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