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 const struct class ublk_chr_class = {
 193         .name = "ublk-char",
 194 };
 195
 196 static DEFINE_IDR(ublk_index_idr);
 197 static DEFINE_SPINLOCK(ublk_idr_lock);
 198 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 199
 200 static DEFINE_MUTEX(ublk_ctl_mutex);
 201
 202 /*
 203  * Max ublk devices allowed to add
 204  *
 205  * It can be extended to one per-user limit in future or even controlled
 206  * by cgroup.
 207  */
 208 static unsigned int ublks_max = 64;
 209 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 210
 211 static struct miscdevice ublk_misc;
 212
 213 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 214 {
 215         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 216                 UBLK_QID_BITS_MASK;
 217 }
 218
 219 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 220 {
 221         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 222 }
 223
 224 static inline unsigned ublk_pos_to_tag(loff_t pos)
 225 {
 226         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 227                 UBLK_TAG_BITS_MASK;
 228 }
 229
 230 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 231 {
 232         struct request_queue *q = ub->ub_disk->queue;
 233         const struct ublk_param_basic *p = &ub->params.basic;
 234
 235         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 236         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 237         blk_queue_io_min(q, 1 << p->io_min_shift);
 238         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 239
 240         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 241                         p->attrs & UBLK_ATTR_FUA);
 242         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 243                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 244         else
 245                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 246
 247         blk_queue_max_hw_sectors(q, p->max_sectors);
 248         blk_queue_chunk_sectors(q, p->chunk_sectors);
 249         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 250
 251         if (p->attrs & UBLK_ATTR_READ_ONLY)
 252                 set_disk_ro(ub->ub_disk, true);
 253
 254         set_capacity(ub->ub_disk, p->dev_sectors);
 255 }
 256
 257 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 258 {
 259         struct request_queue *q = ub->ub_disk->queue;
 260         const struct ublk_param_discard *p = &ub->params.discard;
 261
 262         q->limits.discard_alignment = p->discard_alignment;
 263         q->limits.discard_granularity = p->discard_granularity;
 264         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 265         blk_queue_max_write_zeroes_sectors(q,
 266                         p->max_write_zeroes_sectors);
 267         blk_queue_max_discard_segments(q, p->max_discard_segments);
 268 }
 269
 270 static int ublk_validate_params(const struct ublk_device *ub)
 271 {
 272         /* basic param is the only one which must be set */
 273         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 274                 const struct ublk_param_basic *p = &ub->params.basic;
 275
 276                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 277                         return -EINVAL;
 278
 279                 if (p->logical_bs_shift > p->physical_bs_shift)
 280                         return -EINVAL;
 281
 282                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 283                         return -EINVAL;
 284         } else
 285                 return -EINVAL;
 286
 287         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 288                 const struct ublk_param_discard *p = &ub->params.discard;
 289
 290                 /* So far, only support single segment discard */
 291                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 292                         return -EINVAL;
 293
 294                 if (!p->discard_granularity)
 295                         return -EINVAL;
 296         }
 297
 298         /* dev_t is read-only */
 299         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 300                 return -EINVAL;
 301
 302         return 0;
 303 }
 304
 305 static int ublk_apply_params(struct ublk_device *ub)
 306 {
 307         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 308                 return -EINVAL;
 309
 310         ublk_dev_param_basic_apply(ub);
 311
 312         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 313                 ublk_dev_param_discard_apply(ub);
 314
 315         return 0;
 316 }
 317
 318 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 319 {
 320         return ubq->flags & UBLK_F_USER_COPY;
 321 }
 322
 323 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 324 {
 325         /*
 326          * read()/write() is involved in user copy, so request reference
 327          * has to be grabbed
 328          */
 329         return ublk_support_user_copy(ubq);
 330 }
 331
 332 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 333                 struct request *req)
 334 {
 335         if (ublk_need_req_ref(ubq)) {
 336                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 337
 338                 kref_init(&data->ref);
 339         }
 340 }
 341
 342 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 343                 struct request *req)
 344 {
 345         if (ublk_need_req_ref(ubq)) {
 346                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 347
 348                 return kref_get_unless_zero(&data->ref);
 349         }
 350
 351         return true;
 352 }
 353
 354 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 355                 struct request *req)
 356 {
 357         if (ublk_need_req_ref(ubq)) {
 358                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 359
 360                 kref_put(&data->ref, ublk_complete_rq);
 361         } else {
 362                 __ublk_complete_rq(req);
 363         }
 364 }
 365
 366 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 367 {
 368         return ubq->flags & UBLK_F_NEED_GET_DATA;
 369 }
 370
 371 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 372 {
 373         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 374                 return ub;
 375         return NULL;
 376 }
 377
 378 static void ublk_put_device(struct ublk_device *ub)
 379 {
 380         put_device(&ub->cdev_dev);
 381 }
 382
 383 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 384                 int qid)
 385 {
 386        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 387 }
 388
 389 static inline bool ublk_rq_has_data(const struct request *rq)
 390 {
 391         return bio_has_data(rq->bio);
 392 }
 393
 394 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 395                 int tag)
 396 {
 397         return (struct ublksrv_io_desc *)
 398                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 399 }
 400
 401 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 402 {
 403         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 404 }
 405
 406 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 407 {
 408         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 409
 410         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 411                         PAGE_SIZE);
 412 }
 413
 414 static inline bool ublk_queue_can_use_recovery_reissue(
 415                 struct ublk_queue *ubq)
 416 {
 417         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 418                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 419 }
 420
 421 static inline bool ublk_queue_can_use_recovery(
 422                 struct ublk_queue *ubq)
 423 {
 424         return ubq->flags & UBLK_F_USER_RECOVERY;
 425 }
 426
 427 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 428 {
 429         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 430 }
 431
 432 static void ublk_free_disk(struct gendisk *disk)
 433 {
 434         struct ublk_device *ub = disk->private_data;
 435
 436         clear_bit(UB_STATE_USED, &ub->state);
 437         put_device(&ub->cdev_dev);
 438 }
 439
 440 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 441                 unsigned int *owner_gid)
 442 {
 443         kuid_t uid;
 444         kgid_t gid;
 445
 446         current_uid_gid(&uid, &gid);
 447
 448         *owner_uid = from_kuid(&init_user_ns, uid);
 449         *owner_gid = from_kgid(&init_user_ns, gid);
 450 }
 451
 452 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 453 {
 454         struct ublk_device *ub = disk->private_data;
 455
 456         if (capable(CAP_SYS_ADMIN))
 457                 return 0;
 458
 459         /*
 460          * If it is one unprivileged device, only owner can open
 461          * the disk. Otherwise it could be one trap made by one
 462          * evil user who grants this disk's privileges to other
 463          * users deliberately.
 464          *
 465          * This way is reasonable too given anyone can create
 466          * unprivileged device, and no need other's grant.
 467          */
 468         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 469                 unsigned int curr_uid, curr_gid;
 470
 471                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 472
 473                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 474                                 ub->dev_info.owner_gid)
 475                         return -EPERM;
 476         }
 477
 478         return 0;
 479 }
 480
 481 static const struct block_device_operations ub_fops = {
 482         .owner =        THIS_MODULE,
 483         .open =         ublk_open,
 484         .free_disk =    ublk_free_disk,
 485 };
 486
 487 #define UBLK_MAX_PIN_PAGES      32
 488
 489 struct ublk_io_iter {
 490         struct page *pages[UBLK_MAX_PIN_PAGES];
 491         struct bio *bio;
 492         struct bvec_iter iter;
 493 };
 494
 495 /* return how many pages are copied */
 496 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 497                 size_t total, size_t pg_off, int dir)
 498 {
 499         unsigned done = 0;
 500         unsigned pg_idx = 0;
 501
 502         while (done < total) {
 503                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 504                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 505                                 (unsigned)(PAGE_SIZE - pg_off));
 506                 void *bv_buf = bvec_kmap_local(&bv);
 507                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 508
 509                 if (dir == ITER_DEST)
 510                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 511                 else
 512                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 513
 514                 kunmap_local(pg_buf);
 515                 kunmap_local(bv_buf);
 516
 517                 /* advance page array */
 518                 pg_off += bytes;
 519                 if (pg_off == PAGE_SIZE) {
 520                         pg_idx += 1;
 521                         pg_off = 0;
 522                 }
 523
 524                 done += bytes;
 525
 526                 /* advance bio */
 527                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 528                 if (!data->iter.bi_size) {
 529                         data->bio = data->bio->bi_next;
 530                         if (data->bio == NULL)
 531                                 break;
 532                         data->iter = data->bio->bi_iter;
 533                 }
 534         }
 535 }
 536
 537 static bool ublk_advance_io_iter(const struct request *req,
 538                 struct ublk_io_iter *iter, unsigned int offset)
 539 {
 540         struct bio *bio = req->bio;
 541
 542         for_each_bio(bio) {
 543                 if (bio->bi_iter.bi_size > offset) {
 544                         iter->bio = bio;
 545                         iter->iter = bio->bi_iter;
 546                         bio_advance_iter(iter->bio, &iter->iter, offset);
 547                         return true;
 548                 }
 549                 offset -= bio->bi_iter.bi_size;
 550         }
 551         return false;
 552 }
 553
 554 /*
 555  * Copy data between request pages and io_iter, and 'offset'
 556  * is the start point of linear offset of request.
 557  */
 558 static size_t ublk_copy_user_pages(const struct request *req,
 559                 unsigned offset, struct iov_iter *uiter, int dir)
 560 {
 561         struct ublk_io_iter iter;
 562         size_t done = 0;
 563
 564         if (!ublk_advance_io_iter(req, &iter, offset))
 565                 return 0;
 566
 567         while (iov_iter_count(uiter) && iter.bio) {
 568                 unsigned nr_pages;
 569                 ssize_t len;
 570                 size_t off;
 571                 int i;
 572
 573                 len = iov_iter_get_pages2(uiter, iter.pages,
 574                                 iov_iter_count(uiter),
 575                                 UBLK_MAX_PIN_PAGES, &off);
 576                 if (len <= 0)
 577                         return done;
 578
 579                 ublk_copy_io_pages(&iter, len, off, dir);
 580                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 581                 for (i = 0; i < nr_pages; i++) {
 582                         if (dir == ITER_DEST)
 583                                 set_page_dirty(iter.pages[i]);
 584                         put_page(iter.pages[i]);
 585                 }
 586                 done += len;
 587         }
 588
 589         return done;
 590 }
 591
 592 static inline bool ublk_need_map_req(const struct request *req)
 593 {
 594         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 595 }
 596
 597 static inline bool ublk_need_unmap_req(const struct request *req)
 598 {
 599         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
 600 }
 601
 602 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 603                 struct ublk_io *io)
 604 {
 605         const unsigned int rq_bytes = blk_rq_bytes(req);
 606
 607         if (ublk_support_user_copy(ubq))
 608                 return rq_bytes;
 609
 610         /*
 611          * no zero copy, we delay copy WRITE request data into ublksrv
 612          * context and the big benefit is that pinning pages in current
 613          * context is pretty fast, see ublk_pin_user_pages
 614          */
 615         if (ublk_need_map_req(req)) {
 616                 struct iov_iter iter;
 617                 struct iovec iov;
 618                 const int dir = ITER_DEST;
 619
 620                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 621                                 &iov, &iter);
 622
 623                 return ublk_copy_user_pages(req, 0, &iter, dir);
 624         }
 625         return rq_bytes;
 626 }
 627
 628 static int ublk_unmap_io(const struct ublk_queue *ubq,
 629                 const struct request *req,
 630                 struct ublk_io *io)
 631 {
 632         const unsigned int rq_bytes = blk_rq_bytes(req);
 633
 634         if (ublk_support_user_copy(ubq))
 635                 return rq_bytes;
 636
 637         if (ublk_need_unmap_req(req)) {
 638                 struct iov_iter iter;
 639                 struct iovec iov;
 640                 const int dir = ITER_SOURCE;
 641
 642                 WARN_ON_ONCE(io->res > rq_bytes);
 643
 644                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 645                                 &iov, &iter);
 646                 return ublk_copy_user_pages(req, 0, &iter, dir);
 647         }
 648         return rq_bytes;
 649 }
 650
 651 static inline unsigned int ublk_req_build_flags(struct request *req)
 652 {
 653         unsigned flags = 0;
 654
 655         if (req->cmd_flags & REQ_FAILFAST_DEV)
 656                 flags |= UBLK_IO_F_FAILFAST_DEV;
 657
 658         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 659                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 660
 661         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 662                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 663
 664         if (req->cmd_flags & REQ_META)
 665                 flags |= UBLK_IO_F_META;
 666
 667         if (req->cmd_flags & REQ_FUA)
 668                 flags |= UBLK_IO_F_FUA;
 669
 670         if (req->cmd_flags & REQ_NOUNMAP)
 671                 flags |= UBLK_IO_F_NOUNMAP;
 672
 673         if (req->cmd_flags & REQ_SWAP)
 674                 flags |= UBLK_IO_F_SWAP;
 675
 676         return flags;
 677 }
 678
 679 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 680 {
 681         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 682         struct ublk_io *io = &ubq->ios[req->tag];
 683         u32 ublk_op;
 684
 685         switch (req_op(req)) {
 686         case REQ_OP_READ:
 687                 ublk_op = UBLK_IO_OP_READ;
 688                 break;
 689         case REQ_OP_WRITE:
 690                 ublk_op = UBLK_IO_OP_WRITE;
 691                 break;
 692         case REQ_OP_FLUSH:
 693                 ublk_op = UBLK_IO_OP_FLUSH;
 694                 break;
 695         case REQ_OP_DISCARD:
 696                 ublk_op = UBLK_IO_OP_DISCARD;
 697                 break;
 698         case REQ_OP_WRITE_ZEROES:
 699                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 700                 break;
 701         default:
 702                 return BLK_STS_IOERR;
 703         }
 704
 705         /* need to translate since kernel may change */
 706         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 707         iod->nr_sectors = blk_rq_sectors(req);
 708         iod->start_sector = blk_rq_pos(req);
 709         iod->addr = io->addr;
 710
 711         return BLK_STS_OK;
 712 }
 713
 714 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 715                 struct io_uring_cmd *ioucmd)
 716 {
 717         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 718 }
 719
 720 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 721 {
 722         return ubq->ubq_daemon->flags & PF_EXITING;
 723 }
 724
 725 /* todo: handle partial completion */
 726 static inline void __ublk_complete_rq(struct request *req)
 727 {
 728         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 729         struct ublk_io *io = &ubq->ios[req->tag];
 730         unsigned int unmapped_bytes;
 731         blk_status_t res = BLK_STS_OK;
 732
 733         /* called from ublk_abort_queue() code path */
 734         if (io->flags & UBLK_IO_FLAG_ABORTED) {
 735                 res = BLK_STS_IOERR;
 736                 goto exit;
 737         }
 738
 739         /* failed read IO if nothing is read */
 740         if (!io->res && req_op(req) == REQ_OP_READ)
 741                 io->res = -EIO;
 742
 743         if (io->res < 0) {
 744                 res = errno_to_blk_status(io->res);
 745                 goto exit;
 746         }
 747
 748         /*
 749          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
 750          * directly.
 751          *
 752          * Both the two needn't unmap.
 753          */
 754         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
 755                 goto exit;
 756
 757         /* for READ request, writing data in iod->addr to rq buffers */
 758         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 759
 760         /*
 761          * Extremely impossible since we got data filled in just before
 762          *
 763          * Re-read simply for this unlikely case.
 764          */
 765         if (unlikely(unmapped_bytes < io->res))
 766                 io->res = unmapped_bytes;
 767
 768         if (blk_update_request(req, BLK_STS_OK, io->res))
 769                 blk_mq_requeue_request(req, true);
 770         else
 771                 __blk_mq_end_request(req, BLK_STS_OK);
 772
 773         return;
 774 exit:
 775         blk_mq_end_request(req, res);
 776 }
 777
 778 static void ublk_complete_rq(struct kref *ref)
 779 {
 780         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
 781                         ref);
 782         struct request *req = blk_mq_rq_from_pdu(data);
 783
 784         __ublk_complete_rq(req);
 785 }
 786
 787 /*
 788  * Since __ublk_rq_task_work always fails requests immediately during
 789  * exiting, __ublk_fail_req() is only called from abort context during
 790  * exiting. So lock is unnecessary.
 791  *
 792  * Also aborting may not be started yet, keep in mind that one failed
 793  * request may be issued by block layer again.
 794  */
 795 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
 796                 struct request *req)
 797 {
 798         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 799
 800         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 801                 io->flags |= UBLK_IO_FLAG_ABORTED;
 802                 if (ublk_queue_can_use_recovery_reissue(ubq))
 803                         blk_mq_requeue_request(req, false);
 804                 else
 805                         ublk_put_req_ref(ubq, req);
 806         }
 807 }
 808
 809 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
 810                                 unsigned issue_flags)
 811 {
 812         /* mark this cmd owned by ublksrv */
 813         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 814
 815         /*
 816          * clear ACTIVE since we are done with this sqe/cmd slot
 817          * We can only accept io cmd in case of being not active.
 818          */
 819         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 820
 821         /* tell ublksrv one io request is coming */
 822         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
 823 }
 824
 825 #define UBLK_REQUEUE_DELAY_MS   3
 826
 827 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 828                 struct request *rq)
 829 {
 830         /* We cannot process this rq so just requeue it. */
 831         if (ublk_queue_can_use_recovery(ubq))
 832                 blk_mq_requeue_request(rq, false);
 833         else
 834                 blk_mq_end_request(rq, BLK_STS_IOERR);
 835
 836         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 837 }
 838
 839 static inline void __ublk_rq_task_work(struct request *req,
 840                                        unsigned issue_flags)
 841 {
 842         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 843         int tag = req->tag;
 844         struct ublk_io *io = &ubq->ios[tag];
 845         unsigned int mapped_bytes;
 846
 847         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 848                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 849                         ublk_get_iod(ubq, req->tag)->addr);
 850
 851         /*
 852          * Task is exiting if either:
 853          *
 854          * (1) current != ubq_daemon.
 855          * io_uring_cmd_complete_in_task() tries to run task_work
 856          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 857          *
 858          * (2) current->flags & PF_EXITING.
 859          */
 860         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 861                 __ublk_abort_rq(ubq, req);
 862                 return;
 863         }
 864
 865         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
 866                 /*
 867                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 868                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 869                  * and notify it.
 870                  */
 871                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 872                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 873                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 874                                         __func__, io->cmd->cmd_op, ubq->q_id,
 875                                         req->tag, io->flags);
 876                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
 877                         return;
 878                 }
 879                 /*
 880                  * We have handled UBLK_IO_NEED_GET_DATA command,
 881                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 882                  * do the copy work.
 883                  */
 884                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 885                 /* update iod->addr because ublksrv may have passed a new io buffer */
 886                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 887                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 888                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 889                                 ublk_get_iod(ubq, req->tag)->addr);
 890         }
 891
 892         mapped_bytes = ublk_map_io(ubq, req, io);
 893
 894         /* partially mapped, update io descriptor */
 895         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 896                 /*
 897                  * Nothing mapped, retry until we succeed.
 898                  *
 899                  * We may never succeed in mapping any bytes here because
 900                  * of OOM. TODO: reserve one buffer with single page pinned
 901                  * for providing forward progress guarantee.
 902                  */
 903                 if (unlikely(!mapped_bytes)) {
 904                         blk_mq_requeue_request(req, false);
 905                         blk_mq_delay_kick_requeue_list(req->q,
 906                                         UBLK_REQUEUE_DELAY_MS);
 907                         return;
 908                 }
 909
 910                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 911                         mapped_bytes >> 9;
 912         }
 913
 914         ublk_init_req_ref(ubq, req);
 915         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
 916 }
 917
 918 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
 919                                         unsigned issue_flags)
 920 {
 921         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 922         struct ublk_rq_data *data, *tmp;
 923
 924         io_cmds = llist_reverse_order(io_cmds);
 925         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 926                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
 927 }
 928
 929 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 930 {
 931         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 932         struct ublk_rq_data *data, *tmp;
 933
 934         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 935                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 936 }
 937
 938 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
 939 {
 940         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 941         struct ublk_queue *ubq = pdu->ubq;
 942
 943         ublk_forward_io_cmds(ubq, issue_flags);
 944 }
 945
 946 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 947 {
 948         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 949         struct ublk_io *io;
 950
 951         if (!llist_add(&data->node, &ubq->io_cmds))
 952                 return;
 953
 954         io = &ubq->ios[rq->tag];
 955         /*
 956          * If the check pass, we know that this is a re-issued request aborted
 957          * previously in monitor_work because the ubq_daemon(cmd's task) is
 958          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 959          * because this ioucmd's io_uring context may be freed now if no inflight
 960          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 961          *
 962          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 963          * the tag). Then the request is re-started(allocating the tag) and we are here.
 964          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 965          * guarantees that here is a re-issued request aborted previously.
 966          */
 967         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 968                 ublk_abort_io_cmds(ubq);
 969         } else {
 970                 struct io_uring_cmd *cmd = io->cmd;
 971                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 972
 973                 pdu->ubq = ubq;
 974                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 975         }
 976 }
 977
 978 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
 979 {
 980         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
 981
 982         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
 983                 if (!ubq->timeout) {
 984                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
 985                         ubq->timeout = true;
 986                 }
 987
 988                 return BLK_EH_DONE;
 989         }
 990
 991         return BLK_EH_RESET_TIMER;
 992 }
 993
 994 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 995                 const struct blk_mq_queue_data *bd)
 996 {
 997         struct ublk_queue *ubq = hctx->driver_data;
 998         struct request *rq = bd->rq;
 999         blk_status_t res;
1000
1001         /* fill iod to slot in io cmd buffer */
1002         res = ublk_setup_iod(ubq, rq);
1003         if (unlikely(res != BLK_STS_OK))
1004                 return BLK_STS_IOERR;
1005
1006         /* With recovery feature enabled, force_abort is set in
1007          * ublk_stop_dev() before calling del_gendisk(). We have to
1008          * abort all requeued and new rqs here to let del_gendisk()
1009          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1010          * to avoid UAF on io_uring ctx.
1011          *
1012          * Note: force_abort is guaranteed to be seen because it is set
1013          * before request queue is unqiuesced.
1014          */
1015         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1016                 return BLK_STS_IOERR;
1017
1018         blk_mq_start_request(bd->rq);
1019
1020         if (unlikely(ubq_daemon_is_dying(ubq))) {
1021                 __ublk_abort_rq(ubq, rq);
1022                 return BLK_STS_OK;
1023         }
1024
1025         ublk_queue_cmd(ubq, rq);
1026
1027         return BLK_STS_OK;
1028 }
1029
1030 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1031                 unsigned int hctx_idx)
1032 {
1033         struct ublk_device *ub = driver_data;
1034         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1035
1036         hctx->driver_data = ubq;
1037         return 0;
1038 }
1039
1040 static const struct blk_mq_ops ublk_mq_ops = {
1041         .queue_rq       = ublk_queue_rq,
1042         .init_hctx      = ublk_init_hctx,
1043         .timeout        = ublk_timeout,
1044 };
1045
1046 static int ublk_ch_open(struct inode *inode, struct file *filp)
1047 {
1048         struct ublk_device *ub = container_of(inode->i_cdev,
1049                         struct ublk_device, cdev);
1050
1051         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1052                 return -EBUSY;
1053         filp->private_data = ub;
1054         return 0;
1055 }
1056
1057 static int ublk_ch_release(struct inode *inode, struct file *filp)
1058 {
1059         struct ublk_device *ub = filp->private_data;
1060
1061         clear_bit(UB_STATE_OPEN, &ub->state);
1062         return 0;
1063 }
1064
1065 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1066 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1067 {
1068         struct ublk_device *ub = filp->private_data;
1069         size_t sz = vma->vm_end - vma->vm_start;
1070         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1071         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1072         int q_id, ret = 0;
1073
1074         spin_lock(&ub->mm_lock);
1075         if (!ub->mm)
1076                 ub->mm = current->mm;
1077         if (current->mm != ub->mm)
1078                 ret = -EINVAL;
1079         spin_unlock(&ub->mm_lock);
1080
1081         if (ret)
1082                 return ret;
1083
1084         if (vma->vm_flags & VM_WRITE)
1085                 return -EPERM;
1086
1087         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1088         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1089                 return -EINVAL;
1090
1091         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1092         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1093                         __func__, q_id, current->pid, vma->vm_start,
1094                         phys_off, (unsigned long)sz);
1095
1096         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1097                 return -EINVAL;
1098
1099         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1100         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1101 }
1102
1103 static void ublk_commit_completion(struct ublk_device *ub,
1104                 const struct ublksrv_io_cmd *ub_cmd)
1105 {
1106         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1107         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1108         struct ublk_io *io = &ubq->ios[tag];
1109         struct request *req;
1110
1111         /* now this cmd slot is owned by nbd driver */
1112         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1113         io->res = ub_cmd->result;
1114
1115         /* find the io request and complete */
1116         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1117
1118         if (req && likely(!blk_should_fake_timeout(req->q)))
1119                 ublk_put_req_ref(ubq, req);
1120 }
1121
1122 /*
1123  * When ->ubq_daemon is exiting, either new request is ended immediately,
1124  * or any queued io command is drained, so it is safe to abort queue
1125  * lockless
1126  */
1127 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1128 {
1129         int i;
1130
1131         if (!ublk_get_device(ub))
1132                 return;
1133
1134         for (i = 0; i < ubq->q_depth; i++) {
1135                 struct ublk_io *io = &ubq->ios[i];
1136
1137                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1138                         struct request *rq;
1139
1140                         /*
1141                          * Either we fail the request or ublk_rq_task_work_fn
1142                          * will do it
1143                          */
1144                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1145                         if (rq)
1146                                 __ublk_fail_req(ubq, io, rq);
1147                 }
1148         }
1149         ublk_put_device(ub);
1150 }
1151
1152 static void ublk_daemon_monitor_work(struct work_struct *work)
1153 {
1154         struct ublk_device *ub =
1155                 container_of(work, struct ublk_device, monitor_work.work);
1156         int i;
1157
1158         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1159                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1160
1161                 if (ubq_daemon_is_dying(ubq)) {
1162                         if (ublk_queue_can_use_recovery(ubq))
1163                                 schedule_work(&ub->quiesce_work);
1164                         else
1165                                 schedule_work(&ub->stop_work);
1166
1167                         /* abort queue is for making forward progress */
1168                         ublk_abort_queue(ub, ubq);
1169                 }
1170         }
1171
1172         /*
1173          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1174          * after ublk_remove() or __ublk_quiesce_dev() is started.
1175          *
1176          * No need ub->mutex, monitor work are canceled after state is marked
1177          * as not LIVE, so new state is observed reliably.
1178          */
1179         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1180                 schedule_delayed_work(&ub->monitor_work,
1181                                 UBLK_DAEMON_MONITOR_PERIOD);
1182 }
1183
1184 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1185 {
1186         return ubq->nr_io_ready == ubq->q_depth;
1187 }
1188
1189 static void ublk_cmd_cancel_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1190 {
1191         io_uring_cmd_done(cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1192 }
1193
1194 static void ublk_cancel_queue(struct ublk_queue *ubq)
1195 {
1196         int i;
1197
1198         if (!ublk_queue_ready(ubq))
1199                 return;
1200
1201         for (i = 0; i < ubq->q_depth; i++) {
1202                 struct ublk_io *io = &ubq->ios[i];
1203
1204                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1205                         io_uring_cmd_complete_in_task(io->cmd,
1206                                                       ublk_cmd_cancel_cb);
1207         }
1208
1209         /* all io commands are canceled */
1210         ubq->nr_io_ready = 0;
1211 }
1212
1213 /* Cancel all pending commands, must be called after del_gendisk() returns */
1214 static void ublk_cancel_dev(struct ublk_device *ub)
1215 {
1216         int i;
1217
1218         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1219                 ublk_cancel_queue(ublk_get_queue(ub, i));
1220 }
1221
1222 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1223 {
1224         bool *idle = data;
1225
1226         if (blk_mq_request_started(rq)) {
1227                 *idle = false;
1228                 return false;
1229         }
1230         return true;
1231 }
1232
1233 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1234 {
1235         bool idle;
1236
1237         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1238         while (true) {
1239                 idle = true;
1240                 blk_mq_tagset_busy_iter(&ub->tag_set,
1241                                 ublk_check_inflight_rq, &idle);
1242                 if (idle)
1243                         break;
1244                 msleep(UBLK_REQUEUE_DELAY_MS);
1245         }
1246 }
1247
1248 static void __ublk_quiesce_dev(struct ublk_device *ub)
1249 {
1250         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1251                         __func__, ub->dev_info.dev_id,
1252                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1253                         "LIVE" : "QUIESCED");
1254         blk_mq_quiesce_queue(ub->ub_disk->queue);
1255         ublk_wait_tagset_rqs_idle(ub);
1256         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1257         ublk_cancel_dev(ub);
1258         /* we are going to release task_struct of ubq_daemon and resets
1259          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1260          * Besides, monitor_work is not necessary in QUIESCED state since we have
1261          * already scheduled quiesce_work and quiesced all ubqs.
1262          *
1263          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1264          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1265          */
1266         cancel_delayed_work_sync(&ub->monitor_work);
1267 }
1268
1269 static void ublk_quiesce_work_fn(struct work_struct *work)
1270 {
1271         struct ublk_device *ub =
1272                 container_of(work, struct ublk_device, quiesce_work);
1273
1274         mutex_lock(&ub->mutex);
1275         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1276                 goto unlock;
1277         __ublk_quiesce_dev(ub);
1278  unlock:
1279         mutex_unlock(&ub->mutex);
1280 }
1281
1282 static void ublk_unquiesce_dev(struct ublk_device *ub)
1283 {
1284         int i;
1285
1286         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1287                         __func__, ub->dev_info.dev_id,
1288                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1289                         "LIVE" : "QUIESCED");
1290         /* quiesce_work has run. We let requeued rqs be aborted
1291          * before running fallback_wq. "force_abort" must be seen
1292          * after request queue is unqiuesced. Then del_gendisk()
1293          * can move on.
1294          */
1295         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1296                 ublk_get_queue(ub, i)->force_abort = true;
1297
1298         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1299         /* We may have requeued some rqs in ublk_quiesce_queue() */
1300         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1301 }
1302
1303 static void ublk_stop_dev(struct ublk_device *ub)
1304 {
1305         mutex_lock(&ub->mutex);
1306         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1307                 goto unlock;
1308         if (ublk_can_use_recovery(ub)) {
1309                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1310                         __ublk_quiesce_dev(ub);
1311                 ublk_unquiesce_dev(ub);
1312         }
1313         del_gendisk(ub->ub_disk);
1314         ub->dev_info.state = UBLK_S_DEV_DEAD;
1315         ub->dev_info.ublksrv_pid = -1;
1316         put_disk(ub->ub_disk);
1317         ub->ub_disk = NULL;
1318  unlock:
1319         ublk_cancel_dev(ub);
1320         mutex_unlock(&ub->mutex);
1321         cancel_delayed_work_sync(&ub->monitor_work);
1322 }
1323
1324 /* device can only be started after all IOs are ready */
1325 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1326 {
1327         mutex_lock(&ub->mutex);
1328         ubq->nr_io_ready++;
1329         if (ublk_queue_ready(ubq)) {
1330                 ubq->ubq_daemon = current;
1331                 get_task_struct(ubq->ubq_daemon);
1332                 ub->nr_queues_ready++;
1333
1334                 if (capable(CAP_SYS_ADMIN))
1335                         ub->nr_privileged_daemon++;
1336         }
1337         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1338                 complete_all(&ub->completion);
1339         mutex_unlock(&ub->mutex);
1340 }
1341
1342 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1343                 int tag)
1344 {
1345         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1346         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1347
1348         ublk_queue_cmd(ubq, req);
1349 }
1350
1351 static inline int ublk_check_cmd_op(u32 cmd_op)
1352 {
1353         u32 ioc_type = _IOC_TYPE(cmd_op);
1354
1355         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1356                 return -EOPNOTSUPP;
1357
1358         if (ioc_type != 'u' && ioc_type != 0)
1359                 return -EOPNOTSUPP;
1360
1361         return 0;
1362 }
1363
1364 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1365                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1366 {
1367         io->cmd = cmd;
1368         io->flags |= UBLK_IO_FLAG_ACTIVE;
1369         io->addr = buf_addr;
1370 }
1371
1372 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1373                                unsigned int issue_flags,
1374                                const struct ublksrv_io_cmd *ub_cmd)
1375 {
1376         struct ublk_device *ub = cmd->file->private_data;
1377         struct ublk_queue *ubq;
1378         struct ublk_io *io;
1379         u32 cmd_op = cmd->cmd_op;
1380         unsigned tag = ub_cmd->tag;
1381         int ret = -EINVAL;
1382         struct request *req;
1383
1384         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1385                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1386                         ub_cmd->result);
1387
1388         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1389                 goto out;
1390
1391         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1392         if (!ubq || ub_cmd->q_id != ubq->q_id)
1393                 goto out;
1394
1395         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1396                 goto out;
1397
1398         if (tag >= ubq->q_depth)
1399                 goto out;
1400
1401         io = &ubq->ios[tag];
1402
1403         /* there is pending io cmd, something must be wrong */
1404         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1405                 ret = -EBUSY;
1406                 goto out;
1407         }
1408
1409         /*
1410          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1411          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1412          */
1413         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1414                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1415                 goto out;
1416
1417         if (ublk_support_user_copy(ubq) && ub_cmd->addr) {
1418                 ret = -EINVAL;
1419                 goto out;
1420         }
1421
1422         ret = ublk_check_cmd_op(cmd_op);
1423         if (ret)
1424                 goto out;
1425
1426         ret = -EINVAL;
1427         switch (_IOC_NR(cmd_op)) {
1428         case UBLK_IO_FETCH_REQ:
1429                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1430                 if (ublk_queue_ready(ubq)) {
1431                         ret = -EBUSY;
1432                         goto out;
1433                 }
1434                 /*
1435                  * The io is being handled by server, so COMMIT_RQ is expected
1436                  * instead of FETCH_REQ
1437                  */
1438                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1439                         goto out;
1440
1441                 if (!ublk_support_user_copy(ubq)) {
1442                         /*
1443                          * FETCH_RQ has to provide IO buffer if NEED GET
1444                          * DATA is not enabled
1445                          */
1446                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1447                                 goto out;
1448                 }
1449
1450                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1451                 ublk_mark_io_ready(ub, ubq);
1452                 break;
1453         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1454                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1455
1456                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1457                         goto out;
1458
1459                 if (!ublk_support_user_copy(ubq)) {
1460                         /*
1461                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1462                          * NEED GET DATA is not enabled or it is Read IO.
1463                          */
1464                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1465                                                 req_op(req) == REQ_OP_READ))
1466                                 goto out;
1467                 }
1468                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1469                 ublk_commit_completion(ub, ub_cmd);
1470                 break;
1471         case UBLK_IO_NEED_GET_DATA:
1472                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1473                         goto out;
1474                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1475                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1476                 break;
1477         default:
1478                 goto out;
1479         }
1480         return -EIOCBQUEUED;
1481
1482  out:
1483         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1484         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1485                         __func__, cmd_op, tag, ret, io->flags);
1486         return -EIOCBQUEUED;
1487 }
1488
1489 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1490                 struct ublk_queue *ubq, int tag, size_t offset)
1491 {
1492         struct request *req;
1493
1494         if (!ublk_need_req_ref(ubq))
1495                 return NULL;
1496
1497         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1498         if (!req)
1499                 return NULL;
1500
1501         if (!ublk_get_req_ref(ubq, req))
1502                 return NULL;
1503
1504         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1505                 goto fail_put;
1506
1507         if (!ublk_rq_has_data(req))
1508                 goto fail_put;
1509
1510         if (offset > blk_rq_bytes(req))
1511                 goto fail_put;
1512
1513         return req;
1514 fail_put:
1515         ublk_put_req_ref(ubq, req);
1516         return NULL;
1517 }
1518
1519 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1520 {
1521         /*
1522          * Not necessary for async retry, but let's keep it simple and always
1523          * copy the values to avoid any potential reuse.
1524          */
1525         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1526         const struct ublksrv_io_cmd ub_cmd = {
1527                 .q_id = READ_ONCE(ub_src->q_id),
1528                 .tag = READ_ONCE(ub_src->tag),
1529                 .result = READ_ONCE(ub_src->result),
1530                 .addr = READ_ONCE(ub_src->addr)
1531         };
1532
1533         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1534 }
1535
1536 static inline bool ublk_check_ubuf_dir(const struct request *req,
1537                 int ubuf_dir)
1538 {
1539         /* copy ubuf to request pages */
1540         if (req_op(req) == REQ_OP_READ && ubuf_dir == ITER_SOURCE)
1541                 return true;
1542
1543         /* copy request pages to ubuf */
1544         if (req_op(req) == REQ_OP_WRITE && ubuf_dir == ITER_DEST)
1545                 return true;
1546
1547         return false;
1548 }
1549
1550 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1551                 struct iov_iter *iter, size_t *off, int dir)
1552 {
1553         struct ublk_device *ub = iocb->ki_filp->private_data;
1554         struct ublk_queue *ubq;
1555         struct request *req;
1556         size_t buf_off;
1557         u16 tag, q_id;
1558
1559         if (!ub)
1560                 return ERR_PTR(-EACCES);
1561
1562         if (!user_backed_iter(iter))
1563                 return ERR_PTR(-EACCES);
1564
1565         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1566                 return ERR_PTR(-EACCES);
1567
1568         tag = ublk_pos_to_tag(iocb->ki_pos);
1569         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1570         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1571
1572         if (q_id >= ub->dev_info.nr_hw_queues)
1573                 return ERR_PTR(-EINVAL);
1574
1575         ubq = ublk_get_queue(ub, q_id);
1576         if (!ubq)
1577                 return ERR_PTR(-EINVAL);
1578
1579         if (tag >= ubq->q_depth)
1580                 return ERR_PTR(-EINVAL);
1581
1582         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1583         if (!req)
1584                 return ERR_PTR(-EINVAL);
1585
1586         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1587                 goto fail;
1588
1589         if (!ublk_check_ubuf_dir(req, dir))
1590                 goto fail;
1591
1592         *off = buf_off;
1593         return req;
1594 fail:
1595         ublk_put_req_ref(ubq, req);
1596         return ERR_PTR(-EACCES);
1597 }
1598
1599 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1600 {
1601         struct ublk_queue *ubq;
1602         struct request *req;
1603         size_t buf_off;
1604         size_t ret;
1605
1606         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1607         if (IS_ERR(req))
1608                 return PTR_ERR(req);
1609
1610         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1611         ubq = req->mq_hctx->driver_data;
1612         ublk_put_req_ref(ubq, req);
1613
1614         return ret;
1615 }
1616
1617 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1618 {
1619         struct ublk_queue *ubq;
1620         struct request *req;
1621         size_t buf_off;
1622         size_t ret;
1623
1624         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1625         if (IS_ERR(req))
1626                 return PTR_ERR(req);
1627
1628         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1629         ubq = req->mq_hctx->driver_data;
1630         ublk_put_req_ref(ubq, req);
1631
1632         return ret;
1633 }
1634
1635 static const struct file_operations ublk_ch_fops = {
1636         .owner = THIS_MODULE,
1637         .open = ublk_ch_open,
1638         .release = ublk_ch_release,
1639         .llseek = no_llseek,
1640         .read_iter = ublk_ch_read_iter,
1641         .write_iter = ublk_ch_write_iter,
1642         .uring_cmd = ublk_ch_uring_cmd,
1643         .mmap = ublk_ch_mmap,
1644 };
1645
1646 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1647 {
1648         int size = ublk_queue_cmd_buf_size(ub, q_id);
1649         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1650
1651         if (ubq->ubq_daemon)
1652                 put_task_struct(ubq->ubq_daemon);
1653         if (ubq->io_cmd_buf)
1654                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1655 }
1656
1657 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1658 {
1659         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1660         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1661         void *ptr;
1662         int size;
1663
1664         ubq->flags = ub->dev_info.flags;
1665         ubq->q_id = q_id;
1666         ubq->q_depth = ub->dev_info.queue_depth;
1667         size = ublk_queue_cmd_buf_size(ub, q_id);
1668
1669         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1670         if (!ptr)
1671                 return -ENOMEM;
1672
1673         ubq->io_cmd_buf = ptr;
1674         ubq->dev = ub;
1675         return 0;
1676 }
1677
1678 static void ublk_deinit_queues(struct ublk_device *ub)
1679 {
1680         int nr_queues = ub->dev_info.nr_hw_queues;
1681         int i;
1682
1683         if (!ub->__queues)
1684                 return;
1685
1686         for (i = 0; i < nr_queues; i++)
1687                 ublk_deinit_queue(ub, i);
1688         kfree(ub->__queues);
1689 }
1690
1691 static int ublk_init_queues(struct ublk_device *ub)
1692 {
1693         int nr_queues = ub->dev_info.nr_hw_queues;
1694         int depth = ub->dev_info.queue_depth;
1695         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1696         int i, ret = -ENOMEM;
1697
1698         ub->queue_size = ubq_size;
1699         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1700         if (!ub->__queues)
1701                 return ret;
1702
1703         for (i = 0; i < nr_queues; i++) {
1704                 if (ublk_init_queue(ub, i))
1705                         goto fail;
1706         }
1707
1708         init_completion(&ub->completion);
1709         return 0;
1710
1711  fail:
1712         ublk_deinit_queues(ub);
1713         return ret;
1714 }
1715
1716 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1717 {
1718         int i = idx;
1719         int err;
1720
1721         spin_lock(&ublk_idr_lock);
1722         /* allocate id, if @id >= 0, we're requesting that specific id */
1723         if (i >= 0) {
1724                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1725                 if (err == -ENOSPC)
1726                         err = -EEXIST;
1727         } else {
1728                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1729         }
1730         spin_unlock(&ublk_idr_lock);
1731
1732         if (err >= 0)
1733                 ub->ub_number = err;
1734
1735         return err;
1736 }
1737
1738 static void ublk_free_dev_number(struct ublk_device *ub)
1739 {
1740         spin_lock(&ublk_idr_lock);
1741         idr_remove(&ublk_index_idr, ub->ub_number);
1742         wake_up_all(&ublk_idr_wq);
1743         spin_unlock(&ublk_idr_lock);
1744 }
1745
1746 static void ublk_cdev_rel(struct device *dev)
1747 {
1748         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1749
1750         blk_mq_free_tag_set(&ub->tag_set);
1751         ublk_deinit_queues(ub);
1752         ublk_free_dev_number(ub);
1753         mutex_destroy(&ub->mutex);
1754         kfree(ub);
1755 }
1756
1757 static int ublk_add_chdev(struct ublk_device *ub)
1758 {
1759         struct device *dev = &ub->cdev_dev;
1760         int minor = ub->ub_number;
1761         int ret;
1762
1763         dev->parent = ublk_misc.this_device;
1764         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1765         dev->class = &ublk_chr_class;
1766         dev->release = ublk_cdev_rel;
1767         device_initialize(dev);
1768
1769         ret = dev_set_name(dev, "ublkc%d", minor);
1770         if (ret)
1771                 goto fail;
1772
1773         cdev_init(&ub->cdev, &ublk_ch_fops);
1774         ret = cdev_device_add(&ub->cdev, dev);
1775         if (ret)
1776                 goto fail;
1777
1778         ublks_added++;
1779         return 0;
1780  fail:
1781         put_device(dev);
1782         return ret;
1783 }
1784
1785 static void ublk_stop_work_fn(struct work_struct *work)
1786 {
1787         struct ublk_device *ub =
1788                 container_of(work, struct ublk_device, stop_work);
1789
1790         ublk_stop_dev(ub);
1791 }
1792
1793 /* align max io buffer size with PAGE_SIZE */
1794 static void ublk_align_max_io_size(struct ublk_device *ub)
1795 {
1796         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1797
1798         ub->dev_info.max_io_buf_bytes =
1799                 round_down(max_io_bytes, PAGE_SIZE);
1800 }
1801
1802 static int ublk_add_tag_set(struct ublk_device *ub)
1803 {
1804         ub->tag_set.ops = &ublk_mq_ops;
1805         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1806         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1807         ub->tag_set.numa_node = NUMA_NO_NODE;
1808         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1809         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1810         ub->tag_set.driver_data = ub;
1811         return blk_mq_alloc_tag_set(&ub->tag_set);
1812 }
1813
1814 static void ublk_remove(struct ublk_device *ub)
1815 {
1816         ublk_stop_dev(ub);
1817         cancel_work_sync(&ub->stop_work);
1818         cancel_work_sync(&ub->quiesce_work);
1819         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1820         put_device(&ub->cdev_dev);
1821         ublks_added--;
1822 }
1823
1824 static struct ublk_device *ublk_get_device_from_id(int idx)
1825 {
1826         struct ublk_device *ub = NULL;
1827
1828         if (idx < 0)
1829                 return NULL;
1830
1831         spin_lock(&ublk_idr_lock);
1832         ub = idr_find(&ublk_index_idr, idx);
1833         if (ub)
1834                 ub = ublk_get_device(ub);
1835         spin_unlock(&ublk_idr_lock);
1836
1837         return ub;
1838 }
1839
1840 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1841 {
1842         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1843         int ublksrv_pid = (int)header->data[0];
1844         struct gendisk *disk;
1845         int ret = -EINVAL;
1846
1847         if (ublksrv_pid <= 0)
1848                 return -EINVAL;
1849
1850         if (wait_for_completion_interruptible(&ub->completion) != 0)
1851                 return -EINTR;
1852
1853         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1854
1855         mutex_lock(&ub->mutex);
1856         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1857             test_bit(UB_STATE_USED, &ub->state)) {
1858                 ret = -EEXIST;
1859                 goto out_unlock;
1860         }
1861
1862         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1863         if (IS_ERR(disk)) {
1864                 ret = PTR_ERR(disk);
1865                 goto out_unlock;
1866         }
1867         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1868         disk->fops = &ub_fops;
1869         disk->private_data = ub;
1870
1871         ub->dev_info.ublksrv_pid = ublksrv_pid;
1872         ub->ub_disk = disk;
1873
1874         ret = ublk_apply_params(ub);
1875         if (ret)
1876                 goto out_put_disk;
1877
1878         /* don't probe partitions if any one ubq daemon is un-trusted */
1879         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1880                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1881
1882         get_device(&ub->cdev_dev);
1883         ub->dev_info.state = UBLK_S_DEV_LIVE;
1884         ret = add_disk(disk);
1885         if (ret) {
1886                 /*
1887                  * Has to drop the reference since ->free_disk won't be
1888                  * called in case of add_disk failure.
1889                  */
1890                 ub->dev_info.state = UBLK_S_DEV_DEAD;
1891                 ublk_put_device(ub);
1892                 goto out_put_disk;
1893         }
1894         set_bit(UB_STATE_USED, &ub->state);
1895 out_put_disk:
1896         if (ret)
1897                 put_disk(disk);
1898 out_unlock:
1899         mutex_unlock(&ub->mutex);
1900         return ret;
1901 }
1902
1903 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1904                 struct io_uring_cmd *cmd)
1905 {
1906         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1907         void __user *argp = (void __user *)(unsigned long)header->addr;
1908         cpumask_var_t cpumask;
1909         unsigned long queue;
1910         unsigned int retlen;
1911         unsigned int i;
1912         int ret;
1913
1914         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1915                 return -EINVAL;
1916         if (header->len & (sizeof(unsigned long)-1))
1917                 return -EINVAL;
1918         if (!header->addr)
1919                 return -EINVAL;
1920
1921         queue = header->data[0];
1922         if (queue >= ub->dev_info.nr_hw_queues)
1923                 return -EINVAL;
1924
1925         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1926                 return -ENOMEM;
1927
1928         for_each_possible_cpu(i) {
1929                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1930                         cpumask_set_cpu(i, cpumask);
1931         }
1932
1933         ret = -EFAULT;
1934         retlen = min_t(unsigned short, header->len, cpumask_size());
1935         if (copy_to_user(argp, cpumask, retlen))
1936                 goto out_free_cpumask;
1937         if (retlen != header->len &&
1938             clear_user(argp + retlen, header->len - retlen))
1939                 goto out_free_cpumask;
1940
1941         ret = 0;
1942 out_free_cpumask:
1943         free_cpumask_var(cpumask);
1944         return ret;
1945 }
1946
1947 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1948 {
1949         pr_devel("%s: dev id %d flags %llx\n", __func__,
1950                         info->dev_id, info->flags);
1951         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1952                         info->nr_hw_queues, info->queue_depth);
1953 }
1954
1955 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1956 {
1957         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1958         void __user *argp = (void __user *)(unsigned long)header->addr;
1959         struct ublksrv_ctrl_dev_info info;
1960         struct ublk_device *ub;
1961         int ret = -EINVAL;
1962
1963         if (header->len < sizeof(info) || !header->addr)
1964                 return -EINVAL;
1965         if (header->queue_id != (u16)-1) {
1966                 pr_warn("%s: queue_id is wrong %x\n",
1967                         __func__, header->queue_id);
1968                 return -EINVAL;
1969         }
1970
1971         if (copy_from_user(&info, argp, sizeof(info)))
1972                 return -EFAULT;
1973
1974         if (capable(CAP_SYS_ADMIN))
1975                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1976         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1977                 return -EPERM;
1978
1979         /*
1980          * unprivileged device can't be trusted, but RECOVERY and
1981          * RECOVERY_REISSUE still may hang error handling, so can't
1982          * support recovery features for unprivileged ublk now
1983          *
1984          * TODO: provide forward progress for RECOVERY handler, so that
1985          * unprivileged device can benefit from it
1986          */
1987         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1988                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1989                                 UBLK_F_USER_RECOVERY);
1990
1991         /* the created device is always owned by current user */
1992         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1993
1994         if (header->dev_id != info.dev_id) {
1995                 pr_warn("%s: dev id not match %u %u\n",
1996                         __func__, header->dev_id, info.dev_id);
1997                 return -EINVAL;
1998         }
1999
2000         ublk_dump_dev_info(&info);
2001
2002         ret = mutex_lock_killable(&ublk_ctl_mutex);
2003         if (ret)
2004                 return ret;
2005
2006         ret = -EACCES;
2007         if (ublks_added >= ublks_max)
2008                 goto out_unlock;
2009
2010         ret = -ENOMEM;
2011         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2012         if (!ub)
2013                 goto out_unlock;
2014         mutex_init(&ub->mutex);
2015         spin_lock_init(&ub->mm_lock);
2016         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2017         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2018         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
2019
2020         ret = ublk_alloc_dev_number(ub, header->dev_id);
2021         if (ret < 0)
2022                 goto out_free_ub;
2023
2024         memcpy(&ub->dev_info, &info, sizeof(info));
2025
2026         /* update device id */
2027         ub->dev_info.dev_id = ub->ub_number;
2028
2029         /*
2030          * 64bit flags will be copied back to userspace as feature
2031          * negotiation result, so have to clear flags which driver
2032          * doesn't support yet, then userspace can get correct flags
2033          * (features) to handle.
2034          */
2035         ub->dev_info.flags &= UBLK_F_ALL;
2036
2037         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2038                 UBLK_F_URING_CMD_COMP_IN_TASK;
2039
2040         /* GET_DATA isn't needed any more with USER_COPY */
2041         if (ub->dev_info.flags & UBLK_F_USER_COPY)
2042                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2043
2044         /* We are not ready to support zero copy */
2045         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2046
2047         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2048                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2049         ublk_align_max_io_size(ub);
2050
2051         ret = ublk_init_queues(ub);
2052         if (ret)
2053                 goto out_free_dev_number;
2054
2055         ret = ublk_add_tag_set(ub);
2056         if (ret)
2057                 goto out_deinit_queues;
2058
2059         ret = -EFAULT;
2060         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2061                 goto out_free_tag_set;
2062
2063         /*
2064          * Add the char dev so that ublksrv daemon can be setup.
2065          * ublk_add_chdev() will cleanup everything if it fails.
2066          */
2067         ret = ublk_add_chdev(ub);
2068         goto out_unlock;
2069
2070 out_free_tag_set:
2071         blk_mq_free_tag_set(&ub->tag_set);
2072 out_deinit_queues:
2073         ublk_deinit_queues(ub);
2074 out_free_dev_number:
2075         ublk_free_dev_number(ub);
2076 out_free_ub:
2077         mutex_destroy(&ub->mutex);
2078         kfree(ub);
2079 out_unlock:
2080         mutex_unlock(&ublk_ctl_mutex);
2081         return ret;
2082 }
2083
2084 static inline bool ublk_idr_freed(int id)
2085 {
2086         void *ptr;
2087
2088         spin_lock(&ublk_idr_lock);
2089         ptr = idr_find(&ublk_index_idr, id);
2090         spin_unlock(&ublk_idr_lock);
2091
2092         return ptr == NULL;
2093 }
2094
2095 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2096 {
2097         struct ublk_device *ub = *p_ub;
2098         int idx = ub->ub_number;
2099         int ret;
2100
2101         ret = mutex_lock_killable(&ublk_ctl_mutex);
2102         if (ret)
2103                 return ret;
2104
2105         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2106                 ublk_remove(ub);
2107                 set_bit(UB_STATE_DELETED, &ub->state);
2108         }
2109
2110         /* Mark the reference as consumed */
2111         *p_ub = NULL;
2112         ublk_put_device(ub);
2113         mutex_unlock(&ublk_ctl_mutex);
2114
2115         /*
2116          * Wait until the idr is removed, then it can be reused after
2117          * DEL_DEV command is returned.
2118          *
2119          * If we returns because of user interrupt, future delete command
2120          * may come:
2121          *
2122          * - the device number isn't freed, this device won't or needn't
2123          *   be deleted again, since UB_STATE_DELETED is set, and device
2124          *   will be released after the last reference is dropped
2125          *
2126          * - the device number is freed already, we will not find this
2127          *   device via ublk_get_device_from_id()
2128          */
2129         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
2130
2131         return 0;
2132 }
2133
2134 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2135 {
2136         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2137
2138         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2139                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2140                         header->data[0], header->addr, header->len);
2141 }
2142
2143 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2144 {
2145         ublk_stop_dev(ub);
2146         cancel_work_sync(&ub->stop_work);
2147         cancel_work_sync(&ub->quiesce_work);
2148
2149         return 0;
2150 }
2151
2152 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2153                 struct io_uring_cmd *cmd)
2154 {
2155         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2156         void __user *argp = (void __user *)(unsigned long)header->addr;
2157
2158         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2159                 return -EINVAL;
2160
2161         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2162                 return -EFAULT;
2163
2164         return 0;
2165 }
2166
2167 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2168 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2169 {
2170         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2171         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2172
2173         if (ub->ub_disk) {
2174                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2175                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2176         } else {
2177                 ub->params.devt.disk_major = 0;
2178                 ub->params.devt.disk_minor = 0;
2179         }
2180         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2181 }
2182
2183 static int ublk_ctrl_get_params(struct ublk_device *ub,
2184                 struct io_uring_cmd *cmd)
2185 {
2186         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2187         void __user *argp = (void __user *)(unsigned long)header->addr;
2188         struct ublk_params_header ph;
2189         int ret;
2190
2191         if (header->len <= sizeof(ph) || !header->addr)
2192                 return -EINVAL;
2193
2194         if (copy_from_user(&ph, argp, sizeof(ph)))
2195                 return -EFAULT;
2196
2197         if (ph.len > header->len || !ph.len)
2198                 return -EINVAL;
2199
2200         if (ph.len > sizeof(struct ublk_params))
2201                 ph.len = sizeof(struct ublk_params);
2202
2203         mutex_lock(&ub->mutex);
2204         ublk_ctrl_fill_params_devt(ub);
2205         if (copy_to_user(argp, &ub->params, ph.len))
2206                 ret = -EFAULT;
2207         else
2208                 ret = 0;
2209         mutex_unlock(&ub->mutex);
2210
2211         return ret;
2212 }
2213
2214 static int ublk_ctrl_set_params(struct ublk_device *ub,
2215                 struct io_uring_cmd *cmd)
2216 {
2217         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2218         void __user *argp = (void __user *)(unsigned long)header->addr;
2219         struct ublk_params_header ph;
2220         int ret = -EFAULT;
2221
2222         if (header->len <= sizeof(ph) || !header->addr)
2223                 return -EINVAL;
2224
2225         if (copy_from_user(&ph, argp, sizeof(ph)))
2226                 return -EFAULT;
2227
2228         if (ph.len > header->len || !ph.len || !ph.types)
2229                 return -EINVAL;
2230
2231         if (ph.len > sizeof(struct ublk_params))
2232                 ph.len = sizeof(struct ublk_params);
2233
2234         /* parameters can only be changed when device isn't live */
2235         mutex_lock(&ub->mutex);
2236         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2237                 ret = -EACCES;
2238         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2239                 ret = -EFAULT;
2240         } else {
2241                 /* clear all we don't support yet */
2242                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2243                 ret = ublk_validate_params(ub);
2244                 if (ret)
2245                         ub->params.types = 0;
2246         }
2247         mutex_unlock(&ub->mutex);
2248
2249         return ret;
2250 }
2251
2252 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2253 {
2254         int i;
2255
2256         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2257         /* All old ioucmds have to be completed */
2258         WARN_ON_ONCE(ubq->nr_io_ready);
2259         /* old daemon is PF_EXITING, put it now */
2260         put_task_struct(ubq->ubq_daemon);
2261         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2262         ubq->ubq_daemon = NULL;
2263         ubq->timeout = false;
2264
2265         for (i = 0; i < ubq->q_depth; i++) {
2266                 struct ublk_io *io = &ubq->ios[i];
2267
2268                 /* forget everything now and be ready for new FETCH_REQ */
2269                 io->flags = 0;
2270                 io->cmd = NULL;
2271                 io->addr = 0;
2272         }
2273 }
2274
2275 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2276                 struct io_uring_cmd *cmd)
2277 {
2278         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2279         int ret = -EINVAL;
2280         int i;
2281
2282         mutex_lock(&ub->mutex);
2283         if (!ublk_can_use_recovery(ub))
2284                 goto out_unlock;
2285         /*
2286          * START_RECOVERY is only allowd after:
2287          *
2288          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2289          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2290          *     released.
2291          *
2292          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2293          *     (a)has quiesced request queue
2294          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2295          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2296          *     (d)has completed/camceled all ioucmds owned by ther dying process
2297          */
2298         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2299                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2300                 ret = -EBUSY;
2301                 goto out_unlock;
2302         }
2303         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2304         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2305                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2306         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2307         ub->mm = NULL;
2308         ub->nr_queues_ready = 0;
2309         ub->nr_privileged_daemon = 0;
2310         init_completion(&ub->completion);
2311         ret = 0;
2312  out_unlock:
2313         mutex_unlock(&ub->mutex);
2314         return ret;
2315 }
2316
2317 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2318                 struct io_uring_cmd *cmd)
2319 {
2320         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2321         int ublksrv_pid = (int)header->data[0];
2322         int ret = -EINVAL;
2323
2324         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2325                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2326         /* wait until new ubq_daemon sending all FETCH_REQ */
2327         if (wait_for_completion_interruptible(&ub->completion))
2328                 return -EINTR;
2329
2330         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2331                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2332
2333         mutex_lock(&ub->mutex);
2334         if (!ublk_can_use_recovery(ub))
2335                 goto out_unlock;
2336
2337         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2338                 ret = -EBUSY;
2339                 goto out_unlock;
2340         }
2341         ub->dev_info.ublksrv_pid = ublksrv_pid;
2342         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2343                         __func__, ublksrv_pid, header->dev_id);
2344         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2345         pr_devel("%s: queue unquiesced, dev id %d.\n",
2346                         __func__, header->dev_id);
2347         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2348         ub->dev_info.state = UBLK_S_DEV_LIVE;
2349         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2350         ret = 0;
2351  out_unlock:
2352         mutex_unlock(&ub->mutex);
2353         return ret;
2354 }
2355
2356 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2357 {
2358         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2359         void __user *argp = (void __user *)(unsigned long)header->addr;
2360         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2361
2362         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2363                 return -EINVAL;
2364
2365         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2366                 return -EFAULT;
2367
2368         return 0;
2369 }
2370
2371 /*
2372  * All control commands are sent via /dev/ublk-control, so we have to check
2373  * the destination device's permission
2374  */
2375 static int ublk_char_dev_permission(struct ublk_device *ub,
2376                 const char *dev_path, int mask)
2377 {
2378         int err;
2379         struct path path;
2380         struct kstat stat;
2381
2382         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2383         if (err)
2384                 return err;
2385
2386         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2387         if (err)
2388                 goto exit;
2389
2390         err = -EPERM;
2391         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2392                 goto exit;
2393
2394         err = inode_permission(&nop_mnt_idmap,
2395                         d_backing_inode(path.dentry), mask);
2396 exit:
2397         path_put(&path);
2398         return err;
2399 }
2400
2401 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2402                 struct io_uring_cmd *cmd)
2403 {
2404         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2405         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2406         void __user *argp = (void __user *)(unsigned long)header->addr;
2407         char *dev_path = NULL;
2408         int ret = 0;
2409         int mask;
2410
2411         if (!unprivileged) {
2412                 if (!capable(CAP_SYS_ADMIN))
2413                         return -EPERM;
2414                 /*
2415                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2416                  * char_dev_path in payload too, since userspace may not
2417                  * know if the specified device is created as unprivileged
2418                  * mode.
2419                  */
2420                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2421                         return 0;
2422         }
2423
2424         /*
2425          * User has to provide the char device path for unprivileged ublk
2426          *
2427          * header->addr always points to the dev path buffer, and
2428          * header->dev_path_len records length of dev path buffer.
2429          */
2430         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2431                 return -EINVAL;
2432
2433         if (header->len < header->dev_path_len)
2434                 return -EINVAL;
2435
2436         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2437         if (!dev_path)
2438                 return -ENOMEM;
2439
2440         ret = -EFAULT;
2441         if (copy_from_user(dev_path, argp, header->dev_path_len))
2442                 goto exit;
2443         dev_path[header->dev_path_len] = 0;
2444
2445         ret = -EINVAL;
2446         switch (_IOC_NR(cmd->cmd_op)) {
2447         case UBLK_CMD_GET_DEV_INFO:
2448         case UBLK_CMD_GET_DEV_INFO2:
2449         case UBLK_CMD_GET_QUEUE_AFFINITY:
2450         case UBLK_CMD_GET_PARAMS:
2451         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2452                 mask = MAY_READ;
2453                 break;
2454         case UBLK_CMD_START_DEV:
2455         case UBLK_CMD_STOP_DEV:
2456         case UBLK_CMD_ADD_DEV:
2457         case UBLK_CMD_DEL_DEV:
2458         case UBLK_CMD_SET_PARAMS:
2459         case UBLK_CMD_START_USER_RECOVERY:
2460         case UBLK_CMD_END_USER_RECOVERY:
2461                 mask = MAY_READ | MAY_WRITE;
2462                 break;
2463         default:
2464                 goto exit;
2465         }
2466
2467         ret = ublk_char_dev_permission(ub, dev_path, mask);
2468         if (!ret) {
2469                 header->len -= header->dev_path_len;
2470                 header->addr += header->dev_path_len;
2471         }
2472         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2473                         __func__, ub->ub_number, cmd->cmd_op,
2474                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2475                         dev_path, ret);
2476 exit:
2477         kfree(dev_path);
2478         return ret;
2479 }
2480
2481 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2482                 unsigned int issue_flags)
2483 {
2484         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2485         struct ublk_device *ub = NULL;
2486         u32 cmd_op = cmd->cmd_op;
2487         int ret = -EINVAL;
2488
2489         if (issue_flags & IO_URING_F_NONBLOCK)
2490                 return -EAGAIN;
2491
2492         ublk_ctrl_cmd_dump(cmd);
2493
2494         if (!(issue_flags & IO_URING_F_SQE128))
2495                 goto out;
2496
2497         ret = ublk_check_cmd_op(cmd_op);
2498         if (ret)
2499                 goto out;
2500
2501         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2502                 ret = ublk_ctrl_get_features(cmd);
2503                 goto out;
2504         }
2505
2506         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2507                 ret = -ENODEV;
2508                 ub = ublk_get_device_from_id(header->dev_id);
2509                 if (!ub)
2510                         goto out;
2511
2512                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2513                 if (ret)
2514                         goto put_dev;
2515         }
2516
2517         switch (_IOC_NR(cmd_op)) {
2518         case UBLK_CMD_START_DEV:
2519                 ret = ublk_ctrl_start_dev(ub, cmd);
2520                 break;
2521         case UBLK_CMD_STOP_DEV:
2522                 ret = ublk_ctrl_stop_dev(ub);
2523                 break;
2524         case UBLK_CMD_GET_DEV_INFO:
2525         case UBLK_CMD_GET_DEV_INFO2:
2526                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2527                 break;
2528         case UBLK_CMD_ADD_DEV:
2529                 ret = ublk_ctrl_add_dev(cmd);
2530                 break;
2531         case UBLK_CMD_DEL_DEV:
2532                 ret = ublk_ctrl_del_dev(&ub);
2533                 break;
2534         case UBLK_CMD_GET_QUEUE_AFFINITY:
2535                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2536                 break;
2537         case UBLK_CMD_GET_PARAMS:
2538                 ret = ublk_ctrl_get_params(ub, cmd);
2539                 break;
2540         case UBLK_CMD_SET_PARAMS:
2541                 ret = ublk_ctrl_set_params(ub, cmd);
2542                 break;
2543         case UBLK_CMD_START_USER_RECOVERY:
2544                 ret = ublk_ctrl_start_recovery(ub, cmd);
2545                 break;
2546         case UBLK_CMD_END_USER_RECOVERY:
2547                 ret = ublk_ctrl_end_recovery(ub, cmd);
2548                 break;
2549         default:
2550                 ret = -ENOTSUPP;
2551                 break;
2552         }
2553
2554  put_dev:
2555         if (ub)
2556                 ublk_put_device(ub);
2557  out:
2558         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2559         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2560                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2561         return -EIOCBQUEUED;
2562 }
2563
2564 static const struct file_operations ublk_ctl_fops = {
2565         .open           = nonseekable_open,
2566         .uring_cmd      = ublk_ctrl_uring_cmd,
2567         .owner          = THIS_MODULE,
2568         .llseek         = noop_llseek,
2569 };
2570
2571 static struct miscdevice ublk_misc = {
2572         .minor          = MISC_DYNAMIC_MINOR,
2573         .name           = "ublk-control",
2574         .fops           = &ublk_ctl_fops,
2575 };
2576
2577 static int __init ublk_init(void)
2578 {
2579         int ret;
2580
2581         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2582                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2583
2584         init_waitqueue_head(&ublk_idr_wq);
2585
2586         ret = misc_register(&ublk_misc);
2587         if (ret)
2588                 return ret;
2589
2590         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2591         if (ret)
2592                 goto unregister_mis;
2593
2594         ret = class_register(&ublk_chr_class);
2595         if (ret)
2596                 goto free_chrdev_region;
2597
2598         return 0;
2599
2600 free_chrdev_region:
2601         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2602 unregister_mis:
2603         misc_deregister(&ublk_misc);
2604         return ret;
2605 }
2606
2607 static void __exit ublk_exit(void)
2608 {
2609         struct ublk_device *ub;
2610         int id;
2611
2612         idr_for_each_entry(&ublk_index_idr, ub, id)
2613                 ublk_remove(ub);
2614
2615         class_unregister(&ublk_chr_class);
2616         misc_deregister(&ublk_misc);
2617
2618         idr_destroy(&ublk_index_idr);
2619         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2620 }
2621
2622 module_init(ublk_init);
2623 module_exit(ublk_exit);
2624
2625 module_param(ublks_max, int, 0444);
2626 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2627
2628 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2629 MODULE_LICENSE("GPL");