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 <uapi/linux/ublk_cmd.h>
47 #define UBLK_MINORS (1U << MINORBITS)
49 /* All UBLK_F_* have to be included into UBLK_F_ALL */
50 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
51 | UBLK_F_URING_CMD_COMP_IN_TASK \
52 | UBLK_F_NEED_GET_DATA \
53 | UBLK_F_USER_RECOVERY \
54 | UBLK_F_USER_RECOVERY_REISSUE)
56 /* All UBLK_PARAM_TYPE_* should be included here */
57 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD)
60 struct llist_node node;
61 struct callback_head work;
64 struct ublk_uring_cmd_pdu {
65 struct ublk_queue *ubq;
69 * io command is active: sqe cmd is received, and its cqe isn't done
71 * If the flag is set, the io command is owned by ublk driver, and waited
72 * for incoming blk-mq request from the ublk block device.
74 * If the flag is cleared, the io command will be completed, and owned by
77 #define UBLK_IO_FLAG_ACTIVE 0x01
80 * IO command is completed via cqe, and it is being handled by ublksrv, and
83 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
86 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
89 * IO command is aborted, so this flag is set in case of
90 * !UBLK_IO_FLAG_ACTIVE.
92 * After this flag is observed, any pending or new incoming request
93 * associated with this io command will be failed immediately
95 #define UBLK_IO_FLAG_ABORTED 0x04
98 * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
99 * get data buffer address from ublksrv.
101 * Then, bio data could be copied into this data buffer for a WRITE request
102 * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
104 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
107 /* userspace buffer address from io cmd */
112 struct io_uring_cmd *cmd;
120 struct task_struct *ubq_daemon;
123 struct llist_head io_cmds;
125 unsigned long io_addr; /* mapped vm address */
126 unsigned int max_io_sz;
128 unsigned short nr_io_ready; /* how many ios setup */
129 struct ublk_device *dev;
130 struct ublk_io ios[];
133 #define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
136 struct gendisk *ub_disk;
140 unsigned short queue_size;
141 struct ublksrv_ctrl_dev_info dev_info;
143 struct blk_mq_tag_set tag_set;
146 struct device cdev_dev;
148 #define UB_STATE_OPEN 0
149 #define UB_STATE_USED 1
156 struct mm_struct *mm;
158 struct ublk_params params;
160 struct completion completion;
161 unsigned int nr_queues_ready;
162 atomic_t nr_aborted_queues;
165 * Our ubq->daemon may be killed without any notification, so
166 * monitor each queue's daemon periodically
168 struct delayed_work monitor_work;
169 struct work_struct quiesce_work;
170 struct work_struct stop_work;
173 /* header of ublk_params */
174 struct ublk_params_header {
179 static dev_t ublk_chr_devt;
180 static struct class *ublk_chr_class;
182 static DEFINE_IDR(ublk_index_idr);
183 static DEFINE_SPINLOCK(ublk_idr_lock);
184 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
186 static DEFINE_MUTEX(ublk_ctl_mutex);
188 static struct miscdevice ublk_misc;
190 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
192 struct request_queue *q = ub->ub_disk->queue;
193 const struct ublk_param_basic *p = &ub->params.basic;
195 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
196 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
197 blk_queue_io_min(q, 1 << p->io_min_shift);
198 blk_queue_io_opt(q, 1 << p->io_opt_shift);
200 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
201 p->attrs & UBLK_ATTR_FUA);
202 if (p->attrs & UBLK_ATTR_ROTATIONAL)
203 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
205 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
207 blk_queue_max_hw_sectors(q, p->max_sectors);
208 blk_queue_chunk_sectors(q, p->chunk_sectors);
209 blk_queue_virt_boundary(q, p->virt_boundary_mask);
211 if (p->attrs & UBLK_ATTR_READ_ONLY)
212 set_disk_ro(ub->ub_disk, true);
214 set_capacity(ub->ub_disk, p->dev_sectors);
217 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
219 struct request_queue *q = ub->ub_disk->queue;
220 const struct ublk_param_discard *p = &ub->params.discard;
222 q->limits.discard_alignment = p->discard_alignment;
223 q->limits.discard_granularity = p->discard_granularity;
224 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
225 blk_queue_max_write_zeroes_sectors(q,
226 p->max_write_zeroes_sectors);
227 blk_queue_max_discard_segments(q, p->max_discard_segments);
230 static int ublk_validate_params(const struct ublk_device *ub)
232 /* basic param is the only one which must be set */
233 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
234 const struct ublk_param_basic *p = &ub->params.basic;
236 if (p->logical_bs_shift > PAGE_SHIFT)
239 if (p->logical_bs_shift > p->physical_bs_shift)
242 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
247 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
248 const struct ublk_param_discard *p = &ub->params.discard;
250 /* So far, only support single segment discard */
251 if (p->max_discard_sectors && p->max_discard_segments != 1)
254 if (!p->discard_granularity)
261 static int ublk_apply_params(struct ublk_device *ub)
263 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
266 ublk_dev_param_basic_apply(ub);
268 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
269 ublk_dev_param_discard_apply(ub);
274 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
276 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
277 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
282 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
284 if (ubq->flags & UBLK_F_NEED_GET_DATA)
289 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
291 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
296 static void ublk_put_device(struct ublk_device *ub)
298 put_device(&ub->cdev_dev);
301 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
304 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
307 static inline bool ublk_rq_has_data(const struct request *rq)
309 return rq->bio && bio_has_data(rq->bio);
312 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
315 return (struct ublksrv_io_desc *)
316 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
319 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
321 return ublk_get_queue(ub, q_id)->io_cmd_buf;
324 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
326 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
328 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
332 static inline bool ublk_queue_can_use_recovery_reissue(
333 struct ublk_queue *ubq)
335 if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
336 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
341 static inline bool ublk_queue_can_use_recovery(
342 struct ublk_queue *ubq)
344 if (ubq->flags & UBLK_F_USER_RECOVERY)
349 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
351 if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
356 static void ublk_free_disk(struct gendisk *disk)
358 struct ublk_device *ub = disk->private_data;
360 clear_bit(UB_STATE_USED, &ub->state);
361 put_device(&ub->cdev_dev);
364 static const struct block_device_operations ub_fops = {
365 .owner = THIS_MODULE,
366 .free_disk = ublk_free_disk,
369 #define UBLK_MAX_PIN_PAGES 32
371 struct ublk_map_data {
372 const struct ublk_queue *ubq;
373 const struct request *rq;
374 const struct ublk_io *io;
378 struct ublk_io_iter {
379 struct page *pages[UBLK_MAX_PIN_PAGES];
380 unsigned pg_off; /* offset in the 1st page in pages */
381 int nr_pages; /* how many page pointers in pages */
383 struct bvec_iter iter;
386 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
387 unsigned max_bytes, bool to_vm)
389 const unsigned total = min_t(unsigned, max_bytes,
390 PAGE_SIZE - data->pg_off +
391 ((data->nr_pages - 1) << PAGE_SHIFT));
395 while (done < total) {
396 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
397 const unsigned int bytes = min3(bv.bv_len, total - done,
398 (unsigned)(PAGE_SIZE - data->pg_off));
399 void *bv_buf = bvec_kmap_local(&bv);
400 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
403 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
405 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
407 kunmap_local(pg_buf);
408 kunmap_local(bv_buf);
410 /* advance page array */
411 data->pg_off += bytes;
412 if (data->pg_off == PAGE_SIZE) {
420 bio_advance_iter_single(data->bio, &data->iter, bytes);
421 if (!data->iter.bi_size) {
422 data->bio = data->bio->bi_next;
423 if (data->bio == NULL)
425 data->iter = data->bio->bi_iter;
432 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
435 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
436 const unsigned long start_vm = data->io->addr;
437 unsigned int done = 0;
438 struct ublk_io_iter iter = {
439 .pg_off = start_vm & (PAGE_SIZE - 1),
440 .bio = data->rq->bio,
441 .iter = data->rq->bio->bi_iter,
443 const unsigned int nr_pages = round_up(data->max_bytes +
444 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
446 while (done < nr_pages) {
447 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
451 iter.nr_pages = get_user_pages_fast(start_vm +
452 (done << PAGE_SHIFT), to_pin, gup_flags,
454 if (iter.nr_pages <= 0)
455 return done == 0 ? iter.nr_pages : done;
456 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
457 for (i = 0; i < iter.nr_pages; i++) {
459 set_page_dirty(iter.pages[i]);
460 put_page(iter.pages[i]);
462 data->max_bytes -= len;
463 done += iter.nr_pages;
469 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
472 const unsigned int rq_bytes = blk_rq_bytes(req);
474 * no zero copy, we delay copy WRITE request data into ublksrv
475 * context and the big benefit is that pinning pages in current
476 * context is pretty fast, see ublk_pin_user_pages
478 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
481 if (ublk_rq_has_data(req)) {
482 struct ublk_map_data data = {
486 .max_bytes = rq_bytes,
489 ublk_copy_user_pages(&data, true);
491 return rq_bytes - data.max_bytes;
496 static int ublk_unmap_io(const struct ublk_queue *ubq,
497 const struct request *req,
500 const unsigned int rq_bytes = blk_rq_bytes(req);
502 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
503 struct ublk_map_data data = {
507 .max_bytes = io->res,
510 WARN_ON_ONCE(io->res > rq_bytes);
512 ublk_copy_user_pages(&data, false);
514 return io->res - data.max_bytes;
519 static inline unsigned int ublk_req_build_flags(struct request *req)
523 if (req->cmd_flags & REQ_FAILFAST_DEV)
524 flags |= UBLK_IO_F_FAILFAST_DEV;
526 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
527 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
529 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
530 flags |= UBLK_IO_F_FAILFAST_DRIVER;
532 if (req->cmd_flags & REQ_META)
533 flags |= UBLK_IO_F_META;
535 if (req->cmd_flags & REQ_FUA)
536 flags |= UBLK_IO_F_FUA;
538 if (req->cmd_flags & REQ_NOUNMAP)
539 flags |= UBLK_IO_F_NOUNMAP;
541 if (req->cmd_flags & REQ_SWAP)
542 flags |= UBLK_IO_F_SWAP;
547 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
549 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
550 struct ublk_io *io = &ubq->ios[req->tag];
553 switch (req_op(req)) {
555 ublk_op = UBLK_IO_OP_READ;
558 ublk_op = UBLK_IO_OP_WRITE;
561 ublk_op = UBLK_IO_OP_FLUSH;
564 ublk_op = UBLK_IO_OP_DISCARD;
566 case REQ_OP_WRITE_ZEROES:
567 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
570 return BLK_STS_IOERR;
573 /* need to translate since kernel may change */
574 iod->op_flags = ublk_op | ublk_req_build_flags(req);
575 iod->nr_sectors = blk_rq_sectors(req);
576 iod->start_sector = blk_rq_pos(req);
577 iod->addr = io->addr;
582 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
583 struct io_uring_cmd *ioucmd)
585 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
588 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
590 return ubq->ubq_daemon->flags & PF_EXITING;
593 /* todo: handle partial completion */
594 static void ublk_complete_rq(struct request *req)
596 struct ublk_queue *ubq = req->mq_hctx->driver_data;
597 struct ublk_io *io = &ubq->ios[req->tag];
598 unsigned int unmapped_bytes;
600 /* failed read IO if nothing is read */
601 if (!io->res && req_op(req) == REQ_OP_READ)
605 blk_mq_end_request(req, errno_to_blk_status(io->res));
610 * FLUSH or DISCARD usually won't return bytes returned, so end them
613 * Both the two needn't unmap.
615 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
616 blk_mq_end_request(req, BLK_STS_OK);
620 /* for READ request, writing data in iod->addr to rq buffers */
621 unmapped_bytes = ublk_unmap_io(ubq, req, io);
624 * Extremely impossible since we got data filled in just before
626 * Re-read simply for this unlikely case.
628 if (unlikely(unmapped_bytes < io->res))
629 io->res = unmapped_bytes;
631 if (blk_update_request(req, BLK_STS_OK, io->res))
632 blk_mq_requeue_request(req, true);
634 __blk_mq_end_request(req, BLK_STS_OK);
638 * Since __ublk_rq_task_work always fails requests immediately during
639 * exiting, __ublk_fail_req() is only called from abort context during
640 * exiting. So lock is unnecessary.
642 * Also aborting may not be started yet, keep in mind that one failed
643 * request may be issued by block layer again.
645 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
648 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
650 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
651 io->flags |= UBLK_IO_FLAG_ABORTED;
652 if (ublk_queue_can_use_recovery_reissue(ubq))
653 blk_mq_requeue_request(req, false);
655 blk_mq_end_request(req, BLK_STS_IOERR);
659 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
661 /* mark this cmd owned by ublksrv */
662 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
665 * clear ACTIVE since we are done with this sqe/cmd slot
666 * We can only accept io cmd in case of being not active.
668 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
670 /* tell ublksrv one io request is coming */
671 io_uring_cmd_done(io->cmd, res, 0);
674 #define UBLK_REQUEUE_DELAY_MS 3
676 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
679 /* We cannot process this rq so just requeue it. */
680 if (ublk_queue_can_use_recovery(ubq))
681 blk_mq_requeue_request(rq, false);
683 blk_mq_end_request(rq, BLK_STS_IOERR);
685 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
688 static inline void __ublk_rq_task_work(struct request *req)
690 struct ublk_queue *ubq = req->mq_hctx->driver_data;
692 struct ublk_io *io = &ubq->ios[tag];
693 unsigned int mapped_bytes;
695 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
696 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
697 ublk_get_iod(ubq, req->tag)->addr);
700 * Task is exiting if either:
702 * (1) current != ubq_daemon.
703 * io_uring_cmd_complete_in_task() tries to run task_work
704 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
706 * (2) current->flags & PF_EXITING.
708 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
709 __ublk_abort_rq(ubq, req);
713 if (ublk_need_get_data(ubq) &&
714 (req_op(req) == REQ_OP_WRITE ||
715 req_op(req) == REQ_OP_FLUSH)) {
717 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
718 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
721 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
722 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
723 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
724 __func__, io->cmd->cmd_op, ubq->q_id,
725 req->tag, io->flags);
726 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
730 * We have handled UBLK_IO_NEED_GET_DATA command,
731 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
734 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
735 /* update iod->addr because ublksrv may have passed a new io buffer */
736 ublk_get_iod(ubq, req->tag)->addr = io->addr;
737 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
738 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
739 ublk_get_iod(ubq, req->tag)->addr);
742 mapped_bytes = ublk_map_io(ubq, req, io);
744 /* partially mapped, update io descriptor */
745 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
747 * Nothing mapped, retry until we succeed.
749 * We may never succeed in mapping any bytes here because
750 * of OOM. TODO: reserve one buffer with single page pinned
751 * for providing forward progress guarantee.
753 if (unlikely(!mapped_bytes)) {
754 blk_mq_requeue_request(req, false);
755 blk_mq_delay_kick_requeue_list(req->q,
756 UBLK_REQUEUE_DELAY_MS);
760 ublk_get_iod(ubq, req->tag)->nr_sectors =
764 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
767 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
769 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
770 struct ublk_rq_data *data, *tmp;
772 io_cmds = llist_reverse_order(io_cmds);
773 llist_for_each_entry_safe(data, tmp, io_cmds, node)
774 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
777 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
779 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
780 struct ublk_rq_data *data, *tmp;
782 llist_for_each_entry_safe(data, tmp, io_cmds, node)
783 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
786 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
788 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
789 struct ublk_queue *ubq = pdu->ubq;
791 ublk_forward_io_cmds(ubq);
794 static void ublk_rq_task_work_fn(struct callback_head *work)
796 struct ublk_rq_data *data = container_of(work,
797 struct ublk_rq_data, work);
798 struct request *req = blk_mq_rq_from_pdu(data);
799 struct ublk_queue *ubq = req->mq_hctx->driver_data;
801 ublk_forward_io_cmds(ubq);
804 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
806 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
809 if (!llist_add(&data->node, &ubq->io_cmds))
812 io = &ubq->ios[rq->tag];
814 * If the check pass, we know that this is a re-issued request aborted
815 * previously in monitor_work because the ubq_daemon(cmd's task) is
816 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
817 * because this ioucmd's io_uring context may be freed now if no inflight
818 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
820 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
821 * the tag). Then the request is re-started(allocating the tag) and we are here.
822 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
823 * guarantees that here is a re-issued request aborted previously.
825 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
826 ublk_abort_io_cmds(ubq);
827 } else if (ublk_can_use_task_work(ubq)) {
828 if (task_work_add(ubq->ubq_daemon, &data->work,
830 ublk_abort_io_cmds(ubq);
832 struct io_uring_cmd *cmd = io->cmd;
833 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
836 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
840 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
841 const struct blk_mq_queue_data *bd)
843 struct ublk_queue *ubq = hctx->driver_data;
844 struct request *rq = bd->rq;
847 /* fill iod to slot in io cmd buffer */
848 res = ublk_setup_iod(ubq, rq);
849 if (unlikely(res != BLK_STS_OK))
850 return BLK_STS_IOERR;
852 /* With recovery feature enabled, force_abort is set in
853 * ublk_stop_dev() before calling del_gendisk(). We have to
854 * abort all requeued and new rqs here to let del_gendisk()
855 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
856 * to avoid UAF on io_uring ctx.
858 * Note: force_abort is guaranteed to be seen because it is set
859 * before request queue is unqiuesced.
861 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
862 return BLK_STS_IOERR;
864 blk_mq_start_request(bd->rq);
866 if (unlikely(ubq_daemon_is_dying(ubq))) {
867 __ublk_abort_rq(ubq, rq);
871 ublk_queue_cmd(ubq, rq);
876 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
877 unsigned int hctx_idx)
879 struct ublk_device *ub = driver_data;
880 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
882 hctx->driver_data = ubq;
886 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
887 unsigned int hctx_idx, unsigned int numa_node)
889 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
891 init_task_work(&data->work, ublk_rq_task_work_fn);
895 static const struct blk_mq_ops ublk_mq_ops = {
896 .queue_rq = ublk_queue_rq,
897 .init_hctx = ublk_init_hctx,
898 .init_request = ublk_init_rq,
901 static int ublk_ch_open(struct inode *inode, struct file *filp)
903 struct ublk_device *ub = container_of(inode->i_cdev,
904 struct ublk_device, cdev);
906 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
908 filp->private_data = ub;
912 static int ublk_ch_release(struct inode *inode, struct file *filp)
914 struct ublk_device *ub = filp->private_data;
916 clear_bit(UB_STATE_OPEN, &ub->state);
920 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
921 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
923 struct ublk_device *ub = filp->private_data;
924 size_t sz = vma->vm_end - vma->vm_start;
925 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
926 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
929 spin_lock(&ub->mm_lock);
931 ub->mm = current->mm;
932 if (current->mm != ub->mm)
934 spin_unlock(&ub->mm_lock);
939 if (vma->vm_flags & VM_WRITE)
942 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
943 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
946 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
947 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
948 __func__, q_id, current->pid, vma->vm_start,
949 phys_off, (unsigned long)sz);
951 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
954 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
955 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
958 static void ublk_commit_completion(struct ublk_device *ub,
959 struct ublksrv_io_cmd *ub_cmd)
961 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
962 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
963 struct ublk_io *io = &ubq->ios[tag];
966 /* now this cmd slot is owned by nbd driver */
967 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
968 io->res = ub_cmd->result;
970 /* find the io request and complete */
971 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
973 if (req && likely(!blk_should_fake_timeout(req->q)))
974 ublk_complete_rq(req);
978 * When ->ubq_daemon is exiting, either new request is ended immediately,
979 * or any queued io command is drained, so it is safe to abort queue
982 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
986 if (!ublk_get_device(ub))
989 for (i = 0; i < ubq->q_depth; i++) {
990 struct ublk_io *io = &ubq->ios[i];
992 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
996 * Either we fail the request or ublk_rq_task_work_fn
999 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1001 __ublk_fail_req(ubq, io, rq);
1004 ublk_put_device(ub);
1007 static void ublk_daemon_monitor_work(struct work_struct *work)
1009 struct ublk_device *ub =
1010 container_of(work, struct ublk_device, monitor_work.work);
1013 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1014 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1016 if (ubq_daemon_is_dying(ubq)) {
1017 if (ublk_queue_can_use_recovery(ubq))
1018 schedule_work(&ub->quiesce_work);
1020 schedule_work(&ub->stop_work);
1022 /* abort queue is for making forward progress */
1023 ublk_abort_queue(ub, ubq);
1028 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1029 * after ublk_remove() or __ublk_quiesce_dev() is started.
1031 * No need ub->mutex, monitor work are canceled after state is marked
1032 * as not LIVE, so new state is observed reliably.
1034 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1035 schedule_delayed_work(&ub->monitor_work,
1036 UBLK_DAEMON_MONITOR_PERIOD);
1039 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1041 return ubq->nr_io_ready == ubq->q_depth;
1044 static void ublk_cancel_queue(struct ublk_queue *ubq)
1048 if (!ublk_queue_ready(ubq))
1051 for (i = 0; i < ubq->q_depth; i++) {
1052 struct ublk_io *io = &ubq->ios[i];
1054 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1055 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1058 /* all io commands are canceled */
1059 ubq->nr_io_ready = 0;
1062 /* Cancel all pending commands, must be called after del_gendisk() returns */
1063 static void ublk_cancel_dev(struct ublk_device *ub)
1067 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1068 ublk_cancel_queue(ublk_get_queue(ub, i));
1071 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1075 if (blk_mq_request_started(rq)) {
1082 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1086 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1089 blk_mq_tagset_busy_iter(&ub->tag_set,
1090 ublk_check_inflight_rq, &idle);
1093 msleep(UBLK_REQUEUE_DELAY_MS);
1097 static void __ublk_quiesce_dev(struct ublk_device *ub)
1099 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1100 __func__, ub->dev_info.dev_id,
1101 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1102 "LIVE" : "QUIESCED");
1103 blk_mq_quiesce_queue(ub->ub_disk->queue);
1104 ublk_wait_tagset_rqs_idle(ub);
1105 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1106 ublk_cancel_dev(ub);
1107 /* we are going to release task_struct of ubq_daemon and resets
1108 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1109 * Besides, monitor_work is not necessary in QUIESCED state since we have
1110 * already scheduled quiesce_work and quiesced all ubqs.
1112 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1113 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1115 cancel_delayed_work_sync(&ub->monitor_work);
1118 static void ublk_quiesce_work_fn(struct work_struct *work)
1120 struct ublk_device *ub =
1121 container_of(work, struct ublk_device, quiesce_work);
1123 mutex_lock(&ub->mutex);
1124 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1126 __ublk_quiesce_dev(ub);
1128 mutex_unlock(&ub->mutex);
1131 static void ublk_unquiesce_dev(struct ublk_device *ub)
1135 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1136 __func__, ub->dev_info.dev_id,
1137 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1138 "LIVE" : "QUIESCED");
1139 /* quiesce_work has run. We let requeued rqs be aborted
1140 * before running fallback_wq. "force_abort" must be seen
1141 * after request queue is unqiuesced. Then del_gendisk()
1144 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1145 ublk_get_queue(ub, i)->force_abort = true;
1147 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1148 /* We may have requeued some rqs in ublk_quiesce_queue() */
1149 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1152 static void ublk_stop_dev(struct ublk_device *ub)
1154 mutex_lock(&ub->mutex);
1155 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1157 if (ublk_can_use_recovery(ub)) {
1158 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1159 __ublk_quiesce_dev(ub);
1160 ublk_unquiesce_dev(ub);
1162 del_gendisk(ub->ub_disk);
1163 ub->dev_info.state = UBLK_S_DEV_DEAD;
1164 ub->dev_info.ublksrv_pid = -1;
1165 put_disk(ub->ub_disk);
1168 ublk_cancel_dev(ub);
1169 mutex_unlock(&ub->mutex);
1170 cancel_delayed_work_sync(&ub->monitor_work);
1173 /* device can only be started after all IOs are ready */
1174 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1176 mutex_lock(&ub->mutex);
1178 if (ublk_queue_ready(ubq)) {
1179 ubq->ubq_daemon = current;
1180 get_task_struct(ubq->ubq_daemon);
1181 ub->nr_queues_ready++;
1183 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1184 complete_all(&ub->completion);
1185 mutex_unlock(&ub->mutex);
1188 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1191 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1192 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1194 ublk_queue_cmd(ubq, req);
1197 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1199 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1200 struct ublk_device *ub = cmd->file->private_data;
1201 struct ublk_queue *ubq;
1203 u32 cmd_op = cmd->cmd_op;
1204 unsigned tag = ub_cmd->tag;
1207 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1208 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1211 if (!(issue_flags & IO_URING_F_SQE128))
1214 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1217 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1218 if (!ubq || ub_cmd->q_id != ubq->q_id)
1221 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1224 if (tag >= ubq->q_depth)
1227 io = &ubq->ios[tag];
1229 /* there is pending io cmd, something must be wrong */
1230 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1236 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1237 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1239 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1240 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1244 case UBLK_IO_FETCH_REQ:
1245 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1246 if (ublk_queue_ready(ubq)) {
1251 * The io is being handled by server, so COMMIT_RQ is expected
1252 * instead of FETCH_REQ
1254 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1256 /* FETCH_RQ has to provide IO buffer */
1260 io->flags |= UBLK_IO_FLAG_ACTIVE;
1261 io->addr = ub_cmd->addr;
1263 ublk_mark_io_ready(ub, ubq);
1265 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1266 /* FETCH_RQ has to provide IO buffer */
1269 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1271 io->addr = ub_cmd->addr;
1272 io->flags |= UBLK_IO_FLAG_ACTIVE;
1274 ublk_commit_completion(ub, ub_cmd);
1276 case UBLK_IO_NEED_GET_DATA:
1277 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1279 io->addr = ub_cmd->addr;
1281 io->flags |= UBLK_IO_FLAG_ACTIVE;
1282 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1287 return -EIOCBQUEUED;
1290 io_uring_cmd_done(cmd, ret, 0);
1291 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1292 __func__, cmd_op, tag, ret, io->flags);
1293 return -EIOCBQUEUED;
1296 static const struct file_operations ublk_ch_fops = {
1297 .owner = THIS_MODULE,
1298 .open = ublk_ch_open,
1299 .release = ublk_ch_release,
1300 .llseek = no_llseek,
1301 .uring_cmd = ublk_ch_uring_cmd,
1302 .mmap = ublk_ch_mmap,
1305 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1307 int size = ublk_queue_cmd_buf_size(ub, q_id);
1308 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1310 if (ubq->ubq_daemon)
1311 put_task_struct(ubq->ubq_daemon);
1312 if (ubq->io_cmd_buf)
1313 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1316 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1318 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1319 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1323 ubq->flags = ub->dev_info.flags;
1325 ubq->q_depth = ub->dev_info.queue_depth;
1326 size = ublk_queue_cmd_buf_size(ub, q_id);
1328 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1332 ubq->io_cmd_buf = ptr;
1337 static void ublk_deinit_queues(struct ublk_device *ub)
1339 int nr_queues = ub->dev_info.nr_hw_queues;
1345 for (i = 0; i < nr_queues; i++)
1346 ublk_deinit_queue(ub, i);
1347 kfree(ub->__queues);
1350 static int ublk_init_queues(struct ublk_device *ub)
1352 int nr_queues = ub->dev_info.nr_hw_queues;
1353 int depth = ub->dev_info.queue_depth;
1354 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1355 int i, ret = -ENOMEM;
1357 ub->queue_size = ubq_size;
1358 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1362 for (i = 0; i < nr_queues; i++) {
1363 if (ublk_init_queue(ub, i))
1367 init_completion(&ub->completion);
1371 ublk_deinit_queues(ub);
1375 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1380 spin_lock(&ublk_idr_lock);
1381 /* allocate id, if @id >= 0, we're requesting that specific id */
1383 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1387 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1389 spin_unlock(&ublk_idr_lock);
1392 ub->ub_number = err;
1397 static void ublk_free_dev_number(struct ublk_device *ub)
1399 spin_lock(&ublk_idr_lock);
1400 idr_remove(&ublk_index_idr, ub->ub_number);
1401 wake_up_all(&ublk_idr_wq);
1402 spin_unlock(&ublk_idr_lock);
1405 static void ublk_cdev_rel(struct device *dev)
1407 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1409 blk_mq_free_tag_set(&ub->tag_set);
1410 ublk_deinit_queues(ub);
1411 ublk_free_dev_number(ub);
1412 mutex_destroy(&ub->mutex);
1416 static int ublk_add_chdev(struct ublk_device *ub)
1418 struct device *dev = &ub->cdev_dev;
1419 int minor = ub->ub_number;
1422 dev->parent = ublk_misc.this_device;
1423 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1424 dev->class = ublk_chr_class;
1425 dev->release = ublk_cdev_rel;
1426 device_initialize(dev);
1428 ret = dev_set_name(dev, "ublkc%d", minor);
1432 cdev_init(&ub->cdev, &ublk_ch_fops);
1433 ret = cdev_device_add(&ub->cdev, dev);
1442 static void ublk_stop_work_fn(struct work_struct *work)
1444 struct ublk_device *ub =
1445 container_of(work, struct ublk_device, stop_work);
1450 /* align max io buffer size with PAGE_SIZE */
1451 static void ublk_align_max_io_size(struct ublk_device *ub)
1453 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1455 ub->dev_info.max_io_buf_bytes =
1456 round_down(max_io_bytes, PAGE_SIZE);
1459 static int ublk_add_tag_set(struct ublk_device *ub)
1461 ub->tag_set.ops = &ublk_mq_ops;
1462 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1463 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1464 ub->tag_set.numa_node = NUMA_NO_NODE;
1465 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1466 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1467 ub->tag_set.driver_data = ub;
1468 return blk_mq_alloc_tag_set(&ub->tag_set);
1471 static void ublk_remove(struct ublk_device *ub)
1474 cancel_work_sync(&ub->stop_work);
1475 cancel_work_sync(&ub->quiesce_work);
1476 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1477 put_device(&ub->cdev_dev);
1480 static struct ublk_device *ublk_get_device_from_id(int idx)
1482 struct ublk_device *ub = NULL;
1487 spin_lock(&ublk_idr_lock);
1488 ub = idr_find(&ublk_index_idr, idx);
1490 ub = ublk_get_device(ub);
1491 spin_unlock(&ublk_idr_lock);
1496 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1498 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1499 int ublksrv_pid = (int)header->data[0];
1500 struct ublk_device *ub;
1501 struct gendisk *disk;
1504 if (ublksrv_pid <= 0)
1507 ub = ublk_get_device_from_id(header->dev_id);
1511 wait_for_completion_interruptible(&ub->completion);
1513 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1515 mutex_lock(&ub->mutex);
1516 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1517 test_bit(UB_STATE_USED, &ub->state)) {
1522 disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1524 ret = PTR_ERR(disk);
1527 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1528 disk->fops = &ub_fops;
1529 disk->private_data = ub;
1531 ub->dev_info.ublksrv_pid = ublksrv_pid;
1534 ret = ublk_apply_params(ub);
1538 get_device(&ub->cdev_dev);
1539 ret = add_disk(disk);
1542 * Has to drop the reference since ->free_disk won't be
1543 * called in case of add_disk failure.
1545 ublk_put_device(ub);
1548 set_bit(UB_STATE_USED, &ub->state);
1549 ub->dev_info.state = UBLK_S_DEV_LIVE;
1554 mutex_unlock(&ub->mutex);
1555 ublk_put_device(ub);
1559 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1561 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1562 void __user *argp = (void __user *)(unsigned long)header->addr;
1563 struct ublk_device *ub;
1564 cpumask_var_t cpumask;
1565 unsigned long queue;
1566 unsigned int retlen;
1570 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1572 if (header->len & (sizeof(unsigned long)-1))
1577 ub = ublk_get_device_from_id(header->dev_id);
1581 queue = header->data[0];
1582 if (queue >= ub->dev_info.nr_hw_queues)
1583 goto out_put_device;
1586 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1587 goto out_put_device;
1589 for_each_possible_cpu(i) {
1590 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1591 cpumask_set_cpu(i, cpumask);
1595 retlen = min_t(unsigned short, header->len, cpumask_size());
1596 if (copy_to_user(argp, cpumask, retlen))
1597 goto out_free_cpumask;
1598 if (retlen != header->len &&
1599 clear_user(argp + retlen, header->len - retlen))
1600 goto out_free_cpumask;
1604 free_cpumask_var(cpumask);
1606 ublk_put_device(ub);
1610 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1612 pr_devel("%s: dev id %d flags %llx\n", __func__,
1613 info->dev_id, info->flags);
1614 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1615 info->nr_hw_queues, info->queue_depth);
1618 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1620 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1621 void __user *argp = (void __user *)(unsigned long)header->addr;
1622 struct ublksrv_ctrl_dev_info info;
1623 struct ublk_device *ub;
1626 if (header->len < sizeof(info) || !header->addr)
1628 if (header->queue_id != (u16)-1) {
1629 pr_warn("%s: queue_id is wrong %x\n",
1630 __func__, header->queue_id);
1633 if (copy_from_user(&info, argp, sizeof(info)))
1635 ublk_dump_dev_info(&info);
1636 if (header->dev_id != info.dev_id) {
1637 pr_warn("%s: dev id not match %u %u\n",
1638 __func__, header->dev_id, info.dev_id);
1642 ret = mutex_lock_killable(&ublk_ctl_mutex);
1647 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1650 mutex_init(&ub->mutex);
1651 spin_lock_init(&ub->mm_lock);
1652 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1653 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1654 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1656 ret = ublk_alloc_dev_number(ub, header->dev_id);
1660 memcpy(&ub->dev_info, &info, sizeof(info));
1662 /* update device id */
1663 ub->dev_info.dev_id = ub->ub_number;
1666 * 64bit flags will be copied back to userspace as feature
1667 * negotiation result, so have to clear flags which driver
1668 * doesn't support yet, then userspace can get correct flags
1669 * (features) to handle.
1671 ub->dev_info.flags &= UBLK_F_ALL;
1673 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1674 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1676 /* We are not ready to support zero copy */
1677 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1679 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1680 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1681 ublk_align_max_io_size(ub);
1683 ret = ublk_init_queues(ub);
1685 goto out_free_dev_number;
1687 ret = ublk_add_tag_set(ub);
1689 goto out_deinit_queues;
1692 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1693 goto out_free_tag_set;
1696 * Add the char dev so that ublksrv daemon can be setup.
1697 * ublk_add_chdev() will cleanup everything if it fails.
1699 ret = ublk_add_chdev(ub);
1703 blk_mq_free_tag_set(&ub->tag_set);
1705 ublk_deinit_queues(ub);
1706 out_free_dev_number:
1707 ublk_free_dev_number(ub);
1709 mutex_destroy(&ub->mutex);
1712 mutex_unlock(&ublk_ctl_mutex);
1716 static inline bool ublk_idr_freed(int id)
1720 spin_lock(&ublk_idr_lock);
1721 ptr = idr_find(&ublk_index_idr, id);
1722 spin_unlock(&ublk_idr_lock);
1727 static int ublk_ctrl_del_dev(int idx)
1729 struct ublk_device *ub;
1732 ret = mutex_lock_killable(&ublk_ctl_mutex);
1736 ub = ublk_get_device_from_id(idx);
1739 ublk_put_device(ub);
1746 * Wait until the idr is removed, then it can be reused after
1747 * DEL_DEV command is returned.
1750 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1751 mutex_unlock(&ublk_ctl_mutex);
1756 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1758 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1760 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1761 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1762 header->data[0], header->addr, header->len);
1765 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1767 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1768 struct ublk_device *ub;
1770 ub = ublk_get_device_from_id(header->dev_id);
1775 cancel_work_sync(&ub->stop_work);
1776 cancel_work_sync(&ub->quiesce_work);
1778 ublk_put_device(ub);
1782 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1784 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1785 void __user *argp = (void __user *)(unsigned long)header->addr;
1786 struct ublk_device *ub;
1789 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1792 ub = ublk_get_device_from_id(header->dev_id);
1796 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1798 ublk_put_device(ub);
1803 static int ublk_ctrl_get_params(struct io_uring_cmd *cmd)
1805 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1806 void __user *argp = (void __user *)(unsigned long)header->addr;
1807 struct ublk_params_header ph;
1808 struct ublk_device *ub;
1811 if (header->len <= sizeof(ph) || !header->addr)
1814 if (copy_from_user(&ph, argp, sizeof(ph)))
1817 if (ph.len > header->len || !ph.len)
1820 if (ph.len > sizeof(struct ublk_params))
1821 ph.len = sizeof(struct ublk_params);
1823 ub = ublk_get_device_from_id(header->dev_id);
1827 mutex_lock(&ub->mutex);
1828 if (copy_to_user(argp, &ub->params, ph.len))
1832 mutex_unlock(&ub->mutex);
1834 ublk_put_device(ub);
1838 static int ublk_ctrl_set_params(struct io_uring_cmd *cmd)
1840 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1841 void __user *argp = (void __user *)(unsigned long)header->addr;
1842 struct ublk_params_header ph;
1843 struct ublk_device *ub;
1846 if (header->len <= sizeof(ph) || !header->addr)
1849 if (copy_from_user(&ph, argp, sizeof(ph)))
1852 if (ph.len > header->len || !ph.len || !ph.types)
1855 if (ph.len > sizeof(struct ublk_params))
1856 ph.len = sizeof(struct ublk_params);
1858 ub = ublk_get_device_from_id(header->dev_id);
1862 /* parameters can only be changed when device isn't live */
1863 mutex_lock(&ub->mutex);
1864 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1866 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1869 /* clear all we don't support yet */
1870 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1871 ret = ublk_validate_params(ub);
1873 mutex_unlock(&ub->mutex);
1874 ublk_put_device(ub);
1879 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1883 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1884 /* All old ioucmds have to be completed */
1885 WARN_ON_ONCE(ubq->nr_io_ready);
1886 /* old daemon is PF_EXITING, put it now */
1887 put_task_struct(ubq->ubq_daemon);
1888 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1889 ubq->ubq_daemon = NULL;
1891 for (i = 0; i < ubq->q_depth; i++) {
1892 struct ublk_io *io = &ubq->ios[i];
1894 /* forget everything now and be ready for new FETCH_REQ */
1901 static int ublk_ctrl_start_recovery(struct io_uring_cmd *cmd)
1903 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1904 struct ublk_device *ub;
1908 ub = ublk_get_device_from_id(header->dev_id);
1912 mutex_lock(&ub->mutex);
1913 if (!ublk_can_use_recovery(ub))
1916 * START_RECOVERY is only allowd after:
1918 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1919 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1922 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1923 * (a)has quiesced request queue
1924 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1925 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1926 * (d)has completed/camceled all ioucmds owned by ther dying process
1928 if (test_bit(UB_STATE_OPEN, &ub->state) ||
1929 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1933 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1934 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1935 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1936 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1938 ub->nr_queues_ready = 0;
1939 init_completion(&ub->completion);
1942 mutex_unlock(&ub->mutex);
1943 ublk_put_device(ub);
1947 static int ublk_ctrl_end_recovery(struct io_uring_cmd *cmd)
1949 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1950 int ublksrv_pid = (int)header->data[0];
1951 struct ublk_device *ub;
1954 ub = ublk_get_device_from_id(header->dev_id);
1958 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1959 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1960 /* wait until new ubq_daemon sending all FETCH_REQ */
1961 wait_for_completion_interruptible(&ub->completion);
1962 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
1963 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1965 mutex_lock(&ub->mutex);
1966 if (!ublk_can_use_recovery(ub))
1969 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1973 ub->dev_info.ublksrv_pid = ublksrv_pid;
1974 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
1975 __func__, ublksrv_pid, header->dev_id);
1976 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1977 pr_devel("%s: queue unquiesced, dev id %d.\n",
1978 __func__, header->dev_id);
1979 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1980 ub->dev_info.state = UBLK_S_DEV_LIVE;
1981 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1984 mutex_unlock(&ub->mutex);
1985 ublk_put_device(ub);
1989 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1990 unsigned int issue_flags)
1992 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1995 if (issue_flags & IO_URING_F_NONBLOCK)
1998 ublk_ctrl_cmd_dump(cmd);
2000 if (!(issue_flags & IO_URING_F_SQE128))
2004 if (!capable(CAP_SYS_ADMIN))
2008 switch (cmd->cmd_op) {
2009 case UBLK_CMD_START_DEV:
2010 ret = ublk_ctrl_start_dev(cmd);
2012 case UBLK_CMD_STOP_DEV:
2013 ret = ublk_ctrl_stop_dev(cmd);
2015 case UBLK_CMD_GET_DEV_INFO:
2016 ret = ublk_ctrl_get_dev_info(cmd);
2018 case UBLK_CMD_ADD_DEV:
2019 ret = ublk_ctrl_add_dev(cmd);
2021 case UBLK_CMD_DEL_DEV:
2022 ret = ublk_ctrl_del_dev(header->dev_id);
2024 case UBLK_CMD_GET_QUEUE_AFFINITY:
2025 ret = ublk_ctrl_get_queue_affinity(cmd);
2027 case UBLK_CMD_GET_PARAMS:
2028 ret = ublk_ctrl_get_params(cmd);
2030 case UBLK_CMD_SET_PARAMS:
2031 ret = ublk_ctrl_set_params(cmd);
2033 case UBLK_CMD_START_USER_RECOVERY:
2034 ret = ublk_ctrl_start_recovery(cmd);
2036 case UBLK_CMD_END_USER_RECOVERY:
2037 ret = ublk_ctrl_end_recovery(cmd);
2043 io_uring_cmd_done(cmd, ret, 0);
2044 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2045 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2046 return -EIOCBQUEUED;
2049 static const struct file_operations ublk_ctl_fops = {
2050 .open = nonseekable_open,
2051 .uring_cmd = ublk_ctrl_uring_cmd,
2052 .owner = THIS_MODULE,
2053 .llseek = noop_llseek,
2056 static struct miscdevice ublk_misc = {
2057 .minor = MISC_DYNAMIC_MINOR,
2058 .name = "ublk-control",
2059 .fops = &ublk_ctl_fops,
2062 static int __init ublk_init(void)
2066 init_waitqueue_head(&ublk_idr_wq);
2068 ret = misc_register(&ublk_misc);
2072 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2074 goto unregister_mis;
2076 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2077 if (IS_ERR(ublk_chr_class)) {
2078 ret = PTR_ERR(ublk_chr_class);
2079 goto free_chrdev_region;
2084 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2086 misc_deregister(&ublk_misc);
2090 static void __exit ublk_exit(void)
2092 struct ublk_device *ub;
2095 idr_for_each_entry(&ublk_index_idr, ub, id)
2098 class_destroy(ublk_chr_class);
2099 misc_deregister(&ublk_misc);
2101 idr_destroy(&ublk_index_idr);
2102 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2105 module_init(ublk_init);
2106 module_exit(ublk_exit);
2108 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2109 MODULE_LICENSE("GPL");