drivers/block/ublk_drv.c

   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>
  44 #include <linux/task_work.h>
  45 #include <linux/namei.h>
  46 #include <linux/kref.h>
  47 #include <uapi/linux/ublk_cmd.h>
  48
  49 #define UBLK_MINORS             (1U << MINORBITS)
  50
  51 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  52 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  53                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  54                 | UBLK_F_NEED_GET_DATA \
  55                 | UBLK_F_USER_RECOVERY \
  56                 | UBLK_F_USER_RECOVERY_REISSUE \
  57                 | UBLK_F_UNPRIVILEGED_DEV \
  58                 | UBLK_F_CMD_IOCTL_ENCODE \
  59                 | UBLK_F_USER_COPY)
  60
  61 /* All UBLK_PARAM_TYPE_* should be included here */
  62 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | \
  63                 UBLK_PARAM_TYPE_DISCARD | UBLK_PARAM_TYPE_DEVT)
  64
  65 struct ublk_rq_data {
  66         struct llist_node node;
  67
  68         struct kref ref;
  69 };
  70
  71 struct ublk_uring_cmd_pdu {
  72         struct ublk_queue *ubq;
  73 };
  74
  75 /*
  76  * io command is active: sqe cmd is received, and its cqe isn't done
  77  *
  78  * If the flag is set, the io command is owned by ublk driver, and waited
  79  * for incoming blk-mq request from the ublk block device.
  80  *
  81  * If the flag is cleared, the io command will be completed, and owned by
  82  * ublk server.
  83  */
  84 #define UBLK_IO_FLAG_ACTIVE     0x01
  85
  86 /*
  87  * IO command is completed via cqe, and it is being handled by ublksrv, and
  88  * not committed yet
  89  *
  90  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  91  * cross verification
  92  */
  93 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  94
  95 /*
  96  * IO command is aborted, so this flag is set in case of
  97  * !UBLK_IO_FLAG_ACTIVE.
  98  *
  99  * After this flag is observed, any pending or new incoming request
 100  * associated with this io command will be failed immediately
 101  */
 102 #define UBLK_IO_FLAG_ABORTED 0x04
 103
 104 /*
 105  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
 106  * get data buffer address from ublksrv.
 107  *
 108  * Then, bio data could be copied into this data buffer for a WRITE request
 109  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 110  */
 111 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 112
 113 struct ublk_io {
 114         /* userspace buffer address from io cmd */
 115         __u64   addr;
 116         unsigned int flags;
 117         int res;
 118
 119         struct io_uring_cmd *cmd;
 120 };
 121
 122 struct ublk_queue {
 123         int q_id;
 124         int q_depth;
 125
 126         unsigned long flags;
 127         struct task_struct      *ubq_daemon;
 128         char *io_cmd_buf;
 129
 130         struct llist_head       io_cmds;
 131
 132         unsigned long io_addr;  /* mapped vm address */
 133         unsigned int max_io_sz;
 134         bool force_abort;
 135         bool timeout;
 136         unsigned short nr_io_ready;     /* how many ios setup */
 137         struct ublk_device *dev;
 138         struct ublk_io ios[];
 139 };
 140
 141 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 142
 143 struct ublk_device {
 144         struct gendisk          *ub_disk;
 145
 146         char    *__queues;
 147
 148         unsigned int    queue_size;
 149         struct ublksrv_ctrl_dev_info    dev_info;
 150
 151         struct blk_mq_tag_set   tag_set;
 152
 153         struct cdev             cdev;
 154         struct device           cdev_dev;
 155
 156 #define UB_STATE_OPEN           0
 157 #define UB_STATE_USED           1
 158 #define UB_STATE_DELETED        2
 159         unsigned long           state;
 160         int                     ub_number;
 161
 162         struct mutex            mutex;
 163
 164         spinlock_t              mm_lock;
 165         struct mm_struct        *mm;
 166
 167         struct ublk_params      params;
 168
 169         struct completion       completion;
 170         unsigned int            nr_queues_ready;
 171         unsigned int            nr_privileged_daemon;
 172
 173         /*
 174          * Our ubq->daemon may be killed without any notification, so
 175          * monitor each queue's daemon periodically
 176          */
 177         struct delayed_work     monitor_work;
 178         struct work_struct      quiesce_work;
 179         struct work_struct      stop_work;
 180 };
 181
 182 /* header of ublk_params */
 183 struct ublk_params_header {
 184         __u32   len;
 185         __u32   types;
 186 };
 187
 188 static inline void __ublk_complete_rq(struct request *req);
 189 static void ublk_complete_rq(struct kref *ref);
 190
 191 static dev_t ublk_chr_devt;
 192 static struct class *ublk_chr_class;
 193
 194 static DEFINE_IDR(ublk_index_idr);
 195 static DEFINE_SPINLOCK(ublk_idr_lock);
 196 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 197
 198 static DEFINE_MUTEX(ublk_ctl_mutex);
 199
 200 /*
 201  * Max ublk devices allowed to add
 202  *
 203  * It can be extended to one per-user limit in future or even controlled
 204  * by cgroup.
 205  */
 206 static unsigned int ublks_max = 64;
 207 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 208
 209 static struct miscdevice ublk_misc;
 210
 211 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 212 {
 213         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 214                 UBLK_QID_BITS_MASK;
 215 }
 216
 217 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 218 {
 219         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 220 }
 221
 222 static inline unsigned ublk_pos_to_tag(loff_t pos)
 223 {
 224         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 225                 UBLK_TAG_BITS_MASK;
 226 }
 227
 228 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 229 {
 230         struct request_queue *q = ub->ub_disk->queue;
 231         const struct ublk_param_basic *p = &ub->params.basic;
 232
 233         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 234         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 235         blk_queue_io_min(q, 1 << p->io_min_shift);
 236         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 237
 238         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 239                         p->attrs & UBLK_ATTR_FUA);
 240         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 241                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 242         else
 243                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 244
 245         blk_queue_max_hw_sectors(q, p->max_sectors);
 246         blk_queue_chunk_sectors(q, p->chunk_sectors);
 247         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 248
 249         if (p->attrs & UBLK_ATTR_READ_ONLY)
 250                 set_disk_ro(ub->ub_disk, true);
 251
 252         set_capacity(ub->ub_disk, p->dev_sectors);
 253 }
 254
 255 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 256 {
 257         struct request_queue *q = ub->ub_disk->queue;
 258         const struct ublk_param_discard *p = &ub->params.discard;
 259
 260         q->limits.discard_alignment = p->discard_alignment;
 261         q->limits.discard_granularity = p->discard_granularity;
 262         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 263         blk_queue_max_write_zeroes_sectors(q,
 264                         p->max_write_zeroes_sectors);
 265         blk_queue_max_discard_segments(q, p->max_discard_segments);
 266 }
 267
 268 static int ublk_validate_params(const struct ublk_device *ub)
 269 {
 270         /* basic param is the only one which must be set */
 271         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 272                 const struct ublk_param_basic *p = &ub->params.basic;
 273
 274                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 275                         return -EINVAL;
 276
 277                 if (p->logical_bs_shift > p->physical_bs_shift)
 278                         return -EINVAL;
 279
 280                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 281                         return -EINVAL;
 282         } else
 283                 return -EINVAL;
 284
 285         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 286                 const struct ublk_param_discard *p = &ub->params.discard;
 287
 288                 /* So far, only support single segment discard */
 289                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 290                         return -EINVAL;
 291
 292                 if (!p->discard_granularity)
 293                         return -EINVAL;
 294         }
 295
 296         /* dev_t is read-only */
 297         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 298                 return -EINVAL;
 299
 300         return 0;
 301 }
 302
 303 static int ublk_apply_params(struct ublk_device *ub)
 304 {
 305         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 306                 return -EINVAL;
 307
 308         ublk_dev_param_basic_apply(ub);
 309
 310         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 311                 ublk_dev_param_discard_apply(ub);
 312
 313         return 0;
 314 }
 315
 316 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 317 {
 318         return ubq->flags & UBLK_F_USER_COPY;
 319 }
 320
 321 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 322 {
 323         /*
 324          * read()/write() is involved in user copy, so request reference
 325          * has to be grabbed
 326          */
 327         return ublk_support_user_copy(ubq);
 328 }
 329
 330 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 331                 struct request *req)
 332 {
 333         if (ublk_need_req_ref(ubq)) {
 334                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 335
 336                 kref_init(&data->ref);
 337         }
 338 }
 339
 340 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 341                 struct request *req)
 342 {
 343         if (ublk_need_req_ref(ubq)) {
 344                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 345
 346                 return kref_get_unless_zero(&data->ref);
 347         }
 348
 349         return true;
 350 }
 351
 352 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 353                 struct request *req)
 354 {
 355         if (ublk_need_req_ref(ubq)) {
 356                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 357
 358                 kref_put(&data->ref, ublk_complete_rq);
 359         } else {
 360                 __ublk_complete_rq(req);
 361         }
 362 }
 363
 364 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 365 {
 366         return ubq->flags & UBLK_F_NEED_GET_DATA;
 367 }
 368
 369 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 370 {
 371         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 372                 return ub;
 373         return NULL;
 374 }
 375
 376 static void ublk_put_device(struct ublk_device *ub)
 377 {
 378         put_device(&ub->cdev_dev);
 379 }
 380
 381 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 382                 int qid)
 383 {
 384        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 385 }
 386
 387 static inline bool ublk_rq_has_data(const struct request *rq)
 388 {
 389         return bio_has_data(rq->bio);
 390 }
 391
 392 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 393                 int tag)
 394 {
 395         return (struct ublksrv_io_desc *)
 396                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 397 }
 398
 399 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 400 {
 401         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 402 }
 403
 404 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 405 {
 406         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 407
 408         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 409                         PAGE_SIZE);
 410 }
 411
 412 static inline bool ublk_queue_can_use_recovery_reissue(
 413                 struct ublk_queue *ubq)
 414 {
 415         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 416                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 417 }
 418
 419 static inline bool ublk_queue_can_use_recovery(
 420                 struct ublk_queue *ubq)
 421 {
 422         return ubq->flags & UBLK_F_USER_RECOVERY;
 423 }
 424
 425 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 426 {
 427         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 428 }
 429
 430 static void ublk_free_disk(struct gendisk *disk)
 431 {
 432         struct ublk_device *ub = disk->private_data;
 433
 434         clear_bit(UB_STATE_USED, &ub->state);
 435         put_device(&ub->cdev_dev);
 436 }
 437
 438 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 439                 unsigned int *owner_gid)
 440 {
 441         kuid_t uid;
 442         kgid_t gid;
 443
 444         current_uid_gid(&uid, &gid);
 445
 446         *owner_uid = from_kuid(&init_user_ns, uid);
 447         *owner_gid = from_kgid(&init_user_ns, gid);
 448 }
 449
 450 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 451 {
 452         struct ublk_device *ub = disk->private_data;
 453
 454         if (capable(CAP_SYS_ADMIN))
 455                 return 0;
 456
 457         /*
 458          * If it is one unprivileged device, only owner can open
 459          * the disk. Otherwise it could be one trap made by one
 460          * evil user who grants this disk's privileges to other
 461          * users deliberately.
 462          *
 463          * This way is reasonable too given anyone can create
 464          * unprivileged device, and no need other's grant.
 465          */
 466         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 467                 unsigned int curr_uid, curr_gid;
 468
 469                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 470
 471                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 472                                 ub->dev_info.owner_gid)
 473                         return -EPERM;
 474         }
 475
 476         return 0;
 477 }
 478
 479 static const struct block_device_operations ub_fops = {
 480         .owner =        THIS_MODULE,
 481         .open =         ublk_open,
 482         .free_disk =    ublk_free_disk,
 483 };
 484
 485 #define UBLK_MAX_PIN_PAGES      32
 486
 487 struct ublk_io_iter {
 488         struct page *pages[UBLK_MAX_PIN_PAGES];
 489         struct bio *bio;
 490         struct bvec_iter iter;
 491 };
 492
 493 /* return how many pages are copied */
 494 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 495                 size_t total, size_t pg_off, int dir)
 496 {
 497         unsigned done = 0;
 498         unsigned pg_idx = 0;
 499
 500         while (done < total) {
 501                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 502                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 503                                 (unsigned)(PAGE_SIZE - pg_off));
 504                 void *bv_buf = bvec_kmap_local(&bv);
 505                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 506
 507                 if (dir == ITER_DEST)
 508                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 509                 else
 510                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 511
 512                 kunmap_local(pg_buf);
 513                 kunmap_local(bv_buf);
 514
 515                 /* advance page array */
 516                 pg_off += bytes;
 517                 if (pg_off == PAGE_SIZE) {
 518                         pg_idx += 1;
 519                         pg_off = 0;
 520                 }
 521
 522                 done += bytes;
 523
 524                 /* advance bio */
 525                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 526                 if (!data->iter.bi_size) {
 527                         data->bio = data->bio->bi_next;
 528                         if (data->bio == NULL)
 529                                 break;
 530                         data->iter = data->bio->bi_iter;
 531                 }
 532         }
 533 }
 534
 535 static bool ublk_advance_io_iter(const struct request *req,
 536                 struct ublk_io_iter *iter, unsigned int offset)
 537 {
 538         struct bio *bio = req->bio;
 539
 540         for_each_bio(bio) {
 541                 if (bio->bi_iter.bi_size > offset) {
 542                         iter->bio = bio;
 543                         iter->iter = bio->bi_iter;
 544                         bio_advance_iter(iter->bio, &iter->iter, offset);
 545                         return true;
 546                 }
 547                 offset -= bio->bi_iter.bi_size;
 548         }
 549         return false;
 550 }
 551
 552 /*
 553  * Copy data between request pages and io_iter, and 'offset'
 554  * is the start point of linear offset of request.
 555  */
 556 static size_t ublk_copy_user_pages(const struct request *req,
 557                 unsigned offset, struct iov_iter *uiter, int dir)
 558 {
 559         struct ublk_io_iter iter;
 560         size_t done = 0;
 561
 562         if (!ublk_advance_io_iter(req, &iter, offset))
 563                 return 0;
 564
 565         while (iov_iter_count(uiter) && iter.bio) {
 566                 unsigned nr_pages;
 567                 ssize_t len;
 568                 size_t off;
 569                 int i;
 570
 571                 len = iov_iter_get_pages2(uiter, iter.pages,
 572                                 iov_iter_count(uiter),
 573                                 UBLK_MAX_PIN_PAGES, &off);
 574                 if (len <= 0)
 575                         return done;
 576
 577                 ublk_copy_io_pages(&iter, len, off, dir);
 578                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 579                 for (i = 0; i < nr_pages; i++) {
 580                         if (dir == ITER_DEST)
 581                                 set_page_dirty(iter.pages[i]);
 582                         put_page(iter.pages[i]);
 583                 }
 584                 done += len;
 585         }
 586
 587         return done;
 588 }
 589
 590 static inline bool ublk_need_map_req(const struct request *req)
 591 {
 592         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 593 }
 594
 595 static inline bool ublk_need_unmap_req(const struct request *req)
 596 {
 597         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
 598 }
 599
 600 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 601                 struct ublk_io *io)
 602 {
 603         const unsigned int rq_bytes = blk_rq_bytes(req);
 604
 605         if (ublk_support_user_copy(ubq))
 606                 return rq_bytes;
 607
 608         /*
 609          * no zero copy, we delay copy WRITE request data into ublksrv
 610          * context and the big benefit is that pinning pages in current
 611          * context is pretty fast, see ublk_pin_user_pages
 612          */
 613         if (ublk_need_map_req(req)) {
 614                 struct iov_iter iter;
 615                 struct iovec iov;
 616                 const int dir = ITER_DEST;
 617
 618                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 619                                 &iov, &iter);
 620
 621                 return ublk_copy_user_pages(req, 0, &iter, dir);
 622         }
 623         return rq_bytes;
 624 }
 625
 626 static int ublk_unmap_io(const struct ublk_queue *ubq,
 627                 const struct request *req,
 628                 struct ublk_io *io)
 629 {
 630         const unsigned int rq_bytes = blk_rq_bytes(req);
 631
 632         if (ublk_support_user_copy(ubq))
 633                 return rq_bytes;
 634
 635         if (ublk_need_unmap_req(req)) {
 636                 struct iov_iter iter;
 637                 struct iovec iov;
 638                 const int dir = ITER_SOURCE;
 639
 640                 WARN_ON_ONCE(io->res > rq_bytes);
 641
 642                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 643                                 &iov, &iter);
 644                 return ublk_copy_user_pages(req, 0, &iter, dir);
 645         }
 646         return rq_bytes;
 647 }
 648
 649 static inline unsigned int ublk_req_build_flags(struct request *req)
 650 {
 651         unsigned flags = 0;
 652
 653         if (req->cmd_flags & REQ_FAILFAST_DEV)
 654                 flags |= UBLK_IO_F_FAILFAST_DEV;
 655
 656         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 657                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 658
 659         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 660                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 661
 662         if (req->cmd_flags & REQ_META)
 663                 flags |= UBLK_IO_F_META;
 664
 665         if (req->cmd_flags & REQ_FUA)
 666                 flags |= UBLK_IO_F_FUA;
 667
 668         if (req->cmd_flags & REQ_NOUNMAP)
 669                 flags |= UBLK_IO_F_NOUNMAP;
 670
 671         if (req->cmd_flags & REQ_SWAP)
 672                 flags |= UBLK_IO_F_SWAP;
 673
 674         return flags;
 675 }
 676
 677 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 678 {
 679         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 680         struct ublk_io *io = &ubq->ios[req->tag];
 681         u32 ublk_op;
 682
 683         switch (req_op(req)) {
 684         case REQ_OP_READ:
 685                 ublk_op = UBLK_IO_OP_READ;
 686                 break;
 687         case REQ_OP_WRITE:
 688                 ublk_op = UBLK_IO_OP_WRITE;
 689                 break;
 690         case REQ_OP_FLUSH:
 691                 ublk_op = UBLK_IO_OP_FLUSH;
 692                 break;
 693         case REQ_OP_DISCARD:
 694                 ublk_op = UBLK_IO_OP_DISCARD;
 695                 break;
 696         case REQ_OP_WRITE_ZEROES:
 697                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 698                 break;
 699         default:
 700                 return BLK_STS_IOERR;
 701         }
 702
 703         /* need to translate since kernel may change */
 704         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 705         iod->nr_sectors = blk_rq_sectors(req);
 706         iod->start_sector = blk_rq_pos(req);
 707         iod->addr = io->addr;
 708
 709         return BLK_STS_OK;
 710 }
 711
 712 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 713                 struct io_uring_cmd *ioucmd)
 714 {
 715         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 716 }
 717
 718 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 719 {
 720         return ubq->ubq_daemon->flags & PF_EXITING;
 721 }
 722
 723 /* todo: handle partial completion */
 724 static inline void __ublk_complete_rq(struct request *req)
 725 {
 726         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 727         struct ublk_io *io = &ubq->ios[req->tag];
 728         unsigned int unmapped_bytes;
 729         blk_status_t res = BLK_STS_OK;
 730
 731         /* called from ublk_abort_queue() code path */
 732         if (io->flags & UBLK_IO_FLAG_ABORTED) {
 733                 res = BLK_STS_IOERR;
 734                 goto exit;
 735         }
 736
 737         /* failed read IO if nothing is read */
 738         if (!io->res && req_op(req) == REQ_OP_READ)
 739                 io->res = -EIO;
 740
 741         if (io->res < 0) {
 742                 res = errno_to_blk_status(io->res);
 743                 goto exit;
 744         }
 745
 746         /*
 747          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
 748          * directly.
 749          *
 750          * Both the two needn't unmap.
 751          */
 752         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
 753                 goto exit;
 754
 755         /* for READ request, writing data in iod->addr to rq buffers */
 756         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 757
 758         /*
 759          * Extremely impossible since we got data filled in just before
 760          *
 761          * Re-read simply for this unlikely case.
 762          */
 763         if (unlikely(unmapped_bytes < io->res))
 764                 io->res = unmapped_bytes;
 765
 766         if (blk_update_request(req, BLK_STS_OK, io->res))
 767                 blk_mq_requeue_request(req, true);
 768         else
 769                 __blk_mq_end_request(req, BLK_STS_OK);
 770
 771         return;
 772 exit:
 773         blk_mq_end_request(req, res);
 774 }
 775
 776 static void ublk_complete_rq(struct kref *ref)
 777 {
 778         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
 779                         ref);
 780         struct request *req = blk_mq_rq_from_pdu(data);
 781
 782         __ublk_complete_rq(req);
 783 }
 784
 785 /*
 786  * Since __ublk_rq_task_work always fails requests immediately during
 787  * exiting, __ublk_fail_req() is only called from abort context during
 788  * exiting. So lock is unnecessary.
 789  *
 790  * Also aborting may not be started yet, keep in mind that one failed
 791  * request may be issued by block layer again.
 792  */
 793 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
 794                 struct request *req)
 795 {
 796         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 797
 798         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 799                 io->flags |= UBLK_IO_FLAG_ABORTED;
 800                 if (ublk_queue_can_use_recovery_reissue(ubq))
 801                         blk_mq_requeue_request(req, false);
 802                 else
 803                         ublk_put_req_ref(ubq, req);
 804         }
 805 }
 806
 807 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
 808                                 unsigned issue_flags)
 809 {
 810         /* mark this cmd owned by ublksrv */
 811         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 812
 813         /*
 814          * clear ACTIVE since we are done with this sqe/cmd slot
 815          * We can only accept io cmd in case of being not active.
 816          */
 817         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 818
 819         /* tell ublksrv one io request is coming */
 820         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
 821 }
 822
 823 #define UBLK_REQUEUE_DELAY_MS   3
 824
 825 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 826                 struct request *rq)
 827 {
 828         /* We cannot process this rq so just requeue it. */
 829         if (ublk_queue_can_use_recovery(ubq))
 830                 blk_mq_requeue_request(rq, false);
 831         else
 832                 blk_mq_end_request(rq, BLK_STS_IOERR);
 833
 834         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 835 }
 836
 837 static inline void __ublk_rq_task_work(struct request *req,
 838                                        unsigned issue_flags)
 839 {
 840         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 841         int tag = req->tag;
 842         struct ublk_io *io = &ubq->ios[tag];
 843         unsigned int mapped_bytes;
 844
 845         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 846                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 847                         ublk_get_iod(ubq, req->tag)->addr);
 848
 849         /*
 850          * Task is exiting if either:
 851          *
 852          * (1) current != ubq_daemon.
 853          * io_uring_cmd_complete_in_task() tries to run task_work
 854          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 855          *
 856          * (2) current->flags & PF_EXITING.
 857          */
 858         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 859                 __ublk_abort_rq(ubq, req);
 860                 return;
 861         }
 862
 863         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
 864                 /*
 865                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 866                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 867                  * and notify it.
 868                  */
 869                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 870                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 871                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 872                                         __func__, io->cmd->cmd_op, ubq->q_id,
 873                                         req->tag, io->flags);
 874                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
 875                         return;
 876                 }
 877                 /*
 878                  * We have handled UBLK_IO_NEED_GET_DATA command,
 879                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 880                  * do the copy work.
 881                  */
 882                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 883                 /* update iod->addr because ublksrv may have passed a new io buffer */
 884                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 885                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 886                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 887                                 ublk_get_iod(ubq, req->tag)->addr);
 888         }
 889
 890         mapped_bytes = ublk_map_io(ubq, req, io);
 891
 892         /* partially mapped, update io descriptor */
 893         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 894                 /*
 895                  * Nothing mapped, retry until we succeed.
 896                  *
 897                  * We may never succeed in mapping any bytes here because
 898                  * of OOM. TODO: reserve one buffer with single page pinned
 899                  * for providing forward progress guarantee.
 900                  */
 901                 if (unlikely(!mapped_bytes)) {
 902                         blk_mq_requeue_request(req, false);
 903                         blk_mq_delay_kick_requeue_list(req->q,
 904                                         UBLK_REQUEUE_DELAY_MS);
 905                         return;
 906                 }
 907
 908                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 909                         mapped_bytes >> 9;
 910         }
 911
 912         ublk_init_req_ref(ubq, req);
 913         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
 914 }
 915
 916 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
 917                                         unsigned issue_flags)
 918 {
 919         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 920         struct ublk_rq_data *data, *tmp;
 921
 922         io_cmds = llist_reverse_order(io_cmds);
 923         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 924                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
 925 }
 926
 927 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 928 {
 929         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 930         struct ublk_rq_data *data, *tmp;
 931
 932         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 933                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 934 }
 935
 936 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
 937 {
 938         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 939         struct ublk_queue *ubq = pdu->ubq;
 940
 941         ublk_forward_io_cmds(ubq, issue_flags);
 942 }
 943
 944 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 945 {
 946         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 947         struct ublk_io *io;
 948
 949         if (!llist_add(&data->node, &ubq->io_cmds))
 950                 return;
 951
 952         io = &ubq->ios[rq->tag];
 953         /*
 954          * If the check pass, we know that this is a re-issued request aborted
 955          * previously in monitor_work because the ubq_daemon(cmd's task) is
 956          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 957          * because this ioucmd's io_uring context may be freed now if no inflight
 958          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 959          *
 960          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 961          * the tag). Then the request is re-started(allocating the tag) and we are here.
 962          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 963          * guarantees that here is a re-issued request aborted previously.
 964          */
 965         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 966                 ublk_abort_io_cmds(ubq);
 967         } else {
 968                 struct io_uring_cmd *cmd = io->cmd;
 969                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 970
 971                 pdu->ubq = ubq;
 972                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 973         }
 974 }
 975
 976 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
 977 {
 978         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
 979
 980         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
 981                 if (!ubq->timeout) {
 982                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
 983                         ubq->timeout = true;
 984                 }
 985
 986                 return BLK_EH_DONE;
 987         }
 988
 989         return BLK_EH_RESET_TIMER;
 990 }
 991
 992 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 993                 const struct blk_mq_queue_data *bd)
 994 {
 995         struct ublk_queue *ubq = hctx->driver_data;
 996         struct request *rq = bd->rq;
 997         blk_status_t res;
 998
 999         /* fill iod to slot in io cmd buffer */
1000         res = ublk_setup_iod(ubq, rq);
1001         if (unlikely(res != BLK_STS_OK))
1002                 return BLK_STS_IOERR;
1003
1004         /* With recovery feature enabled, force_abort is set in
1005          * ublk_stop_dev() before calling del_gendisk(). We have to
1006          * abort all requeued and new rqs here to let del_gendisk()
1007          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1008          * to avoid UAF on io_uring ctx.
1009          *
1010          * Note: force_abort is guaranteed to be seen because it is set
1011          * before request queue is unqiuesced.
1012          */
1013         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1014                 return BLK_STS_IOERR;
1015
1016         blk_mq_start_request(bd->rq);
1017
1018         if (unlikely(ubq_daemon_is_dying(ubq))) {
1019                 __ublk_abort_rq(ubq, rq);
1020                 return BLK_STS_OK;
1021         }
1022
1023         ublk_queue_cmd(ubq, rq);
1024
1025         return BLK_STS_OK;
1026 }
1027
1028 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1029                 unsigned int hctx_idx)
1030 {
1031         struct ublk_device *ub = driver_data;
1032         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1033
1034         hctx->driver_data = ubq;
1035         return 0;
1036 }
1037
1038 static const struct blk_mq_ops ublk_mq_ops = {
1039         .queue_rq       = ublk_queue_rq,
1040         .init_hctx      = ublk_init_hctx,
1041         .timeout        = ublk_timeout,
1042 };
1043
1044 static int ublk_ch_open(struct inode *inode, struct file *filp)
1045 {
1046         struct ublk_device *ub = container_of(inode->i_cdev,
1047                         struct ublk_device, cdev);
1048
1049         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1050                 return -EBUSY;
1051         filp->private_data = ub;
1052         return 0;
1053 }
1054
1055 static int ublk_ch_release(struct inode *inode, struct file *filp)
1056 {
1057         struct ublk_device *ub = filp->private_data;
1058
1059         clear_bit(UB_STATE_OPEN, &ub->state);
1060         return 0;
1061 }
1062
1063 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1064 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1065 {
1066         struct ublk_device *ub = filp->private_data;
1067         size_t sz = vma->vm_end - vma->vm_start;
1068         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1069         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1070         int q_id, ret = 0;
1071
1072         spin_lock(&ub->mm_lock);
1073         if (!ub->mm)
1074                 ub->mm = current->mm;
1075         if (current->mm != ub->mm)
1076                 ret = -EINVAL;
1077         spin_unlock(&ub->mm_lock);
1078
1079         if (ret)
1080                 return ret;
1081
1082         if (vma->vm_flags & VM_WRITE)
1083                 return -EPERM;
1084
1085         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1086         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1087                 return -EINVAL;
1088
1089         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1090         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1091                         __func__, q_id, current->pid, vma->vm_start,
1092                         phys_off, (unsigned long)sz);
1093
1094         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1095                 return -EINVAL;
1096
1097         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1098         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1099 }
1100
1101 static void ublk_commit_completion(struct ublk_device *ub,
1102                 const struct ublksrv_io_cmd *ub_cmd)
1103 {
1104         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1105         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1106         struct ublk_io *io = &ubq->ios[tag];
1107         struct request *req;
1108
1109         /* now this cmd slot is owned by nbd driver */
1110         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1111         io->res = ub_cmd->result;
1112
1113         /* find the io request and complete */
1114         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1115
1116         if (req && likely(!blk_should_fake_timeout(req->q)))
1117                 ublk_put_req_ref(ubq, req);
1118 }
1119
1120 /*
1121  * When ->ubq_daemon is exiting, either new request is ended immediately,
1122  * or any queued io command is drained, so it is safe to abort queue
1123  * lockless
1124  */
1125 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1126 {
1127         int i;
1128
1129         if (!ublk_get_device(ub))
1130                 return;
1131
1132         for (i = 0; i < ubq->q_depth; i++) {
1133                 struct ublk_io *io = &ubq->ios[i];
1134
1135                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1136                         struct request *rq;
1137
1138                         /*
1139                          * Either we fail the request or ublk_rq_task_work_fn
1140                          * will do it
1141                          */
1142                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1143                         if (rq)
1144                                 __ublk_fail_req(ubq, io, rq);
1145                 }
1146         }
1147         ublk_put_device(ub);
1148 }
1149
1150 static void ublk_daemon_monitor_work(struct work_struct *work)
1151 {
1152         struct ublk_device *ub =
1153                 container_of(work, struct ublk_device, monitor_work.work);
1154         int i;
1155
1156         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1157                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1158
1159                 if (ubq_daemon_is_dying(ubq)) {
1160                         if (ublk_queue_can_use_recovery(ubq))
1161                                 schedule_work(&ub->quiesce_work);
1162                         else
1163                                 schedule_work(&ub->stop_work);
1164
1165                         /* abort queue is for making forward progress */
1166                         ublk_abort_queue(ub, ubq);
1167                 }
1168         }
1169
1170         /*
1171          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1172          * after ublk_remove() or __ublk_quiesce_dev() is started.
1173          *
1174          * No need ub->mutex, monitor work are canceled after state is marked
1175          * as not LIVE, so new state is observed reliably.
1176          */
1177         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1178                 schedule_delayed_work(&ub->monitor_work,
1179                                 UBLK_DAEMON_MONITOR_PERIOD);
1180 }
1181
1182 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1183 {
1184         return ubq->nr_io_ready == ubq->q_depth;
1185 }
1186
1187 static void ublk_cancel_queue(struct ublk_queue *ubq)
1188 {
1189         int i;
1190
1191         if (!ublk_queue_ready(ubq))
1192                 return;
1193
1194         for (i = 0; i < ubq->q_depth; i++) {
1195                 struct ublk_io *io = &ubq->ios[i];
1196
1197                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1198                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1199                                                 IO_URING_F_UNLOCKED);
1200         }
1201
1202         /* all io commands are canceled */
1203         ubq->nr_io_ready = 0;
1204 }
1205
1206 /* Cancel all pending commands, must be called after del_gendisk() returns */
1207 static void ublk_cancel_dev(struct ublk_device *ub)
1208 {
1209         int i;
1210
1211         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1212                 ublk_cancel_queue(ublk_get_queue(ub, i));
1213 }
1214
1215 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1216 {
1217         bool *idle = data;
1218
1219         if (blk_mq_request_started(rq)) {
1220                 *idle = false;
1221                 return false;
1222         }
1223         return true;
1224 }
1225
1226 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1227 {
1228         bool idle;
1229
1230         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1231         while (true) {
1232                 idle = true;
1233                 blk_mq_tagset_busy_iter(&ub->tag_set,
1234                                 ublk_check_inflight_rq, &idle);
1235                 if (idle)
1236                         break;
1237                 msleep(UBLK_REQUEUE_DELAY_MS);
1238         }
1239 }
1240
1241 static void __ublk_quiesce_dev(struct ublk_device *ub)
1242 {
1243         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1244                         __func__, ub->dev_info.dev_id,
1245                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1246                         "LIVE" : "QUIESCED");
1247         blk_mq_quiesce_queue(ub->ub_disk->queue);
1248         ublk_wait_tagset_rqs_idle(ub);
1249         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1250         ublk_cancel_dev(ub);
1251         /* we are going to release task_struct of ubq_daemon and resets
1252          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1253          * Besides, monitor_work is not necessary in QUIESCED state since we have
1254          * already scheduled quiesce_work and quiesced all ubqs.
1255          *
1256          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1257          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1258          */
1259         cancel_delayed_work_sync(&ub->monitor_work);
1260 }
1261
1262 static void ublk_quiesce_work_fn(struct work_struct *work)
1263 {
1264         struct ublk_device *ub =
1265                 container_of(work, struct ublk_device, quiesce_work);
1266
1267         mutex_lock(&ub->mutex);
1268         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1269                 goto unlock;
1270         __ublk_quiesce_dev(ub);
1271  unlock:
1272         mutex_unlock(&ub->mutex);
1273 }
1274
1275 static void ublk_unquiesce_dev(struct ublk_device *ub)
1276 {
1277         int i;
1278
1279         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1280                         __func__, ub->dev_info.dev_id,
1281                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1282                         "LIVE" : "QUIESCED");
1283         /* quiesce_work has run. We let requeued rqs be aborted
1284          * before running fallback_wq. "force_abort" must be seen
1285          * after request queue is unqiuesced. Then del_gendisk()
1286          * can move on.
1287          */
1288         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1289                 ublk_get_queue(ub, i)->force_abort = true;
1290
1291         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1292         /* We may have requeued some rqs in ublk_quiesce_queue() */
1293         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1294 }
1295
1296 static void ublk_stop_dev(struct ublk_device *ub)
1297 {
1298         mutex_lock(&ub->mutex);
1299         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1300                 goto unlock;
1301         if (ublk_can_use_recovery(ub)) {
1302                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1303                         __ublk_quiesce_dev(ub);
1304                 ublk_unquiesce_dev(ub);
1305         }
1306         del_gendisk(ub->ub_disk);
1307         ub->dev_info.state = UBLK_S_DEV_DEAD;
1308         ub->dev_info.ublksrv_pid = -1;
1309         put_disk(ub->ub_disk);
1310         ub->ub_disk = NULL;
1311  unlock:
1312         ublk_cancel_dev(ub);
1313         mutex_unlock(&ub->mutex);
1314         cancel_delayed_work_sync(&ub->monitor_work);
1315 }
1316
1317 /* device can only be started after all IOs are ready */
1318 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1319 {
1320         mutex_lock(&ub->mutex);
1321         ubq->nr_io_ready++;
1322         if (ublk_queue_ready(ubq)) {
1323                 ubq->ubq_daemon = current;
1324                 get_task_struct(ubq->ubq_daemon);
1325                 ub->nr_queues_ready++;
1326
1327                 if (capable(CAP_SYS_ADMIN))
1328                         ub->nr_privileged_daemon++;
1329         }
1330         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1331                 complete_all(&ub->completion);
1332         mutex_unlock(&ub->mutex);
1333 }
1334
1335 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1336                 int tag)
1337 {
1338         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1339         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1340
1341         ublk_queue_cmd(ubq, req);
1342 }
1343
1344 static inline int ublk_check_cmd_op(u32 cmd_op)
1345 {
1346         u32 ioc_type = _IOC_TYPE(cmd_op);
1347
1348         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1349                 return -EOPNOTSUPP;
1350
1351         if (ioc_type != 'u' && ioc_type != 0)
1352                 return -EOPNOTSUPP;
1353
1354         return 0;
1355 }
1356
1357 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1358                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1359 {
1360         io->cmd = cmd;
1361         io->flags |= UBLK_IO_FLAG_ACTIVE;
1362         io->addr = buf_addr;
1363 }
1364
1365 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1366                                unsigned int issue_flags,
1367                                const struct ublksrv_io_cmd *ub_cmd)
1368 {
1369         struct ublk_device *ub = cmd->file->private_data;
1370         struct ublk_queue *ubq;
1371         struct ublk_io *io;
1372         u32 cmd_op = cmd->cmd_op;
1373         unsigned tag = ub_cmd->tag;
1374         int ret = -EINVAL;
1375         struct request *req;
1376
1377         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1378                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1379                         ub_cmd->result);
1380
1381         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1382                 goto out;
1383
1384         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1385         if (!ubq || ub_cmd->q_id != ubq->q_id)
1386                 goto out;
1387
1388         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1389                 goto out;
1390
1391         if (tag >= ubq->q_depth)
1392                 goto out;
1393
1394         io = &ubq->ios[tag];
1395
1396         /* there is pending io cmd, something must be wrong */
1397         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1398                 ret = -EBUSY;
1399                 goto out;
1400         }
1401
1402         /*
1403          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1404          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1405          */
1406         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1407                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1408                 goto out;
1409
1410         if (ublk_support_user_copy(ubq) && ub_cmd->addr) {
1411                 ret = -EINVAL;
1412                 goto out;
1413         }
1414
1415         ret = ublk_check_cmd_op(cmd_op);
1416         if (ret)
1417                 goto out;
1418
1419         ret = -EINVAL;
1420         switch (_IOC_NR(cmd_op)) {
1421         case UBLK_IO_FETCH_REQ:
1422                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1423                 if (ublk_queue_ready(ubq)) {
1424                         ret = -EBUSY;
1425                         goto out;
1426                 }
1427                 /*
1428                  * The io is being handled by server, so COMMIT_RQ is expected
1429                  * instead of FETCH_REQ
1430                  */
1431                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1432                         goto out;
1433
1434                 if (!ublk_support_user_copy(ubq)) {
1435                         /*
1436                          * FETCH_RQ has to provide IO buffer if NEED GET
1437                          * DATA is not enabled
1438                          */
1439                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1440                                 goto out;
1441                 }
1442
1443                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1444                 ublk_mark_io_ready(ub, ubq);
1445                 break;
1446         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1447                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1448
1449                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1450                         goto out;
1451
1452                 if (!ublk_support_user_copy(ubq)) {
1453                         /*
1454                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1455                          * NEED GET DATA is not enabled or it is Read IO.
1456                          */
1457                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1458                                                 req_op(req) == REQ_OP_READ))
1459                                 goto out;
1460                 }
1461                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1462                 ublk_commit_completion(ub, ub_cmd);
1463                 break;
1464         case UBLK_IO_NEED_GET_DATA:
1465                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1466                         goto out;
1467                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1468                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1469                 break;
1470         default:
1471                 goto out;
1472         }
1473         return -EIOCBQUEUED;
1474
1475  out:
1476         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1477         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1478                         __func__, cmd_op, tag, ret, io->flags);
1479         return -EIOCBQUEUED;
1480 }
1481
1482 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1483                 struct ublk_queue *ubq, int tag, size_t offset)
1484 {
1485         struct request *req;
1486
1487         if (!ublk_need_req_ref(ubq))
1488                 return NULL;
1489
1490         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1491         if (!req)
1492                 return NULL;
1493
1494         if (!ublk_get_req_ref(ubq, req))
1495                 return NULL;
1496
1497         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1498                 goto fail_put;
1499
1500         if (!ublk_rq_has_data(req))
1501                 goto fail_put;
1502
1503         if (offset > blk_rq_bytes(req))
1504                 goto fail_put;
1505
1506         return req;
1507 fail_put:
1508         ublk_put_req_ref(ubq, req);
1509         return NULL;
1510 }
1511
1512 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1513 {
1514         /*
1515          * Not necessary for async retry, but let's keep it simple and always
1516          * copy the values to avoid any potential reuse.
1517          */
1518         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1519         const struct ublksrv_io_cmd ub_cmd = {
1520                 .q_id = READ_ONCE(ub_src->q_id),
1521                 .tag = READ_ONCE(ub_src->tag),
1522                 .result = READ_ONCE(ub_src->result),
1523                 .addr = READ_ONCE(ub_src->addr)
1524         };
1525
1526         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1527 }
1528
1529 static inline bool ublk_check_ubuf_dir(const struct request *req,
1530                 int ubuf_dir)
1531 {
1532         /* copy ubuf to request pages */
1533         if (req_op(req) == REQ_OP_READ && ubuf_dir == ITER_SOURCE)
1534                 return true;
1535
1536         /* copy request pages to ubuf */
1537         if (req_op(req) == REQ_OP_WRITE && ubuf_dir == ITER_DEST)
1538                 return true;
1539
1540         return false;
1541 }
1542
1543 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1544                 struct iov_iter *iter, size_t *off, int dir)
1545 {
1546         struct ublk_device *ub = iocb->ki_filp->private_data;
1547         struct ublk_queue *ubq;
1548         struct request *req;
1549         size_t buf_off;
1550         u16 tag, q_id;
1551
1552         if (!ub)
1553                 return ERR_PTR(-EACCES);
1554
1555         if (!user_backed_iter(iter))
1556                 return ERR_PTR(-EACCES);
1557
1558         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1559                 return ERR_PTR(-EACCES);
1560
1561         tag = ublk_pos_to_tag(iocb->ki_pos);
1562         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1563         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1564
1565         if (q_id >= ub->dev_info.nr_hw_queues)
1566                 return ERR_PTR(-EINVAL);
1567
1568         ubq = ublk_get_queue(ub, q_id);
1569         if (!ubq)
1570                 return ERR_PTR(-EINVAL);
1571
1572         if (tag >= ubq->q_depth)
1573                 return ERR_PTR(-EINVAL);
1574
1575         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1576         if (!req)
1577                 return ERR_PTR(-EINVAL);
1578
1579         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1580                 goto fail;
1581
1582         if (!ublk_check_ubuf_dir(req, dir))
1583                 goto fail;
1584
1585         *off = buf_off;
1586         return req;
1587 fail:
1588         ublk_put_req_ref(ubq, req);
1589         return ERR_PTR(-EACCES);
1590 }
1591
1592 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1593 {
1594         struct ublk_queue *ubq;
1595         struct request *req;
1596         size_t buf_off;
1597         size_t ret;
1598
1599         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1600         if (IS_ERR(req))
1601                 return PTR_ERR(req);
1602
1603         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1604         ubq = req->mq_hctx->driver_data;
1605         ublk_put_req_ref(ubq, req);
1606
1607         return ret;
1608 }
1609
1610 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1611 {
1612         struct ublk_queue *ubq;
1613         struct request *req;
1614         size_t buf_off;
1615         size_t ret;
1616
1617         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1618         if (IS_ERR(req))
1619                 return PTR_ERR(req);
1620
1621         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1622         ubq = req->mq_hctx->driver_data;
1623         ublk_put_req_ref(ubq, req);
1624
1625         return ret;
1626 }
1627
1628 static const struct file_operations ublk_ch_fops = {
1629         .owner = THIS_MODULE,
1630         .open = ublk_ch_open,
1631         .release = ublk_ch_release,
1632         .llseek = no_llseek,
1633         .read_iter = ublk_ch_read_iter,
1634         .write_iter = ublk_ch_write_iter,
1635         .uring_cmd = ublk_ch_uring_cmd,
1636         .mmap = ublk_ch_mmap,
1637 };
1638
1639 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1640 {
1641         int size = ublk_queue_cmd_buf_size(ub, q_id);
1642         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1643
1644         if (ubq->ubq_daemon)
1645                 put_task_struct(ubq->ubq_daemon);
1646         if (ubq->io_cmd_buf)
1647                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1648 }
1649
1650 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1651 {
1652         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1653         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1654         void *ptr;
1655         int size;
1656
1657         ubq->flags = ub->dev_info.flags;
1658         ubq->q_id = q_id;
1659         ubq->q_depth = ub->dev_info.queue_depth;
1660         size = ublk_queue_cmd_buf_size(ub, q_id);
1661
1662         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1663         if (!ptr)
1664                 return -ENOMEM;
1665
1666         ubq->io_cmd_buf = ptr;
1667         ubq->dev = ub;
1668         return 0;
1669 }
1670
1671 static void ublk_deinit_queues(struct ublk_device *ub)
1672 {
1673         int nr_queues = ub->dev_info.nr_hw_queues;
1674         int i;
1675
1676         if (!ub->__queues)
1677                 return;
1678
1679         for (i = 0; i < nr_queues; i++)
1680                 ublk_deinit_queue(ub, i);
1681         kfree(ub->__queues);
1682 }
1683
1684 static int ublk_init_queues(struct ublk_device *ub)
1685 {
1686         int nr_queues = ub->dev_info.nr_hw_queues;
1687         int depth = ub->dev_info.queue_depth;
1688         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1689         int i, ret = -ENOMEM;
1690
1691         ub->queue_size = ubq_size;
1692         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1693         if (!ub->__queues)
1694                 return ret;
1695
1696         for (i = 0; i < nr_queues; i++) {
1697                 if (ublk_init_queue(ub, i))
1698                         goto fail;
1699         }
1700
1701         init_completion(&ub->completion);
1702         return 0;
1703
1704  fail:
1705         ublk_deinit_queues(ub);
1706         return ret;
1707 }
1708
1709 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1710 {
1711         int i = idx;
1712         int err;
1713
1714         spin_lock(&ublk_idr_lock);
1715         /* allocate id, if @id >= 0, we're requesting that specific id */
1716         if (i >= 0) {
1717                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1718                 if (err == -ENOSPC)
1719                         err = -EEXIST;
1720         } else {
1721                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1722         }
1723         spin_unlock(&ublk_idr_lock);
1724
1725         if (err >= 0)
1726                 ub->ub_number = err;
1727
1728         return err;
1729 }
1730
1731 static void ublk_free_dev_number(struct ublk_device *ub)
1732 {
1733         spin_lock(&ublk_idr_lock);
1734         idr_remove(&ublk_index_idr, ub->ub_number);
1735         wake_up_all(&ublk_idr_wq);
1736         spin_unlock(&ublk_idr_lock);
1737 }
1738
1739 static void ublk_cdev_rel(struct device *dev)
1740 {
1741         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1742
1743         blk_mq_free_tag_set(&ub->tag_set);
1744         ublk_deinit_queues(ub);
1745         ublk_free_dev_number(ub);
1746         mutex_destroy(&ub->mutex);
1747         kfree(ub);
1748 }
1749
1750 static int ublk_add_chdev(struct ublk_device *ub)
1751 {
1752         struct device *dev = &ub->cdev_dev;
1753         int minor = ub->ub_number;
1754         int ret;
1755
1756         dev->parent = ublk_misc.this_device;
1757         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1758         dev->class = ublk_chr_class;
1759         dev->release = ublk_cdev_rel;
1760         device_initialize(dev);
1761
1762         ret = dev_set_name(dev, "ublkc%d", minor);
1763         if (ret)
1764                 goto fail;
1765
1766         cdev_init(&ub->cdev, &ublk_ch_fops);
1767         ret = cdev_device_add(&ub->cdev, dev);
1768         if (ret)
1769                 goto fail;
1770
1771         ublks_added++;
1772         return 0;
1773  fail:
1774         put_device(dev);
1775         return ret;
1776 }
1777
1778 static void ublk_stop_work_fn(struct work_struct *work)
1779 {
1780         struct ublk_device *ub =
1781                 container_of(work, struct ublk_device, stop_work);
1782
1783         ublk_stop_dev(ub);
1784 }
1785
1786 /* align max io buffer size with PAGE_SIZE */
1787 static void ublk_align_max_io_size(struct ublk_device *ub)
1788 {
1789         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1790
1791         ub->dev_info.max_io_buf_bytes =
1792                 round_down(max_io_bytes, PAGE_SIZE);
1793 }
1794
1795 static int ublk_add_tag_set(struct ublk_device *ub)
1796 {
1797         ub->tag_set.ops = &ublk_mq_ops;
1798         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1799         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1800         ub->tag_set.numa_node = NUMA_NO_NODE;
1801         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1802         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1803         ub->tag_set.driver_data = ub;
1804         return blk_mq_alloc_tag_set(&ub->tag_set);
1805 }
1806
1807 static void ublk_remove(struct ublk_device *ub)
1808 {
1809         ublk_stop_dev(ub);
1810         cancel_work_sync(&ub->stop_work);
1811         cancel_work_sync(&ub->quiesce_work);
1812         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1813         put_device(&ub->cdev_dev);
1814         ublks_added--;
1815 }
1816
1817 static struct ublk_device *ublk_get_device_from_id(int idx)
1818 {
1819         struct ublk_device *ub = NULL;
1820
1821         if (idx < 0)
1822                 return NULL;
1823
1824         spin_lock(&ublk_idr_lock);
1825         ub = idr_find(&ublk_index_idr, idx);
1826         if (ub)
1827                 ub = ublk_get_device(ub);
1828         spin_unlock(&ublk_idr_lock);
1829
1830         return ub;
1831 }
1832
1833 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1834 {
1835         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1836         int ublksrv_pid = (int)header->data[0];
1837         struct gendisk *disk;
1838         int ret = -EINVAL;
1839
1840         if (ublksrv_pid <= 0)
1841                 return -EINVAL;
1842
1843         wait_for_completion_interruptible(&ub->completion);
1844
1845         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1846
1847         mutex_lock(&ub->mutex);
1848         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1849             test_bit(UB_STATE_USED, &ub->state)) {
1850                 ret = -EEXIST;
1851                 goto out_unlock;
1852         }
1853
1854         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1855         if (IS_ERR(disk)) {
1856                 ret = PTR_ERR(disk);
1857                 goto out_unlock;
1858         }
1859         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1860         disk->fops = &ub_fops;
1861         disk->private_data = ub;
1862
1863         ub->dev_info.ublksrv_pid = ublksrv_pid;
1864         ub->ub_disk = disk;
1865
1866         ret = ublk_apply_params(ub);
1867         if (ret)
1868                 goto out_put_disk;
1869
1870         /* don't probe partitions if any one ubq daemon is un-trusted */
1871         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1872                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1873
1874         get_device(&ub->cdev_dev);
1875         ub->dev_info.state = UBLK_S_DEV_LIVE;
1876         ret = add_disk(disk);
1877         if (ret) {
1878                 /*
1879                  * Has to drop the reference since ->free_disk won't be
1880                  * called in case of add_disk failure.
1881                  */
1882                 ub->dev_info.state = UBLK_S_DEV_DEAD;
1883                 ublk_put_device(ub);
1884                 goto out_put_disk;
1885         }
1886         set_bit(UB_STATE_USED, &ub->state);
1887 out_put_disk:
1888         if (ret)
1889                 put_disk(disk);
1890 out_unlock:
1891         mutex_unlock(&ub->mutex);
1892         return ret;
1893 }
1894
1895 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1896                 struct io_uring_cmd *cmd)
1897 {
1898         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1899         void __user *argp = (void __user *)(unsigned long)header->addr;
1900         cpumask_var_t cpumask;
1901         unsigned long queue;
1902         unsigned int retlen;
1903         unsigned int i;
1904         int ret;
1905
1906         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1907                 return -EINVAL;
1908         if (header->len & (sizeof(unsigned long)-1))
1909                 return -EINVAL;
1910         if (!header->addr)
1911                 return -EINVAL;
1912
1913         queue = header->data[0];
1914         if (queue >= ub->dev_info.nr_hw_queues)
1915                 return -EINVAL;
1916
1917         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1918                 return -ENOMEM;
1919
1920         for_each_possible_cpu(i) {
1921                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1922                         cpumask_set_cpu(i, cpumask);
1923         }
1924
1925         ret = -EFAULT;
1926         retlen = min_t(unsigned short, header->len, cpumask_size());
1927         if (copy_to_user(argp, cpumask, retlen))
1928                 goto out_free_cpumask;
1929         if (retlen != header->len &&
1930             clear_user(argp + retlen, header->len - retlen))
1931                 goto out_free_cpumask;
1932
1933         ret = 0;
1934 out_free_cpumask:
1935         free_cpumask_var(cpumask);
1936         return ret;
1937 }
1938
1939 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1940 {
1941         pr_devel("%s: dev id %d flags %llx\n", __func__,
1942                         info->dev_id, info->flags);
1943         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1944                         info->nr_hw_queues, info->queue_depth);
1945 }
1946
1947 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1948 {
1949         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1950         void __user *argp = (void __user *)(unsigned long)header->addr;
1951         struct ublksrv_ctrl_dev_info info;
1952         struct ublk_device *ub;
1953         int ret = -EINVAL;
1954
1955         if (header->len < sizeof(info) || !header->addr)
1956                 return -EINVAL;
1957         if (header->queue_id != (u16)-1) {
1958                 pr_warn("%s: queue_id is wrong %x\n",
1959                         __func__, header->queue_id);
1960                 return -EINVAL;
1961         }
1962
1963         if (copy_from_user(&info, argp, sizeof(info)))
1964                 return -EFAULT;
1965
1966         if (capable(CAP_SYS_ADMIN))
1967                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1968         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1969                 return -EPERM;
1970
1971         /*
1972          * unprivileged device can't be trusted, but RECOVERY and
1973          * RECOVERY_REISSUE still may hang error handling, so can't
1974          * support recovery features for unprivileged ublk now
1975          *
1976          * TODO: provide forward progress for RECOVERY handler, so that
1977          * unprivileged device can benefit from it
1978          */
1979         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1980                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1981                                 UBLK_F_USER_RECOVERY);
1982
1983         /* the created device is always owned by current user */
1984         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1985
1986         if (header->dev_id != info.dev_id) {
1987                 pr_warn("%s: dev id not match %u %u\n",
1988                         __func__, header->dev_id, info.dev_id);
1989                 return -EINVAL;
1990         }
1991
1992         ublk_dump_dev_info(&info);
1993
1994         ret = mutex_lock_killable(&ublk_ctl_mutex);
1995         if (ret)
1996                 return ret;
1997
1998         ret = -EACCES;
1999         if (ublks_added >= ublks_max)
2000                 goto out_unlock;
2001
2002         ret = -ENOMEM;
2003         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2004         if (!ub)
2005                 goto out_unlock;
2006         mutex_init(&ub->mutex);
2007         spin_lock_init(&ub->mm_lock);
2008         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2009         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2010         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
2011
2012         ret = ublk_alloc_dev_number(ub, header->dev_id);
2013         if (ret < 0)
2014                 goto out_free_ub;
2015
2016         memcpy(&ub->dev_info, &info, sizeof(info));
2017
2018         /* update device id */
2019         ub->dev_info.dev_id = ub->ub_number;
2020
2021         /*
2022          * 64bit flags will be copied back to userspace as feature
2023          * negotiation result, so have to clear flags which driver
2024          * doesn't support yet, then userspace can get correct flags
2025          * (features) to handle.
2026          */
2027         ub->dev_info.flags &= UBLK_F_ALL;
2028
2029         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2030                 UBLK_F_URING_CMD_COMP_IN_TASK;
2031
2032         /* GET_DATA isn't needed any more with USER_COPY */
2033         if (ub->dev_info.flags & UBLK_F_USER_COPY)
2034                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2035
2036         /* We are not ready to support zero copy */
2037         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2038
2039         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2040                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2041         ublk_align_max_io_size(ub);
2042
2043         ret = ublk_init_queues(ub);
2044         if (ret)
2045                 goto out_free_dev_number;
2046
2047         ret = ublk_add_tag_set(ub);
2048         if (ret)
2049                 goto out_deinit_queues;
2050
2051         ret = -EFAULT;
2052         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2053                 goto out_free_tag_set;
2054
2055         /*
2056          * Add the char dev so that ublksrv daemon can be setup.
2057          * ublk_add_chdev() will cleanup everything if it fails.
2058          */
2059         ret = ublk_add_chdev(ub);
2060         goto out_unlock;
2061
2062 out_free_tag_set:
2063         blk_mq_free_tag_set(&ub->tag_set);
2064 out_deinit_queues:
2065         ublk_deinit_queues(ub);
2066 out_free_dev_number:
2067         ublk_free_dev_number(ub);
2068 out_free_ub:
2069         mutex_destroy(&ub->mutex);
2070         kfree(ub);
2071 out_unlock:
2072         mutex_unlock(&ublk_ctl_mutex);
2073         return ret;
2074 }
2075
2076 static inline bool ublk_idr_freed(int id)
2077 {
2078         void *ptr;
2079
2080         spin_lock(&ublk_idr_lock);
2081         ptr = idr_find(&ublk_index_idr, id);
2082         spin_unlock(&ublk_idr_lock);
2083
2084         return ptr == NULL;
2085 }
2086
2087 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2088 {
2089         struct ublk_device *ub = *p_ub;
2090         int idx = ub->ub_number;
2091         int ret;
2092
2093         ret = mutex_lock_killable(&ublk_ctl_mutex);
2094         if (ret)
2095                 return ret;
2096
2097         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2098                 ublk_remove(ub);
2099                 set_bit(UB_STATE_DELETED, &ub->state);
2100         }
2101
2102         /* Mark the reference as consumed */
2103         *p_ub = NULL;
2104         ublk_put_device(ub);
2105         mutex_unlock(&ublk_ctl_mutex);
2106
2107         /*
2108          * Wait until the idr is removed, then it can be reused after
2109          * DEL_DEV command is returned.
2110          *
2111          * If we returns because of user interrupt, future delete command
2112          * may come:
2113          *
2114          * - the device number isn't freed, this device won't or needn't
2115          *   be deleted again, since UB_STATE_DELETED is set, and device
2116          *   will be released after the last reference is dropped
2117          *
2118          * - the device number is freed already, we will not find this
2119          *   device via ublk_get_device_from_id()
2120          */
2121         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
2122
2123         return 0;
2124 }
2125
2126 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2127 {
2128         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2129
2130         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2131                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2132                         header->data[0], header->addr, header->len);
2133 }
2134
2135 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2136 {
2137         ublk_stop_dev(ub);
2138         cancel_work_sync(&ub->stop_work);
2139         cancel_work_sync(&ub->quiesce_work);
2140
2141         return 0;
2142 }
2143
2144 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2145                 struct io_uring_cmd *cmd)
2146 {
2147         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2148         void __user *argp = (void __user *)(unsigned long)header->addr;
2149
2150         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2151                 return -EINVAL;
2152
2153         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2154                 return -EFAULT;
2155
2156         return 0;
2157 }
2158
2159 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2160 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2161 {
2162         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2163         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2164
2165         if (ub->ub_disk) {
2166                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2167                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2168         } else {
2169                 ub->params.devt.disk_major = 0;
2170                 ub->params.devt.disk_minor = 0;
2171         }
2172         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2173 }
2174
2175 static int ublk_ctrl_get_params(struct ublk_device *ub,
2176                 struct io_uring_cmd *cmd)
2177 {
2178         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2179         void __user *argp = (void __user *)(unsigned long)header->addr;
2180         struct ublk_params_header ph;
2181         int ret;
2182
2183         if (header->len <= sizeof(ph) || !header->addr)
2184                 return -EINVAL;
2185
2186         if (copy_from_user(&ph, argp, sizeof(ph)))
2187                 return -EFAULT;
2188
2189         if (ph.len > header->len || !ph.len)
2190                 return -EINVAL;
2191
2192         if (ph.len > sizeof(struct ublk_params))
2193                 ph.len = sizeof(struct ublk_params);
2194
2195         mutex_lock(&ub->mutex);
2196         ublk_ctrl_fill_params_devt(ub);
2197         if (copy_to_user(argp, &ub->params, ph.len))
2198                 ret = -EFAULT;
2199         else
2200                 ret = 0;
2201         mutex_unlock(&ub->mutex);
2202
2203         return ret;
2204 }
2205
2206 static int ublk_ctrl_set_params(struct ublk_device *ub,
2207                 struct io_uring_cmd *cmd)
2208 {
2209         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2210         void __user *argp = (void __user *)(unsigned long)header->addr;
2211         struct ublk_params_header ph;
2212         int ret = -EFAULT;
2213
2214         if (header->len <= sizeof(ph) || !header->addr)
2215                 return -EINVAL;
2216
2217         if (copy_from_user(&ph, argp, sizeof(ph)))
2218                 return -EFAULT;
2219
2220         if (ph.len > header->len || !ph.len || !ph.types)
2221                 return -EINVAL;
2222
2223         if (ph.len > sizeof(struct ublk_params))
2224                 ph.len = sizeof(struct ublk_params);
2225
2226         /* parameters can only be changed when device isn't live */
2227         mutex_lock(&ub->mutex);
2228         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2229                 ret = -EACCES;
2230         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2231                 ret = -EFAULT;
2232         } else {
2233                 /* clear all we don't support yet */
2234                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2235                 ret = ublk_validate_params(ub);
2236                 if (ret)
2237                         ub->params.types = 0;
2238         }
2239         mutex_unlock(&ub->mutex);
2240
2241         return ret;
2242 }
2243
2244 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2245 {
2246         int i;
2247
2248         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2249         /* All old ioucmds have to be completed */
2250         WARN_ON_ONCE(ubq->nr_io_ready);
2251         /* old daemon is PF_EXITING, put it now */
2252         put_task_struct(ubq->ubq_daemon);
2253         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2254         ubq->ubq_daemon = NULL;
2255         ubq->timeout = false;
2256
2257         for (i = 0; i < ubq->q_depth; i++) {
2258                 struct ublk_io *io = &ubq->ios[i];
2259
2260                 /* forget everything now and be ready for new FETCH_REQ */
2261                 io->flags = 0;
2262                 io->cmd = NULL;
2263                 io->addr = 0;
2264         }
2265 }
2266
2267 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2268                 struct io_uring_cmd *cmd)
2269 {
2270         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2271         int ret = -EINVAL;
2272         int i;
2273
2274         mutex_lock(&ub->mutex);
2275         if (!ublk_can_use_recovery(ub))
2276                 goto out_unlock;
2277         /*
2278          * START_RECOVERY is only allowd after:
2279          *
2280          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2281          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2282          *     released.
2283          *
2284          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2285          *     (a)has quiesced request queue
2286          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2287          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2288          *     (d)has completed/camceled all ioucmds owned by ther dying process
2289          */
2290         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2291                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2292                 ret = -EBUSY;
2293                 goto out_unlock;
2294         }
2295         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2296         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2297                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2298         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2299         ub->mm = NULL;
2300         ub->nr_queues_ready = 0;
2301         ub->nr_privileged_daemon = 0;
2302         init_completion(&ub->completion);
2303         ret = 0;
2304  out_unlock:
2305         mutex_unlock(&ub->mutex);
2306         return ret;
2307 }
2308
2309 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2310                 struct io_uring_cmd *cmd)
2311 {
2312         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2313         int ublksrv_pid = (int)header->data[0];
2314         int ret = -EINVAL;
2315
2316         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2317                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2318         /* wait until new ubq_daemon sending all FETCH_REQ */
2319         wait_for_completion_interruptible(&ub->completion);
2320         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2321                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2322
2323         mutex_lock(&ub->mutex);
2324         if (!ublk_can_use_recovery(ub))
2325                 goto out_unlock;
2326
2327         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2328                 ret = -EBUSY;
2329                 goto out_unlock;
2330         }
2331         ub->dev_info.ublksrv_pid = ublksrv_pid;
2332         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2333                         __func__, ublksrv_pid, header->dev_id);
2334         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2335         pr_devel("%s: queue unquiesced, dev id %d.\n",
2336                         __func__, header->dev_id);
2337         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2338         ub->dev_info.state = UBLK_S_DEV_LIVE;
2339         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2340         ret = 0;
2341  out_unlock:
2342         mutex_unlock(&ub->mutex);
2343         return ret;
2344 }
2345
2346 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2347 {
2348         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2349         void __user *argp = (void __user *)(unsigned long)header->addr;
2350         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2351
2352         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2353                 return -EINVAL;
2354
2355         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2356                 return -EFAULT;
2357
2358         return 0;
2359 }
2360
2361 /*
2362  * All control commands are sent via /dev/ublk-control, so we have to check
2363  * the destination device's permission
2364  */
2365 static int ublk_char_dev_permission(struct ublk_device *ub,
2366                 const char *dev_path, int mask)
2367 {
2368         int err;
2369         struct path path;
2370         struct kstat stat;
2371
2372         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2373         if (err)
2374                 return err;
2375
2376         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2377         if (err)
2378                 goto exit;
2379
2380         err = -EPERM;
2381         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2382                 goto exit;
2383
2384         err = inode_permission(&nop_mnt_idmap,
2385                         d_backing_inode(path.dentry), mask);
2386 exit:
2387         path_put(&path);
2388         return err;
2389 }
2390
2391 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2392                 struct io_uring_cmd *cmd)
2393 {
2394         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2395         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2396         void __user *argp = (void __user *)(unsigned long)header->addr;
2397         char *dev_path = NULL;
2398         int ret = 0;
2399         int mask;
2400
2401         if (!unprivileged) {
2402                 if (!capable(CAP_SYS_ADMIN))
2403                         return -EPERM;
2404                 /*
2405                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2406                  * char_dev_path in payload too, since userspace may not
2407                  * know if the specified device is created as unprivileged
2408                  * mode.
2409                  */
2410                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2411                         return 0;
2412         }
2413
2414         /*
2415          * User has to provide the char device path for unprivileged ublk
2416          *
2417          * header->addr always points to the dev path buffer, and
2418          * header->dev_path_len records length of dev path buffer.
2419          */
2420         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2421                 return -EINVAL;
2422
2423         if (header->len < header->dev_path_len)
2424                 return -EINVAL;
2425
2426         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2427         if (!dev_path)
2428                 return -ENOMEM;
2429
2430         ret = -EFAULT;
2431         if (copy_from_user(dev_path, argp, header->dev_path_len))
2432                 goto exit;
2433         dev_path[header->dev_path_len] = 0;
2434
2435         ret = -EINVAL;
2436         switch (_IOC_NR(cmd->cmd_op)) {
2437         case UBLK_CMD_GET_DEV_INFO:
2438         case UBLK_CMD_GET_DEV_INFO2:
2439         case UBLK_CMD_GET_QUEUE_AFFINITY:
2440         case UBLK_CMD_GET_PARAMS:
2441         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2442                 mask = MAY_READ;
2443                 break;
2444         case UBLK_CMD_START_DEV:
2445         case UBLK_CMD_STOP_DEV:
2446         case UBLK_CMD_ADD_DEV:
2447         case UBLK_CMD_DEL_DEV:
2448         case UBLK_CMD_SET_PARAMS:
2449         case UBLK_CMD_START_USER_RECOVERY:
2450         case UBLK_CMD_END_USER_RECOVERY:
2451                 mask = MAY_READ | MAY_WRITE;
2452                 break;
2453         default:
2454                 goto exit;
2455         }
2456
2457         ret = ublk_char_dev_permission(ub, dev_path, mask);
2458         if (!ret) {
2459                 header->len -= header->dev_path_len;
2460                 header->addr += header->dev_path_len;
2461         }
2462         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2463                         __func__, ub->ub_number, cmd->cmd_op,
2464                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2465                         dev_path, ret);
2466 exit:
2467         kfree(dev_path);
2468         return ret;
2469 }
2470
2471 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2472                 unsigned int issue_flags)
2473 {
2474         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2475         struct ublk_device *ub = NULL;
2476         u32 cmd_op = cmd->cmd_op;
2477         int ret = -EINVAL;
2478
2479         if (issue_flags & IO_URING_F_NONBLOCK)
2480                 return -EAGAIN;
2481
2482         ublk_ctrl_cmd_dump(cmd);
2483
2484         if (!(issue_flags & IO_URING_F_SQE128))
2485                 goto out;
2486
2487         ret = ublk_check_cmd_op(cmd_op);
2488         if (ret)
2489                 goto out;
2490
2491         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2492                 ret = ublk_ctrl_get_features(cmd);
2493                 goto out;
2494         }
2495
2496         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2497                 ret = -ENODEV;
2498                 ub = ublk_get_device_from_id(header->dev_id);
2499                 if (!ub)
2500                         goto out;
2501
2502                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2503                 if (ret)
2504                         goto put_dev;
2505         }
2506
2507         switch (_IOC_NR(cmd_op)) {
2508         case UBLK_CMD_START_DEV:
2509                 ret = ublk_ctrl_start_dev(ub, cmd);
2510                 break;
2511         case UBLK_CMD_STOP_DEV:
2512                 ret = ublk_ctrl_stop_dev(ub);
2513                 break;
2514         case UBLK_CMD_GET_DEV_INFO:
2515         case UBLK_CMD_GET_DEV_INFO2:
2516                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2517                 break;
2518         case UBLK_CMD_ADD_DEV:
2519                 ret = ublk_ctrl_add_dev(cmd);
2520                 break;
2521         case UBLK_CMD_DEL_DEV:
2522                 ret = ublk_ctrl_del_dev(&ub);
2523                 break;
2524         case UBLK_CMD_GET_QUEUE_AFFINITY:
2525                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2526                 break;
2527         case UBLK_CMD_GET_PARAMS:
2528                 ret = ublk_ctrl_get_params(ub, cmd);
2529                 break;
2530         case UBLK_CMD_SET_PARAMS:
2531                 ret = ublk_ctrl_set_params(ub, cmd);
2532                 break;
2533         case UBLK_CMD_START_USER_RECOVERY:
2534                 ret = ublk_ctrl_start_recovery(ub, cmd);
2535                 break;
2536         case UBLK_CMD_END_USER_RECOVERY:
2537                 ret = ublk_ctrl_end_recovery(ub, cmd);
2538                 break;
2539         default:
2540                 ret = -ENOTSUPP;
2541                 break;
2542         }
2543
2544  put_dev:
2545         if (ub)
2546                 ublk_put_device(ub);
2547  out:
2548         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2549         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2550                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2551         return -EIOCBQUEUED;
2552 }
2553
2554 static const struct file_operations ublk_ctl_fops = {
2555         .open           = nonseekable_open,
2556         .uring_cmd      = ublk_ctrl_uring_cmd,
2557         .owner          = THIS_MODULE,
2558         .llseek         = noop_llseek,
2559 };
2560
2561 static struct miscdevice ublk_misc = {
2562         .minor          = MISC_DYNAMIC_MINOR,
2563         .name           = "ublk-control",
2564         .fops           = &ublk_ctl_fops,
2565 };
2566
2567 static int __init ublk_init(void)
2568 {
2569         int ret;
2570
2571         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2572                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2573
2574         init_waitqueue_head(&ublk_idr_wq);
2575
2576         ret = misc_register(&ublk_misc);
2577         if (ret)
2578                 return ret;
2579
2580         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2581         if (ret)
2582                 goto unregister_mis;
2583
2584         ublk_chr_class = class_create("ublk-char");
2585         if (IS_ERR(ublk_chr_class)) {
2586                 ret = PTR_ERR(ublk_chr_class);
2587                 goto free_chrdev_region;
2588         }
2589         return 0;
2590
2591 free_chrdev_region:
2592         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2593 unregister_mis:
2594         misc_deregister(&ublk_misc);
2595         return ret;
2596 }
2597
2598 static void __exit ublk_exit(void)
2599 {
2600         struct ublk_device *ub;
2601         int id;
2602
2603         idr_for_each_entry(&ublk_index_idr, ub, id)
2604                 ublk_remove(ub);
2605
2606         class_destroy(ublk_chr_class);
2607         misc_deregister(&ublk_misc);
2608
2609         idr_destroy(&ublk_index_idr);
2610         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2611 }
2612
2613 module_init(ublk_init);
2614 module_exit(ublk_exit);
2615
2616 module_param(ublks_max, int, 0444);
2617 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2618
2619 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2620 MODULE_LICENSE("GPL");