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_map_data {
416 const struct request *rq;
421 struct ublk_io_iter {
422 struct page *pages[UBLK_MAX_PIN_PAGES];
423 unsigned pg_off; /* offset in the 1st page in pages */
424 int nr_pages; /* how many page pointers in pages */
426 struct bvec_iter iter;
429 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
430 unsigned max_bytes, bool to_vm)
432 const unsigned total = min_t(unsigned, max_bytes,
433 PAGE_SIZE - data->pg_off +
434 ((data->nr_pages - 1) << PAGE_SHIFT));
438 while (done < total) {
439 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
440 const unsigned int bytes = min3(bv.bv_len, total - done,
441 (unsigned)(PAGE_SIZE - data->pg_off));
442 void *bv_buf = bvec_kmap_local(&bv);
443 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
446 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
448 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
450 kunmap_local(pg_buf);
451 kunmap_local(bv_buf);
453 /* advance page array */
454 data->pg_off += bytes;
455 if (data->pg_off == PAGE_SIZE) {
463 bio_advance_iter_single(data->bio, &data->iter, bytes);
464 if (!data->iter.bi_size) {
465 data->bio = data->bio->bi_next;
466 if (data->bio == NULL)
468 data->iter = data->bio->bi_iter;
475 static int ublk_copy_user_pages(struct ublk_map_data *data, bool to_vm)
477 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
478 const unsigned long start_vm = data->ubuf;
479 unsigned int done = 0;
480 struct ublk_io_iter iter = {
481 .pg_off = start_vm & (PAGE_SIZE - 1),
482 .bio = data->rq->bio,
483 .iter = data->rq->bio->bi_iter,
485 const unsigned int nr_pages = round_up(data->len +
486 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
488 while (done < nr_pages) {
489 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
493 iter.nr_pages = get_user_pages_fast(start_vm +
494 (done << PAGE_SHIFT), to_pin, gup_flags,
496 if (iter.nr_pages <= 0)
497 return done == 0 ? iter.nr_pages : done;
498 len = ublk_copy_io_pages(&iter, data->len, to_vm);
499 for (i = 0; i < iter.nr_pages; i++) {
501 set_page_dirty(iter.pages[i]);
502 put_page(iter.pages[i]);
505 done += iter.nr_pages;
511 static inline bool ublk_need_map_req(const struct request *req)
513 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
516 static inline bool ublk_need_unmap_req(const struct request *req)
518 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
521 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
524 const unsigned int rq_bytes = blk_rq_bytes(req);
527 * no zero copy, we delay copy WRITE request data into ublksrv
528 * context and the big benefit is that pinning pages in current
529 * context is pretty fast, see ublk_pin_user_pages
531 if (ublk_need_map_req(req)) {
532 struct ublk_map_data data = {
538 ublk_copy_user_pages(&data, true);
540 return rq_bytes - data.len;
545 static int ublk_unmap_io(const struct ublk_queue *ubq,
546 const struct request *req,
549 const unsigned int rq_bytes = blk_rq_bytes(req);
551 if (ublk_need_unmap_req(req)) {
552 struct ublk_map_data data = {
558 WARN_ON_ONCE(io->res > rq_bytes);
560 ublk_copy_user_pages(&data, false);
562 return io->res - data.len;
567 static inline unsigned int ublk_req_build_flags(struct request *req)
571 if (req->cmd_flags & REQ_FAILFAST_DEV)
572 flags |= UBLK_IO_F_FAILFAST_DEV;
574 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
575 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
577 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
578 flags |= UBLK_IO_F_FAILFAST_DRIVER;
580 if (req->cmd_flags & REQ_META)
581 flags |= UBLK_IO_F_META;
583 if (req->cmd_flags & REQ_FUA)
584 flags |= UBLK_IO_F_FUA;
586 if (req->cmd_flags & REQ_NOUNMAP)
587 flags |= UBLK_IO_F_NOUNMAP;
589 if (req->cmd_flags & REQ_SWAP)
590 flags |= UBLK_IO_F_SWAP;
595 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
597 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
598 struct ublk_io *io = &ubq->ios[req->tag];
601 switch (req_op(req)) {
603 ublk_op = UBLK_IO_OP_READ;
606 ublk_op = UBLK_IO_OP_WRITE;
609 ublk_op = UBLK_IO_OP_FLUSH;
612 ublk_op = UBLK_IO_OP_DISCARD;
614 case REQ_OP_WRITE_ZEROES:
615 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
618 return BLK_STS_IOERR;
621 /* need to translate since kernel may change */
622 iod->op_flags = ublk_op | ublk_req_build_flags(req);
623 iod->nr_sectors = blk_rq_sectors(req);
624 iod->start_sector = blk_rq_pos(req);
625 iod->addr = io->addr;
630 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
631 struct io_uring_cmd *ioucmd)
633 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
636 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
638 return ubq->ubq_daemon->flags & PF_EXITING;
641 /* todo: handle partial completion */
642 static void ublk_complete_rq(struct request *req)
644 struct ublk_queue *ubq = req->mq_hctx->driver_data;
645 struct ublk_io *io = &ubq->ios[req->tag];
646 unsigned int unmapped_bytes;
647 blk_status_t res = BLK_STS_OK;
649 /* failed read IO if nothing is read */
650 if (!io->res && req_op(req) == REQ_OP_READ)
654 res = errno_to_blk_status(io->res);
659 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
662 * Both the two needn't unmap.
664 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
667 /* for READ request, writing data in iod->addr to rq buffers */
668 unmapped_bytes = ublk_unmap_io(ubq, req, io);
671 * Extremely impossible since we got data filled in just before
673 * Re-read simply for this unlikely case.
675 if (unlikely(unmapped_bytes < io->res))
676 io->res = unmapped_bytes;
678 if (blk_update_request(req, BLK_STS_OK, io->res))
679 blk_mq_requeue_request(req, true);
681 __blk_mq_end_request(req, BLK_STS_OK);
685 blk_mq_end_request(req, res);
689 * Since __ublk_rq_task_work always fails requests immediately during
690 * exiting, __ublk_fail_req() is only called from abort context during
691 * exiting. So lock is unnecessary.
693 * Also aborting may not be started yet, keep in mind that one failed
694 * request may be issued by block layer again.
696 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
699 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
701 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
702 io->flags |= UBLK_IO_FLAG_ABORTED;
703 if (ublk_queue_can_use_recovery_reissue(ubq))
704 blk_mq_requeue_request(req, false);
706 blk_mq_end_request(req, BLK_STS_IOERR);
710 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
711 unsigned issue_flags)
713 /* mark this cmd owned by ublksrv */
714 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
717 * clear ACTIVE since we are done with this sqe/cmd slot
718 * We can only accept io cmd in case of being not active.
720 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
722 /* tell ublksrv one io request is coming */
723 io_uring_cmd_done(io->cmd, res, 0, issue_flags);
726 #define UBLK_REQUEUE_DELAY_MS 3
728 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
731 /* We cannot process this rq so just requeue it. */
732 if (ublk_queue_can_use_recovery(ubq))
733 blk_mq_requeue_request(rq, false);
735 blk_mq_end_request(rq, BLK_STS_IOERR);
737 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
740 static inline void __ublk_rq_task_work(struct request *req,
741 unsigned issue_flags)
743 struct ublk_queue *ubq = req->mq_hctx->driver_data;
745 struct ublk_io *io = &ubq->ios[tag];
746 unsigned int mapped_bytes;
748 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
749 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
750 ublk_get_iod(ubq, req->tag)->addr);
753 * Task is exiting if either:
755 * (1) current != ubq_daemon.
756 * io_uring_cmd_complete_in_task() tries to run task_work
757 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
759 * (2) current->flags & PF_EXITING.
761 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
762 __ublk_abort_rq(ubq, req);
766 if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
768 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
769 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
772 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
773 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
774 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
775 __func__, io->cmd->cmd_op, ubq->q_id,
776 req->tag, io->flags);
777 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
781 * We have handled UBLK_IO_NEED_GET_DATA command,
782 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
785 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
786 /* update iod->addr because ublksrv may have passed a new io buffer */
787 ublk_get_iod(ubq, req->tag)->addr = io->addr;
788 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
789 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
790 ublk_get_iod(ubq, req->tag)->addr);
793 mapped_bytes = ublk_map_io(ubq, req, io);
795 /* partially mapped, update io descriptor */
796 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
798 * Nothing mapped, retry until we succeed.
800 * We may never succeed in mapping any bytes here because
801 * of OOM. TODO: reserve one buffer with single page pinned
802 * for providing forward progress guarantee.
804 if (unlikely(!mapped_bytes)) {
805 blk_mq_requeue_request(req, false);
806 blk_mq_delay_kick_requeue_list(req->q,
807 UBLK_REQUEUE_DELAY_MS);
811 ublk_get_iod(ubq, req->tag)->nr_sectors =
815 ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
818 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
819 unsigned issue_flags)
821 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
822 struct ublk_rq_data *data, *tmp;
824 io_cmds = llist_reverse_order(io_cmds);
825 llist_for_each_entry_safe(data, tmp, io_cmds, node)
826 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
829 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
831 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
832 struct ublk_rq_data *data, *tmp;
834 llist_for_each_entry_safe(data, tmp, io_cmds, node)
835 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
838 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
840 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
841 struct ublk_queue *ubq = pdu->ubq;
843 ublk_forward_io_cmds(ubq, issue_flags);
846 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
848 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
851 if (!llist_add(&data->node, &ubq->io_cmds))
854 io = &ubq->ios[rq->tag];
856 * If the check pass, we know that this is a re-issued request aborted
857 * previously in monitor_work because the ubq_daemon(cmd's task) is
858 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
859 * because this ioucmd's io_uring context may be freed now if no inflight
860 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
862 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
863 * the tag). Then the request is re-started(allocating the tag) and we are here.
864 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
865 * guarantees that here is a re-issued request aborted previously.
867 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
868 ublk_abort_io_cmds(ubq);
870 struct io_uring_cmd *cmd = io->cmd;
871 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
874 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
878 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
880 struct ublk_queue *ubq = rq->mq_hctx->driver_data;
882 if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
884 send_sig(SIGKILL, ubq->ubq_daemon, 0);
891 return BLK_EH_RESET_TIMER;
894 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
895 const struct blk_mq_queue_data *bd)
897 struct ublk_queue *ubq = hctx->driver_data;
898 struct request *rq = bd->rq;
901 /* fill iod to slot in io cmd buffer */
902 res = ublk_setup_iod(ubq, rq);
903 if (unlikely(res != BLK_STS_OK))
904 return BLK_STS_IOERR;
906 /* With recovery feature enabled, force_abort is set in
907 * ublk_stop_dev() before calling del_gendisk(). We have to
908 * abort all requeued and new rqs here to let del_gendisk()
909 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
910 * to avoid UAF on io_uring ctx.
912 * Note: force_abort is guaranteed to be seen because it is set
913 * before request queue is unqiuesced.
915 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
916 return BLK_STS_IOERR;
918 blk_mq_start_request(bd->rq);
920 if (unlikely(ubq_daemon_is_dying(ubq))) {
921 __ublk_abort_rq(ubq, rq);
925 ublk_queue_cmd(ubq, rq);
930 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
931 unsigned int hctx_idx)
933 struct ublk_device *ub = driver_data;
934 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
936 hctx->driver_data = ubq;
940 static const struct blk_mq_ops ublk_mq_ops = {
941 .queue_rq = ublk_queue_rq,
942 .init_hctx = ublk_init_hctx,
943 .timeout = ublk_timeout,
946 static int ublk_ch_open(struct inode *inode, struct file *filp)
948 struct ublk_device *ub = container_of(inode->i_cdev,
949 struct ublk_device, cdev);
951 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
953 filp->private_data = ub;
957 static int ublk_ch_release(struct inode *inode, struct file *filp)
959 struct ublk_device *ub = filp->private_data;
961 clear_bit(UB_STATE_OPEN, &ub->state);
965 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
966 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
968 struct ublk_device *ub = filp->private_data;
969 size_t sz = vma->vm_end - vma->vm_start;
970 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
971 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
974 spin_lock(&ub->mm_lock);
976 ub->mm = current->mm;
977 if (current->mm != ub->mm)
979 spin_unlock(&ub->mm_lock);
984 if (vma->vm_flags & VM_WRITE)
987 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
988 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
991 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
992 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
993 __func__, q_id, current->pid, vma->vm_start,
994 phys_off, (unsigned long)sz);
996 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
999 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1000 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1003 static void ublk_commit_completion(struct ublk_device *ub,
1004 const struct ublksrv_io_cmd *ub_cmd)
1006 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1007 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1008 struct ublk_io *io = &ubq->ios[tag];
1009 struct request *req;
1011 /* now this cmd slot is owned by nbd driver */
1012 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1013 io->res = ub_cmd->result;
1015 /* find the io request and complete */
1016 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1018 if (req && likely(!blk_should_fake_timeout(req->q)))
1019 ublk_complete_rq(req);
1023 * When ->ubq_daemon is exiting, either new request is ended immediately,
1024 * or any queued io command is drained, so it is safe to abort queue
1027 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1031 if (!ublk_get_device(ub))
1034 for (i = 0; i < ubq->q_depth; i++) {
1035 struct ublk_io *io = &ubq->ios[i];
1037 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1041 * Either we fail the request or ublk_rq_task_work_fn
1044 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1046 __ublk_fail_req(ubq, io, rq);
1049 ublk_put_device(ub);
1052 static void ublk_daemon_monitor_work(struct work_struct *work)
1054 struct ublk_device *ub =
1055 container_of(work, struct ublk_device, monitor_work.work);
1058 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1059 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1061 if (ubq_daemon_is_dying(ubq)) {
1062 if (ublk_queue_can_use_recovery(ubq))
1063 schedule_work(&ub->quiesce_work);
1065 schedule_work(&ub->stop_work);
1067 /* abort queue is for making forward progress */
1068 ublk_abort_queue(ub, ubq);
1073 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1074 * after ublk_remove() or __ublk_quiesce_dev() is started.
1076 * No need ub->mutex, monitor work are canceled after state is marked
1077 * as not LIVE, so new state is observed reliably.
1079 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1080 schedule_delayed_work(&ub->monitor_work,
1081 UBLK_DAEMON_MONITOR_PERIOD);
1084 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1086 return ubq->nr_io_ready == ubq->q_depth;
1089 static void ublk_cancel_queue(struct ublk_queue *ubq)
1093 if (!ublk_queue_ready(ubq))
1096 for (i = 0; i < ubq->q_depth; i++) {
1097 struct ublk_io *io = &ubq->ios[i];
1099 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1100 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1101 IO_URING_F_UNLOCKED);
1104 /* all io commands are canceled */
1105 ubq->nr_io_ready = 0;
1108 /* Cancel all pending commands, must be called after del_gendisk() returns */
1109 static void ublk_cancel_dev(struct ublk_device *ub)
1113 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1114 ublk_cancel_queue(ublk_get_queue(ub, i));
1117 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1121 if (blk_mq_request_started(rq)) {
1128 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1132 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1135 blk_mq_tagset_busy_iter(&ub->tag_set,
1136 ublk_check_inflight_rq, &idle);
1139 msleep(UBLK_REQUEUE_DELAY_MS);
1143 static void __ublk_quiesce_dev(struct ublk_device *ub)
1145 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1146 __func__, ub->dev_info.dev_id,
1147 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1148 "LIVE" : "QUIESCED");
1149 blk_mq_quiesce_queue(ub->ub_disk->queue);
1150 ublk_wait_tagset_rqs_idle(ub);
1151 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1152 ublk_cancel_dev(ub);
1153 /* we are going to release task_struct of ubq_daemon and resets
1154 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1155 * Besides, monitor_work is not necessary in QUIESCED state since we have
1156 * already scheduled quiesce_work and quiesced all ubqs.
1158 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1159 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1161 cancel_delayed_work_sync(&ub->monitor_work);
1164 static void ublk_quiesce_work_fn(struct work_struct *work)
1166 struct ublk_device *ub =
1167 container_of(work, struct ublk_device, quiesce_work);
1169 mutex_lock(&ub->mutex);
1170 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1172 __ublk_quiesce_dev(ub);
1174 mutex_unlock(&ub->mutex);
1177 static void ublk_unquiesce_dev(struct ublk_device *ub)
1181 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1182 __func__, ub->dev_info.dev_id,
1183 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1184 "LIVE" : "QUIESCED");
1185 /* quiesce_work has run. We let requeued rqs be aborted
1186 * before running fallback_wq. "force_abort" must be seen
1187 * after request queue is unqiuesced. Then del_gendisk()
1190 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1191 ublk_get_queue(ub, i)->force_abort = true;
1193 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1194 /* We may have requeued some rqs in ublk_quiesce_queue() */
1195 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1198 static void ublk_stop_dev(struct ublk_device *ub)
1200 mutex_lock(&ub->mutex);
1201 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1203 if (ublk_can_use_recovery(ub)) {
1204 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1205 __ublk_quiesce_dev(ub);
1206 ublk_unquiesce_dev(ub);
1208 del_gendisk(ub->ub_disk);
1209 ub->dev_info.state = UBLK_S_DEV_DEAD;
1210 ub->dev_info.ublksrv_pid = -1;
1211 put_disk(ub->ub_disk);
1214 ublk_cancel_dev(ub);
1215 mutex_unlock(&ub->mutex);
1216 cancel_delayed_work_sync(&ub->monitor_work);
1219 /* device can only be started after all IOs are ready */
1220 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1222 mutex_lock(&ub->mutex);
1224 if (ublk_queue_ready(ubq)) {
1225 ubq->ubq_daemon = current;
1226 get_task_struct(ubq->ubq_daemon);
1227 ub->nr_queues_ready++;
1229 if (capable(CAP_SYS_ADMIN))
1230 ub->nr_privileged_daemon++;
1232 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1233 complete_all(&ub->completion);
1234 mutex_unlock(&ub->mutex);
1237 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1240 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1241 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1243 ublk_queue_cmd(ubq, req);
1246 static inline int ublk_check_cmd_op(u32 cmd_op)
1248 u32 ioc_type = _IOC_TYPE(cmd_op);
1250 if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1253 if (ioc_type != 'u' && ioc_type != 0)
1259 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1260 struct io_uring_cmd *cmd, unsigned long buf_addr)
1263 io->flags |= UBLK_IO_FLAG_ACTIVE;
1264 io->addr = buf_addr;
1267 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1268 unsigned int issue_flags,
1269 const struct ublksrv_io_cmd *ub_cmd)
1271 struct ublk_device *ub = cmd->file->private_data;
1272 struct ublk_queue *ubq;
1274 u32 cmd_op = cmd->cmd_op;
1275 unsigned tag = ub_cmd->tag;
1277 struct request *req;
1279 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1280 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1283 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1286 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1287 if (!ubq || ub_cmd->q_id != ubq->q_id)
1290 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1293 if (tag >= ubq->q_depth)
1296 io = &ubq->ios[tag];
1298 /* there is pending io cmd, something must be wrong */
1299 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1305 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1306 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1308 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1309 ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1312 ret = ublk_check_cmd_op(cmd_op);
1317 switch (_IOC_NR(cmd_op)) {
1318 case UBLK_IO_FETCH_REQ:
1319 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1320 if (ublk_queue_ready(ubq)) {
1325 * The io is being handled by server, so COMMIT_RQ is expected
1326 * instead of FETCH_REQ
1328 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1330 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1331 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1334 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1335 ublk_mark_io_ready(ub, ubq);
1337 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1338 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1340 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1341 * not enabled or it is Read IO.
1343 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1345 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1347 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1348 ublk_commit_completion(ub, ub_cmd);
1350 case UBLK_IO_NEED_GET_DATA:
1351 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1353 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1354 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1359 return -EIOCBQUEUED;
1362 io_uring_cmd_done(cmd, ret, 0, issue_flags);
1363 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1364 __func__, cmd_op, tag, ret, io->flags);
1365 return -EIOCBQUEUED;
1368 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1371 * Not necessary for async retry, but let's keep it simple and always
1372 * copy the values to avoid any potential reuse.
1374 const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1375 const struct ublksrv_io_cmd ub_cmd = {
1376 .q_id = READ_ONCE(ub_src->q_id),
1377 .tag = READ_ONCE(ub_src->tag),
1378 .result = READ_ONCE(ub_src->result),
1379 .addr = READ_ONCE(ub_src->addr)
1382 return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1385 static const struct file_operations ublk_ch_fops = {
1386 .owner = THIS_MODULE,
1387 .open = ublk_ch_open,
1388 .release = ublk_ch_release,
1389 .llseek = no_llseek,
1390 .uring_cmd = ublk_ch_uring_cmd,
1391 .mmap = ublk_ch_mmap,
1394 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1396 int size = ublk_queue_cmd_buf_size(ub, q_id);
1397 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1399 if (ubq->ubq_daemon)
1400 put_task_struct(ubq->ubq_daemon);
1401 if (ubq->io_cmd_buf)
1402 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1405 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1407 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1408 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1412 ubq->flags = ub->dev_info.flags;
1414 ubq->q_depth = ub->dev_info.queue_depth;
1415 size = ublk_queue_cmd_buf_size(ub, q_id);
1417 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1421 ubq->io_cmd_buf = ptr;
1426 static void ublk_deinit_queues(struct ublk_device *ub)
1428 int nr_queues = ub->dev_info.nr_hw_queues;
1434 for (i = 0; i < nr_queues; i++)
1435 ublk_deinit_queue(ub, i);
1436 kfree(ub->__queues);
1439 static int ublk_init_queues(struct ublk_device *ub)
1441 int nr_queues = ub->dev_info.nr_hw_queues;
1442 int depth = ub->dev_info.queue_depth;
1443 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1444 int i, ret = -ENOMEM;
1446 ub->queue_size = ubq_size;
1447 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1451 for (i = 0; i < nr_queues; i++) {
1452 if (ublk_init_queue(ub, i))
1456 init_completion(&ub->completion);
1460 ublk_deinit_queues(ub);
1464 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1469 spin_lock(&ublk_idr_lock);
1470 /* allocate id, if @id >= 0, we're requesting that specific id */
1472 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1476 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1478 spin_unlock(&ublk_idr_lock);
1481 ub->ub_number = err;
1486 static void ublk_free_dev_number(struct ublk_device *ub)
1488 spin_lock(&ublk_idr_lock);
1489 idr_remove(&ublk_index_idr, ub->ub_number);
1490 wake_up_all(&ublk_idr_wq);
1491 spin_unlock(&ublk_idr_lock);
1494 static void ublk_cdev_rel(struct device *dev)
1496 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1498 blk_mq_free_tag_set(&ub->tag_set);
1499 ublk_deinit_queues(ub);
1500 ublk_free_dev_number(ub);
1501 mutex_destroy(&ub->mutex);
1505 static int ublk_add_chdev(struct ublk_device *ub)
1507 struct device *dev = &ub->cdev_dev;
1508 int minor = ub->ub_number;
1511 dev->parent = ublk_misc.this_device;
1512 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1513 dev->class = ublk_chr_class;
1514 dev->release = ublk_cdev_rel;
1515 device_initialize(dev);
1517 ret = dev_set_name(dev, "ublkc%d", minor);
1521 cdev_init(&ub->cdev, &ublk_ch_fops);
1522 ret = cdev_device_add(&ub->cdev, dev);
1533 static void ublk_stop_work_fn(struct work_struct *work)
1535 struct ublk_device *ub =
1536 container_of(work, struct ublk_device, stop_work);
1541 /* align max io buffer size with PAGE_SIZE */
1542 static void ublk_align_max_io_size(struct ublk_device *ub)
1544 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1546 ub->dev_info.max_io_buf_bytes =
1547 round_down(max_io_bytes, PAGE_SIZE);
1550 static int ublk_add_tag_set(struct ublk_device *ub)
1552 ub->tag_set.ops = &ublk_mq_ops;
1553 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1554 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1555 ub->tag_set.numa_node = NUMA_NO_NODE;
1556 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1557 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1558 ub->tag_set.driver_data = ub;
1559 return blk_mq_alloc_tag_set(&ub->tag_set);
1562 static void ublk_remove(struct ublk_device *ub)
1565 cancel_work_sync(&ub->stop_work);
1566 cancel_work_sync(&ub->quiesce_work);
1567 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1568 put_device(&ub->cdev_dev);
1572 static struct ublk_device *ublk_get_device_from_id(int idx)
1574 struct ublk_device *ub = NULL;
1579 spin_lock(&ublk_idr_lock);
1580 ub = idr_find(&ublk_index_idr, idx);
1582 ub = ublk_get_device(ub);
1583 spin_unlock(&ublk_idr_lock);
1588 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1590 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1591 int ublksrv_pid = (int)header->data[0];
1592 struct gendisk *disk;
1595 if (ublksrv_pid <= 0)
1598 wait_for_completion_interruptible(&ub->completion);
1600 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1602 mutex_lock(&ub->mutex);
1603 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1604 test_bit(UB_STATE_USED, &ub->state)) {
1609 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1611 ret = PTR_ERR(disk);
1614 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1615 disk->fops = &ub_fops;
1616 disk->private_data = ub;
1618 ub->dev_info.ublksrv_pid = ublksrv_pid;
1621 ret = ublk_apply_params(ub);
1625 /* don't probe partitions if any one ubq daemon is un-trusted */
1626 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1627 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1629 get_device(&ub->cdev_dev);
1630 ub->dev_info.state = UBLK_S_DEV_LIVE;
1631 ret = add_disk(disk);
1634 * Has to drop the reference since ->free_disk won't be
1635 * called in case of add_disk failure.
1637 ub->dev_info.state = UBLK_S_DEV_DEAD;
1638 ublk_put_device(ub);
1641 set_bit(UB_STATE_USED, &ub->state);
1646 mutex_unlock(&ub->mutex);
1650 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1651 struct io_uring_cmd *cmd)
1653 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1654 void __user *argp = (void __user *)(unsigned long)header->addr;
1655 cpumask_var_t cpumask;
1656 unsigned long queue;
1657 unsigned int retlen;
1661 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1663 if (header->len & (sizeof(unsigned long)-1))
1668 queue = header->data[0];
1669 if (queue >= ub->dev_info.nr_hw_queues)
1672 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1675 for_each_possible_cpu(i) {
1676 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1677 cpumask_set_cpu(i, cpumask);
1681 retlen = min_t(unsigned short, header->len, cpumask_size());
1682 if (copy_to_user(argp, cpumask, retlen))
1683 goto out_free_cpumask;
1684 if (retlen != header->len &&
1685 clear_user(argp + retlen, header->len - retlen))
1686 goto out_free_cpumask;
1690 free_cpumask_var(cpumask);
1694 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1696 pr_devel("%s: dev id %d flags %llx\n", __func__,
1697 info->dev_id, info->flags);
1698 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1699 info->nr_hw_queues, info->queue_depth);
1702 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1704 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1705 void __user *argp = (void __user *)(unsigned long)header->addr;
1706 struct ublksrv_ctrl_dev_info info;
1707 struct ublk_device *ub;
1710 if (header->len < sizeof(info) || !header->addr)
1712 if (header->queue_id != (u16)-1) {
1713 pr_warn("%s: queue_id is wrong %x\n",
1714 __func__, header->queue_id);
1718 if (copy_from_user(&info, argp, sizeof(info)))
1721 if (capable(CAP_SYS_ADMIN))
1722 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1723 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1727 * unprivileged device can't be trusted, but RECOVERY and
1728 * RECOVERY_REISSUE still may hang error handling, so can't
1729 * support recovery features for unprivileged ublk now
1731 * TODO: provide forward progress for RECOVERY handler, so that
1732 * unprivileged device can benefit from it
1734 if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1735 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1736 UBLK_F_USER_RECOVERY);
1738 /* the created device is always owned by current user */
1739 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1741 if (header->dev_id != info.dev_id) {
1742 pr_warn("%s: dev id not match %u %u\n",
1743 __func__, header->dev_id, info.dev_id);
1747 ublk_dump_dev_info(&info);
1749 ret = mutex_lock_killable(&ublk_ctl_mutex);
1754 if (ublks_added >= ublks_max)
1758 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1761 mutex_init(&ub->mutex);
1762 spin_lock_init(&ub->mm_lock);
1763 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1764 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1765 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1767 ret = ublk_alloc_dev_number(ub, header->dev_id);
1771 memcpy(&ub->dev_info, &info, sizeof(info));
1773 /* update device id */
1774 ub->dev_info.dev_id = ub->ub_number;
1777 * 64bit flags will be copied back to userspace as feature
1778 * negotiation result, so have to clear flags which driver
1779 * doesn't support yet, then userspace can get correct flags
1780 * (features) to handle.
1782 ub->dev_info.flags &= UBLK_F_ALL;
1784 ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
1785 UBLK_F_URING_CMD_COMP_IN_TASK;
1787 /* We are not ready to support zero copy */
1788 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1790 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1791 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1792 ublk_align_max_io_size(ub);
1794 ret = ublk_init_queues(ub);
1796 goto out_free_dev_number;
1798 ret = ublk_add_tag_set(ub);
1800 goto out_deinit_queues;
1803 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1804 goto out_free_tag_set;
1807 * Add the char dev so that ublksrv daemon can be setup.
1808 * ublk_add_chdev() will cleanup everything if it fails.
1810 ret = ublk_add_chdev(ub);
1814 blk_mq_free_tag_set(&ub->tag_set);
1816 ublk_deinit_queues(ub);
1817 out_free_dev_number:
1818 ublk_free_dev_number(ub);
1820 mutex_destroy(&ub->mutex);
1823 mutex_unlock(&ublk_ctl_mutex);
1827 static inline bool ublk_idr_freed(int id)
1831 spin_lock(&ublk_idr_lock);
1832 ptr = idr_find(&ublk_index_idr, id);
1833 spin_unlock(&ublk_idr_lock);
1838 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1840 struct ublk_device *ub = *p_ub;
1841 int idx = ub->ub_number;
1844 ret = mutex_lock_killable(&ublk_ctl_mutex);
1848 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1850 set_bit(UB_STATE_DELETED, &ub->state);
1853 /* Mark the reference as consumed */
1855 ublk_put_device(ub);
1856 mutex_unlock(&ublk_ctl_mutex);
1859 * Wait until the idr is removed, then it can be reused after
1860 * DEL_DEV command is returned.
1862 * If we returns because of user interrupt, future delete command
1865 * - the device number isn't freed, this device won't or needn't
1866 * be deleted again, since UB_STATE_DELETED is set, and device
1867 * will be released after the last reference is dropped
1869 * - the device number is freed already, we will not find this
1870 * device via ublk_get_device_from_id()
1872 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1877 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1879 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1881 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1882 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1883 header->data[0], header->addr, header->len);
1886 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1889 cancel_work_sync(&ub->stop_work);
1890 cancel_work_sync(&ub->quiesce_work);
1895 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1896 struct io_uring_cmd *cmd)
1898 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1899 void __user *argp = (void __user *)(unsigned long)header->addr;
1901 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1904 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1910 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1911 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1913 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1914 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1917 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1918 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1920 ub->params.devt.disk_major = 0;
1921 ub->params.devt.disk_minor = 0;
1923 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1926 static int ublk_ctrl_get_params(struct ublk_device *ub,
1927 struct io_uring_cmd *cmd)
1929 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1930 void __user *argp = (void __user *)(unsigned long)header->addr;
1931 struct ublk_params_header ph;
1934 if (header->len <= sizeof(ph) || !header->addr)
1937 if (copy_from_user(&ph, argp, sizeof(ph)))
1940 if (ph.len > header->len || !ph.len)
1943 if (ph.len > sizeof(struct ublk_params))
1944 ph.len = sizeof(struct ublk_params);
1946 mutex_lock(&ub->mutex);
1947 ublk_ctrl_fill_params_devt(ub);
1948 if (copy_to_user(argp, &ub->params, ph.len))
1952 mutex_unlock(&ub->mutex);
1957 static int ublk_ctrl_set_params(struct ublk_device *ub,
1958 struct io_uring_cmd *cmd)
1960 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1961 void __user *argp = (void __user *)(unsigned long)header->addr;
1962 struct ublk_params_header ph;
1965 if (header->len <= sizeof(ph) || !header->addr)
1968 if (copy_from_user(&ph, argp, sizeof(ph)))
1971 if (ph.len > header->len || !ph.len || !ph.types)
1974 if (ph.len > sizeof(struct ublk_params))
1975 ph.len = sizeof(struct ublk_params);
1977 /* parameters can only be changed when device isn't live */
1978 mutex_lock(&ub->mutex);
1979 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1981 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1984 /* clear all we don't support yet */
1985 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1986 ret = ublk_validate_params(ub);
1988 ub->params.types = 0;
1990 mutex_unlock(&ub->mutex);
1995 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1999 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2000 /* All old ioucmds have to be completed */
2001 WARN_ON_ONCE(ubq->nr_io_ready);
2002 /* old daemon is PF_EXITING, put it now */
2003 put_task_struct(ubq->ubq_daemon);
2004 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2005 ubq->ubq_daemon = NULL;
2006 ubq->timeout = false;
2008 for (i = 0; i < ubq->q_depth; i++) {
2009 struct ublk_io *io = &ubq->ios[i];
2011 /* forget everything now and be ready for new FETCH_REQ */
2018 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2019 struct io_uring_cmd *cmd)
2021 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2025 mutex_lock(&ub->mutex);
2026 if (!ublk_can_use_recovery(ub))
2029 * START_RECOVERY is only allowd after:
2031 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2032 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
2035 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2036 * (a)has quiesced request queue
2037 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
2038 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2039 * (d)has completed/camceled all ioucmds owned by ther dying process
2041 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2042 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2046 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2047 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2048 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2049 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2051 ub->nr_queues_ready = 0;
2052 ub->nr_privileged_daemon = 0;
2053 init_completion(&ub->completion);
2056 mutex_unlock(&ub->mutex);
2060 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2061 struct io_uring_cmd *cmd)
2063 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2064 int ublksrv_pid = (int)header->data[0];
2067 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2068 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2069 /* wait until new ubq_daemon sending all FETCH_REQ */
2070 wait_for_completion_interruptible(&ub->completion);
2071 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2072 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2074 mutex_lock(&ub->mutex);
2075 if (!ublk_can_use_recovery(ub))
2078 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2082 ub->dev_info.ublksrv_pid = ublksrv_pid;
2083 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2084 __func__, ublksrv_pid, header->dev_id);
2085 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2086 pr_devel("%s: queue unquiesced, dev id %d.\n",
2087 __func__, header->dev_id);
2088 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2089 ub->dev_info.state = UBLK_S_DEV_LIVE;
2090 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2093 mutex_unlock(&ub->mutex);
2098 * All control commands are sent via /dev/ublk-control, so we have to check
2099 * the destination device's permission
2101 static int ublk_char_dev_permission(struct ublk_device *ub,
2102 const char *dev_path, int mask)
2108 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2112 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2117 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2120 err = inode_permission(&nop_mnt_idmap,
2121 d_backing_inode(path.dentry), mask);
2127 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2128 struct io_uring_cmd *cmd)
2130 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2131 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2132 void __user *argp = (void __user *)(unsigned long)header->addr;
2133 char *dev_path = NULL;
2137 if (!unprivileged) {
2138 if (!capable(CAP_SYS_ADMIN))
2141 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2142 * char_dev_path in payload too, since userspace may not
2143 * know if the specified device is created as unprivileged
2146 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2151 * User has to provide the char device path for unprivileged ublk
2153 * header->addr always points to the dev path buffer, and
2154 * header->dev_path_len records length of dev path buffer.
2156 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2159 if (header->len < header->dev_path_len)
2162 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2167 if (copy_from_user(dev_path, argp, header->dev_path_len))
2169 dev_path[header->dev_path_len] = 0;
2172 switch (_IOC_NR(cmd->cmd_op)) {
2173 case UBLK_CMD_GET_DEV_INFO:
2174 case UBLK_CMD_GET_DEV_INFO2:
2175 case UBLK_CMD_GET_QUEUE_AFFINITY:
2176 case UBLK_CMD_GET_PARAMS:
2179 case UBLK_CMD_START_DEV:
2180 case UBLK_CMD_STOP_DEV:
2181 case UBLK_CMD_ADD_DEV:
2182 case UBLK_CMD_DEL_DEV:
2183 case UBLK_CMD_SET_PARAMS:
2184 case UBLK_CMD_START_USER_RECOVERY:
2185 case UBLK_CMD_END_USER_RECOVERY:
2186 mask = MAY_READ | MAY_WRITE;
2192 ret = ublk_char_dev_permission(ub, dev_path, mask);
2194 header->len -= header->dev_path_len;
2195 header->addr += header->dev_path_len;
2197 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2198 __func__, ub->ub_number, cmd->cmd_op,
2199 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2206 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2207 unsigned int issue_flags)
2209 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2210 struct ublk_device *ub = NULL;
2211 u32 cmd_op = cmd->cmd_op;
2214 if (issue_flags & IO_URING_F_NONBLOCK)
2217 ublk_ctrl_cmd_dump(cmd);
2219 if (!(issue_flags & IO_URING_F_SQE128))
2222 ret = ublk_check_cmd_op(cmd_op);
2226 if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2228 ub = ublk_get_device_from_id(header->dev_id);
2232 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2237 switch (_IOC_NR(cmd_op)) {
2238 case UBLK_CMD_START_DEV:
2239 ret = ublk_ctrl_start_dev(ub, cmd);
2241 case UBLK_CMD_STOP_DEV:
2242 ret = ublk_ctrl_stop_dev(ub);
2244 case UBLK_CMD_GET_DEV_INFO:
2245 case UBLK_CMD_GET_DEV_INFO2:
2246 ret = ublk_ctrl_get_dev_info(ub, cmd);
2248 case UBLK_CMD_ADD_DEV:
2249 ret = ublk_ctrl_add_dev(cmd);
2251 case UBLK_CMD_DEL_DEV:
2252 ret = ublk_ctrl_del_dev(&ub);
2254 case UBLK_CMD_GET_QUEUE_AFFINITY:
2255 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2257 case UBLK_CMD_GET_PARAMS:
2258 ret = ublk_ctrl_get_params(ub, cmd);
2260 case UBLK_CMD_SET_PARAMS:
2261 ret = ublk_ctrl_set_params(ub, cmd);
2263 case UBLK_CMD_START_USER_RECOVERY:
2264 ret = ublk_ctrl_start_recovery(ub, cmd);
2266 case UBLK_CMD_END_USER_RECOVERY:
2267 ret = ublk_ctrl_end_recovery(ub, cmd);
2276 ublk_put_device(ub);
2278 io_uring_cmd_done(cmd, ret, 0, issue_flags);
2279 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2280 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2281 return -EIOCBQUEUED;
2284 static const struct file_operations ublk_ctl_fops = {
2285 .open = nonseekable_open,
2286 .uring_cmd = ublk_ctrl_uring_cmd,
2287 .owner = THIS_MODULE,
2288 .llseek = noop_llseek,
2291 static struct miscdevice ublk_misc = {
2292 .minor = MISC_DYNAMIC_MINOR,
2293 .name = "ublk-control",
2294 .fops = &ublk_ctl_fops,
2297 static int __init ublk_init(void)
2301 init_waitqueue_head(&ublk_idr_wq);
2303 ret = misc_register(&ublk_misc);
2307 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2309 goto unregister_mis;
2311 ublk_chr_class = class_create("ublk-char");
2312 if (IS_ERR(ublk_chr_class)) {
2313 ret = PTR_ERR(ublk_chr_class);
2314 goto free_chrdev_region;
2319 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2321 misc_deregister(&ublk_misc);
2325 static void __exit ublk_exit(void)
2327 struct ublk_device *ub;
2330 idr_for_each_entry(&ublk_index_idr, ub, id)
2333 class_destroy(ublk_chr_class);
2334 misc_deregister(&ublk_misc);
2336 idr_destroy(&ublk_index_idr);
2337 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2340 module_init(ublk_init);
2341 module_exit(ublk_exit);
2343 module_param(ublks_max, int, 0444);
2344 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2346 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2347 MODULE_LICENSE("GPL");