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)
 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 {
 720         /* mark this cmd owned by ublksrv */
 721         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 722
 723         /*
 724          * clear ACTIVE since we are done with this sqe/cmd slot
 725          * We can only accept io cmd in case of being not active.
 726          */
 727         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 728
 729         /* tell ublksrv one io request is coming */
 730         io_uring_cmd_done(io->cmd, res, 0);
 731 }
 732
 733 #define UBLK_REQUEUE_DELAY_MS   3
 734
 735 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 736                 struct request *rq)
 737 {
 738         /* We cannot process this rq so just requeue it. */
 739         if (ublk_queue_can_use_recovery(ubq))
 740                 blk_mq_requeue_request(rq, false);
 741         else
 742                 blk_mq_end_request(rq, BLK_STS_IOERR);
 743
 744         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 745 }
 746
 747 static inline void __ublk_rq_task_work(struct request *req)
 748 {
 749         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 750         int tag = req->tag;
 751         struct ublk_io *io = &ubq->ios[tag];
 752         unsigned int mapped_bytes;
 753
 754         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 755                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 756                         ublk_get_iod(ubq, req->tag)->addr);
 757
 758         /*
 759          * Task is exiting if either:
 760          *
 761          * (1) current != ubq_daemon.
 762          * io_uring_cmd_complete_in_task() tries to run task_work
 763          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 764          *
 765          * (2) current->flags & PF_EXITING.
 766          */
 767         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 768                 __ublk_abort_rq(ubq, req);
 769                 return;
 770         }
 771
 772         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
 773                 /*
 774                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 775                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 776                  * and notify it.
 777                  */
 778                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 779                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 780                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 781                                         __func__, io->cmd->cmd_op, ubq->q_id,
 782                                         req->tag, io->flags);
 783                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
 784                         return;
 785                 }
 786                 /*
 787                  * We have handled UBLK_IO_NEED_GET_DATA command,
 788                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 789                  * do the copy work.
 790                  */
 791                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 792                 /* update iod->addr because ublksrv may have passed a new io buffer */
 793                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 794                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 795                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 796                                 ublk_get_iod(ubq, req->tag)->addr);
 797         }
 798
 799         mapped_bytes = ublk_map_io(ubq, req, io);
 800
 801         /* partially mapped, update io descriptor */
 802         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 803                 /*
 804                  * Nothing mapped, retry until we succeed.
 805                  *
 806                  * We may never succeed in mapping any bytes here because
 807                  * of OOM. TODO: reserve one buffer with single page pinned
 808                  * for providing forward progress guarantee.
 809                  */
 810                 if (unlikely(!mapped_bytes)) {
 811                         blk_mq_requeue_request(req, false);
 812                         blk_mq_delay_kick_requeue_list(req->q,
 813                                         UBLK_REQUEUE_DELAY_MS);
 814                         return;
 815                 }
 816
 817                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 818                         mapped_bytes >> 9;
 819         }
 820
 821         ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
 822 }
 823
 824 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
 825 {
 826         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 827         struct ublk_rq_data *data, *tmp;
 828
 829         io_cmds = llist_reverse_order(io_cmds);
 830         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 831                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
 832 }
 833
 834 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 835 {
 836         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 837         struct ublk_rq_data *data, *tmp;
 838
 839         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 840                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 841 }
 842
 843 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
 844 {
 845         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 846         struct ublk_queue *ubq = pdu->ubq;
 847
 848         ublk_forward_io_cmds(ubq);
 849 }
 850
 851 static void ublk_rq_task_work_fn(struct callback_head *work)
 852 {
 853         struct ublk_rq_data *data = container_of(work,
 854                         struct ublk_rq_data, work);
 855         struct request *req = blk_mq_rq_from_pdu(data);
 856         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 857
 858         ublk_forward_io_cmds(ubq);
 859 }
 860
 861 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 862 {
 863         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 864         struct ublk_io *io;
 865
 866         if (!llist_add(&data->node, &ubq->io_cmds))
 867                 return;
 868
 869         io = &ubq->ios[rq->tag];
 870         /*
 871          * If the check pass, we know that this is a re-issued request aborted
 872          * previously in monitor_work because the ubq_daemon(cmd's task) is
 873          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 874          * because this ioucmd's io_uring context may be freed now if no inflight
 875          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 876          *
 877          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 878          * the tag). Then the request is re-started(allocating the tag) and we are here.
 879          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 880          * guarantees that here is a re-issued request aborted previously.
 881          */
 882         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 883                 ublk_abort_io_cmds(ubq);
 884         } else if (ublk_can_use_task_work(ubq)) {
 885                 if (task_work_add(ubq->ubq_daemon, &data->work,
 886                                         TWA_SIGNAL_NO_IPI))
 887                         ublk_abort_io_cmds(ubq);
 888         } else {
 889                 struct io_uring_cmd *cmd = io->cmd;
 890                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 891
 892                 pdu->ubq = ubq;
 893                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 894         }
 895 }
 896
 897 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 898                 const struct blk_mq_queue_data *bd)
 899 {
 900         struct ublk_queue *ubq = hctx->driver_data;
 901         struct request *rq = bd->rq;
 902         blk_status_t res;
 903
 904         /* fill iod to slot in io cmd buffer */
 905         res = ublk_setup_iod(ubq, rq);
 906         if (unlikely(res != BLK_STS_OK))
 907                 return BLK_STS_IOERR;
 908
 909         /* With recovery feature enabled, force_abort is set in
 910          * ublk_stop_dev() before calling del_gendisk(). We have to
 911          * abort all requeued and new rqs here to let del_gendisk()
 912          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 913          * to avoid UAF on io_uring ctx.
 914          *
 915          * Note: force_abort is guaranteed to be seen because it is set
 916          * before request queue is unqiuesced.
 917          */
 918         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 919                 return BLK_STS_IOERR;
 920
 921         blk_mq_start_request(bd->rq);
 922
 923         if (unlikely(ubq_daemon_is_dying(ubq))) {
 924                 __ublk_abort_rq(ubq, rq);
 925                 return BLK_STS_OK;
 926         }
 927
 928         ublk_queue_cmd(ubq, rq);
 929
 930         return BLK_STS_OK;
 931 }
 932
 933 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 934                 unsigned int hctx_idx)
 935 {
 936         struct ublk_device *ub = driver_data;
 937         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 938
 939         hctx->driver_data = ubq;
 940         return 0;
 941 }
 942
 943 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 944                 unsigned int hctx_idx, unsigned int numa_node)
 945 {
 946         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 947
 948         init_task_work(&data->work, ublk_rq_task_work_fn);
 949         return 0;
 950 }
 951
 952 static const struct blk_mq_ops ublk_mq_ops = {
 953         .queue_rq       = ublk_queue_rq,
 954         .init_hctx      = ublk_init_hctx,
 955         .init_request   = ublk_init_rq,
 956 };
 957
 958 static int ublk_ch_open(struct inode *inode, struct file *filp)
 959 {
 960         struct ublk_device *ub = container_of(inode->i_cdev,
 961                         struct ublk_device, cdev);
 962
 963         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 964                 return -EBUSY;
 965         filp->private_data = ub;
 966         return 0;
 967 }
 968
 969 static int ublk_ch_release(struct inode *inode, struct file *filp)
 970 {
 971         struct ublk_device *ub = filp->private_data;
 972
 973         clear_bit(UB_STATE_OPEN, &ub->state);
 974         return 0;
 975 }
 976
 977 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
 978 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
 979 {
 980         struct ublk_device *ub = filp->private_data;
 981         size_t sz = vma->vm_end - vma->vm_start;
 982         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
 983         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
 984         int q_id, ret = 0;
 985
 986         spin_lock(&ub->mm_lock);
 987         if (!ub->mm)
 988                 ub->mm = current->mm;
 989         if (current->mm != ub->mm)
 990                 ret = -EINVAL;
 991         spin_unlock(&ub->mm_lock);
 992
 993         if (ret)
 994                 return ret;
 995
 996         if (vma->vm_flags & VM_WRITE)
 997                 return -EPERM;
 998
 999         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1000         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1001                 return -EINVAL;
1002
1003         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1004         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1005                         __func__, q_id, current->pid, vma->vm_start,
1006                         phys_off, (unsigned long)sz);
1007
1008         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1009                 return -EINVAL;
1010
1011         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1012         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1013 }
1014
1015 static void ublk_commit_completion(struct ublk_device *ub,
1016                 struct ublksrv_io_cmd *ub_cmd)
1017 {
1018         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1019         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1020         struct ublk_io *io = &ubq->ios[tag];
1021         struct request *req;
1022
1023         /* now this cmd slot is owned by nbd driver */
1024         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1025         io->res = ub_cmd->result;
1026
1027         /* find the io request and complete */
1028         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1029
1030         if (req && likely(!blk_should_fake_timeout(req->q)))
1031                 ublk_complete_rq(req);
1032 }
1033
1034 /*
1035  * When ->ubq_daemon is exiting, either new request is ended immediately,
1036  * or any queued io command is drained, so it is safe to abort queue
1037  * lockless
1038  */
1039 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1040 {
1041         int i;
1042
1043         if (!ublk_get_device(ub))
1044                 return;
1045
1046         for (i = 0; i < ubq->q_depth; i++) {
1047                 struct ublk_io *io = &ubq->ios[i];
1048
1049                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1050                         struct request *rq;
1051
1052                         /*
1053                          * Either we fail the request or ublk_rq_task_work_fn
1054                          * will do it
1055                          */
1056                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1057                         if (rq)
1058                                 __ublk_fail_req(ubq, io, rq);
1059                 }
1060         }
1061         ublk_put_device(ub);
1062 }
1063
1064 static void ublk_daemon_monitor_work(struct work_struct *work)
1065 {
1066         struct ublk_device *ub =
1067                 container_of(work, struct ublk_device, monitor_work.work);
1068         int i;
1069
1070         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1071                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1072
1073                 if (ubq_daemon_is_dying(ubq)) {
1074                         if (ublk_queue_can_use_recovery(ubq))
1075                                 schedule_work(&ub->quiesce_work);
1076                         else
1077                                 schedule_work(&ub->stop_work);
1078
1079                         /* abort queue is for making forward progress */
1080                         ublk_abort_queue(ub, ubq);
1081                 }
1082         }
1083
1084         /*
1085          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1086          * after ublk_remove() or __ublk_quiesce_dev() is started.
1087          *
1088          * No need ub->mutex, monitor work are canceled after state is marked
1089          * as not LIVE, so new state is observed reliably.
1090          */
1091         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1092                 schedule_delayed_work(&ub->monitor_work,
1093                                 UBLK_DAEMON_MONITOR_PERIOD);
1094 }
1095
1096 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1097 {
1098         return ubq->nr_io_ready == ubq->q_depth;
1099 }
1100
1101 static void ublk_cancel_queue(struct ublk_queue *ubq)
1102 {
1103         int i;
1104
1105         if (!ublk_queue_ready(ubq))
1106                 return;
1107
1108         for (i = 0; i < ubq->q_depth; i++) {
1109                 struct ublk_io *io = &ubq->ios[i];
1110
1111                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1112                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1113         }
1114
1115         /* all io commands are canceled */
1116         ubq->nr_io_ready = 0;
1117 }
1118
1119 /* Cancel all pending commands, must be called after del_gendisk() returns */
1120 static void ublk_cancel_dev(struct ublk_device *ub)
1121 {
1122         int i;
1123
1124         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1125                 ublk_cancel_queue(ublk_get_queue(ub, i));
1126 }
1127
1128 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1129 {
1130         bool *idle = data;
1131
1132         if (blk_mq_request_started(rq)) {
1133                 *idle = false;
1134                 return false;
1135         }
1136         return true;
1137 }
1138
1139 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1140 {
1141         bool idle;
1142
1143         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1144         while (true) {
1145                 idle = true;
1146                 blk_mq_tagset_busy_iter(&ub->tag_set,
1147                                 ublk_check_inflight_rq, &idle);
1148                 if (idle)
1149                         break;
1150                 msleep(UBLK_REQUEUE_DELAY_MS);
1151         }
1152 }
1153
1154 static void __ublk_quiesce_dev(struct ublk_device *ub)
1155 {
1156         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1157                         __func__, ub->dev_info.dev_id,
1158                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1159                         "LIVE" : "QUIESCED");
1160         blk_mq_quiesce_queue(ub->ub_disk->queue);
1161         ublk_wait_tagset_rqs_idle(ub);
1162         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1163         ublk_cancel_dev(ub);
1164         /* we are going to release task_struct of ubq_daemon and resets
1165          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1166          * Besides, monitor_work is not necessary in QUIESCED state since we have
1167          * already scheduled quiesce_work and quiesced all ubqs.
1168          *
1169          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1170          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1171          */
1172         cancel_delayed_work_sync(&ub->monitor_work);
1173 }
1174
1175 static void ublk_quiesce_work_fn(struct work_struct *work)
1176 {
1177         struct ublk_device *ub =
1178                 container_of(work, struct ublk_device, quiesce_work);
1179
1180         mutex_lock(&ub->mutex);
1181         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1182                 goto unlock;
1183         __ublk_quiesce_dev(ub);
1184  unlock:
1185         mutex_unlock(&ub->mutex);
1186 }
1187
1188 static void ublk_unquiesce_dev(struct ublk_device *ub)
1189 {
1190         int i;
1191
1192         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1193                         __func__, ub->dev_info.dev_id,
1194                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1195                         "LIVE" : "QUIESCED");
1196         /* quiesce_work has run. We let requeued rqs be aborted
1197          * before running fallback_wq. "force_abort" must be seen
1198          * after request queue is unqiuesced. Then del_gendisk()
1199          * can move on.
1200          */
1201         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1202                 ublk_get_queue(ub, i)->force_abort = true;
1203
1204         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1205         /* We may have requeued some rqs in ublk_quiesce_queue() */
1206         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1207 }
1208
1209 static void ublk_stop_dev(struct ublk_device *ub)
1210 {
1211         mutex_lock(&ub->mutex);
1212         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1213                 goto unlock;
1214         if (ublk_can_use_recovery(ub)) {
1215                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1216                         __ublk_quiesce_dev(ub);
1217                 ublk_unquiesce_dev(ub);
1218         }
1219         del_gendisk(ub->ub_disk);
1220         ub->dev_info.state = UBLK_S_DEV_DEAD;
1221         ub->dev_info.ublksrv_pid = -1;
1222         put_disk(ub->ub_disk);
1223         ub->ub_disk = NULL;
1224  unlock:
1225         ublk_cancel_dev(ub);
1226         mutex_unlock(&ub->mutex);
1227         cancel_delayed_work_sync(&ub->monitor_work);
1228 }
1229
1230 /* device can only be started after all IOs are ready */
1231 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1232 {
1233         mutex_lock(&ub->mutex);
1234         ubq->nr_io_ready++;
1235         if (ublk_queue_ready(ubq)) {
1236                 ubq->ubq_daemon = current;
1237                 get_task_struct(ubq->ubq_daemon);
1238                 ub->nr_queues_ready++;
1239
1240                 if (capable(CAP_SYS_ADMIN))
1241                         ub->nr_privileged_daemon++;
1242         }
1243         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1244                 complete_all(&ub->completion);
1245         mutex_unlock(&ub->mutex);
1246 }
1247
1248 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1249                 int tag)
1250 {
1251         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1252         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1253
1254         ublk_queue_cmd(ubq, req);
1255 }
1256
1257 static inline int ublk_check_cmd_op(u32 cmd_op)
1258 {
1259         u32 ioc_type = _IOC_TYPE(cmd_op);
1260
1261         if (IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1262                 return -EOPNOTSUPP;
1263
1264         if (ioc_type != 'u' && ioc_type != 0)
1265                 return -EOPNOTSUPP;
1266
1267         return 0;
1268 }
1269
1270 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1271 {
1272         struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1273         struct ublk_device *ub = cmd->file->private_data;
1274         struct ublk_queue *ubq;
1275         struct ublk_io *io;
1276         u32 cmd_op = cmd->cmd_op;
1277         unsigned tag = ub_cmd->tag;
1278         int ret = -EINVAL;
1279         struct request *req;
1280
1281         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1282                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1283                         ub_cmd->result);
1284
1285         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1286                 goto out;
1287
1288         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1289         if (!ubq || ub_cmd->q_id != ubq->q_id)
1290                 goto out;
1291
1292         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1293                 goto out;
1294
1295         if (tag >= ubq->q_depth)
1296                 goto out;
1297
1298         io = &ubq->ios[tag];
1299
1300         /* there is pending io cmd, something must be wrong */
1301         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1302                 ret = -EBUSY;
1303                 goto out;
1304         }
1305
1306         /*
1307          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1308          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1309          */
1310         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1311                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1312                 goto out;
1313
1314         ret = ublk_check_cmd_op(cmd_op);
1315         if (ret)
1316                 goto out;
1317
1318         ret = -EINVAL;
1319         switch (_IOC_NR(cmd_op)) {
1320         case UBLK_IO_FETCH_REQ:
1321                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1322                 if (ublk_queue_ready(ubq)) {
1323                         ret = -EBUSY;
1324                         goto out;
1325                 }
1326                 /*
1327                  * The io is being handled by server, so COMMIT_RQ is expected
1328                  * instead of FETCH_REQ
1329                  */
1330                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1331                         goto out;
1332                 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1333                 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1334                         goto out;
1335                 io->cmd = cmd;
1336                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1337                 io->addr = ub_cmd->addr;
1338
1339                 ublk_mark_io_ready(ub, ubq);
1340                 break;
1341         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1342                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1343                 /*
1344                  * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1345                  * not enabled or it is Read IO.
1346                  */
1347                 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1348                         goto out;
1349                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1350                         goto out;
1351                 io->addr = ub_cmd->addr;
1352                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1353                 io->cmd = cmd;
1354                 ublk_commit_completion(ub, ub_cmd);
1355                 break;
1356         case UBLK_IO_NEED_GET_DATA:
1357                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1358                         goto out;
1359                 io->addr = ub_cmd->addr;
1360                 io->cmd = cmd;
1361                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1362                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1363                 break;
1364         default:
1365                 goto out;
1366         }
1367         return -EIOCBQUEUED;
1368
1369  out:
1370         io_uring_cmd_done(cmd, ret, 0);
1371         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1372                         __func__, cmd_op, tag, ret, io->flags);
1373         return -EIOCBQUEUED;
1374 }
1375
1376 static const struct file_operations ublk_ch_fops = {
1377         .owner = THIS_MODULE,
1378         .open = ublk_ch_open,
1379         .release = ublk_ch_release,
1380         .llseek = no_llseek,
1381         .uring_cmd = ublk_ch_uring_cmd,
1382         .mmap = ublk_ch_mmap,
1383 };
1384
1385 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1386 {
1387         int size = ublk_queue_cmd_buf_size(ub, q_id);
1388         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1389
1390         if (ubq->ubq_daemon)
1391                 put_task_struct(ubq->ubq_daemon);
1392         if (ubq->io_cmd_buf)
1393                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1394 }
1395
1396 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1397 {
1398         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1399         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1400         void *ptr;
1401         int size;
1402
1403         ubq->flags = ub->dev_info.flags;
1404         ubq->q_id = q_id;
1405         ubq->q_depth = ub->dev_info.queue_depth;
1406         size = ublk_queue_cmd_buf_size(ub, q_id);
1407
1408         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1409         if (!ptr)
1410                 return -ENOMEM;
1411
1412         ubq->io_cmd_buf = ptr;
1413         ubq->dev = ub;
1414         return 0;
1415 }
1416
1417 static void ublk_deinit_queues(struct ublk_device *ub)
1418 {
1419         int nr_queues = ub->dev_info.nr_hw_queues;
1420         int i;
1421
1422         if (!ub->__queues)
1423                 return;
1424
1425         for (i = 0; i < nr_queues; i++)
1426                 ublk_deinit_queue(ub, i);
1427         kfree(ub->__queues);
1428 }
1429
1430 static int ublk_init_queues(struct ublk_device *ub)
1431 {
1432         int nr_queues = ub->dev_info.nr_hw_queues;
1433         int depth = ub->dev_info.queue_depth;
1434         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1435         int i, ret = -ENOMEM;
1436
1437         ub->queue_size = ubq_size;
1438         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1439         if (!ub->__queues)
1440                 return ret;
1441
1442         for (i = 0; i < nr_queues; i++) {
1443                 if (ublk_init_queue(ub, i))
1444                         goto fail;
1445         }
1446
1447         init_completion(&ub->completion);
1448         return 0;
1449
1450  fail:
1451         ublk_deinit_queues(ub);
1452         return ret;
1453 }
1454
1455 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1456 {
1457         int i = idx;
1458         int err;
1459
1460         spin_lock(&ublk_idr_lock);
1461         /* allocate id, if @id >= 0, we're requesting that specific id */
1462         if (i >= 0) {
1463                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1464                 if (err == -ENOSPC)
1465                         err = -EEXIST;
1466         } else {
1467                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1468         }
1469         spin_unlock(&ublk_idr_lock);
1470
1471         if (err >= 0)
1472                 ub->ub_number = err;
1473
1474         return err;
1475 }
1476
1477 static void ublk_free_dev_number(struct ublk_device *ub)
1478 {
1479         spin_lock(&ublk_idr_lock);
1480         idr_remove(&ublk_index_idr, ub->ub_number);
1481         wake_up_all(&ublk_idr_wq);
1482         spin_unlock(&ublk_idr_lock);
1483 }
1484
1485 static void ublk_cdev_rel(struct device *dev)
1486 {
1487         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1488
1489         blk_mq_free_tag_set(&ub->tag_set);
1490         ublk_deinit_queues(ub);
1491         ublk_free_dev_number(ub);
1492         mutex_destroy(&ub->mutex);
1493         kfree(ub);
1494 }
1495
1496 static int ublk_add_chdev(struct ublk_device *ub)
1497 {
1498         struct device *dev = &ub->cdev_dev;
1499         int minor = ub->ub_number;
1500         int ret;
1501
1502         dev->parent = ublk_misc.this_device;
1503         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1504         dev->class = ublk_chr_class;
1505         dev->release = ublk_cdev_rel;
1506         device_initialize(dev);
1507
1508         ret = dev_set_name(dev, "ublkc%d", minor);
1509         if (ret)
1510                 goto fail;
1511
1512         cdev_init(&ub->cdev, &ublk_ch_fops);
1513         ret = cdev_device_add(&ub->cdev, dev);
1514         if (ret)
1515                 goto fail;
1516
1517         ublks_added++;
1518         return 0;
1519  fail:
1520         put_device(dev);
1521         return ret;
1522 }
1523
1524 static void ublk_stop_work_fn(struct work_struct *work)
1525 {
1526         struct ublk_device *ub =
1527                 container_of(work, struct ublk_device, stop_work);
1528
1529         ublk_stop_dev(ub);
1530 }
1531
1532 /* align max io buffer size with PAGE_SIZE */
1533 static void ublk_align_max_io_size(struct ublk_device *ub)
1534 {
1535         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1536
1537         ub->dev_info.max_io_buf_bytes =
1538                 round_down(max_io_bytes, PAGE_SIZE);
1539 }
1540
1541 static int ublk_add_tag_set(struct ublk_device *ub)
1542 {
1543         ub->tag_set.ops = &ublk_mq_ops;
1544         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1545         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1546         ub->tag_set.numa_node = NUMA_NO_NODE;
1547         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1548         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1549         ub->tag_set.driver_data = ub;
1550         return blk_mq_alloc_tag_set(&ub->tag_set);
1551 }
1552
1553 static void ublk_remove(struct ublk_device *ub)
1554 {
1555         ublk_stop_dev(ub);
1556         cancel_work_sync(&ub->stop_work);
1557         cancel_work_sync(&ub->quiesce_work);
1558         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1559         put_device(&ub->cdev_dev);
1560         ublks_added--;
1561 }
1562
1563 static struct ublk_device *ublk_get_device_from_id(int idx)
1564 {
1565         struct ublk_device *ub = NULL;
1566
1567         if (idx < 0)
1568                 return NULL;
1569
1570         spin_lock(&ublk_idr_lock);
1571         ub = idr_find(&ublk_index_idr, idx);
1572         if (ub)
1573                 ub = ublk_get_device(ub);
1574         spin_unlock(&ublk_idr_lock);
1575
1576         return ub;
1577 }
1578
1579 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1580 {
1581         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1582         int ublksrv_pid = (int)header->data[0];
1583         struct gendisk *disk;
1584         int ret = -EINVAL;
1585
1586         if (ublksrv_pid <= 0)
1587                 return -EINVAL;
1588
1589         wait_for_completion_interruptible(&ub->completion);
1590
1591         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1592
1593         mutex_lock(&ub->mutex);
1594         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1595             test_bit(UB_STATE_USED, &ub->state)) {
1596                 ret = -EEXIST;
1597                 goto out_unlock;
1598         }
1599
1600         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1601         if (IS_ERR(disk)) {
1602                 ret = PTR_ERR(disk);
1603                 goto out_unlock;
1604         }
1605         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1606         disk->fops = &ub_fops;
1607         disk->private_data = ub;
1608
1609         ub->dev_info.ublksrv_pid = ublksrv_pid;
1610         ub->ub_disk = disk;
1611
1612         ret = ublk_apply_params(ub);
1613         if (ret)
1614                 goto out_put_disk;
1615
1616         /* don't probe partitions if any one ubq daemon is un-trusted */
1617         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1618                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1619
1620         get_device(&ub->cdev_dev);
1621         ret = add_disk(disk);
1622         if (ret) {
1623                 /*
1624                  * Has to drop the reference since ->free_disk won't be
1625                  * called in case of add_disk failure.
1626                  */
1627                 ublk_put_device(ub);
1628                 goto out_put_disk;
1629         }
1630         set_bit(UB_STATE_USED, &ub->state);
1631         ub->dev_info.state = UBLK_S_DEV_LIVE;
1632 out_put_disk:
1633         if (ret)
1634                 put_disk(disk);
1635 out_unlock:
1636         mutex_unlock(&ub->mutex);
1637         return ret;
1638 }
1639
1640 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1641                 struct io_uring_cmd *cmd)
1642 {
1643         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1644         void __user *argp = (void __user *)(unsigned long)header->addr;
1645         cpumask_var_t cpumask;
1646         unsigned long queue;
1647         unsigned int retlen;
1648         unsigned int i;
1649         int ret;
1650
1651         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1652                 return -EINVAL;
1653         if (header->len & (sizeof(unsigned long)-1))
1654                 return -EINVAL;
1655         if (!header->addr)
1656                 return -EINVAL;
1657
1658         queue = header->data[0];
1659         if (queue >= ub->dev_info.nr_hw_queues)
1660                 return -EINVAL;
1661
1662         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1663                 return -ENOMEM;
1664
1665         for_each_possible_cpu(i) {
1666                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1667                         cpumask_set_cpu(i, cpumask);
1668         }
1669
1670         ret = -EFAULT;
1671         retlen = min_t(unsigned short, header->len, cpumask_size());
1672         if (copy_to_user(argp, cpumask, retlen))
1673                 goto out_free_cpumask;
1674         if (retlen != header->len &&
1675             clear_user(argp + retlen, header->len - retlen))
1676                 goto out_free_cpumask;
1677
1678         ret = 0;
1679 out_free_cpumask:
1680         free_cpumask_var(cpumask);
1681         return ret;
1682 }
1683
1684 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1685 {
1686         pr_devel("%s: dev id %d flags %llx\n", __func__,
1687                         info->dev_id, info->flags);
1688         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1689                         info->nr_hw_queues, info->queue_depth);
1690 }
1691
1692 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1693 {
1694         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1695         void __user *argp = (void __user *)(unsigned long)header->addr;
1696         struct ublksrv_ctrl_dev_info info;
1697         struct ublk_device *ub;
1698         int ret = -EINVAL;
1699
1700         if (header->len < sizeof(info) || !header->addr)
1701                 return -EINVAL;
1702         if (header->queue_id != (u16)-1) {
1703                 pr_warn("%s: queue_id is wrong %x\n",
1704                         __func__, header->queue_id);
1705                 return -EINVAL;
1706         }
1707
1708         if (copy_from_user(&info, argp, sizeof(info)))
1709                 return -EFAULT;
1710
1711         if (capable(CAP_SYS_ADMIN))
1712                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1713         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1714                 return -EPERM;
1715
1716         /* the created device is always owned by current user */
1717         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1718
1719         if (header->dev_id != info.dev_id) {
1720                 pr_warn("%s: dev id not match %u %u\n",
1721                         __func__, header->dev_id, info.dev_id);
1722                 return -EINVAL;
1723         }
1724
1725         ublk_dump_dev_info(&info);
1726
1727         ret = mutex_lock_killable(&ublk_ctl_mutex);
1728         if (ret)
1729                 return ret;
1730
1731         ret = -EACCES;
1732         if (ublks_added >= ublks_max)
1733                 goto out_unlock;
1734
1735         ret = -ENOMEM;
1736         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1737         if (!ub)
1738                 goto out_unlock;
1739         mutex_init(&ub->mutex);
1740         spin_lock_init(&ub->mm_lock);
1741         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1742         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1743         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1744
1745         ret = ublk_alloc_dev_number(ub, header->dev_id);
1746         if (ret < 0)
1747                 goto out_free_ub;
1748
1749         memcpy(&ub->dev_info, &info, sizeof(info));
1750
1751         /* update device id */
1752         ub->dev_info.dev_id = ub->ub_number;
1753
1754         /*
1755          * 64bit flags will be copied back to userspace as feature
1756          * negotiation result, so have to clear flags which driver
1757          * doesn't support yet, then userspace can get correct flags
1758          * (features) to handle.
1759          */
1760         ub->dev_info.flags &= UBLK_F_ALL;
1761
1762         if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1763                 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1764
1765         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE;
1766
1767         /* We are not ready to support zero copy */
1768         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1769
1770         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1771                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1772         ublk_align_max_io_size(ub);
1773
1774         ret = ublk_init_queues(ub);
1775         if (ret)
1776                 goto out_free_dev_number;
1777
1778         ret = ublk_add_tag_set(ub);
1779         if (ret)
1780                 goto out_deinit_queues;
1781
1782         ret = -EFAULT;
1783         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1784                 goto out_free_tag_set;
1785
1786         /*
1787          * Add the char dev so that ublksrv daemon can be setup.
1788          * ublk_add_chdev() will cleanup everything if it fails.
1789          */
1790         ret = ublk_add_chdev(ub);
1791         goto out_unlock;
1792
1793 out_free_tag_set:
1794         blk_mq_free_tag_set(&ub->tag_set);
1795 out_deinit_queues:
1796         ublk_deinit_queues(ub);
1797 out_free_dev_number:
1798         ublk_free_dev_number(ub);
1799 out_free_ub:
1800         mutex_destroy(&ub->mutex);
1801         kfree(ub);
1802 out_unlock:
1803         mutex_unlock(&ublk_ctl_mutex);
1804         return ret;
1805 }
1806
1807 static inline bool ublk_idr_freed(int id)
1808 {
1809         void *ptr;
1810
1811         spin_lock(&ublk_idr_lock);
1812         ptr = idr_find(&ublk_index_idr, id);
1813         spin_unlock(&ublk_idr_lock);
1814
1815         return ptr == NULL;
1816 }
1817
1818 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1819 {
1820         struct ublk_device *ub = *p_ub;
1821         int idx = ub->ub_number;
1822         int ret;
1823
1824         ret = mutex_lock_killable(&ublk_ctl_mutex);
1825         if (ret)
1826                 return ret;
1827
1828         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1829                 ublk_remove(ub);
1830                 set_bit(UB_STATE_DELETED, &ub->state);
1831         }
1832
1833         /* Mark the reference as consumed */
1834         *p_ub = NULL;
1835         ublk_put_device(ub);
1836         mutex_unlock(&ublk_ctl_mutex);
1837
1838         /*
1839          * Wait until the idr is removed, then it can be reused after
1840          * DEL_DEV command is returned.
1841          *
1842          * If we returns because of user interrupt, future delete command
1843          * may come:
1844          *
1845          * - the device number isn't freed, this device won't or needn't
1846          *   be deleted again, since UB_STATE_DELETED is set, and device
1847          *   will be released after the last reference is dropped
1848          *
1849          * - the device number is freed already, we will not find this
1850          *   device via ublk_get_device_from_id()
1851          */
1852         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1853
1854         return 0;
1855 }
1856
1857 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1858 {
1859         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1860
1861         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1862                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1863                         header->data[0], header->addr, header->len);
1864 }
1865
1866 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1867 {
1868         ublk_stop_dev(ub);
1869         cancel_work_sync(&ub->stop_work);
1870         cancel_work_sync(&ub->quiesce_work);
1871
1872         return 0;
1873 }
1874
1875 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1876                 struct io_uring_cmd *cmd)
1877 {
1878         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1879         void __user *argp = (void __user *)(unsigned long)header->addr;
1880
1881         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1882                 return -EINVAL;
1883
1884         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1885                 return -EFAULT;
1886
1887         return 0;
1888 }
1889
1890 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1891 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1892 {
1893         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1894         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1895
1896         if (ub->ub_disk) {
1897                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1898                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1899         } else {
1900                 ub->params.devt.disk_major = 0;
1901                 ub->params.devt.disk_minor = 0;
1902         }
1903         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1904 }
1905
1906 static int ublk_ctrl_get_params(struct ublk_device *ub,
1907                 struct io_uring_cmd *cmd)
1908 {
1909         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1910         void __user *argp = (void __user *)(unsigned long)header->addr;
1911         struct ublk_params_header ph;
1912         int ret;
1913
1914         if (header->len <= sizeof(ph) || !header->addr)
1915                 return -EINVAL;
1916
1917         if (copy_from_user(&ph, argp, sizeof(ph)))
1918                 return -EFAULT;
1919
1920         if (ph.len > header->len || !ph.len)
1921                 return -EINVAL;
1922
1923         if (ph.len > sizeof(struct ublk_params))
1924                 ph.len = sizeof(struct ublk_params);
1925
1926         mutex_lock(&ub->mutex);
1927         ublk_ctrl_fill_params_devt(ub);
1928         if (copy_to_user(argp, &ub->params, ph.len))
1929                 ret = -EFAULT;
1930         else
1931                 ret = 0;
1932         mutex_unlock(&ub->mutex);
1933
1934         return ret;
1935 }
1936
1937 static int ublk_ctrl_set_params(struct ublk_device *ub,
1938                 struct io_uring_cmd *cmd)
1939 {
1940         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1941         void __user *argp = (void __user *)(unsigned long)header->addr;
1942         struct ublk_params_header ph;
1943         int ret = -EFAULT;
1944
1945         if (header->len <= sizeof(ph) || !header->addr)
1946                 return -EINVAL;
1947
1948         if (copy_from_user(&ph, argp, sizeof(ph)))
1949                 return -EFAULT;
1950
1951         if (ph.len > header->len || !ph.len || !ph.types)
1952                 return -EINVAL;
1953
1954         if (ph.len > sizeof(struct ublk_params))
1955                 ph.len = sizeof(struct ublk_params);
1956
1957         /* parameters can only be changed when device isn't live */
1958         mutex_lock(&ub->mutex);
1959         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1960                 ret = -EACCES;
1961         } else if (copy_from_user(&ub->params, argp, ph.len)) {
1962                 ret = -EFAULT;
1963         } else {
1964                 /* clear all we don't support yet */
1965                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1966                 ret = ublk_validate_params(ub);
1967         }
1968         mutex_unlock(&ub->mutex);
1969
1970         return ret;
1971 }
1972
1973 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1974 {
1975         int i;
1976
1977         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1978         /* All old ioucmds have to be completed */
1979         WARN_ON_ONCE(ubq->nr_io_ready);
1980         /* old daemon is PF_EXITING, put it now */
1981         put_task_struct(ubq->ubq_daemon);
1982         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1983         ubq->ubq_daemon = NULL;
1984
1985         for (i = 0; i < ubq->q_depth; i++) {
1986                 struct ublk_io *io = &ubq->ios[i];
1987
1988                 /* forget everything now and be ready for new FETCH_REQ */
1989                 io->flags = 0;
1990                 io->cmd = NULL;
1991                 io->addr = 0;
1992         }
1993 }
1994
1995 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
1996                 struct io_uring_cmd *cmd)
1997 {
1998         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1999         int ret = -EINVAL;
2000         int i;
2001
2002         mutex_lock(&ub->mutex);
2003         if (!ublk_can_use_recovery(ub))
2004                 goto out_unlock;
2005         /*
2006          * START_RECOVERY is only allowd after:
2007          *
2008          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2009          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2010          *     released.
2011          *
2012          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2013          *     (a)has quiesced request queue
2014          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2015          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2016          *     (d)has completed/camceled all ioucmds owned by ther dying process
2017          */
2018         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2019                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2020                 ret = -EBUSY;
2021                 goto out_unlock;
2022         }
2023         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2024         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2025                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2026         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2027         ub->mm = NULL;
2028         ub->nr_queues_ready = 0;
2029         ub->nr_privileged_daemon = 0;
2030         init_completion(&ub->completion);
2031         ret = 0;
2032  out_unlock:
2033         mutex_unlock(&ub->mutex);
2034         return ret;
2035 }
2036
2037 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2038                 struct io_uring_cmd *cmd)
2039 {
2040         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2041         int ublksrv_pid = (int)header->data[0];
2042         int ret = -EINVAL;
2043
2044         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2045                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2046         /* wait until new ubq_daemon sending all FETCH_REQ */
2047         wait_for_completion_interruptible(&ub->completion);
2048         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2049                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2050
2051         mutex_lock(&ub->mutex);
2052         if (!ublk_can_use_recovery(ub))
2053                 goto out_unlock;
2054
2055         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2056                 ret = -EBUSY;
2057                 goto out_unlock;
2058         }
2059         ub->dev_info.ublksrv_pid = ublksrv_pid;
2060         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2061                         __func__, ublksrv_pid, header->dev_id);
2062         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2063         pr_devel("%s: queue unquiesced, dev id %d.\n",
2064                         __func__, header->dev_id);
2065         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2066         ub->dev_info.state = UBLK_S_DEV_LIVE;
2067         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2068         ret = 0;
2069  out_unlock:
2070         mutex_unlock(&ub->mutex);
2071         return ret;
2072 }
2073
2074 /*
2075  * All control commands are sent via /dev/ublk-control, so we have to check
2076  * the destination device's permission
2077  */
2078 static int ublk_char_dev_permission(struct ublk_device *ub,
2079                 const char *dev_path, int mask)
2080 {
2081         int err;
2082         struct path path;
2083         struct kstat stat;
2084
2085         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2086         if (err)
2087                 return err;
2088
2089         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2090         if (err)
2091                 goto exit;
2092
2093         err = -EPERM;
2094         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2095                 goto exit;
2096
2097         err = inode_permission(&nop_mnt_idmap,
2098                         d_backing_inode(path.dentry), mask);
2099 exit:
2100         path_put(&path);
2101         return err;
2102 }
2103
2104 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2105                 struct io_uring_cmd *cmd)
2106 {
2107         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2108         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2109         void __user *argp = (void __user *)(unsigned long)header->addr;
2110         char *dev_path = NULL;
2111         int ret = 0;
2112         int mask;
2113
2114         if (!unprivileged) {
2115                 if (!capable(CAP_SYS_ADMIN))
2116                         return -EPERM;
2117                 /*
2118                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2119                  * char_dev_path in payload too, since userspace may not
2120                  * know if the specified device is created as unprivileged
2121                  * mode.
2122                  */
2123                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2124                         return 0;
2125         }
2126
2127         /*
2128          * User has to provide the char device path for unprivileged ublk
2129          *
2130          * header->addr always points to the dev path buffer, and
2131          * header->dev_path_len records length of dev path buffer.
2132          */
2133         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2134                 return -EINVAL;
2135
2136         if (header->len < header->dev_path_len)
2137                 return -EINVAL;
2138
2139         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2140         if (!dev_path)
2141                 return -ENOMEM;
2142
2143         ret = -EFAULT;
2144         if (copy_from_user(dev_path, argp, header->dev_path_len))
2145                 goto exit;
2146         dev_path[header->dev_path_len] = 0;
2147
2148         ret = -EINVAL;
2149         switch (_IOC_NR(cmd->cmd_op)) {
2150         case UBLK_CMD_GET_DEV_INFO:
2151         case UBLK_CMD_GET_DEV_INFO2:
2152         case UBLK_CMD_GET_QUEUE_AFFINITY:
2153         case UBLK_CMD_GET_PARAMS:
2154                 mask = MAY_READ;
2155                 break;
2156         case UBLK_CMD_START_DEV:
2157         case UBLK_CMD_STOP_DEV:
2158         case UBLK_CMD_ADD_DEV:
2159         case UBLK_CMD_DEL_DEV:
2160         case UBLK_CMD_SET_PARAMS:
2161         case UBLK_CMD_START_USER_RECOVERY:
2162         case UBLK_CMD_END_USER_RECOVERY:
2163                 mask = MAY_READ | MAY_WRITE;
2164                 break;
2165         default:
2166                 goto exit;
2167         }
2168
2169         ret = ublk_char_dev_permission(ub, dev_path, mask);
2170         if (!ret) {
2171                 header->len -= header->dev_path_len;
2172                 header->addr += header->dev_path_len;
2173         }
2174         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2175                         __func__, ub->ub_number, cmd->cmd_op,
2176                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2177                         dev_path, ret);
2178 exit:
2179         kfree(dev_path);
2180         return ret;
2181 }
2182
2183 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2184                 unsigned int issue_flags)
2185 {
2186         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2187         struct ublk_device *ub = NULL;
2188         u32 cmd_op = cmd->cmd_op;
2189         int ret = -EINVAL;
2190
2191         if (issue_flags & IO_URING_F_NONBLOCK)
2192                 return -EAGAIN;
2193
2194         ublk_ctrl_cmd_dump(cmd);
2195
2196         if (!(issue_flags & IO_URING_F_SQE128))
2197                 goto out;
2198
2199         ret = ublk_check_cmd_op(cmd_op);
2200         if (ret)
2201                 goto out;
2202
2203         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2204                 ret = -ENODEV;
2205                 ub = ublk_get_device_from_id(header->dev_id);
2206                 if (!ub)
2207                         goto out;
2208
2209                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2210                 if (ret)
2211                         goto put_dev;
2212         }
2213
2214         switch (_IOC_NR(cmd_op)) {
2215         case UBLK_CMD_START_DEV:
2216                 ret = ublk_ctrl_start_dev(ub, cmd);
2217                 break;
2218         case UBLK_CMD_STOP_DEV:
2219                 ret = ublk_ctrl_stop_dev(ub);
2220                 break;
2221         case UBLK_CMD_GET_DEV_INFO:
2222         case UBLK_CMD_GET_DEV_INFO2:
2223                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2224                 break;
2225         case UBLK_CMD_ADD_DEV:
2226                 ret = ublk_ctrl_add_dev(cmd);
2227                 break;
2228         case UBLK_CMD_DEL_DEV:
2229                 ret = ublk_ctrl_del_dev(&ub);
2230                 break;
2231         case UBLK_CMD_GET_QUEUE_AFFINITY:
2232                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2233                 break;
2234         case UBLK_CMD_GET_PARAMS:
2235                 ret = ublk_ctrl_get_params(ub, cmd);
2236                 break;
2237         case UBLK_CMD_SET_PARAMS:
2238                 ret = ublk_ctrl_set_params(ub, cmd);
2239                 break;
2240         case UBLK_CMD_START_USER_RECOVERY:
2241                 ret = ublk_ctrl_start_recovery(ub, cmd);
2242                 break;
2243         case UBLK_CMD_END_USER_RECOVERY:
2244                 ret = ublk_ctrl_end_recovery(ub, cmd);
2245                 break;
2246         default:
2247                 ret = -ENOTSUPP;
2248                 break;
2249         }
2250
2251  put_dev:
2252         if (ub)
2253                 ublk_put_device(ub);
2254  out:
2255         io_uring_cmd_done(cmd, ret, 0);
2256         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2257                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2258         return -EIOCBQUEUED;
2259 }
2260
2261 static const struct file_operations ublk_ctl_fops = {
2262         .open           = nonseekable_open,
2263         .uring_cmd      = ublk_ctrl_uring_cmd,
2264         .owner          = THIS_MODULE,
2265         .llseek         = noop_llseek,
2266 };
2267
2268 static struct miscdevice ublk_misc = {
2269         .minor          = MISC_DYNAMIC_MINOR,
2270         .name           = "ublk-control",
2271         .fops           = &ublk_ctl_fops,
2272 };
2273
2274 static int __init ublk_init(void)
2275 {
2276         int ret;
2277
2278         init_waitqueue_head(&ublk_idr_wq);
2279
2280         ret = misc_register(&ublk_misc);
2281         if (ret)
2282                 return ret;
2283
2284         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2285         if (ret)
2286                 goto unregister_mis;
2287
2288         ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2289         if (IS_ERR(ublk_chr_class)) {
2290                 ret = PTR_ERR(ublk_chr_class);
2291                 goto free_chrdev_region;
2292         }
2293         return 0;
2294
2295 free_chrdev_region:
2296         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2297 unregister_mis:
2298         misc_deregister(&ublk_misc);
2299         return ret;
2300 }
2301
2302 static void __exit ublk_exit(void)
2303 {
2304         struct ublk_device *ub;
2305         int id;
2306
2307         idr_for_each_entry(&ublk_index_idr, ub, id)
2308                 ublk_remove(ub);
2309
2310         class_destroy(ublk_chr_class);
2311         misc_deregister(&ublk_misc);
2312
2313         idr_destroy(&ublk_index_idr);
2314         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2315 }
2316
2317 module_init(ublk_init);
2318 module_exit(ublk_exit);
2319
2320 module_param(ublks_max, int, 0444);
2321 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2322
2323 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2324 MODULE_LICENSE("GPL");