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)
58 /* All UBLK_PARAM_TYPE_* should be included here */
59 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | \
60 UBLK_PARAM_TYPE_DISCARD | UBLK_PARAM_TYPE_DEVT)
63 struct llist_node node;
64 struct callback_head work;
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;
131 unsigned short nr_io_ready; /* how many ios setup */
132 struct ublk_device *dev;
133 struct ublk_io ios[];
136 #define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
139 struct gendisk *ub_disk;
143 unsigned short queue_size;
144 struct ublksrv_ctrl_dev_info dev_info;
146 struct blk_mq_tag_set tag_set;
149 struct device cdev_dev;
151 #define UB_STATE_OPEN 0
152 #define UB_STATE_USED 1
159 struct mm_struct *mm;
161 struct ublk_params params;
163 struct completion completion;
164 unsigned int nr_queues_ready;
165 unsigned int nr_privileged_daemon;
168 * Our ubq->daemon may be killed without any notification, so
169 * monitor each queue's daemon periodically
171 struct delayed_work monitor_work;
172 struct work_struct quiesce_work;
173 struct work_struct stop_work;
176 /* header of ublk_params */
177 struct ublk_params_header {
182 static dev_t ublk_chr_devt;
183 static struct class *ublk_chr_class;
185 static DEFINE_IDR(ublk_index_idr);
186 static DEFINE_SPINLOCK(ublk_idr_lock);
187 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
189 static DEFINE_MUTEX(ublk_ctl_mutex);
192 * Max ublk devices allowed to add
194 * It can be extended to one per-user limit in future or even controlled
197 static unsigned int ublks_max = 64;
198 static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
200 static struct miscdevice ublk_misc;
202 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
204 struct request_queue *q = ub->ub_disk->queue;
205 const struct ublk_param_basic *p = &ub->params.basic;
207 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
208 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
209 blk_queue_io_min(q, 1 << p->io_min_shift);
210 blk_queue_io_opt(q, 1 << p->io_opt_shift);
212 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
213 p->attrs & UBLK_ATTR_FUA);
214 if (p->attrs & UBLK_ATTR_ROTATIONAL)
215 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
217 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
219 blk_queue_max_hw_sectors(q, p->max_sectors);
220 blk_queue_chunk_sectors(q, p->chunk_sectors);
221 blk_queue_virt_boundary(q, p->virt_boundary_mask);
223 if (p->attrs & UBLK_ATTR_READ_ONLY)
224 set_disk_ro(ub->ub_disk, true);
226 set_capacity(ub->ub_disk, p->dev_sectors);
229 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
231 struct request_queue *q = ub->ub_disk->queue;
232 const struct ublk_param_discard *p = &ub->params.discard;
234 q->limits.discard_alignment = p->discard_alignment;
235 q->limits.discard_granularity = p->discard_granularity;
236 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
237 blk_queue_max_write_zeroes_sectors(q,
238 p->max_write_zeroes_sectors);
239 blk_queue_max_discard_segments(q, p->max_discard_segments);
242 static int ublk_validate_params(const struct ublk_device *ub)
244 /* basic param is the only one which must be set */
245 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
246 const struct ublk_param_basic *p = &ub->params.basic;
248 if (p->logical_bs_shift > PAGE_SHIFT)
251 if (p->logical_bs_shift > p->physical_bs_shift)
254 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
259 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
260 const struct ublk_param_discard *p = &ub->params.discard;
262 /* So far, only support single segment discard */
263 if (p->max_discard_sectors && p->max_discard_segments != 1)
266 if (!p->discard_granularity)
270 /* dev_t is read-only */
271 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
277 static int ublk_apply_params(struct ublk_device *ub)
279 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
282 ublk_dev_param_basic_apply(ub);
284 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
285 ublk_dev_param_discard_apply(ub);
290 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
292 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
293 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
298 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
300 if (ubq->flags & UBLK_F_NEED_GET_DATA)
305 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
307 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
312 static void ublk_put_device(struct ublk_device *ub)
314 put_device(&ub->cdev_dev);
317 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
320 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
323 static inline bool ublk_rq_has_data(const struct request *rq)
325 return rq->bio && bio_has_data(rq->bio);
328 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
331 return (struct ublksrv_io_desc *)
332 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
335 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
337 return ublk_get_queue(ub, q_id)->io_cmd_buf;
340 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
342 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
344 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
348 static inline bool ublk_queue_can_use_recovery_reissue(
349 struct ublk_queue *ubq)
351 if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
352 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
357 static inline bool ublk_queue_can_use_recovery(
358 struct ublk_queue *ubq)
360 if (ubq->flags & UBLK_F_USER_RECOVERY)
365 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
367 if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
372 static void ublk_free_disk(struct gendisk *disk)
374 struct ublk_device *ub = disk->private_data;
376 clear_bit(UB_STATE_USED, &ub->state);
377 put_device(&ub->cdev_dev);
380 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
381 unsigned int *owner_gid)
386 current_uid_gid(&uid, &gid);
388 *owner_uid = from_kuid(&init_user_ns, uid);
389 *owner_gid = from_kgid(&init_user_ns, gid);
392 static int ublk_open(struct block_device *bdev, fmode_t mode)
394 struct ublk_device *ub = bdev->bd_disk->private_data;
396 if (capable(CAP_SYS_ADMIN))
400 * If it is one unprivileged device, only owner can open
401 * the disk. Otherwise it could be one trap made by one
402 * evil user who grants this disk's privileges to other
403 * users deliberately.
405 * This way is reasonable too given anyone can create
406 * unprivileged device, and no need other's grant.
408 if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
409 unsigned int curr_uid, curr_gid;
411 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
413 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
414 ub->dev_info.owner_gid)
421 static const struct block_device_operations ub_fops = {
422 .owner = THIS_MODULE,
424 .free_disk = ublk_free_disk,
427 #define UBLK_MAX_PIN_PAGES 32
429 struct ublk_map_data {
430 const struct ublk_queue *ubq;
431 const struct request *rq;
432 const struct ublk_io *io;
436 struct ublk_io_iter {
437 struct page *pages[UBLK_MAX_PIN_PAGES];
438 unsigned pg_off; /* offset in the 1st page in pages */
439 int nr_pages; /* how many page pointers in pages */
441 struct bvec_iter iter;
444 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
445 unsigned max_bytes, bool to_vm)
447 const unsigned total = min_t(unsigned, max_bytes,
448 PAGE_SIZE - data->pg_off +
449 ((data->nr_pages - 1) << PAGE_SHIFT));
453 while (done < total) {
454 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
455 const unsigned int bytes = min3(bv.bv_len, total - done,
456 (unsigned)(PAGE_SIZE - data->pg_off));
457 void *bv_buf = bvec_kmap_local(&bv);
458 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
461 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
463 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
465 kunmap_local(pg_buf);
466 kunmap_local(bv_buf);
468 /* advance page array */
469 data->pg_off += bytes;
470 if (data->pg_off == PAGE_SIZE) {
478 bio_advance_iter_single(data->bio, &data->iter, bytes);
479 if (!data->iter.bi_size) {
480 data->bio = data->bio->bi_next;
481 if (data->bio == NULL)
483 data->iter = data->bio->bi_iter;
490 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
493 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
494 const unsigned long start_vm = data->io->addr;
495 unsigned int done = 0;
496 struct ublk_io_iter iter = {
497 .pg_off = start_vm & (PAGE_SIZE - 1),
498 .bio = data->rq->bio,
499 .iter = data->rq->bio->bi_iter,
501 const unsigned int nr_pages = round_up(data->max_bytes +
502 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
504 while (done < nr_pages) {
505 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
509 iter.nr_pages = get_user_pages_fast(start_vm +
510 (done << PAGE_SHIFT), to_pin, gup_flags,
512 if (iter.nr_pages <= 0)
513 return done == 0 ? iter.nr_pages : done;
514 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
515 for (i = 0; i < iter.nr_pages; i++) {
517 set_page_dirty(iter.pages[i]);
518 put_page(iter.pages[i]);
520 data->max_bytes -= len;
521 done += iter.nr_pages;
527 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
530 const unsigned int rq_bytes = blk_rq_bytes(req);
532 * no zero copy, we delay copy WRITE request data into ublksrv
533 * context and the big benefit is that pinning pages in current
534 * context is pretty fast, see ublk_pin_user_pages
536 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
539 if (ublk_rq_has_data(req)) {
540 struct ublk_map_data data = {
544 .max_bytes = rq_bytes,
547 ublk_copy_user_pages(&data, true);
549 return rq_bytes - data.max_bytes;
554 static int ublk_unmap_io(const struct ublk_queue *ubq,
555 const struct request *req,
558 const unsigned int rq_bytes = blk_rq_bytes(req);
560 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
561 struct ublk_map_data data = {
565 .max_bytes = io->res,
568 WARN_ON_ONCE(io->res > rq_bytes);
570 ublk_copy_user_pages(&data, false);
572 return io->res - data.max_bytes;
577 static inline unsigned int ublk_req_build_flags(struct request *req)
581 if (req->cmd_flags & REQ_FAILFAST_DEV)
582 flags |= UBLK_IO_F_FAILFAST_DEV;
584 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
585 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
587 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
588 flags |= UBLK_IO_F_FAILFAST_DRIVER;
590 if (req->cmd_flags & REQ_META)
591 flags |= UBLK_IO_F_META;
593 if (req->cmd_flags & REQ_FUA)
594 flags |= UBLK_IO_F_FUA;
596 if (req->cmd_flags & REQ_NOUNMAP)
597 flags |= UBLK_IO_F_NOUNMAP;
599 if (req->cmd_flags & REQ_SWAP)
600 flags |= UBLK_IO_F_SWAP;
605 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
607 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
608 struct ublk_io *io = &ubq->ios[req->tag];
611 switch (req_op(req)) {
613 ublk_op = UBLK_IO_OP_READ;
616 ublk_op = UBLK_IO_OP_WRITE;
619 ublk_op = UBLK_IO_OP_FLUSH;
622 ublk_op = UBLK_IO_OP_DISCARD;
624 case REQ_OP_WRITE_ZEROES:
625 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
628 return BLK_STS_IOERR;
631 /* need to translate since kernel may change */
632 iod->op_flags = ublk_op | ublk_req_build_flags(req);
633 iod->nr_sectors = blk_rq_sectors(req);
634 iod->start_sector = blk_rq_pos(req);
635 iod->addr = io->addr;
640 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
641 struct io_uring_cmd *ioucmd)
643 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
646 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
648 return ubq->ubq_daemon->flags & PF_EXITING;
651 /* todo: handle partial completion */
652 static void ublk_complete_rq(struct request *req)
654 struct ublk_queue *ubq = req->mq_hctx->driver_data;
655 struct ublk_io *io = &ubq->ios[req->tag];
656 unsigned int unmapped_bytes;
658 /* failed read IO if nothing is read */
659 if (!io->res && req_op(req) == REQ_OP_READ)
663 blk_mq_end_request(req, errno_to_blk_status(io->res));
668 * FLUSH or DISCARD usually won't return bytes returned, so end them
671 * Both the two needn't unmap.
673 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
674 blk_mq_end_request(req, BLK_STS_OK);
678 /* for READ request, writing data in iod->addr to rq buffers */
679 unmapped_bytes = ublk_unmap_io(ubq, req, io);
682 * Extremely impossible since we got data filled in just before
684 * Re-read simply for this unlikely case.
686 if (unlikely(unmapped_bytes < io->res))
687 io->res = unmapped_bytes;
689 if (blk_update_request(req, BLK_STS_OK, io->res))
690 blk_mq_requeue_request(req, true);
692 __blk_mq_end_request(req, BLK_STS_OK);
696 * Since __ublk_rq_task_work always fails requests immediately during
697 * exiting, __ublk_fail_req() is only called from abort context during
698 * exiting. So lock is unnecessary.
700 * Also aborting may not be started yet, keep in mind that one failed
701 * request may be issued by block layer again.
703 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
706 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
708 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
709 io->flags |= UBLK_IO_FLAG_ABORTED;
710 if (ublk_queue_can_use_recovery_reissue(ubq))
711 blk_mq_requeue_request(req, false);
713 blk_mq_end_request(req, BLK_STS_IOERR);
717 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
719 /* mark this cmd owned by ublksrv */
720 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
723 * clear ACTIVE since we are done with this sqe/cmd slot
724 * We can only accept io cmd in case of being not active.
726 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
728 /* tell ublksrv one io request is coming */
729 io_uring_cmd_done(io->cmd, res, 0);
732 #define UBLK_REQUEUE_DELAY_MS 3
734 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
737 /* We cannot process this rq so just requeue it. */
738 if (ublk_queue_can_use_recovery(ubq))
739 blk_mq_requeue_request(rq, false);
741 blk_mq_end_request(rq, BLK_STS_IOERR);
743 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
746 static inline void __ublk_rq_task_work(struct request *req)
748 struct ublk_queue *ubq = req->mq_hctx->driver_data;
750 struct ublk_io *io = &ubq->ios[tag];
751 unsigned int mapped_bytes;
753 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
754 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
755 ublk_get_iod(ubq, req->tag)->addr);
758 * Task is exiting if either:
760 * (1) current != ubq_daemon.
761 * io_uring_cmd_complete_in_task() tries to run task_work
762 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
764 * (2) current->flags & PF_EXITING.
766 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
767 __ublk_abort_rq(ubq, req);
771 if (ublk_need_get_data(ubq) &&
772 (req_op(req) == REQ_OP_WRITE ||
773 req_op(req) == REQ_OP_FLUSH)) {
775 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
776 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
779 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
780 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
781 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
782 __func__, io->cmd->cmd_op, ubq->q_id,
783 req->tag, io->flags);
784 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
788 * We have handled UBLK_IO_NEED_GET_DATA command,
789 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
792 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
793 /* update iod->addr because ublksrv may have passed a new io buffer */
794 ublk_get_iod(ubq, req->tag)->addr = io->addr;
795 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
796 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
797 ublk_get_iod(ubq, req->tag)->addr);
800 mapped_bytes = ublk_map_io(ubq, req, io);
802 /* partially mapped, update io descriptor */
803 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
805 * Nothing mapped, retry until we succeed.
807 * We may never succeed in mapping any bytes here because
808 * of OOM. TODO: reserve one buffer with single page pinned
809 * for providing forward progress guarantee.
811 if (unlikely(!mapped_bytes)) {
812 blk_mq_requeue_request(req, false);
813 blk_mq_delay_kick_requeue_list(req->q,
814 UBLK_REQUEUE_DELAY_MS);
818 ublk_get_iod(ubq, req->tag)->nr_sectors =
822 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
825 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
827 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
828 struct ublk_rq_data *data, *tmp;
830 io_cmds = llist_reverse_order(io_cmds);
831 llist_for_each_entry_safe(data, tmp, io_cmds, node)
832 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
835 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
837 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
838 struct ublk_rq_data *data, *tmp;
840 llist_for_each_entry_safe(data, tmp, io_cmds, node)
841 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
844 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
846 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
847 struct ublk_queue *ubq = pdu->ubq;
849 ublk_forward_io_cmds(ubq);
852 static void ublk_rq_task_work_fn(struct callback_head *work)
854 struct ublk_rq_data *data = container_of(work,
855 struct ublk_rq_data, work);
856 struct request *req = blk_mq_rq_from_pdu(data);
857 struct ublk_queue *ubq = req->mq_hctx->driver_data;
859 ublk_forward_io_cmds(ubq);
862 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
864 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
867 if (!llist_add(&data->node, &ubq->io_cmds))
870 io = &ubq->ios[rq->tag];
872 * If the check pass, we know that this is a re-issued request aborted
873 * previously in monitor_work because the ubq_daemon(cmd's task) is
874 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
875 * because this ioucmd's io_uring context may be freed now if no inflight
876 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
878 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
879 * the tag). Then the request is re-started(allocating the tag) and we are here.
880 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
881 * guarantees that here is a re-issued request aborted previously.
883 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
884 ublk_abort_io_cmds(ubq);
885 } else if (ublk_can_use_task_work(ubq)) {
886 if (task_work_add(ubq->ubq_daemon, &data->work,
888 ublk_abort_io_cmds(ubq);
890 struct io_uring_cmd *cmd = io->cmd;
891 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
894 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
898 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
899 const struct blk_mq_queue_data *bd)
901 struct ublk_queue *ubq = hctx->driver_data;
902 struct request *rq = bd->rq;
905 /* fill iod to slot in io cmd buffer */
906 res = ublk_setup_iod(ubq, rq);
907 if (unlikely(res != BLK_STS_OK))
908 return BLK_STS_IOERR;
910 /* With recovery feature enabled, force_abort is set in
911 * ublk_stop_dev() before calling del_gendisk(). We have to
912 * abort all requeued and new rqs here to let del_gendisk()
913 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
914 * to avoid UAF on io_uring ctx.
916 * Note: force_abort is guaranteed to be seen because it is set
917 * before request queue is unqiuesced.
919 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
920 return BLK_STS_IOERR;
922 blk_mq_start_request(bd->rq);
924 if (unlikely(ubq_daemon_is_dying(ubq))) {
925 __ublk_abort_rq(ubq, rq);
929 ublk_queue_cmd(ubq, rq);
934 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
935 unsigned int hctx_idx)
937 struct ublk_device *ub = driver_data;
938 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
940 hctx->driver_data = ubq;
944 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
945 unsigned int hctx_idx, unsigned int numa_node)
947 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
949 init_task_work(&data->work, ublk_rq_task_work_fn);
953 static const struct blk_mq_ops ublk_mq_ops = {
954 .queue_rq = ublk_queue_rq,
955 .init_hctx = ublk_init_hctx,
956 .init_request = ublk_init_rq,
959 static int ublk_ch_open(struct inode *inode, struct file *filp)
961 struct ublk_device *ub = container_of(inode->i_cdev,
962 struct ublk_device, cdev);
964 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
966 filp->private_data = ub;
970 static int ublk_ch_release(struct inode *inode, struct file *filp)
972 struct ublk_device *ub = filp->private_data;
974 clear_bit(UB_STATE_OPEN, &ub->state);
978 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
979 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
981 struct ublk_device *ub = filp->private_data;
982 size_t sz = vma->vm_end - vma->vm_start;
983 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
984 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
987 spin_lock(&ub->mm_lock);
989 ub->mm = current->mm;
990 if (current->mm != ub->mm)
992 spin_unlock(&ub->mm_lock);
997 if (vma->vm_flags & VM_WRITE)
1000 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1001 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1004 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1005 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1006 __func__, q_id, current->pid, vma->vm_start,
1007 phys_off, (unsigned long)sz);
1009 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1012 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1013 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1016 static void ublk_commit_completion(struct ublk_device *ub,
1017 struct ublksrv_io_cmd *ub_cmd)
1019 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1020 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1021 struct ublk_io *io = &ubq->ios[tag];
1022 struct request *req;
1024 /* now this cmd slot is owned by nbd driver */
1025 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1026 io->res = ub_cmd->result;
1028 /* find the io request and complete */
1029 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1031 if (req && likely(!blk_should_fake_timeout(req->q)))
1032 ublk_complete_rq(req);
1036 * When ->ubq_daemon is exiting, either new request is ended immediately,
1037 * or any queued io command is drained, so it is safe to abort queue
1040 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1044 if (!ublk_get_device(ub))
1047 for (i = 0; i < ubq->q_depth; i++) {
1048 struct ublk_io *io = &ubq->ios[i];
1050 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1054 * Either we fail the request or ublk_rq_task_work_fn
1057 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1059 __ublk_fail_req(ubq, io, rq);
1062 ublk_put_device(ub);
1065 static void ublk_daemon_monitor_work(struct work_struct *work)
1067 struct ublk_device *ub =
1068 container_of(work, struct ublk_device, monitor_work.work);
1071 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1072 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1074 if (ubq_daemon_is_dying(ubq)) {
1075 if (ublk_queue_can_use_recovery(ubq))
1076 schedule_work(&ub->quiesce_work);
1078 schedule_work(&ub->stop_work);
1080 /* abort queue is for making forward progress */
1081 ublk_abort_queue(ub, ubq);
1086 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1087 * after ublk_remove() or __ublk_quiesce_dev() is started.
1089 * No need ub->mutex, monitor work are canceled after state is marked
1090 * as not LIVE, so new state is observed reliably.
1092 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1093 schedule_delayed_work(&ub->monitor_work,
1094 UBLK_DAEMON_MONITOR_PERIOD);
1097 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1099 return ubq->nr_io_ready == ubq->q_depth;
1102 static void ublk_cancel_queue(struct ublk_queue *ubq)
1106 if (!ublk_queue_ready(ubq))
1109 for (i = 0; i < ubq->q_depth; i++) {
1110 struct ublk_io *io = &ubq->ios[i];
1112 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1113 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1116 /* all io commands are canceled */
1117 ubq->nr_io_ready = 0;
1120 /* Cancel all pending commands, must be called after del_gendisk() returns */
1121 static void ublk_cancel_dev(struct ublk_device *ub)
1125 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1126 ublk_cancel_queue(ublk_get_queue(ub, i));
1129 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1133 if (blk_mq_request_started(rq)) {
1140 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1144 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1147 blk_mq_tagset_busy_iter(&ub->tag_set,
1148 ublk_check_inflight_rq, &idle);
1151 msleep(UBLK_REQUEUE_DELAY_MS);
1155 static void __ublk_quiesce_dev(struct ublk_device *ub)
1157 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1158 __func__, ub->dev_info.dev_id,
1159 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1160 "LIVE" : "QUIESCED");
1161 blk_mq_quiesce_queue(ub->ub_disk->queue);
1162 ublk_wait_tagset_rqs_idle(ub);
1163 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1164 ublk_cancel_dev(ub);
1165 /* we are going to release task_struct of ubq_daemon and resets
1166 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1167 * Besides, monitor_work is not necessary in QUIESCED state since we have
1168 * already scheduled quiesce_work and quiesced all ubqs.
1170 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1171 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1173 cancel_delayed_work_sync(&ub->monitor_work);
1176 static void ublk_quiesce_work_fn(struct work_struct *work)
1178 struct ublk_device *ub =
1179 container_of(work, struct ublk_device, quiesce_work);
1181 mutex_lock(&ub->mutex);
1182 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1184 __ublk_quiesce_dev(ub);
1186 mutex_unlock(&ub->mutex);
1189 static void ublk_unquiesce_dev(struct ublk_device *ub)
1193 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1194 __func__, ub->dev_info.dev_id,
1195 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1196 "LIVE" : "QUIESCED");
1197 /* quiesce_work has run. We let requeued rqs be aborted
1198 * before running fallback_wq. "force_abort" must be seen
1199 * after request queue is unqiuesced. Then del_gendisk()
1202 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1203 ublk_get_queue(ub, i)->force_abort = true;
1205 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1206 /* We may have requeued some rqs in ublk_quiesce_queue() */
1207 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1210 static void ublk_stop_dev(struct ublk_device *ub)
1212 mutex_lock(&ub->mutex);
1213 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1215 if (ublk_can_use_recovery(ub)) {
1216 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1217 __ublk_quiesce_dev(ub);
1218 ublk_unquiesce_dev(ub);
1220 del_gendisk(ub->ub_disk);
1221 ub->dev_info.state = UBLK_S_DEV_DEAD;
1222 ub->dev_info.ublksrv_pid = -1;
1223 put_disk(ub->ub_disk);
1226 ublk_cancel_dev(ub);
1227 mutex_unlock(&ub->mutex);
1228 cancel_delayed_work_sync(&ub->monitor_work);
1231 /* device can only be started after all IOs are ready */
1232 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1234 mutex_lock(&ub->mutex);
1236 if (ublk_queue_ready(ubq)) {
1237 ubq->ubq_daemon = current;
1238 get_task_struct(ubq->ubq_daemon);
1239 ub->nr_queues_ready++;
1241 if (capable(CAP_SYS_ADMIN))
1242 ub->nr_privileged_daemon++;
1244 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1245 complete_all(&ub->completion);
1246 mutex_unlock(&ub->mutex);
1249 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1252 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1253 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1255 ublk_queue_cmd(ubq, req);
1258 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1260 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1261 struct ublk_device *ub = cmd->file->private_data;
1262 struct ublk_queue *ubq;
1264 u32 cmd_op = cmd->cmd_op;
1265 unsigned tag = ub_cmd->tag;
1268 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1269 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1272 if (!(issue_flags & IO_URING_F_SQE128))
1275 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1278 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1279 if (!ubq || ub_cmd->q_id != ubq->q_id)
1282 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1285 if (tag >= ubq->q_depth)
1288 io = &ubq->ios[tag];
1290 /* there is pending io cmd, something must be wrong */
1291 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1297 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1298 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1300 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1301 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1305 case UBLK_IO_FETCH_REQ:
1306 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1307 if (ublk_queue_ready(ubq)) {
1312 * The io is being handled by server, so COMMIT_RQ is expected
1313 * instead of FETCH_REQ
1315 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1317 /* FETCH_RQ has to provide IO buffer */
1321 io->flags |= UBLK_IO_FLAG_ACTIVE;
1322 io->addr = ub_cmd->addr;
1324 ublk_mark_io_ready(ub, ubq);
1326 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1327 /* FETCH_RQ has to provide IO buffer */
1330 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1332 io->addr = ub_cmd->addr;
1333 io->flags |= UBLK_IO_FLAG_ACTIVE;
1335 ublk_commit_completion(ub, ub_cmd);
1337 case UBLK_IO_NEED_GET_DATA:
1338 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1340 io->addr = ub_cmd->addr;
1342 io->flags |= UBLK_IO_FLAG_ACTIVE;
1343 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1348 return -EIOCBQUEUED;
1351 io_uring_cmd_done(cmd, ret, 0);
1352 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1353 __func__, cmd_op, tag, ret, io->flags);
1354 return -EIOCBQUEUED;
1357 static const struct file_operations ublk_ch_fops = {
1358 .owner = THIS_MODULE,
1359 .open = ublk_ch_open,
1360 .release = ublk_ch_release,
1361 .llseek = no_llseek,
1362 .uring_cmd = ublk_ch_uring_cmd,
1363 .mmap = ublk_ch_mmap,
1366 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1368 int size = ublk_queue_cmd_buf_size(ub, q_id);
1369 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1371 if (ubq->ubq_daemon)
1372 put_task_struct(ubq->ubq_daemon);
1373 if (ubq->io_cmd_buf)
1374 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1377 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1379 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1380 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1384 ubq->flags = ub->dev_info.flags;
1386 ubq->q_depth = ub->dev_info.queue_depth;
1387 size = ublk_queue_cmd_buf_size(ub, q_id);
1389 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1393 ubq->io_cmd_buf = ptr;
1398 static void ublk_deinit_queues(struct ublk_device *ub)
1400 int nr_queues = ub->dev_info.nr_hw_queues;
1406 for (i = 0; i < nr_queues; i++)
1407 ublk_deinit_queue(ub, i);
1408 kfree(ub->__queues);
1411 static int ublk_init_queues(struct ublk_device *ub)
1413 int nr_queues = ub->dev_info.nr_hw_queues;
1414 int depth = ub->dev_info.queue_depth;
1415 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1416 int i, ret = -ENOMEM;
1418 ub->queue_size = ubq_size;
1419 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1423 for (i = 0; i < nr_queues; i++) {
1424 if (ublk_init_queue(ub, i))
1428 init_completion(&ub->completion);
1432 ublk_deinit_queues(ub);
1436 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1441 spin_lock(&ublk_idr_lock);
1442 /* allocate id, if @id >= 0, we're requesting that specific id */
1444 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1448 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1450 spin_unlock(&ublk_idr_lock);
1453 ub->ub_number = err;
1458 static void ublk_free_dev_number(struct ublk_device *ub)
1460 spin_lock(&ublk_idr_lock);
1461 idr_remove(&ublk_index_idr, ub->ub_number);
1462 wake_up_all(&ublk_idr_wq);
1463 spin_unlock(&ublk_idr_lock);
1466 static void ublk_cdev_rel(struct device *dev)
1468 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1470 blk_mq_free_tag_set(&ub->tag_set);
1471 ublk_deinit_queues(ub);
1472 ublk_free_dev_number(ub);
1473 mutex_destroy(&ub->mutex);
1477 static int ublk_add_chdev(struct ublk_device *ub)
1479 struct device *dev = &ub->cdev_dev;
1480 int minor = ub->ub_number;
1483 dev->parent = ublk_misc.this_device;
1484 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1485 dev->class = ublk_chr_class;
1486 dev->release = ublk_cdev_rel;
1487 device_initialize(dev);
1489 ret = dev_set_name(dev, "ublkc%d", minor);
1493 cdev_init(&ub->cdev, &ublk_ch_fops);
1494 ret = cdev_device_add(&ub->cdev, dev);
1505 static void ublk_stop_work_fn(struct work_struct *work)
1507 struct ublk_device *ub =
1508 container_of(work, struct ublk_device, stop_work);
1513 /* align max io buffer size with PAGE_SIZE */
1514 static void ublk_align_max_io_size(struct ublk_device *ub)
1516 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1518 ub->dev_info.max_io_buf_bytes =
1519 round_down(max_io_bytes, PAGE_SIZE);
1522 static int ublk_add_tag_set(struct ublk_device *ub)
1524 ub->tag_set.ops = &ublk_mq_ops;
1525 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1526 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1527 ub->tag_set.numa_node = NUMA_NO_NODE;
1528 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1529 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1530 ub->tag_set.driver_data = ub;
1531 return blk_mq_alloc_tag_set(&ub->tag_set);
1534 static void ublk_remove(struct ublk_device *ub)
1537 cancel_work_sync(&ub->stop_work);
1538 cancel_work_sync(&ub->quiesce_work);
1539 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1540 put_device(&ub->cdev_dev);
1544 static struct ublk_device *ublk_get_device_from_id(int idx)
1546 struct ublk_device *ub = NULL;
1551 spin_lock(&ublk_idr_lock);
1552 ub = idr_find(&ublk_index_idr, idx);
1554 ub = ublk_get_device(ub);
1555 spin_unlock(&ublk_idr_lock);
1560 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1562 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1563 int ublksrv_pid = (int)header->data[0];
1564 struct gendisk *disk;
1567 if (ublksrv_pid <= 0)
1570 wait_for_completion_interruptible(&ub->completion);
1572 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1574 mutex_lock(&ub->mutex);
1575 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1576 test_bit(UB_STATE_USED, &ub->state)) {
1581 disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1583 ret = PTR_ERR(disk);
1586 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1587 disk->fops = &ub_fops;
1588 disk->private_data = ub;
1590 ub->dev_info.ublksrv_pid = ublksrv_pid;
1593 ret = ublk_apply_params(ub);
1597 /* don't probe partitions if any one ubq daemon is un-trusted */
1598 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1599 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1601 get_device(&ub->cdev_dev);
1602 ret = add_disk(disk);
1605 * Has to drop the reference since ->free_disk won't be
1606 * called in case of add_disk failure.
1608 ublk_put_device(ub);
1611 set_bit(UB_STATE_USED, &ub->state);
1612 ub->dev_info.state = UBLK_S_DEV_LIVE;
1617 mutex_unlock(&ub->mutex);
1621 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1622 struct io_uring_cmd *cmd)
1624 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1625 void __user *argp = (void __user *)(unsigned long)header->addr;
1626 cpumask_var_t cpumask;
1627 unsigned long queue;
1628 unsigned int retlen;
1632 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1634 if (header->len & (sizeof(unsigned long)-1))
1639 queue = header->data[0];
1640 if (queue >= ub->dev_info.nr_hw_queues)
1643 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1646 for_each_possible_cpu(i) {
1647 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1648 cpumask_set_cpu(i, cpumask);
1652 retlen = min_t(unsigned short, header->len, cpumask_size());
1653 if (copy_to_user(argp, cpumask, retlen))
1654 goto out_free_cpumask;
1655 if (retlen != header->len &&
1656 clear_user(argp + retlen, header->len - retlen))
1657 goto out_free_cpumask;
1661 free_cpumask_var(cpumask);
1665 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1667 pr_devel("%s: dev id %d flags %llx\n", __func__,
1668 info->dev_id, info->flags);
1669 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1670 info->nr_hw_queues, info->queue_depth);
1673 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1675 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1676 void __user *argp = (void __user *)(unsigned long)header->addr;
1677 struct ublksrv_ctrl_dev_info info;
1678 struct ublk_device *ub;
1681 if (header->len < sizeof(info) || !header->addr)
1683 if (header->queue_id != (u16)-1) {
1684 pr_warn("%s: queue_id is wrong %x\n",
1685 __func__, header->queue_id);
1689 if (copy_from_user(&info, argp, sizeof(info)))
1692 if (capable(CAP_SYS_ADMIN))
1693 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1694 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1697 /* the created device is always owned by current user */
1698 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1700 if (header->dev_id != info.dev_id) {
1701 pr_warn("%s: dev id not match %u %u\n",
1702 __func__, header->dev_id, info.dev_id);
1706 ublk_dump_dev_info(&info);
1708 ret = mutex_lock_killable(&ublk_ctl_mutex);
1713 if (ublks_added >= ublks_max)
1717 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1720 mutex_init(&ub->mutex);
1721 spin_lock_init(&ub->mm_lock);
1722 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1723 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1724 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1726 ret = ublk_alloc_dev_number(ub, header->dev_id);
1730 memcpy(&ub->dev_info, &info, sizeof(info));
1732 /* update device id */
1733 ub->dev_info.dev_id = ub->ub_number;
1736 * 64bit flags will be copied back to userspace as feature
1737 * negotiation result, so have to clear flags which driver
1738 * doesn't support yet, then userspace can get correct flags
1739 * (features) to handle.
1741 ub->dev_info.flags &= UBLK_F_ALL;
1743 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1744 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1746 /* We are not ready to support zero copy */
1747 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1749 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1750 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1751 ublk_align_max_io_size(ub);
1753 ret = ublk_init_queues(ub);
1755 goto out_free_dev_number;
1757 ret = ublk_add_tag_set(ub);
1759 goto out_deinit_queues;
1762 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1763 goto out_free_tag_set;
1766 * Add the char dev so that ublksrv daemon can be setup.
1767 * ublk_add_chdev() will cleanup everything if it fails.
1769 ret = ublk_add_chdev(ub);
1773 blk_mq_free_tag_set(&ub->tag_set);
1775 ublk_deinit_queues(ub);
1776 out_free_dev_number:
1777 ublk_free_dev_number(ub);
1779 mutex_destroy(&ub->mutex);
1782 mutex_unlock(&ublk_ctl_mutex);
1786 static inline bool ublk_idr_freed(int id)
1790 spin_lock(&ublk_idr_lock);
1791 ptr = idr_find(&ublk_index_idr, id);
1792 spin_unlock(&ublk_idr_lock);
1797 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1799 struct ublk_device *ub = *p_ub;
1800 int idx = ub->ub_number;
1803 ret = mutex_lock_killable(&ublk_ctl_mutex);
1809 /* Mark the reference as consumed */
1811 ublk_put_device(ub);
1814 * Wait until the idr is removed, then it can be reused after
1815 * DEL_DEV command is returned.
1817 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1818 mutex_unlock(&ublk_ctl_mutex);
1823 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1825 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1827 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1828 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1829 header->data[0], header->addr, header->len);
1832 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1835 cancel_work_sync(&ub->stop_work);
1836 cancel_work_sync(&ub->quiesce_work);
1841 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1842 struct io_uring_cmd *cmd)
1844 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1845 void __user *argp = (void __user *)(unsigned long)header->addr;
1847 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1850 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1856 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1857 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1859 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1860 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1863 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1864 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1866 ub->params.devt.disk_major = 0;
1867 ub->params.devt.disk_minor = 0;
1869 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1872 static int ublk_ctrl_get_params(struct ublk_device *ub,
1873 struct io_uring_cmd *cmd)
1875 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1876 void __user *argp = (void __user *)(unsigned long)header->addr;
1877 struct ublk_params_header ph;
1880 if (header->len <= sizeof(ph) || !header->addr)
1883 if (copy_from_user(&ph, argp, sizeof(ph)))
1886 if (ph.len > header->len || !ph.len)
1889 if (ph.len > sizeof(struct ublk_params))
1890 ph.len = sizeof(struct ublk_params);
1892 mutex_lock(&ub->mutex);
1893 ublk_ctrl_fill_params_devt(ub);
1894 if (copy_to_user(argp, &ub->params, ph.len))
1898 mutex_unlock(&ub->mutex);
1903 static int ublk_ctrl_set_params(struct ublk_device *ub,
1904 struct io_uring_cmd *cmd)
1906 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1907 void __user *argp = (void __user *)(unsigned long)header->addr;
1908 struct ublk_params_header ph;
1911 if (header->len <= sizeof(ph) || !header->addr)
1914 if (copy_from_user(&ph, argp, sizeof(ph)))
1917 if (ph.len > header->len || !ph.len || !ph.types)
1920 if (ph.len > sizeof(struct ublk_params))
1921 ph.len = sizeof(struct ublk_params);
1923 /* parameters can only be changed when device isn't live */
1924 mutex_lock(&ub->mutex);
1925 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1927 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1930 /* clear all we don't support yet */
1931 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1932 ret = ublk_validate_params(ub);
1934 mutex_unlock(&ub->mutex);
1939 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1943 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1944 /* All old ioucmds have to be completed */
1945 WARN_ON_ONCE(ubq->nr_io_ready);
1946 /* old daemon is PF_EXITING, put it now */
1947 put_task_struct(ubq->ubq_daemon);
1948 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1949 ubq->ubq_daemon = NULL;
1951 for (i = 0; i < ubq->q_depth; i++) {
1952 struct ublk_io *io = &ubq->ios[i];
1954 /* forget everything now and be ready for new FETCH_REQ */
1961 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1962 struct io_uring_cmd *cmd)
1964 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1968 mutex_lock(&ub->mutex);
1969 if (!ublk_can_use_recovery(ub))
1972 * START_RECOVERY is only allowd after:
1974 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1975 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1978 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1979 * (a)has quiesced request queue
1980 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1981 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1982 * (d)has completed/camceled all ioucmds owned by ther dying process
1984 if (test_bit(UB_STATE_OPEN, &ub->state) ||
1985 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1989 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1990 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1991 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1992 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1994 ub->nr_queues_ready = 0;
1995 ub->nr_privileged_daemon = 0;
1996 init_completion(&ub->completion);
1999 mutex_unlock(&ub->mutex);
2003 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2004 struct io_uring_cmd *cmd)
2006 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2007 int ublksrv_pid = (int)header->data[0];
2010 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2011 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2012 /* wait until new ubq_daemon sending all FETCH_REQ */
2013 wait_for_completion_interruptible(&ub->completion);
2014 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2015 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2017 mutex_lock(&ub->mutex);
2018 if (!ublk_can_use_recovery(ub))
2021 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2025 ub->dev_info.ublksrv_pid = ublksrv_pid;
2026 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2027 __func__, ublksrv_pid, header->dev_id);
2028 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2029 pr_devel("%s: queue unquiesced, dev id %d.\n",
2030 __func__, header->dev_id);
2031 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2032 ub->dev_info.state = UBLK_S_DEV_LIVE;
2033 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2036 mutex_unlock(&ub->mutex);
2041 * All control commands are sent via /dev/ublk-control, so we have to check
2042 * the destination device's permission
2044 static int ublk_char_dev_permission(struct ublk_device *ub,
2045 const char *dev_path, int mask)
2051 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2055 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2060 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2063 err = inode_permission(&init_user_ns,
2064 d_backing_inode(path.dentry), mask);
2070 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2071 struct io_uring_cmd *cmd)
2073 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2074 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2075 void __user *argp = (void __user *)(unsigned long)header->addr;
2076 char *dev_path = NULL;
2080 if (!unprivileged) {
2081 if (!capable(CAP_SYS_ADMIN))
2084 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2085 * char_dev_path in payload too, since userspace may not
2086 * know if the specified device is created as unprivileged
2089 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2094 * User has to provide the char device path for unprivileged ublk
2096 * header->addr always points to the dev path buffer, and
2097 * header->dev_path_len records length of dev path buffer.
2099 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2102 if (header->len < header->dev_path_len)
2105 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2110 if (copy_from_user(dev_path, argp, header->dev_path_len))
2112 dev_path[header->dev_path_len] = 0;
2115 switch (cmd->cmd_op) {
2116 case UBLK_CMD_GET_DEV_INFO:
2117 case UBLK_CMD_GET_DEV_INFO2:
2118 case UBLK_CMD_GET_QUEUE_AFFINITY:
2119 case UBLK_CMD_GET_PARAMS:
2122 case UBLK_CMD_START_DEV:
2123 case UBLK_CMD_STOP_DEV:
2124 case UBLK_CMD_ADD_DEV:
2125 case UBLK_CMD_DEL_DEV:
2126 case UBLK_CMD_SET_PARAMS:
2127 case UBLK_CMD_START_USER_RECOVERY:
2128 case UBLK_CMD_END_USER_RECOVERY:
2129 mask = MAY_READ | MAY_WRITE;
2135 ret = ublk_char_dev_permission(ub, dev_path, mask);
2137 header->len -= header->dev_path_len;
2138 header->addr += header->dev_path_len;
2140 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2141 __func__, ub->ub_number, cmd->cmd_op,
2142 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2149 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2150 unsigned int issue_flags)
2152 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2153 struct ublk_device *ub = NULL;
2156 if (issue_flags & IO_URING_F_NONBLOCK)
2159 ublk_ctrl_cmd_dump(cmd);
2161 if (!(issue_flags & IO_URING_F_SQE128))
2164 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2166 ub = ublk_get_device_from_id(header->dev_id);
2170 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2172 /* ADD_DEV permission check is done in command handler */
2179 switch (cmd->cmd_op) {
2180 case UBLK_CMD_START_DEV:
2181 ret = ublk_ctrl_start_dev(ub, cmd);
2183 case UBLK_CMD_STOP_DEV:
2184 ret = ublk_ctrl_stop_dev(ub);
2186 case UBLK_CMD_GET_DEV_INFO:
2187 case UBLK_CMD_GET_DEV_INFO2:
2188 ret = ublk_ctrl_get_dev_info(ub, cmd);
2190 case UBLK_CMD_ADD_DEV:
2191 ret = ublk_ctrl_add_dev(cmd);
2193 case UBLK_CMD_DEL_DEV:
2194 ret = ublk_ctrl_del_dev(&ub);
2196 case UBLK_CMD_GET_QUEUE_AFFINITY:
2197 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2199 case UBLK_CMD_GET_PARAMS:
2200 ret = ublk_ctrl_get_params(ub, cmd);
2202 case UBLK_CMD_SET_PARAMS:
2203 ret = ublk_ctrl_set_params(ub, cmd);
2205 case UBLK_CMD_START_USER_RECOVERY:
2206 ret = ublk_ctrl_start_recovery(ub, cmd);
2208 case UBLK_CMD_END_USER_RECOVERY:
2209 ret = ublk_ctrl_end_recovery(ub, cmd);
2218 ublk_put_device(ub);
2220 io_uring_cmd_done(cmd, ret, 0);
2221 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2222 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2223 return -EIOCBQUEUED;
2226 static const struct file_operations ublk_ctl_fops = {
2227 .open = nonseekable_open,
2228 .uring_cmd = ublk_ctrl_uring_cmd,
2229 .owner = THIS_MODULE,
2230 .llseek = noop_llseek,
2233 static struct miscdevice ublk_misc = {
2234 .minor = MISC_DYNAMIC_MINOR,
2235 .name = "ublk-control",
2236 .fops = &ublk_ctl_fops,
2239 static int __init ublk_init(void)
2243 init_waitqueue_head(&ublk_idr_wq);
2245 ret = misc_register(&ublk_misc);
2249 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2251 goto unregister_mis;
2253 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2254 if (IS_ERR(ublk_chr_class)) {
2255 ret = PTR_ERR(ublk_chr_class);
2256 goto free_chrdev_region;
2261 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2263 misc_deregister(&ublk_misc);
2267 static void __exit ublk_exit(void)
2269 struct ublk_device *ub;
2272 idr_for_each_entry(&ublk_index_idr, ub, id)
2275 class_destroy(ublk_chr_class);
2276 misc_deregister(&ublk_misc);
2278 idr_destroy(&ublk_index_idr);
2279 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2282 module_init(ublk_init);
2283 module_exit(ublk_exit);
2285 module_param(ublks_max, int, 0444);
2286 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2288 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2289 MODULE_LICENSE("GPL");