1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * RDMA Network Block Driver
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.
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
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>
21 MODULE_DESCRIPTION("RDMA Network Block Device Client");
22 MODULE_LICENSE("GPL");
24 static int rnbd_client_major;
25 static DEFINE_IDA(index_ida);
26 static DEFINE_MUTEX(ida_lock);
27 static DEFINE_MUTEX(sess_lock);
28 static LIST_HEAD(sess_list);
31 * Maximum number of partitions an instance can have.
32 * 6 bits = 64 minors = 63 partitions (one minor is used for the device itself)
34 #define RNBD_PART_BITS 6
36 static inline bool rnbd_clt_get_sess(struct rnbd_clt_session *sess)
38 return refcount_inc_not_zero(&sess->refcount);
41 static void free_sess(struct rnbd_clt_session *sess);
43 static void rnbd_clt_put_sess(struct rnbd_clt_session *sess)
47 if (refcount_dec_and_test(&sess->refcount))
51 static void rnbd_clt_put_dev(struct rnbd_clt_dev *dev)
55 if (!refcount_dec_and_test(&dev->refcount))
58 mutex_lock(&ida_lock);
59 ida_simple_remove(&index_ida, dev->clt_device_id);
60 mutex_unlock(&ida_lock);
61 kfree(dev->hw_queues);
63 rnbd_clt_put_sess(dev->sess);
64 mutex_destroy(&dev->lock);
68 static inline bool rnbd_clt_get_dev(struct rnbd_clt_dev *dev)
70 return refcount_inc_not_zero(&dev->refcount);
73 static int rnbd_clt_set_dev_attr(struct rnbd_clt_dev *dev,
74 const struct rnbd_msg_open_rsp *rsp)
76 struct rnbd_clt_session *sess = dev->sess;
78 if (!rsp->logical_block_size)
81 dev->device_id = le32_to_cpu(rsp->device_id);
82 dev->nsectors = le64_to_cpu(rsp->nsectors);
83 dev->logical_block_size = le16_to_cpu(rsp->logical_block_size);
84 dev->physical_block_size = le16_to_cpu(rsp->physical_block_size);
85 dev->max_write_same_sectors = le32_to_cpu(rsp->max_write_same_sectors);
86 dev->max_discard_sectors = le32_to_cpu(rsp->max_discard_sectors);
87 dev->discard_granularity = le32_to_cpu(rsp->discard_granularity);
88 dev->discard_alignment = le32_to_cpu(rsp->discard_alignment);
89 dev->secure_discard = le16_to_cpu(rsp->secure_discard);
90 dev->rotational = rsp->rotational;
91 dev->wc = !!(rsp->cache_policy & RNBD_WRITEBACK);
92 dev->fua = !!(rsp->cache_policy & RNBD_FUA);
94 dev->max_hw_sectors = sess->max_io_size / SECTOR_SIZE;
95 dev->max_segments = BMAX_SEGMENTS;
100 static int rnbd_clt_change_capacity(struct rnbd_clt_dev *dev,
103 rnbd_clt_info(dev, "Device size changed from %zu to %zu sectors\n",
104 dev->nsectors, new_nsectors);
105 dev->nsectors = new_nsectors;
106 set_capacity_and_notify(dev->gd, dev->nsectors);
110 static int process_msg_open_rsp(struct rnbd_clt_dev *dev,
111 struct rnbd_msg_open_rsp *rsp)
113 struct kobject *gd_kobj;
116 mutex_lock(&dev->lock);
117 if (dev->dev_state == DEV_STATE_UNMAPPED) {
119 "Ignoring Open-Response message from server for unmapped device\n");
123 if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED) {
124 u64 nsectors = le64_to_cpu(rsp->nsectors);
127 * If the device was remapped and the size changed in the
128 * meantime we need to revalidate it
130 if (dev->nsectors != nsectors)
131 rnbd_clt_change_capacity(dev, nsectors);
132 gd_kobj = &disk_to_dev(dev->gd)->kobj;
133 kobject_uevent(gd_kobj, KOBJ_ONLINE);
134 rnbd_clt_info(dev, "Device online, device remapped successfully\n");
136 err = rnbd_clt_set_dev_attr(dev, rsp);
139 dev->dev_state = DEV_STATE_MAPPED;
142 mutex_unlock(&dev->lock);
147 int rnbd_clt_resize_disk(struct rnbd_clt_dev *dev, size_t newsize)
151 mutex_lock(&dev->lock);
152 if (dev->dev_state != DEV_STATE_MAPPED) {
153 pr_err("Failed to set new size of the device, device is not opened\n");
157 ret = rnbd_clt_change_capacity(dev, newsize);
160 mutex_unlock(&dev->lock);
165 static inline void rnbd_clt_dev_requeue(struct rnbd_queue *q)
167 if (WARN_ON(!q->hctx))
170 /* We can come here from interrupt, thus async=true */
171 blk_mq_run_hw_queue(q->hctx, true);
175 RNBD_DELAY_IFBUSY = -1,
179 * rnbd_get_cpu_qlist() - finds a list with HW queues to be rerun
180 * @sess: Session to find a queue for
181 * @cpu: Cpu to start the search from
184 * Each CPU has a list of HW queues, which needs to be rerun. If a list
185 * is not empty - it is marked with a bit. This function finds first
186 * set bit in a bitmap and returns corresponding CPU list.
188 static struct rnbd_cpu_qlist *
189 rnbd_get_cpu_qlist(struct rnbd_clt_session *sess, int cpu)
193 /* Search from cpu to nr_cpu_ids */
194 bit = find_next_bit(sess->cpu_queues_bm, nr_cpu_ids, cpu);
195 if (bit < nr_cpu_ids) {
196 return per_cpu_ptr(sess->cpu_queues, bit);
197 } else if (cpu != 0) {
198 /* Search from 0 to cpu */
199 bit = find_next_bit(sess->cpu_queues_bm, cpu, 0);
201 return per_cpu_ptr(sess->cpu_queues, bit);
207 static inline int nxt_cpu(int cpu)
209 return (cpu + 1) % nr_cpu_ids;
213 * rnbd_rerun_if_needed() - rerun next queue marked as stopped
214 * @sess: Session to rerun a queue on
217 * Each CPU has it's own list of HW queues, which should be rerun.
218 * Function finds such list with HW queues, takes a list lock, picks up
219 * the first HW queue out of the list and requeues it.
222 * True if the queue was requeued, false otherwise.
227 static bool rnbd_rerun_if_needed(struct rnbd_clt_session *sess)
229 struct rnbd_queue *q = NULL;
230 struct rnbd_cpu_qlist *cpu_q;
235 * To keep fairness and not to let other queues starve we always
236 * try to wake up someone else in round-robin manner. That of course
237 * increases latency but queues always have a chance to be executed.
239 cpup = get_cpu_ptr(sess->cpu_rr);
240 for (cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(*cpup)); cpu_q;
241 cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(cpu_q->cpu))) {
242 if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
244 if (unlikely(!test_bit(cpu_q->cpu, sess->cpu_queues_bm)))
246 q = list_first_entry_or_null(&cpu_q->requeue_list,
247 typeof(*q), requeue_list);
250 list_del_init(&q->requeue_list);
251 clear_bit_unlock(0, &q->in_list);
253 if (list_empty(&cpu_q->requeue_list)) {
254 /* Clear bit if nothing is left */
256 clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
259 spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
266 * Saves the CPU that is going to be requeued on the per-cpu var. Just
267 * incrementing it doesn't work because rnbd_get_cpu_qlist() will
268 * always return the first CPU with something on the queue list when the
269 * value stored on the var is greater than the last CPU with something
274 put_cpu_var(sess->cpu_rr);
277 rnbd_clt_dev_requeue(q);
283 * rnbd_rerun_all_if_idle() - rerun all queues left in the list if
284 * session is idling (there are no requests
286 * @sess: Session to rerun the queues on
289 * This function tries to rerun all stopped queues if there are no
290 * requests in-flight anymore. This function tries to solve an obvious
291 * problem, when number of tags < than number of queues (hctx), which
292 * are stopped and put to sleep. If last permit, which has been just put,
293 * does not wake up all left queues (hctxs), IO requests hang forever.
295 * That can happen when all number of permits, say N, have been exhausted
296 * from one CPU, and we have many block devices per session, say M.
297 * Each block device has it's own queue (hctx) for each CPU, so eventually
298 * we can put that number of queues (hctxs) to sleep: M x nr_cpu_ids.
299 * If number of permits N < M x nr_cpu_ids finally we will get an IO hang.
301 * To avoid this hang last caller of rnbd_put_permit() (last caller is the
302 * one who observes sess->busy == 0) must wake up all remaining queues.
307 static void rnbd_rerun_all_if_idle(struct rnbd_clt_session *sess)
312 requeued = rnbd_rerun_if_needed(sess);
313 } while (atomic_read(&sess->busy) == 0 && requeued);
316 static struct rtrs_permit *rnbd_get_permit(struct rnbd_clt_session *sess,
317 enum rtrs_clt_con_type con_type,
320 struct rtrs_permit *permit;
322 permit = rtrs_clt_get_permit(sess->rtrs, con_type, wait);
324 /* We have a subtle rare case here, when all permits can be
325 * consumed before busy counter increased. This is safe,
326 * because loser will get NULL as a permit, observe 0 busy
327 * counter and immediately restart the queue himself.
329 atomic_inc(&sess->busy);
334 static void rnbd_put_permit(struct rnbd_clt_session *sess,
335 struct rtrs_permit *permit)
337 rtrs_clt_put_permit(sess->rtrs, permit);
338 atomic_dec(&sess->busy);
339 /* Paired with rnbd_clt_dev_add_to_requeue(). Decrement first
340 * and then check queue bits.
342 smp_mb__after_atomic();
343 rnbd_rerun_all_if_idle(sess);
346 static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
347 enum rtrs_clt_con_type con_type,
351 struct rtrs_permit *permit;
353 iu = kzalloc(sizeof(*iu), GFP_KERNEL);
358 permit = rnbd_get_permit(sess, con_type, wait);
359 if (unlikely(!permit)) {
366 * 1st reference is dropped after finishing sending a "user" message,
367 * 2nd reference is dropped after confirmation with the response is
369 * 1st and 2nd can happen in any order, so the rnbd_iu should be
370 * released (rtrs_permit returned to rtrs) only after both
373 atomic_set(&iu->refcount, 2);
374 init_waitqueue_head(&iu->comp.wait);
375 iu->comp.errno = INT_MAX;
377 if (sg_alloc_table(&iu->sgt, 1, GFP_KERNEL)) {
378 rnbd_put_permit(sess, permit);
386 static void rnbd_put_iu(struct rnbd_clt_session *sess, struct rnbd_iu *iu)
388 if (atomic_dec_and_test(&iu->refcount)) {
389 sg_free_table(&iu->sgt);
390 rnbd_put_permit(sess, iu->permit);
395 static void rnbd_softirq_done_fn(struct request *rq)
397 struct rnbd_clt_dev *dev = rq->rq_disk->private_data;
398 struct rnbd_clt_session *sess = dev->sess;
401 iu = blk_mq_rq_to_pdu(rq);
402 sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
403 rnbd_put_permit(sess, iu->permit);
404 blk_mq_end_request(rq, errno_to_blk_status(iu->errno));
407 static void msg_io_conf(void *priv, int errno)
409 struct rnbd_iu *iu = priv;
410 struct rnbd_clt_dev *dev = iu->dev;
411 struct request *rq = iu->rq;
412 int rw = rq_data_dir(rq);
416 blk_mq_complete_request(rq);
419 rnbd_clt_info_rl(dev, "%s I/O failed with err: %d\n",
420 rw == READ ? "read" : "write", errno);
423 static void wake_up_iu_comp(struct rnbd_iu *iu, int errno)
425 iu->comp.errno = errno;
426 wake_up(&iu->comp.wait);
429 static void msg_conf(void *priv, int errno)
431 struct rnbd_iu *iu = priv;
434 schedule_work(&iu->work);
437 static int send_usr_msg(struct rtrs_clt *rtrs, int dir,
438 struct rnbd_iu *iu, struct kvec *vec,
439 size_t len, struct scatterlist *sg, unsigned int sg_len,
440 void (*conf)(struct work_struct *work),
441 int *errno, int wait)
444 struct rtrs_clt_req_ops req_ops;
446 INIT_WORK(&iu->work, conf);
447 req_ops = (struct rtrs_clt_req_ops) {
451 err = rtrs_clt_request(dir, &req_ops, rtrs, iu->permit,
452 vec, 1, len, sg, sg_len);
454 wait_event(iu->comp.wait, iu->comp.errno != INT_MAX);
455 *errno = iu->comp.errno;
463 static void msg_close_conf(struct work_struct *work)
465 struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
466 struct rnbd_clt_dev *dev = iu->dev;
468 wake_up_iu_comp(iu, iu->errno);
469 rnbd_put_iu(dev->sess, iu);
470 rnbd_clt_put_dev(dev);
473 static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id,
476 struct rnbd_clt_session *sess = dev->sess;
477 struct rnbd_msg_close msg;
481 .iov_len = sizeof(msg)
485 iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
492 msg.hdr.type = cpu_to_le16(RNBD_MSG_CLOSE);
493 msg.device_id = cpu_to_le32(device_id);
495 WARN_ON(!rnbd_clt_get_dev(dev));
496 err = send_usr_msg(sess->rtrs, WRITE, iu, &vec, 0, NULL, 0,
497 msg_close_conf, &errno, wait);
499 rnbd_clt_put_dev(dev);
500 rnbd_put_iu(sess, iu);
505 rnbd_put_iu(sess, iu);
509 static void msg_open_conf(struct work_struct *work)
511 struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
512 struct rnbd_msg_open_rsp *rsp = iu->buf;
513 struct rnbd_clt_dev *dev = iu->dev;
514 int errno = iu->errno;
518 "Opening failed, server responded: %d\n",
521 errno = process_msg_open_rsp(dev, rsp);
523 u32 device_id = le32_to_cpu(rsp->device_id);
525 * If server thinks its fine, but we fail to process
526 * then be nice and send a close to server.
528 send_msg_close(dev, device_id, RTRS_PERMIT_NOWAIT);
532 wake_up_iu_comp(iu, errno);
533 rnbd_put_iu(dev->sess, iu);
534 rnbd_clt_put_dev(dev);
537 static void msg_sess_info_conf(struct work_struct *work)
539 struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
540 struct rnbd_msg_sess_info_rsp *rsp = iu->buf;
541 struct rnbd_clt_session *sess = iu->sess;
544 sess->ver = min_t(u8, rsp->ver, RNBD_PROTO_VER_MAJOR);
547 wake_up_iu_comp(iu, iu->errno);
548 rnbd_put_iu(sess, iu);
549 rnbd_clt_put_sess(sess);
552 static int send_msg_open(struct rnbd_clt_dev *dev, enum wait_type wait)
554 struct rnbd_clt_session *sess = dev->sess;
555 struct rnbd_msg_open_rsp *rsp;
556 struct rnbd_msg_open msg;
560 .iov_len = sizeof(msg)
564 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
568 iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
577 sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
579 msg.hdr.type = cpu_to_le16(RNBD_MSG_OPEN);
580 msg.access_mode = dev->access_mode;
581 strlcpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
583 WARN_ON(!rnbd_clt_get_dev(dev));
584 err = send_usr_msg(sess->rtrs, READ, iu,
585 &vec, sizeof(*rsp), iu->sgt.sgl, 1,
586 msg_open_conf, &errno, wait);
588 rnbd_clt_put_dev(dev);
589 rnbd_put_iu(sess, iu);
595 rnbd_put_iu(sess, iu);
599 static int send_msg_sess_info(struct rnbd_clt_session *sess, enum wait_type wait)
601 struct rnbd_msg_sess_info_rsp *rsp;
602 struct rnbd_msg_sess_info msg;
606 .iov_len = sizeof(msg)
610 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
614 iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
622 sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
624 msg.hdr.type = cpu_to_le16(RNBD_MSG_SESS_INFO);
625 msg.ver = RNBD_PROTO_VER_MAJOR;
627 if (!rnbd_clt_get_sess(sess)) {
629 * That can happen only in one case, when RTRS has restablished
630 * the connection and link_ev() is called, but session is almost
631 * dead, last reference on session is put and caller is waiting
632 * for RTRS to close everything.
637 err = send_usr_msg(sess->rtrs, READ, iu,
638 &vec, sizeof(*rsp), iu->sgt.sgl, 1,
639 msg_sess_info_conf, &errno, wait);
641 rnbd_clt_put_sess(sess);
643 rnbd_put_iu(sess, iu);
648 rnbd_put_iu(sess, iu);
652 static void set_dev_states_to_disconnected(struct rnbd_clt_session *sess)
654 struct rnbd_clt_dev *dev;
655 struct kobject *gd_kobj;
657 mutex_lock(&sess->lock);
658 list_for_each_entry(dev, &sess->devs_list, list) {
659 rnbd_clt_err(dev, "Device disconnected.\n");
661 mutex_lock(&dev->lock);
662 if (dev->dev_state == DEV_STATE_MAPPED) {
663 dev->dev_state = DEV_STATE_MAPPED_DISCONNECTED;
664 gd_kobj = &disk_to_dev(dev->gd)->kobj;
665 kobject_uevent(gd_kobj, KOBJ_OFFLINE);
667 mutex_unlock(&dev->lock);
669 mutex_unlock(&sess->lock);
672 static void remap_devs(struct rnbd_clt_session *sess)
674 struct rnbd_clt_dev *dev;
675 struct rtrs_attrs attrs;
679 * Careful here: we are called from RTRS link event directly,
680 * thus we can't send any RTRS request and wait for response
681 * or RTRS will not be able to complete request with failure
682 * if something goes wrong (failing of outstanding requests
683 * happens exactly from the context where we are blocking now).
685 * So to avoid deadlocks each usr message sent from here must
689 err = send_msg_sess_info(sess, RTRS_PERMIT_NOWAIT);
691 pr_err("send_msg_sess_info(\"%s\"): %d\n", sess->sessname, err);
695 rtrs_clt_query(sess->rtrs, &attrs);
696 mutex_lock(&sess->lock);
697 sess->max_io_size = attrs.max_io_size;
699 list_for_each_entry(dev, &sess->devs_list, list) {
702 mutex_lock(&dev->lock);
703 skip = (dev->dev_state == DEV_STATE_INIT);
704 mutex_unlock(&dev->lock);
707 * When device is establishing connection for the first
708 * time - do not remap, it will be closed soon.
712 rnbd_clt_info(dev, "session reconnected, remapping device\n");
713 err = send_msg_open(dev, RTRS_PERMIT_NOWAIT);
715 rnbd_clt_err(dev, "send_msg_open(): %d\n", err);
719 mutex_unlock(&sess->lock);
722 static void rnbd_clt_link_ev(void *priv, enum rtrs_clt_link_ev ev)
724 struct rnbd_clt_session *sess = priv;
727 case RTRS_CLT_LINK_EV_DISCONNECTED:
728 set_dev_states_to_disconnected(sess);
730 case RTRS_CLT_LINK_EV_RECONNECTED:
734 pr_err("Unknown session event received (%d), session: %s\n",
739 static void rnbd_init_cpu_qlists(struct rnbd_cpu_qlist __percpu *cpu_queues)
742 struct rnbd_cpu_qlist *cpu_q;
744 for_each_possible_cpu(cpu) {
745 cpu_q = per_cpu_ptr(cpu_queues, cpu);
748 INIT_LIST_HEAD(&cpu_q->requeue_list);
749 spin_lock_init(&cpu_q->requeue_lock);
753 static void destroy_mq_tags(struct rnbd_clt_session *sess)
755 if (sess->tag_set.tags)
756 blk_mq_free_tag_set(&sess->tag_set);
759 static inline void wake_up_rtrs_waiters(struct rnbd_clt_session *sess)
761 sess->rtrs_ready = true;
762 wake_up_all(&sess->rtrs_waitq);
765 static void close_rtrs(struct rnbd_clt_session *sess)
769 if (!IS_ERR_OR_NULL(sess->rtrs)) {
770 rtrs_clt_close(sess->rtrs);
772 wake_up_rtrs_waiters(sess);
776 static void free_sess(struct rnbd_clt_session *sess)
778 WARN_ON(!list_empty(&sess->devs_list));
783 destroy_mq_tags(sess);
784 if (!list_empty(&sess->list)) {
785 mutex_lock(&sess_lock);
786 list_del(&sess->list);
787 mutex_unlock(&sess_lock);
789 free_percpu(sess->cpu_queues);
790 free_percpu(sess->cpu_rr);
791 mutex_destroy(&sess->lock);
795 static struct rnbd_clt_session *alloc_sess(const char *sessname)
797 struct rnbd_clt_session *sess;
800 sess = kzalloc_node(sizeof(*sess), GFP_KERNEL, NUMA_NO_NODE);
802 return ERR_PTR(-ENOMEM);
803 strlcpy(sess->sessname, sessname, sizeof(sess->sessname));
804 atomic_set(&sess->busy, 0);
805 mutex_init(&sess->lock);
806 INIT_LIST_HEAD(&sess->devs_list);
807 INIT_LIST_HEAD(&sess->list);
808 bitmap_zero(sess->cpu_queues_bm, NR_CPUS);
809 init_waitqueue_head(&sess->rtrs_waitq);
810 refcount_set(&sess->refcount, 1);
812 sess->cpu_queues = alloc_percpu(struct rnbd_cpu_qlist);
813 if (!sess->cpu_queues) {
817 rnbd_init_cpu_qlists(sess->cpu_queues);
820 * That is simple percpu variable which stores cpu indices, which are
821 * incremented on each access. We need that for the sake of fairness
822 * to wake up queues in a round-robin manner.
824 sess->cpu_rr = alloc_percpu(int);
829 for_each_possible_cpu(cpu)
830 * per_cpu_ptr(sess->cpu_rr, cpu) = cpu;
840 static int wait_for_rtrs_connection(struct rnbd_clt_session *sess)
842 wait_event(sess->rtrs_waitq, sess->rtrs_ready);
843 if (IS_ERR_OR_NULL(sess->rtrs))
849 static void wait_for_rtrs_disconnection(struct rnbd_clt_session *sess)
850 __releases(&sess_lock)
851 __acquires(&sess_lock)
855 prepare_to_wait(&sess->rtrs_waitq, &wait, TASK_UNINTERRUPTIBLE);
856 if (IS_ERR_OR_NULL(sess->rtrs)) {
857 finish_wait(&sess->rtrs_waitq, &wait);
860 mutex_unlock(&sess_lock);
861 /* loop in caller, see __find_and_get_sess().
862 * You can't leave mutex locked and call schedule(), you will catch a
863 * deadlock with a caller of free_sess(), which has just put the last
864 * reference and is about to take the sess_lock in order to delete
865 * the session from the list.
868 mutex_lock(&sess_lock);
871 static struct rnbd_clt_session *__find_and_get_sess(const char *sessname)
872 __releases(&sess_lock)
873 __acquires(&sess_lock)
875 struct rnbd_clt_session *sess, *sn;
879 list_for_each_entry_safe(sess, sn, &sess_list, list) {
880 if (strcmp(sessname, sess->sessname))
883 if (sess->rtrs_ready && IS_ERR_OR_NULL(sess->rtrs))
885 * No RTRS connection, session is dying.
889 if (rnbd_clt_get_sess(sess)) {
891 * Alive session is found, wait for RTRS connection.
893 mutex_unlock(&sess_lock);
894 err = wait_for_rtrs_connection(sess);
896 rnbd_clt_put_sess(sess);
897 mutex_lock(&sess_lock);
900 /* Session is dying, repeat the loop */
906 * Ref is 0, session is dying, wait for RTRS disconnect
907 * in order to avoid session names clashes.
909 wait_for_rtrs_disconnection(sess);
911 * RTRS is disconnected and soon session will be freed,
920 /* caller is responsible for initializing 'first' to false */
922 rnbd_clt_session *find_or_create_sess(const char *sessname, bool *first)
924 struct rnbd_clt_session *sess = NULL;
926 mutex_lock(&sess_lock);
927 sess = __find_and_get_sess(sessname);
929 sess = alloc_sess(sessname);
931 mutex_unlock(&sess_lock);
934 list_add(&sess->list, &sess_list);
937 mutex_unlock(&sess_lock);
942 static int rnbd_client_open(struct block_device *block_device, fmode_t mode)
944 struct rnbd_clt_dev *dev = block_device->bd_disk->private_data;
946 if (dev->read_only && (mode & FMODE_WRITE))
949 if (dev->dev_state == DEV_STATE_UNMAPPED ||
950 !rnbd_clt_get_dev(dev))
956 static void rnbd_client_release(struct gendisk *gen, fmode_t mode)
958 struct rnbd_clt_dev *dev = gen->private_data;
960 rnbd_clt_put_dev(dev);
963 static int rnbd_client_getgeo(struct block_device *block_device,
964 struct hd_geometry *geo)
967 struct rnbd_clt_dev *dev;
969 dev = block_device->bd_disk->private_data;
970 size = dev->size * (dev->logical_block_size / SECTOR_SIZE);
971 geo->cylinders = size >> 6; /* size/64 */
979 static const struct block_device_operations rnbd_client_ops = {
980 .owner = THIS_MODULE,
981 .open = rnbd_client_open,
982 .release = rnbd_client_release,
983 .getgeo = rnbd_client_getgeo
986 /* The amount of data that belongs to an I/O and the amount of data that
987 * should be read or written to the disk (bi_size) can differ.
989 * E.g. When WRITE_SAME is used, only a small amount of data is
990 * transferred that is then written repeatedly over a lot of sectors.
992 * Get the size of data to be transferred via RTRS by summing up the size
993 * of the scather-gather list entries.
995 static size_t rnbd_clt_get_sg_size(struct scatterlist *sglist, u32 len)
997 struct scatterlist *sg;
1001 for_each_sg(sglist, sg, len, i)
1002 tsize += sg->length;
1006 static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
1010 struct rtrs_clt *rtrs = dev->sess->rtrs;
1011 struct rtrs_permit *permit = iu->permit;
1012 struct rnbd_msg_io msg;
1013 struct rtrs_clt_req_ops req_ops;
1014 unsigned int sg_cnt = 0;
1021 msg.sector = cpu_to_le64(blk_rq_pos(rq));
1022 msg.bi_size = cpu_to_le32(blk_rq_bytes(rq));
1023 msg.rw = cpu_to_le32(rq_to_rnbd_flags(rq));
1024 msg.prio = cpu_to_le16(req_get_ioprio(rq));
1027 * We only support discards with single segment for now.
1030 if (req_op(rq) != REQ_OP_DISCARD)
1031 sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sgt.sgl);
1034 sg_mark_end(&iu->sgt.sgl[0]);
1036 msg.hdr.type = cpu_to_le16(RNBD_MSG_IO);
1037 msg.device_id = cpu_to_le32(dev->device_id);
1039 vec = (struct kvec) {
1041 .iov_len = sizeof(msg)
1043 size = rnbd_clt_get_sg_size(iu->sgt.sgl, sg_cnt);
1044 req_ops = (struct rtrs_clt_req_ops) {
1046 .conf_fn = msg_io_conf,
1048 err = rtrs_clt_request(rq_data_dir(rq), &req_ops, rtrs, permit,
1049 &vec, 1, size, iu->sgt.sgl, sg_cnt);
1050 if (unlikely(err)) {
1051 rnbd_clt_err_rl(dev, "RTRS failed to transfer IO, err: %d\n",
1060 * rnbd_clt_dev_add_to_requeue() - add device to requeue if session is busy
1061 * @dev: Device to be checked
1062 * @q: Queue to be added to the requeue list if required
1065 * If session is busy, that means someone will requeue us when resources
1066 * are freed. If session is not doing anything - device is not added to
1067 * the list and @false is returned.
1069 static bool rnbd_clt_dev_add_to_requeue(struct rnbd_clt_dev *dev,
1070 struct rnbd_queue *q)
1072 struct rnbd_clt_session *sess = dev->sess;
1073 struct rnbd_cpu_qlist *cpu_q;
1074 unsigned long flags;
1078 cpu_q = get_cpu_ptr(sess->cpu_queues);
1079 spin_lock_irqsave(&cpu_q->requeue_lock, flags);
1081 if (likely(!test_and_set_bit_lock(0, &q->in_list))) {
1082 if (WARN_ON(!list_empty(&q->requeue_list)))
1085 need_set = !test_bit(cpu_q->cpu, sess->cpu_queues_bm);
1087 set_bit(cpu_q->cpu, sess->cpu_queues_bm);
1088 /* Paired with rnbd_put_permit(). Set a bit first
1089 * and then observe the busy counter.
1091 smp_mb__before_atomic();
1093 if (likely(atomic_read(&sess->busy))) {
1094 list_add_tail(&q->requeue_list, &cpu_q->requeue_list);
1096 /* Very unlikely, but possible: busy counter was
1097 * observed as zero. Drop all bits and return
1098 * false to restart the queue by ourselves.
1101 clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
1102 clear_bit_unlock(0, &q->in_list);
1107 spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
1108 put_cpu_ptr(sess->cpu_queues);
1113 static void rnbd_clt_dev_kick_mq_queue(struct rnbd_clt_dev *dev,
1114 struct blk_mq_hw_ctx *hctx,
1117 struct rnbd_queue *q = hctx->driver_data;
1119 if (delay != RNBD_DELAY_IFBUSY)
1120 blk_mq_delay_run_hw_queue(hctx, delay);
1121 else if (unlikely(!rnbd_clt_dev_add_to_requeue(dev, q)))
1123 * If session is not busy we have to restart
1124 * the queue ourselves.
1126 blk_mq_delay_run_hw_queue(hctx, 10/*ms*/);
1129 static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1130 const struct blk_mq_queue_data *bd)
1132 struct request *rq = bd->rq;
1133 struct rnbd_clt_dev *dev = rq->rq_disk->private_data;
1134 struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
1136 blk_status_t ret = BLK_STS_IOERR;
1138 if (unlikely(dev->dev_state != DEV_STATE_MAPPED))
1139 return BLK_STS_IOERR;
1141 iu->permit = rnbd_get_permit(dev->sess, RTRS_IO_CON,
1142 RTRS_PERMIT_NOWAIT);
1143 if (unlikely(!iu->permit)) {
1144 rnbd_clt_dev_kick_mq_queue(dev, hctx, RNBD_DELAY_IFBUSY);
1145 return BLK_STS_RESOURCE;
1148 iu->sgt.sgl = iu->first_sgl;
1149 err = sg_alloc_table_chained(&iu->sgt,
1150 /* Even-if the request has no segment,
1151 * sglist must have one entry at least */
1152 blk_rq_nr_phys_segments(rq) ? : 1,
1154 RNBD_INLINE_SG_CNT);
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;
1162 blk_mq_start_request(rq);
1163 err = rnbd_client_xfer_request(dev, rq, iu);
1164 if (likely(err == 0))
1166 if (unlikely(err == -EAGAIN || err == -ENOMEM)) {
1167 rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1168 ret = BLK_STS_RESOURCE;
1170 sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
1171 rnbd_put_permit(dev->sess, iu->permit);
1175 static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx)
1177 struct rnbd_queue *q = hctx->driver_data;
1178 struct rnbd_clt_dev *dev = q->dev;
1181 cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
1185 static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
1187 struct rnbd_clt_session *sess = set->driver_data;
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]);
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",
1205 set->map[HCTX_TYPE_DEFAULT].nr_queues,
1206 set->map[HCTX_TYPE_READ].nr_queues,
1207 set->map[HCTX_TYPE_POLL].nr_queues);
1209 pr_info("[session=%s] mapped %d/%d default/read queues.\n",
1211 set->map[HCTX_TYPE_DEFAULT].nr_queues,
1212 set->map[HCTX_TYPE_READ].nr_queues);
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,
1225 static int setup_mq_tags(struct rnbd_clt_session *sess)
1227 struct blk_mq_tag_set *tag_set = &sess->tag_set;
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;
1237 /* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
1238 tag_set->nr_maps = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1240 * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
1241 * others are for HCTX_TYPE_POLL
1243 tag_set->nr_hw_queues = num_online_cpus() + sess->nr_poll_queues;
1244 tag_set->driver_data = sess;
1246 return blk_mq_alloc_tag_set(tag_set);
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)
1254 struct rnbd_clt_session *sess;
1255 struct rtrs_attrs attrs;
1258 struct rtrs_clt_ops rtrs_ops;
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)) {
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.
1276 pr_err("Session %s not found, and path parameter not given", sessname);
1281 rtrs_ops = (struct rtrs_clt_ops) {
1283 .link_ev = rnbd_clt_link_ev,
1286 * Nothing was found, establish rtrs connection and proceed further.
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, BMAX_SEGMENTS,
1292 BLK_MAX_SEGMENT_SIZE,
1293 MAX_RECONNECTS, nr_poll_queues);
1294 if (IS_ERR(sess->rtrs)) {
1295 err = PTR_ERR(sess->rtrs);
1296 goto wake_up_and_put;
1298 rtrs_clt_query(sess->rtrs, &attrs);
1299 sess->max_io_size = attrs.max_io_size;
1300 sess->queue_depth = attrs.queue_depth;
1301 sess->nr_poll_queues = nr_poll_queues;
1303 err = setup_mq_tags(sess);
1307 err = send_msg_sess_info(sess, RTRS_PERMIT_WAIT);
1311 wake_up_rtrs_waiters(sess);
1318 rnbd_clt_put_sess(sess);
1320 return ERR_PTR(err);
1323 wake_up_rtrs_waiters(sess);
1327 static inline void rnbd_init_hw_queue(struct rnbd_clt_dev *dev,
1328 struct rnbd_queue *q,
1329 struct blk_mq_hw_ctx *hctx)
1331 INIT_LIST_HEAD(&q->requeue_list);
1336 static void rnbd_init_mq_hw_queues(struct rnbd_clt_dev *dev)
1339 struct blk_mq_hw_ctx *hctx;
1340 struct rnbd_queue *q;
1342 queue_for_each_hw_ctx(dev->queue, hctx, i) {
1343 q = &dev->hw_queues[i];
1344 rnbd_init_hw_queue(dev, q, hctx);
1345 hctx->driver_data = q;
1349 static int setup_mq_dev(struct rnbd_clt_dev *dev)
1351 dev->queue = blk_mq_init_queue(&dev->sess->tag_set);
1352 if (IS_ERR(dev->queue)) {
1353 rnbd_clt_err(dev, "Initializing multiqueue queue failed, err: %ld\n",
1354 PTR_ERR(dev->queue));
1355 return PTR_ERR(dev->queue);
1357 rnbd_init_mq_hw_queues(dev);
1361 static void setup_request_queue(struct rnbd_clt_dev *dev)
1363 blk_queue_logical_block_size(dev->queue, dev->logical_block_size);
1364 blk_queue_physical_block_size(dev->queue, dev->physical_block_size);
1365 blk_queue_max_hw_sectors(dev->queue, dev->max_hw_sectors);
1366 blk_queue_max_write_same_sectors(dev->queue,
1367 dev->max_write_same_sectors);
1370 * we don't support discards to "discontiguous" segments
1373 blk_queue_max_discard_segments(dev->queue, 1);
1375 blk_queue_max_discard_sectors(dev->queue, dev->max_discard_sectors);
1376 dev->queue->limits.discard_granularity = dev->discard_granularity;
1377 dev->queue->limits.discard_alignment = dev->discard_alignment;
1378 if (dev->max_discard_sectors)
1379 blk_queue_flag_set(QUEUE_FLAG_DISCARD, dev->queue);
1380 if (dev->secure_discard)
1381 blk_queue_flag_set(QUEUE_FLAG_SECERASE, dev->queue);
1383 blk_queue_flag_set(QUEUE_FLAG_SAME_COMP, dev->queue);
1384 blk_queue_flag_set(QUEUE_FLAG_SAME_FORCE, dev->queue);
1385 blk_queue_max_segments(dev->queue, dev->max_segments);
1386 blk_queue_io_opt(dev->queue, dev->sess->max_io_size);
1387 blk_queue_virt_boundary(dev->queue, SZ_4K - 1);
1388 blk_queue_write_cache(dev->queue, dev->wc, dev->fua);
1389 dev->queue->queuedata = dev;
1392 static void rnbd_clt_setup_gen_disk(struct rnbd_clt_dev *dev, int idx)
1394 dev->gd->major = rnbd_client_major;
1395 dev->gd->first_minor = idx << RNBD_PART_BITS;
1396 dev->gd->fops = &rnbd_client_ops;
1397 dev->gd->queue = dev->queue;
1398 dev->gd->private_data = dev;
1399 snprintf(dev->gd->disk_name, sizeof(dev->gd->disk_name), "rnbd%d",
1401 pr_debug("disk_name=%s, capacity=%zu\n",
1403 dev->nsectors * (dev->logical_block_size / SECTOR_SIZE)
1406 set_capacity(dev->gd, dev->nsectors);
1408 if (dev->access_mode == RNBD_ACCESS_RO) {
1409 dev->read_only = true;
1410 set_disk_ro(dev->gd, true);
1412 dev->read_only = false;
1415 if (!dev->rotational)
1416 blk_queue_flag_set(QUEUE_FLAG_NONROT, dev->queue);
1420 static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
1422 int err, idx = dev->clt_device_id;
1424 dev->size = dev->nsectors * dev->logical_block_size;
1426 err = setup_mq_dev(dev);
1430 setup_request_queue(dev);
1432 dev->gd = alloc_disk_node(1 << RNBD_PART_BITS, NUMA_NO_NODE);
1434 blk_cleanup_queue(dev->queue);
1438 rnbd_clt_setup_gen_disk(dev, idx);
1443 static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
1444 enum rnbd_access_mode access_mode,
1445 const char *pathname,
1448 struct rnbd_clt_dev *dev;
1451 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, NUMA_NO_NODE);
1453 return ERR_PTR(-ENOMEM);
1456 * nr_cpu_ids: the number of softirq queues
1457 * nr_poll_queues: the number of polling queues
1459 dev->hw_queues = kcalloc(nr_cpu_ids + nr_poll_queues,
1460 sizeof(*dev->hw_queues),
1462 if (!dev->hw_queues) {
1467 mutex_lock(&ida_lock);
1468 ret = ida_simple_get(&index_ida, 0, 1 << (MINORBITS - RNBD_PART_BITS),
1470 mutex_unlock(&ida_lock);
1472 pr_err("Failed to initialize device '%s' from session %s, allocating idr failed, err: %d\n",
1473 pathname, sess->sessname, ret);
1477 dev->pathname = kstrdup(pathname, GFP_KERNEL);
1478 if (!dev->pathname) {
1483 dev->clt_device_id = ret;
1485 dev->access_mode = access_mode;
1486 dev->nr_poll_queues = nr_poll_queues;
1487 mutex_init(&dev->lock);
1488 refcount_set(&dev->refcount, 1);
1489 dev->dev_state = DEV_STATE_INIT;
1492 * Here we called from sysfs entry, thus clt-sysfs is
1493 * responsible that session will not disappear.
1495 WARN_ON(!rnbd_clt_get_sess(sess));
1500 kfree(dev->hw_queues);
1503 return ERR_PTR(ret);
1506 static bool __exists_dev(const char *pathname, const char *sessname)
1508 struct rnbd_clt_session *sess;
1509 struct rnbd_clt_dev *dev;
1512 list_for_each_entry(sess, &sess_list, list) {
1513 if (sessname && strncmp(sess->sessname, sessname,
1514 sizeof(sess->sessname)))
1516 mutex_lock(&sess->lock);
1517 list_for_each_entry(dev, &sess->devs_list, list) {
1518 if (strlen(dev->pathname) == strlen(pathname) &&
1519 !strcmp(dev->pathname, pathname)) {
1524 mutex_unlock(&sess->lock);
1532 static bool exists_devpath(const char *pathname, const char *sessname)
1536 mutex_lock(&sess_lock);
1537 found = __exists_dev(pathname, sessname);
1538 mutex_unlock(&sess_lock);
1543 static bool insert_dev_if_not_exists_devpath(struct rnbd_clt_dev *dev)
1546 struct rnbd_clt_session *sess = dev->sess;
1548 mutex_lock(&sess_lock);
1549 found = __exists_dev(dev->pathname, sess->sessname);
1551 mutex_lock(&sess->lock);
1552 list_add_tail(&dev->list, &sess->devs_list);
1553 mutex_unlock(&sess->lock);
1555 mutex_unlock(&sess_lock);
1560 static void delete_dev(struct rnbd_clt_dev *dev)
1562 struct rnbd_clt_session *sess = dev->sess;
1564 mutex_lock(&sess->lock);
1565 list_del(&dev->list);
1566 mutex_unlock(&sess->lock);
1569 struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
1570 struct rtrs_addr *paths,
1571 size_t path_cnt, u16 port_nr,
1572 const char *pathname,
1573 enum rnbd_access_mode access_mode,
1576 struct rnbd_clt_session *sess;
1577 struct rnbd_clt_dev *dev;
1580 if (unlikely(exists_devpath(pathname, sessname)))
1581 return ERR_PTR(-EEXIST);
1583 sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
1585 return ERR_CAST(sess);
1587 dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
1589 pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
1590 pathname, sess->sessname, PTR_ERR(dev));
1594 if (insert_dev_if_not_exists_devpath(dev)) {
1598 ret = send_msg_open(dev, RTRS_PERMIT_WAIT);
1601 "map_device: failed, can't open remote device, err: %d\n",
1605 mutex_lock(&dev->lock);
1606 pr_debug("Opened remote device: session=%s, path='%s'\n",
1607 sess->sessname, pathname);
1608 ret = rnbd_client_setup_device(dev);
1611 "map_device: Failed to configure device, err: %d\n",
1613 mutex_unlock(&dev->lock);
1618 "map_device: Device mapped as %s (nsectors: %zu, logical_block_size: %d, physical_block_size: %d, max_write_same_sectors: %d, max_discard_sectors: %d, discard_granularity: %d, discard_alignment: %d, secure_discard: %d, max_segments: %d, max_hw_sectors: %d, rotational: %d, wc: %d, fua: %d)\n",
1619 dev->gd->disk_name, dev->nsectors,
1620 dev->logical_block_size, dev->physical_block_size,
1621 dev->max_write_same_sectors, dev->max_discard_sectors,
1622 dev->discard_granularity, dev->discard_alignment,
1623 dev->secure_discard, dev->max_segments,
1624 dev->max_hw_sectors, dev->rotational, dev->wc, dev->fua);
1626 mutex_unlock(&dev->lock);
1627 rnbd_clt_put_sess(sess);
1632 send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1636 rnbd_clt_put_dev(dev);
1638 rnbd_clt_put_sess(sess);
1640 return ERR_PTR(ret);
1643 static void destroy_gen_disk(struct rnbd_clt_dev *dev)
1645 del_gendisk(dev->gd);
1646 blk_cleanup_queue(dev->queue);
1650 static void destroy_sysfs(struct rnbd_clt_dev *dev,
1651 const struct attribute *sysfs_self)
1653 rnbd_clt_remove_dev_symlink(dev);
1654 if (dev->kobj.state_initialized) {
1656 /* To avoid deadlock firstly remove itself */
1657 sysfs_remove_file_self(&dev->kobj, sysfs_self);
1658 kobject_del(&dev->kobj);
1659 kobject_put(&dev->kobj);
1663 int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
1664 const struct attribute *sysfs_self)
1666 struct rnbd_clt_session *sess = dev->sess;
1667 int refcount, ret = 0;
1670 mutex_lock(&dev->lock);
1671 if (dev->dev_state == DEV_STATE_UNMAPPED) {
1672 rnbd_clt_info(dev, "Device is already being unmapped\n");
1676 refcount = refcount_read(&dev->refcount);
1677 if (!force && refcount > 1) {
1679 "Closing device failed, device is in use, (%d device users)\n",
1684 was_mapped = (dev->dev_state == DEV_STATE_MAPPED);
1685 dev->dev_state = DEV_STATE_UNMAPPED;
1686 mutex_unlock(&dev->lock);
1689 destroy_sysfs(dev, sysfs_self);
1690 destroy_gen_disk(dev);
1691 if (was_mapped && sess->rtrs)
1692 send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1694 rnbd_clt_info(dev, "Device is unmapped\n");
1696 /* Likely last reference put */
1697 rnbd_clt_put_dev(dev);
1700 * Here device and session can be vanished!
1705 mutex_unlock(&dev->lock);
1710 int rnbd_clt_remap_device(struct rnbd_clt_dev *dev)
1714 mutex_lock(&dev->lock);
1715 if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED)
1717 else if (dev->dev_state == DEV_STATE_UNMAPPED)
1719 else if (dev->dev_state == DEV_STATE_MAPPED)
1723 mutex_unlock(&dev->lock);
1725 rnbd_clt_info(dev, "Remapping device.\n");
1726 err = send_msg_open(dev, RTRS_PERMIT_WAIT);
1728 rnbd_clt_err(dev, "remap_device: %d\n", err);
1734 static void unmap_device_work(struct work_struct *work)
1736 struct rnbd_clt_dev *dev;
1738 dev = container_of(work, typeof(*dev), unmap_on_rmmod_work);
1739 rnbd_clt_unmap_device(dev, true, NULL);
1742 static void rnbd_destroy_sessions(void)
1744 struct rnbd_clt_session *sess, *sn;
1745 struct rnbd_clt_dev *dev, *tn;
1747 /* Firstly forbid access through sysfs interface */
1748 rnbd_clt_destroy_sysfs_files();
1751 * Here at this point there is no any concurrent access to sessions
1752 * list and devices list:
1753 * 1. New session or device can't be created - session sysfs files
1755 * 2. Device or session can't be removed - module reference is taken
1756 * into account in unmap device sysfs callback.
1757 * 3. No IO requests inflight - each file open of block_dev increases
1758 * module reference in get_disk().
1760 * But still there can be user requests inflights, which are sent by
1761 * asynchronous send_msg_*() functions, thus before unmapping devices
1762 * RTRS session must be explicitly closed.
1765 list_for_each_entry_safe(sess, sn, &sess_list, list) {
1766 if (!rnbd_clt_get_sess(sess))
1769 list_for_each_entry_safe(dev, tn, &sess->devs_list, list) {
1771 * Here unmap happens in parallel for only one reason:
1772 * blk_cleanup_queue() takes around half a second, so
1773 * on huge amount of devices the whole module unload
1774 * procedure takes minutes.
1776 INIT_WORK(&dev->unmap_on_rmmod_work, unmap_device_work);
1777 queue_work(system_long_wq, &dev->unmap_on_rmmod_work);
1779 rnbd_clt_put_sess(sess);
1781 /* Wait for all scheduled unmap works */
1782 flush_workqueue(system_long_wq);
1783 WARN_ON(!list_empty(&sess_list));
1786 static int __init rnbd_client_init(void)
1790 BUILD_BUG_ON(sizeof(struct rnbd_msg_hdr) != 4);
1791 BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info) != 36);
1792 BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info_rsp) != 36);
1793 BUILD_BUG_ON(sizeof(struct rnbd_msg_open) != 264);
1794 BUILD_BUG_ON(sizeof(struct rnbd_msg_close) != 8);
1795 BUILD_BUG_ON(sizeof(struct rnbd_msg_open_rsp) != 56);
1796 rnbd_client_major = register_blkdev(rnbd_client_major, "rnbd");
1797 if (rnbd_client_major <= 0) {
1798 pr_err("Failed to load module, block device registration failed\n");
1802 err = rnbd_clt_create_sysfs_files();
1804 pr_err("Failed to load module, creating sysfs device files failed, err: %d\n",
1806 unregister_blkdev(rnbd_client_major, "rnbd");
1812 static void __exit rnbd_client_exit(void)
1814 rnbd_destroy_sessions();
1815 unregister_blkdev(rnbd_client_major, "rnbd");
1816 ida_destroy(&index_ida);
1819 module_init(rnbd_client_init);
1820 module_exit(rnbd_client_exit);