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;
659 /* failed read IO if nothing is read */
660 if (!io->res && req_op(req) == REQ_OP_READ)
664 blk_mq_end_request(req, errno_to_blk_status(io->res));
669 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
672 * Both the two needn't unmap.
674 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
675 blk_mq_end_request(req, BLK_STS_OK);
679 /* for READ request, writing data in iod->addr to rq buffers */
680 unmapped_bytes = ublk_unmap_io(ubq, req, io);
683 * Extremely impossible since we got data filled in just before
685 * Re-read simply for this unlikely case.
687 if (unlikely(unmapped_bytes < io->res))
688 io->res = unmapped_bytes;
690 if (blk_update_request(req, BLK_STS_OK, io->res))
691 blk_mq_requeue_request(req, true);
693 __blk_mq_end_request(req, BLK_STS_OK);
697 * Since __ublk_rq_task_work always fails requests immediately during
698 * exiting, __ublk_fail_req() is only called from abort context during
699 * exiting. So lock is unnecessary.
701 * Also aborting may not be started yet, keep in mind that one failed
702 * request may be issued by block layer again.
704 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
707 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
709 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
710 io->flags |= UBLK_IO_FLAG_ABORTED;
711 if (ublk_queue_can_use_recovery_reissue(ubq))
712 blk_mq_requeue_request(req, false);
714 blk_mq_end_request(req, BLK_STS_IOERR);
718 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
719 unsigned issue_flags)
721 /* mark this cmd owned by ublksrv */
722 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
725 * clear ACTIVE since we are done with this sqe/cmd slot
726 * We can only accept io cmd in case of being not active.
728 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
730 /* tell ublksrv one io request is coming */
731 io_uring_cmd_done(io->cmd, res, 0, issue_flags);
734 #define UBLK_REQUEUE_DELAY_MS 3
736 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
739 /* We cannot process this rq so just requeue it. */
740 if (ublk_queue_can_use_recovery(ubq))
741 blk_mq_requeue_request(rq, false);
743 blk_mq_end_request(rq, BLK_STS_IOERR);
745 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
748 static inline void __ublk_rq_task_work(struct request *req,
749 unsigned issue_flags)
751 struct ublk_queue *ubq = req->mq_hctx->driver_data;
753 struct ublk_io *io = &ubq->ios[tag];
754 unsigned int mapped_bytes;
756 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
757 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
758 ublk_get_iod(ubq, req->tag)->addr);
761 * Task is exiting if either:
763 * (1) current != ubq_daemon.
764 * io_uring_cmd_complete_in_task() tries to run task_work
765 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
767 * (2) current->flags & PF_EXITING.
769 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
770 __ublk_abort_rq(ubq, req);
774 if (ublk_need_get_data(ubq) &&
775 (req_op(req) == REQ_OP_WRITE ||
776 req_op(req) == REQ_OP_FLUSH)) {
778 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
779 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
782 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
783 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
784 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
785 __func__, io->cmd->cmd_op, ubq->q_id,
786 req->tag, io->flags);
787 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
791 * We have handled UBLK_IO_NEED_GET_DATA command,
792 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
795 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
796 /* update iod->addr because ublksrv may have passed a new io buffer */
797 ublk_get_iod(ubq, req->tag)->addr = io->addr;
798 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
799 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
800 ublk_get_iod(ubq, req->tag)->addr);
803 mapped_bytes = ublk_map_io(ubq, req, io);
805 /* partially mapped, update io descriptor */
806 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
808 * Nothing mapped, retry until we succeed.
810 * We may never succeed in mapping any bytes here because
811 * of OOM. TODO: reserve one buffer with single page pinned
812 * for providing forward progress guarantee.
814 if (unlikely(!mapped_bytes)) {
815 blk_mq_requeue_request(req, false);
816 blk_mq_delay_kick_requeue_list(req->q,
817 UBLK_REQUEUE_DELAY_MS);
821 ublk_get_iod(ubq, req->tag)->nr_sectors =
825 ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
828 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
829 unsigned issue_flags)
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), issue_flags);
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, unsigned issue_flags)
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, issue_flags);
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;
862 unsigned issue_flags = IO_URING_F_UNLOCKED;
864 ublk_forward_io_cmds(ubq, issue_flags);
867 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
869 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
872 if (!llist_add(&data->node, &ubq->io_cmds))
875 io = &ubq->ios[rq->tag];
877 * If the check pass, we know that this is a re-issued request aborted
878 * previously in monitor_work because the ubq_daemon(cmd's task) is
879 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
880 * because this ioucmd's io_uring context may be freed now if no inflight
881 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
883 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
884 * the tag). Then the request is re-started(allocating the tag) and we are here.
885 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
886 * guarantees that here is a re-issued request aborted previously.
888 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
889 ublk_abort_io_cmds(ubq);
890 } else if (ublk_can_use_task_work(ubq)) {
891 if (task_work_add(ubq->ubq_daemon, &data->work,
893 ublk_abort_io_cmds(ubq);
895 struct io_uring_cmd *cmd = io->cmd;
896 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
899 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
903 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
904 const struct blk_mq_queue_data *bd)
906 struct ublk_queue *ubq = hctx->driver_data;
907 struct request *rq = bd->rq;
910 /* fill iod to slot in io cmd buffer */
911 res = ublk_setup_iod(ubq, rq);
912 if (unlikely(res != BLK_STS_OK))
913 return BLK_STS_IOERR;
915 /* With recovery feature enabled, force_abort is set in
916 * ublk_stop_dev() before calling del_gendisk(). We have to
917 * abort all requeued and new rqs here to let del_gendisk()
918 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
919 * to avoid UAF on io_uring ctx.
921 * Note: force_abort is guaranteed to be seen because it is set
922 * before request queue is unqiuesced.
924 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
925 return BLK_STS_IOERR;
927 blk_mq_start_request(bd->rq);
929 if (unlikely(ubq_daemon_is_dying(ubq))) {
930 __ublk_abort_rq(ubq, rq);
934 ublk_queue_cmd(ubq, rq);
939 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
940 unsigned int hctx_idx)
942 struct ublk_device *ub = driver_data;
943 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
945 hctx->driver_data = ubq;
949 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
950 unsigned int hctx_idx, unsigned int numa_node)
952 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
954 init_task_work(&data->work, ublk_rq_task_work_fn);
958 static const struct blk_mq_ops ublk_mq_ops = {
959 .queue_rq = ublk_queue_rq,
960 .init_hctx = ublk_init_hctx,
961 .init_request = ublk_init_rq,
964 static int ublk_ch_open(struct inode *inode, struct file *filp)
966 struct ublk_device *ub = container_of(inode->i_cdev,
967 struct ublk_device, cdev);
969 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
971 filp->private_data = ub;
975 static int ublk_ch_release(struct inode *inode, struct file *filp)
977 struct ublk_device *ub = filp->private_data;
979 clear_bit(UB_STATE_OPEN, &ub->state);
983 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
984 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
986 struct ublk_device *ub = filp->private_data;
987 size_t sz = vma->vm_end - vma->vm_start;
988 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
989 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
992 spin_lock(&ub->mm_lock);
994 ub->mm = current->mm;
995 if (current->mm != ub->mm)
997 spin_unlock(&ub->mm_lock);
1002 if (vma->vm_flags & VM_WRITE)
1005 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1006 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1009 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1010 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1011 __func__, q_id, current->pid, vma->vm_start,
1012 phys_off, (unsigned long)sz);
1014 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1017 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1018 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1021 static void ublk_commit_completion(struct ublk_device *ub,
1022 struct ublksrv_io_cmd *ub_cmd)
1024 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1025 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1026 struct ublk_io *io = &ubq->ios[tag];
1027 struct request *req;
1029 /* now this cmd slot is owned by nbd driver */
1030 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1031 io->res = ub_cmd->result;
1033 /* find the io request and complete */
1034 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1036 if (req && likely(!blk_should_fake_timeout(req->q)))
1037 ublk_complete_rq(req);
1041 * When ->ubq_daemon is exiting, either new request is ended immediately,
1042 * or any queued io command is drained, so it is safe to abort queue
1045 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1049 if (!ublk_get_device(ub))
1052 for (i = 0; i < ubq->q_depth; i++) {
1053 struct ublk_io *io = &ubq->ios[i];
1055 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1059 * Either we fail the request or ublk_rq_task_work_fn
1062 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1064 __ublk_fail_req(ubq, io, rq);
1067 ublk_put_device(ub);
1070 static void ublk_daemon_monitor_work(struct work_struct *work)
1072 struct ublk_device *ub =
1073 container_of(work, struct ublk_device, monitor_work.work);
1076 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1077 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1079 if (ubq_daemon_is_dying(ubq)) {
1080 if (ublk_queue_can_use_recovery(ubq))
1081 schedule_work(&ub->quiesce_work);
1083 schedule_work(&ub->stop_work);
1085 /* abort queue is for making forward progress */
1086 ublk_abort_queue(ub, ubq);
1091 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1092 * after ublk_remove() or __ublk_quiesce_dev() is started.
1094 * No need ub->mutex, monitor work are canceled after state is marked
1095 * as not LIVE, so new state is observed reliably.
1097 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1098 schedule_delayed_work(&ub->monitor_work,
1099 UBLK_DAEMON_MONITOR_PERIOD);
1102 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1104 return ubq->nr_io_ready == ubq->q_depth;
1107 static void ublk_cancel_queue(struct ublk_queue *ubq)
1111 if (!ublk_queue_ready(ubq))
1114 for (i = 0; i < ubq->q_depth; i++) {
1115 struct ublk_io *io = &ubq->ios[i];
1117 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1118 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1119 IO_URING_F_UNLOCKED);
1122 /* all io commands are canceled */
1123 ubq->nr_io_ready = 0;
1126 /* Cancel all pending commands, must be called after del_gendisk() returns */
1127 static void ublk_cancel_dev(struct ublk_device *ub)
1131 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1132 ublk_cancel_queue(ublk_get_queue(ub, i));
1135 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1139 if (blk_mq_request_started(rq)) {
1146 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1150 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1153 blk_mq_tagset_busy_iter(&ub->tag_set,
1154 ublk_check_inflight_rq, &idle);
1157 msleep(UBLK_REQUEUE_DELAY_MS);
1161 static void __ublk_quiesce_dev(struct ublk_device *ub)
1163 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1164 __func__, ub->dev_info.dev_id,
1165 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1166 "LIVE" : "QUIESCED");
1167 blk_mq_quiesce_queue(ub->ub_disk->queue);
1168 ublk_wait_tagset_rqs_idle(ub);
1169 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1170 ublk_cancel_dev(ub);
1171 /* we are going to release task_struct of ubq_daemon and resets
1172 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1173 * Besides, monitor_work is not necessary in QUIESCED state since we have
1174 * already scheduled quiesce_work and quiesced all ubqs.
1176 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1177 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1179 cancel_delayed_work_sync(&ub->monitor_work);
1182 static void ublk_quiesce_work_fn(struct work_struct *work)
1184 struct ublk_device *ub =
1185 container_of(work, struct ublk_device, quiesce_work);
1187 mutex_lock(&ub->mutex);
1188 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1190 __ublk_quiesce_dev(ub);
1192 mutex_unlock(&ub->mutex);
1195 static void ublk_unquiesce_dev(struct ublk_device *ub)
1199 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1200 __func__, ub->dev_info.dev_id,
1201 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1202 "LIVE" : "QUIESCED");
1203 /* quiesce_work has run. We let requeued rqs be aborted
1204 * before running fallback_wq. "force_abort" must be seen
1205 * after request queue is unqiuesced. Then del_gendisk()
1208 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1209 ublk_get_queue(ub, i)->force_abort = true;
1211 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1212 /* We may have requeued some rqs in ublk_quiesce_queue() */
1213 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1216 static void ublk_stop_dev(struct ublk_device *ub)
1218 mutex_lock(&ub->mutex);
1219 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1221 if (ublk_can_use_recovery(ub)) {
1222 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1223 __ublk_quiesce_dev(ub);
1224 ublk_unquiesce_dev(ub);
1226 del_gendisk(ub->ub_disk);
1227 ub->dev_info.state = UBLK_S_DEV_DEAD;
1228 ub->dev_info.ublksrv_pid = -1;
1229 put_disk(ub->ub_disk);
1232 ublk_cancel_dev(ub);
1233 mutex_unlock(&ub->mutex);
1234 cancel_delayed_work_sync(&ub->monitor_work);
1237 /* device can only be started after all IOs are ready */
1238 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1240 mutex_lock(&ub->mutex);
1242 if (ublk_queue_ready(ubq)) {
1243 ubq->ubq_daemon = current;
1244 get_task_struct(ubq->ubq_daemon);
1245 ub->nr_queues_ready++;
1247 if (capable(CAP_SYS_ADMIN))
1248 ub->nr_privileged_daemon++;
1250 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1251 complete_all(&ub->completion);
1252 mutex_unlock(&ub->mutex);
1255 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1258 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1259 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1261 ublk_queue_cmd(ubq, req);
1264 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1265 unsigned int issue_flags,
1266 struct ublksrv_io_cmd *ub_cmd)
1268 struct ublk_device *ub = cmd->file->private_data;
1269 struct ublk_queue *ubq;
1271 u32 cmd_op = cmd->cmd_op;
1272 unsigned tag = ub_cmd->tag;
1274 struct request *req;
1276 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1277 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1280 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1283 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1284 if (!ubq || ub_cmd->q_id != ubq->q_id)
1287 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1290 if (tag >= ubq->q_depth)
1293 io = &ubq->ios[tag];
1295 /* there is pending io cmd, something must be wrong */
1296 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1302 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1303 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1305 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1306 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1310 case UBLK_IO_FETCH_REQ:
1311 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1312 if (ublk_queue_ready(ubq)) {
1317 * The io is being handled by server, so COMMIT_RQ is expected
1318 * instead of FETCH_REQ
1320 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1322 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1323 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1326 io->flags |= UBLK_IO_FLAG_ACTIVE;
1327 io->addr = ub_cmd->addr;
1329 ublk_mark_io_ready(ub, ubq);
1331 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1332 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1334 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1335 * not enabled or it is Read IO.
1337 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1339 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1341 io->addr = ub_cmd->addr;
1342 io->flags |= UBLK_IO_FLAG_ACTIVE;
1344 ublk_commit_completion(ub, ub_cmd);
1346 case UBLK_IO_NEED_GET_DATA:
1347 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1349 io->addr = ub_cmd->addr;
1351 io->flags |= UBLK_IO_FLAG_ACTIVE;
1352 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1357 return -EIOCBQUEUED;
1360 io_uring_cmd_done(cmd, ret, 0, issue_flags);
1361 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1362 __func__, cmd_op, tag, ret, io->flags);
1363 return -EIOCBQUEUED;
1366 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1368 struct ublksrv_io_cmd *ub_src = (struct ublksrv_io_cmd *) cmd->cmd;
1369 struct ublksrv_io_cmd ub_cmd;
1372 * Not necessary for async retry, but let's keep it simple and always
1373 * copy the values to avoid any potential reuse.
1375 ub_cmd.q_id = READ_ONCE(ub_src->q_id);
1376 ub_cmd.tag = READ_ONCE(ub_src->tag);
1377 ub_cmd.result = READ_ONCE(ub_src->result);
1378 ub_cmd.addr = READ_ONCE(ub_src->addr);
1380 return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1383 static const struct file_operations ublk_ch_fops = {
1384 .owner = THIS_MODULE,
1385 .open = ublk_ch_open,
1386 .release = ublk_ch_release,
1387 .llseek = no_llseek,
1388 .uring_cmd = ublk_ch_uring_cmd,
1389 .mmap = ublk_ch_mmap,
1392 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1394 int size = ublk_queue_cmd_buf_size(ub, q_id);
1395 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1397 if (ubq->ubq_daemon)
1398 put_task_struct(ubq->ubq_daemon);
1399 if (ubq->io_cmd_buf)
1400 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1403 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1405 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1406 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1410 ubq->flags = ub->dev_info.flags;
1412 ubq->q_depth = ub->dev_info.queue_depth;
1413 size = ublk_queue_cmd_buf_size(ub, q_id);
1415 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1419 ubq->io_cmd_buf = ptr;
1424 static void ublk_deinit_queues(struct ublk_device *ub)
1426 int nr_queues = ub->dev_info.nr_hw_queues;
1432 for (i = 0; i < nr_queues; i++)
1433 ublk_deinit_queue(ub, i);
1434 kfree(ub->__queues);
1437 static int ublk_init_queues(struct ublk_device *ub)
1439 int nr_queues = ub->dev_info.nr_hw_queues;
1440 int depth = ub->dev_info.queue_depth;
1441 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1442 int i, ret = -ENOMEM;
1444 ub->queue_size = ubq_size;
1445 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1449 for (i = 0; i < nr_queues; i++) {
1450 if (ublk_init_queue(ub, i))
1454 init_completion(&ub->completion);
1458 ublk_deinit_queues(ub);
1462 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1467 spin_lock(&ublk_idr_lock);
1468 /* allocate id, if @id >= 0, we're requesting that specific id */
1470 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1474 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1476 spin_unlock(&ublk_idr_lock);
1479 ub->ub_number = err;
1484 static void ublk_free_dev_number(struct ublk_device *ub)
1486 spin_lock(&ublk_idr_lock);
1487 idr_remove(&ublk_index_idr, ub->ub_number);
1488 wake_up_all(&ublk_idr_wq);
1489 spin_unlock(&ublk_idr_lock);
1492 static void ublk_cdev_rel(struct device *dev)
1494 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1496 blk_mq_free_tag_set(&ub->tag_set);
1497 ublk_deinit_queues(ub);
1498 ublk_free_dev_number(ub);
1499 mutex_destroy(&ub->mutex);
1503 static int ublk_add_chdev(struct ublk_device *ub)
1505 struct device *dev = &ub->cdev_dev;
1506 int minor = ub->ub_number;
1509 dev->parent = ublk_misc.this_device;
1510 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1511 dev->class = ublk_chr_class;
1512 dev->release = ublk_cdev_rel;
1513 device_initialize(dev);
1515 ret = dev_set_name(dev, "ublkc%d", minor);
1519 cdev_init(&ub->cdev, &ublk_ch_fops);
1520 ret = cdev_device_add(&ub->cdev, dev);
1531 static void ublk_stop_work_fn(struct work_struct *work)
1533 struct ublk_device *ub =
1534 container_of(work, struct ublk_device, stop_work);
1539 /* align max io buffer size with PAGE_SIZE */
1540 static void ublk_align_max_io_size(struct ublk_device *ub)
1542 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1544 ub->dev_info.max_io_buf_bytes =
1545 round_down(max_io_bytes, PAGE_SIZE);
1548 static int ublk_add_tag_set(struct ublk_device *ub)
1550 ub->tag_set.ops = &ublk_mq_ops;
1551 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1552 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1553 ub->tag_set.numa_node = NUMA_NO_NODE;
1554 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1555 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1556 ub->tag_set.driver_data = ub;
1557 return blk_mq_alloc_tag_set(&ub->tag_set);
1560 static void ublk_remove(struct ublk_device *ub)
1563 cancel_work_sync(&ub->stop_work);
1564 cancel_work_sync(&ub->quiesce_work);
1565 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1566 put_device(&ub->cdev_dev);
1570 static struct ublk_device *ublk_get_device_from_id(int idx)
1572 struct ublk_device *ub = NULL;
1577 spin_lock(&ublk_idr_lock);
1578 ub = idr_find(&ublk_index_idr, idx);
1580 ub = ublk_get_device(ub);
1581 spin_unlock(&ublk_idr_lock);
1586 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1588 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1589 int ublksrv_pid = (int)header->data[0];
1590 struct gendisk *disk;
1593 if (ublksrv_pid <= 0)
1596 wait_for_completion_interruptible(&ub->completion);
1598 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1600 mutex_lock(&ub->mutex);
1601 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1602 test_bit(UB_STATE_USED, &ub->state)) {
1607 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1609 ret = PTR_ERR(disk);
1612 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1613 disk->fops = &ub_fops;
1614 disk->private_data = ub;
1616 ub->dev_info.ublksrv_pid = ublksrv_pid;
1619 ret = ublk_apply_params(ub);
1623 /* don't probe partitions if any one ubq daemon is un-trusted */
1624 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1625 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1627 get_device(&ub->cdev_dev);
1628 ub->dev_info.state = UBLK_S_DEV_LIVE;
1629 ret = add_disk(disk);
1632 * Has to drop the reference since ->free_disk won't be
1633 * called in case of add_disk failure.
1635 ub->dev_info.state = UBLK_S_DEV_DEAD;
1636 ublk_put_device(ub);
1639 set_bit(UB_STATE_USED, &ub->state);
1644 mutex_unlock(&ub->mutex);
1648 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1649 struct io_uring_cmd *cmd)
1651 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1652 void __user *argp = (void __user *)(unsigned long)header->addr;
1653 cpumask_var_t cpumask;
1654 unsigned long queue;
1655 unsigned int retlen;
1659 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1661 if (header->len & (sizeof(unsigned long)-1))
1666 queue = header->data[0];
1667 if (queue >= ub->dev_info.nr_hw_queues)
1670 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1673 for_each_possible_cpu(i) {
1674 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1675 cpumask_set_cpu(i, cpumask);
1679 retlen = min_t(unsigned short, header->len, cpumask_size());
1680 if (copy_to_user(argp, cpumask, retlen))
1681 goto out_free_cpumask;
1682 if (retlen != header->len &&
1683 clear_user(argp + retlen, header->len - retlen))
1684 goto out_free_cpumask;
1688 free_cpumask_var(cpumask);
1692 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1694 pr_devel("%s: dev id %d flags %llx\n", __func__,
1695 info->dev_id, info->flags);
1696 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1697 info->nr_hw_queues, info->queue_depth);
1700 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1702 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1703 void __user *argp = (void __user *)(unsigned long)header->addr;
1704 struct ublksrv_ctrl_dev_info info;
1705 struct ublk_device *ub;
1708 if (header->len < sizeof(info) || !header->addr)
1710 if (header->queue_id != (u16)-1) {
1711 pr_warn("%s: queue_id is wrong %x\n",
1712 __func__, header->queue_id);
1716 if (copy_from_user(&info, argp, sizeof(info)))
1719 if (capable(CAP_SYS_ADMIN))
1720 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1721 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1724 /* the created device is always owned by current user */
1725 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1727 if (header->dev_id != info.dev_id) {
1728 pr_warn("%s: dev id not match %u %u\n",
1729 __func__, header->dev_id, info.dev_id);
1733 ublk_dump_dev_info(&info);
1735 ret = mutex_lock_killable(&ublk_ctl_mutex);
1740 if (ublks_added >= ublks_max)
1744 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1747 mutex_init(&ub->mutex);
1748 spin_lock_init(&ub->mm_lock);
1749 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1750 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1751 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1753 ret = ublk_alloc_dev_number(ub, header->dev_id);
1757 memcpy(&ub->dev_info, &info, sizeof(info));
1759 /* update device id */
1760 ub->dev_info.dev_id = ub->ub_number;
1763 * 64bit flags will be copied back to userspace as feature
1764 * negotiation result, so have to clear flags which driver
1765 * doesn't support yet, then userspace can get correct flags
1766 * (features) to handle.
1768 ub->dev_info.flags &= UBLK_F_ALL;
1770 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1771 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1773 /* We are not ready to support zero copy */
1774 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1776 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1777 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1778 ublk_align_max_io_size(ub);
1780 ret = ublk_init_queues(ub);
1782 goto out_free_dev_number;
1784 ret = ublk_add_tag_set(ub);
1786 goto out_deinit_queues;
1789 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1790 goto out_free_tag_set;
1793 * Add the char dev so that ublksrv daemon can be setup.
1794 * ublk_add_chdev() will cleanup everything if it fails.
1796 ret = ublk_add_chdev(ub);
1800 blk_mq_free_tag_set(&ub->tag_set);
1802 ublk_deinit_queues(ub);
1803 out_free_dev_number:
1804 ublk_free_dev_number(ub);
1806 mutex_destroy(&ub->mutex);
1809 mutex_unlock(&ublk_ctl_mutex);
1813 static inline bool ublk_idr_freed(int id)
1817 spin_lock(&ublk_idr_lock);
1818 ptr = idr_find(&ublk_index_idr, id);
1819 spin_unlock(&ublk_idr_lock);
1824 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1826 struct ublk_device *ub = *p_ub;
1827 int idx = ub->ub_number;
1830 ret = mutex_lock_killable(&ublk_ctl_mutex);
1834 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1836 set_bit(UB_STATE_DELETED, &ub->state);
1839 /* Mark the reference as consumed */
1841 ublk_put_device(ub);
1842 mutex_unlock(&ublk_ctl_mutex);
1845 * Wait until the idr is removed, then it can be reused after
1846 * DEL_DEV command is returned.
1848 * If we returns because of user interrupt, future delete command
1851 * - the device number isn't freed, this device won't or needn't
1852 * be deleted again, since UB_STATE_DELETED is set, and device
1853 * will be released after the last reference is dropped
1855 * - the device number is freed already, we will not find this
1856 * device via ublk_get_device_from_id()
1858 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1863 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1865 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1867 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1868 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1869 header->data[0], header->addr, header->len);
1872 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1875 cancel_work_sync(&ub->stop_work);
1876 cancel_work_sync(&ub->quiesce_work);
1881 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1882 struct io_uring_cmd *cmd)
1884 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1885 void __user *argp = (void __user *)(unsigned long)header->addr;
1887 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1890 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1896 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1897 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1899 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1900 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1903 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1904 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1906 ub->params.devt.disk_major = 0;
1907 ub->params.devt.disk_minor = 0;
1909 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1912 static int ublk_ctrl_get_params(struct ublk_device *ub,
1913 struct io_uring_cmd *cmd)
1915 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1916 void __user *argp = (void __user *)(unsigned long)header->addr;
1917 struct ublk_params_header ph;
1920 if (header->len <= sizeof(ph) || !header->addr)
1923 if (copy_from_user(&ph, argp, sizeof(ph)))
1926 if (ph.len > header->len || !ph.len)
1929 if (ph.len > sizeof(struct ublk_params))
1930 ph.len = sizeof(struct ublk_params);
1932 mutex_lock(&ub->mutex);
1933 ublk_ctrl_fill_params_devt(ub);
1934 if (copy_to_user(argp, &ub->params, ph.len))
1938 mutex_unlock(&ub->mutex);
1943 static int ublk_ctrl_set_params(struct ublk_device *ub,
1944 struct io_uring_cmd *cmd)
1946 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1947 void __user *argp = (void __user *)(unsigned long)header->addr;
1948 struct ublk_params_header ph;
1951 if (header->len <= sizeof(ph) || !header->addr)
1954 if (copy_from_user(&ph, argp, sizeof(ph)))
1957 if (ph.len > header->len || !ph.len || !ph.types)
1960 if (ph.len > sizeof(struct ublk_params))
1961 ph.len = sizeof(struct ublk_params);
1963 /* parameters can only be changed when device isn't live */
1964 mutex_lock(&ub->mutex);
1965 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1967 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1970 /* clear all we don't support yet */
1971 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1972 ret = ublk_validate_params(ub);
1974 mutex_unlock(&ub->mutex);
1979 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1983 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1984 /* All old ioucmds have to be completed */
1985 WARN_ON_ONCE(ubq->nr_io_ready);
1986 /* old daemon is PF_EXITING, put it now */
1987 put_task_struct(ubq->ubq_daemon);
1988 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1989 ubq->ubq_daemon = NULL;
1991 for (i = 0; i < ubq->q_depth; i++) {
1992 struct ublk_io *io = &ubq->ios[i];
1994 /* forget everything now and be ready for new FETCH_REQ */
2001 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2002 struct io_uring_cmd *cmd)
2004 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2008 mutex_lock(&ub->mutex);
2009 if (!ublk_can_use_recovery(ub))
2012 * START_RECOVERY is only allowd after:
2014 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2015 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
2018 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2019 * (a)has quiesced request queue
2020 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
2021 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2022 * (d)has completed/camceled all ioucmds owned by ther dying process
2024 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2025 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2029 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2030 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2031 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2032 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2034 ub->nr_queues_ready = 0;
2035 ub->nr_privileged_daemon = 0;
2036 init_completion(&ub->completion);
2039 mutex_unlock(&ub->mutex);
2043 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2044 struct io_uring_cmd *cmd)
2046 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2047 int ublksrv_pid = (int)header->data[0];
2050 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2051 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2052 /* wait until new ubq_daemon sending all FETCH_REQ */
2053 wait_for_completion_interruptible(&ub->completion);
2054 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2055 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2057 mutex_lock(&ub->mutex);
2058 if (!ublk_can_use_recovery(ub))
2061 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2065 ub->dev_info.ublksrv_pid = ublksrv_pid;
2066 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2067 __func__, ublksrv_pid, header->dev_id);
2068 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2069 pr_devel("%s: queue unquiesced, dev id %d.\n",
2070 __func__, header->dev_id);
2071 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2072 ub->dev_info.state = UBLK_S_DEV_LIVE;
2073 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2076 mutex_unlock(&ub->mutex);
2081 * All control commands are sent via /dev/ublk-control, so we have to check
2082 * the destination device's permission
2084 static int ublk_char_dev_permission(struct ublk_device *ub,
2085 const char *dev_path, int mask)
2091 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2095 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2100 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2103 err = inode_permission(&nop_mnt_idmap,
2104 d_backing_inode(path.dentry), mask);
2110 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2111 struct io_uring_cmd *cmd)
2113 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2114 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2115 void __user *argp = (void __user *)(unsigned long)header->addr;
2116 char *dev_path = NULL;
2120 if (!unprivileged) {
2121 if (!capable(CAP_SYS_ADMIN))
2124 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2125 * char_dev_path in payload too, since userspace may not
2126 * know if the specified device is created as unprivileged
2129 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2134 * User has to provide the char device path for unprivileged ublk
2136 * header->addr always points to the dev path buffer, and
2137 * header->dev_path_len records length of dev path buffer.
2139 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2142 if (header->len < header->dev_path_len)
2145 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2150 if (copy_from_user(dev_path, argp, header->dev_path_len))
2152 dev_path[header->dev_path_len] = 0;
2155 switch (cmd->cmd_op) {
2156 case UBLK_CMD_GET_DEV_INFO:
2157 case UBLK_CMD_GET_DEV_INFO2:
2158 case UBLK_CMD_GET_QUEUE_AFFINITY:
2159 case UBLK_CMD_GET_PARAMS:
2162 case UBLK_CMD_START_DEV:
2163 case UBLK_CMD_STOP_DEV:
2164 case UBLK_CMD_ADD_DEV:
2165 case UBLK_CMD_DEL_DEV:
2166 case UBLK_CMD_SET_PARAMS:
2167 case UBLK_CMD_START_USER_RECOVERY:
2168 case UBLK_CMD_END_USER_RECOVERY:
2169 mask = MAY_READ | MAY_WRITE;
2175 ret = ublk_char_dev_permission(ub, dev_path, mask);
2177 header->len -= header->dev_path_len;
2178 header->addr += header->dev_path_len;
2180 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2181 __func__, ub->ub_number, cmd->cmd_op,
2182 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2189 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2190 unsigned int issue_flags)
2192 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2193 struct ublk_device *ub = NULL;
2196 if (issue_flags & IO_URING_F_NONBLOCK)
2199 ublk_ctrl_cmd_dump(cmd);
2201 if (!(issue_flags & IO_URING_F_SQE128))
2204 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2206 ub = ublk_get_device_from_id(header->dev_id);
2210 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2212 /* ADD_DEV permission check is done in command handler */
2219 switch (cmd->cmd_op) {
2220 case UBLK_CMD_START_DEV:
2221 ret = ublk_ctrl_start_dev(ub, cmd);
2223 case UBLK_CMD_STOP_DEV:
2224 ret = ublk_ctrl_stop_dev(ub);
2226 case UBLK_CMD_GET_DEV_INFO:
2227 case UBLK_CMD_GET_DEV_INFO2:
2228 ret = ublk_ctrl_get_dev_info(ub, cmd);
2230 case UBLK_CMD_ADD_DEV:
2231 ret = ublk_ctrl_add_dev(cmd);
2233 case UBLK_CMD_DEL_DEV:
2234 ret = ublk_ctrl_del_dev(&ub);
2236 case UBLK_CMD_GET_QUEUE_AFFINITY:
2237 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2239 case UBLK_CMD_GET_PARAMS:
2240 ret = ublk_ctrl_get_params(ub, cmd);
2242 case UBLK_CMD_SET_PARAMS:
2243 ret = ublk_ctrl_set_params(ub, cmd);
2245 case UBLK_CMD_START_USER_RECOVERY:
2246 ret = ublk_ctrl_start_recovery(ub, cmd);
2248 case UBLK_CMD_END_USER_RECOVERY:
2249 ret = ublk_ctrl_end_recovery(ub, cmd);
2258 ublk_put_device(ub);
2260 io_uring_cmd_done(cmd, ret, 0, issue_flags);
2261 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2262 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2263 return -EIOCBQUEUED;
2266 static const struct file_operations ublk_ctl_fops = {
2267 .open = nonseekable_open,
2268 .uring_cmd = ublk_ctrl_uring_cmd,
2269 .owner = THIS_MODULE,
2270 .llseek = noop_llseek,
2273 static struct miscdevice ublk_misc = {
2274 .minor = MISC_DYNAMIC_MINOR,
2275 .name = "ublk-control",
2276 .fops = &ublk_ctl_fops,
2279 static int __init ublk_init(void)
2283 init_waitqueue_head(&ublk_idr_wq);
2285 ret = misc_register(&ublk_misc);
2289 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2291 goto unregister_mis;
2293 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2294 if (IS_ERR(ublk_chr_class)) {
2295 ret = PTR_ERR(ublk_chr_class);
2296 goto free_chrdev_region;
2301 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2303 misc_deregister(&ublk_misc);
2307 static void __exit ublk_exit(void)
2309 struct ublk_device *ub;
2312 idr_for_each_entry(&ublk_index_idr, ub, id)
2315 class_destroy(ublk_chr_class);
2316 misc_deregister(&ublk_misc);
2318 idr_destroy(&ublk_index_idr);
2319 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2322 module_init(ublk_init);
2323 module_exit(ublk_exit);
2325 module_param(ublks_max, int, 0444);
2326 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2328 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2329 MODULE_LICENSE("GPL");