xref: /openbmc/linux/drivers/nvme/target/loop.c (revision c5017e85705bfea721732e153305d1988ff965c2)
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 /*
27  * We handle AEN commands ourselves and don't even let the
28  * block layer know about them.
29  */
30 #define NVME_LOOP_NR_AEN_COMMANDS	1
31 #define NVME_LOOP_AQ_BLKMQ_DEPTH	\
32 	(NVME_AQ_DEPTH - NVME_LOOP_NR_AEN_COMMANDS)
33 
34 struct nvme_loop_iod {
35 	struct nvme_request	nvme_req;
36 	struct nvme_command	cmd;
37 	struct nvme_completion	rsp;
38 	struct nvmet_req	req;
39 	struct nvme_loop_queue	*queue;
40 	struct work_struct	work;
41 	struct sg_table		sg_table;
42 	struct scatterlist	first_sgl[];
43 };
44 
45 struct nvme_loop_ctrl {
46 	struct nvme_loop_queue	*queues;
47 
48 	struct blk_mq_tag_set	admin_tag_set;
49 
50 	struct list_head	list;
51 	struct blk_mq_tag_set	tag_set;
52 	struct nvme_loop_iod	async_event_iod;
53 	struct nvme_ctrl	ctrl;
54 
55 	struct nvmet_ctrl	*target_ctrl;
56 };
57 
58 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
59 {
60 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
61 }
62 
63 struct nvme_loop_queue {
64 	struct nvmet_cq		nvme_cq;
65 	struct nvmet_sq		nvme_sq;
66 	struct nvme_loop_ctrl	*ctrl;
67 };
68 
69 static struct nvmet_port *nvmet_loop_port;
70 
71 static LIST_HEAD(nvme_loop_ctrl_list);
72 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
73 
74 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
75 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
76 
77 static struct nvmet_fabrics_ops nvme_loop_ops;
78 
79 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
80 {
81 	return queue - queue->ctrl->queues;
82 }
83 
84 static void nvme_loop_complete_rq(struct request *req)
85 {
86 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
87 
88 	nvme_cleanup_cmd(req);
89 	sg_free_table_chained(&iod->sg_table, true);
90 	nvme_complete_rq(req);
91 }
92 
93 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
94 {
95 	u32 queue_idx = nvme_loop_queue_idx(queue);
96 
97 	if (queue_idx == 0)
98 		return queue->ctrl->admin_tag_set.tags[queue_idx];
99 	return queue->ctrl->tag_set.tags[queue_idx - 1];
100 }
101 
102 static void nvme_loop_queue_response(struct nvmet_req *req)
103 {
104 	struct nvme_loop_queue *queue =
105 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
106 	struct nvme_completion *cqe = req->rsp;
107 
108 	/*
109 	 * AEN requests are special as they don't time out and can
110 	 * survive any kind of queue freeze and often don't respond to
111 	 * aborts.  We don't even bother to allocate a struct request
112 	 * for them but rather special case them here.
113 	 */
114 	if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
115 			cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
116 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
117 				&cqe->result);
118 	} else {
119 		struct request *rq;
120 
121 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
122 		if (!rq) {
123 			dev_err(queue->ctrl->ctrl.device,
124 				"tag 0x%x on queue %d not found\n",
125 				cqe->command_id, nvme_loop_queue_idx(queue));
126 			return;
127 		}
128 
129 		nvme_end_request(rq, cqe->status, cqe->result);
130 	}
131 }
132 
133 static void nvme_loop_execute_work(struct work_struct *work)
134 {
135 	struct nvme_loop_iod *iod =
136 		container_of(work, struct nvme_loop_iod, work);
137 
138 	iod->req.execute(&iod->req);
139 }
140 
141 static enum blk_eh_timer_return
142 nvme_loop_timeout(struct request *rq, bool reserved)
143 {
144 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
145 
146 	/* queue error recovery */
147 	nvme_reset_ctrl(&iod->queue->ctrl->ctrl);
148 
149 	/* fail with DNR on admin cmd timeout */
150 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
151 
152 	return BLK_EH_HANDLED;
153 }
154 
155 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
156 		const struct blk_mq_queue_data *bd)
157 {
158 	struct nvme_ns *ns = hctx->queue->queuedata;
159 	struct nvme_loop_queue *queue = hctx->driver_data;
160 	struct request *req = bd->rq;
161 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
162 	blk_status_t ret;
163 
164 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
165 	if (ret)
166 		return ret;
167 
168 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
169 	iod->req.port = nvmet_loop_port;
170 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
171 			&queue->nvme_sq, &nvme_loop_ops)) {
172 		nvme_cleanup_cmd(req);
173 		blk_mq_start_request(req);
174 		nvme_loop_queue_response(&iod->req);
175 		return BLK_STS_OK;
176 	}
177 
178 	if (blk_rq_bytes(req)) {
179 		iod->sg_table.sgl = iod->first_sgl;
180 		if (sg_alloc_table_chained(&iod->sg_table,
181 				blk_rq_nr_phys_segments(req),
182 				iod->sg_table.sgl))
183 			return BLK_STS_RESOURCE;
184 
185 		iod->req.sg = iod->sg_table.sgl;
186 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
187 	}
188 
189 	blk_mq_start_request(req);
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, int aer_idx)
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_LOOP_AQ_BLKMQ_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 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
278 	blk_cleanup_queue(ctrl->ctrl.admin_q);
279 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
280 }
281 
282 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
283 {
284 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
285 
286 	if (list_empty(&ctrl->list))
287 		goto free_ctrl;
288 
289 	mutex_lock(&nvme_loop_ctrl_mutex);
290 	list_del(&ctrl->list);
291 	mutex_unlock(&nvme_loop_ctrl_mutex);
292 
293 	if (nctrl->tagset) {
294 		blk_cleanup_queue(ctrl->ctrl.connect_q);
295 		blk_mq_free_tag_set(&ctrl->tag_set);
296 	}
297 	kfree(ctrl->queues);
298 	nvmf_free_options(nctrl->opts);
299 free_ctrl:
300 	kfree(ctrl);
301 }
302 
303 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
304 {
305 	int i;
306 
307 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
308 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
309 }
310 
311 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
312 {
313 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
314 	unsigned int nr_io_queues;
315 	int ret, i;
316 
317 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
318 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
319 	if (ret || !nr_io_queues)
320 		return ret;
321 
322 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
323 
324 	for (i = 1; i <= nr_io_queues; i++) {
325 		ctrl->queues[i].ctrl = ctrl;
326 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
327 		if (ret)
328 			goto out_destroy_queues;
329 
330 		ctrl->ctrl.queue_count++;
331 	}
332 
333 	return 0;
334 
335 out_destroy_queues:
336 	nvme_loop_destroy_io_queues(ctrl);
337 	return ret;
338 }
339 
340 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
341 {
342 	int i, ret;
343 
344 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
345 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
346 		if (ret)
347 			return ret;
348 	}
349 
350 	return 0;
351 }
352 
353 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
354 {
355 	int error;
356 
357 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
358 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
359 	ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
360 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
361 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
362 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
363 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
364 	ctrl->admin_tag_set.driver_data = ctrl;
365 	ctrl->admin_tag_set.nr_hw_queues = 1;
366 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
367 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
368 
369 	ctrl->queues[0].ctrl = ctrl;
370 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
371 	if (error)
372 		return error;
373 	ctrl->ctrl.queue_count = 1;
374 
375 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
376 	if (error)
377 		goto out_free_sq;
378 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
379 
380 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
381 	if (IS_ERR(ctrl->ctrl.admin_q)) {
382 		error = PTR_ERR(ctrl->ctrl.admin_q);
383 		goto out_free_tagset;
384 	}
385 
386 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
387 	if (error)
388 		goto out_cleanup_queue;
389 
390 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
391 	if (error) {
392 		dev_err(ctrl->ctrl.device,
393 			"prop_get NVME_REG_CAP failed\n");
394 		goto out_cleanup_queue;
395 	}
396 
397 	ctrl->ctrl.sqsize =
398 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
399 
400 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
401 	if (error)
402 		goto out_cleanup_queue;
403 
404 	ctrl->ctrl.max_hw_sectors =
405 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
406 
407 	error = nvme_init_identify(&ctrl->ctrl);
408 	if (error)
409 		goto out_cleanup_queue;
410 
411 	return 0;
412 
413 out_cleanup_queue:
414 	blk_cleanup_queue(ctrl->ctrl.admin_q);
415 out_free_tagset:
416 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
417 out_free_sq:
418 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
419 	return error;
420 }
421 
422 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
423 {
424 	if (ctrl->ctrl.queue_count > 1) {
425 		nvme_stop_queues(&ctrl->ctrl);
426 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
427 					nvme_cancel_request, &ctrl->ctrl);
428 		nvme_loop_destroy_io_queues(ctrl);
429 	}
430 
431 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
432 		nvme_shutdown_ctrl(&ctrl->ctrl);
433 
434 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
435 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
436 				nvme_cancel_request, &ctrl->ctrl);
437 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
438 	nvme_loop_destroy_admin_queue(ctrl);
439 }
440 
441 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
442 {
443 	nvme_stop_ctrl(ctrl);
444 	nvme_remove_namespaces(ctrl);
445 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
446 	nvme_uninit_ctrl(ctrl);
447 	nvme_put_ctrl(ctrl);
448 }
449 
450 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
451 {
452 	struct nvme_loop_ctrl *ctrl;
453 
454 	mutex_lock(&nvme_loop_ctrl_mutex);
455 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
456 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
457 			nvme_delete_ctrl(&ctrl->ctrl);
458 	}
459 	mutex_unlock(&nvme_loop_ctrl_mutex);
460 }
461 
462 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
463 {
464 	struct nvme_loop_ctrl *ctrl =
465 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
466 	bool changed;
467 	int ret;
468 
469 	nvme_stop_ctrl(&ctrl->ctrl);
470 	nvme_loop_shutdown_ctrl(ctrl);
471 
472 	ret = nvme_loop_configure_admin_queue(ctrl);
473 	if (ret)
474 		goto out_disable;
475 
476 	ret = nvme_loop_init_io_queues(ctrl);
477 	if (ret)
478 		goto out_destroy_admin;
479 
480 	ret = nvme_loop_connect_io_queues(ctrl);
481 	if (ret)
482 		goto out_destroy_io;
483 
484 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
485 			ctrl->ctrl.queue_count - 1);
486 
487 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
488 	WARN_ON_ONCE(!changed);
489 
490 	nvme_start_ctrl(&ctrl->ctrl);
491 
492 	return;
493 
494 out_destroy_io:
495 	nvme_loop_destroy_io_queues(ctrl);
496 out_destroy_admin:
497 	nvme_loop_destroy_admin_queue(ctrl);
498 out_disable:
499 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
500 	nvme_uninit_ctrl(&ctrl->ctrl);
501 	nvme_put_ctrl(&ctrl->ctrl);
502 }
503 
504 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
505 	.name			= "loop",
506 	.module			= THIS_MODULE,
507 	.flags			= NVME_F_FABRICS,
508 	.reg_read32		= nvmf_reg_read32,
509 	.reg_read64		= nvmf_reg_read64,
510 	.reg_write32		= nvmf_reg_write32,
511 	.free_ctrl		= nvme_loop_free_ctrl,
512 	.submit_async_event	= nvme_loop_submit_async_event,
513 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
514 };
515 
516 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
517 {
518 	int ret;
519 
520 	ret = nvme_loop_init_io_queues(ctrl);
521 	if (ret)
522 		return ret;
523 
524 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
525 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
526 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
527 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
528 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
529 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
530 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
531 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
532 	ctrl->tag_set.driver_data = ctrl;
533 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
534 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
535 	ctrl->ctrl.tagset = &ctrl->tag_set;
536 
537 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
538 	if (ret)
539 		goto out_destroy_queues;
540 
541 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
542 	if (IS_ERR(ctrl->ctrl.connect_q)) {
543 		ret = PTR_ERR(ctrl->ctrl.connect_q);
544 		goto out_free_tagset;
545 	}
546 
547 	ret = nvme_loop_connect_io_queues(ctrl);
548 	if (ret)
549 		goto out_cleanup_connect_q;
550 
551 	return 0;
552 
553 out_cleanup_connect_q:
554 	blk_cleanup_queue(ctrl->ctrl.connect_q);
555 out_free_tagset:
556 	blk_mq_free_tag_set(&ctrl->tag_set);
557 out_destroy_queues:
558 	nvme_loop_destroy_io_queues(ctrl);
559 	return ret;
560 }
561 
562 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
563 		struct nvmf_ctrl_options *opts)
564 {
565 	struct nvme_loop_ctrl *ctrl;
566 	bool changed;
567 	int ret;
568 
569 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
570 	if (!ctrl)
571 		return ERR_PTR(-ENOMEM);
572 	ctrl->ctrl.opts = opts;
573 	INIT_LIST_HEAD(&ctrl->list);
574 
575 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
576 
577 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
578 				0 /* no quirks, we're perfect! */);
579 	if (ret)
580 		goto out_put_ctrl;
581 
582 	ret = -ENOMEM;
583 
584 	ctrl->ctrl.sqsize = opts->queue_size - 1;
585 	ctrl->ctrl.kato = opts->kato;
586 
587 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
588 			GFP_KERNEL);
589 	if (!ctrl->queues)
590 		goto out_uninit_ctrl;
591 
592 	ret = nvme_loop_configure_admin_queue(ctrl);
593 	if (ret)
594 		goto out_free_queues;
595 
596 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
597 		/* warn if maxcmd is lower than queue_size */
598 		dev_warn(ctrl->ctrl.device,
599 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
600 			opts->queue_size, ctrl->ctrl.maxcmd);
601 		opts->queue_size = ctrl->ctrl.maxcmd;
602 	}
603 
604 	if (opts->nr_io_queues) {
605 		ret = nvme_loop_create_io_queues(ctrl);
606 		if (ret)
607 			goto out_remove_admin_queue;
608 	}
609 
610 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
611 
612 	dev_info(ctrl->ctrl.device,
613 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
614 
615 	nvme_get_ctrl(&ctrl->ctrl);
616 
617 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
618 	WARN_ON_ONCE(!changed);
619 
620 	mutex_lock(&nvme_loop_ctrl_mutex);
621 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
622 	mutex_unlock(&nvme_loop_ctrl_mutex);
623 
624 	nvme_start_ctrl(&ctrl->ctrl);
625 
626 	return &ctrl->ctrl;
627 
628 out_remove_admin_queue:
629 	nvme_loop_destroy_admin_queue(ctrl);
630 out_free_queues:
631 	kfree(ctrl->queues);
632 out_uninit_ctrl:
633 	nvme_uninit_ctrl(&ctrl->ctrl);
634 out_put_ctrl:
635 	nvme_put_ctrl(&ctrl->ctrl);
636 	if (ret > 0)
637 		ret = -EIO;
638 	return ERR_PTR(ret);
639 }
640 
641 static int nvme_loop_add_port(struct nvmet_port *port)
642 {
643 	/*
644 	 * XXX: disalow adding more than one port so
645 	 * there is no connection rejections when a
646 	 * a subsystem is assigned to a port for which
647 	 * loop doesn't have a pointer.
648 	 * This scenario would be possible if we allowed
649 	 * more than one port to be added and a subsystem
650 	 * was assigned to a port other than nvmet_loop_port.
651 	 */
652 
653 	if (nvmet_loop_port)
654 		return -EPERM;
655 
656 	nvmet_loop_port = port;
657 	return 0;
658 }
659 
660 static void nvme_loop_remove_port(struct nvmet_port *port)
661 {
662 	if (port == nvmet_loop_port)
663 		nvmet_loop_port = NULL;
664 }
665 
666 static struct nvmet_fabrics_ops nvme_loop_ops = {
667 	.owner		= THIS_MODULE,
668 	.type		= NVMF_TRTYPE_LOOP,
669 	.add_port	= nvme_loop_add_port,
670 	.remove_port	= nvme_loop_remove_port,
671 	.queue_response = nvme_loop_queue_response,
672 	.delete_ctrl	= nvme_loop_delete_ctrl,
673 };
674 
675 static struct nvmf_transport_ops nvme_loop_transport = {
676 	.name		= "loop",
677 	.create_ctrl	= nvme_loop_create_ctrl,
678 };
679 
680 static int __init nvme_loop_init_module(void)
681 {
682 	int ret;
683 
684 	ret = nvmet_register_transport(&nvme_loop_ops);
685 	if (ret)
686 		return ret;
687 
688 	ret = nvmf_register_transport(&nvme_loop_transport);
689 	if (ret)
690 		nvmet_unregister_transport(&nvme_loop_ops);
691 
692 	return ret;
693 }
694 
695 static void __exit nvme_loop_cleanup_module(void)
696 {
697 	struct nvme_loop_ctrl *ctrl, *next;
698 
699 	nvmf_unregister_transport(&nvme_loop_transport);
700 	nvmet_unregister_transport(&nvme_loop_ops);
701 
702 	mutex_lock(&nvme_loop_ctrl_mutex);
703 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
704 		nvme_delete_ctrl(&ctrl->ctrl);
705 	mutex_unlock(&nvme_loop_ctrl_mutex);
706 
707 	flush_workqueue(nvme_wq);
708 }
709 
710 module_init(nvme_loop_init_module);
711 module_exit(nvme_loop_cleanup_module);
712 
713 MODULE_LICENSE("GPL v2");
714 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
715