projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ublk_drv: add mechanism for supporting unprivileged ublk device
[linux-block.git] / drivers / block / ublk_drv.c
CommitLineData
71f28f31
ML		 1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Userspace block device - block device which IO is handled from userspace
4 *
5 * Take full use of io_uring passthrough command for communicating with
6 * ublk userspace daemon(ublksrvd) for handling basic IO request.
7 *
8 * Copyright 2022 Ming Lei <ming.lei@redhat.com>
9 *
10 * (part of code stolen from loop.c)
11 */
12#include <linux/module.h>
13#include <linux/moduleparam.h>
14#include <linux/sched.h>
15#include <linux/fs.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>
42#include <linux/mm.h>
43#include <asm/page.h>
0edb3696 44#include <linux/task_work.h>
4093cb5a 45#include <linux/namei.h>
71f28f31
ML		 46#include <uapi/linux/ublk_cmd.h>
47
48#define UBLK_MINORS (1U << MINORBITS)
49
6d8c5afc 50/* All UBLK_F_* have to be included into UBLK_F_ALL */
c86019ff
Z		 51#define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
52 | UBLK_F_URING_CMD_COMP_IN_TASK \
77a440e2 53 | UBLK_F_NEED_GET_DATA \
a0d41dc1 54 | UBLK_F_USER_RECOVERY \
4093cb5a
ML		 55 | UBLK_F_USER_RECOVERY_REISSUE \
56 | UBLK_F_UNPRIVILEGED_DEV)
6d8c5afc 57
0aa73170 58/* All UBLK_PARAM_TYPE_* should be included here */
abb864d3
ML		 59#define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | \
60 UBLK_PARAM_TYPE_DISCARD | UBLK_PARAM_TYPE_DEVT)
0aa73170 61
0edb3696 62struct ublk_rq_data {
7d4a9317
ML		 63 struct llist_node node;
64 struct callback_head work;
0edb3696
ML		 65};
66
71f28f31 67struct ublk_uring_cmd_pdu {
3ab6e94c 68 struct ublk_queue *ubq;
71f28f31
ML		 69};
70
71/*
72 * io command is active: sqe cmd is received, and its cqe isn't done
73 *
74 * If the flag is set, the io command is owned by ublk driver, and waited
75 * for incoming blk-mq request from the ublk block device.
76 *
77 * If the flag is cleared, the io command will be completed, and owned by
78 * ublk server.
79 */
80#define UBLK_IO_FLAG_ACTIVE 0x01
81
82/*
83 * IO command is completed via cqe, and it is being handled by ublksrv, and
84 * not committed yet
85 *
86 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
87 * cross verification
88 */
89#define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
90
91/*
92 * IO command is aborted, so this flag is set in case of
93 * !UBLK_IO_FLAG_ACTIVE.
94 *
95 * After this flag is observed, any pending or new incoming request
96 * associated with this io command will be failed immediately
97 */
98#define UBLK_IO_FLAG_ABORTED 0x04
99
c86019ff
Z		 100/*
101 * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
102 * get data buffer address from ublksrv.
103 *
104 * Then, bio data could be copied into this data buffer for a WRITE request
105 * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
106 */
107#define UBLK_IO_FLAG_NEED_GET_DATA 0x08
108
71f28f31
ML		 109struct ublk_io {
110 /* userspace buffer address from io cmd */
111 __u64 addr;
112 unsigned int flags;
113 int res;
114
115 struct io_uring_cmd *cmd;
116};
117
118struct ublk_queue {
119 int q_id;
120 int q_depth;
121
0edb3696 122 unsigned long flags;
71f28f31
ML		 123 struct task_struct *ubq_daemon;
124 char *io_cmd_buf;
125
3ab6e94c
ML		 126 struct llist_head io_cmds;
127
71f28f31
ML		 128 unsigned long io_addr; /* mapped vm address */
129 unsigned int max_io_sz;
bbae8d1f 130 bool force_abort;
71f28f31
ML		 131 unsigned short nr_io_ready; /* how many ios setup */
132 struct ublk_device *dev;
72495b5a 133 struct ublk_io ios[];
71f28f31
ML		 134};
135
136#define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
137
138struct ublk_device {
139 struct gendisk *ub_disk;
71f28f31
ML		 140
141 char *__queues;
142
143 unsigned short queue_size;
71f28f31
ML		 144 struct ublksrv_ctrl_dev_info dev_info;
145
146 struct blk_mq_tag_set tag_set;
147
148 struct cdev cdev;
149 struct device cdev_dev;
150
8d9fdb60
DC		 151#define UB_STATE_OPEN 0
152#define UB_STATE_USED 1
fa362045 153 unsigned long state;
71f28f31
ML		 154 int ub_number;
155
156 struct mutex mutex;
157
e94eb459 158 spinlock_t mm_lock;
71f28f31
ML		 159 struct mm_struct *mm;
160
0aa73170
ML		 161 struct ublk_params params;
162
71f28f31
ML		 163 struct completion completion;
164 unsigned int nr_queues_ready;
73a166d9 165 unsigned int nr_privileged_daemon;
71f28f31
ML		 166
167 /*
168 * Our ubq->daemon may be killed without any notification, so
169 * monitor each queue's daemon periodically
170 */
171 struct delayed_work monitor_work;
bbae8d1f 172 struct work_struct quiesce_work;
71f28f31
ML		 173 struct work_struct stop_work;
174};
175
0aa73170
ML		 176/* header of ublk_params */
177struct ublk_params_header {
178 __u32 len;
179 __u32 types;
180};
181
71f28f31
ML		 182static dev_t ublk_chr_devt;
183static struct class *ublk_chr_class;
184
185static DEFINE_IDR(ublk_index_idr);
186static DEFINE_SPINLOCK(ublk_idr_lock);
187static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
188
189static DEFINE_MUTEX(ublk_ctl_mutex);
190
403ebc87
ML		 191/*
192 * Max ublk devices allowed to add
193 *
194 * It can be extended to one per-user limit in future or even controlled
195 * by cgroup.
196 */
197static unsigned int ublks_max = 64;
198static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
199
71f28f31
ML		 200static struct miscdevice ublk_misc;
201
0aa73170
ML		 202static void ublk_dev_param_basic_apply(struct ublk_device *ub)
203{
204 struct request_queue *q = ub->ub_disk->queue;
205 const struct ublk_param_basic *p = &ub->params.basic;
206
207 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
208 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
209 blk_queue_io_min(q, 1 << p->io_min_shift);
210 blk_queue_io_opt(q, 1 << p->io_opt_shift);
211
212 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
213 p->attrs & UBLK_ATTR_FUA);
214 if (p->attrs & UBLK_ATTR_ROTATIONAL)
215 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
216 else
217 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
218
219 blk_queue_max_hw_sectors(q, p->max_sectors);
220 blk_queue_chunk_sectors(q, p->chunk_sectors);
221 blk_queue_virt_boundary(q, p->virt_boundary_mask);
222
223 if (p->attrs & UBLK_ATTR_READ_ONLY)
224 set_disk_ro(ub->ub_disk, true);
225
226 set_capacity(ub->ub_disk, p->dev_sectors);
227}
228
229static void ublk_dev_param_discard_apply(struct ublk_device *ub)
230{
231 struct request_queue *q = ub->ub_disk->queue;
232 const struct ublk_param_discard *p = &ub->params.discard;
233
234 q->limits.discard_alignment = p->discard_alignment;
235 q->limits.discard_granularity = p->discard_granularity;
236 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
237 blk_queue_max_write_zeroes_sectors(q,
238 p->max_write_zeroes_sectors);
239 blk_queue_max_discard_segments(q, p->max_discard_segments);
240}
241
242static int ublk_validate_params(const struct ublk_device *ub)
243{
244 /* basic param is the only one which must be set */
245 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
246 const struct ublk_param_basic *p = &ub->params.basic;
247
248 if (p->logical_bs_shift > PAGE_SHIFT)
249 return -EINVAL;
250
251 if (p->logical_bs_shift > p->physical_bs_shift)
252 return -EINVAL;
253
4bf9cbf3 254 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
0aa73170
ML		 255 return -EINVAL;
256 } else
257 return -EINVAL;
258
259 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
260 const struct ublk_param_discard *p = &ub->params.discard;
261
262 /* So far, only support single segment discard */
263 if (p->max_discard_sectors && p->max_discard_segments != 1)
264 return -EINVAL;
265
266 if (!p->discard_granularity)
267 return -EINVAL;
268 }
269
abb864d3
ML		 270 /* dev_t is read-only */
271 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
272 return -EINVAL;
273
0aa73170
ML		 274 return 0;
275}
276
277static int ublk_apply_params(struct ublk_device *ub)
278{
279 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
280 return -EINVAL;
281
282 ublk_dev_param_basic_apply(ub);
283
284 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
285 ublk_dev_param_discard_apply(ub);
286
287 return 0;
288}
289
0edb3696
ML		 290static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
291{
292 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
293 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
294 return true;
295 return false;
296}
297
c86019ff
Z		 298static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
299{
300 if (ubq->flags & UBLK_F_NEED_GET_DATA)
301 return true;
302 return false;
303}
304
71f28f31
ML		 305static struct ublk_device *ublk_get_device(struct ublk_device *ub)
306{
307 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
308 return ub;
309 return NULL;
310}
311
312static void ublk_put_device(struct ublk_device *ub)
313{
314 put_device(&ub->cdev_dev);
315}
316
317static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
318 int qid)
319{
320 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
321}
322
323static inline bool ublk_rq_has_data(const struct request *rq)
324{
325 return rq->bio && bio_has_data(rq->bio);
326}
327
328static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
329 int tag)
330{
331 return (struct ublksrv_io_desc *)
332 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
333}
334
335static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
336{
337 return ublk_get_queue(ub, q_id)->io_cmd_buf;
338}
339
340static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
341{
342 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
343
344 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
345 PAGE_SIZE);
346}
347
a0d41dc1
Z		 348static inline bool ublk_queue_can_use_recovery_reissue(
349 struct ublk_queue *ubq)
350{
351 if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
352 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
353 return true;
354 return false;
355}
356
77a440e2
Z		 357static inline bool ublk_queue_can_use_recovery(
358 struct ublk_queue *ubq)
359{
360 if (ubq->flags & UBLK_F_USER_RECOVERY)
361 return true;
362 return false;
363}
364
365static inline bool ublk_can_use_recovery(struct ublk_device *ub)
366{
367 if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
368 return true;
369 return false;
370}
371
6d9e6dfd
CH		 372static void ublk_free_disk(struct gendisk *disk)
373{
374 struct ublk_device *ub = disk->private_data;
375
376 clear_bit(UB_STATE_USED, &ub->state);
377 put_device(&ub->cdev_dev);
378}
379
71f28f31
ML		 380static const struct block_device_operations ub_fops = {
381 .owner = THIS_MODULE,
6d9e6dfd 382 .free_disk = ublk_free_disk,
71f28f31
ML		 383};
384
385#define UBLK_MAX_PIN_PAGES 32
386
387struct ublk_map_data {
388 const struct ublk_queue *ubq;
389 const struct request *rq;
390 const struct ublk_io *io;
391 unsigned max_bytes;
392};
393
394struct ublk_io_iter {
395 struct page *pages[UBLK_MAX_PIN_PAGES];
396 unsigned pg_off; /* offset in the 1st page in pages */
397 int nr_pages; /* how many page pointers in pages */
398 struct bio *bio;
399 struct bvec_iter iter;
400};
401
402static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
403 unsigned max_bytes, bool to_vm)
404{
405 const unsigned total = min_t(unsigned, max_bytes,
406 PAGE_SIZE - data->pg_off +
407 ((data->nr_pages - 1) << PAGE_SHIFT));
408 unsigned done = 0;
409 unsigned pg_idx = 0;
410
411 while (done < total) {
412 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
413 const unsigned int bytes = min3(bv.bv_len, total - done,
414 (unsigned)(PAGE_SIZE - data->pg_off));
415 void *bv_buf = bvec_kmap_local(&bv);
416 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
417
418 if (to_vm)
419 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
420 else
421 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
422
423 kunmap_local(pg_buf);
424 kunmap_local(bv_buf);
425
426 /* advance page array */
427 data->pg_off += bytes;
428 if (data->pg_off == PAGE_SIZE) {
429 pg_idx += 1;
430 data->pg_off = 0;
431 }
432
433 done += bytes;
434
435 /* advance bio */
436 bio_advance_iter_single(data->bio, &data->iter, bytes);
437 if (!data->iter.bi_size) {
438 data->bio = data->bio->bi_next;
439 if (data->bio == NULL)
440 break;
441 data->iter = data->bio->bi_iter;
442 }
443 }
444
445 return done;
446}
447
448static inline int ublk_copy_user_pages(struct ublk_map_data *data,
449 bool to_vm)
450{
451 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
452 const unsigned long start_vm = data->io->addr;
453 unsigned int done = 0;
454 struct ublk_io_iter iter = {
455 .pg_off = start_vm & (PAGE_SIZE - 1),
456 .bio = data->rq->bio,
457 .iter = data->rq->bio->bi_iter,
458 };
459 const unsigned int nr_pages = round_up(data->max_bytes +
460 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
461
462 while (done < nr_pages) {
463 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
464 nr_pages - done);
465 unsigned i, len;
466
467 iter.nr_pages = get_user_pages_fast(start_vm +
468 (done << PAGE_SHIFT), to_pin, gup_flags,
469 iter.pages);
470 if (iter.nr_pages <= 0)
471 return done == 0 ? iter.nr_pages : done;
472 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
473 for (i = 0; i < iter.nr_pages; i++) {
474 if (to_vm)
475 set_page_dirty(iter.pages[i]);
476 put_page(iter.pages[i]);
477 }
478 data->max_bytes -= len;
479 done += iter.nr_pages;
480 }
481
482 return done;
483}
484
485static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
486 struct ublk_io *io)
487{
488 const unsigned int rq_bytes = blk_rq_bytes(req);
489 /*
490 * no zero copy, we delay copy WRITE request data into ublksrv
491 * context and the big benefit is that pinning pages in current
492 * context is pretty fast, see ublk_pin_user_pages
493 */
494 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
495 return rq_bytes;
496
497 if (ublk_rq_has_data(req)) {
498 struct ublk_map_data data = {
499 .ubq = ubq,
500 .rq = req,
501 .io = io,
502 .max_bytes = rq_bytes,
503 };
504
505 ublk_copy_user_pages(&data, true);
506
507 return rq_bytes - data.max_bytes;
508 }
509 return rq_bytes;
510}
511
512static int ublk_unmap_io(const struct ublk_queue *ubq,
513 const struct request *req,
514 struct ublk_io *io)
515{
516 const unsigned int rq_bytes = blk_rq_bytes(req);
517
518 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
519 struct ublk_map_data data = {
520 .ubq = ubq,
521 .rq = req,
522 .io = io,
523 .max_bytes = io->res,
524 };
525
526 WARN_ON_ONCE(io->res > rq_bytes);
527
528 ublk_copy_user_pages(&data, false);
529
530 return io->res - data.max_bytes;
531 }
532 return rq_bytes;
533}
534
535static inline unsigned int ublk_req_build_flags(struct request *req)
536{
537 unsigned flags = 0;
538
539 if (req->cmd_flags & REQ_FAILFAST_DEV)
540 flags |= UBLK_IO_F_FAILFAST_DEV;
541
542 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
543 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
544
545 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
546 flags |= UBLK_IO_F_FAILFAST_DRIVER;
547
548 if (req->cmd_flags & REQ_META)
549 flags |= UBLK_IO_F_META;
550
71f28f31
ML		 551 if (req->cmd_flags & REQ_FUA)
552 flags |= UBLK_IO_F_FUA;
553
71f28f31
ML		 554 if (req->cmd_flags & REQ_NOUNMAP)
555 flags |= UBLK_IO_F_NOUNMAP;
556
557 if (req->cmd_flags & REQ_SWAP)
558 flags |= UBLK_IO_F_SWAP;
559
560 return flags;
561}
562
f2450f8a 563static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
71f28f31
ML		 564{
565 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
566 struct ublk_io *io = &ubq->ios[req->tag];
567 u32 ublk_op;
568
569 switch (req_op(req)) {
570 case REQ_OP_READ:
571 ublk_op = UBLK_IO_OP_READ;
572 break;
573 case REQ_OP_WRITE:
574 ublk_op = UBLK_IO_OP_WRITE;
575 break;
576 case REQ_OP_FLUSH:
577 ublk_op = UBLK_IO_OP_FLUSH;
578 break;
579 case REQ_OP_DISCARD:
580 ublk_op = UBLK_IO_OP_DISCARD;
581 break;
582 case REQ_OP_WRITE_ZEROES:
583 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
584 break;
585 default:
586 return BLK_STS_IOERR;
587 }
588
589 /* need to translate since kernel may change */
590 iod->op_flags = ublk_op | ublk_req_build_flags(req);
591 iod->nr_sectors = blk_rq_sectors(req);
592 iod->start_sector = blk_rq_pos(req);
593 iod->addr = io->addr;
594
595 return BLK_STS_OK;
596}
597
598static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
599 struct io_uring_cmd *ioucmd)
600{
601 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
602}
603
966120b5 604static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
71f28f31
ML		 605{
606 return ubq->ubq_daemon->flags & PF_EXITING;
607}
608
609/* todo: handle partial completion */
610static void ublk_complete_rq(struct request *req)
611{
612 struct ublk_queue *ubq = req->mq_hctx->driver_data;
613 struct ublk_io *io = &ubq->ios[req->tag];
614 unsigned int unmapped_bytes;
615
616 /* failed read IO if nothing is read */
617 if (!io->res && req_op(req) == REQ_OP_READ)
618 io->res = -EIO;
619
620 if (io->res < 0) {
621 blk_mq_end_request(req, errno_to_blk_status(io->res));
622 return;
623 }
624
625 /*
626 * FLUSH or DISCARD usually won't return bytes returned, so end them
627 * directly.
628 *
629 * Both the two needn't unmap.
630 */
631 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
632 blk_mq_end_request(req, BLK_STS_OK);
633 return;
634 }
635
636 /* for READ request, writing data in iod->addr to rq buffers */
637 unmapped_bytes = ublk_unmap_io(ubq, req, io);
638
639 /*
640 * Extremely impossible since we got data filled in just before
641 *
642 * Re-read simply for this unlikely case.
643 */
644 if (unlikely(unmapped_bytes < io->res))
645 io->res = unmapped_bytes;
646
647 if (blk_update_request(req, BLK_STS_OK, io->res))
648 blk_mq_requeue_request(req, true);
649 else
650 __blk_mq_end_request(req, BLK_STS_OK);
651}
652
653/*
bb241747
Z		 654 * Since __ublk_rq_task_work always fails requests immediately during
655 * exiting, __ublk_fail_req() is only called from abort context during
656 * exiting. So lock is unnecessary.
71f28f31
ML		 657 *
658 * Also aborting may not be started yet, keep in mind that one failed
659 * request may be issued by block layer again.
660 */
a0d41dc1
Z		 661static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
662 struct request *req)
71f28f31
ML		 663{
664 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
665
666 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
667 io->flags |= UBLK_IO_FLAG_ABORTED;
a0d41dc1
Z		 668 if (ublk_queue_can_use_recovery_reissue(ubq))
669 blk_mq_requeue_request(req, false);
670 else
671 blk_mq_end_request(req, BLK_STS_IOERR);
71f28f31
ML		 672 }
673}
674
c86019ff
Z		 675static void ubq_complete_io_cmd(struct ublk_io *io, int res)
676{
677 /* mark this cmd owned by ublksrv */
678 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
679
680 /*
681 * clear ACTIVE since we are done with this sqe/cmd slot
682 * We can only accept io cmd in case of being not active.
683 */
684 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
685
686 /* tell ublksrv one io request is coming */
687 io_uring_cmd_done(io->cmd, res, 0);
688}
689
71f28f31
ML		 690#define UBLK_REQUEUE_DELAY_MS 3
691
42cf5fc5
Z		 692static inline void __ublk_abort_rq(struct ublk_queue *ubq,
693 struct request *rq)
694{
695 /* We cannot process this rq so just requeue it. */
696 if (ublk_queue_can_use_recovery(ubq))
697 blk_mq_requeue_request(rq, false);
698 else
699 blk_mq_end_request(rq, BLK_STS_IOERR);
700
701 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
702}
703
0edb3696 704static inline void __ublk_rq_task_work(struct request *req)
71f28f31 705{
71f28f31
ML		 706 struct ublk_queue *ubq = req->mq_hctx->driver_data;
707 int tag = req->tag;
708 struct ublk_io *io = &ubq->ios[tag];
71f28f31
ML		 709 unsigned int mapped_bytes;
710
711 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
712 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
713 ublk_get_iod(ubq, req->tag)->addr);
714
ae3f7193
Z		 715 /*
716 * Task is exiting if either:
717 *
718 * (1) current != ubq_daemon.
719 * io_uring_cmd_complete_in_task() tries to run task_work
720 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
721 *
722 * (2) current->flags & PF_EXITING.
723 */
724 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
42cf5fc5 725 __ublk_abort_rq(ubq, req);
71f28f31
ML		 726 return;
727 }
728
c86019ff
Z		 729 if (ublk_need_get_data(ubq) &&
730 (req_op(req) == REQ_OP_WRITE ||
731 req_op(req) == REQ_OP_FLUSH)) {
732 /*
733 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
734 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
735 * and notify it.
736 */
737 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
738 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
739 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
740 __func__, io->cmd->cmd_op, ubq->q_id,
741 req->tag, io->flags);
742 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
743 return;
744 }
745 /*
746 * We have handled UBLK_IO_NEED_GET_DATA command,
747 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
748 * do the copy work.
749 */
750 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
92cb6e2e
Z		 751 /* update iod->addr because ublksrv may have passed a new io buffer */
752 ublk_get_iod(ubq, req->tag)->addr = io->addr;
753 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
754 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
755 ublk_get_iod(ubq, req->tag)->addr);
c86019ff
Z		 756 }
757
71f28f31
ML		 758 mapped_bytes = ublk_map_io(ubq, req, io);
759
760 /* partially mapped, update io descriptor */
761 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
762 /*
763 * Nothing mapped, retry until we succeed.
764 *
765 * We may never succeed in mapping any bytes here because
766 * of OOM. TODO: reserve one buffer with single page pinned
767 * for providing forward progress guarantee.
768 */
769 if (unlikely(!mapped_bytes)) {
770 blk_mq_requeue_request(req, false);
771 blk_mq_delay_kick_requeue_list(req->q,
772 UBLK_REQUEUE_DELAY_MS);
773 return;
774 }
775
776 ublk_get_iod(ubq, req->tag)->nr_sectors =
777 mapped_bytes >> 9;
778 }
779
c86019ff 780 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
71f28f31
ML		 781}
782
7d4a9317
ML		 783static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
784{
785 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
786 struct ublk_rq_data *data, *tmp;
787
788 io_cmds = llist_reverse_order(io_cmds);
789 llist_for_each_entry_safe(data, tmp, io_cmds, node)
790 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
791}
792
793static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
794{
795 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
796 struct ublk_rq_data *data, *tmp;
797
798 llist_for_each_entry_safe(data, tmp, io_cmds, node)
799 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
800}
801
0edb3696
ML		 802static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
803{
804 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
3ab6e94c 805 struct ublk_queue *ubq = pdu->ubq;
0edb3696 806
7d4a9317 807 ublk_forward_io_cmds(ubq);
0edb3696
ML		 808}
809
810static void ublk_rq_task_work_fn(struct callback_head *work)
811{
812 struct ublk_rq_data *data = container_of(work,
813 struct ublk_rq_data, work);
814 struct request *req = blk_mq_rq_from_pdu(data);
7d4a9317 815 struct ublk_queue *ubq = req->mq_hctx->driver_data;
0edb3696 816
7d4a9317 817 ublk_forward_io_cmds(ubq);
0edb3696
ML		 818}
819
7d4a9317 820static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
3ab6e94c 821{
7d4a9317
ML		 822 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
823 struct ublk_io *io;
3ab6e94c 824
7d4a9317
ML		 825 if (!llist_add(&data->node, &ubq->io_cmds))
826 return;
827
828 io = &ubq->ios[rq->tag];
3ab6e94c
ML		 829 /*
830 * If the check pass, we know that this is a re-issued request aborted
831 * previously in monitor_work because the ubq_daemon(cmd's task) is
832 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
833 * because this ioucmd's io_uring context may be freed now if no inflight
834 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
835 *
836 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
837 * the tag). Then the request is re-started(allocating the tag) and we are here.
838 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
839 * guarantees that here is a re-issued request aborted previously.
840 */
841 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
7d4a9317
ML		 842 ublk_abort_io_cmds(ubq);
843 } else if (ublk_can_use_task_work(ubq)) {
844 if (task_work_add(ubq->ubq_daemon, &data->work,
845 TWA_SIGNAL_NO_IPI))
846 ublk_abort_io_cmds(ubq);
3ab6e94c
ML		 847 } else {
848 struct io_uring_cmd *cmd = io->cmd;
849 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
850
851 pdu->ubq = ubq;
852 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
853 }
854}
855
71f28f31
ML		 856static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
857 const struct blk_mq_queue_data *bd)
858{
859 struct ublk_queue *ubq = hctx->driver_data;
860 struct request *rq = bd->rq;
71f28f31
ML		 861 blk_status_t res;
862
863 /* fill iod to slot in io cmd buffer */
864 res = ublk_setup_iod(ubq, rq);
865 if (unlikely(res != BLK_STS_OK))
866 return BLK_STS_IOERR;
3ab6e94c 867
bbae8d1f
Z		 868 /* With recovery feature enabled, force_abort is set in
869 * ublk_stop_dev() before calling del_gendisk(). We have to
870 * abort all requeued and new rqs here to let del_gendisk()
871 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
872 * to avoid UAF on io_uring ctx.
873 *
874 * Note: force_abort is guaranteed to be seen because it is set
875 * before request queue is unqiuesced.
876 */
877 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
878 return BLK_STS_IOERR;
71f28f31
ML		 879
880 blk_mq_start_request(bd->rq);
881
882 if (unlikely(ubq_daemon_is_dying(ubq))) {
42cf5fc5
Z		 883 __ublk_abort_rq(ubq, rq);
884 return BLK_STS_OK;
71f28f31
ML		 885 }
886
7d4a9317 887 ublk_queue_cmd(ubq, rq);
0edb3696 888
71f28f31
ML		 889 return BLK_STS_OK;
890}
891
71f28f31
ML		 892static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
893 unsigned int hctx_idx)
894{
cebbe577 895 struct ublk_device *ub = driver_data;
71f28f31
ML		 896 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
897
898 hctx->driver_data = ubq;
899 return 0;
900}
901
0edb3696
ML		 902static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
903 unsigned int hctx_idx, unsigned int numa_node)
904{
905 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
906
907 init_task_work(&data->work, ublk_rq_task_work_fn);
908 return 0;
909}
910
71f28f31
ML		 911static const struct blk_mq_ops ublk_mq_ops = {
912 .queue_rq = ublk_queue_rq,
913 .init_hctx = ublk_init_hctx,
0edb3696 914 .init_request = ublk_init_rq,
71f28f31
ML		 915};
916
917static int ublk_ch_open(struct inode *inode, struct file *filp)
918{
919 struct ublk_device *ub = container_of(inode->i_cdev,
920 struct ublk_device, cdev);
921
fa362045
CH		 922 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
923 return -EBUSY;
924 filp->private_data = ub;
925 return 0;
71f28f31
ML		 926}
927
928static int ublk_ch_release(struct inode *inode, struct file *filp)
929{
930 struct ublk_device *ub = filp->private_data;
931
fa362045 932 clear_bit(UB_STATE_OPEN, &ub->state);
71f28f31
ML		 933 return 0;
934}
935
936/* map pre-allocated per-queue cmd buffer to ublksrv daemon */
937static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
938{
939 struct ublk_device *ub = filp->private_data;
940 size_t sz = vma->vm_end - vma->vm_start;
941 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
942 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
943 int q_id, ret = 0;
944
e94eb459 945 spin_lock(&ub->mm_lock);
71f28f31
ML		 946 if (!ub->mm)
947 ub->mm = current->mm;
948 if (current->mm != ub->mm)
949 ret = -EINVAL;
e94eb459 950 spin_unlock(&ub->mm_lock);
71f28f31
ML		 951
952 if (ret)
953 return ret;
954
955 if (vma->vm_flags & VM_WRITE)
956 return -EPERM;
957
958 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
959 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
960 return -EINVAL;
961
962 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
963 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
964 __func__, q_id, current->pid, vma->vm_start,
965 phys_off, (unsigned long)sz);
966
967 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
968 return -EINVAL;
969
970 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
971 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
972}
973
974static void ublk_commit_completion(struct ublk_device *ub,
975 struct ublksrv_io_cmd *ub_cmd)
976{
977 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
978 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
979 struct ublk_io *io = &ubq->ios[tag];
980 struct request *req;
981
982 /* now this cmd slot is owned by nbd driver */
983 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
984 io->res = ub_cmd->result;
985
986 /* find the io request and complete */
987 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
988
989 if (req && likely(!blk_should_fake_timeout(req->q)))
990 ublk_complete_rq(req);
991}
992
993/*
994 * When ->ubq_daemon is exiting, either new request is ended immediately,
995 * or any queued io command is drained, so it is safe to abort queue
996 * lockless
997 */
998static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
999{
1000 int i;
1001
1002 if (!ublk_get_device(ub))
1003 return;
1004
1005 for (i = 0; i < ubq->q_depth; i++) {
1006 struct ublk_io *io = &ubq->ios[i];
1007
1008 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1009 struct request *rq;
1010
1011 /*
1012 * Either we fail the request or ublk_rq_task_work_fn
1013 * will do it
1014 */
1015 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1016 if (rq)
a0d41dc1 1017 __ublk_fail_req(ubq, io, rq);
71f28f31
ML		 1018 }
1019 }
1020 ublk_put_device(ub);
1021}
1022
1023static void ublk_daemon_monitor_work(struct work_struct *work)
1024{
1025 struct ublk_device *ub =
1026 container_of(work, struct ublk_device, monitor_work.work);
1027 int i;
1028
1029 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1030 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1031
1032 if (ubq_daemon_is_dying(ubq)) {
bbae8d1f
Z		 1033 if (ublk_queue_can_use_recovery(ubq))
1034 schedule_work(&ub->quiesce_work);
1035 else
1036 schedule_work(&ub->stop_work);
71f28f31
ML		 1037
1038 /* abort queue is for making forward progress */
1039 ublk_abort_queue(ub, ubq);
1040 }
1041 }
1042
1043 /*
bbae8d1f
Z		 1044 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1045 * after ublk_remove() or __ublk_quiesce_dev() is started.
71f28f31
ML		 1046 *
1047 * No need ub->mutex, monitor work are canceled after state is marked
bbae8d1f 1048 * as not LIVE, so new state is observed reliably.
71f28f31 1049 */
bbae8d1f 1050 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
71f28f31
ML		 1051 schedule_delayed_work(&ub->monitor_work,
1052 UBLK_DAEMON_MONITOR_PERIOD);
1053}
1054
a8ce5f52
ML		 1055static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1056{
1057 return ubq->nr_io_ready == ubq->q_depth;
1058}
1059
71f28f31
ML		 1060static void ublk_cancel_queue(struct ublk_queue *ubq)
1061{
1062 int i;
1063
a8ce5f52
ML		 1064 if (!ublk_queue_ready(ubq))
1065 return;
1066
71f28f31
ML		 1067 for (i = 0; i < ubq->q_depth; i++) {
1068 struct ublk_io *io = &ubq->ios[i];
1069
1070 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1071 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1072 }
a8ce5f52
ML		 1073
1074 /* all io commands are canceled */
1075 ubq->nr_io_ready = 0;
71f28f31
ML		 1076}
1077
1078/* Cancel all pending commands, must be called after del_gendisk() returns */
1079static void ublk_cancel_dev(struct ublk_device *ub)
1080{
1081 int i;
1082
1083 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1084 ublk_cancel_queue(ublk_get_queue(ub, i));
1085}
1086
bbae8d1f
Z		 1087static bool ublk_check_inflight_rq(struct request *rq, void *data)
1088{
1089 bool *idle = data;
1090
1091 if (blk_mq_request_started(rq)) {
1092 *idle = false;
1093 return false;
1094 }
1095 return true;
1096}
1097
1098static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1099{
1100 bool idle;
1101
1102 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1103 while (true) {
1104 idle = true;
1105 blk_mq_tagset_busy_iter(&ub->tag_set,
1106 ublk_check_inflight_rq, &idle);
1107 if (idle)
1108 break;
1109 msleep(UBLK_REQUEUE_DELAY_MS);
1110 }
1111}
1112
1113static void __ublk_quiesce_dev(struct ublk_device *ub)
71f28f31 1114{
bbae8d1f
Z		 1115 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1116 __func__, ub->dev_info.dev_id,
1117 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1118 "LIVE" : "QUIESCED");
1119 blk_mq_quiesce_queue(ub->ub_disk->queue);
1120 ublk_wait_tagset_rqs_idle(ub);
1121 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1122 ublk_cancel_dev(ub);
1123 /* we are going to release task_struct of ubq_daemon and resets
1124 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1125 * Besides, monitor_work is not necessary in QUIESCED state since we have
1126 * already scheduled quiesce_work and quiesced all ubqs.
1127 *
1128 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1129 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1130 */
1131 cancel_delayed_work_sync(&ub->monitor_work);
1132}
1133
1134static void ublk_quiesce_work_fn(struct work_struct *work)
1135{
1136 struct ublk_device *ub =
1137 container_of(work, struct ublk_device, quiesce_work);
1138
71f28f31 1139 mutex_lock(&ub->mutex);
6d9e6dfd 1140 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
71f28f31 1141 goto unlock;
bbae8d1f
Z		 1142 __ublk_quiesce_dev(ub);
1143 unlock:
1144 mutex_unlock(&ub->mutex);
1145}
71f28f31 1146
bbae8d1f
Z		 1147static void ublk_unquiesce_dev(struct ublk_device *ub)
1148{
1149 int i;
1150
1151 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1152 __func__, ub->dev_info.dev_id,
1153 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1154 "LIVE" : "QUIESCED");
1155 /* quiesce_work has run. We let requeued rqs be aborted
1156 * before running fallback_wq. "force_abort" must be seen
1157 * after request queue is unqiuesced. Then del_gendisk()
1158 * can move on.
1159 */
1160 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1161 ublk_get_queue(ub, i)->force_abort = true;
1162
1163 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1164 /* We may have requeued some rqs in ublk_quiesce_queue() */
1165 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1166}
1167
1168static void ublk_stop_dev(struct ublk_device *ub)
1169{
1170 mutex_lock(&ub->mutex);
1171 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1172 goto unlock;
1173 if (ublk_can_use_recovery(ub)) {
1174 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1175 __ublk_quiesce_dev(ub);
1176 ublk_unquiesce_dev(ub);
1177 }
71f28f31
ML		 1178 del_gendisk(ub->ub_disk);
1179 ub->dev_info.state = UBLK_S_DEV_DEAD;
1180 ub->dev_info.ublksrv_pid = -1;
6d9e6dfd
CH		 1181 put_disk(ub->ub_disk);
1182 ub->ub_disk = NULL;
71f28f31 1183 unlock:
a8ce5f52 1184 ublk_cancel_dev(ub);
71f28f31
ML		 1185 mutex_unlock(&ub->mutex);
1186 cancel_delayed_work_sync(&ub->monitor_work);
1187}
1188
71f28f31
ML		 1189/* device can only be started after all IOs are ready */
1190static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1191{
1192 mutex_lock(&ub->mutex);
1193 ubq->nr_io_ready++;
1194 if (ublk_queue_ready(ubq)) {
1195 ubq->ubq_daemon = current;
1196 get_task_struct(ubq->ubq_daemon);
1197 ub->nr_queues_ready++;
73a166d9
ML		 1198
1199 if (capable(CAP_SYS_ADMIN))
1200 ub->nr_privileged_daemon++;
71f28f31
ML		 1201 }
1202 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1203 complete_all(&ub->completion);
1204 mutex_unlock(&ub->mutex);
1205}
1206
c86019ff 1207static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
fee32f31 1208 int tag)
c86019ff
Z		 1209{
1210 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1211 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1212
7d4a9317 1213 ublk_queue_cmd(ubq, req);
c86019ff
Z		 1214}
1215
71f28f31
ML		 1216static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1217{
1218 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1219 struct ublk_device *ub = cmd->file->private_data;
1220 struct ublk_queue *ubq;
1221 struct ublk_io *io;
1222 u32 cmd_op = cmd->cmd_op;
1223 unsigned tag = ub_cmd->tag;
1224 int ret = -EINVAL;
1225
1226 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1227 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1228 ub_cmd->result);
1229
1230 if (!(issue_flags & IO_URING_F_SQE128))
1231 goto out;
1232
1233 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1234 goto out;
1235
1236 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1237 if (!ubq || ub_cmd->q_id != ubq->q_id)
1238 goto out;
1239
1240 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1241 goto out;
1242
1243 if (tag >= ubq->q_depth)
1244 goto out;
1245
1246 io = &ubq->ios[tag];
1247
1248 /* there is pending io cmd, something must be wrong */
1249 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1250 ret = -EBUSY;
1251 goto out;
1252 }
1253
c86019ff
Z		 1254 /*
1255 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1256 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1257 */
1258 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1259 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1260 goto out;
1261
71f28f31
ML		 1262 switch (cmd_op) {
1263 case UBLK_IO_FETCH_REQ:
1264 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1265 if (ublk_queue_ready(ubq)) {
1266 ret = -EBUSY;
1267 goto out;
1268 }
1269 /*
1270 * The io is being handled by server, so COMMIT_RQ is expected
1271 * instead of FETCH_REQ
1272 */
1273 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1274 goto out;
1275 /* FETCH_RQ has to provide IO buffer */
1276 if (!ub_cmd->addr)
1277 goto out;
1278 io->cmd = cmd;
1279 io->flags |= UBLK_IO_FLAG_ACTIVE;
1280 io->addr = ub_cmd->addr;
1281
1282 ublk_mark_io_ready(ub, ubq);
1283 break;
1284 case UBLK_IO_COMMIT_AND_FETCH_REQ:
1285 /* FETCH_RQ has to provide IO buffer */
1286 if (!ub_cmd->addr)
1287 goto out;
1288 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1289 goto out;
1290 io->addr = ub_cmd->addr;
1291 io->flags |= UBLK_IO_FLAG_ACTIVE;
1292 io->cmd = cmd;
1293 ublk_commit_completion(ub, ub_cmd);
1294 break;
c86019ff
Z		 1295 case UBLK_IO_NEED_GET_DATA:
1296 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1297 goto out;
1298 io->addr = ub_cmd->addr;
1299 io->cmd = cmd;
1300 io->flags |= UBLK_IO_FLAG_ACTIVE;
fee32f31 1301 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
c86019ff 1302 break;
71f28f31
ML		 1303 default:
1304 goto out;
1305 }
1306 return -EIOCBQUEUED;
1307
1308 out:
71f28f31
ML		 1309 io_uring_cmd_done(cmd, ret, 0);
1310 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1311 __func__, cmd_op, tag, ret, io->flags);
1312 return -EIOCBQUEUED;
1313}
1314
1315static const struct file_operations ublk_ch_fops = {
1316 .owner = THIS_MODULE,
1317 .open = ublk_ch_open,
1318 .release = ublk_ch_release,
1319 .llseek = no_llseek,
1320 .uring_cmd = ublk_ch_uring_cmd,
1321 .mmap = ublk_ch_mmap,
1322};
1323
1324static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1325{
1326 int size = ublk_queue_cmd_buf_size(ub, q_id);
1327 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1328
1329 if (ubq->ubq_daemon)
1330 put_task_struct(ubq->ubq_daemon);
1331 if (ubq->io_cmd_buf)
1332 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1333}
1334
1335static int ublk_init_queue(struct ublk_device *ub, int q_id)
1336{
1337 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1338 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1339 void *ptr;
1340 int size;
1341
6d8c5afc 1342 ubq->flags = ub->dev_info.flags;
71f28f31
ML		 1343 ubq->q_id = q_id;
1344 ubq->q_depth = ub->dev_info.queue_depth;
1345 size = ublk_queue_cmd_buf_size(ub, q_id);
1346
1347 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1348 if (!ptr)
1349 return -ENOMEM;
1350
1351 ubq->io_cmd_buf = ptr;
1352 ubq->dev = ub;
1353 return 0;
1354}
1355
1356static void ublk_deinit_queues(struct ublk_device *ub)
1357{
1358 int nr_queues = ub->dev_info.nr_hw_queues;
1359 int i;
1360
1361 if (!ub->__queues)
1362 return;
1363
1364 for (i = 0; i < nr_queues; i++)
1365 ublk_deinit_queue(ub, i);
1366 kfree(ub->__queues);
1367}
1368
1369static int ublk_init_queues(struct ublk_device *ub)
1370{
1371 int nr_queues = ub->dev_info.nr_hw_queues;
1372 int depth = ub->dev_info.queue_depth;
1373 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1374 int i, ret = -ENOMEM;
1375
1376 ub->queue_size = ubq_size;
1377 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1378 if (!ub->__queues)
1379 return ret;
1380
1381 for (i = 0; i < nr_queues; i++) {
1382 if (ublk_init_queue(ub, i))
1383 goto fail;
1384 }
1385
1386 init_completion(&ub->completion);
1387 return 0;
1388
1389 fail:
1390 ublk_deinit_queues(ub);
1391 return ret;
1392}
1393
fa9482e0 1394static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
71f28f31
ML		 1395{
1396 int i = idx;
1397 int err;
1398
1399 spin_lock(&ublk_idr_lock);
1400 /* allocate id, if @id >= 0, we're requesting that specific id */
1401 if (i >= 0) {
1402 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1403 if (err == -ENOSPC)
1404 err = -EEXIST;
1405 } else {
1406 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1407 }
1408 spin_unlock(&ublk_idr_lock);
1409
1410 if (err >= 0)
1411 ub->ub_number = err;
1412
1413 return err;
1414}
1415
fa9482e0 1416static void ublk_free_dev_number(struct ublk_device *ub)
71f28f31
ML		 1417{
1418 spin_lock(&ublk_idr_lock);
1419 idr_remove(&ublk_index_idr, ub->ub_number);
1420 wake_up_all(&ublk_idr_wq);
1421 spin_unlock(&ublk_idr_lock);
71f28f31
ML		 1422}
1423
1424static void ublk_cdev_rel(struct device *dev)
1425{
1426 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1427
71f28f31 1428 blk_mq_free_tag_set(&ub->tag_set);
71f28f31 1429 ublk_deinit_queues(ub);
fa9482e0
CH		 1430 ublk_free_dev_number(ub);
1431 mutex_destroy(&ub->mutex);
1432 kfree(ub);
71f28f31
ML		 1433}
1434
1435static int ublk_add_chdev(struct ublk_device *ub)
1436{
1437 struct device *dev = &ub->cdev_dev;
1438 int minor = ub->ub_number;
1439 int ret;
1440
1441 dev->parent = ublk_misc.this_device;
1442 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1443 dev->class = ublk_chr_class;
1444 dev->release = ublk_cdev_rel;
1445 device_initialize(dev);
1446
1447 ret = dev_set_name(dev, "ublkc%d", minor);
1448 if (ret)
1449 goto fail;
1450
1451 cdev_init(&ub->cdev, &ublk_ch_fops);
1452 ret = cdev_device_add(&ub->cdev, dev);
1453 if (ret)
1454 goto fail;
403ebc87
ML		 1455
1456 ublks_added++;
71f28f31
ML		 1457 return 0;
1458 fail:
1459 put_device(dev);
1460 return ret;
1461}
1462
1463static void ublk_stop_work_fn(struct work_struct *work)
1464{
1465 struct ublk_device *ub =
1466 container_of(work, struct ublk_device, stop_work);
1467
1468 ublk_stop_dev(ub);
1469}
1470
4bf9cbf3 1471/* align max io buffer size with PAGE_SIZE */
6d9e6dfd 1472static void ublk_align_max_io_size(struct ublk_device *ub)
71f28f31 1473{
4bf9cbf3 1474 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
71f28f31 1475
4bf9cbf3
ML		 1476 ub->dev_info.max_io_buf_bytes =
1477 round_down(max_io_bytes, PAGE_SIZE);
71f28f31
ML		 1478}
1479
fa9482e0 1480static int ublk_add_tag_set(struct ublk_device *ub)
71f28f31 1481{
71f28f31
ML		 1482 ub->tag_set.ops = &ublk_mq_ops;
1483 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1484 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1485 ub->tag_set.numa_node = NUMA_NO_NODE;
0edb3696 1486 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
71f28f31
ML		 1487 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1488 ub->tag_set.driver_data = ub;
fa9482e0 1489 return blk_mq_alloc_tag_set(&ub->tag_set);
71f28f31
ML		 1490}
1491
1492static void ublk_remove(struct ublk_device *ub)
1493{
34d8f2be
CH		 1494 ublk_stop_dev(ub);
1495 cancel_work_sync(&ub->stop_work);
bbae8d1f 1496 cancel_work_sync(&ub->quiesce_work);
71f28f31
ML		 1497 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1498 put_device(&ub->cdev_dev);
403ebc87 1499 ublks_added--;
71f28f31
ML		 1500}
1501
1502static struct ublk_device *ublk_get_device_from_id(int idx)
1503{
1504 struct ublk_device *ub = NULL;
1505
1506 if (idx < 0)
1507 return NULL;
1508
1509 spin_lock(&ublk_idr_lock);
1510 ub = idr_find(&ublk_index_idr, idx);
1511 if (ub)
1512 ub = ublk_get_device(ub);
1513 spin_unlock(&ublk_idr_lock);
1514
1515 return ub;
1516}
1517
bfbcef03 1518static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
71f28f31
ML		 1519{
1520 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
71f28f31 1521 int ublksrv_pid = (int)header->data[0];
6d9e6dfd 1522 struct gendisk *disk;
34d8f2be 1523 int ret = -EINVAL;
71f28f31
ML		 1524
1525 if (ublksrv_pid <= 0)
34d8f2be
CH		 1526 return -EINVAL;
1527
71f28f31
ML		 1528 wait_for_completion_interruptible(&ub->completion);
1529
1530 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1531
1532 mutex_lock(&ub->mutex);
6d9e6dfd
CH		 1533 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1534 test_bit(UB_STATE_USED, &ub->state)) {
71f28f31 1535 ret = -EEXIST;
34d8f2be 1536 goto out_unlock;
71f28f31 1537 }
71f28f31 1538
6d9e6dfd
CH		 1539 disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1540 if (IS_ERR(disk)) {
1541 ret = PTR_ERR(disk);
1542 goto out_unlock;
34d8f2be 1543 }
6d9e6dfd
CH		 1544 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1545 disk->fops = &ub_fops;
1546 disk->private_data = ub;
1547
34d8f2be 1548 ub->dev_info.ublksrv_pid = ublksrv_pid;
6d9e6dfd 1549 ub->ub_disk = disk;
0aa73170
ML		 1550
1551 ret = ublk_apply_params(ub);
1552 if (ret)
1553 goto out_put_disk;
1554
73a166d9
ML		 1555 /* don't probe partitions if any one ubq daemon is un-trusted */
1556 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1557 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1558
6d9e6dfd
CH		 1559 get_device(&ub->cdev_dev);
1560 ret = add_disk(disk);
1561 if (ret) {
93d71ec8
ML		 1562 /*
1563 * Has to drop the reference since ->free_disk won't be
1564 * called in case of add_disk failure.
1565 */
1566 ublk_put_device(ub);
0aa73170 1567 goto out_put_disk;
6d9e6dfd
CH		 1568 }
1569 set_bit(UB_STATE_USED, &ub->state);
34d8f2be 1570 ub->dev_info.state = UBLK_S_DEV_LIVE;
0aa73170
ML		 1571out_put_disk:
1572 if (ret)
1573 put_disk(disk);
34d8f2be
CH		 1574out_unlock:
1575 mutex_unlock(&ub->mutex);
71f28f31
ML		 1576 return ret;
1577}
1578
bfbcef03
ML		 1579static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1580 struct io_uring_cmd *cmd)
71f28f31
ML		 1581{
1582 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1583 void __user *argp = (void __user *)(unsigned long)header->addr;
c50061f0 1584 cpumask_var_t cpumask;
71f28f31
ML		 1585 unsigned long queue;
1586 unsigned int retlen;
c50061f0 1587 unsigned int i;
bfbcef03
ML		 1588 int ret;
1589
34d8f2be
CH		 1590 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1591 return -EINVAL;
1592 if (header->len & (sizeof(unsigned long)-1))
1593 return -EINVAL;
1594 if (!header->addr)
1595 return -EINVAL;
71f28f31 1596
71f28f31
ML		 1597 queue = header->data[0];
1598 if (queue >= ub->dev_info.nr_hw_queues)
bfbcef03 1599 return -EINVAL;
71f28f31 1600
c50061f0 1601 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
bfbcef03 1602 return -ENOMEM;
c50061f0
CH		 1603
1604 for_each_possible_cpu(i) {
1605 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1606 cpumask_set_cpu(i, cpumask);
71f28f31 1607 }
c50061f0
CH		 1608
1609 ret = -EFAULT;
1610 retlen = min_t(unsigned short, header->len, cpumask_size());
1611 if (copy_to_user(argp, cpumask, retlen))
1612 goto out_free_cpumask;
1613 if (retlen != header->len &&
1614 clear_user(argp + retlen, header->len - retlen))
1615 goto out_free_cpumask;
1616
71f28f31 1617 ret = 0;
c50061f0
CH		 1618out_free_cpumask:
1619 free_cpumask_var(cpumask);
71f28f31
ML		 1620 return ret;
1621}
1622
4093cb5a
ML		 1623static void ublk_store_owner_uid_gid(struct ublksrv_ctrl_dev_info *info)
1624{
1625 kuid_t uid;
1626 kgid_t gid;
1627
1628 current_uid_gid(&uid, &gid);
1629
1630 info->owner_uid = from_kuid(&init_user_ns, uid);
1631 info->owner_gid = from_kgid(&init_user_ns, gid);
1632}
1633
34d8f2be 1634static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
71f28f31 1635{
34d8f2be 1636 pr_devel("%s: dev id %d flags %llx\n", __func__,
6d8c5afc 1637 info->dev_id, info->flags);
4bf9cbf3
ML		 1638 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1639 info->nr_hw_queues, info->queue_depth);
34d8f2be
CH		 1640}
1641
1642static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1643{
1644 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1645 void __user *argp = (void __user *)(unsigned long)header->addr;
1646 struct ublksrv_ctrl_dev_info info;
71f28f31 1647 struct ublk_device *ub;
34d8f2be
CH		 1648 int ret = -EINVAL;
1649
1650 if (header->len < sizeof(info) || !header->addr)
1651 return -EINVAL;
1652 if (header->queue_id != (u16)-1) {
1653 pr_warn("%s: queue_id is wrong %x\n",
1654 __func__, header->queue_id);
1655 return -EINVAL;
1656 }
4093cb5a 1657
34d8f2be
CH		 1658 if (copy_from_user(&info, argp, sizeof(info)))
1659 return -EFAULT;
4093cb5a
ML		 1660
1661 if (capable(CAP_SYS_ADMIN))
1662 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1663 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1664 return -EPERM;
1665
1666 /* the created device is always owned by current user */
1667 ublk_store_owner_uid_gid(&info);
1668
34d8f2be
CH		 1669 if (header->dev_id != info.dev_id) {
1670 pr_warn("%s: dev id not match %u %u\n",
1671 __func__, header->dev_id, info.dev_id);
1672 return -EINVAL;
1673 }
71f28f31 1674
4093cb5a
ML		 1675 ublk_dump_dev_info(&info);
1676
71f28f31
ML		 1677 ret = mutex_lock_killable(&ublk_ctl_mutex);
1678 if (ret)
1679 return ret;
1680
403ebc87
ML		 1681 ret = -EACCES;
1682 if (ublks_added >= ublks_max)
1683 goto out_unlock;
1684
cfee7e4d
CH		 1685 ret = -ENOMEM;
1686 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1687 if (!ub)
1688 goto out_unlock;
fa9482e0
CH		 1689 mutex_init(&ub->mutex);
1690 spin_lock_init(&ub->mm_lock);
bbae8d1f 1691 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
fa9482e0
CH		 1692 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1693 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
cfee7e4d 1694
fa9482e0
CH		 1695 ret = ublk_alloc_dev_number(ub, header->dev_id);
1696 if (ret < 0)
1697 goto out_free_ub;
71f28f31 1698
34d8f2be 1699 memcpy(&ub->dev_info, &info, sizeof(info));
71f28f31 1700
34d8f2be
CH		 1701 /* update device id */
1702 ub->dev_info.dev_id = ub->ub_number;
1703
6d8c5afc
ML		 1704 /*
1705 * 64bit flags will be copied back to userspace as feature
1706 * negotiation result, so have to clear flags which driver
1707 * doesn't support yet, then userspace can get correct flags
1708 * (features) to handle.
1709 */
1710 ub->dev_info.flags &= UBLK_F_ALL;
1711
224e858f
ML		 1712 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1713 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1714
fa9482e0 1715 /* We are not ready to support zero copy */
6d8c5afc 1716 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
fa9482e0 1717
fa9482e0
CH		 1718 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1719 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1720 ublk_align_max_io_size(ub);
1721
1722 ret = ublk_init_queues(ub);
34d8f2be 1723 if (ret)
fa9482e0 1724 goto out_free_dev_number;
34d8f2be 1725
fa9482e0
CH		 1726 ret = ublk_add_tag_set(ub);
1727 if (ret)
1728 goto out_deinit_queues;
1729
1730 ret = -EFAULT;
1731 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1732 goto out_free_tag_set;
1733
1734 /*
1735 * Add the char dev so that ublksrv daemon can be setup.
1736 * ublk_add_chdev() will cleanup everything if it fails.
1737 */
1738 ret = ublk_add_chdev(ub);
1739 goto out_unlock;
1740
1741out_free_tag_set:
1742 blk_mq_free_tag_set(&ub->tag_set);
1743out_deinit_queues:
1744 ublk_deinit_queues(ub);
1745out_free_dev_number:
1746 ublk_free_dev_number(ub);
1747out_free_ub:
1748 mutex_destroy(&ub->mutex);
1749 kfree(ub);
34d8f2be 1750out_unlock:
71f28f31 1751 mutex_unlock(&ublk_ctl_mutex);
71f28f31
ML		 1752 return ret;
1753}
1754
1755static inline bool ublk_idr_freed(int id)
1756{
1757 void *ptr;
1758
1759 spin_lock(&ublk_idr_lock);
1760 ptr = idr_find(&ublk_index_idr, id);
1761 spin_unlock(&ublk_idr_lock);
1762
1763 return ptr == NULL;
1764}
1765
bfbcef03 1766static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
71f28f31 1767{
bfbcef03
ML		 1768 struct ublk_device *ub = *p_ub;
1769 int idx = ub->ub_number;
71f28f31
ML		 1770 int ret;
1771
1772 ret = mutex_lock_killable(&ublk_ctl_mutex);
1773 if (ret)
1774 return ret;
1775
bfbcef03
ML		 1776 ublk_remove(ub);
1777
1778 /* Mark the reference as consumed */
1779 *p_ub = NULL;
1780 ublk_put_device(ub);
71f28f31
ML		 1781
1782 /*
1783 * Wait until the idr is removed, then it can be reused after
1784 * DEL_DEV command is returned.
1785 */
bfbcef03 1786 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
71f28f31
ML		 1787 mutex_unlock(&ublk_ctl_mutex);
1788
1789 return ret;
1790}
1791
71f28f31
ML		 1792static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1793{
1794 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1795
1796 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1797 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1798 header->data[0], header->addr, header->len);
1799}
1800
bfbcef03 1801static int ublk_ctrl_stop_dev(struct ublk_device *ub)
71f28f31 1802{
34d8f2be
CH		 1803 ublk_stop_dev(ub);
1804 cancel_work_sync(&ub->stop_work);
bbae8d1f 1805 cancel_work_sync(&ub->quiesce_work);
71f28f31
ML		 1806
1807 return 0;
1808}
1809
bfbcef03
ML		 1810static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1811 struct io_uring_cmd *cmd)
71f28f31
ML		 1812{
1813 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1814 void __user *argp = (void __user *)(unsigned long)header->addr;
34d8f2be
CH		 1815
1816 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1817 return -EINVAL;
1818
34d8f2be 1819 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
bfbcef03 1820 return -EFAULT;
34d8f2be 1821
bfbcef03 1822 return 0;
34d8f2be
CH		 1823}
1824
abb864d3
ML		 1825/* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1826static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1827{
1828 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1829 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1830
1831 if (ub->ub_disk) {
1832 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1833 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1834 } else {
1835 ub->params.devt.disk_major = 0;
1836 ub->params.devt.disk_minor = 0;
1837 }
1838 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1839}
1840
bfbcef03
ML		 1841static int ublk_ctrl_get_params(struct ublk_device *ub,
1842 struct io_uring_cmd *cmd)
0aa73170
ML		 1843{
1844 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1845 void __user *argp = (void __user *)(unsigned long)header->addr;
1846 struct ublk_params_header ph;
0aa73170
ML		 1847 int ret;
1848
1849 if (header->len <= sizeof(ph) || !header->addr)
1850 return -EINVAL;
1851
1852 if (copy_from_user(&ph, argp, sizeof(ph)))
1853 return -EFAULT;
1854
1855 if (ph.len > header->len || !ph.len)
1856 return -EINVAL;
1857
1858 if (ph.len > sizeof(struct ublk_params))
1859 ph.len = sizeof(struct ublk_params);
1860
0aa73170 1861 mutex_lock(&ub->mutex);
abb864d3 1862 ublk_ctrl_fill_params_devt(ub);
0aa73170
ML		 1863 if (copy_to_user(argp, &ub->params, ph.len))
1864 ret = -EFAULT;
1865 else
1866 ret = 0;
1867 mutex_unlock(&ub->mutex);
1868
0aa73170
ML		 1869 return ret;
1870}
1871
bfbcef03
ML		 1872static int ublk_ctrl_set_params(struct ublk_device *ub,
1873 struct io_uring_cmd *cmd)
0aa73170
ML		 1874{
1875 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1876 void __user *argp = (void __user *)(unsigned long)header->addr;
1877 struct ublk_params_header ph;
0aa73170
ML		 1878 int ret = -EFAULT;
1879
1880 if (header->len <= sizeof(ph) || !header->addr)
1881 return -EINVAL;
1882
1883 if (copy_from_user(&ph, argp, sizeof(ph)))
1884 return -EFAULT;
1885
1886 if (ph.len > header->len || !ph.len || !ph.types)
1887 return -EINVAL;
1888
1889 if (ph.len > sizeof(struct ublk_params))
1890 ph.len = sizeof(struct ublk_params);
1891
0aa73170
ML		 1892 /* parameters can only be changed when device isn't live */
1893 mutex_lock(&ub->mutex);
1894 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1895 ret = -EACCES;
1896 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1897 ret = -EFAULT;
1898 } else {
1899 /* clear all we don't support yet */
1900 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1901 ret = ublk_validate_params(ub);
1902 }
1903 mutex_unlock(&ub->mutex);
0aa73170
ML		 1904
1905 return ret;
1906}
1907
c732a852
Z		 1908static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1909{
1910 int i;
1911
1912 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1913 /* All old ioucmds have to be completed */
1914 WARN_ON_ONCE(ubq->nr_io_ready);
1915 /* old daemon is PF_EXITING, put it now */
1916 put_task_struct(ubq->ubq_daemon);
1917 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1918 ubq->ubq_daemon = NULL;
1919
1920 for (i = 0; i < ubq->q_depth; i++) {
1921 struct ublk_io *io = &ubq->ios[i];
1922
1923 /* forget everything now and be ready for new FETCH_REQ */
1924 io->flags = 0;
1925 io->cmd = NULL;
1926 io->addr = 0;
1927 }
1928}
1929
bfbcef03
ML		 1930static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1931 struct io_uring_cmd *cmd)
c732a852
Z		 1932{
1933 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
c732a852
Z		 1934 int ret = -EINVAL;
1935 int i;
1936
c732a852
Z		 1937 mutex_lock(&ub->mutex);
1938 if (!ublk_can_use_recovery(ub))
1939 goto out_unlock;
1940 /*
1941 * START_RECOVERY is only allowd after:
1942 *
1943 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1944 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1945 * released.
1946 *
1947 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1948 * (a)has quiesced request queue
1949 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1950 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1951 * (d)has completed/camceled all ioucmds owned by ther dying process
1952 */
1953 if (test_bit(UB_STATE_OPEN, &ub->state) ||
1954 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1955 ret = -EBUSY;
1956 goto out_unlock;
1957 }
1958 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1959 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1960 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1961 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1962 ub->mm = NULL;
1963 ub->nr_queues_ready = 0;
73a166d9 1964 ub->nr_privileged_daemon = 0;
c732a852
Z		 1965 init_completion(&ub->completion);
1966 ret = 0;
1967 out_unlock:
1968 mutex_unlock(&ub->mutex);
c732a852
Z		 1969 return ret;
1970}
1971
bfbcef03
ML		 1972static int ublk_ctrl_end_recovery(struct ublk_device *ub,
1973 struct io_uring_cmd *cmd)
c732a852
Z		 1974{
1975 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1976 int ublksrv_pid = (int)header->data[0];
c732a852
Z		 1977 int ret = -EINVAL;
1978
c732a852
Z		 1979 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1980 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1981 /* wait until new ubq_daemon sending all FETCH_REQ */
1982 wait_for_completion_interruptible(&ub->completion);
1983 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
1984 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1985
1986 mutex_lock(&ub->mutex);
1987 if (!ublk_can_use_recovery(ub))
1988 goto out_unlock;
1989
1990 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1991 ret = -EBUSY;
1992 goto out_unlock;
1993 }
1994 ub->dev_info.ublksrv_pid = ublksrv_pid;
1995 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
1996 __func__, ublksrv_pid, header->dev_id);
1997 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1998 pr_devel("%s: queue unquiesced, dev id %d.\n",
1999 __func__, header->dev_id);
2000 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2001 ub->dev_info.state = UBLK_S_DEV_LIVE;
2002 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2003 ret = 0;
2004 out_unlock:
2005 mutex_unlock(&ub->mutex);
c732a852
Z		 2006 return ret;
2007}
2008
4093cb5a
ML		 2009/*
2010 * All control commands are sent via /dev/ublk-control, so we have to check
2011 * the destination device's permission
2012 */
2013static int ublk_char_dev_permission(struct ublk_device *ub,
2014 const char *dev_path, int mask)
2015{
2016 int err;
2017 struct path path;
2018 struct kstat stat;
2019
2020 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2021 if (err)
2022 return err;
2023
2024 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2025 if (err)
2026 goto exit;
2027
2028 err = -EPERM;
2029 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2030 goto exit;
2031
2032 err = inode_permission(&init_user_ns,
2033 d_backing_inode(path.dentry), mask);
2034exit:
2035 path_put(&path);
2036 return err;
2037}
2038
2039static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2040 struct io_uring_cmd *cmd)
2041{
2042 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2043 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2044 void __user *argp = (void __user *)(unsigned long)header->addr;
2045 char *dev_path = NULL;
2046 int ret = 0;
2047 int mask;
2048
2049 if (!unprivileged) {
2050 if (!capable(CAP_SYS_ADMIN))
2051 return -EPERM;
2052 /*
2053 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2054 * char_dev_path in payload too, since userspace may not
2055 * know if the specified device is created as unprivileged
2056 * mode.
2057 */
2058 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2059 return 0;
2060 }
2061
2062 /*
2063 * User has to provide the char device path for unprivileged ublk
2064 *
2065 * header->addr always points to the dev path buffer, and
2066 * header->dev_path_len records length of dev path buffer.
2067 */
2068 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2069 return -EINVAL;
2070
2071 if (header->len < header->dev_path_len)
2072 return -EINVAL;
2073
2074 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2075 if (!dev_path)
2076 return -ENOMEM;
2077
2078 ret = -EFAULT;
2079 if (copy_from_user(dev_path, argp, header->dev_path_len))
2080 goto exit;
2081 dev_path[header->dev_path_len] = 0;
2082
2083 ret = -EINVAL;
2084 switch (cmd->cmd_op) {
2085 case UBLK_CMD_GET_DEV_INFO:
2086 case UBLK_CMD_GET_DEV_INFO2:
2087 case UBLK_CMD_GET_QUEUE_AFFINITY:
2088 case UBLK_CMD_GET_PARAMS:
2089 mask = MAY_READ;
2090 break;
2091 case UBLK_CMD_START_DEV:
2092 case UBLK_CMD_STOP_DEV:
2093 case UBLK_CMD_ADD_DEV:
2094 case UBLK_CMD_DEL_DEV:
2095 case UBLK_CMD_SET_PARAMS:
2096 case UBLK_CMD_START_USER_RECOVERY:
2097 case UBLK_CMD_END_USER_RECOVERY:
2098 mask = MAY_READ | MAY_WRITE;
2099 break;
2100 default:
2101 goto exit;
2102 }
2103
2104 ret = ublk_char_dev_permission(ub, dev_path, mask);
2105 if (!ret) {
2106 header->len -= header->dev_path_len;
2107 header->addr += header->dev_path_len;
2108 }
2109 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2110 __func__, ub->ub_number, cmd->cmd_op,
2111 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2112 dev_path, ret);
2113exit:
2114 kfree(dev_path);
2115 return ret;
2116}
2117
34d8f2be
CH		 2118static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2119 unsigned int issue_flags)
2120{
2121 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
bfbcef03 2122 struct ublk_device *ub = NULL;
71f28f31
ML		 2123 int ret = -EINVAL;
2124
fa8e442e
ML		 2125 if (issue_flags & IO_URING_F_NONBLOCK)
2126 return -EAGAIN;
2127
71f28f31
ML		 2128 ublk_ctrl_cmd_dump(cmd);
2129
2130 if (!(issue_flags & IO_URING_F_SQE128))
2131 goto out;
2132
bfbcef03
ML		 2133 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2134 ret = -ENODEV;
2135 ub = ublk_get_device_from_id(header->dev_id);
2136 if (!ub)
2137 goto out;
4093cb5a
ML		 2138
2139 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2140 } else {
2141 /* ADD_DEV permission check is done in command handler */
2142 ret = 0;
bfbcef03
ML		 2143 }
2144
4093cb5a
ML		 2145 if (ret)
2146 goto put_dev;
2147
34d8f2be 2148 switch (cmd->cmd_op) {
71f28f31 2149 case UBLK_CMD_START_DEV:
bfbcef03 2150 ret = ublk_ctrl_start_dev(ub, cmd);
71f28f31
ML		 2151 break;
2152 case UBLK_CMD_STOP_DEV:
bfbcef03 2153 ret = ublk_ctrl_stop_dev(ub);
71f28f31
ML		 2154 break;
2155 case UBLK_CMD_GET_DEV_INFO:
4093cb5a 2156 case UBLK_CMD_GET_DEV_INFO2:
bfbcef03 2157 ret = ublk_ctrl_get_dev_info(ub, cmd);
71f28f31
ML		 2158 break;
2159 case UBLK_CMD_ADD_DEV:
34d8f2be 2160 ret = ublk_ctrl_add_dev(cmd);
71f28f31
ML		 2161 break;
2162 case UBLK_CMD_DEL_DEV:
bfbcef03 2163 ret = ublk_ctrl_del_dev(&ub);
71f28f31
ML		 2164 break;
2165 case UBLK_CMD_GET_QUEUE_AFFINITY:
bfbcef03 2166 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
71f28f31 2167 break;
0aa73170 2168 case UBLK_CMD_GET_PARAMS:
bfbcef03 2169 ret = ublk_ctrl_get_params(ub, cmd);
0aa73170
ML		 2170 break;
2171 case UBLK_CMD_SET_PARAMS:
bfbcef03 2172 ret = ublk_ctrl_set_params(ub, cmd);
0aa73170 2173 break;
c732a852 2174 case UBLK_CMD_START_USER_RECOVERY:
bfbcef03 2175 ret = ublk_ctrl_start_recovery(ub, cmd);
c732a852
Z		 2176 break;
2177 case UBLK_CMD_END_USER_RECOVERY:
bfbcef03 2178 ret = ublk_ctrl_end_recovery(ub, cmd);
c732a852 2179 break;
71f28f31 2180 default:
bfbcef03 2181 ret = -ENOTSUPP;
71f28f31 2182 break;
6b1439d2 2183 }
4093cb5a
ML		 2184
2185 put_dev:
bfbcef03
ML		 2186 if (ub)
2187 ublk_put_device(ub);
71f28f31
ML		 2188 out:
2189 io_uring_cmd_done(cmd, ret, 0);
2190 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2191 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2192 return -EIOCBQUEUED;
2193}
2194
2195static const struct file_operations ublk_ctl_fops = {
2196 .open = nonseekable_open,
2197 .uring_cmd = ublk_ctrl_uring_cmd,
2198 .owner = THIS_MODULE,
2199 .llseek = noop_llseek,
2200};
2201
2202static struct miscdevice ublk_misc = {
2203 .minor = MISC_DYNAMIC_MINOR,
2204 .name = "ublk-control",
2205 .fops = &ublk_ctl_fops,
2206};
2207
2208static int __init ublk_init(void)
2209{
2210 int ret;
2211
2212 init_waitqueue_head(&ublk_idr_wq);
2213
2214 ret = misc_register(&ublk_misc);
2215 if (ret)
2216 return ret;
2217
2218 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2219 if (ret)
2220 goto unregister_mis;
2221
2222 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2223 if (IS_ERR(ublk_chr_class)) {
2224 ret = PTR_ERR(ublk_chr_class);
2225 goto free_chrdev_region;
2226 }
2227 return 0;
2228
2229free_chrdev_region:
2230 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2231unregister_mis:
2232 misc_deregister(&ublk_misc);
2233 return ret;
2234}
2235
2236static void __exit ublk_exit(void)
2237{
2238 struct ublk_device *ub;
2239 int id;
2240
71f28f31
ML		 2241 idr_for_each_entry(&ublk_index_idr, ub, id)
2242 ublk_remove(ub);
2243
8e4ff684
ML		 2244 class_destroy(ublk_chr_class);
2245 misc_deregister(&ublk_misc);
2246
71f28f31
ML		 2247 idr_destroy(&ublk_index_idr);
2248 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2249}
2250
2251module_init(ublk_init);
2252module_exit(ublk_exit);
2253
403ebc87
ML		 2254module_param(ublks_max, int, 0444);
2255MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2256
71f28f31
ML		 2257MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2258MODULE_LICENSE("GPL");
Linux 6.x block layer and io_uring tree(s)