1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Userspace block device - block device which IO is handled from userspace
5 * Take full use of io_uring passthrough command for communicating with
6 * ublk userspace daemon(ublksrvd) for handling basic IO request.
8 * Copyright 2022 Ming Lei <ming.lei@redhat.com>
10 * (part of code stolen from loop.c)
12 #include <linux/module.h>
13 #include <linux/moduleparam.h>
14 #include <linux/sched.h>
16 #include <linux/pagemap.h>
17 #include <linux/file.h>
18 #include <linux/stat.h>
19 #include <linux/errno.h>
20 #include <linux/major.h>
21 #include <linux/wait.h>
22 #include <linux/blkdev.h>
23 #include <linux/init.h>
24 #include <linux/swap.h>
25 #include <linux/slab.h>
26 #include <linux/compat.h>
27 #include <linux/mutex.h>
28 #include <linux/writeback.h>
29 #include <linux/completion.h>
30 #include <linux/highmem.h>
31 #include <linux/sysfs.h>
32 #include <linux/miscdevice.h>
33 #include <linux/falloc.h>
34 #include <linux/uio.h>
35 #include <linux/ioprio.h>
36 #include <linux/sched/mm.h>
37 #include <linux/uaccess.h>
38 #include <linux/cdev.h>
39 #include <linux/io_uring.h>
40 #include <linux/blk-mq.h>
41 #include <linux/delay.h>
44 #include <linux/task_work.h>
45 #include <linux/namei.h>
46 #include <uapi/linux/ublk_cmd.h>
48 #define UBLK_MINORS (1U << MINORBITS)
50 /* All UBLK_F_* have to be included into UBLK_F_ALL */
51 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
52 | UBLK_F_URING_CMD_COMP_IN_TASK \
53 | UBLK_F_NEED_GET_DATA \
54 | UBLK_F_USER_RECOVERY \
55 | UBLK_F_USER_RECOVERY_REISSUE \
56 | UBLK_F_UNPRIVILEGED_DEV \
57 | UBLK_F_CMD_IOCTL_ENCODE)
59 /* All UBLK_PARAM_TYPE_* should be included here */
60 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | \
61 UBLK_PARAM_TYPE_DISCARD | UBLK_PARAM_TYPE_DEVT)
64 struct llist_node node;
67 struct ublk_uring_cmd_pdu {
68 struct ublk_queue *ubq;
72 * io command is active: sqe cmd is received, and its cqe isn't done
74 * If the flag is set, the io command is owned by ublk driver, and waited
75 * for incoming blk-mq request from the ublk block device.
77 * If the flag is cleared, the io command will be completed, and owned by
80 #define UBLK_IO_FLAG_ACTIVE 0x01
83 * IO command is completed via cqe, and it is being handled by ublksrv, and
86 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
89 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
92 * IO command is aborted, so this flag is set in case of
93 * !UBLK_IO_FLAG_ACTIVE.
95 * After this flag is observed, any pending or new incoming request
96 * associated with this io command will be failed immediately
98 #define UBLK_IO_FLAG_ABORTED 0x04
101 * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
102 * get data buffer address from ublksrv.
104 * Then, bio data could be copied into this data buffer for a WRITE request
105 * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
107 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
110 /* userspace buffer address from io cmd */
115 struct io_uring_cmd *cmd;
123 struct task_struct *ubq_daemon;
126 struct llist_head io_cmds;
128 unsigned long io_addr; /* mapped vm address */
129 unsigned int max_io_sz;
132 unsigned short nr_io_ready; /* how many ios setup */
133 struct ublk_device *dev;
134 struct ublk_io ios[];
137 #define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
140 struct gendisk *ub_disk;
144 unsigned int queue_size;
145 struct ublksrv_ctrl_dev_info dev_info;
147 struct blk_mq_tag_set tag_set;
150 struct device cdev_dev;
152 #define UB_STATE_OPEN 0
153 #define UB_STATE_USED 1
154 #define UB_STATE_DELETED 2
161 struct mm_struct *mm;
163 struct ublk_params params;
165 struct completion completion;
166 unsigned int nr_queues_ready;
167 unsigned int nr_privileged_daemon;
170 * Our ubq->daemon may be killed without any notification, so
171 * monitor each queue's daemon periodically
173 struct delayed_work monitor_work;
174 struct work_struct quiesce_work;
175 struct work_struct stop_work;
178 /* header of ublk_params */
179 struct ublk_params_header {
184 static dev_t ublk_chr_devt;
185 static struct class *ublk_chr_class;
187 static DEFINE_IDR(ublk_index_idr);
188 static DEFINE_SPINLOCK(ublk_idr_lock);
189 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
191 static DEFINE_MUTEX(ublk_ctl_mutex);
194 * Max ublk devices allowed to add
196 * It can be extended to one per-user limit in future or even controlled
199 static unsigned int ublks_max = 64;
200 static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
202 static struct miscdevice ublk_misc;
204 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
206 struct request_queue *q = ub->ub_disk->queue;
207 const struct ublk_param_basic *p = &ub->params.basic;
209 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
210 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
211 blk_queue_io_min(q, 1 << p->io_min_shift);
212 blk_queue_io_opt(q, 1 << p->io_opt_shift);
214 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
215 p->attrs & UBLK_ATTR_FUA);
216 if (p->attrs & UBLK_ATTR_ROTATIONAL)
217 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
219 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
221 blk_queue_max_hw_sectors(q, p->max_sectors);
222 blk_queue_chunk_sectors(q, p->chunk_sectors);
223 blk_queue_virt_boundary(q, p->virt_boundary_mask);
225 if (p->attrs & UBLK_ATTR_READ_ONLY)
226 set_disk_ro(ub->ub_disk, true);
228 set_capacity(ub->ub_disk, p->dev_sectors);
231 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
233 struct request_queue *q = ub->ub_disk->queue;
234 const struct ublk_param_discard *p = &ub->params.discard;
236 q->limits.discard_alignment = p->discard_alignment;
237 q->limits.discard_granularity = p->discard_granularity;
238 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
239 blk_queue_max_write_zeroes_sectors(q,
240 p->max_write_zeroes_sectors);
241 blk_queue_max_discard_segments(q, p->max_discard_segments);
244 static int ublk_validate_params(const struct ublk_device *ub)
246 /* basic param is the only one which must be set */
247 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
248 const struct ublk_param_basic *p = &ub->params.basic;
250 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
253 if (p->logical_bs_shift > p->physical_bs_shift)
256 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
261 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
262 const struct ublk_param_discard *p = &ub->params.discard;
264 /* So far, only support single segment discard */
265 if (p->max_discard_sectors && p->max_discard_segments != 1)
268 if (!p->discard_granularity)
272 /* dev_t is read-only */
273 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
279 static int ublk_apply_params(struct ublk_device *ub)
281 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
284 ublk_dev_param_basic_apply(ub);
286 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
287 ublk_dev_param_discard_apply(ub);
292 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
294 return ubq->flags & UBLK_F_NEED_GET_DATA;
297 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
299 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
304 static void ublk_put_device(struct ublk_device *ub)
306 put_device(&ub->cdev_dev);
309 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
312 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
315 static inline bool ublk_rq_has_data(const struct request *rq)
317 return bio_has_data(rq->bio);
320 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
323 return (struct ublksrv_io_desc *)
324 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
327 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
329 return ublk_get_queue(ub, q_id)->io_cmd_buf;
332 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
334 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
336 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
340 static inline bool ublk_queue_can_use_recovery_reissue(
341 struct ublk_queue *ubq)
343 return (ubq->flags & UBLK_F_USER_RECOVERY) &&
344 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
347 static inline bool ublk_queue_can_use_recovery(
348 struct ublk_queue *ubq)
350 return ubq->flags & UBLK_F_USER_RECOVERY;
353 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
355 return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
358 static void ublk_free_disk(struct gendisk *disk)
360 struct ublk_device *ub = disk->private_data;
362 clear_bit(UB_STATE_USED, &ub->state);
363 put_device(&ub->cdev_dev);
366 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
367 unsigned int *owner_gid)
372 current_uid_gid(&uid, &gid);
374 *owner_uid = from_kuid(&init_user_ns, uid);
375 *owner_gid = from_kgid(&init_user_ns, gid);
378 static int ublk_open(struct block_device *bdev, fmode_t mode)
380 struct ublk_device *ub = bdev->bd_disk->private_data;
382 if (capable(CAP_SYS_ADMIN))
386 * If it is one unprivileged device, only owner can open
387 * the disk. Otherwise it could be one trap made by one
388 * evil user who grants this disk's privileges to other
389 * users deliberately.
391 * This way is reasonable too given anyone can create
392 * unprivileged device, and no need other's grant.
394 if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
395 unsigned int curr_uid, curr_gid;
397 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
399 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
400 ub->dev_info.owner_gid)
407 static const struct block_device_operations ub_fops = {
408 .owner = THIS_MODULE,
410 .free_disk = ublk_free_disk,
413 #define UBLK_MAX_PIN_PAGES 32
415 struct ublk_io_iter {
416 struct page *pages[UBLK_MAX_PIN_PAGES];
418 struct bvec_iter iter;
421 /* return how many pages are copied */
422 static void ublk_copy_io_pages(struct ublk_io_iter *data,
423 size_t total, size_t pg_off, int dir)
428 while (done < total) {
429 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
430 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
431 (unsigned)(PAGE_SIZE - pg_off));
432 void *bv_buf = bvec_kmap_local(&bv);
433 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
435 if (dir == ITER_DEST)
436 memcpy(pg_buf + pg_off, bv_buf, bytes);
438 memcpy(bv_buf, pg_buf + pg_off, bytes);
440 kunmap_local(pg_buf);
441 kunmap_local(bv_buf);
443 /* advance page array */
445 if (pg_off == PAGE_SIZE) {
453 bio_advance_iter_single(data->bio, &data->iter, bytes);
454 if (!data->iter.bi_size) {
455 data->bio = data->bio->bi_next;
456 if (data->bio == NULL)
458 data->iter = data->bio->bi_iter;
464 * Copy data between request pages and io_iter, and 'offset'
465 * is the start point of linear offset of request.
467 static size_t ublk_copy_user_pages(const struct request *req,
468 struct iov_iter *uiter, int dir)
470 struct ublk_io_iter iter = {
472 .iter = req->bio->bi_iter,
476 while (iov_iter_count(uiter) && iter.bio) {
481 len = iov_iter_get_pages2(uiter, iter.pages,
482 iov_iter_count(uiter),
483 UBLK_MAX_PIN_PAGES, &off);
487 ublk_copy_io_pages(&iter, len, off, dir);
488 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
489 for (i = 0; i < nr_pages; i++) {
490 if (dir == ITER_DEST)
491 set_page_dirty(iter.pages[i]);
492 put_page(iter.pages[i]);
500 static inline bool ublk_need_map_req(const struct request *req)
502 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
505 static inline bool ublk_need_unmap_req(const struct request *req)
507 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
510 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
513 const unsigned int rq_bytes = blk_rq_bytes(req);
516 * no zero copy, we delay copy WRITE request data into ublksrv
517 * context and the big benefit is that pinning pages in current
518 * context is pretty fast, see ublk_pin_user_pages
520 if (ublk_need_map_req(req)) {
521 struct iov_iter iter;
523 const int dir = ITER_DEST;
525 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
528 return ublk_copy_user_pages(req, &iter, dir);
533 static int ublk_unmap_io(const struct ublk_queue *ubq,
534 const struct request *req,
537 const unsigned int rq_bytes = blk_rq_bytes(req);
539 if (ublk_need_unmap_req(req)) {
540 struct iov_iter iter;
542 const int dir = ITER_SOURCE;
544 WARN_ON_ONCE(io->res > rq_bytes);
546 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
548 return ublk_copy_user_pages(req, &iter, dir);
553 static inline unsigned int ublk_req_build_flags(struct request *req)
557 if (req->cmd_flags & REQ_FAILFAST_DEV)
558 flags |= UBLK_IO_F_FAILFAST_DEV;
560 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
561 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
563 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
564 flags |= UBLK_IO_F_FAILFAST_DRIVER;
566 if (req->cmd_flags & REQ_META)
567 flags |= UBLK_IO_F_META;
569 if (req->cmd_flags & REQ_FUA)
570 flags |= UBLK_IO_F_FUA;
572 if (req->cmd_flags & REQ_NOUNMAP)
573 flags |= UBLK_IO_F_NOUNMAP;
575 if (req->cmd_flags & REQ_SWAP)
576 flags |= UBLK_IO_F_SWAP;
581 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
583 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
584 struct ublk_io *io = &ubq->ios[req->tag];
587 switch (req_op(req)) {
589 ublk_op = UBLK_IO_OP_READ;
592 ublk_op = UBLK_IO_OP_WRITE;
595 ublk_op = UBLK_IO_OP_FLUSH;
598 ublk_op = UBLK_IO_OP_DISCARD;
600 case REQ_OP_WRITE_ZEROES:
601 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
604 return BLK_STS_IOERR;
607 /* need to translate since kernel may change */
608 iod->op_flags = ublk_op | ublk_req_build_flags(req);
609 iod->nr_sectors = blk_rq_sectors(req);
610 iod->start_sector = blk_rq_pos(req);
611 iod->addr = io->addr;
616 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
617 struct io_uring_cmd *ioucmd)
619 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
622 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
624 return ubq->ubq_daemon->flags & PF_EXITING;
627 /* todo: handle partial completion */
628 static void ublk_complete_rq(struct request *req)
630 struct ublk_queue *ubq = req->mq_hctx->driver_data;
631 struct ublk_io *io = &ubq->ios[req->tag];
632 unsigned int unmapped_bytes;
633 blk_status_t res = BLK_STS_OK;
635 /* failed read IO if nothing is read */
636 if (!io->res && req_op(req) == REQ_OP_READ)
640 res = errno_to_blk_status(io->res);
645 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
648 * Both the two needn't unmap.
650 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
653 /* for READ request, writing data in iod->addr to rq buffers */
654 unmapped_bytes = ublk_unmap_io(ubq, req, io);
657 * Extremely impossible since we got data filled in just before
659 * Re-read simply for this unlikely case.
661 if (unlikely(unmapped_bytes < io->res))
662 io->res = unmapped_bytes;
664 if (blk_update_request(req, BLK_STS_OK, io->res))
665 blk_mq_requeue_request(req, true);
667 __blk_mq_end_request(req, BLK_STS_OK);
671 blk_mq_end_request(req, res);
675 * Since __ublk_rq_task_work always fails requests immediately during
676 * exiting, __ublk_fail_req() is only called from abort context during
677 * exiting. So lock is unnecessary.
679 * Also aborting may not be started yet, keep in mind that one failed
680 * request may be issued by block layer again.
682 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
685 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
687 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
688 io->flags |= UBLK_IO_FLAG_ABORTED;
689 if (ublk_queue_can_use_recovery_reissue(ubq))
690 blk_mq_requeue_request(req, false);
692 blk_mq_end_request(req, BLK_STS_IOERR);
696 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
697 unsigned issue_flags)
699 /* mark this cmd owned by ublksrv */
700 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
703 * clear ACTIVE since we are done with this sqe/cmd slot
704 * We can only accept io cmd in case of being not active.
706 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
708 /* tell ublksrv one io request is coming */
709 io_uring_cmd_done(io->cmd, res, 0, issue_flags);
712 #define UBLK_REQUEUE_DELAY_MS 3
714 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
717 /* We cannot process this rq so just requeue it. */
718 if (ublk_queue_can_use_recovery(ubq))
719 blk_mq_requeue_request(rq, false);
721 blk_mq_end_request(rq, BLK_STS_IOERR);
723 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
726 static inline void __ublk_rq_task_work(struct request *req,
727 unsigned issue_flags)
729 struct ublk_queue *ubq = req->mq_hctx->driver_data;
731 struct ublk_io *io = &ubq->ios[tag];
732 unsigned int mapped_bytes;
734 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
735 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
736 ublk_get_iod(ubq, req->tag)->addr);
739 * Task is exiting if either:
741 * (1) current != ubq_daemon.
742 * io_uring_cmd_complete_in_task() tries to run task_work
743 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
745 * (2) current->flags & PF_EXITING.
747 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
748 __ublk_abort_rq(ubq, req);
752 if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
754 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
755 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
758 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
759 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
760 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
761 __func__, io->cmd->cmd_op, ubq->q_id,
762 req->tag, io->flags);
763 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
767 * We have handled UBLK_IO_NEED_GET_DATA command,
768 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
771 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
772 /* update iod->addr because ublksrv may have passed a new io buffer */
773 ublk_get_iod(ubq, req->tag)->addr = io->addr;
774 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
775 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
776 ublk_get_iod(ubq, req->tag)->addr);
779 mapped_bytes = ublk_map_io(ubq, req, io);
781 /* partially mapped, update io descriptor */
782 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
784 * Nothing mapped, retry until we succeed.
786 * We may never succeed in mapping any bytes here because
787 * of OOM. TODO: reserve one buffer with single page pinned
788 * for providing forward progress guarantee.
790 if (unlikely(!mapped_bytes)) {
791 blk_mq_requeue_request(req, false);
792 blk_mq_delay_kick_requeue_list(req->q,
793 UBLK_REQUEUE_DELAY_MS);
797 ublk_get_iod(ubq, req->tag)->nr_sectors =
801 ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
804 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
805 unsigned issue_flags)
807 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
808 struct ublk_rq_data *data, *tmp;
810 io_cmds = llist_reverse_order(io_cmds);
811 llist_for_each_entry_safe(data, tmp, io_cmds, node)
812 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
815 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
817 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
818 struct ublk_rq_data *data, *tmp;
820 llist_for_each_entry_safe(data, tmp, io_cmds, node)
821 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
824 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
826 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
827 struct ublk_queue *ubq = pdu->ubq;
829 ublk_forward_io_cmds(ubq, issue_flags);
832 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
834 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
837 if (!llist_add(&data->node, &ubq->io_cmds))
840 io = &ubq->ios[rq->tag];
842 * If the check pass, we know that this is a re-issued request aborted
843 * previously in monitor_work because the ubq_daemon(cmd's task) is
844 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
845 * because this ioucmd's io_uring context may be freed now if no inflight
846 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
848 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
849 * the tag). Then the request is re-started(allocating the tag) and we are here.
850 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
851 * guarantees that here is a re-issued request aborted previously.
853 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
854 ublk_abort_io_cmds(ubq);
856 struct io_uring_cmd *cmd = io->cmd;
857 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
860 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
864 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
866 struct ublk_queue *ubq = rq->mq_hctx->driver_data;
868 if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
870 send_sig(SIGKILL, ubq->ubq_daemon, 0);
877 return BLK_EH_RESET_TIMER;
880 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
881 const struct blk_mq_queue_data *bd)
883 struct ublk_queue *ubq = hctx->driver_data;
884 struct request *rq = bd->rq;
887 /* fill iod to slot in io cmd buffer */
888 res = ublk_setup_iod(ubq, rq);
889 if (unlikely(res != BLK_STS_OK))
890 return BLK_STS_IOERR;
892 /* With recovery feature enabled, force_abort is set in
893 * ublk_stop_dev() before calling del_gendisk(). We have to
894 * abort all requeued and new rqs here to let del_gendisk()
895 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
896 * to avoid UAF on io_uring ctx.
898 * Note: force_abort is guaranteed to be seen because it is set
899 * before request queue is unqiuesced.
901 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
902 return BLK_STS_IOERR;
904 blk_mq_start_request(bd->rq);
906 if (unlikely(ubq_daemon_is_dying(ubq))) {
907 __ublk_abort_rq(ubq, rq);
911 ublk_queue_cmd(ubq, rq);
916 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
917 unsigned int hctx_idx)
919 struct ublk_device *ub = driver_data;
920 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
922 hctx->driver_data = ubq;
926 static const struct blk_mq_ops ublk_mq_ops = {
927 .queue_rq = ublk_queue_rq,
928 .init_hctx = ublk_init_hctx,
929 .timeout = ublk_timeout,
932 static int ublk_ch_open(struct inode *inode, struct file *filp)
934 struct ublk_device *ub = container_of(inode->i_cdev,
935 struct ublk_device, cdev);
937 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
939 filp->private_data = ub;
943 static int ublk_ch_release(struct inode *inode, struct file *filp)
945 struct ublk_device *ub = filp->private_data;
947 clear_bit(UB_STATE_OPEN, &ub->state);
951 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
952 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
954 struct ublk_device *ub = filp->private_data;
955 size_t sz = vma->vm_end - vma->vm_start;
956 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
957 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
960 spin_lock(&ub->mm_lock);
962 ub->mm = current->mm;
963 if (current->mm != ub->mm)
965 spin_unlock(&ub->mm_lock);
970 if (vma->vm_flags & VM_WRITE)
973 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
974 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
977 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
978 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
979 __func__, q_id, current->pid, vma->vm_start,
980 phys_off, (unsigned long)sz);
982 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
985 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
986 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
989 static void ublk_commit_completion(struct ublk_device *ub,
990 const struct ublksrv_io_cmd *ub_cmd)
992 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
993 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
994 struct ublk_io *io = &ubq->ios[tag];
997 /* now this cmd slot is owned by nbd driver */
998 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
999 io->res = ub_cmd->result;
1001 /* find the io request and complete */
1002 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1004 if (req && likely(!blk_should_fake_timeout(req->q)))
1005 ublk_complete_rq(req);
1009 * When ->ubq_daemon is exiting, either new request is ended immediately,
1010 * or any queued io command is drained, so it is safe to abort queue
1013 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1017 if (!ublk_get_device(ub))
1020 for (i = 0; i < ubq->q_depth; i++) {
1021 struct ublk_io *io = &ubq->ios[i];
1023 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1027 * Either we fail the request or ublk_rq_task_work_fn
1030 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1032 __ublk_fail_req(ubq, io, rq);
1035 ublk_put_device(ub);
1038 static void ublk_daemon_monitor_work(struct work_struct *work)
1040 struct ublk_device *ub =
1041 container_of(work, struct ublk_device, monitor_work.work);
1044 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1045 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1047 if (ubq_daemon_is_dying(ubq)) {
1048 if (ublk_queue_can_use_recovery(ubq))
1049 schedule_work(&ub->quiesce_work);
1051 schedule_work(&ub->stop_work);
1053 /* abort queue is for making forward progress */
1054 ublk_abort_queue(ub, ubq);
1059 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1060 * after ublk_remove() or __ublk_quiesce_dev() is started.
1062 * No need ub->mutex, monitor work are canceled after state is marked
1063 * as not LIVE, so new state is observed reliably.
1065 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1066 schedule_delayed_work(&ub->monitor_work,
1067 UBLK_DAEMON_MONITOR_PERIOD);
1070 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1072 return ubq->nr_io_ready == ubq->q_depth;
1075 static void ublk_cancel_queue(struct ublk_queue *ubq)
1079 if (!ublk_queue_ready(ubq))
1082 for (i = 0; i < ubq->q_depth; i++) {
1083 struct ublk_io *io = &ubq->ios[i];
1085 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1086 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1087 IO_URING_F_UNLOCKED);
1090 /* all io commands are canceled */
1091 ubq->nr_io_ready = 0;
1094 /* Cancel all pending commands, must be called after del_gendisk() returns */
1095 static void ublk_cancel_dev(struct ublk_device *ub)
1099 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1100 ublk_cancel_queue(ublk_get_queue(ub, i));
1103 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1107 if (blk_mq_request_started(rq)) {
1114 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1118 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1121 blk_mq_tagset_busy_iter(&ub->tag_set,
1122 ublk_check_inflight_rq, &idle);
1125 msleep(UBLK_REQUEUE_DELAY_MS);
1129 static void __ublk_quiesce_dev(struct ublk_device *ub)
1131 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1132 __func__, ub->dev_info.dev_id,
1133 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1134 "LIVE" : "QUIESCED");
1135 blk_mq_quiesce_queue(ub->ub_disk->queue);
1136 ublk_wait_tagset_rqs_idle(ub);
1137 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1138 ublk_cancel_dev(ub);
1139 /* we are going to release task_struct of ubq_daemon and resets
1140 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1141 * Besides, monitor_work is not necessary in QUIESCED state since we have
1142 * already scheduled quiesce_work and quiesced all ubqs.
1144 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1145 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1147 cancel_delayed_work_sync(&ub->monitor_work);
1150 static void ublk_quiesce_work_fn(struct work_struct *work)
1152 struct ublk_device *ub =
1153 container_of(work, struct ublk_device, quiesce_work);
1155 mutex_lock(&ub->mutex);
1156 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1158 __ublk_quiesce_dev(ub);
1160 mutex_unlock(&ub->mutex);
1163 static void ublk_unquiesce_dev(struct ublk_device *ub)
1167 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1168 __func__, ub->dev_info.dev_id,
1169 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1170 "LIVE" : "QUIESCED");
1171 /* quiesce_work has run. We let requeued rqs be aborted
1172 * before running fallback_wq. "force_abort" must be seen
1173 * after request queue is unqiuesced. Then del_gendisk()
1176 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1177 ublk_get_queue(ub, i)->force_abort = true;
1179 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1180 /* We may have requeued some rqs in ublk_quiesce_queue() */
1181 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1184 static void ublk_stop_dev(struct ublk_device *ub)
1186 mutex_lock(&ub->mutex);
1187 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1189 if (ublk_can_use_recovery(ub)) {
1190 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1191 __ublk_quiesce_dev(ub);
1192 ublk_unquiesce_dev(ub);
1194 del_gendisk(ub->ub_disk);
1195 ub->dev_info.state = UBLK_S_DEV_DEAD;
1196 ub->dev_info.ublksrv_pid = -1;
1197 put_disk(ub->ub_disk);
1200 ublk_cancel_dev(ub);
1201 mutex_unlock(&ub->mutex);
1202 cancel_delayed_work_sync(&ub->monitor_work);
1205 /* device can only be started after all IOs are ready */
1206 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1208 mutex_lock(&ub->mutex);
1210 if (ublk_queue_ready(ubq)) {
1211 ubq->ubq_daemon = current;
1212 get_task_struct(ubq->ubq_daemon);
1213 ub->nr_queues_ready++;
1215 if (capable(CAP_SYS_ADMIN))
1216 ub->nr_privileged_daemon++;
1218 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1219 complete_all(&ub->completion);
1220 mutex_unlock(&ub->mutex);
1223 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1226 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1227 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1229 ublk_queue_cmd(ubq, req);
1232 static inline int ublk_check_cmd_op(u32 cmd_op)
1234 u32 ioc_type = _IOC_TYPE(cmd_op);
1236 if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1239 if (ioc_type != 'u' && ioc_type != 0)
1245 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1246 struct io_uring_cmd *cmd, unsigned long buf_addr)
1249 io->flags |= UBLK_IO_FLAG_ACTIVE;
1250 io->addr = buf_addr;
1253 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1254 unsigned int issue_flags,
1255 const struct ublksrv_io_cmd *ub_cmd)
1257 struct ublk_device *ub = cmd->file->private_data;
1258 struct ublk_queue *ubq;
1260 u32 cmd_op = cmd->cmd_op;
1261 unsigned tag = ub_cmd->tag;
1263 struct request *req;
1265 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1266 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1269 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1272 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1273 if (!ubq || ub_cmd->q_id != ubq->q_id)
1276 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1279 if (tag >= ubq->q_depth)
1282 io = &ubq->ios[tag];
1284 /* there is pending io cmd, something must be wrong */
1285 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1291 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1292 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1294 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1295 ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1298 ret = ublk_check_cmd_op(cmd_op);
1303 switch (_IOC_NR(cmd_op)) {
1304 case UBLK_IO_FETCH_REQ:
1305 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1306 if (ublk_queue_ready(ubq)) {
1311 * The io is being handled by server, so COMMIT_RQ is expected
1312 * instead of FETCH_REQ
1314 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1316 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1317 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1320 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1321 ublk_mark_io_ready(ub, ubq);
1323 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1324 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1326 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1327 * not enabled or it is Read IO.
1329 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1331 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1333 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1334 ublk_commit_completion(ub, ub_cmd);
1336 case UBLK_IO_NEED_GET_DATA:
1337 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1339 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1340 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1345 return -EIOCBQUEUED;
1348 io_uring_cmd_done(cmd, ret, 0, issue_flags);
1349 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1350 __func__, cmd_op, tag, ret, io->flags);
1351 return -EIOCBQUEUED;
1354 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1357 * Not necessary for async retry, but let's keep it simple and always
1358 * copy the values to avoid any potential reuse.
1360 const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1361 const struct ublksrv_io_cmd ub_cmd = {
1362 .q_id = READ_ONCE(ub_src->q_id),
1363 .tag = READ_ONCE(ub_src->tag),
1364 .result = READ_ONCE(ub_src->result),
1365 .addr = READ_ONCE(ub_src->addr)
1368 return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1371 static const struct file_operations ublk_ch_fops = {
1372 .owner = THIS_MODULE,
1373 .open = ublk_ch_open,
1374 .release = ublk_ch_release,
1375 .llseek = no_llseek,
1376 .uring_cmd = ublk_ch_uring_cmd,
1377 .mmap = ublk_ch_mmap,
1380 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1382 int size = ublk_queue_cmd_buf_size(ub, q_id);
1383 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1385 if (ubq->ubq_daemon)
1386 put_task_struct(ubq->ubq_daemon);
1387 if (ubq->io_cmd_buf)
1388 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1391 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1393 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1394 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1398 ubq->flags = ub->dev_info.flags;
1400 ubq->q_depth = ub->dev_info.queue_depth;
1401 size = ublk_queue_cmd_buf_size(ub, q_id);
1403 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1407 ubq->io_cmd_buf = ptr;
1412 static void ublk_deinit_queues(struct ublk_device *ub)
1414 int nr_queues = ub->dev_info.nr_hw_queues;
1420 for (i = 0; i < nr_queues; i++)
1421 ublk_deinit_queue(ub, i);
1422 kfree(ub->__queues);
1425 static int ublk_init_queues(struct ublk_device *ub)
1427 int nr_queues = ub->dev_info.nr_hw_queues;
1428 int depth = ub->dev_info.queue_depth;
1429 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1430 int i, ret = -ENOMEM;
1432 ub->queue_size = ubq_size;
1433 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1437 for (i = 0; i < nr_queues; i++) {
1438 if (ublk_init_queue(ub, i))
1442 init_completion(&ub->completion);
1446 ublk_deinit_queues(ub);
1450 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1455 spin_lock(&ublk_idr_lock);
1456 /* allocate id, if @id >= 0, we're requesting that specific id */
1458 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1462 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1464 spin_unlock(&ublk_idr_lock);
1467 ub->ub_number = err;
1472 static void ublk_free_dev_number(struct ublk_device *ub)
1474 spin_lock(&ublk_idr_lock);
1475 idr_remove(&ublk_index_idr, ub->ub_number);
1476 wake_up_all(&ublk_idr_wq);
1477 spin_unlock(&ublk_idr_lock);
1480 static void ublk_cdev_rel(struct device *dev)
1482 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1484 blk_mq_free_tag_set(&ub->tag_set);
1485 ublk_deinit_queues(ub);
1486 ublk_free_dev_number(ub);
1487 mutex_destroy(&ub->mutex);
1491 static int ublk_add_chdev(struct ublk_device *ub)
1493 struct device *dev = &ub->cdev_dev;
1494 int minor = ub->ub_number;
1497 dev->parent = ublk_misc.this_device;
1498 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1499 dev->class = ublk_chr_class;
1500 dev->release = ublk_cdev_rel;
1501 device_initialize(dev);
1503 ret = dev_set_name(dev, "ublkc%d", minor);
1507 cdev_init(&ub->cdev, &ublk_ch_fops);
1508 ret = cdev_device_add(&ub->cdev, dev);
1519 static void ublk_stop_work_fn(struct work_struct *work)
1521 struct ublk_device *ub =
1522 container_of(work, struct ublk_device, stop_work);
1527 /* align max io buffer size with PAGE_SIZE */
1528 static void ublk_align_max_io_size(struct ublk_device *ub)
1530 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1532 ub->dev_info.max_io_buf_bytes =
1533 round_down(max_io_bytes, PAGE_SIZE);
1536 static int ublk_add_tag_set(struct ublk_device *ub)
1538 ub->tag_set.ops = &ublk_mq_ops;
1539 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1540 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1541 ub->tag_set.numa_node = NUMA_NO_NODE;
1542 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1543 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1544 ub->tag_set.driver_data = ub;
1545 return blk_mq_alloc_tag_set(&ub->tag_set);
1548 static void ublk_remove(struct ublk_device *ub)
1551 cancel_work_sync(&ub->stop_work);
1552 cancel_work_sync(&ub->quiesce_work);
1553 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1554 put_device(&ub->cdev_dev);
1558 static struct ublk_device *ublk_get_device_from_id(int idx)
1560 struct ublk_device *ub = NULL;
1565 spin_lock(&ublk_idr_lock);
1566 ub = idr_find(&ublk_index_idr, idx);
1568 ub = ublk_get_device(ub);
1569 spin_unlock(&ublk_idr_lock);
1574 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1576 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1577 int ublksrv_pid = (int)header->data[0];
1578 struct gendisk *disk;
1581 if (ublksrv_pid <= 0)
1584 wait_for_completion_interruptible(&ub->completion);
1586 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1588 mutex_lock(&ub->mutex);
1589 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1590 test_bit(UB_STATE_USED, &ub->state)) {
1595 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1597 ret = PTR_ERR(disk);
1600 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1601 disk->fops = &ub_fops;
1602 disk->private_data = ub;
1604 ub->dev_info.ublksrv_pid = ublksrv_pid;
1607 ret = ublk_apply_params(ub);
1611 /* don't probe partitions if any one ubq daemon is un-trusted */
1612 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1613 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1615 get_device(&ub->cdev_dev);
1616 ub->dev_info.state = UBLK_S_DEV_LIVE;
1617 ret = add_disk(disk);
1620 * Has to drop the reference since ->free_disk won't be
1621 * called in case of add_disk failure.
1623 ub->dev_info.state = UBLK_S_DEV_DEAD;
1624 ublk_put_device(ub);
1627 set_bit(UB_STATE_USED, &ub->state);
1632 mutex_unlock(&ub->mutex);
1636 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1637 struct io_uring_cmd *cmd)
1639 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1640 void __user *argp = (void __user *)(unsigned long)header->addr;
1641 cpumask_var_t cpumask;
1642 unsigned long queue;
1643 unsigned int retlen;
1647 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1649 if (header->len & (sizeof(unsigned long)-1))
1654 queue = header->data[0];
1655 if (queue >= ub->dev_info.nr_hw_queues)
1658 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1661 for_each_possible_cpu(i) {
1662 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1663 cpumask_set_cpu(i, cpumask);
1667 retlen = min_t(unsigned short, header->len, cpumask_size());
1668 if (copy_to_user(argp, cpumask, retlen))
1669 goto out_free_cpumask;
1670 if (retlen != header->len &&
1671 clear_user(argp + retlen, header->len - retlen))
1672 goto out_free_cpumask;
1676 free_cpumask_var(cpumask);
1680 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1682 pr_devel("%s: dev id %d flags %llx\n", __func__,
1683 info->dev_id, info->flags);
1684 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1685 info->nr_hw_queues, info->queue_depth);
1688 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1690 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1691 void __user *argp = (void __user *)(unsigned long)header->addr;
1692 struct ublksrv_ctrl_dev_info info;
1693 struct ublk_device *ub;
1696 if (header->len < sizeof(info) || !header->addr)
1698 if (header->queue_id != (u16)-1) {
1699 pr_warn("%s: queue_id is wrong %x\n",
1700 __func__, header->queue_id);
1704 if (copy_from_user(&info, argp, sizeof(info)))
1707 if (capable(CAP_SYS_ADMIN))
1708 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1709 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1713 * unprivileged device can't be trusted, but RECOVERY and
1714 * RECOVERY_REISSUE still may hang error handling, so can't
1715 * support recovery features for unprivileged ublk now
1717 * TODO: provide forward progress for RECOVERY handler, so that
1718 * unprivileged device can benefit from it
1720 if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1721 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1722 UBLK_F_USER_RECOVERY);
1724 /* the created device is always owned by current user */
1725 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1727 if (header->dev_id != info.dev_id) {
1728 pr_warn("%s: dev id not match %u %u\n",
1729 __func__, header->dev_id, info.dev_id);
1733 ublk_dump_dev_info(&info);
1735 ret = mutex_lock_killable(&ublk_ctl_mutex);
1740 if (ublks_added >= ublks_max)
1744 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1747 mutex_init(&ub->mutex);
1748 spin_lock_init(&ub->mm_lock);
1749 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1750 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1751 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1753 ret = ublk_alloc_dev_number(ub, header->dev_id);
1757 memcpy(&ub->dev_info, &info, sizeof(info));
1759 /* update device id */
1760 ub->dev_info.dev_id = ub->ub_number;
1763 * 64bit flags will be copied back to userspace as feature
1764 * negotiation result, so have to clear flags which driver
1765 * doesn't support yet, then userspace can get correct flags
1766 * (features) to handle.
1768 ub->dev_info.flags &= UBLK_F_ALL;
1770 ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
1771 UBLK_F_URING_CMD_COMP_IN_TASK;
1773 /* We are not ready to support zero copy */
1774 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1776 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1777 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1778 ublk_align_max_io_size(ub);
1780 ret = ublk_init_queues(ub);
1782 goto out_free_dev_number;
1784 ret = ublk_add_tag_set(ub);
1786 goto out_deinit_queues;
1789 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1790 goto out_free_tag_set;
1793 * Add the char dev so that ublksrv daemon can be setup.
1794 * ublk_add_chdev() will cleanup everything if it fails.
1796 ret = ublk_add_chdev(ub);
1800 blk_mq_free_tag_set(&ub->tag_set);
1802 ublk_deinit_queues(ub);
1803 out_free_dev_number:
1804 ublk_free_dev_number(ub);
1806 mutex_destroy(&ub->mutex);
1809 mutex_unlock(&ublk_ctl_mutex);
1813 static inline bool ublk_idr_freed(int id)
1817 spin_lock(&ublk_idr_lock);
1818 ptr = idr_find(&ublk_index_idr, id);
1819 spin_unlock(&ublk_idr_lock);
1824 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1826 struct ublk_device *ub = *p_ub;
1827 int idx = ub->ub_number;
1830 ret = mutex_lock_killable(&ublk_ctl_mutex);
1834 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1836 set_bit(UB_STATE_DELETED, &ub->state);
1839 /* Mark the reference as consumed */
1841 ublk_put_device(ub);
1842 mutex_unlock(&ublk_ctl_mutex);
1845 * Wait until the idr is removed, then it can be reused after
1846 * DEL_DEV command is returned.
1848 * If we returns because of user interrupt, future delete command
1851 * - the device number isn't freed, this device won't or needn't
1852 * be deleted again, since UB_STATE_DELETED is set, and device
1853 * will be released after the last reference is dropped
1855 * - the device number is freed already, we will not find this
1856 * device via ublk_get_device_from_id()
1858 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1863 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1865 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1867 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1868 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1869 header->data[0], header->addr, header->len);
1872 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1875 cancel_work_sync(&ub->stop_work);
1876 cancel_work_sync(&ub->quiesce_work);
1881 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1882 struct io_uring_cmd *cmd)
1884 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1885 void __user *argp = (void __user *)(unsigned long)header->addr;
1887 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1890 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1896 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1897 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1899 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1900 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1903 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1904 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1906 ub->params.devt.disk_major = 0;
1907 ub->params.devt.disk_minor = 0;
1909 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1912 static int ublk_ctrl_get_params(struct ublk_device *ub,
1913 struct io_uring_cmd *cmd)
1915 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1916 void __user *argp = (void __user *)(unsigned long)header->addr;
1917 struct ublk_params_header ph;
1920 if (header->len <= sizeof(ph) || !header->addr)
1923 if (copy_from_user(&ph, argp, sizeof(ph)))
1926 if (ph.len > header->len || !ph.len)
1929 if (ph.len > sizeof(struct ublk_params))
1930 ph.len = sizeof(struct ublk_params);
1932 mutex_lock(&ub->mutex);
1933 ublk_ctrl_fill_params_devt(ub);
1934 if (copy_to_user(argp, &ub->params, ph.len))
1938 mutex_unlock(&ub->mutex);
1943 static int ublk_ctrl_set_params(struct ublk_device *ub,
1944 struct io_uring_cmd *cmd)
1946 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1947 void __user *argp = (void __user *)(unsigned long)header->addr;
1948 struct ublk_params_header ph;
1951 if (header->len <= sizeof(ph) || !header->addr)
1954 if (copy_from_user(&ph, argp, sizeof(ph)))
1957 if (ph.len > header->len || !ph.len || !ph.types)
1960 if (ph.len > sizeof(struct ublk_params))
1961 ph.len = sizeof(struct ublk_params);
1963 /* parameters can only be changed when device isn't live */
1964 mutex_lock(&ub->mutex);
1965 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1967 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1970 /* clear all we don't support yet */
1971 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1972 ret = ublk_validate_params(ub);
1974 ub->params.types = 0;
1976 mutex_unlock(&ub->mutex);
1981 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1985 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1986 /* All old ioucmds have to be completed */
1987 WARN_ON_ONCE(ubq->nr_io_ready);
1988 /* old daemon is PF_EXITING, put it now */
1989 put_task_struct(ubq->ubq_daemon);
1990 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1991 ubq->ubq_daemon = NULL;
1992 ubq->timeout = false;
1994 for (i = 0; i < ubq->q_depth; i++) {
1995 struct ublk_io *io = &ubq->ios[i];
1997 /* forget everything now and be ready for new FETCH_REQ */
2004 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2005 struct io_uring_cmd *cmd)
2007 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2011 mutex_lock(&ub->mutex);
2012 if (!ublk_can_use_recovery(ub))
2015 * START_RECOVERY is only allowd after:
2017 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2018 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
2021 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2022 * (a)has quiesced request queue
2023 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
2024 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2025 * (d)has completed/camceled all ioucmds owned by ther dying process
2027 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2028 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2032 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2033 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2034 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2035 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2037 ub->nr_queues_ready = 0;
2038 ub->nr_privileged_daemon = 0;
2039 init_completion(&ub->completion);
2042 mutex_unlock(&ub->mutex);
2046 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2047 struct io_uring_cmd *cmd)
2049 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2050 int ublksrv_pid = (int)header->data[0];
2053 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2054 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2055 /* wait until new ubq_daemon sending all FETCH_REQ */
2056 wait_for_completion_interruptible(&ub->completion);
2057 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2058 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2060 mutex_lock(&ub->mutex);
2061 if (!ublk_can_use_recovery(ub))
2064 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2068 ub->dev_info.ublksrv_pid = ublksrv_pid;
2069 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2070 __func__, ublksrv_pid, header->dev_id);
2071 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2072 pr_devel("%s: queue unquiesced, dev id %d.\n",
2073 __func__, header->dev_id);
2074 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2075 ub->dev_info.state = UBLK_S_DEV_LIVE;
2076 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2079 mutex_unlock(&ub->mutex);
2084 * All control commands are sent via /dev/ublk-control, so we have to check
2085 * the destination device's permission
2087 static int ublk_char_dev_permission(struct ublk_device *ub,
2088 const char *dev_path, int mask)
2094 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2098 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2103 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2106 err = inode_permission(&nop_mnt_idmap,
2107 d_backing_inode(path.dentry), mask);
2113 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2114 struct io_uring_cmd *cmd)
2116 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2117 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2118 void __user *argp = (void __user *)(unsigned long)header->addr;
2119 char *dev_path = NULL;
2123 if (!unprivileged) {
2124 if (!capable(CAP_SYS_ADMIN))
2127 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2128 * char_dev_path in payload too, since userspace may not
2129 * know if the specified device is created as unprivileged
2132 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2137 * User has to provide the char device path for unprivileged ublk
2139 * header->addr always points to the dev path buffer, and
2140 * header->dev_path_len records length of dev path buffer.
2142 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2145 if (header->len < header->dev_path_len)
2148 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2153 if (copy_from_user(dev_path, argp, header->dev_path_len))
2155 dev_path[header->dev_path_len] = 0;
2158 switch (_IOC_NR(cmd->cmd_op)) {
2159 case UBLK_CMD_GET_DEV_INFO:
2160 case UBLK_CMD_GET_DEV_INFO2:
2161 case UBLK_CMD_GET_QUEUE_AFFINITY:
2162 case UBLK_CMD_GET_PARAMS:
2165 case UBLK_CMD_START_DEV:
2166 case UBLK_CMD_STOP_DEV:
2167 case UBLK_CMD_ADD_DEV:
2168 case UBLK_CMD_DEL_DEV:
2169 case UBLK_CMD_SET_PARAMS:
2170 case UBLK_CMD_START_USER_RECOVERY:
2171 case UBLK_CMD_END_USER_RECOVERY:
2172 mask = MAY_READ | MAY_WRITE;
2178 ret = ublk_char_dev_permission(ub, dev_path, mask);
2180 header->len -= header->dev_path_len;
2181 header->addr += header->dev_path_len;
2183 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2184 __func__, ub->ub_number, cmd->cmd_op,
2185 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2192 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2193 unsigned int issue_flags)
2195 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2196 struct ublk_device *ub = NULL;
2197 u32 cmd_op = cmd->cmd_op;
2200 if (issue_flags & IO_URING_F_NONBLOCK)
2203 ublk_ctrl_cmd_dump(cmd);
2205 if (!(issue_flags & IO_URING_F_SQE128))
2208 ret = ublk_check_cmd_op(cmd_op);
2212 if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2214 ub = ublk_get_device_from_id(header->dev_id);
2218 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2223 switch (_IOC_NR(cmd_op)) {
2224 case UBLK_CMD_START_DEV:
2225 ret = ublk_ctrl_start_dev(ub, cmd);
2227 case UBLK_CMD_STOP_DEV:
2228 ret = ublk_ctrl_stop_dev(ub);
2230 case UBLK_CMD_GET_DEV_INFO:
2231 case UBLK_CMD_GET_DEV_INFO2:
2232 ret = ublk_ctrl_get_dev_info(ub, cmd);
2234 case UBLK_CMD_ADD_DEV:
2235 ret = ublk_ctrl_add_dev(cmd);
2237 case UBLK_CMD_DEL_DEV:
2238 ret = ublk_ctrl_del_dev(&ub);
2240 case UBLK_CMD_GET_QUEUE_AFFINITY:
2241 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2243 case UBLK_CMD_GET_PARAMS:
2244 ret = ublk_ctrl_get_params(ub, cmd);
2246 case UBLK_CMD_SET_PARAMS:
2247 ret = ublk_ctrl_set_params(ub, cmd);
2249 case UBLK_CMD_START_USER_RECOVERY:
2250 ret = ublk_ctrl_start_recovery(ub, cmd);
2252 case UBLK_CMD_END_USER_RECOVERY:
2253 ret = ublk_ctrl_end_recovery(ub, cmd);
2262 ublk_put_device(ub);
2264 io_uring_cmd_done(cmd, ret, 0, issue_flags);
2265 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2266 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2267 return -EIOCBQUEUED;
2270 static const struct file_operations ublk_ctl_fops = {
2271 .open = nonseekable_open,
2272 .uring_cmd = ublk_ctrl_uring_cmd,
2273 .owner = THIS_MODULE,
2274 .llseek = noop_llseek,
2277 static struct miscdevice ublk_misc = {
2278 .minor = MISC_DYNAMIC_MINOR,
2279 .name = "ublk-control",
2280 .fops = &ublk_ctl_fops,
2283 static int __init ublk_init(void)
2287 init_waitqueue_head(&ublk_idr_wq);
2289 ret = misc_register(&ublk_misc);
2293 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2295 goto unregister_mis;
2297 ublk_chr_class = class_create("ublk-char");
2298 if (IS_ERR(ublk_chr_class)) {
2299 ret = PTR_ERR(ublk_chr_class);
2300 goto free_chrdev_region;
2305 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2307 misc_deregister(&ublk_misc);
2311 static void __exit ublk_exit(void)
2313 struct ublk_device *ub;
2316 idr_for_each_entry(&ublk_index_idr, ub, id)
2319 class_destroy(ublk_chr_class);
2320 misc_deregister(&ublk_misc);
2322 idr_destroy(&ublk_index_idr);
2323 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2326 module_init(ublk_init);
2327 module_exit(ublk_exit);
2329 module_param(ublks_max, int, 0444);
2330 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2332 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2333 MODULE_LICENSE("GPL");