projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ublk: remove the empty open and release block device operations
[linux-2.6-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>
71f28f31
ML		 45#include <uapi/linux/ublk_cmd.h>
46
47#define UBLK_MINORS (1U << MINORBITS)
48
0edb3696
ML		 49struct ublk_rq_data {
50 struct callback_head work;
51};
52
71f28f31
ML		 53struct ublk_uring_cmd_pdu {
54 struct request *req;
55};
56
57/*
58 * io command is active: sqe cmd is received, and its cqe isn't done
59 *
60 * If the flag is set, the io command is owned by ublk driver, and waited
61 * for incoming blk-mq request from the ublk block device.
62 *
63 * If the flag is cleared, the io command will be completed, and owned by
64 * ublk server.
65 */
66#define UBLK_IO_FLAG_ACTIVE 0x01
67
68/*
69 * IO command is completed via cqe, and it is being handled by ublksrv, and
70 * not committed yet
71 *
72 * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
73 * cross verification
74 */
75#define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
76
77/*
78 * IO command is aborted, so this flag is set in case of
79 * !UBLK_IO_FLAG_ACTIVE.
80 *
81 * After this flag is observed, any pending or new incoming request
82 * associated with this io command will be failed immediately
83 */
84#define UBLK_IO_FLAG_ABORTED 0x04
85
86struct ublk_io {
87 /* userspace buffer address from io cmd */
88 __u64 addr;
89 unsigned int flags;
90 int res;
91
92 struct io_uring_cmd *cmd;
93};
94
95struct ublk_queue {
96 int q_id;
97 int q_depth;
98
0edb3696 99 unsigned long flags;
71f28f31
ML		 100 struct task_struct *ubq_daemon;
101 char *io_cmd_buf;
102
103 unsigned long io_addr; /* mapped vm address */
104 unsigned int max_io_sz;
105 bool abort_work_pending;
106 unsigned short nr_io_ready; /* how many ios setup */
107 struct ublk_device *dev;
108 struct ublk_io ios[0];
109};
110
111#define UBLK_DAEMON_MONITOR_PERIOD (5 * HZ)
112
113struct ublk_device {
114 struct gendisk *ub_disk;
115 struct request_queue *ub_queue;
116
117 char *__queues;
118
119 unsigned short queue_size;
120 unsigned short bs_shift;
121 struct ublksrv_ctrl_dev_info dev_info;
122
123 struct blk_mq_tag_set tag_set;
124
125 struct cdev cdev;
126 struct device cdev_dev;
127
128 atomic_t ch_open_cnt;
129 int ub_number;
130
131 struct mutex mutex;
132
133 struct mm_struct *mm;
134
135 struct completion completion;
136 unsigned int nr_queues_ready;
137 atomic_t nr_aborted_queues;
138
139 /*
140 * Our ubq->daemon may be killed without any notification, so
141 * monitor each queue's daemon periodically
142 */
143 struct delayed_work monitor_work;
144 struct work_struct stop_work;
145};
146
147static dev_t ublk_chr_devt;
148static struct class *ublk_chr_class;
149
150static DEFINE_IDR(ublk_index_idr);
151static DEFINE_SPINLOCK(ublk_idr_lock);
152static wait_queue_head_t ublk_idr_wq; /* wait until one idr is freed */
153
154static DEFINE_MUTEX(ublk_ctl_mutex);
155
156static struct miscdevice ublk_misc;
157
cebbe577
ML		 158static struct lock_class_key ublk_bio_compl_lkclass;
159
0edb3696
ML		 160static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
161{
162 if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
163 !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
164 return true;
165 return false;
166}
167
71f28f31
ML		 168static struct ublk_device *ublk_get_device(struct ublk_device *ub)
169{
170 if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
171 return ub;
172 return NULL;
173}
174
175static void ublk_put_device(struct ublk_device *ub)
176{
177 put_device(&ub->cdev_dev);
178}
179
180static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
181 int qid)
182{
183 return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
184}
185
186static inline bool ublk_rq_has_data(const struct request *rq)
187{
188 return rq->bio && bio_has_data(rq->bio);
189}
190
191static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
192 int tag)
193{
194 return (struct ublksrv_io_desc *)
195 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
196}
197
198static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
199{
200 return ublk_get_queue(ub, q_id)->io_cmd_buf;
201}
202
203static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
204{
205 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
206
207 return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
208 PAGE_SIZE);
209}
210
71f28f31
ML		 211static const struct block_device_operations ub_fops = {
212 .owner = THIS_MODULE,
71f28f31
ML		 213};
214
215#define UBLK_MAX_PIN_PAGES 32
216
217struct ublk_map_data {
218 const struct ublk_queue *ubq;
219 const struct request *rq;
220 const struct ublk_io *io;
221 unsigned max_bytes;
222};
223
224struct ublk_io_iter {
225 struct page *pages[UBLK_MAX_PIN_PAGES];
226 unsigned pg_off; /* offset in the 1st page in pages */
227 int nr_pages; /* how many page pointers in pages */
228 struct bio *bio;
229 struct bvec_iter iter;
230};
231
232static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
233 unsigned max_bytes, bool to_vm)
234{
235 const unsigned total = min_t(unsigned, max_bytes,
236 PAGE_SIZE - data->pg_off +
237 ((data->nr_pages - 1) << PAGE_SHIFT));
238 unsigned done = 0;
239 unsigned pg_idx = 0;
240
241 while (done < total) {
242 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
243 const unsigned int bytes = min3(bv.bv_len, total - done,
244 (unsigned)(PAGE_SIZE - data->pg_off));
245 void *bv_buf = bvec_kmap_local(&bv);
246 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
247
248 if (to_vm)
249 memcpy(pg_buf + data->pg_off, bv_buf, bytes);
250 else
251 memcpy(bv_buf, pg_buf + data->pg_off, bytes);
252
253 kunmap_local(pg_buf);
254 kunmap_local(bv_buf);
255
256 /* advance page array */
257 data->pg_off += bytes;
258 if (data->pg_off == PAGE_SIZE) {
259 pg_idx += 1;
260 data->pg_off = 0;
261 }
262
263 done += bytes;
264
265 /* advance bio */
266 bio_advance_iter_single(data->bio, &data->iter, bytes);
267 if (!data->iter.bi_size) {
268 data->bio = data->bio->bi_next;
269 if (data->bio == NULL)
270 break;
271 data->iter = data->bio->bi_iter;
272 }
273 }
274
275 return done;
276}
277
278static inline int ublk_copy_user_pages(struct ublk_map_data *data,
279 bool to_vm)
280{
281 const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
282 const unsigned long start_vm = data->io->addr;
283 unsigned int done = 0;
284 struct ublk_io_iter iter = {
285 .pg_off = start_vm & (PAGE_SIZE - 1),
286 .bio = data->rq->bio,
287 .iter = data->rq->bio->bi_iter,
288 };
289 const unsigned int nr_pages = round_up(data->max_bytes +
290 (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
291
292 while (done < nr_pages) {
293 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
294 nr_pages - done);
295 unsigned i, len;
296
297 iter.nr_pages = get_user_pages_fast(start_vm +
298 (done << PAGE_SHIFT), to_pin, gup_flags,
299 iter.pages);
300 if (iter.nr_pages <= 0)
301 return done == 0 ? iter.nr_pages : done;
302 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
303 for (i = 0; i < iter.nr_pages; i++) {
304 if (to_vm)
305 set_page_dirty(iter.pages[i]);
306 put_page(iter.pages[i]);
307 }
308 data->max_bytes -= len;
309 done += iter.nr_pages;
310 }
311
312 return done;
313}
314
315static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
316 struct ublk_io *io)
317{
318 const unsigned int rq_bytes = blk_rq_bytes(req);
319 /*
320 * no zero copy, we delay copy WRITE request data into ublksrv
321 * context and the big benefit is that pinning pages in current
322 * context is pretty fast, see ublk_pin_user_pages
323 */
324 if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
325 return rq_bytes;
326
327 if (ublk_rq_has_data(req)) {
328 struct ublk_map_data data = {
329 .ubq = ubq,
330 .rq = req,
331 .io = io,
332 .max_bytes = rq_bytes,
333 };
334
335 ublk_copy_user_pages(&data, true);
336
337 return rq_bytes - data.max_bytes;
338 }
339 return rq_bytes;
340}
341
342static int ublk_unmap_io(const struct ublk_queue *ubq,
343 const struct request *req,
344 struct ublk_io *io)
345{
346 const unsigned int rq_bytes = blk_rq_bytes(req);
347
348 if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
349 struct ublk_map_data data = {
350 .ubq = ubq,
351 .rq = req,
352 .io = io,
353 .max_bytes = io->res,
354 };
355
356 WARN_ON_ONCE(io->res > rq_bytes);
357
358 ublk_copy_user_pages(&data, false);
359
360 return io->res - data.max_bytes;
361 }
362 return rq_bytes;
363}
364
365static inline unsigned int ublk_req_build_flags(struct request *req)
366{
367 unsigned flags = 0;
368
369 if (req->cmd_flags & REQ_FAILFAST_DEV)
370 flags |= UBLK_IO_F_FAILFAST_DEV;
371
372 if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
373 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
374
375 if (req->cmd_flags & REQ_FAILFAST_DRIVER)
376 flags |= UBLK_IO_F_FAILFAST_DRIVER;
377
378 if (req->cmd_flags & REQ_META)
379 flags |= UBLK_IO_F_META;
380
71f28f31
ML		 381 if (req->cmd_flags & REQ_FUA)
382 flags |= UBLK_IO_F_FUA;
383
71f28f31
ML		 384 if (req->cmd_flags & REQ_NOUNMAP)
385 flags |= UBLK_IO_F_NOUNMAP;
386
387 if (req->cmd_flags & REQ_SWAP)
388 flags |= UBLK_IO_F_SWAP;
389
390 return flags;
391}
392
f2450f8a 393static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
71f28f31
ML		 394{
395 struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
396 struct ublk_io *io = &ubq->ios[req->tag];
397 u32 ublk_op;
398
399 switch (req_op(req)) {
400 case REQ_OP_READ:
401 ublk_op = UBLK_IO_OP_READ;
402 break;
403 case REQ_OP_WRITE:
404 ublk_op = UBLK_IO_OP_WRITE;
405 break;
406 case REQ_OP_FLUSH:
407 ublk_op = UBLK_IO_OP_FLUSH;
408 break;
409 case REQ_OP_DISCARD:
410 ublk_op = UBLK_IO_OP_DISCARD;
411 break;
412 case REQ_OP_WRITE_ZEROES:
413 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
414 break;
415 default:
416 return BLK_STS_IOERR;
417 }
418
419 /* need to translate since kernel may change */
420 iod->op_flags = ublk_op | ublk_req_build_flags(req);
421 iod->nr_sectors = blk_rq_sectors(req);
422 iod->start_sector = blk_rq_pos(req);
423 iod->addr = io->addr;
424
425 return BLK_STS_OK;
426}
427
428static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
429 struct io_uring_cmd *ioucmd)
430{
431 return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
432}
433
434static bool ubq_daemon_is_dying(struct ublk_queue *ubq)
435{
436 return ubq->ubq_daemon->flags & PF_EXITING;
437}
438
439/* todo: handle partial completion */
440static void ublk_complete_rq(struct request *req)
441{
442 struct ublk_queue *ubq = req->mq_hctx->driver_data;
443 struct ublk_io *io = &ubq->ios[req->tag];
444 unsigned int unmapped_bytes;
445
446 /* failed read IO if nothing is read */
447 if (!io->res && req_op(req) == REQ_OP_READ)
448 io->res = -EIO;
449
450 if (io->res < 0) {
451 blk_mq_end_request(req, errno_to_blk_status(io->res));
452 return;
453 }
454
455 /*
456 * FLUSH or DISCARD usually won't return bytes returned, so end them
457 * directly.
458 *
459 * Both the two needn't unmap.
460 */
461 if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
462 blk_mq_end_request(req, BLK_STS_OK);
463 return;
464 }
465
466 /* for READ request, writing data in iod->addr to rq buffers */
467 unmapped_bytes = ublk_unmap_io(ubq, req, io);
468
469 /*
470 * Extremely impossible since we got data filled in just before
471 *
472 * Re-read simply for this unlikely case.
473 */
474 if (unlikely(unmapped_bytes < io->res))
475 io->res = unmapped_bytes;
476
477 if (blk_update_request(req, BLK_STS_OK, io->res))
478 blk_mq_requeue_request(req, true);
479 else
480 __blk_mq_end_request(req, BLK_STS_OK);
481}
482
483/*
484 * __ublk_fail_req() may be called from abort context or ->ubq_daemon
485 * context during exiting, so lock is required.
486 *
487 * Also aborting may not be started yet, keep in mind that one failed
488 * request may be issued by block layer again.
489 */
490static void __ublk_fail_req(struct ublk_io *io, struct request *req)
491{
492 WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
493
494 if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
495 io->flags |= UBLK_IO_FLAG_ABORTED;
496 blk_mq_end_request(req, BLK_STS_IOERR);
497 }
498}
499
500#define UBLK_REQUEUE_DELAY_MS 3
501
0edb3696 502static inline void __ublk_rq_task_work(struct request *req)
71f28f31 503{
71f28f31 504 struct ublk_queue *ubq = req->mq_hctx->driver_data;
0edb3696 505 struct ublk_device *ub = ubq->dev;
71f28f31
ML		 506 int tag = req->tag;
507 struct ublk_io *io = &ubq->ios[tag];
508 bool task_exiting = current != ubq->ubq_daemon ||
509 (current->flags & PF_EXITING);
510 unsigned int mapped_bytes;
511
512 pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
513 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
514 ublk_get_iod(ubq, req->tag)->addr);
515
516 if (unlikely(task_exiting)) {
517 blk_mq_end_request(req, BLK_STS_IOERR);
518 mod_delayed_work(system_wq, &ub->monitor_work, 0);
519 return;
520 }
521
522 mapped_bytes = ublk_map_io(ubq, req, io);
523
524 /* partially mapped, update io descriptor */
525 if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
526 /*
527 * Nothing mapped, retry until we succeed.
528 *
529 * We may never succeed in mapping any bytes here because
530 * of OOM. TODO: reserve one buffer with single page pinned
531 * for providing forward progress guarantee.
532 */
533 if (unlikely(!mapped_bytes)) {
534 blk_mq_requeue_request(req, false);
535 blk_mq_delay_kick_requeue_list(req->q,
536 UBLK_REQUEUE_DELAY_MS);
537 return;
538 }
539
540 ublk_get_iod(ubq, req->tag)->nr_sectors =
541 mapped_bytes >> 9;
542 }
543
544 /* mark this cmd owned by ublksrv */
545 io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
546
547 /*
548 * clear ACTIVE since we are done with this sqe/cmd slot
549 * We can only accept io cmd in case of being not active.
550 */
551 io->flags &= ~UBLK_IO_FLAG_ACTIVE;
552
553 /* tell ublksrv one io request is coming */
554 io_uring_cmd_done(io->cmd, UBLK_IO_RES_OK, 0);
555}
556
0edb3696
ML		 557static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
558{
559 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
560
561 __ublk_rq_task_work(pdu->req);
562}
563
564static void ublk_rq_task_work_fn(struct callback_head *work)
565{
566 struct ublk_rq_data *data = container_of(work,
567 struct ublk_rq_data, work);
568 struct request *req = blk_mq_rq_from_pdu(data);
569
570 __ublk_rq_task_work(req);
571}
572
71f28f31
ML		 573static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
574 const struct blk_mq_queue_data *bd)
575{
576 struct ublk_queue *ubq = hctx->driver_data;
577 struct request *rq = bd->rq;
71f28f31
ML		 578 blk_status_t res;
579
580 /* fill iod to slot in io cmd buffer */
581 res = ublk_setup_iod(ubq, rq);
582 if (unlikely(res != BLK_STS_OK))
583 return BLK_STS_IOERR;
584
585 blk_mq_start_request(bd->rq);
586
587 if (unlikely(ubq_daemon_is_dying(ubq))) {
0edb3696 588 fail:
71f28f31
ML		 589 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
590 return BLK_STS_IOERR;
591 }
592
0edb3696
ML		 593 if (ublk_can_use_task_work(ubq)) {
594 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
595 enum task_work_notify_mode notify_mode = bd->last ?
596 TWA_SIGNAL_NO_IPI : TWA_NONE;
597
598 if (task_work_add(ubq->ubq_daemon, &data->work, notify_mode))
599 goto fail;
600 } else {
601 struct io_uring_cmd *cmd = ubq->ios[rq->tag].cmd;
602 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
603
604 pdu->req = rq;
605 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
606 }
71f28f31
ML		 607
608 return BLK_STS_OK;
609}
610
0edb3696
ML		 611static void ublk_commit_rqs(struct blk_mq_hw_ctx *hctx)
612{
613 struct ublk_queue *ubq = hctx->driver_data;
614
615 if (ublk_can_use_task_work(ubq))
616 __set_notify_signal(ubq->ubq_daemon);
617}
71f28f31
ML		 618
619static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
620 unsigned int hctx_idx)
621{
cebbe577 622 struct ublk_device *ub = driver_data;
71f28f31
ML		 623 struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
624
625 hctx->driver_data = ubq;
626 return 0;
627}
628
0edb3696
ML		 629static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
630 unsigned int hctx_idx, unsigned int numa_node)
631{
632 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
633
634 init_task_work(&data->work, ublk_rq_task_work_fn);
635 return 0;
636}
637
71f28f31
ML		 638static const struct blk_mq_ops ublk_mq_ops = {
639 .queue_rq = ublk_queue_rq,
0edb3696 640 .commit_rqs = ublk_commit_rqs,
71f28f31 641 .init_hctx = ublk_init_hctx,
0edb3696 642 .init_request = ublk_init_rq,
71f28f31
ML		 643};
644
645static int ublk_ch_open(struct inode *inode, struct file *filp)
646{
647 struct ublk_device *ub = container_of(inode->i_cdev,
648 struct ublk_device, cdev);
649
650 if (atomic_cmpxchg(&ub->ch_open_cnt, 0, 1) == 0) {
651 filp->private_data = ub;
652 return 0;
653 }
654 return -EBUSY;
655}
656
657static int ublk_ch_release(struct inode *inode, struct file *filp)
658{
659 struct ublk_device *ub = filp->private_data;
660
661 while (atomic_cmpxchg(&ub->ch_open_cnt, 1, 0) != 1)
662 cpu_relax();
663
664 filp->private_data = NULL;
665 return 0;
666}
667
668/* map pre-allocated per-queue cmd buffer to ublksrv daemon */
669static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
670{
671 struct ublk_device *ub = filp->private_data;
672 size_t sz = vma->vm_end - vma->vm_start;
673 unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
674 unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
675 int q_id, ret = 0;
676
677 mutex_lock(&ub->mutex);
678 if (!ub->mm)
679 ub->mm = current->mm;
680 if (current->mm != ub->mm)
681 ret = -EINVAL;
682 mutex_unlock(&ub->mutex);
683
684 if (ret)
685 return ret;
686
687 if (vma->vm_flags & VM_WRITE)
688 return -EPERM;
689
690 end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
691 if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
692 return -EINVAL;
693
694 q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
695 pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
696 __func__, q_id, current->pid, vma->vm_start,
697 phys_off, (unsigned long)sz);
698
699 if (sz != ublk_queue_cmd_buf_size(ub, q_id))
700 return -EINVAL;
701
702 pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
703 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
704}
705
706static void ublk_commit_completion(struct ublk_device *ub,
707 struct ublksrv_io_cmd *ub_cmd)
708{
709 u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
710 struct ublk_queue *ubq = ublk_get_queue(ub, qid);
711 struct ublk_io *io = &ubq->ios[tag];
712 struct request *req;
713
714 /* now this cmd slot is owned by nbd driver */
715 io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
716 io->res = ub_cmd->result;
717
718 /* find the io request and complete */
719 req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
720
721 if (req && likely(!blk_should_fake_timeout(req->q)))
722 ublk_complete_rq(req);
723}
724
725/*
726 * When ->ubq_daemon is exiting, either new request is ended immediately,
727 * or any queued io command is drained, so it is safe to abort queue
728 * lockless
729 */
730static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
731{
732 int i;
733
734 if (!ublk_get_device(ub))
735 return;
736
737 for (i = 0; i < ubq->q_depth; i++) {
738 struct ublk_io *io = &ubq->ios[i];
739
740 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
741 struct request *rq;
742
743 /*
744 * Either we fail the request or ublk_rq_task_work_fn
745 * will do it
746 */
747 rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
748 if (rq)
749 __ublk_fail_req(io, rq);
750 }
751 }
752 ublk_put_device(ub);
753}
754
755static void ublk_daemon_monitor_work(struct work_struct *work)
756{
757 struct ublk_device *ub =
758 container_of(work, struct ublk_device, monitor_work.work);
759 int i;
760
761 for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
762 struct ublk_queue *ubq = ublk_get_queue(ub, i);
763
764 if (ubq_daemon_is_dying(ubq)) {
765 schedule_work(&ub->stop_work);
766
767 /* abort queue is for making forward progress */
768 ublk_abort_queue(ub, ubq);
769 }
770 }
771
772 /*
773 * We can't schedule monitor work after ublk_remove() is started.
774 *
775 * No need ub->mutex, monitor work are canceled after state is marked
776 * as DEAD, so DEAD state is observed reliably.
777 */
778 if (ub->dev_info.state != UBLK_S_DEV_DEAD)
779 schedule_delayed_work(&ub->monitor_work,
780 UBLK_DAEMON_MONITOR_PERIOD);
781}
782
783static void ublk_cancel_queue(struct ublk_queue *ubq)
784{
785 int i;
786
787 for (i = 0; i < ubq->q_depth; i++) {
788 struct ublk_io *io = &ubq->ios[i];
789
790 if (io->flags & UBLK_IO_FLAG_ACTIVE)
791 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
792 }
793}
794
795/* Cancel all pending commands, must be called after del_gendisk() returns */
796static void ublk_cancel_dev(struct ublk_device *ub)
797{
798 int i;
799
800 for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
801 ublk_cancel_queue(ublk_get_queue(ub, i));
802}
803
804static void ublk_stop_dev(struct ublk_device *ub)
805{
806 mutex_lock(&ub->mutex);
807 if (!disk_live(ub->ub_disk))
808 goto unlock;
809
810 del_gendisk(ub->ub_disk);
811 ub->dev_info.state = UBLK_S_DEV_DEAD;
812 ub->dev_info.ublksrv_pid = -1;
813 ublk_cancel_dev(ub);
814 unlock:
815 mutex_unlock(&ub->mutex);
816 cancel_delayed_work_sync(&ub->monitor_work);
817}
818
819static int ublk_ctrl_stop_dev(struct ublk_device *ub)
820{
821 ublk_stop_dev(ub);
822 cancel_work_sync(&ub->stop_work);
823 return 0;
824}
825
826static inline bool ublk_queue_ready(struct ublk_queue *ubq)
827{
828 return ubq->nr_io_ready == ubq->q_depth;
829}
830
831/* device can only be started after all IOs are ready */
832static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
833{
834 mutex_lock(&ub->mutex);
835 ubq->nr_io_ready++;
836 if (ublk_queue_ready(ubq)) {
837 ubq->ubq_daemon = current;
838 get_task_struct(ubq->ubq_daemon);
839 ub->nr_queues_ready++;
840 }
841 if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
842 complete_all(&ub->completion);
843 mutex_unlock(&ub->mutex);
844}
845
846static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
847{
848 struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
849 struct ublk_device *ub = cmd->file->private_data;
850 struct ublk_queue *ubq;
851 struct ublk_io *io;
852 u32 cmd_op = cmd->cmd_op;
853 unsigned tag = ub_cmd->tag;
854 int ret = -EINVAL;
855
856 pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
857 __func__, cmd->cmd_op, ub_cmd->q_id, tag,
858 ub_cmd->result);
859
860 if (!(issue_flags & IO_URING_F_SQE128))
861 goto out;
862
863 if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
864 goto out;
865
866 ubq = ublk_get_queue(ub, ub_cmd->q_id);
867 if (!ubq || ub_cmd->q_id != ubq->q_id)
868 goto out;
869
870 if (ubq->ubq_daemon && ubq->ubq_daemon != current)
871 goto out;
872
873 if (tag >= ubq->q_depth)
874 goto out;
875
876 io = &ubq->ios[tag];
877
878 /* there is pending io cmd, something must be wrong */
879 if (io->flags & UBLK_IO_FLAG_ACTIVE) {
880 ret = -EBUSY;
881 goto out;
882 }
883
884 switch (cmd_op) {
885 case UBLK_IO_FETCH_REQ:
886 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
887 if (ublk_queue_ready(ubq)) {
888 ret = -EBUSY;
889 goto out;
890 }
891 /*
892 * The io is being handled by server, so COMMIT_RQ is expected
893 * instead of FETCH_REQ
894 */
895 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
896 goto out;
897 /* FETCH_RQ has to provide IO buffer */
898 if (!ub_cmd->addr)
899 goto out;
900 io->cmd = cmd;
901 io->flags |= UBLK_IO_FLAG_ACTIVE;
902 io->addr = ub_cmd->addr;
903
904 ublk_mark_io_ready(ub, ubq);
905 break;
906 case UBLK_IO_COMMIT_AND_FETCH_REQ:
907 /* FETCH_RQ has to provide IO buffer */
908 if (!ub_cmd->addr)
909 goto out;
910 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
911 goto out;
912 io->addr = ub_cmd->addr;
913 io->flags |= UBLK_IO_FLAG_ACTIVE;
914 io->cmd = cmd;
915 ublk_commit_completion(ub, ub_cmd);
916 break;
917 default:
918 goto out;
919 }
920 return -EIOCBQUEUED;
921
922 out:
71f28f31
ML		 923 io_uring_cmd_done(cmd, ret, 0);
924 pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
925 __func__, cmd_op, tag, ret, io->flags);
926 return -EIOCBQUEUED;
927}
928
929static const struct file_operations ublk_ch_fops = {
930 .owner = THIS_MODULE,
931 .open = ublk_ch_open,
932 .release = ublk_ch_release,
933 .llseek = no_llseek,
934 .uring_cmd = ublk_ch_uring_cmd,
935 .mmap = ublk_ch_mmap,
936};
937
938static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
939{
940 int size = ublk_queue_cmd_buf_size(ub, q_id);
941 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
942
943 if (ubq->ubq_daemon)
944 put_task_struct(ubq->ubq_daemon);
945 if (ubq->io_cmd_buf)
946 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
947}
948
949static int ublk_init_queue(struct ublk_device *ub, int q_id)
950{
951 struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
952 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
953 void *ptr;
954 int size;
955
0edb3696 956 ubq->flags = ub->dev_info.flags[0];
71f28f31
ML		 957 ubq->q_id = q_id;
958 ubq->q_depth = ub->dev_info.queue_depth;
959 size = ublk_queue_cmd_buf_size(ub, q_id);
960
961 ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
962 if (!ptr)
963 return -ENOMEM;
964
965 ubq->io_cmd_buf = ptr;
966 ubq->dev = ub;
967 return 0;
968}
969
970static void ublk_deinit_queues(struct ublk_device *ub)
971{
972 int nr_queues = ub->dev_info.nr_hw_queues;
973 int i;
974
975 if (!ub->__queues)
976 return;
977
978 for (i = 0; i < nr_queues; i++)
979 ublk_deinit_queue(ub, i);
980 kfree(ub->__queues);
981}
982
983static int ublk_init_queues(struct ublk_device *ub)
984{
985 int nr_queues = ub->dev_info.nr_hw_queues;
986 int depth = ub->dev_info.queue_depth;
987 int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
988 int i, ret = -ENOMEM;
989
990 ub->queue_size = ubq_size;
991 ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
992 if (!ub->__queues)
993 return ret;
994
995 for (i = 0; i < nr_queues; i++) {
996 if (ublk_init_queue(ub, i))
997 goto fail;
998 }
999
1000 init_completion(&ub->completion);
1001 return 0;
1002
1003 fail:
1004 ublk_deinit_queues(ub);
1005 return ret;
1006}
1007
1008static int __ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1009{
1010 int i = idx;
1011 int err;
1012
1013 spin_lock(&ublk_idr_lock);
1014 /* allocate id, if @id >= 0, we're requesting that specific id */
1015 if (i >= 0) {
1016 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1017 if (err == -ENOSPC)
1018 err = -EEXIST;
1019 } else {
1020 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1021 }
1022 spin_unlock(&ublk_idr_lock);
1023
1024 if (err >= 0)
1025 ub->ub_number = err;
1026
1027 return err;
1028}
1029
1030static struct ublk_device *__ublk_create_dev(int idx)
1031{
1032 struct ublk_device *ub = NULL;
1033 int ret;
1034
1035 ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1036 if (!ub)
1037 return ERR_PTR(-ENOMEM);
1038
1039 ret = __ublk_alloc_dev_number(ub, idx);
1040 if (ret < 0) {
1041 kfree(ub);
1042 return ERR_PTR(ret);
1043 }
1044 return ub;
1045}
1046
1047static void __ublk_destroy_dev(struct ublk_device *ub)
1048{
1049 spin_lock(&ublk_idr_lock);
1050 idr_remove(&ublk_index_idr, ub->ub_number);
1051 wake_up_all(&ublk_idr_wq);
1052 spin_unlock(&ublk_idr_lock);
1053
1054 mutex_destroy(&ub->mutex);
1055
1056 kfree(ub);
1057}
1058
1059static void ublk_cdev_rel(struct device *dev)
1060{
1061 struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1062
cebbe577
ML		 1063 blk_mq_destroy_queue(ub->ub_queue);
1064
71f28f31
ML		 1065 put_disk(ub->ub_disk);
1066
1067 blk_mq_free_tag_set(&ub->tag_set);
1068
1069 ublk_deinit_queues(ub);
1070
1071 __ublk_destroy_dev(ub);
1072}
1073
1074static int ublk_add_chdev(struct ublk_device *ub)
1075{
1076 struct device *dev = &ub->cdev_dev;
1077 int minor = ub->ub_number;
1078 int ret;
1079
1080 dev->parent = ublk_misc.this_device;
1081 dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1082 dev->class = ublk_chr_class;
1083 dev->release = ublk_cdev_rel;
1084 device_initialize(dev);
1085
1086 ret = dev_set_name(dev, "ublkc%d", minor);
1087 if (ret)
1088 goto fail;
1089
1090 cdev_init(&ub->cdev, &ublk_ch_fops);
1091 ret = cdev_device_add(&ub->cdev, dev);
1092 if (ret)
1093 goto fail;
1094 return 0;
1095 fail:
1096 put_device(dev);
1097 return ret;
1098}
1099
1100static void ublk_stop_work_fn(struct work_struct *work)
1101{
1102 struct ublk_device *ub =
1103 container_of(work, struct ublk_device, stop_work);
1104
1105 ublk_stop_dev(ub);
1106}
1107
1108static void ublk_update_capacity(struct ublk_device *ub)
1109{
1110 unsigned int max_rq_bytes;
1111
1112 /* make max request buffer size aligned with PAGE_SIZE */
1113 max_rq_bytes = round_down(ub->dev_info.rq_max_blocks <<
1114 ub->bs_shift, PAGE_SIZE);
1115 ub->dev_info.rq_max_blocks = max_rq_bytes >> ub->bs_shift;
1116
1117 set_capacity(ub->ub_disk, ub->dev_info.dev_blocks << (ub->bs_shift - 9));
1118}
1119
1120/* add disk & cdev, cleanup everything in case of failure */
1121static int ublk_add_dev(struct ublk_device *ub)
1122{
1123 struct gendisk *disk;
1124 int err = -ENOMEM;
1125 int bsize;
1126
1127 /* We are not ready to support zero copy */
1128 ub->dev_info.flags[0] &= ~UBLK_F_SUPPORT_ZERO_COPY;
1129
1130 bsize = ub->dev_info.block_size;
1131 ub->bs_shift = ilog2(bsize);
1132
1133 ub->dev_info.nr_hw_queues = min_t(unsigned int,
1134 ub->dev_info.nr_hw_queues, nr_cpu_ids);
1135
1136 INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1137 INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1138
1139 if (ublk_init_queues(ub))
1140 goto out_destroy_dev;
1141
1142 ub->tag_set.ops = &ublk_mq_ops;
1143 ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1144 ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1145 ub->tag_set.numa_node = NUMA_NO_NODE;
0edb3696 1146 ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
71f28f31
ML		 1147 ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1148 ub->tag_set.driver_data = ub;
1149
1150 err = blk_mq_alloc_tag_set(&ub->tag_set);
1151 if (err)
1152 goto out_deinit_queues;
1153
cebbe577 1154 ub->ub_queue = blk_mq_init_queue(&ub->tag_set);
f50e5d67
YY		 1155 if (IS_ERR(ub->ub_queue)) {
1156 err = PTR_ERR(ub->ub_queue);
cebbe577 1157 goto out_cleanup_tags;
f50e5d67 1158 }
cebbe577
ML		 1159 ub->ub_queue->queuedata = ub;
1160
1161 disk = ub->ub_disk = blk_mq_alloc_disk_for_queue(ub->ub_queue,
1162 &ublk_bio_compl_lkclass);
fe3333f6
DC		 1163 if (!disk) {
1164 err = -ENOMEM;
cebbe577 1165 goto out_free_request_queue;
71f28f31 1166 }
71f28f31
ML		 1167
1168 blk_queue_logical_block_size(ub->ub_queue, bsize);
1169 blk_queue_physical_block_size(ub->ub_queue, bsize);
1170 blk_queue_io_min(ub->ub_queue, bsize);
1171
1172 blk_queue_max_hw_sectors(ub->ub_queue, ub->dev_info.rq_max_blocks <<
1173 (ub->bs_shift - 9));
1174
1175 ub->ub_queue->limits.discard_granularity = PAGE_SIZE;
1176
1177 blk_queue_max_discard_sectors(ub->ub_queue, UINT_MAX >> 9);
1178 blk_queue_max_write_zeroes_sectors(ub->ub_queue, UINT_MAX >> 9);
1179
1180 ublk_update_capacity(ub);
1181
1182 disk->fops = &ub_fops;
1183 disk->private_data = ub;
1184 disk->queue = ub->ub_queue;
1185 sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1186
1187 mutex_init(&ub->mutex);
1188
1189 /* add char dev so that ublksrv daemon can be setup */
1190 err = ublk_add_chdev(ub);
1191 if (err)
1192 return err;
1193
1194 /* don't expose disk now until we got start command from cdev */
1195
1196 return 0;
1197
cebbe577
ML		 1198out_free_request_queue:
1199 blk_mq_destroy_queue(ub->ub_queue);
71f28f31
ML		 1200out_cleanup_tags:
1201 blk_mq_free_tag_set(&ub->tag_set);
1202out_deinit_queues:
1203 ublk_deinit_queues(ub);
1204out_destroy_dev:
1205 __ublk_destroy_dev(ub);
1206 return err;
1207}
1208
1209static void ublk_remove(struct ublk_device *ub)
1210{
1211 ublk_ctrl_stop_dev(ub);
1212
1213 cdev_device_del(&ub->cdev, &ub->cdev_dev);
1214 put_device(&ub->cdev_dev);
1215}
1216
1217static struct ublk_device *ublk_get_device_from_id(int idx)
1218{
1219 struct ublk_device *ub = NULL;
1220
1221 if (idx < 0)
1222 return NULL;
1223
1224 spin_lock(&ublk_idr_lock);
1225 ub = idr_find(&ublk_index_idr, idx);
1226 if (ub)
1227 ub = ublk_get_device(ub);
1228 spin_unlock(&ublk_idr_lock);
1229
1230 return ub;
1231}
1232
1233static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1234{
1235 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1236 int ret = -EINVAL;
1237 int ublksrv_pid = (int)header->data[0];
1238 unsigned long dev_blocks = header->data[1];
1239
1240 if (ublksrv_pid <= 0)
1241 return ret;
1242
1243 wait_for_completion_interruptible(&ub->completion);
1244
1245 schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1246
1247 mutex_lock(&ub->mutex);
1248 if (!disk_live(ub->ub_disk)) {
1249 /* We may get disk size updated */
1250 if (dev_blocks) {
1251 ub->dev_info.dev_blocks = dev_blocks;
1252 ublk_update_capacity(ub);
1253 }
1254 ub->dev_info.ublksrv_pid = ublksrv_pid;
1255 ret = add_disk(ub->ub_disk);
1256 if (!ret)
1257 ub->dev_info.state = UBLK_S_DEV_LIVE;
1258 } else {
1259 ret = -EEXIST;
1260 }
1261 mutex_unlock(&ub->mutex);
1262
1263 return ret;
1264}
1265
1266static struct blk_mq_hw_ctx *ublk_get_hw_queue(struct ublk_device *ub,
1267 unsigned int index)
1268{
1269 struct blk_mq_hw_ctx *hctx;
1270 unsigned long i;
1271
1272 queue_for_each_hw_ctx(ub->ub_queue, hctx, i)
1273 if (hctx->queue_num == index)
1274 return hctx;
1275 return NULL;
1276}
1277
1278static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1279{
1280 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1281 void __user *argp = (void __user *)(unsigned long)header->addr;
1282 struct blk_mq_hw_ctx *hctx;
1283 struct ublk_device *ub;
1284 unsigned long queue;
1285 unsigned int retlen;
f2450f8a 1286 int ret = -EINVAL;
71f28f31
ML		 1287
1288 ub = ublk_get_device_from_id(header->dev_id);
1289 if (!ub)
1290 goto out;
1291
71f28f31
ML		 1292 queue = header->data[0];
1293 if (queue >= ub->dev_info.nr_hw_queues)
1294 goto out;
1295 hctx = ublk_get_hw_queue(ub, queue);
1296 if (!hctx)
1297 goto out;
1298
1299 retlen = min_t(unsigned short, header->len, cpumask_size());
1300 if (copy_to_user(argp, hctx->cpumask, retlen)) {
1301 ret = -EFAULT;
1302 goto out;
1303 }
1304 if (retlen != header->len) {
1305 if (clear_user(argp + retlen, header->len - retlen)) {
1306 ret = -EFAULT;
1307 goto out;
1308 }
1309 }
1310 ret = 0;
1311 out:
1312 if (ub)
1313 ublk_put_device(ub);
1314 return ret;
1315}
1316
1317static int ublk_ctrl_add_dev(const struct ublksrv_ctrl_dev_info *info,
1318 void __user *argp, int idx)
1319{
1320 struct ublk_device *ub;
1321 int ret;
1322
1323 ret = mutex_lock_killable(&ublk_ctl_mutex);
1324 if (ret)
1325 return ret;
1326
1327 ub = __ublk_create_dev(idx);
1328 if (!IS_ERR_OR_NULL(ub)) {
1329 memcpy(&ub->dev_info, info, sizeof(*info));
1330
1331 /* update device id */
1332 ub->dev_info.dev_id = ub->ub_number;
1333
1334 ret = ublk_add_dev(ub);
1335 if (!ret) {
1336 if (copy_to_user(argp, &ub->dev_info, sizeof(*info))) {
1337 ublk_remove(ub);
1338 ret = -EFAULT;
1339 }
1340 }
1341 } else {
1342 if (IS_ERR(ub))
1343 ret = PTR_ERR(ub);
1344 else
1345 ret = -ENOMEM;
1346 }
1347 mutex_unlock(&ublk_ctl_mutex);
1348
1349 return ret;
1350}
1351
1352static inline bool ublk_idr_freed(int id)
1353{
1354 void *ptr;
1355
1356 spin_lock(&ublk_idr_lock);
1357 ptr = idr_find(&ublk_index_idr, id);
1358 spin_unlock(&ublk_idr_lock);
1359
1360 return ptr == NULL;
1361}
1362
1363static int ublk_ctrl_del_dev(int idx)
1364{
1365 struct ublk_device *ub;
1366 int ret;
1367
1368 ret = mutex_lock_killable(&ublk_ctl_mutex);
1369 if (ret)
1370 return ret;
1371
1372 ub = ublk_get_device_from_id(idx);
1373 if (ub) {
1374 ublk_remove(ub);
1375 ublk_put_device(ub);
1376 ret = 0;
1377 } else {
1378 ret = -ENODEV;
1379 }
1380
1381 /*
1382 * Wait until the idr is removed, then it can be reused after
1383 * DEL_DEV command is returned.
1384 */
1385 if (!ret)
1386 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1387 mutex_unlock(&ublk_ctl_mutex);
1388
1389 return ret;
1390}
1391
1392
1393static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1394{
1395 pr_devel("%s: dev id %d flags %llx\n", __func__,
1396 info->dev_id, info->flags[0]);
1397 pr_devel("\t nr_hw_queues %d queue_depth %d block size %d dev_capacity %lld\n",
1398 info->nr_hw_queues, info->queue_depth,
1399 info->block_size, info->dev_blocks);
1400}
1401
1402static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1403{
1404 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1405
1406 pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1407 __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1408 header->data[0], header->addr, header->len);
1409}
1410
1411static int ublk_ctrl_cmd_validate(struct io_uring_cmd *cmd,
1412 struct ublksrv_ctrl_dev_info *info)
1413{
1414 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1415 u32 cmd_op = cmd->cmd_op;
1416 void __user *argp = (void __user *)(unsigned long)header->addr;
1417
1418 if (!capable(CAP_SYS_ADMIN))
1419 return -EPERM;
1420
1421 switch (cmd_op) {
1422 case UBLK_CMD_GET_DEV_INFO:
1423 if (header->len < sizeof(*info) || !header->addr)
1424 return -EINVAL;
1425 break;
1426 case UBLK_CMD_ADD_DEV:
1427 if (header->len < sizeof(*info) || !header->addr)
1428 return -EINVAL;
1429 if (copy_from_user(info, argp, sizeof(*info)) != 0)
1430 return -EFAULT;
1431 ublk_dump_dev_info(info);
1432 if (header->dev_id != info->dev_id) {
1433 printk(KERN_WARNING "%s: cmd %x, dev id not match %u %u\n",
1434 __func__, cmd_op, header->dev_id,
1435 info->dev_id);
1436 return -EINVAL;
1437 }
1438 if (header->queue_id != (u16)-1) {
1439 printk(KERN_WARNING "%s: cmd %x queue_id is wrong %x\n",
1440 __func__, cmd_op, header->queue_id);
1441 return -EINVAL;
1442 }
1443 break;
1444 case UBLK_CMD_GET_QUEUE_AFFINITY:
1445 if ((header->len * BITS_PER_BYTE) < nr_cpu_ids)
1446 return -EINVAL;
1447 if (header->len & (sizeof(unsigned long)-1))
1448 return -EINVAL;
1449 if (!header->addr)
1450 return -EINVAL;
6b1439d2 1451 }
71f28f31
ML		 1452
1453 return 0;
1454}
1455
1456static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1457 unsigned int issue_flags)
1458{
1459 struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1460 void __user *argp = (void __user *)(unsigned long)header->addr;
1461 struct ublksrv_ctrl_dev_info info;
1462 u32 cmd_op = cmd->cmd_op;
1463 struct ublk_device *ub;
1464 int ret = -EINVAL;
1465
1466 ublk_ctrl_cmd_dump(cmd);
1467
1468 if (!(issue_flags & IO_URING_F_SQE128))
1469 goto out;
1470
1471 ret = ublk_ctrl_cmd_validate(cmd, &info);
1472 if (ret)
1473 goto out;
1474
1475 ret = -ENODEV;
1476 switch (cmd_op) {
1477 case UBLK_CMD_START_DEV:
1478 ub = ublk_get_device_from_id(header->dev_id);
1479 if (ub) {
1480 ret = ublk_ctrl_start_dev(ub, cmd);
1481 ublk_put_device(ub);
1482 }
1483 break;
1484 case UBLK_CMD_STOP_DEV:
1485 ub = ublk_get_device_from_id(header->dev_id);
1486 if (ub) {
1487 ret = ublk_ctrl_stop_dev(ub);
1488 ublk_put_device(ub);
1489 }
1490 break;
1491 case UBLK_CMD_GET_DEV_INFO:
1492 ub = ublk_get_device_from_id(header->dev_id);
1493 if (ub) {
1494 if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1495 ret = -EFAULT;
1496 else
1497 ret = 0;
1498 ublk_put_device(ub);
1499 }
1500 break;
1501 case UBLK_CMD_ADD_DEV:
1502 ret = ublk_ctrl_add_dev(&info, argp, header->dev_id);
1503 break;
1504 case UBLK_CMD_DEL_DEV:
1505 ret = ublk_ctrl_del_dev(header->dev_id);
1506 break;
1507 case UBLK_CMD_GET_QUEUE_AFFINITY:
1508 ret = ublk_ctrl_get_queue_affinity(cmd);
1509 break;
1510 default:
1511 break;
6b1439d2 1512 }
71f28f31
ML		 1513 out:
1514 io_uring_cmd_done(cmd, ret, 0);
1515 pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
1516 __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
1517 return -EIOCBQUEUED;
1518}
1519
1520static const struct file_operations ublk_ctl_fops = {
1521 .open = nonseekable_open,
1522 .uring_cmd = ublk_ctrl_uring_cmd,
1523 .owner = THIS_MODULE,
1524 .llseek = noop_llseek,
1525};
1526
1527static struct miscdevice ublk_misc = {
1528 .minor = MISC_DYNAMIC_MINOR,
1529 .name = "ublk-control",
1530 .fops = &ublk_ctl_fops,
1531};
1532
1533static int __init ublk_init(void)
1534{
1535 int ret;
1536
1537 init_waitqueue_head(&ublk_idr_wq);
1538
1539 ret = misc_register(&ublk_misc);
1540 if (ret)
1541 return ret;
1542
1543 ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
1544 if (ret)
1545 goto unregister_mis;
1546
1547 ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
1548 if (IS_ERR(ublk_chr_class)) {
1549 ret = PTR_ERR(ublk_chr_class);
1550 goto free_chrdev_region;
1551 }
1552 return 0;
1553
1554free_chrdev_region:
1555 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
1556unregister_mis:
1557 misc_deregister(&ublk_misc);
1558 return ret;
1559}
1560
1561static void __exit ublk_exit(void)
1562{
1563 struct ublk_device *ub;
1564 int id;
1565
1566 class_destroy(ublk_chr_class);
1567
1568 misc_deregister(&ublk_misc);
1569
1570 idr_for_each_entry(&ublk_index_idr, ub, id)
1571 ublk_remove(ub);
1572
1573 idr_destroy(&ublk_index_idr);
1574 unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
1575}
1576
1577module_init(ublk_init);
1578module_exit(ublk_exit);
1579
1580MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
1581MODULE_LICENSE("GPL");
Linux 6.x block layer and io_uring tree(s)