xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 04a7279f)
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 };
49 
50 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
51 {
52 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
53 }
54 
55 struct nvme_loop_queue {
56 	struct nvmet_cq		nvme_cq;
57 	struct nvmet_sq		nvme_sq;
58 	struct nvme_loop_ctrl	*ctrl;
59 };
60 
61 static struct nvmet_port *nvmet_loop_port;
62 
63 static LIST_HEAD(nvme_loop_ctrl_list);
64 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
65 
66 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
67 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
68 
69 static struct nvmet_fabrics_ops nvme_loop_ops;
70 
71 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
72 {
73 	return queue - queue->ctrl->queues;
74 }
75 
76 static void nvme_loop_complete_rq(struct request *req)
77 {
78 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
79 
80 	nvme_cleanup_cmd(req);
81 	sg_free_table_chained(&iod->sg_table, true);
82 	nvme_complete_rq(req);
83 }
84 
85 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
86 {
87 	u32 queue_idx = nvme_loop_queue_idx(queue);
88 
89 	if (queue_idx == 0)
90 		return queue->ctrl->admin_tag_set.tags[queue_idx];
91 	return queue->ctrl->tag_set.tags[queue_idx - 1];
92 }
93 
94 static void nvme_loop_queue_response(struct nvmet_req *req)
95 {
96 	struct nvme_loop_queue *queue =
97 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
98 	struct nvme_completion *cqe = req->rsp;
99 
100 	/*
101 	 * AEN requests are special as they don't time out and can
102 	 * survive any kind of queue freeze and often don't respond to
103 	 * aborts.  We don't even bother to allocate a struct request
104 	 * for them but rather special case them here.
105 	 */
106 	if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
107 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
108 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
109 				&cqe->result);
110 	} else {
111 		struct request *rq;
112 
113 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
114 		if (!rq) {
115 			dev_err(queue->ctrl->ctrl.device,
116 				"tag 0x%x on queue %d not found\n",
117 				cqe->command_id, nvme_loop_queue_idx(queue));
118 			return;
119 		}
120 
121 		nvme_end_request(rq, cqe->status, cqe->result);
122 	}
123 }
124 
125 static void nvme_loop_execute_work(struct work_struct *work)
126 {
127 	struct nvme_loop_iod *iod =
128 		container_of(work, struct nvme_loop_iod, work);
129 
130 	nvmet_req_execute(&iod->req);
131 }
132 
133 static enum blk_eh_timer_return
134 nvme_loop_timeout(struct request *rq, bool reserved)
135 {
136 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
137 
138 	/* queue error recovery */
139 	nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
140 
141 	/* fail with DNR on admin cmd timeout */
142 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
143 
144 	return BLK_EH_HANDLED;
145 }
146 
147 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
148 		const struct blk_mq_queue_data *bd)
149 {
150 	struct nvme_ns *ns = hctx->queue->queuedata;
151 	struct nvme_loop_queue *queue = hctx->driver_data;
152 	struct request *req = bd->rq;
153 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
154 	blk_status_t ret;
155 
156 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
157 	if (ret)
158 		return ret;
159 
160 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
161 	iod->req.port = nvmet_loop_port;
162 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
163 			&queue->nvme_sq, &nvme_loop_ops)) {
164 		nvme_cleanup_cmd(req);
165 		blk_mq_start_request(req);
166 		nvme_loop_queue_response(&iod->req);
167 		return BLK_STS_OK;
168 	}
169 
170 	if (blk_rq_bytes(req)) {
171 		iod->sg_table.sgl = iod->first_sgl;
172 		if (sg_alloc_table_chained(&iod->sg_table,
173 				blk_rq_nr_phys_segments(req),
174 				iod->sg_table.sgl))
175 			return BLK_STS_RESOURCE;
176 
177 		iod->req.sg = iod->sg_table.sgl;
178 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
179 		iod->req.transfer_len = blk_rq_bytes(req);
180 	}
181 
182 	blk_mq_start_request(req);
183 
184 	schedule_work(&iod->work);
185 	return BLK_STS_OK;
186 }
187 
188 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
189 {
190 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
191 	struct nvme_loop_queue *queue = &ctrl->queues[0];
192 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
193 
194 	memset(&iod->cmd, 0, sizeof(iod->cmd));
195 	iod->cmd.common.opcode = nvme_admin_async_event;
196 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
197 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
198 
199 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
200 			&nvme_loop_ops)) {
201 		dev_err(ctrl->ctrl.device, "failed async event work\n");
202 		return;
203 	}
204 
205 	schedule_work(&iod->work);
206 }
207 
208 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
209 		struct nvme_loop_iod *iod, unsigned int queue_idx)
210 {
211 	iod->req.cmd = &iod->cmd;
212 	iod->req.rsp = &iod->rsp;
213 	iod->queue = &ctrl->queues[queue_idx];
214 	INIT_WORK(&iod->work, nvme_loop_execute_work);
215 	return 0;
216 }
217 
218 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
219 		struct request *req, unsigned int hctx_idx,
220 		unsigned int numa_node)
221 {
222 	struct nvme_loop_ctrl *ctrl = set->driver_data;
223 
224 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
225 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
226 }
227 
228 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
229 		unsigned int hctx_idx)
230 {
231 	struct nvme_loop_ctrl *ctrl = data;
232 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
233 
234 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
235 
236 	hctx->driver_data = queue;
237 	return 0;
238 }
239 
240 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
241 		unsigned int hctx_idx)
242 {
243 	struct nvme_loop_ctrl *ctrl = data;
244 	struct nvme_loop_queue *queue = &ctrl->queues[0];
245 
246 	BUG_ON(hctx_idx != 0);
247 
248 	hctx->driver_data = queue;
249 	return 0;
250 }
251 
252 static const struct blk_mq_ops nvme_loop_mq_ops = {
253 	.queue_rq	= nvme_loop_queue_rq,
254 	.complete	= nvme_loop_complete_rq,
255 	.init_request	= nvme_loop_init_request,
256 	.init_hctx	= nvme_loop_init_hctx,
257 	.timeout	= nvme_loop_timeout,
258 };
259 
260 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
261 	.queue_rq	= nvme_loop_queue_rq,
262 	.complete	= nvme_loop_complete_rq,
263 	.init_request	= nvme_loop_init_request,
264 	.init_hctx	= nvme_loop_init_admin_hctx,
265 	.timeout	= nvme_loop_timeout,
266 };
267 
268 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
269 {
270 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
271 	blk_cleanup_queue(ctrl->ctrl.admin_q);
272 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
273 }
274 
275 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
276 {
277 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
278 
279 	if (list_empty(&ctrl->list))
280 		goto free_ctrl;
281 
282 	mutex_lock(&nvme_loop_ctrl_mutex);
283 	list_del(&ctrl->list);
284 	mutex_unlock(&nvme_loop_ctrl_mutex);
285 
286 	if (nctrl->tagset) {
287 		blk_cleanup_queue(ctrl->ctrl.connect_q);
288 		blk_mq_free_tag_set(&ctrl->tag_set);
289 	}
290 	kfree(ctrl->queues);
291 	nvmf_free_options(nctrl->opts);
292 free_ctrl:
293 	kfree(ctrl);
294 }
295 
296 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
297 {
298 	int i;
299 
300 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
301 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
302 }
303 
304 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
305 {
306 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
307 	unsigned int nr_io_queues;
308 	int ret, i;
309 
310 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
311 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
312 	if (ret || !nr_io_queues)
313 		return ret;
314 
315 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
316 
317 	for (i = 1; i <= nr_io_queues; i++) {
318 		ctrl->queues[i].ctrl = ctrl;
319 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
320 		if (ret)
321 			goto out_destroy_queues;
322 
323 		ctrl->ctrl.queue_count++;
324 	}
325 
326 	return 0;
327 
328 out_destroy_queues:
329 	nvme_loop_destroy_io_queues(ctrl);
330 	return ret;
331 }
332 
333 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
334 {
335 	int i, ret;
336 
337 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
338 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
339 		if (ret)
340 			return ret;
341 	}
342 
343 	return 0;
344 }
345 
346 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
347 {
348 	int error;
349 
350 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
351 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
352 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
353 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
354 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
355 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
356 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
357 	ctrl->admin_tag_set.driver_data = ctrl;
358 	ctrl->admin_tag_set.nr_hw_queues = 1;
359 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
360 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
361 
362 	ctrl->queues[0].ctrl = ctrl;
363 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
364 	if (error)
365 		return error;
366 	ctrl->ctrl.queue_count = 1;
367 
368 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
369 	if (error)
370 		goto out_free_sq;
371 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
372 
373 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
374 	if (IS_ERR(ctrl->ctrl.admin_q)) {
375 		error = PTR_ERR(ctrl->ctrl.admin_q);
376 		goto out_free_tagset;
377 	}
378 
379 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
380 	if (error)
381 		goto out_cleanup_queue;
382 
383 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
384 	if (error) {
385 		dev_err(ctrl->ctrl.device,
386 			"prop_get NVME_REG_CAP failed\n");
387 		goto out_cleanup_queue;
388 	}
389 
390 	ctrl->ctrl.sqsize =
391 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
392 
393 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
394 	if (error)
395 		goto out_cleanup_queue;
396 
397 	ctrl->ctrl.max_hw_sectors =
398 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
399 
400 	error = nvme_init_identify(&ctrl->ctrl);
401 	if (error)
402 		goto out_cleanup_queue;
403 
404 	return 0;
405 
406 out_cleanup_queue:
407 	blk_cleanup_queue(ctrl->ctrl.admin_q);
408 out_free_tagset:
409 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
410 out_free_sq:
411 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
412 	return error;
413 }
414 
415 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
416 {
417 	if (ctrl->ctrl.queue_count > 1) {
418 		nvme_stop_queues(&ctrl->ctrl);
419 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
420 					nvme_cancel_request, &ctrl->ctrl);
421 		nvme_loop_destroy_io_queues(ctrl);
422 	}
423 
424 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
425 		nvme_shutdown_ctrl(&ctrl->ctrl);
426 
427 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
428 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
429 				nvme_cancel_request, &ctrl->ctrl);
430 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
431 	nvme_loop_destroy_admin_queue(ctrl);
432 }
433 
434 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
435 {
436 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
437 }
438 
439 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
440 {
441 	struct nvme_loop_ctrl *ctrl;
442 
443 	mutex_lock(&nvme_loop_ctrl_mutex);
444 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
445 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
446 			nvme_delete_ctrl(&ctrl->ctrl);
447 	}
448 	mutex_unlock(&nvme_loop_ctrl_mutex);
449 }
450 
451 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
452 {
453 	struct nvme_loop_ctrl *ctrl =
454 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
455 	bool changed;
456 	int ret;
457 
458 	nvme_stop_ctrl(&ctrl->ctrl);
459 	nvme_loop_shutdown_ctrl(ctrl);
460 
461 	ret = nvme_loop_configure_admin_queue(ctrl);
462 	if (ret)
463 		goto out_disable;
464 
465 	ret = nvme_loop_init_io_queues(ctrl);
466 	if (ret)
467 		goto out_destroy_admin;
468 
469 	ret = nvme_loop_connect_io_queues(ctrl);
470 	if (ret)
471 		goto out_destroy_io;
472 
473 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
474 			ctrl->ctrl.queue_count - 1);
475 
476 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
477 	WARN_ON_ONCE(!changed);
478 
479 	nvme_start_ctrl(&ctrl->ctrl);
480 
481 	return;
482 
483 out_destroy_io:
484 	nvme_loop_destroy_io_queues(ctrl);
485 out_destroy_admin:
486 	nvme_loop_destroy_admin_queue(ctrl);
487 out_disable:
488 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
489 	nvme_uninit_ctrl(&ctrl->ctrl);
490 	nvme_put_ctrl(&ctrl->ctrl);
491 }
492 
493 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
494 	.name			= "loop",
495 	.module			= THIS_MODULE,
496 	.flags			= NVME_F_FABRICS,
497 	.reg_read32		= nvmf_reg_read32,
498 	.reg_read64		= nvmf_reg_read64,
499 	.reg_write32		= nvmf_reg_write32,
500 	.free_ctrl		= nvme_loop_free_ctrl,
501 	.submit_async_event	= nvme_loop_submit_async_event,
502 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
503 };
504 
505 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
506 {
507 	int ret;
508 
509 	ret = nvme_loop_init_io_queues(ctrl);
510 	if (ret)
511 		return ret;
512 
513 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
514 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
515 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
516 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
517 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
518 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
519 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
520 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
521 	ctrl->tag_set.driver_data = ctrl;
522 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
523 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
524 	ctrl->ctrl.tagset = &ctrl->tag_set;
525 
526 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
527 	if (ret)
528 		goto out_destroy_queues;
529 
530 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
531 	if (IS_ERR(ctrl->ctrl.connect_q)) {
532 		ret = PTR_ERR(ctrl->ctrl.connect_q);
533 		goto out_free_tagset;
534 	}
535 
536 	ret = nvme_loop_connect_io_queues(ctrl);
537 	if (ret)
538 		goto out_cleanup_connect_q;
539 
540 	return 0;
541 
542 out_cleanup_connect_q:
543 	blk_cleanup_queue(ctrl->ctrl.connect_q);
544 out_free_tagset:
545 	blk_mq_free_tag_set(&ctrl->tag_set);
546 out_destroy_queues:
547 	nvme_loop_destroy_io_queues(ctrl);
548 	return ret;
549 }
550 
551 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
552 		struct nvmf_ctrl_options *opts)
553 {
554 	struct nvme_loop_ctrl *ctrl;
555 	bool changed;
556 	int ret;
557 
558 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
559 	if (!ctrl)
560 		return ERR_PTR(-ENOMEM);
561 	ctrl->ctrl.opts = opts;
562 	INIT_LIST_HEAD(&ctrl->list);
563 
564 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
565 
566 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
567 				0 /* no quirks, we're perfect! */);
568 	if (ret)
569 		goto out_put_ctrl;
570 
571 	ret = -ENOMEM;
572 
573 	ctrl->ctrl.sqsize = opts->queue_size - 1;
574 	ctrl->ctrl.kato = opts->kato;
575 
576 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
577 			GFP_KERNEL);
578 	if (!ctrl->queues)
579 		goto out_uninit_ctrl;
580 
581 	ret = nvme_loop_configure_admin_queue(ctrl);
582 	if (ret)
583 		goto out_free_queues;
584 
585 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
586 		/* warn if maxcmd is lower than queue_size */
587 		dev_warn(ctrl->ctrl.device,
588 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
589 			opts->queue_size, ctrl->ctrl.maxcmd);
590 		opts->queue_size = ctrl->ctrl.maxcmd;
591 	}
592 
593 	if (opts->nr_io_queues) {
594 		ret = nvme_loop_create_io_queues(ctrl);
595 		if (ret)
596 			goto out_remove_admin_queue;
597 	}
598 
599 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
600 
601 	dev_info(ctrl->ctrl.device,
602 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
603 
604 	nvme_get_ctrl(&ctrl->ctrl);
605 
606 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
607 	WARN_ON_ONCE(!changed);
608 
609 	mutex_lock(&nvme_loop_ctrl_mutex);
610 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
611 	mutex_unlock(&nvme_loop_ctrl_mutex);
612 
613 	nvme_start_ctrl(&ctrl->ctrl);
614 
615 	return &ctrl->ctrl;
616 
617 out_remove_admin_queue:
618 	nvme_loop_destroy_admin_queue(ctrl);
619 out_free_queues:
620 	kfree(ctrl->queues);
621 out_uninit_ctrl:
622 	nvme_uninit_ctrl(&ctrl->ctrl);
623 out_put_ctrl:
624 	nvme_put_ctrl(&ctrl->ctrl);
625 	if (ret > 0)
626 		ret = -EIO;
627 	return ERR_PTR(ret);
628 }
629 
630 static int nvme_loop_add_port(struct nvmet_port *port)
631 {
632 	/*
633 	 * XXX: disalow adding more than one port so
634 	 * there is no connection rejections when a
635 	 * a subsystem is assigned to a port for which
636 	 * loop doesn't have a pointer.
637 	 * This scenario would be possible if we allowed
638 	 * more than one port to be added and a subsystem
639 	 * was assigned to a port other than nvmet_loop_port.
640 	 */
641 
642 	if (nvmet_loop_port)
643 		return -EPERM;
644 
645 	nvmet_loop_port = port;
646 	return 0;
647 }
648 
649 static void nvme_loop_remove_port(struct nvmet_port *port)
650 {
651 	if (port == nvmet_loop_port)
652 		nvmet_loop_port = NULL;
653 }
654 
655 static struct nvmet_fabrics_ops nvme_loop_ops = {
656 	.owner		= THIS_MODULE,
657 	.type		= NVMF_TRTYPE_LOOP,
658 	.add_port	= nvme_loop_add_port,
659 	.remove_port	= nvme_loop_remove_port,
660 	.queue_response = nvme_loop_queue_response,
661 	.delete_ctrl	= nvme_loop_delete_ctrl,
662 };
663 
664 static struct nvmf_transport_ops nvme_loop_transport = {
665 	.name		= "loop",
666 	.create_ctrl	= nvme_loop_create_ctrl,
667 };
668 
669 static int __init nvme_loop_init_module(void)
670 {
671 	int ret;
672 
673 	ret = nvmet_register_transport(&nvme_loop_ops);
674 	if (ret)
675 		return ret;
676 
677 	ret = nvmf_register_transport(&nvme_loop_transport);
678 	if (ret)
679 		nvmet_unregister_transport(&nvme_loop_ops);
680 
681 	return ret;
682 }
683 
684 static void __exit nvme_loop_cleanup_module(void)
685 {
686 	struct nvme_loop_ctrl *ctrl, *next;
687 
688 	nvmf_unregister_transport(&nvme_loop_transport);
689 	nvmet_unregister_transport(&nvme_loop_ops);
690 
691 	mutex_lock(&nvme_loop_ctrl_mutex);
692 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
693 		nvme_delete_ctrl(&ctrl->ctrl);
694 	mutex_unlock(&nvme_loop_ctrl_mutex);
695 
696 	flush_workqueue(nvme_wq);
697 }
698 
699 module_init(nvme_loop_init_module);
700 module_exit(nvme_loop_cleanup_module);
701 
702 MODULE_LICENSE("GPL v2");
703 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
704