xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 7f877908)
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 	sg_free_table_chained(&iod->sg_table, NVME_INLINE_SG_CNT);
80 	nvme_complete_rq(req);
81 }
82 
83 static struct blk_mq_tags *nvme_loop_tagset(struct nvme_loop_queue *queue)
84 {
85 	u32 queue_idx = nvme_loop_queue_idx(queue);
86 
87 	if (queue_idx == 0)
88 		return queue->ctrl->admin_tag_set.tags[queue_idx];
89 	return queue->ctrl->tag_set.tags[queue_idx - 1];
90 }
91 
92 static void nvme_loop_queue_response(struct nvmet_req *req)
93 {
94 	struct nvme_loop_queue *queue =
95 		container_of(req->sq, struct nvme_loop_queue, nvme_sq);
96 	struct nvme_completion *cqe = req->cqe;
97 
98 	/*
99 	 * AEN requests are special as they don't time out and can
100 	 * survive any kind of queue freeze and often don't respond to
101 	 * aborts.  We don't even bother to allocate a struct request
102 	 * for them but rather special case them here.
103 	 */
104 	if (unlikely(nvme_is_aen_req(nvme_loop_queue_idx(queue),
105 				     cqe->command_id))) {
106 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
107 				&cqe->result);
108 	} else {
109 		struct request *rq;
110 
111 		rq = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
112 		if (!rq) {
113 			dev_err(queue->ctrl->ctrl.device,
114 				"tag 0x%x on queue %d not found\n",
115 				cqe->command_id, nvme_loop_queue_idx(queue));
116 			return;
117 		}
118 
119 		nvme_end_request(rq, cqe->status, cqe->result);
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 (!nvmf_check_ready(&queue->ctrl->ctrl, req, queue_ready))
142 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, req);
143 
144 	ret = nvme_setup_cmd(ns, req, &iod->cmd);
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 
209 	nvme_req(req)->ctrl = &ctrl->ctrl;
210 	return nvme_loop_init_iod(ctrl, blk_mq_rq_to_pdu(req),
211 			(set == &ctrl->tag_set) ? hctx_idx + 1 : 0);
212 }
213 
214 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
215 		unsigned int hctx_idx)
216 {
217 	struct nvme_loop_ctrl *ctrl = data;
218 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
219 
220 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
221 
222 	hctx->driver_data = queue;
223 	return 0;
224 }
225 
226 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
227 		unsigned int hctx_idx)
228 {
229 	struct nvme_loop_ctrl *ctrl = data;
230 	struct nvme_loop_queue *queue = &ctrl->queues[0];
231 
232 	BUG_ON(hctx_idx != 0);
233 
234 	hctx->driver_data = queue;
235 	return 0;
236 }
237 
238 static const struct blk_mq_ops nvme_loop_mq_ops = {
239 	.queue_rq	= nvme_loop_queue_rq,
240 	.complete	= nvme_loop_complete_rq,
241 	.init_request	= nvme_loop_init_request,
242 	.init_hctx	= nvme_loop_init_hctx,
243 };
244 
245 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
246 	.queue_rq	= nvme_loop_queue_rq,
247 	.complete	= nvme_loop_complete_rq,
248 	.init_request	= nvme_loop_init_request,
249 	.init_hctx	= nvme_loop_init_admin_hctx,
250 };
251 
252 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
253 {
254 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
255 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
256 	blk_cleanup_queue(ctrl->ctrl.admin_q);
257 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
258 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
259 }
260 
261 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
262 {
263 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
264 
265 	if (list_empty(&ctrl->list))
266 		goto free_ctrl;
267 
268 	mutex_lock(&nvme_loop_ctrl_mutex);
269 	list_del(&ctrl->list);
270 	mutex_unlock(&nvme_loop_ctrl_mutex);
271 
272 	if (nctrl->tagset) {
273 		blk_cleanup_queue(ctrl->ctrl.connect_q);
274 		blk_mq_free_tag_set(&ctrl->tag_set);
275 	}
276 	kfree(ctrl->queues);
277 	nvmf_free_options(nctrl->opts);
278 free_ctrl:
279 	kfree(ctrl);
280 }
281 
282 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
283 {
284 	int i;
285 
286 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
287 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
288 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
289 	}
290 }
291 
292 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
293 {
294 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
295 	unsigned int nr_io_queues;
296 	int ret, i;
297 
298 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
299 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
300 	if (ret || !nr_io_queues)
301 		return ret;
302 
303 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
304 
305 	for (i = 1; i <= nr_io_queues; i++) {
306 		ctrl->queues[i].ctrl = ctrl;
307 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
308 		if (ret)
309 			goto out_destroy_queues;
310 
311 		ctrl->ctrl.queue_count++;
312 	}
313 
314 	return 0;
315 
316 out_destroy_queues:
317 	nvme_loop_destroy_io_queues(ctrl);
318 	return ret;
319 }
320 
321 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
322 {
323 	int i, ret;
324 
325 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
326 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
327 		if (ret)
328 			return ret;
329 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
330 	}
331 
332 	return 0;
333 }
334 
335 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
336 {
337 	int error;
338 
339 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
340 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
341 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
342 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
343 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
344 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
345 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
346 	ctrl->admin_tag_set.driver_data = ctrl;
347 	ctrl->admin_tag_set.nr_hw_queues = 1;
348 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
349 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
350 
351 	ctrl->queues[0].ctrl = ctrl;
352 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
353 	if (error)
354 		return error;
355 	ctrl->ctrl.queue_count = 1;
356 
357 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
358 	if (error)
359 		goto out_free_sq;
360 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
361 
362 	ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
363 	if (IS_ERR(ctrl->ctrl.fabrics_q)) {
364 		error = PTR_ERR(ctrl->ctrl.fabrics_q);
365 		goto out_free_tagset;
366 	}
367 
368 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
369 	if (IS_ERR(ctrl->ctrl.admin_q)) {
370 		error = PTR_ERR(ctrl->ctrl.admin_q);
371 		goto out_cleanup_fabrics_q;
372 	}
373 
374 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
375 	if (error)
376 		goto out_cleanup_queue;
377 
378 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
379 
380 	error = nvme_enable_ctrl(&ctrl->ctrl);
381 	if (error)
382 		goto out_cleanup_queue;
383 
384 	ctrl->ctrl.max_hw_sectors =
385 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
386 
387 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
388 
389 	error = nvme_init_identify(&ctrl->ctrl);
390 	if (error)
391 		goto out_cleanup_queue;
392 
393 	return 0;
394 
395 out_cleanup_queue:
396 	blk_cleanup_queue(ctrl->ctrl.admin_q);
397 out_cleanup_fabrics_q:
398 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
399 out_free_tagset:
400 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
401 out_free_sq:
402 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
403 	return error;
404 }
405 
406 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
407 {
408 	if (ctrl->ctrl.queue_count > 1) {
409 		nvme_stop_queues(&ctrl->ctrl);
410 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
411 					nvme_cancel_request, &ctrl->ctrl);
412 		blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
413 		nvme_loop_destroy_io_queues(ctrl);
414 	}
415 
416 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
417 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
418 		nvme_shutdown_ctrl(&ctrl->ctrl);
419 
420 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
421 				nvme_cancel_request, &ctrl->ctrl);
422 	blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
423 	nvme_loop_destroy_admin_queue(ctrl);
424 }
425 
426 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
427 {
428 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
429 }
430 
431 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
432 {
433 	struct nvme_loop_ctrl *ctrl;
434 
435 	mutex_lock(&nvme_loop_ctrl_mutex);
436 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
437 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
438 			nvme_delete_ctrl(&ctrl->ctrl);
439 	}
440 	mutex_unlock(&nvme_loop_ctrl_mutex);
441 }
442 
443 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
444 {
445 	struct nvme_loop_ctrl *ctrl =
446 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
447 	bool changed;
448 	int ret;
449 
450 	nvme_stop_ctrl(&ctrl->ctrl);
451 	nvme_loop_shutdown_ctrl(ctrl);
452 
453 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
454 		/* state change failure should never happen */
455 		WARN_ON_ONCE(1);
456 		return;
457 	}
458 
459 	ret = nvme_loop_configure_admin_queue(ctrl);
460 	if (ret)
461 		goto out_disable;
462 
463 	ret = nvme_loop_init_io_queues(ctrl);
464 	if (ret)
465 		goto out_destroy_admin;
466 
467 	ret = nvme_loop_connect_io_queues(ctrl);
468 	if (ret)
469 		goto out_destroy_io;
470 
471 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
472 			ctrl->ctrl.queue_count - 1);
473 
474 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
475 	WARN_ON_ONCE(!changed);
476 
477 	nvme_start_ctrl(&ctrl->ctrl);
478 
479 	return;
480 
481 out_destroy_io:
482 	nvme_loop_destroy_io_queues(ctrl);
483 out_destroy_admin:
484 	nvme_loop_destroy_admin_queue(ctrl);
485 out_disable:
486 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
487 	nvme_uninit_ctrl(&ctrl->ctrl);
488 	nvme_put_ctrl(&ctrl->ctrl);
489 }
490 
491 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
492 	.name			= "loop",
493 	.module			= THIS_MODULE,
494 	.flags			= NVME_F_FABRICS,
495 	.reg_read32		= nvmf_reg_read32,
496 	.reg_read64		= nvmf_reg_read64,
497 	.reg_write32		= nvmf_reg_write32,
498 	.free_ctrl		= nvme_loop_free_ctrl,
499 	.submit_async_event	= nvme_loop_submit_async_event,
500 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
501 	.get_address		= nvmf_get_address,
502 };
503 
504 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
505 {
506 	int ret;
507 
508 	ret = nvme_loop_init_io_queues(ctrl);
509 	if (ret)
510 		return ret;
511 
512 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
513 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
514 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
515 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
516 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
517 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
518 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
519 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
520 	ctrl->tag_set.driver_data = ctrl;
521 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
522 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
523 	ctrl->ctrl.tagset = &ctrl->tag_set;
524 
525 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
526 	if (ret)
527 		goto out_destroy_queues;
528 
529 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
530 	if (IS_ERR(ctrl->ctrl.connect_q)) {
531 		ret = PTR_ERR(ctrl->ctrl.connect_q);
532 		goto out_free_tagset;
533 	}
534 
535 	ret = nvme_loop_connect_io_queues(ctrl);
536 	if (ret)
537 		goto out_cleanup_connect_q;
538 
539 	return 0;
540 
541 out_cleanup_connect_q:
542 	blk_cleanup_queue(ctrl->ctrl.connect_q);
543 out_free_tagset:
544 	blk_mq_free_tag_set(&ctrl->tag_set);
545 out_destroy_queues:
546 	nvme_loop_destroy_io_queues(ctrl);
547 	return ret;
548 }
549 
550 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
551 {
552 	struct nvmet_port *p, *found = NULL;
553 
554 	mutex_lock(&nvme_loop_ports_mutex);
555 	list_for_each_entry(p, &nvme_loop_ports, entry) {
556 		/* if no transport address is specified use the first port */
557 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
558 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
559 			continue;
560 		found = p;
561 		break;
562 	}
563 	mutex_unlock(&nvme_loop_ports_mutex);
564 	return found;
565 }
566 
567 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
568 		struct nvmf_ctrl_options *opts)
569 {
570 	struct nvme_loop_ctrl *ctrl;
571 	bool changed;
572 	int ret;
573 
574 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
575 	if (!ctrl)
576 		return ERR_PTR(-ENOMEM);
577 	ctrl->ctrl.opts = opts;
578 	INIT_LIST_HEAD(&ctrl->list);
579 
580 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
581 
582 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
583 				0 /* no quirks, we're perfect! */);
584 	if (ret)
585 		goto out_put_ctrl;
586 
587 	ret = -ENOMEM;
588 
589 	ctrl->ctrl.sqsize = opts->queue_size - 1;
590 	ctrl->ctrl.kato = opts->kato;
591 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
592 
593 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
594 			GFP_KERNEL);
595 	if (!ctrl->queues)
596 		goto out_uninit_ctrl;
597 
598 	ret = nvme_loop_configure_admin_queue(ctrl);
599 	if (ret)
600 		goto out_free_queues;
601 
602 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
603 		/* warn if maxcmd is lower than queue_size */
604 		dev_warn(ctrl->ctrl.device,
605 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
606 			opts->queue_size, ctrl->ctrl.maxcmd);
607 		opts->queue_size = ctrl->ctrl.maxcmd;
608 	}
609 
610 	if (opts->nr_io_queues) {
611 		ret = nvme_loop_create_io_queues(ctrl);
612 		if (ret)
613 			goto out_remove_admin_queue;
614 	}
615 
616 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
617 
618 	dev_info(ctrl->ctrl.device,
619 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
620 
621 	nvme_get_ctrl(&ctrl->ctrl);
622 
623 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
624 	WARN_ON_ONCE(!changed);
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_loop_destroy_admin_queue(ctrl);
636 out_free_queues:
637 	kfree(ctrl->queues);
638 out_uninit_ctrl:
639 	nvme_uninit_ctrl(&ctrl->ctrl);
640 out_put_ctrl:
641 	nvme_put_ctrl(&ctrl->ctrl);
642 	if (ret > 0)
643 		ret = -EIO;
644 	return ERR_PTR(ret);
645 }
646 
647 static int nvme_loop_add_port(struct nvmet_port *port)
648 {
649 	mutex_lock(&nvme_loop_ports_mutex);
650 	list_add_tail(&port->entry, &nvme_loop_ports);
651 	mutex_unlock(&nvme_loop_ports_mutex);
652 	return 0;
653 }
654 
655 static void nvme_loop_remove_port(struct nvmet_port *port)
656 {
657 	mutex_lock(&nvme_loop_ports_mutex);
658 	list_del_init(&port->entry);
659 	mutex_unlock(&nvme_loop_ports_mutex);
660 
661 	/*
662 	 * Ensure any ctrls that are in the process of being
663 	 * deleted are in fact deleted before we return
664 	 * and free the port. This is to prevent active
665 	 * ctrls from using a port after it's freed.
666 	 */
667 	flush_workqueue(nvme_delete_wq);
668 }
669 
670 static const struct nvmet_fabrics_ops nvme_loop_ops = {
671 	.owner		= THIS_MODULE,
672 	.type		= NVMF_TRTYPE_LOOP,
673 	.add_port	= nvme_loop_add_port,
674 	.remove_port	= nvme_loop_remove_port,
675 	.queue_response = nvme_loop_queue_response,
676 	.delete_ctrl	= nvme_loop_delete_ctrl,
677 };
678 
679 static struct nvmf_transport_ops nvme_loop_transport = {
680 	.name		= "loop",
681 	.module		= THIS_MODULE,
682 	.create_ctrl	= nvme_loop_create_ctrl,
683 	.allowed_opts	= NVMF_OPT_TRADDR,
684 };
685 
686 static int __init nvme_loop_init_module(void)
687 {
688 	int ret;
689 
690 	ret = nvmet_register_transport(&nvme_loop_ops);
691 	if (ret)
692 		return ret;
693 
694 	ret = nvmf_register_transport(&nvme_loop_transport);
695 	if (ret)
696 		nvmet_unregister_transport(&nvme_loop_ops);
697 
698 	return ret;
699 }
700 
701 static void __exit nvme_loop_cleanup_module(void)
702 {
703 	struct nvme_loop_ctrl *ctrl, *next;
704 
705 	nvmf_unregister_transport(&nvme_loop_transport);
706 	nvmet_unregister_transport(&nvme_loop_ops);
707 
708 	mutex_lock(&nvme_loop_ctrl_mutex);
709 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
710 		nvme_delete_ctrl(&ctrl->ctrl);
711 	mutex_unlock(&nvme_loop_ctrl_mutex);
712 
713 	flush_workqueue(nvme_delete_wq);
714 }
715 
716 module_init(nvme_loop_init_module);
717 module_exit(nvme_loop_cleanup_module);
718 
719 MODULE_LICENSE("GPL v2");
720 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
721