xref: /openbmc/linux/drivers/nvme/target/loop.c (revision ba61bb17)
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 	bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags);
162 	blk_status_t ret;
163 
164 	if (!nvmf_check_ready(&queue->ctrl->ctrl, req, queue_ready))
165 		return nvmf_fail_nonready_command(req);
166 
167 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
168 	if (ret)
169 		return ret;
170 
171 	blk_mq_start_request(req);
172 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
173 	iod->req.port = queue->ctrl->port;
174 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
175 			&queue->nvme_sq, &nvme_loop_ops))
176 		return BLK_STS_OK;
177 
178 	if (blk_rq_nr_phys_segments(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 		iod->req.transfer_len = blk_rq_payload_bytes(req);
188 	}
189 
190 	schedule_work(&iod->work);
191 	return BLK_STS_OK;
192 }
193 
194 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
195 {
196 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
197 	struct nvme_loop_queue *queue = &ctrl->queues[0];
198 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
199 
200 	memset(&iod->cmd, 0, sizeof(iod->cmd));
201 	iod->cmd.common.opcode = nvme_admin_async_event;
202 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
203 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
204 
205 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
206 			&nvme_loop_ops)) {
207 		dev_err(ctrl->ctrl.device, "failed async event work\n");
208 		return;
209 	}
210 
211 	schedule_work(&iod->work);
212 }
213 
214 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
215 		struct nvme_loop_iod *iod, unsigned int queue_idx)
216 {
217 	iod->req.cmd = &iod->cmd;
218 	iod->req.rsp = &iod->rsp;
219 	iod->queue = &ctrl->queues[queue_idx];
220 	INIT_WORK(&iod->work, nvme_loop_execute_work);
221 	return 0;
222 }
223 
224 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
225 		struct request *req, unsigned int hctx_idx,
226 		unsigned int numa_node)
227 {
228 	struct nvme_loop_ctrl *ctrl = set->driver_data;
229 
230 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
231 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
232 }
233 
234 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
235 		unsigned int hctx_idx)
236 {
237 	struct nvme_loop_ctrl *ctrl = data;
238 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
239 
240 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
241 
242 	hctx->driver_data = queue;
243 	return 0;
244 }
245 
246 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
247 		unsigned int hctx_idx)
248 {
249 	struct nvme_loop_ctrl *ctrl = data;
250 	struct nvme_loop_queue *queue = &ctrl->queues[0];
251 
252 	BUG_ON(hctx_idx != 0);
253 
254 	hctx->driver_data = queue;
255 	return 0;
256 }
257 
258 static const struct blk_mq_ops nvme_loop_mq_ops = {
259 	.queue_rq	= nvme_loop_queue_rq,
260 	.complete	= nvme_loop_complete_rq,
261 	.init_request	= nvme_loop_init_request,
262 	.init_hctx	= nvme_loop_init_hctx,
263 	.timeout	= nvme_loop_timeout,
264 };
265 
266 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
267 	.queue_rq	= nvme_loop_queue_rq,
268 	.complete	= nvme_loop_complete_rq,
269 	.init_request	= nvme_loop_init_request,
270 	.init_hctx	= nvme_loop_init_admin_hctx,
271 	.timeout	= nvme_loop_timeout,
272 };
273 
274 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
275 {
276 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
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 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
309 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
310 	}
311 }
312 
313 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
314 {
315 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
316 	unsigned int nr_io_queues;
317 	int ret, i;
318 
319 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
320 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
321 	if (ret || !nr_io_queues)
322 		return ret;
323 
324 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
325 
326 	for (i = 1; i <= nr_io_queues; i++) {
327 		ctrl->queues[i].ctrl = ctrl;
328 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
329 		if (ret)
330 			goto out_destroy_queues;
331 
332 		ctrl->ctrl.queue_count++;
333 	}
334 
335 	return 0;
336 
337 out_destroy_queues:
338 	nvme_loop_destroy_io_queues(ctrl);
339 	return ret;
340 }
341 
342 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
343 {
344 	int i, ret;
345 
346 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
347 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
348 		if (ret)
349 			return ret;
350 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
351 	}
352 
353 	return 0;
354 }
355 
356 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
357 {
358 	int error;
359 
360 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
361 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
362 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
363 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
364 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
365 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
366 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
367 	ctrl->admin_tag_set.driver_data = ctrl;
368 	ctrl->admin_tag_set.nr_hw_queues = 1;
369 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
370 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
371 
372 	ctrl->queues[0].ctrl = ctrl;
373 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
374 	if (error)
375 		return error;
376 	ctrl->ctrl.queue_count = 1;
377 
378 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
379 	if (error)
380 		goto out_free_sq;
381 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
382 
383 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
384 	if (IS_ERR(ctrl->ctrl.admin_q)) {
385 		error = PTR_ERR(ctrl->ctrl.admin_q);
386 		goto out_free_tagset;
387 	}
388 
389 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
390 	if (error)
391 		goto out_cleanup_queue;
392 
393 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
394 
395 	error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->ctrl.cap);
396 	if (error) {
397 		dev_err(ctrl->ctrl.device,
398 			"prop_get NVME_REG_CAP failed\n");
399 		goto out_cleanup_queue;
400 	}
401 
402 	ctrl->ctrl.sqsize =
403 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
404 
405 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
406 	if (error)
407 		goto out_cleanup_queue;
408 
409 	ctrl->ctrl.max_hw_sectors =
410 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
411 
412 	error = nvme_init_identify(&ctrl->ctrl);
413 	if (error)
414 		goto out_cleanup_queue;
415 
416 	return 0;
417 
418 out_cleanup_queue:
419 	blk_cleanup_queue(ctrl->ctrl.admin_q);
420 out_free_tagset:
421 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
422 out_free_sq:
423 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
424 	return error;
425 }
426 
427 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
428 {
429 	if (ctrl->ctrl.queue_count > 1) {
430 		nvme_stop_queues(&ctrl->ctrl);
431 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
432 					nvme_cancel_request, &ctrl->ctrl);
433 		nvme_loop_destroy_io_queues(ctrl);
434 	}
435 
436 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
437 		nvme_shutdown_ctrl(&ctrl->ctrl);
438 
439 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
440 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
441 				nvme_cancel_request, &ctrl->ctrl);
442 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
443 	nvme_loop_destroy_admin_queue(ctrl);
444 }
445 
446 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
447 {
448 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
449 }
450 
451 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
452 {
453 	struct nvme_loop_ctrl *ctrl;
454 
455 	mutex_lock(&nvme_loop_ctrl_mutex);
456 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
457 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
458 			nvme_delete_ctrl(&ctrl->ctrl);
459 	}
460 	mutex_unlock(&nvme_loop_ctrl_mutex);
461 }
462 
463 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
464 {
465 	struct nvme_loop_ctrl *ctrl =
466 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
467 	bool changed;
468 	int ret;
469 
470 	nvme_stop_ctrl(&ctrl->ctrl);
471 	nvme_loop_shutdown_ctrl(ctrl);
472 
473 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
474 		/* state change failure should never happen */
475 		WARN_ON_ONCE(1);
476 		return;
477 	}
478 
479 	ret = nvme_loop_configure_admin_queue(ctrl);
480 	if (ret)
481 		goto out_disable;
482 
483 	ret = nvme_loop_init_io_queues(ctrl);
484 	if (ret)
485 		goto out_destroy_admin;
486 
487 	ret = nvme_loop_connect_io_queues(ctrl);
488 	if (ret)
489 		goto out_destroy_io;
490 
491 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
492 			ctrl->ctrl.queue_count - 1);
493 
494 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
495 	WARN_ON_ONCE(!changed);
496 
497 	nvme_start_ctrl(&ctrl->ctrl);
498 
499 	return;
500 
501 out_destroy_io:
502 	nvme_loop_destroy_io_queues(ctrl);
503 out_destroy_admin:
504 	nvme_loop_destroy_admin_queue(ctrl);
505 out_disable:
506 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
507 	nvme_uninit_ctrl(&ctrl->ctrl);
508 	nvme_put_ctrl(&ctrl->ctrl);
509 }
510 
511 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
512 	.name			= "loop",
513 	.module			= THIS_MODULE,
514 	.flags			= NVME_F_FABRICS,
515 	.reg_read32		= nvmf_reg_read32,
516 	.reg_read64		= nvmf_reg_read64,
517 	.reg_write32		= nvmf_reg_write32,
518 	.free_ctrl		= nvme_loop_free_ctrl,
519 	.submit_async_event	= nvme_loop_submit_async_event,
520 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
521 	.get_address		= nvmf_get_address,
522 };
523 
524 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
525 {
526 	int ret;
527 
528 	ret = nvme_loop_init_io_queues(ctrl);
529 	if (ret)
530 		return ret;
531 
532 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
533 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
534 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
535 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
536 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
537 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
538 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
539 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
540 	ctrl->tag_set.driver_data = ctrl;
541 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
542 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
543 	ctrl->ctrl.tagset = &ctrl->tag_set;
544 
545 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
546 	if (ret)
547 		goto out_destroy_queues;
548 
549 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
550 	if (IS_ERR(ctrl->ctrl.connect_q)) {
551 		ret = PTR_ERR(ctrl->ctrl.connect_q);
552 		goto out_free_tagset;
553 	}
554 
555 	ret = nvme_loop_connect_io_queues(ctrl);
556 	if (ret)
557 		goto out_cleanup_connect_q;
558 
559 	return 0;
560 
561 out_cleanup_connect_q:
562 	blk_cleanup_queue(ctrl->ctrl.connect_q);
563 out_free_tagset:
564 	blk_mq_free_tag_set(&ctrl->tag_set);
565 out_destroy_queues:
566 	nvme_loop_destroy_io_queues(ctrl);
567 	return ret;
568 }
569 
570 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
571 {
572 	struct nvmet_port *p, *found = NULL;
573 
574 	mutex_lock(&nvme_loop_ports_mutex);
575 	list_for_each_entry(p, &nvme_loop_ports, entry) {
576 		/* if no transport address is specified use the first port */
577 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
578 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
579 			continue;
580 		found = p;
581 		break;
582 	}
583 	mutex_unlock(&nvme_loop_ports_mutex);
584 	return found;
585 }
586 
587 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
588 		struct nvmf_ctrl_options *opts)
589 {
590 	struct nvme_loop_ctrl *ctrl;
591 	bool changed;
592 	int ret;
593 
594 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
595 	if (!ctrl)
596 		return ERR_PTR(-ENOMEM);
597 	ctrl->ctrl.opts = opts;
598 	INIT_LIST_HEAD(&ctrl->list);
599 
600 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
601 
602 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
603 				0 /* no quirks, we're perfect! */);
604 	if (ret)
605 		goto out_put_ctrl;
606 
607 	ret = -ENOMEM;
608 
609 	ctrl->ctrl.sqsize = opts->queue_size - 1;
610 	ctrl->ctrl.kato = opts->kato;
611 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
612 
613 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
614 			GFP_KERNEL);
615 	if (!ctrl->queues)
616 		goto out_uninit_ctrl;
617 
618 	ret = nvme_loop_configure_admin_queue(ctrl);
619 	if (ret)
620 		goto out_free_queues;
621 
622 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
623 		/* warn if maxcmd is lower than queue_size */
624 		dev_warn(ctrl->ctrl.device,
625 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
626 			opts->queue_size, ctrl->ctrl.maxcmd);
627 		opts->queue_size = ctrl->ctrl.maxcmd;
628 	}
629 
630 	if (opts->nr_io_queues) {
631 		ret = nvme_loop_create_io_queues(ctrl);
632 		if (ret)
633 			goto out_remove_admin_queue;
634 	}
635 
636 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
637 
638 	dev_info(ctrl->ctrl.device,
639 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
640 
641 	nvme_get_ctrl(&ctrl->ctrl);
642 
643 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
644 	WARN_ON_ONCE(!changed);
645 
646 	mutex_lock(&nvme_loop_ctrl_mutex);
647 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
648 	mutex_unlock(&nvme_loop_ctrl_mutex);
649 
650 	nvme_start_ctrl(&ctrl->ctrl);
651 
652 	return &ctrl->ctrl;
653 
654 out_remove_admin_queue:
655 	nvme_loop_destroy_admin_queue(ctrl);
656 out_free_queues:
657 	kfree(ctrl->queues);
658 out_uninit_ctrl:
659 	nvme_uninit_ctrl(&ctrl->ctrl);
660 out_put_ctrl:
661 	nvme_put_ctrl(&ctrl->ctrl);
662 	if (ret > 0)
663 		ret = -EIO;
664 	return ERR_PTR(ret);
665 }
666 
667 static int nvme_loop_add_port(struct nvmet_port *port)
668 {
669 	mutex_lock(&nvme_loop_ports_mutex);
670 	list_add_tail(&port->entry, &nvme_loop_ports);
671 	mutex_unlock(&nvme_loop_ports_mutex);
672 	return 0;
673 }
674 
675 static void nvme_loop_remove_port(struct nvmet_port *port)
676 {
677 	mutex_lock(&nvme_loop_ports_mutex);
678 	list_del_init(&port->entry);
679 	mutex_unlock(&nvme_loop_ports_mutex);
680 }
681 
682 static const struct nvmet_fabrics_ops nvme_loop_ops = {
683 	.owner		= THIS_MODULE,
684 	.type		= NVMF_TRTYPE_LOOP,
685 	.add_port	= nvme_loop_add_port,
686 	.remove_port	= nvme_loop_remove_port,
687 	.queue_response = nvme_loop_queue_response,
688 	.delete_ctrl	= nvme_loop_delete_ctrl,
689 };
690 
691 static struct nvmf_transport_ops nvme_loop_transport = {
692 	.name		= "loop",
693 	.module		= THIS_MODULE,
694 	.create_ctrl	= nvme_loop_create_ctrl,
695 	.allowed_opts	= NVMF_OPT_TRADDR,
696 };
697 
698 static int __init nvme_loop_init_module(void)
699 {
700 	int ret;
701 
702 	ret = nvmet_register_transport(&nvme_loop_ops);
703 	if (ret)
704 		return ret;
705 
706 	ret = nvmf_register_transport(&nvme_loop_transport);
707 	if (ret)
708 		nvmet_unregister_transport(&nvme_loop_ops);
709 
710 	return ret;
711 }
712 
713 static void __exit nvme_loop_cleanup_module(void)
714 {
715 	struct nvme_loop_ctrl *ctrl, *next;
716 
717 	nvmf_unregister_transport(&nvme_loop_transport);
718 	nvmet_unregister_transport(&nvme_loop_ops);
719 
720 	mutex_lock(&nvme_loop_ctrl_mutex);
721 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
722 		nvme_delete_ctrl(&ctrl->ctrl);
723 	mutex_unlock(&nvme_loop_ctrl_mutex);
724 
725 	flush_workqueue(nvme_delete_wq);
726 }
727 
728 module_init(nvme_loop_init_module);
729 module_exit(nvme_loop_cleanup_module);
730 
731 MODULE_LICENSE("GPL v2");
732 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
733