xref: /linux/drivers/nvme/target/loop.c (revision 0ce92d548b44649a8de706f9bb9e74a4ed2f18a7)
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_ctrl	ctrl;
37 
38 	struct nvmet_port	*port;
39 
40 	/* Must be last --ends in a flexible-array member. */
41 	struct nvme_loop_iod	async_event_iod;
42 };
43 
44 static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
45 {
46 	return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
47 }
48 
49 enum nvme_loop_queue_flags {
50 	NVME_LOOP_Q_LIVE	= 0,
51 };
52 
53 struct nvme_loop_queue {
54 	struct nvmet_cq		nvme_cq;
55 	struct nvmet_sq		nvme_sq;
56 	struct nvme_loop_ctrl	*ctrl;
57 	unsigned long		flags;
58 };
59 
60 static LIST_HEAD(nvme_loop_ports);
61 static DEFINE_MUTEX(nvme_loop_ports_mutex);
62 
63 static LIST_HEAD(nvme_loop_ctrl_list);
64 static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
65 
66 static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
67 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
68 
69 static const struct nvmet_fabrics_ops nvme_loop_ops;
70 
71 static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
72 {
73 	return queue - queue->ctrl->queues;
74 }
75 
76 static void nvme_loop_complete_rq(struct request *req)
77 {
78 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
79 
80 	sg_free_table_chained(&iod->sg_table, NVME_INLINE_SG_CNT);
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_is_aen_req(nvme_loop_queue_idx(queue),
106 				     cqe->command_id))) {
107 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
108 				&cqe->result);
109 	} else {
110 		struct request *rq;
111 
112 		rq = nvme_find_rq(nvme_loop_tagset(queue), cqe->command_id);
113 		if (!rq) {
114 			dev_err(queue->ctrl->ctrl.device,
115 				"got bad command_id %#x on queue %d\n",
116 				cqe->command_id, nvme_loop_queue_idx(queue));
117 			return;
118 		}
119 
120 		if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
121 			nvme_loop_complete_rq(rq);
122 	}
123 }
124 
125 static void nvme_loop_execute_work(struct work_struct *work)
126 {
127 	struct nvme_loop_iod *iod =
128 		container_of(work, struct nvme_loop_iod, work);
129 
130 	iod->req.execute(&iod->req);
131 }
132 
133 static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
134 		const struct blk_mq_queue_data *bd)
135 {
136 	struct nvme_ns *ns = hctx->queue->queuedata;
137 	struct nvme_loop_queue *queue = hctx->driver_data;
138 	struct request *req = bd->rq;
139 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
140 	bool queue_ready = test_bit(NVME_LOOP_Q_LIVE, &queue->flags);
141 	blk_status_t ret;
142 
143 	if (!nvme_check_ready(&queue->ctrl->ctrl, req, queue_ready))
144 		return nvme_fail_nonready_command(&queue->ctrl->ctrl, req);
145 
146 	ret = nvme_setup_cmd(ns, req);
147 	if (ret)
148 		return ret;
149 
150 	nvme_start_request(req);
151 	iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
152 	iod->req.port = queue->ctrl->port;
153 	if (!nvmet_req_init(&iod->req, &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, NVME_INLINE_SG_CNT)) {
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, iod->sg_table.sgl);
167 		iod->req.transfer_len = blk_rq_payload_bytes(req);
168 	}
169 
170 	queue_work(nvmet_wq, &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_sq, &nvme_loop_ops)) {
186 		dev_err(ctrl->ctrl.device, "failed async event work\n");
187 		return;
188 	}
189 
190 	queue_work(nvmet_wq, &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 = to_loop_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 = to_loop_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 = to_loop_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 	/*
269 	 * It's possible that some requests might have been added
270 	 * after admin queue is stopped/quiesced. So now start the
271 	 * queue to flush these requests to the completion.
272 	 */
273 	nvme_unquiesce_admin_queue(&ctrl->ctrl);
274 
275 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
276 	nvmet_cq_put(&ctrl->queues[0].nvme_cq);
277 	nvme_remove_admin_tag_set(&ctrl->ctrl);
278 }
279 
280 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
281 {
282 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
283 
284 	if (list_empty(&ctrl->list))
285 		goto free_ctrl;
286 
287 	mutex_lock(&nvme_loop_ctrl_mutex);
288 	list_del(&ctrl->list);
289 	mutex_unlock(&nvme_loop_ctrl_mutex);
290 
291 	if (nctrl->tagset)
292 		nvme_remove_io_tag_set(nctrl);
293 	kfree(ctrl->queues);
294 	nvmf_free_options(nctrl->opts);
295 free_ctrl:
296 	kfree(ctrl);
297 }
298 
299 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
300 {
301 	int i;
302 
303 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
304 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
305 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
306 		nvmet_cq_put(&ctrl->queues[i].nvme_cq);
307 	}
308 	ctrl->ctrl.queue_count = 1;
309 	/*
310 	 * It's possible that some requests might have been added
311 	 * after io queue is stopped/quiesced. So now start the
312 	 * queue to flush these requests to the completion.
313 	 */
314 	nvme_unquiesce_io_queues(&ctrl->ctrl);
315 }
316 
317 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
318 {
319 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
320 	unsigned int nr_io_queues;
321 	int ret, i;
322 
323 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
324 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
325 	if (ret || !nr_io_queues)
326 		return ret;
327 
328 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
329 
330 	for (i = 1; i <= nr_io_queues; i++) {
331 		ctrl->queues[i].ctrl = ctrl;
332 		nvmet_cq_init(&ctrl->queues[i].nvme_cq);
333 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq,
334 				&ctrl->queues[i].nvme_cq);
335 		if (ret) {
336 			nvmet_cq_put(&ctrl->queues[i].nvme_cq);
337 			goto out_destroy_queues;
338 		}
339 
340 		ctrl->ctrl.queue_count++;
341 	}
342 
343 	return 0;
344 
345 out_destroy_queues:
346 	nvme_loop_destroy_io_queues(ctrl);
347 	return ret;
348 }
349 
350 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
351 {
352 	int i, ret;
353 
354 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
355 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
356 		if (ret)
357 			return ret;
358 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
359 	}
360 
361 	return 0;
362 }
363 
364 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
365 {
366 	int error;
367 
368 	ctrl->queues[0].ctrl = ctrl;
369 	nvmet_cq_init(&ctrl->queues[0].nvme_cq);
370 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq,
371 			&ctrl->queues[0].nvme_cq);
372 	if (error) {
373 		nvmet_cq_put(&ctrl->queues[0].nvme_cq);
374 		return error;
375 	}
376 	ctrl->ctrl.queue_count = 1;
377 
378 	error = nvme_alloc_admin_tag_set(&ctrl->ctrl, &ctrl->admin_tag_set,
379 			&nvme_loop_admin_mq_ops,
380 			sizeof(struct nvme_loop_iod) +
381 			NVME_INLINE_SG_CNT * sizeof(struct scatterlist));
382 	if (error)
383 		goto out_free_sq;
384 
385 	/* reset stopped state for the fresh admin queue */
386 	clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->ctrl.flags);
387 
388 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
389 	if (error)
390 		goto out_cleanup_tagset;
391 
392 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
393 
394 	error = nvme_enable_ctrl(&ctrl->ctrl);
395 	if (error)
396 		goto out_cleanup_tagset;
397 
398 	ctrl->ctrl.max_hw_sectors =
399 		(NVME_LOOP_MAX_SEGMENTS - 1) << PAGE_SECTORS_SHIFT;
400 
401 	nvme_unquiesce_admin_queue(&ctrl->ctrl);
402 
403 	error = nvme_init_ctrl_finish(&ctrl->ctrl, false);
404 	if (error)
405 		goto out_cleanup_tagset;
406 
407 	return 0;
408 
409 out_cleanup_tagset:
410 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
411 	nvme_remove_admin_tag_set(&ctrl->ctrl);
412 out_free_sq:
413 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
414 	nvmet_cq_put(&ctrl->queues[0].nvme_cq);
415 	return error;
416 }
417 
418 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
419 {
420 	if (ctrl->ctrl.queue_count > 1) {
421 		nvme_quiesce_io_queues(&ctrl->ctrl);
422 		nvme_cancel_tagset(&ctrl->ctrl);
423 		nvme_loop_destroy_io_queues(ctrl);
424 	}
425 
426 	nvme_quiesce_admin_queue(&ctrl->ctrl);
427 	if (nvme_ctrl_state(&ctrl->ctrl) == NVME_CTRL_LIVE)
428 		nvme_disable_ctrl(&ctrl->ctrl, true);
429 
430 	nvme_cancel_admin_tagset(&ctrl->ctrl);
431 	nvme_loop_destroy_admin_queue(ctrl);
432 }
433 
434 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
435 {
436 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
437 }
438 
439 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
440 {
441 	struct nvme_loop_ctrl *ctrl;
442 
443 	mutex_lock(&nvme_loop_ctrl_mutex);
444 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
445 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
446 			nvme_delete_ctrl(&ctrl->ctrl);
447 	}
448 	mutex_unlock(&nvme_loop_ctrl_mutex);
449 }
450 
451 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
452 {
453 	struct nvme_loop_ctrl *ctrl =
454 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
455 	int ret;
456 
457 	nvme_stop_ctrl(&ctrl->ctrl);
458 	nvme_loop_shutdown_ctrl(ctrl);
459 
460 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
461 		enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
462 
463 		if (state != NVME_CTRL_DELETING &&
464 		    state != NVME_CTRL_DELETING_NOIO)
465 			/* state change failure for non-deleted ctrl? */
466 			WARN_ON_ONCE(1);
467 		return;
468 	}
469 
470 	ret = nvme_loop_configure_admin_queue(ctrl);
471 	if (ret)
472 		goto out_disable;
473 
474 	ret = nvme_loop_init_io_queues(ctrl);
475 	if (ret)
476 		goto out_destroy_admin;
477 
478 	ret = nvme_loop_connect_io_queues(ctrl);
479 	if (ret)
480 		goto out_destroy_io;
481 
482 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
483 			ctrl->ctrl.queue_count - 1);
484 
485 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
486 		WARN_ON_ONCE(1);
487 
488 	nvme_start_ctrl(&ctrl->ctrl);
489 
490 	return;
491 
492 out_destroy_io:
493 	nvme_loop_destroy_io_queues(ctrl);
494 out_destroy_admin:
495 	nvme_quiesce_admin_queue(&ctrl->ctrl);
496 	nvme_cancel_admin_tagset(&ctrl->ctrl);
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 }
502 
503 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
504 	.name			= "loop",
505 	.module			= THIS_MODULE,
506 	.flags			= NVME_F_FABRICS,
507 	.reg_read32		= nvmf_reg_read32,
508 	.reg_read64		= nvmf_reg_read64,
509 	.reg_write32		= nvmf_reg_write32,
510 	.free_ctrl		= nvme_loop_free_ctrl,
511 	.submit_async_event	= nvme_loop_submit_async_event,
512 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
513 	.get_address		= nvmf_get_address,
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 	ret = nvme_alloc_io_tag_set(&ctrl->ctrl, &ctrl->tag_set,
525 			&nvme_loop_mq_ops, 1,
526 			sizeof(struct nvme_loop_iod) +
527 			NVME_INLINE_SG_CNT * sizeof(struct scatterlist));
528 	if (ret)
529 		goto out_destroy_queues;
530 
531 	ret = nvme_loop_connect_io_queues(ctrl);
532 	if (ret)
533 		goto out_cleanup_tagset;
534 
535 	return 0;
536 
537 out_cleanup_tagset:
538 	nvme_remove_io_tag_set(&ctrl->ctrl);
539 out_destroy_queues:
540 	nvme_loop_destroy_io_queues(ctrl);
541 	return ret;
542 }
543 
544 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
545 {
546 	struct nvmet_port *p, *found = NULL;
547 
548 	mutex_lock(&nvme_loop_ports_mutex);
549 	list_for_each_entry(p, &nvme_loop_ports, entry) {
550 		/* if no transport address is specified use the first port */
551 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
552 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
553 			continue;
554 		found = p;
555 		break;
556 	}
557 	mutex_unlock(&nvme_loop_ports_mutex);
558 	return found;
559 }
560 
561 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
562 		struct nvmf_ctrl_options *opts)
563 {
564 	struct nvme_loop_ctrl *ctrl;
565 	int ret;
566 
567 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
568 	if (!ctrl)
569 		return ERR_PTR(-ENOMEM);
570 	ctrl->ctrl.opts = opts;
571 	INIT_LIST_HEAD(&ctrl->list);
572 
573 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
574 
575 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
576 				0 /* no quirks, we're perfect! */);
577 	if (ret) {
578 		kfree(ctrl);
579 		goto out;
580 	}
581 
582 	ret = nvme_add_ctrl(&ctrl->ctrl);
583 	if (ret)
584 		goto out_put_ctrl;
585 
586 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
587 		WARN_ON_ONCE(1);
588 
589 	ret = -ENOMEM;
590 
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 	ctrl->ctrl.sqsize = opts->queue_size - 1;
611 
612 	if (opts->nr_io_queues) {
613 		ret = nvme_loop_create_io_queues(ctrl);
614 		if (ret)
615 			goto out_remove_admin_queue;
616 	}
617 
618 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
619 
620 	dev_info(ctrl->ctrl.device,
621 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
622 
623 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
624 		WARN_ON_ONCE(1);
625 
626 	mutex_lock(&nvme_loop_ctrl_mutex);
627 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
628 	mutex_unlock(&nvme_loop_ctrl_mutex);
629 
630 	nvme_start_ctrl(&ctrl->ctrl);
631 
632 	return &ctrl->ctrl;
633 
634 out_remove_admin_queue:
635 	nvme_quiesce_admin_queue(&ctrl->ctrl);
636 	nvme_cancel_admin_tagset(&ctrl->ctrl);
637 	nvme_loop_destroy_admin_queue(ctrl);
638 out_free_queues:
639 	kfree(ctrl->queues);
640 out_uninit_ctrl:
641 	nvme_uninit_ctrl(&ctrl->ctrl);
642 out_put_ctrl:
643 	nvme_put_ctrl(&ctrl->ctrl);
644 out:
645 	if (ret > 0)
646 		ret = -EIO;
647 	return ERR_PTR(ret);
648 }
649 
650 static int nvme_loop_add_port(struct nvmet_port *port)
651 {
652 	mutex_lock(&nvme_loop_ports_mutex);
653 	list_add_tail(&port->entry, &nvme_loop_ports);
654 	mutex_unlock(&nvme_loop_ports_mutex);
655 	return 0;
656 }
657 
658 static void nvme_loop_remove_port(struct nvmet_port *port)
659 {
660 	mutex_lock(&nvme_loop_ports_mutex);
661 	list_del_init(&port->entry);
662 	mutex_unlock(&nvme_loop_ports_mutex);
663 
664 	/*
665 	 * Ensure any ctrls that are in the process of being
666 	 * deleted are in fact deleted before we return
667 	 * and free the port. This is to prevent active
668 	 * ctrls from using a port after it's freed.
669 	 */
670 	flush_workqueue(nvme_delete_wq);
671 }
672 
673 static const struct nvmet_fabrics_ops nvme_loop_ops = {
674 	.owner		= THIS_MODULE,
675 	.type		= NVMF_TRTYPE_LOOP,
676 	.add_port	= nvme_loop_add_port,
677 	.remove_port	= nvme_loop_remove_port,
678 	.queue_response = nvme_loop_queue_response,
679 	.delete_ctrl	= nvme_loop_delete_ctrl,
680 };
681 
682 static struct nvmf_transport_ops nvme_loop_transport = {
683 	.name		= "loop",
684 	.module		= THIS_MODULE,
685 	.create_ctrl	= nvme_loop_create_ctrl,
686 	.allowed_opts	= NVMF_OPT_TRADDR,
687 };
688 
689 static int __init nvme_loop_init_module(void)
690 {
691 	int ret;
692 
693 	ret = nvmet_register_transport(&nvme_loop_ops);
694 	if (ret)
695 		return ret;
696 
697 	ret = nvmf_register_transport(&nvme_loop_transport);
698 	if (ret)
699 		nvmet_unregister_transport(&nvme_loop_ops);
700 
701 	return ret;
702 }
703 
704 static void __exit nvme_loop_cleanup_module(void)
705 {
706 	struct nvme_loop_ctrl *ctrl, *next;
707 
708 	nvmf_unregister_transport(&nvme_loop_transport);
709 	nvmet_unregister_transport(&nvme_loop_ops);
710 
711 	mutex_lock(&nvme_loop_ctrl_mutex);
712 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
713 		nvme_delete_ctrl(&ctrl->ctrl);
714 	mutex_unlock(&nvme_loop_ctrl_mutex);
715 
716 	flush_workqueue(nvme_delete_wq);
717 }
718 
719 module_init(nvme_loop_init_module);
720 module_exit(nvme_loop_cleanup_module);
721 
722 MODULE_DESCRIPTION("NVMe target loop transport driver");
723 MODULE_LICENSE("GPL v2");
724 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
725