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 int 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
153 #define UB_STATE_DELETED 2
160 struct mm_struct *mm;
162 struct ublk_params params;
164 struct completion completion;
165 unsigned int nr_queues_ready;
166 unsigned int nr_privileged_daemon;
169 * Our ubq->daemon may be killed without any notification, so
170 * monitor each queue's daemon periodically
172 struct delayed_work monitor_work;
173 struct work_struct quiesce_work;
174 struct work_struct stop_work;
177 /* header of ublk_params */
178 struct ublk_params_header {
183 static dev_t ublk_chr_devt;
184 static struct class *ublk_chr_class;
186 static DEFINE_IDR(ublk_index_idr);
187 static DEFINE_SPINLOCK(ublk_idr_lock);
188 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
190 static DEFINE_MUTEX(ublk_ctl_mutex);
193 * Max ublk devices allowed to add
195 * It can be extended to one per-user limit in future or even controlled
198 static unsigned int ublks_max = 64;
199 static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
201 static struct miscdevice ublk_misc;
203 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
205 struct request_queue *q = ub->ub_disk->queue;
206 const struct ublk_param_basic *p = &ub->params.basic;
208 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
209 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
210 blk_queue_io_min(q, 1 << p->io_min_shift);
211 blk_queue_io_opt(q, 1 << p->io_opt_shift);
213 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
214 p->attrs & UBLK_ATTR_FUA);
215 if (p->attrs & UBLK_ATTR_ROTATIONAL)
216 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
218 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
220 blk_queue_max_hw_sectors(q, p->max_sectors);
221 blk_queue_chunk_sectors(q, p->chunk_sectors);
222 blk_queue_virt_boundary(q, p->virt_boundary_mask);
224 if (p->attrs & UBLK_ATTR_READ_ONLY)
225 set_disk_ro(ub->ub_disk, true);
227 set_capacity(ub->ub_disk, p->dev_sectors);
230 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
232 struct request_queue *q = ub->ub_disk->queue;
233 const struct ublk_param_discard *p = &ub->params.discard;
235 q->limits.discard_alignment = p->discard_alignment;
236 q->limits.discard_granularity = p->discard_granularity;
237 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
238 blk_queue_max_write_zeroes_sectors(q,
239 p->max_write_zeroes_sectors);
240 blk_queue_max_discard_segments(q, p->max_discard_segments);
243 static int ublk_validate_params(const struct ublk_device *ub)
245 /* basic param is the only one which must be set */
246 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
247 const struct ublk_param_basic *p = &ub->params.basic;
249 if (p->logical_bs_shift > PAGE_SHIFT)
252 if (p->logical_bs_shift > p->physical_bs_shift)
255 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
260 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
261 const struct ublk_param_discard *p = &ub->params.discard;
263 /* So far, only support single segment discard */
264 if (p->max_discard_sectors && p->max_discard_segments != 1)
267 if (!p->discard_granularity)
271 /* dev_t is read-only */
272 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
278 static int ublk_apply_params(struct ublk_device *ub)
280 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
283 ublk_dev_param_basic_apply(ub);
285 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
286 ublk_dev_param_discard_apply(ub);
291 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
293 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
294 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
299 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
301 if (ubq->flags & UBLK_F_NEED_GET_DATA)
306 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
308 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
313 static void ublk_put_device(struct ublk_device *ub)
315 put_device(&ub->cdev_dev);
318 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
321 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
324 static inline bool ublk_rq_has_data(const struct request *rq)
326 return bio_has_data(rq->bio);
329 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
332 return (struct ublksrv_io_desc *)
333 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
336 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
338 return ublk_get_queue(ub, q_id)->io_cmd_buf;
341 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
343 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
345 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
349 static inline bool ublk_queue_can_use_recovery_reissue(
350 struct ublk_queue *ubq)
352 if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
353 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
358 static inline bool ublk_queue_can_use_recovery(
359 struct ublk_queue *ubq)
361 if (ubq->flags & UBLK_F_USER_RECOVERY)
366 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
368 if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
373 static void ublk_free_disk(struct gendisk *disk)
375 struct ublk_device *ub = disk->private_data;
377 clear_bit(UB_STATE_USED, &ub->state);
378 put_device(&ub->cdev_dev);
381 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
382 unsigned int *owner_gid)
387 current_uid_gid(&uid, &gid);
389 *owner_uid = from_kuid(&init_user_ns, uid);
390 *owner_gid = from_kgid(&init_user_ns, gid);
393 static int ublk_open(struct block_device *bdev, fmode_t mode)
395 struct ublk_device *ub = bdev->bd_disk->private_data;
397 if (capable(CAP_SYS_ADMIN))
401 * If it is one unprivileged device, only owner can open
402 * the disk. Otherwise it could be one trap made by one
403 * evil user who grants this disk's privileges to other
404 * users deliberately.
406 * This way is reasonable too given anyone can create
407 * unprivileged device, and no need other's grant.
409 if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
410 unsigned int curr_uid, curr_gid;
412 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
414 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
415 ub->dev_info.owner_gid)
422 static const struct block_device_operations ub_fops = {
423 .owner = THIS_MODULE,
425 .free_disk = ublk_free_disk,
428 #define UBLK_MAX_PIN_PAGES 32
430 struct ublk_map_data {
431 const struct ublk_queue *ubq;
432 const struct request *rq;
433 const struct ublk_io *io;
437 struct ublk_io_iter {
438 struct page *pages[UBLK_MAX_PIN_PAGES];
439 unsigned pg_off; /* offset in the 1st page in pages */
440 int nr_pages; /* how many page pointers in pages */
442 struct bvec_iter iter;
445 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
446 unsigned max_bytes, bool to_vm)
448 const unsigned total = min_t(unsigned, max_bytes,
449 PAGE_SIZE - data->pg_off +
450 ((data->nr_pages - 1) << PAGE_SHIFT));
454 while (done < total) {
455 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
456 const unsigned int bytes = min3(bv.bv_len, total - done,
457 (unsigned)(PAGE_SIZE - data->pg_off));
458 void *bv_buf = bvec_kmap_local(&bv);
459 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
462 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
464 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
466 kunmap_local(pg_buf);
467 kunmap_local(bv_buf);
469 /* advance page array */
470 data->pg_off += bytes;
471 if (data->pg_off == PAGE_SIZE) {
479 bio_advance_iter_single(data->bio, &data->iter, bytes);
480 if (!data->iter.bi_size) {
481 data->bio = data->bio->bi_next;
482 if (data->bio == NULL)
484 data->iter = data->bio->bi_iter;
491 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
494 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
495 const unsigned long start_vm = data->io->addr;
496 unsigned int done = 0;
497 struct ublk_io_iter iter = {
498 .pg_off = start_vm & (PAGE_SIZE - 1),
499 .bio = data->rq->bio,
500 .iter = data->rq->bio->bi_iter,
502 const unsigned int nr_pages = round_up(data->max_bytes +
503 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
505 while (done < nr_pages) {
506 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
510 iter.nr_pages = get_user_pages_fast(start_vm +
511 (done << PAGE_SHIFT), to_pin, gup_flags,
513 if (iter.nr_pages <= 0)
514 return done == 0 ? iter.nr_pages : done;
515 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
516 for (i = 0; i < iter.nr_pages; i++) {
518 set_page_dirty(iter.pages[i]);
519 put_page(iter.pages[i]);
521 data->max_bytes -= len;
522 done += iter.nr_pages;
528 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
531 const unsigned int rq_bytes = blk_rq_bytes(req);
533 * no zero copy, we delay copy WRITE request data into ublksrv
534 * context and the big benefit is that pinning pages in current
535 * context is pretty fast, see ublk_pin_user_pages
537 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
540 if (ublk_rq_has_data(req)) {
541 struct ublk_map_data data = {
545 .max_bytes = rq_bytes,
548 ublk_copy_user_pages(&data, true);
550 return rq_bytes - data.max_bytes;
555 static int ublk_unmap_io(const struct ublk_queue *ubq,
556 const struct request *req,
559 const unsigned int rq_bytes = blk_rq_bytes(req);
561 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
562 struct ublk_map_data data = {
566 .max_bytes = io->res,
569 WARN_ON_ONCE(io->res > rq_bytes);
571 ublk_copy_user_pages(&data, false);
573 return io->res - data.max_bytes;
578 static inline unsigned int ublk_req_build_flags(struct request *req)
582 if (req->cmd_flags & REQ_FAILFAST_DEV)
583 flags |= UBLK_IO_F_FAILFAST_DEV;
585 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
586 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
588 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
589 flags |= UBLK_IO_F_FAILFAST_DRIVER;
591 if (req->cmd_flags & REQ_META)
592 flags |= UBLK_IO_F_META;
594 if (req->cmd_flags & REQ_FUA)
595 flags |= UBLK_IO_F_FUA;
597 if (req->cmd_flags & REQ_NOUNMAP)
598 flags |= UBLK_IO_F_NOUNMAP;
600 if (req->cmd_flags & REQ_SWAP)
601 flags |= UBLK_IO_F_SWAP;
606 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
608 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
609 struct ublk_io *io = &ubq->ios[req->tag];
612 switch (req_op(req)) {
614 ublk_op = UBLK_IO_OP_READ;
617 ublk_op = UBLK_IO_OP_WRITE;
620 ublk_op = UBLK_IO_OP_FLUSH;
623 ublk_op = UBLK_IO_OP_DISCARD;
625 case REQ_OP_WRITE_ZEROES:
626 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
629 return BLK_STS_IOERR;
632 /* need to translate since kernel may change */
633 iod->op_flags = ublk_op | ublk_req_build_flags(req);
634 iod->nr_sectors = blk_rq_sectors(req);
635 iod->start_sector = blk_rq_pos(req);
636 iod->addr = io->addr;
641 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
642 struct io_uring_cmd *ioucmd)
644 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
647 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
649 return ubq->ubq_daemon->flags & PF_EXITING;
652 /* todo: handle partial completion */
653 static void ublk_complete_rq(struct request *req)
655 struct ublk_queue *ubq = req->mq_hctx->driver_data;
656 struct ublk_io *io = &ubq->ios[req->tag];
657 unsigned int unmapped_bytes;
658 blk_status_t res = BLK_STS_OK;
660 /* failed read IO if nothing is read */
661 if (!io->res && req_op(req) == REQ_OP_READ)
665 res = errno_to_blk_status(io->res);
670 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
673 * Both the two needn't unmap.
675 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
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 blk_mq_end_request(req, res);
700 * Since __ublk_rq_task_work always fails requests immediately during
701 * exiting, __ublk_fail_req() is only called from abort context during
702 * exiting. So lock is unnecessary.
704 * Also aborting may not be started yet, keep in mind that one failed
705 * request may be issued by block layer again.
707 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
710 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
712 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
713 io->flags |= UBLK_IO_FLAG_ABORTED;
714 if (ublk_queue_can_use_recovery_reissue(ubq))
715 blk_mq_requeue_request(req, false);
717 blk_mq_end_request(req, BLK_STS_IOERR);
721 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
723 /* mark this cmd owned by ublksrv */
724 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
727 * clear ACTIVE since we are done with this sqe/cmd slot
728 * We can only accept io cmd in case of being not active.
730 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
732 /* tell ublksrv one io request is coming */
733 io_uring_cmd_done(io->cmd, res, 0);
736 #define UBLK_REQUEUE_DELAY_MS 3
738 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
741 /* We cannot process this rq so just requeue it. */
742 if (ublk_queue_can_use_recovery(ubq))
743 blk_mq_requeue_request(rq, false);
745 blk_mq_end_request(rq, BLK_STS_IOERR);
747 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
750 static inline void __ublk_rq_task_work(struct request *req)
752 struct ublk_queue *ubq = req->mq_hctx->driver_data;
754 struct ublk_io *io = &ubq->ios[tag];
755 unsigned int mapped_bytes;
757 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
758 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
759 ublk_get_iod(ubq, req->tag)->addr);
762 * Task is exiting if either:
764 * (1) current != ubq_daemon.
765 * io_uring_cmd_complete_in_task() tries to run task_work
766 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
768 * (2) current->flags & PF_EXITING.
770 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
771 __ublk_abort_rq(ubq, req);
775 if (ublk_need_get_data(ubq) &&
776 (req_op(req) == REQ_OP_WRITE ||
777 req_op(req) == REQ_OP_FLUSH)) {
779 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
780 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
783 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
784 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
785 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
786 __func__, io->cmd->cmd_op, ubq->q_id,
787 req->tag, io->flags);
788 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
792 * We have handled UBLK_IO_NEED_GET_DATA command,
793 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
796 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
797 /* update iod->addr because ublksrv may have passed a new io buffer */
798 ublk_get_iod(ubq, req->tag)->addr = io->addr;
799 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
800 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
801 ublk_get_iod(ubq, req->tag)->addr);
804 mapped_bytes = ublk_map_io(ubq, req, io);
806 /* partially mapped, update io descriptor */
807 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
809 * Nothing mapped, retry until we succeed.
811 * We may never succeed in mapping any bytes here because
812 * of OOM. TODO: reserve one buffer with single page pinned
813 * for providing forward progress guarantee.
815 if (unlikely(!mapped_bytes)) {
816 blk_mq_requeue_request(req, false);
817 blk_mq_delay_kick_requeue_list(req->q,
818 UBLK_REQUEUE_DELAY_MS);
822 ublk_get_iod(ubq, req->tag)->nr_sectors =
826 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
829 static inline void ublk_forward_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 io_cmds = llist_reverse_order(io_cmds);
835 llist_for_each_entry_safe(data, tmp, io_cmds, node)
836 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
839 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
841 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
842 struct ublk_rq_data *data, *tmp;
844 llist_for_each_entry_safe(data, tmp, io_cmds, node)
845 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
848 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
850 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
851 struct ublk_queue *ubq = pdu->ubq;
853 ublk_forward_io_cmds(ubq);
856 static void ublk_rq_task_work_fn(struct callback_head *work)
858 struct ublk_rq_data *data = container_of(work,
859 struct ublk_rq_data, work);
860 struct request *req = blk_mq_rq_from_pdu(data);
861 struct ublk_queue *ubq = req->mq_hctx->driver_data;
863 ublk_forward_io_cmds(ubq);
866 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
868 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
871 if (!llist_add(&data->node, &ubq->io_cmds))
874 io = &ubq->ios[rq->tag];
876 * If the check pass, we know that this is a re-issued request aborted
877 * previously in monitor_work because the ubq_daemon(cmd's task) is
878 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
879 * because this ioucmd's io_uring context may be freed now if no inflight
880 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
882 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
883 * the tag). Then the request is re-started(allocating the tag) and we are here.
884 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
885 * guarantees that here is a re-issued request aborted previously.
887 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
888 ublk_abort_io_cmds(ubq);
889 } else if (ublk_can_use_task_work(ubq)) {
890 if (task_work_add(ubq->ubq_daemon, &data->work,
892 ublk_abort_io_cmds(ubq);
894 struct io_uring_cmd *cmd = io->cmd;
895 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
898 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
902 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
903 const struct blk_mq_queue_data *bd)
905 struct ublk_queue *ubq = hctx->driver_data;
906 struct request *rq = bd->rq;
909 /* fill iod to slot in io cmd buffer */
910 res = ublk_setup_iod(ubq, rq);
911 if (unlikely(res != BLK_STS_OK))
912 return BLK_STS_IOERR;
914 /* With recovery feature enabled, force_abort is set in
915 * ublk_stop_dev() before calling del_gendisk(). We have to
916 * abort all requeued and new rqs here to let del_gendisk()
917 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
918 * to avoid UAF on io_uring ctx.
920 * Note: force_abort is guaranteed to be seen because it is set
921 * before request queue is unqiuesced.
923 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
924 return BLK_STS_IOERR;
926 blk_mq_start_request(bd->rq);
928 if (unlikely(ubq_daemon_is_dying(ubq))) {
929 __ublk_abort_rq(ubq, rq);
933 ublk_queue_cmd(ubq, rq);
938 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
939 unsigned int hctx_idx)
941 struct ublk_device *ub = driver_data;
942 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
944 hctx->driver_data = ubq;
948 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
949 unsigned int hctx_idx, unsigned int numa_node)
951 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
953 init_task_work(&data->work, ublk_rq_task_work_fn);
957 static const struct blk_mq_ops ublk_mq_ops = {
958 .queue_rq = ublk_queue_rq,
959 .init_hctx = ublk_init_hctx,
960 .init_request = ublk_init_rq,
963 static int ublk_ch_open(struct inode *inode, struct file *filp)
965 struct ublk_device *ub = container_of(inode->i_cdev,
966 struct ublk_device, cdev);
968 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
970 filp->private_data = ub;
974 static int ublk_ch_release(struct inode *inode, struct file *filp)
976 struct ublk_device *ub = filp->private_data;
978 clear_bit(UB_STATE_OPEN, &ub->state);
982 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
983 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
985 struct ublk_device *ub = filp->private_data;
986 size_t sz = vma->vm_end - vma->vm_start;
987 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
988 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
991 spin_lock(&ub->mm_lock);
993 ub->mm = current->mm;
994 if (current->mm != ub->mm)
996 spin_unlock(&ub->mm_lock);
1001 if (vma->vm_flags & VM_WRITE)
1004 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1005 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1008 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1009 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1010 __func__, q_id, current->pid, vma->vm_start,
1011 phys_off, (unsigned long)sz);
1013 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1016 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1017 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1020 static void ublk_commit_completion(struct ublk_device *ub,
1021 struct ublksrv_io_cmd *ub_cmd)
1023 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1024 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1025 struct ublk_io *io = &ubq->ios[tag];
1026 struct request *req;
1028 /* now this cmd slot is owned by nbd driver */
1029 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1030 io->res = ub_cmd->result;
1032 /* find the io request and complete */
1033 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1035 if (req && likely(!blk_should_fake_timeout(req->q)))
1036 ublk_complete_rq(req);
1040 * When ->ubq_daemon is exiting, either new request is ended immediately,
1041 * or any queued io command is drained, so it is safe to abort queue
1044 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1048 if (!ublk_get_device(ub))
1051 for (i = 0; i < ubq->q_depth; i++) {
1052 struct ublk_io *io = &ubq->ios[i];
1054 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1058 * Either we fail the request or ublk_rq_task_work_fn
1061 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1063 __ublk_fail_req(ubq, io, rq);
1066 ublk_put_device(ub);
1069 static void ublk_daemon_monitor_work(struct work_struct *work)
1071 struct ublk_device *ub =
1072 container_of(work, struct ublk_device, monitor_work.work);
1075 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1076 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1078 if (ubq_daemon_is_dying(ubq)) {
1079 if (ublk_queue_can_use_recovery(ubq))
1080 schedule_work(&ub->quiesce_work);
1082 schedule_work(&ub->stop_work);
1084 /* abort queue is for making forward progress */
1085 ublk_abort_queue(ub, ubq);
1090 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1091 * after ublk_remove() or __ublk_quiesce_dev() is started.
1093 * No need ub->mutex, monitor work are canceled after state is marked
1094 * as not LIVE, so new state is observed reliably.
1096 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1097 schedule_delayed_work(&ub->monitor_work,
1098 UBLK_DAEMON_MONITOR_PERIOD);
1101 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1103 return ubq->nr_io_ready == ubq->q_depth;
1106 static void ublk_cancel_queue(struct ublk_queue *ubq)
1110 if (!ublk_queue_ready(ubq))
1113 for (i = 0; i < ubq->q_depth; i++) {
1114 struct ublk_io *io = &ubq->ios[i];
1116 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1117 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1120 /* all io commands are canceled */
1121 ubq->nr_io_ready = 0;
1124 /* Cancel all pending commands, must be called after del_gendisk() returns */
1125 static void ublk_cancel_dev(struct ublk_device *ub)
1129 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1130 ublk_cancel_queue(ublk_get_queue(ub, i));
1133 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1137 if (blk_mq_request_started(rq)) {
1144 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1148 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1151 blk_mq_tagset_busy_iter(&ub->tag_set,
1152 ublk_check_inflight_rq, &idle);
1155 msleep(UBLK_REQUEUE_DELAY_MS);
1159 static void __ublk_quiesce_dev(struct ublk_device *ub)
1161 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1162 __func__, ub->dev_info.dev_id,
1163 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1164 "LIVE" : "QUIESCED");
1165 blk_mq_quiesce_queue(ub->ub_disk->queue);
1166 ublk_wait_tagset_rqs_idle(ub);
1167 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1168 ublk_cancel_dev(ub);
1169 /* we are going to release task_struct of ubq_daemon and resets
1170 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1171 * Besides, monitor_work is not necessary in QUIESCED state since we have
1172 * already scheduled quiesce_work and quiesced all ubqs.
1174 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1175 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1177 cancel_delayed_work_sync(&ub->monitor_work);
1180 static void ublk_quiesce_work_fn(struct work_struct *work)
1182 struct ublk_device *ub =
1183 container_of(work, struct ublk_device, quiesce_work);
1185 mutex_lock(&ub->mutex);
1186 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1188 __ublk_quiesce_dev(ub);
1190 mutex_unlock(&ub->mutex);
1193 static void ublk_unquiesce_dev(struct ublk_device *ub)
1197 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1198 __func__, ub->dev_info.dev_id,
1199 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1200 "LIVE" : "QUIESCED");
1201 /* quiesce_work has run. We let requeued rqs be aborted
1202 * before running fallback_wq. "force_abort" must be seen
1203 * after request queue is unqiuesced. Then del_gendisk()
1206 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1207 ublk_get_queue(ub, i)->force_abort = true;
1209 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1210 /* We may have requeued some rqs in ublk_quiesce_queue() */
1211 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1214 static void ublk_stop_dev(struct ublk_device *ub)
1216 mutex_lock(&ub->mutex);
1217 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1219 if (ublk_can_use_recovery(ub)) {
1220 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1221 __ublk_quiesce_dev(ub);
1222 ublk_unquiesce_dev(ub);
1224 del_gendisk(ub->ub_disk);
1225 ub->dev_info.state = UBLK_S_DEV_DEAD;
1226 ub->dev_info.ublksrv_pid = -1;
1227 put_disk(ub->ub_disk);
1230 ublk_cancel_dev(ub);
1231 mutex_unlock(&ub->mutex);
1232 cancel_delayed_work_sync(&ub->monitor_work);
1235 /* device can only be started after all IOs are ready */
1236 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1238 mutex_lock(&ub->mutex);
1240 if (ublk_queue_ready(ubq)) {
1241 ubq->ubq_daemon = current;
1242 get_task_struct(ubq->ubq_daemon);
1243 ub->nr_queues_ready++;
1245 if (capable(CAP_SYS_ADMIN))
1246 ub->nr_privileged_daemon++;
1248 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1249 complete_all(&ub->completion);
1250 mutex_unlock(&ub->mutex);
1253 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1256 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1257 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1259 ublk_queue_cmd(ubq, req);
1262 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1264 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1265 struct ublk_device *ub = cmd->file->private_data;
1266 struct ublk_queue *ubq;
1268 u32 cmd_op = cmd->cmd_op;
1269 unsigned tag = ub_cmd->tag;
1271 struct request *req;
1273 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1274 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1277 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1280 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1281 if (!ubq || ub_cmd->q_id != ubq->q_id)
1284 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1287 if (tag >= ubq->q_depth)
1290 io = &ubq->ios[tag];
1292 /* there is pending io cmd, something must be wrong */
1293 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1299 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1300 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1302 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1303 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1307 case UBLK_IO_FETCH_REQ:
1308 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1309 if (ublk_queue_ready(ubq)) {
1314 * The io is being handled by server, so COMMIT_RQ is expected
1315 * instead of FETCH_REQ
1317 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1319 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1320 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1323 io->flags |= UBLK_IO_FLAG_ACTIVE;
1324 io->addr = ub_cmd->addr;
1326 ublk_mark_io_ready(ub, ubq);
1328 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1329 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1331 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1332 * not enabled or it is Read IO.
1334 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1336 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1338 io->addr = ub_cmd->addr;
1339 io->flags |= UBLK_IO_FLAG_ACTIVE;
1341 ublk_commit_completion(ub, ub_cmd);
1343 case UBLK_IO_NEED_GET_DATA:
1344 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1346 io->addr = ub_cmd->addr;
1348 io->flags |= UBLK_IO_FLAG_ACTIVE;
1349 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1354 return -EIOCBQUEUED;
1357 io_uring_cmd_done(cmd, ret, 0);
1358 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1359 __func__, cmd_op, tag, ret, io->flags);
1360 return -EIOCBQUEUED;
1363 static const struct file_operations ublk_ch_fops = {
1364 .owner = THIS_MODULE,
1365 .open = ublk_ch_open,
1366 .release = ublk_ch_release,
1367 .llseek = no_llseek,
1368 .uring_cmd = ublk_ch_uring_cmd,
1369 .mmap = ublk_ch_mmap,
1372 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1374 int size = ublk_queue_cmd_buf_size(ub, q_id);
1375 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1377 if (ubq->ubq_daemon)
1378 put_task_struct(ubq->ubq_daemon);
1379 if (ubq->io_cmd_buf)
1380 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1383 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1385 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1386 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1390 ubq->flags = ub->dev_info.flags;
1392 ubq->q_depth = ub->dev_info.queue_depth;
1393 size = ublk_queue_cmd_buf_size(ub, q_id);
1395 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1399 ubq->io_cmd_buf = ptr;
1404 static void ublk_deinit_queues(struct ublk_device *ub)
1406 int nr_queues = ub->dev_info.nr_hw_queues;
1412 for (i = 0; i < nr_queues; i++)
1413 ublk_deinit_queue(ub, i);
1414 kfree(ub->__queues);
1417 static int ublk_init_queues(struct ublk_device *ub)
1419 int nr_queues = ub->dev_info.nr_hw_queues;
1420 int depth = ub->dev_info.queue_depth;
1421 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1422 int i, ret = -ENOMEM;
1424 ub->queue_size = ubq_size;
1425 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1429 for (i = 0; i < nr_queues; i++) {
1430 if (ublk_init_queue(ub, i))
1434 init_completion(&ub->completion);
1438 ublk_deinit_queues(ub);
1442 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1447 spin_lock(&ublk_idr_lock);
1448 /* allocate id, if @id >= 0, we're requesting that specific id */
1450 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1454 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1456 spin_unlock(&ublk_idr_lock);
1459 ub->ub_number = err;
1464 static void ublk_free_dev_number(struct ublk_device *ub)
1466 spin_lock(&ublk_idr_lock);
1467 idr_remove(&ublk_index_idr, ub->ub_number);
1468 wake_up_all(&ublk_idr_wq);
1469 spin_unlock(&ublk_idr_lock);
1472 static void ublk_cdev_rel(struct device *dev)
1474 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1476 blk_mq_free_tag_set(&ub->tag_set);
1477 ublk_deinit_queues(ub);
1478 ublk_free_dev_number(ub);
1479 mutex_destroy(&ub->mutex);
1483 static int ublk_add_chdev(struct ublk_device *ub)
1485 struct device *dev = &ub->cdev_dev;
1486 int minor = ub->ub_number;
1489 dev->parent = ublk_misc.this_device;
1490 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1491 dev->class = ublk_chr_class;
1492 dev->release = ublk_cdev_rel;
1493 device_initialize(dev);
1495 ret = dev_set_name(dev, "ublkc%d", minor);
1499 cdev_init(&ub->cdev, &ublk_ch_fops);
1500 ret = cdev_device_add(&ub->cdev, dev);
1511 static void ublk_stop_work_fn(struct work_struct *work)
1513 struct ublk_device *ub =
1514 container_of(work, struct ublk_device, stop_work);
1519 /* align max io buffer size with PAGE_SIZE */
1520 static void ublk_align_max_io_size(struct ublk_device *ub)
1522 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1524 ub->dev_info.max_io_buf_bytes =
1525 round_down(max_io_bytes, PAGE_SIZE);
1528 static int ublk_add_tag_set(struct ublk_device *ub)
1530 ub->tag_set.ops = &ublk_mq_ops;
1531 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1532 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1533 ub->tag_set.numa_node = NUMA_NO_NODE;
1534 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1535 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1536 ub->tag_set.driver_data = ub;
1537 return blk_mq_alloc_tag_set(&ub->tag_set);
1540 static void ublk_remove(struct ublk_device *ub)
1543 cancel_work_sync(&ub->stop_work);
1544 cancel_work_sync(&ub->quiesce_work);
1545 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1546 put_device(&ub->cdev_dev);
1550 static struct ublk_device *ublk_get_device_from_id(int idx)
1552 struct ublk_device *ub = NULL;
1557 spin_lock(&ublk_idr_lock);
1558 ub = idr_find(&ublk_index_idr, idx);
1560 ub = ublk_get_device(ub);
1561 spin_unlock(&ublk_idr_lock);
1566 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1568 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1569 int ublksrv_pid = (int)header->data[0];
1570 struct gendisk *disk;
1573 if (ublksrv_pid <= 0)
1576 wait_for_completion_interruptible(&ub->completion);
1578 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1580 mutex_lock(&ub->mutex);
1581 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1582 test_bit(UB_STATE_USED, &ub->state)) {
1587 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1589 ret = PTR_ERR(disk);
1592 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1593 disk->fops = &ub_fops;
1594 disk->private_data = ub;
1596 ub->dev_info.ublksrv_pid = ublksrv_pid;
1599 ret = ublk_apply_params(ub);
1603 /* don't probe partitions if any one ubq daemon is un-trusted */
1604 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1605 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1607 get_device(&ub->cdev_dev);
1608 ret = add_disk(disk);
1611 * Has to drop the reference since ->free_disk won't be
1612 * called in case of add_disk failure.
1614 ublk_put_device(ub);
1617 set_bit(UB_STATE_USED, &ub->state);
1618 ub->dev_info.state = UBLK_S_DEV_LIVE;
1623 mutex_unlock(&ub->mutex);
1627 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1628 struct io_uring_cmd *cmd)
1630 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1631 void __user *argp = (void __user *)(unsigned long)header->addr;
1632 cpumask_var_t cpumask;
1633 unsigned long queue;
1634 unsigned int retlen;
1638 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1640 if (header->len & (sizeof(unsigned long)-1))
1645 queue = header->data[0];
1646 if (queue >= ub->dev_info.nr_hw_queues)
1649 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1652 for_each_possible_cpu(i) {
1653 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1654 cpumask_set_cpu(i, cpumask);
1658 retlen = min_t(unsigned short, header->len, cpumask_size());
1659 if (copy_to_user(argp, cpumask, retlen))
1660 goto out_free_cpumask;
1661 if (retlen != header->len &&
1662 clear_user(argp + retlen, header->len - retlen))
1663 goto out_free_cpumask;
1667 free_cpumask_var(cpumask);
1671 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1673 pr_devel("%s: dev id %d flags %llx\n", __func__,
1674 info->dev_id, info->flags);
1675 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1676 info->nr_hw_queues, info->queue_depth);
1679 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1681 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1682 void __user *argp = (void __user *)(unsigned long)header->addr;
1683 struct ublksrv_ctrl_dev_info info;
1684 struct ublk_device *ub;
1687 if (header->len < sizeof(info) || !header->addr)
1689 if (header->queue_id != (u16)-1) {
1690 pr_warn("%s: queue_id is wrong %x\n",
1691 __func__, header->queue_id);
1695 if (copy_from_user(&info, argp, sizeof(info)))
1698 if (capable(CAP_SYS_ADMIN))
1699 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1700 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1703 /* the created device is always owned by current user */
1704 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1706 if (header->dev_id != info.dev_id) {
1707 pr_warn("%s: dev id not match %u %u\n",
1708 __func__, header->dev_id, info.dev_id);
1712 ublk_dump_dev_info(&info);
1714 ret = mutex_lock_killable(&ublk_ctl_mutex);
1719 if (ublks_added >= ublks_max)
1723 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1726 mutex_init(&ub->mutex);
1727 spin_lock_init(&ub->mm_lock);
1728 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1729 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1730 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1732 ret = ublk_alloc_dev_number(ub, header->dev_id);
1736 memcpy(&ub->dev_info, &info, sizeof(info));
1738 /* update device id */
1739 ub->dev_info.dev_id = ub->ub_number;
1742 * 64bit flags will be copied back to userspace as feature
1743 * negotiation result, so have to clear flags which driver
1744 * doesn't support yet, then userspace can get correct flags
1745 * (features) to handle.
1747 ub->dev_info.flags &= UBLK_F_ALL;
1749 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1750 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1752 /* We are not ready to support zero copy */
1753 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1755 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1756 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1757 ublk_align_max_io_size(ub);
1759 ret = ublk_init_queues(ub);
1761 goto out_free_dev_number;
1763 ret = ublk_add_tag_set(ub);
1765 goto out_deinit_queues;
1768 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1769 goto out_free_tag_set;
1772 * Add the char dev so that ublksrv daemon can be setup.
1773 * ublk_add_chdev() will cleanup everything if it fails.
1775 ret = ublk_add_chdev(ub);
1779 blk_mq_free_tag_set(&ub->tag_set);
1781 ublk_deinit_queues(ub);
1782 out_free_dev_number:
1783 ublk_free_dev_number(ub);
1785 mutex_destroy(&ub->mutex);
1788 mutex_unlock(&ublk_ctl_mutex);
1792 static inline bool ublk_idr_freed(int id)
1796 spin_lock(&ublk_idr_lock);
1797 ptr = idr_find(&ublk_index_idr, id);
1798 spin_unlock(&ublk_idr_lock);
1803 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1805 struct ublk_device *ub = *p_ub;
1806 int idx = ub->ub_number;
1809 ret = mutex_lock_killable(&ublk_ctl_mutex);
1813 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1815 set_bit(UB_STATE_DELETED, &ub->state);
1818 /* Mark the reference as consumed */
1820 ublk_put_device(ub);
1821 mutex_unlock(&ublk_ctl_mutex);
1824 * Wait until the idr is removed, then it can be reused after
1825 * DEL_DEV command is returned.
1827 * If we returns because of user interrupt, future delete command
1830 * - the device number isn't freed, this device won't or needn't
1831 * be deleted again, since UB_STATE_DELETED is set, and device
1832 * will be released after the last reference is dropped
1834 * - the device number is freed already, we will not find this
1835 * device via ublk_get_device_from_id()
1837 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1842 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1844 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1846 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1847 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1848 header->data[0], header->addr, header->len);
1851 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1854 cancel_work_sync(&ub->stop_work);
1855 cancel_work_sync(&ub->quiesce_work);
1860 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1861 struct io_uring_cmd *cmd)
1863 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1864 void __user *argp = (void __user *)(unsigned long)header->addr;
1866 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1869 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1875 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1876 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1878 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1879 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1882 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1883 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1885 ub->params.devt.disk_major = 0;
1886 ub->params.devt.disk_minor = 0;
1888 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1891 static int ublk_ctrl_get_params(struct ublk_device *ub,
1892 struct io_uring_cmd *cmd)
1894 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1895 void __user *argp = (void __user *)(unsigned long)header->addr;
1896 struct ublk_params_header ph;
1899 if (header->len <= sizeof(ph) || !header->addr)
1902 if (copy_from_user(&ph, argp, sizeof(ph)))
1905 if (ph.len > header->len || !ph.len)
1908 if (ph.len > sizeof(struct ublk_params))
1909 ph.len = sizeof(struct ublk_params);
1911 mutex_lock(&ub->mutex);
1912 ublk_ctrl_fill_params_devt(ub);
1913 if (copy_to_user(argp, &ub->params, ph.len))
1917 mutex_unlock(&ub->mutex);
1922 static int ublk_ctrl_set_params(struct ublk_device *ub,
1923 struct io_uring_cmd *cmd)
1925 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1926 void __user *argp = (void __user *)(unsigned long)header->addr;
1927 struct ublk_params_header ph;
1930 if (header->len <= sizeof(ph) || !header->addr)
1933 if (copy_from_user(&ph, argp, sizeof(ph)))
1936 if (ph.len > header->len || !ph.len || !ph.types)
1939 if (ph.len > sizeof(struct ublk_params))
1940 ph.len = sizeof(struct ublk_params);
1942 /* parameters can only be changed when device isn't live */
1943 mutex_lock(&ub->mutex);
1944 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1946 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1949 /* clear all we don't support yet */
1950 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1951 ret = ublk_validate_params(ub);
1953 mutex_unlock(&ub->mutex);
1958 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1962 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1963 /* All old ioucmds have to be completed */
1964 WARN_ON_ONCE(ubq->nr_io_ready);
1965 /* old daemon is PF_EXITING, put it now */
1966 put_task_struct(ubq->ubq_daemon);
1967 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1968 ubq->ubq_daemon = NULL;
1970 for (i = 0; i < ubq->q_depth; i++) {
1971 struct ublk_io *io = &ubq->ios[i];
1973 /* forget everything now and be ready for new FETCH_REQ */
1980 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1981 struct io_uring_cmd *cmd)
1983 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1987 mutex_lock(&ub->mutex);
1988 if (!ublk_can_use_recovery(ub))
1991 * START_RECOVERY is only allowd after:
1993 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1994 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1997 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1998 * (a)has quiesced request queue
1999 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
2000 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2001 * (d)has completed/camceled all ioucmds owned by ther dying process
2003 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2004 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2008 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2009 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2010 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2011 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2013 ub->nr_queues_ready = 0;
2014 ub->nr_privileged_daemon = 0;
2015 init_completion(&ub->completion);
2018 mutex_unlock(&ub->mutex);
2022 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2023 struct io_uring_cmd *cmd)
2025 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2026 int ublksrv_pid = (int)header->data[0];
2029 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2030 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2031 /* wait until new ubq_daemon sending all FETCH_REQ */
2032 wait_for_completion_interruptible(&ub->completion);
2033 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2034 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2036 mutex_lock(&ub->mutex);
2037 if (!ublk_can_use_recovery(ub))
2040 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2044 ub->dev_info.ublksrv_pid = ublksrv_pid;
2045 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2046 __func__, ublksrv_pid, header->dev_id);
2047 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2048 pr_devel("%s: queue unquiesced, dev id %d.\n",
2049 __func__, header->dev_id);
2050 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2051 ub->dev_info.state = UBLK_S_DEV_LIVE;
2052 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2055 mutex_unlock(&ub->mutex);
2060 * All control commands are sent via /dev/ublk-control, so we have to check
2061 * the destination device's permission
2063 static int ublk_char_dev_permission(struct ublk_device *ub,
2064 const char *dev_path, int mask)
2070 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2074 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2079 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2082 err = inode_permission(&nop_mnt_idmap,
2083 d_backing_inode(path.dentry), mask);
2089 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2090 struct io_uring_cmd *cmd)
2092 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2093 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2094 void __user *argp = (void __user *)(unsigned long)header->addr;
2095 char *dev_path = NULL;
2099 if (!unprivileged) {
2100 if (!capable(CAP_SYS_ADMIN))
2103 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2104 * char_dev_path in payload too, since userspace may not
2105 * know if the specified device is created as unprivileged
2108 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2113 * User has to provide the char device path for unprivileged ublk
2115 * header->addr always points to the dev path buffer, and
2116 * header->dev_path_len records length of dev path buffer.
2118 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2121 if (header->len < header->dev_path_len)
2124 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2129 if (copy_from_user(dev_path, argp, header->dev_path_len))
2131 dev_path[header->dev_path_len] = 0;
2134 switch (cmd->cmd_op) {
2135 case UBLK_CMD_GET_DEV_INFO:
2136 case UBLK_CMD_GET_DEV_INFO2:
2137 case UBLK_CMD_GET_QUEUE_AFFINITY:
2138 case UBLK_CMD_GET_PARAMS:
2141 case UBLK_CMD_START_DEV:
2142 case UBLK_CMD_STOP_DEV:
2143 case UBLK_CMD_ADD_DEV:
2144 case UBLK_CMD_DEL_DEV:
2145 case UBLK_CMD_SET_PARAMS:
2146 case UBLK_CMD_START_USER_RECOVERY:
2147 case UBLK_CMD_END_USER_RECOVERY:
2148 mask = MAY_READ | MAY_WRITE;
2154 ret = ublk_char_dev_permission(ub, dev_path, mask);
2156 header->len -= header->dev_path_len;
2157 header->addr += header->dev_path_len;
2159 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2160 __func__, ub->ub_number, cmd->cmd_op,
2161 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2168 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2169 unsigned int issue_flags)
2171 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2172 struct ublk_device *ub = NULL;
2175 if (issue_flags & IO_URING_F_NONBLOCK)
2178 ublk_ctrl_cmd_dump(cmd);
2180 if (!(issue_flags & IO_URING_F_SQE128))
2183 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2185 ub = ublk_get_device_from_id(header->dev_id);
2189 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2191 /* ADD_DEV permission check is done in command handler */
2198 switch (cmd->cmd_op) {
2199 case UBLK_CMD_START_DEV:
2200 ret = ublk_ctrl_start_dev(ub, cmd);
2202 case UBLK_CMD_STOP_DEV:
2203 ret = ublk_ctrl_stop_dev(ub);
2205 case UBLK_CMD_GET_DEV_INFO:
2206 case UBLK_CMD_GET_DEV_INFO2:
2207 ret = ublk_ctrl_get_dev_info(ub, cmd);
2209 case UBLK_CMD_ADD_DEV:
2210 ret = ublk_ctrl_add_dev(cmd);
2212 case UBLK_CMD_DEL_DEV:
2213 ret = ublk_ctrl_del_dev(&ub);
2215 case UBLK_CMD_GET_QUEUE_AFFINITY:
2216 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2218 case UBLK_CMD_GET_PARAMS:
2219 ret = ublk_ctrl_get_params(ub, cmd);
2221 case UBLK_CMD_SET_PARAMS:
2222 ret = ublk_ctrl_set_params(ub, cmd);
2224 case UBLK_CMD_START_USER_RECOVERY:
2225 ret = ublk_ctrl_start_recovery(ub, cmd);
2227 case UBLK_CMD_END_USER_RECOVERY:
2228 ret = ublk_ctrl_end_recovery(ub, cmd);
2237 ublk_put_device(ub);
2239 io_uring_cmd_done(cmd, ret, 0);
2240 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2241 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2242 return -EIOCBQUEUED;
2245 static const struct file_operations ublk_ctl_fops = {
2246 .open = nonseekable_open,
2247 .uring_cmd = ublk_ctrl_uring_cmd,
2248 .owner = THIS_MODULE,
2249 .llseek = noop_llseek,
2252 static struct miscdevice ublk_misc = {
2253 .minor = MISC_DYNAMIC_MINOR,
2254 .name = "ublk-control",
2255 .fops = &ublk_ctl_fops,
2258 static int __init ublk_init(void)
2262 init_waitqueue_head(&ublk_idr_wq);
2264 ret = misc_register(&ublk_misc);
2268 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2270 goto unregister_mis;
2272 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2273 if (IS_ERR(ublk_chr_class)) {
2274 ret = PTR_ERR(ublk_chr_class);
2275 goto free_chrdev_region;
2280 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2282 misc_deregister(&ublk_misc);
2286 static void __exit ublk_exit(void)
2288 struct ublk_device *ub;
2291 idr_for_each_entry(&ublk_index_idr, ub, id)
2294 class_destroy(ublk_chr_class);
2295 misc_deregister(&ublk_misc);
2297 idr_destroy(&ublk_index_idr);
2298 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2301 module_init(ublk_init);
2302 module_exit(ublk_exit);
2304 module_param(ublks_max, int, 0444);
2305 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2307 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2308 MODULE_LICENSE("GPL");