projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
block: ublk_drv: cleanup 'struct ublk_map_data'
[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
29baef78 143 unsigned int 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
0abe39de 153#define UB_STATE_DELETED 2
fa362045 154 unsigned long state;
71f28f31
ML		 155 int ub_number;
156
157 struct mutex mutex;
158
e94eb459 159 spinlock_t mm_lock;
71f28f31
ML		 160 struct mm_struct *mm;
161
0aa73170
ML		 162 struct ublk_params params;
163
71f28f31
ML		 164 struct completion completion;
165 unsigned int nr_queues_ready;
73a166d9 166 unsigned int nr_privileged_daemon;
71f28f31
ML		 167
168 /*
169 * Our ubq->daemon may be killed without any notification, so
170 * monitor each queue's daemon periodically
171 */
172 struct delayed_work monitor_work;
bbae8d1f 173 struct work_struct quiesce_work;
71f28f31
ML		 174 struct work_struct stop_work;
175};
176
0aa73170
ML		 177/* header of ublk_params */
178struct ublk_params_header {
179 __u32 len;
180 __u32 types;
181};
182
71f28f31
ML		 183static dev_t ublk_chr_devt;
184static struct class *ublk_chr_class;
185
186static DEFINE_IDR(ublk_index_idr);
187static DEFINE_SPINLOCK(ublk_idr_lock);
188static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
189
190static DEFINE_MUTEX(ublk_ctl_mutex);
191
403ebc87
ML		 192/*
193 * Max ublk devices allowed to add
194 *
195 * It can be extended to one per-user limit in future or even controlled
196 * by cgroup.
197 */
198static unsigned int ublks_max = 64;
199static unsigned int ublks_added; /* protected by ublk_ctl_mutex */
200
71f28f31
ML		 201static struct miscdevice ublk_misc;
202
0aa73170
ML		 203static void ublk_dev_param_basic_apply(struct ublk_device *ub)
204{
205 struct request_queue *q = ub->ub_disk->queue;
206 const struct ublk_param_basic *p = &ub->params.basic;
207
208 blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
209 blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
210 blk_queue_io_min(q, 1 << p->io_min_shift);
211 blk_queue_io_opt(q, 1 << p->io_opt_shift);
212
213 blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
214 p->attrs & UBLK_ATTR_FUA);
215 if (p->attrs & UBLK_ATTR_ROTATIONAL)
216 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
217 else
218 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
219
220 blk_queue_max_hw_sectors(q, p->max_sectors);
221 blk_queue_chunk_sectors(q, p->chunk_sectors);
222 blk_queue_virt_boundary(q, p->virt_boundary_mask);
223
224 if (p->attrs & UBLK_ATTR_READ_ONLY)
225 set_disk_ro(ub->ub_disk, true);
226
227 set_capacity(ub->ub_disk, p->dev_sectors);
228}
229
230static void ublk_dev_param_discard_apply(struct ublk_device *ub)
231{
232 struct request_queue *q = ub->ub_disk->queue;
233 const struct ublk_param_discard *p = &ub->params.discard;
234
235 q->limits.discard_alignment = p->discard_alignment;
236 q->limits.discard_granularity = p->discard_granularity;
237 blk_queue_max_discard_sectors(q, p->max_discard_sectors);
238 blk_queue_max_write_zeroes_sectors(q,
239 p->max_write_zeroes_sectors);
240 blk_queue_max_discard_segments(q, p->max_discard_segments);
241}
242
243static int ublk_validate_params(const struct ublk_device *ub)
244{
245 /* basic param is the only one which must be set */
246 if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
247 const struct ublk_param_basic *p = &ub->params.basic;
248
249 if (p->logical_bs_shift > PAGE_SHIFT)
250 return -EINVAL;
251
252 if (p->logical_bs_shift > p->physical_bs_shift)
253 return -EINVAL;
254
4bf9cbf3 255 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
0aa73170
ML		 256 return -EINVAL;
257 } else
258 return -EINVAL;
259
260 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
261 const struct ublk_param_discard *p = &ub->params.discard;
262
263 /* So far, only support single segment discard */
264 if (p->max_discard_sectors && p->max_discard_segments != 1)
265 return -EINVAL;
266
267 if (!p->discard_granularity)
268 return -EINVAL;
269 }
270
abb864d3
ML		 271 /* dev_t is read-only */
272 if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
273 return -EINVAL;
274
0aa73170
ML		 275 return 0;
276}
277
278static int ublk_apply_params(struct ublk_device *ub)
279{
280 if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
281 return -EINVAL;
282
283 ublk_dev_param_basic_apply(ub);
284
285 if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
286 ublk_dev_param_discard_apply(ub);
287
288 return 0;
289}
290
0edb3696
ML		 291static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
292{
293 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
294 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
295 return true;
296 return false;
297}
298
c86019ff
Z		 299static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
300{
96cf2f54 301 return ubq->flags & UBLK_F_NEED_GET_DATA;
c86019ff
Z		 302}
303
71f28f31
ML		 304static struct ublk_device *ublk_get_device(struct ublk_device *ub)
305{
306 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
307 return ub;
308 return NULL;
309}
310
311static void ublk_put_device(struct ublk_device *ub)
312{
313 put_device(&ub->cdev_dev);
314}
315
316static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
317 int qid)
318{
319 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
320}
321
322static inline bool ublk_rq_has_data(const struct request *rq)
323{
731e208d 324 return bio_has_data(rq->bio);
71f28f31
ML		 325}
326
327static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
328 int tag)
329{
330 return (struct ublksrv_io_desc *)
331 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
332}
333
334static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
335{
336 return ublk_get_queue(ub, q_id)->io_cmd_buf;
337}
338
339static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
340{
341 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
342
343 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
344 PAGE_SIZE);
345}
346
a0d41dc1
Z		 347static inline bool ublk_queue_can_use_recovery_reissue(
348 struct ublk_queue *ubq)
349{
96cf2f54
ML		 350 return (ubq->flags & UBLK_F_USER_RECOVERY) &&
351 (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
a0d41dc1
Z		 352}
353
77a440e2
Z		 354static inline bool ublk_queue_can_use_recovery(
355 struct ublk_queue *ubq)
356{
96cf2f54 357 return ubq->flags & UBLK_F_USER_RECOVERY;
77a440e2
Z		 358}
359
360static inline bool ublk_can_use_recovery(struct ublk_device *ub)
361{
96cf2f54 362 return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
77a440e2
Z		 363}
364
6d9e6dfd
CH		 365static void ublk_free_disk(struct gendisk *disk)
366{
367 struct ublk_device *ub = disk->private_data;
368
369 clear_bit(UB_STATE_USED, &ub->state);
370 put_device(&ub->cdev_dev);
371}
372
48a90519
ML		 373static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
374 unsigned int *owner_gid)
375{
376 kuid_t uid;
377 kgid_t gid;
378
379 current_uid_gid(&uid, &gid);
380
381 *owner_uid = from_kuid(&init_user_ns, uid);
382 *owner_gid = from_kgid(&init_user_ns, gid);
383}
384
385static int ublk_open(struct block_device *bdev, fmode_t mode)
386{
387 struct ublk_device *ub = bdev->bd_disk->private_data;
388
389 if (capable(CAP_SYS_ADMIN))
390 return 0;
391
392 /*
393 * If it is one unprivileged device, only owner can open
394 * the disk. Otherwise it could be one trap made by one
395 * evil user who grants this disk's privileges to other
396 * users deliberately.
397 *
398 * This way is reasonable too given anyone can create
399 * unprivileged device, and no need other's grant.
400 */
401 if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
402 unsigned int curr_uid, curr_gid;
403
404 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
405
406 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
407 ub->dev_info.owner_gid)
408 return -EPERM;
409 }
410
411 return 0;
412}
413
71f28f31
ML		 414static const struct block_device_operations ub_fops = {
415 .owner = THIS_MODULE,
48a90519 416 .open = ublk_open,
6d9e6dfd 417 .free_disk = ublk_free_disk,
71f28f31
ML		 418};
419
420#define UBLK_MAX_PIN_PAGES 32
421
422struct ublk_map_data {
71f28f31 423 const struct request *rq;
ae9f5cce
ML		 424 unsigned long ubuf;
425 unsigned int len;
71f28f31
ML		 426};
427
428struct ublk_io_iter {
429 struct page *pages[UBLK_MAX_PIN_PAGES];
430 unsigned pg_off; /* offset in the 1st page in pages */
431 int nr_pages; /* how many page pointers in pages */
432 struct bio *bio;
433 struct bvec_iter iter;
434};
435
436static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
437 unsigned max_bytes, bool to_vm)
438{
439 const unsigned total = min_t(unsigned, max_bytes,
440 PAGE_SIZE - data->pg_off +
441 ((data->nr_pages - 1) << PAGE_SHIFT));
442 unsigned done = 0;
443 unsigned pg_idx = 0;
444
445 while (done < total) {
446 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
447 const unsigned int bytes = min3(bv.bv_len, total - done,
448 (unsigned)(PAGE_SIZE - data->pg_off));
449 void *bv_buf = bvec_kmap_local(&bv);
450 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
451
452 if (to_vm)
453 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
454 else
455 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
456
457 kunmap_local(pg_buf);
458 kunmap_local(bv_buf);
459
460 /* advance page array */
461 data->pg_off += bytes;
462 if (data->pg_off == PAGE_SIZE) {
463 pg_idx += 1;
464 data->pg_off = 0;
465 }
466
467 done += bytes;
468
469 /* advance bio */
470 bio_advance_iter_single(data->bio, &data->iter, bytes);
471 if (!data->iter.bi_size) {
472 data->bio = data->bio->bi_next;
473 if (data->bio == NULL)
474 break;
475 data->iter = data->bio->bi_iter;
476 }
477 }
478
479 return done;
480}
481
2f3af723 482static int ublk_copy_user_pages(struct ublk_map_data *data, bool to_vm)
71f28f31
ML		 483{
484 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
ae9f5cce 485 const unsigned long start_vm = data->ubuf;
71f28f31
ML		 486 unsigned int done = 0;
487 struct ublk_io_iter iter = {
488 .pg_off = start_vm & (PAGE_SIZE - 1),
489 .bio = data->rq->bio,
490 .iter = data->rq->bio->bi_iter,
491 };
ae9f5cce 492 const unsigned int nr_pages = round_up(data->len +
71f28f31
ML		 493 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
494
495 while (done < nr_pages) {
496 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
497 nr_pages - done);
498 unsigned i, len;
499
500 iter.nr_pages = get_user_pages_fast(start_vm +
501 (done << PAGE_SHIFT), to_pin, gup_flags,
502 iter.pages);
503 if (iter.nr_pages <= 0)
504 return done == 0 ? iter.nr_pages : done;
ae9f5cce 505 len = ublk_copy_io_pages(&iter, data->len, to_vm);
71f28f31
ML		 506 for (i = 0; i < iter.nr_pages; i++) {
507 if (to_vm)
508 set_page_dirty(iter.pages[i]);
509 put_page(iter.pages[i]);
510 }
ae9f5cce 511 data->len -= len;
71f28f31
ML		 512 done += iter.nr_pages;
513 }
514
515 return done;
516}
517
2f3af723
ML		 518static inline bool ublk_need_map_req(const struct request *req)
519{
520 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
521}
522
523static inline bool ublk_need_unmap_req(const struct request *req)
524{
525 return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
526}
527
71f28f31
ML		 528static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
529 struct ublk_io *io)
530{
531 const unsigned int rq_bytes = blk_rq_bytes(req);
23ef8220 532
71f28f31
ML		 533 /*
534 * no zero copy, we delay copy WRITE request data into ublksrv
535 * context and the big benefit is that pinning pages in current
536 * context is pretty fast, see ublk_pin_user_pages
537 */
2f3af723 538 if (ublk_need_map_req(req)) {
71f28f31 539 struct ublk_map_data data = {
71f28f31 540 .rq = req,
ae9f5cce
ML		 541 .ubuf = io->addr,
542 .len = rq_bytes,
71f28f31
ML		 543 };
544
545 ublk_copy_user_pages(&data, true);
546
ae9f5cce 547 return rq_bytes - data.len;
71f28f31
ML		 548 }
549 return rq_bytes;
550}
551
552static int ublk_unmap_io(const struct ublk_queue *ubq,
553 const struct request *req,
554 struct ublk_io *io)
555{
556 const unsigned int rq_bytes = blk_rq_bytes(req);
557
2f3af723 558 if (ublk_need_unmap_req(req)) {
71f28f31 559 struct ublk_map_data data = {
71f28f31 560 .rq = req,
ae9f5cce
ML		 561 .ubuf = io->addr,
562 .len = io->res,
71f28f31
ML		 563 };
564
565 WARN_ON_ONCE(io->res > rq_bytes);
566
567 ublk_copy_user_pages(&data, false);
568
ae9f5cce 569 return io->res - data.len;
71f28f31
ML		 570 }
571 return rq_bytes;
572}
573
574static inline unsigned int ublk_req_build_flags(struct request *req)
575{
576 unsigned flags = 0;
577
578 if (req->cmd_flags & REQ_FAILFAST_DEV)
579 flags |= UBLK_IO_F_FAILFAST_DEV;
580
581 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
582 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
583
584 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
585 flags |= UBLK_IO_F_FAILFAST_DRIVER;
586
587 if (req->cmd_flags & REQ_META)
588 flags |= UBLK_IO_F_META;
589
71f28f31
ML		 590 if (req->cmd_flags & REQ_FUA)
591 flags |= UBLK_IO_F_FUA;
592
71f28f31
ML		 593 if (req->cmd_flags & REQ_NOUNMAP)
594 flags |= UBLK_IO_F_NOUNMAP;
595
596 if (req->cmd_flags & REQ_SWAP)
597 flags |= UBLK_IO_F_SWAP;
598
599 return flags;
600}
601
f2450f8a 602static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
71f28f31
ML		 603{
604 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
605 struct ublk_io *io = &ubq->ios[req->tag];
606 u32 ublk_op;
607
608 switch (req_op(req)) {
609 case REQ_OP_READ:
610 ublk_op = UBLK_IO_OP_READ;
611 break;
612 case REQ_OP_WRITE:
613 ublk_op = UBLK_IO_OP_WRITE;
614 break;
615 case REQ_OP_FLUSH:
616 ublk_op = UBLK_IO_OP_FLUSH;
617 break;
618 case REQ_OP_DISCARD:
619 ublk_op = UBLK_IO_OP_DISCARD;
620 break;
621 case REQ_OP_WRITE_ZEROES:
622 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
623 break;
624 default:
625 return BLK_STS_IOERR;
626 }
627
628 /* need to translate since kernel may change */
629 iod->op_flags = ublk_op | ublk_req_build_flags(req);
630 iod->nr_sectors = blk_rq_sectors(req);
631 iod->start_sector = blk_rq_pos(req);
632 iod->addr = io->addr;
633
634 return BLK_STS_OK;
635}
636
637static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
638 struct io_uring_cmd *ioucmd)
639{
640 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
641}
642
966120b5 643static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
71f28f31
ML		 644{
645 return ubq->ubq_daemon->flags & PF_EXITING;
646}
647
648/* todo: handle partial completion */
649static void ublk_complete_rq(struct request *req)
650{
651 struct ublk_queue *ubq = req->mq_hctx->driver_data;
652 struct ublk_io *io = &ubq->ios[req->tag];
653 unsigned int unmapped_bytes;
903f8aee 654 blk_status_t res = BLK_STS_OK;
71f28f31
ML		 655
656 /* failed read IO if nothing is read */
657 if (!io->res && req_op(req) == REQ_OP_READ)
658 io->res = -EIO;
659
660 if (io->res < 0) {
903f8aee
ML		 661 res = errno_to_blk_status(io->res);
662 goto exit;
71f28f31
ML		 663 }
664
665 /*
b352389e 666 * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
71f28f31
ML		 667 * directly.
668 *
669 * Both the two needn't unmap.
670 */
903f8aee
ML		 671 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
672 goto exit;
71f28f31
ML		 673
674 /* for READ request, writing data in iod->addr to rq buffers */
675 unmapped_bytes = ublk_unmap_io(ubq, req, io);
676
677 /*
678 * Extremely impossible since we got data filled in just before
679 *
680 * Re-read simply for this unlikely case.
681 */
682 if (unlikely(unmapped_bytes < io->res))
683 io->res = unmapped_bytes;
684
685 if (blk_update_request(req, BLK_STS_OK, io->res))
686 blk_mq_requeue_request(req, true);
687 else
688 __blk_mq_end_request(req, BLK_STS_OK);
903f8aee
ML		 689
690 return;
691exit:
692 blk_mq_end_request(req, res);
71f28f31
ML		 693}
694
695/*
bb241747
Z		 696 * Since __ublk_rq_task_work always fails requests immediately during
697 * exiting, __ublk_fail_req() is only called from abort context during
698 * exiting. So lock is unnecessary.
71f28f31
ML		 699 *
700 * Also aborting may not be started yet, keep in mind that one failed
701 * request may be issued by block layer again.
702 */
a0d41dc1
Z		 703static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
704 struct request *req)
71f28f31
ML		 705{
706 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
707
708 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
709 io->flags |= UBLK_IO_FLAG_ABORTED;
a0d41dc1
Z		 710 if (ublk_queue_can_use_recovery_reissue(ubq))
711 blk_mq_requeue_request(req, false);
712 else
713 blk_mq_end_request(req, BLK_STS_IOERR);
71f28f31
ML		 714 }
715}
716
c86019ff
Z		 717static void ubq_complete_io_cmd(struct ublk_io *io, int res)
718{
719 /* mark this cmd owned by ublksrv */
720 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
721
722 /*
723 * clear ACTIVE since we are done with this sqe/cmd slot
724 * We can only accept io cmd in case of being not active.
725 */
726 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
727
728 /* tell ublksrv one io request is coming */
729 io_uring_cmd_done(io->cmd, res, 0);
730}
731
71f28f31
ML		 732#define UBLK_REQUEUE_DELAY_MS 3
733
42cf5fc5
Z		 734static inline void __ublk_abort_rq(struct ublk_queue *ubq,
735 struct request *rq)
736{
737 /* We cannot process this rq so just requeue it. */
738 if (ublk_queue_can_use_recovery(ubq))
739 blk_mq_requeue_request(rq, false);
740 else
741 blk_mq_end_request(rq, BLK_STS_IOERR);
742
743 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
744}
745
0edb3696 746static inline void __ublk_rq_task_work(struct request *req)
71f28f31 747{
71f28f31
ML		 748 struct ublk_queue *ubq = req->mq_hctx->driver_data;
749 int tag = req->tag;
750 struct ublk_io *io = &ubq->ios[tag];
71f28f31
ML		 751 unsigned int mapped_bytes;
752
753 pr_devel("%s: complete: 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);
756
ae3f7193
Z		 757 /*
758 * Task is exiting if either:
759 *
760 * (1) current != ubq_daemon.
761 * io_uring_cmd_complete_in_task() tries to run task_work
762 * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
763 *
764 * (2) current->flags & PF_EXITING.
765 */
766 if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
42cf5fc5 767 __ublk_abort_rq(ubq, req);
71f28f31
ML		 768 return;
769 }
770
2f3af723 771 if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
c86019ff
Z		 772 /*
773 * We have not handled UBLK_IO_NEED_GET_DATA command yet,
774 * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
775 * and notify it.
776 */
777 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
778 io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
779 pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
780 __func__, io->cmd->cmd_op, ubq->q_id,
781 req->tag, io->flags);
782 ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
783 return;
784 }
785 /*
786 * We have handled UBLK_IO_NEED_GET_DATA command,
787 * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
788 * do the copy work.
789 */
790 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
92cb6e2e
Z		 791 /* update iod->addr because ublksrv may have passed a new io buffer */
792 ublk_get_iod(ubq, req->tag)->addr = io->addr;
793 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
794 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
795 ublk_get_iod(ubq, req->tag)->addr);
c86019ff
Z		 796 }
797
71f28f31
ML		 798 mapped_bytes = ublk_map_io(ubq, req, io);
799
800 /* partially mapped, update io descriptor */
801 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
802 /*
803 * Nothing mapped, retry until we succeed.
804 *
805 * We may never succeed in mapping any bytes here because
806 * of OOM. TODO: reserve one buffer with single page pinned
807 * for providing forward progress guarantee.
808 */
809 if (unlikely(!mapped_bytes)) {
810 blk_mq_requeue_request(req, false);
811 blk_mq_delay_kick_requeue_list(req->q,
812 UBLK_REQUEUE_DELAY_MS);
813 return;
814 }
815
816 ublk_get_iod(ubq, req->tag)->nr_sectors =
817 mapped_bytes >> 9;
818 }
819
c86019ff 820 ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
71f28f31
ML		 821}
822
7d4a9317
ML		 823static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
824{
825 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
826 struct ublk_rq_data *data, *tmp;
827
828 io_cmds = llist_reverse_order(io_cmds);
829 llist_for_each_entry_safe(data, tmp, io_cmds, node)
830 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
831}
832
833static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
834{
835 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
836 struct ublk_rq_data *data, *tmp;
837
838 llist_for_each_entry_safe(data, tmp, io_cmds, node)
839 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
840}
841
0edb3696
ML		 842static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
843{
844 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
3ab6e94c 845 struct ublk_queue *ubq = pdu->ubq;
0edb3696 846
7d4a9317 847 ublk_forward_io_cmds(ubq);
0edb3696
ML		 848}
849
850static void ublk_rq_task_work_fn(struct callback_head *work)
851{
852 struct ublk_rq_data *data = container_of(work,
853 struct ublk_rq_data, work);
854 struct request *req = blk_mq_rq_from_pdu(data);
7d4a9317 855 struct ublk_queue *ubq = req->mq_hctx->driver_data;
0edb3696 856
7d4a9317 857 ublk_forward_io_cmds(ubq);
0edb3696
ML		 858}
859
7d4a9317 860static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
3ab6e94c 861{
7d4a9317
ML		 862 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
863 struct ublk_io *io;
3ab6e94c 864
7d4a9317
ML		 865 if (!llist_add(&data->node, &ubq->io_cmds))
866 return;
867
868 io = &ubq->ios[rq->tag];
3ab6e94c
ML		 869 /*
870 * If the check pass, we know that this is a re-issued request aborted
871 * previously in monitor_work because the ubq_daemon(cmd's task) is
872 * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
873 * because this ioucmd's io_uring context may be freed now if no inflight
874 * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
875 *
876 * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
877 * the tag). Then the request is re-started(allocating the tag) and we are here.
878 * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
879 * guarantees that here is a re-issued request aborted previously.
880 */
881 if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
7d4a9317
ML		 882 ublk_abort_io_cmds(ubq);
883 } else if (ublk_can_use_task_work(ubq)) {
884 if (task_work_add(ubq->ubq_daemon, &data->work,
885 TWA_SIGNAL_NO_IPI))
886 ublk_abort_io_cmds(ubq);
3ab6e94c
ML		 887 } else {
888 struct io_uring_cmd *cmd = io->cmd;
889 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
890
891 pdu->ubq = ubq;
892 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
893 }
894}
895
71f28f31
ML		 896static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
897 const struct blk_mq_queue_data *bd)
898{
899 struct ublk_queue *ubq = hctx->driver_data;
900 struct request *rq = bd->rq;
71f28f31
ML		 901 blk_status_t res;
902
903 /* fill iod to slot in io cmd buffer */
904 res = ublk_setup_iod(ubq, rq);
905 if (unlikely(res != BLK_STS_OK))
906 return BLK_STS_IOERR;
3ab6e94c 907
bbae8d1f
Z		 908 /* With recovery feature enabled, force_abort is set in
909 * ublk_stop_dev() before calling del_gendisk(). We have to
910 * abort all requeued and new rqs here to let del_gendisk()
911 * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
912 * to avoid UAF on io_uring ctx.
913 *
914 * Note: force_abort is guaranteed to be seen because it is set
915 * before request queue is unqiuesced.
916 */
917 if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
918 return BLK_STS_IOERR;
71f28f31
ML		 919
920 blk_mq_start_request(bd->rq);
921
922 if (unlikely(ubq_daemon_is_dying(ubq))) {
42cf5fc5
Z		 923 __ublk_abort_rq(ubq, rq);
924 return BLK_STS_OK;
71f28f31
ML		 925 }
926
7d4a9317 927 ublk_queue_cmd(ubq, rq);
0edb3696 928
71f28f31
ML		 929 return BLK_STS_OK;
930}
931
71f28f31
ML		 932static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
933 unsigned int hctx_idx)
934{
cebbe577 935 struct ublk_device *ub = driver_data;
71f28f31
ML		 936 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
937
938 hctx->driver_data = ubq;
939 return 0;
940}
941
0edb3696
ML		 942static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
943 unsigned int hctx_idx, unsigned int numa_node)
944{
945 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
946
947 init_task_work(&data->work, ublk_rq_task_work_fn);
948 return 0;
949}
950
71f28f31
ML		 951static const struct blk_mq_ops ublk_mq_ops = {
952 .queue_rq = ublk_queue_rq,
953 .init_hctx = ublk_init_hctx,
0edb3696 954 .init_request = ublk_init_rq,
71f28f31
ML		 955};
956
957static int ublk_ch_open(struct inode *inode, struct file *filp)
958{
959 struct ublk_device *ub = container_of(inode->i_cdev,
960 struct ublk_device, cdev);
961
fa362045
CH		 962 if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
963 return -EBUSY;
964 filp->private_data = ub;
965 return 0;
71f28f31
ML		 966}
967
968static int ublk_ch_release(struct inode *inode, struct file *filp)
969{
970 struct ublk_device *ub = filp->private_data;
971
fa362045 972 clear_bit(UB_STATE_OPEN, &ub->state);
71f28f31
ML		 973 return 0;
974}
975
976/* map pre-allocated per-queue cmd buffer to ublksrv daemon */
977static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
978{
979 struct ublk_device *ub = filp->private_data;
980 size_t sz = vma->vm_end - vma->vm_start;
981 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
982 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
983 int q_id, ret = 0;
984
e94eb459 985 spin_lock(&ub->mm_lock);
71f28f31
ML		 986 if (!ub->mm)
987 ub->mm = current->mm;
988 if (current->mm != ub->mm)
989 ret = -EINVAL;
e94eb459 990 spin_unlock(&ub->mm_lock);
71f28f31
ML		 991
992 if (ret)
993 return ret;
994
995 if (vma->vm_flags & VM_WRITE)
996 return -EPERM;
997
998 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
999 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1000 return -EINVAL;
1001
1002 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1003 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1004 __func__, q_id, current->pid, vma->vm_start,
1005 phys_off, (unsigned long)sz);
1006
1007 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1008 return -EINVAL;
1009
1010 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1011 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1012}
1013
1014static void ublk_commit_completion(struct ublk_device *ub,
1015 struct ublksrv_io_cmd *ub_cmd)
1016{
1017 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1018 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1019 struct ublk_io *io = &ubq->ios[tag];
1020 struct request *req;
1021
1022 /* now this cmd slot is owned by nbd driver */
1023 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1024 io->res = ub_cmd->result;
1025
1026 /* find the io request and complete */
1027 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1028
1029 if (req && likely(!blk_should_fake_timeout(req->q)))
1030 ublk_complete_rq(req);
1031}
1032
1033/*
1034 * When ->ubq_daemon is exiting, either new request is ended immediately,
1035 * or any queued io command is drained, so it is safe to abort queue
1036 * lockless
1037 */
1038static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1039{
1040 int i;
1041
1042 if (!ublk_get_device(ub))
1043 return;
1044
1045 for (i = 0; i < ubq->q_depth; i++) {
1046 struct ublk_io *io = &ubq->ios[i];
1047
1048 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1049 struct request *rq;
1050
1051 /*
1052 * Either we fail the request or ublk_rq_task_work_fn
1053 * will do it
1054 */
1055 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1056 if (rq)
a0d41dc1 1057 __ublk_fail_req(ubq, io, rq);
71f28f31
ML		 1058 }
1059 }
1060 ublk_put_device(ub);
1061}
1062
1063static void ublk_daemon_monitor_work(struct work_struct *work)
1064{
1065 struct ublk_device *ub =
1066 container_of(work, struct ublk_device, monitor_work.work);
1067 int i;
1068
1069 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1070 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1071
1072 if (ubq_daemon_is_dying(ubq)) {
bbae8d1f
Z		 1073 if (ublk_queue_can_use_recovery(ubq))
1074 schedule_work(&ub->quiesce_work);
1075 else
1076 schedule_work(&ub->stop_work);
71f28f31
ML		 1077
1078 /* abort queue is for making forward progress */
1079 ublk_abort_queue(ub, ubq);
1080 }
1081 }
1082
1083 /*
bbae8d1f
Z		 1084 * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1085 * after ublk_remove() or __ublk_quiesce_dev() is started.
71f28f31
ML		 1086 *
1087 * No need ub->mutex, monitor work are canceled after state is marked
bbae8d1f 1088 * as not LIVE, so new state is observed reliably.
71f28f31 1089 */
bbae8d1f 1090 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
71f28f31
ML		 1091 schedule_delayed_work(&ub->monitor_work,
1092 UBLK_DAEMON_MONITOR_PERIOD);
1093}
1094
a8ce5f52
ML		 1095static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1096{
1097 return ubq->nr_io_ready == ubq->q_depth;
1098}
1099
71f28f31
ML		 1100static void ublk_cancel_queue(struct ublk_queue *ubq)
1101{
1102 int i;
1103
a8ce5f52
ML		 1104 if (!ublk_queue_ready(ubq))
1105 return;
1106
71f28f31
ML		 1107 for (i = 0; i < ubq->q_depth; i++) {
1108 struct ublk_io *io = &ubq->ios[i];
1109
1110 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1111 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1112 }
a8ce5f52
ML		 1113
1114 /* all io commands are canceled */
1115 ubq->nr_io_ready = 0;
71f28f31
ML		 1116}
1117
1118/* Cancel all pending commands, must be called after del_gendisk() returns */
1119static void ublk_cancel_dev(struct ublk_device *ub)
1120{
1121 int i;
1122
1123 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1124 ublk_cancel_queue(ublk_get_queue(ub, i));
1125}
1126
bbae8d1f
Z		 1127static bool ublk_check_inflight_rq(struct request *rq, void *data)
1128{
1129 bool *idle = data;
1130
1131 if (blk_mq_request_started(rq)) {
1132 *idle = false;
1133 return false;
1134 }
1135 return true;
1136}
1137
1138static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1139{
1140 bool idle;
1141
1142 WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1143 while (true) {
1144 idle = true;
1145 blk_mq_tagset_busy_iter(&ub->tag_set,
1146 ublk_check_inflight_rq, &idle);
1147 if (idle)
1148 break;
1149 msleep(UBLK_REQUEUE_DELAY_MS);
1150 }
1151}
1152
1153static void __ublk_quiesce_dev(struct ublk_device *ub)
71f28f31 1154{
bbae8d1f
Z		 1155 pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1156 __func__, ub->dev_info.dev_id,
1157 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1158 "LIVE" : "QUIESCED");
1159 blk_mq_quiesce_queue(ub->ub_disk->queue);
1160 ublk_wait_tagset_rqs_idle(ub);
1161 ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1162 ublk_cancel_dev(ub);
1163 /* we are going to release task_struct of ubq_daemon and resets
1164 * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1165 * Besides, monitor_work is not necessary in QUIESCED state since we have
1166 * already scheduled quiesce_work and quiesced all ubqs.
1167 *
1168 * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1169 * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1170 */
1171 cancel_delayed_work_sync(&ub->monitor_work);
1172}
1173
1174static void ublk_quiesce_work_fn(struct work_struct *work)
1175{
1176 struct ublk_device *ub =
1177 container_of(work, struct ublk_device, quiesce_work);
1178
71f28f31 1179 mutex_lock(&ub->mutex);
6d9e6dfd 1180 if (ub->dev_info.state != UBLK_S_DEV_LIVE)
71f28f31 1181 goto unlock;
bbae8d1f
Z		 1182 __ublk_quiesce_dev(ub);
1183 unlock:
1184 mutex_unlock(&ub->mutex);
1185}
71f28f31 1186
bbae8d1f
Z		 1187static void ublk_unquiesce_dev(struct ublk_device *ub)
1188{
1189 int i;
1190
1191 pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1192 __func__, ub->dev_info.dev_id,
1193 ub->dev_info.state == UBLK_S_DEV_LIVE ?
1194 "LIVE" : "QUIESCED");
1195 /* quiesce_work has run. We let requeued rqs be aborted
1196 * before running fallback_wq. "force_abort" must be seen
1197 * after request queue is unqiuesced. Then del_gendisk()
1198 * can move on.
1199 */
1200 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1201 ublk_get_queue(ub, i)->force_abort = true;
1202
1203 blk_mq_unquiesce_queue(ub->ub_disk->queue);
1204 /* We may have requeued some rqs in ublk_quiesce_queue() */
1205 blk_mq_kick_requeue_list(ub->ub_disk->queue);
1206}
1207
1208static void ublk_stop_dev(struct ublk_device *ub)
1209{
1210 mutex_lock(&ub->mutex);
1211 if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1212 goto unlock;
1213 if (ublk_can_use_recovery(ub)) {
1214 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1215 __ublk_quiesce_dev(ub);
1216 ublk_unquiesce_dev(ub);
1217 }
71f28f31
ML		 1218 del_gendisk(ub->ub_disk);
1219 ub->dev_info.state = UBLK_S_DEV_DEAD;
1220 ub->dev_info.ublksrv_pid = -1;
6d9e6dfd
CH		 1221 put_disk(ub->ub_disk);
1222 ub->ub_disk = NULL;
71f28f31 1223 unlock:
a8ce5f52 1224 ublk_cancel_dev(ub);
71f28f31
ML		 1225 mutex_unlock(&ub->mutex);
1226 cancel_delayed_work_sync(&ub->monitor_work);
1227}
1228
71f28f31
ML		 1229/* device can only be started after all IOs are ready */
1230static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1231{
1232 mutex_lock(&ub->mutex);
1233 ubq->nr_io_ready++;
1234 if (ublk_queue_ready(ubq)) {
1235 ubq->ubq_daemon = current;
1236 get_task_struct(ubq->ubq_daemon);
1237 ub->nr_queues_ready++;
73a166d9
ML		 1238
1239 if (capable(CAP_SYS_ADMIN))
1240 ub->nr_privileged_daemon++;
71f28f31
ML		 1241 }
1242 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1243 complete_all(&ub->completion);
1244 mutex_unlock(&ub->mutex);
1245}
1246
c86019ff 1247static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
fee32f31 1248 int tag)
c86019ff
Z		 1249{
1250 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1251 struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1252
7d4a9317 1253 ublk_queue_cmd(ubq, req);
c86019ff
Z		 1254}
1255
71f28f31
ML		 1256static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1257{
1258 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1259 struct ublk_device *ub = cmd->file->private_data;
1260 struct ublk_queue *ubq;
1261 struct ublk_io *io;
1262 u32 cmd_op = cmd->cmd_op;
1263 unsigned tag = ub_cmd->tag;
1264 int ret = -EINVAL;
2f1e07dd 1265 struct request *req;
71f28f31
ML		 1266
1267 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1268 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1269 ub_cmd->result);
1270
71f28f31
ML		 1271 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1272 goto out;
1273
1274 ubq = ublk_get_queue(ub, ub_cmd->q_id);
1275 if (!ubq || ub_cmd->q_id != ubq->q_id)
1276 goto out;
1277
1278 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1279 goto out;
1280
1281 if (tag >= ubq->q_depth)
1282 goto out;
1283
1284 io = &ubq->ios[tag];
1285
1286 /* there is pending io cmd, something must be wrong */
1287 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1288 ret = -EBUSY;
1289 goto out;
1290 }
1291
c86019ff
Z		 1292 /*
1293 * ensure that the user issues UBLK_IO_NEED_GET_DATA
1294 * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1295 */
1296 if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1297 ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1298 goto out;
1299
71f28f31
ML		 1300 switch (cmd_op) {
1301 case UBLK_IO_FETCH_REQ:
1302 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1303 if (ublk_queue_ready(ubq)) {
1304 ret = -EBUSY;
1305 goto out;
1306 }
1307 /*
1308 * The io is being handled by server, so COMMIT_RQ is expected
1309 * instead of FETCH_REQ
1310 */
1311 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1312 goto out;
2f1e07dd
LX		 1313 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1314 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
71f28f31
ML		 1315 goto out;
1316 io->cmd = cmd;
1317 io->flags |= UBLK_IO_FLAG_ACTIVE;
1318 io->addr = ub_cmd->addr;
1319
1320 ublk_mark_io_ready(ub, ubq);
1321 break;
1322 case UBLK_IO_COMMIT_AND_FETCH_REQ:
2f1e07dd
LX		 1323 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1324 /*
1325 * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1326 * not enabled or it is Read IO.
1327 */
1328 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
71f28f31
ML		 1329 goto out;
1330 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1331 goto out;
1332 io->addr = ub_cmd->addr;
1333 io->flags |= UBLK_IO_FLAG_ACTIVE;
1334 io->cmd = cmd;
1335 ublk_commit_completion(ub, ub_cmd);
1336 break;
c86019ff
Z		 1337 case UBLK_IO_NEED_GET_DATA:
1338 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1339 goto out;
1340 io->addr = ub_cmd->addr;
1341 io->cmd = cmd;
1342 io->flags |= UBLK_IO_FLAG_ACTIVE;
fee32f31 1343 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
c86019ff 1344 break;
71f28f31
ML		 1345 default:
1346 goto out;
1347 }
1348 return -EIOCBQUEUED;
1349
1350 out:
71f28f31
ML		 1351 io_uring_cmd_done(cmd, ret, 0);
1352 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1353 __func__, cmd_op, tag, ret, io->flags);
1354 return -EIOCBQUEUED;
1355}
1356
1357static const struct file_operations ublk_ch_fops = {
1358 .owner = THIS_MODULE,
1359 .open = ublk_ch_open,
1360 .release = ublk_ch_release,
1361 .llseek = no_llseek,
1362 .uring_cmd = ublk_ch_uring_cmd,
1363 .mmap = ublk_ch_mmap,
1364};
1365
1366static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1367{
1368 int size = ublk_queue_cmd_buf_size(ub, q_id);
1369 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1370
1371 if (ubq->ubq_daemon)
1372 put_task_struct(ubq->ubq_daemon);
1373 if (ubq->io_cmd_buf)
1374 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1375}
1376
1377static int ublk_init_queue(struct ublk_device *ub, int q_id)
1378{
1379 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1380 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1381 void *ptr;
1382 int size;
1383
6d8c5afc 1384 ubq->flags = ub->dev_info.flags;
71f28f31
ML		 1385 ubq->q_id = q_id;
1386 ubq->q_depth = ub->dev_info.queue_depth;
1387 size = ublk_queue_cmd_buf_size(ub, q_id);
1388
1389 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1390 if (!ptr)
1391 return -ENOMEM;
1392
1393 ubq->io_cmd_buf = ptr;
1394 ubq->dev = ub;
1395 return 0;
1396}
1397
1398static void ublk_deinit_queues(struct ublk_device *ub)
1399{
1400 int nr_queues = ub->dev_info.nr_hw_queues;
1401 int i;
1402
1403 if (!ub->__queues)
1404 return;
1405
1406 for (i = 0; i < nr_queues; i++)
1407 ublk_deinit_queue(ub, i);
1408 kfree(ub->__queues);
1409}
1410
1411static int ublk_init_queues(struct ublk_device *ub)
1412{
1413 int nr_queues = ub->dev_info.nr_hw_queues;
1414 int depth = ub->dev_info.queue_depth;
1415 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1416 int i, ret = -ENOMEM;
1417
1418 ub->queue_size = ubq_size;
1419 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1420 if (!ub->__queues)
1421 return ret;
1422
1423 for (i = 0; i < nr_queues; i++) {
1424 if (ublk_init_queue(ub, i))
1425 goto fail;
1426 }
1427
1428 init_completion(&ub->completion);
1429 return 0;
1430
1431 fail:
1432 ublk_deinit_queues(ub);
1433 return ret;
1434}
1435
fa9482e0 1436static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
71f28f31
ML		 1437{
1438 int i = idx;
1439 int err;
1440
1441 spin_lock(&ublk_idr_lock);
1442 /* allocate id, if @id >= 0, we're requesting that specific id */
1443 if (i >= 0) {
1444 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1445 if (err == -ENOSPC)
1446 err = -EEXIST;
1447 } else {
1448 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1449 }
1450 spin_unlock(&ublk_idr_lock);
1451
1452 if (err >= 0)
1453 ub->ub_number = err;
1454
1455 return err;
1456}
1457
fa9482e0 1458static void ublk_free_dev_number(struct ublk_device *ub)
71f28f31
ML		 1459{
1460 spin_lock(&ublk_idr_lock);
1461 idr_remove(&ublk_index_idr, ub->ub_number);
1462 wake_up_all(&ublk_idr_wq);
1463 spin_unlock(&ublk_idr_lock);
71f28f31
ML		 1464}
1465
1466static void ublk_cdev_rel(struct device *dev)
1467{
1468 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1469
71f28f31 1470 blk_mq_free_tag_set(&ub->tag_set);
71f28f31 1471 ublk_deinit_queues(ub);
fa9482e0
CH		 1472 ublk_free_dev_number(ub);
1473 mutex_destroy(&ub->mutex);
1474 kfree(ub);
71f28f31
ML		 1475}
1476
1477static int ublk_add_chdev(struct ublk_device *ub)
1478{
1479 struct device *dev = &ub->cdev_dev;
1480 int minor = ub->ub_number;
1481 int ret;
1482
1483 dev->parent = ublk_misc.this_device;
1484 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1485 dev->class = ublk_chr_class;
1486 dev->release = ublk_cdev_rel;
1487 device_initialize(dev);
1488
1489 ret = dev_set_name(dev, "ublkc%d", minor);
1490 if (ret)
1491 goto fail;
1492
1493 cdev_init(&ub->cdev, &ublk_ch_fops);
1494 ret = cdev_device_add(&ub->cdev, dev);
1495 if (ret)
1496 goto fail;
403ebc87
ML		 1497
1498 ublks_added++;
71f28f31
ML		 1499 return 0;
1500 fail:
1501 put_device(dev);
1502 return ret;
1503}
1504
1505static void ublk_stop_work_fn(struct work_struct *work)
1506{
1507 struct ublk_device *ub =
1508 container_of(work, struct ublk_device, stop_work);
1509
1510 ublk_stop_dev(ub);
1511}
1512
4bf9cbf3 1513/* align max io buffer size with PAGE_SIZE */
6d9e6dfd 1514static void ublk_align_max_io_size(struct ublk_device *ub)
71f28f31 1515{
4bf9cbf3 1516 unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
71f28f31 1517
4bf9cbf3
ML		 1518 ub->dev_info.max_io_buf_bytes =
1519 round_down(max_io_bytes, PAGE_SIZE);
71f28f31
ML		 1520}
1521
fa9482e0 1522static int ublk_add_tag_set(struct ublk_device *ub)
71f28f31 1523{
71f28f31
ML		 1524 ub->tag_set.ops = &ublk_mq_ops;
1525 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1526 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1527 ub->tag_set.numa_node = NUMA_NO_NODE;
0edb3696 1528 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
71f28f31
ML		 1529 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1530 ub->tag_set.driver_data = ub;
fa9482e0 1531 return blk_mq_alloc_tag_set(&ub->tag_set);
71f28f31
ML		 1532}
1533
1534static void ublk_remove(struct ublk_device *ub)
1535{
34d8f2be
CH		 1536 ublk_stop_dev(ub);
1537 cancel_work_sync(&ub->stop_work);
bbae8d1f 1538 cancel_work_sync(&ub->quiesce_work);
71f28f31
ML		 1539 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1540 put_device(&ub->cdev_dev);
403ebc87 1541 ublks_added--;
71f28f31
ML		 1542}
1543
1544static struct ublk_device *ublk_get_device_from_id(int idx)
1545{
1546 struct ublk_device *ub = NULL;
1547
1548 if (idx < 0)
1549 return NULL;
1550
1551 spin_lock(&ublk_idr_lock);
1552 ub = idr_find(&ublk_index_idr, idx);
1553 if (ub)
1554 ub = ublk_get_device(ub);
1555 spin_unlock(&ublk_idr_lock);
1556
1557 return ub;
1558}
1559
bfbcef03 1560static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
71f28f31
ML		 1561{
1562 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
71f28f31 1563 int ublksrv_pid = (int)header->data[0];
6d9e6dfd 1564 struct gendisk *disk;
34d8f2be 1565 int ret = -EINVAL;
71f28f31
ML		 1566
1567 if (ublksrv_pid <= 0)
34d8f2be
CH		 1568 return -EINVAL;
1569
71f28f31
ML		 1570 wait_for_completion_interruptible(&ub->completion);
1571
1572 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1573
1574 mutex_lock(&ub->mutex);
6d9e6dfd
CH		 1575 if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1576 test_bit(UB_STATE_USED, &ub->state)) {
71f28f31 1577 ret = -EEXIST;
34d8f2be 1578 goto out_unlock;
71f28f31 1579 }
71f28f31 1580
1972d038 1581 disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
6d9e6dfd
CH		 1582 if (IS_ERR(disk)) {
1583 ret = PTR_ERR(disk);
1584 goto out_unlock;
34d8f2be 1585 }
6d9e6dfd
CH		 1586 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1587 disk->fops = &ub_fops;
1588 disk->private_data = ub;
1589
34d8f2be 1590 ub->dev_info.ublksrv_pid = ublksrv_pid;
6d9e6dfd 1591 ub->ub_disk = disk;
0aa73170
ML		 1592
1593 ret = ublk_apply_params(ub);
1594 if (ret)
1595 goto out_put_disk;
1596
73a166d9
ML		 1597 /* don't probe partitions if any one ubq daemon is un-trusted */
1598 if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1599 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1600
6d9e6dfd
CH		 1601 get_device(&ub->cdev_dev);
1602 ret = add_disk(disk);
1603 if (ret) {
93d71ec8
ML		 1604 /*
1605 * Has to drop the reference since ->free_disk won't be
1606 * called in case of add_disk failure.
1607 */
1608 ublk_put_device(ub);
0aa73170 1609 goto out_put_disk;
6d9e6dfd
CH		 1610 }
1611 set_bit(UB_STATE_USED, &ub->state);
34d8f2be 1612 ub->dev_info.state = UBLK_S_DEV_LIVE;
0aa73170
ML		 1613out_put_disk:
1614 if (ret)
1615 put_disk(disk);
34d8f2be
CH		 1616out_unlock:
1617 mutex_unlock(&ub->mutex);
71f28f31
ML		 1618 return ret;
1619}
1620
bfbcef03
ML		 1621static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1622 struct io_uring_cmd *cmd)
71f28f31
ML		 1623{
1624 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1625 void __user *argp = (void __user *)(unsigned long)header->addr;
c50061f0 1626 cpumask_var_t cpumask;
71f28f31
ML		 1627 unsigned long queue;
1628 unsigned int retlen;
c50061f0 1629 unsigned int i;
bfbcef03
ML		 1630 int ret;
1631
34d8f2be
CH		 1632 if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1633 return -EINVAL;
1634 if (header->len & (sizeof(unsigned long)-1))
1635 return -EINVAL;
1636 if (!header->addr)
1637 return -EINVAL;
71f28f31 1638
71f28f31
ML		 1639 queue = header->data[0];
1640 if (queue >= ub->dev_info.nr_hw_queues)
bfbcef03 1641 return -EINVAL;
71f28f31 1642
c50061f0 1643 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
bfbcef03 1644 return -ENOMEM;
c50061f0
CH		 1645
1646 for_each_possible_cpu(i) {
1647 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1648 cpumask_set_cpu(i, cpumask);
71f28f31 1649 }
c50061f0
CH		 1650
1651 ret = -EFAULT;
1652 retlen = min_t(unsigned short, header->len, cpumask_size());
1653 if (copy_to_user(argp, cpumask, retlen))
1654 goto out_free_cpumask;
1655 if (retlen != header->len &&
1656 clear_user(argp + retlen, header->len - retlen))
1657 goto out_free_cpumask;
1658
71f28f31 1659 ret = 0;
c50061f0
CH		 1660out_free_cpumask:
1661 free_cpumask_var(cpumask);
71f28f31
ML		 1662 return ret;
1663}
1664
34d8f2be 1665static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
71f28f31 1666{
34d8f2be 1667 pr_devel("%s: dev id %d flags %llx\n", __func__,
6d8c5afc 1668 info->dev_id, info->flags);
4bf9cbf3
ML		 1669 pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1670 info->nr_hw_queues, info->queue_depth);
34d8f2be
CH		 1671}
1672
1673static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1674{
1675 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1676 void __user *argp = (void __user *)(unsigned long)header->addr;
1677 struct ublksrv_ctrl_dev_info info;
71f28f31 1678 struct ublk_device *ub;
34d8f2be
CH		 1679 int ret = -EINVAL;
1680
1681 if (header->len < sizeof(info) || !header->addr)
1682 return -EINVAL;
1683 if (header->queue_id != (u16)-1) {
1684 pr_warn("%s: queue_id is wrong %x\n",
1685 __func__, header->queue_id);
1686 return -EINVAL;
1687 }
4093cb5a 1688
34d8f2be
CH		 1689 if (copy_from_user(&info, argp, sizeof(info)))
1690 return -EFAULT;
4093cb5a
ML		 1691
1692 if (capable(CAP_SYS_ADMIN))
1693 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1694 else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1695 return -EPERM;
1696
1697 /* the created device is always owned by current user */
48a90519 1698 ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
4093cb5a 1699
34d8f2be
CH		 1700 if (header->dev_id != info.dev_id) {
1701 pr_warn("%s: dev id not match %u %u\n",
1702 __func__, header->dev_id, info.dev_id);
1703 return -EINVAL;
1704 }
71f28f31 1705
4093cb5a
ML		 1706 ublk_dump_dev_info(&info);
1707
71f28f31
ML		 1708 ret = mutex_lock_killable(&ublk_ctl_mutex);
1709 if (ret)
1710 return ret;
1711
403ebc87
ML		 1712 ret = -EACCES;
1713 if (ublks_added >= ublks_max)
1714 goto out_unlock;
1715
cfee7e4d
CH		 1716 ret = -ENOMEM;
1717 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1718 if (!ub)
1719 goto out_unlock;
fa9482e0
CH		 1720 mutex_init(&ub->mutex);
1721 spin_lock_init(&ub->mm_lock);
bbae8d1f 1722 INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
fa9482e0
CH		 1723 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1724 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
cfee7e4d 1725
fa9482e0
CH		 1726 ret = ublk_alloc_dev_number(ub, header->dev_id);
1727 if (ret < 0)
1728 goto out_free_ub;
71f28f31 1729
34d8f2be 1730 memcpy(&ub->dev_info, &info, sizeof(info));
71f28f31 1731
34d8f2be
CH		 1732 /* update device id */
1733 ub->dev_info.dev_id = ub->ub_number;
1734
6d8c5afc
ML		 1735 /*
1736 * 64bit flags will be copied back to userspace as feature
1737 * negotiation result, so have to clear flags which driver
1738 * doesn't support yet, then userspace can get correct flags
1739 * (features) to handle.
1740 */
1741 ub->dev_info.flags &= UBLK_F_ALL;
1742
224e858f
ML		 1743 if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1744 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1745
fa9482e0 1746 /* We are not ready to support zero copy */
6d8c5afc 1747 ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
fa9482e0 1748
fa9482e0
CH		 1749 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1750 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1751 ublk_align_max_io_size(ub);
1752
1753 ret = ublk_init_queues(ub);
34d8f2be 1754 if (ret)
fa9482e0 1755 goto out_free_dev_number;
34d8f2be 1756
fa9482e0
CH		 1757 ret = ublk_add_tag_set(ub);
1758 if (ret)
1759 goto out_deinit_queues;
1760
1761 ret = -EFAULT;
1762 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1763 goto out_free_tag_set;
1764
1765 /*
1766 * Add the char dev so that ublksrv daemon can be setup.
1767 * ublk_add_chdev() will cleanup everything if it fails.
1768 */
1769 ret = ublk_add_chdev(ub);
1770 goto out_unlock;
1771
1772out_free_tag_set:
1773 blk_mq_free_tag_set(&ub->tag_set);
1774out_deinit_queues:
1775 ublk_deinit_queues(ub);
1776out_free_dev_number:
1777 ublk_free_dev_number(ub);
1778out_free_ub:
1779 mutex_destroy(&ub->mutex);
1780 kfree(ub);
34d8f2be 1781out_unlock:
71f28f31 1782 mutex_unlock(&ublk_ctl_mutex);
71f28f31
ML		 1783 return ret;
1784}
1785
1786static inline bool ublk_idr_freed(int id)
1787{
1788 void *ptr;
1789
1790 spin_lock(&ublk_idr_lock);
1791 ptr = idr_find(&ublk_index_idr, id);
1792 spin_unlock(&ublk_idr_lock);
1793
1794 return ptr == NULL;
1795}
1796
bfbcef03 1797static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
71f28f31 1798{
bfbcef03
ML		 1799 struct ublk_device *ub = *p_ub;
1800 int idx = ub->ub_number;
71f28f31
ML		 1801 int ret;
1802
1803 ret = mutex_lock_killable(&ublk_ctl_mutex);
1804 if (ret)
1805 return ret;
1806
0abe39de 1807 if (!test_bit(UB_STATE_DELETED, &ub->state)) {
71f28f31 1808 ublk_remove(ub);
0abe39de 1809 set_bit(UB_STATE_DELETED, &ub->state);
71f28f31
ML		 1810 }
1811
bfbcef03
ML		 1812 /* Mark the reference as consumed */
1813 *p_ub = NULL;
1814 ublk_put_device(ub);
0abe39de 1815 mutex_unlock(&ublk_ctl_mutex);
71f28f31
ML		 1816
1817 /*
1818 * Wait until the idr is removed, then it can be reused after
1819 * DEL_DEV command is returned.
0abe39de
ML		 1820 *
1821 * If we returns because of user interrupt, future delete command
1822 * may come:
1823 *
1824 * - the device number isn't freed, this device won't or needn't
1825 * be deleted again, since UB_STATE_DELETED is set, and device
1826 * will be released after the last reference is dropped
1827 *
1828 * - the device number is freed already, we will not find this
1829 * device via ublk_get_device_from_id()
71f28f31 1830 */
0abe39de 1831 wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
71f28f31 1832
0abe39de 1833 return 0;
71f28f31
ML		 1834}
1835
71f28f31
ML		 1836static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1837{
1838 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1839
1840 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1841 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1842 header->data[0], header->addr, header->len);
1843}
1844
bfbcef03 1845static int ublk_ctrl_stop_dev(struct ublk_device *ub)
71f28f31 1846{
34d8f2be
CH		 1847 ublk_stop_dev(ub);
1848 cancel_work_sync(&ub->stop_work);
bbae8d1f 1849 cancel_work_sync(&ub->quiesce_work);
71f28f31
ML		 1850
1851 return 0;
1852}
1853
bfbcef03
ML		 1854static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1855 struct io_uring_cmd *cmd)
71f28f31
ML		 1856{
1857 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1858 void __user *argp = (void __user *)(unsigned long)header->addr;
34d8f2be
CH		 1859
1860 if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1861 return -EINVAL;
1862
34d8f2be 1863 if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
bfbcef03 1864 return -EFAULT;
34d8f2be 1865
bfbcef03 1866 return 0;
34d8f2be
CH		 1867}
1868
abb864d3
ML		 1869/* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1870static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1871{
1872 ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1873 ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1874
1875 if (ub->ub_disk) {
1876 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1877 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1878 } else {
1879 ub->params.devt.disk_major = 0;
1880 ub->params.devt.disk_minor = 0;
1881 }
1882 ub->params.types |= UBLK_PARAM_TYPE_DEVT;
34d8f2be
CH		 1883}
1884
bfbcef03
ML		 1885static int ublk_ctrl_get_params(struct ublk_device *ub,
1886 struct io_uring_cmd *cmd)
0aa73170
ML		 1887{
1888 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1889 void __user *argp = (void __user *)(unsigned long)header->addr;
1890 struct ublk_params_header ph;
0aa73170
ML		 1891 int ret;
1892
1893 if (header->len <= sizeof(ph) || !header->addr)
1894 return -EINVAL;
1895
1896 if (copy_from_user(&ph, argp, sizeof(ph)))
1897 return -EFAULT;
1898
1899 if (ph.len > header->len || !ph.len)
1900 return -EINVAL;
1901
1902 if (ph.len > sizeof(struct ublk_params))
1903 ph.len = sizeof(struct ublk_params);
1904
0aa73170 1905 mutex_lock(&ub->mutex);
abb864d3 1906 ublk_ctrl_fill_params_devt(ub);
0aa73170
ML		 1907 if (copy_to_user(argp, &ub->params, ph.len))
1908 ret = -EFAULT;
1909 else
1910 ret = 0;
1911 mutex_unlock(&ub->mutex);
1912
0aa73170
ML		 1913 return ret;
1914}
1915
bfbcef03
ML		 1916static int ublk_ctrl_set_params(struct ublk_device *ub,
1917 struct io_uring_cmd *cmd)
0aa73170
ML		 1918{
1919 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1920 void __user *argp = (void __user *)(unsigned long)header->addr;
1921 struct ublk_params_header ph;
0aa73170
ML		 1922 int ret = -EFAULT;
1923
1924 if (header->len <= sizeof(ph) || !header->addr)
1925 return -EINVAL;
1926
1927 if (copy_from_user(&ph, argp, sizeof(ph)))
1928 return -EFAULT;
1929
1930 if (ph.len > header->len || !ph.len || !ph.types)
1931 return -EINVAL;
1932
1933 if (ph.len > sizeof(struct ublk_params))
1934 ph.len = sizeof(struct ublk_params);
1935
0aa73170
ML		 1936 /* parameters can only be changed when device isn't live */
1937 mutex_lock(&ub->mutex);
1938 if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1939 ret = -EACCES;
1940 } else if (copy_from_user(&ub->params, argp, ph.len)) {
1941 ret = -EFAULT;
1942 } else {
1943 /* clear all we don't support yet */
1944 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1945 ret = ublk_validate_params(ub);
1946 }
1947 mutex_unlock(&ub->mutex);
0aa73170
ML		 1948
1949 return ret;
1950}
1951
c732a852
Z		 1952static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1953{
1954 int i;
1955
1956 WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1957 /* All old ioucmds have to be completed */
1958 WARN_ON_ONCE(ubq->nr_io_ready);
1959 /* old daemon is PF_EXITING, put it now */
1960 put_task_struct(ubq->ubq_daemon);
1961 /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1962 ubq->ubq_daemon = NULL;
1963
1964 for (i = 0; i < ubq->q_depth; i++) {
1965 struct ublk_io *io = &ubq->ios[i];
1966
1967 /* forget everything now and be ready for new FETCH_REQ */
1968 io->flags = 0;
1969 io->cmd = NULL;
1970 io->addr = 0;
1971 }
1972}
1973
bfbcef03
ML		 1974static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1975 struct io_uring_cmd *cmd)
c732a852
Z		 1976{
1977 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
c732a852
Z		 1978 int ret = -EINVAL;
1979 int i;
1980
c732a852
Z		 1981 mutex_lock(&ub->mutex);
1982 if (!ublk_can_use_recovery(ub))
1983 goto out_unlock;
1984 /*
1985 * START_RECOVERY is only allowd after:
1986 *
1987 * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1988 * and related io_uring ctx is freed so file struct of /dev/ublkcX is
1989 * released.
1990 *
1991 * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1992 * (a)has quiesced request queue
1993 * (b)has requeued every inflight rqs whose io_flags is ACTIVE
1994 * (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1995 * (d)has completed/camceled all ioucmds owned by ther dying process
1996 */
1997 if (test_bit(UB_STATE_OPEN, &ub->state) ||
1998 ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1999 ret = -EBUSY;
2000 goto out_unlock;
2001 }
2002 pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2003 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2004 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2005 /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2006 ub->mm = NULL;
2007 ub->nr_queues_ready = 0;
73a166d9 2008 ub->nr_privileged_daemon = 0;
c732a852
Z		 2009 init_completion(&ub->completion);
2010 ret = 0;
2011 out_unlock:
2012 mutex_unlock(&ub->mutex);
c732a852
Z		 2013 return ret;
2014}
2015
bfbcef03
ML		 2016static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2017 struct io_uring_cmd *cmd)
c732a852
Z		 2018{
2019 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2020 int ublksrv_pid = (int)header->data[0];
c732a852
Z		 2021 int ret = -EINVAL;
2022
c732a852
Z		 2023 pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2024 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2025 /* wait until new ubq_daemon sending all FETCH_REQ */
2026 wait_for_completion_interruptible(&ub->completion);
2027 pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2028 __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2029
2030 mutex_lock(&ub->mutex);
2031 if (!ublk_can_use_recovery(ub))
2032 goto out_unlock;
2033
2034 if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2035 ret = -EBUSY;
2036 goto out_unlock;
2037 }
2038 ub->dev_info.ublksrv_pid = ublksrv_pid;
2039 pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2040 __func__, ublksrv_pid, header->dev_id);
2041 blk_mq_unquiesce_queue(ub->ub_disk->queue);
2042 pr_devel("%s: queue unquiesced, dev id %d.\n",
2043 __func__, header->dev_id);
2044 blk_mq_kick_requeue_list(ub->ub_disk->queue);
2045 ub->dev_info.state = UBLK_S_DEV_LIVE;
2046 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2047 ret = 0;
2048 out_unlock:
2049 mutex_unlock(&ub->mutex);
c732a852
Z		 2050 return ret;
2051}
2052
4093cb5a
ML		 2053/*
2054 * All control commands are sent via /dev/ublk-control, so we have to check
2055 * the destination device's permission
2056 */
2057static int ublk_char_dev_permission(struct ublk_device *ub,
2058 const char *dev_path, int mask)
2059{
2060 int err;
2061 struct path path;
2062 struct kstat stat;
2063
2064 err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2065 if (err)
2066 return err;
2067
2068 err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2069 if (err)
2070 goto exit;
2071
2072 err = -EPERM;
2073 if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2074 goto exit;
2075
5b0ed596 2076 err = inode_permission(&nop_mnt_idmap,
4093cb5a
ML		 2077 d_backing_inode(path.dentry), mask);
2078exit:
2079 path_put(&path);
2080 return err;
2081}
2082
2083static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2084 struct io_uring_cmd *cmd)
2085{
2086 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2087 bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2088 void __user *argp = (void __user *)(unsigned long)header->addr;
2089 char *dev_path = NULL;
2090 int ret = 0;
2091 int mask;
2092
2093 if (!unprivileged) {
2094 if (!capable(CAP_SYS_ADMIN))
2095 return -EPERM;
2096 /*
2097 * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2098 * char_dev_path in payload too, since userspace may not
2099 * know if the specified device is created as unprivileged
2100 * mode.
2101 */
2102 if (cmd->cmd_op != UBLK_CMD_GET_DEV_INFO2)
2103 return 0;
2104 }
2105
2106 /*
2107 * User has to provide the char device path for unprivileged ublk
2108 *
2109 * header->addr always points to the dev path buffer, and
2110 * header->dev_path_len records length of dev path buffer.
2111 */
2112 if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2113 return -EINVAL;
2114
2115 if (header->len < header->dev_path_len)
2116 return -EINVAL;
2117
2118 dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2119 if (!dev_path)
2120 return -ENOMEM;
2121
2122 ret = -EFAULT;
2123 if (copy_from_user(dev_path, argp, header->dev_path_len))
2124 goto exit;
2125 dev_path[header->dev_path_len] = 0;
2126
2127 ret = -EINVAL;
2128 switch (cmd->cmd_op) {
2129 case UBLK_CMD_GET_DEV_INFO:
2130 case UBLK_CMD_GET_DEV_INFO2:
2131 case UBLK_CMD_GET_QUEUE_AFFINITY:
2132 case UBLK_CMD_GET_PARAMS:
2133 mask = MAY_READ;
2134 break;
2135 case UBLK_CMD_START_DEV:
2136 case UBLK_CMD_STOP_DEV:
2137 case UBLK_CMD_ADD_DEV:
2138 case UBLK_CMD_DEL_DEV:
2139 case UBLK_CMD_SET_PARAMS:
2140 case UBLK_CMD_START_USER_RECOVERY:
2141 case UBLK_CMD_END_USER_RECOVERY:
2142 mask = MAY_READ | MAY_WRITE;
2143 break;
2144 default:
2145 goto exit;
2146 }
2147
2148 ret = ublk_char_dev_permission(ub, dev_path, mask);
2149 if (!ret) {
2150 header->len -= header->dev_path_len;
2151 header->addr += header->dev_path_len;
2152 }
2153 pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2154 __func__, ub->ub_number, cmd->cmd_op,
2155 ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2156 dev_path, ret);
2157exit:
2158 kfree(dev_path);
c732a852
Z		 2159 return ret;
2160}
2161
34d8f2be
CH		 2162static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2163 unsigned int issue_flags)
2164{
2165 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
bfbcef03 2166 struct ublk_device *ub = NULL;
71f28f31
ML		 2167 int ret = -EINVAL;
2168
fa8e442e
ML		 2169 if (issue_flags & IO_URING_F_NONBLOCK)
2170 return -EAGAIN;
2171
71f28f31
ML		 2172 ublk_ctrl_cmd_dump(cmd);
2173
2174 if (!(issue_flags & IO_URING_F_SQE128))
2175 goto out;
2176
bfbcef03
ML		 2177 if (cmd->cmd_op != UBLK_CMD_ADD_DEV) {
2178 ret = -ENODEV;
2179 ub = ublk_get_device_from_id(header->dev_id);
2180 if (!ub)
2181 goto out;
4093cb5a
ML		 2182
2183 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2184 } else {
2185 /* ADD_DEV permission check is done in command handler */
2186 ret = 0;
bfbcef03
ML		 2187 }
2188
4093cb5a
ML		 2189 if (ret)
2190 goto put_dev;
71f28f31 2191
34d8f2be 2192 switch (cmd->cmd_op) {
71f28f31 2193 case UBLK_CMD_START_DEV:
bfbcef03 2194 ret = ublk_ctrl_start_dev(ub, cmd);
71f28f31
ML		 2195 break;
2196 case UBLK_CMD_STOP_DEV:
bfbcef03 2197 ret = ublk_ctrl_stop_dev(ub);
71f28f31
ML		 2198 break;
2199 case UBLK_CMD_GET_DEV_INFO:
4093cb5a 2200 case UBLK_CMD_GET_DEV_INFO2:
bfbcef03 2201 ret = ublk_ctrl_get_dev_info(ub, cmd);
71f28f31
ML		 2202 break;
2203 case UBLK_CMD_ADD_DEV:
34d8f2be 2204 ret = ublk_ctrl_add_dev(cmd);
71f28f31
ML		 2205 break;
2206 case UBLK_CMD_DEL_DEV:
bfbcef03 2207 ret = ublk_ctrl_del_dev(&ub);
71f28f31
ML		 2208 break;
2209 case UBLK_CMD_GET_QUEUE_AFFINITY:
bfbcef03 2210 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
71f28f31 2211 break;
0aa73170 2212 case UBLK_CMD_GET_PARAMS:
bfbcef03 2213 ret = ublk_ctrl_get_params(ub, cmd);
0aa73170
ML		 2214 break;
2215 case UBLK_CMD_SET_PARAMS:
bfbcef03 2216 ret = ublk_ctrl_set_params(ub, cmd);
0aa73170 2217 break;
c732a852 2218 case UBLK_CMD_START_USER_RECOVERY:
bfbcef03 2219 ret = ublk_ctrl_start_recovery(ub, cmd);
c732a852
Z		 2220 break;
2221 case UBLK_CMD_END_USER_RECOVERY:
bfbcef03 2222 ret = ublk_ctrl_end_recovery(ub, cmd);
c732a852 2223 break;
71f28f31 2224 default:
bfbcef03 2225 ret = -ENOTSUPP;
71f28f31 2226 break;
6b1439d2 2227 }
4093cb5a
ML		 2228
2229 put_dev:
bfbcef03
ML		 2230 if (ub)
2231 ublk_put_device(ub);
71f28f31
ML		 2232 out:
2233 io_uring_cmd_done(cmd, ret, 0);
2234 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2235 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2236 return -EIOCBQUEUED;
2237}
2238
2239static const struct file_operations ublk_ctl_fops = {
2240 .open = nonseekable_open,
2241 .uring_cmd = ublk_ctrl_uring_cmd,
2242 .owner = THIS_MODULE,
2243 .llseek = noop_llseek,
2244};
2245
2246static struct miscdevice ublk_misc = {
2247 .minor = MISC_DYNAMIC_MINOR,
2248 .name = "ublk-control",
2249 .fops = &ublk_ctl_fops,
2250};
2251
2252static int __init ublk_init(void)
2253{
2254 int ret;
2255
2256 init_waitqueue_head(&ublk_idr_wq);
2257
2258 ret = misc_register(&ublk_misc);
2259 if (ret)
2260 return ret;
2261
2262 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2263 if (ret)
2264 goto unregister_mis;
2265
2266 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2267 if (IS_ERR(ublk_chr_class)) {
2268 ret = PTR_ERR(ublk_chr_class);
2269 goto free_chrdev_region;
2270 }
2271 return 0;
2272
2273free_chrdev_region:
2274 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2275unregister_mis:
2276 misc_deregister(&ublk_misc);
2277 return ret;
2278}
2279
2280static void __exit ublk_exit(void)
2281{
2282 struct ublk_device *ub;
2283 int id;
2284
71f28f31
ML		 2285 idr_for_each_entry(&ublk_index_idr, ub, id)
2286 ublk_remove(ub);
2287
8e4ff684
ML		 2288 class_destroy(ublk_chr_class);
2289 misc_deregister(&ublk_misc);
2290
71f28f31
ML		 2291 idr_destroy(&ublk_index_idr);
2292 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2293}
2294
2295module_init(ublk_init);
2296module_exit(ublk_exit);
2297
403ebc87
ML		 2298module_param(ublks_max, int, 0444);
2299MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2300
71f28f31
ML		 2301MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2302MODULE_LICENSE("GPL");
Linux 6.x block layer and io_uring tree(s)