xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 42bc47b3)
1 /*
2  * NVMe over Fabrics loopback device.
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/scatterlist.h>
16 #include <linux/blk-mq.h>
17 #include <linux/nvme.h>
18 #include <linux/module.h>
19 #include <linux/parser.h>
20 #include "nvmet.h"
21 #include "../host/nvme.h"
22 #include "../host/fabrics.h"
23 
24 #define NVME_LOOP_MAX_SEGMENTS		256
25 
26 struct nvme_loop_iod {
27 	struct nvme_request	nvme_req;
28 	struct nvme_command	cmd;
29 	struct nvme_completion	rsp;
30 	struct nvmet_req	req;
31 	struct nvme_loop_queue	*queue;
32 	struct work_struct	work;
33 	struct sg_table		sg_table;
34 	struct scatterlist	first_sgl[];
35 };
36 
37 struct nvme_loop_ctrl {
38 	struct nvme_loop_queue	*queues;
39 
40 	struct blk_mq_tag_set	admin_tag_set;
41 
42 	struct list_head	list;
43 	struct blk_mq_tag_set	tag_set;
44 	struct nvme_loop_iod	async_event_iod;
45 	struct nvme_ctrl	ctrl;
46 
47 	struct nvmet_ctrl	*target_ctrl;
48 	struct nvmet_port	*port;
49 };
50 
51 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
52 {
53 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
54 }
55 
56 enum nvme_loop_queue_flags {
57 	NVME_LOOP_Q_LIVE	= 0,
58 };
59 
60 struct nvme_loop_queue {
61 	struct nvmet_cq		nvme_cq;
62 	struct nvmet_sq		nvme_sq;
63 	struct nvme_loop_ctrl	*ctrl;
64 	unsigned long		flags;
65 };
66 
67 static LIST_HEAD(nvme_loop_ports);
68 static DEFINE_MUTEX(nvme_loop_ports_mutex);
69 
70 static LIST_HEAD(nvme_loop_ctrl_list);
71 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
72 
73 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
74 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
75 
76 static const struct nvmet_fabrics_ops nvme_loop_ops;
77 
78 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
79 {
80 	return queue - queue->ctrl->queues;
81 }
82 
83 static void nvme_loop_complete_rq(struct request *req)
84 {
85 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
86 
87 	nvme_cleanup_cmd(req);
88 	sg_free_table_chained(&iod->sg_table, true);
89 	nvme_complete_rq(req);
90 }
91 
92 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
93 {
94 	u32 queue_idx = nvme_loop_queue_idx(queue);
95 
96 	if (queue_idx == 0)
97 		return queue->ctrl->admin_tag_set.tags[queue_idx];
98 	return queue->ctrl->tag_set.tags[queue_idx - 1];
99 }
100 
101 static void nvme_loop_queue_response(struct nvmet_req *req)
102 {
103 	struct nvme_loop_queue *queue =
104 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
105 	struct nvme_completion *cqe = req->rsp;
106 
107 	/*
108 	 * AEN requests are special as they don't time out and can
109 	 * survive any kind of queue freeze and often don't respond to
110 	 * aborts.  We don't even bother to allocate a struct request
111 	 * for them but rather special case them here.
112 	 */
113 	if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
114 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
115 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
116 				&cqe->result);
117 	} else {
118 		struct request *rq;
119 
120 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
121 		if (!rq) {
122 			dev_err(queue->ctrl->ctrl.device,
123 				"tag 0x%x on queue %d not found\n",
124 				cqe->command_id, nvme_loop_queue_idx(queue));
125 			return;
126 		}
127 
128 		nvme_end_request(rq, cqe->status, cqe->result);
129 	}
130 }
131 
132 static void nvme_loop_execute_work(struct work_struct *work)
133 {
134 	struct nvme_loop_iod *iod =
135 		container_of(work, struct nvme_loop_iod, work);
136 
137 	nvmet_req_execute(&iod->req);
138 }
139 
140 static enum blk_eh_timer_return
141 nvme_loop_timeout(struct request *rq, bool reserved)
142 {
143 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
144 
145 	/* queue error recovery */
146 	nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
147 
148 	/* fail with DNR on admin cmd timeout */
149 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
150 
151 	return BLK_EH_DONE;
152 }
153 
154 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
155 		const struct blk_mq_queue_data *bd)
156 {
157 	struct nvme_ns *ns = hctx->queue->queuedata;
158 	struct nvme_loop_queue *queue = hctx->driver_data;
159 	struct request *req = bd->rq;
160 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
161 	blk_status_t ret;
162 
163 	ret = nvmf_check_if_ready(&queue->ctrl->ctrl, req,
164 		test_bit(NVME_LOOP_Q_LIVE, &queue->flags), true);
165 	if (unlikely(ret))
166 		return ret;
167 
168 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
169 	if (ret)
170 		return ret;
171 
172 	blk_mq_start_request(req);
173 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
174 	iod->req.port = queue->ctrl->port;
175 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
176 			&queue->nvme_sq, &nvme_loop_ops))
177 		return BLK_STS_OK;
178 
179 	if (blk_rq_nr_phys_segments(req)) {
180 		iod->sg_table.sgl = iod->first_sgl;
181 		if (sg_alloc_table_chained(&iod->sg_table,
182 				blk_rq_nr_phys_segments(req),
183 				iod->sg_table.sgl))
184 			return BLK_STS_RESOURCE;
185 
186 		iod->req.sg = iod->sg_table.sgl;
187 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
188 		iod->req.transfer_len = blk_rq_payload_bytes(req);
189 	}
190 
191 	schedule_work(&iod->work);
192 	return BLK_STS_OK;
193 }
194 
195 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
196 {
197 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
198 	struct nvme_loop_queue *queue = &ctrl->queues[0];
199 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
200 
201 	memset(&iod->cmd, 0, sizeof(iod->cmd));
202 	iod->cmd.common.opcode = nvme_admin_async_event;
203 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
204 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
205 
206 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
207 			&nvme_loop_ops)) {
208 		dev_err(ctrl->ctrl.device, "failed async event work\n");
209 		return;
210 	}
211 
212 	schedule_work(&iod->work);
213 }
214 
215 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
216 		struct nvme_loop_iod *iod, unsigned int queue_idx)
217 {
218 	iod->req.cmd = &iod->cmd;
219 	iod->req.rsp = &iod->rsp;
220 	iod->queue = &ctrl->queues[queue_idx];
221 	INIT_WORK(&iod->work, nvme_loop_execute_work);
222 	return 0;
223 }
224 
225 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
226 		struct request *req, unsigned int hctx_idx,
227 		unsigned int numa_node)
228 {
229 	struct nvme_loop_ctrl *ctrl = set->driver_data;
230 
231 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
232 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
233 }
234 
235 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
236 		unsigned int hctx_idx)
237 {
238 	struct nvme_loop_ctrl *ctrl = data;
239 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
240 
241 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
242 
243 	hctx->driver_data = queue;
244 	return 0;
245 }
246 
247 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
248 		unsigned int hctx_idx)
249 {
250 	struct nvme_loop_ctrl *ctrl = data;
251 	struct nvme_loop_queue *queue = &ctrl->queues[0];
252 
253 	BUG_ON(hctx_idx != 0);
254 
255 	hctx->driver_data = queue;
256 	return 0;
257 }
258 
259 static const struct blk_mq_ops nvme_loop_mq_ops = {
260 	.queue_rq	= nvme_loop_queue_rq,
261 	.complete	= nvme_loop_complete_rq,
262 	.init_request	= nvme_loop_init_request,
263 	.init_hctx	= nvme_loop_init_hctx,
264 	.timeout	= nvme_loop_timeout,
265 };
266 
267 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
268 	.queue_rq	= nvme_loop_queue_rq,
269 	.complete	= nvme_loop_complete_rq,
270 	.init_request	= nvme_loop_init_request,
271 	.init_hctx	= nvme_loop_init_admin_hctx,
272 	.timeout	= nvme_loop_timeout,
273 };
274 
275 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
276 {
277 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
278 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
279 	blk_cleanup_queue(ctrl->ctrl.admin_q);
280 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
281 }
282 
283 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
284 {
285 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
286 
287 	if (list_empty(&ctrl->list))
288 		goto free_ctrl;
289 
290 	mutex_lock(&nvme_loop_ctrl_mutex);
291 	list_del(&ctrl->list);
292 	mutex_unlock(&nvme_loop_ctrl_mutex);
293 
294 	if (nctrl->tagset) {
295 		blk_cleanup_queue(ctrl->ctrl.connect_q);
296 		blk_mq_free_tag_set(&ctrl->tag_set);
297 	}
298 	kfree(ctrl->queues);
299 	nvmf_free_options(nctrl->opts);
300 free_ctrl:
301 	kfree(ctrl);
302 }
303 
304 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
305 {
306 	int i;
307 
308 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
309 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
310 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
311 	}
312 }
313 
314 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
315 {
316 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
317 	unsigned int nr_io_queues;
318 	int ret, i;
319 
320 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
321 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
322 	if (ret || !nr_io_queues)
323 		return ret;
324 
325 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
326 
327 	for (i = 1; i <= nr_io_queues; i++) {
328 		ctrl->queues[i].ctrl = ctrl;
329 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
330 		if (ret)
331 			goto out_destroy_queues;
332 
333 		ctrl->ctrl.queue_count++;
334 	}
335 
336 	return 0;
337 
338 out_destroy_queues:
339 	nvme_loop_destroy_io_queues(ctrl);
340 	return ret;
341 }
342 
343 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
344 {
345 	int i, ret;
346 
347 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
348 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
349 		if (ret)
350 			return ret;
351 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
352 	}
353 
354 	return 0;
355 }
356 
357 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
358 {
359 	int error;
360 
361 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
362 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
363 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
364 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
365 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
366 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
367 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
368 	ctrl->admin_tag_set.driver_data = ctrl;
369 	ctrl->admin_tag_set.nr_hw_queues = 1;
370 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
371 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
372 
373 	ctrl->queues[0].ctrl = ctrl;
374 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
375 	if (error)
376 		return error;
377 	ctrl->ctrl.queue_count = 1;
378 
379 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
380 	if (error)
381 		goto out_free_sq;
382 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
383 
384 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
385 	if (IS_ERR(ctrl->ctrl.admin_q)) {
386 		error = PTR_ERR(ctrl->ctrl.admin_q);
387 		goto out_free_tagset;
388 	}
389 
390 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
391 	if (error)
392 		goto out_cleanup_queue;
393 
394 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
395 
396 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
397 	if (error) {
398 		dev_err(ctrl->ctrl.device,
399 			"prop_get NVME_REG_CAP failed\n");
400 		goto out_cleanup_queue;
401 	}
402 
403 	ctrl->ctrl.sqsize =
404 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
405 
406 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
407 	if (error)
408 		goto out_cleanup_queue;
409 
410 	ctrl->ctrl.max_hw_sectors =
411 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
412 
413 	error = nvme_init_identify(&ctrl->ctrl);
414 	if (error)
415 		goto out_cleanup_queue;
416 
417 	return 0;
418 
419 out_cleanup_queue:
420 	blk_cleanup_queue(ctrl->ctrl.admin_q);
421 out_free_tagset:
422 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
423 out_free_sq:
424 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
425 	return error;
426 }
427 
428 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
429 {
430 	if (ctrl->ctrl.queue_count > 1) {
431 		nvme_stop_queues(&ctrl->ctrl);
432 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
433 					nvme_cancel_request, &ctrl->ctrl);
434 		nvme_loop_destroy_io_queues(ctrl);
435 	}
436 
437 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
438 		nvme_shutdown_ctrl(&ctrl->ctrl);
439 
440 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
441 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
442 				nvme_cancel_request, &ctrl->ctrl);
443 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
444 	nvme_loop_destroy_admin_queue(ctrl);
445 }
446 
447 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
448 {
449 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
450 }
451 
452 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
453 {
454 	struct nvme_loop_ctrl *ctrl;
455 
456 	mutex_lock(&nvme_loop_ctrl_mutex);
457 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
458 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
459 			nvme_delete_ctrl(&ctrl->ctrl);
460 	}
461 	mutex_unlock(&nvme_loop_ctrl_mutex);
462 }
463 
464 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
465 {
466 	struct nvme_loop_ctrl *ctrl =
467 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
468 	bool changed;
469 	int ret;
470 
471 	nvme_stop_ctrl(&ctrl->ctrl);
472 	nvme_loop_shutdown_ctrl(ctrl);
473 
474 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
475 		/* state change failure should never happen */
476 		WARN_ON_ONCE(1);
477 		return;
478 	}
479 
480 	ret = nvme_loop_configure_admin_queue(ctrl);
481 	if (ret)
482 		goto out_disable;
483 
484 	ret = nvme_loop_init_io_queues(ctrl);
485 	if (ret)
486 		goto out_destroy_admin;
487 
488 	ret = nvme_loop_connect_io_queues(ctrl);
489 	if (ret)
490 		goto out_destroy_io;
491 
492 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
493 			ctrl->ctrl.queue_count - 1);
494 
495 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
496 	WARN_ON_ONCE(!changed);
497 
498 	nvme_start_ctrl(&ctrl->ctrl);
499 
500 	return;
501 
502 out_destroy_io:
503 	nvme_loop_destroy_io_queues(ctrl);
504 out_destroy_admin:
505 	nvme_loop_destroy_admin_queue(ctrl);
506 out_disable:
507 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
508 	nvme_uninit_ctrl(&ctrl->ctrl);
509 	nvme_put_ctrl(&ctrl->ctrl);
510 }
511 
512 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
513 	.name			= "loop",
514 	.module			= THIS_MODULE,
515 	.flags			= NVME_F_FABRICS,
516 	.reg_read32		= nvmf_reg_read32,
517 	.reg_read64		= nvmf_reg_read64,
518 	.reg_write32		= nvmf_reg_write32,
519 	.free_ctrl		= nvme_loop_free_ctrl,
520 	.submit_async_event	= nvme_loop_submit_async_event,
521 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
522 	.get_address		= nvmf_get_address,
523 };
524 
525 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
526 {
527 	int ret;
528 
529 	ret = nvme_loop_init_io_queues(ctrl);
530 	if (ret)
531 		return ret;
532 
533 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
534 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
535 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
536 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
537 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
538 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
539 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
540 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
541 	ctrl->tag_set.driver_data = ctrl;
542 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
543 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
544 	ctrl->ctrl.tagset = &ctrl->tag_set;
545 
546 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
547 	if (ret)
548 		goto out_destroy_queues;
549 
550 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
551 	if (IS_ERR(ctrl->ctrl.connect_q)) {
552 		ret = PTR_ERR(ctrl->ctrl.connect_q);
553 		goto out_free_tagset;
554 	}
555 
556 	ret = nvme_loop_connect_io_queues(ctrl);
557 	if (ret)
558 		goto out_cleanup_connect_q;
559 
560 	return 0;
561 
562 out_cleanup_connect_q:
563 	blk_cleanup_queue(ctrl->ctrl.connect_q);
564 out_free_tagset:
565 	blk_mq_free_tag_set(&ctrl->tag_set);
566 out_destroy_queues:
567 	nvme_loop_destroy_io_queues(ctrl);
568 	return ret;
569 }
570 
571 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
572 {
573 	struct nvmet_port *p, *found = NULL;
574 
575 	mutex_lock(&nvme_loop_ports_mutex);
576 	list_for_each_entry(p, &nvme_loop_ports, entry) {
577 		/* if no transport address is specified use the first port */
578 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
579 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
580 			continue;
581 		found = p;
582 		break;
583 	}
584 	mutex_unlock(&nvme_loop_ports_mutex);
585 	return found;
586 }
587 
588 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
589 		struct nvmf_ctrl_options *opts)
590 {
591 	struct nvme_loop_ctrl *ctrl;
592 	bool changed;
593 	int ret;
594 
595 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
596 	if (!ctrl)
597 		return ERR_PTR(-ENOMEM);
598 	ctrl->ctrl.opts = opts;
599 	INIT_LIST_HEAD(&ctrl->list);
600 
601 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
602 
603 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
604 				0 /* no quirks, we're perfect! */);
605 	if (ret)
606 		goto out_put_ctrl;
607 
608 	ret = -ENOMEM;
609 
610 	ctrl->ctrl.sqsize = opts->queue_size - 1;
611 	ctrl->ctrl.kato = opts->kato;
612 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
613 
614 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
615 			GFP_KERNEL);
616 	if (!ctrl->queues)
617 		goto out_uninit_ctrl;
618 
619 	ret = nvme_loop_configure_admin_queue(ctrl);
620 	if (ret)
621 		goto out_free_queues;
622 
623 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
624 		/* warn if maxcmd is lower than queue_size */
625 		dev_warn(ctrl->ctrl.device,
626 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
627 			opts->queue_size, ctrl->ctrl.maxcmd);
628 		opts->queue_size = ctrl->ctrl.maxcmd;
629 	}
630 
631 	if (opts->nr_io_queues) {
632 		ret = nvme_loop_create_io_queues(ctrl);
633 		if (ret)
634 			goto out_remove_admin_queue;
635 	}
636 
637 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
638 
639 	dev_info(ctrl->ctrl.device,
640 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
641 
642 	nvme_get_ctrl(&ctrl->ctrl);
643 
644 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
645 	WARN_ON_ONCE(!changed);
646 
647 	mutex_lock(&nvme_loop_ctrl_mutex);
648 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
649 	mutex_unlock(&nvme_loop_ctrl_mutex);
650 
651 	nvme_start_ctrl(&ctrl->ctrl);
652 
653 	return &ctrl->ctrl;
654 
655 out_remove_admin_queue:
656 	nvme_loop_destroy_admin_queue(ctrl);
657 out_free_queues:
658 	kfree(ctrl->queues);
659 out_uninit_ctrl:
660 	nvme_uninit_ctrl(&ctrl->ctrl);
661 out_put_ctrl:
662 	nvme_put_ctrl(&ctrl->ctrl);
663 	if (ret > 0)
664 		ret = -EIO;
665 	return ERR_PTR(ret);
666 }
667 
668 static int nvme_loop_add_port(struct nvmet_port *port)
669 {
670 	mutex_lock(&nvme_loop_ports_mutex);
671 	list_add_tail(&port->entry, &nvme_loop_ports);
672 	mutex_unlock(&nvme_loop_ports_mutex);
673 	return 0;
674 }
675 
676 static void nvme_loop_remove_port(struct nvmet_port *port)
677 {
678 	mutex_lock(&nvme_loop_ports_mutex);
679 	list_del_init(&port->entry);
680 	mutex_unlock(&nvme_loop_ports_mutex);
681 }
682 
683 static const struct nvmet_fabrics_ops nvme_loop_ops = {
684 	.owner		= THIS_MODULE,
685 	.type		= NVMF_TRTYPE_LOOP,
686 	.add_port	= nvme_loop_add_port,
687 	.remove_port	= nvme_loop_remove_port,
688 	.queue_response = nvme_loop_queue_response,
689 	.delete_ctrl	= nvme_loop_delete_ctrl,
690 };
691 
692 static struct nvmf_transport_ops nvme_loop_transport = {
693 	.name		= "loop",
694 	.module		= THIS_MODULE,
695 	.create_ctrl	= nvme_loop_create_ctrl,
696 	.allowed_opts	= NVMF_OPT_TRADDR,
697 };
698 
699 static int __init nvme_loop_init_module(void)
700 {
701 	int ret;
702 
703 	ret = nvmet_register_transport(&nvme_loop_ops);
704 	if (ret)
705 		return ret;
706 
707 	ret = nvmf_register_transport(&nvme_loop_transport);
708 	if (ret)
709 		nvmet_unregister_transport(&nvme_loop_ops);
710 
711 	return ret;
712 }
713 
714 static void __exit nvme_loop_cleanup_module(void)
715 {
716 	struct nvme_loop_ctrl *ctrl, *next;
717 
718 	nvmf_unregister_transport(&nvme_loop_transport);
719 	nvmet_unregister_transport(&nvme_loop_ops);
720 
721 	mutex_lock(&nvme_loop_ctrl_mutex);
722 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
723 		nvme_delete_ctrl(&ctrl->ctrl);
724 	mutex_unlock(&nvme_loop_ctrl_mutex);
725 
726 	flush_workqueue(nvme_delete_wq);
727 }
728 
729 module_init(nvme_loop_init_module);
730 module_exit(nvme_loop_cleanup_module);
731 
732 MODULE_LICENSE("GPL v2");
733 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
734