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