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)
720 /* mark this cmd owned by ublksrv */
721 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
724 * clear ACTIVE since we are done with this sqe/cmd slot
725 * We can only accept io cmd in case of being not active.
727 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
729 /* tell ublksrv one io request is coming */
730 io_uring_cmd_done(io->cmd, res, 0);
733 #define UBLK_REQUEUE_DELAY_MS 3
735 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
738 /* We cannot process this rq so just requeue it. */
739 if (ublk_queue_can_use_recovery(ubq))
740 blk_mq_requeue_request(rq, false);
742 blk_mq_end_request(rq, BLK_STS_IOERR);
744 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
747 static inline void __ublk_rq_task_work(struct request *req)
749 struct ublk_queue *ubq = req->mq_hctx->driver_data;
751 struct ublk_io *io = &ubq->ios[tag];
752 unsigned int mapped_bytes;
754 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
755 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
756 ublk_get_iod(ubq, req->tag)->addr);
759 * Task is exiting if either:
761 * (1) current != ubq_daemon.
762 * io_uring_cmd_complete_in_task() tries to run task_work
763 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
765 * (2) current->flags & PF_EXITING.
767 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
768 __ublk_abort_rq(ubq, req);
772 if (ublk_need_get_data(ubq) &&
773 (req_op(req) == REQ_OP_WRITE ||
774 req_op(req) == REQ_OP_FLUSH)) {
776 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
777 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
780 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
781 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
782 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
783 __func__, io->cmd->cmd_op, ubq->q_id,
784 req->tag, io->flags);
785 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
789 * We have handled UBLK_IO_NEED_GET_DATA command,
790 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
793 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
794 /* update iod->addr because ublksrv may have passed a new io buffer */
795 ublk_get_iod(ubq, req->tag)->addr = io->addr;
796 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
797 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
798 ublk_get_iod(ubq, req->tag)->addr);
801 mapped_bytes = ublk_map_io(ubq, req, io);
803 /* partially mapped, update io descriptor */
804 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
806 * Nothing mapped, retry until we succeed.
808 * We may never succeed in mapping any bytes here because
809 * of OOM. TODO: reserve one buffer with single page pinned
810 * for providing forward progress guarantee.
812 if (unlikely(!mapped_bytes)) {
813 blk_mq_requeue_request(req, false);
814 blk_mq_delay_kick_requeue_list(req->q,
815 UBLK_REQUEUE_DELAY_MS);
819 ublk_get_iod(ubq, req->tag)->nr_sectors =
823 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
826 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
828 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
829 struct ublk_rq_data *data, *tmp;
831 io_cmds = llist_reverse_order(io_cmds);
832 llist_for_each_entry_safe(data, tmp, io_cmds, node)
833 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
836 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
838 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
839 struct ublk_rq_data *data, *tmp;
841 llist_for_each_entry_safe(data, tmp, io_cmds, node)
842 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
845 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
847 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
848 struct ublk_queue *ubq = pdu->ubq;
850 ublk_forward_io_cmds(ubq);
853 static void ublk_rq_task_work_fn(struct callback_head *work)
855 struct ublk_rq_data *data = container_of(work,
856 struct ublk_rq_data, work);
857 struct request *req = blk_mq_rq_from_pdu(data);
858 struct ublk_queue *ubq = req->mq_hctx->driver_data;
860 ublk_forward_io_cmds(ubq);
863 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
865 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
868 if (!llist_add(&data->node, &ubq->io_cmds))
871 io = &ubq->ios[rq->tag];
873 * If the check pass, we know that this is a re-issued request aborted
874 * previously in monitor_work because the ubq_daemon(cmd's task) is
875 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
876 * because this ioucmd's io_uring context may be freed now if no inflight
877 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
879 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
880 * the tag). Then the request is re-started(allocating the tag) and we are here.
881 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
882 * guarantees that here is a re-issued request aborted previously.
884 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
885 ublk_abort_io_cmds(ubq);
886 } else if (ublk_can_use_task_work(ubq)) {
887 if (task_work_add(ubq->ubq_daemon, &data->work,
889 ublk_abort_io_cmds(ubq);
891 struct io_uring_cmd *cmd = io->cmd;
892 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
895 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
899 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
900 const struct blk_mq_queue_data *bd)
902 struct ublk_queue *ubq = hctx->driver_data;
903 struct request *rq = bd->rq;
906 /* fill iod to slot in io cmd buffer */
907 res = ublk_setup_iod(ubq, rq);
908 if (unlikely(res != BLK_STS_OK))
909 return BLK_STS_IOERR;
911 /* With recovery feature enabled, force_abort is set in
912 * ublk_stop_dev() before calling del_gendisk(). We have to
913 * abort all requeued and new rqs here to let del_gendisk()
914 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
915 * to avoid UAF on io_uring ctx.
917 * Note: force_abort is guaranteed to be seen because it is set
918 * before request queue is unqiuesced.
920 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
921 return BLK_STS_IOERR;
923 blk_mq_start_request(bd->rq);
925 if (unlikely(ubq_daemon_is_dying(ubq))) {
926 __ublk_abort_rq(ubq, rq);
930 ublk_queue_cmd(ubq, rq);
935 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
936 unsigned int hctx_idx)
938 struct ublk_device *ub = driver_data;
939 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
941 hctx->driver_data = ubq;
945 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
946 unsigned int hctx_idx, unsigned int numa_node)
948 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
950 init_task_work(&data->work, ublk_rq_task_work_fn);
954 static const struct blk_mq_ops ublk_mq_ops = {
955 .queue_rq = ublk_queue_rq,
956 .init_hctx = ublk_init_hctx,
957 .init_request = ublk_init_rq,
960 static int ublk_ch_open(struct inode *inode, struct file *filp)
962 struct ublk_device *ub = container_of(inode->i_cdev,
963 struct ublk_device, cdev);
965 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
967 filp->private_data = ub;
971 static int ublk_ch_release(struct inode *inode, struct file *filp)
973 struct ublk_device *ub = filp->private_data;
975 clear_bit(UB_STATE_OPEN, &ub->state);
979 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
980 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
982 struct ublk_device *ub = filp->private_data;
983 size_t sz = vma->vm_end - vma->vm_start;
984 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
985 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
988 spin_lock(&ub->mm_lock);
990 ub->mm = current->mm;
991 if (current->mm != ub->mm)
993 spin_unlock(&ub->mm_lock);
998 if (vma->vm_flags & VM_WRITE)
1001 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1002 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1005 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1006 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1007 __func__, q_id, current->pid, vma->vm_start,
1008 phys_off, (unsigned long)sz);
1010 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1013 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1014 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1017 static void ublk_commit_completion(struct ublk_device *ub,
1018 struct ublksrv_io_cmd *ub_cmd)
1020 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1021 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1022 struct ublk_io *io = &ubq->ios[tag];
1023 struct request *req;
1025 /* now this cmd slot is owned by nbd driver */
1026 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1027 io->res = ub_cmd->result;
1029 /* find the io request and complete */
1030 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1032 if (req && likely(!blk_should_fake_timeout(req->q)))
1033 ublk_complete_rq(req);
1037 * When ->ubq_daemon is exiting, either new request is ended immediately,
1038 * or any queued io command is drained, so it is safe to abort queue
1041 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1045 if (!ublk_get_device(ub))
1048 for (i = 0; i < ubq->q_depth; i++) {
1049 struct ublk_io *io = &ubq->ios[i];
1051 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1055 * Either we fail the request or ublk_rq_task_work_fn
1058 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1060 __ublk_fail_req(ubq, io, rq);
1063 ublk_put_device(ub);
1066 static void ublk_daemon_monitor_work(struct work_struct *work)
1068 struct ublk_device *ub =
1069 container_of(work, struct ublk_device, monitor_work.work);
1072 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1073 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1075 if (ubq_daemon_is_dying(ubq)) {
1076 if (ublk_queue_can_use_recovery(ubq))
1077 schedule_work(&ub->quiesce_work);
1079 schedule_work(&ub->stop_work);
1081 /* abort queue is for making forward progress */
1082 ublk_abort_queue(ub, ubq);
1087 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1088 * after ublk_remove() or __ublk_quiesce_dev() is started.
1090 * No need ub->mutex, monitor work are canceled after state is marked
1091 * as not LIVE, so new state is observed reliably.
1093 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1094 schedule_delayed_work(&ub->monitor_work,
1095 UBLK_DAEMON_MONITOR_PERIOD);
1098 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1100 return ubq->nr_io_ready == ubq->q_depth;
1103 static void ublk_cancel_queue(struct ublk_queue *ubq)
1107 if (!ublk_queue_ready(ubq))
1110 for (i = 0; i < ubq->q_depth; i++) {
1111 struct ublk_io *io = &ubq->ios[i];
1113 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1114 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1117 /* all io commands are canceled */
1118 ubq->nr_io_ready = 0;
1121 /* Cancel all pending commands, must be called after del_gendisk() returns */
1122 static void ublk_cancel_dev(struct ublk_device *ub)
1126 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1127 ublk_cancel_queue(ublk_get_queue(ub, i));
1130 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1134 if (blk_mq_request_started(rq)) {
1141 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1145 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1148 blk_mq_tagset_busy_iter(&ub->tag_set,
1149 ublk_check_inflight_rq, &idle);
1152 msleep(UBLK_REQUEUE_DELAY_MS);
1156 static void __ublk_quiesce_dev(struct ublk_device *ub)
1158 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1159 __func__, ub->dev_info.dev_id,
1160 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1161 "LIVE" : "QUIESCED");
1162 blk_mq_quiesce_queue(ub->ub_disk->queue);
1163 ublk_wait_tagset_rqs_idle(ub);
1164 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1165 ublk_cancel_dev(ub);
1166 /* we are going to release task_struct of ubq_daemon and resets
1167 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1168 * Besides, monitor_work is not necessary in QUIESCED state since we have
1169 * already scheduled quiesce_work and quiesced all ubqs.
1171 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1172 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1174 cancel_delayed_work_sync(&ub->monitor_work);
1177 static void ublk_quiesce_work_fn(struct work_struct *work)
1179 struct ublk_device *ub =
1180 container_of(work, struct ublk_device, quiesce_work);
1182 mutex_lock(&ub->mutex);
1183 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1185 __ublk_quiesce_dev(ub);
1187 mutex_unlock(&ub->mutex);
1190 static void ublk_unquiesce_dev(struct ublk_device *ub)
1194 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1195 __func__, ub->dev_info.dev_id,
1196 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1197 "LIVE" : "QUIESCED");
1198 /* quiesce_work has run. We let requeued rqs be aborted
1199 * before running fallback_wq. "force_abort" must be seen
1200 * after request queue is unqiuesced. Then del_gendisk()
1203 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1204 ublk_get_queue(ub, i)->force_abort = true;
1206 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1207 /* We may have requeued some rqs in ublk_quiesce_queue() */
1208 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1211 static void ublk_stop_dev(struct ublk_device *ub)
1213 mutex_lock(&ub->mutex);
1214 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1216 if (ublk_can_use_recovery(ub)) {
1217 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1218 __ublk_quiesce_dev(ub);
1219 ublk_unquiesce_dev(ub);
1221 del_gendisk(ub->ub_disk);
1222 ub->dev_info.state = UBLK_S_DEV_DEAD;
1223 ub->dev_info.ublksrv_pid = -1;
1224 put_disk(ub->ub_disk);
1227 ublk_cancel_dev(ub);
1228 mutex_unlock(&ub->mutex);
1229 cancel_delayed_work_sync(&ub->monitor_work);
1232 /* device can only be started after all IOs are ready */
1233 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1235 mutex_lock(&ub->mutex);
1237 if (ublk_queue_ready(ubq)) {
1238 ubq->ubq_daemon = current;
1239 get_task_struct(ubq->ubq_daemon);
1240 ub->nr_queues_ready++;
1242 if (capable(CAP_SYS_ADMIN))
1243 ub->nr_privileged_daemon++;
1245 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1246 complete_all(&ub->completion);
1247 mutex_unlock(&ub->mutex);
1250 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1253 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1254 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1256 ublk_queue_cmd(ubq, req);
1259 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1261 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1262 struct ublk_device *ub = cmd->file->private_data;
1263 struct ublk_queue *ubq;
1265 u32 cmd_op = cmd->cmd_op;
1266 unsigned tag = ub_cmd->tag;
1268 struct request *req;
1270 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1271 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1274 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1277 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1278 if (!ubq || ub_cmd->q_id != ubq->q_id)
1281 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1284 if (tag >= ubq->q_depth)
1287 io = &ubq->ios[tag];
1289 /* there is pending io cmd, something must be wrong */
1290 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1296 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1297 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1299 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1300 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1304 case UBLK_IO_FETCH_REQ:
1305 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1306 if (ublk_queue_ready(ubq)) {
1311 * The io is being handled by server, so COMMIT_RQ is expected
1312 * instead of FETCH_REQ
1314 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1316 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1317 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1320 io->flags |= UBLK_IO_FLAG_ACTIVE;
1321 io->addr = ub_cmd->addr;
1323 ublk_mark_io_ready(ub, ubq);
1325 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1326 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1328 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1329 * not enabled or it is Read IO.
1331 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1333 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1335 io->addr = ub_cmd->addr;
1336 io->flags |= UBLK_IO_FLAG_ACTIVE;
1338 ublk_commit_completion(ub, ub_cmd);
1340 case UBLK_IO_NEED_GET_DATA:
1341 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1343 io->addr = ub_cmd->addr;
1345 io->flags |= UBLK_IO_FLAG_ACTIVE;
1346 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1351 return -EIOCBQUEUED;
1354 io_uring_cmd_done(cmd, ret, 0);
1355 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1356 __func__, cmd_op, tag, ret, io->flags);
1357 return -EIOCBQUEUED;
1360 static const struct file_operations ublk_ch_fops = {
1361 .owner = THIS_MODULE,
1362 .open = ublk_ch_open,
1363 .release = ublk_ch_release,
1364 .llseek = no_llseek,
1365 .uring_cmd = ublk_ch_uring_cmd,
1366 .mmap = ublk_ch_mmap,
1369 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1371 int size = ublk_queue_cmd_buf_size(ub, q_id);
1372 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1374 if (ubq->ubq_daemon)
1375 put_task_struct(ubq->ubq_daemon);
1376 if (ubq->io_cmd_buf)
1377 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1380 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1382 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1383 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1387 ubq->flags = ub->dev_info.flags;
1389 ubq->q_depth = ub->dev_info.queue_depth;
1390 size = ublk_queue_cmd_buf_size(ub, q_id);
1392 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1396 ubq->io_cmd_buf = ptr;
1401 static void ublk_deinit_queues(struct ublk_device *ub)
1403 int nr_queues = ub->dev_info.nr_hw_queues;
1409 for (i = 0; i < nr_queues; i++)
1410 ublk_deinit_queue(ub, i);
1411 kfree(ub->__queues);
1414 static int ublk_init_queues(struct ublk_device *ub)
1416 int nr_queues = ub->dev_info.nr_hw_queues;
1417 int depth = ub->dev_info.queue_depth;
1418 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1419 int i, ret = -ENOMEM;
1421 ub->queue_size = ubq_size;
1422 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1426 for (i = 0; i < nr_queues; i++) {
1427 if (ublk_init_queue(ub, i))
1431 init_completion(&ub->completion);
1435 ublk_deinit_queues(ub);
1439 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1444 spin_lock(&ublk_idr_lock);
1445 /* allocate id, if @id >= 0, we're requesting that specific id */
1447 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1451 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1453 spin_unlock(&ublk_idr_lock);
1456 ub->ub_number = err;
1461 static void ublk_free_dev_number(struct ublk_device *ub)
1463 spin_lock(&ublk_idr_lock);
1464 idr_remove(&ublk_index_idr, ub->ub_number);
1465 wake_up_all(&ublk_idr_wq);
1466 spin_unlock(&ublk_idr_lock);
1469 static void ublk_cdev_rel(struct device *dev)
1471 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1473 blk_mq_free_tag_set(&ub->tag_set);
1474 ublk_deinit_queues(ub);
1475 ublk_free_dev_number(ub);
1476 mutex_destroy(&ub->mutex);
1480 static int ublk_add_chdev(struct ublk_device *ub)
1482 struct device *dev = &ub->cdev_dev;
1483 int minor = ub->ub_number;
1486 dev->parent = ublk_misc.this_device;
1487 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1488 dev->class = ublk_chr_class;
1489 dev->release = ublk_cdev_rel;
1490 device_initialize(dev);
1492 ret = dev_set_name(dev, "ublkc%d", minor);
1496 cdev_init(&ub->cdev, &ublk_ch_fops);
1497 ret = cdev_device_add(&ub->cdev, dev);
1508 static void ublk_stop_work_fn(struct work_struct *work)
1510 struct ublk_device *ub =
1511 container_of(work, struct ublk_device, stop_work);
1516 /* align max io buffer size with PAGE_SIZE */
1517 static void ublk_align_max_io_size(struct ublk_device *ub)
1519 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1521 ub->dev_info.max_io_buf_bytes =
1522 round_down(max_io_bytes, PAGE_SIZE);
1525 static int ublk_add_tag_set(struct ublk_device *ub)
1527 ub->tag_set.ops = &ublk_mq_ops;
1528 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1529 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1530 ub->tag_set.numa_node = NUMA_NO_NODE;
1531 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1532 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1533 ub->tag_set.driver_data = ub;
1534 return blk_mq_alloc_tag_set(&ub->tag_set);
1537 static void ublk_remove(struct ublk_device *ub)
1540 cancel_work_sync(&ub->stop_work);
1541 cancel_work_sync(&ub->quiesce_work);
1542 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1543 put_device(&ub->cdev_dev);
1547 static struct ublk_device *ublk_get_device_from_id(int idx)
1549 struct ublk_device *ub = NULL;
1554 spin_lock(&ublk_idr_lock);
1555 ub = idr_find(&ublk_index_idr, idx);
1557 ub = ublk_get_device(ub);
1558 spin_unlock(&ublk_idr_lock);
1563 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1565 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1566 int ublksrv_pid = (int)header->data[0];
1567 struct gendisk *disk;
1570 if (ublksrv_pid <= 0)
1573 wait_for_completion_interruptible(&ub->completion);
1575 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1577 mutex_lock(&ub->mutex);
1578 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1579 test_bit(UB_STATE_USED, &ub->state)) {
1584 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1586 ret = PTR_ERR(disk);
1589 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1590 disk->fops = &ub_fops;
1591 disk->private_data = ub;
1593 ub->dev_info.ublksrv_pid = ublksrv_pid;
1596 ret = ublk_apply_params(ub);
1600 /* don't probe partitions if any one ubq daemon is un-trusted */
1601 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1602 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1604 get_device(&ub->cdev_dev);
1605 ret = add_disk(disk);
1608 * Has to drop the reference since ->free_disk won't be
1609 * called in case of add_disk failure.
1611 ublk_put_device(ub);
1614 set_bit(UB_STATE_USED, &ub->state);
1615 ub->dev_info.state = UBLK_S_DEV_LIVE;
1620 mutex_unlock(&ub->mutex);
1624 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1625 struct io_uring_cmd *cmd)
1627 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1628 void __user *argp = (void __user *)(unsigned long)header->addr;
1629 cpumask_var_t cpumask;
1630 unsigned long queue;
1631 unsigned int retlen;
1635 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1637 if (header->len & (sizeof(unsigned long)-1))
1642 queue = header->data[0];
1643 if (queue >= ub->dev_info.nr_hw_queues)
1646 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1649 for_each_possible_cpu(i) {
1650 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1651 cpumask_set_cpu(i, cpumask);
1655 retlen = min_t(unsigned short, header->len, cpumask_size());
1656 if (copy_to_user(argp, cpumask, retlen))
1657 goto out_free_cpumask;
1658 if (retlen != header->len &&
1659 clear_user(argp + retlen, header->len - retlen))
1660 goto out_free_cpumask;
1664 free_cpumask_var(cpumask);
1668 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1670 pr_devel("%s: dev id %d flags %llx\n", __func__,
1671 info->dev_id, info->flags);
1672 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1673 info->nr_hw_queues, info->queue_depth);
1676 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1678 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1679 void __user *argp = (void __user *)(unsigned long)header->addr;
1680 struct ublksrv_ctrl_dev_info info;
1681 struct ublk_device *ub;
1684 if (header->len < sizeof(info) || !header->addr)
1686 if (header->queue_id != (u16)-1) {
1687 pr_warn("%s: queue_id is wrong %x\n",
1688 __func__, header->queue_id);
1692 if (copy_from_user(&info, argp, sizeof(info)))
1695 if (capable(CAP_SYS_ADMIN))
1696 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1697 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1700 /* the created device is always owned by current user */
1701 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1703 if (header->dev_id != info.dev_id) {
1704 pr_warn("%s: dev id not match %u %u\n",
1705 __func__, header->dev_id, info.dev_id);
1709 ublk_dump_dev_info(&info);
1711 ret = mutex_lock_killable(&ublk_ctl_mutex);
1716 if (ublks_added >= ublks_max)
1720 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1723 mutex_init(&ub->mutex);
1724 spin_lock_init(&ub->mm_lock);
1725 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1726 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1727 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1729 ret = ublk_alloc_dev_number(ub, header->dev_id);
1733 memcpy(&ub->dev_info, &info, sizeof(info));
1735 /* update device id */
1736 ub->dev_info.dev_id = ub->ub_number;
1739 * 64bit flags will be copied back to userspace as feature
1740 * negotiation result, so have to clear flags which driver
1741 * doesn't support yet, then userspace can get correct flags
1742 * (features) to handle.
1744 ub->dev_info.flags &= UBLK_F_ALL;
1746 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1747 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1749 /* We are not ready to support zero copy */
1750 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1752 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1753 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1754 ublk_align_max_io_size(ub);
1756 ret = ublk_init_queues(ub);
1758 goto out_free_dev_number;
1760 ret = ublk_add_tag_set(ub);
1762 goto out_deinit_queues;
1765 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1766 goto out_free_tag_set;
1769 * Add the char dev so that ublksrv daemon can be setup.
1770 * ublk_add_chdev() will cleanup everything if it fails.
1772 ret = ublk_add_chdev(ub);
1776 blk_mq_free_tag_set(&ub->tag_set);
1778 ublk_deinit_queues(ub);
1779 out_free_dev_number:
1780 ublk_free_dev_number(ub);
1782 mutex_destroy(&ub->mutex);
1785 mutex_unlock(&ublk_ctl_mutex);
1789 static inline bool ublk_idr_freed(int id)
1793 spin_lock(&ublk_idr_lock);
1794 ptr = idr_find(&ublk_index_idr, id);
1795 spin_unlock(&ublk_idr_lock);
1800 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1802 struct ublk_device *ub = *p_ub;
1803 int idx = ub->ub_number;
1806 ret = mutex_lock_killable(&ublk_ctl_mutex);
1810 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1812 set_bit(UB_STATE_DELETED, &ub->state);
1815 /* Mark the reference as consumed */
1817 ublk_put_device(ub);
1818 mutex_unlock(&ublk_ctl_mutex);
1821 * Wait until the idr is removed, then it can be reused after
1822 * DEL_DEV command is returned.
1824 * If we returns because of user interrupt, future delete command
1827 * - the device number isn't freed, this device won't or needn't
1828 * be deleted again, since UB_STATE_DELETED is set, and device
1829 * will be released after the last reference is dropped
1831 * - the device number is freed already, we will not find this
1832 * device via ublk_get_device_from_id()
1834 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1839 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1841 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1843 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1844 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1845 header->data[0], header->addr, header->len);
1848 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1851 cancel_work_sync(&ub->stop_work);
1852 cancel_work_sync(&ub->quiesce_work);
1857 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1858 struct io_uring_cmd *cmd)
1860 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1861 void __user *argp = (void __user *)(unsigned long)header->addr;
1863 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1866 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1872 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1873 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1875 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1876 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1879 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1880 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1882 ub->params.devt.disk_major = 0;
1883 ub->params.devt.disk_minor = 0;
1885 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1888 static int ublk_ctrl_get_params(struct ublk_device *ub,
1889 struct io_uring_cmd *cmd)
1891 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1892 void __user *argp = (void __user *)(unsigned long)header->addr;
1893 struct ublk_params_header ph;
1896 if (header->len <= sizeof(ph) || !header->addr)
1899 if (copy_from_user(&ph, argp, sizeof(ph)))
1902 if (ph.len > header->len || !ph.len)
1905 if (ph.len > sizeof(struct ublk_params))
1906 ph.len = sizeof(struct ublk_params);
1908 mutex_lock(&ub->mutex);
1909 ublk_ctrl_fill_params_devt(ub);
1910 if (copy_to_user(argp, &ub->params, ph.len))
1914 mutex_unlock(&ub->mutex);
1919 static int ublk_ctrl_set_params(struct ublk_device *ub,
1920 struct io_uring_cmd *cmd)
1922 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1923 void __user *argp = (void __user *)(unsigned long)header->addr;
1924 struct ublk_params_header ph;
1927 if (header->len <= sizeof(ph) || !header->addr)
1930 if (copy_from_user(&ph, argp, sizeof(ph)))
1933 if (ph.len > header->len || !ph.len || !ph.types)
1936 if (ph.len > sizeof(struct ublk_params))
1937 ph.len = sizeof(struct ublk_params);
1939 /* parameters can only be changed when device isn't live */
1940 mutex_lock(&ub->mutex);
1941 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1943 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1946 /* clear all we don't support yet */
1947 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1948 ret = ublk_validate_params(ub);
1950 mutex_unlock(&ub->mutex);
1955 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1959 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1960 /* All old ioucmds have to be completed */
1961 WARN_ON_ONCE(ubq->nr_io_ready);
1962 /* old daemon is PF_EXITING, put it now */
1963 put_task_struct(ubq->ubq_daemon);
1964 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1965 ubq->ubq_daemon = NULL;
1967 for (i = 0; i < ubq->q_depth; i++) {
1968 struct ublk_io *io = &ubq->ios[i];
1970 /* forget everything now and be ready for new FETCH_REQ */
1977 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1978 struct io_uring_cmd *cmd)
1980 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1984 mutex_lock(&ub->mutex);
1985 if (!ublk_can_use_recovery(ub))
1988 * START_RECOVERY is only allowd after:
1990 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1991 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1994 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1995 * (a)has quiesced request queue
1996 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1997 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1998 * (d)has completed/camceled all ioucmds owned by ther dying process
2000 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2001 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2005 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2006 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2007 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2008 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2010 ub->nr_queues_ready = 0;
2011 ub->nr_privileged_daemon = 0;
2012 init_completion(&ub->completion);
2015 mutex_unlock(&ub->mutex);
2019 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2020 struct io_uring_cmd *cmd)
2022 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2023 int ublksrv_pid = (int)header->data[0];
2026 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2027 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2028 /* wait until new ubq_daemon sending all FETCH_REQ */
2029 wait_for_completion_interruptible(&ub->completion);
2030 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2031 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2033 mutex_lock(&ub->mutex);
2034 if (!ublk_can_use_recovery(ub))
2037 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2041 ub->dev_info.ublksrv_pid = ublksrv_pid;
2042 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2043 __func__, ublksrv_pid, header->dev_id);
2044 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2045 pr_devel("%s: queue unquiesced, dev id %d.\n",
2046 __func__, header->dev_id);
2047 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2048 ub->dev_info.state = UBLK_S_DEV_LIVE;
2049 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2052 mutex_unlock(&ub->mutex);
2057 * All control commands are sent via /dev/ublk-control, so we have to check
2058 * the destination device's permission
2060 static int ublk_char_dev_permission(struct ublk_device *ub,
2061 const char *dev_path, int mask)
2067 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2071 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2076 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2079 err = inode_permission(&nop_mnt_idmap,
2080 d_backing_inode(path.dentry), mask);
2086 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2087 struct io_uring_cmd *cmd)
2089 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2090 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2091 void __user *argp = (void __user *)(unsigned long)header->addr;
2092 char *dev_path = NULL;
2096 if (!unprivileged) {
2097 if (!capable(CAP_SYS_ADMIN))
2100 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2101 * char_dev_path in payload too, since userspace may not
2102 * know if the specified device is created as unprivileged
2105 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2110 * User has to provide the char device path for unprivileged ublk
2112 * header->addr always points to the dev path buffer, and
2113 * header->dev_path_len records length of dev path buffer.
2115 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2118 if (header->len < header->dev_path_len)
2121 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2126 if (copy_from_user(dev_path, argp, header->dev_path_len))
2128 dev_path[header->dev_path_len] = 0;
2131 switch (cmd->cmd_op) {
2132 case UBLK_CMD_GET_DEV_INFO:
2133 case UBLK_CMD_GET_DEV_INFO2:
2134 case UBLK_CMD_GET_QUEUE_AFFINITY:
2135 case UBLK_CMD_GET_PARAMS:
2138 case UBLK_CMD_START_DEV:
2139 case UBLK_CMD_STOP_DEV:
2140 case UBLK_CMD_ADD_DEV:
2141 case UBLK_CMD_DEL_DEV:
2142 case UBLK_CMD_SET_PARAMS:
2143 case UBLK_CMD_START_USER_RECOVERY:
2144 case UBLK_CMD_END_USER_RECOVERY:
2145 mask = MAY_READ | MAY_WRITE;
2151 ret = ublk_char_dev_permission(ub, dev_path, mask);
2153 header->len -= header->dev_path_len;
2154 header->addr += header->dev_path_len;
2156 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2157 __func__, ub->ub_number, cmd->cmd_op,
2158 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2165 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2166 unsigned int issue_flags)
2168 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2169 struct ublk_device *ub = NULL;
2172 if (issue_flags & IO_URING_F_NONBLOCK)
2175 ublk_ctrl_cmd_dump(cmd);
2177 if (!(issue_flags & IO_URING_F_SQE128))
2180 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2182 ub = ublk_get_device_from_id(header->dev_id);
2186 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2188 /* ADD_DEV permission check is done in command handler */
2195 switch (cmd->cmd_op) {
2196 case UBLK_CMD_START_DEV:
2197 ret = ublk_ctrl_start_dev(ub, cmd);
2199 case UBLK_CMD_STOP_DEV:
2200 ret = ublk_ctrl_stop_dev(ub);
2202 case UBLK_CMD_GET_DEV_INFO:
2203 case UBLK_CMD_GET_DEV_INFO2:
2204 ret = ublk_ctrl_get_dev_info(ub, cmd);
2206 case UBLK_CMD_ADD_DEV:
2207 ret = ublk_ctrl_add_dev(cmd);
2209 case UBLK_CMD_DEL_DEV:
2210 ret = ublk_ctrl_del_dev(&ub);
2212 case UBLK_CMD_GET_QUEUE_AFFINITY:
2213 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2215 case UBLK_CMD_GET_PARAMS:
2216 ret = ublk_ctrl_get_params(ub, cmd);
2218 case UBLK_CMD_SET_PARAMS:
2219 ret = ublk_ctrl_set_params(ub, cmd);
2221 case UBLK_CMD_START_USER_RECOVERY:
2222 ret = ublk_ctrl_start_recovery(ub, cmd);
2224 case UBLK_CMD_END_USER_RECOVERY:
2225 ret = ublk_ctrl_end_recovery(ub, cmd);
2234 ublk_put_device(ub);
2236 io_uring_cmd_done(cmd, ret, 0);
2237 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2238 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2239 return -EIOCBQUEUED;
2242 static const struct file_operations ublk_ctl_fops = {
2243 .open = nonseekable_open,
2244 .uring_cmd = ublk_ctrl_uring_cmd,
2245 .owner = THIS_MODULE,
2246 .llseek = noop_llseek,
2249 static struct miscdevice ublk_misc = {
2250 .minor = MISC_DYNAMIC_MINOR,
2251 .name = "ublk-control",
2252 .fops = &ublk_ctl_fops,
2255 static int __init ublk_init(void)
2259 init_waitqueue_head(&ublk_idr_wq);
2261 ret = misc_register(&ublk_misc);
2265 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2267 goto unregister_mis;
2269 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2270 if (IS_ERR(ublk_chr_class)) {
2271 ret = PTR_ERR(ublk_chr_class);
2272 goto free_chrdev_region;
2277 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2279 misc_deregister(&ublk_misc);
2283 static void __exit ublk_exit(void)
2285 struct ublk_device *ub;
2288 idr_for_each_entry(&ublk_index_idr, ub, id)
2291 class_destroy(ublk_chr_class);
2292 misc_deregister(&ublk_misc);
2294 idr_destroy(&ublk_index_idr);
2295 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2298 module_init(ublk_init);
2299 module_exit(ublk_exit);
2301 module_param(ublks_max, int, 0444);
2302 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2304 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2305 MODULE_LICENSE("GPL");