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/cmd.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 <linux/kref.h>
47 #include <uapi/linux/ublk_cmd.h>
49 #define UBLK_MINORS (1U << MINORBITS)
51 /* All UBLK_F_* have to be included into UBLK_F_ALL */
52 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
53 | UBLK_F_URING_CMD_COMP_IN_TASK \
54 | UBLK_F_NEED_GET_DATA \
55 | UBLK_F_USER_RECOVERY \
56 | UBLK_F_USER_RECOVERY_REISSUE \
57 | UBLK_F_UNPRIVILEGED_DEV \
58 | UBLK_F_CMD_IOCTL_ENCODE \
62 /* All UBLK_PARAM_TYPE_* should be included here */
63 #define UBLK_PARAM_TYPE_ALL \
64 (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD | \
65 UBLK_PARAM_TYPE_DEVT | UBLK_PARAM_TYPE_ZONED)
68 struct llist_node node;
76 struct ublk_uring_cmd_pdu {
77 struct ublk_queue *ubq;
82 * io command is active: sqe cmd is received, and its cqe isn't done
84 * If the flag is set, the io command is owned by ublk driver, and waited
85 * for incoming blk-mq request from the ublk block device.
87 * If the flag is cleared, the io command will be completed, and owned by
90 #define UBLK_IO_FLAG_ACTIVE 0x01
93 * IO command is completed via cqe, and it is being handled by ublksrv, and
96 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
99 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
102 * IO command is aborted, so this flag is set in case of
103 * !UBLK_IO_FLAG_ACTIVE.
105 * After this flag is observed, any pending or new incoming request
106 * associated with this io command will be failed immediately
108 #define UBLK_IO_FLAG_ABORTED 0x04
111 * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
112 * get data buffer address from ublksrv.
114 * Then, bio data could be copied into this data buffer for a WRITE request
115 * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
117 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
119 /* atomic RW with ubq->cancel_lock */
120 #define UBLK_IO_FLAG_CANCELED 0x80000000
123 /* userspace buffer address from io cmd */
128 struct io_uring_cmd *cmd;
136 struct task_struct *ubq_daemon;
139 struct llist_head io_cmds;
141 unsigned long io_addr; /* mapped vm address */
142 unsigned int max_io_sz;
146 unsigned short nr_io_ready; /* how many ios setup */
147 spinlock_t cancel_lock;
148 struct ublk_device *dev;
149 struct ublk_io ios[];
153 struct gendisk *ub_disk;
157 unsigned int queue_size;
158 struct ublksrv_ctrl_dev_info dev_info;
160 struct blk_mq_tag_set tag_set;
163 struct device cdev_dev;
165 #define UB_STATE_OPEN 0
166 #define UB_STATE_USED 1
167 #define UB_STATE_DELETED 2
174 struct mm_struct *mm;
176 struct ublk_params params;
178 struct completion completion;
179 unsigned int nr_queues_ready;
180 unsigned int nr_privileged_daemon;
182 struct work_struct quiesce_work;
183 struct work_struct stop_work;
186 /* header of ublk_params */
187 struct ublk_params_header {
192 static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq);
194 static inline unsigned int ublk_req_build_flags(struct request *req);
195 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
197 static inline bool ublk_dev_is_user_copy(const struct ublk_device *ub)
199 return ub->dev_info.flags & UBLK_F_USER_COPY;
202 static inline bool ublk_dev_is_zoned(const struct ublk_device *ub)
204 return ub->dev_info.flags & UBLK_F_ZONED;
207 static inline bool ublk_queue_is_zoned(struct ublk_queue *ubq)
209 return ubq->flags & UBLK_F_ZONED;
212 #ifdef CONFIG_BLK_DEV_ZONED
214 static int ublk_get_nr_zones(const struct ublk_device *ub)
216 const struct ublk_param_basic *p = &ub->params.basic;
218 /* Zone size is a power of 2 */
219 return p->dev_sectors >> ilog2(p->chunk_sectors);
222 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
224 return blk_revalidate_disk_zones(ub->ub_disk, NULL);
227 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
229 const struct ublk_param_zoned *p = &ub->params.zoned;
232 if (!ublk_dev_is_zoned(ub))
235 if (!p->max_zone_append_sectors)
238 nr_zones = ublk_get_nr_zones(ub);
240 if (p->max_active_zones > nr_zones)
243 if (p->max_open_zones > nr_zones)
249 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
251 const struct ublk_param_zoned *p = &ub->params.zoned;
253 disk_set_zoned(ub->ub_disk);
254 blk_queue_flag_set(QUEUE_FLAG_ZONE_RESETALL, ub->ub_disk->queue);
255 blk_queue_required_elevator_features(ub->ub_disk->queue,
256 ELEVATOR_F_ZBD_SEQ_WRITE);
257 disk_set_max_active_zones(ub->ub_disk, p->max_active_zones);
258 disk_set_max_open_zones(ub->ub_disk, p->max_open_zones);
259 blk_queue_max_zone_append_sectors(ub->ub_disk->queue, p->max_zone_append_sectors);
261 ub->ub_disk->nr_zones = ublk_get_nr_zones(ub);
266 /* Based on virtblk_alloc_report_buffer */
267 static void *ublk_alloc_report_buffer(struct ublk_device *ublk,
268 unsigned int nr_zones, size_t *buflen)
270 struct request_queue *q = ublk->ub_disk->queue;
274 nr_zones = min_t(unsigned int, nr_zones,
275 ublk->ub_disk->nr_zones);
277 bufsize = nr_zones * sizeof(struct blk_zone);
279 min_t(size_t, bufsize, queue_max_hw_sectors(q) << SECTOR_SHIFT);
281 while (bufsize >= sizeof(struct blk_zone)) {
282 buf = kvmalloc(bufsize, GFP_KERNEL | __GFP_NORETRY);
294 static int ublk_report_zones(struct gendisk *disk, sector_t sector,
295 unsigned int nr_zones, report_zones_cb cb, void *data)
297 struct ublk_device *ub = disk->private_data;
298 unsigned int zone_size_sectors = disk->queue->limits.chunk_sectors;
299 unsigned int first_zone = sector >> ilog2(zone_size_sectors);
300 unsigned int done_zones = 0;
301 unsigned int max_zones_per_request;
303 struct blk_zone *buffer;
304 size_t buffer_length;
306 nr_zones = min_t(unsigned int, ub->ub_disk->nr_zones - first_zone,
309 buffer = ublk_alloc_report_buffer(ub, nr_zones, &buffer_length);
313 max_zones_per_request = buffer_length / sizeof(struct blk_zone);
315 while (done_zones < nr_zones) {
316 unsigned int remaining_zones = nr_zones - done_zones;
317 unsigned int zones_in_request =
318 min_t(unsigned int, remaining_zones, max_zones_per_request);
320 struct ublk_rq_data *pdu;
323 memset(buffer, 0, buffer_length);
325 req = blk_mq_alloc_request(disk->queue, REQ_OP_DRV_IN, 0);
331 pdu = blk_mq_rq_to_pdu(req);
332 pdu->operation = UBLK_IO_OP_REPORT_ZONES;
333 pdu->sector = sector;
334 pdu->nr_zones = zones_in_request;
336 ret = blk_rq_map_kern(disk->queue, req, buffer, buffer_length,
339 blk_mq_free_request(req);
343 status = blk_execute_rq(req, 0);
344 ret = blk_status_to_errno(status);
345 blk_mq_free_request(req);
349 for (unsigned int i = 0; i < zones_in_request; i++) {
350 struct blk_zone *zone = buffer + i;
352 /* A zero length zone means no more zones in this response */
356 ret = cb(zone, i, data);
361 sector += zone_size_sectors;
373 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
376 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
377 struct ublk_io *io = &ubq->ios[req->tag];
378 struct ublk_rq_data *pdu = blk_mq_rq_to_pdu(req);
381 switch (req_op(req)) {
382 case REQ_OP_ZONE_OPEN:
383 ublk_op = UBLK_IO_OP_ZONE_OPEN;
385 case REQ_OP_ZONE_CLOSE:
386 ublk_op = UBLK_IO_OP_ZONE_CLOSE;
388 case REQ_OP_ZONE_FINISH:
389 ublk_op = UBLK_IO_OP_ZONE_FINISH;
391 case REQ_OP_ZONE_RESET:
392 ublk_op = UBLK_IO_OP_ZONE_RESET;
394 case REQ_OP_ZONE_APPEND:
395 ublk_op = UBLK_IO_OP_ZONE_APPEND;
397 case REQ_OP_ZONE_RESET_ALL:
398 ublk_op = UBLK_IO_OP_ZONE_RESET_ALL;
401 ublk_op = pdu->operation;
403 case UBLK_IO_OP_REPORT_ZONES:
404 iod->op_flags = ublk_op | ublk_req_build_flags(req);
405 iod->nr_zones = pdu->nr_zones;
406 iod->start_sector = pdu->sector;
409 return BLK_STS_IOERR;
412 /* We do not support drv_out */
413 return BLK_STS_NOTSUPP;
415 return BLK_STS_IOERR;
418 iod->op_flags = ublk_op | ublk_req_build_flags(req);
419 iod->nr_sectors = blk_rq_sectors(req);
420 iod->start_sector = blk_rq_pos(req);
421 iod->addr = io->addr;
428 #define ublk_report_zones (NULL)
430 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
435 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
440 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
445 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
448 return BLK_STS_NOTSUPP;
453 static inline void __ublk_complete_rq(struct request *req);
454 static void ublk_complete_rq(struct kref *ref);
456 static dev_t ublk_chr_devt;
457 static const struct class ublk_chr_class = {
461 static DEFINE_IDR(ublk_index_idr);
462 static DEFINE_SPINLOCK(ublk_idr_lock);
463 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
465 static DEFINE_MUTEX(ublk_ctl_mutex);
468 * Max ublk devices allowed to add
470 * It can be extended to one per-user limit in future or even controlled
473 #define UBLK_MAX_UBLKS UBLK_MINORS
474 static unsigned int ublks_max = 64;
475 static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
477 static struct miscdevice ublk_misc;
479 static inline unsigned ublk_pos_to_hwq(loff_t pos)
481 return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
485 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
487 return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
490 static inline unsigned ublk_pos_to_tag(loff_t pos)
492 return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
496 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
498 struct request_queue *q = ub->ub_disk->queue;
499 const struct ublk_param_basic *p = &ub->params.basic;
501 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
502 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
503 blk_queue_io_min(q, 1 << p->io_min_shift);
504 blk_queue_io_opt(q, 1 << p->io_opt_shift);
506 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
507 p->attrs & UBLK_ATTR_FUA);
508 if (p->attrs & UBLK_ATTR_ROTATIONAL)
509 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
511 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
513 blk_queue_max_hw_sectors(q, p->max_sectors);
514 blk_queue_chunk_sectors(q, p->chunk_sectors);
515 blk_queue_virt_boundary(q, p->virt_boundary_mask);
517 if (p->attrs & UBLK_ATTR_READ_ONLY)
518 set_disk_ro(ub->ub_disk, true);
520 set_capacity(ub->ub_disk, p->dev_sectors);
523 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
525 struct request_queue *q = ub->ub_disk->queue;
526 const struct ublk_param_discard *p = &ub->params.discard;
528 q->limits.discard_alignment = p->discard_alignment;
529 q->limits.discard_granularity = p->discard_granularity;
530 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
531 blk_queue_max_write_zeroes_sectors(q,
532 p->max_write_zeroes_sectors);
533 blk_queue_max_discard_segments(q, p->max_discard_segments);
536 static int ublk_validate_params(const struct ublk_device *ub)
538 /* basic param is the only one which must be set */
539 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
540 const struct ublk_param_basic *p = &ub->params.basic;
542 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
545 if (p->logical_bs_shift > p->physical_bs_shift)
548 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
551 if (ublk_dev_is_zoned(ub) && !p->chunk_sectors)
556 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
557 const struct ublk_param_discard *p = &ub->params.discard;
559 /* So far, only support single segment discard */
560 if (p->max_discard_sectors && p->max_discard_segments != 1)
563 if (!p->discard_granularity)
567 /* dev_t is read-only */
568 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
571 if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
572 return ublk_dev_param_zoned_validate(ub);
573 else if (ublk_dev_is_zoned(ub))
579 static int ublk_apply_params(struct ublk_device *ub)
581 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
584 ublk_dev_param_basic_apply(ub);
586 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
587 ublk_dev_param_discard_apply(ub);
589 if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
590 return ublk_dev_param_zoned_apply(ub);
595 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
597 return ubq->flags & UBLK_F_USER_COPY;
600 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
603 * read()/write() is involved in user copy, so request reference
606 return ublk_support_user_copy(ubq);
609 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
612 if (ublk_need_req_ref(ubq)) {
613 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
615 kref_init(&data->ref);
619 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
622 if (ublk_need_req_ref(ubq)) {
623 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
625 return kref_get_unless_zero(&data->ref);
631 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
634 if (ublk_need_req_ref(ubq)) {
635 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
637 kref_put(&data->ref, ublk_complete_rq);
639 __ublk_complete_rq(req);
643 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
645 return ubq->flags & UBLK_F_NEED_GET_DATA;
648 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
650 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
655 static void ublk_put_device(struct ublk_device *ub)
657 put_device(&ub->cdev_dev);
660 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
663 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
666 static inline bool ublk_rq_has_data(const struct request *rq)
668 return bio_has_data(rq->bio);
671 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
674 return (struct ublksrv_io_desc *)
675 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
678 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
680 return ublk_get_queue(ub, q_id)->io_cmd_buf;
683 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
685 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
687 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
691 static inline bool ublk_queue_can_use_recovery_reissue(
692 struct ublk_queue *ubq)
694 return (ubq->flags & UBLK_F_USER_RECOVERY) &&
695 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
698 static inline bool ublk_queue_can_use_recovery(
699 struct ublk_queue *ubq)
701 return ubq->flags & UBLK_F_USER_RECOVERY;
704 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
706 return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
709 static void ublk_free_disk(struct gendisk *disk)
711 struct ublk_device *ub = disk->private_data;
713 clear_bit(UB_STATE_USED, &ub->state);
714 put_device(&ub->cdev_dev);
717 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
718 unsigned int *owner_gid)
723 current_uid_gid(&uid, &gid);
725 *owner_uid = from_kuid(&init_user_ns, uid);
726 *owner_gid = from_kgid(&init_user_ns, gid);
729 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
731 struct ublk_device *ub = disk->private_data;
733 if (capable(CAP_SYS_ADMIN))
737 * If it is one unprivileged device, only owner can open
738 * the disk. Otherwise it could be one trap made by one
739 * evil user who grants this disk's privileges to other
740 * users deliberately.
742 * This way is reasonable too given anyone can create
743 * unprivileged device, and no need other's grant.
745 if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
746 unsigned int curr_uid, curr_gid;
748 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
750 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
751 ub->dev_info.owner_gid)
758 static const struct block_device_operations ub_fops = {
759 .owner = THIS_MODULE,
761 .free_disk = ublk_free_disk,
762 .report_zones = ublk_report_zones,
765 #define UBLK_MAX_PIN_PAGES 32
767 struct ublk_io_iter {
768 struct page *pages[UBLK_MAX_PIN_PAGES];
770 struct bvec_iter iter;
773 /* return how many pages are copied */
774 static void ublk_copy_io_pages(struct ublk_io_iter *data,
775 size_t total, size_t pg_off, int dir)
780 while (done < total) {
781 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
782 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
783 (unsigned)(PAGE_SIZE - pg_off));
784 void *bv_buf = bvec_kmap_local(&bv);
785 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
787 if (dir == ITER_DEST)
788 memcpy(pg_buf + pg_off, bv_buf, bytes);
790 memcpy(bv_buf, pg_buf + pg_off, bytes);
792 kunmap_local(pg_buf);
793 kunmap_local(bv_buf);
795 /* advance page array */
797 if (pg_off == PAGE_SIZE) {
805 bio_advance_iter_single(data->bio, &data->iter, bytes);
806 if (!data->iter.bi_size) {
807 data->bio = data->bio->bi_next;
808 if (data->bio == NULL)
810 data->iter = data->bio->bi_iter;
815 static bool ublk_advance_io_iter(const struct request *req,
816 struct ublk_io_iter *iter, unsigned int offset)
818 struct bio *bio = req->bio;
821 if (bio->bi_iter.bi_size > offset) {
823 iter->iter = bio->bi_iter;
824 bio_advance_iter(iter->bio, &iter->iter, offset);
827 offset -= bio->bi_iter.bi_size;
833 * Copy data between request pages and io_iter, and 'offset'
834 * is the start point of linear offset of request.
836 static size_t ublk_copy_user_pages(const struct request *req,
837 unsigned offset, struct iov_iter *uiter, int dir)
839 struct ublk_io_iter iter;
842 if (!ublk_advance_io_iter(req, &iter, offset))
845 while (iov_iter_count(uiter) && iter.bio) {
851 len = iov_iter_get_pages2(uiter, iter.pages,
852 iov_iter_count(uiter),
853 UBLK_MAX_PIN_PAGES, &off);
857 ublk_copy_io_pages(&iter, len, off, dir);
858 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
859 for (i = 0; i < nr_pages; i++) {
860 if (dir == ITER_DEST)
861 set_page_dirty(iter.pages[i]);
862 put_page(iter.pages[i]);
870 static inline bool ublk_need_map_req(const struct request *req)
872 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
875 static inline bool ublk_need_unmap_req(const struct request *req)
877 return ublk_rq_has_data(req) &&
878 (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN);
881 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
884 const unsigned int rq_bytes = blk_rq_bytes(req);
886 if (ublk_support_user_copy(ubq))
890 * no zero copy, we delay copy WRITE request data into ublksrv
891 * context and the big benefit is that pinning pages in current
892 * context is pretty fast, see ublk_pin_user_pages
894 if (ublk_need_map_req(req)) {
895 struct iov_iter iter;
896 const int dir = ITER_DEST;
898 import_ubuf(dir, u64_to_user_ptr(io->addr), rq_bytes, &iter);
899 return ublk_copy_user_pages(req, 0, &iter, dir);
904 static int ublk_unmap_io(const struct ublk_queue *ubq,
905 const struct request *req,
908 const unsigned int rq_bytes = blk_rq_bytes(req);
910 if (ublk_support_user_copy(ubq))
913 if (ublk_need_unmap_req(req)) {
914 struct iov_iter iter;
915 const int dir = ITER_SOURCE;
917 WARN_ON_ONCE(io->res > rq_bytes);
919 import_ubuf(dir, u64_to_user_ptr(io->addr), io->res, &iter);
920 return ublk_copy_user_pages(req, 0, &iter, dir);
925 static inline unsigned int ublk_req_build_flags(struct request *req)
929 if (req->cmd_flags & REQ_FAILFAST_DEV)
930 flags |= UBLK_IO_F_FAILFAST_DEV;
932 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
933 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
935 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
936 flags |= UBLK_IO_F_FAILFAST_DRIVER;
938 if (req->cmd_flags & REQ_META)
939 flags |= UBLK_IO_F_META;
941 if (req->cmd_flags & REQ_FUA)
942 flags |= UBLK_IO_F_FUA;
944 if (req->cmd_flags & REQ_NOUNMAP)
945 flags |= UBLK_IO_F_NOUNMAP;
947 if (req->cmd_flags & REQ_SWAP)
948 flags |= UBLK_IO_F_SWAP;
953 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
955 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
956 struct ublk_io *io = &ubq->ios[req->tag];
957 enum req_op op = req_op(req);
960 if (!ublk_queue_is_zoned(ubq) &&
961 (op_is_zone_mgmt(op) || op == REQ_OP_ZONE_APPEND))
962 return BLK_STS_IOERR;
964 switch (req_op(req)) {
966 ublk_op = UBLK_IO_OP_READ;
969 ublk_op = UBLK_IO_OP_WRITE;
972 ublk_op = UBLK_IO_OP_FLUSH;
975 ublk_op = UBLK_IO_OP_DISCARD;
977 case REQ_OP_WRITE_ZEROES:
978 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
981 if (ublk_queue_is_zoned(ubq))
982 return ublk_setup_iod_zoned(ubq, req);
983 return BLK_STS_IOERR;
986 /* need to translate since kernel may change */
987 iod->op_flags = ublk_op | ublk_req_build_flags(req);
988 iod->nr_sectors = blk_rq_sectors(req);
989 iod->start_sector = blk_rq_pos(req);
990 iod->addr = io->addr;
995 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
996 struct io_uring_cmd *ioucmd)
998 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
1001 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
1003 return ubq->ubq_daemon->flags & PF_EXITING;
1006 /* todo: handle partial completion */
1007 static inline void __ublk_complete_rq(struct request *req)
1009 struct ublk_queue *ubq = req->mq_hctx->driver_data;
1010 struct ublk_io *io = &ubq->ios[req->tag];
1011 unsigned int unmapped_bytes;
1012 blk_status_t res = BLK_STS_OK;
1014 /* called from ublk_abort_queue() code path */
1015 if (io->flags & UBLK_IO_FLAG_ABORTED) {
1016 res = BLK_STS_IOERR;
1020 /* failed read IO if nothing is read */
1021 if (!io->res && req_op(req) == REQ_OP_READ)
1025 res = errno_to_blk_status(io->res);
1030 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
1033 * Both the two needn't unmap.
1035 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE &&
1036 req_op(req) != REQ_OP_DRV_IN)
1039 /* for READ request, writing data in iod->addr to rq buffers */
1040 unmapped_bytes = ublk_unmap_io(ubq, req, io);
1043 * Extremely impossible since we got data filled in just before
1045 * Re-read simply for this unlikely case.
1047 if (unlikely(unmapped_bytes < io->res))
1048 io->res = unmapped_bytes;
1050 if (blk_update_request(req, BLK_STS_OK, io->res))
1051 blk_mq_requeue_request(req, true);
1053 __blk_mq_end_request(req, BLK_STS_OK);
1057 blk_mq_end_request(req, res);
1060 static void ublk_complete_rq(struct kref *ref)
1062 struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
1064 struct request *req = blk_mq_rq_from_pdu(data);
1066 __ublk_complete_rq(req);
1070 * Since __ublk_rq_task_work always fails requests immediately during
1071 * exiting, __ublk_fail_req() is only called from abort context during
1072 * exiting. So lock is unnecessary.
1074 * Also aborting may not be started yet, keep in mind that one failed
1075 * request may be issued by block layer again.
1077 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
1078 struct request *req)
1080 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
1082 if (ublk_queue_can_use_recovery_reissue(ubq))
1083 blk_mq_requeue_request(req, false);
1085 ublk_put_req_ref(ubq, req);
1088 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
1089 unsigned issue_flags)
1091 /* mark this cmd owned by ublksrv */
1092 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
1095 * clear ACTIVE since we are done with this sqe/cmd slot
1096 * We can only accept io cmd in case of being not active.
1098 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
1100 /* tell ublksrv one io request is coming */
1101 io_uring_cmd_done(io->cmd, res, 0, issue_flags);
1104 #define UBLK_REQUEUE_DELAY_MS 3
1106 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
1109 /* We cannot process this rq so just requeue it. */
1110 if (ublk_queue_can_use_recovery(ubq))
1111 blk_mq_requeue_request(rq, false);
1113 blk_mq_end_request(rq, BLK_STS_IOERR);
1116 static inline void __ublk_rq_task_work(struct request *req,
1117 unsigned issue_flags)
1119 struct ublk_queue *ubq = req->mq_hctx->driver_data;
1121 struct ublk_io *io = &ubq->ios[tag];
1122 unsigned int mapped_bytes;
1124 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
1125 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1126 ublk_get_iod(ubq, req->tag)->addr);
1129 * Task is exiting if either:
1131 * (1) current != ubq_daemon.
1132 * io_uring_cmd_complete_in_task() tries to run task_work
1133 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
1135 * (2) current->flags & PF_EXITING.
1137 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
1138 __ublk_abort_rq(ubq, req);
1142 if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
1144 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
1145 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
1148 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
1149 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
1150 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
1151 __func__, io->cmd->cmd_op, ubq->q_id,
1152 req->tag, io->flags);
1153 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
1157 * We have handled UBLK_IO_NEED_GET_DATA command,
1158 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
1161 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
1162 /* update iod->addr because ublksrv may have passed a new io buffer */
1163 ublk_get_iod(ubq, req->tag)->addr = io->addr;
1164 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
1165 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1166 ublk_get_iod(ubq, req->tag)->addr);
1169 mapped_bytes = ublk_map_io(ubq, req, io);
1171 /* partially mapped, update io descriptor */
1172 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
1174 * Nothing mapped, retry until we succeed.
1176 * We may never succeed in mapping any bytes here because
1177 * of OOM. TODO: reserve one buffer with single page pinned
1178 * for providing forward progress guarantee.
1180 if (unlikely(!mapped_bytes)) {
1181 blk_mq_requeue_request(req, false);
1182 blk_mq_delay_kick_requeue_list(req->q,
1183 UBLK_REQUEUE_DELAY_MS);
1187 ublk_get_iod(ubq, req->tag)->nr_sectors =
1191 ublk_init_req_ref(ubq, req);
1192 ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
1195 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
1196 unsigned issue_flags)
1198 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1199 struct ublk_rq_data *data, *tmp;
1201 io_cmds = llist_reverse_order(io_cmds);
1202 llist_for_each_entry_safe(data, tmp, io_cmds, node)
1203 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
1206 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1208 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1209 struct ublk_queue *ubq = pdu->ubq;
1211 ublk_forward_io_cmds(ubq, issue_flags);
1214 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
1216 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
1218 if (llist_add(&data->node, &ubq->io_cmds)) {
1219 struct ublk_io *io = &ubq->ios[rq->tag];
1221 io_uring_cmd_complete_in_task(io->cmd, ublk_rq_task_work_cb);
1225 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
1227 struct ublk_queue *ubq = rq->mq_hctx->driver_data;
1228 unsigned int nr_inflight = 0;
1231 if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
1232 if (!ubq->timeout) {
1233 send_sig(SIGKILL, ubq->ubq_daemon, 0);
1234 ubq->timeout = true;
1240 if (!ubq_daemon_is_dying(ubq))
1241 return BLK_EH_RESET_TIMER;
1243 for (i = 0; i < ubq->q_depth; i++) {
1244 struct ublk_io *io = &ubq->ios[i];
1246 if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1250 /* cancelable uring_cmd can't help us if all commands are in-flight */
1251 if (nr_inflight == ubq->q_depth) {
1252 struct ublk_device *ub = ubq->dev;
1254 if (ublk_abort_requests(ub, ubq)) {
1255 if (ublk_can_use_recovery(ub))
1256 schedule_work(&ub->quiesce_work);
1258 schedule_work(&ub->stop_work);
1263 return BLK_EH_RESET_TIMER;
1266 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
1267 const struct blk_mq_queue_data *bd)
1269 struct ublk_queue *ubq = hctx->driver_data;
1270 struct request *rq = bd->rq;
1273 /* fill iod to slot in io cmd buffer */
1274 res = ublk_setup_iod(ubq, rq);
1275 if (unlikely(res != BLK_STS_OK))
1276 return BLK_STS_IOERR;
1278 /* With recovery feature enabled, force_abort is set in
1279 * ublk_stop_dev() before calling del_gendisk(). We have to
1280 * abort all requeued and new rqs here to let del_gendisk()
1281 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1282 * to avoid UAF on io_uring ctx.
1284 * Note: force_abort is guaranteed to be seen because it is set
1285 * before request queue is unqiuesced.
1287 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1288 return BLK_STS_IOERR;
1290 if (unlikely(ubq->canceling)) {
1291 __ublk_abort_rq(ubq, rq);
1295 blk_mq_start_request(bd->rq);
1296 ublk_queue_cmd(ubq, rq);
1301 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1302 unsigned int hctx_idx)
1304 struct ublk_device *ub = driver_data;
1305 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1307 hctx->driver_data = ubq;
1311 static const struct blk_mq_ops ublk_mq_ops = {
1312 .queue_rq = ublk_queue_rq,
1313 .init_hctx = ublk_init_hctx,
1314 .timeout = ublk_timeout,
1317 static int ublk_ch_open(struct inode *inode, struct file *filp)
1319 struct ublk_device *ub = container_of(inode->i_cdev,
1320 struct ublk_device, cdev);
1322 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1324 filp->private_data = ub;
1328 static int ublk_ch_release(struct inode *inode, struct file *filp)
1330 struct ublk_device *ub = filp->private_data;
1332 clear_bit(UB_STATE_OPEN, &ub->state);
1336 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1337 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1339 struct ublk_device *ub = filp->private_data;
1340 size_t sz = vma->vm_end - vma->vm_start;
1341 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1342 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1345 spin_lock(&ub->lock);
1347 ub->mm = current->mm;
1348 if (current->mm != ub->mm)
1350 spin_unlock(&ub->lock);
1355 if (vma->vm_flags & VM_WRITE)
1358 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1359 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1362 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1363 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1364 __func__, q_id, current->pid, vma->vm_start,
1365 phys_off, (unsigned long)sz);
1367 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1370 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1371 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1374 static void ublk_commit_completion(struct ublk_device *ub,
1375 const struct ublksrv_io_cmd *ub_cmd)
1377 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1378 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1379 struct ublk_io *io = &ubq->ios[tag];
1380 struct request *req;
1382 /* now this cmd slot is owned by nbd driver */
1383 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1384 io->res = ub_cmd->result;
1386 /* find the io request and complete */
1387 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1388 if (WARN_ON_ONCE(unlikely(!req)))
1391 if (req_op(req) == REQ_OP_ZONE_APPEND)
1392 req->__sector = ub_cmd->zone_append_lba;
1394 if (likely(!blk_should_fake_timeout(req->q)))
1395 ublk_put_req_ref(ubq, req);
1399 * Called from ubq_daemon context via cancel fn, meantime quiesce ublk
1400 * blk-mq queue, so we are called exclusively with blk-mq and ubq_daemon
1401 * context, so everything is serialized.
1403 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1407 for (i = 0; i < ubq->q_depth; i++) {
1408 struct ublk_io *io = &ubq->ios[i];
1410 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1414 * Either we fail the request or ublk_rq_task_work_fn
1417 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1418 if (rq && blk_mq_request_started(rq)) {
1419 io->flags |= UBLK_IO_FLAG_ABORTED;
1420 __ublk_fail_req(ubq, io, rq);
1426 static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq)
1428 struct gendisk *disk;
1430 spin_lock(&ubq->cancel_lock);
1431 if (ubq->canceling) {
1432 spin_unlock(&ubq->cancel_lock);
1435 ubq->canceling = true;
1436 spin_unlock(&ubq->cancel_lock);
1438 spin_lock(&ub->lock);
1441 get_device(disk_to_dev(disk));
1442 spin_unlock(&ub->lock);
1444 /* Our disk has been dead */
1448 /* Now we are serialized with ublk_queue_rq() */
1449 blk_mq_quiesce_queue(disk->queue);
1450 /* abort queue is for making forward progress */
1451 ublk_abort_queue(ub, ubq);
1452 blk_mq_unquiesce_queue(disk->queue);
1453 put_device(disk_to_dev(disk));
1458 static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io,
1459 unsigned int issue_flags)
1463 if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1466 spin_lock(&ubq->cancel_lock);
1467 done = !!(io->flags & UBLK_IO_FLAG_CANCELED);
1469 io->flags |= UBLK_IO_FLAG_CANCELED;
1470 spin_unlock(&ubq->cancel_lock);
1473 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1477 * The ublk char device won't be closed when calling cancel fn, so both
1478 * ublk device and queue are guaranteed to be live
1480 static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd,
1481 unsigned int issue_flags)
1483 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1484 struct ublk_queue *ubq = pdu->ubq;
1485 struct task_struct *task;
1486 struct ublk_device *ub;
1490 if (WARN_ON_ONCE(!ubq))
1493 if (WARN_ON_ONCE(pdu->tag >= ubq->q_depth))
1496 task = io_uring_cmd_get_task(cmd);
1497 if (WARN_ON_ONCE(task && task != ubq->ubq_daemon))
1501 need_schedule = ublk_abort_requests(ub, ubq);
1503 io = &ubq->ios[pdu->tag];
1504 WARN_ON_ONCE(io->cmd != cmd);
1505 ublk_cancel_cmd(ubq, io, issue_flags);
1507 if (need_schedule) {
1508 if (ublk_can_use_recovery(ub))
1509 schedule_work(&ub->quiesce_work);
1511 schedule_work(&ub->stop_work);
1515 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1517 return ubq->nr_io_ready == ubq->q_depth;
1520 static void ublk_cancel_queue(struct ublk_queue *ubq)
1524 for (i = 0; i < ubq->q_depth; i++)
1525 ublk_cancel_cmd(ubq, &ubq->ios[i], IO_URING_F_UNLOCKED);
1528 /* Cancel all pending commands, must be called after del_gendisk() returns */
1529 static void ublk_cancel_dev(struct ublk_device *ub)
1533 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1534 ublk_cancel_queue(ublk_get_queue(ub, i));
1537 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1541 if (blk_mq_request_started(rq)) {
1548 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1552 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1555 blk_mq_tagset_busy_iter(&ub->tag_set,
1556 ublk_check_inflight_rq, &idle);
1559 msleep(UBLK_REQUEUE_DELAY_MS);
1563 static void __ublk_quiesce_dev(struct ublk_device *ub)
1565 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1566 __func__, ub->dev_info.dev_id,
1567 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1568 "LIVE" : "QUIESCED");
1569 blk_mq_quiesce_queue(ub->ub_disk->queue);
1570 ublk_wait_tagset_rqs_idle(ub);
1571 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1574 static void ublk_quiesce_work_fn(struct work_struct *work)
1576 struct ublk_device *ub =
1577 container_of(work, struct ublk_device, quiesce_work);
1579 mutex_lock(&ub->mutex);
1580 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1582 __ublk_quiesce_dev(ub);
1584 mutex_unlock(&ub->mutex);
1585 ublk_cancel_dev(ub);
1588 static void ublk_unquiesce_dev(struct ublk_device *ub)
1592 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1593 __func__, ub->dev_info.dev_id,
1594 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1595 "LIVE" : "QUIESCED");
1596 /* quiesce_work has run. We let requeued rqs be aborted
1597 * before running fallback_wq. "force_abort" must be seen
1598 * after request queue is unqiuesced. Then del_gendisk()
1601 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1602 ublk_get_queue(ub, i)->force_abort = true;
1604 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1605 /* We may have requeued some rqs in ublk_quiesce_queue() */
1606 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1609 static void ublk_stop_dev(struct ublk_device *ub)
1611 struct gendisk *disk;
1613 mutex_lock(&ub->mutex);
1614 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1616 if (ublk_can_use_recovery(ub)) {
1617 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1618 __ublk_quiesce_dev(ub);
1619 ublk_unquiesce_dev(ub);
1621 del_gendisk(ub->ub_disk);
1623 /* Sync with ublk_abort_queue() by holding the lock */
1624 spin_lock(&ub->lock);
1626 ub->dev_info.state = UBLK_S_DEV_DEAD;
1627 ub->dev_info.ublksrv_pid = -1;
1629 spin_unlock(&ub->lock);
1632 mutex_unlock(&ub->mutex);
1633 ublk_cancel_dev(ub);
1636 /* device can only be started after all IOs are ready */
1637 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1639 mutex_lock(&ub->mutex);
1641 if (ublk_queue_ready(ubq)) {
1642 ubq->ubq_daemon = current;
1643 get_task_struct(ubq->ubq_daemon);
1644 ub->nr_queues_ready++;
1646 if (capable(CAP_SYS_ADMIN))
1647 ub->nr_privileged_daemon++;
1649 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1650 complete_all(&ub->completion);
1651 mutex_unlock(&ub->mutex);
1654 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1657 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1658 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1660 ublk_queue_cmd(ubq, req);
1663 static inline int ublk_check_cmd_op(u32 cmd_op)
1665 u32 ioc_type = _IOC_TYPE(cmd_op);
1667 if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1670 if (ioc_type != 'u' && ioc_type != 0)
1676 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1677 struct io_uring_cmd *cmd, unsigned long buf_addr)
1680 io->flags |= UBLK_IO_FLAG_ACTIVE;
1681 io->addr = buf_addr;
1684 static inline void ublk_prep_cancel(struct io_uring_cmd *cmd,
1685 unsigned int issue_flags,
1686 struct ublk_queue *ubq, unsigned int tag)
1688 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1691 * Safe to refer to @ubq since ublk_queue won't be died until its
1692 * commands are completed
1696 io_uring_cmd_mark_cancelable(cmd, issue_flags);
1699 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1700 unsigned int issue_flags,
1701 const struct ublksrv_io_cmd *ub_cmd)
1703 struct ublk_device *ub = cmd->file->private_data;
1704 struct ublk_queue *ubq;
1706 u32 cmd_op = cmd->cmd_op;
1707 unsigned tag = ub_cmd->tag;
1709 struct request *req;
1711 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1712 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1715 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1718 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1719 if (!ubq || ub_cmd->q_id != ubq->q_id)
1722 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1725 if (tag >= ubq->q_depth)
1728 io = &ubq->ios[tag];
1730 /* there is pending io cmd, something must be wrong */
1731 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1737 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1738 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1740 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1741 ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1744 ret = ublk_check_cmd_op(cmd_op);
1749 switch (_IOC_NR(cmd_op)) {
1750 case UBLK_IO_FETCH_REQ:
1751 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1752 if (ublk_queue_ready(ubq)) {
1757 * The io is being handled by server, so COMMIT_RQ is expected
1758 * instead of FETCH_REQ
1760 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1763 if (!ublk_support_user_copy(ubq)) {
1765 * FETCH_RQ has to provide IO buffer if NEED GET
1766 * DATA is not enabled
1768 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1770 } else if (ub_cmd->addr) {
1771 /* User copy requires addr to be unset */
1776 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1777 ublk_mark_io_ready(ub, ubq);
1779 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1780 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1782 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1785 if (!ublk_support_user_copy(ubq)) {
1787 * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1788 * NEED GET DATA is not enabled or it is Read IO.
1790 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1791 req_op(req) == REQ_OP_READ))
1793 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1795 * User copy requires addr to be unset when command is
1802 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1803 ublk_commit_completion(ub, ub_cmd);
1805 case UBLK_IO_NEED_GET_DATA:
1806 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1808 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1809 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1814 ublk_prep_cancel(cmd, issue_flags, ubq, tag);
1815 return -EIOCBQUEUED;
1818 io_uring_cmd_done(cmd, ret, 0, issue_flags);
1819 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1820 __func__, cmd_op, tag, ret, io->flags);
1821 return -EIOCBQUEUED;
1824 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1825 struct ublk_queue *ubq, int tag, size_t offset)
1827 struct request *req;
1829 if (!ublk_need_req_ref(ubq))
1832 req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1836 if (!ublk_get_req_ref(ubq, req))
1839 if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1842 if (!ublk_rq_has_data(req))
1845 if (offset > blk_rq_bytes(req))
1850 ublk_put_req_ref(ubq, req);
1854 static inline int ublk_ch_uring_cmd_local(struct io_uring_cmd *cmd,
1855 unsigned int issue_flags)
1858 * Not necessary for async retry, but let's keep it simple and always
1859 * copy the values to avoid any potential reuse.
1861 const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1862 const struct ublksrv_io_cmd ub_cmd = {
1863 .q_id = READ_ONCE(ub_src->q_id),
1864 .tag = READ_ONCE(ub_src->tag),
1865 .result = READ_ONCE(ub_src->result),
1866 .addr = READ_ONCE(ub_src->addr)
1869 WARN_ON_ONCE(issue_flags & IO_URING_F_UNLOCKED);
1871 return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1874 static void ublk_ch_uring_cmd_cb(struct io_uring_cmd *cmd,
1875 unsigned int issue_flags)
1877 ublk_ch_uring_cmd_local(cmd, issue_flags);
1880 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1882 if (unlikely(issue_flags & IO_URING_F_CANCEL)) {
1883 ublk_uring_cmd_cancel_fn(cmd, issue_flags);
1887 /* well-implemented server won't run into unlocked */
1888 if (unlikely(issue_flags & IO_URING_F_UNLOCKED)) {
1889 io_uring_cmd_complete_in_task(cmd, ublk_ch_uring_cmd_cb);
1890 return -EIOCBQUEUED;
1893 return ublk_ch_uring_cmd_local(cmd, issue_flags);
1896 static inline bool ublk_check_ubuf_dir(const struct request *req,
1899 /* copy ubuf to request pages */
1900 if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1901 ubuf_dir == ITER_SOURCE)
1904 /* copy request pages to ubuf */
1905 if ((req_op(req) == REQ_OP_WRITE ||
1906 req_op(req) == REQ_OP_ZONE_APPEND) &&
1907 ubuf_dir == ITER_DEST)
1913 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1914 struct iov_iter *iter, size_t *off, int dir)
1916 struct ublk_device *ub = iocb->ki_filp->private_data;
1917 struct ublk_queue *ubq;
1918 struct request *req;
1923 return ERR_PTR(-EACCES);
1925 if (!user_backed_iter(iter))
1926 return ERR_PTR(-EACCES);
1928 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1929 return ERR_PTR(-EACCES);
1931 tag = ublk_pos_to_tag(iocb->ki_pos);
1932 q_id = ublk_pos_to_hwq(iocb->ki_pos);
1933 buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1935 if (q_id >= ub->dev_info.nr_hw_queues)
1936 return ERR_PTR(-EINVAL);
1938 ubq = ublk_get_queue(ub, q_id);
1940 return ERR_PTR(-EINVAL);
1942 if (tag >= ubq->q_depth)
1943 return ERR_PTR(-EINVAL);
1945 req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1947 return ERR_PTR(-EINVAL);
1949 if (!req->mq_hctx || !req->mq_hctx->driver_data)
1952 if (!ublk_check_ubuf_dir(req, dir))
1958 ublk_put_req_ref(ubq, req);
1959 return ERR_PTR(-EACCES);
1962 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1964 struct ublk_queue *ubq;
1965 struct request *req;
1969 req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1971 return PTR_ERR(req);
1973 ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1974 ubq = req->mq_hctx->driver_data;
1975 ublk_put_req_ref(ubq, req);
1980 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1982 struct ublk_queue *ubq;
1983 struct request *req;
1987 req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1989 return PTR_ERR(req);
1991 ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1992 ubq = req->mq_hctx->driver_data;
1993 ublk_put_req_ref(ubq, req);
1998 static const struct file_operations ublk_ch_fops = {
1999 .owner = THIS_MODULE,
2000 .open = ublk_ch_open,
2001 .release = ublk_ch_release,
2002 .llseek = no_llseek,
2003 .read_iter = ublk_ch_read_iter,
2004 .write_iter = ublk_ch_write_iter,
2005 .uring_cmd = ublk_ch_uring_cmd,
2006 .mmap = ublk_ch_mmap,
2009 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
2011 int size = ublk_queue_cmd_buf_size(ub, q_id);
2012 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
2014 if (ubq->ubq_daemon)
2015 put_task_struct(ubq->ubq_daemon);
2016 if (ubq->io_cmd_buf)
2017 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
2020 static int ublk_init_queue(struct ublk_device *ub, int q_id)
2022 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
2023 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
2027 spin_lock_init(&ubq->cancel_lock);
2028 ubq->flags = ub->dev_info.flags;
2030 ubq->q_depth = ub->dev_info.queue_depth;
2031 size = ublk_queue_cmd_buf_size(ub, q_id);
2033 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
2037 ubq->io_cmd_buf = ptr;
2042 static void ublk_deinit_queues(struct ublk_device *ub)
2044 int nr_queues = ub->dev_info.nr_hw_queues;
2050 for (i = 0; i < nr_queues; i++)
2051 ublk_deinit_queue(ub, i);
2052 kfree(ub->__queues);
2055 static int ublk_init_queues(struct ublk_device *ub)
2057 int nr_queues = ub->dev_info.nr_hw_queues;
2058 int depth = ub->dev_info.queue_depth;
2059 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
2060 int i, ret = -ENOMEM;
2062 ub->queue_size = ubq_size;
2063 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
2067 for (i = 0; i < nr_queues; i++) {
2068 if (ublk_init_queue(ub, i))
2072 init_completion(&ub->completion);
2076 ublk_deinit_queues(ub);
2080 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2085 spin_lock(&ublk_idr_lock);
2086 /* allocate id, if @id >= 0, we're requesting that specific id */
2088 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2092 err = idr_alloc(&ublk_index_idr, ub, 0, UBLK_MAX_UBLKS,
2095 spin_unlock(&ublk_idr_lock);
2098 ub->ub_number = err;
2103 static void ublk_free_dev_number(struct ublk_device *ub)
2105 spin_lock(&ublk_idr_lock);
2106 idr_remove(&ublk_index_idr, ub->ub_number);
2107 wake_up_all(&ublk_idr_wq);
2108 spin_unlock(&ublk_idr_lock);
2111 static void ublk_cdev_rel(struct device *dev)
2113 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2115 blk_mq_free_tag_set(&ub->tag_set);
2116 ublk_deinit_queues(ub);
2117 ublk_free_dev_number(ub);
2118 mutex_destroy(&ub->mutex);
2122 static int ublk_add_chdev(struct ublk_device *ub)
2124 struct device *dev = &ub->cdev_dev;
2125 int minor = ub->ub_number;
2128 dev->parent = ublk_misc.this_device;
2129 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2130 dev->class = &ublk_chr_class;
2131 dev->release = ublk_cdev_rel;
2132 device_initialize(dev);
2134 ret = dev_set_name(dev, "ublkc%d", minor);
2138 cdev_init(&ub->cdev, &ublk_ch_fops);
2139 ret = cdev_device_add(&ub->cdev, dev);
2150 static void ublk_stop_work_fn(struct work_struct *work)
2152 struct ublk_device *ub =
2153 container_of(work, struct ublk_device, stop_work);
2158 /* align max io buffer size with PAGE_SIZE */
2159 static void ublk_align_max_io_size(struct ublk_device *ub)
2161 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2163 ub->dev_info.max_io_buf_bytes =
2164 round_down(max_io_bytes, PAGE_SIZE);
2167 static int ublk_add_tag_set(struct ublk_device *ub)
2169 ub->tag_set.ops = &ublk_mq_ops;
2170 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2171 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2172 ub->tag_set.numa_node = NUMA_NO_NODE;
2173 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2174 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2175 ub->tag_set.driver_data = ub;
2176 return blk_mq_alloc_tag_set(&ub->tag_set);
2179 static void ublk_remove(struct ublk_device *ub)
2182 cancel_work_sync(&ub->stop_work);
2183 cancel_work_sync(&ub->quiesce_work);
2184 cdev_device_del(&ub->cdev, &ub->cdev_dev);
2185 put_device(&ub->cdev_dev);
2189 static struct ublk_device *ublk_get_device_from_id(int idx)
2191 struct ublk_device *ub = NULL;
2196 spin_lock(&ublk_idr_lock);
2197 ub = idr_find(&ublk_index_idr, idx);
2199 ub = ublk_get_device(ub);
2200 spin_unlock(&ublk_idr_lock);
2205 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2207 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2208 int ublksrv_pid = (int)header->data[0];
2209 struct gendisk *disk;
2212 if (ublksrv_pid <= 0)
2215 if (wait_for_completion_interruptible(&ub->completion) != 0)
2218 mutex_lock(&ub->mutex);
2219 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2220 test_bit(UB_STATE_USED, &ub->state)) {
2225 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
2227 ret = PTR_ERR(disk);
2230 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2231 disk->fops = &ub_fops;
2232 disk->private_data = ub;
2234 ub->dev_info.ublksrv_pid = ublksrv_pid;
2237 ret = ublk_apply_params(ub);
2241 /* don't probe partitions if any one ubq daemon is un-trusted */
2242 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2243 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2245 get_device(&ub->cdev_dev);
2246 ub->dev_info.state = UBLK_S_DEV_LIVE;
2248 if (ublk_dev_is_zoned(ub)) {
2249 ret = ublk_revalidate_disk_zones(ub);
2254 ret = add_disk(disk);
2258 set_bit(UB_STATE_USED, &ub->state);
2262 ub->dev_info.state = UBLK_S_DEV_DEAD;
2263 ublk_put_device(ub);
2269 mutex_unlock(&ub->mutex);
2273 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2274 struct io_uring_cmd *cmd)
2276 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2277 void __user *argp = (void __user *)(unsigned long)header->addr;
2278 cpumask_var_t cpumask;
2279 unsigned long queue;
2280 unsigned int retlen;
2284 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2286 if (header->len & (sizeof(unsigned long)-1))
2291 queue = header->data[0];
2292 if (queue >= ub->dev_info.nr_hw_queues)
2295 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2298 for_each_possible_cpu(i) {
2299 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2300 cpumask_set_cpu(i, cpumask);
2304 retlen = min_t(unsigned short, header->len, cpumask_size());
2305 if (copy_to_user(argp, cpumask, retlen))
2306 goto out_free_cpumask;
2307 if (retlen != header->len &&
2308 clear_user(argp + retlen, header->len - retlen))
2309 goto out_free_cpumask;
2313 free_cpumask_var(cpumask);
2317 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2319 pr_devel("%s: dev id %d flags %llx\n", __func__,
2320 info->dev_id, info->flags);
2321 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2322 info->nr_hw_queues, info->queue_depth);
2325 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2327 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2328 void __user *argp = (void __user *)(unsigned long)header->addr;
2329 struct ublksrv_ctrl_dev_info info;
2330 struct ublk_device *ub;
2333 if (header->len < sizeof(info) || !header->addr)
2335 if (header->queue_id != (u16)-1) {
2336 pr_warn("%s: queue_id is wrong %x\n",
2337 __func__, header->queue_id);
2341 if (copy_from_user(&info, argp, sizeof(info)))
2344 if (capable(CAP_SYS_ADMIN))
2345 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2346 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2350 * unprivileged device can't be trusted, but RECOVERY and
2351 * RECOVERY_REISSUE still may hang error handling, so can't
2352 * support recovery features for unprivileged ublk now
2354 * TODO: provide forward progress for RECOVERY handler, so that
2355 * unprivileged device can benefit from it
2357 if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2358 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2359 UBLK_F_USER_RECOVERY);
2361 /* the created device is always owned by current user */
2362 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2364 if (header->dev_id != info.dev_id) {
2365 pr_warn("%s: dev id not match %u %u\n",
2366 __func__, header->dev_id, info.dev_id);
2370 if (header->dev_id != U32_MAX && header->dev_id >= UBLK_MAX_UBLKS) {
2371 pr_warn("%s: dev id is too large. Max supported is %d\n",
2372 __func__, UBLK_MAX_UBLKS - 1);
2376 ublk_dump_dev_info(&info);
2378 ret = mutex_lock_killable(&ublk_ctl_mutex);
2383 if (ublks_added >= ublks_max)
2387 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2390 mutex_init(&ub->mutex);
2391 spin_lock_init(&ub->lock);
2392 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2393 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2395 ret = ublk_alloc_dev_number(ub, header->dev_id);
2399 memcpy(&ub->dev_info, &info, sizeof(info));
2401 /* update device id */
2402 ub->dev_info.dev_id = ub->ub_number;
2405 * 64bit flags will be copied back to userspace as feature
2406 * negotiation result, so have to clear flags which driver
2407 * doesn't support yet, then userspace can get correct flags
2408 * (features) to handle.
2410 ub->dev_info.flags &= UBLK_F_ALL;
2412 ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2413 UBLK_F_URING_CMD_COMP_IN_TASK;
2415 /* GET_DATA isn't needed any more with USER_COPY */
2416 if (ublk_dev_is_user_copy(ub))
2417 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2419 /* Zoned storage support requires user copy feature */
2420 if (ublk_dev_is_zoned(ub) &&
2421 (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2423 goto out_free_dev_number;
2426 /* We are not ready to support zero copy */
2427 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2429 ub->dev_info.nr_hw_queues = min_t(unsigned int,
2430 ub->dev_info.nr_hw_queues, nr_cpu_ids);
2431 ublk_align_max_io_size(ub);
2433 ret = ublk_init_queues(ub);
2435 goto out_free_dev_number;
2437 ret = ublk_add_tag_set(ub);
2439 goto out_deinit_queues;
2442 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2443 goto out_free_tag_set;
2446 * Add the char dev so that ublksrv daemon can be setup.
2447 * ublk_add_chdev() will cleanup everything if it fails.
2449 ret = ublk_add_chdev(ub);
2453 blk_mq_free_tag_set(&ub->tag_set);
2455 ublk_deinit_queues(ub);
2456 out_free_dev_number:
2457 ublk_free_dev_number(ub);
2459 mutex_destroy(&ub->mutex);
2462 mutex_unlock(&ublk_ctl_mutex);
2466 static inline bool ublk_idr_freed(int id)
2470 spin_lock(&ublk_idr_lock);
2471 ptr = idr_find(&ublk_index_idr, id);
2472 spin_unlock(&ublk_idr_lock);
2477 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2479 struct ublk_device *ub = *p_ub;
2480 int idx = ub->ub_number;
2483 ret = mutex_lock_killable(&ublk_ctl_mutex);
2487 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2489 set_bit(UB_STATE_DELETED, &ub->state);
2492 /* Mark the reference as consumed */
2494 ublk_put_device(ub);
2495 mutex_unlock(&ublk_ctl_mutex);
2498 * Wait until the idr is removed, then it can be reused after
2499 * DEL_DEV command is returned.
2501 * If we returns because of user interrupt, future delete command
2504 * - the device number isn't freed, this device won't or needn't
2505 * be deleted again, since UB_STATE_DELETED is set, and device
2506 * will be released after the last reference is dropped
2508 * - the device number is freed already, we will not find this
2509 * device via ublk_get_device_from_id()
2511 if (wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx)))
2516 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2518 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2520 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2521 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2522 header->data[0], header->addr, header->len);
2525 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2528 cancel_work_sync(&ub->stop_work);
2529 cancel_work_sync(&ub->quiesce_work);
2534 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2535 struct io_uring_cmd *cmd)
2537 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2538 void __user *argp = (void __user *)(unsigned long)header->addr;
2540 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2543 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2549 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2550 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2552 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2553 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2556 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2557 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2559 ub->params.devt.disk_major = 0;
2560 ub->params.devt.disk_minor = 0;
2562 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2565 static int ublk_ctrl_get_params(struct ublk_device *ub,
2566 struct io_uring_cmd *cmd)
2568 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2569 void __user *argp = (void __user *)(unsigned long)header->addr;
2570 struct ublk_params_header ph;
2573 if (header->len <= sizeof(ph) || !header->addr)
2576 if (copy_from_user(&ph, argp, sizeof(ph)))
2579 if (ph.len > header->len || !ph.len)
2582 if (ph.len > sizeof(struct ublk_params))
2583 ph.len = sizeof(struct ublk_params);
2585 mutex_lock(&ub->mutex);
2586 ublk_ctrl_fill_params_devt(ub);
2587 if (copy_to_user(argp, &ub->params, ph.len))
2591 mutex_unlock(&ub->mutex);
2596 static int ublk_ctrl_set_params(struct ublk_device *ub,
2597 struct io_uring_cmd *cmd)
2599 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2600 void __user *argp = (void __user *)(unsigned long)header->addr;
2601 struct ublk_params_header ph;
2604 if (header->len <= sizeof(ph) || !header->addr)
2607 if (copy_from_user(&ph, argp, sizeof(ph)))
2610 if (ph.len > header->len || !ph.len || !ph.types)
2613 if (ph.len > sizeof(struct ublk_params))
2614 ph.len = sizeof(struct ublk_params);
2616 /* parameters can only be changed when device isn't live */
2617 mutex_lock(&ub->mutex);
2618 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2620 } else if (copy_from_user(&ub->params, argp, ph.len)) {
2623 /* clear all we don't support yet */
2624 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2625 ret = ublk_validate_params(ub);
2627 ub->params.types = 0;
2629 mutex_unlock(&ub->mutex);
2634 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2638 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2640 /* All old ioucmds have to be completed */
2641 ubq->nr_io_ready = 0;
2642 /* old daemon is PF_EXITING, put it now */
2643 put_task_struct(ubq->ubq_daemon);
2644 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2645 ubq->ubq_daemon = NULL;
2646 ubq->timeout = false;
2647 ubq->canceling = false;
2649 for (i = 0; i < ubq->q_depth; i++) {
2650 struct ublk_io *io = &ubq->ios[i];
2652 /* forget everything now and be ready for new FETCH_REQ */
2659 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2660 struct io_uring_cmd *cmd)
2662 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2666 mutex_lock(&ub->mutex);
2667 if (!ublk_can_use_recovery(ub))
2670 * START_RECOVERY is only allowd after:
2672 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2673 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
2676 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2677 * (a)has quiesced request queue
2678 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
2679 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2680 * (d)has completed/camceled all ioucmds owned by ther dying process
2682 if (test_bit(UB_STATE_OPEN, &ub->state) ||
2683 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2687 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2688 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2689 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2690 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2692 ub->nr_queues_ready = 0;
2693 ub->nr_privileged_daemon = 0;
2694 init_completion(&ub->completion);
2697 mutex_unlock(&ub->mutex);
2701 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2702 struct io_uring_cmd *cmd)
2704 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2705 int ublksrv_pid = (int)header->data[0];
2708 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2709 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2710 /* wait until new ubq_daemon sending all FETCH_REQ */
2711 if (wait_for_completion_interruptible(&ub->completion))
2714 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2715 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2717 mutex_lock(&ub->mutex);
2718 if (!ublk_can_use_recovery(ub))
2721 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2725 ub->dev_info.ublksrv_pid = ublksrv_pid;
2726 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2727 __func__, ublksrv_pid, header->dev_id);
2728 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2729 pr_devel("%s: queue unquiesced, dev id %d.\n",
2730 __func__, header->dev_id);
2731 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2732 ub->dev_info.state = UBLK_S_DEV_LIVE;
2735 mutex_unlock(&ub->mutex);
2739 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2741 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2742 void __user *argp = (void __user *)(unsigned long)header->addr;
2743 u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2745 if (header->len != UBLK_FEATURES_LEN || !header->addr)
2748 if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2755 * All control commands are sent via /dev/ublk-control, so we have to check
2756 * the destination device's permission
2758 static int ublk_char_dev_permission(struct ublk_device *ub,
2759 const char *dev_path, int mask)
2765 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2769 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2774 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2777 err = inode_permission(&nop_mnt_idmap,
2778 d_backing_inode(path.dentry), mask);
2784 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2785 struct io_uring_cmd *cmd)
2787 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2788 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2789 void __user *argp = (void __user *)(unsigned long)header->addr;
2790 char *dev_path = NULL;
2794 if (!unprivileged) {
2795 if (!capable(CAP_SYS_ADMIN))
2798 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2799 * char_dev_path in payload too, since userspace may not
2800 * know if the specified device is created as unprivileged
2803 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2808 * User has to provide the char device path for unprivileged ublk
2810 * header->addr always points to the dev path buffer, and
2811 * header->dev_path_len records length of dev path buffer.
2813 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2816 if (header->len < header->dev_path_len)
2819 dev_path = memdup_user_nul(argp, header->dev_path_len);
2820 if (IS_ERR(dev_path))
2821 return PTR_ERR(dev_path);
2824 switch (_IOC_NR(cmd->cmd_op)) {
2825 case UBLK_CMD_GET_DEV_INFO:
2826 case UBLK_CMD_GET_DEV_INFO2:
2827 case UBLK_CMD_GET_QUEUE_AFFINITY:
2828 case UBLK_CMD_GET_PARAMS:
2829 case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2832 case UBLK_CMD_START_DEV:
2833 case UBLK_CMD_STOP_DEV:
2834 case UBLK_CMD_ADD_DEV:
2835 case UBLK_CMD_DEL_DEV:
2836 case UBLK_CMD_SET_PARAMS:
2837 case UBLK_CMD_START_USER_RECOVERY:
2838 case UBLK_CMD_END_USER_RECOVERY:
2839 mask = MAY_READ | MAY_WRITE;
2845 ret = ublk_char_dev_permission(ub, dev_path, mask);
2847 header->len -= header->dev_path_len;
2848 header->addr += header->dev_path_len;
2850 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2851 __func__, ub->ub_number, cmd->cmd_op,
2852 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2859 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2860 unsigned int issue_flags)
2862 const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2863 struct ublk_device *ub = NULL;
2864 u32 cmd_op = cmd->cmd_op;
2867 if (issue_flags & IO_URING_F_NONBLOCK)
2870 ublk_ctrl_cmd_dump(cmd);
2872 if (!(issue_flags & IO_URING_F_SQE128))
2875 ret = ublk_check_cmd_op(cmd_op);
2879 if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2880 ret = ublk_ctrl_get_features(cmd);
2884 if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2886 ub = ublk_get_device_from_id(header->dev_id);
2890 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2895 switch (_IOC_NR(cmd_op)) {
2896 case UBLK_CMD_START_DEV:
2897 ret = ublk_ctrl_start_dev(ub, cmd);
2899 case UBLK_CMD_STOP_DEV:
2900 ret = ublk_ctrl_stop_dev(ub);
2902 case UBLK_CMD_GET_DEV_INFO:
2903 case UBLK_CMD_GET_DEV_INFO2:
2904 ret = ublk_ctrl_get_dev_info(ub, cmd);
2906 case UBLK_CMD_ADD_DEV:
2907 ret = ublk_ctrl_add_dev(cmd);
2909 case UBLK_CMD_DEL_DEV:
2910 ret = ublk_ctrl_del_dev(&ub);
2912 case UBLK_CMD_GET_QUEUE_AFFINITY:
2913 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2915 case UBLK_CMD_GET_PARAMS:
2916 ret = ublk_ctrl_get_params(ub, cmd);
2918 case UBLK_CMD_SET_PARAMS:
2919 ret = ublk_ctrl_set_params(ub, cmd);
2921 case UBLK_CMD_START_USER_RECOVERY:
2922 ret = ublk_ctrl_start_recovery(ub, cmd);
2924 case UBLK_CMD_END_USER_RECOVERY:
2925 ret = ublk_ctrl_end_recovery(ub, cmd);
2934 ublk_put_device(ub);
2936 io_uring_cmd_done(cmd, ret, 0, issue_flags);
2937 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2938 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2939 return -EIOCBQUEUED;
2942 static const struct file_operations ublk_ctl_fops = {
2943 .open = nonseekable_open,
2944 .uring_cmd = ublk_ctrl_uring_cmd,
2945 .owner = THIS_MODULE,
2946 .llseek = noop_llseek,
2949 static struct miscdevice ublk_misc = {
2950 .minor = MISC_DYNAMIC_MINOR,
2951 .name = "ublk-control",
2952 .fops = &ublk_ctl_fops,
2955 static int __init ublk_init(void)
2959 BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2960 UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2962 init_waitqueue_head(&ublk_idr_wq);
2964 ret = misc_register(&ublk_misc);
2968 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2970 goto unregister_mis;
2972 ret = class_register(&ublk_chr_class);
2974 goto free_chrdev_region;
2979 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2981 misc_deregister(&ublk_misc);
2985 static void __exit ublk_exit(void)
2987 struct ublk_device *ub;
2990 idr_for_each_entry(&ublk_index_idr, ub, id)
2993 class_unregister(&ublk_chr_class);
2994 misc_deregister(&ublk_misc);
2996 idr_destroy(&ublk_index_idr);
2997 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
3000 module_init(ublk_init);
3001 module_exit(ublk_exit);
3003 static int ublk_set_max_ublks(const char *buf, const struct kernel_param *kp)
3005 return param_set_uint_minmax(buf, kp, 0, UBLK_MAX_UBLKS);
3008 static int ublk_get_max_ublks(char *buf, const struct kernel_param *kp)
3010 return sysfs_emit(buf, "%u\n", ublks_max);
3013 static const struct kernel_param_ops ublk_max_ublks_ops = {
3014 .set = ublk_set_max_ublks,
3015 .get = ublk_get_max_ublks,
3018 module_param_cb(ublks_max, &ublk_max_ublks_ops, &ublks_max, 0644);
3019 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
3021 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
3022 MODULE_LICENSE("GPL");