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)
61 struct callback_head work;
62 struct llist_node node;
66 struct ublk_uring_cmd_pdu {
67 struct ublk_queue *ubq;
71 * io command is active: sqe cmd is received, and its cqe isn't done
73 * If the flag is set, the io command is owned by ublk driver, and waited
74 * for incoming blk-mq request from the ublk block device.
76 * If the flag is cleared, the io command will be completed, and owned by
79 #define UBLK_IO_FLAG_ACTIVE 0x01
82 * IO command is completed via cqe, and it is being handled by ublksrv, and
85 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
88 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
91 * IO command is aborted, so this flag is set in case of
92 * !UBLK_IO_FLAG_ACTIVE.
94 * After this flag is observed, any pending or new incoming request
95 * associated with this io command will be failed immediately
97 #define UBLK_IO_FLAG_ABORTED 0x04
100 * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
101 * get data buffer address from ublksrv.
103 * Then, bio data could be copied into this data buffer for a WRITE request
104 * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
106 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
109 /* userspace buffer address from io cmd */
114 struct io_uring_cmd *cmd;
122 struct task_struct *ubq_daemon;
125 struct llist_head io_cmds;
127 unsigned long io_addr; /* mapped vm address */
128 unsigned int max_io_sz;
130 unsigned short nr_io_ready; /* how many ios setup */
131 struct ublk_device *dev;
132 struct ublk_io ios[];
135 #define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
138 struct gendisk *ub_disk;
142 unsigned short queue_size;
143 struct ublksrv_ctrl_dev_info dev_info;
145 struct blk_mq_tag_set tag_set;
148 struct device cdev_dev;
150 #define UB_STATE_OPEN 0
151 #define UB_STATE_USED 1
158 struct mm_struct *mm;
160 struct ublk_params params;
162 struct completion completion;
163 unsigned int nr_queues_ready;
164 atomic_t nr_aborted_queues;
167 * Our ubq->daemon may be killed without any notification, so
168 * monitor each queue's daemon periodically
170 struct delayed_work monitor_work;
171 struct work_struct quiesce_work;
172 struct work_struct stop_work;
175 /* header of ublk_params */
176 struct ublk_params_header {
181 static dev_t ublk_chr_devt;
182 static struct class *ublk_chr_class;
184 static DEFINE_IDR(ublk_index_idr);
185 static DEFINE_SPINLOCK(ublk_idr_lock);
186 static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
188 static DEFINE_MUTEX(ublk_ctl_mutex);
190 static struct miscdevice ublk_misc;
192 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
194 struct request_queue *q = ub->ub_disk->queue;
195 const struct ublk_param_basic *p = &ub->params.basic;
197 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
198 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
199 blk_queue_io_min(q, 1 << p->io_min_shift);
200 blk_queue_io_opt(q, 1 << p->io_opt_shift);
202 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
203 p->attrs & UBLK_ATTR_FUA);
204 if (p->attrs & UBLK_ATTR_ROTATIONAL)
205 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
207 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
209 blk_queue_max_hw_sectors(q, p->max_sectors);
210 blk_queue_chunk_sectors(q, p->chunk_sectors);
211 blk_queue_virt_boundary(q, p->virt_boundary_mask);
213 if (p->attrs & UBLK_ATTR_READ_ONLY)
214 set_disk_ro(ub->ub_disk, true);
216 set_capacity(ub->ub_disk, p->dev_sectors);
219 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
221 struct request_queue *q = ub->ub_disk->queue;
222 const struct ublk_param_discard *p = &ub->params.discard;
224 q->limits.discard_alignment = p->discard_alignment;
225 q->limits.discard_granularity = p->discard_granularity;
226 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
227 blk_queue_max_write_zeroes_sectors(q,
228 p->max_write_zeroes_sectors);
229 blk_queue_max_discard_segments(q, p->max_discard_segments);
232 static int ublk_validate_params(const struct ublk_device *ub)
234 /* basic param is the only one which must be set */
235 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
236 const struct ublk_param_basic *p = &ub->params.basic;
238 if (p->logical_bs_shift > PAGE_SHIFT)
241 if (p->logical_bs_shift > p->physical_bs_shift)
244 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
249 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
250 const struct ublk_param_discard *p = &ub->params.discard;
252 /* So far, only support single segment discard */
253 if (p->max_discard_sectors && p->max_discard_segments != 1)
256 if (!p->discard_granularity)
263 static int ublk_apply_params(struct ublk_device *ub)
265 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
268 ublk_dev_param_basic_apply(ub);
270 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
271 ublk_dev_param_discard_apply(ub);
276 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
278 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
279 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
284 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
286 if (ubq->flags & UBLK_F_NEED_GET_DATA)
291 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
293 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
298 static void ublk_put_device(struct ublk_device *ub)
300 put_device(&ub->cdev_dev);
303 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
306 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
309 static inline bool ublk_rq_has_data(const struct request *rq)
311 return rq->bio && bio_has_data(rq->bio);
314 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
317 return (struct ublksrv_io_desc *)
318 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
321 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
323 return ublk_get_queue(ub, q_id)->io_cmd_buf;
326 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
328 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
330 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
334 static inline bool ublk_queue_can_use_recovery_reissue(
335 struct ublk_queue *ubq)
337 if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
338 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
343 static inline bool ublk_queue_can_use_recovery(
344 struct ublk_queue *ubq)
346 if (ubq->flags & UBLK_F_USER_RECOVERY)
351 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
353 if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
358 static void ublk_free_disk(struct gendisk *disk)
360 struct ublk_device *ub = disk->private_data;
362 clear_bit(UB_STATE_USED, &ub->state);
363 put_device(&ub->cdev_dev);
366 static const struct block_device_operations ub_fops = {
367 .owner = THIS_MODULE,
368 .free_disk = ublk_free_disk,
371 #define UBLK_MAX_PIN_PAGES 32
373 struct ublk_map_data {
374 const struct ublk_queue *ubq;
375 const struct request *rq;
376 const struct ublk_io *io;
380 struct ublk_io_iter {
381 struct page *pages[UBLK_MAX_PIN_PAGES];
382 unsigned pg_off; /* offset in the 1st page in pages */
383 int nr_pages; /* how many page pointers in pages */
385 struct bvec_iter iter;
388 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
389 unsigned max_bytes, bool to_vm)
391 const unsigned total = min_t(unsigned, max_bytes,
392 PAGE_SIZE - data->pg_off +
393 ((data->nr_pages - 1) << PAGE_SHIFT));
397 while (done < total) {
398 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
399 const unsigned int bytes = min3(bv.bv_len, total - done,
400 (unsigned)(PAGE_SIZE - data->pg_off));
401 void *bv_buf = bvec_kmap_local(&bv);
402 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
405 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
407 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
409 kunmap_local(pg_buf);
410 kunmap_local(bv_buf);
412 /* advance page array */
413 data->pg_off += bytes;
414 if (data->pg_off == PAGE_SIZE) {
422 bio_advance_iter_single(data->bio, &data->iter, bytes);
423 if (!data->iter.bi_size) {
424 data->bio = data->bio->bi_next;
425 if (data->bio == NULL)
427 data->iter = data->bio->bi_iter;
434 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
437 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
438 const unsigned long start_vm = data->io->addr;
439 unsigned int done = 0;
440 struct ublk_io_iter iter = {
441 .pg_off = start_vm & (PAGE_SIZE - 1),
442 .bio = data->rq->bio,
443 .iter = data->rq->bio->bi_iter,
445 const unsigned int nr_pages = round_up(data->max_bytes +
446 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
448 while (done < nr_pages) {
449 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
453 iter.nr_pages = get_user_pages_fast(start_vm +
454 (done << PAGE_SHIFT), to_pin, gup_flags,
456 if (iter.nr_pages <= 0)
457 return done == 0 ? iter.nr_pages : done;
458 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
459 for (i = 0; i < iter.nr_pages; i++) {
461 set_page_dirty(iter.pages[i]);
462 put_page(iter.pages[i]);
464 data->max_bytes -= len;
465 done += iter.nr_pages;
471 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
474 const unsigned int rq_bytes = blk_rq_bytes(req);
476 * no zero copy, we delay copy WRITE request data into ublksrv
477 * context and the big benefit is that pinning pages in current
478 * context is pretty fast, see ublk_pin_user_pages
480 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
483 if (ublk_rq_has_data(req)) {
484 struct ublk_map_data data = {
488 .max_bytes = rq_bytes,
491 ublk_copy_user_pages(&data, true);
493 return rq_bytes - data.max_bytes;
498 static int ublk_unmap_io(const struct ublk_queue *ubq,
499 const struct request *req,
502 const unsigned int rq_bytes = blk_rq_bytes(req);
504 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
505 struct ublk_map_data data = {
509 .max_bytes = io->res,
512 WARN_ON_ONCE(io->res > rq_bytes);
514 ublk_copy_user_pages(&data, false);
516 return io->res - data.max_bytes;
521 static inline unsigned int ublk_req_build_flags(struct request *req)
525 if (req->cmd_flags & REQ_FAILFAST_DEV)
526 flags |= UBLK_IO_F_FAILFAST_DEV;
528 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
529 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
531 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
532 flags |= UBLK_IO_F_FAILFAST_DRIVER;
534 if (req->cmd_flags & REQ_META)
535 flags |= UBLK_IO_F_META;
537 if (req->cmd_flags & REQ_FUA)
538 flags |= UBLK_IO_F_FUA;
540 if (req->cmd_flags & REQ_NOUNMAP)
541 flags |= UBLK_IO_F_NOUNMAP;
543 if (req->cmd_flags & REQ_SWAP)
544 flags |= UBLK_IO_F_SWAP;
549 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
551 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
552 struct ublk_io *io = &ubq->ios[req->tag];
555 switch (req_op(req)) {
557 ublk_op = UBLK_IO_OP_READ;
560 ublk_op = UBLK_IO_OP_WRITE;
563 ublk_op = UBLK_IO_OP_FLUSH;
566 ublk_op = UBLK_IO_OP_DISCARD;
568 case REQ_OP_WRITE_ZEROES:
569 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
572 return BLK_STS_IOERR;
575 /* need to translate since kernel may change */
576 iod->op_flags = ublk_op | ublk_req_build_flags(req);
577 iod->nr_sectors = blk_rq_sectors(req);
578 iod->start_sector = blk_rq_pos(req);
579 iod->addr = io->addr;
584 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
585 struct io_uring_cmd *ioucmd)
587 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
590 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
592 return ubq->ubq_daemon->flags & PF_EXITING;
595 /* todo: handle partial completion */
596 static void ublk_complete_rq(struct request *req)
598 struct ublk_queue *ubq = req->mq_hctx->driver_data;
599 struct ublk_io *io = &ubq->ios[req->tag];
600 unsigned int unmapped_bytes;
602 /* failed read IO if nothing is read */
603 if (!io->res && req_op(req) == REQ_OP_READ)
607 blk_mq_end_request(req, errno_to_blk_status(io->res));
612 * FLUSH or DISCARD usually won't return bytes returned, so end them
615 * Both the two needn't unmap.
617 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
618 blk_mq_end_request(req, BLK_STS_OK);
622 /* for READ request, writing data in iod->addr to rq buffers */
623 unmapped_bytes = ublk_unmap_io(ubq, req, io);
626 * Extremely impossible since we got data filled in just before
628 * Re-read simply for this unlikely case.
630 if (unlikely(unmapped_bytes < io->res))
631 io->res = unmapped_bytes;
633 if (blk_update_request(req, BLK_STS_OK, io->res))
634 blk_mq_requeue_request(req, true);
636 __blk_mq_end_request(req, BLK_STS_OK);
640 * Since __ublk_rq_task_work always fails requests immediately during
641 * exiting, __ublk_fail_req() is only called from abort context during
642 * exiting. So lock is unnecessary.
644 * Also aborting may not be started yet, keep in mind that one failed
645 * request may be issued by block layer again.
647 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
650 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
652 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
653 io->flags |= UBLK_IO_FLAG_ABORTED;
654 if (ublk_queue_can_use_recovery_reissue(ubq))
655 blk_mq_requeue_request(req, false);
657 blk_mq_end_request(req, BLK_STS_IOERR);
661 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
663 /* mark this cmd owned by ublksrv */
664 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
667 * clear ACTIVE since we are done with this sqe/cmd slot
668 * We can only accept io cmd in case of being not active.
670 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
672 /* tell ublksrv one io request is coming */
673 io_uring_cmd_done(io->cmd, res, 0);
676 #define UBLK_REQUEUE_DELAY_MS 3
678 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
681 /* We cannot process this rq so just requeue it. */
682 if (ublk_queue_can_use_recovery(ubq))
683 blk_mq_requeue_request(rq, false);
685 blk_mq_end_request(rq, BLK_STS_IOERR);
687 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
690 static inline void __ublk_rq_task_work(struct request *req)
692 struct ublk_queue *ubq = req->mq_hctx->driver_data;
694 struct ublk_io *io = &ubq->ios[tag];
695 unsigned int mapped_bytes;
697 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
698 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
699 ublk_get_iod(ubq, req->tag)->addr);
702 * Task is exiting if either:
704 * (1) current != ubq_daemon.
705 * io_uring_cmd_complete_in_task() tries to run task_work
706 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
708 * (2) current->flags & PF_EXITING.
710 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
711 __ublk_abort_rq(ubq, req);
715 if (ublk_need_get_data(ubq) &&
716 (req_op(req) == REQ_OP_WRITE ||
717 req_op(req) == REQ_OP_FLUSH)) {
719 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
720 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
723 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
724 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
725 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
726 __func__, io->cmd->cmd_op, ubq->q_id,
727 req->tag, io->flags);
728 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
732 * We have handled UBLK_IO_NEED_GET_DATA command,
733 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
736 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
737 /* update iod->addr because ublksrv may have passed a new io buffer */
738 ublk_get_iod(ubq, req->tag)->addr = io->addr;
739 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
740 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
741 ublk_get_iod(ubq, req->tag)->addr);
744 mapped_bytes = ublk_map_io(ubq, req, io);
746 /* partially mapped, update io descriptor */
747 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
749 * Nothing mapped, retry until we succeed.
751 * We may never succeed in mapping any bytes here because
752 * of OOM. TODO: reserve one buffer with single page pinned
753 * for providing forward progress guarantee.
755 if (unlikely(!mapped_bytes)) {
756 blk_mq_requeue_request(req, false);
757 blk_mq_delay_kick_requeue_list(req->q,
758 UBLK_REQUEUE_DELAY_MS);
762 ublk_get_iod(ubq, req->tag)->nr_sectors =
766 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
769 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
771 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
772 struct ublk_queue *ubq = pdu->ubq;
773 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
774 struct ublk_rq_data *data;
776 llist_for_each_entry(data, io_cmds, node)
777 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
780 static void ublk_rq_task_work_fn(struct callback_head *work)
782 struct ublk_rq_data *data = container_of(work,
783 struct ublk_rq_data, work);
784 struct request *req = blk_mq_rq_from_pdu(data);
786 __ublk_rq_task_work(req);
789 static void ublk_submit_cmd(struct ublk_queue *ubq, const struct request *rq)
791 struct ublk_io *io = &ubq->ios[rq->tag];
794 * If the check pass, we know that this is a re-issued request aborted
795 * previously in monitor_work because the ubq_daemon(cmd's task) is
796 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
797 * because this ioucmd's io_uring context may be freed now if no inflight
798 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
800 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
801 * the tag). Then the request is re-started(allocating the tag) and we are here.
802 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
803 * guarantees that here is a re-issued request aborted previously.
805 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
806 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
807 struct ublk_rq_data *data;
809 llist_for_each_entry(data, io_cmds, node)
810 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
812 struct io_uring_cmd *cmd = io->cmd;
813 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
816 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
820 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq,
823 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
825 if (ublk_can_use_task_work(ubq)) {
826 enum task_work_notify_mode notify_mode = last ?
827 TWA_SIGNAL_NO_IPI : TWA_NONE;
829 if (task_work_add(ubq->ubq_daemon, &data->work, notify_mode))
830 __ublk_abort_rq(ubq, rq);
832 if (llist_add(&data->node, &ubq->io_cmds))
833 ublk_submit_cmd(ubq, rq);
837 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
838 const struct blk_mq_queue_data *bd)
840 struct ublk_queue *ubq = hctx->driver_data;
841 struct request *rq = bd->rq;
844 /* fill iod to slot in io cmd buffer */
845 res = ublk_setup_iod(ubq, rq);
846 if (unlikely(res != BLK_STS_OK))
847 return BLK_STS_IOERR;
849 /* With recovery feature enabled, force_abort is set in
850 * ublk_stop_dev() before calling del_gendisk(). We have to
851 * abort all requeued and new rqs here to let del_gendisk()
852 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
853 * to avoid UAF on io_uring ctx.
855 * Note: force_abort is guaranteed to be seen because it is set
856 * before request queue is unqiuesced.
858 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
859 return BLK_STS_IOERR;
861 blk_mq_start_request(bd->rq);
863 if (unlikely(ubq_daemon_is_dying(ubq))) {
864 __ublk_abort_rq(ubq, rq);
868 ublk_queue_cmd(ubq, rq, bd->last);
873 static void ublk_commit_rqs(struct blk_mq_hw_ctx *hctx)
875 struct ublk_queue *ubq = hctx->driver_data;
877 if (ublk_can_use_task_work(ubq))
878 __set_notify_signal(ubq->ubq_daemon);
881 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
882 unsigned int hctx_idx)
884 struct ublk_device *ub = driver_data;
885 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
887 hctx->driver_data = ubq;
891 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
892 unsigned int hctx_idx, unsigned int numa_node)
894 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
896 init_task_work(&data->work, ublk_rq_task_work_fn);
900 static const struct blk_mq_ops ublk_mq_ops = {
901 .queue_rq = ublk_queue_rq,
902 .commit_rqs = ublk_commit_rqs,
903 .init_hctx = ublk_init_hctx,
904 .init_request = ublk_init_rq,
907 static int ublk_ch_open(struct inode *inode, struct file *filp)
909 struct ublk_device *ub = container_of(inode->i_cdev,
910 struct ublk_device, cdev);
912 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
914 filp->private_data = ub;
918 static int ublk_ch_release(struct inode *inode, struct file *filp)
920 struct ublk_device *ub = filp->private_data;
922 clear_bit(UB_STATE_OPEN, &ub->state);
926 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
927 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
929 struct ublk_device *ub = filp->private_data;
930 size_t sz = vma->vm_end - vma->vm_start;
931 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
932 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
935 spin_lock(&ub->mm_lock);
937 ub->mm = current->mm;
938 if (current->mm != ub->mm)
940 spin_unlock(&ub->mm_lock);
945 if (vma->vm_flags & VM_WRITE)
948 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
949 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
952 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
953 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
954 __func__, q_id, current->pid, vma->vm_start,
955 phys_off, (unsigned long)sz);
957 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
960 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
961 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
964 static void ublk_commit_completion(struct ublk_device *ub,
965 struct ublksrv_io_cmd *ub_cmd)
967 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
968 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
969 struct ublk_io *io = &ubq->ios[tag];
972 /* now this cmd slot is owned by nbd driver */
973 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
974 io->res = ub_cmd->result;
976 /* find the io request and complete */
977 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
979 if (req && likely(!blk_should_fake_timeout(req->q)))
980 ublk_complete_rq(req);
984 * When ->ubq_daemon is exiting, either new request is ended immediately,
985 * or any queued io command is drained, so it is safe to abort queue
988 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
992 if (!ublk_get_device(ub))
995 for (i = 0; i < ubq->q_depth; i++) {
996 struct ublk_io *io = &ubq->ios[i];
998 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1002 * Either we fail the request or ublk_rq_task_work_fn
1005 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1007 __ublk_fail_req(ubq, io, rq);
1010 ublk_put_device(ub);
1013 static void ublk_daemon_monitor_work(struct work_struct *work)
1015 struct ublk_device *ub =
1016 container_of(work, struct ublk_device, monitor_work.work);
1019 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1020 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1022 if (ubq_daemon_is_dying(ubq)) {
1023 if (ublk_queue_can_use_recovery(ubq))
1024 schedule_work(&ub->quiesce_work);
1026 schedule_work(&ub->stop_work);
1028 /* abort queue is for making forward progress */
1029 ublk_abort_queue(ub, ubq);
1034 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1035 * after ublk_remove() or __ublk_quiesce_dev() is started.
1037 * No need ub->mutex, monitor work are canceled after state is marked
1038 * as not LIVE, so new state is observed reliably.
1040 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1041 schedule_delayed_work(&ub->monitor_work,
1042 UBLK_DAEMON_MONITOR_PERIOD);
1045 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1047 return ubq->nr_io_ready == ubq->q_depth;
1050 static void ublk_cancel_queue(struct ublk_queue *ubq)
1054 if (!ublk_queue_ready(ubq))
1057 for (i = 0; i < ubq->q_depth; i++) {
1058 struct ublk_io *io = &ubq->ios[i];
1060 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1061 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1064 /* all io commands are canceled */
1065 ubq->nr_io_ready = 0;
1068 /* Cancel all pending commands, must be called after del_gendisk() returns */
1069 static void ublk_cancel_dev(struct ublk_device *ub)
1073 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1074 ublk_cancel_queue(ublk_get_queue(ub, i));
1077 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1081 if (blk_mq_request_started(rq)) {
1088 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1092 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1095 blk_mq_tagset_busy_iter(&ub->tag_set,
1096 ublk_check_inflight_rq, &idle);
1099 msleep(UBLK_REQUEUE_DELAY_MS);
1103 static void __ublk_quiesce_dev(struct ublk_device *ub)
1105 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1106 __func__, ub->dev_info.dev_id,
1107 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1108 "LIVE" : "QUIESCED");
1109 blk_mq_quiesce_queue(ub->ub_disk->queue);
1110 ublk_wait_tagset_rqs_idle(ub);
1111 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1112 ublk_cancel_dev(ub);
1113 /* we are going to release task_struct of ubq_daemon and resets
1114 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1115 * Besides, monitor_work is not necessary in QUIESCED state since we have
1116 * already scheduled quiesce_work and quiesced all ubqs.
1118 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1119 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1121 cancel_delayed_work_sync(&ub->monitor_work);
1124 static void ublk_quiesce_work_fn(struct work_struct *work)
1126 struct ublk_device *ub =
1127 container_of(work, struct ublk_device, quiesce_work);
1129 mutex_lock(&ub->mutex);
1130 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1132 __ublk_quiesce_dev(ub);
1134 mutex_unlock(&ub->mutex);
1137 static void ublk_unquiesce_dev(struct ublk_device *ub)
1141 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1142 __func__, ub->dev_info.dev_id,
1143 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1144 "LIVE" : "QUIESCED");
1145 /* quiesce_work has run. We let requeued rqs be aborted
1146 * before running fallback_wq. "force_abort" must be seen
1147 * after request queue is unqiuesced. Then del_gendisk()
1150 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1151 ublk_get_queue(ub, i)->force_abort = true;
1153 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1154 /* We may have requeued some rqs in ublk_quiesce_queue() */
1155 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1158 static void ublk_stop_dev(struct ublk_device *ub)
1160 mutex_lock(&ub->mutex);
1161 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1163 if (ublk_can_use_recovery(ub)) {
1164 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1165 __ublk_quiesce_dev(ub);
1166 ublk_unquiesce_dev(ub);
1168 del_gendisk(ub->ub_disk);
1169 ub->dev_info.state = UBLK_S_DEV_DEAD;
1170 ub->dev_info.ublksrv_pid = -1;
1171 put_disk(ub->ub_disk);
1174 ublk_cancel_dev(ub);
1175 mutex_unlock(&ub->mutex);
1176 cancel_delayed_work_sync(&ub->monitor_work);
1179 /* device can only be started after all IOs are ready */
1180 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1182 mutex_lock(&ub->mutex);
1184 if (ublk_queue_ready(ubq)) {
1185 ubq->ubq_daemon = current;
1186 get_task_struct(ubq->ubq_daemon);
1187 ub->nr_queues_ready++;
1189 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1190 complete_all(&ub->completion);
1191 mutex_unlock(&ub->mutex);
1194 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1197 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1198 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1200 ublk_queue_cmd(ubq, req, true);
1203 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1205 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1206 struct ublk_device *ub = cmd->file->private_data;
1207 struct ublk_queue *ubq;
1209 u32 cmd_op = cmd->cmd_op;
1210 unsigned tag = ub_cmd->tag;
1213 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1214 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1217 if (!(issue_flags & IO_URING_F_SQE128))
1220 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1223 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1224 if (!ubq || ub_cmd->q_id != ubq->q_id)
1227 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1230 if (tag >= ubq->q_depth)
1233 io = &ubq->ios[tag];
1235 /* there is pending io cmd, something must be wrong */
1236 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1242 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1243 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1245 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1246 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1250 case UBLK_IO_FETCH_REQ:
1251 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1252 if (ublk_queue_ready(ubq)) {
1257 * The io is being handled by server, so COMMIT_RQ is expected
1258 * instead of FETCH_REQ
1260 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1262 /* FETCH_RQ has to provide IO buffer */
1266 io->flags |= UBLK_IO_FLAG_ACTIVE;
1267 io->addr = ub_cmd->addr;
1269 ublk_mark_io_ready(ub, ubq);
1271 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1272 /* FETCH_RQ has to provide IO buffer */
1275 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1277 io->addr = ub_cmd->addr;
1278 io->flags |= UBLK_IO_FLAG_ACTIVE;
1280 ublk_commit_completion(ub, ub_cmd);
1282 case UBLK_IO_NEED_GET_DATA:
1283 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1285 io->addr = ub_cmd->addr;
1287 io->flags |= UBLK_IO_FLAG_ACTIVE;
1288 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1293 return -EIOCBQUEUED;
1296 io_uring_cmd_done(cmd, ret, 0);
1297 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1298 __func__, cmd_op, tag, ret, io->flags);
1299 return -EIOCBQUEUED;
1302 static const struct file_operations ublk_ch_fops = {
1303 .owner = THIS_MODULE,
1304 .open = ublk_ch_open,
1305 .release = ublk_ch_release,
1306 .llseek = no_llseek,
1307 .uring_cmd = ublk_ch_uring_cmd,
1308 .mmap = ublk_ch_mmap,
1311 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1313 int size = ublk_queue_cmd_buf_size(ub, q_id);
1314 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1316 if (ubq->ubq_daemon)
1317 put_task_struct(ubq->ubq_daemon);
1318 if (ubq->io_cmd_buf)
1319 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1322 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1324 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1325 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1329 ubq->flags = ub->dev_info.flags;
1331 ubq->q_depth = ub->dev_info.queue_depth;
1332 size = ublk_queue_cmd_buf_size(ub, q_id);
1334 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1338 ubq->io_cmd_buf = ptr;
1343 static void ublk_deinit_queues(struct ublk_device *ub)
1345 int nr_queues = ub->dev_info.nr_hw_queues;
1351 for (i = 0; i < nr_queues; i++)
1352 ublk_deinit_queue(ub, i);
1353 kfree(ub->__queues);
1356 static int ublk_init_queues(struct ublk_device *ub)
1358 int nr_queues = ub->dev_info.nr_hw_queues;
1359 int depth = ub->dev_info.queue_depth;
1360 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1361 int i, ret = -ENOMEM;
1363 ub->queue_size = ubq_size;
1364 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1368 for (i = 0; i < nr_queues; i++) {
1369 if (ublk_init_queue(ub, i))
1373 init_completion(&ub->completion);
1377 ublk_deinit_queues(ub);
1381 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1386 spin_lock(&ublk_idr_lock);
1387 /* allocate id, if @id >= 0, we're requesting that specific id */
1389 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1393 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1395 spin_unlock(&ublk_idr_lock);
1398 ub->ub_number = err;
1403 static void ublk_free_dev_number(struct ublk_device *ub)
1405 spin_lock(&ublk_idr_lock);
1406 idr_remove(&ublk_index_idr, ub->ub_number);
1407 wake_up_all(&ublk_idr_wq);
1408 spin_unlock(&ublk_idr_lock);
1411 static void ublk_cdev_rel(struct device *dev)
1413 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1415 blk_mq_free_tag_set(&ub->tag_set);
1416 ublk_deinit_queues(ub);
1417 ublk_free_dev_number(ub);
1418 mutex_destroy(&ub->mutex);
1422 static int ublk_add_chdev(struct ublk_device *ub)
1424 struct device *dev = &ub->cdev_dev;
1425 int minor = ub->ub_number;
1428 dev->parent = ublk_misc.this_device;
1429 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1430 dev->class = ublk_chr_class;
1431 dev->release = ublk_cdev_rel;
1432 device_initialize(dev);
1434 ret = dev_set_name(dev, "ublkc%d", minor);
1438 cdev_init(&ub->cdev, &ublk_ch_fops);
1439 ret = cdev_device_add(&ub->cdev, dev);
1448 static void ublk_stop_work_fn(struct work_struct *work)
1450 struct ublk_device *ub =
1451 container_of(work, struct ublk_device, stop_work);
1456 /* align max io buffer size with PAGE_SIZE */
1457 static void ublk_align_max_io_size(struct ublk_device *ub)
1459 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1461 ub->dev_info.max_io_buf_bytes =
1462 round_down(max_io_bytes, PAGE_SIZE);
1465 static int ublk_add_tag_set(struct ublk_device *ub)
1467 ub->tag_set.ops = &ublk_mq_ops;
1468 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1469 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1470 ub->tag_set.numa_node = NUMA_NO_NODE;
1471 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1472 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1473 ub->tag_set.driver_data = ub;
1474 return blk_mq_alloc_tag_set(&ub->tag_set);
1477 static void ublk_remove(struct ublk_device *ub)
1480 cancel_work_sync(&ub->stop_work);
1481 cancel_work_sync(&ub->quiesce_work);
1482 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1483 put_device(&ub->cdev_dev);
1486 static struct ublk_device *ublk_get_device_from_id(int idx)
1488 struct ublk_device *ub = NULL;
1493 spin_lock(&ublk_idr_lock);
1494 ub = idr_find(&ublk_index_idr, idx);
1496 ub = ublk_get_device(ub);
1497 spin_unlock(&ublk_idr_lock);
1502 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1504 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1505 int ublksrv_pid = (int)header->data[0];
1506 struct ublk_device *ub;
1507 struct gendisk *disk;
1510 if (ublksrv_pid <= 0)
1513 ub = ublk_get_device_from_id(header->dev_id);
1517 wait_for_completion_interruptible(&ub->completion);
1519 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1521 mutex_lock(&ub->mutex);
1522 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1523 test_bit(UB_STATE_USED, &ub->state)) {
1528 disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1530 ret = PTR_ERR(disk);
1533 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1534 disk->fops = &ub_fops;
1535 disk->private_data = ub;
1537 ub->dev_info.ublksrv_pid = ublksrv_pid;
1540 ret = ublk_apply_params(ub);
1544 get_device(&ub->cdev_dev);
1545 ret = add_disk(disk);
1548 * Has to drop the reference since ->free_disk won't be
1549 * called in case of add_disk failure.
1551 ublk_put_device(ub);
1554 set_bit(UB_STATE_USED, &ub->state);
1555 ub->dev_info.state = UBLK_S_DEV_LIVE;
1560 mutex_unlock(&ub->mutex);
1561 ublk_put_device(ub);
1565 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1567 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1568 void __user *argp = (void __user *)(unsigned long)header->addr;
1569 struct ublk_device *ub;
1570 cpumask_var_t cpumask;
1571 unsigned long queue;
1572 unsigned int retlen;
1576 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1578 if (header->len & (sizeof(unsigned long)-1))
1583 ub = ublk_get_device_from_id(header->dev_id);
1587 queue = header->data[0];
1588 if (queue >= ub->dev_info.nr_hw_queues)
1589 goto out_put_device;
1592 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1593 goto out_put_device;
1595 for_each_possible_cpu(i) {
1596 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1597 cpumask_set_cpu(i, cpumask);
1601 retlen = min_t(unsigned short, header->len, cpumask_size());
1602 if (copy_to_user(argp, cpumask, retlen))
1603 goto out_free_cpumask;
1604 if (retlen != header->len &&
1605 clear_user(argp + retlen, header->len - retlen))
1606 goto out_free_cpumask;
1610 free_cpumask_var(cpumask);
1612 ublk_put_device(ub);
1616 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1618 pr_devel("%s: dev id %d flags %llx\n", __func__,
1619 info->dev_id, info->flags);
1620 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1621 info->nr_hw_queues, info->queue_depth);
1624 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1626 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1627 void __user *argp = (void __user *)(unsigned long)header->addr;
1628 struct ublksrv_ctrl_dev_info info;
1629 struct ublk_device *ub;
1632 if (header->len < sizeof(info) || !header->addr)
1634 if (header->queue_id != (u16)-1) {
1635 pr_warn("%s: queue_id is wrong %x\n",
1636 __func__, header->queue_id);
1639 if (copy_from_user(&info, argp, sizeof(info)))
1641 ublk_dump_dev_info(&info);
1642 if (header->dev_id != info.dev_id) {
1643 pr_warn("%s: dev id not match %u %u\n",
1644 __func__, header->dev_id, info.dev_id);
1648 ret = mutex_lock_killable(&ublk_ctl_mutex);
1653 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1656 mutex_init(&ub->mutex);
1657 spin_lock_init(&ub->mm_lock);
1658 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1659 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1660 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1662 ret = ublk_alloc_dev_number(ub, header->dev_id);
1666 memcpy(&ub->dev_info, &info, sizeof(info));
1668 /* update device id */
1669 ub->dev_info.dev_id = ub->ub_number;
1672 * 64bit flags will be copied back to userspace as feature
1673 * negotiation result, so have to clear flags which driver
1674 * doesn't support yet, then userspace can get correct flags
1675 * (features) to handle.
1677 ub->dev_info.flags &= UBLK_F_ALL;
1679 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1680 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1682 /* We are not ready to support zero copy */
1683 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1685 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1686 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1687 ublk_align_max_io_size(ub);
1689 ret = ublk_init_queues(ub);
1691 goto out_free_dev_number;
1693 ret = ublk_add_tag_set(ub);
1695 goto out_deinit_queues;
1698 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1699 goto out_free_tag_set;
1702 * Add the char dev so that ublksrv daemon can be setup.
1703 * ublk_add_chdev() will cleanup everything if it fails.
1705 ret = ublk_add_chdev(ub);
1709 blk_mq_free_tag_set(&ub->tag_set);
1711 ublk_deinit_queues(ub);
1712 out_free_dev_number:
1713 ublk_free_dev_number(ub);
1715 mutex_destroy(&ub->mutex);
1718 mutex_unlock(&ublk_ctl_mutex);
1722 static inline bool ublk_idr_freed(int id)
1726 spin_lock(&ublk_idr_lock);
1727 ptr = idr_find(&ublk_index_idr, id);
1728 spin_unlock(&ublk_idr_lock);
1733 static int ublk_ctrl_del_dev(int idx)
1735 struct ublk_device *ub;
1738 ret = mutex_lock_killable(&ublk_ctl_mutex);
1742 ub = ublk_get_device_from_id(idx);
1745 ublk_put_device(ub);
1752 * Wait until the idr is removed, then it can be reused after
1753 * DEL_DEV command is returned.
1756 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1757 mutex_unlock(&ublk_ctl_mutex);
1762 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1764 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1766 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1767 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1768 header->data[0], header->addr, header->len);
1771 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1773 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1774 struct ublk_device *ub;
1776 ub = ublk_get_device_from_id(header->dev_id);
1781 cancel_work_sync(&ub->stop_work);
1782 cancel_work_sync(&ub->quiesce_work);
1784 ublk_put_device(ub);
1788 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1790 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1791 void __user *argp = (void __user *)(unsigned long)header->addr;
1792 struct ublk_device *ub;
1795 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1798 ub = ublk_get_device_from_id(header->dev_id);
1802 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1804 ublk_put_device(ub);
1809 static int ublk_ctrl_get_params(struct io_uring_cmd *cmd)
1811 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1812 void __user *argp = (void __user *)(unsigned long)header->addr;
1813 struct ublk_params_header ph;
1814 struct ublk_device *ub;
1817 if (header->len <= sizeof(ph) || !header->addr)
1820 if (copy_from_user(&ph, argp, sizeof(ph)))
1823 if (ph.len > header->len || !ph.len)
1826 if (ph.len > sizeof(struct ublk_params))
1827 ph.len = sizeof(struct ublk_params);
1829 ub = ublk_get_device_from_id(header->dev_id);
1833 mutex_lock(&ub->mutex);
1834 if (copy_to_user(argp, &ub->params, ph.len))
1838 mutex_unlock(&ub->mutex);
1840 ublk_put_device(ub);
1844 static int ublk_ctrl_set_params(struct io_uring_cmd *cmd)
1846 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1847 void __user *argp = (void __user *)(unsigned long)header->addr;
1848 struct ublk_params_header ph;
1849 struct ublk_device *ub;
1852 if (header->len <= sizeof(ph) || !header->addr)
1855 if (copy_from_user(&ph, argp, sizeof(ph)))
1858 if (ph.len > header->len || !ph.len || !ph.types)
1861 if (ph.len > sizeof(struct ublk_params))
1862 ph.len = sizeof(struct ublk_params);
1864 ub = ublk_get_device_from_id(header->dev_id);
1868 /* parameters can only be changed when device isn't live */
1869 mutex_lock(&ub->mutex);
1870 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1872 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1875 /* clear all we don't support yet */
1876 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1877 ret = ublk_validate_params(ub);
1879 mutex_unlock(&ub->mutex);
1880 ublk_put_device(ub);
1885 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1889 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1890 /* All old ioucmds have to be completed */
1891 WARN_ON_ONCE(ubq->nr_io_ready);
1892 /* old daemon is PF_EXITING, put it now */
1893 put_task_struct(ubq->ubq_daemon);
1894 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1895 ubq->ubq_daemon = NULL;
1897 for (i = 0; i < ubq->q_depth; i++) {
1898 struct ublk_io *io = &ubq->ios[i];
1900 /* forget everything now and be ready for new FETCH_REQ */
1907 static int ublk_ctrl_start_recovery(struct io_uring_cmd *cmd)
1909 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1910 struct ublk_device *ub;
1914 ub = ublk_get_device_from_id(header->dev_id);
1918 mutex_lock(&ub->mutex);
1919 if (!ublk_can_use_recovery(ub))
1922 * START_RECOVERY is only allowd after:
1924 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1925 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1928 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1929 * (a)has quiesced request queue
1930 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1931 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1932 * (d)has completed/camceled all ioucmds owned by ther dying process
1934 if (test_bit(UB_STATE_OPEN, &ub->state) ||
1935 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1939 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1940 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1941 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1942 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1944 ub->nr_queues_ready = 0;
1945 init_completion(&ub->completion);
1948 mutex_unlock(&ub->mutex);
1949 ublk_put_device(ub);
1953 static int ublk_ctrl_end_recovery(struct io_uring_cmd *cmd)
1955 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1956 int ublksrv_pid = (int)header->data[0];
1957 struct ublk_device *ub;
1960 ub = ublk_get_device_from_id(header->dev_id);
1964 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1965 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1966 /* wait until new ubq_daemon sending all FETCH_REQ */
1967 wait_for_completion_interruptible(&ub->completion);
1968 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
1969 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1971 mutex_lock(&ub->mutex);
1972 if (!ublk_can_use_recovery(ub))
1975 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1979 ub->dev_info.ublksrv_pid = ublksrv_pid;
1980 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
1981 __func__, ublksrv_pid, header->dev_id);
1982 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1983 pr_devel("%s: queue unquiesced, dev id %d.\n",
1984 __func__, header->dev_id);
1985 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1986 ub->dev_info.state = UBLK_S_DEV_LIVE;
1987 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1990 mutex_unlock(&ub->mutex);
1991 ublk_put_device(ub);
1995 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1996 unsigned int issue_flags)
1998 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2001 ublk_ctrl_cmd_dump(cmd);
2003 if (!(issue_flags & IO_URING_F_SQE128))
2007 if (!capable(CAP_SYS_ADMIN))
2011 switch (cmd->cmd_op) {
2012 case UBLK_CMD_START_DEV:
2013 ret = ublk_ctrl_start_dev(cmd);
2015 case UBLK_CMD_STOP_DEV:
2016 ret = ublk_ctrl_stop_dev(cmd);
2018 case UBLK_CMD_GET_DEV_INFO:
2019 ret = ublk_ctrl_get_dev_info(cmd);
2021 case UBLK_CMD_ADD_DEV:
2022 ret = ublk_ctrl_add_dev(cmd);
2024 case UBLK_CMD_DEL_DEV:
2025 ret = ublk_ctrl_del_dev(header->dev_id);
2027 case UBLK_CMD_GET_QUEUE_AFFINITY:
2028 ret = ublk_ctrl_get_queue_affinity(cmd);
2030 case UBLK_CMD_GET_PARAMS:
2031 ret = ublk_ctrl_get_params(cmd);
2033 case UBLK_CMD_SET_PARAMS:
2034 ret = ublk_ctrl_set_params(cmd);
2036 case UBLK_CMD_START_USER_RECOVERY:
2037 ret = ublk_ctrl_start_recovery(cmd);
2039 case UBLK_CMD_END_USER_RECOVERY:
2040 ret = ublk_ctrl_end_recovery(cmd);
2046 io_uring_cmd_done(cmd, ret, 0);
2047 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2048 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2049 return -EIOCBQUEUED;
2052 static const struct file_operations ublk_ctl_fops = {
2053 .open = nonseekable_open,
2054 .uring_cmd = ublk_ctrl_uring_cmd,
2055 .owner = THIS_MODULE,
2056 .llseek = noop_llseek,
2059 static struct miscdevice ublk_misc = {
2060 .minor = MISC_DYNAMIC_MINOR,
2061 .name = "ublk-control",
2062 .fops = &ublk_ctl_fops,
2065 static int __init ublk_init(void)
2069 init_waitqueue_head(&ublk_idr_wq);
2071 ret = misc_register(&ublk_misc);
2075 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2077 goto unregister_mis;
2079 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2080 if (IS_ERR(ublk_chr_class)) {
2081 ret = PTR_ERR(ublk_chr_class);
2082 goto free_chrdev_region;
2087 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2089 misc_deregister(&ublk_misc);
2093 static void __exit ublk_exit(void)
2095 struct ublk_device *ub;
2098 class_destroy(ublk_chr_class);
2100 misc_deregister(&ublk_misc);
2102 idr_for_each_entry(&ublk_index_idr, ub, id)
2105 idr_destroy(&ublk_index_idr);
2106 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2109 module_init(ublk_init);
2110 module_exit(ublk_exit);
2112 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2113 MODULE_LICENSE("GPL");