xref: /openbmc/linux/drivers/nvme/target/core.c (revision 9be08a27)
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 struct workqueue_struct *buffered_io_wq;
22 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
23 static DEFINE_IDA(cntlid_ida);
24 
25 /*
26  * This read/write semaphore is used to synchronize access to configuration
27  * information on a target system that will result in discovery log page
28  * information change for at least one host.
29  * The full list of resources to protected by this semaphore is:
30  *
31  *  - subsystems list
32  *  - per-subsystem allowed hosts list
33  *  - allow_any_host subsystem attribute
34  *  - nvmet_genctr
35  *  - the nvmet_transports array
36  *
37  * When updating any of those lists/structures write lock should be obtained,
38  * while when reading (popolating discovery log page or checking host-subsystem
39  * link) read lock is obtained to allow concurrent reads.
40  */
41 DECLARE_RWSEM(nvmet_config_sem);
42 
43 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
44 u64 nvmet_ana_chgcnt;
45 DECLARE_RWSEM(nvmet_ana_sem);
46 
47 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
48 		const char *subsysnqn);
49 
50 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
51 		size_t len)
52 {
53 	if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
54 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
55 	return 0;
56 }
57 
58 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
59 {
60 	if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len)
61 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
62 	return 0;
63 }
64 
65 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
66 {
67 	if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len)
68 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
69 	return 0;
70 }
71 
72 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
73 {
74 	struct nvmet_ns *ns;
75 
76 	if (list_empty(&subsys->namespaces))
77 		return 0;
78 
79 	ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
80 	return ns->nsid;
81 }
82 
83 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
84 {
85 	return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
86 }
87 
88 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
89 {
90 	struct nvmet_req *req;
91 
92 	while (1) {
93 		mutex_lock(&ctrl->lock);
94 		if (!ctrl->nr_async_event_cmds) {
95 			mutex_unlock(&ctrl->lock);
96 			return;
97 		}
98 
99 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
100 		mutex_unlock(&ctrl->lock);
101 		nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
102 	}
103 }
104 
105 static void nvmet_async_event_work(struct work_struct *work)
106 {
107 	struct nvmet_ctrl *ctrl =
108 		container_of(work, struct nvmet_ctrl, async_event_work);
109 	struct nvmet_async_event *aen;
110 	struct nvmet_req *req;
111 
112 	while (1) {
113 		mutex_lock(&ctrl->lock);
114 		aen = list_first_entry_or_null(&ctrl->async_events,
115 				struct nvmet_async_event, entry);
116 		if (!aen || !ctrl->nr_async_event_cmds) {
117 			mutex_unlock(&ctrl->lock);
118 			return;
119 		}
120 
121 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
122 		nvmet_set_result(req, nvmet_async_event_result(aen));
123 
124 		list_del(&aen->entry);
125 		kfree(aen);
126 
127 		mutex_unlock(&ctrl->lock);
128 		nvmet_req_complete(req, 0);
129 	}
130 }
131 
132 static void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
133 		u8 event_info, u8 log_page)
134 {
135 	struct nvmet_async_event *aen;
136 
137 	aen = kmalloc(sizeof(*aen), GFP_KERNEL);
138 	if (!aen)
139 		return;
140 
141 	aen->event_type = event_type;
142 	aen->event_info = event_info;
143 	aen->log_page = log_page;
144 
145 	mutex_lock(&ctrl->lock);
146 	list_add_tail(&aen->entry, &ctrl->async_events);
147 	mutex_unlock(&ctrl->lock);
148 
149 	schedule_work(&ctrl->async_event_work);
150 }
151 
152 static bool nvmet_aen_disabled(struct nvmet_ctrl *ctrl, u32 aen)
153 {
154 	if (!(READ_ONCE(ctrl->aen_enabled) & aen))
155 		return true;
156 	return test_and_set_bit(aen, &ctrl->aen_masked);
157 }
158 
159 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
160 {
161 	u32 i;
162 
163 	mutex_lock(&ctrl->lock);
164 	if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
165 		goto out_unlock;
166 
167 	for (i = 0; i < ctrl->nr_changed_ns; i++) {
168 		if (ctrl->changed_ns_list[i] == nsid)
169 			goto out_unlock;
170 	}
171 
172 	if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
173 		ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
174 		ctrl->nr_changed_ns = U32_MAX;
175 		goto out_unlock;
176 	}
177 
178 	ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
179 out_unlock:
180 	mutex_unlock(&ctrl->lock);
181 }
182 
183 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
184 {
185 	struct nvmet_ctrl *ctrl;
186 
187 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
188 		nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
189 		if (nvmet_aen_disabled(ctrl, NVME_AEN_CFG_NS_ATTR))
190 			continue;
191 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
192 				NVME_AER_NOTICE_NS_CHANGED,
193 				NVME_LOG_CHANGED_NS);
194 	}
195 }
196 
197 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
198 		struct nvmet_port *port)
199 {
200 	struct nvmet_ctrl *ctrl;
201 
202 	mutex_lock(&subsys->lock);
203 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
204 		if (port && ctrl->port != port)
205 			continue;
206 		if (nvmet_aen_disabled(ctrl, NVME_AEN_CFG_ANA_CHANGE))
207 			continue;
208 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
209 				NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
210 	}
211 	mutex_unlock(&subsys->lock);
212 }
213 
214 void nvmet_port_send_ana_event(struct nvmet_port *port)
215 {
216 	struct nvmet_subsys_link *p;
217 
218 	down_read(&nvmet_config_sem);
219 	list_for_each_entry(p, &port->subsystems, entry)
220 		nvmet_send_ana_event(p->subsys, port);
221 	up_read(&nvmet_config_sem);
222 }
223 
224 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
225 {
226 	int ret = 0;
227 
228 	down_write(&nvmet_config_sem);
229 	if (nvmet_transports[ops->type])
230 		ret = -EINVAL;
231 	else
232 		nvmet_transports[ops->type] = ops;
233 	up_write(&nvmet_config_sem);
234 
235 	return ret;
236 }
237 EXPORT_SYMBOL_GPL(nvmet_register_transport);
238 
239 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
240 {
241 	down_write(&nvmet_config_sem);
242 	nvmet_transports[ops->type] = NULL;
243 	up_write(&nvmet_config_sem);
244 }
245 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
246 
247 int nvmet_enable_port(struct nvmet_port *port)
248 {
249 	const struct nvmet_fabrics_ops *ops;
250 	int ret;
251 
252 	lockdep_assert_held(&nvmet_config_sem);
253 
254 	ops = nvmet_transports[port->disc_addr.trtype];
255 	if (!ops) {
256 		up_write(&nvmet_config_sem);
257 		request_module("nvmet-transport-%d", port->disc_addr.trtype);
258 		down_write(&nvmet_config_sem);
259 		ops = nvmet_transports[port->disc_addr.trtype];
260 		if (!ops) {
261 			pr_err("transport type %d not supported\n",
262 				port->disc_addr.trtype);
263 			return -EINVAL;
264 		}
265 	}
266 
267 	if (!try_module_get(ops->owner))
268 		return -EINVAL;
269 
270 	ret = ops->add_port(port);
271 	if (ret) {
272 		module_put(ops->owner);
273 		return ret;
274 	}
275 
276 	/* If the transport didn't set inline_data_size, then disable it. */
277 	if (port->inline_data_size < 0)
278 		port->inline_data_size = 0;
279 
280 	port->enabled = true;
281 	return 0;
282 }
283 
284 void nvmet_disable_port(struct nvmet_port *port)
285 {
286 	const struct nvmet_fabrics_ops *ops;
287 
288 	lockdep_assert_held(&nvmet_config_sem);
289 
290 	port->enabled = false;
291 
292 	ops = nvmet_transports[port->disc_addr.trtype];
293 	ops->remove_port(port);
294 	module_put(ops->owner);
295 }
296 
297 static void nvmet_keep_alive_timer(struct work_struct *work)
298 {
299 	struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
300 			struct nvmet_ctrl, ka_work);
301 
302 	pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
303 		ctrl->cntlid, ctrl->kato);
304 
305 	nvmet_ctrl_fatal_error(ctrl);
306 }
307 
308 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
309 {
310 	pr_debug("ctrl %d start keep-alive timer for %d secs\n",
311 		ctrl->cntlid, ctrl->kato);
312 
313 	INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
314 	schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
315 }
316 
317 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
318 {
319 	pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
320 
321 	cancel_delayed_work_sync(&ctrl->ka_work);
322 }
323 
324 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
325 		__le32 nsid)
326 {
327 	struct nvmet_ns *ns;
328 
329 	list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
330 		if (ns->nsid == le32_to_cpu(nsid))
331 			return ns;
332 	}
333 
334 	return NULL;
335 }
336 
337 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
338 {
339 	struct nvmet_ns *ns;
340 
341 	rcu_read_lock();
342 	ns = __nvmet_find_namespace(ctrl, nsid);
343 	if (ns)
344 		percpu_ref_get(&ns->ref);
345 	rcu_read_unlock();
346 
347 	return ns;
348 }
349 
350 static void nvmet_destroy_namespace(struct percpu_ref *ref)
351 {
352 	struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
353 
354 	complete(&ns->disable_done);
355 }
356 
357 void nvmet_put_namespace(struct nvmet_ns *ns)
358 {
359 	percpu_ref_put(&ns->ref);
360 }
361 
362 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
363 {
364 	nvmet_bdev_ns_disable(ns);
365 	nvmet_file_ns_disable(ns);
366 }
367 
368 int nvmet_ns_enable(struct nvmet_ns *ns)
369 {
370 	struct nvmet_subsys *subsys = ns->subsys;
371 	int ret;
372 
373 	mutex_lock(&subsys->lock);
374 	ret = -EMFILE;
375 	if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
376 		goto out_unlock;
377 	ret = 0;
378 	if (ns->enabled)
379 		goto out_unlock;
380 
381 	ret = nvmet_bdev_ns_enable(ns);
382 	if (ret == -ENOTBLK)
383 		ret = nvmet_file_ns_enable(ns);
384 	if (ret)
385 		goto out_unlock;
386 
387 	ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
388 				0, GFP_KERNEL);
389 	if (ret)
390 		goto out_dev_put;
391 
392 	if (ns->nsid > subsys->max_nsid)
393 		subsys->max_nsid = ns->nsid;
394 
395 	/*
396 	 * The namespaces list needs to be sorted to simplify the implementation
397 	 * of the Identify Namepace List subcommand.
398 	 */
399 	if (list_empty(&subsys->namespaces)) {
400 		list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
401 	} else {
402 		struct nvmet_ns *old;
403 
404 		list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
405 			BUG_ON(ns->nsid == old->nsid);
406 			if (ns->nsid < old->nsid)
407 				break;
408 		}
409 
410 		list_add_tail_rcu(&ns->dev_link, &old->dev_link);
411 	}
412 	subsys->nr_namespaces++;
413 
414 	nvmet_ns_changed(subsys, ns->nsid);
415 	ns->enabled = true;
416 	ret = 0;
417 out_unlock:
418 	mutex_unlock(&subsys->lock);
419 	return ret;
420 out_dev_put:
421 	nvmet_ns_dev_disable(ns);
422 	goto out_unlock;
423 }
424 
425 void nvmet_ns_disable(struct nvmet_ns *ns)
426 {
427 	struct nvmet_subsys *subsys = ns->subsys;
428 
429 	mutex_lock(&subsys->lock);
430 	if (!ns->enabled)
431 		goto out_unlock;
432 
433 	ns->enabled = false;
434 	list_del_rcu(&ns->dev_link);
435 	if (ns->nsid == subsys->max_nsid)
436 		subsys->max_nsid = nvmet_max_nsid(subsys);
437 	mutex_unlock(&subsys->lock);
438 
439 	/*
440 	 * Now that we removed the namespaces from the lookup list, we
441 	 * can kill the per_cpu ref and wait for any remaining references
442 	 * to be dropped, as well as a RCU grace period for anyone only
443 	 * using the namepace under rcu_read_lock().  Note that we can't
444 	 * use call_rcu here as we need to ensure the namespaces have
445 	 * been fully destroyed before unloading the module.
446 	 */
447 	percpu_ref_kill(&ns->ref);
448 	synchronize_rcu();
449 	wait_for_completion(&ns->disable_done);
450 	percpu_ref_exit(&ns->ref);
451 
452 	mutex_lock(&subsys->lock);
453 	subsys->nr_namespaces--;
454 	nvmet_ns_changed(subsys, ns->nsid);
455 	nvmet_ns_dev_disable(ns);
456 out_unlock:
457 	mutex_unlock(&subsys->lock);
458 }
459 
460 void nvmet_ns_free(struct nvmet_ns *ns)
461 {
462 	nvmet_ns_disable(ns);
463 
464 	down_write(&nvmet_ana_sem);
465 	nvmet_ana_group_enabled[ns->anagrpid]--;
466 	up_write(&nvmet_ana_sem);
467 
468 	kfree(ns->device_path);
469 	kfree(ns);
470 }
471 
472 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
473 {
474 	struct nvmet_ns *ns;
475 
476 	ns = kzalloc(sizeof(*ns), GFP_KERNEL);
477 	if (!ns)
478 		return NULL;
479 
480 	INIT_LIST_HEAD(&ns->dev_link);
481 	init_completion(&ns->disable_done);
482 
483 	ns->nsid = nsid;
484 	ns->subsys = subsys;
485 
486 	down_write(&nvmet_ana_sem);
487 	ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
488 	nvmet_ana_group_enabled[ns->anagrpid]++;
489 	up_write(&nvmet_ana_sem);
490 
491 	uuid_gen(&ns->uuid);
492 	ns->buffered_io = false;
493 
494 	return ns;
495 }
496 
497 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
498 {
499 	u32 old_sqhd, new_sqhd;
500 	u16 sqhd;
501 
502 	if (status)
503 		nvmet_set_status(req, status);
504 
505 	if (req->sq->size) {
506 		do {
507 			old_sqhd = req->sq->sqhd;
508 			new_sqhd = (old_sqhd + 1) % req->sq->size;
509 		} while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
510 					old_sqhd);
511 	}
512 	sqhd = req->sq->sqhd & 0x0000FFFF;
513 	req->rsp->sq_head = cpu_to_le16(sqhd);
514 	req->rsp->sq_id = cpu_to_le16(req->sq->qid);
515 	req->rsp->command_id = req->cmd->common.command_id;
516 
517 	if (req->ns)
518 		nvmet_put_namespace(req->ns);
519 	req->ops->queue_response(req);
520 }
521 
522 void nvmet_req_complete(struct nvmet_req *req, u16 status)
523 {
524 	__nvmet_req_complete(req, status);
525 	percpu_ref_put(&req->sq->ref);
526 }
527 EXPORT_SYMBOL_GPL(nvmet_req_complete);
528 
529 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
530 		u16 qid, u16 size)
531 {
532 	cq->qid = qid;
533 	cq->size = size;
534 
535 	ctrl->cqs[qid] = cq;
536 }
537 
538 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
539 		u16 qid, u16 size)
540 {
541 	sq->sqhd = 0;
542 	sq->qid = qid;
543 	sq->size = size;
544 
545 	ctrl->sqs[qid] = sq;
546 }
547 
548 static void nvmet_confirm_sq(struct percpu_ref *ref)
549 {
550 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
551 
552 	complete(&sq->confirm_done);
553 }
554 
555 void nvmet_sq_destroy(struct nvmet_sq *sq)
556 {
557 	/*
558 	 * If this is the admin queue, complete all AERs so that our
559 	 * queue doesn't have outstanding requests on it.
560 	 */
561 	if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
562 		nvmet_async_events_free(sq->ctrl);
563 	percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
564 	wait_for_completion(&sq->confirm_done);
565 	wait_for_completion(&sq->free_done);
566 	percpu_ref_exit(&sq->ref);
567 
568 	if (sq->ctrl) {
569 		nvmet_ctrl_put(sq->ctrl);
570 		sq->ctrl = NULL; /* allows reusing the queue later */
571 	}
572 }
573 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
574 
575 static void nvmet_sq_free(struct percpu_ref *ref)
576 {
577 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
578 
579 	complete(&sq->free_done);
580 }
581 
582 int nvmet_sq_init(struct nvmet_sq *sq)
583 {
584 	int ret;
585 
586 	ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
587 	if (ret) {
588 		pr_err("percpu_ref init failed!\n");
589 		return ret;
590 	}
591 	init_completion(&sq->free_done);
592 	init_completion(&sq->confirm_done);
593 
594 	return 0;
595 }
596 EXPORT_SYMBOL_GPL(nvmet_sq_init);
597 
598 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
599 		struct nvmet_ns *ns)
600 {
601 	enum nvme_ana_state state = port->ana_state[ns->anagrpid];
602 
603 	if (unlikely(state == NVME_ANA_INACCESSIBLE))
604 		return NVME_SC_ANA_INACCESSIBLE;
605 	if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
606 		return NVME_SC_ANA_PERSISTENT_LOSS;
607 	if (unlikely(state == NVME_ANA_CHANGE))
608 		return NVME_SC_ANA_TRANSITION;
609 	return 0;
610 }
611 
612 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
613 {
614 	if (unlikely(req->ns->readonly)) {
615 		switch (req->cmd->common.opcode) {
616 		case nvme_cmd_read:
617 		case nvme_cmd_flush:
618 			break;
619 		default:
620 			return NVME_SC_NS_WRITE_PROTECTED;
621 		}
622 	}
623 
624 	return 0;
625 }
626 
627 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
628 {
629 	struct nvme_command *cmd = req->cmd;
630 	u16 ret;
631 
632 	ret = nvmet_check_ctrl_status(req, cmd);
633 	if (unlikely(ret))
634 		return ret;
635 
636 	req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
637 	if (unlikely(!req->ns))
638 		return NVME_SC_INVALID_NS | NVME_SC_DNR;
639 	ret = nvmet_check_ana_state(req->port, req->ns);
640 	if (unlikely(ret))
641 		return ret;
642 	ret = nvmet_io_cmd_check_access(req);
643 	if (unlikely(ret))
644 		return ret;
645 
646 	if (req->ns->file)
647 		return nvmet_file_parse_io_cmd(req);
648 	else
649 		return nvmet_bdev_parse_io_cmd(req);
650 }
651 
652 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
653 		struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
654 {
655 	u8 flags = req->cmd->common.flags;
656 	u16 status;
657 
658 	req->cq = cq;
659 	req->sq = sq;
660 	req->ops = ops;
661 	req->sg = NULL;
662 	req->sg_cnt = 0;
663 	req->transfer_len = 0;
664 	req->rsp->status = 0;
665 	req->ns = NULL;
666 
667 	/* no support for fused commands yet */
668 	if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
669 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
670 		goto fail;
671 	}
672 
673 	/*
674 	 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
675 	 * contains an address of a single contiguous physical buffer that is
676 	 * byte aligned.
677 	 */
678 	if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
679 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
680 		goto fail;
681 	}
682 
683 	if (unlikely(!req->sq->ctrl))
684 		/* will return an error for any Non-connect command: */
685 		status = nvmet_parse_connect_cmd(req);
686 	else if (likely(req->sq->qid != 0))
687 		status = nvmet_parse_io_cmd(req);
688 	else if (req->cmd->common.opcode == nvme_fabrics_command)
689 		status = nvmet_parse_fabrics_cmd(req);
690 	else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
691 		status = nvmet_parse_discovery_cmd(req);
692 	else
693 		status = nvmet_parse_admin_cmd(req);
694 
695 	if (status)
696 		goto fail;
697 
698 	if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
699 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
700 		goto fail;
701 	}
702 
703 	return true;
704 
705 fail:
706 	__nvmet_req_complete(req, status);
707 	return false;
708 }
709 EXPORT_SYMBOL_GPL(nvmet_req_init);
710 
711 void nvmet_req_uninit(struct nvmet_req *req)
712 {
713 	percpu_ref_put(&req->sq->ref);
714 	if (req->ns)
715 		nvmet_put_namespace(req->ns);
716 }
717 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
718 
719 void nvmet_req_execute(struct nvmet_req *req)
720 {
721 	if (unlikely(req->data_len != req->transfer_len))
722 		nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
723 	else
724 		req->execute(req);
725 }
726 EXPORT_SYMBOL_GPL(nvmet_req_execute);
727 
728 static inline bool nvmet_cc_en(u32 cc)
729 {
730 	return (cc >> NVME_CC_EN_SHIFT) & 0x1;
731 }
732 
733 static inline u8 nvmet_cc_css(u32 cc)
734 {
735 	return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
736 }
737 
738 static inline u8 nvmet_cc_mps(u32 cc)
739 {
740 	return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
741 }
742 
743 static inline u8 nvmet_cc_ams(u32 cc)
744 {
745 	return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
746 }
747 
748 static inline u8 nvmet_cc_shn(u32 cc)
749 {
750 	return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
751 }
752 
753 static inline u8 nvmet_cc_iosqes(u32 cc)
754 {
755 	return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
756 }
757 
758 static inline u8 nvmet_cc_iocqes(u32 cc)
759 {
760 	return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
761 }
762 
763 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
764 {
765 	lockdep_assert_held(&ctrl->lock);
766 
767 	if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
768 	    nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
769 	    nvmet_cc_mps(ctrl->cc) != 0 ||
770 	    nvmet_cc_ams(ctrl->cc) != 0 ||
771 	    nvmet_cc_css(ctrl->cc) != 0) {
772 		ctrl->csts = NVME_CSTS_CFS;
773 		return;
774 	}
775 
776 	ctrl->csts = NVME_CSTS_RDY;
777 
778 	/*
779 	 * Controllers that are not yet enabled should not really enforce the
780 	 * keep alive timeout, but we still want to track a timeout and cleanup
781 	 * in case a host died before it enabled the controller.  Hence, simply
782 	 * reset the keep alive timer when the controller is enabled.
783 	 */
784 	mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
785 }
786 
787 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
788 {
789 	lockdep_assert_held(&ctrl->lock);
790 
791 	/* XXX: tear down queues? */
792 	ctrl->csts &= ~NVME_CSTS_RDY;
793 	ctrl->cc = 0;
794 }
795 
796 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
797 {
798 	u32 old;
799 
800 	mutex_lock(&ctrl->lock);
801 	old = ctrl->cc;
802 	ctrl->cc = new;
803 
804 	if (nvmet_cc_en(new) && !nvmet_cc_en(old))
805 		nvmet_start_ctrl(ctrl);
806 	if (!nvmet_cc_en(new) && nvmet_cc_en(old))
807 		nvmet_clear_ctrl(ctrl);
808 	if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
809 		nvmet_clear_ctrl(ctrl);
810 		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
811 	}
812 	if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
813 		ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
814 	mutex_unlock(&ctrl->lock);
815 }
816 
817 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
818 {
819 	/* command sets supported: NVMe command set: */
820 	ctrl->cap = (1ULL << 37);
821 	/* CC.EN timeout in 500msec units: */
822 	ctrl->cap |= (15ULL << 24);
823 	/* maximum queue entries supported: */
824 	ctrl->cap |= NVMET_QUEUE_SIZE - 1;
825 }
826 
827 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
828 		struct nvmet_req *req, struct nvmet_ctrl **ret)
829 {
830 	struct nvmet_subsys *subsys;
831 	struct nvmet_ctrl *ctrl;
832 	u16 status = 0;
833 
834 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
835 	if (!subsys) {
836 		pr_warn("connect request for invalid subsystem %s!\n",
837 			subsysnqn);
838 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
839 		return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
840 	}
841 
842 	mutex_lock(&subsys->lock);
843 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
844 		if (ctrl->cntlid == cntlid) {
845 			if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
846 				pr_warn("hostnqn mismatch.\n");
847 				continue;
848 			}
849 			if (!kref_get_unless_zero(&ctrl->ref))
850 				continue;
851 
852 			*ret = ctrl;
853 			goto out;
854 		}
855 	}
856 
857 	pr_warn("could not find controller %d for subsys %s / host %s\n",
858 		cntlid, subsysnqn, hostnqn);
859 	req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
860 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
861 
862 out:
863 	mutex_unlock(&subsys->lock);
864 	nvmet_subsys_put(subsys);
865 	return status;
866 }
867 
868 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
869 {
870 	if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
871 		pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
872 		       cmd->common.opcode, req->sq->qid);
873 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
874 	}
875 
876 	if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
877 		pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
878 		       cmd->common.opcode, req->sq->qid);
879 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
880 	}
881 	return 0;
882 }
883 
884 static bool __nvmet_host_allowed(struct nvmet_subsys *subsys,
885 		const char *hostnqn)
886 {
887 	struct nvmet_host_link *p;
888 
889 	if (subsys->allow_any_host)
890 		return true;
891 
892 	list_for_each_entry(p, &subsys->hosts, entry) {
893 		if (!strcmp(nvmet_host_name(p->host), hostnqn))
894 			return true;
895 	}
896 
897 	return false;
898 }
899 
900 static bool nvmet_host_discovery_allowed(struct nvmet_req *req,
901 		const char *hostnqn)
902 {
903 	struct nvmet_subsys_link *s;
904 
905 	list_for_each_entry(s, &req->port->subsystems, entry) {
906 		if (__nvmet_host_allowed(s->subsys, hostnqn))
907 			return true;
908 	}
909 
910 	return false;
911 }
912 
913 bool nvmet_host_allowed(struct nvmet_req *req, struct nvmet_subsys *subsys,
914 		const char *hostnqn)
915 {
916 	lockdep_assert_held(&nvmet_config_sem);
917 
918 	if (subsys->type == NVME_NQN_DISC)
919 		return nvmet_host_discovery_allowed(req, hostnqn);
920 	else
921 		return __nvmet_host_allowed(subsys, hostnqn);
922 }
923 
924 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
925 		struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
926 {
927 	struct nvmet_subsys *subsys;
928 	struct nvmet_ctrl *ctrl;
929 	int ret;
930 	u16 status;
931 
932 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
933 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
934 	if (!subsys) {
935 		pr_warn("connect request for invalid subsystem %s!\n",
936 			subsysnqn);
937 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
938 		goto out;
939 	}
940 
941 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
942 	down_read(&nvmet_config_sem);
943 	if (!nvmet_host_allowed(req, subsys, hostnqn)) {
944 		pr_info("connect by host %s for subsystem %s not allowed\n",
945 			hostnqn, subsysnqn);
946 		req->rsp->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
947 		up_read(&nvmet_config_sem);
948 		status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
949 		goto out_put_subsystem;
950 	}
951 	up_read(&nvmet_config_sem);
952 
953 	status = NVME_SC_INTERNAL;
954 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
955 	if (!ctrl)
956 		goto out_put_subsystem;
957 	mutex_init(&ctrl->lock);
958 
959 	nvmet_init_cap(ctrl);
960 
961 	ctrl->port = req->port;
962 
963 	INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
964 	INIT_LIST_HEAD(&ctrl->async_events);
965 
966 	memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
967 	memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
968 
969 	kref_init(&ctrl->ref);
970 	ctrl->subsys = subsys;
971 	WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
972 
973 	ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
974 			sizeof(__le32), GFP_KERNEL);
975 	if (!ctrl->changed_ns_list)
976 		goto out_free_ctrl;
977 
978 	ctrl->cqs = kcalloc(subsys->max_qid + 1,
979 			sizeof(struct nvmet_cq *),
980 			GFP_KERNEL);
981 	if (!ctrl->cqs)
982 		goto out_free_changed_ns_list;
983 
984 	ctrl->sqs = kcalloc(subsys->max_qid + 1,
985 			sizeof(struct nvmet_sq *),
986 			GFP_KERNEL);
987 	if (!ctrl->sqs)
988 		goto out_free_cqs;
989 
990 	ret = ida_simple_get(&cntlid_ida,
991 			     NVME_CNTLID_MIN, NVME_CNTLID_MAX,
992 			     GFP_KERNEL);
993 	if (ret < 0) {
994 		status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
995 		goto out_free_sqs;
996 	}
997 	ctrl->cntlid = ret;
998 
999 	ctrl->ops = req->ops;
1000 	if (ctrl->subsys->type == NVME_NQN_DISC) {
1001 		/* Don't accept keep-alive timeout for discovery controllers */
1002 		if (kato) {
1003 			status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
1004 			goto out_remove_ida;
1005 		}
1006 
1007 		/*
1008 		 * Discovery controllers use some arbitrary high value in order
1009 		 * to cleanup stale discovery sessions
1010 		 *
1011 		 * From the latest base diff RC:
1012 		 * "The Keep Alive command is not supported by
1013 		 * Discovery controllers. A transport may specify a
1014 		 * fixed Discovery controller activity timeout value
1015 		 * (e.g., 2 minutes).  If no commands are received
1016 		 * by a Discovery controller within that time
1017 		 * period, the controller may perform the
1018 		 * actions for Keep Alive Timer expiration".
1019 		 */
1020 		ctrl->kato = NVMET_DISC_KATO;
1021 	} else {
1022 		/* keep-alive timeout in seconds */
1023 		ctrl->kato = DIV_ROUND_UP(kato, 1000);
1024 	}
1025 	nvmet_start_keep_alive_timer(ctrl);
1026 
1027 	mutex_lock(&subsys->lock);
1028 	list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1029 	mutex_unlock(&subsys->lock);
1030 
1031 	*ctrlp = ctrl;
1032 	return 0;
1033 
1034 out_remove_ida:
1035 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1036 out_free_sqs:
1037 	kfree(ctrl->sqs);
1038 out_free_cqs:
1039 	kfree(ctrl->cqs);
1040 out_free_changed_ns_list:
1041 	kfree(ctrl->changed_ns_list);
1042 out_free_ctrl:
1043 	kfree(ctrl);
1044 out_put_subsystem:
1045 	nvmet_subsys_put(subsys);
1046 out:
1047 	return status;
1048 }
1049 
1050 static void nvmet_ctrl_free(struct kref *ref)
1051 {
1052 	struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1053 	struct nvmet_subsys *subsys = ctrl->subsys;
1054 
1055 	mutex_lock(&subsys->lock);
1056 	list_del(&ctrl->subsys_entry);
1057 	mutex_unlock(&subsys->lock);
1058 
1059 	nvmet_stop_keep_alive_timer(ctrl);
1060 
1061 	flush_work(&ctrl->async_event_work);
1062 	cancel_work_sync(&ctrl->fatal_err_work);
1063 
1064 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1065 
1066 	kfree(ctrl->sqs);
1067 	kfree(ctrl->cqs);
1068 	kfree(ctrl->changed_ns_list);
1069 	kfree(ctrl);
1070 
1071 	nvmet_subsys_put(subsys);
1072 }
1073 
1074 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1075 {
1076 	kref_put(&ctrl->ref, nvmet_ctrl_free);
1077 }
1078 
1079 static void nvmet_fatal_error_handler(struct work_struct *work)
1080 {
1081 	struct nvmet_ctrl *ctrl =
1082 			container_of(work, struct nvmet_ctrl, fatal_err_work);
1083 
1084 	pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1085 	ctrl->ops->delete_ctrl(ctrl);
1086 }
1087 
1088 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1089 {
1090 	mutex_lock(&ctrl->lock);
1091 	if (!(ctrl->csts & NVME_CSTS_CFS)) {
1092 		ctrl->csts |= NVME_CSTS_CFS;
1093 		INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1094 		schedule_work(&ctrl->fatal_err_work);
1095 	}
1096 	mutex_unlock(&ctrl->lock);
1097 }
1098 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1099 
1100 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1101 		const char *subsysnqn)
1102 {
1103 	struct nvmet_subsys_link *p;
1104 
1105 	if (!port)
1106 		return NULL;
1107 
1108 	if (!strncmp(NVME_DISC_SUBSYS_NAME, subsysnqn,
1109 			NVMF_NQN_SIZE)) {
1110 		if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1111 			return NULL;
1112 		return nvmet_disc_subsys;
1113 	}
1114 
1115 	down_read(&nvmet_config_sem);
1116 	list_for_each_entry(p, &port->subsystems, entry) {
1117 		if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1118 				NVMF_NQN_SIZE)) {
1119 			if (!kref_get_unless_zero(&p->subsys->ref))
1120 				break;
1121 			up_read(&nvmet_config_sem);
1122 			return p->subsys;
1123 		}
1124 	}
1125 	up_read(&nvmet_config_sem);
1126 	return NULL;
1127 }
1128 
1129 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1130 		enum nvme_subsys_type type)
1131 {
1132 	struct nvmet_subsys *subsys;
1133 
1134 	subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1135 	if (!subsys)
1136 		return NULL;
1137 
1138 	subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1139 	/* generate a random serial number as our controllers are ephemeral: */
1140 	get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1141 
1142 	switch (type) {
1143 	case NVME_NQN_NVME:
1144 		subsys->max_qid = NVMET_NR_QUEUES;
1145 		break;
1146 	case NVME_NQN_DISC:
1147 		subsys->max_qid = 0;
1148 		break;
1149 	default:
1150 		pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1151 		kfree(subsys);
1152 		return NULL;
1153 	}
1154 	subsys->type = type;
1155 	subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1156 			GFP_KERNEL);
1157 	if (!subsys->subsysnqn) {
1158 		kfree(subsys);
1159 		return NULL;
1160 	}
1161 
1162 	kref_init(&subsys->ref);
1163 
1164 	mutex_init(&subsys->lock);
1165 	INIT_LIST_HEAD(&subsys->namespaces);
1166 	INIT_LIST_HEAD(&subsys->ctrls);
1167 	INIT_LIST_HEAD(&subsys->hosts);
1168 
1169 	return subsys;
1170 }
1171 
1172 static void nvmet_subsys_free(struct kref *ref)
1173 {
1174 	struct nvmet_subsys *subsys =
1175 		container_of(ref, struct nvmet_subsys, ref);
1176 
1177 	WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1178 
1179 	kfree(subsys->subsysnqn);
1180 	kfree(subsys);
1181 }
1182 
1183 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1184 {
1185 	struct nvmet_ctrl *ctrl;
1186 
1187 	mutex_lock(&subsys->lock);
1188 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1189 		ctrl->ops->delete_ctrl(ctrl);
1190 	mutex_unlock(&subsys->lock);
1191 }
1192 
1193 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1194 {
1195 	kref_put(&subsys->ref, nvmet_subsys_free);
1196 }
1197 
1198 static int __init nvmet_init(void)
1199 {
1200 	int error;
1201 
1202 	nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1203 
1204 	buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1205 			WQ_MEM_RECLAIM, 0);
1206 	if (!buffered_io_wq) {
1207 		error = -ENOMEM;
1208 		goto out;
1209 	}
1210 
1211 	error = nvmet_init_discovery();
1212 	if (error)
1213 		goto out;
1214 
1215 	error = nvmet_init_configfs();
1216 	if (error)
1217 		goto out_exit_discovery;
1218 	return 0;
1219 
1220 out_exit_discovery:
1221 	nvmet_exit_discovery();
1222 out:
1223 	return error;
1224 }
1225 
1226 static void __exit nvmet_exit(void)
1227 {
1228 	nvmet_exit_configfs();
1229 	nvmet_exit_discovery();
1230 	ida_destroy(&cntlid_ida);
1231 	destroy_workqueue(buffered_io_wq);
1232 
1233 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1234 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1235 }
1236 
1237 module_init(nvmet_init);
1238 module_exit(nvmet_exit);
1239 
1240 MODULE_LICENSE("GPL v2");
1241