xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 94d964e5)
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_port	*port;
40 };
41 
42 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
43 {
44 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
45 }
46 
47 enum nvme_loop_queue_flags {
48 	NVME_LOOP_Q_LIVE	= 0,
49 };
50 
51 struct nvme_loop_queue {
52 	struct nvmet_cq		nvme_cq;
53 	struct nvmet_sq		nvme_sq;
54 	struct nvme_loop_ctrl	*ctrl;
55 	unsigned long		flags;
56 };
57 
58 static LIST_HEAD(nvme_loop_ports);
59 static DEFINE_MUTEX(nvme_loop_ports_mutex);
60 
61 static LIST_HEAD(nvme_loop_ctrl_list);
62 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
63 
64 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
65 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
66 
67 static const struct nvmet_fabrics_ops nvme_loop_ops;
68 
69 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
70 {
71 	return queue - queue->ctrl->queues;
72 }
73 
74 static void nvme_loop_complete_rq(struct request *req)
75 {
76 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
77 
78 	sg_free_table_chained(&iod->sg_table, NVME_INLINE_SG_CNT);
79 	nvme_complete_rq(req);
80 }
81 
82 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
83 {
84 	u32 queue_idx = nvme_loop_queue_idx(queue);
85 
86 	if (queue_idx == 0)
87 		return queue->ctrl->admin_tag_set.tags[queue_idx];
88 	return queue->ctrl->tag_set.tags[queue_idx - 1];
89 }
90 
91 static void nvme_loop_queue_response(struct nvmet_req *req)
92 {
93 	struct nvme_loop_queue *queue =
94 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
95 	struct nvme_completion *cqe = req->cqe;
96 
97 	/*
98 	 * AEN requests are special as they don't time out and can
99 	 * survive any kind of queue freeze and often don't respond to
100 	 * aborts.  We don't even bother to allocate a struct request
101 	 * for them but rather special case them here.
102 	 */
103 	if (unlikely(nvme_is_aen_req(nvme_loop_queue_idx(queue),
104 				     cqe->command_id))) {
105 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
106 				&cqe->result);
107 	} else {
108 		struct request *rq;
109 
110 		rq = nvme_find_rq(nvme_loop_tagset(queue), cqe->command_id);
111 		if (!rq) {
112 			dev_err(queue->ctrl->ctrl.device,
113 				"got bad command_id %#x on queue %d\n",
114 				cqe->command_id, nvme_loop_queue_idx(queue));
115 			return;
116 		}
117 
118 		if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
119 			nvme_loop_complete_rq(rq);
120 	}
121 }
122 
123 static void nvme_loop_execute_work(struct work_struct *work)
124 {
125 	struct nvme_loop_iod *iod =
126 		container_of(work, struct nvme_loop_iod, work);
127 
128 	iod->req.execute(&iod->req);
129 }
130 
131 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
132 		const struct blk_mq_queue_data *bd)
133 {
134 	struct nvme_ns *ns = hctx->queue->queuedata;
135 	struct nvme_loop_queue *queue = hctx->driver_data;
136 	struct request *req = bd->rq;
137 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
138 	bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags);
139 	blk_status_t ret;
140 
141 	if (!nvme_check_ready(&queue->ctrl->ctrl, req, queue_ready))
142 		return nvme_fail_nonready_command(&queue->ctrl->ctrl, req);
143 
144 	ret = nvme_setup_cmd(ns, req);
145 	if (ret)
146 		return ret;
147 
148 	blk_mq_start_request(req);
149 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
150 	iod->req.port = queue->ctrl->port;
151 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
152 			&queue->nvme_sq, &nvme_loop_ops))
153 		return BLK_STS_OK;
154 
155 	if (blk_rq_nr_phys_segments(req)) {
156 		iod->sg_table.sgl = iod->first_sgl;
157 		if (sg_alloc_table_chained(&iod->sg_table,
158 				blk_rq_nr_phys_segments(req),
159 				iod->sg_table.sgl, NVME_INLINE_SG_CNT)) {
160 			nvme_cleanup_cmd(req);
161 			return BLK_STS_RESOURCE;
162 		}
163 
164 		iod->req.sg = iod->sg_table.sgl;
165 		iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
166 		iod->req.transfer_len = blk_rq_payload_bytes(req);
167 	}
168 
169 	schedule_work(&iod->work);
170 	return BLK_STS_OK;
171 }
172 
173 static void nvme_loop_submit_async_event(struct nvme_ctrl *arg)
174 {
175 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
176 	struct nvme_loop_queue *queue = &ctrl->queues[0];
177 	struct nvme_loop_iod *iod = &ctrl->async_event_iod;
178 
179 	memset(&iod->cmd, 0, sizeof(iod->cmd));
180 	iod->cmd.common.opcode = nvme_admin_async_event;
181 	iod->cmd.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
182 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
183 
184 	if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
185 			&nvme_loop_ops)) {
186 		dev_err(ctrl->ctrl.device, "failed async event work\n");
187 		return;
188 	}
189 
190 	schedule_work(&iod->work);
191 }
192 
193 static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
194 		struct nvme_loop_iod *iod, unsigned int queue_idx)
195 {
196 	iod->req.cmd = &iod->cmd;
197 	iod->req.cqe = &iod->cqe;
198 	iod->queue = &ctrl->queues[queue_idx];
199 	INIT_WORK(&iod->work, nvme_loop_execute_work);
200 	return 0;
201 }
202 
203 static int nvme_loop_init_request(struct blk_mq_tag_set *set,
204 		struct request *req, unsigned int hctx_idx,
205 		unsigned int numa_node)
206 {
207 	struct nvme_loop_ctrl *ctrl = set->driver_data;
208 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
209 
210 	nvme_req(req)->ctrl = &ctrl->ctrl;
211 	nvme_req(req)->cmd = &iod->cmd;
212 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
213 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
214 }
215 
216 static struct lock_class_key loop_hctx_fq_lock_key;
217 
218 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
219 		unsigned int hctx_idx)
220 {
221 	struct nvme_loop_ctrl *ctrl = data;
222 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
223 
224 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
225 
226 	/*
227 	 * flush_end_io() can be called recursively for us, so use our own
228 	 * lock class key for avoiding lockdep possible recursive locking,
229 	 * then we can remove the dynamically allocated lock class for each
230 	 * flush queue, that way may cause horrible boot delay.
231 	 */
232 	blk_mq_hctx_set_fq_lock_class(hctx, &loop_hctx_fq_lock_key);
233 
234 	hctx->driver_data = queue;
235 	return 0;
236 }
237 
238 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
239 		unsigned int hctx_idx)
240 {
241 	struct nvme_loop_ctrl *ctrl = data;
242 	struct nvme_loop_queue *queue = &ctrl->queues[0];
243 
244 	BUG_ON(hctx_idx != 0);
245 
246 	hctx->driver_data = queue;
247 	return 0;
248 }
249 
250 static const struct blk_mq_ops nvme_loop_mq_ops = {
251 	.queue_rq	= nvme_loop_queue_rq,
252 	.complete	= nvme_loop_complete_rq,
253 	.init_request	= nvme_loop_init_request,
254 	.init_hctx	= nvme_loop_init_hctx,
255 };
256 
257 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
258 	.queue_rq	= nvme_loop_queue_rq,
259 	.complete	= nvme_loop_complete_rq,
260 	.init_request	= nvme_loop_init_request,
261 	.init_hctx	= nvme_loop_init_admin_hctx,
262 };
263 
264 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
265 {
266 	if (!test_and_clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags))
267 		return;
268 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
269 	blk_cleanup_queue(ctrl->ctrl.admin_q);
270 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
271 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
272 }
273 
274 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
275 {
276 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
277 
278 	if (list_empty(&ctrl->list))
279 		goto free_ctrl;
280 
281 	mutex_lock(&nvme_loop_ctrl_mutex);
282 	list_del(&ctrl->list);
283 	mutex_unlock(&nvme_loop_ctrl_mutex);
284 
285 	if (nctrl->tagset) {
286 		blk_cleanup_queue(ctrl->ctrl.connect_q);
287 		blk_mq_free_tag_set(&ctrl->tag_set);
288 	}
289 	kfree(ctrl->queues);
290 	nvmf_free_options(nctrl->opts);
291 free_ctrl:
292 	kfree(ctrl);
293 }
294 
295 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
296 {
297 	int i;
298 
299 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
300 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
301 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
302 	}
303 	ctrl->ctrl.queue_count = 1;
304 }
305 
306 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
307 {
308 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
309 	unsigned int nr_io_queues;
310 	int ret, i;
311 
312 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
313 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
314 	if (ret || !nr_io_queues)
315 		return ret;
316 
317 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
318 
319 	for (i = 1; i <= nr_io_queues; i++) {
320 		ctrl->queues[i].ctrl = ctrl;
321 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
322 		if (ret)
323 			goto out_destroy_queues;
324 
325 		ctrl->ctrl.queue_count++;
326 	}
327 
328 	return 0;
329 
330 out_destroy_queues:
331 	nvme_loop_destroy_io_queues(ctrl);
332 	return ret;
333 }
334 
335 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
336 {
337 	int i, ret;
338 
339 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
340 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
341 		if (ret)
342 			return ret;
343 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
344 	}
345 
346 	return 0;
347 }
348 
349 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
350 {
351 	int error;
352 
353 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
354 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
355 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
356 	ctrl->admin_tag_set.reserved_tags = NVMF_RESERVED_TAGS;
357 	ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
358 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
359 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
360 	ctrl->admin_tag_set.driver_data = ctrl;
361 	ctrl->admin_tag_set.nr_hw_queues = 1;
362 	ctrl->admin_tag_set.timeout = NVME_ADMIN_TIMEOUT;
363 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
364 
365 	ctrl->queues[0].ctrl = ctrl;
366 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
367 	if (error)
368 		return error;
369 	ctrl->ctrl.queue_count = 1;
370 
371 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
372 	if (error)
373 		goto out_free_sq;
374 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
375 
376 	ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
377 	if (IS_ERR(ctrl->ctrl.fabrics_q)) {
378 		error = PTR_ERR(ctrl->ctrl.fabrics_q);
379 		goto out_free_tagset;
380 	}
381 
382 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
383 	if (IS_ERR(ctrl->ctrl.admin_q)) {
384 		error = PTR_ERR(ctrl->ctrl.admin_q);
385 		goto out_cleanup_fabrics_q;
386 	}
387 	/* reset stopped state for the fresh admin queue */
388 	clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->ctrl.flags);
389 
390 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
391 	if (error)
392 		goto out_cleanup_queue;
393 
394 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
395 
396 	error = nvme_enable_ctrl(&ctrl->ctrl);
397 	if (error)
398 		goto out_cleanup_queue;
399 
400 	ctrl->ctrl.max_hw_sectors =
401 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
402 
403 	nvme_start_admin_queue(&ctrl->ctrl);
404 
405 	error = nvme_init_ctrl_finish(&ctrl->ctrl);
406 	if (error)
407 		goto out_cleanup_queue;
408 
409 	return 0;
410 
411 out_cleanup_queue:
412 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
413 	blk_cleanup_queue(ctrl->ctrl.admin_q);
414 out_cleanup_fabrics_q:
415 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
416 out_free_tagset:
417 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
418 out_free_sq:
419 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
420 	return error;
421 }
422 
423 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
424 {
425 	if (ctrl->ctrl.queue_count > 1) {
426 		nvme_stop_queues(&ctrl->ctrl);
427 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
428 					nvme_cancel_request, &ctrl->ctrl);
429 		blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
430 		nvme_loop_destroy_io_queues(ctrl);
431 	}
432 
433 	nvme_stop_admin_queue(&ctrl->ctrl);
434 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
435 		nvme_shutdown_ctrl(&ctrl->ctrl);
436 
437 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
438 				nvme_cancel_request, &ctrl->ctrl);
439 	blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
440 	nvme_loop_destroy_admin_queue(ctrl);
441 }
442 
443 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
444 {
445 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
446 }
447 
448 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
449 {
450 	struct nvme_loop_ctrl *ctrl;
451 
452 	mutex_lock(&nvme_loop_ctrl_mutex);
453 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
454 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
455 			nvme_delete_ctrl(&ctrl->ctrl);
456 	}
457 	mutex_unlock(&nvme_loop_ctrl_mutex);
458 }
459 
460 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
461 {
462 	struct nvme_loop_ctrl *ctrl =
463 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
464 	int ret;
465 
466 	nvme_stop_ctrl(&ctrl->ctrl);
467 	nvme_loop_shutdown_ctrl(ctrl);
468 
469 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
470 		if (ctrl->ctrl.state != NVME_CTRL_DELETING &&
471 		    ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO)
472 			/* state change failure for non-deleted ctrl? */
473 			WARN_ON_ONCE(1);
474 		return;
475 	}
476 
477 	ret = nvme_loop_configure_admin_queue(ctrl);
478 	if (ret)
479 		goto out_disable;
480 
481 	ret = nvme_loop_init_io_queues(ctrl);
482 	if (ret)
483 		goto out_destroy_admin;
484 
485 	ret = nvme_loop_connect_io_queues(ctrl);
486 	if (ret)
487 		goto out_destroy_io;
488 
489 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
490 			ctrl->ctrl.queue_count - 1);
491 
492 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
493 		WARN_ON_ONCE(1);
494 
495 	nvme_start_ctrl(&ctrl->ctrl);
496 
497 	return;
498 
499 out_destroy_io:
500 	nvme_loop_destroy_io_queues(ctrl);
501 out_destroy_admin:
502 	nvme_loop_destroy_admin_queue(ctrl);
503 out_disable:
504 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
505 	nvme_uninit_ctrl(&ctrl->ctrl);
506 }
507 
508 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
509 	.name			= "loop",
510 	.module			= THIS_MODULE,
511 	.flags			= NVME_F_FABRICS,
512 	.reg_read32		= nvmf_reg_read32,
513 	.reg_read64		= nvmf_reg_read64,
514 	.reg_write32		= nvmf_reg_write32,
515 	.free_ctrl		= nvme_loop_free_ctrl,
516 	.submit_async_event	= nvme_loop_submit_async_event,
517 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
518 	.get_address		= nvmf_get_address,
519 };
520 
521 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
522 {
523 	int ret;
524 
525 	ret = nvme_loop_init_io_queues(ctrl);
526 	if (ret)
527 		return ret;
528 
529 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
530 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
531 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
532 	ctrl->tag_set.reserved_tags = NVMF_RESERVED_TAGS;
533 	ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
534 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
535 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
536 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
537 	ctrl->tag_set.driver_data = ctrl;
538 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
539 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
540 	ctrl->ctrl.tagset = &ctrl->tag_set;
541 
542 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
543 	if (ret)
544 		goto out_destroy_queues;
545 
546 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
547 	if (IS_ERR(ctrl->ctrl.connect_q)) {
548 		ret = PTR_ERR(ctrl->ctrl.connect_q);
549 		goto out_free_tagset;
550 	}
551 
552 	ret = nvme_loop_connect_io_queues(ctrl);
553 	if (ret)
554 		goto out_cleanup_connect_q;
555 
556 	return 0;
557 
558 out_cleanup_connect_q:
559 	blk_cleanup_queue(ctrl->ctrl.connect_q);
560 out_free_tagset:
561 	blk_mq_free_tag_set(&ctrl->tag_set);
562 out_destroy_queues:
563 	nvme_loop_destroy_io_queues(ctrl);
564 	return ret;
565 }
566 
567 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
568 {
569 	struct nvmet_port *p, *found = NULL;
570 
571 	mutex_lock(&nvme_loop_ports_mutex);
572 	list_for_each_entry(p, &nvme_loop_ports, entry) {
573 		/* if no transport address is specified use the first port */
574 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
575 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
576 			continue;
577 		found = p;
578 		break;
579 	}
580 	mutex_unlock(&nvme_loop_ports_mutex);
581 	return found;
582 }
583 
584 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
585 		struct nvmf_ctrl_options *opts)
586 {
587 	struct nvme_loop_ctrl *ctrl;
588 	int ret;
589 
590 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
591 	if (!ctrl)
592 		return ERR_PTR(-ENOMEM);
593 	ctrl->ctrl.opts = opts;
594 	INIT_LIST_HEAD(&ctrl->list);
595 
596 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
597 
598 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
599 				0 /* no quirks, we're perfect! */);
600 	if (ret) {
601 		kfree(ctrl);
602 		goto out;
603 	}
604 
605 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
606 		WARN_ON_ONCE(1);
607 
608 	ret = -ENOMEM;
609 
610 	ctrl->ctrl.sqsize = opts->queue_size - 1;
611 	ctrl->ctrl.kato = opts->kato;
612 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
613 
614 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
615 			GFP_KERNEL);
616 	if (!ctrl->queues)
617 		goto out_uninit_ctrl;
618 
619 	ret = nvme_loop_configure_admin_queue(ctrl);
620 	if (ret)
621 		goto out_free_queues;
622 
623 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
624 		/* warn if maxcmd is lower than queue_size */
625 		dev_warn(ctrl->ctrl.device,
626 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
627 			opts->queue_size, ctrl->ctrl.maxcmd);
628 		opts->queue_size = ctrl->ctrl.maxcmd;
629 	}
630 
631 	if (opts->nr_io_queues) {
632 		ret = nvme_loop_create_io_queues(ctrl);
633 		if (ret)
634 			goto out_remove_admin_queue;
635 	}
636 
637 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
638 
639 	dev_info(ctrl->ctrl.device,
640 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
641 
642 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
643 		WARN_ON_ONCE(1);
644 
645 	mutex_lock(&nvme_loop_ctrl_mutex);
646 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
647 	mutex_unlock(&nvme_loop_ctrl_mutex);
648 
649 	nvme_start_ctrl(&ctrl->ctrl);
650 
651 	return &ctrl->ctrl;
652 
653 out_remove_admin_queue:
654 	nvme_loop_destroy_admin_queue(ctrl);
655 out_free_queues:
656 	kfree(ctrl->queues);
657 out_uninit_ctrl:
658 	nvme_uninit_ctrl(&ctrl->ctrl);
659 	nvme_put_ctrl(&ctrl->ctrl);
660 out:
661 	if (ret > 0)
662 		ret = -EIO;
663 	return ERR_PTR(ret);
664 }
665 
666 static int nvme_loop_add_port(struct nvmet_port *port)
667 {
668 	mutex_lock(&nvme_loop_ports_mutex);
669 	list_add_tail(&port->entry, &nvme_loop_ports);
670 	mutex_unlock(&nvme_loop_ports_mutex);
671 	return 0;
672 }
673 
674 static void nvme_loop_remove_port(struct nvmet_port *port)
675 {
676 	mutex_lock(&nvme_loop_ports_mutex);
677 	list_del_init(&port->entry);
678 	mutex_unlock(&nvme_loop_ports_mutex);
679 
680 	/*
681 	 * Ensure any ctrls that are in the process of being
682 	 * deleted are in fact deleted before we return
683 	 * and free the port. This is to prevent active
684 	 * ctrls from using a port after it's freed.
685 	 */
686 	flush_workqueue(nvme_delete_wq);
687 }
688 
689 static const struct nvmet_fabrics_ops nvme_loop_ops = {
690 	.owner		= THIS_MODULE,
691 	.type		= NVMF_TRTYPE_LOOP,
692 	.add_port	= nvme_loop_add_port,
693 	.remove_port	= nvme_loop_remove_port,
694 	.queue_response = nvme_loop_queue_response,
695 	.delete_ctrl	= nvme_loop_delete_ctrl,
696 };
697 
698 static struct nvmf_transport_ops nvme_loop_transport = {
699 	.name		= "loop",
700 	.module		= THIS_MODULE,
701 	.create_ctrl	= nvme_loop_create_ctrl,
702 	.allowed_opts	= NVMF_OPT_TRADDR,
703 };
704 
705 static int __init nvme_loop_init_module(void)
706 {
707 	int ret;
708 
709 	ret = nvmet_register_transport(&nvme_loop_ops);
710 	if (ret)
711 		return ret;
712 
713 	ret = nvmf_register_transport(&nvme_loop_transport);
714 	if (ret)
715 		nvmet_unregister_transport(&nvme_loop_ops);
716 
717 	return ret;
718 }
719 
720 static void __exit nvme_loop_cleanup_module(void)
721 {
722 	struct nvme_loop_ctrl *ctrl, *next;
723 
724 	nvmf_unregister_transport(&nvme_loop_transport);
725 	nvmet_unregister_transport(&nvme_loop_ops);
726 
727 	mutex_lock(&nvme_loop_ctrl_mutex);
728 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
729 		nvme_delete_ctrl(&ctrl->ctrl);
730 	mutex_unlock(&nvme_loop_ctrl_mutex);
731 
732 	flush_workqueue(nvme_delete_wq);
733 }
734 
735 module_init(nvme_loop_init_module);
736 module_exit(nvme_loop_cleanup_module);
737 
738 MODULE_LICENSE("GPL v2");
739 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
740