xref: /openbmc/linux/drivers/block/rnbd/rnbd-clt.c (revision e961f8c6)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * RDMA Network Block Driver
4  *
5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
8  */
9 
10 #undef pr_fmt
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
12 
13 #include <linux/module.h>
14 #include <linux/blkdev.h>
15 #include <linux/hdreg.h>
16 #include <linux/scatterlist.h>
17 #include <linux/idr.h>
18 
19 #include "rnbd-clt.h"
20 
21 MODULE_DESCRIPTION("RDMA Network Block Device Client");
22 MODULE_LICENSE("GPL");
23 
24 static int rnbd_client_major;
25 static DEFINE_IDA(index_ida);
26 static DEFINE_MUTEX(sess_lock);
27 static LIST_HEAD(sess_list);
28 static struct workqueue_struct *rnbd_clt_wq;
29 
30 /*
31  * Maximum number of partitions an instance can have.
32  * 6 bits = 64 minors = 63 partitions (one minor is used for the device itself)
33  */
34 #define RNBD_PART_BITS		6
35 
36 static inline bool rnbd_clt_get_sess(struct rnbd_clt_session *sess)
37 {
38 	return refcount_inc_not_zero(&sess->refcount);
39 }
40 
41 static void free_sess(struct rnbd_clt_session *sess);
42 
43 static void rnbd_clt_put_sess(struct rnbd_clt_session *sess)
44 {
45 	might_sleep();
46 
47 	if (refcount_dec_and_test(&sess->refcount))
48 		free_sess(sess);
49 }
50 
51 static void rnbd_clt_put_dev(struct rnbd_clt_dev *dev)
52 {
53 	might_sleep();
54 
55 	if (!refcount_dec_and_test(&dev->refcount))
56 		return;
57 
58 	ida_free(&index_ida, dev->clt_device_id);
59 	kfree(dev->hw_queues);
60 	kfree(dev->pathname);
61 	rnbd_clt_put_sess(dev->sess);
62 	mutex_destroy(&dev->lock);
63 	kfree(dev);
64 }
65 
66 static inline bool rnbd_clt_get_dev(struct rnbd_clt_dev *dev)
67 {
68 	return refcount_inc_not_zero(&dev->refcount);
69 }
70 
71 static void rnbd_clt_change_capacity(struct rnbd_clt_dev *dev,
72 				    sector_t new_nsectors)
73 {
74 	if (get_capacity(dev->gd) == new_nsectors)
75 		return;
76 
77 	/*
78 	 * If the size changed, we need to revalidate it
79 	 */
80 	rnbd_clt_info(dev, "Device size changed from %llu to %llu sectors\n",
81 		      get_capacity(dev->gd), new_nsectors);
82 	set_capacity_and_notify(dev->gd, new_nsectors);
83 }
84 
85 static int process_msg_open_rsp(struct rnbd_clt_dev *dev,
86 				struct rnbd_msg_open_rsp *rsp)
87 {
88 	struct kobject *gd_kobj;
89 	int err = 0;
90 
91 	mutex_lock(&dev->lock);
92 	if (dev->dev_state == DEV_STATE_UNMAPPED) {
93 		rnbd_clt_info(dev,
94 			       "Ignoring Open-Response message from server for  unmapped device\n");
95 		err = -ENOENT;
96 		goto out;
97 	}
98 	if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED) {
99 		u64 nsectors = le64_to_cpu(rsp->nsectors);
100 
101 		rnbd_clt_change_capacity(dev, nsectors);
102 		gd_kobj = &disk_to_dev(dev->gd)->kobj;
103 		kobject_uevent(gd_kobj, KOBJ_ONLINE);
104 		rnbd_clt_info(dev, "Device online, device remapped successfully\n");
105 	}
106 	if (!rsp->logical_block_size) {
107 		err = -EINVAL;
108 		goto out;
109 	}
110 	dev->device_id = le32_to_cpu(rsp->device_id);
111 	dev->dev_state = DEV_STATE_MAPPED;
112 
113 out:
114 	mutex_unlock(&dev->lock);
115 
116 	return err;
117 }
118 
119 int rnbd_clt_resize_disk(struct rnbd_clt_dev *dev, sector_t newsize)
120 {
121 	int ret = 0;
122 
123 	mutex_lock(&dev->lock);
124 	if (dev->dev_state != DEV_STATE_MAPPED) {
125 		pr_err("Failed to set new size of the device, device is not opened\n");
126 		ret = -ENOENT;
127 		goto out;
128 	}
129 	rnbd_clt_change_capacity(dev, newsize);
130 
131 out:
132 	mutex_unlock(&dev->lock);
133 
134 	return ret;
135 }
136 
137 static inline void rnbd_clt_dev_requeue(struct rnbd_queue *q)
138 {
139 	if (WARN_ON(!q->hctx))
140 		return;
141 
142 	/* We can come here from interrupt, thus async=true */
143 	blk_mq_run_hw_queue(q->hctx, true);
144 }
145 
146 enum {
147 	RNBD_DELAY_IFBUSY = -1,
148 };
149 
150 /**
151  * rnbd_get_cpu_qlist() - finds a list with HW queues to be rerun
152  * @sess:	Session to find a queue for
153  * @cpu:	Cpu to start the search from
154  *
155  * Description:
156  *     Each CPU has a list of HW queues, which needs to be rerun.  If a list
157  *     is not empty - it is marked with a bit.  This function finds first
158  *     set bit in a bitmap and returns corresponding CPU list.
159  */
160 static struct rnbd_cpu_qlist *
161 rnbd_get_cpu_qlist(struct rnbd_clt_session *sess, int cpu)
162 {
163 	int bit;
164 
165 	/* Search from cpu to nr_cpu_ids */
166 	bit = find_next_bit(sess->cpu_queues_bm, nr_cpu_ids, cpu);
167 	if (bit < nr_cpu_ids) {
168 		return per_cpu_ptr(sess->cpu_queues, bit);
169 	} else if (cpu != 0) {
170 		/* Search from 0 to cpu */
171 		bit = find_first_bit(sess->cpu_queues_bm, cpu);
172 		if (bit < cpu)
173 			return per_cpu_ptr(sess->cpu_queues, bit);
174 	}
175 
176 	return NULL;
177 }
178 
179 static inline int nxt_cpu(int cpu)
180 {
181 	return (cpu + 1) % nr_cpu_ids;
182 }
183 
184 /**
185  * rnbd_rerun_if_needed() - rerun next queue marked as stopped
186  * @sess:	Session to rerun a queue on
187  *
188  * Description:
189  *     Each CPU has it's own list of HW queues, which should be rerun.
190  *     Function finds such list with HW queues, takes a list lock, picks up
191  *     the first HW queue out of the list and requeues it.
192  *
193  * Return:
194  *     True if the queue was requeued, false otherwise.
195  *
196  * Context:
197  *     Does not matter.
198  */
199 static bool rnbd_rerun_if_needed(struct rnbd_clt_session *sess)
200 {
201 	struct rnbd_queue *q = NULL;
202 	struct rnbd_cpu_qlist *cpu_q;
203 	unsigned long flags;
204 	int *cpup;
205 
206 	/*
207 	 * To keep fairness and not to let other queues starve we always
208 	 * try to wake up someone else in round-robin manner.  That of course
209 	 * increases latency but queues always have a chance to be executed.
210 	 */
211 	cpup = get_cpu_ptr(sess->cpu_rr);
212 	for (cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(*cpup)); cpu_q;
213 	     cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(cpu_q->cpu))) {
214 		if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
215 			continue;
216 		if (!test_bit(cpu_q->cpu, sess->cpu_queues_bm))
217 			goto unlock;
218 		q = list_first_entry_or_null(&cpu_q->requeue_list,
219 					     typeof(*q), requeue_list);
220 		if (WARN_ON(!q))
221 			goto clear_bit;
222 		list_del_init(&q->requeue_list);
223 		clear_bit_unlock(0, &q->in_list);
224 
225 		if (list_empty(&cpu_q->requeue_list)) {
226 			/* Clear bit if nothing is left */
227 clear_bit:
228 			clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
229 		}
230 unlock:
231 		spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
232 
233 		if (q)
234 			break;
235 	}
236 
237 	/**
238 	 * Saves the CPU that is going to be requeued on the per-cpu var. Just
239 	 * incrementing it doesn't work because rnbd_get_cpu_qlist() will
240 	 * always return the first CPU with something on the queue list when the
241 	 * value stored on the var is greater than the last CPU with something
242 	 * on the list.
243 	 */
244 	if (cpu_q)
245 		*cpup = cpu_q->cpu;
246 	put_cpu_ptr(sess->cpu_rr);
247 
248 	if (q)
249 		rnbd_clt_dev_requeue(q);
250 
251 	return q;
252 }
253 
254 /**
255  * rnbd_rerun_all_if_idle() - rerun all queues left in the list if
256  *				 session is idling (there are no requests
257  *				 in-flight).
258  * @sess:	Session to rerun the queues on
259  *
260  * Description:
261  *     This function tries to rerun all stopped queues if there are no
262  *     requests in-flight anymore.  This function tries to solve an obvious
263  *     problem, when number of tags < than number of queues (hctx), which
264  *     are stopped and put to sleep.  If last permit, which has been just put,
265  *     does not wake up all left queues (hctxs), IO requests hang forever.
266  *
267  *     That can happen when all number of permits, say N, have been exhausted
268  *     from one CPU, and we have many block devices per session, say M.
269  *     Each block device has it's own queue (hctx) for each CPU, so eventually
270  *     we can put that number of queues (hctxs) to sleep: M x nr_cpu_ids.
271  *     If number of permits N < M x nr_cpu_ids finally we will get an IO hang.
272  *
273  *     To avoid this hang last caller of rnbd_put_permit() (last caller is the
274  *     one who observes sess->busy == 0) must wake up all remaining queues.
275  *
276  * Context:
277  *     Does not matter.
278  */
279 static void rnbd_rerun_all_if_idle(struct rnbd_clt_session *sess)
280 {
281 	bool requeued;
282 
283 	do {
284 		requeued = rnbd_rerun_if_needed(sess);
285 	} while (atomic_read(&sess->busy) == 0 && requeued);
286 }
287 
288 static struct rtrs_permit *rnbd_get_permit(struct rnbd_clt_session *sess,
289 					     enum rtrs_clt_con_type con_type,
290 					     enum wait_type wait)
291 {
292 	struct rtrs_permit *permit;
293 
294 	permit = rtrs_clt_get_permit(sess->rtrs, con_type, wait);
295 	if (permit)
296 		/* We have a subtle rare case here, when all permits can be
297 		 * consumed before busy counter increased.  This is safe,
298 		 * because loser will get NULL as a permit, observe 0 busy
299 		 * counter and immediately restart the queue himself.
300 		 */
301 		atomic_inc(&sess->busy);
302 
303 	return permit;
304 }
305 
306 static void rnbd_put_permit(struct rnbd_clt_session *sess,
307 			     struct rtrs_permit *permit)
308 {
309 	rtrs_clt_put_permit(sess->rtrs, permit);
310 	atomic_dec(&sess->busy);
311 	/* Paired with rnbd_clt_dev_add_to_requeue().  Decrement first
312 	 * and then check queue bits.
313 	 */
314 	smp_mb__after_atomic();
315 	rnbd_rerun_all_if_idle(sess);
316 }
317 
318 static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
319 				     enum rtrs_clt_con_type con_type,
320 				     enum wait_type wait)
321 {
322 	struct rnbd_iu *iu;
323 	struct rtrs_permit *permit;
324 
325 	iu = kzalloc(sizeof(*iu), GFP_KERNEL);
326 	if (!iu)
327 		return NULL;
328 
329 	permit = rnbd_get_permit(sess, con_type, wait);
330 	if (!permit) {
331 		kfree(iu);
332 		return NULL;
333 	}
334 
335 	iu->permit = permit;
336 	/*
337 	 * 1st reference is dropped after finishing sending a "user" message,
338 	 * 2nd reference is dropped after confirmation with the response is
339 	 * returned.
340 	 * 1st and 2nd can happen in any order, so the rnbd_iu should be
341 	 * released (rtrs_permit returned to rtrs) only after both
342 	 * are finished.
343 	 */
344 	atomic_set(&iu->refcount, 2);
345 	init_waitqueue_head(&iu->comp.wait);
346 	iu->comp.errno = INT_MAX;
347 
348 	if (sg_alloc_table(&iu->sgt, 1, GFP_KERNEL)) {
349 		rnbd_put_permit(sess, permit);
350 		kfree(iu);
351 		return NULL;
352 	}
353 
354 	return iu;
355 }
356 
357 static void rnbd_put_iu(struct rnbd_clt_session *sess, struct rnbd_iu *iu)
358 {
359 	if (atomic_dec_and_test(&iu->refcount)) {
360 		sg_free_table(&iu->sgt);
361 		rnbd_put_permit(sess, iu->permit);
362 		kfree(iu);
363 	}
364 }
365 
366 static void rnbd_softirq_done_fn(struct request *rq)
367 {
368 	struct rnbd_clt_dev *dev	= rq->q->disk->private_data;
369 	struct rnbd_clt_session *sess	= dev->sess;
370 	struct rnbd_iu *iu;
371 
372 	iu = blk_mq_rq_to_pdu(rq);
373 	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
374 	rnbd_put_permit(sess, iu->permit);
375 	blk_mq_end_request(rq, errno_to_blk_status(iu->errno));
376 }
377 
378 static void msg_io_conf(void *priv, int errno)
379 {
380 	struct rnbd_iu *iu = priv;
381 	struct rnbd_clt_dev *dev = iu->dev;
382 	struct request *rq = iu->rq;
383 	int rw = rq_data_dir(rq);
384 
385 	iu->errno = errno;
386 
387 	blk_mq_complete_request(rq);
388 
389 	if (errno)
390 		rnbd_clt_info_rl(dev, "%s I/O failed with err: %d\n",
391 				 rw == READ ? "read" : "write", errno);
392 }
393 
394 static void wake_up_iu_comp(struct rnbd_iu *iu, int errno)
395 {
396 	iu->comp.errno = errno;
397 	wake_up(&iu->comp.wait);
398 }
399 
400 static void msg_conf(void *priv, int errno)
401 {
402 	struct rnbd_iu *iu = priv;
403 
404 	iu->errno = errno;
405 	schedule_work(&iu->work);
406 }
407 
408 static int send_usr_msg(struct rtrs_clt_sess *rtrs, int dir,
409 			struct rnbd_iu *iu, struct kvec *vec,
410 			size_t len, struct scatterlist *sg, unsigned int sg_len,
411 			void (*conf)(struct work_struct *work),
412 			int *errno, int wait)
413 {
414 	int err;
415 	struct rtrs_clt_req_ops req_ops;
416 
417 	INIT_WORK(&iu->work, conf);
418 	req_ops = (struct rtrs_clt_req_ops) {
419 		.priv = iu,
420 		.conf_fn = msg_conf,
421 	};
422 	err = rtrs_clt_request(dir, &req_ops, rtrs, iu->permit,
423 				vec, 1, len, sg, sg_len);
424 	if (!err && wait) {
425 		wait_event(iu->comp.wait, iu->comp.errno != INT_MAX);
426 		*errno = iu->comp.errno;
427 	} else {
428 		*errno = 0;
429 	}
430 
431 	return err;
432 }
433 
434 static void msg_close_conf(struct work_struct *work)
435 {
436 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
437 	struct rnbd_clt_dev *dev = iu->dev;
438 
439 	wake_up_iu_comp(iu, iu->errno);
440 	rnbd_put_iu(dev->sess, iu);
441 	rnbd_clt_put_dev(dev);
442 }
443 
444 static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id,
445 			  enum wait_type wait)
446 {
447 	struct rnbd_clt_session *sess = dev->sess;
448 	struct rnbd_msg_close msg;
449 	struct rnbd_iu *iu;
450 	struct kvec vec = {
451 		.iov_base = &msg,
452 		.iov_len  = sizeof(msg)
453 	};
454 	int err, errno;
455 
456 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
457 	if (!iu)
458 		return -ENOMEM;
459 
460 	iu->buf = NULL;
461 	iu->dev = dev;
462 
463 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_CLOSE);
464 	msg.device_id	= cpu_to_le32(device_id);
465 
466 	WARN_ON(!rnbd_clt_get_dev(dev));
467 	err = send_usr_msg(sess->rtrs, WRITE, iu, &vec, 0, NULL, 0,
468 			   msg_close_conf, &errno, wait);
469 	if (err) {
470 		rnbd_clt_put_dev(dev);
471 		rnbd_put_iu(sess, iu);
472 	} else {
473 		err = errno;
474 	}
475 
476 	rnbd_put_iu(sess, iu);
477 	return err;
478 }
479 
480 static void msg_open_conf(struct work_struct *work)
481 {
482 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
483 	struct rnbd_msg_open_rsp *rsp = iu->buf;
484 	struct rnbd_clt_dev *dev = iu->dev;
485 	int errno = iu->errno;
486 	bool from_map = false;
487 
488 	/* INIT state is only triggered from rnbd_clt_map_device */
489 	if (dev->dev_state == DEV_STATE_INIT)
490 		from_map = true;
491 
492 	if (errno) {
493 		rnbd_clt_err(dev,
494 			      "Opening failed, server responded: %d\n",
495 			      errno);
496 	} else {
497 		errno = process_msg_open_rsp(dev, rsp);
498 		if (errno) {
499 			u32 device_id = le32_to_cpu(rsp->device_id);
500 			/*
501 			 * If server thinks its fine, but we fail to process
502 			 * then be nice and send a close to server.
503 			 */
504 			send_msg_close(dev, device_id, RTRS_PERMIT_NOWAIT);
505 		}
506 	}
507 	/* We free rsp in rnbd_clt_map_device for map scenario */
508 	if (!from_map)
509 		kfree(rsp);
510 	wake_up_iu_comp(iu, errno);
511 	rnbd_put_iu(dev->sess, iu);
512 	rnbd_clt_put_dev(dev);
513 }
514 
515 static void msg_sess_info_conf(struct work_struct *work)
516 {
517 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
518 	struct rnbd_msg_sess_info_rsp *rsp = iu->buf;
519 	struct rnbd_clt_session *sess = iu->sess;
520 
521 	if (!iu->errno)
522 		sess->ver = min_t(u8, rsp->ver, RNBD_PROTO_VER_MAJOR);
523 
524 	kfree(rsp);
525 	wake_up_iu_comp(iu, iu->errno);
526 	rnbd_put_iu(sess, iu);
527 	rnbd_clt_put_sess(sess);
528 }
529 
530 static int send_msg_open(struct rnbd_clt_dev *dev, enum wait_type wait)
531 {
532 	struct rnbd_clt_session *sess = dev->sess;
533 	struct rnbd_msg_open_rsp *rsp;
534 	struct rnbd_msg_open msg;
535 	struct rnbd_iu *iu;
536 	struct kvec vec = {
537 		.iov_base = &msg,
538 		.iov_len  = sizeof(msg)
539 	};
540 	int err, errno;
541 
542 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
543 	if (!rsp)
544 		return -ENOMEM;
545 
546 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
547 	if (!iu) {
548 		kfree(rsp);
549 		return -ENOMEM;
550 	}
551 
552 	iu->buf = rsp;
553 	iu->dev = dev;
554 
555 	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
556 
557 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_OPEN);
558 	msg.access_mode	= dev->access_mode;
559 	strscpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
560 
561 	WARN_ON(!rnbd_clt_get_dev(dev));
562 	err = send_usr_msg(sess->rtrs, READ, iu,
563 			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
564 			   msg_open_conf, &errno, wait);
565 	if (err) {
566 		rnbd_clt_put_dev(dev);
567 		rnbd_put_iu(sess, iu);
568 		kfree(rsp);
569 	} else {
570 		err = errno;
571 	}
572 
573 	rnbd_put_iu(sess, iu);
574 	return err;
575 }
576 
577 static int send_msg_sess_info(struct rnbd_clt_session *sess, enum wait_type wait)
578 {
579 	struct rnbd_msg_sess_info_rsp *rsp;
580 	struct rnbd_msg_sess_info msg;
581 	struct rnbd_iu *iu;
582 	struct kvec vec = {
583 		.iov_base = &msg,
584 		.iov_len  = sizeof(msg)
585 	};
586 	int err, errno;
587 
588 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
589 	if (!rsp)
590 		return -ENOMEM;
591 
592 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
593 	if (!iu) {
594 		kfree(rsp);
595 		return -ENOMEM;
596 	}
597 
598 	iu->buf = rsp;
599 	iu->sess = sess;
600 	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
601 
602 	msg.hdr.type = cpu_to_le16(RNBD_MSG_SESS_INFO);
603 	msg.ver      = RNBD_PROTO_VER_MAJOR;
604 
605 	if (!rnbd_clt_get_sess(sess)) {
606 		/*
607 		 * That can happen only in one case, when RTRS has restablished
608 		 * the connection and link_ev() is called, but session is almost
609 		 * dead, last reference on session is put and caller is waiting
610 		 * for RTRS to close everything.
611 		 */
612 		err = -ENODEV;
613 		goto put_iu;
614 	}
615 	err = send_usr_msg(sess->rtrs, READ, iu,
616 			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
617 			   msg_sess_info_conf, &errno, wait);
618 	if (err) {
619 		rnbd_clt_put_sess(sess);
620 put_iu:
621 		rnbd_put_iu(sess, iu);
622 		kfree(rsp);
623 	} else {
624 		err = errno;
625 	}
626 	rnbd_put_iu(sess, iu);
627 	return err;
628 }
629 
630 static void set_dev_states_to_disconnected(struct rnbd_clt_session *sess)
631 {
632 	struct rnbd_clt_dev *dev;
633 	struct kobject *gd_kobj;
634 
635 	mutex_lock(&sess->lock);
636 	list_for_each_entry(dev, &sess->devs_list, list) {
637 		rnbd_clt_err(dev, "Device disconnected.\n");
638 
639 		mutex_lock(&dev->lock);
640 		if (dev->dev_state == DEV_STATE_MAPPED) {
641 			dev->dev_state = DEV_STATE_MAPPED_DISCONNECTED;
642 			gd_kobj = &disk_to_dev(dev->gd)->kobj;
643 			kobject_uevent(gd_kobj, KOBJ_OFFLINE);
644 		}
645 		mutex_unlock(&dev->lock);
646 	}
647 	mutex_unlock(&sess->lock);
648 }
649 
650 static void remap_devs(struct rnbd_clt_session *sess)
651 {
652 	struct rnbd_clt_dev *dev;
653 	struct rtrs_attrs attrs;
654 	int err;
655 
656 	/*
657 	 * Careful here: we are called from RTRS link event directly,
658 	 * thus we can't send any RTRS request and wait for response
659 	 * or RTRS will not be able to complete request with failure
660 	 * if something goes wrong (failing of outstanding requests
661 	 * happens exactly from the context where we are blocking now).
662 	 *
663 	 * So to avoid deadlocks each usr message sent from here must
664 	 * be asynchronous.
665 	 */
666 
667 	err = send_msg_sess_info(sess, RTRS_PERMIT_NOWAIT);
668 	if (err) {
669 		pr_err("send_msg_sess_info(\"%s\"): %d\n", sess->sessname, err);
670 		return;
671 	}
672 
673 	err = rtrs_clt_query(sess->rtrs, &attrs);
674 	if (err) {
675 		pr_err("rtrs_clt_query(\"%s\"): %d\n", sess->sessname, err);
676 		return;
677 	}
678 	mutex_lock(&sess->lock);
679 	sess->max_io_size = attrs.max_io_size;
680 
681 	list_for_each_entry(dev, &sess->devs_list, list) {
682 		bool skip;
683 
684 		mutex_lock(&dev->lock);
685 		skip = (dev->dev_state == DEV_STATE_INIT);
686 		mutex_unlock(&dev->lock);
687 		if (skip)
688 			/*
689 			 * When device is establishing connection for the first
690 			 * time - do not remap, it will be closed soon.
691 			 */
692 			continue;
693 
694 		rnbd_clt_info(dev, "session reconnected, remapping device\n");
695 		err = send_msg_open(dev, RTRS_PERMIT_NOWAIT);
696 		if (err) {
697 			rnbd_clt_err(dev, "send_msg_open(): %d\n", err);
698 			break;
699 		}
700 	}
701 	mutex_unlock(&sess->lock);
702 }
703 
704 static void rnbd_clt_link_ev(void *priv, enum rtrs_clt_link_ev ev)
705 {
706 	struct rnbd_clt_session *sess = priv;
707 
708 	switch (ev) {
709 	case RTRS_CLT_LINK_EV_DISCONNECTED:
710 		set_dev_states_to_disconnected(sess);
711 		break;
712 	case RTRS_CLT_LINK_EV_RECONNECTED:
713 		remap_devs(sess);
714 		break;
715 	default:
716 		pr_err("Unknown session event received (%d), session: %s\n",
717 		       ev, sess->sessname);
718 	}
719 }
720 
721 static void rnbd_init_cpu_qlists(struct rnbd_cpu_qlist __percpu *cpu_queues)
722 {
723 	unsigned int cpu;
724 	struct rnbd_cpu_qlist *cpu_q;
725 
726 	for_each_possible_cpu(cpu) {
727 		cpu_q = per_cpu_ptr(cpu_queues, cpu);
728 
729 		cpu_q->cpu = cpu;
730 		INIT_LIST_HEAD(&cpu_q->requeue_list);
731 		spin_lock_init(&cpu_q->requeue_lock);
732 	}
733 }
734 
735 static void destroy_mq_tags(struct rnbd_clt_session *sess)
736 {
737 	if (sess->tag_set.tags)
738 		blk_mq_free_tag_set(&sess->tag_set);
739 }
740 
741 static inline void wake_up_rtrs_waiters(struct rnbd_clt_session *sess)
742 {
743 	sess->rtrs_ready = true;
744 	wake_up_all(&sess->rtrs_waitq);
745 }
746 
747 static void close_rtrs(struct rnbd_clt_session *sess)
748 {
749 	might_sleep();
750 
751 	if (!IS_ERR_OR_NULL(sess->rtrs)) {
752 		rtrs_clt_close(sess->rtrs);
753 		sess->rtrs = NULL;
754 		wake_up_rtrs_waiters(sess);
755 	}
756 }
757 
758 static void free_sess(struct rnbd_clt_session *sess)
759 {
760 	WARN_ON(!list_empty(&sess->devs_list));
761 
762 	might_sleep();
763 
764 	close_rtrs(sess);
765 	destroy_mq_tags(sess);
766 	if (!list_empty(&sess->list)) {
767 		mutex_lock(&sess_lock);
768 		list_del(&sess->list);
769 		mutex_unlock(&sess_lock);
770 	}
771 	free_percpu(sess->cpu_queues);
772 	free_percpu(sess->cpu_rr);
773 	mutex_destroy(&sess->lock);
774 	kfree(sess);
775 }
776 
777 static struct rnbd_clt_session *alloc_sess(const char *sessname)
778 {
779 	struct rnbd_clt_session *sess;
780 	int err, cpu;
781 
782 	sess = kzalloc_node(sizeof(*sess), GFP_KERNEL, NUMA_NO_NODE);
783 	if (!sess)
784 		return ERR_PTR(-ENOMEM);
785 	strscpy(sess->sessname, sessname, sizeof(sess->sessname));
786 	atomic_set(&sess->busy, 0);
787 	mutex_init(&sess->lock);
788 	INIT_LIST_HEAD(&sess->devs_list);
789 	INIT_LIST_HEAD(&sess->list);
790 	bitmap_zero(sess->cpu_queues_bm, num_possible_cpus());
791 	init_waitqueue_head(&sess->rtrs_waitq);
792 	refcount_set(&sess->refcount, 1);
793 
794 	sess->cpu_queues = alloc_percpu(struct rnbd_cpu_qlist);
795 	if (!sess->cpu_queues) {
796 		err = -ENOMEM;
797 		goto err;
798 	}
799 	rnbd_init_cpu_qlists(sess->cpu_queues);
800 
801 	/*
802 	 * That is simple percpu variable which stores cpu indices, which are
803 	 * incremented on each access.  We need that for the sake of fairness
804 	 * to wake up queues in a round-robin manner.
805 	 */
806 	sess->cpu_rr = alloc_percpu(int);
807 	if (!sess->cpu_rr) {
808 		err = -ENOMEM;
809 		goto err;
810 	}
811 	for_each_possible_cpu(cpu)
812 		* per_cpu_ptr(sess->cpu_rr, cpu) = cpu;
813 
814 	return sess;
815 
816 err:
817 	free_sess(sess);
818 
819 	return ERR_PTR(err);
820 }
821 
822 static int wait_for_rtrs_connection(struct rnbd_clt_session *sess)
823 {
824 	wait_event(sess->rtrs_waitq, sess->rtrs_ready);
825 	if (IS_ERR_OR_NULL(sess->rtrs))
826 		return -ECONNRESET;
827 
828 	return 0;
829 }
830 
831 static void wait_for_rtrs_disconnection(struct rnbd_clt_session *sess)
832 	__releases(&sess_lock)
833 	__acquires(&sess_lock)
834 {
835 	DEFINE_WAIT(wait);
836 
837 	prepare_to_wait(&sess->rtrs_waitq, &wait, TASK_UNINTERRUPTIBLE);
838 	if (IS_ERR_OR_NULL(sess->rtrs)) {
839 		finish_wait(&sess->rtrs_waitq, &wait);
840 		return;
841 	}
842 	mutex_unlock(&sess_lock);
843 	/* loop in caller, see __find_and_get_sess().
844 	 * You can't leave mutex locked and call schedule(), you will catch a
845 	 * deadlock with a caller of free_sess(), which has just put the last
846 	 * reference and is about to take the sess_lock in order to delete
847 	 * the session from the list.
848 	 */
849 	schedule();
850 	mutex_lock(&sess_lock);
851 }
852 
853 static struct rnbd_clt_session *__find_and_get_sess(const char *sessname)
854 	__releases(&sess_lock)
855 	__acquires(&sess_lock)
856 {
857 	struct rnbd_clt_session *sess, *sn;
858 	int err;
859 
860 again:
861 	list_for_each_entry_safe(sess, sn, &sess_list, list) {
862 		if (strcmp(sessname, sess->sessname))
863 			continue;
864 
865 		if (sess->rtrs_ready && IS_ERR_OR_NULL(sess->rtrs))
866 			/*
867 			 * No RTRS connection, session is dying.
868 			 */
869 			continue;
870 
871 		if (rnbd_clt_get_sess(sess)) {
872 			/*
873 			 * Alive session is found, wait for RTRS connection.
874 			 */
875 			mutex_unlock(&sess_lock);
876 			err = wait_for_rtrs_connection(sess);
877 			if (err)
878 				rnbd_clt_put_sess(sess);
879 			mutex_lock(&sess_lock);
880 
881 			if (err)
882 				/* Session is dying, repeat the loop */
883 				goto again;
884 
885 			return sess;
886 		}
887 		/*
888 		 * Ref is 0, session is dying, wait for RTRS disconnect
889 		 * in order to avoid session names clashes.
890 		 */
891 		wait_for_rtrs_disconnection(sess);
892 		/*
893 		 * RTRS is disconnected and soon session will be freed,
894 		 * so repeat a loop.
895 		 */
896 		goto again;
897 	}
898 
899 	return NULL;
900 }
901 
902 /* caller is responsible for initializing 'first' to false */
903 static struct
904 rnbd_clt_session *find_or_create_sess(const char *sessname, bool *first)
905 {
906 	struct rnbd_clt_session *sess = NULL;
907 
908 	mutex_lock(&sess_lock);
909 	sess = __find_and_get_sess(sessname);
910 	if (!sess) {
911 		sess = alloc_sess(sessname);
912 		if (IS_ERR(sess)) {
913 			mutex_unlock(&sess_lock);
914 			return sess;
915 		}
916 		list_add(&sess->list, &sess_list);
917 		*first = true;
918 	}
919 	mutex_unlock(&sess_lock);
920 
921 	return sess;
922 }
923 
924 static int rnbd_client_open(struct gendisk *disk, blk_mode_t mode)
925 {
926 	struct rnbd_clt_dev *dev = disk->private_data;
927 
928 	if (get_disk_ro(dev->gd) && (mode & BLK_OPEN_WRITE))
929 		return -EPERM;
930 
931 	if (dev->dev_state == DEV_STATE_UNMAPPED ||
932 	    !rnbd_clt_get_dev(dev))
933 		return -EIO;
934 
935 	return 0;
936 }
937 
938 static void rnbd_client_release(struct gendisk *gen)
939 {
940 	struct rnbd_clt_dev *dev = gen->private_data;
941 
942 	rnbd_clt_put_dev(dev);
943 }
944 
945 static int rnbd_client_getgeo(struct block_device *block_device,
946 			      struct hd_geometry *geo)
947 {
948 	u64 size;
949 	struct rnbd_clt_dev *dev = block_device->bd_disk->private_data;
950 	struct queue_limits *limit = &dev->queue->limits;
951 
952 	size = dev->size * (limit->logical_block_size / SECTOR_SIZE);
953 	geo->cylinders	= size >> 6;	/* size/64 */
954 	geo->heads	= 4;
955 	geo->sectors	= 16;
956 	geo->start	= 0;
957 
958 	return 0;
959 }
960 
961 static const struct block_device_operations rnbd_client_ops = {
962 	.owner		= THIS_MODULE,
963 	.open		= rnbd_client_open,
964 	.release	= rnbd_client_release,
965 	.getgeo		= rnbd_client_getgeo
966 };
967 
968 /* The amount of data that belongs to an I/O and the amount of data that
969  * should be read or written to the disk (bi_size) can differ.
970  *
971  * E.g. When WRITE_SAME is used, only a small amount of data is
972  * transferred that is then written repeatedly over a lot of sectors.
973  *
974  * Get the size of data to be transferred via RTRS by summing up the size
975  * of the scather-gather list entries.
976  */
977 static size_t rnbd_clt_get_sg_size(struct scatterlist *sglist, u32 len)
978 {
979 	struct scatterlist *sg;
980 	size_t tsize = 0;
981 	int i;
982 
983 	for_each_sg(sglist, sg, len, i)
984 		tsize += sg->length;
985 	return tsize;
986 }
987 
988 static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
989 				     struct request *rq,
990 				     struct rnbd_iu *iu)
991 {
992 	struct rtrs_clt_sess *rtrs = dev->sess->rtrs;
993 	struct rtrs_permit *permit = iu->permit;
994 	struct rnbd_msg_io msg;
995 	struct rtrs_clt_req_ops req_ops;
996 	unsigned int sg_cnt = 0;
997 	struct kvec vec;
998 	size_t size;
999 	int err;
1000 
1001 	iu->rq		= rq;
1002 	iu->dev		= dev;
1003 	msg.sector	= cpu_to_le64(blk_rq_pos(rq));
1004 	msg.bi_size	= cpu_to_le32(blk_rq_bytes(rq));
1005 	msg.rw		= cpu_to_le32(rq_to_rnbd_flags(rq));
1006 	msg.prio	= cpu_to_le16(req_get_ioprio(rq));
1007 
1008 	/*
1009 	 * We only support discards with single segment for now.
1010 	 * See queue limits.
1011 	 */
1012 	if (req_op(rq) != REQ_OP_DISCARD)
1013 		sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sgt.sgl);
1014 
1015 	if (sg_cnt == 0)
1016 		sg_mark_end(&iu->sgt.sgl[0]);
1017 
1018 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_IO);
1019 	msg.device_id	= cpu_to_le32(dev->device_id);
1020 
1021 	vec = (struct kvec) {
1022 		.iov_base = &msg,
1023 		.iov_len  = sizeof(msg)
1024 	};
1025 	size = rnbd_clt_get_sg_size(iu->sgt.sgl, sg_cnt);
1026 	req_ops = (struct rtrs_clt_req_ops) {
1027 		.priv = iu,
1028 		.conf_fn = msg_io_conf,
1029 	};
1030 	err = rtrs_clt_request(rq_data_dir(rq), &req_ops, rtrs, permit,
1031 			       &vec, 1, size, iu->sgt.sgl, sg_cnt);
1032 	if (err) {
1033 		rnbd_clt_err_rl(dev, "RTRS failed to transfer IO, err: %d\n",
1034 				 err);
1035 		return err;
1036 	}
1037 
1038 	return 0;
1039 }
1040 
1041 /**
1042  * rnbd_clt_dev_add_to_requeue() - add device to requeue if session is busy
1043  * @dev:	Device to be checked
1044  * @q:		Queue to be added to the requeue list if required
1045  *
1046  * Description:
1047  *     If session is busy, that means someone will requeue us when resources
1048  *     are freed.  If session is not doing anything - device is not added to
1049  *     the list and @false is returned.
1050  */
1051 static bool rnbd_clt_dev_add_to_requeue(struct rnbd_clt_dev *dev,
1052 						struct rnbd_queue *q)
1053 {
1054 	struct rnbd_clt_session *sess = dev->sess;
1055 	struct rnbd_cpu_qlist *cpu_q;
1056 	unsigned long flags;
1057 	bool added = true;
1058 	bool need_set;
1059 
1060 	cpu_q = get_cpu_ptr(sess->cpu_queues);
1061 	spin_lock_irqsave(&cpu_q->requeue_lock, flags);
1062 
1063 	if (!test_and_set_bit_lock(0, &q->in_list)) {
1064 		if (WARN_ON(!list_empty(&q->requeue_list)))
1065 			goto unlock;
1066 
1067 		need_set = !test_bit(cpu_q->cpu, sess->cpu_queues_bm);
1068 		if (need_set) {
1069 			set_bit(cpu_q->cpu, sess->cpu_queues_bm);
1070 			/* Paired with rnbd_put_permit(). Set a bit first
1071 			 * and then observe the busy counter.
1072 			 */
1073 			smp_mb__before_atomic();
1074 		}
1075 		if (atomic_read(&sess->busy)) {
1076 			list_add_tail(&q->requeue_list, &cpu_q->requeue_list);
1077 		} else {
1078 			/* Very unlikely, but possible: busy counter was
1079 			 * observed as zero.  Drop all bits and return
1080 			 * false to restart the queue by ourselves.
1081 			 */
1082 			if (need_set)
1083 				clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
1084 			clear_bit_unlock(0, &q->in_list);
1085 			added = false;
1086 		}
1087 	}
1088 unlock:
1089 	spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
1090 	put_cpu_ptr(sess->cpu_queues);
1091 
1092 	return added;
1093 }
1094 
1095 static void rnbd_clt_dev_kick_mq_queue(struct rnbd_clt_dev *dev,
1096 					struct blk_mq_hw_ctx *hctx,
1097 					int delay)
1098 {
1099 	struct rnbd_queue *q = hctx->driver_data;
1100 
1101 	if (delay != RNBD_DELAY_IFBUSY)
1102 		blk_mq_delay_run_hw_queue(hctx, delay);
1103 	else if (!rnbd_clt_dev_add_to_requeue(dev, q))
1104 		/*
1105 		 * If session is not busy we have to restart
1106 		 * the queue ourselves.
1107 		 */
1108 		blk_mq_delay_run_hw_queue(hctx, 10/*ms*/);
1109 }
1110 
1111 static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1112 				   const struct blk_mq_queue_data *bd)
1113 {
1114 	struct request *rq = bd->rq;
1115 	struct rnbd_clt_dev *dev = rq->q->disk->private_data;
1116 	struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
1117 	int err;
1118 	blk_status_t ret = BLK_STS_IOERR;
1119 
1120 	if (dev->dev_state != DEV_STATE_MAPPED)
1121 		return BLK_STS_IOERR;
1122 
1123 	iu->permit = rnbd_get_permit(dev->sess, RTRS_IO_CON,
1124 				      RTRS_PERMIT_NOWAIT);
1125 	if (!iu->permit) {
1126 		rnbd_clt_dev_kick_mq_queue(dev, hctx, RNBD_DELAY_IFBUSY);
1127 		return BLK_STS_RESOURCE;
1128 	}
1129 
1130 	iu->sgt.sgl = iu->first_sgl;
1131 	err = sg_alloc_table_chained(&iu->sgt,
1132 				     /* Even-if the request has no segment,
1133 				      * sglist must have one entry at least.
1134 				      */
1135 				     blk_rq_nr_phys_segments(rq) ? : 1,
1136 				     iu->sgt.sgl,
1137 				     RNBD_INLINE_SG_CNT);
1138 	if (err) {
1139 		rnbd_clt_err_rl(dev, "sg_alloc_table_chained ret=%d\n", err);
1140 		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1141 		rnbd_put_permit(dev->sess, iu->permit);
1142 		return BLK_STS_RESOURCE;
1143 	}
1144 
1145 	blk_mq_start_request(rq);
1146 	err = rnbd_client_xfer_request(dev, rq, iu);
1147 	if (err == 0)
1148 		return BLK_STS_OK;
1149 	if (err == -EAGAIN || err == -ENOMEM) {
1150 		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1151 		ret = BLK_STS_RESOURCE;
1152 	}
1153 	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
1154 	rnbd_put_permit(dev->sess, iu->permit);
1155 	return ret;
1156 }
1157 
1158 static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1159 {
1160 	struct rnbd_queue *q = hctx->driver_data;
1161 	struct rnbd_clt_dev *dev = q->dev;
1162 
1163 	return rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
1164 }
1165 
1166 static void rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
1167 {
1168 	struct rnbd_clt_session *sess = set->driver_data;
1169 
1170 	/* shared read/write queues */
1171 	set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
1172 	set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1173 	set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
1174 	set->map[HCTX_TYPE_READ].queue_offset = 0;
1175 	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1176 	blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
1177 
1178 	if (sess->nr_poll_queues) {
1179 		/* dedicated queue for poll */
1180 		set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
1181 		set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
1182 			set->map[HCTX_TYPE_READ].nr_queues;
1183 		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1184 		pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
1185 			sess->sessname,
1186 			set->map[HCTX_TYPE_DEFAULT].nr_queues,
1187 			set->map[HCTX_TYPE_READ].nr_queues,
1188 			set->map[HCTX_TYPE_POLL].nr_queues);
1189 	} else {
1190 		pr_info("[session=%s] mapped %d/%d default/read queues.\n",
1191 			sess->sessname,
1192 			set->map[HCTX_TYPE_DEFAULT].nr_queues,
1193 			set->map[HCTX_TYPE_READ].nr_queues);
1194 	}
1195 }
1196 
1197 static struct blk_mq_ops rnbd_mq_ops = {
1198 	.queue_rq	= rnbd_queue_rq,
1199 	.complete	= rnbd_softirq_done_fn,
1200 	.map_queues     = rnbd_rdma_map_queues,
1201 	.poll           = rnbd_rdma_poll,
1202 };
1203 
1204 static int setup_mq_tags(struct rnbd_clt_session *sess)
1205 {
1206 	struct blk_mq_tag_set *tag_set = &sess->tag_set;
1207 
1208 	memset(tag_set, 0, sizeof(*tag_set));
1209 	tag_set->ops		= &rnbd_mq_ops;
1210 	tag_set->queue_depth	= sess->queue_depth;
1211 	tag_set->numa_node		= NUMA_NO_NODE;
1212 	tag_set->flags		= BLK_MQ_F_SHOULD_MERGE |
1213 				  BLK_MQ_F_TAG_QUEUE_SHARED;
1214 	tag_set->cmd_size	= sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
1215 
1216 	/* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
1217 	tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1218 	/*
1219 	 * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
1220 	 * others are for HCTX_TYPE_POLL
1221 	 */
1222 	tag_set->nr_hw_queues	= num_online_cpus() + sess->nr_poll_queues;
1223 	tag_set->driver_data    = sess;
1224 
1225 	return blk_mq_alloc_tag_set(tag_set);
1226 }
1227 
1228 static struct rnbd_clt_session *
1229 find_and_get_or_create_sess(const char *sessname,
1230 			    const struct rtrs_addr *paths,
1231 			    size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
1232 {
1233 	struct rnbd_clt_session *sess;
1234 	struct rtrs_attrs attrs;
1235 	int err;
1236 	bool first = false;
1237 	struct rtrs_clt_ops rtrs_ops;
1238 
1239 	sess = find_or_create_sess(sessname, &first);
1240 	if (sess == ERR_PTR(-ENOMEM)) {
1241 		return ERR_PTR(-ENOMEM);
1242 	} else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
1243 		/*
1244 		 * A device MUST have its own session to use the polling-mode.
1245 		 * It must fail to map new device with the same session.
1246 		 */
1247 		err = -EINVAL;
1248 		goto put_sess;
1249 	}
1250 
1251 	if (!first)
1252 		return sess;
1253 
1254 	if (!path_cnt) {
1255 		pr_err("Session %s not found, and path parameter not given", sessname);
1256 		err = -ENXIO;
1257 		goto put_sess;
1258 	}
1259 
1260 	rtrs_ops = (struct rtrs_clt_ops) {
1261 		.priv = sess,
1262 		.link_ev = rnbd_clt_link_ev,
1263 	};
1264 	/*
1265 	 * Nothing was found, establish rtrs connection and proceed further.
1266 	 */
1267 	sess->rtrs = rtrs_clt_open(&rtrs_ops, sessname,
1268 				   paths, path_cnt, port_nr,
1269 				   0, /* Do not use pdu of rtrs */
1270 				   RECONNECT_DELAY,
1271 				   MAX_RECONNECTS, nr_poll_queues);
1272 	if (IS_ERR(sess->rtrs)) {
1273 		err = PTR_ERR(sess->rtrs);
1274 		goto wake_up_and_put;
1275 	}
1276 
1277 	err = rtrs_clt_query(sess->rtrs, &attrs);
1278 	if (err)
1279 		goto close_rtrs;
1280 
1281 	sess->max_io_size = attrs.max_io_size;
1282 	sess->queue_depth = attrs.queue_depth;
1283 	sess->nr_poll_queues = nr_poll_queues;
1284 	sess->max_segments = attrs.max_segments;
1285 
1286 	err = setup_mq_tags(sess);
1287 	if (err)
1288 		goto close_rtrs;
1289 
1290 	err = send_msg_sess_info(sess, RTRS_PERMIT_WAIT);
1291 	if (err)
1292 		goto close_rtrs;
1293 
1294 	wake_up_rtrs_waiters(sess);
1295 
1296 	return sess;
1297 
1298 close_rtrs:
1299 	close_rtrs(sess);
1300 put_sess:
1301 	rnbd_clt_put_sess(sess);
1302 
1303 	return ERR_PTR(err);
1304 
1305 wake_up_and_put:
1306 	wake_up_rtrs_waiters(sess);
1307 	goto put_sess;
1308 }
1309 
1310 static inline void rnbd_init_hw_queue(struct rnbd_clt_dev *dev,
1311 				       struct rnbd_queue *q,
1312 				       struct blk_mq_hw_ctx *hctx)
1313 {
1314 	INIT_LIST_HEAD(&q->requeue_list);
1315 	q->dev  = dev;
1316 	q->hctx = hctx;
1317 }
1318 
1319 static void rnbd_init_mq_hw_queues(struct rnbd_clt_dev *dev)
1320 {
1321 	unsigned long i;
1322 	struct blk_mq_hw_ctx *hctx;
1323 	struct rnbd_queue *q;
1324 
1325 	queue_for_each_hw_ctx(dev->queue, hctx, i) {
1326 		q = &dev->hw_queues[i];
1327 		rnbd_init_hw_queue(dev, q, hctx);
1328 		hctx->driver_data = q;
1329 	}
1330 }
1331 
1332 static void setup_request_queue(struct rnbd_clt_dev *dev,
1333 				struct rnbd_msg_open_rsp *rsp)
1334 {
1335 	blk_queue_logical_block_size(dev->queue,
1336 				     le16_to_cpu(rsp->logical_block_size));
1337 	blk_queue_physical_block_size(dev->queue,
1338 				      le16_to_cpu(rsp->physical_block_size));
1339 	blk_queue_max_hw_sectors(dev->queue,
1340 				 dev->sess->max_io_size / SECTOR_SIZE);
1341 
1342 	/*
1343 	 * we don't support discards to "discontiguous" segments
1344 	 * in on request
1345 	 */
1346 	blk_queue_max_discard_segments(dev->queue, 1);
1347 
1348 	blk_queue_max_discard_sectors(dev->queue,
1349 				      le32_to_cpu(rsp->max_discard_sectors));
1350 	dev->queue->limits.discard_granularity =
1351 					le32_to_cpu(rsp->discard_granularity);
1352 	dev->queue->limits.discard_alignment =
1353 					le32_to_cpu(rsp->discard_alignment);
1354 	if (le16_to_cpu(rsp->secure_discard))
1355 		blk_queue_max_secure_erase_sectors(dev->queue,
1356 					le32_to_cpu(rsp->max_discard_sectors));
1357 	blk_queue_flag_set(QUEUE_FLAG_SAME_COMP, dev->queue);
1358 	blk_queue_flag_set(QUEUE_FLAG_SAME_FORCE, dev->queue);
1359 	blk_queue_max_segments(dev->queue, dev->sess->max_segments);
1360 	blk_queue_io_opt(dev->queue, dev->sess->max_io_size);
1361 	blk_queue_virt_boundary(dev->queue, SZ_4K - 1);
1362 	blk_queue_write_cache(dev->queue,
1363 			      !!(rsp->cache_policy & RNBD_WRITEBACK),
1364 			      !!(rsp->cache_policy & RNBD_FUA));
1365 }
1366 
1367 static int rnbd_clt_setup_gen_disk(struct rnbd_clt_dev *dev,
1368 				   struct rnbd_msg_open_rsp *rsp, int idx)
1369 {
1370 	int err;
1371 
1372 	dev->gd->major		= rnbd_client_major;
1373 	dev->gd->first_minor	= idx << RNBD_PART_BITS;
1374 	dev->gd->minors		= 1 << RNBD_PART_BITS;
1375 	dev->gd->fops		= &rnbd_client_ops;
1376 	dev->gd->queue		= dev->queue;
1377 	dev->gd->private_data	= dev;
1378 	snprintf(dev->gd->disk_name, sizeof(dev->gd->disk_name), "rnbd%d",
1379 		 idx);
1380 	pr_debug("disk_name=%s, capacity=%llu\n",
1381 		 dev->gd->disk_name,
1382 		 le64_to_cpu(rsp->nsectors) *
1383 		 (le16_to_cpu(rsp->logical_block_size) / SECTOR_SIZE));
1384 
1385 	set_capacity(dev->gd, le64_to_cpu(rsp->nsectors));
1386 
1387 	if (dev->access_mode == RNBD_ACCESS_RO)
1388 		set_disk_ro(dev->gd, true);
1389 
1390 	/*
1391 	 * Network device does not need rotational
1392 	 */
1393 	blk_queue_flag_set(QUEUE_FLAG_NONROT, dev->queue);
1394 	err = add_disk(dev->gd);
1395 	if (err)
1396 		put_disk(dev->gd);
1397 
1398 	return err;
1399 }
1400 
1401 static int rnbd_client_setup_device(struct rnbd_clt_dev *dev,
1402 				    struct rnbd_msg_open_rsp *rsp)
1403 {
1404 	int idx = dev->clt_device_id;
1405 
1406 	dev->size = le64_to_cpu(rsp->nsectors) *
1407 			le16_to_cpu(rsp->logical_block_size);
1408 
1409 	dev->gd = blk_mq_alloc_disk(&dev->sess->tag_set, dev);
1410 	if (IS_ERR(dev->gd))
1411 		return PTR_ERR(dev->gd);
1412 	dev->queue = dev->gd->queue;
1413 	rnbd_init_mq_hw_queues(dev);
1414 
1415 	setup_request_queue(dev, rsp);
1416 	return rnbd_clt_setup_gen_disk(dev, rsp, idx);
1417 }
1418 
1419 static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
1420 				      enum rnbd_access_mode access_mode,
1421 				      const char *pathname,
1422 				      u32 nr_poll_queues)
1423 {
1424 	struct rnbd_clt_dev *dev;
1425 	int ret;
1426 
1427 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, NUMA_NO_NODE);
1428 	if (!dev)
1429 		return ERR_PTR(-ENOMEM);
1430 
1431 	/*
1432 	 * nr_cpu_ids: the number of softirq queues
1433 	 * nr_poll_queues: the number of polling queues
1434 	 */
1435 	dev->hw_queues = kcalloc(nr_cpu_ids + nr_poll_queues,
1436 				 sizeof(*dev->hw_queues),
1437 				 GFP_KERNEL);
1438 	if (!dev->hw_queues) {
1439 		ret = -ENOMEM;
1440 		goto out_alloc;
1441 	}
1442 
1443 	ret = ida_alloc_max(&index_ida, (1 << (MINORBITS - RNBD_PART_BITS)) - 1,
1444 			    GFP_KERNEL);
1445 	if (ret < 0) {
1446 		pr_err("Failed to initialize device '%s' from session %s, allocating idr failed, err: %d\n",
1447 		       pathname, sess->sessname, ret);
1448 		goto out_queues;
1449 	}
1450 
1451 	dev->pathname = kstrdup(pathname, GFP_KERNEL);
1452 	if (!dev->pathname) {
1453 		ret = -ENOMEM;
1454 		goto out_queues;
1455 	}
1456 
1457 	dev->clt_device_id	= ret;
1458 	dev->sess		= sess;
1459 	dev->access_mode	= access_mode;
1460 	dev->nr_poll_queues	= nr_poll_queues;
1461 	mutex_init(&dev->lock);
1462 	refcount_set(&dev->refcount, 1);
1463 	dev->dev_state = DEV_STATE_INIT;
1464 
1465 	/*
1466 	 * Here we called from sysfs entry, thus clt-sysfs is
1467 	 * responsible that session will not disappear.
1468 	 */
1469 	WARN_ON(!rnbd_clt_get_sess(sess));
1470 
1471 	return dev;
1472 
1473 out_queues:
1474 	kfree(dev->hw_queues);
1475 out_alloc:
1476 	kfree(dev);
1477 	return ERR_PTR(ret);
1478 }
1479 
1480 static bool __exists_dev(const char *pathname, const char *sessname)
1481 {
1482 	struct rnbd_clt_session *sess;
1483 	struct rnbd_clt_dev *dev;
1484 	bool found = false;
1485 
1486 	list_for_each_entry(sess, &sess_list, list) {
1487 		if (sessname && strncmp(sess->sessname, sessname,
1488 					sizeof(sess->sessname)))
1489 			continue;
1490 		mutex_lock(&sess->lock);
1491 		list_for_each_entry(dev, &sess->devs_list, list) {
1492 			if (strlen(dev->pathname) == strlen(pathname) &&
1493 			    !strcmp(dev->pathname, pathname)) {
1494 				found = true;
1495 				break;
1496 			}
1497 		}
1498 		mutex_unlock(&sess->lock);
1499 		if (found)
1500 			break;
1501 	}
1502 
1503 	return found;
1504 }
1505 
1506 static bool exists_devpath(const char *pathname, const char *sessname)
1507 {
1508 	bool found;
1509 
1510 	mutex_lock(&sess_lock);
1511 	found = __exists_dev(pathname, sessname);
1512 	mutex_unlock(&sess_lock);
1513 
1514 	return found;
1515 }
1516 
1517 static bool insert_dev_if_not_exists_devpath(struct rnbd_clt_dev *dev)
1518 {
1519 	bool found;
1520 	struct rnbd_clt_session *sess = dev->sess;
1521 
1522 	mutex_lock(&sess_lock);
1523 	found = __exists_dev(dev->pathname, sess->sessname);
1524 	if (!found) {
1525 		mutex_lock(&sess->lock);
1526 		list_add_tail(&dev->list, &sess->devs_list);
1527 		mutex_unlock(&sess->lock);
1528 	}
1529 	mutex_unlock(&sess_lock);
1530 
1531 	return found;
1532 }
1533 
1534 static void delete_dev(struct rnbd_clt_dev *dev)
1535 {
1536 	struct rnbd_clt_session *sess = dev->sess;
1537 
1538 	mutex_lock(&sess->lock);
1539 	list_del(&dev->list);
1540 	mutex_unlock(&sess->lock);
1541 }
1542 
1543 struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
1544 					   struct rtrs_addr *paths,
1545 					   size_t path_cnt, u16 port_nr,
1546 					   const char *pathname,
1547 					   enum rnbd_access_mode access_mode,
1548 					   u32 nr_poll_queues)
1549 {
1550 	struct rnbd_clt_session *sess;
1551 	struct rnbd_clt_dev *dev;
1552 	int ret, errno;
1553 	struct rnbd_msg_open_rsp *rsp;
1554 	struct rnbd_msg_open msg;
1555 	struct rnbd_iu *iu;
1556 	struct kvec vec = {
1557 		.iov_base = &msg,
1558 		.iov_len  = sizeof(msg)
1559 	};
1560 
1561 	if (exists_devpath(pathname, sessname))
1562 		return ERR_PTR(-EEXIST);
1563 
1564 	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
1565 	if (IS_ERR(sess))
1566 		return ERR_CAST(sess);
1567 
1568 	dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
1569 	if (IS_ERR(dev)) {
1570 		pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
1571 		       pathname, sess->sessname, PTR_ERR(dev));
1572 		ret = PTR_ERR(dev);
1573 		goto put_sess;
1574 	}
1575 	if (insert_dev_if_not_exists_devpath(dev)) {
1576 		ret = -EEXIST;
1577 		goto put_dev;
1578 	}
1579 
1580 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
1581 	if (!rsp) {
1582 		ret = -ENOMEM;
1583 		goto del_dev;
1584 	}
1585 
1586 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
1587 	if (!iu) {
1588 		ret = -ENOMEM;
1589 		kfree(rsp);
1590 		goto del_dev;
1591 	}
1592 	iu->buf = rsp;
1593 	iu->dev = dev;
1594 	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
1595 
1596 	msg.hdr.type    = cpu_to_le16(RNBD_MSG_OPEN);
1597 	msg.access_mode = dev->access_mode;
1598 	strscpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
1599 
1600 	WARN_ON(!rnbd_clt_get_dev(dev));
1601 	ret = send_usr_msg(sess->rtrs, READ, iu,
1602 			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
1603 			   msg_open_conf, &errno, RTRS_PERMIT_WAIT);
1604 	if (ret) {
1605 		rnbd_clt_put_dev(dev);
1606 		rnbd_put_iu(sess, iu);
1607 	} else {
1608 		ret = errno;
1609 	}
1610 	if (ret) {
1611 		rnbd_clt_err(dev,
1612 			      "map_device: failed, can't open remote device, err: %d\n",
1613 			      ret);
1614 		goto put_iu;
1615 	}
1616 	mutex_lock(&dev->lock);
1617 	pr_debug("Opened remote device: session=%s, path='%s'\n",
1618 		 sess->sessname, pathname);
1619 	ret = rnbd_client_setup_device(dev, rsp);
1620 	if (ret) {
1621 		rnbd_clt_err(dev,
1622 			      "map_device: Failed to configure device, err: %d\n",
1623 			      ret);
1624 		mutex_unlock(&dev->lock);
1625 		goto send_close;
1626 	}
1627 
1628 	rnbd_clt_info(dev,
1629 		       "map_device: Device mapped as %s (nsectors: %llu, logical_block_size: %d, physical_block_size: %d, max_discard_sectors: %d, discard_granularity: %d, discard_alignment: %d, secure_discard: %d, max_segments: %d, max_hw_sectors: %d, wc: %d, fua: %d)\n",
1630 		       dev->gd->disk_name, le64_to_cpu(rsp->nsectors),
1631 		       le16_to_cpu(rsp->logical_block_size),
1632 		       le16_to_cpu(rsp->physical_block_size),
1633 		       le32_to_cpu(rsp->max_discard_sectors),
1634 		       le32_to_cpu(rsp->discard_granularity),
1635 		       le32_to_cpu(rsp->discard_alignment),
1636 		       le16_to_cpu(rsp->secure_discard),
1637 		       sess->max_segments, sess->max_io_size / SECTOR_SIZE,
1638 		       !!(rsp->cache_policy & RNBD_WRITEBACK),
1639 		       !!(rsp->cache_policy & RNBD_FUA));
1640 
1641 	mutex_unlock(&dev->lock);
1642 	kfree(rsp);
1643 	rnbd_put_iu(sess, iu);
1644 	rnbd_clt_put_sess(sess);
1645 
1646 	return dev;
1647 
1648 send_close:
1649 	send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1650 put_iu:
1651 	kfree(rsp);
1652 	rnbd_put_iu(sess, iu);
1653 del_dev:
1654 	delete_dev(dev);
1655 put_dev:
1656 	rnbd_clt_put_dev(dev);
1657 put_sess:
1658 	rnbd_clt_put_sess(sess);
1659 
1660 	return ERR_PTR(ret);
1661 }
1662 
1663 static void destroy_gen_disk(struct rnbd_clt_dev *dev)
1664 {
1665 	del_gendisk(dev->gd);
1666 	put_disk(dev->gd);
1667 }
1668 
1669 static void destroy_sysfs(struct rnbd_clt_dev *dev,
1670 			  const struct attribute *sysfs_self)
1671 {
1672 	rnbd_clt_remove_dev_symlink(dev);
1673 	if (dev->kobj.state_initialized) {
1674 		if (sysfs_self)
1675 			/* To avoid deadlock firstly remove itself */
1676 			sysfs_remove_file_self(&dev->kobj, sysfs_self);
1677 		kobject_del(&dev->kobj);
1678 		kobject_put(&dev->kobj);
1679 	}
1680 }
1681 
1682 int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
1683 			   const struct attribute *sysfs_self)
1684 {
1685 	struct rnbd_clt_session *sess = dev->sess;
1686 	int refcount, ret = 0;
1687 	bool was_mapped;
1688 
1689 	mutex_lock(&dev->lock);
1690 	if (dev->dev_state == DEV_STATE_UNMAPPED) {
1691 		rnbd_clt_info(dev, "Device is already being unmapped\n");
1692 		ret = -EALREADY;
1693 		goto err;
1694 	}
1695 	refcount = refcount_read(&dev->refcount);
1696 	if (!force && refcount > 1) {
1697 		rnbd_clt_err(dev,
1698 			      "Closing device failed, device is in use, (%d device users)\n",
1699 			      refcount - 1);
1700 		ret = -EBUSY;
1701 		goto err;
1702 	}
1703 	was_mapped = (dev->dev_state == DEV_STATE_MAPPED);
1704 	dev->dev_state = DEV_STATE_UNMAPPED;
1705 	mutex_unlock(&dev->lock);
1706 
1707 	delete_dev(dev);
1708 	destroy_sysfs(dev, sysfs_self);
1709 	destroy_gen_disk(dev);
1710 	if (was_mapped && sess->rtrs)
1711 		send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1712 
1713 	rnbd_clt_info(dev, "Device is unmapped\n");
1714 
1715 	/* Likely last reference put */
1716 	rnbd_clt_put_dev(dev);
1717 
1718 	/*
1719 	 * Here device and session can be vanished!
1720 	 */
1721 
1722 	return 0;
1723 err:
1724 	mutex_unlock(&dev->lock);
1725 
1726 	return ret;
1727 }
1728 
1729 int rnbd_clt_remap_device(struct rnbd_clt_dev *dev)
1730 {
1731 	int err;
1732 
1733 	mutex_lock(&dev->lock);
1734 	if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED)
1735 		err = 0;
1736 	else if (dev->dev_state == DEV_STATE_UNMAPPED)
1737 		err = -ENODEV;
1738 	else if (dev->dev_state == DEV_STATE_MAPPED)
1739 		err = -EALREADY;
1740 	else
1741 		err = -EBUSY;
1742 	mutex_unlock(&dev->lock);
1743 	if (!err) {
1744 		rnbd_clt_info(dev, "Remapping device.\n");
1745 		err = send_msg_open(dev, RTRS_PERMIT_WAIT);
1746 		if (err)
1747 			rnbd_clt_err(dev, "remap_device: %d\n", err);
1748 	}
1749 
1750 	return err;
1751 }
1752 
1753 static void unmap_device_work(struct work_struct *work)
1754 {
1755 	struct rnbd_clt_dev *dev;
1756 
1757 	dev = container_of(work, typeof(*dev), unmap_on_rmmod_work);
1758 	rnbd_clt_unmap_device(dev, true, NULL);
1759 }
1760 
1761 static void rnbd_destroy_sessions(void)
1762 {
1763 	struct rnbd_clt_session *sess, *sn;
1764 	struct rnbd_clt_dev *dev, *tn;
1765 
1766 	/* Firstly forbid access through sysfs interface */
1767 	rnbd_clt_destroy_sysfs_files();
1768 
1769 	/*
1770 	 * Here at this point there is no any concurrent access to sessions
1771 	 * list and devices list:
1772 	 *   1. New session or device can't be created - session sysfs files
1773 	 *      are removed.
1774 	 *   2. Device or session can't be removed - module reference is taken
1775 	 *      into account in unmap device sysfs callback.
1776 	 *   3. No IO requests inflight - each file open of block_dev increases
1777 	 *      module reference in get_disk().
1778 	 *
1779 	 * But still there can be user requests inflights, which are sent by
1780 	 * asynchronous send_msg_*() functions, thus before unmapping devices
1781 	 * RTRS session must be explicitly closed.
1782 	 */
1783 
1784 	list_for_each_entry_safe(sess, sn, &sess_list, list) {
1785 		if (!rnbd_clt_get_sess(sess))
1786 			continue;
1787 		close_rtrs(sess);
1788 		list_for_each_entry_safe(dev, tn, &sess->devs_list, list) {
1789 			/*
1790 			 * Here unmap happens in parallel for only one reason:
1791 			 * del_gendisk() takes around half a second, so
1792 			 * on huge amount of devices the whole module unload
1793 			 * procedure takes minutes.
1794 			 */
1795 			INIT_WORK(&dev->unmap_on_rmmod_work, unmap_device_work);
1796 			queue_work(rnbd_clt_wq, &dev->unmap_on_rmmod_work);
1797 		}
1798 		rnbd_clt_put_sess(sess);
1799 	}
1800 	/* Wait for all scheduled unmap works */
1801 	flush_workqueue(rnbd_clt_wq);
1802 	WARN_ON(!list_empty(&sess_list));
1803 }
1804 
1805 static int __init rnbd_client_init(void)
1806 {
1807 	int err = 0;
1808 
1809 	BUILD_BUG_ON(sizeof(struct rnbd_msg_hdr) != 4);
1810 	BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info) != 36);
1811 	BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info_rsp) != 36);
1812 	BUILD_BUG_ON(sizeof(struct rnbd_msg_open) != 264);
1813 	BUILD_BUG_ON(sizeof(struct rnbd_msg_close) != 8);
1814 	BUILD_BUG_ON(sizeof(struct rnbd_msg_open_rsp) != 56);
1815 	rnbd_client_major = register_blkdev(rnbd_client_major, "rnbd");
1816 	if (rnbd_client_major <= 0) {
1817 		pr_err("Failed to load module, block device registration failed\n");
1818 		return -EBUSY;
1819 	}
1820 
1821 	err = rnbd_clt_create_sysfs_files();
1822 	if (err) {
1823 		pr_err("Failed to load module, creating sysfs device files failed, err: %d\n",
1824 		       err);
1825 		unregister_blkdev(rnbd_client_major, "rnbd");
1826 		return err;
1827 	}
1828 	rnbd_clt_wq = alloc_workqueue("rnbd_clt_wq", 0, 0);
1829 	if (!rnbd_clt_wq) {
1830 		pr_err("Failed to load module, alloc_workqueue failed.\n");
1831 		rnbd_clt_destroy_sysfs_files();
1832 		unregister_blkdev(rnbd_client_major, "rnbd");
1833 		err = -ENOMEM;
1834 	}
1835 
1836 	return err;
1837 }
1838 
1839 static void __exit rnbd_client_exit(void)
1840 {
1841 	rnbd_destroy_sessions();
1842 	unregister_blkdev(rnbd_client_major, "rnbd");
1843 	ida_destroy(&index_ida);
1844 	destroy_workqueue(rnbd_clt_wq);
1845 }
1846 
1847 module_init(rnbd_client_init);
1848 module_exit(rnbd_client_exit);
1849