526 	}
1527 
1528 	pr_info("enabling port %d (%pISpcs)\n",
1529 		le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1530 	port->priv = cm_id;
1531 	return 0;
1532 
1533 out_destroy_id:
1534 	rdma_destroy_id(cm_id);
1535 	return ret;
1536 }
1537 
1538 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1539 {
1540 	struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1541 
1542 	if (cm_id)
1543 		rdma_destroy_id(cm_id);
1544 }
1545 
1546 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1547 		struct nvmet_port *port, char *traddr)
1548 {
1549 	struct rdma_cm_id *cm_id = port->priv;
1550 
1551 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1552 		struct nvmet_rdma_rsp *rsp =
1553 			container_of(req, struct nvmet_rdma_rsp, req);
1554 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1555 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1556 
1557 		sprintf(traddr, "%pISc", addr);
1558 	} else {
1559 		memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1560 	}
1561 }
1562 
1563 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1564 	.owner			= THIS_MODULE,
1565 	.type			= NVMF_TRTYPE_RDMA,
1566 	.msdbd			= 1,
1567 	.has_keyed_sgls		= 1,
1568 	.add_port		= nvmet_rdma_add_port,
1569 	.remove_port		= nvmet_rdma_remove_port,
1570 	.queue_response		= nvmet_rdma_queue_response,
1571 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
1572 	.disc_traddr		= nvmet_rdma_disc_port_addr,
1573 };
1574 
1575 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1576 {
1577 	struct nvmet_rdma_queue *queue, *tmp;
1578 	struct nvmet_rdma_device *ndev;
1579 	bool found = false;
1580 
1581 	mutex_lock(&device_list_mutex);
1582 	list_for_each_entry(ndev, &device_list, entry) {
1583 		if (ndev->device == ib_device) {
1584 			found = true;
1585 			break;
1586 		}
1587 	}
1588 	mutex_unlock(&device_list_mutex);
1589 
1590 	if (!found)
1591 		return;
1592 
1593 	/*
1594 	 * IB Device that is used by nvmet controllers is being removed,
1595 	 * delete all queues using this device.
1596 	 */
1597 	mutex_lock(&nvmet_rdma_queue_mutex);
1598 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1599 				 queue_list) {
1600 		if (queue->dev->device != ib_device)
1601 			continue;
1602 
1603 		pr_info("Removing queue %d\n", queue->idx);
1604 		list_del_init(&queue->queue_list);
1605 		__nvmet_rdma_queue_disconnect(queue);
1606 	}
1607 	mutex_unlock(&nvmet_rdma_queue_mutex);
1608 
1609 	flush_scheduled_work();
1610 }
1611 
1612 static struct ib_client nvmet_rdma_ib_client = {
1613 	.name   = "nvmet_rdma",
1614 	.remove = nvmet_rdma_remove_one
1615 };
1616 
1617 static int __init nvmet_rdma_init(void)
1618 {
1619 	int ret;
1620 
1621 	ret = ib_register_client(&nvmet_rdma_ib_client);
1622 	if (ret)
1623 		return ret;
1624 
1625 	ret = nvmet_register_transport(&nvmet_rdma_ops);
1626 	if (ret)
1627 		goto err_ib_client;
1628 
1629 	return 0;
1630 
1631 err_ib_client:
1632 	ib_unregister_client(&nvmet_rdma_ib_client);
1633 	return ret;
1634 }
1635 
1636 static void __exit nvmet_rdma_exit(void)
1637 {
1638 	nvmet_unregister_transport(&nvmet_rdma_ops);
1639 	ib_unregister_client(&nvmet_rdma_ib_client);
1640 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1641 	ida_destroy(&nvmet_rdma_queue_ida);
1642 }
1643 
1644 module_init(nvmet_rdma_init);
1645 module_exit(nvmet_rdma_exit);
1646 
1647 MODULE_LICENSE("GPL v2");
1648 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
1649