xref: /linux/drivers/nvme/target/loop.c (revision 4b99990cdf9560e8a071640baf19f312e6ae02f4)
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 		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 }
278 
279 static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
280 {
281 	struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
282 
283 	if (list_empty(&ctrl->list))
284 		goto free_ctrl;
285 
286 	mutex_lock(&nvme_loop_ctrl_mutex);
287 	list_del(&ctrl->list);
288 	mutex_unlock(&nvme_loop_ctrl_mutex);
289 
290 	if (nctrl->tagset)
291 		nvme_remove_io_tag_set(nctrl);
292 	kfree(ctrl->queues);
293 	nvmf_free_options(nctrl->opts);
294 free_ctrl:
295 	kfree(ctrl);
296 }
297 
298 static void nvme_loop_destroy_io_queues(struct nvme_loop_ctrl *ctrl)
299 {
300 	int i;
301 
302 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
303 		clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
304 		nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
305 		nvmet_cq_put(&ctrl->queues[i].nvme_cq);
306 	}
307 	ctrl->ctrl.queue_count = 1;
308 	/*
309 	 * It's possible that some requests might have been added
310 	 * after io queue is stopped/quiesced. So now start the
311 	 * queue to flush these requests to the completion.
312 	 */
313 	nvme_unquiesce_io_queues(&ctrl->ctrl);
314 }
315 
316 static int nvme_loop_init_io_queues(struct nvme_loop_ctrl *ctrl)
317 {
318 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
319 	unsigned int nr_io_queues;
320 	int ret, i;
321 
322 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
323 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
324 	if (ret || !nr_io_queues)
325 		return ret;
326 
327 	dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n", nr_io_queues);
328 
329 	for (i = 1; i <= nr_io_queues; i++) {
330 		ctrl->queues[i].ctrl = ctrl;
331 		nvmet_cq_init(&ctrl->queues[i].nvme_cq);
332 		ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq,
333 				&ctrl->queues[i].nvme_cq);
334 		if (ret) {
335 			nvmet_cq_put(&ctrl->queues[i].nvme_cq);
336 			goto out_destroy_queues;
337 		}
338 
339 		ctrl->ctrl.queue_count++;
340 	}
341 
342 	return 0;
343 
344 out_destroy_queues:
345 	nvme_loop_destroy_io_queues(ctrl);
346 	return ret;
347 }
348 
349 static int nvme_loop_connect_io_queues(struct nvme_loop_ctrl *ctrl)
350 {
351 	int i, ret;
352 
353 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
354 		ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
355 		if (ret)
356 			return ret;
357 		set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[i].flags);
358 	}
359 
360 	return 0;
361 }
362 
363 static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
364 {
365 	int error;
366 
367 	ctrl->queues[0].ctrl = ctrl;
368 	nvmet_cq_init(&ctrl->queues[0].nvme_cq);
369 	error = nvmet_sq_init(&ctrl->queues[0].nvme_sq,
370 			&ctrl->queues[0].nvme_cq);
371 	if (error) {
372 		nvmet_cq_put(&ctrl->queues[0].nvme_cq);
373 		return error;
374 	}
375 	ctrl->ctrl.queue_count = 1;
376 
377 	/* reset stopped state for the fresh admin queue */
378 	clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->ctrl.flags);
379 
380 	error = nvmf_connect_admin_queue(&ctrl->ctrl);
381 	if (error)
382 		goto out_free_sq;
383 
384 	set_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
385 
386 	error = nvme_enable_ctrl(&ctrl->ctrl);
387 	if (error)
388 		goto out_free_sq;
389 
390 	ctrl->ctrl.max_hw_sectors =
391 		(NVME_LOOP_MAX_SEGMENTS - 1) << PAGE_SECTORS_SHIFT;
392 
393 	nvme_unquiesce_admin_queue(&ctrl->ctrl);
394 
395 	error = nvme_init_ctrl_finish(&ctrl->ctrl, false);
396 	if (error)
397 		goto out_free_sq;
398 
399 	return 0;
400 
401 out_free_sq:
402 	clear_bit(NVME_LOOP_Q_LIVE, &ctrl->queues[0].flags);
403 	nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
404 	nvmet_cq_put(&ctrl->queues[0].nvme_cq);
405 	return error;
406 }
407 
408 static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
409 {
410 	if (ctrl->ctrl.queue_count > 1) {
411 		nvme_quiesce_io_queues(&ctrl->ctrl);
412 		nvme_loop_destroy_io_queues(ctrl);
413 	}
414 
415 	nvme_quiesce_admin_queue(&ctrl->ctrl);
416 	if (nvme_ctrl_state(&ctrl->ctrl) == NVME_CTRL_LIVE)
417 		nvme_disable_ctrl(&ctrl->ctrl, true);
418 
419 	nvme_loop_destroy_admin_queue(ctrl);
420 }
421 
422 static void nvme_loop_delete_ctrl_host(struct nvme_ctrl *ctrl)
423 {
424 	nvme_loop_shutdown_ctrl(to_loop_ctrl(ctrl));
425 	nvme_remove_admin_tag_set(ctrl);
426 }
427 
428 static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
429 {
430 	struct nvme_loop_ctrl *ctrl;
431 
432 	mutex_lock(&nvme_loop_ctrl_mutex);
433 	list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
434 		if (ctrl->ctrl.cntlid == nctrl->cntlid)
435 			nvme_delete_ctrl(&ctrl->ctrl);
436 	}
437 	mutex_unlock(&nvme_loop_ctrl_mutex);
438 }
439 
440 static void nvme_loop_reset_ctrl_work(struct work_struct *work)
441 {
442 	struct nvme_loop_ctrl *ctrl =
443 		container_of(work, struct nvme_loop_ctrl, ctrl.reset_work);
444 	int ret;
445 
446 	nvme_stop_ctrl(&ctrl->ctrl);
447 	nvme_loop_shutdown_ctrl(ctrl);
448 
449 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
450 		enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
451 
452 		if (state != NVME_CTRL_DELETING &&
453 		    state != NVME_CTRL_DELETING_NOIO)
454 			/* state change failure for non-deleted ctrl? */
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 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
475 		WARN_ON_ONCE(1);
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_quiesce_admin_queue(&ctrl->ctrl);
485 	nvme_cancel_admin_tagset(&ctrl->ctrl);
486 	nvme_loop_destroy_admin_queue(ctrl);
487 out_disable:
488 	nvme_remove_admin_tag_set(&ctrl->ctrl);
489 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
490 	nvme_uninit_ctrl(&ctrl->ctrl);
491 }
492 
493 static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
494 	.name			= "loop",
495 	.module			= THIS_MODULE,
496 	.flags			= NVME_F_FABRICS,
497 	.reg_read32		= nvmf_reg_read32,
498 	.reg_read64		= nvmf_reg_read64,
499 	.reg_write32		= nvmf_reg_write32,
500 	.free_ctrl		= nvme_loop_free_ctrl,
501 	.submit_async_event	= nvme_loop_submit_async_event,
502 	.delete_ctrl		= nvme_loop_delete_ctrl_host,
503 	.get_address		= nvmf_get_address,
504 	.get_virt_boundary	= nvme_get_virt_boundary,
505 };
506 
507 static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
508 {
509 	int ret;
510 
511 	ret = nvme_loop_init_io_queues(ctrl);
512 	if (ret)
513 		return ret;
514 
515 	ret = nvme_alloc_io_tag_set(&ctrl->ctrl, &ctrl->tag_set,
516 			&nvme_loop_mq_ops, 1,
517 			sizeof(struct nvme_loop_iod) +
518 			NVME_INLINE_SG_CNT * sizeof(struct scatterlist));
519 	if (ret)
520 		goto out_destroy_queues;
521 
522 	ret = nvme_loop_connect_io_queues(ctrl);
523 	if (ret)
524 		goto out_cleanup_tagset;
525 
526 	return 0;
527 
528 out_cleanup_tagset:
529 	nvme_remove_io_tag_set(&ctrl->ctrl);
530 out_destroy_queues:
531 	nvme_loop_destroy_io_queues(ctrl);
532 	return ret;
533 }
534 
535 static struct nvmet_port *nvme_loop_find_port(struct nvme_ctrl *ctrl)
536 {
537 	struct nvmet_port *p, *found = NULL;
538 
539 	mutex_lock(&nvme_loop_ports_mutex);
540 	list_for_each_entry(p, &nvme_loop_ports, entry) {
541 		/* if no transport address is specified use the first port */
542 		if ((ctrl->opts->mask & NVMF_OPT_TRADDR) &&
543 		    strcmp(ctrl->opts->traddr, p->disc_addr.traddr))
544 			continue;
545 		found = p;
546 		break;
547 	}
548 	mutex_unlock(&nvme_loop_ports_mutex);
549 	return found;
550 }
551 
552 static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
553 		struct nvmf_ctrl_options *opts)
554 {
555 	struct nvme_loop_ctrl *ctrl;
556 	int ret;
557 
558 	ctrl = kzalloc_obj(*ctrl);
559 	if (!ctrl)
560 		return ERR_PTR(-ENOMEM);
561 	ctrl->ctrl.opts = opts;
562 	INIT_LIST_HEAD(&ctrl->list);
563 
564 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_loop_reset_ctrl_work);
565 
566 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
567 				0 /* no quirks, we're perfect! */);
568 	if (ret) {
569 		kfree(ctrl);
570 		goto out;
571 	}
572 
573 	ret = nvme_add_ctrl(&ctrl->ctrl);
574 	if (ret)
575 		goto out_put_ctrl;
576 
577 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING))
578 		WARN_ON_ONCE(1);
579 
580 	ret = -ENOMEM;
581 
582 	ctrl->ctrl.kato = opts->kato;
583 	ctrl->port = nvme_loop_find_port(&ctrl->ctrl);
584 
585 	ctrl->queues = kzalloc_objs(*ctrl->queues, opts->nr_io_queues + 1);
586 	if (!ctrl->queues)
587 		goto out_uninit_ctrl;
588 
589 	ret = nvme_alloc_admin_tag_set(&ctrl->ctrl, &ctrl->admin_tag_set,
590 			&nvme_loop_admin_mq_ops,
591 			sizeof(struct nvme_loop_iod) +
592 			NVME_INLINE_SG_CNT * sizeof(struct scatterlist));
593 	if (ret)
594 		goto out_free_queues;
595 
596 	ret = nvme_loop_configure_admin_queue(ctrl);
597 	if (ret)
598 		goto out_remove_admin_tagset;
599 
600 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
601 		/* warn if maxcmd is lower than queue_size */
602 		dev_warn(ctrl->ctrl.device,
603 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
604 			opts->queue_size, ctrl->ctrl.maxcmd);
605 		opts->queue_size = ctrl->ctrl.maxcmd;
606 	}
607 	ctrl->ctrl.sqsize = opts->queue_size - 1;
608 
609 	if (opts->nr_io_queues) {
610 		ret = nvme_loop_create_io_queues(ctrl);
611 		if (ret)
612 			goto out_remove_admin_queue;
613 	}
614 
615 	nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
616 
617 	dev_info(ctrl->ctrl.device,
618 		 "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
619 
620 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE))
621 		WARN_ON_ONCE(1);
622 
623 	mutex_lock(&nvme_loop_ctrl_mutex);
624 	list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
625 	mutex_unlock(&nvme_loop_ctrl_mutex);
626 
627 	nvme_start_ctrl(&ctrl->ctrl);
628 
629 	return &ctrl->ctrl;
630 
631 out_remove_admin_queue:
632 	nvme_quiesce_admin_queue(&ctrl->ctrl);
633 	nvme_cancel_admin_tagset(&ctrl->ctrl);
634 	nvme_loop_destroy_admin_queue(ctrl);
635 out_remove_admin_tagset:
636 	nvme_remove_admin_tag_set(&ctrl->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 out:
644 	if (ret > 0)
645 		ret = -EIO;
646 	return ERR_PTR(ret);
647 }
648 
649 static int nvme_loop_add_port(struct nvmet_port *port)
650 {
651 	mutex_lock(&nvme_loop_ports_mutex);
652 	list_add_tail(&port->entry, &nvme_loop_ports);
653 	mutex_unlock(&nvme_loop_ports_mutex);
654 	return 0;
655 }
656 
657 static void nvme_loop_remove_port(struct nvmet_port *port)
658 {
659 	mutex_lock(&nvme_loop_ports_mutex);
660 	list_del_init(&port->entry);
661 	mutex_unlock(&nvme_loop_ports_mutex);
662 
663 	/*
664 	 * Ensure any ctrls that are in the process of being
665 	 * deleted are in fact deleted before we return
666 	 * and free the port. This is to prevent active
667 	 * ctrls from using a port after it's freed.
668 	 */
669 	flush_workqueue(nvme_delete_wq);
670 }
671 
672 static const struct nvmet_fabrics_ops nvme_loop_ops = {
673 	.owner		= THIS_MODULE,
674 	.type		= NVMF_TRTYPE_LOOP,
675 	.add_port	= nvme_loop_add_port,
676 	.remove_port	= nvme_loop_remove_port,
677 	.queue_response = nvme_loop_queue_response,
678 	.delete_ctrl	= nvme_loop_delete_ctrl,
679 };
680 
681 static struct nvmf_transport_ops nvme_loop_transport = {
682 	.name		= "loop",
683 	.module		= THIS_MODULE,
684 	.create_ctrl	= nvme_loop_create_ctrl,
685 	.allowed_opts	= NVMF_OPT_TRADDR,
686 };
687 
688 static int __init nvme_loop_init_module(void)
689 {
690 	int ret;
691 
692 	ret = nvmet_register_transport(&nvme_loop_ops);
693 	if (ret)
694 		return ret;
695 
696 	ret = nvmf_register_transport(&nvme_loop_transport);
697 	if (ret)
698 		nvmet_unregister_transport(&nvme_loop_ops);
699 
700 	return ret;
701 }
702 
703 static void __exit nvme_loop_cleanup_module(void)
704 {
705 	struct nvme_loop_ctrl *ctrl, *next;
706 
707 	nvmf_unregister_transport(&nvme_loop_transport);
708 	nvmet_unregister_transport(&nvme_loop_ops);
709 
710 	mutex_lock(&nvme_loop_ctrl_mutex);
711 	list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
712 		nvme_delete_ctrl(&ctrl->ctrl);
713 	mutex_unlock(&nvme_loop_ctrl_mutex);
714 
715 	flush_workqueue(nvme_delete_wq);
716 }
717 
718 module_init(nvme_loop_init_module);
719 module_exit(nvme_loop_cleanup_module);
720 
721 MODULE_DESCRIPTION("NVMe target loop transport driver");
722 MODULE_LICENSE("GPL v2");
723 MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */
724