xref: /openbmc/linux/drivers/nvme/target/core.c (revision 77201135)
1 /*
2  * Common code for the NVMe target.
3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  */
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/module.h>
16 #include <linux/random.h>
17 #include <linux/rculist.h>
18 
19 #include "nvmet.h"
20 
21 static struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
22 static DEFINE_IDA(cntlid_ida);
23 
24 /*
25  * This read/write semaphore is used to synchronize access to configuration
26  * information on a target system that will result in discovery log page
27  * information change for at least one host.
28  * The full list of resources to protected by this semaphore is:
29  *
30  *  - subsystems list
31  *  - per-subsystem allowed hosts list
32  *  - allow_any_host subsystem attribute
33  *  - nvmet_genctr
34  *  - the nvmet_transports array
35  *
36  * When updating any of those lists/structures write lock should be obtained,
37  * while when reading (popolating discovery log page or checking host-subsystem
38  * link) read lock is obtained to allow concurrent reads.
39  */
40 DECLARE_RWSEM(nvmet_config_sem);
41 
42 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
43 		const char *subsysnqn);
44 
45 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
46 		size_t len)
47 {
48 	if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
49 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
50 	return 0;
51 }
52 
53 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
54 {
55 	if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
56 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
57 	return 0;
58 }
59 
60 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
61 {
62 	return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
63 }
64 
65 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
66 {
67 	struct nvmet_req *req;
68 
69 	while (1) {
70 		mutex_lock(&ctrl->lock);
71 		if (!ctrl->nr_async_event_cmds) {
72 			mutex_unlock(&ctrl->lock);
73 			return;
74 		}
75 
76 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
77 		mutex_unlock(&ctrl->lock);
78 		nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
79 	}
80 }
81 
82 static void nvmet_async_event_work(struct work_struct *work)
83 {
84 	struct nvmet_ctrl *ctrl =
85 		container_of(work, struct nvmet_ctrl, async_event_work);
86 	struct nvmet_async_event *aen;
87 	struct nvmet_req *req;
88 
89 	while (1) {
90 		mutex_lock(&ctrl->lock);
91 		aen = list_first_entry_or_null(&ctrl->async_events,
92 				struct nvmet_async_event, entry);
93 		if (!aen || !ctrl->nr_async_event_cmds) {
94 			mutex_unlock(&ctrl->lock);
95 			return;
96 		}
97 
98 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
99 		nvmet_set_result(req, nvmet_async_event_result(aen));
100 
101 		list_del(&aen->entry);
102 		kfree(aen);
103 
104 		mutex_unlock(&ctrl->lock);
105 		nvmet_req_complete(req, 0);
106 	}
107 }
108 
109 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
110 		u8 event_info, u8 log_page)
111 {
112 	struct nvmet_async_event *aen;
113 
114 	aen = kmalloc(sizeof(*aen), GFP_KERNEL);
115 	if (!aen)
116 		return;
117 
118 	aen->event_type = event_type;
119 	aen->event_info = event_info;
120 	aen->log_page = log_page;
121 
122 	mutex_lock(&ctrl->lock);
123 	list_add_tail(&aen->entry, &ctrl->async_events);
124 	mutex_unlock(&ctrl->lock);
125 
126 	schedule_work(&ctrl->async_event_work);
127 }
128 
129 int nvmet_register_transport(struct nvmet_fabrics_ops *ops)
130 {
131 	int ret = 0;
132 
133 	down_write(&nvmet_config_sem);
134 	if (nvmet_transports[ops->type])
135 		ret = -EINVAL;
136 	else
137 		nvmet_transports[ops->type] = ops;
138 	up_write(&nvmet_config_sem);
139 
140 	return ret;
141 }
142 EXPORT_SYMBOL_GPL(nvmet_register_transport);
143 
144 void nvmet_unregister_transport(struct nvmet_fabrics_ops *ops)
145 {
146 	down_write(&nvmet_config_sem);
147 	nvmet_transports[ops->type] = NULL;
148 	up_write(&nvmet_config_sem);
149 }
150 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
151 
152 int nvmet_enable_port(struct nvmet_port *port)
153 {
154 	struct nvmet_fabrics_ops *ops;
155 	int ret;
156 
157 	lockdep_assert_held(&nvmet_config_sem);
158 
159 	ops = nvmet_transports[port->disc_addr.trtype];
160 	if (!ops) {
161 		up_write(&nvmet_config_sem);
162 		request_module("nvmet-transport-%d", port->disc_addr.trtype);
163 		down_write(&nvmet_config_sem);
164 		ops = nvmet_transports[port->disc_addr.trtype];
165 		if (!ops) {
166 			pr_err("transport type %d not supported\n",
167 				port->disc_addr.trtype);
168 			return -EINVAL;
169 		}
170 	}
171 
172 	if (!try_module_get(ops->owner))
173 		return -EINVAL;
174 
175 	ret = ops->add_port(port);
176 	if (ret) {
177 		module_put(ops->owner);
178 		return ret;
179 	}
180 
181 	port->enabled = true;
182 	return 0;
183 }
184 
185 void nvmet_disable_port(struct nvmet_port *port)
186 {
187 	struct nvmet_fabrics_ops *ops;
188 
189 	lockdep_assert_held(&nvmet_config_sem);
190 
191 	port->enabled = false;
192 
193 	ops = nvmet_transports[port->disc_addr.trtype];
194 	ops->remove_port(port);
195 	module_put(ops->owner);
196 }
197 
198 static void nvmet_keep_alive_timer(struct work_struct *work)
199 {
200 	struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
201 			struct nvmet_ctrl, ka_work);
202 
203 	pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
204 		ctrl->cntlid, ctrl->kato);
205 
206 	nvmet_ctrl_fatal_error(ctrl);
207 }
208 
209 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
210 {
211 	pr_debug("ctrl %d start keep-alive timer for %d secs\n",
212 		ctrl->cntlid, ctrl->kato);
213 
214 	INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
215 	schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
216 }
217 
218 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
219 {
220 	pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
221 
222 	cancel_delayed_work_sync(&ctrl->ka_work);
223 }
224 
225 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
226 		__le32 nsid)
227 {
228 	struct nvmet_ns *ns;
229 
230 	list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
231 		if (ns->nsid == le32_to_cpu(nsid))
232 			return ns;
233 	}
234 
235 	return NULL;
236 }
237 
238 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
239 {
240 	struct nvmet_ns *ns;
241 
242 	rcu_read_lock();
243 	ns = __nvmet_find_namespace(ctrl, nsid);
244 	if (ns)
245 		percpu_ref_get(&ns->ref);
246 	rcu_read_unlock();
247 
248 	return ns;
249 }
250 
251 static void nvmet_destroy_namespace(struct percpu_ref *ref)
252 {
253 	struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
254 
255 	complete(&ns->disable_done);
256 }
257 
258 void nvmet_put_namespace(struct nvmet_ns *ns)
259 {
260 	percpu_ref_put(&ns->ref);
261 }
262 
263 int nvmet_ns_enable(struct nvmet_ns *ns)
264 {
265 	struct nvmet_subsys *subsys = ns->subsys;
266 	struct nvmet_ctrl *ctrl;
267 	int ret = 0;
268 
269 	mutex_lock(&subsys->lock);
270 	if (ns->enabled)
271 		goto out_unlock;
272 
273 	ns->bdev = blkdev_get_by_path(ns->device_path, FMODE_READ | FMODE_WRITE,
274 			NULL);
275 	if (IS_ERR(ns->bdev)) {
276 		pr_err("failed to open block device %s: (%ld)\n",
277 		       ns->device_path, PTR_ERR(ns->bdev));
278 		ret = PTR_ERR(ns->bdev);
279 		ns->bdev = NULL;
280 		goto out_unlock;
281 	}
282 
283 	ns->size = i_size_read(ns->bdev->bd_inode);
284 	ns->blksize_shift = blksize_bits(bdev_logical_block_size(ns->bdev));
285 
286 	ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
287 				0, GFP_KERNEL);
288 	if (ret)
289 		goto out_blkdev_put;
290 
291 	if (ns->nsid > subsys->max_nsid)
292 		subsys->max_nsid = ns->nsid;
293 
294 	/*
295 	 * The namespaces list needs to be sorted to simplify the implementation
296 	 * of the Identify Namepace List subcommand.
297 	 */
298 	if (list_empty(&subsys->namespaces)) {
299 		list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
300 	} else {
301 		struct nvmet_ns *old;
302 
303 		list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
304 			BUG_ON(ns->nsid == old->nsid);
305 			if (ns->nsid < old->nsid)
306 				break;
307 		}
308 
309 		list_add_tail_rcu(&ns->dev_link, &old->dev_link);
310 	}
311 
312 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
313 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
314 
315 	ns->enabled = true;
316 	ret = 0;
317 out_unlock:
318 	mutex_unlock(&subsys->lock);
319 	return ret;
320 out_blkdev_put:
321 	blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
322 	ns->bdev = NULL;
323 	goto out_unlock;
324 }
325 
326 void nvmet_ns_disable(struct nvmet_ns *ns)
327 {
328 	struct nvmet_subsys *subsys = ns->subsys;
329 	struct nvmet_ctrl *ctrl;
330 
331 	mutex_lock(&subsys->lock);
332 	if (!ns->enabled)
333 		goto out_unlock;
334 
335 	ns->enabled = false;
336 	list_del_rcu(&ns->dev_link);
337 	mutex_unlock(&subsys->lock);
338 
339 	/*
340 	 * Now that we removed the namespaces from the lookup list, we
341 	 * can kill the per_cpu ref and wait for any remaining references
342 	 * to be dropped, as well as a RCU grace period for anyone only
343 	 * using the namepace under rcu_read_lock().  Note that we can't
344 	 * use call_rcu here as we need to ensure the namespaces have
345 	 * been fully destroyed before unloading the module.
346 	 */
347 	percpu_ref_kill(&ns->ref);
348 	synchronize_rcu();
349 	wait_for_completion(&ns->disable_done);
350 	percpu_ref_exit(&ns->ref);
351 
352 	mutex_lock(&subsys->lock);
353 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
354 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
355 
356 	if (ns->bdev)
357 		blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
358 out_unlock:
359 	mutex_unlock(&subsys->lock);
360 }
361 
362 void nvmet_ns_free(struct nvmet_ns *ns)
363 {
364 	nvmet_ns_disable(ns);
365 
366 	kfree(ns->device_path);
367 	kfree(ns);
368 }
369 
370 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
371 {
372 	struct nvmet_ns *ns;
373 
374 	ns = kzalloc(sizeof(*ns), GFP_KERNEL);
375 	if (!ns)
376 		return NULL;
377 
378 	INIT_LIST_HEAD(&ns->dev_link);
379 	init_completion(&ns->disable_done);
380 
381 	ns->nsid = nsid;
382 	ns->subsys = subsys;
383 
384 	return ns;
385 }
386 
387 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
388 {
389 	if (status)
390 		nvmet_set_status(req, status);
391 
392 	/* XXX: need to fill in something useful for sq_head */
393 	req->rsp->sq_head = 0;
394 	if (likely(req->sq)) /* may happen during early failure */
395 		req->rsp->sq_id = cpu_to_le16(req->sq->qid);
396 	req->rsp->command_id = req->cmd->common.command_id;
397 
398 	if (req->ns)
399 		nvmet_put_namespace(req->ns);
400 	req->ops->queue_response(req);
401 }
402 
403 void nvmet_req_complete(struct nvmet_req *req, u16 status)
404 {
405 	__nvmet_req_complete(req, status);
406 	percpu_ref_put(&req->sq->ref);
407 }
408 EXPORT_SYMBOL_GPL(nvmet_req_complete);
409 
410 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
411 		u16 qid, u16 size)
412 {
413 	cq->qid = qid;
414 	cq->size = size;
415 
416 	ctrl->cqs[qid] = cq;
417 }
418 
419 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
420 		u16 qid, u16 size)
421 {
422 	sq->qid = qid;
423 	sq->size = size;
424 
425 	ctrl->sqs[qid] = sq;
426 }
427 
428 static void nvmet_confirm_sq(struct percpu_ref *ref)
429 {
430 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
431 
432 	complete(&sq->confirm_done);
433 }
434 
435 void nvmet_sq_destroy(struct nvmet_sq *sq)
436 {
437 	/*
438 	 * If this is the admin queue, complete all AERs so that our
439 	 * queue doesn't have outstanding requests on it.
440 	 */
441 	if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
442 		nvmet_async_events_free(sq->ctrl);
443 	percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
444 	wait_for_completion(&sq->confirm_done);
445 	wait_for_completion(&sq->free_done);
446 	percpu_ref_exit(&sq->ref);
447 
448 	if (sq->ctrl) {
449 		nvmet_ctrl_put(sq->ctrl);
450 		sq->ctrl = NULL; /* allows reusing the queue later */
451 	}
452 }
453 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
454 
455 static void nvmet_sq_free(struct percpu_ref *ref)
456 {
457 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
458 
459 	complete(&sq->free_done);
460 }
461 
462 int nvmet_sq_init(struct nvmet_sq *sq)
463 {
464 	int ret;
465 
466 	ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
467 	if (ret) {
468 		pr_err("percpu_ref init failed!\n");
469 		return ret;
470 	}
471 	init_completion(&sq->free_done);
472 	init_completion(&sq->confirm_done);
473 
474 	return 0;
475 }
476 EXPORT_SYMBOL_GPL(nvmet_sq_init);
477 
478 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
479 		struct nvmet_sq *sq, struct nvmet_fabrics_ops *ops)
480 {
481 	u8 flags = req->cmd->common.flags;
482 	u16 status;
483 
484 	req->cq = cq;
485 	req->sq = sq;
486 	req->ops = ops;
487 	req->sg = NULL;
488 	req->sg_cnt = 0;
489 	req->rsp->status = 0;
490 
491 	/* no support for fused commands yet */
492 	if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
493 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
494 		goto fail;
495 	}
496 
497 	/* either variant of SGLs is fine, as we don't support metadata */
498 	if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF &&
499 		     (flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METASEG)) {
500 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
501 		goto fail;
502 	}
503 
504 	if (unlikely(!req->sq->ctrl))
505 		/* will return an error for any Non-connect command: */
506 		status = nvmet_parse_connect_cmd(req);
507 	else if (likely(req->sq->qid != 0))
508 		status = nvmet_parse_io_cmd(req);
509 	else if (req->cmd->common.opcode == nvme_fabrics_command)
510 		status = nvmet_parse_fabrics_cmd(req);
511 	else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
512 		status = nvmet_parse_discovery_cmd(req);
513 	else
514 		status = nvmet_parse_admin_cmd(req);
515 
516 	if (status)
517 		goto fail;
518 
519 	if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
520 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
521 		goto fail;
522 	}
523 
524 	return true;
525 
526 fail:
527 	__nvmet_req_complete(req, status);
528 	return false;
529 }
530 EXPORT_SYMBOL_GPL(nvmet_req_init);
531 
532 void nvmet_req_uninit(struct nvmet_req *req)
533 {
534 	percpu_ref_put(&req->sq->ref);
535 }
536 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
537 
538 static inline bool nvmet_cc_en(u32 cc)
539 {
540 	return cc & 0x1;
541 }
542 
543 static inline u8 nvmet_cc_css(u32 cc)
544 {
545 	return (cc >> 4) & 0x7;
546 }
547 
548 static inline u8 nvmet_cc_mps(u32 cc)
549 {
550 	return (cc >> 7) & 0xf;
551 }
552 
553 static inline u8 nvmet_cc_ams(u32 cc)
554 {
555 	return (cc >> 11) & 0x7;
556 }
557 
558 static inline u8 nvmet_cc_shn(u32 cc)
559 {
560 	return (cc >> 14) & 0x3;
561 }
562 
563 static inline u8 nvmet_cc_iosqes(u32 cc)
564 {
565 	return (cc >> 16) & 0xf;
566 }
567 
568 static inline u8 nvmet_cc_iocqes(u32 cc)
569 {
570 	return (cc >> 20) & 0xf;
571 }
572 
573 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
574 {
575 	lockdep_assert_held(&ctrl->lock);
576 
577 	if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
578 	    nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
579 	    nvmet_cc_mps(ctrl->cc) != 0 ||
580 	    nvmet_cc_ams(ctrl->cc) != 0 ||
581 	    nvmet_cc_css(ctrl->cc) != 0) {
582 		ctrl->csts = NVME_CSTS_CFS;
583 		return;
584 	}
585 
586 	ctrl->csts = NVME_CSTS_RDY;
587 }
588 
589 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
590 {
591 	lockdep_assert_held(&ctrl->lock);
592 
593 	/* XXX: tear down queues? */
594 	ctrl->csts &= ~NVME_CSTS_RDY;
595 	ctrl->cc = 0;
596 }
597 
598 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
599 {
600 	u32 old;
601 
602 	mutex_lock(&ctrl->lock);
603 	old = ctrl->cc;
604 	ctrl->cc = new;
605 
606 	if (nvmet_cc_en(new) && !nvmet_cc_en(old))
607 		nvmet_start_ctrl(ctrl);
608 	if (!nvmet_cc_en(new) && nvmet_cc_en(old))
609 		nvmet_clear_ctrl(ctrl);
610 	if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
611 		nvmet_clear_ctrl(ctrl);
612 		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
613 	}
614 	if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
615 		ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
616 	mutex_unlock(&ctrl->lock);
617 }
618 
619 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
620 {
621 	/* command sets supported: NVMe command set: */
622 	ctrl->cap = (1ULL << 37);
623 	/* CC.EN timeout in 500msec units: */
624 	ctrl->cap |= (15ULL << 24);
625 	/* maximum queue entries supported: */
626 	ctrl->cap |= NVMET_QUEUE_SIZE - 1;
627 }
628 
629 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
630 		struct nvmet_req *req, struct nvmet_ctrl **ret)
631 {
632 	struct nvmet_subsys *subsys;
633 	struct nvmet_ctrl *ctrl;
634 	u16 status = 0;
635 
636 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
637 	if (!subsys) {
638 		pr_warn("connect request for invalid subsystem %s!\n",
639 			subsysnqn);
640 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
641 		return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
642 	}
643 
644 	mutex_lock(&subsys->lock);
645 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
646 		if (ctrl->cntlid == cntlid) {
647 			if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
648 				pr_warn("hostnqn mismatch.\n");
649 				continue;
650 			}
651 			if (!kref_get_unless_zero(&ctrl->ref))
652 				continue;
653 
654 			*ret = ctrl;
655 			goto out;
656 		}
657 	}
658 
659 	pr_warn("could not find controller %d for subsys %s / host %s\n",
660 		cntlid, subsysnqn, hostnqn);
661 	req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
662 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
663 
664 out:
665 	mutex_unlock(&subsys->lock);
666 	nvmet_subsys_put(subsys);
667 	return status;
668 }
669 
670 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
671 {
672 	if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
673 		pr_err("got io cmd %d while CC.EN == 0 on qid = %d\n",
674 		       cmd->common.opcode, req->sq->qid);
675 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
676 	}
677 
678 	if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
679 		pr_err("got io cmd %d while CSTS.RDY == 0 on qid = %d\n",
680 		       cmd->common.opcode, req->sq->qid);
681 		req->ns = NULL;
682 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
683 	}
684 	return 0;
685 }
686 
687 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
688 		const char *hostnqn)
689 {
690 	struct nvmet_host_link *p;
691 
692 	if (subsys->allow_any_host)
693 		return true;
694 
695 	list_for_each_entry(p, &subsys->hosts, entry) {
696 		if (!strcmp(nvmet_host_name(p->host), hostnqn))
697 			return true;
698 	}
699 
700 	return false;
701 }
702 
703 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
704 		const char *hostnqn)
705 {
706 	struct nvmet_subsys_link *s;
707 
708 	list_for_each_entry(s, &req->port->subsystems, entry) {
709 		if (__nvmet_host_allowed(s->subsys, hostnqn))
710 			return true;
711 	}
712 
713 	return false;
714 }
715 
716 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
717 		const char *hostnqn)
718 {
719 	lockdep_assert_held(&nvmet_config_sem);
720 
721 	if (subsys->type == NVME_NQN_DISC)
722 		return nvmet_host_discovery_allowed(req, hostnqn);
723 	else
724 		return __nvmet_host_allowed(subsys, hostnqn);
725 }
726 
727 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
728 		struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
729 {
730 	struct nvmet_subsys *subsys;
731 	struct nvmet_ctrl *ctrl;
732 	int ret;
733 	u16 status;
734 
735 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
736 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
737 	if (!subsys) {
738 		pr_warn("connect request for invalid subsystem %s!\n",
739 			subsysnqn);
740 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
741 		goto out;
742 	}
743 
744 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
745 	down_read(&nvmet_config_sem);
746 	if (!nvmet_host_allowed(req, subsys, hostnqn)) {
747 		pr_info("connect by host %s for subsystem %s not allowed\n",
748 			hostnqn, subsysnqn);
749 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
750 		up_read(&nvmet_config_sem);
751 		goto out_put_subsystem;
752 	}
753 	up_read(&nvmet_config_sem);
754 
755 	status = NVME_SC_INTERNAL;
756 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
757 	if (!ctrl)
758 		goto out_put_subsystem;
759 	mutex_init(&ctrl->lock);
760 
761 	nvmet_init_cap(ctrl);
762 
763 	INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
764 	INIT_LIST_HEAD(&ctrl->async_events);
765 
766 	memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
767 	memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
768 
769 	/* generate a random serial number as our controllers are ephemeral: */
770 	get_random_bytes(&ctrl->serial, sizeof(ctrl->serial));
771 
772 	kref_init(&ctrl->ref);
773 	ctrl->subsys = subsys;
774 
775 	ctrl->cqs = kcalloc(subsys->max_qid + 1,
776 			sizeof(struct nvmet_cq *),
777 			GFP_KERNEL);
778 	if (!ctrl->cqs)
779 		goto out_free_ctrl;
780 
781 	ctrl->sqs = kcalloc(subsys->max_qid + 1,
782 			sizeof(struct nvmet_sq *),
783 			GFP_KERNEL);
784 	if (!ctrl->sqs)
785 		goto out_free_cqs;
786 
787 	ret = ida_simple_get(&cntlid_ida,
788 			     NVME_CNTLID_MIN, NVME_CNTLID_MAX,
789 			     GFP_KERNEL);
790 	if (ret < 0) {
791 		status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
792 		goto out_free_sqs;
793 	}
794 	ctrl->cntlid = ret;
795 
796 	ctrl->ops = req->ops;
797 	if (ctrl->subsys->type == NVME_NQN_DISC) {
798 		/* Don't accept keep-alive timeout for discovery controllers */
799 		if (kato) {
800 			status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
801 			goto out_free_sqs;
802 		}
803 
804 		/*
805 		 * Discovery controllers use some arbitrary high value in order
806 		 * to cleanup stale discovery sessions
807 		 *
808 		 * From the latest base diff RC:
809 		 * "The Keep Alive command is not supported by
810 		 * Discovery controllers. A transport may specify a
811 		 * fixed Discovery controller activity timeout value
812 		 * (e.g., 2 minutes).  If no commands are received
813 		 * by a Discovery controller within that time
814 		 * period, the controller may perform the
815 		 * actions for Keep Alive Timer expiration".
816 		 */
817 		ctrl->kato = NVMET_DISC_KATO;
818 	} else {
819 		/* keep-alive timeout in seconds */
820 		ctrl->kato = DIV_ROUND_UP(kato, 1000);
821 	}
822 	nvmet_start_keep_alive_timer(ctrl);
823 
824 	mutex_lock(&subsys->lock);
825 	list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
826 	mutex_unlock(&subsys->lock);
827 
828 	*ctrlp = ctrl;
829 	return 0;
830 
831 out_free_sqs:
832 	kfree(ctrl->sqs);
833 out_free_cqs:
834 	kfree(ctrl->cqs);
835 out_free_ctrl:
836 	kfree(ctrl);
837 out_put_subsystem:
838 	nvmet_subsys_put(subsys);
839 out:
840 	return status;
841 }
842 
843 static void nvmet_ctrl_free(struct kref *ref)
844 {
845 	struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
846 	struct nvmet_subsys *subsys = ctrl->subsys;
847 
848 	nvmet_stop_keep_alive_timer(ctrl);
849 
850 	mutex_lock(&subsys->lock);
851 	list_del(&ctrl->subsys_entry);
852 	mutex_unlock(&subsys->lock);
853 
854 	flush_work(&ctrl->async_event_work);
855 	cancel_work_sync(&ctrl->fatal_err_work);
856 
857 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
858 	nvmet_subsys_put(subsys);
859 
860 	kfree(ctrl->sqs);
861 	kfree(ctrl->cqs);
862 	kfree(ctrl);
863 }
864 
865 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
866 {
867 	kref_put(&ctrl->ref, nvmet_ctrl_free);
868 }
869 
870 static void nvmet_fatal_error_handler(struct work_struct *work)
871 {
872 	struct nvmet_ctrl *ctrl =
873 			container_of(work, struct nvmet_ctrl, fatal_err_work);
874 
875 	pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
876 	ctrl->ops->delete_ctrl(ctrl);
877 }
878 
879 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
880 {
881 	mutex_lock(&ctrl->lock);
882 	if (!(ctrl->csts & NVME_CSTS_CFS)) {
883 		ctrl->csts |= NVME_CSTS_CFS;
884 		INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
885 		schedule_work(&ctrl->fatal_err_work);
886 	}
887 	mutex_unlock(&ctrl->lock);
888 }
889 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
890 
891 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
892 		const char *subsysnqn)
893 {
894 	struct nvmet_subsys_link *p;
895 
896 	if (!port)
897 		return NULL;
898 
899 	if (!strncmp(NVME_DISC_SUBSYS_NAME, subsysnqn,
900 			NVMF_NQN_SIZE)) {
901 		if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
902 			return NULL;
903 		return nvmet_disc_subsys;
904 	}
905 
906 	down_read(&nvmet_config_sem);
907 	list_for_each_entry(p, &port->subsystems, entry) {
908 		if (!strncmp(p->subsys->subsysnqn, subsysnqn,
909 				NVMF_NQN_SIZE)) {
910 			if (!kref_get_unless_zero(&p->subsys->ref))
911 				break;
912 			up_read(&nvmet_config_sem);
913 			return p->subsys;
914 		}
915 	}
916 	up_read(&nvmet_config_sem);
917 	return NULL;
918 }
919 
920 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
921 		enum nvme_subsys_type type)
922 {
923 	struct nvmet_subsys *subsys;
924 
925 	subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
926 	if (!subsys)
927 		return NULL;
928 
929 	subsys->ver = NVME_VS(1, 2, 1); /* NVMe 1.2.1 */
930 
931 	switch (type) {
932 	case NVME_NQN_NVME:
933 		subsys->max_qid = NVMET_NR_QUEUES;
934 		break;
935 	case NVME_NQN_DISC:
936 		subsys->max_qid = 0;
937 		break;
938 	default:
939 		pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
940 		kfree(subsys);
941 		return NULL;
942 	}
943 	subsys->type = type;
944 	subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
945 			GFP_KERNEL);
946 	if (!subsys->subsysnqn) {
947 		kfree(subsys);
948 		return NULL;
949 	}
950 
951 	kref_init(&subsys->ref);
952 
953 	mutex_init(&subsys->lock);
954 	INIT_LIST_HEAD(&subsys->namespaces);
955 	INIT_LIST_HEAD(&subsys->ctrls);
956 	INIT_LIST_HEAD(&subsys->hosts);
957 
958 	return subsys;
959 }
960 
961 static void nvmet_subsys_free(struct kref *ref)
962 {
963 	struct nvmet_subsys *subsys =
964 		container_of(ref, struct nvmet_subsys, ref);
965 
966 	WARN_ON_ONCE(!list_empty(&subsys->namespaces));
967 
968 	kfree(subsys->subsysnqn);
969 	kfree(subsys);
970 }
971 
972 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
973 {
974 	struct nvmet_ctrl *ctrl;
975 
976 	mutex_lock(&subsys->lock);
977 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
978 		ctrl->ops->delete_ctrl(ctrl);
979 	mutex_unlock(&subsys->lock);
980 }
981 
982 void nvmet_subsys_put(struct nvmet_subsys *subsys)
983 {
984 	kref_put(&subsys->ref, nvmet_subsys_free);
985 }
986 
987 static int __init nvmet_init(void)
988 {
989 	int error;
990 
991 	error = nvmet_init_discovery();
992 	if (error)
993 		goto out;
994 
995 	error = nvmet_init_configfs();
996 	if (error)
997 		goto out_exit_discovery;
998 	return 0;
999 
1000 out_exit_discovery:
1001 	nvmet_exit_discovery();
1002 out:
1003 	return error;
1004 }
1005 
1006 static void __exit nvmet_exit(void)
1007 {
1008 	nvmet_exit_configfs();
1009 	nvmet_exit_discovery();
1010 	ida_destroy(&cntlid_ida);
1011 
1012 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1013 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1014 }
1015 
1016 module_init(nvmet_init);
1017 module_exit(nvmet_exit);
1018 
1019 MODULE_LICENSE("GPL v2");
1020