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