xref: /openbmc/linux/drivers/block/rnbd/rnbd-clt.c (revision 7ac3945d)
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 int rnbd_clt_set_dev_attr(struct rnbd_clt_dev *dev,
72 				 const struct rnbd_msg_open_rsp *rsp)
73 {
74 	struct rnbd_clt_session *sess = dev->sess;
75 
76 	if (!rsp->logical_block_size)
77 		return -EINVAL;
78 
79 	dev->device_id		    = le32_to_cpu(rsp->device_id);
80 	dev->nsectors		    = le64_to_cpu(rsp->nsectors);
81 	dev->logical_block_size	    = le16_to_cpu(rsp->logical_block_size);
82 	dev->physical_block_size    = le16_to_cpu(rsp->physical_block_size);
83 	dev->max_discard_sectors    = le32_to_cpu(rsp->max_discard_sectors);
84 	dev->discard_granularity    = le32_to_cpu(rsp->discard_granularity);
85 	dev->discard_alignment	    = le32_to_cpu(rsp->discard_alignment);
86 	dev->secure_discard	    = le16_to_cpu(rsp->secure_discard);
87 	dev->wc			    = !!(rsp->cache_policy & RNBD_WRITEBACK);
88 	dev->fua		    = !!(rsp->cache_policy & RNBD_FUA);
89 
90 	dev->max_hw_sectors = sess->max_io_size / SECTOR_SIZE;
91 	dev->max_segments = sess->max_segments;
92 
93 	return 0;
94 }
95 
96 static int rnbd_clt_change_capacity(struct rnbd_clt_dev *dev,
97 				    size_t new_nsectors)
98 {
99 	rnbd_clt_info(dev, "Device size changed from %zu to %zu sectors\n",
100 		       dev->nsectors, new_nsectors);
101 	dev->nsectors = new_nsectors;
102 	set_capacity_and_notify(dev->gd, dev->nsectors);
103 	return 0;
104 }
105 
106 static int process_msg_open_rsp(struct rnbd_clt_dev *dev,
107 				struct rnbd_msg_open_rsp *rsp)
108 {
109 	struct kobject *gd_kobj;
110 	int err = 0;
111 
112 	mutex_lock(&dev->lock);
113 	if (dev->dev_state == DEV_STATE_UNMAPPED) {
114 		rnbd_clt_info(dev,
115 			       "Ignoring Open-Response message from server for  unmapped device\n");
116 		err = -ENOENT;
117 		goto out;
118 	}
119 	if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED) {
120 		u64 nsectors = le64_to_cpu(rsp->nsectors);
121 
122 		/*
123 		 * If the device was remapped and the size changed in the
124 		 * meantime we need to revalidate it
125 		 */
126 		if (dev->nsectors != nsectors)
127 			rnbd_clt_change_capacity(dev, nsectors);
128 		gd_kobj = &disk_to_dev(dev->gd)->kobj;
129 		kobject_uevent(gd_kobj, KOBJ_ONLINE);
130 		rnbd_clt_info(dev, "Device online, device remapped successfully\n");
131 	}
132 	err = rnbd_clt_set_dev_attr(dev, rsp);
133 	if (err)
134 		goto out;
135 	dev->dev_state = DEV_STATE_MAPPED;
136 
137 out:
138 	mutex_unlock(&dev->lock);
139 
140 	return err;
141 }
142 
143 int rnbd_clt_resize_disk(struct rnbd_clt_dev *dev, size_t newsize)
144 {
145 	int ret = 0;
146 
147 	mutex_lock(&dev->lock);
148 	if (dev->dev_state != DEV_STATE_MAPPED) {
149 		pr_err("Failed to set new size of the device, device is not opened\n");
150 		ret = -ENOENT;
151 		goto out;
152 	}
153 	ret = rnbd_clt_change_capacity(dev, newsize);
154 
155 out:
156 	mutex_unlock(&dev->lock);
157 
158 	return ret;
159 }
160 
161 static inline void rnbd_clt_dev_requeue(struct rnbd_queue *q)
162 {
163 	if (WARN_ON(!q->hctx))
164 		return;
165 
166 	/* We can come here from interrupt, thus async=true */
167 	blk_mq_run_hw_queue(q->hctx, true);
168 }
169 
170 enum {
171 	RNBD_DELAY_IFBUSY = -1,
172 };
173 
174 /**
175  * rnbd_get_cpu_qlist() - finds a list with HW queues to be rerun
176  * @sess:	Session to find a queue for
177  * @cpu:	Cpu to start the search from
178  *
179  * Description:
180  *     Each CPU has a list of HW queues, which needs to be rerun.  If a list
181  *     is not empty - it is marked with a bit.  This function finds first
182  *     set bit in a bitmap and returns corresponding CPU list.
183  */
184 static struct rnbd_cpu_qlist *
185 rnbd_get_cpu_qlist(struct rnbd_clt_session *sess, int cpu)
186 {
187 	int bit;
188 
189 	/* Search from cpu to nr_cpu_ids */
190 	bit = find_next_bit(sess->cpu_queues_bm, nr_cpu_ids, cpu);
191 	if (bit < nr_cpu_ids) {
192 		return per_cpu_ptr(sess->cpu_queues, bit);
193 	} else if (cpu != 0) {
194 		/* Search from 0 to cpu */
195 		bit = find_first_bit(sess->cpu_queues_bm, cpu);
196 		if (bit < cpu)
197 			return per_cpu_ptr(sess->cpu_queues, bit);
198 	}
199 
200 	return NULL;
201 }
202 
203 static inline int nxt_cpu(int cpu)
204 {
205 	return (cpu + 1) % nr_cpu_ids;
206 }
207 
208 /**
209  * rnbd_rerun_if_needed() - rerun next queue marked as stopped
210  * @sess:	Session to rerun a queue on
211  *
212  * Description:
213  *     Each CPU has it's own list of HW queues, which should be rerun.
214  *     Function finds such list with HW queues, takes a list lock, picks up
215  *     the first HW queue out of the list and requeues it.
216  *
217  * Return:
218  *     True if the queue was requeued, false otherwise.
219  *
220  * Context:
221  *     Does not matter.
222  */
223 static bool rnbd_rerun_if_needed(struct rnbd_clt_session *sess)
224 {
225 	struct rnbd_queue *q = NULL;
226 	struct rnbd_cpu_qlist *cpu_q;
227 	unsigned long flags;
228 	int *cpup;
229 
230 	/*
231 	 * To keep fairness and not to let other queues starve we always
232 	 * try to wake up someone else in round-robin manner.  That of course
233 	 * increases latency but queues always have a chance to be executed.
234 	 */
235 	cpup = get_cpu_ptr(sess->cpu_rr);
236 	for (cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(*cpup)); cpu_q;
237 	     cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(cpu_q->cpu))) {
238 		if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
239 			continue;
240 		if (!test_bit(cpu_q->cpu, sess->cpu_queues_bm))
241 			goto unlock;
242 		q = list_first_entry_or_null(&cpu_q->requeue_list,
243 					     typeof(*q), requeue_list);
244 		if (WARN_ON(!q))
245 			goto clear_bit;
246 		list_del_init(&q->requeue_list);
247 		clear_bit_unlock(0, &q->in_list);
248 
249 		if (list_empty(&cpu_q->requeue_list)) {
250 			/* Clear bit if nothing is left */
251 clear_bit:
252 			clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
253 		}
254 unlock:
255 		spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
256 
257 		if (q)
258 			break;
259 	}
260 
261 	/**
262 	 * Saves the CPU that is going to be requeued on the per-cpu var. Just
263 	 * incrementing it doesn't work because rnbd_get_cpu_qlist() will
264 	 * always return the first CPU with something on the queue list when the
265 	 * value stored on the var is greater than the last CPU with something
266 	 * on the list.
267 	 */
268 	if (cpu_q)
269 		*cpup = cpu_q->cpu;
270 	put_cpu_ptr(sess->cpu_rr);
271 
272 	if (q)
273 		rnbd_clt_dev_requeue(q);
274 
275 	return q;
276 }
277 
278 /**
279  * rnbd_rerun_all_if_idle() - rerun all queues left in the list if
280  *				 session is idling (there are no requests
281  *				 in-flight).
282  * @sess:	Session to rerun the queues on
283  *
284  * Description:
285  *     This function tries to rerun all stopped queues if there are no
286  *     requests in-flight anymore.  This function tries to solve an obvious
287  *     problem, when number of tags < than number of queues (hctx), which
288  *     are stopped and put to sleep.  If last permit, which has been just put,
289  *     does not wake up all left queues (hctxs), IO requests hang forever.
290  *
291  *     That can happen when all number of permits, say N, have been exhausted
292  *     from one CPU, and we have many block devices per session, say M.
293  *     Each block device has it's own queue (hctx) for each CPU, so eventually
294  *     we can put that number of queues (hctxs) to sleep: M x nr_cpu_ids.
295  *     If number of permits N < M x nr_cpu_ids finally we will get an IO hang.
296  *
297  *     To avoid this hang last caller of rnbd_put_permit() (last caller is the
298  *     one who observes sess->busy == 0) must wake up all remaining queues.
299  *
300  * Context:
301  *     Does not matter.
302  */
303 static void rnbd_rerun_all_if_idle(struct rnbd_clt_session *sess)
304 {
305 	bool requeued;
306 
307 	do {
308 		requeued = rnbd_rerun_if_needed(sess);
309 	} while (atomic_read(&sess->busy) == 0 && requeued);
310 }
311 
312 static struct rtrs_permit *rnbd_get_permit(struct rnbd_clt_session *sess,
313 					     enum rtrs_clt_con_type con_type,
314 					     enum wait_type wait)
315 {
316 	struct rtrs_permit *permit;
317 
318 	permit = rtrs_clt_get_permit(sess->rtrs, con_type, wait);
319 	if (permit)
320 		/* We have a subtle rare case here, when all permits can be
321 		 * consumed before busy counter increased.  This is safe,
322 		 * because loser will get NULL as a permit, observe 0 busy
323 		 * counter and immediately restart the queue himself.
324 		 */
325 		atomic_inc(&sess->busy);
326 
327 	return permit;
328 }
329 
330 static void rnbd_put_permit(struct rnbd_clt_session *sess,
331 			     struct rtrs_permit *permit)
332 {
333 	rtrs_clt_put_permit(sess->rtrs, permit);
334 	atomic_dec(&sess->busy);
335 	/* Paired with rnbd_clt_dev_add_to_requeue().  Decrement first
336 	 * and then check queue bits.
337 	 */
338 	smp_mb__after_atomic();
339 	rnbd_rerun_all_if_idle(sess);
340 }
341 
342 static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
343 				     enum rtrs_clt_con_type con_type,
344 				     enum wait_type wait)
345 {
346 	struct rnbd_iu *iu;
347 	struct rtrs_permit *permit;
348 
349 	iu = kzalloc(sizeof(*iu), GFP_KERNEL);
350 	if (!iu)
351 		return NULL;
352 
353 	permit = rnbd_get_permit(sess, con_type, wait);
354 	if (!permit) {
355 		kfree(iu);
356 		return NULL;
357 	}
358 
359 	iu->permit = permit;
360 	/*
361 	 * 1st reference is dropped after finishing sending a "user" message,
362 	 * 2nd reference is dropped after confirmation with the response is
363 	 * returned.
364 	 * 1st and 2nd can happen in any order, so the rnbd_iu should be
365 	 * released (rtrs_permit returned to rtrs) only after both
366 	 * are finished.
367 	 */
368 	atomic_set(&iu->refcount, 2);
369 	init_waitqueue_head(&iu->comp.wait);
370 	iu->comp.errno = INT_MAX;
371 
372 	if (sg_alloc_table(&iu->sgt, 1, GFP_KERNEL)) {
373 		rnbd_put_permit(sess, permit);
374 		kfree(iu);
375 		return NULL;
376 	}
377 
378 	return iu;
379 }
380 
381 static void rnbd_put_iu(struct rnbd_clt_session *sess, struct rnbd_iu *iu)
382 {
383 	if (atomic_dec_and_test(&iu->refcount)) {
384 		sg_free_table(&iu->sgt);
385 		rnbd_put_permit(sess, iu->permit);
386 		kfree(iu);
387 	}
388 }
389 
390 static void rnbd_softirq_done_fn(struct request *rq)
391 {
392 	struct rnbd_clt_dev *dev	= rq->q->disk->private_data;
393 	struct rnbd_clt_session *sess	= dev->sess;
394 	struct rnbd_iu *iu;
395 
396 	iu = blk_mq_rq_to_pdu(rq);
397 	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
398 	rnbd_put_permit(sess, iu->permit);
399 	blk_mq_end_request(rq, errno_to_blk_status(iu->errno));
400 }
401 
402 static void msg_io_conf(void *priv, int errno)
403 {
404 	struct rnbd_iu *iu = priv;
405 	struct rnbd_clt_dev *dev = iu->dev;
406 	struct request *rq = iu->rq;
407 	int rw = rq_data_dir(rq);
408 
409 	iu->errno = errno;
410 
411 	blk_mq_complete_request(rq);
412 
413 	if (errno)
414 		rnbd_clt_info_rl(dev, "%s I/O failed with err: %d\n",
415 				 rw == READ ? "read" : "write", errno);
416 }
417 
418 static void wake_up_iu_comp(struct rnbd_iu *iu, int errno)
419 {
420 	iu->comp.errno = errno;
421 	wake_up(&iu->comp.wait);
422 }
423 
424 static void msg_conf(void *priv, int errno)
425 {
426 	struct rnbd_iu *iu = priv;
427 
428 	iu->errno = errno;
429 	schedule_work(&iu->work);
430 }
431 
432 static int send_usr_msg(struct rtrs_clt_sess *rtrs, int dir,
433 			struct rnbd_iu *iu, struct kvec *vec,
434 			size_t len, struct scatterlist *sg, unsigned int sg_len,
435 			void (*conf)(struct work_struct *work),
436 			int *errno, int wait)
437 {
438 	int err;
439 	struct rtrs_clt_req_ops req_ops;
440 
441 	INIT_WORK(&iu->work, conf);
442 	req_ops = (struct rtrs_clt_req_ops) {
443 		.priv = iu,
444 		.conf_fn = msg_conf,
445 	};
446 	err = rtrs_clt_request(dir, &req_ops, rtrs, iu->permit,
447 				vec, 1, len, sg, sg_len);
448 	if (!err && wait) {
449 		wait_event(iu->comp.wait, iu->comp.errno != INT_MAX);
450 		*errno = iu->comp.errno;
451 	} else {
452 		*errno = 0;
453 	}
454 
455 	return err;
456 }
457 
458 static void msg_close_conf(struct work_struct *work)
459 {
460 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
461 	struct rnbd_clt_dev *dev = iu->dev;
462 
463 	wake_up_iu_comp(iu, iu->errno);
464 	rnbd_put_iu(dev->sess, iu);
465 	rnbd_clt_put_dev(dev);
466 }
467 
468 static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id,
469 			  enum wait_type wait)
470 {
471 	struct rnbd_clt_session *sess = dev->sess;
472 	struct rnbd_msg_close msg;
473 	struct rnbd_iu *iu;
474 	struct kvec vec = {
475 		.iov_base = &msg,
476 		.iov_len  = sizeof(msg)
477 	};
478 	int err, errno;
479 
480 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
481 	if (!iu)
482 		return -ENOMEM;
483 
484 	iu->buf = NULL;
485 	iu->dev = dev;
486 
487 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_CLOSE);
488 	msg.device_id	= cpu_to_le32(device_id);
489 
490 	WARN_ON(!rnbd_clt_get_dev(dev));
491 	err = send_usr_msg(sess->rtrs, WRITE, iu, &vec, 0, NULL, 0,
492 			   msg_close_conf, &errno, wait);
493 	if (err) {
494 		rnbd_clt_put_dev(dev);
495 		rnbd_put_iu(sess, iu);
496 	} else {
497 		err = errno;
498 	}
499 
500 	rnbd_put_iu(sess, iu);
501 	return err;
502 }
503 
504 static void msg_open_conf(struct work_struct *work)
505 {
506 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
507 	struct rnbd_msg_open_rsp *rsp = iu->buf;
508 	struct rnbd_clt_dev *dev = iu->dev;
509 	int errno = iu->errno;
510 
511 	if (errno) {
512 		rnbd_clt_err(dev,
513 			      "Opening failed, server responded: %d\n",
514 			      errno);
515 	} else {
516 		errno = process_msg_open_rsp(dev, rsp);
517 		if (errno) {
518 			u32 device_id = le32_to_cpu(rsp->device_id);
519 			/*
520 			 * If server thinks its fine, but we fail to process
521 			 * then be nice and send a close to server.
522 			 */
523 			send_msg_close(dev, device_id, RTRS_PERMIT_NOWAIT);
524 		}
525 	}
526 	kfree(rsp);
527 	wake_up_iu_comp(iu, errno);
528 	rnbd_put_iu(dev->sess, iu);
529 	rnbd_clt_put_dev(dev);
530 }
531 
532 static void msg_sess_info_conf(struct work_struct *work)
533 {
534 	struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
535 	struct rnbd_msg_sess_info_rsp *rsp = iu->buf;
536 	struct rnbd_clt_session *sess = iu->sess;
537 
538 	if (!iu->errno)
539 		sess->ver = min_t(u8, rsp->ver, RNBD_PROTO_VER_MAJOR);
540 
541 	kfree(rsp);
542 	wake_up_iu_comp(iu, iu->errno);
543 	rnbd_put_iu(sess, iu);
544 	rnbd_clt_put_sess(sess);
545 }
546 
547 static int send_msg_open(struct rnbd_clt_dev *dev, enum wait_type wait)
548 {
549 	struct rnbd_clt_session *sess = dev->sess;
550 	struct rnbd_msg_open_rsp *rsp;
551 	struct rnbd_msg_open msg;
552 	struct rnbd_iu *iu;
553 	struct kvec vec = {
554 		.iov_base = &msg,
555 		.iov_len  = sizeof(msg)
556 	};
557 	int err, errno;
558 
559 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
560 	if (!rsp)
561 		return -ENOMEM;
562 
563 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
564 	if (!iu) {
565 		kfree(rsp);
566 		return -ENOMEM;
567 	}
568 
569 	iu->buf = rsp;
570 	iu->dev = dev;
571 
572 	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
573 
574 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_OPEN);
575 	msg.access_mode	= dev->access_mode;
576 	strscpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
577 
578 	WARN_ON(!rnbd_clt_get_dev(dev));
579 	err = send_usr_msg(sess->rtrs, READ, iu,
580 			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
581 			   msg_open_conf, &errno, wait);
582 	if (err) {
583 		rnbd_clt_put_dev(dev);
584 		rnbd_put_iu(sess, iu);
585 		kfree(rsp);
586 	} else {
587 		err = errno;
588 	}
589 
590 	rnbd_put_iu(sess, iu);
591 	return err;
592 }
593 
594 static int send_msg_sess_info(struct rnbd_clt_session *sess, enum wait_type wait)
595 {
596 	struct rnbd_msg_sess_info_rsp *rsp;
597 	struct rnbd_msg_sess_info msg;
598 	struct rnbd_iu *iu;
599 	struct kvec vec = {
600 		.iov_base = &msg,
601 		.iov_len  = sizeof(msg)
602 	};
603 	int err, errno;
604 
605 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
606 	if (!rsp)
607 		return -ENOMEM;
608 
609 	iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
610 	if (!iu) {
611 		kfree(rsp);
612 		return -ENOMEM;
613 	}
614 
615 	iu->buf = rsp;
616 	iu->sess = sess;
617 	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
618 
619 	msg.hdr.type = cpu_to_le16(RNBD_MSG_SESS_INFO);
620 	msg.ver      = RNBD_PROTO_VER_MAJOR;
621 
622 	if (!rnbd_clt_get_sess(sess)) {
623 		/*
624 		 * That can happen only in one case, when RTRS has restablished
625 		 * the connection and link_ev() is called, but session is almost
626 		 * dead, last reference on session is put and caller is waiting
627 		 * for RTRS to close everything.
628 		 */
629 		err = -ENODEV;
630 		goto put_iu;
631 	}
632 	err = send_usr_msg(sess->rtrs, READ, iu,
633 			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
634 			   msg_sess_info_conf, &errno, wait);
635 	if (err) {
636 		rnbd_clt_put_sess(sess);
637 put_iu:
638 		rnbd_put_iu(sess, iu);
639 		kfree(rsp);
640 	} else {
641 		err = errno;
642 	}
643 	rnbd_put_iu(sess, iu);
644 	return err;
645 }
646 
647 static void set_dev_states_to_disconnected(struct rnbd_clt_session *sess)
648 {
649 	struct rnbd_clt_dev *dev;
650 	struct kobject *gd_kobj;
651 
652 	mutex_lock(&sess->lock);
653 	list_for_each_entry(dev, &sess->devs_list, list) {
654 		rnbd_clt_err(dev, "Device disconnected.\n");
655 
656 		mutex_lock(&dev->lock);
657 		if (dev->dev_state == DEV_STATE_MAPPED) {
658 			dev->dev_state = DEV_STATE_MAPPED_DISCONNECTED;
659 			gd_kobj = &disk_to_dev(dev->gd)->kobj;
660 			kobject_uevent(gd_kobj, KOBJ_OFFLINE);
661 		}
662 		mutex_unlock(&dev->lock);
663 	}
664 	mutex_unlock(&sess->lock);
665 }
666 
667 static void remap_devs(struct rnbd_clt_session *sess)
668 {
669 	struct rnbd_clt_dev *dev;
670 	struct rtrs_attrs attrs;
671 	int err;
672 
673 	/*
674 	 * Careful here: we are called from RTRS link event directly,
675 	 * thus we can't send any RTRS request and wait for response
676 	 * or RTRS will not be able to complete request with failure
677 	 * if something goes wrong (failing of outstanding requests
678 	 * happens exactly from the context where we are blocking now).
679 	 *
680 	 * So to avoid deadlocks each usr message sent from here must
681 	 * be asynchronous.
682 	 */
683 
684 	err = send_msg_sess_info(sess, RTRS_PERMIT_NOWAIT);
685 	if (err) {
686 		pr_err("send_msg_sess_info(\"%s\"): %d\n", sess->sessname, err);
687 		return;
688 	}
689 
690 	err = rtrs_clt_query(sess->rtrs, &attrs);
691 	if (err) {
692 		pr_err("rtrs_clt_query(\"%s\"): %d\n", sess->sessname, err);
693 		return;
694 	}
695 	mutex_lock(&sess->lock);
696 	sess->max_io_size = attrs.max_io_size;
697 
698 	list_for_each_entry(dev, &sess->devs_list, list) {
699 		bool skip;
700 
701 		mutex_lock(&dev->lock);
702 		skip = (dev->dev_state == DEV_STATE_INIT);
703 		mutex_unlock(&dev->lock);
704 		if (skip)
705 			/*
706 			 * When device is establishing connection for the first
707 			 * time - do not remap, it will be closed soon.
708 			 */
709 			continue;
710 
711 		rnbd_clt_info(dev, "session reconnected, remapping device\n");
712 		err = send_msg_open(dev, RTRS_PERMIT_NOWAIT);
713 		if (err) {
714 			rnbd_clt_err(dev, "send_msg_open(): %d\n", err);
715 			break;
716 		}
717 	}
718 	mutex_unlock(&sess->lock);
719 }
720 
721 static void rnbd_clt_link_ev(void *priv, enum rtrs_clt_link_ev ev)
722 {
723 	struct rnbd_clt_session *sess = priv;
724 
725 	switch (ev) {
726 	case RTRS_CLT_LINK_EV_DISCONNECTED:
727 		set_dev_states_to_disconnected(sess);
728 		break;
729 	case RTRS_CLT_LINK_EV_RECONNECTED:
730 		remap_devs(sess);
731 		break;
732 	default:
733 		pr_err("Unknown session event received (%d), session: %s\n",
734 		       ev, sess->sessname);
735 	}
736 }
737 
738 static void rnbd_init_cpu_qlists(struct rnbd_cpu_qlist __percpu *cpu_queues)
739 {
740 	unsigned int cpu;
741 	struct rnbd_cpu_qlist *cpu_q;
742 
743 	for_each_possible_cpu(cpu) {
744 		cpu_q = per_cpu_ptr(cpu_queues, cpu);
745 
746 		cpu_q->cpu = cpu;
747 		INIT_LIST_HEAD(&cpu_q->requeue_list);
748 		spin_lock_init(&cpu_q->requeue_lock);
749 	}
750 }
751 
752 static void destroy_mq_tags(struct rnbd_clt_session *sess)
753 {
754 	if (sess->tag_set.tags)
755 		blk_mq_free_tag_set(&sess->tag_set);
756 }
757 
758 static inline void wake_up_rtrs_waiters(struct rnbd_clt_session *sess)
759 {
760 	sess->rtrs_ready = true;
761 	wake_up_all(&sess->rtrs_waitq);
762 }
763 
764 static void close_rtrs(struct rnbd_clt_session *sess)
765 {
766 	might_sleep();
767 
768 	if (!IS_ERR_OR_NULL(sess->rtrs)) {
769 		rtrs_clt_close(sess->rtrs);
770 		sess->rtrs = NULL;
771 		wake_up_rtrs_waiters(sess);
772 	}
773 }
774 
775 static void free_sess(struct rnbd_clt_session *sess)
776 {
777 	WARN_ON(!list_empty(&sess->devs_list));
778 
779 	might_sleep();
780 
781 	close_rtrs(sess);
782 	destroy_mq_tags(sess);
783 	if (!list_empty(&sess->list)) {
784 		mutex_lock(&sess_lock);
785 		list_del(&sess->list);
786 		mutex_unlock(&sess_lock);
787 	}
788 	free_percpu(sess->cpu_queues);
789 	free_percpu(sess->cpu_rr);
790 	mutex_destroy(&sess->lock);
791 	kfree(sess);
792 }
793 
794 static struct rnbd_clt_session *alloc_sess(const char *sessname)
795 {
796 	struct rnbd_clt_session *sess;
797 	int err, cpu;
798 
799 	sess = kzalloc_node(sizeof(*sess), GFP_KERNEL, NUMA_NO_NODE);
800 	if (!sess)
801 		return ERR_PTR(-ENOMEM);
802 	strscpy(sess->sessname, sessname, sizeof(sess->sessname));
803 	atomic_set(&sess->busy, 0);
804 	mutex_init(&sess->lock);
805 	INIT_LIST_HEAD(&sess->devs_list);
806 	INIT_LIST_HEAD(&sess->list);
807 	bitmap_zero(sess->cpu_queues_bm, num_possible_cpus());
808 	init_waitqueue_head(&sess->rtrs_waitq);
809 	refcount_set(&sess->refcount, 1);
810 
811 	sess->cpu_queues = alloc_percpu(struct rnbd_cpu_qlist);
812 	if (!sess->cpu_queues) {
813 		err = -ENOMEM;
814 		goto err;
815 	}
816 	rnbd_init_cpu_qlists(sess->cpu_queues);
817 
818 	/*
819 	 * That is simple percpu variable which stores cpu indices, which are
820 	 * incremented on each access.  We need that for the sake of fairness
821 	 * to wake up queues in a round-robin manner.
822 	 */
823 	sess->cpu_rr = alloc_percpu(int);
824 	if (!sess->cpu_rr) {
825 		err = -ENOMEM;
826 		goto err;
827 	}
828 	for_each_possible_cpu(cpu)
829 		* per_cpu_ptr(sess->cpu_rr, cpu) = cpu;
830 
831 	return sess;
832 
833 err:
834 	free_sess(sess);
835 
836 	return ERR_PTR(err);
837 }
838 
839 static int wait_for_rtrs_connection(struct rnbd_clt_session *sess)
840 {
841 	wait_event(sess->rtrs_waitq, sess->rtrs_ready);
842 	if (IS_ERR_OR_NULL(sess->rtrs))
843 		return -ECONNRESET;
844 
845 	return 0;
846 }
847 
848 static void wait_for_rtrs_disconnection(struct rnbd_clt_session *sess)
849 	__releases(&sess_lock)
850 	__acquires(&sess_lock)
851 {
852 	DEFINE_WAIT(wait);
853 
854 	prepare_to_wait(&sess->rtrs_waitq, &wait, TASK_UNINTERRUPTIBLE);
855 	if (IS_ERR_OR_NULL(sess->rtrs)) {
856 		finish_wait(&sess->rtrs_waitq, &wait);
857 		return;
858 	}
859 	mutex_unlock(&sess_lock);
860 	/* loop in caller, see __find_and_get_sess().
861 	 * You can't leave mutex locked and call schedule(), you will catch a
862 	 * deadlock with a caller of free_sess(), which has just put the last
863 	 * reference and is about to take the sess_lock in order to delete
864 	 * the session from the list.
865 	 */
866 	schedule();
867 	mutex_lock(&sess_lock);
868 }
869 
870 static struct rnbd_clt_session *__find_and_get_sess(const char *sessname)
871 	__releases(&sess_lock)
872 	__acquires(&sess_lock)
873 {
874 	struct rnbd_clt_session *sess, *sn;
875 	int err;
876 
877 again:
878 	list_for_each_entry_safe(sess, sn, &sess_list, list) {
879 		if (strcmp(sessname, sess->sessname))
880 			continue;
881 
882 		if (sess->rtrs_ready && IS_ERR_OR_NULL(sess->rtrs))
883 			/*
884 			 * No RTRS connection, session is dying.
885 			 */
886 			continue;
887 
888 		if (rnbd_clt_get_sess(sess)) {
889 			/*
890 			 * Alive session is found, wait for RTRS connection.
891 			 */
892 			mutex_unlock(&sess_lock);
893 			err = wait_for_rtrs_connection(sess);
894 			if (err)
895 				rnbd_clt_put_sess(sess);
896 			mutex_lock(&sess_lock);
897 
898 			if (err)
899 				/* Session is dying, repeat the loop */
900 				goto again;
901 
902 			return sess;
903 		}
904 		/*
905 		 * Ref is 0, session is dying, wait for RTRS disconnect
906 		 * in order to avoid session names clashes.
907 		 */
908 		wait_for_rtrs_disconnection(sess);
909 		/*
910 		 * RTRS is disconnected and soon session will be freed,
911 		 * so repeat a loop.
912 		 */
913 		goto again;
914 	}
915 
916 	return NULL;
917 }
918 
919 /* caller is responsible for initializing 'first' to false */
920 static struct
921 rnbd_clt_session *find_or_create_sess(const char *sessname, bool *first)
922 {
923 	struct rnbd_clt_session *sess = NULL;
924 
925 	mutex_lock(&sess_lock);
926 	sess = __find_and_get_sess(sessname);
927 	if (!sess) {
928 		sess = alloc_sess(sessname);
929 		if (IS_ERR(sess)) {
930 			mutex_unlock(&sess_lock);
931 			return sess;
932 		}
933 		list_add(&sess->list, &sess_list);
934 		*first = true;
935 	}
936 	mutex_unlock(&sess_lock);
937 
938 	return sess;
939 }
940 
941 static int rnbd_client_open(struct block_device *block_device, fmode_t mode)
942 {
943 	struct rnbd_clt_dev *dev = block_device->bd_disk->private_data;
944 
945 	if (dev->read_only && (mode & FMODE_WRITE))
946 		return -EPERM;
947 
948 	if (dev->dev_state == DEV_STATE_UNMAPPED ||
949 	    !rnbd_clt_get_dev(dev))
950 		return -EIO;
951 
952 	return 0;
953 }
954 
955 static void rnbd_client_release(struct gendisk *gen, fmode_t mode)
956 {
957 	struct rnbd_clt_dev *dev = gen->private_data;
958 
959 	rnbd_clt_put_dev(dev);
960 }
961 
962 static int rnbd_client_getgeo(struct block_device *block_device,
963 			      struct hd_geometry *geo)
964 {
965 	u64 size;
966 	struct rnbd_clt_dev *dev;
967 
968 	dev = block_device->bd_disk->private_data;
969 	size = dev->size * (dev->logical_block_size / SECTOR_SIZE);
970 	geo->cylinders	= size >> 6;	/* size/64 */
971 	geo->heads	= 4;
972 	geo->sectors	= 16;
973 	geo->start	= 0;
974 
975 	return 0;
976 }
977 
978 static const struct block_device_operations rnbd_client_ops = {
979 	.owner		= THIS_MODULE,
980 	.open		= rnbd_client_open,
981 	.release	= rnbd_client_release,
982 	.getgeo		= rnbd_client_getgeo
983 };
984 
985 /* The amount of data that belongs to an I/O and the amount of data that
986  * should be read or written to the disk (bi_size) can differ.
987  *
988  * E.g. When WRITE_SAME is used, only a small amount of data is
989  * transferred that is then written repeatedly over a lot of sectors.
990  *
991  * Get the size of data to be transferred via RTRS by summing up the size
992  * of the scather-gather list entries.
993  */
994 static size_t rnbd_clt_get_sg_size(struct scatterlist *sglist, u32 len)
995 {
996 	struct scatterlist *sg;
997 	size_t tsize = 0;
998 	int i;
999 
1000 	for_each_sg(sglist, sg, len, i)
1001 		tsize += sg->length;
1002 	return tsize;
1003 }
1004 
1005 static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
1006 				     struct request *rq,
1007 				     struct rnbd_iu *iu)
1008 {
1009 	struct rtrs_clt_sess *rtrs = dev->sess->rtrs;
1010 	struct rtrs_permit *permit = iu->permit;
1011 	struct rnbd_msg_io msg;
1012 	struct rtrs_clt_req_ops req_ops;
1013 	unsigned int sg_cnt = 0;
1014 	struct kvec vec;
1015 	size_t size;
1016 	int err;
1017 
1018 	iu->rq		= rq;
1019 	iu->dev		= dev;
1020 	msg.sector	= cpu_to_le64(blk_rq_pos(rq));
1021 	msg.bi_size	= cpu_to_le32(blk_rq_bytes(rq));
1022 	msg.rw		= cpu_to_le32(rq_to_rnbd_flags(rq));
1023 	msg.prio	= cpu_to_le16(req_get_ioprio(rq));
1024 
1025 	/*
1026 	 * We only support discards with single segment for now.
1027 	 * See queue limits.
1028 	 */
1029 	if (req_op(rq) != REQ_OP_DISCARD)
1030 		sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sgt.sgl);
1031 
1032 	if (sg_cnt == 0)
1033 		sg_mark_end(&iu->sgt.sgl[0]);
1034 
1035 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_IO);
1036 	msg.device_id	= cpu_to_le32(dev->device_id);
1037 
1038 	vec = (struct kvec) {
1039 		.iov_base = &msg,
1040 		.iov_len  = sizeof(msg)
1041 	};
1042 	size = rnbd_clt_get_sg_size(iu->sgt.sgl, sg_cnt);
1043 	req_ops = (struct rtrs_clt_req_ops) {
1044 		.priv = iu,
1045 		.conf_fn = msg_io_conf,
1046 	};
1047 	err = rtrs_clt_request(rq_data_dir(rq), &req_ops, rtrs, permit,
1048 			       &vec, 1, size, iu->sgt.sgl, sg_cnt);
1049 	if (err) {
1050 		rnbd_clt_err_rl(dev, "RTRS failed to transfer IO, err: %d\n",
1051 				 err);
1052 		return err;
1053 	}
1054 
1055 	return 0;
1056 }
1057 
1058 /**
1059  * rnbd_clt_dev_add_to_requeue() - add device to requeue if session is busy
1060  * @dev:	Device to be checked
1061  * @q:		Queue to be added to the requeue list if required
1062  *
1063  * Description:
1064  *     If session is busy, that means someone will requeue us when resources
1065  *     are freed.  If session is not doing anything - device is not added to
1066  *     the list and @false is returned.
1067  */
1068 static bool rnbd_clt_dev_add_to_requeue(struct rnbd_clt_dev *dev,
1069 						struct rnbd_queue *q)
1070 {
1071 	struct rnbd_clt_session *sess = dev->sess;
1072 	struct rnbd_cpu_qlist *cpu_q;
1073 	unsigned long flags;
1074 	bool added = true;
1075 	bool need_set;
1076 
1077 	cpu_q = get_cpu_ptr(sess->cpu_queues);
1078 	spin_lock_irqsave(&cpu_q->requeue_lock, flags);
1079 
1080 	if (!test_and_set_bit_lock(0, &q->in_list)) {
1081 		if (WARN_ON(!list_empty(&q->requeue_list)))
1082 			goto unlock;
1083 
1084 		need_set = !test_bit(cpu_q->cpu, sess->cpu_queues_bm);
1085 		if (need_set) {
1086 			set_bit(cpu_q->cpu, sess->cpu_queues_bm);
1087 			/* Paired with rnbd_put_permit(). Set a bit first
1088 			 * and then observe the busy counter.
1089 			 */
1090 			smp_mb__before_atomic();
1091 		}
1092 		if (atomic_read(&sess->busy)) {
1093 			list_add_tail(&q->requeue_list, &cpu_q->requeue_list);
1094 		} else {
1095 			/* Very unlikely, but possible: busy counter was
1096 			 * observed as zero.  Drop all bits and return
1097 			 * false to restart the queue by ourselves.
1098 			 */
1099 			if (need_set)
1100 				clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
1101 			clear_bit_unlock(0, &q->in_list);
1102 			added = false;
1103 		}
1104 	}
1105 unlock:
1106 	spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
1107 	put_cpu_ptr(sess->cpu_queues);
1108 
1109 	return added;
1110 }
1111 
1112 static void rnbd_clt_dev_kick_mq_queue(struct rnbd_clt_dev *dev,
1113 					struct blk_mq_hw_ctx *hctx,
1114 					int delay)
1115 {
1116 	struct rnbd_queue *q = hctx->driver_data;
1117 
1118 	if (delay != RNBD_DELAY_IFBUSY)
1119 		blk_mq_delay_run_hw_queue(hctx, delay);
1120 	else if (!rnbd_clt_dev_add_to_requeue(dev, q))
1121 		/*
1122 		 * If session is not busy we have to restart
1123 		 * the queue ourselves.
1124 		 */
1125 		blk_mq_delay_run_hw_queue(hctx, 10/*ms*/);
1126 }
1127 
1128 static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1129 				   const struct blk_mq_queue_data *bd)
1130 {
1131 	struct request *rq = bd->rq;
1132 	struct rnbd_clt_dev *dev = rq->q->disk->private_data;
1133 	struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
1134 	int err;
1135 	blk_status_t ret = BLK_STS_IOERR;
1136 
1137 	if (dev->dev_state != DEV_STATE_MAPPED)
1138 		return BLK_STS_IOERR;
1139 
1140 	iu->permit = rnbd_get_permit(dev->sess, RTRS_IO_CON,
1141 				      RTRS_PERMIT_NOWAIT);
1142 	if (!iu->permit) {
1143 		rnbd_clt_dev_kick_mq_queue(dev, hctx, RNBD_DELAY_IFBUSY);
1144 		return BLK_STS_RESOURCE;
1145 	}
1146 
1147 	iu->sgt.sgl = iu->first_sgl;
1148 	err = sg_alloc_table_chained(&iu->sgt,
1149 				     /* Even-if the request has no segment,
1150 				      * sglist must have one entry at least.
1151 				      */
1152 				     blk_rq_nr_phys_segments(rq) ? : 1,
1153 				     iu->sgt.sgl,
1154 				     RNBD_INLINE_SG_CNT);
1155 	if (err) {
1156 		rnbd_clt_err_rl(dev, "sg_alloc_table_chained ret=%d\n", err);
1157 		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1158 		rnbd_put_permit(dev->sess, iu->permit);
1159 		return BLK_STS_RESOURCE;
1160 	}
1161 
1162 	blk_mq_start_request(rq);
1163 	err = rnbd_client_xfer_request(dev, rq, iu);
1164 	if (err == 0)
1165 		return BLK_STS_OK;
1166 	if (err == -EAGAIN || err == -ENOMEM) {
1167 		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1168 		ret = BLK_STS_RESOURCE;
1169 	}
1170 	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
1171 	rnbd_put_permit(dev->sess, iu->permit);
1172 	return ret;
1173 }
1174 
1175 static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
1176 {
1177 	struct rnbd_queue *q = hctx->driver_data;
1178 	struct rnbd_clt_dev *dev = q->dev;
1179 	int cnt;
1180 
1181 	cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
1182 	return cnt;
1183 }
1184 
1185 static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
1186 {
1187 	struct rnbd_clt_session *sess = set->driver_data;
1188 
1189 	/* shared read/write queues */
1190 	set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
1191 	set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1192 	set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
1193 	set->map[HCTX_TYPE_READ].queue_offset = 0;
1194 	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1195 	blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
1196 
1197 	if (sess->nr_poll_queues) {
1198 		/* dedicated queue for poll */
1199 		set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
1200 		set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
1201 			set->map[HCTX_TYPE_READ].nr_queues;
1202 		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1203 		pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
1204 			sess->sessname,
1205 			set->map[HCTX_TYPE_DEFAULT].nr_queues,
1206 			set->map[HCTX_TYPE_READ].nr_queues,
1207 			set->map[HCTX_TYPE_POLL].nr_queues);
1208 	} else {
1209 		pr_info("[session=%s] mapped %d/%d default/read queues.\n",
1210 			sess->sessname,
1211 			set->map[HCTX_TYPE_DEFAULT].nr_queues,
1212 			set->map[HCTX_TYPE_READ].nr_queues);
1213 	}
1214 
1215 	return 0;
1216 }
1217 
1218 static struct blk_mq_ops rnbd_mq_ops = {
1219 	.queue_rq	= rnbd_queue_rq,
1220 	.complete	= rnbd_softirq_done_fn,
1221 	.map_queues     = rnbd_rdma_map_queues,
1222 	.poll           = rnbd_rdma_poll,
1223 };
1224 
1225 static int setup_mq_tags(struct rnbd_clt_session *sess)
1226 {
1227 	struct blk_mq_tag_set *tag_set = &sess->tag_set;
1228 
1229 	memset(tag_set, 0, sizeof(*tag_set));
1230 	tag_set->ops		= &rnbd_mq_ops;
1231 	tag_set->queue_depth	= sess->queue_depth;
1232 	tag_set->numa_node		= NUMA_NO_NODE;
1233 	tag_set->flags		= BLK_MQ_F_SHOULD_MERGE |
1234 				  BLK_MQ_F_TAG_QUEUE_SHARED;
1235 	tag_set->cmd_size	= sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
1236 
1237 	/* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
1238 	tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1239 	/*
1240 	 * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
1241 	 * others are for HCTX_TYPE_POLL
1242 	 */
1243 	tag_set->nr_hw_queues	= num_online_cpus() + sess->nr_poll_queues;
1244 	tag_set->driver_data    = sess;
1245 
1246 	return blk_mq_alloc_tag_set(tag_set);
1247 }
1248 
1249 static struct rnbd_clt_session *
1250 find_and_get_or_create_sess(const char *sessname,
1251 			    const struct rtrs_addr *paths,
1252 			    size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
1253 {
1254 	struct rnbd_clt_session *sess;
1255 	struct rtrs_attrs attrs;
1256 	int err;
1257 	bool first = false;
1258 	struct rtrs_clt_ops rtrs_ops;
1259 
1260 	sess = find_or_create_sess(sessname, &first);
1261 	if (sess == ERR_PTR(-ENOMEM)) {
1262 		return ERR_PTR(-ENOMEM);
1263 	} else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
1264 		/*
1265 		 * A device MUST have its own session to use the polling-mode.
1266 		 * It must fail to map new device with the same session.
1267 		 */
1268 		err = -EINVAL;
1269 		goto put_sess;
1270 	}
1271 
1272 	if (!first)
1273 		return sess;
1274 
1275 	if (!path_cnt) {
1276 		pr_err("Session %s not found, and path parameter not given", sessname);
1277 		err = -ENXIO;
1278 		goto put_sess;
1279 	}
1280 
1281 	rtrs_ops = (struct rtrs_clt_ops) {
1282 		.priv = sess,
1283 		.link_ev = rnbd_clt_link_ev,
1284 	};
1285 	/*
1286 	 * Nothing was found, establish rtrs connection and proceed further.
1287 	 */
1288 	sess->rtrs = rtrs_clt_open(&rtrs_ops, sessname,
1289 				   paths, path_cnt, port_nr,
1290 				   0, /* Do not use pdu of rtrs */
1291 				   RECONNECT_DELAY,
1292 				   MAX_RECONNECTS, nr_poll_queues);
1293 	if (IS_ERR(sess->rtrs)) {
1294 		err = PTR_ERR(sess->rtrs);
1295 		goto wake_up_and_put;
1296 	}
1297 
1298 	err = rtrs_clt_query(sess->rtrs, &attrs);
1299 	if (err)
1300 		goto close_rtrs;
1301 
1302 	sess->max_io_size = attrs.max_io_size;
1303 	sess->queue_depth = attrs.queue_depth;
1304 	sess->nr_poll_queues = nr_poll_queues;
1305 	sess->max_segments = attrs.max_segments;
1306 
1307 	err = setup_mq_tags(sess);
1308 	if (err)
1309 		goto close_rtrs;
1310 
1311 	err = send_msg_sess_info(sess, RTRS_PERMIT_WAIT);
1312 	if (err)
1313 		goto close_rtrs;
1314 
1315 	wake_up_rtrs_waiters(sess);
1316 
1317 	return sess;
1318 
1319 close_rtrs:
1320 	close_rtrs(sess);
1321 put_sess:
1322 	rnbd_clt_put_sess(sess);
1323 
1324 	return ERR_PTR(err);
1325 
1326 wake_up_and_put:
1327 	wake_up_rtrs_waiters(sess);
1328 	goto put_sess;
1329 }
1330 
1331 static inline void rnbd_init_hw_queue(struct rnbd_clt_dev *dev,
1332 				       struct rnbd_queue *q,
1333 				       struct blk_mq_hw_ctx *hctx)
1334 {
1335 	INIT_LIST_HEAD(&q->requeue_list);
1336 	q->dev  = dev;
1337 	q->hctx = hctx;
1338 }
1339 
1340 static void rnbd_init_mq_hw_queues(struct rnbd_clt_dev *dev)
1341 {
1342 	unsigned long i;
1343 	struct blk_mq_hw_ctx *hctx;
1344 	struct rnbd_queue *q;
1345 
1346 	queue_for_each_hw_ctx(dev->queue, hctx, i) {
1347 		q = &dev->hw_queues[i];
1348 		rnbd_init_hw_queue(dev, q, hctx);
1349 		hctx->driver_data = q;
1350 	}
1351 }
1352 
1353 static void setup_request_queue(struct rnbd_clt_dev *dev)
1354 {
1355 	blk_queue_logical_block_size(dev->queue, dev->logical_block_size);
1356 	blk_queue_physical_block_size(dev->queue, dev->physical_block_size);
1357 	blk_queue_max_hw_sectors(dev->queue, dev->max_hw_sectors);
1358 
1359 	/*
1360 	 * we don't support discards to "discontiguous" segments
1361 	 * in on request
1362 	 */
1363 	blk_queue_max_discard_segments(dev->queue, 1);
1364 
1365 	blk_queue_max_discard_sectors(dev->queue, dev->max_discard_sectors);
1366 	dev->queue->limits.discard_granularity	= dev->discard_granularity;
1367 	dev->queue->limits.discard_alignment	= dev->discard_alignment;
1368 	if (dev->secure_discard)
1369 		blk_queue_max_secure_erase_sectors(dev->queue,
1370 				dev->max_discard_sectors);
1371 	blk_queue_flag_set(QUEUE_FLAG_SAME_COMP, dev->queue);
1372 	blk_queue_flag_set(QUEUE_FLAG_SAME_FORCE, dev->queue);
1373 	blk_queue_max_segments(dev->queue, dev->max_segments);
1374 	blk_queue_io_opt(dev->queue, dev->sess->max_io_size);
1375 	blk_queue_virt_boundary(dev->queue, SZ_4K - 1);
1376 	blk_queue_write_cache(dev->queue, dev->wc, dev->fua);
1377 }
1378 
1379 static int rnbd_clt_setup_gen_disk(struct rnbd_clt_dev *dev, int idx)
1380 {
1381 	int err;
1382 
1383 	dev->gd->major		= rnbd_client_major;
1384 	dev->gd->first_minor	= idx << RNBD_PART_BITS;
1385 	dev->gd->minors		= 1 << RNBD_PART_BITS;
1386 	dev->gd->fops		= &rnbd_client_ops;
1387 	dev->gd->queue		= dev->queue;
1388 	dev->gd->private_data	= dev;
1389 	snprintf(dev->gd->disk_name, sizeof(dev->gd->disk_name), "rnbd%d",
1390 		 idx);
1391 	pr_debug("disk_name=%s, capacity=%zu\n",
1392 		 dev->gd->disk_name,
1393 		 dev->nsectors * (dev->logical_block_size / SECTOR_SIZE)
1394 		 );
1395 
1396 	set_capacity(dev->gd, dev->nsectors);
1397 
1398 	if (dev->access_mode == RNBD_ACCESS_RO) {
1399 		dev->read_only = true;
1400 		set_disk_ro(dev->gd, true);
1401 	} else {
1402 		dev->read_only = false;
1403 	}
1404 
1405 	/*
1406 	 * Network device does not need rotational
1407 	 */
1408 	blk_queue_flag_set(QUEUE_FLAG_NONROT, dev->queue);
1409 	err = add_disk(dev->gd);
1410 	if (err)
1411 		blk_cleanup_disk(dev->gd);
1412 
1413 	return err;
1414 }
1415 
1416 static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
1417 {
1418 	int idx = dev->clt_device_id;
1419 
1420 	dev->size = dev->nsectors * dev->logical_block_size;
1421 
1422 	dev->gd = blk_mq_alloc_disk(&dev->sess->tag_set, dev);
1423 	if (IS_ERR(dev->gd))
1424 		return PTR_ERR(dev->gd);
1425 	dev->queue = dev->gd->queue;
1426 	rnbd_init_mq_hw_queues(dev);
1427 
1428 	setup_request_queue(dev);
1429 	return rnbd_clt_setup_gen_disk(dev, idx);
1430 }
1431 
1432 static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
1433 				      enum rnbd_access_mode access_mode,
1434 				      const char *pathname,
1435 				      u32 nr_poll_queues)
1436 {
1437 	struct rnbd_clt_dev *dev;
1438 	int ret;
1439 
1440 	dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, NUMA_NO_NODE);
1441 	if (!dev)
1442 		return ERR_PTR(-ENOMEM);
1443 
1444 	/*
1445 	 * nr_cpu_ids: the number of softirq queues
1446 	 * nr_poll_queues: the number of polling queues
1447 	 */
1448 	dev->hw_queues = kcalloc(nr_cpu_ids + nr_poll_queues,
1449 				 sizeof(*dev->hw_queues),
1450 				 GFP_KERNEL);
1451 	if (!dev->hw_queues) {
1452 		ret = -ENOMEM;
1453 		goto out_alloc;
1454 	}
1455 
1456 	ret = ida_alloc_max(&index_ida, 1 << (MINORBITS - RNBD_PART_BITS),
1457 			    GFP_KERNEL);
1458 	if (ret < 0) {
1459 		pr_err("Failed to initialize device '%s' from session %s, allocating idr failed, err: %d\n",
1460 		       pathname, sess->sessname, ret);
1461 		goto out_queues;
1462 	}
1463 
1464 	dev->pathname = kstrdup(pathname, GFP_KERNEL);
1465 	if (!dev->pathname) {
1466 		ret = -ENOMEM;
1467 		goto out_queues;
1468 	}
1469 
1470 	dev->clt_device_id	= ret;
1471 	dev->sess		= sess;
1472 	dev->access_mode	= access_mode;
1473 	dev->nr_poll_queues	= nr_poll_queues;
1474 	mutex_init(&dev->lock);
1475 	refcount_set(&dev->refcount, 1);
1476 	dev->dev_state = DEV_STATE_INIT;
1477 
1478 	/*
1479 	 * Here we called from sysfs entry, thus clt-sysfs is
1480 	 * responsible that session will not disappear.
1481 	 */
1482 	WARN_ON(!rnbd_clt_get_sess(sess));
1483 
1484 	return dev;
1485 
1486 out_queues:
1487 	kfree(dev->hw_queues);
1488 out_alloc:
1489 	kfree(dev);
1490 	return ERR_PTR(ret);
1491 }
1492 
1493 static bool __exists_dev(const char *pathname, const char *sessname)
1494 {
1495 	struct rnbd_clt_session *sess;
1496 	struct rnbd_clt_dev *dev;
1497 	bool found = false;
1498 
1499 	list_for_each_entry(sess, &sess_list, list) {
1500 		if (sessname && strncmp(sess->sessname, sessname,
1501 					sizeof(sess->sessname)))
1502 			continue;
1503 		mutex_lock(&sess->lock);
1504 		list_for_each_entry(dev, &sess->devs_list, list) {
1505 			if (strlen(dev->pathname) == strlen(pathname) &&
1506 			    !strcmp(dev->pathname, pathname)) {
1507 				found = true;
1508 				break;
1509 			}
1510 		}
1511 		mutex_unlock(&sess->lock);
1512 		if (found)
1513 			break;
1514 	}
1515 
1516 	return found;
1517 }
1518 
1519 static bool exists_devpath(const char *pathname, const char *sessname)
1520 {
1521 	bool found;
1522 
1523 	mutex_lock(&sess_lock);
1524 	found = __exists_dev(pathname, sessname);
1525 	mutex_unlock(&sess_lock);
1526 
1527 	return found;
1528 }
1529 
1530 static bool insert_dev_if_not_exists_devpath(struct rnbd_clt_dev *dev)
1531 {
1532 	bool found;
1533 	struct rnbd_clt_session *sess = dev->sess;
1534 
1535 	mutex_lock(&sess_lock);
1536 	found = __exists_dev(dev->pathname, sess->sessname);
1537 	if (!found) {
1538 		mutex_lock(&sess->lock);
1539 		list_add_tail(&dev->list, &sess->devs_list);
1540 		mutex_unlock(&sess->lock);
1541 	}
1542 	mutex_unlock(&sess_lock);
1543 
1544 	return found;
1545 }
1546 
1547 static void delete_dev(struct rnbd_clt_dev *dev)
1548 {
1549 	struct rnbd_clt_session *sess = dev->sess;
1550 
1551 	mutex_lock(&sess->lock);
1552 	list_del(&dev->list);
1553 	mutex_unlock(&sess->lock);
1554 }
1555 
1556 struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
1557 					   struct rtrs_addr *paths,
1558 					   size_t path_cnt, u16 port_nr,
1559 					   const char *pathname,
1560 					   enum rnbd_access_mode access_mode,
1561 					   u32 nr_poll_queues)
1562 {
1563 	struct rnbd_clt_session *sess;
1564 	struct rnbd_clt_dev *dev;
1565 	int ret;
1566 
1567 	if (exists_devpath(pathname, sessname))
1568 		return ERR_PTR(-EEXIST);
1569 
1570 	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
1571 	if (IS_ERR(sess))
1572 		return ERR_CAST(sess);
1573 
1574 	dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
1575 	if (IS_ERR(dev)) {
1576 		pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
1577 		       pathname, sess->sessname, PTR_ERR(dev));
1578 		ret = PTR_ERR(dev);
1579 		goto put_sess;
1580 	}
1581 	if (insert_dev_if_not_exists_devpath(dev)) {
1582 		ret = -EEXIST;
1583 		goto put_dev;
1584 	}
1585 	ret = send_msg_open(dev, RTRS_PERMIT_WAIT);
1586 	if (ret) {
1587 		rnbd_clt_err(dev,
1588 			      "map_device: failed, can't open remote device, err: %d\n",
1589 			      ret);
1590 		goto del_dev;
1591 	}
1592 	mutex_lock(&dev->lock);
1593 	pr_debug("Opened remote device: session=%s, path='%s'\n",
1594 		 sess->sessname, pathname);
1595 	ret = rnbd_client_setup_device(dev);
1596 	if (ret) {
1597 		rnbd_clt_err(dev,
1598 			      "map_device: Failed to configure device, err: %d\n",
1599 			      ret);
1600 		mutex_unlock(&dev->lock);
1601 		goto send_close;
1602 	}
1603 
1604 	rnbd_clt_info(dev,
1605 		       "map_device: Device mapped as %s (nsectors: %zu, 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",
1606 		       dev->gd->disk_name, dev->nsectors,
1607 		       dev->logical_block_size, dev->physical_block_size,
1608 		       dev->max_discard_sectors,
1609 		       dev->discard_granularity, dev->discard_alignment,
1610 		       dev->secure_discard, dev->max_segments,
1611 		       dev->max_hw_sectors, dev->wc, dev->fua);
1612 
1613 	mutex_unlock(&dev->lock);
1614 	rnbd_clt_put_sess(sess);
1615 
1616 	return dev;
1617 
1618 send_close:
1619 	send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1620 del_dev:
1621 	delete_dev(dev);
1622 put_dev:
1623 	rnbd_clt_put_dev(dev);
1624 put_sess:
1625 	rnbd_clt_put_sess(sess);
1626 
1627 	return ERR_PTR(ret);
1628 }
1629 
1630 static void destroy_gen_disk(struct rnbd_clt_dev *dev)
1631 {
1632 	del_gendisk(dev->gd);
1633 	blk_cleanup_disk(dev->gd);
1634 }
1635 
1636 static void destroy_sysfs(struct rnbd_clt_dev *dev,
1637 			  const struct attribute *sysfs_self)
1638 {
1639 	rnbd_clt_remove_dev_symlink(dev);
1640 	if (dev->kobj.state_initialized) {
1641 		if (sysfs_self)
1642 			/* To avoid deadlock firstly remove itself */
1643 			sysfs_remove_file_self(&dev->kobj, sysfs_self);
1644 		kobject_del(&dev->kobj);
1645 		kobject_put(&dev->kobj);
1646 	}
1647 }
1648 
1649 int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
1650 			   const struct attribute *sysfs_self)
1651 {
1652 	struct rnbd_clt_session *sess = dev->sess;
1653 	int refcount, ret = 0;
1654 	bool was_mapped;
1655 
1656 	mutex_lock(&dev->lock);
1657 	if (dev->dev_state == DEV_STATE_UNMAPPED) {
1658 		rnbd_clt_info(dev, "Device is already being unmapped\n");
1659 		ret = -EALREADY;
1660 		goto err;
1661 	}
1662 	refcount = refcount_read(&dev->refcount);
1663 	if (!force && refcount > 1) {
1664 		rnbd_clt_err(dev,
1665 			      "Closing device failed, device is in use, (%d device users)\n",
1666 			      refcount - 1);
1667 		ret = -EBUSY;
1668 		goto err;
1669 	}
1670 	was_mapped = (dev->dev_state == DEV_STATE_MAPPED);
1671 	dev->dev_state = DEV_STATE_UNMAPPED;
1672 	mutex_unlock(&dev->lock);
1673 
1674 	delete_dev(dev);
1675 	destroy_sysfs(dev, sysfs_self);
1676 	destroy_gen_disk(dev);
1677 	if (was_mapped && sess->rtrs)
1678 		send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1679 
1680 	rnbd_clt_info(dev, "Device is unmapped\n");
1681 
1682 	/* Likely last reference put */
1683 	rnbd_clt_put_dev(dev);
1684 
1685 	/*
1686 	 * Here device and session can be vanished!
1687 	 */
1688 
1689 	return 0;
1690 err:
1691 	mutex_unlock(&dev->lock);
1692 
1693 	return ret;
1694 }
1695 
1696 int rnbd_clt_remap_device(struct rnbd_clt_dev *dev)
1697 {
1698 	int err;
1699 
1700 	mutex_lock(&dev->lock);
1701 	if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED)
1702 		err = 0;
1703 	else if (dev->dev_state == DEV_STATE_UNMAPPED)
1704 		err = -ENODEV;
1705 	else if (dev->dev_state == DEV_STATE_MAPPED)
1706 		err = -EALREADY;
1707 	else
1708 		err = -EBUSY;
1709 	mutex_unlock(&dev->lock);
1710 	if (!err) {
1711 		rnbd_clt_info(dev, "Remapping device.\n");
1712 		err = send_msg_open(dev, RTRS_PERMIT_WAIT);
1713 		if (err)
1714 			rnbd_clt_err(dev, "remap_device: %d\n", err);
1715 	}
1716 
1717 	return err;
1718 }
1719 
1720 static void unmap_device_work(struct work_struct *work)
1721 {
1722 	struct rnbd_clt_dev *dev;
1723 
1724 	dev = container_of(work, typeof(*dev), unmap_on_rmmod_work);
1725 	rnbd_clt_unmap_device(dev, true, NULL);
1726 }
1727 
1728 static void rnbd_destroy_sessions(void)
1729 {
1730 	struct rnbd_clt_session *sess, *sn;
1731 	struct rnbd_clt_dev *dev, *tn;
1732 
1733 	/* Firstly forbid access through sysfs interface */
1734 	rnbd_clt_destroy_sysfs_files();
1735 
1736 	/*
1737 	 * Here at this point there is no any concurrent access to sessions
1738 	 * list and devices list:
1739 	 *   1. New session or device can't be created - session sysfs files
1740 	 *      are removed.
1741 	 *   2. Device or session can't be removed - module reference is taken
1742 	 *      into account in unmap device sysfs callback.
1743 	 *   3. No IO requests inflight - each file open of block_dev increases
1744 	 *      module reference in get_disk().
1745 	 *
1746 	 * But still there can be user requests inflights, which are sent by
1747 	 * asynchronous send_msg_*() functions, thus before unmapping devices
1748 	 * RTRS session must be explicitly closed.
1749 	 */
1750 
1751 	list_for_each_entry_safe(sess, sn, &sess_list, list) {
1752 		if (!rnbd_clt_get_sess(sess))
1753 			continue;
1754 		close_rtrs(sess);
1755 		list_for_each_entry_safe(dev, tn, &sess->devs_list, list) {
1756 			/*
1757 			 * Here unmap happens in parallel for only one reason:
1758 			 * blk_cleanup_queue() takes around half a second, so
1759 			 * on huge amount of devices the whole module unload
1760 			 * procedure takes minutes.
1761 			 */
1762 			INIT_WORK(&dev->unmap_on_rmmod_work, unmap_device_work);
1763 			queue_work(rnbd_clt_wq, &dev->unmap_on_rmmod_work);
1764 		}
1765 		rnbd_clt_put_sess(sess);
1766 	}
1767 	/* Wait for all scheduled unmap works */
1768 	flush_workqueue(rnbd_clt_wq);
1769 	WARN_ON(!list_empty(&sess_list));
1770 }
1771 
1772 static int __init rnbd_client_init(void)
1773 {
1774 	int err = 0;
1775 
1776 	BUILD_BUG_ON(sizeof(struct rnbd_msg_hdr) != 4);
1777 	BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info) != 36);
1778 	BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info_rsp) != 36);
1779 	BUILD_BUG_ON(sizeof(struct rnbd_msg_open) != 264);
1780 	BUILD_BUG_ON(sizeof(struct rnbd_msg_close) != 8);
1781 	BUILD_BUG_ON(sizeof(struct rnbd_msg_open_rsp) != 56);
1782 	rnbd_client_major = register_blkdev(rnbd_client_major, "rnbd");
1783 	if (rnbd_client_major <= 0) {
1784 		pr_err("Failed to load module, block device registration failed\n");
1785 		return -EBUSY;
1786 	}
1787 
1788 	err = rnbd_clt_create_sysfs_files();
1789 	if (err) {
1790 		pr_err("Failed to load module, creating sysfs device files failed, err: %d\n",
1791 		       err);
1792 		unregister_blkdev(rnbd_client_major, "rnbd");
1793 		return err;
1794 	}
1795 	rnbd_clt_wq = alloc_workqueue("rnbd_clt_wq", 0, 0);
1796 	if (!rnbd_clt_wq) {
1797 		pr_err("Failed to load module, alloc_workqueue failed.\n");
1798 		rnbd_clt_destroy_sysfs_files();
1799 		unregister_blkdev(rnbd_client_major, "rnbd");
1800 		err = -ENOMEM;
1801 	}
1802 
1803 	return err;
1804 }
1805 
1806 static void __exit rnbd_client_exit(void)
1807 {
1808 	rnbd_destroy_sessions();
1809 	unregister_blkdev(rnbd_client_major, "rnbd");
1810 	ida_destroy(&index_ida);
1811 	destroy_workqueue(rnbd_clt_wq);
1812 }
1813 
1814 module_init(rnbd_client_init);
1815 module_exit(rnbd_client_exit);
1816