xref: /openbmc/linux/drivers/nvme/target/core.c (revision b830f94f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Common code for the NVMe target.
4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5  */
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/random.h>
9 #include <linux/rculist.h>
10 #include <linux/pci-p2pdma.h>
11 #include <linux/scatterlist.h>
12 
13 #define CREATE_TRACE_POINTS
14 #include "trace.h"
15 
16 #include "nvmet.h"
17 
18 struct workqueue_struct *buffered_io_wq;
19 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
20 static DEFINE_IDA(cntlid_ida);
21 
22 /*
23  * This read/write semaphore is used to synchronize access to configuration
24  * information on a target system that will result in discovery log page
25  * information change for at least one host.
26  * The full list of resources to protected by this semaphore is:
27  *
28  *  - subsystems list
29  *  - per-subsystem allowed hosts list
30  *  - allow_any_host subsystem attribute
31  *  - nvmet_genctr
32  *  - the nvmet_transports array
33  *
34  * When updating any of those lists/structures write lock should be obtained,
35  * while when reading (popolating discovery log page or checking host-subsystem
36  * link) read lock is obtained to allow concurrent reads.
37  */
38 DECLARE_RWSEM(nvmet_config_sem);
39 
40 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
41 u64 nvmet_ana_chgcnt;
42 DECLARE_RWSEM(nvmet_ana_sem);
43 
44 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
45 {
46 	u16 status;
47 
48 	switch (errno) {
49 	case -ENOSPC:
50 		req->error_loc = offsetof(struct nvme_rw_command, length);
51 		status = NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
52 		break;
53 	case -EREMOTEIO:
54 		req->error_loc = offsetof(struct nvme_rw_command, slba);
55 		status = NVME_SC_LBA_RANGE | NVME_SC_DNR;
56 		break;
57 	case -EOPNOTSUPP:
58 		req->error_loc = offsetof(struct nvme_common_command, opcode);
59 		switch (req->cmd->common.opcode) {
60 		case nvme_cmd_dsm:
61 		case nvme_cmd_write_zeroes:
62 			status = NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
63 			break;
64 		default:
65 			status = NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
66 		}
67 		break;
68 	case -ENODATA:
69 		req->error_loc = offsetof(struct nvme_rw_command, nsid);
70 		status = NVME_SC_ACCESS_DENIED;
71 		break;
72 	case -EIO:
73 		/* FALLTHRU */
74 	default:
75 		req->error_loc = offsetof(struct nvme_common_command, opcode);
76 		status = NVME_SC_INTERNAL | NVME_SC_DNR;
77 	}
78 
79 	return status;
80 }
81 
82 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
83 		const char *subsysnqn);
84 
85 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
86 		size_t len)
87 {
88 	if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
89 		req->error_loc = offsetof(struct nvme_common_command, dptr);
90 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
91 	}
92 	return 0;
93 }
94 
95 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
96 {
97 	if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
98 		req->error_loc = offsetof(struct nvme_common_command, dptr);
99 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
100 	}
101 	return 0;
102 }
103 
104 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
105 {
106 	if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
107 		req->error_loc = offsetof(struct nvme_common_command, dptr);
108 		return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
109 	}
110 	return 0;
111 }
112 
113 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
114 {
115 	struct nvmet_ns *ns;
116 
117 	if (list_empty(&subsys->namespaces))
118 		return 0;
119 
120 	ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
121 	return ns->nsid;
122 }
123 
124 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
125 {
126 	return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
127 }
128 
129 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
130 {
131 	struct nvmet_req *req;
132 
133 	while (1) {
134 		mutex_lock(&ctrl->lock);
135 		if (!ctrl->nr_async_event_cmds) {
136 			mutex_unlock(&ctrl->lock);
137 			return;
138 		}
139 
140 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
141 		mutex_unlock(&ctrl->lock);
142 		nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
143 	}
144 }
145 
146 static void nvmet_async_event_work(struct work_struct *work)
147 {
148 	struct nvmet_ctrl *ctrl =
149 		container_of(work, struct nvmet_ctrl, async_event_work);
150 	struct nvmet_async_event *aen;
151 	struct nvmet_req *req;
152 
153 	while (1) {
154 		mutex_lock(&ctrl->lock);
155 		aen = list_first_entry_or_null(&ctrl->async_events,
156 				struct nvmet_async_event, entry);
157 		if (!aen || !ctrl->nr_async_event_cmds) {
158 			mutex_unlock(&ctrl->lock);
159 			return;
160 		}
161 
162 		req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
163 		nvmet_set_result(req, nvmet_async_event_result(aen));
164 
165 		list_del(&aen->entry);
166 		kfree(aen);
167 
168 		mutex_unlock(&ctrl->lock);
169 		nvmet_req_complete(req, 0);
170 	}
171 }
172 
173 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
174 		u8 event_info, u8 log_page)
175 {
176 	struct nvmet_async_event *aen;
177 
178 	aen = kmalloc(sizeof(*aen), GFP_KERNEL);
179 	if (!aen)
180 		return;
181 
182 	aen->event_type = event_type;
183 	aen->event_info = event_info;
184 	aen->log_page = log_page;
185 
186 	mutex_lock(&ctrl->lock);
187 	list_add_tail(&aen->entry, &ctrl->async_events);
188 	mutex_unlock(&ctrl->lock);
189 
190 	schedule_work(&ctrl->async_event_work);
191 }
192 
193 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
194 {
195 	u32 i;
196 
197 	mutex_lock(&ctrl->lock);
198 	if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
199 		goto out_unlock;
200 
201 	for (i = 0; i < ctrl->nr_changed_ns; i++) {
202 		if (ctrl->changed_ns_list[i] == nsid)
203 			goto out_unlock;
204 	}
205 
206 	if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
207 		ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
208 		ctrl->nr_changed_ns = U32_MAX;
209 		goto out_unlock;
210 	}
211 
212 	ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
213 out_unlock:
214 	mutex_unlock(&ctrl->lock);
215 }
216 
217 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
218 {
219 	struct nvmet_ctrl *ctrl;
220 
221 	lockdep_assert_held(&subsys->lock);
222 
223 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
224 		nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
225 		if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
226 			continue;
227 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
228 				NVME_AER_NOTICE_NS_CHANGED,
229 				NVME_LOG_CHANGED_NS);
230 	}
231 }
232 
233 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
234 		struct nvmet_port *port)
235 {
236 	struct nvmet_ctrl *ctrl;
237 
238 	mutex_lock(&subsys->lock);
239 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
240 		if (port && ctrl->port != port)
241 			continue;
242 		if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
243 			continue;
244 		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
245 				NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
246 	}
247 	mutex_unlock(&subsys->lock);
248 }
249 
250 void nvmet_port_send_ana_event(struct nvmet_port *port)
251 {
252 	struct nvmet_subsys_link *p;
253 
254 	down_read(&nvmet_config_sem);
255 	list_for_each_entry(p, &port->subsystems, entry)
256 		nvmet_send_ana_event(p->subsys, port);
257 	up_read(&nvmet_config_sem);
258 }
259 
260 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
261 {
262 	int ret = 0;
263 
264 	down_write(&nvmet_config_sem);
265 	if (nvmet_transports[ops->type])
266 		ret = -EINVAL;
267 	else
268 		nvmet_transports[ops->type] = ops;
269 	up_write(&nvmet_config_sem);
270 
271 	return ret;
272 }
273 EXPORT_SYMBOL_GPL(nvmet_register_transport);
274 
275 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
276 {
277 	down_write(&nvmet_config_sem);
278 	nvmet_transports[ops->type] = NULL;
279 	up_write(&nvmet_config_sem);
280 }
281 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
282 
283 int nvmet_enable_port(struct nvmet_port *port)
284 {
285 	const struct nvmet_fabrics_ops *ops;
286 	int ret;
287 
288 	lockdep_assert_held(&nvmet_config_sem);
289 
290 	ops = nvmet_transports[port->disc_addr.trtype];
291 	if (!ops) {
292 		up_write(&nvmet_config_sem);
293 		request_module("nvmet-transport-%d", port->disc_addr.trtype);
294 		down_write(&nvmet_config_sem);
295 		ops = nvmet_transports[port->disc_addr.trtype];
296 		if (!ops) {
297 			pr_err("transport type %d not supported\n",
298 				port->disc_addr.trtype);
299 			return -EINVAL;
300 		}
301 	}
302 
303 	if (!try_module_get(ops->owner))
304 		return -EINVAL;
305 
306 	ret = ops->add_port(port);
307 	if (ret) {
308 		module_put(ops->owner);
309 		return ret;
310 	}
311 
312 	/* If the transport didn't set inline_data_size, then disable it. */
313 	if (port->inline_data_size < 0)
314 		port->inline_data_size = 0;
315 
316 	port->enabled = true;
317 	port->tr_ops = ops;
318 	return 0;
319 }
320 
321 void nvmet_disable_port(struct nvmet_port *port)
322 {
323 	const struct nvmet_fabrics_ops *ops;
324 
325 	lockdep_assert_held(&nvmet_config_sem);
326 
327 	port->enabled = false;
328 	port->tr_ops = NULL;
329 
330 	ops = nvmet_transports[port->disc_addr.trtype];
331 	ops->remove_port(port);
332 	module_put(ops->owner);
333 }
334 
335 static void nvmet_keep_alive_timer(struct work_struct *work)
336 {
337 	struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
338 			struct nvmet_ctrl, ka_work);
339 	bool cmd_seen = ctrl->cmd_seen;
340 
341 	ctrl->cmd_seen = false;
342 	if (cmd_seen) {
343 		pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
344 			ctrl->cntlid);
345 		schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
346 		return;
347 	}
348 
349 	pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
350 		ctrl->cntlid, ctrl->kato);
351 
352 	nvmet_ctrl_fatal_error(ctrl);
353 }
354 
355 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
356 {
357 	pr_debug("ctrl %d start keep-alive timer for %d secs\n",
358 		ctrl->cntlid, ctrl->kato);
359 
360 	INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
361 	schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
362 }
363 
364 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
365 {
366 	pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
367 
368 	cancel_delayed_work_sync(&ctrl->ka_work);
369 }
370 
371 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
372 		__le32 nsid)
373 {
374 	struct nvmet_ns *ns;
375 
376 	list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
377 		if (ns->nsid == le32_to_cpu(nsid))
378 			return ns;
379 	}
380 
381 	return NULL;
382 }
383 
384 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
385 {
386 	struct nvmet_ns *ns;
387 
388 	rcu_read_lock();
389 	ns = __nvmet_find_namespace(ctrl, nsid);
390 	if (ns)
391 		percpu_ref_get(&ns->ref);
392 	rcu_read_unlock();
393 
394 	return ns;
395 }
396 
397 static void nvmet_destroy_namespace(struct percpu_ref *ref)
398 {
399 	struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
400 
401 	complete(&ns->disable_done);
402 }
403 
404 void nvmet_put_namespace(struct nvmet_ns *ns)
405 {
406 	percpu_ref_put(&ns->ref);
407 }
408 
409 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
410 {
411 	nvmet_bdev_ns_disable(ns);
412 	nvmet_file_ns_disable(ns);
413 }
414 
415 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
416 {
417 	int ret;
418 	struct pci_dev *p2p_dev;
419 
420 	if (!ns->use_p2pmem)
421 		return 0;
422 
423 	if (!ns->bdev) {
424 		pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
425 		return -EINVAL;
426 	}
427 
428 	if (!blk_queue_pci_p2pdma(ns->bdev->bd_queue)) {
429 		pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
430 		       ns->device_path);
431 		return -EINVAL;
432 	}
433 
434 	if (ns->p2p_dev) {
435 		ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
436 		if (ret < 0)
437 			return -EINVAL;
438 	} else {
439 		/*
440 		 * Right now we just check that there is p2pmem available so
441 		 * we can report an error to the user right away if there
442 		 * is not. We'll find the actual device to use once we
443 		 * setup the controller when the port's device is available.
444 		 */
445 
446 		p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
447 		if (!p2p_dev) {
448 			pr_err("no peer-to-peer memory is available for %s\n",
449 			       ns->device_path);
450 			return -EINVAL;
451 		}
452 
453 		pci_dev_put(p2p_dev);
454 	}
455 
456 	return 0;
457 }
458 
459 /*
460  * Note: ctrl->subsys->lock should be held when calling this function
461  */
462 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
463 				    struct nvmet_ns *ns)
464 {
465 	struct device *clients[2];
466 	struct pci_dev *p2p_dev;
467 	int ret;
468 
469 	if (!ctrl->p2p_client || !ns->use_p2pmem)
470 		return;
471 
472 	if (ns->p2p_dev) {
473 		ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
474 		if (ret < 0)
475 			return;
476 
477 		p2p_dev = pci_dev_get(ns->p2p_dev);
478 	} else {
479 		clients[0] = ctrl->p2p_client;
480 		clients[1] = nvmet_ns_dev(ns);
481 
482 		p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
483 		if (!p2p_dev) {
484 			pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
485 			       dev_name(ctrl->p2p_client), ns->device_path);
486 			return;
487 		}
488 	}
489 
490 	ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
491 	if (ret < 0)
492 		pci_dev_put(p2p_dev);
493 
494 	pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
495 		ns->nsid);
496 }
497 
498 int nvmet_ns_enable(struct nvmet_ns *ns)
499 {
500 	struct nvmet_subsys *subsys = ns->subsys;
501 	struct nvmet_ctrl *ctrl;
502 	int ret;
503 
504 	mutex_lock(&subsys->lock);
505 	ret = 0;
506 	if (ns->enabled)
507 		goto out_unlock;
508 
509 	ret = -EMFILE;
510 	if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
511 		goto out_unlock;
512 
513 	ret = nvmet_bdev_ns_enable(ns);
514 	if (ret == -ENOTBLK)
515 		ret = nvmet_file_ns_enable(ns);
516 	if (ret)
517 		goto out_unlock;
518 
519 	ret = nvmet_p2pmem_ns_enable(ns);
520 	if (ret)
521 		goto out_dev_disable;
522 
523 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
524 		nvmet_p2pmem_ns_add_p2p(ctrl, ns);
525 
526 	ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
527 				0, GFP_KERNEL);
528 	if (ret)
529 		goto out_dev_put;
530 
531 	if (ns->nsid > subsys->max_nsid)
532 		subsys->max_nsid = ns->nsid;
533 
534 	/*
535 	 * The namespaces list needs to be sorted to simplify the implementation
536 	 * of the Identify Namepace List subcommand.
537 	 */
538 	if (list_empty(&subsys->namespaces)) {
539 		list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
540 	} else {
541 		struct nvmet_ns *old;
542 
543 		list_for_each_entry_rcu(old, &subsys->namespaces, dev_link) {
544 			BUG_ON(ns->nsid == old->nsid);
545 			if (ns->nsid < old->nsid)
546 				break;
547 		}
548 
549 		list_add_tail_rcu(&ns->dev_link, &old->dev_link);
550 	}
551 	subsys->nr_namespaces++;
552 
553 	nvmet_ns_changed(subsys, ns->nsid);
554 	ns->enabled = true;
555 	ret = 0;
556 out_unlock:
557 	mutex_unlock(&subsys->lock);
558 	return ret;
559 out_dev_put:
560 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
561 		pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
562 out_dev_disable:
563 	nvmet_ns_dev_disable(ns);
564 	goto out_unlock;
565 }
566 
567 void nvmet_ns_disable(struct nvmet_ns *ns)
568 {
569 	struct nvmet_subsys *subsys = ns->subsys;
570 	struct nvmet_ctrl *ctrl;
571 
572 	mutex_lock(&subsys->lock);
573 	if (!ns->enabled)
574 		goto out_unlock;
575 
576 	ns->enabled = false;
577 	list_del_rcu(&ns->dev_link);
578 	if (ns->nsid == subsys->max_nsid)
579 		subsys->max_nsid = nvmet_max_nsid(subsys);
580 
581 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
582 		pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
583 
584 	mutex_unlock(&subsys->lock);
585 
586 	/*
587 	 * Now that we removed the namespaces from the lookup list, we
588 	 * can kill the per_cpu ref and wait for any remaining references
589 	 * to be dropped, as well as a RCU grace period for anyone only
590 	 * using the namepace under rcu_read_lock().  Note that we can't
591 	 * use call_rcu here as we need to ensure the namespaces have
592 	 * been fully destroyed before unloading the module.
593 	 */
594 	percpu_ref_kill(&ns->ref);
595 	synchronize_rcu();
596 	wait_for_completion(&ns->disable_done);
597 	percpu_ref_exit(&ns->ref);
598 
599 	mutex_lock(&subsys->lock);
600 
601 	subsys->nr_namespaces--;
602 	nvmet_ns_changed(subsys, ns->nsid);
603 	nvmet_ns_dev_disable(ns);
604 out_unlock:
605 	mutex_unlock(&subsys->lock);
606 }
607 
608 void nvmet_ns_free(struct nvmet_ns *ns)
609 {
610 	nvmet_ns_disable(ns);
611 
612 	down_write(&nvmet_ana_sem);
613 	nvmet_ana_group_enabled[ns->anagrpid]--;
614 	up_write(&nvmet_ana_sem);
615 
616 	kfree(ns->device_path);
617 	kfree(ns);
618 }
619 
620 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
621 {
622 	struct nvmet_ns *ns;
623 
624 	ns = kzalloc(sizeof(*ns), GFP_KERNEL);
625 	if (!ns)
626 		return NULL;
627 
628 	INIT_LIST_HEAD(&ns->dev_link);
629 	init_completion(&ns->disable_done);
630 
631 	ns->nsid = nsid;
632 	ns->subsys = subsys;
633 
634 	down_write(&nvmet_ana_sem);
635 	ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
636 	nvmet_ana_group_enabled[ns->anagrpid]++;
637 	up_write(&nvmet_ana_sem);
638 
639 	uuid_gen(&ns->uuid);
640 	ns->buffered_io = false;
641 
642 	return ns;
643 }
644 
645 static void nvmet_update_sq_head(struct nvmet_req *req)
646 {
647 	if (req->sq->size) {
648 		u32 old_sqhd, new_sqhd;
649 
650 		do {
651 			old_sqhd = req->sq->sqhd;
652 			new_sqhd = (old_sqhd + 1) % req->sq->size;
653 		} while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
654 					old_sqhd);
655 	}
656 	req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
657 }
658 
659 static void nvmet_set_error(struct nvmet_req *req, u16 status)
660 {
661 	struct nvmet_ctrl *ctrl = req->sq->ctrl;
662 	struct nvme_error_slot *new_error_slot;
663 	unsigned long flags;
664 
665 	req->cqe->status = cpu_to_le16(status << 1);
666 
667 	if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
668 		return;
669 
670 	spin_lock_irqsave(&ctrl->error_lock, flags);
671 	ctrl->err_counter++;
672 	new_error_slot =
673 		&ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
674 
675 	new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
676 	new_error_slot->sqid = cpu_to_le16(req->sq->qid);
677 	new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
678 	new_error_slot->status_field = cpu_to_le16(status << 1);
679 	new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
680 	new_error_slot->lba = cpu_to_le64(req->error_slba);
681 	new_error_slot->nsid = req->cmd->common.nsid;
682 	spin_unlock_irqrestore(&ctrl->error_lock, flags);
683 
684 	/* set the more bit for this request */
685 	req->cqe->status |= cpu_to_le16(1 << 14);
686 }
687 
688 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
689 {
690 	if (!req->sq->sqhd_disabled)
691 		nvmet_update_sq_head(req);
692 	req->cqe->sq_id = cpu_to_le16(req->sq->qid);
693 	req->cqe->command_id = req->cmd->common.command_id;
694 
695 	if (unlikely(status))
696 		nvmet_set_error(req, status);
697 
698 	trace_nvmet_req_complete(req);
699 
700 	if (req->ns)
701 		nvmet_put_namespace(req->ns);
702 	req->ops->queue_response(req);
703 }
704 
705 void nvmet_req_complete(struct nvmet_req *req, u16 status)
706 {
707 	__nvmet_req_complete(req, status);
708 	percpu_ref_put(&req->sq->ref);
709 }
710 EXPORT_SYMBOL_GPL(nvmet_req_complete);
711 
712 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
713 		u16 qid, u16 size)
714 {
715 	cq->qid = qid;
716 	cq->size = size;
717 
718 	ctrl->cqs[qid] = cq;
719 }
720 
721 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
722 		u16 qid, u16 size)
723 {
724 	sq->sqhd = 0;
725 	sq->qid = qid;
726 	sq->size = size;
727 
728 	ctrl->sqs[qid] = sq;
729 }
730 
731 static void nvmet_confirm_sq(struct percpu_ref *ref)
732 {
733 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
734 
735 	complete(&sq->confirm_done);
736 }
737 
738 void nvmet_sq_destroy(struct nvmet_sq *sq)
739 {
740 	/*
741 	 * If this is the admin queue, complete all AERs so that our
742 	 * queue doesn't have outstanding requests on it.
743 	 */
744 	if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
745 		nvmet_async_events_free(sq->ctrl);
746 	percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
747 	wait_for_completion(&sq->confirm_done);
748 	wait_for_completion(&sq->free_done);
749 	percpu_ref_exit(&sq->ref);
750 
751 	if (sq->ctrl) {
752 		nvmet_ctrl_put(sq->ctrl);
753 		sq->ctrl = NULL; /* allows reusing the queue later */
754 	}
755 }
756 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
757 
758 static void nvmet_sq_free(struct percpu_ref *ref)
759 {
760 	struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
761 
762 	complete(&sq->free_done);
763 }
764 
765 int nvmet_sq_init(struct nvmet_sq *sq)
766 {
767 	int ret;
768 
769 	ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
770 	if (ret) {
771 		pr_err("percpu_ref init failed!\n");
772 		return ret;
773 	}
774 	init_completion(&sq->free_done);
775 	init_completion(&sq->confirm_done);
776 
777 	return 0;
778 }
779 EXPORT_SYMBOL_GPL(nvmet_sq_init);
780 
781 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
782 		struct nvmet_ns *ns)
783 {
784 	enum nvme_ana_state state = port->ana_state[ns->anagrpid];
785 
786 	if (unlikely(state == NVME_ANA_INACCESSIBLE))
787 		return NVME_SC_ANA_INACCESSIBLE;
788 	if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
789 		return NVME_SC_ANA_PERSISTENT_LOSS;
790 	if (unlikely(state == NVME_ANA_CHANGE))
791 		return NVME_SC_ANA_TRANSITION;
792 	return 0;
793 }
794 
795 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
796 {
797 	if (unlikely(req->ns->readonly)) {
798 		switch (req->cmd->common.opcode) {
799 		case nvme_cmd_read:
800 		case nvme_cmd_flush:
801 			break;
802 		default:
803 			return NVME_SC_NS_WRITE_PROTECTED;
804 		}
805 	}
806 
807 	return 0;
808 }
809 
810 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
811 {
812 	struct nvme_command *cmd = req->cmd;
813 	u16 ret;
814 
815 	ret = nvmet_check_ctrl_status(req, cmd);
816 	if (unlikely(ret))
817 		return ret;
818 
819 	req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
820 	if (unlikely(!req->ns)) {
821 		req->error_loc = offsetof(struct nvme_common_command, nsid);
822 		return NVME_SC_INVALID_NS | NVME_SC_DNR;
823 	}
824 	ret = nvmet_check_ana_state(req->port, req->ns);
825 	if (unlikely(ret)) {
826 		req->error_loc = offsetof(struct nvme_common_command, nsid);
827 		return ret;
828 	}
829 	ret = nvmet_io_cmd_check_access(req);
830 	if (unlikely(ret)) {
831 		req->error_loc = offsetof(struct nvme_common_command, nsid);
832 		return ret;
833 	}
834 
835 	if (req->ns->file)
836 		return nvmet_file_parse_io_cmd(req);
837 	else
838 		return nvmet_bdev_parse_io_cmd(req);
839 }
840 
841 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
842 		struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
843 {
844 	u8 flags = req->cmd->common.flags;
845 	u16 status;
846 
847 	req->cq = cq;
848 	req->sq = sq;
849 	req->ops = ops;
850 	req->sg = NULL;
851 	req->sg_cnt = 0;
852 	req->transfer_len = 0;
853 	req->cqe->status = 0;
854 	req->cqe->sq_head = 0;
855 	req->ns = NULL;
856 	req->error_loc = NVMET_NO_ERROR_LOC;
857 	req->error_slba = 0;
858 
859 	trace_nvmet_req_init(req, req->cmd);
860 
861 	/* no support for fused commands yet */
862 	if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
863 		req->error_loc = offsetof(struct nvme_common_command, flags);
864 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
865 		goto fail;
866 	}
867 
868 	/*
869 	 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
870 	 * contains an address of a single contiguous physical buffer that is
871 	 * byte aligned.
872 	 */
873 	if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
874 		req->error_loc = offsetof(struct nvme_common_command, flags);
875 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
876 		goto fail;
877 	}
878 
879 	if (unlikely(!req->sq->ctrl))
880 		/* will return an error for any Non-connect command: */
881 		status = nvmet_parse_connect_cmd(req);
882 	else if (likely(req->sq->qid != 0))
883 		status = nvmet_parse_io_cmd(req);
884 	else if (nvme_is_fabrics(req->cmd))
885 		status = nvmet_parse_fabrics_cmd(req);
886 	else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
887 		status = nvmet_parse_discovery_cmd(req);
888 	else
889 		status = nvmet_parse_admin_cmd(req);
890 
891 	if (status)
892 		goto fail;
893 
894 	if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
895 		status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
896 		goto fail;
897 	}
898 
899 	if (sq->ctrl)
900 		sq->ctrl->cmd_seen = true;
901 
902 	return true;
903 
904 fail:
905 	__nvmet_req_complete(req, status);
906 	return false;
907 }
908 EXPORT_SYMBOL_GPL(nvmet_req_init);
909 
910 void nvmet_req_uninit(struct nvmet_req *req)
911 {
912 	percpu_ref_put(&req->sq->ref);
913 	if (req->ns)
914 		nvmet_put_namespace(req->ns);
915 }
916 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
917 
918 void nvmet_req_execute(struct nvmet_req *req)
919 {
920 	if (unlikely(req->data_len != req->transfer_len)) {
921 		req->error_loc = offsetof(struct nvme_common_command, dptr);
922 		nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
923 	} else
924 		req->execute(req);
925 }
926 EXPORT_SYMBOL_GPL(nvmet_req_execute);
927 
928 int nvmet_req_alloc_sgl(struct nvmet_req *req)
929 {
930 	struct pci_dev *p2p_dev = NULL;
931 
932 	if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
933 		if (req->sq->ctrl && req->ns)
934 			p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
935 						    req->ns->nsid);
936 
937 		req->p2p_dev = NULL;
938 		if (req->sq->qid && p2p_dev) {
939 			req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
940 						       req->transfer_len);
941 			if (req->sg) {
942 				req->p2p_dev = p2p_dev;
943 				return 0;
944 			}
945 		}
946 
947 		/*
948 		 * If no P2P memory was available we fallback to using
949 		 * regular memory
950 		 */
951 	}
952 
953 	req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
954 	if (!req->sg)
955 		return -ENOMEM;
956 
957 	return 0;
958 }
959 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
960 
961 void nvmet_req_free_sgl(struct nvmet_req *req)
962 {
963 	if (req->p2p_dev)
964 		pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
965 	else
966 		sgl_free(req->sg);
967 
968 	req->sg = NULL;
969 	req->sg_cnt = 0;
970 }
971 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
972 
973 static inline bool nvmet_cc_en(u32 cc)
974 {
975 	return (cc >> NVME_CC_EN_SHIFT) & 0x1;
976 }
977 
978 static inline u8 nvmet_cc_css(u32 cc)
979 {
980 	return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
981 }
982 
983 static inline u8 nvmet_cc_mps(u32 cc)
984 {
985 	return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
986 }
987 
988 static inline u8 nvmet_cc_ams(u32 cc)
989 {
990 	return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
991 }
992 
993 static inline u8 nvmet_cc_shn(u32 cc)
994 {
995 	return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
996 }
997 
998 static inline u8 nvmet_cc_iosqes(u32 cc)
999 {
1000 	return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1001 }
1002 
1003 static inline u8 nvmet_cc_iocqes(u32 cc)
1004 {
1005 	return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1006 }
1007 
1008 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1009 {
1010 	lockdep_assert_held(&ctrl->lock);
1011 
1012 	if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1013 	    nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1014 	    nvmet_cc_mps(ctrl->cc) != 0 ||
1015 	    nvmet_cc_ams(ctrl->cc) != 0 ||
1016 	    nvmet_cc_css(ctrl->cc) != 0) {
1017 		ctrl->csts = NVME_CSTS_CFS;
1018 		return;
1019 	}
1020 
1021 	ctrl->csts = NVME_CSTS_RDY;
1022 
1023 	/*
1024 	 * Controllers that are not yet enabled should not really enforce the
1025 	 * keep alive timeout, but we still want to track a timeout and cleanup
1026 	 * in case a host died before it enabled the controller.  Hence, simply
1027 	 * reset the keep alive timer when the controller is enabled.
1028 	 */
1029 	mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1030 }
1031 
1032 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1033 {
1034 	lockdep_assert_held(&ctrl->lock);
1035 
1036 	/* XXX: tear down queues? */
1037 	ctrl->csts &= ~NVME_CSTS_RDY;
1038 	ctrl->cc = 0;
1039 }
1040 
1041 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1042 {
1043 	u32 old;
1044 
1045 	mutex_lock(&ctrl->lock);
1046 	old = ctrl->cc;
1047 	ctrl->cc = new;
1048 
1049 	if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1050 		nvmet_start_ctrl(ctrl);
1051 	if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1052 		nvmet_clear_ctrl(ctrl);
1053 	if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1054 		nvmet_clear_ctrl(ctrl);
1055 		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1056 	}
1057 	if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1058 		ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1059 	mutex_unlock(&ctrl->lock);
1060 }
1061 
1062 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1063 {
1064 	/* command sets supported: NVMe command set: */
1065 	ctrl->cap = (1ULL << 37);
1066 	/* CC.EN timeout in 500msec units: */
1067 	ctrl->cap |= (15ULL << 24);
1068 	/* maximum queue entries supported: */
1069 	ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1070 }
1071 
1072 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1073 		struct nvmet_req *req, struct nvmet_ctrl **ret)
1074 {
1075 	struct nvmet_subsys *subsys;
1076 	struct nvmet_ctrl *ctrl;
1077 	u16 status = 0;
1078 
1079 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1080 	if (!subsys) {
1081 		pr_warn("connect request for invalid subsystem %s!\n",
1082 			subsysnqn);
1083 		req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1084 		return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1085 	}
1086 
1087 	mutex_lock(&subsys->lock);
1088 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1089 		if (ctrl->cntlid == cntlid) {
1090 			if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1091 				pr_warn("hostnqn mismatch.\n");
1092 				continue;
1093 			}
1094 			if (!kref_get_unless_zero(&ctrl->ref))
1095 				continue;
1096 
1097 			*ret = ctrl;
1098 			goto out;
1099 		}
1100 	}
1101 
1102 	pr_warn("could not find controller %d for subsys %s / host %s\n",
1103 		cntlid, subsysnqn, hostnqn);
1104 	req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1105 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1106 
1107 out:
1108 	mutex_unlock(&subsys->lock);
1109 	nvmet_subsys_put(subsys);
1110 	return status;
1111 }
1112 
1113 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1114 {
1115 	if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1116 		pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1117 		       cmd->common.opcode, req->sq->qid);
1118 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1119 	}
1120 
1121 	if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1122 		pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1123 		       cmd->common.opcode, req->sq->qid);
1124 		return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1125 	}
1126 	return 0;
1127 }
1128 
1129 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1130 {
1131 	struct nvmet_host_link *p;
1132 
1133 	lockdep_assert_held(&nvmet_config_sem);
1134 
1135 	if (subsys->allow_any_host)
1136 		return true;
1137 
1138 	if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1139 		return true;
1140 
1141 	list_for_each_entry(p, &subsys->hosts, entry) {
1142 		if (!strcmp(nvmet_host_name(p->host), hostnqn))
1143 			return true;
1144 	}
1145 
1146 	return false;
1147 }
1148 
1149 /*
1150  * Note: ctrl->subsys->lock should be held when calling this function
1151  */
1152 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1153 		struct nvmet_req *req)
1154 {
1155 	struct nvmet_ns *ns;
1156 
1157 	if (!req->p2p_client)
1158 		return;
1159 
1160 	ctrl->p2p_client = get_device(req->p2p_client);
1161 
1162 	list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link)
1163 		nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1164 }
1165 
1166 /*
1167  * Note: ctrl->subsys->lock should be held when calling this function
1168  */
1169 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1170 {
1171 	struct radix_tree_iter iter;
1172 	void __rcu **slot;
1173 
1174 	radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1175 		pci_dev_put(radix_tree_deref_slot(slot));
1176 
1177 	put_device(ctrl->p2p_client);
1178 }
1179 
1180 static void nvmet_fatal_error_handler(struct work_struct *work)
1181 {
1182 	struct nvmet_ctrl *ctrl =
1183 			container_of(work, struct nvmet_ctrl, fatal_err_work);
1184 
1185 	pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1186 	ctrl->ops->delete_ctrl(ctrl);
1187 }
1188 
1189 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1190 		struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1191 {
1192 	struct nvmet_subsys *subsys;
1193 	struct nvmet_ctrl *ctrl;
1194 	int ret;
1195 	u16 status;
1196 
1197 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1198 	subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1199 	if (!subsys) {
1200 		pr_warn("connect request for invalid subsystem %s!\n",
1201 			subsysnqn);
1202 		req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1203 		goto out;
1204 	}
1205 
1206 	status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1207 	down_read(&nvmet_config_sem);
1208 	if (!nvmet_host_allowed(subsys, hostnqn)) {
1209 		pr_info("connect by host %s for subsystem %s not allowed\n",
1210 			hostnqn, subsysnqn);
1211 		req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1212 		up_read(&nvmet_config_sem);
1213 		status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1214 		goto out_put_subsystem;
1215 	}
1216 	up_read(&nvmet_config_sem);
1217 
1218 	status = NVME_SC_INTERNAL;
1219 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1220 	if (!ctrl)
1221 		goto out_put_subsystem;
1222 	mutex_init(&ctrl->lock);
1223 
1224 	nvmet_init_cap(ctrl);
1225 
1226 	ctrl->port = req->port;
1227 
1228 	INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1229 	INIT_LIST_HEAD(&ctrl->async_events);
1230 	INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1231 	INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1232 
1233 	memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1234 	memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1235 
1236 	kref_init(&ctrl->ref);
1237 	ctrl->subsys = subsys;
1238 	WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1239 
1240 	ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1241 			sizeof(__le32), GFP_KERNEL);
1242 	if (!ctrl->changed_ns_list)
1243 		goto out_free_ctrl;
1244 
1245 	ctrl->cqs = kcalloc(subsys->max_qid + 1,
1246 			sizeof(struct nvmet_cq *),
1247 			GFP_KERNEL);
1248 	if (!ctrl->cqs)
1249 		goto out_free_changed_ns_list;
1250 
1251 	ctrl->sqs = kcalloc(subsys->max_qid + 1,
1252 			sizeof(struct nvmet_sq *),
1253 			GFP_KERNEL);
1254 	if (!ctrl->sqs)
1255 		goto out_free_cqs;
1256 
1257 	ret = ida_simple_get(&cntlid_ida,
1258 			     NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1259 			     GFP_KERNEL);
1260 	if (ret < 0) {
1261 		status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1262 		goto out_free_sqs;
1263 	}
1264 	ctrl->cntlid = ret;
1265 
1266 	ctrl->ops = req->ops;
1267 
1268 	/*
1269 	 * Discovery controllers may use some arbitrary high value
1270 	 * in order to cleanup stale discovery sessions
1271 	 */
1272 	if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1273 		kato = NVMET_DISC_KATO_MS;
1274 
1275 	/* keep-alive timeout in seconds */
1276 	ctrl->kato = DIV_ROUND_UP(kato, 1000);
1277 
1278 	ctrl->err_counter = 0;
1279 	spin_lock_init(&ctrl->error_lock);
1280 
1281 	nvmet_start_keep_alive_timer(ctrl);
1282 
1283 	mutex_lock(&subsys->lock);
1284 	list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1285 	nvmet_setup_p2p_ns_map(ctrl, req);
1286 	mutex_unlock(&subsys->lock);
1287 
1288 	*ctrlp = ctrl;
1289 	return 0;
1290 
1291 out_free_sqs:
1292 	kfree(ctrl->sqs);
1293 out_free_cqs:
1294 	kfree(ctrl->cqs);
1295 out_free_changed_ns_list:
1296 	kfree(ctrl->changed_ns_list);
1297 out_free_ctrl:
1298 	kfree(ctrl);
1299 out_put_subsystem:
1300 	nvmet_subsys_put(subsys);
1301 out:
1302 	return status;
1303 }
1304 
1305 static void nvmet_ctrl_free(struct kref *ref)
1306 {
1307 	struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1308 	struct nvmet_subsys *subsys = ctrl->subsys;
1309 
1310 	mutex_lock(&subsys->lock);
1311 	nvmet_release_p2p_ns_map(ctrl);
1312 	list_del(&ctrl->subsys_entry);
1313 	mutex_unlock(&subsys->lock);
1314 
1315 	nvmet_stop_keep_alive_timer(ctrl);
1316 
1317 	flush_work(&ctrl->async_event_work);
1318 	cancel_work_sync(&ctrl->fatal_err_work);
1319 
1320 	ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1321 
1322 	kfree(ctrl->sqs);
1323 	kfree(ctrl->cqs);
1324 	kfree(ctrl->changed_ns_list);
1325 	kfree(ctrl);
1326 
1327 	nvmet_subsys_put(subsys);
1328 }
1329 
1330 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1331 {
1332 	kref_put(&ctrl->ref, nvmet_ctrl_free);
1333 }
1334 
1335 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1336 {
1337 	mutex_lock(&ctrl->lock);
1338 	if (!(ctrl->csts & NVME_CSTS_CFS)) {
1339 		ctrl->csts |= NVME_CSTS_CFS;
1340 		schedule_work(&ctrl->fatal_err_work);
1341 	}
1342 	mutex_unlock(&ctrl->lock);
1343 }
1344 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1345 
1346 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1347 		const char *subsysnqn)
1348 {
1349 	struct nvmet_subsys_link *p;
1350 
1351 	if (!port)
1352 		return NULL;
1353 
1354 	if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1355 		if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1356 			return NULL;
1357 		return nvmet_disc_subsys;
1358 	}
1359 
1360 	down_read(&nvmet_config_sem);
1361 	list_for_each_entry(p, &port->subsystems, entry) {
1362 		if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1363 				NVMF_NQN_SIZE)) {
1364 			if (!kref_get_unless_zero(&p->subsys->ref))
1365 				break;
1366 			up_read(&nvmet_config_sem);
1367 			return p->subsys;
1368 		}
1369 	}
1370 	up_read(&nvmet_config_sem);
1371 	return NULL;
1372 }
1373 
1374 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1375 		enum nvme_subsys_type type)
1376 {
1377 	struct nvmet_subsys *subsys;
1378 
1379 	subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1380 	if (!subsys)
1381 		return ERR_PTR(-ENOMEM);
1382 
1383 	subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1384 	/* generate a random serial number as our controllers are ephemeral: */
1385 	get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1386 
1387 	switch (type) {
1388 	case NVME_NQN_NVME:
1389 		subsys->max_qid = NVMET_NR_QUEUES;
1390 		break;
1391 	case NVME_NQN_DISC:
1392 		subsys->max_qid = 0;
1393 		break;
1394 	default:
1395 		pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1396 		kfree(subsys);
1397 		return ERR_PTR(-EINVAL);
1398 	}
1399 	subsys->type = type;
1400 	subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1401 			GFP_KERNEL);
1402 	if (!subsys->subsysnqn) {
1403 		kfree(subsys);
1404 		return ERR_PTR(-ENOMEM);
1405 	}
1406 
1407 	kref_init(&subsys->ref);
1408 
1409 	mutex_init(&subsys->lock);
1410 	INIT_LIST_HEAD(&subsys->namespaces);
1411 	INIT_LIST_HEAD(&subsys->ctrls);
1412 	INIT_LIST_HEAD(&subsys->hosts);
1413 
1414 	return subsys;
1415 }
1416 
1417 static void nvmet_subsys_free(struct kref *ref)
1418 {
1419 	struct nvmet_subsys *subsys =
1420 		container_of(ref, struct nvmet_subsys, ref);
1421 
1422 	WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1423 
1424 	kfree(subsys->subsysnqn);
1425 	kfree(subsys);
1426 }
1427 
1428 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1429 {
1430 	struct nvmet_ctrl *ctrl;
1431 
1432 	mutex_lock(&subsys->lock);
1433 	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1434 		ctrl->ops->delete_ctrl(ctrl);
1435 	mutex_unlock(&subsys->lock);
1436 }
1437 
1438 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1439 {
1440 	kref_put(&subsys->ref, nvmet_subsys_free);
1441 }
1442 
1443 static int __init nvmet_init(void)
1444 {
1445 	int error;
1446 
1447 	nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1448 
1449 	buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1450 			WQ_MEM_RECLAIM, 0);
1451 	if (!buffered_io_wq) {
1452 		error = -ENOMEM;
1453 		goto out;
1454 	}
1455 
1456 	error = nvmet_init_discovery();
1457 	if (error)
1458 		goto out_free_work_queue;
1459 
1460 	error = nvmet_init_configfs();
1461 	if (error)
1462 		goto out_exit_discovery;
1463 	return 0;
1464 
1465 out_exit_discovery:
1466 	nvmet_exit_discovery();
1467 out_free_work_queue:
1468 	destroy_workqueue(buffered_io_wq);
1469 out:
1470 	return error;
1471 }
1472 
1473 static void __exit nvmet_exit(void)
1474 {
1475 	nvmet_exit_configfs();
1476 	nvmet_exit_discovery();
1477 	ida_destroy(&cntlid_ida);
1478 	destroy_workqueue(buffered_io_wq);
1479 
1480 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1481 	BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1482 }
1483 
1484 module_init(nvmet_init);
1485 module_exit(nvmet_exit);
1486 
1487 MODULE_LICENSE("GPL v2");
1488