xref: /openbmc/linux/drivers/nvme/target/rdma.c (revision 1830dad34c070161fda2ff1db77b39ffa78aa380)
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 	bool			allocated;
70 	u8			n_rdma;
71 	u32			flags;
72 	u32			invalidate_rkey;
73 
74 	struct list_head	wait_list;
75 	struct list_head	free_list;
76 };
77 
78 enum nvmet_rdma_queue_state {
79 	NVMET_RDMA_Q_CONNECTING,
80 	NVMET_RDMA_Q_LIVE,
81 	NVMET_RDMA_Q_DISCONNECTING,
82 };
83 
84 struct nvmet_rdma_queue {
85 	struct rdma_cm_id	*cm_id;
86 	struct nvmet_port	*port;
87 	struct ib_cq		*cq;
88 	atomic_t		sq_wr_avail;
89 	struct nvmet_rdma_device *dev;
90 	spinlock_t		state_lock;
91 	enum nvmet_rdma_queue_state state;
92 	struct nvmet_cq		nvme_cq;
93 	struct nvmet_sq		nvme_sq;
94 
95 	struct nvmet_rdma_rsp	*rsps;
96 	struct list_head	free_rsps;
97 	spinlock_t		rsps_lock;
98 	struct nvmet_rdma_cmd	*cmds;
99 
100 	struct work_struct	release_work;
101 	struct list_head	rsp_wait_list;
102 	struct list_head	rsp_wr_wait_list;
103 	spinlock_t		rsp_wr_wait_lock;
104 
105 	int			idx;
106 	int			host_qid;
107 	int			recv_queue_size;
108 	int			send_queue_size;
109 
110 	struct list_head	queue_list;
111 };
112 
113 struct nvmet_rdma_device {
114 	struct ib_device	*device;
115 	struct ib_pd		*pd;
116 	struct ib_srq		*srq;
117 	struct nvmet_rdma_cmd	*srq_cmds;
118 	size_t			srq_size;
119 	struct kref		ref;
120 	struct list_head	entry;
121 	int			inline_data_size;
122 	int			inline_page_count;
123 };
124 
125 static bool nvmet_rdma_use_srq;
126 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
127 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
128 
129 static DEFINE_IDA(nvmet_rdma_queue_ida);
130 static LIST_HEAD(nvmet_rdma_queue_list);
131 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
132 
133 static LIST_HEAD(device_list);
134 static DEFINE_MUTEX(device_list_mutex);
135 
136 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
137 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
138 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
139 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
140 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
141 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
142 
143 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
144 
145 static int num_pages(int len)
146 {
147 	return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
148 }
149 
150 /* XXX: really should move to a generic header sooner or later.. */
151 static inline u32 get_unaligned_le24(const u8 *p)
152 {
153 	return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
154 }
155 
156 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
157 {
158 	return nvme_is_write(rsp->req.cmd) &&
159 		rsp->req.transfer_len &&
160 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
161 }
162 
163 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
164 {
165 	return !nvme_is_write(rsp->req.cmd) &&
166 		rsp->req.transfer_len &&
167 		!rsp->req.rsp->status &&
168 		!(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
169 }
170 
171 static inline struct nvmet_rdma_rsp *
172 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
173 {
174 	struct nvmet_rdma_rsp *rsp;
175 	unsigned long flags;
176 
177 	spin_lock_irqsave(&queue->rsps_lock, flags);
178 	rsp = list_first_entry_or_null(&queue->free_rsps,
179 				struct nvmet_rdma_rsp, free_list);
180 	if (likely(rsp))
181 		list_del(&rsp->free_list);
182 	spin_unlock_irqrestore(&queue->rsps_lock, flags);
183 
184 	if (unlikely(!rsp)) {
185 		rsp = kmalloc(sizeof(*rsp), GFP_KERNEL);
186 		if (unlikely(!rsp))
187 			return NULL;
188 		rsp->allocated = true;
189 	}
190 
191 	return rsp;
192 }
193 
194 static inline void
195 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
196 {
197 	unsigned long flags;
198 
199 	if (rsp->allocated) {
200 		kfree(rsp);
201 		return;
202 	}
203 
204 	spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
205 	list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
206 	spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
207 }
208 
209 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
210 				struct nvmet_rdma_cmd *c)
211 {
212 	struct scatterlist *sg;
213 	struct ib_sge *sge;
214 	int i;
215 
216 	if (!ndev->inline_data_size)
217 		return;
218 
219 	sg = c->inline_sg;
220 	sge = &c->sge[1];
221 
222 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
223 		if (sge->length)
224 			ib_dma_unmap_page(ndev->device, sge->addr,
225 					sge->length, DMA_FROM_DEVICE);
226 		if (sg_page(sg))
227 			__free_page(sg_page(sg));
228 	}
229 }
230 
231 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
232 				struct nvmet_rdma_cmd *c)
233 {
234 	struct scatterlist *sg;
235 	struct ib_sge *sge;
236 	struct page *pg;
237 	int len;
238 	int i;
239 
240 	if (!ndev->inline_data_size)
241 		return 0;
242 
243 	sg = c->inline_sg;
244 	sg_init_table(sg, ndev->inline_page_count);
245 	sge = &c->sge[1];
246 	len = ndev->inline_data_size;
247 
248 	for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
249 		pg = alloc_page(GFP_KERNEL);
250 		if (!pg)
251 			goto out_err;
252 		sg_assign_page(sg, pg);
253 		sge->addr = ib_dma_map_page(ndev->device,
254 			pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
255 		if (ib_dma_mapping_error(ndev->device, sge->addr))
256 			goto out_err;
257 		sge->length = min_t(int, len, PAGE_SIZE);
258 		sge->lkey = ndev->pd->local_dma_lkey;
259 		len -= sge->length;
260 	}
261 
262 	return 0;
263 out_err:
264 	for (; i >= 0; i--, sg--, sge--) {
265 		if (sge->length)
266 			ib_dma_unmap_page(ndev->device, sge->addr,
267 					sge->length, DMA_FROM_DEVICE);
268 		if (sg_page(sg))
269 			__free_page(sg_page(sg));
270 	}
271 	return -ENOMEM;
272 }
273 
274 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
275 			struct nvmet_rdma_cmd *c, bool admin)
276 {
277 	/* NVMe command / RDMA RECV */
278 	c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
279 	if (!c->nvme_cmd)
280 		goto out;
281 
282 	c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
283 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
284 	if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
285 		goto out_free_cmd;
286 
287 	c->sge[0].length = sizeof(*c->nvme_cmd);
288 	c->sge[0].lkey = ndev->pd->local_dma_lkey;
289 
290 	if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
291 		goto out_unmap_cmd;
292 
293 	c->cqe.done = nvmet_rdma_recv_done;
294 
295 	c->wr.wr_cqe = &c->cqe;
296 	c->wr.sg_list = c->sge;
297 	c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
298 
299 	return 0;
300 
301 out_unmap_cmd:
302 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
303 			sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
304 out_free_cmd:
305 	kfree(c->nvme_cmd);
306 
307 out:
308 	return -ENOMEM;
309 }
310 
311 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
312 		struct nvmet_rdma_cmd *c, bool admin)
313 {
314 	if (!admin)
315 		nvmet_rdma_free_inline_pages(ndev, c);
316 	ib_dma_unmap_single(ndev->device, c->sge[0].addr,
317 				sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
318 	kfree(c->nvme_cmd);
319 }
320 
321 static struct nvmet_rdma_cmd *
322 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
323 		int nr_cmds, bool admin)
324 {
325 	struct nvmet_rdma_cmd *cmds;
326 	int ret = -EINVAL, i;
327 
328 	cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
329 	if (!cmds)
330 		goto out;
331 
332 	for (i = 0; i < nr_cmds; i++) {
333 		ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
334 		if (ret)
335 			goto out_free;
336 	}
337 
338 	return cmds;
339 
340 out_free:
341 	while (--i >= 0)
342 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
343 	kfree(cmds);
344 out:
345 	return ERR_PTR(ret);
346 }
347 
348 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
349 		struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
350 {
351 	int i;
352 
353 	for (i = 0; i < nr_cmds; i++)
354 		nvmet_rdma_free_cmd(ndev, cmds + i, admin);
355 	kfree(cmds);
356 }
357 
358 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
359 		struct nvmet_rdma_rsp *r)
360 {
361 	/* NVMe CQE / RDMA SEND */
362 	r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
363 	if (!r->req.rsp)
364 		goto out;
365 
366 	r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
367 			sizeof(*r->req.rsp), DMA_TO_DEVICE);
368 	if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
369 		goto out_free_rsp;
370 
371 	r->send_sge.length = sizeof(*r->req.rsp);
372 	r->send_sge.lkey = ndev->pd->local_dma_lkey;
373 
374 	r->send_cqe.done = nvmet_rdma_send_done;
375 
376 	r->send_wr.wr_cqe = &r->send_cqe;
377 	r->send_wr.sg_list = &r->send_sge;
378 	r->send_wr.num_sge = 1;
379 	r->send_wr.send_flags = IB_SEND_SIGNALED;
380 
381 	/* Data In / RDMA READ */
382 	r->read_cqe.done = nvmet_rdma_read_data_done;
383 	return 0;
384 
385 out_free_rsp:
386 	kfree(r->req.rsp);
387 out:
388 	return -ENOMEM;
389 }
390 
391 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
392 		struct nvmet_rdma_rsp *r)
393 {
394 	ib_dma_unmap_single(ndev->device, r->send_sge.addr,
395 				sizeof(*r->req.rsp), DMA_TO_DEVICE);
396 	kfree(r->req.rsp);
397 }
398 
399 static int
400 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
401 {
402 	struct nvmet_rdma_device *ndev = queue->dev;
403 	int nr_rsps = queue->recv_queue_size * 2;
404 	int ret = -EINVAL, i;
405 
406 	queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
407 			GFP_KERNEL);
408 	if (!queue->rsps)
409 		goto out;
410 
411 	for (i = 0; i < nr_rsps; i++) {
412 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
413 
414 		ret = nvmet_rdma_alloc_rsp(ndev, rsp);
415 		if (ret)
416 			goto out_free;
417 
418 		list_add_tail(&rsp->free_list, &queue->free_rsps);
419 	}
420 
421 	return 0;
422 
423 out_free:
424 	while (--i >= 0) {
425 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
426 
427 		list_del(&rsp->free_list);
428 		nvmet_rdma_free_rsp(ndev, rsp);
429 	}
430 	kfree(queue->rsps);
431 out:
432 	return ret;
433 }
434 
435 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
436 {
437 	struct nvmet_rdma_device *ndev = queue->dev;
438 	int i, nr_rsps = queue->recv_queue_size * 2;
439 
440 	for (i = 0; i < nr_rsps; i++) {
441 		struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
442 
443 		list_del(&rsp->free_list);
444 		nvmet_rdma_free_rsp(ndev, rsp);
445 	}
446 	kfree(queue->rsps);
447 }
448 
449 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
450 		struct nvmet_rdma_cmd *cmd)
451 {
452 	int ret;
453 
454 	ib_dma_sync_single_for_device(ndev->device,
455 		cmd->sge[0].addr, cmd->sge[0].length,
456 		DMA_FROM_DEVICE);
457 
458 	if (ndev->srq)
459 		ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
460 	else
461 		ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
462 
463 	if (unlikely(ret))
464 		pr_err("post_recv cmd failed\n");
465 
466 	return ret;
467 }
468 
469 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
470 {
471 	spin_lock(&queue->rsp_wr_wait_lock);
472 	while (!list_empty(&queue->rsp_wr_wait_list)) {
473 		struct nvmet_rdma_rsp *rsp;
474 		bool ret;
475 
476 		rsp = list_entry(queue->rsp_wr_wait_list.next,
477 				struct nvmet_rdma_rsp, wait_list);
478 		list_del(&rsp->wait_list);
479 
480 		spin_unlock(&queue->rsp_wr_wait_lock);
481 		ret = nvmet_rdma_execute_command(rsp);
482 		spin_lock(&queue->rsp_wr_wait_lock);
483 
484 		if (!ret) {
485 			list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
486 			break;
487 		}
488 	}
489 	spin_unlock(&queue->rsp_wr_wait_lock);
490 }
491 
492 
493 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
494 {
495 	struct nvmet_rdma_queue *queue = rsp->queue;
496 
497 	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
498 
499 	if (rsp->n_rdma) {
500 		rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
501 				queue->cm_id->port_num, rsp->req.sg,
502 				rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
503 	}
504 
505 	if (rsp->req.sg != rsp->cmd->inline_sg)
506 		sgl_free(rsp->req.sg);
507 
508 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
509 		nvmet_rdma_process_wr_wait_list(queue);
510 
511 	nvmet_rdma_put_rsp(rsp);
512 }
513 
514 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
515 {
516 	if (queue->nvme_sq.ctrl) {
517 		nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
518 	} else {
519 		/*
520 		 * we didn't setup the controller yet in case
521 		 * of admin connect error, just disconnect and
522 		 * cleanup the queue
523 		 */
524 		nvmet_rdma_queue_disconnect(queue);
525 	}
526 }
527 
528 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
529 {
530 	struct nvmet_rdma_rsp *rsp =
531 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
532 
533 	nvmet_rdma_release_rsp(rsp);
534 
535 	if (unlikely(wc->status != IB_WC_SUCCESS &&
536 		     wc->status != IB_WC_WR_FLUSH_ERR)) {
537 		pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
538 			wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
539 		nvmet_rdma_error_comp(rsp->queue);
540 	}
541 }
542 
543 static void nvmet_rdma_queue_response(struct nvmet_req *req)
544 {
545 	struct nvmet_rdma_rsp *rsp =
546 		container_of(req, struct nvmet_rdma_rsp, req);
547 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
548 	struct ib_send_wr *first_wr;
549 
550 	if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
551 		rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
552 		rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
553 	} else {
554 		rsp->send_wr.opcode = IB_WR_SEND;
555 	}
556 
557 	if (nvmet_rdma_need_data_out(rsp))
558 		first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
559 				cm_id->port_num, NULL, &rsp->send_wr);
560 	else
561 		first_wr = &rsp->send_wr;
562 
563 	nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
564 
565 	ib_dma_sync_single_for_device(rsp->queue->dev->device,
566 		rsp->send_sge.addr, rsp->send_sge.length,
567 		DMA_TO_DEVICE);
568 
569 	if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
570 		pr_err("sending cmd response failed\n");
571 		nvmet_rdma_release_rsp(rsp);
572 	}
573 }
574 
575 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
576 {
577 	struct nvmet_rdma_rsp *rsp =
578 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
579 	struct nvmet_rdma_queue *queue = cq->cq_context;
580 
581 	WARN_ON(rsp->n_rdma <= 0);
582 	atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
583 	rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
584 			queue->cm_id->port_num, rsp->req.sg,
585 			rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
586 	rsp->n_rdma = 0;
587 
588 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
589 		nvmet_req_uninit(&rsp->req);
590 		nvmet_rdma_release_rsp(rsp);
591 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
592 			pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
593 				wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
594 			nvmet_rdma_error_comp(queue);
595 		}
596 		return;
597 	}
598 
599 	nvmet_req_execute(&rsp->req);
600 }
601 
602 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
603 		u64 off)
604 {
605 	int sg_count = num_pages(len);
606 	struct scatterlist *sg;
607 	int i;
608 
609 	sg = rsp->cmd->inline_sg;
610 	for (i = 0; i < sg_count; i++, sg++) {
611 		if (i < sg_count - 1)
612 			sg_unmark_end(sg);
613 		else
614 			sg_mark_end(sg);
615 		sg->offset = off;
616 		sg->length = min_t(int, len, PAGE_SIZE - off);
617 		len -= sg->length;
618 		if (!i)
619 			off = 0;
620 	}
621 
622 	rsp->req.sg = rsp->cmd->inline_sg;
623 	rsp->req.sg_cnt = sg_count;
624 }
625 
626 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
627 {
628 	struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
629 	u64 off = le64_to_cpu(sgl->addr);
630 	u32 len = le32_to_cpu(sgl->length);
631 
632 	if (!nvme_is_write(rsp->req.cmd))
633 		return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
634 
635 	if (off + len > rsp->queue->dev->inline_data_size) {
636 		pr_err("invalid inline data offset!\n");
637 		return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
638 	}
639 
640 	/* no data command? */
641 	if (!len)
642 		return 0;
643 
644 	nvmet_rdma_use_inline_sg(rsp, len, off);
645 	rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
646 	rsp->req.transfer_len += len;
647 	return 0;
648 }
649 
650 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
651 		struct nvme_keyed_sgl_desc *sgl, bool invalidate)
652 {
653 	struct rdma_cm_id *cm_id = rsp->queue->cm_id;
654 	u64 addr = le64_to_cpu(sgl->addr);
655 	u32 len = get_unaligned_le24(sgl->length);
656 	u32 key = get_unaligned_le32(sgl->key);
657 	int ret;
658 
659 	/* no data command? */
660 	if (!len)
661 		return 0;
662 
663 	rsp->req.sg = sgl_alloc(len, GFP_KERNEL, &rsp->req.sg_cnt);
664 	if (!rsp->req.sg)
665 		return NVME_SC_INTERNAL;
666 
667 	ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
668 			rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
669 			nvmet_data_dir(&rsp->req));
670 	if (ret < 0)
671 		return NVME_SC_INTERNAL;
672 	rsp->req.transfer_len += len;
673 	rsp->n_rdma += ret;
674 
675 	if (invalidate) {
676 		rsp->invalidate_rkey = key;
677 		rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
678 	}
679 
680 	return 0;
681 }
682 
683 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
684 {
685 	struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
686 
687 	switch (sgl->type >> 4) {
688 	case NVME_SGL_FMT_DATA_DESC:
689 		switch (sgl->type & 0xf) {
690 		case NVME_SGL_FMT_OFFSET:
691 			return nvmet_rdma_map_sgl_inline(rsp);
692 		default:
693 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
694 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
695 		}
696 	case NVME_KEY_SGL_FMT_DATA_DESC:
697 		switch (sgl->type & 0xf) {
698 		case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
699 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
700 		case NVME_SGL_FMT_ADDRESS:
701 			return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
702 		default:
703 			pr_err("invalid SGL subtype: %#x\n", sgl->type);
704 			return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
705 		}
706 	default:
707 		pr_err("invalid SGL type: %#x\n", sgl->type);
708 		return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
709 	}
710 }
711 
712 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
713 {
714 	struct nvmet_rdma_queue *queue = rsp->queue;
715 
716 	if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
717 			&queue->sq_wr_avail) < 0)) {
718 		pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
719 				1 + rsp->n_rdma, queue->idx,
720 				queue->nvme_sq.ctrl->cntlid);
721 		atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
722 		return false;
723 	}
724 
725 	if (nvmet_rdma_need_data_in(rsp)) {
726 		if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
727 				queue->cm_id->port_num, &rsp->read_cqe, NULL))
728 			nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
729 	} else {
730 		nvmet_req_execute(&rsp->req);
731 	}
732 
733 	return true;
734 }
735 
736 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
737 		struct nvmet_rdma_rsp *cmd)
738 {
739 	u16 status;
740 
741 	ib_dma_sync_single_for_cpu(queue->dev->device,
742 		cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
743 		DMA_FROM_DEVICE);
744 	ib_dma_sync_single_for_cpu(queue->dev->device,
745 		cmd->send_sge.addr, cmd->send_sge.length,
746 		DMA_TO_DEVICE);
747 
748 	if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
749 			&queue->nvme_sq, &nvmet_rdma_ops))
750 		return;
751 
752 	status = nvmet_rdma_map_sgl(cmd);
753 	if (status)
754 		goto out_err;
755 
756 	if (unlikely(!nvmet_rdma_execute_command(cmd))) {
757 		spin_lock(&queue->rsp_wr_wait_lock);
758 		list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
759 		spin_unlock(&queue->rsp_wr_wait_lock);
760 	}
761 
762 	return;
763 
764 out_err:
765 	nvmet_req_complete(&cmd->req, status);
766 }
767 
768 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
769 {
770 	struct nvmet_rdma_cmd *cmd =
771 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
772 	struct nvmet_rdma_queue *queue = cq->cq_context;
773 	struct nvmet_rdma_rsp *rsp;
774 
775 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
776 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
777 			pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
778 				wc->wr_cqe, ib_wc_status_msg(wc->status),
779 				wc->status);
780 			nvmet_rdma_error_comp(queue);
781 		}
782 		return;
783 	}
784 
785 	if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
786 		pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
787 		nvmet_rdma_error_comp(queue);
788 		return;
789 	}
790 
791 	cmd->queue = queue;
792 	rsp = nvmet_rdma_get_rsp(queue);
793 	if (unlikely(!rsp)) {
794 		/*
795 		 * we get here only under memory pressure,
796 		 * silently drop and have the host retry
797 		 * as we can't even fail it.
798 		 */
799 		nvmet_rdma_post_recv(queue->dev, cmd);
800 		return;
801 	}
802 	rsp->queue = queue;
803 	rsp->cmd = cmd;
804 	rsp->flags = 0;
805 	rsp->req.cmd = cmd->nvme_cmd;
806 	rsp->req.port = queue->port;
807 	rsp->n_rdma = 0;
808 
809 	if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
810 		unsigned long flags;
811 
812 		spin_lock_irqsave(&queue->state_lock, flags);
813 		if (queue->state == NVMET_RDMA_Q_CONNECTING)
814 			list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
815 		else
816 			nvmet_rdma_put_rsp(rsp);
817 		spin_unlock_irqrestore(&queue->state_lock, flags);
818 		return;
819 	}
820 
821 	nvmet_rdma_handle_command(queue, rsp);
822 }
823 
824 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
825 {
826 	if (!ndev->srq)
827 		return;
828 
829 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
830 	ib_destroy_srq(ndev->srq);
831 }
832 
833 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
834 {
835 	struct ib_srq_init_attr srq_attr = { NULL, };
836 	struct ib_srq *srq;
837 	size_t srq_size;
838 	int ret, i;
839 
840 	srq_size = 4095;	/* XXX: tune */
841 
842 	srq_attr.attr.max_wr = srq_size;
843 	srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
844 	srq_attr.attr.srq_limit = 0;
845 	srq_attr.srq_type = IB_SRQT_BASIC;
846 	srq = ib_create_srq(ndev->pd, &srq_attr);
847 	if (IS_ERR(srq)) {
848 		/*
849 		 * If SRQs aren't supported we just go ahead and use normal
850 		 * non-shared receive queues.
851 		 */
852 		pr_info("SRQ requested but not supported.\n");
853 		return 0;
854 	}
855 
856 	ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
857 	if (IS_ERR(ndev->srq_cmds)) {
858 		ret = PTR_ERR(ndev->srq_cmds);
859 		goto out_destroy_srq;
860 	}
861 
862 	ndev->srq = srq;
863 	ndev->srq_size = srq_size;
864 
865 	for (i = 0; i < srq_size; i++) {
866 		ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
867 		if (ret)
868 			goto out_free_cmds;
869 	}
870 
871 	return 0;
872 
873 out_free_cmds:
874 	nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
875 out_destroy_srq:
876 	ib_destroy_srq(srq);
877 	return ret;
878 }
879 
880 static void nvmet_rdma_free_dev(struct kref *ref)
881 {
882 	struct nvmet_rdma_device *ndev =
883 		container_of(ref, struct nvmet_rdma_device, ref);
884 
885 	mutex_lock(&device_list_mutex);
886 	list_del(&ndev->entry);
887 	mutex_unlock(&device_list_mutex);
888 
889 	nvmet_rdma_destroy_srq(ndev);
890 	ib_dealloc_pd(ndev->pd);
891 
892 	kfree(ndev);
893 }
894 
895 static struct nvmet_rdma_device *
896 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
897 {
898 	struct nvmet_port *port = cm_id->context;
899 	struct nvmet_rdma_device *ndev;
900 	int inline_page_count;
901 	int inline_sge_count;
902 	int ret;
903 
904 	mutex_lock(&device_list_mutex);
905 	list_for_each_entry(ndev, &device_list, entry) {
906 		if (ndev->device->node_guid == cm_id->device->node_guid &&
907 		    kref_get_unless_zero(&ndev->ref))
908 			goto out_unlock;
909 	}
910 
911 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
912 	if (!ndev)
913 		goto out_err;
914 
915 	inline_page_count = num_pages(port->inline_data_size);
916 	inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
917 				cm_id->device->attrs.max_recv_sge) - 1;
918 	if (inline_page_count > inline_sge_count) {
919 		pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
920 			port->inline_data_size, cm_id->device->name,
921 			inline_sge_count * PAGE_SIZE);
922 		port->inline_data_size = inline_sge_count * PAGE_SIZE;
923 		inline_page_count = inline_sge_count;
924 	}
925 	ndev->inline_data_size = port->inline_data_size;
926 	ndev->inline_page_count = inline_page_count;
927 	ndev->device = cm_id->device;
928 	kref_init(&ndev->ref);
929 
930 	ndev->pd = ib_alloc_pd(ndev->device, 0);
931 	if (IS_ERR(ndev->pd))
932 		goto out_free_dev;
933 
934 	if (nvmet_rdma_use_srq) {
935 		ret = nvmet_rdma_init_srq(ndev);
936 		if (ret)
937 			goto out_free_pd;
938 	}
939 
940 	list_add(&ndev->entry, &device_list);
941 out_unlock:
942 	mutex_unlock(&device_list_mutex);
943 	pr_debug("added %s.\n", ndev->device->name);
944 	return ndev;
945 
946 out_free_pd:
947 	ib_dealloc_pd(ndev->pd);
948 out_free_dev:
949 	kfree(ndev);
950 out_err:
951 	mutex_unlock(&device_list_mutex);
952 	return NULL;
953 }
954 
955 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
956 {
957 	struct ib_qp_init_attr qp_attr;
958 	struct nvmet_rdma_device *ndev = queue->dev;
959 	int comp_vector, nr_cqe, ret, i;
960 
961 	/*
962 	 * Spread the io queues across completion vectors,
963 	 * but still keep all admin queues on vector 0.
964 	 */
965 	comp_vector = !queue->host_qid ? 0 :
966 		queue->idx % ndev->device->num_comp_vectors;
967 
968 	/*
969 	 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
970 	 */
971 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
972 
973 	queue->cq = ib_alloc_cq(ndev->device, queue,
974 			nr_cqe + 1, comp_vector,
975 			IB_POLL_WORKQUEUE);
976 	if (IS_ERR(queue->cq)) {
977 		ret = PTR_ERR(queue->cq);
978 		pr_err("failed to create CQ cqe= %d ret= %d\n",
979 		       nr_cqe + 1, ret);
980 		goto out;
981 	}
982 
983 	memset(&qp_attr, 0, sizeof(qp_attr));
984 	qp_attr.qp_context = queue;
985 	qp_attr.event_handler = nvmet_rdma_qp_event;
986 	qp_attr.send_cq = queue->cq;
987 	qp_attr.recv_cq = queue->cq;
988 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
989 	qp_attr.qp_type = IB_QPT_RC;
990 	/* +1 for drain */
991 	qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
992 	qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
993 	qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
994 					ndev->device->attrs.max_send_sge);
995 
996 	if (ndev->srq) {
997 		qp_attr.srq = ndev->srq;
998 	} else {
999 		/* +1 for drain */
1000 		qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1001 		qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1002 	}
1003 
1004 	ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1005 	if (ret) {
1006 		pr_err("failed to create_qp ret= %d\n", ret);
1007 		goto err_destroy_cq;
1008 	}
1009 
1010 	atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1011 
1012 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1013 		 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1014 		 qp_attr.cap.max_send_wr, queue->cm_id);
1015 
1016 	if (!ndev->srq) {
1017 		for (i = 0; i < queue->recv_queue_size; i++) {
1018 			queue->cmds[i].queue = queue;
1019 			ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1020 			if (ret)
1021 				goto err_destroy_qp;
1022 		}
1023 	}
1024 
1025 out:
1026 	return ret;
1027 
1028 err_destroy_qp:
1029 	rdma_destroy_qp(queue->cm_id);
1030 err_destroy_cq:
1031 	ib_free_cq(queue->cq);
1032 	goto out;
1033 }
1034 
1035 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1036 {
1037 	struct ib_qp *qp = queue->cm_id->qp;
1038 
1039 	ib_drain_qp(qp);
1040 	rdma_destroy_id(queue->cm_id);
1041 	ib_destroy_qp(qp);
1042 	ib_free_cq(queue->cq);
1043 }
1044 
1045 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1046 {
1047 	pr_debug("freeing queue %d\n", queue->idx);
1048 
1049 	nvmet_sq_destroy(&queue->nvme_sq);
1050 
1051 	nvmet_rdma_destroy_queue_ib(queue);
1052 	if (!queue->dev->srq) {
1053 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1054 				queue->recv_queue_size,
1055 				!queue->host_qid);
1056 	}
1057 	nvmet_rdma_free_rsps(queue);
1058 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1059 	kfree(queue);
1060 }
1061 
1062 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1063 {
1064 	struct nvmet_rdma_queue *queue =
1065 		container_of(w, struct nvmet_rdma_queue, release_work);
1066 	struct nvmet_rdma_device *dev = queue->dev;
1067 
1068 	nvmet_rdma_free_queue(queue);
1069 
1070 	kref_put(&dev->ref, nvmet_rdma_free_dev);
1071 }
1072 
1073 static int
1074 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1075 				struct nvmet_rdma_queue *queue)
1076 {
1077 	struct nvme_rdma_cm_req *req;
1078 
1079 	req = (struct nvme_rdma_cm_req *)conn->private_data;
1080 	if (!req || conn->private_data_len == 0)
1081 		return NVME_RDMA_CM_INVALID_LEN;
1082 
1083 	if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1084 		return NVME_RDMA_CM_INVALID_RECFMT;
1085 
1086 	queue->host_qid = le16_to_cpu(req->qid);
1087 
1088 	/*
1089 	 * req->hsqsize corresponds to our recv queue size plus 1
1090 	 * req->hrqsize corresponds to our send queue size
1091 	 */
1092 	queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1093 	queue->send_queue_size = le16_to_cpu(req->hrqsize);
1094 
1095 	if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1096 		return NVME_RDMA_CM_INVALID_HSQSIZE;
1097 
1098 	/* XXX: Should we enforce some kind of max for IO queues? */
1099 
1100 	return 0;
1101 }
1102 
1103 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1104 				enum nvme_rdma_cm_status status)
1105 {
1106 	struct nvme_rdma_cm_rej rej;
1107 
1108 	pr_debug("rejecting connect request: status %d (%s)\n",
1109 		 status, nvme_rdma_cm_msg(status));
1110 
1111 	rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1112 	rej.sts = cpu_to_le16(status);
1113 
1114 	return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1115 }
1116 
1117 static struct nvmet_rdma_queue *
1118 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1119 		struct rdma_cm_id *cm_id,
1120 		struct rdma_cm_event *event)
1121 {
1122 	struct nvmet_rdma_queue *queue;
1123 	int ret;
1124 
1125 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1126 	if (!queue) {
1127 		ret = NVME_RDMA_CM_NO_RSC;
1128 		goto out_reject;
1129 	}
1130 
1131 	ret = nvmet_sq_init(&queue->nvme_sq);
1132 	if (ret) {
1133 		ret = NVME_RDMA_CM_NO_RSC;
1134 		goto out_free_queue;
1135 	}
1136 
1137 	ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1138 	if (ret)
1139 		goto out_destroy_sq;
1140 
1141 	/*
1142 	 * Schedules the actual release because calling rdma_destroy_id from
1143 	 * inside a CM callback would trigger a deadlock. (great API design..)
1144 	 */
1145 	INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1146 	queue->dev = ndev;
1147 	queue->cm_id = cm_id;
1148 
1149 	spin_lock_init(&queue->state_lock);
1150 	queue->state = NVMET_RDMA_Q_CONNECTING;
1151 	INIT_LIST_HEAD(&queue->rsp_wait_list);
1152 	INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1153 	spin_lock_init(&queue->rsp_wr_wait_lock);
1154 	INIT_LIST_HEAD(&queue->free_rsps);
1155 	spin_lock_init(&queue->rsps_lock);
1156 	INIT_LIST_HEAD(&queue->queue_list);
1157 
1158 	queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1159 	if (queue->idx < 0) {
1160 		ret = NVME_RDMA_CM_NO_RSC;
1161 		goto out_destroy_sq;
1162 	}
1163 
1164 	ret = nvmet_rdma_alloc_rsps(queue);
1165 	if (ret) {
1166 		ret = NVME_RDMA_CM_NO_RSC;
1167 		goto out_ida_remove;
1168 	}
1169 
1170 	if (!ndev->srq) {
1171 		queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1172 				queue->recv_queue_size,
1173 				!queue->host_qid);
1174 		if (IS_ERR(queue->cmds)) {
1175 			ret = NVME_RDMA_CM_NO_RSC;
1176 			goto out_free_responses;
1177 		}
1178 	}
1179 
1180 	ret = nvmet_rdma_create_queue_ib(queue);
1181 	if (ret) {
1182 		pr_err("%s: creating RDMA queue failed (%d).\n",
1183 			__func__, ret);
1184 		ret = NVME_RDMA_CM_NO_RSC;
1185 		goto out_free_cmds;
1186 	}
1187 
1188 	return queue;
1189 
1190 out_free_cmds:
1191 	if (!ndev->srq) {
1192 		nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1193 				queue->recv_queue_size,
1194 				!queue->host_qid);
1195 	}
1196 out_free_responses:
1197 	nvmet_rdma_free_rsps(queue);
1198 out_ida_remove:
1199 	ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1200 out_destroy_sq:
1201 	nvmet_sq_destroy(&queue->nvme_sq);
1202 out_free_queue:
1203 	kfree(queue);
1204 out_reject:
1205 	nvmet_rdma_cm_reject(cm_id, ret);
1206 	return NULL;
1207 }
1208 
1209 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1210 {
1211 	struct nvmet_rdma_queue *queue = priv;
1212 
1213 	switch (event->event) {
1214 	case IB_EVENT_COMM_EST:
1215 		rdma_notify(queue->cm_id, event->event);
1216 		break;
1217 	default:
1218 		pr_err("received IB QP event: %s (%d)\n",
1219 		       ib_event_msg(event->event), event->event);
1220 		break;
1221 	}
1222 }
1223 
1224 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1225 		struct nvmet_rdma_queue *queue,
1226 		struct rdma_conn_param *p)
1227 {
1228 	struct rdma_conn_param  param = { };
1229 	struct nvme_rdma_cm_rep priv = { };
1230 	int ret = -ENOMEM;
1231 
1232 	param.rnr_retry_count = 7;
1233 	param.flow_control = 1;
1234 	param.initiator_depth = min_t(u8, p->initiator_depth,
1235 		queue->dev->device->attrs.max_qp_init_rd_atom);
1236 	param.private_data = &priv;
1237 	param.private_data_len = sizeof(priv);
1238 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1239 	priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1240 
1241 	ret = rdma_accept(cm_id, &param);
1242 	if (ret)
1243 		pr_err("rdma_accept failed (error code = %d)\n", ret);
1244 
1245 	return ret;
1246 }
1247 
1248 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1249 		struct rdma_cm_event *event)
1250 {
1251 	struct nvmet_rdma_device *ndev;
1252 	struct nvmet_rdma_queue *queue;
1253 	int ret = -EINVAL;
1254 
1255 	ndev = nvmet_rdma_find_get_device(cm_id);
1256 	if (!ndev) {
1257 		nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1258 		return -ECONNREFUSED;
1259 	}
1260 
1261 	queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1262 	if (!queue) {
1263 		ret = -ENOMEM;
1264 		goto put_device;
1265 	}
1266 	queue->port = cm_id->context;
1267 
1268 	if (queue->host_qid == 0) {
1269 		/* Let inflight controller teardown complete */
1270 		flush_scheduled_work();
1271 	}
1272 
1273 	ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1274 	if (ret) {
1275 		schedule_work(&queue->release_work);
1276 		/* Destroying rdma_cm id is not needed here */
1277 		return 0;
1278 	}
1279 
1280 	mutex_lock(&nvmet_rdma_queue_mutex);
1281 	list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1282 	mutex_unlock(&nvmet_rdma_queue_mutex);
1283 
1284 	return 0;
1285 
1286 put_device:
1287 	kref_put(&ndev->ref, nvmet_rdma_free_dev);
1288 
1289 	return ret;
1290 }
1291 
1292 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1293 {
1294 	unsigned long flags;
1295 
1296 	spin_lock_irqsave(&queue->state_lock, flags);
1297 	if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1298 		pr_warn("trying to establish a connected queue\n");
1299 		goto out_unlock;
1300 	}
1301 	queue->state = NVMET_RDMA_Q_LIVE;
1302 
1303 	while (!list_empty(&queue->rsp_wait_list)) {
1304 		struct nvmet_rdma_rsp *cmd;
1305 
1306 		cmd = list_first_entry(&queue->rsp_wait_list,
1307 					struct nvmet_rdma_rsp, wait_list);
1308 		list_del(&cmd->wait_list);
1309 
1310 		spin_unlock_irqrestore(&queue->state_lock, flags);
1311 		nvmet_rdma_handle_command(queue, cmd);
1312 		spin_lock_irqsave(&queue->state_lock, flags);
1313 	}
1314 
1315 out_unlock:
1316 	spin_unlock_irqrestore(&queue->state_lock, flags);
1317 }
1318 
1319 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1320 {
1321 	bool disconnect = false;
1322 	unsigned long flags;
1323 
1324 	pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1325 
1326 	spin_lock_irqsave(&queue->state_lock, flags);
1327 	switch (queue->state) {
1328 	case NVMET_RDMA_Q_CONNECTING:
1329 	case NVMET_RDMA_Q_LIVE:
1330 		queue->state = NVMET_RDMA_Q_DISCONNECTING;
1331 		disconnect = true;
1332 		break;
1333 	case NVMET_RDMA_Q_DISCONNECTING:
1334 		break;
1335 	}
1336 	spin_unlock_irqrestore(&queue->state_lock, flags);
1337 
1338 	if (disconnect) {
1339 		rdma_disconnect(queue->cm_id);
1340 		schedule_work(&queue->release_work);
1341 	}
1342 }
1343 
1344 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1345 {
1346 	bool disconnect = false;
1347 
1348 	mutex_lock(&nvmet_rdma_queue_mutex);
1349 	if (!list_empty(&queue->queue_list)) {
1350 		list_del_init(&queue->queue_list);
1351 		disconnect = true;
1352 	}
1353 	mutex_unlock(&nvmet_rdma_queue_mutex);
1354 
1355 	if (disconnect)
1356 		__nvmet_rdma_queue_disconnect(queue);
1357 }
1358 
1359 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1360 		struct nvmet_rdma_queue *queue)
1361 {
1362 	WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1363 
1364 	mutex_lock(&nvmet_rdma_queue_mutex);
1365 	if (!list_empty(&queue->queue_list))
1366 		list_del_init(&queue->queue_list);
1367 	mutex_unlock(&nvmet_rdma_queue_mutex);
1368 
1369 	pr_err("failed to connect queue %d\n", queue->idx);
1370 	schedule_work(&queue->release_work);
1371 }
1372 
1373 /**
1374  * nvme_rdma_device_removal() - Handle RDMA device removal
1375  * @cm_id:	rdma_cm id, used for nvmet port
1376  * @queue:      nvmet rdma queue (cm id qp_context)
1377  *
1378  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1379  * to unplug. Note that this event can be generated on a normal
1380  * queue cm_id and/or a device bound listener cm_id (where in this
1381  * case queue will be null).
1382  *
1383  * We registered an ib_client to handle device removal for queues,
1384  * so we only need to handle the listening port cm_ids. In this case
1385  * we nullify the priv to prevent double cm_id destruction and destroying
1386  * the cm_id implicitely by returning a non-zero rc to the callout.
1387  */
1388 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1389 		struct nvmet_rdma_queue *queue)
1390 {
1391 	struct nvmet_port *port;
1392 
1393 	if (queue) {
1394 		/*
1395 		 * This is a queue cm_id. we have registered
1396 		 * an ib_client to handle queues removal
1397 		 * so don't interfear and just return.
1398 		 */
1399 		return 0;
1400 	}
1401 
1402 	port = cm_id->context;
1403 
1404 	/*
1405 	 * This is a listener cm_id. Make sure that
1406 	 * future remove_port won't invoke a double
1407 	 * cm_id destroy. use atomic xchg to make sure
1408 	 * we don't compete with remove_port.
1409 	 */
1410 	if (xchg(&port->priv, NULL) != cm_id)
1411 		return 0;
1412 
1413 	/*
1414 	 * We need to return 1 so that the core will destroy
1415 	 * it's own ID.  What a great API design..
1416 	 */
1417 	return 1;
1418 }
1419 
1420 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1421 		struct rdma_cm_event *event)
1422 {
1423 	struct nvmet_rdma_queue *queue = NULL;
1424 	int ret = 0;
1425 
1426 	if (cm_id->qp)
1427 		queue = cm_id->qp->qp_context;
1428 
1429 	pr_debug("%s (%d): status %d id %p\n",
1430 		rdma_event_msg(event->event), event->event,
1431 		event->status, cm_id);
1432 
1433 	switch (event->event) {
1434 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1435 		ret = nvmet_rdma_queue_connect(cm_id, event);
1436 		break;
1437 	case RDMA_CM_EVENT_ESTABLISHED:
1438 		nvmet_rdma_queue_established(queue);
1439 		break;
1440 	case RDMA_CM_EVENT_ADDR_CHANGE:
1441 	case RDMA_CM_EVENT_DISCONNECTED:
1442 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1443 		nvmet_rdma_queue_disconnect(queue);
1444 		break;
1445 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1446 		ret = nvmet_rdma_device_removal(cm_id, queue);
1447 		break;
1448 	case RDMA_CM_EVENT_REJECTED:
1449 		pr_debug("Connection rejected: %s\n",
1450 			 rdma_reject_msg(cm_id, event->status));
1451 		/* FALLTHROUGH */
1452 	case RDMA_CM_EVENT_UNREACHABLE:
1453 	case RDMA_CM_EVENT_CONNECT_ERROR:
1454 		nvmet_rdma_queue_connect_fail(cm_id, queue);
1455 		break;
1456 	default:
1457 		pr_err("received unrecognized RDMA CM event %d\n",
1458 			event->event);
1459 		break;
1460 	}
1461 
1462 	return ret;
1463 }
1464 
1465 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1466 {
1467 	struct nvmet_rdma_queue *queue;
1468 
1469 restart:
1470 	mutex_lock(&nvmet_rdma_queue_mutex);
1471 	list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1472 		if (queue->nvme_sq.ctrl == ctrl) {
1473 			list_del_init(&queue->queue_list);
1474 			mutex_unlock(&nvmet_rdma_queue_mutex);
1475 
1476 			__nvmet_rdma_queue_disconnect(queue);
1477 			goto restart;
1478 		}
1479 	}
1480 	mutex_unlock(&nvmet_rdma_queue_mutex);
1481 }
1482 
1483 static int nvmet_rdma_add_port(struct nvmet_port *port)
1484 {
1485 	struct rdma_cm_id *cm_id;
1486 	struct sockaddr_storage addr = { };
1487 	__kernel_sa_family_t af;
1488 	int ret;
1489 
1490 	switch (port->disc_addr.adrfam) {
1491 	case NVMF_ADDR_FAMILY_IP4:
1492 		af = AF_INET;
1493 		break;
1494 	case NVMF_ADDR_FAMILY_IP6:
1495 		af = AF_INET6;
1496 		break;
1497 	default:
1498 		pr_err("address family %d not supported\n",
1499 				port->disc_addr.adrfam);
1500 		return -EINVAL;
1501 	}
1502 
1503 	if (port->inline_data_size < 0) {
1504 		port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1505 	} else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1506 		pr_warn("inline_data_size %u is too large, reducing to %u\n",
1507 			port->inline_data_size,
1508 			NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1509 		port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1510 	}
1511 
1512 	ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1513 			port->disc_addr.trsvcid, &addr);
1514 	if (ret) {
1515 		pr_err("malformed ip/port passed: %s:%s\n",
1516 			port->disc_addr.traddr, port->disc_addr.trsvcid);
1517 		return ret;
1518 	}
1519 
1520 	cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1521 			RDMA_PS_TCP, IB_QPT_RC);
1522 	if (IS_ERR(cm_id)) {
1523 		pr_err("CM ID creation failed\n");
1524 		return PTR_ERR(cm_id);
1525 	}
1526 
1527 	/*
1528 	 * Allow both IPv4 and IPv6 sockets to bind a single port
1529 	 * at the same time.
1530 	 */
1531 	ret = rdma_set_afonly(cm_id, 1);
1532 	if (ret) {
1533 		pr_err("rdma_set_afonly failed (%d)\n", ret);
1534 		goto out_destroy_id;
1535 	}
1536 
1537 	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1538 	if (ret) {
1539 		pr_err("binding CM ID to %pISpcs failed (%d)\n",
1540 			(struct sockaddr *)&addr, ret);
1541 		goto out_destroy_id;
1542 	}
1543 
1544 	ret = rdma_listen(cm_id, 128);
1545 	if (ret) {
1546 		pr_err("listening to %pISpcs failed (%d)\n",
1547 			(struct sockaddr *)&addr, ret);
1548 		goto out_destroy_id;
1549 	}
1550 
1551 	pr_info("enabling port %d (%pISpcs)\n",
1552 		le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1553 	port->priv = cm_id;
1554 	return 0;
1555 
1556 out_destroy_id:
1557 	rdma_destroy_id(cm_id);
1558 	return ret;
1559 }
1560 
1561 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1562 {
1563 	struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1564 
1565 	if (cm_id)
1566 		rdma_destroy_id(cm_id);
1567 }
1568 
1569 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1570 		struct nvmet_port *port, char *traddr)
1571 {
1572 	struct rdma_cm_id *cm_id = port->priv;
1573 
1574 	if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1575 		struct nvmet_rdma_rsp *rsp =
1576 			container_of(req, struct nvmet_rdma_rsp, req);
1577 		struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1578 		struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1579 
1580 		sprintf(traddr, "%pISc", addr);
1581 	} else {
1582 		memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1583 	}
1584 }
1585 
1586 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1587 	.owner			= THIS_MODULE,
1588 	.type			= NVMF_TRTYPE_RDMA,
1589 	.msdbd			= 1,
1590 	.has_keyed_sgls		= 1,
1591 	.add_port		= nvmet_rdma_add_port,
1592 	.remove_port		= nvmet_rdma_remove_port,
1593 	.queue_response		= nvmet_rdma_queue_response,
1594 	.delete_ctrl		= nvmet_rdma_delete_ctrl,
1595 	.disc_traddr		= nvmet_rdma_disc_port_addr,
1596 };
1597 
1598 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1599 {
1600 	struct nvmet_rdma_queue *queue, *tmp;
1601 	struct nvmet_rdma_device *ndev;
1602 	bool found = false;
1603 
1604 	mutex_lock(&device_list_mutex);
1605 	list_for_each_entry(ndev, &device_list, entry) {
1606 		if (ndev->device == ib_device) {
1607 			found = true;
1608 			break;
1609 		}
1610 	}
1611 	mutex_unlock(&device_list_mutex);
1612 
1613 	if (!found)
1614 		return;
1615 
1616 	/*
1617 	 * IB Device that is used by nvmet controllers is being removed,
1618 	 * delete all queues using this device.
1619 	 */
1620 	mutex_lock(&nvmet_rdma_queue_mutex);
1621 	list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1622 				 queue_list) {
1623 		if (queue->dev->device != ib_device)
1624 			continue;
1625 
1626 		pr_info("Removing queue %d\n", queue->idx);
1627 		list_del_init(&queue->queue_list);
1628 		__nvmet_rdma_queue_disconnect(queue);
1629 	}
1630 	mutex_unlock(&nvmet_rdma_queue_mutex);
1631 
1632 	flush_scheduled_work();
1633 }
1634 
1635 static struct ib_client nvmet_rdma_ib_client = {
1636 	.name   = "nvmet_rdma",
1637 	.remove = nvmet_rdma_remove_one
1638 };
1639 
1640 static int __init nvmet_rdma_init(void)
1641 {
1642 	int ret;
1643 
1644 	ret = ib_register_client(&nvmet_rdma_ib_client);
1645 	if (ret)
1646 		return ret;
1647 
1648 	ret = nvmet_register_transport(&nvmet_rdma_ops);
1649 	if (ret)
1650 		goto err_ib_client;
1651 
1652 	return 0;
1653 
1654 err_ib_client:
1655 	ib_unregister_client(&nvmet_rdma_ib_client);
1656 	return ret;
1657 }
1658 
1659 static void __exit nvmet_rdma_exit(void)
1660 {
1661 	nvmet_unregister_transport(&nvmet_rdma_ops);
1662 	ib_unregister_client(&nvmet_rdma_ib_client);
1663 	WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1664 	ida_destroy(&nvmet_rdma_queue_ida);
1665 }
1666 
1667 module_init(nvmet_rdma_init);
1668 module_exit(nvmet_rdma_exit);
1669 
1670 MODULE_LICENSE("GPL v2");
1671 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */
1672