xref: /openbmc/linux/drivers/nvme/target/loop.c (revision 84b102f5)
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 = blk_mq_tag_to_rq(nvme_loop_tagset(queue), cqe->command_id);
111 		if (!rq) {
112 			dev_err(queue->ctrl->ctrl.device,
113 				"tag 0x%x on queue %d not found\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 (!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 struct lock_class_key loop_hctx_fq_lock_key;
215 
216 static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
217 		unsigned int hctx_idx)
218 {
219 	struct nvme_loop_ctrl *ctrl = data;
220 	struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
221 
222 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
223 
224 	/*
225 	 * flush_end_io() can be called recursively for us, so use our own
226 	 * lock class key for avoiding lockdep possible recursive locking,
227 	 * then we can remove the dynamically allocated lock class for each
228 	 * flush queue, that way may cause horrible boot delay.
229 	 */
230 	blk_mq_hctx_set_fq_lock_class(hctx, &loop_hctx_fq_lock_key);
231 
232 	hctx->driver_data = queue;
233 	return 0;
234 }
235 
236 static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
237 		unsigned int hctx_idx)
238 {
239 	struct nvme_loop_ctrl *ctrl = data;
240 	struct nvme_loop_queue *queue = &ctrl->queues[0];
241 
242 	BUG_ON(hctx_idx != 0);
243 
244 	hctx->driver_data = queue;
245 	return 0;
246 }
247 
248 static const struct blk_mq_ops nvme_loop_mq_ops = {
249 	.queue_rq	= nvme_loop_queue_rq,
250 	.complete	= nvme_loop_complete_rq,
251 	.init_request	= nvme_loop_init_request,
252 	.init_hctx	= nvme_loop_init_hctx,
253 };
254 
255 static const struct blk_mq_ops nvme_loop_admin_mq_ops = {
256 	.queue_rq	= nvme_loop_queue_rq,
257 	.complete	= nvme_loop_complete_rq,
258 	.init_request	= nvme_loop_init_request,
259 	.init_hctx	= nvme_loop_init_admin_hctx,
260 };
261 
262 static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
263 {
264 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
265 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
266 	blk_cleanup_queue(ctrl->ctrl.admin_q);
267 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
268 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
269 }
270 
271 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
272 {
273 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
274 
275 	if (list_empty(&ctrl->list))
276 		goto free_ctrl;
277 
278 	mutex_lock(&nvme_loop_ctrl_mutex);
279 	list_del(&ctrl->list);
280 	mutex_unlock(&nvme_loop_ctrl_mutex);
281 
282 	if (nctrl->tagset) {
283 		blk_cleanup_queue(ctrl->ctrl.connect_q);
284 		blk_mq_free_tag_set(&ctrl->tag_set);
285 	}
286 	kfree(ctrl->queues);
287 	nvmf_free_options(nctrl->opts);
288 free_ctrl:
289 	kfree(ctrl);
290 }
291 
292 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
293 {
294 	int i;
295 
296 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
297 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
298 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
299 	}
300 }
301 
302 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
303 {
304 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
305 	unsigned int nr_io_queues;
306 	int ret, i;
307 
308 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
309 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
310 	if (ret || !nr_io_queues)
311 		return ret;
312 
313 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
314 
315 	for (i = 1; i <= nr_io_queues; i++) {
316 		ctrl->queues[i].ctrl = ctrl;
317 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
318 		if (ret)
319 			goto out_destroy_queues;
320 
321 		ctrl->ctrl.queue_count++;
322 	}
323 
324 	return 0;
325 
326 out_destroy_queues:
327 	nvme_loop_destroy_io_queues(ctrl);
328 	return ret;
329 }
330 
331 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
332 {
333 	int i, ret;
334 
335 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
336 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i, false);
337 		if (ret)
338 			return ret;
339 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
340 	}
341 
342 	return 0;
343 }
344 
345 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
346 {
347 	int error;
348 
349 	memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
350 	ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
351 	ctrl->admin_tag_set.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
352 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
353 	ctrl->admin_tag_set.numa_node = ctrl->ctrl.numa_node;
354 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
355 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
356 	ctrl->admin_tag_set.driver_data = ctrl;
357 	ctrl->admin_tag_set.nr_hw_queues = 1;
358 	ctrl->admin_tag_set.timeout = NVME_ADMIN_TIMEOUT;
359 	ctrl->admin_tag_set.flags = BLK_MQ_F_NO_SCHED;
360 
361 	ctrl->queues[0].ctrl = ctrl;
362 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
363 	if (error)
364 		return error;
365 	ctrl->ctrl.queue_count = 1;
366 
367 	error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
368 	if (error)
369 		goto out_free_sq;
370 	ctrl->ctrl.admin_tagset = &ctrl->admin_tag_set;
371 
372 	ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
373 	if (IS_ERR(ctrl->ctrl.fabrics_q)) {
374 		error = PTR_ERR(ctrl->ctrl.fabrics_q);
375 		goto out_free_tagset;
376 	}
377 
378 	ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
379 	if (IS_ERR(ctrl->ctrl.admin_q)) {
380 		error = PTR_ERR(ctrl->ctrl.admin_q);
381 		goto out_cleanup_fabrics_q;
382 	}
383 
384 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
385 	if (error)
386 		goto out_cleanup_queue;
387 
388 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
389 
390 	error = nvme_enable_ctrl(&ctrl->ctrl);
391 	if (error)
392 		goto out_cleanup_queue;
393 
394 	ctrl->ctrl.max_hw_sectors =
395 		(NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
396 
397 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
398 
399 	error = nvme_init_identify(&ctrl->ctrl);
400 	if (error)
401 		goto out_cleanup_queue;
402 
403 	return 0;
404 
405 out_cleanup_queue:
406 	blk_cleanup_queue(ctrl->ctrl.admin_q);
407 out_cleanup_fabrics_q:
408 	blk_cleanup_queue(ctrl->ctrl.fabrics_q);
409 out_free_tagset:
410 	blk_mq_free_tag_set(&ctrl->admin_tag_set);
411 out_free_sq:
412 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
413 	return error;
414 }
415 
416 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
417 {
418 	if (ctrl->ctrl.queue_count > 1) {
419 		nvme_stop_queues(&ctrl->ctrl);
420 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
421 					nvme_cancel_request, &ctrl->ctrl);
422 		blk_mq_tagset_wait_completed_request(&ctrl->tag_set);
423 		nvme_loop_destroy_io_queues(ctrl);
424 	}
425 
426 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
427 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
428 		nvme_shutdown_ctrl(&ctrl->ctrl);
429 
430 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
431 				nvme_cancel_request, &ctrl->ctrl);
432 	blk_mq_tagset_wait_completed_request(&ctrl->admin_tag_set);
433 	nvme_loop_destroy_admin_queue(ctrl);
434 }
435 
436 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
437 {
438 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
439 }
440 
441 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
442 {
443 	struct nvme_loop_ctrl *ctrl;
444 
445 	mutex_lock(&nvme_loop_ctrl_mutex);
446 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
447 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
448 			nvme_delete_ctrl(&ctrl->ctrl);
449 	}
450 	mutex_unlock(&nvme_loop_ctrl_mutex);
451 }
452 
453 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
454 {
455 	struct nvme_loop_ctrl *ctrl =
456 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
457 	int ret;
458 
459 	nvme_stop_ctrl(&ctrl->ctrl);
460 	nvme_loop_shutdown_ctrl(ctrl);
461 
462 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
463 		/* state change failure should never happen */
464 		WARN_ON_ONCE(1);
465 		return;
466 	}
467 
468 	ret = nvme_loop_configure_admin_queue(ctrl);
469 	if (ret)
470 		goto out_disable;
471 
472 	ret = nvme_loop_init_io_queues(ctrl);
473 	if (ret)
474 		goto out_destroy_admin;
475 
476 	ret = nvme_loop_connect_io_queues(ctrl);
477 	if (ret)
478 		goto out_destroy_io;
479 
480 	blk_mq_update_nr_hw_queues(&ctrl->tag_set,
481 			ctrl->ctrl.queue_count - 1);
482 
483 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
484 		WARN_ON_ONCE(1);
485 
486 	nvme_start_ctrl(&ctrl->ctrl);
487 
488 	return;
489 
490 out_destroy_io:
491 	nvme_loop_destroy_io_queues(ctrl);
492 out_destroy_admin:
493 	nvme_loop_destroy_admin_queue(ctrl);
494 out_disable:
495 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
496 	nvme_uninit_ctrl(&ctrl->ctrl);
497 }
498 
499 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
500 	.name			= "loop",
501 	.module			= THIS_MODULE,
502 	.flags			= NVME_F_FABRICS,
503 	.reg_read32		= nvmf_reg_read32,
504 	.reg_read64		= nvmf_reg_read64,
505 	.reg_write32		= nvmf_reg_write32,
506 	.free_ctrl		= nvme_loop_free_ctrl,
507 	.submit_async_event	= nvme_loop_submit_async_event,
508 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
509 	.get_address		= nvmf_get_address,
510 };
511 
512 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
513 {
514 	int ret;
515 
516 	ret = nvme_loop_init_io_queues(ctrl);
517 	if (ret)
518 		return ret;
519 
520 	memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
521 	ctrl->tag_set.ops = &nvme_loop_mq_ops;
522 	ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
523 	ctrl->tag_set.reserved_tags = 1; /* fabric connect */
524 	ctrl->tag_set.numa_node = ctrl->ctrl.numa_node;
525 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
526 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
527 		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
528 	ctrl->tag_set.driver_data = ctrl;
529 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
530 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
531 	ctrl->ctrl.tagset = &ctrl->tag_set;
532 
533 	ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
534 	if (ret)
535 		goto out_destroy_queues;
536 
537 	ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
538 	if (IS_ERR(ctrl->ctrl.connect_q)) {
539 		ret = PTR_ERR(ctrl->ctrl.connect_q);
540 		goto out_free_tagset;
541 	}
542 
543 	ret = nvme_loop_connect_io_queues(ctrl);
544 	if (ret)
545 		goto out_cleanup_connect_q;
546 
547 	return 0;
548 
549 out_cleanup_connect_q:
550 	blk_cleanup_queue(ctrl->ctrl.connect_q);
551 out_free_tagset:
552 	blk_mq_free_tag_set(&ctrl->tag_set);
553 out_destroy_queues:
554 	nvme_loop_destroy_io_queues(ctrl);
555 	return ret;
556 }
557 
558 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
559 {
560 	struct nvmet_port *p, *found = NULL;
561 
562 	mutex_lock(&nvme_loop_ports_mutex);
563 	list_for_each_entry(p, &nvme_loop_ports, entry) {
564 		/* if no transport address is specified use the first port */
565 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
566 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
567 			continue;
568 		found = p;
569 		break;
570 	}
571 	mutex_unlock(&nvme_loop_ports_mutex);
572 	return found;
573 }
574 
575 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
576 		struct nvmf_ctrl_options *opts)
577 {
578 	struct nvme_loop_ctrl *ctrl;
579 	int ret;
580 
581 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
582 	if (!ctrl)
583 		return ERR_PTR(-ENOMEM);
584 	ctrl->ctrl.opts = opts;
585 	INIT_LIST_HEAD(&ctrl->list);
586 
587 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
588 
589 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
590 				0 /* no quirks, we're perfect! */);
591 	if (ret)
592 		goto out;
593 
594 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
595 		WARN_ON_ONCE(1);
596 
597 	ret = -ENOMEM;
598 
599 	ctrl->ctrl.sqsize = opts->queue_size - 1;
600 	ctrl->ctrl.kato = opts->kato;
601 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
602 
603 	ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
604 			GFP_KERNEL);
605 	if (!ctrl->queues)
606 		goto out_uninit_ctrl;
607 
608 	ret = nvme_loop_configure_admin_queue(ctrl);
609 	if (ret)
610 		goto out_free_queues;
611 
612 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
613 		/* warn if maxcmd is lower than queue_size */
614 		dev_warn(ctrl->ctrl.device,
615 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
616 			opts->queue_size, ctrl->ctrl.maxcmd);
617 		opts->queue_size = ctrl->ctrl.maxcmd;
618 	}
619 
620 	if (opts->nr_io_queues) {
621 		ret = nvme_loop_create_io_queues(ctrl);
622 		if (ret)
623 			goto out_remove_admin_queue;
624 	}
625 
626 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
627 
628 	dev_info(ctrl->ctrl.device,
629 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
630 
631 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
632 		WARN_ON_ONCE(1);
633 
634 	mutex_lock(&nvme_loop_ctrl_mutex);
635 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
636 	mutex_unlock(&nvme_loop_ctrl_mutex);
637 
638 	nvme_start_ctrl(&ctrl->ctrl);
639 
640 	return &ctrl->ctrl;
641 
642 out_remove_admin_queue:
643 	nvme_loop_destroy_admin_queue(ctrl);
644 out_free_queues:
645 	kfree(ctrl->queues);
646 out_uninit_ctrl:
647 	nvme_uninit_ctrl(&ctrl->ctrl);
648 	nvme_put_ctrl(&ctrl->ctrl);
649 out:
650 	if (ret > 0)
651 		ret = -EIO;
652 	return ERR_PTR(ret);
653 }
654 
655 static int nvme_loop_add_port(struct nvmet_port *port)
656 {
657 	mutex_lock(&nvme_loop_ports_mutex);
658 	list_add_tail(&port->entry, &nvme_loop_ports);
659 	mutex_unlock(&nvme_loop_ports_mutex);
660 	return 0;
661 }
662 
663 static void nvme_loop_remove_port(struct nvmet_port *port)
664 {
665 	mutex_lock(&nvme_loop_ports_mutex);
666 	list_del_init(&port->entry);
667 	mutex_unlock(&nvme_loop_ports_mutex);
668 
669 	/*
670 	 * Ensure any ctrls that are in the process of being
671 	 * deleted are in fact deleted before we return
672 	 * and free the port. This is to prevent active
673 	 * ctrls from using a port after it's freed.
674 	 */
675 	flush_workqueue(nvme_delete_wq);
676 }
677 
678 static const struct nvmet_fabrics_ops nvme_loop_ops = {
679 	.owner		= THIS_MODULE,
680 	.type		= NVMF_TRTYPE_LOOP,
681 	.add_port	= nvme_loop_add_port,
682 	.remove_port	= nvme_loop_remove_port,
683 	.queue_response = nvme_loop_queue_response,
684 	.delete_ctrl	= nvme_loop_delete_ctrl,
685 };
686 
687 static struct nvmf_transport_ops nvme_loop_transport = {
688 	.name		= "loop",
689 	.module		= THIS_MODULE,
690 	.create_ctrl	= nvme_loop_create_ctrl,
691 	.allowed_opts	= NVMF_OPT_TRADDR,
692 };
693 
694 static int __init nvme_loop_init_module(void)
695 {
696 	int ret;
697 
698 	ret = nvmet_register_transport(&nvme_loop_ops);
699 	if (ret)
700 		return ret;
701 
702 	ret = nvmf_register_transport(&nvme_loop_transport);
703 	if (ret)
704 		nvmet_unregister_transport(&nvme_loop_ops);
705 
706 	return ret;
707 }
708 
709 static void __exit nvme_loop_cleanup_module(void)
710 {
711 	struct nvme_loop_ctrl *ctrl, *next;
712 
713 	nvmf_unregister_transport(&nvme_loop_transport);
714 	nvmet_unregister_transport(&nvme_loop_ops);
715 
716 	mutex_lock(&nvme_loop_ctrl_mutex);
717 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
718 		nvme_delete_ctrl(&ctrl->ctrl);
719 	mutex_unlock(&nvme_loop_ctrl_mutex);
720 
721 	flush_workqueue(nvme_delete_wq);
722 }
723 
724 module_init(nvme_loop_init_module);
725 module_exit(nvme_loop_cleanup_module);
726 
727 MODULE_LICENSE("GPL v2");
728 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
729