xref: /openbmc/linux/drivers/nvme/target/core.c (revision ddc141e5)
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 unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
61 {
62 	struct nvmet_ns *ns;
63 
64 	if (list_empty(&subsys->namespaces))
65 		return 0;
66 
67 	ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
68 	return ns->nsid;
69 }
70 
71 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
72 {
73 	return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
74 }
75 
76 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
77 {
78 	struct nvmet_req *req;
79 
80 	while (1) {
81 		mutex_lock(&ctrl->lock);
82 		if (!ctrl->nr_async_event_cmds) {
83 			mutex_unlock(&ctrl->lock);
84 			return;
85 		}
86 
87 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
88 		mutex_unlock(&ctrl->lock);
89 		nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
90 	}
91 }
92 
93 static void nvmet_async_event_work(struct work_struct *work)
94 {
95 	struct nvmet_ctrl *ctrl =
96 		container_of(work, struct nvmet_ctrl, async_event_work);
97 	struct nvmet_async_event *aen;
98 	struct nvmet_req *req;
99 
100 	while (1) {
101 		mutex_lock(&ctrl->lock);
102 		aen = list_first_entry_or_null(&ctrl->async_events,
103 				struct nvmet_async_event, entry);
104 		if (!aen || !ctrl->nr_async_event_cmds) {
105 			mutex_unlock(&ctrl->lock);
106 			return;
107 		}
108 
109 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
110 		nvmet_set_result(req, nvmet_async_event_result(aen));
111 
112 		list_del(&aen->entry);
113 		kfree(aen);
114 
115 		mutex_unlock(&ctrl->lock);
116 		nvmet_req_complete(req, 0);
117 	}
118 }
119 
120 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
121 		u8 event_info, u8 log_page)
122 {
123 	struct nvmet_async_event *aen;
124 
125 	aen = kmalloc(sizeof(*aen), GFP_KERNEL);
126 	if (!aen)
127 		return;
128 
129 	aen->event_type = event_type;
130 	aen->event_info = event_info;
131 	aen->log_page = log_page;
132 
133 	mutex_lock(&ctrl->lock);
134 	list_add_tail(&aen->entry, &ctrl->async_events);
135 	mutex_unlock(&ctrl->lock);
136 
137 	schedule_work(&ctrl->async_event_work);
138 }
139 
140 int nvmet_register_transport(struct nvmet_fabrics_ops *ops)
141 {
142 	int ret = 0;
143 
144 	down_write(&nvmet_config_sem);
145 	if (nvmet_transports[ops->type])
146 		ret = -EINVAL;
147 	else
148 		nvmet_transports[ops->type] = ops;
149 	up_write(&nvmet_config_sem);
150 
151 	return ret;
152 }
153 EXPORT_SYMBOL_GPL(nvmet_register_transport);
154 
155 void nvmet_unregister_transport(struct nvmet_fabrics_ops *ops)
156 {
157 	down_write(&nvmet_config_sem);
158 	nvmet_transports[ops->type] = NULL;
159 	up_write(&nvmet_config_sem);
160 }
161 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
162 
163 int nvmet_enable_port(struct nvmet_port *port)
164 {
165 	struct nvmet_fabrics_ops *ops;
166 	int ret;
167 
168 	lockdep_assert_held(&nvmet_config_sem);
169 
170 	ops = nvmet_transports[port->disc_addr.trtype];
171 	if (!ops) {
172 		up_write(&nvmet_config_sem);
173 		request_module("nvmet-transport-%d", port->disc_addr.trtype);
174 		down_write(&nvmet_config_sem);
175 		ops = nvmet_transports[port->disc_addr.trtype];
176 		if (!ops) {
177 			pr_err("transport type %d not supported\n",
178 				port->disc_addr.trtype);
179 			return -EINVAL;
180 		}
181 	}
182 
183 	if (!try_module_get(ops->owner))
184 		return -EINVAL;
185 
186 	ret = ops->add_port(port);
187 	if (ret) {
188 		module_put(ops->owner);
189 		return ret;
190 	}
191 
192 	port->enabled = true;
193 	return 0;
194 }
195 
196 void nvmet_disable_port(struct nvmet_port *port)
197 {
198 	struct nvmet_fabrics_ops *ops;
199 
200 	lockdep_assert_held(&nvmet_config_sem);
201 
202 	port->enabled = false;
203 
204 	ops = nvmet_transports[port->disc_addr.trtype];
205 	ops->remove_port(port);
206 	module_put(ops->owner);
207 }
208 
209 static void nvmet_keep_alive_timer(struct work_struct *work)
210 {
211 	struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
212 			struct nvmet_ctrl, ka_work);
213 
214 	pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
215 		ctrl->cntlid, ctrl->kato);
216 
217 	nvmet_ctrl_fatal_error(ctrl);
218 }
219 
220 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
221 {
222 	pr_debug("ctrl %d start keep-alive timer for %d secs\n",
223 		ctrl->cntlid, ctrl->kato);
224 
225 	INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
226 	schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
227 }
228 
229 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
230 {
231 	pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
232 
233 	cancel_delayed_work_sync(&ctrl->ka_work);
234 }
235 
236 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
237 		__le32 nsid)
238 {
239 	struct nvmet_ns *ns;
240 
241 	list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
242 		if (ns->nsid == le32_to_cpu(nsid))
243 			return ns;
244 	}
245 
246 	return NULL;
247 }
248 
249 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
250 {
251 	struct nvmet_ns *ns;
252 
253 	rcu_read_lock();
254 	ns = __nvmet_find_namespace(ctrl, nsid);
255 	if (ns)
256 		percpu_ref_get(&ns->ref);
257 	rcu_read_unlock();
258 
259 	return ns;
260 }
261 
262 static void nvmet_destroy_namespace(struct percpu_ref *ref)
263 {
264 	struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
265 
266 	complete(&ns->disable_done);
267 }
268 
269 void nvmet_put_namespace(struct nvmet_ns *ns)
270 {
271 	percpu_ref_put(&ns->ref);
272 }
273 
274 int nvmet_ns_enable(struct nvmet_ns *ns)
275 {
276 	struct nvmet_subsys *subsys = ns->subsys;
277 	struct nvmet_ctrl *ctrl;
278 	int ret = 0;
279 
280 	mutex_lock(&subsys->lock);
281 	if (ns->enabled)
282 		goto out_unlock;
283 
284 	ns->bdev = blkdev_get_by_path(ns->device_path, FMODE_READ | FMODE_WRITE,
285 			NULL);
286 	if (IS_ERR(ns->bdev)) {
287 		pr_err("failed to open block device %s: (%ld)\n",
288 		       ns->device_path, PTR_ERR(ns->bdev));
289 		ret = PTR_ERR(ns->bdev);
290 		ns->bdev = NULL;
291 		goto out_unlock;
292 	}
293 
294 	ns->size = i_size_read(ns->bdev->bd_inode);
295 	ns->blksize_shift = blksize_bits(bdev_logical_block_size(ns->bdev));
296 
297 	ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
298 				0, GFP_KERNEL);
299 	if (ret)
300 		goto out_blkdev_put;
301 
302 	if (ns->nsid > subsys->max_nsid)
303 		subsys->max_nsid = ns->nsid;
304 
305 	/*
306 	 * The namespaces list needs to be sorted to simplify the implementation
307 	 * of the Identify Namepace List subcommand.
308 	 */
309 	if (list_empty(&subsys->namespaces)) {
310 		list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
311 	} else {
312 		struct nvmet_ns *old;
313 
314 		list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
315 			BUG_ON(ns->nsid == old->nsid);
316 			if (ns->nsid < old->nsid)
317 				break;
318 		}
319 
320 		list_add_tail_rcu(&ns->dev_link, &old->dev_link);
321 	}
322 
323 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
324 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
325 
326 	ns->enabled = true;
327 	ret = 0;
328 out_unlock:
329 	mutex_unlock(&subsys->lock);
330 	return ret;
331 out_blkdev_put:
332 	blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
333 	ns->bdev = NULL;
334 	goto out_unlock;
335 }
336 
337 void nvmet_ns_disable(struct nvmet_ns *ns)
338 {
339 	struct nvmet_subsys *subsys = ns->subsys;
340 	struct nvmet_ctrl *ctrl;
341 
342 	mutex_lock(&subsys->lock);
343 	if (!ns->enabled)
344 		goto out_unlock;
345 
346 	ns->enabled = false;
347 	list_del_rcu(&ns->dev_link);
348 	if (ns->nsid == subsys->max_nsid)
349 		subsys->max_nsid = nvmet_max_nsid(subsys);
350 	mutex_unlock(&subsys->lock);
351 
352 	/*
353 	 * Now that we removed the namespaces from the lookup list, we
354 	 * can kill the per_cpu ref and wait for any remaining references
355 	 * to be dropped, as well as a RCU grace period for anyone only
356 	 * using the namepace under rcu_read_lock().  Note that we can't
357 	 * use call_rcu here as we need to ensure the namespaces have
358 	 * been fully destroyed before unloading the module.
359 	 */
360 	percpu_ref_kill(&ns->ref);
361 	synchronize_rcu();
362 	wait_for_completion(&ns->disable_done);
363 	percpu_ref_exit(&ns->ref);
364 
365 	mutex_lock(&subsys->lock);
366 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
367 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
368 
369 	if (ns->bdev)
370 		blkdev_put(ns->bdev, FMODE_WRITE|FMODE_READ);
371 out_unlock:
372 	mutex_unlock(&subsys->lock);
373 }
374 
375 void nvmet_ns_free(struct nvmet_ns *ns)
376 {
377 	nvmet_ns_disable(ns);
378 
379 	kfree(ns->device_path);
380 	kfree(ns);
381 }
382 
383 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
384 {
385 	struct nvmet_ns *ns;
386 
387 	ns = kzalloc(sizeof(*ns), GFP_KERNEL);
388 	if (!ns)
389 		return NULL;
390 
391 	INIT_LIST_HEAD(&ns->dev_link);
392 	init_completion(&ns->disable_done);
393 
394 	ns->nsid = nsid;
395 	ns->subsys = subsys;
396 	uuid_gen(&ns->uuid);
397 
398 	return ns;
399 }
400 
401 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
402 {
403 	u32 old_sqhd, new_sqhd;
404 	u16 sqhd;
405 
406 	if (status)
407 		nvmet_set_status(req, status);
408 
409 	if (req->sq->size) {
410 		do {
411 			old_sqhd = req->sq->sqhd;
412 			new_sqhd = (old_sqhd + 1) % req->sq->size;
413 		} while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
414 					old_sqhd);
415 	}
416 	sqhd = req->sq->sqhd & 0x0000FFFF;
417 	req->rsp->sq_head = cpu_to_le16(sqhd);
418 	req->rsp->sq_id = cpu_to_le16(req->sq->qid);
419 	req->rsp->command_id = req->cmd->common.command_id;
420 
421 	if (req->ns)
422 		nvmet_put_namespace(req->ns);
423 	req->ops->queue_response(req);
424 }
425 
426 void nvmet_req_complete(struct nvmet_req *req, u16 status)
427 {
428 	__nvmet_req_complete(req, status);
429 	percpu_ref_put(&req->sq->ref);
430 }
431 EXPORT_SYMBOL_GPL(nvmet_req_complete);
432 
433 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
434 		u16 qid, u16 size)
435 {
436 	cq->qid = qid;
437 	cq->size = size;
438 
439 	ctrl->cqs[qid] = cq;
440 }
441 
442 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
443 		u16 qid, u16 size)
444 {
445 	sq->sqhd = 0;
446 	sq->qid = qid;
447 	sq->size = size;
448 
449 	ctrl->sqs[qid] = sq;
450 }
451 
452 static void nvmet_confirm_sq(struct percpu_ref *ref)
453 {
454 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
455 
456 	complete(&sq->confirm_done);
457 }
458 
459 void nvmet_sq_destroy(struct nvmet_sq *sq)
460 {
461 	/*
462 	 * If this is the admin queue, complete all AERs so that our
463 	 * queue doesn't have outstanding requests on it.
464 	 */
465 	if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
466 		nvmet_async_events_free(sq->ctrl);
467 	percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
468 	wait_for_completion(&sq->confirm_done);
469 	wait_for_completion(&sq->free_done);
470 	percpu_ref_exit(&sq->ref);
471 
472 	if (sq->ctrl) {
473 		nvmet_ctrl_put(sq->ctrl);
474 		sq->ctrl = NULL; /* allows reusing the queue later */
475 	}
476 }
477 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
478 
479 static void nvmet_sq_free(struct percpu_ref *ref)
480 {
481 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
482 
483 	complete(&sq->free_done);
484 }
485 
486 int nvmet_sq_init(struct nvmet_sq *sq)
487 {
488 	int ret;
489 
490 	ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
491 	if (ret) {
492 		pr_err("percpu_ref init failed!\n");
493 		return ret;
494 	}
495 	init_completion(&sq->free_done);
496 	init_completion(&sq->confirm_done);
497 
498 	return 0;
499 }
500 EXPORT_SYMBOL_GPL(nvmet_sq_init);
501 
502 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
503 		struct nvmet_sq *sq, struct nvmet_fabrics_ops *ops)
504 {
505 	u8 flags = req->cmd->common.flags;
506 	u16 status;
507 
508 	req->cq = cq;
509 	req->sq = sq;
510 	req->ops = ops;
511 	req->sg = NULL;
512 	req->sg_cnt = 0;
513 	req->transfer_len = 0;
514 	req->rsp->status = 0;
515 	req->ns = NULL;
516 
517 	/* no support for fused commands yet */
518 	if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
519 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
520 		goto fail;
521 	}
522 
523 	/*
524 	 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
525 	 * contains an address of a single contiguous physical buffer that is
526 	 * byte aligned.
527 	 */
528 	if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
529 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
530 		goto fail;
531 	}
532 
533 	if (unlikely(!req->sq->ctrl))
534 		/* will return an error for any Non-connect command: */
535 		status = nvmet_parse_connect_cmd(req);
536 	else if (likely(req->sq->qid != 0))
537 		status = nvmet_parse_io_cmd(req);
538 	else if (req->cmd->common.opcode == nvme_fabrics_command)
539 		status = nvmet_parse_fabrics_cmd(req);
540 	else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
541 		status = nvmet_parse_discovery_cmd(req);
542 	else
543 		status = nvmet_parse_admin_cmd(req);
544 
545 	if (status)
546 		goto fail;
547 
548 	if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
549 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
550 		goto fail;
551 	}
552 
553 	return true;
554 
555 fail:
556 	__nvmet_req_complete(req, status);
557 	return false;
558 }
559 EXPORT_SYMBOL_GPL(nvmet_req_init);
560 
561 void nvmet_req_uninit(struct nvmet_req *req)
562 {
563 	percpu_ref_put(&req->sq->ref);
564 	if (req->ns)
565 		nvmet_put_namespace(req->ns);
566 }
567 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
568 
569 void nvmet_req_execute(struct nvmet_req *req)
570 {
571 	if (unlikely(req->data_len != req->transfer_len))
572 		nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
573 	else
574 		req->execute(req);
575 }
576 EXPORT_SYMBOL_GPL(nvmet_req_execute);
577 
578 static inline bool nvmet_cc_en(u32 cc)
579 {
580 	return (cc >> NVME_CC_EN_SHIFT) & 0x1;
581 }
582 
583 static inline u8 nvmet_cc_css(u32 cc)
584 {
585 	return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
586 }
587 
588 static inline u8 nvmet_cc_mps(u32 cc)
589 {
590 	return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
591 }
592 
593 static inline u8 nvmet_cc_ams(u32 cc)
594 {
595 	return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
596 }
597 
598 static inline u8 nvmet_cc_shn(u32 cc)
599 {
600 	return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
601 }
602 
603 static inline u8 nvmet_cc_iosqes(u32 cc)
604 {
605 	return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
606 }
607 
608 static inline u8 nvmet_cc_iocqes(u32 cc)
609 {
610 	return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
611 }
612 
613 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
614 {
615 	lockdep_assert_held(&ctrl->lock);
616 
617 	if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
618 	    nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
619 	    nvmet_cc_mps(ctrl->cc) != 0 ||
620 	    nvmet_cc_ams(ctrl->cc) != 0 ||
621 	    nvmet_cc_css(ctrl->cc) != 0) {
622 		ctrl->csts = NVME_CSTS_CFS;
623 		return;
624 	}
625 
626 	ctrl->csts = NVME_CSTS_RDY;
627 }
628 
629 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
630 {
631 	lockdep_assert_held(&ctrl->lock);
632 
633 	/* XXX: tear down queues? */
634 	ctrl->csts &= ~NVME_CSTS_RDY;
635 	ctrl->cc = 0;
636 }
637 
638 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
639 {
640 	u32 old;
641 
642 	mutex_lock(&ctrl->lock);
643 	old = ctrl->cc;
644 	ctrl->cc = new;
645 
646 	if (nvmet_cc_en(new) && !nvmet_cc_en(old))
647 		nvmet_start_ctrl(ctrl);
648 	if (!nvmet_cc_en(new) && nvmet_cc_en(old))
649 		nvmet_clear_ctrl(ctrl);
650 	if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
651 		nvmet_clear_ctrl(ctrl);
652 		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
653 	}
654 	if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
655 		ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
656 	mutex_unlock(&ctrl->lock);
657 }
658 
659 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
660 {
661 	/* command sets supported: NVMe command set: */
662 	ctrl->cap = (1ULL << 37);
663 	/* CC.EN timeout in 500msec units: */
664 	ctrl->cap |= (15ULL << 24);
665 	/* maximum queue entries supported: */
666 	ctrl->cap |= NVMET_QUEUE_SIZE - 1;
667 }
668 
669 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
670 		struct nvmet_req *req, struct nvmet_ctrl **ret)
671 {
672 	struct nvmet_subsys *subsys;
673 	struct nvmet_ctrl *ctrl;
674 	u16 status = 0;
675 
676 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
677 	if (!subsys) {
678 		pr_warn("connect request for invalid subsystem %s!\n",
679 			subsysnqn);
680 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
681 		return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
682 	}
683 
684 	mutex_lock(&subsys->lock);
685 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
686 		if (ctrl->cntlid == cntlid) {
687 			if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
688 				pr_warn("hostnqn mismatch.\n");
689 				continue;
690 			}
691 			if (!kref_get_unless_zero(&ctrl->ref))
692 				continue;
693 
694 			*ret = ctrl;
695 			goto out;
696 		}
697 	}
698 
699 	pr_warn("could not find controller %d for subsys %s / host %s\n",
700 		cntlid, subsysnqn, hostnqn);
701 	req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
702 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
703 
704 out:
705 	mutex_unlock(&subsys->lock);
706 	nvmet_subsys_put(subsys);
707 	return status;
708 }
709 
710 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
711 {
712 	if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
713 		pr_err("got io cmd %d while CC.EN == 0 on qid = %d\n",
714 		       cmd->common.opcode, req->sq->qid);
715 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
716 	}
717 
718 	if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
719 		pr_err("got io cmd %d while CSTS.RDY == 0 on qid = %d\n",
720 		       cmd->common.opcode, req->sq->qid);
721 		req->ns = NULL;
722 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
723 	}
724 	return 0;
725 }
726 
727 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
728 		const char *hostnqn)
729 {
730 	struct nvmet_host_link *p;
731 
732 	if (subsys->allow_any_host)
733 		return true;
734 
735 	list_for_each_entry(p, &subsys->hosts, entry) {
736 		if (!strcmp(nvmet_host_name(p->host), hostnqn))
737 			return true;
738 	}
739 
740 	return false;
741 }
742 
743 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
744 		const char *hostnqn)
745 {
746 	struct nvmet_subsys_link *s;
747 
748 	list_for_each_entry(s, &req->port->subsystems, entry) {
749 		if (__nvmet_host_allowed(s->subsys, hostnqn))
750 			return true;
751 	}
752 
753 	return false;
754 }
755 
756 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
757 		const char *hostnqn)
758 {
759 	lockdep_assert_held(&nvmet_config_sem);
760 
761 	if (subsys->type == NVME_NQN_DISC)
762 		return nvmet_host_discovery_allowed(req, hostnqn);
763 	else
764 		return __nvmet_host_allowed(subsys, hostnqn);
765 }
766 
767 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
768 		struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
769 {
770 	struct nvmet_subsys *subsys;
771 	struct nvmet_ctrl *ctrl;
772 	int ret;
773 	u16 status;
774 
775 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
776 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
777 	if (!subsys) {
778 		pr_warn("connect request for invalid subsystem %s!\n",
779 			subsysnqn);
780 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
781 		goto out;
782 	}
783 
784 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
785 	down_read(&nvmet_config_sem);
786 	if (!nvmet_host_allowed(req, subsys, hostnqn)) {
787 		pr_info("connect by host %s for subsystem %s not allowed\n",
788 			hostnqn, subsysnqn);
789 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
790 		up_read(&nvmet_config_sem);
791 		status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
792 		goto out_put_subsystem;
793 	}
794 	up_read(&nvmet_config_sem);
795 
796 	status = NVME_SC_INTERNAL;
797 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
798 	if (!ctrl)
799 		goto out_put_subsystem;
800 	mutex_init(&ctrl->lock);
801 
802 	nvmet_init_cap(ctrl);
803 
804 	INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
805 	INIT_LIST_HEAD(&ctrl->async_events);
806 
807 	memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
808 	memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
809 
810 	kref_init(&ctrl->ref);
811 	ctrl->subsys = subsys;
812 
813 	ctrl->cqs = kcalloc(subsys->max_qid + 1,
814 			sizeof(struct nvmet_cq *),
815 			GFP_KERNEL);
816 	if (!ctrl->cqs)
817 		goto out_free_ctrl;
818 
819 	ctrl->sqs = kcalloc(subsys->max_qid + 1,
820 			sizeof(struct nvmet_sq *),
821 			GFP_KERNEL);
822 	if (!ctrl->sqs)
823 		goto out_free_cqs;
824 
825 	ret = ida_simple_get(&cntlid_ida,
826 			     NVME_CNTLID_MIN, NVME_CNTLID_MAX,
827 			     GFP_KERNEL);
828 	if (ret < 0) {
829 		status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
830 		goto out_free_sqs;
831 	}
832 	ctrl->cntlid = ret;
833 
834 	ctrl->ops = req->ops;
835 	if (ctrl->subsys->type == NVME_NQN_DISC) {
836 		/* Don't accept keep-alive timeout for discovery controllers */
837 		if (kato) {
838 			status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
839 			goto out_remove_ida;
840 		}
841 
842 		/*
843 		 * Discovery controllers use some arbitrary high value in order
844 		 * to cleanup stale discovery sessions
845 		 *
846 		 * From the latest base diff RC:
847 		 * "The Keep Alive command is not supported by
848 		 * Discovery controllers. A transport may specify a
849 		 * fixed Discovery controller activity timeout value
850 		 * (e.g., 2 minutes).  If no commands are received
851 		 * by a Discovery controller within that time
852 		 * period, the controller may perform the
853 		 * actions for Keep Alive Timer expiration".
854 		 */
855 		ctrl->kato = NVMET_DISC_KATO;
856 	} else {
857 		/* keep-alive timeout in seconds */
858 		ctrl->kato = DIV_ROUND_UP(kato, 1000);
859 	}
860 	nvmet_start_keep_alive_timer(ctrl);
861 
862 	mutex_lock(&subsys->lock);
863 	list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
864 	mutex_unlock(&subsys->lock);
865 
866 	*ctrlp = ctrl;
867 	return 0;
868 
869 out_remove_ida:
870 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
871 out_free_sqs:
872 	kfree(ctrl->sqs);
873 out_free_cqs:
874 	kfree(ctrl->cqs);
875 out_free_ctrl:
876 	kfree(ctrl);
877 out_put_subsystem:
878 	nvmet_subsys_put(subsys);
879 out:
880 	return status;
881 }
882 
883 static void nvmet_ctrl_free(struct kref *ref)
884 {
885 	struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
886 	struct nvmet_subsys *subsys = ctrl->subsys;
887 
888 	mutex_lock(&subsys->lock);
889 	list_del(&ctrl->subsys_entry);
890 	mutex_unlock(&subsys->lock);
891 
892 	nvmet_stop_keep_alive_timer(ctrl);
893 
894 	flush_work(&ctrl->async_event_work);
895 	cancel_work_sync(&ctrl->fatal_err_work);
896 
897 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
898 
899 	kfree(ctrl->sqs);
900 	kfree(ctrl->cqs);
901 	kfree(ctrl);
902 
903 	nvmet_subsys_put(subsys);
904 }
905 
906 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
907 {
908 	kref_put(&ctrl->ref, nvmet_ctrl_free);
909 }
910 
911 static void nvmet_fatal_error_handler(struct work_struct *work)
912 {
913 	struct nvmet_ctrl *ctrl =
914 			container_of(work, struct nvmet_ctrl, fatal_err_work);
915 
916 	pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
917 	ctrl->ops->delete_ctrl(ctrl);
918 }
919 
920 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
921 {
922 	mutex_lock(&ctrl->lock);
923 	if (!(ctrl->csts & NVME_CSTS_CFS)) {
924 		ctrl->csts |= NVME_CSTS_CFS;
925 		INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
926 		schedule_work(&ctrl->fatal_err_work);
927 	}
928 	mutex_unlock(&ctrl->lock);
929 }
930 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
931 
932 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
933 		const char *subsysnqn)
934 {
935 	struct nvmet_subsys_link *p;
936 
937 	if (!port)
938 		return NULL;
939 
940 	if (!strncmp(NVME_DISC_SUBSYS_NAME, subsysnqn,
941 			NVMF_NQN_SIZE)) {
942 		if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
943 			return NULL;
944 		return nvmet_disc_subsys;
945 	}
946 
947 	down_read(&nvmet_config_sem);
948 	list_for_each_entry(p, &port->subsystems, entry) {
949 		if (!strncmp(p->subsys->subsysnqn, subsysnqn,
950 				NVMF_NQN_SIZE)) {
951 			if (!kref_get_unless_zero(&p->subsys->ref))
952 				break;
953 			up_read(&nvmet_config_sem);
954 			return p->subsys;
955 		}
956 	}
957 	up_read(&nvmet_config_sem);
958 	return NULL;
959 }
960 
961 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
962 		enum nvme_subsys_type type)
963 {
964 	struct nvmet_subsys *subsys;
965 
966 	subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
967 	if (!subsys)
968 		return NULL;
969 
970 	subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
971 	/* generate a random serial number as our controllers are ephemeral: */
972 	get_random_bytes(&subsys->serial, sizeof(subsys->serial));
973 
974 	switch (type) {
975 	case NVME_NQN_NVME:
976 		subsys->max_qid = NVMET_NR_QUEUES;
977 		break;
978 	case NVME_NQN_DISC:
979 		subsys->max_qid = 0;
980 		break;
981 	default:
982 		pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
983 		kfree(subsys);
984 		return NULL;
985 	}
986 	subsys->type = type;
987 	subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
988 			GFP_KERNEL);
989 	if (!subsys->subsysnqn) {
990 		kfree(subsys);
991 		return NULL;
992 	}
993 
994 	kref_init(&subsys->ref);
995 
996 	mutex_init(&subsys->lock);
997 	INIT_LIST_HEAD(&subsys->namespaces);
998 	INIT_LIST_HEAD(&subsys->ctrls);
999 	INIT_LIST_HEAD(&subsys->hosts);
1000 
1001 	return subsys;
1002 }
1003 
1004 static void nvmet_subsys_free(struct kref *ref)
1005 {
1006 	struct nvmet_subsys *subsys =
1007 		container_of(ref, struct nvmet_subsys, ref);
1008 
1009 	WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1010 
1011 	kfree(subsys->subsysnqn);
1012 	kfree(subsys);
1013 }
1014 
1015 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1016 {
1017 	struct nvmet_ctrl *ctrl;
1018 
1019 	mutex_lock(&subsys->lock);
1020 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1021 		ctrl->ops->delete_ctrl(ctrl);
1022 	mutex_unlock(&subsys->lock);
1023 }
1024 
1025 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1026 {
1027 	kref_put(&subsys->ref, nvmet_subsys_free);
1028 }
1029 
1030 static int __init nvmet_init(void)
1031 {
1032 	int error;
1033 
1034 	error = nvmet_init_discovery();
1035 	if (error)
1036 		goto out;
1037 
1038 	error = nvmet_init_configfs();
1039 	if (error)
1040 		goto out_exit_discovery;
1041 	return 0;
1042 
1043 out_exit_discovery:
1044 	nvmet_exit_discovery();
1045 out:
1046 	return error;
1047 }
1048 
1049 static void __exit nvmet_exit(void)
1050 {
1051 	nvmet_exit_configfs();
1052 	nvmet_exit_discovery();
1053 	ida_destroy(&cntlid_ida);
1054 
1055 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1056 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1057 }
1058 
1059 module_init(nvmet_init);
1060 module_exit(nvmet_exit);
1061 
1062 MODULE_LICENSE("GPL v2");
1063