xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 402613f3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe over Fabrics loopback device.
4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/scatterlist.h>
8 #include <linux/blk-mq.h>
9 #include <linux/nvme.h>
10 #include <linux/module.h>
11 #include <linux/parser.h>
12 #include "nvmet.h"
13 #include "../host/nvme.h"
14 #include "../host/fabrics.h"
15 
16 #define NVME_LOOP_MAX_SEGMENTS		256
17 
18 struct nvme_loop_iod {
19 	struct nvme_request	nvme_req;
20 	struct nvme_command	cmd;
21 	struct nvme_completion	cqe;
22 	struct nvmet_req	req;
23 	struct nvme_loop_queue	*queue;
24 	struct work_struct	work;
25 	struct sg_table		sg_table;
26 	struct scatterlist	first_sgl[];
27 };
28 
29 struct nvme_loop_ctrl {
30 	struct nvme_loop_queue	*queues;
31 
32 	struct blk_mq_tag_set	admin_tag_set;
33 
34 	struct list_head	list;
35 	struct blk_mq_tag_set	tag_set;
36 	struct nvme_loop_iod	async_event_iod;
37 	struct nvme_ctrl	ctrl;
38 
39 	struct nvmet_ctrl	*target_ctrl;
40 	struct nvmet_port	*port;
41 };
42 
43 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
44 {
45 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
46 }
47 
48 enum nvme_loop_queue_flags {
49 	NVME_LOOP_Q_LIVE	= 0,
50 };
51 
52 struct nvme_loop_queue {
53 	struct nvmet_cq		nvme_cq;
54 	struct nvmet_sq		nvme_sq;
55 	struct nvme_loop_ctrl	*ctrl;
56 	unsigned long		flags;
57 };
58 
59 static LIST_HEAD(nvme_loop_ports);
60 static DEFINE_MUTEX(nvme_loop_ports_mutex);
61 
62 static LIST_HEAD(nvme_loop_ctrl_list);
63 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
64 
65 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
66 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
67 
68 static const struct nvmet_fabrics_ops nvme_loop_ops;
69 
70 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
71 {
72 	return queue - queue->ctrl->queues;
73 }
74 
75 static void nvme_loop_complete_rq(struct request *req)
76 {
77 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
78 
79 	nvme_cleanup_cmd(req);
80 	sg_free_table_chained(&iod->sg_table, SG_CHUNK_SIZE);
81 	nvme_complete_rq(req);
82 }
83 
84 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
85 {
86 	u32 queue_idx = nvme_loop_queue_idx(queue);
87 
88 	if (queue_idx == 0)
89 		return queue->ctrl->admin_tag_set.tags[queue_idx];
90 	return queue->ctrl->tag_set.tags[queue_idx - 1];
91 }
92 
93 static void nvme_loop_queue_response(struct nvmet_req *req)
94 {
95 	struct nvme_loop_queue *queue =
96 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
97 	struct nvme_completion *cqe = req->cqe;
98 
99 	/*
100 	 * AEN requests are special as they don't time out and can
101 	 * survive any kind of queue freeze and often don't respond to
102 	 * aborts.  We don't even bother to allocate a struct request
103 	 * for them but rather special case them here.
104 	 */
105 	if (unlikely(nvme_loop_queue_idx(queue) == 0 &&
106 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH)) {
107 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
108 				&cqe->result);
109 	} else {
110 		struct request *rq;
111 
112 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
113 		if (!rq) {
114 			dev_err(queue->ctrl->ctrl.device,
115 				"tag 0x%x on queue %d not found\n",
116 				cqe->command_id, nvme_loop_queue_idx(queue));
117 			return;
118 		}
119 
120 		nvme_end_request(rq, cqe->status, cqe->result);
121 	}
122 }
123 
124 static void nvme_loop_execute_work(struct work_struct *work)
125 {
126 	struct nvme_loop_iod *iod =
127 		container_of(work, struct nvme_loop_iod, work);
128 
129 	nvmet_req_execute(&iod->req);
130 }
131 
132 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
133 		const struct blk_mq_queue_data *bd)
134 {
135 	struct nvme_ns *ns = hctx->queue->queuedata;
136 	struct nvme_loop_queue *queue = hctx->driver_data;
137 	struct request *req = bd->rq;
138 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
139 	bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags);
140 	blk_status_t ret;
141 
142 	if (!nvmf_check_ready(&queue->ctrl->ctrl, req, queue_ready))
143 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, req);
144 
145 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
146 	if (ret)
147 		return ret;
148 
149 	blk_mq_start_request(req);
150 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
151 	iod->req.port = queue->ctrl->port;
152 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
153 			&queue->nvme_sq, &nvme_loop_ops))
154 		return BLK_STS_OK;
155 
156 	if (blk_rq_nr_phys_segments(req)) {
157 		iod->sg_table.sgl = iod->first_sgl;
158 		if (sg_alloc_table_chained(&iod->sg_table,
159 				blk_rq_nr_phys_segments(req),
160 				iod->sg_table.sgl, SG_CHUNK_SIZE)) {
161 			nvme_cleanup_cmd(req);
162 			return BLK_STS_RESOURCE;
163 		}
164 
165 		iod->req.sg = iod->sg_table.sgl;
166 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
167 		iod->req.transfer_len = blk_rq_payload_bytes(req);
168 	}
169 
170 	schedule_work(&iod->work);
171 	return BLK_STS_OK;
172 }
173 
174 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
175 {
176 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
177 	struct nvme_loop_queue *queue = &ctrl->queues[0];
178 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
179 
180 	memset(&iod->cmd, 0, sizeof(iod->cmd));
181 	iod->cmd.common.opcode = nvme_admin_async_event;
182 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
183 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
184 
185 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
186 			&nvme_loop_ops)) {
187 		dev_err(ctrl->ctrl.device, "failed async event work\n");
188 		return;
189 	}
190 
191 	schedule_work(&iod->work);
192 }
193 
194 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
195 		struct nvme_loop_iod *iod, unsigned int queue_idx)
196 {
197 	iod->req.cmd = &iod->cmd;
198 	iod->req.cqe = &iod->cqe;
199 	iod->queue = &ctrl->queues[queue_idx];
200 	INIT_WORK(&iod->work, nvme_loop_execute_work);
201 	return 0;
202 }
203 
204 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
205 		struct request *req, unsigned int hctx_idx,
206 		unsigned int numa_node)
207 {
208 	struct nvme_loop_ctrl *ctrl = set->driver_data;
209 
210 	nvme_req(req)->ctrl = &ctrl->ctrl;
211 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
212 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
213 }
214 
215 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
216 		unsigned int hctx_idx)
217 {
218 	struct nvme_loop_ctrl *ctrl = data;
219 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
220 
221 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
222 
223 	hctx->driver_data = queue;
224 	return 0;
225 }
226 
227 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
228 		unsigned int hctx_idx)
229 {
230 	struct nvme_loop_ctrl *ctrl = data;
231 	struct nvme_loop_queue *queue = &ctrl->queues[0];
232 
233 	BUG_ON(hctx_idx != 0);
234 
235 	hctx->driver_data = queue;
236 	return 0;
237 }
238 
239 static const struct blk_mq_ops nvme_loop_mq_ops = {
240 	.queue_rq	= nvme_loop_queue_rq,
241 	.complete	= nvme_loop_complete_rq,
242 	.init_request	= nvme_loop_init_request,
243 	.init_hctx	= nvme_loop_init_hctx,
244 };
245 
246 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
247 	.queue_rq	= nvme_loop_queue_rq,
248 	.complete	= nvme_loop_complete_rq,
249 	.init_request	= nvme_loop_init_request,
250 	.init_hctx	= nvme_loop_init_admin_hctx,
251 };
252 
253 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
254 {
255 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
256 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
257 	blk_cleanup_queue(ctrl->ctrl.admin_q);
258 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
259 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
260 }
261 
262 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
263 {
264 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
265 
266 	if (list_empty(&ctrl->list))
267 		goto free_ctrl;
268 
269 	mutex_lock(&nvme_loop_ctrl_mutex);
270 	list_del(&ctrl->list);
271 	mutex_unlock(&nvme_loop_ctrl_mutex);
272 
273 	if (nctrl->tagset) {
274 		blk_cleanup_queue(ctrl->ctrl.connect_q);
275 		blk_mq_free_tag_set(&ctrl->tag_set);
276 	}
277 	kfree(ctrl->queues);
278 	nvmf_free_options(nctrl->opts);
279 free_ctrl:
280 	kfree(ctrl);
281 }
282 
283 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
284 {
285 	int i;
286 
287 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
288 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
289 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
290 	}
291 }
292 
293 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
294 {
295 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
296 	unsigned int nr_io_queues;
297 	int ret, i;
298 
299 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
300 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
301 	if (ret || !nr_io_queues)
302 		return ret;
303 
304 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
305 
306 	for (i = 1; i <= nr_io_queues; i++) {
307 		ctrl->queues[i].ctrl = ctrl;
308 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
309 		if (ret)
310 			goto out_destroy_queues;
311 
312 		ctrl->ctrl.queue_count++;
313 	}
314 
315 	return 0;
316 
317 out_destroy_queues:
318 	nvme_loop_destroy_io_queues(ctrl);
319 	return ret;
320 }
321 
322 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
323 {
324 	int i, ret;
325 
326 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
327 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
328 		if (ret)
329 			return ret;
330 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
331 	}
332 
333 	return 0;
334 }
335 
336 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
337 {
338 	int error;
339 
340 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
341 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
342 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
343 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
344 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
345 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
346 		SG_CHUNK_SIZE * sizeof(struct scatterlist);
347 	ctrl->admin_tag_set.driver_data = ctrl;
348 	ctrl->admin_tag_set.nr_hw_queues = 1;
349 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
350 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
351 
352 	ctrl->queues[0].ctrl = ctrl;
353 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
354 	if (error)
355 		return error;
356 	ctrl->ctrl.queue_count = 1;
357 
358 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
359 	if (error)
360 		goto out_free_sq;
361 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
362 
363 	ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
364 	if (IS_ERR(ctrl->ctrl.fabrics_q)) {
365 		error = PTR_ERR(ctrl->ctrl.fabrics_q);
366 		goto out_free_tagset;
367 	}
368 
369 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
370 	if (IS_ERR(ctrl->ctrl.admin_q)) {
371 		error = PTR_ERR(ctrl->ctrl.admin_q);
372 		goto out_cleanup_fabrics_q;
373 	}
374 
375 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
376 	if (error)
377 		goto out_cleanup_queue;
378 
379 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
380 
381 	error = nvme_enable_ctrl(&ctrl->ctrl);
382 	if (error)
383 		goto out_cleanup_queue;
384 
385 	ctrl->ctrl.max_hw_sectors =
386 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
387 
388 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
389 
390 	error = nvme_init_identify(&ctrl->ctrl);
391 	if (error)
392 		goto out_cleanup_queue;
393 
394 	return 0;
395 
396 out_cleanup_queue:
397 	blk_cleanup_queue(ctrl->ctrl.admin_q);
398 out_cleanup_fabrics_q:
399 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
400 out_free_tagset:
401 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
402 out_free_sq:
403 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
404 	return error;
405 }
406 
407 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
408 {
409 	if (ctrl->ctrl.queue_count > 1) {
410 		nvme_stop_queues(&ctrl->ctrl);
411 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
412 					nvme_cancel_request, &ctrl->ctrl);
413 		blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
414 		nvme_loop_destroy_io_queues(ctrl);
415 	}
416 
417 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
418 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
419 		nvme_shutdown_ctrl(&ctrl->ctrl);
420 
421 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
422 				nvme_cancel_request, &ctrl->ctrl);
423 	blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
424 	nvme_loop_destroy_admin_queue(ctrl);
425 }
426 
427 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
428 {
429 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
430 }
431 
432 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
433 {
434 	struct nvme_loop_ctrl *ctrl;
435 
436 	mutex_lock(&nvme_loop_ctrl_mutex);
437 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
438 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
439 			nvme_delete_ctrl(&ctrl->ctrl);
440 	}
441 	mutex_unlock(&nvme_loop_ctrl_mutex);
442 }
443 
444 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
445 {
446 	struct nvme_loop_ctrl *ctrl =
447 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
448 	bool changed;
449 	int ret;
450 
451 	nvme_stop_ctrl(&ctrl->ctrl);
452 	nvme_loop_shutdown_ctrl(ctrl);
453 
454 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
455 		/* state change failure should never happen */
456 		WARN_ON_ONCE(1);
457 		return;
458 	}
459 
460 	ret = nvme_loop_configure_admin_queue(ctrl);
461 	if (ret)
462 		goto out_disable;
463 
464 	ret = nvme_loop_init_io_queues(ctrl);
465 	if (ret)
466 		goto out_destroy_admin;
467 
468 	ret = nvme_loop_connect_io_queues(ctrl);
469 	if (ret)
470 		goto out_destroy_io;
471 
472 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
473 			ctrl->ctrl.queue_count - 1);
474 
475 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
476 	WARN_ON_ONCE(!changed);
477 
478 	nvme_start_ctrl(&ctrl->ctrl);
479 
480 	return;
481 
482 out_destroy_io:
483 	nvme_loop_destroy_io_queues(ctrl);
484 out_destroy_admin:
485 	nvme_loop_destroy_admin_queue(ctrl);
486 out_disable:
487 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
488 	nvme_uninit_ctrl(&ctrl->ctrl);
489 	nvme_put_ctrl(&ctrl->ctrl);
490 }
491 
492 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
493 	.name			= "loop",
494 	.module			= THIS_MODULE,
495 	.flags			= NVME_F_FABRICS,
496 	.reg_read32		= nvmf_reg_read32,
497 	.reg_read64		= nvmf_reg_read64,
498 	.reg_write32		= nvmf_reg_write32,
499 	.free_ctrl		= nvme_loop_free_ctrl,
500 	.submit_async_event	= nvme_loop_submit_async_event,
501 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
502 	.get_address		= nvmf_get_address,
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 nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
552 {
553 	struct nvmet_port *p, *found = NULL;
554 
555 	mutex_lock(&nvme_loop_ports_mutex);
556 	list_for_each_entry(p, &nvme_loop_ports, entry) {
557 		/* if no transport address is specified use the first port */
558 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
559 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
560 			continue;
561 		found = p;
562 		break;
563 	}
564 	mutex_unlock(&nvme_loop_ports_mutex);
565 	return found;
566 }
567 
568 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
569 		struct nvmf_ctrl_options *opts)
570 {
571 	struct nvme_loop_ctrl *ctrl;
572 	bool changed;
573 	int ret;
574 
575 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
576 	if (!ctrl)
577 		return ERR_PTR(-ENOMEM);
578 	ctrl->ctrl.opts = opts;
579 	INIT_LIST_HEAD(&ctrl->list);
580 
581 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
582 
583 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
584 				0 /* no quirks, we're perfect! */);
585 	if (ret)
586 		goto out_put_ctrl;
587 
588 	ret = -ENOMEM;
589 
590 	ctrl->ctrl.sqsize = opts->queue_size - 1;
591 	ctrl->ctrl.kato = opts->kato;
592 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
593 
594 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
595 			GFP_KERNEL);
596 	if (!ctrl->queues)
597 		goto out_uninit_ctrl;
598 
599 	ret = nvme_loop_configure_admin_queue(ctrl);
600 	if (ret)
601 		goto out_free_queues;
602 
603 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
604 		/* warn if maxcmd is lower than queue_size */
605 		dev_warn(ctrl->ctrl.device,
606 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
607 			opts->queue_size, ctrl->ctrl.maxcmd);
608 		opts->queue_size = ctrl->ctrl.maxcmd;
609 	}
610 
611 	if (opts->nr_io_queues) {
612 		ret = nvme_loop_create_io_queues(ctrl);
613 		if (ret)
614 			goto out_remove_admin_queue;
615 	}
616 
617 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
618 
619 	dev_info(ctrl->ctrl.device,
620 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
621 
622 	nvme_get_ctrl(&ctrl->ctrl);
623 
624 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
625 	WARN_ON_ONCE(!changed);
626 
627 	mutex_lock(&nvme_loop_ctrl_mutex);
628 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
629 	mutex_unlock(&nvme_loop_ctrl_mutex);
630 
631 	nvme_start_ctrl(&ctrl->ctrl);
632 
633 	return &ctrl->ctrl;
634 
635 out_remove_admin_queue:
636 	nvme_loop_destroy_admin_queue(ctrl);
637 out_free_queues:
638 	kfree(ctrl->queues);
639 out_uninit_ctrl:
640 	nvme_uninit_ctrl(&ctrl->ctrl);
641 out_put_ctrl:
642 	nvme_put_ctrl(&ctrl->ctrl);
643 	if (ret > 0)
644 		ret = -EIO;
645 	return ERR_PTR(ret);
646 }
647 
648 static int nvme_loop_add_port(struct nvmet_port *port)
649 {
650 	mutex_lock(&nvme_loop_ports_mutex);
651 	list_add_tail(&port->entry, &nvme_loop_ports);
652 	mutex_unlock(&nvme_loop_ports_mutex);
653 	return 0;
654 }
655 
656 static void nvme_loop_remove_port(struct nvmet_port *port)
657 {
658 	mutex_lock(&nvme_loop_ports_mutex);
659 	list_del_init(&port->entry);
660 	mutex_unlock(&nvme_loop_ports_mutex);
661 
662 	/*
663 	 * Ensure any ctrls that are in the process of being
664 	 * deleted are in fact deleted before we return
665 	 * and free the port. This is to prevent active
666 	 * ctrls from using a port after it's freed.
667 	 */
668 	flush_workqueue(nvme_delete_wq);
669 }
670 
671 static const struct nvmet_fabrics_ops nvme_loop_ops = {
672 	.owner		= THIS_MODULE,
673 	.type		= NVMF_TRTYPE_LOOP,
674 	.add_port	= nvme_loop_add_port,
675 	.remove_port	= nvme_loop_remove_port,
676 	.queue_response = nvme_loop_queue_response,
677 	.delete_ctrl	= nvme_loop_delete_ctrl,
678 };
679 
680 static struct nvmf_transport_ops nvme_loop_transport = {
681 	.name		= "loop",
682 	.module		= THIS_MODULE,
683 	.create_ctrl	= nvme_loop_create_ctrl,
684 	.allowed_opts	= NVMF_OPT_TRADDR,
685 };
686 
687 static int __init nvme_loop_init_module(void)
688 {
689 	int ret;
690 
691 	ret = nvmet_register_transport(&nvme_loop_ops);
692 	if (ret)
693 		return ret;
694 
695 	ret = nvmf_register_transport(&nvme_loop_transport);
696 	if (ret)
697 		nvmet_unregister_transport(&nvme_loop_ops);
698 
699 	return ret;
700 }
701 
702 static void __exit nvme_loop_cleanup_module(void)
703 {
704 	struct nvme_loop_ctrl *ctrl, *next;
705 
706 	nvmf_unregister_transport(&nvme_loop_transport);
707 	nvmet_unregister_transport(&nvme_loop_ops);
708 
709 	mutex_lock(&nvme_loop_ctrl_mutex);
710 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
711 		nvme_delete_ctrl(&ctrl->ctrl);
712 	mutex_unlock(&nvme_loop_ctrl_mutex);
713 
714 	flush_workqueue(nvme_delete_wq);
715 }
716 
717 module_init(nvme_loop_init_module);
718 module_exit(nvme_loop_cleanup_module);
719 
720 MODULE_LICENSE("GPL v2");
721 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
722