xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 65417d9f)
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 enum nvme_loop_queue_flags {
56 	NVME_LOOP_Q_LIVE	= 0,
57 };
58 
59 struct nvme_loop_queue {
60 	struct nvmet_cq		nvme_cq;
61 	struct nvmet_sq		nvme_sq;
62 	struct nvme_loop_ctrl	*ctrl;
63 	unsigned long		flags;
64 };
65 
66 static struct nvmet_port *nvmet_loop_port;
67 
68 static LIST_HEAD(nvme_loop_ctrl_list);
69 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
70 
71 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
72 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
73 
74 static struct nvmet_fabrics_ops nvme_loop_ops;
75 
76 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
77 {
78 	return queue - queue->ctrl->queues;
79 }
80 
81 static void nvme_loop_complete_rq(struct request *req)
82 {
83 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
84 
85 	nvme_cleanup_cmd(req);
86 	sg_free_table_chained(&iod->sg_table, true);
87 	nvme_complete_rq(req);
88 }
89 
90 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
91 {
92 	u32 queue_idx = nvme_loop_queue_idx(queue);
93 
94 	if (queue_idx == 0)
95 		return queue->ctrl->admin_tag_set.tags[queue_idx];
96 	return queue->ctrl->tag_set.tags[queue_idx - 1];
97 }
98 
99 static void nvme_loop_queue_response(struct nvmet_req *req)
100 {
101 	struct nvme_loop_queue *queue =
102 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
103 	struct nvme_completion *cqe = req->rsp;
104 
105 	/*
106 	 * AEN requests are special as they don't time out and can
107 	 * survive any kind of queue freeze and often don't respond to
108 	 * aborts.  We don't even bother to allocate a struct request
109 	 * for them but rather special case them here.
110 	 */
111 	if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
112 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
113 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
114 				&cqe->result);
115 	} else {
116 		struct request *rq;
117 
118 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
119 		if (!rq) {
120 			dev_err(queue->ctrl->ctrl.device,
121 				"tag 0x%x on queue %d not found\n",
122 				cqe->command_id, nvme_loop_queue_idx(queue));
123 			return;
124 		}
125 
126 		nvme_end_request(rq, cqe->status, cqe->result);
127 	}
128 }
129 
130 static void nvme_loop_execute_work(struct work_struct *work)
131 {
132 	struct nvme_loop_iod *iod =
133 		container_of(work, struct nvme_loop_iod, work);
134 
135 	nvmet_req_execute(&iod->req);
136 }
137 
138 static enum blk_eh_timer_return
139 nvme_loop_timeout(struct request *rq, bool reserved)
140 {
141 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
142 
143 	/* queue error recovery */
144 	nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
145 
146 	/* fail with DNR on admin cmd timeout */
147 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
148 
149 	return BLK_EH_HANDLED;
150 }
151 
152 static inline blk_status_t nvme_loop_is_ready(struct nvme_loop_queue *queue,
153 		struct request *rq)
154 {
155 	if (unlikely(!test_bit(NVME_LOOP_Q_LIVE, &queue->flags)))
156 		return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
157 	return BLK_STS_OK;
158 }
159 
160 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
161 		const struct blk_mq_queue_data *bd)
162 {
163 	struct nvme_ns *ns = hctx->queue->queuedata;
164 	struct nvme_loop_queue *queue = hctx->driver_data;
165 	struct request *req = bd->rq;
166 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
167 	blk_status_t ret;
168 
169 	ret = nvme_loop_is_ready(queue, req);
170 	if (unlikely(ret))
171 		return ret;
172 
173 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
174 	if (ret)
175 		return ret;
176 
177 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
178 	iod->req.port = nvmet_loop_port;
179 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
180 			&queue->nvme_sq, &nvme_loop_ops)) {
181 		nvme_cleanup_cmd(req);
182 		blk_mq_start_request(req);
183 		nvme_loop_queue_response(&iod->req);
184 		return BLK_STS_OK;
185 	}
186 
187 	if (blk_rq_bytes(req)) {
188 		iod->sg_table.sgl = iod->first_sgl;
189 		if (sg_alloc_table_chained(&iod->sg_table,
190 				blk_rq_nr_phys_segments(req),
191 				iod->sg_table.sgl))
192 			return BLK_STS_RESOURCE;
193 
194 		iod->req.sg = iod->sg_table.sgl;
195 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
196 		iod->req.transfer_len = blk_rq_bytes(req);
197 	}
198 
199 	blk_mq_start_request(req);
200 
201 	schedule_work(&iod->work);
202 	return BLK_STS_OK;
203 }
204 
205 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
206 {
207 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
208 	struct nvme_loop_queue *queue = &ctrl->queues[0];
209 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
210 
211 	memset(&iod->cmd, 0, sizeof(iod->cmd));
212 	iod->cmd.common.opcode = nvme_admin_async_event;
213 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
214 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
215 
216 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
217 			&nvme_loop_ops)) {
218 		dev_err(ctrl->ctrl.device, "failed async event work\n");
219 		return;
220 	}
221 
222 	schedule_work(&iod->work);
223 }
224 
225 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
226 		struct nvme_loop_iod *iod, unsigned int queue_idx)
227 {
228 	iod->req.cmd = &iod->cmd;
229 	iod->req.rsp = &iod->rsp;
230 	iod->queue = &ctrl->queues[queue_idx];
231 	INIT_WORK(&iod->work, nvme_loop_execute_work);
232 	return 0;
233 }
234 
235 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
236 		struct request *req, unsigned int hctx_idx,
237 		unsigned int numa_node)
238 {
239 	struct nvme_loop_ctrl *ctrl = set->driver_data;
240 
241 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
242 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
243 }
244 
245 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
246 		unsigned int hctx_idx)
247 {
248 	struct nvme_loop_ctrl *ctrl = data;
249 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
250 
251 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
252 
253 	hctx->driver_data = queue;
254 	return 0;
255 }
256 
257 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
258 		unsigned int hctx_idx)
259 {
260 	struct nvme_loop_ctrl *ctrl = data;
261 	struct nvme_loop_queue *queue = &ctrl->queues[0];
262 
263 	BUG_ON(hctx_idx != 0);
264 
265 	hctx->driver_data = queue;
266 	return 0;
267 }
268 
269 static const struct blk_mq_ops nvme_loop_mq_ops = {
270 	.queue_rq	= nvme_loop_queue_rq,
271 	.complete	= nvme_loop_complete_rq,
272 	.init_request	= nvme_loop_init_request,
273 	.init_hctx	= nvme_loop_init_hctx,
274 	.timeout	= nvme_loop_timeout,
275 };
276 
277 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
278 	.queue_rq	= nvme_loop_queue_rq,
279 	.complete	= nvme_loop_complete_rq,
280 	.init_request	= nvme_loop_init_request,
281 	.init_hctx	= nvme_loop_init_admin_hctx,
282 	.timeout	= nvme_loop_timeout,
283 };
284 
285 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
286 {
287 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
288 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
289 	blk_cleanup_queue(ctrl->ctrl.admin_q);
290 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
291 }
292 
293 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
294 {
295 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
296 
297 	if (list_empty(&ctrl->list))
298 		goto free_ctrl;
299 
300 	mutex_lock(&nvme_loop_ctrl_mutex);
301 	list_del(&ctrl->list);
302 	mutex_unlock(&nvme_loop_ctrl_mutex);
303 
304 	if (nctrl->tagset) {
305 		blk_cleanup_queue(ctrl->ctrl.connect_q);
306 		blk_mq_free_tag_set(&ctrl->tag_set);
307 	}
308 	kfree(ctrl->queues);
309 	nvmf_free_options(nctrl->opts);
310 free_ctrl:
311 	kfree(ctrl);
312 }
313 
314 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
315 {
316 	int i;
317 
318 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
319 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
320 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
321 	}
322 }
323 
324 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
325 {
326 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
327 	unsigned int nr_io_queues;
328 	int ret, i;
329 
330 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
331 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
332 	if (ret || !nr_io_queues)
333 		return ret;
334 
335 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
336 
337 	for (i = 1; i <= nr_io_queues; i++) {
338 		ctrl->queues[i].ctrl = ctrl;
339 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
340 		if (ret)
341 			goto out_destroy_queues;
342 
343 		ctrl->ctrl.queue_count++;
344 	}
345 
346 	return 0;
347 
348 out_destroy_queues:
349 	nvme_loop_destroy_io_queues(ctrl);
350 	return ret;
351 }
352 
353 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
354 {
355 	int i, ret;
356 
357 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
358 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
359 		if (ret)
360 			return ret;
361 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
362 	}
363 
364 	return 0;
365 }
366 
367 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
368 {
369 	int error;
370 
371 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
372 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
373 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
374 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
375 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
376 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
377 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
378 	ctrl->admin_tag_set.driver_data = ctrl;
379 	ctrl->admin_tag_set.nr_hw_queues = 1;
380 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
381 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
382 
383 	ctrl->queues[0].ctrl = ctrl;
384 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
385 	if (error)
386 		return error;
387 	ctrl->ctrl.queue_count = 1;
388 
389 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
390 	if (error)
391 		goto out_free_sq;
392 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
393 
394 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
395 	if (IS_ERR(ctrl->ctrl.admin_q)) {
396 		error = PTR_ERR(ctrl->ctrl.admin_q);
397 		goto out_free_tagset;
398 	}
399 
400 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
401 	if (error)
402 		goto out_cleanup_queue;
403 
404 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
405 
406 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
407 	if (error) {
408 		dev_err(ctrl->ctrl.device,
409 			"prop_get NVME_REG_CAP failed\n");
410 		goto out_cleanup_queue;
411 	}
412 
413 	ctrl->ctrl.sqsize =
414 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
415 
416 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
417 	if (error)
418 		goto out_cleanup_queue;
419 
420 	ctrl->ctrl.max_hw_sectors =
421 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
422 
423 	error = nvme_init_identify(&ctrl->ctrl);
424 	if (error)
425 		goto out_cleanup_queue;
426 
427 	return 0;
428 
429 out_cleanup_queue:
430 	blk_cleanup_queue(ctrl->ctrl.admin_q);
431 out_free_tagset:
432 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
433 out_free_sq:
434 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
435 	return error;
436 }
437 
438 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
439 {
440 	if (ctrl->ctrl.queue_count > 1) {
441 		nvme_stop_queues(&ctrl->ctrl);
442 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
443 					nvme_cancel_request, &ctrl->ctrl);
444 		nvme_loop_destroy_io_queues(ctrl);
445 	}
446 
447 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
448 		nvme_shutdown_ctrl(&ctrl->ctrl);
449 
450 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
451 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
452 				nvme_cancel_request, &ctrl->ctrl);
453 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
454 	nvme_loop_destroy_admin_queue(ctrl);
455 }
456 
457 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
458 {
459 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
460 }
461 
462 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
463 {
464 	struct nvme_loop_ctrl *ctrl;
465 
466 	mutex_lock(&nvme_loop_ctrl_mutex);
467 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
468 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
469 			nvme_delete_ctrl(&ctrl->ctrl);
470 	}
471 	mutex_unlock(&nvme_loop_ctrl_mutex);
472 }
473 
474 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
475 {
476 	struct nvme_loop_ctrl *ctrl =
477 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
478 	bool changed;
479 	int ret;
480 
481 	nvme_stop_ctrl(&ctrl->ctrl);
482 	nvme_loop_shutdown_ctrl(ctrl);
483 
484 	ret = nvme_loop_configure_admin_queue(ctrl);
485 	if (ret)
486 		goto out_disable;
487 
488 	ret = nvme_loop_init_io_queues(ctrl);
489 	if (ret)
490 		goto out_destroy_admin;
491 
492 	ret = nvme_loop_connect_io_queues(ctrl);
493 	if (ret)
494 		goto out_destroy_io;
495 
496 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
497 			ctrl->ctrl.queue_count - 1);
498 
499 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
500 	WARN_ON_ONCE(!changed);
501 
502 	nvme_start_ctrl(&ctrl->ctrl);
503 
504 	return;
505 
506 out_destroy_io:
507 	nvme_loop_destroy_io_queues(ctrl);
508 out_destroy_admin:
509 	nvme_loop_destroy_admin_queue(ctrl);
510 out_disable:
511 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
512 	nvme_uninit_ctrl(&ctrl->ctrl);
513 	nvme_put_ctrl(&ctrl->ctrl);
514 }
515 
516 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
517 	.name			= "loop",
518 	.module			= THIS_MODULE,
519 	.flags			= NVME_F_FABRICS,
520 	.reg_read32		= nvmf_reg_read32,
521 	.reg_read64		= nvmf_reg_read64,
522 	.reg_write32		= nvmf_reg_write32,
523 	.free_ctrl		= nvme_loop_free_ctrl,
524 	.submit_async_event	= nvme_loop_submit_async_event,
525 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
526 };
527 
528 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
529 {
530 	int ret;
531 
532 	ret = nvme_loop_init_io_queues(ctrl);
533 	if (ret)
534 		return ret;
535 
536 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
537 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
538 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
539 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
540 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
541 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
542 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
543 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
544 	ctrl->tag_set.driver_data = ctrl;
545 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
546 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
547 	ctrl->ctrl.tagset = &ctrl->tag_set;
548 
549 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
550 	if (ret)
551 		goto out_destroy_queues;
552 
553 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
554 	if (IS_ERR(ctrl->ctrl.connect_q)) {
555 		ret = PTR_ERR(ctrl->ctrl.connect_q);
556 		goto out_free_tagset;
557 	}
558 
559 	ret = nvme_loop_connect_io_queues(ctrl);
560 	if (ret)
561 		goto out_cleanup_connect_q;
562 
563 	return 0;
564 
565 out_cleanup_connect_q:
566 	blk_cleanup_queue(ctrl->ctrl.connect_q);
567 out_free_tagset:
568 	blk_mq_free_tag_set(&ctrl->tag_set);
569 out_destroy_queues:
570 	nvme_loop_destroy_io_queues(ctrl);
571 	return ret;
572 }
573 
574 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
575 		struct nvmf_ctrl_options *opts)
576 {
577 	struct nvme_loop_ctrl *ctrl;
578 	bool changed;
579 	int ret;
580 
581 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
582 	if (!ctrl)
583 		return ERR_PTR(-ENOMEM);
584 	ctrl->ctrl.opts = opts;
585 	INIT_LIST_HEAD(&ctrl->list);
586 
587 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
588 
589 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
590 				0 /* no quirks, we're perfect! */);
591 	if (ret)
592 		goto out_put_ctrl;
593 
594 	ret = -ENOMEM;
595 
596 	ctrl->ctrl.sqsize = opts->queue_size - 1;
597 	ctrl->ctrl.kato = opts->kato;
598 
599 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
600 			GFP_KERNEL);
601 	if (!ctrl->queues)
602 		goto out_uninit_ctrl;
603 
604 	ret = nvme_loop_configure_admin_queue(ctrl);
605 	if (ret)
606 		goto out_free_queues;
607 
608 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
609 		/* warn if maxcmd is lower than queue_size */
610 		dev_warn(ctrl->ctrl.device,
611 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
612 			opts->queue_size, ctrl->ctrl.maxcmd);
613 		opts->queue_size = ctrl->ctrl.maxcmd;
614 	}
615 
616 	if (opts->nr_io_queues) {
617 		ret = nvme_loop_create_io_queues(ctrl);
618 		if (ret)
619 			goto out_remove_admin_queue;
620 	}
621 
622 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
623 
624 	dev_info(ctrl->ctrl.device,
625 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
626 
627 	nvme_get_ctrl(&ctrl->ctrl);
628 
629 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
630 	WARN_ON_ONCE(!changed);
631 
632 	mutex_lock(&nvme_loop_ctrl_mutex);
633 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
634 	mutex_unlock(&nvme_loop_ctrl_mutex);
635 
636 	nvme_start_ctrl(&ctrl->ctrl);
637 
638 	return &ctrl->ctrl;
639 
640 out_remove_admin_queue:
641 	nvme_loop_destroy_admin_queue(ctrl);
642 out_free_queues:
643 	kfree(ctrl->queues);
644 out_uninit_ctrl:
645 	nvme_uninit_ctrl(&ctrl->ctrl);
646 out_put_ctrl:
647 	nvme_put_ctrl(&ctrl->ctrl);
648 	if (ret > 0)
649 		ret = -EIO;
650 	return ERR_PTR(ret);
651 }
652 
653 static int nvme_loop_add_port(struct nvmet_port *port)
654 {
655 	/*
656 	 * XXX: disalow adding more than one port so
657 	 * there is no connection rejections when a
658 	 * a subsystem is assigned to a port for which
659 	 * loop doesn't have a pointer.
660 	 * This scenario would be possible if we allowed
661 	 * more than one port to be added and a subsystem
662 	 * was assigned to a port other than nvmet_loop_port.
663 	 */
664 
665 	if (nvmet_loop_port)
666 		return -EPERM;
667 
668 	nvmet_loop_port = port;
669 	return 0;
670 }
671 
672 static void nvme_loop_remove_port(struct nvmet_port *port)
673 {
674 	if (port == nvmet_loop_port)
675 		nvmet_loop_port = NULL;
676 }
677 
678 static struct nvmet_fabrics_ops nvme_loop_ops = {
679 	.owner		= THIS_MODULE,
680 	.type		= NVMF_TRTYPE_LOOP,
681 	.add_port	= nvme_loop_add_port,
682 	.remove_port	= nvme_loop_remove_port,
683 	.queue_response = nvme_loop_queue_response,
684 	.delete_ctrl	= nvme_loop_delete_ctrl,
685 };
686 
687 static struct nvmf_transport_ops nvme_loop_transport = {
688 	.name		= "loop",
689 	.create_ctrl	= nvme_loop_create_ctrl,
690 };
691 
692 static int __init nvme_loop_init_module(void)
693 {
694 	int ret;
695 
696 	ret = nvmet_register_transport(&nvme_loop_ops);
697 	if (ret)
698 		return ret;
699 
700 	ret = nvmf_register_transport(&nvme_loop_transport);
701 	if (ret)
702 		nvmet_unregister_transport(&nvme_loop_ops);
703 
704 	return ret;
705 }
706 
707 static void __exit nvme_loop_cleanup_module(void)
708 {
709 	struct nvme_loop_ctrl *ctrl, *next;
710 
711 	nvmf_unregister_transport(&nvme_loop_transport);
712 	nvmet_unregister_transport(&nvme_loop_ops);
713 
714 	mutex_lock(&nvme_loop_ctrl_mutex);
715 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
716 		nvme_delete_ctrl(&ctrl->ctrl);
717 	mutex_unlock(&nvme_loop_ctrl_mutex);
718 
719 	flush_workqueue(nvme_wq);
720 }
721 
722 module_init(nvme_loop_init_module);
723 module_exit(nvme_loop_cleanup_module);
724 
725 MODULE_LICENSE("GPL v2");
726 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
727