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 || p->logical_bs_shift < 9)
 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                                 unsigned issue_flags)
 721 {
 722         /* mark this cmd owned by ublksrv */
 723         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 724
 725         /*
 726          * clear ACTIVE since we are done with this sqe/cmd slot
 727          * We can only accept io cmd in case of being not active.
 728          */
 729         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 730
 731         /* tell ublksrv one io request is coming */
 732         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
 733 }
 734
 735 #define UBLK_REQUEUE_DELAY_MS   3
 736
 737 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 738                 struct request *rq)
 739 {
 740         /* We cannot process this rq so just requeue it. */
 741         if (ublk_queue_can_use_recovery(ubq))
 742                 blk_mq_requeue_request(rq, false);
 743         else
 744                 blk_mq_end_request(rq, BLK_STS_IOERR);
 745
 746         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 747 }
 748
 749 static inline void __ublk_rq_task_work(struct request *req,
 750                                        unsigned issue_flags)
 751 {
 752         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 753         int tag = req->tag;
 754         struct ublk_io *io = &ubq->ios[tag];
 755         unsigned int mapped_bytes;
 756
 757         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 758                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 759                         ublk_get_iod(ubq, req->tag)->addr);
 760
 761         /*
 762          * Task is exiting if either:
 763          *
 764          * (1) current != ubq_daemon.
 765          * io_uring_cmd_complete_in_task() tries to run task_work
 766          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 767          *
 768          * (2) current->flags & PF_EXITING.
 769          */
 770         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 771                 __ublk_abort_rq(ubq, req);
 772                 return;
 773         }
 774
 775         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
 776                 /*
 777                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 778                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 779                  * and notify it.
 780                  */
 781                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 782                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 783                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 784                                         __func__, io->cmd->cmd_op, ubq->q_id,
 785                                         req->tag, io->flags);
 786                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
 787                         return;
 788                 }
 789                 /*
 790                  * We have handled UBLK_IO_NEED_GET_DATA command,
 791                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 792                  * do the copy work.
 793                  */
 794                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 795                 /* update iod->addr because ublksrv may have passed a new io buffer */
 796                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 797                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 798                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 799                                 ublk_get_iod(ubq, req->tag)->addr);
 800         }
 801
 802         mapped_bytes = ublk_map_io(ubq, req, io);
 803
 804         /* partially mapped, update io descriptor */
 805         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 806                 /*
 807                  * Nothing mapped, retry until we succeed.
 808                  *
 809                  * We may never succeed in mapping any bytes here because
 810                  * of OOM. TODO: reserve one buffer with single page pinned
 811                  * for providing forward progress guarantee.
 812                  */
 813                 if (unlikely(!mapped_bytes)) {
 814                         blk_mq_requeue_request(req, false);
 815                         blk_mq_delay_kick_requeue_list(req->q,
 816                                         UBLK_REQUEUE_DELAY_MS);
 817                         return;
 818                 }
 819
 820                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 821                         mapped_bytes >> 9;
 822         }
 823
 824         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
 825 }
 826
 827 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
 828                                         unsigned issue_flags)
 829 {
 830         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 831         struct ublk_rq_data *data, *tmp;
 832
 833         io_cmds = llist_reverse_order(io_cmds);
 834         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 835                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
 836 }
 837
 838 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 839 {
 840         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 841         struct ublk_rq_data *data, *tmp;
 842
 843         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 844                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 845 }
 846
 847 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
 848 {
 849         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 850         struct ublk_queue *ubq = pdu->ubq;
 851
 852         ublk_forward_io_cmds(ubq, issue_flags);
 853 }
 854
 855 static void ublk_rq_task_work_fn(struct callback_head *work)
 856 {
 857         struct ublk_rq_data *data = container_of(work,
 858                         struct ublk_rq_data, work);
 859         struct request *req = blk_mq_rq_from_pdu(data);
 860         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 861         unsigned issue_flags = IO_URING_F_UNLOCKED;
 862
 863         ublk_forward_io_cmds(ubq, issue_flags);
 864 }
 865
 866 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 867 {
 868         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 869         struct ublk_io *io;
 870
 871         if (!llist_add(&data->node, &ubq->io_cmds))
 872                 return;
 873
 874         io = &ubq->ios[rq->tag];
 875         /*
 876          * If the check pass, we know that this is a re-issued request aborted
 877          * previously in monitor_work because the ubq_daemon(cmd's task) is
 878          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 879          * because this ioucmd's io_uring context may be freed now if no inflight
 880          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 881          *
 882          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 883          * the tag). Then the request is re-started(allocating the tag) and we are here.
 884          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 885          * guarantees that here is a re-issued request aborted previously.
 886          */
 887         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 888                 ublk_abort_io_cmds(ubq);
 889         } else if (ublk_can_use_task_work(ubq)) {
 890                 if (task_work_add(ubq->ubq_daemon, &data->work,
 891                                         TWA_SIGNAL_NO_IPI))
 892                         ublk_abort_io_cmds(ubq);
 893         } else {
 894                 struct io_uring_cmd *cmd = io->cmd;
 895                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 896
 897                 pdu->ubq = ubq;
 898                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 899         }
 900 }
 901
 902 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
 903 {
 904         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
 905
 906         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
 907                 if (!ubq->timeout) {
 908                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
 909                         ubq->timeout = true;
 910                 }
 911
 912                 return BLK_EH_DONE;
 913         }
 914
 915         return BLK_EH_RESET_TIMER;
 916 }
 917
 918 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 919                 const struct blk_mq_queue_data *bd)
 920 {
 921         struct ublk_queue *ubq = hctx->driver_data;
 922         struct request *rq = bd->rq;
 923         blk_status_t res;
 924
 925         /* fill iod to slot in io cmd buffer */
 926         res = ublk_setup_iod(ubq, rq);
 927         if (unlikely(res != BLK_STS_OK))
 928                 return BLK_STS_IOERR;
 929
 930         /* With recovery feature enabled, force_abort is set in
 931          * ublk_stop_dev() before calling del_gendisk(). We have to
 932          * abort all requeued and new rqs here to let del_gendisk()
 933          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 934          * to avoid UAF on io_uring ctx.
 935          *
 936          * Note: force_abort is guaranteed to be seen because it is set
 937          * before request queue is unqiuesced.
 938          */
 939         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 940                 return BLK_STS_IOERR;
 941
 942         blk_mq_start_request(bd->rq);
 943
 944         if (unlikely(ubq_daemon_is_dying(ubq))) {
 945                 __ublk_abort_rq(ubq, rq);
 946                 return BLK_STS_OK;
 947         }
 948
 949         ublk_queue_cmd(ubq, rq);
 950
 951         return BLK_STS_OK;
 952 }
 953
 954 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 955                 unsigned int hctx_idx)
 956 {
 957         struct ublk_device *ub = driver_data;
 958         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 959
 960         hctx->driver_data = ubq;
 961         return 0;
 962 }
 963
 964 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 965                 unsigned int hctx_idx, unsigned int numa_node)
 966 {
 967         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 968
 969         init_task_work(&data->work, ublk_rq_task_work_fn);
 970         return 0;
 971 }
 972
 973 static const struct blk_mq_ops ublk_mq_ops = {
 974         .queue_rq       = ublk_queue_rq,
 975         .init_hctx      = ublk_init_hctx,
 976         .init_request   = ublk_init_rq,
 977         .timeout        = ublk_timeout,
 978 };
 979
 980 static int ublk_ch_open(struct inode *inode, struct file *filp)
 981 {
 982         struct ublk_device *ub = container_of(inode->i_cdev,
 983                         struct ublk_device, cdev);
 984
 985         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 986                 return -EBUSY;
 987         filp->private_data = ub;
 988         return 0;
 989 }
 990
 991 static int ublk_ch_release(struct inode *inode, struct file *filp)
 992 {
 993         struct ublk_device *ub = filp->private_data;
 994
 995         clear_bit(UB_STATE_OPEN, &ub->state);
 996         return 0;
 997 }
 998
 999 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1000 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1001 {
1002         struct ublk_device *ub = filp->private_data;
1003         size_t sz = vma->vm_end - vma->vm_start;
1004         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1005         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1006         int q_id, ret = 0;
1007
1008         spin_lock(&ub->mm_lock);
1009         if (!ub->mm)
1010                 ub->mm = current->mm;
1011         if (current->mm != ub->mm)
1012                 ret = -EINVAL;
1013         spin_unlock(&ub->mm_lock);
1014
1015         if (ret)
1016                 return ret;
1017
1018         if (vma->vm_flags & VM_WRITE)
1019                 return -EPERM;
1020
1021         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1022         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1023                 return -EINVAL;
1024
1025         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1026         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1027                         __func__, q_id, current->pid, vma->vm_start,
1028                         phys_off, (unsigned long)sz);
1029
1030         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1031                 return -EINVAL;
1032
1033         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1034         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1035 }
1036
1037 static void ublk_commit_completion(struct ublk_device *ub,
1038                 const struct ublksrv_io_cmd *ub_cmd)
1039 {
1040         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1041         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1042         struct ublk_io *io = &ubq->ios[tag];
1043         struct request *req;
1044
1045         /* now this cmd slot is owned by nbd driver */
1046         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1047         io->res = ub_cmd->result;
1048
1049         /* find the io request and complete */
1050         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1051
1052         if (req && likely(!blk_should_fake_timeout(req->q)))
1053                 ublk_complete_rq(req);
1054 }
1055
1056 /*
1057  * When ->ubq_daemon is exiting, either new request is ended immediately,
1058  * or any queued io command is drained, so it is safe to abort queue
1059  * lockless
1060  */
1061 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1062 {
1063         int i;
1064
1065         if (!ublk_get_device(ub))
1066                 return;
1067
1068         for (i = 0; i < ubq->q_depth; i++) {
1069                 struct ublk_io *io = &ubq->ios[i];
1070
1071                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1072                         struct request *rq;
1073
1074                         /*
1075                          * Either we fail the request or ublk_rq_task_work_fn
1076                          * will do it
1077                          */
1078                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1079                         if (rq)
1080                                 __ublk_fail_req(ubq, io, rq);
1081                 }
1082         }
1083         ublk_put_device(ub);
1084 }
1085
1086 static void ublk_daemon_monitor_work(struct work_struct *work)
1087 {
1088         struct ublk_device *ub =
1089                 container_of(work, struct ublk_device, monitor_work.work);
1090         int i;
1091
1092         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1093                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1094
1095                 if (ubq_daemon_is_dying(ubq)) {
1096                         if (ublk_queue_can_use_recovery(ubq))
1097                                 schedule_work(&ub->quiesce_work);
1098                         else
1099                                 schedule_work(&ub->stop_work);
1100
1101                         /* abort queue is for making forward progress */
1102                         ublk_abort_queue(ub, ubq);
1103                 }
1104         }
1105
1106         /*
1107          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1108          * after ublk_remove() or __ublk_quiesce_dev() is started.
1109          *
1110          * No need ub->mutex, monitor work are canceled after state is marked
1111          * as not LIVE, so new state is observed reliably.
1112          */
1113         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1114                 schedule_delayed_work(&ub->monitor_work,
1115                                 UBLK_DAEMON_MONITOR_PERIOD);
1116 }
1117
1118 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1119 {
1120         return ubq->nr_io_ready == ubq->q_depth;
1121 }
1122
1123 static void ublk_cmd_cancel_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1124 {
1125         io_uring_cmd_done(cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1126 }
1127
1128 static void ublk_cancel_queue(struct ublk_queue *ubq)
1129 {
1130         int i;
1131
1132         if (!ublk_queue_ready(ubq))
1133                 return;
1134
1135         for (i = 0; i < ubq->q_depth; i++) {
1136                 struct ublk_io *io = &ubq->ios[i];
1137
1138                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1139                         io_uring_cmd_complete_in_task(io->cmd,
1140                                                       ublk_cmd_cancel_cb);
1141         }
1142
1143         /* all io commands are canceled */
1144         ubq->nr_io_ready = 0;
1145 }
1146
1147 /* Cancel all pending commands, must be called after del_gendisk() returns */
1148 static void ublk_cancel_dev(struct ublk_device *ub)
1149 {
1150         int i;
1151
1152         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1153                 ublk_cancel_queue(ublk_get_queue(ub, i));
1154 }
1155
1156 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1157 {
1158         bool *idle = data;
1159
1160         if (blk_mq_request_started(rq)) {
1161                 *idle = false;
1162                 return false;
1163         }
1164         return true;
1165 }
1166
1167 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1168 {
1169         bool idle;
1170
1171         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1172         while (true) {
1173                 idle = true;
1174                 blk_mq_tagset_busy_iter(&ub->tag_set,
1175                                 ublk_check_inflight_rq, &idle);
1176                 if (idle)
1177                         break;
1178                 msleep(UBLK_REQUEUE_DELAY_MS);
1179         }
1180 }
1181
1182 static void __ublk_quiesce_dev(struct ublk_device *ub)
1183 {
1184         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1185                         __func__, ub->dev_info.dev_id,
1186                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1187                         "LIVE" : "QUIESCED");
1188         blk_mq_quiesce_queue(ub->ub_disk->queue);
1189         ublk_wait_tagset_rqs_idle(ub);
1190         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1191         ublk_cancel_dev(ub);
1192         /* we are going to release task_struct of ubq_daemon and resets
1193          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1194          * Besides, monitor_work is not necessary in QUIESCED state since we have
1195          * already scheduled quiesce_work and quiesced all ubqs.
1196          *
1197          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1198          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1199          */
1200         cancel_delayed_work_sync(&ub->monitor_work);
1201 }
1202
1203 static void ublk_quiesce_work_fn(struct work_struct *work)
1204 {
1205         struct ublk_device *ub =
1206                 container_of(work, struct ublk_device, quiesce_work);
1207
1208         mutex_lock(&ub->mutex);
1209         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1210                 goto unlock;
1211         __ublk_quiesce_dev(ub);
1212  unlock:
1213         mutex_unlock(&ub->mutex);
1214 }
1215
1216 static void ublk_unquiesce_dev(struct ublk_device *ub)
1217 {
1218         int i;
1219
1220         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1221                         __func__, ub->dev_info.dev_id,
1222                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1223                         "LIVE" : "QUIESCED");
1224         /* quiesce_work has run. We let requeued rqs be aborted
1225          * before running fallback_wq. "force_abort" must be seen
1226          * after request queue is unqiuesced. Then del_gendisk()
1227          * can move on.
1228          */
1229         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1230                 ublk_get_queue(ub, i)->force_abort = true;
1231
1232         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1233         /* We may have requeued some rqs in ublk_quiesce_queue() */
1234         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1235 }
1236
1237 static void ublk_stop_dev(struct ublk_device *ub)
1238 {
1239         mutex_lock(&ub->mutex);
1240         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1241                 goto unlock;
1242         if (ublk_can_use_recovery(ub)) {
1243                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1244                         __ublk_quiesce_dev(ub);
1245                 ublk_unquiesce_dev(ub);
1246         }
1247         del_gendisk(ub->ub_disk);
1248         ub->dev_info.state = UBLK_S_DEV_DEAD;
1249         ub->dev_info.ublksrv_pid = -1;
1250         put_disk(ub->ub_disk);
1251         ub->ub_disk = NULL;
1252  unlock:
1253         ublk_cancel_dev(ub);
1254         mutex_unlock(&ub->mutex);
1255         cancel_delayed_work_sync(&ub->monitor_work);
1256 }
1257
1258 /* device can only be started after all IOs are ready */
1259 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1260 {
1261         mutex_lock(&ub->mutex);
1262         ubq->nr_io_ready++;
1263         if (ublk_queue_ready(ubq)) {
1264                 ubq->ubq_daemon = current;
1265                 get_task_struct(ubq->ubq_daemon);
1266                 ub->nr_queues_ready++;
1267
1268                 if (capable(CAP_SYS_ADMIN))
1269                         ub->nr_privileged_daemon++;
1270         }
1271         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1272                 complete_all(&ub->completion);
1273         mutex_unlock(&ub->mutex);
1274 }
1275
1276 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1277                 int tag)
1278 {
1279         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1280         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1281
1282         ublk_queue_cmd(ubq, req);
1283 }
1284
1285 static inline int ublk_check_cmd_op(u32 cmd_op)
1286 {
1287         u32 ioc_type = _IOC_TYPE(cmd_op);
1288
1289         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1290                 return -EOPNOTSUPP;
1291
1292         if (ioc_type != 'u' && ioc_type != 0)
1293                 return -EOPNOTSUPP;
1294
1295         return 0;
1296 }
1297
1298 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1299                                unsigned int issue_flags,
1300                                const struct ublksrv_io_cmd *ub_cmd)
1301 {
1302         struct ublk_device *ub = cmd->file->private_data;
1303         struct ublk_queue *ubq;
1304         struct ublk_io *io;
1305         u32 cmd_op = cmd->cmd_op;
1306         unsigned tag = ub_cmd->tag;
1307         int ret = -EINVAL;
1308         struct request *req;
1309
1310         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1311                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1312                         ub_cmd->result);
1313
1314         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1315                 goto out;
1316
1317         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1318         if (!ubq || ub_cmd->q_id != ubq->q_id)
1319                 goto out;
1320
1321         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1322                 goto out;
1323
1324         if (tag >= ubq->q_depth)
1325                 goto out;
1326
1327         io = &ubq->ios[tag];
1328
1329         /* there is pending io cmd, something must be wrong */
1330         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1331                 ret = -EBUSY;
1332                 goto out;
1333         }
1334
1335         /*
1336          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1337          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1338          */
1339         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1340                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1341                 goto out;
1342
1343         ret = ublk_check_cmd_op(cmd_op);
1344         if (ret)
1345                 goto out;
1346
1347         ret = -EINVAL;
1348         switch (_IOC_NR(cmd_op)) {
1349         case UBLK_IO_FETCH_REQ:
1350                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1351                 if (ublk_queue_ready(ubq)) {
1352                         ret = -EBUSY;
1353                         goto out;
1354                 }
1355                 /*
1356                  * The io is being handled by server, so COMMIT_RQ is expected
1357                  * instead of FETCH_REQ
1358                  */
1359                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1360                         goto out;
1361                 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1362                 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1363                         goto out;
1364                 io->cmd = cmd;
1365                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1366                 io->addr = ub_cmd->addr;
1367
1368                 ublk_mark_io_ready(ub, ubq);
1369                 break;
1370         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1371                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1372                 /*
1373                  * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1374                  * not enabled or it is Read IO.
1375                  */
1376                 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1377                         goto out;
1378                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1379                         goto out;
1380                 io->addr = ub_cmd->addr;
1381                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1382                 io->cmd = cmd;
1383                 ublk_commit_completion(ub, ub_cmd);
1384                 break;
1385         case UBLK_IO_NEED_GET_DATA:
1386                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1387                         goto out;
1388                 io->addr = ub_cmd->addr;
1389                 io->cmd = cmd;
1390                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1391                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1392                 break;
1393         default:
1394                 goto out;
1395         }
1396         return -EIOCBQUEUED;
1397
1398  out:
1399         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1400         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1401                         __func__, cmd_op, tag, ret, io->flags);
1402         return -EIOCBQUEUED;
1403 }
1404
1405 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1406 {
1407         /*
1408          * Not necessary for async retry, but let's keep it simple and always
1409          * copy the values to avoid any potential reuse.
1410          */
1411         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1412         const struct ublksrv_io_cmd ub_cmd = {
1413                 .q_id = READ_ONCE(ub_src->q_id),
1414                 .tag = READ_ONCE(ub_src->tag),
1415                 .result = READ_ONCE(ub_src->result),
1416                 .addr = READ_ONCE(ub_src->addr)
1417         };
1418
1419         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1420 }
1421
1422 static const struct file_operations ublk_ch_fops = {
1423         .owner = THIS_MODULE,
1424         .open = ublk_ch_open,
1425         .release = ublk_ch_release,
1426         .llseek = no_llseek,
1427         .uring_cmd = ublk_ch_uring_cmd,
1428         .mmap = ublk_ch_mmap,
1429 };
1430
1431 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1432 {
1433         int size = ublk_queue_cmd_buf_size(ub, q_id);
1434         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1435
1436         if (ubq->ubq_daemon)
1437                 put_task_struct(ubq->ubq_daemon);
1438         if (ubq->io_cmd_buf)
1439                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1440 }
1441
1442 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1443 {
1444         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1445         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1446         void *ptr;
1447         int size;
1448
1449         ubq->flags = ub->dev_info.flags;
1450         ubq->q_id = q_id;
1451         ubq->q_depth = ub->dev_info.queue_depth;
1452         size = ublk_queue_cmd_buf_size(ub, q_id);
1453
1454         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1455         if (!ptr)
1456                 return -ENOMEM;
1457
1458         ubq->io_cmd_buf = ptr;
1459         ubq->dev = ub;
1460         return 0;
1461 }
1462
1463 static void ublk_deinit_queues(struct ublk_device *ub)
1464 {
1465         int nr_queues = ub->dev_info.nr_hw_queues;
1466         int i;
1467
1468         if (!ub->__queues)
1469                 return;
1470
1471         for (i = 0; i < nr_queues; i++)
1472                 ublk_deinit_queue(ub, i);
1473         kfree(ub->__queues);
1474 }
1475
1476 static int ublk_init_queues(struct ublk_device *ub)
1477 {
1478         int nr_queues = ub->dev_info.nr_hw_queues;
1479         int depth = ub->dev_info.queue_depth;
1480         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1481         int i, ret = -ENOMEM;
1482
1483         ub->queue_size = ubq_size;
1484         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1485         if (!ub->__queues)
1486                 return ret;
1487
1488         for (i = 0; i < nr_queues; i++) {
1489                 if (ublk_init_queue(ub, i))
1490                         goto fail;
1491         }
1492
1493         init_completion(&ub->completion);
1494         return 0;
1495
1496  fail:
1497         ublk_deinit_queues(ub);
1498         return ret;
1499 }
1500
1501 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1502 {
1503         int i = idx;
1504         int err;
1505
1506         spin_lock(&ublk_idr_lock);
1507         /* allocate id, if @id >= 0, we're requesting that specific id */
1508         if (i >= 0) {
1509                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1510                 if (err == -ENOSPC)
1511                         err = -EEXIST;
1512         } else {
1513                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1514         }
1515         spin_unlock(&ublk_idr_lock);
1516
1517         if (err >= 0)
1518                 ub->ub_number = err;
1519
1520         return err;
1521 }
1522
1523 static void ublk_free_dev_number(struct ublk_device *ub)
1524 {
1525         spin_lock(&ublk_idr_lock);
1526         idr_remove(&ublk_index_idr, ub->ub_number);
1527         wake_up_all(&ublk_idr_wq);
1528         spin_unlock(&ublk_idr_lock);
1529 }
1530
1531 static void ublk_cdev_rel(struct device *dev)
1532 {
1533         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1534
1535         blk_mq_free_tag_set(&ub->tag_set);
1536         ublk_deinit_queues(ub);
1537         ublk_free_dev_number(ub);
1538         mutex_destroy(&ub->mutex);
1539         kfree(ub);
1540 }
1541
1542 static int ublk_add_chdev(struct ublk_device *ub)
1543 {
1544         struct device *dev = &ub->cdev_dev;
1545         int minor = ub->ub_number;
1546         int ret;
1547
1548         dev->parent = ublk_misc.this_device;
1549         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1550         dev->class = ublk_chr_class;
1551         dev->release = ublk_cdev_rel;
1552         device_initialize(dev);
1553
1554         ret = dev_set_name(dev, "ublkc%d", minor);
1555         if (ret)
1556                 goto fail;
1557
1558         cdev_init(&ub->cdev, &ublk_ch_fops);
1559         ret = cdev_device_add(&ub->cdev, dev);
1560         if (ret)
1561                 goto fail;
1562
1563         ublks_added++;
1564         return 0;
1565  fail:
1566         put_device(dev);
1567         return ret;
1568 }
1569
1570 static void ublk_stop_work_fn(struct work_struct *work)
1571 {
1572         struct ublk_device *ub =
1573                 container_of(work, struct ublk_device, stop_work);
1574
1575         ublk_stop_dev(ub);
1576 }
1577
1578 /* align max io buffer size with PAGE_SIZE */
1579 static void ublk_align_max_io_size(struct ublk_device *ub)
1580 {
1581         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1582
1583         ub->dev_info.max_io_buf_bytes =
1584                 round_down(max_io_bytes, PAGE_SIZE);
1585 }
1586
1587 static int ublk_add_tag_set(struct ublk_device *ub)
1588 {
1589         ub->tag_set.ops = &ublk_mq_ops;
1590         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1591         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1592         ub->tag_set.numa_node = NUMA_NO_NODE;
1593         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1594         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1595         ub->tag_set.driver_data = ub;
1596         return blk_mq_alloc_tag_set(&ub->tag_set);
1597 }
1598
1599 static void ublk_remove(struct ublk_device *ub)
1600 {
1601         ublk_stop_dev(ub);
1602         cancel_work_sync(&ub->stop_work);
1603         cancel_work_sync(&ub->quiesce_work);
1604         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1605         put_device(&ub->cdev_dev);
1606         ublks_added--;
1607 }
1608
1609 static struct ublk_device *ublk_get_device_from_id(int idx)
1610 {
1611         struct ublk_device *ub = NULL;
1612
1613         if (idx < 0)
1614                 return NULL;
1615
1616         spin_lock(&ublk_idr_lock);
1617         ub = idr_find(&ublk_index_idr, idx);
1618         if (ub)
1619                 ub = ublk_get_device(ub);
1620         spin_unlock(&ublk_idr_lock);
1621
1622         return ub;
1623 }
1624
1625 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1626 {
1627         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1628         int ublksrv_pid = (int)header->data[0];
1629         struct gendisk *disk;
1630         int ret = -EINVAL;
1631
1632         if (ublksrv_pid <= 0)
1633                 return -EINVAL;
1634
1635         wait_for_completion_interruptible(&ub->completion);
1636
1637         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1638
1639         mutex_lock(&ub->mutex);
1640         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1641             test_bit(UB_STATE_USED, &ub->state)) {
1642                 ret = -EEXIST;
1643                 goto out_unlock;
1644         }
1645
1646         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1647         if (IS_ERR(disk)) {
1648                 ret = PTR_ERR(disk);
1649                 goto out_unlock;
1650         }
1651         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1652         disk->fops = &ub_fops;
1653         disk->private_data = ub;
1654
1655         ub->dev_info.ublksrv_pid = ublksrv_pid;
1656         ub->ub_disk = disk;
1657
1658         ret = ublk_apply_params(ub);
1659         if (ret)
1660                 goto out_put_disk;
1661
1662         /* don't probe partitions if any one ubq daemon is un-trusted */
1663         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1664                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1665
1666         get_device(&ub->cdev_dev);
1667         ub->dev_info.state = UBLK_S_DEV_LIVE;
1668         ret = add_disk(disk);
1669         if (ret) {
1670                 /*
1671                  * Has to drop the reference since ->free_disk won't be
1672                  * called in case of add_disk failure.
1673                  */
1674                 ub->dev_info.state = UBLK_S_DEV_DEAD;
1675                 ublk_put_device(ub);
1676                 goto out_put_disk;
1677         }
1678         set_bit(UB_STATE_USED, &ub->state);
1679 out_put_disk:
1680         if (ret)
1681                 put_disk(disk);
1682 out_unlock:
1683         mutex_unlock(&ub->mutex);
1684         return ret;
1685 }
1686
1687 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1688                 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         cpumask_var_t cpumask;
1693         unsigned long queue;
1694         unsigned int retlen;
1695         unsigned int i;
1696         int ret;
1697
1698         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1699                 return -EINVAL;
1700         if (header->len & (sizeof(unsigned long)-1))
1701                 return -EINVAL;
1702         if (!header->addr)
1703                 return -EINVAL;
1704
1705         queue = header->data[0];
1706         if (queue >= ub->dev_info.nr_hw_queues)
1707                 return -EINVAL;
1708
1709         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1710                 return -ENOMEM;
1711
1712         for_each_possible_cpu(i) {
1713                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1714                         cpumask_set_cpu(i, cpumask);
1715         }
1716
1717         ret = -EFAULT;
1718         retlen = min_t(unsigned short, header->len, cpumask_size());
1719         if (copy_to_user(argp, cpumask, retlen))
1720                 goto out_free_cpumask;
1721         if (retlen != header->len &&
1722             clear_user(argp + retlen, header->len - retlen))
1723                 goto out_free_cpumask;
1724
1725         ret = 0;
1726 out_free_cpumask:
1727         free_cpumask_var(cpumask);
1728         return ret;
1729 }
1730
1731 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1732 {
1733         pr_devel("%s: dev id %d flags %llx\n", __func__,
1734                         info->dev_id, info->flags);
1735         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1736                         info->nr_hw_queues, info->queue_depth);
1737 }
1738
1739 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1740 {
1741         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1742         void __user *argp = (void __user *)(unsigned long)header->addr;
1743         struct ublksrv_ctrl_dev_info info;
1744         struct ublk_device *ub;
1745         int ret = -EINVAL;
1746
1747         if (header->len < sizeof(info) || !header->addr)
1748                 return -EINVAL;
1749         if (header->queue_id != (u16)-1) {
1750                 pr_warn("%s: queue_id is wrong %x\n",
1751                         __func__, header->queue_id);
1752                 return -EINVAL;
1753         }
1754
1755         if (copy_from_user(&info, argp, sizeof(info)))
1756                 return -EFAULT;
1757
1758         if (capable(CAP_SYS_ADMIN))
1759                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1760         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1761                 return -EPERM;
1762
1763         /*
1764          * unprivileged device can't be trusted, but RECOVERY and
1765          * RECOVERY_REISSUE still may hang error handling, so can't
1766          * support recovery features for unprivileged ublk now
1767          *
1768          * TODO: provide forward progress for RECOVERY handler, so that
1769          * unprivileged device can benefit from it
1770          */
1771         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1772                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1773                                 UBLK_F_USER_RECOVERY);
1774
1775         /* the created device is always owned by current user */
1776         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1777
1778         if (header->dev_id != info.dev_id) {
1779                 pr_warn("%s: dev id not match %u %u\n",
1780                         __func__, header->dev_id, info.dev_id);
1781                 return -EINVAL;
1782         }
1783
1784         ublk_dump_dev_info(&info);
1785
1786         ret = mutex_lock_killable(&ublk_ctl_mutex);
1787         if (ret)
1788                 return ret;
1789
1790         ret = -EACCES;
1791         if (ublks_added >= ublks_max)
1792                 goto out_unlock;
1793
1794         ret = -ENOMEM;
1795         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1796         if (!ub)
1797                 goto out_unlock;
1798         mutex_init(&ub->mutex);
1799         spin_lock_init(&ub->mm_lock);
1800         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1801         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1802         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1803
1804         ret = ublk_alloc_dev_number(ub, header->dev_id);
1805         if (ret < 0)
1806                 goto out_free_ub;
1807
1808         memcpy(&ub->dev_info, &info, sizeof(info));
1809
1810         /* update device id */
1811         ub->dev_info.dev_id = ub->ub_number;
1812
1813         /*
1814          * 64bit flags will be copied back to userspace as feature
1815          * negotiation result, so have to clear flags which driver
1816          * doesn't support yet, then userspace can get correct flags
1817          * (features) to handle.
1818          */
1819         ub->dev_info.flags &= UBLK_F_ALL;
1820
1821         if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1822                 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1823
1824         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE;
1825
1826         /* We are not ready to support zero copy */
1827         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1828
1829         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1830                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1831         ublk_align_max_io_size(ub);
1832
1833         ret = ublk_init_queues(ub);
1834         if (ret)
1835                 goto out_free_dev_number;
1836
1837         ret = ublk_add_tag_set(ub);
1838         if (ret)
1839                 goto out_deinit_queues;
1840
1841         ret = -EFAULT;
1842         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1843                 goto out_free_tag_set;
1844
1845         /*
1846          * Add the char dev so that ublksrv daemon can be setup.
1847          * ublk_add_chdev() will cleanup everything if it fails.
1848          */
1849         ret = ublk_add_chdev(ub);
1850         goto out_unlock;
1851
1852 out_free_tag_set:
1853         blk_mq_free_tag_set(&ub->tag_set);
1854 out_deinit_queues:
1855         ublk_deinit_queues(ub);
1856 out_free_dev_number:
1857         ublk_free_dev_number(ub);
1858 out_free_ub:
1859         mutex_destroy(&ub->mutex);
1860         kfree(ub);
1861 out_unlock:
1862         mutex_unlock(&ublk_ctl_mutex);
1863         return ret;
1864 }
1865
1866 static inline bool ublk_idr_freed(int id)
1867 {
1868         void *ptr;
1869
1870         spin_lock(&ublk_idr_lock);
1871         ptr = idr_find(&ublk_index_idr, id);
1872         spin_unlock(&ublk_idr_lock);
1873
1874         return ptr == NULL;
1875 }
1876
1877 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1878 {
1879         struct ublk_device *ub = *p_ub;
1880         int idx = ub->ub_number;
1881         int ret;
1882
1883         ret = mutex_lock_killable(&ublk_ctl_mutex);
1884         if (ret)
1885                 return ret;
1886
1887         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1888                 ublk_remove(ub);
1889                 set_bit(UB_STATE_DELETED, &ub->state);
1890         }
1891
1892         /* Mark the reference as consumed */
1893         *p_ub = NULL;
1894         ublk_put_device(ub);
1895         mutex_unlock(&ublk_ctl_mutex);
1896
1897         /*
1898          * Wait until the idr is removed, then it can be reused after
1899          * DEL_DEV command is returned.
1900          *
1901          * If we returns because of user interrupt, future delete command
1902          * may come:
1903          *
1904          * - the device number isn't freed, this device won't or needn't
1905          *   be deleted again, since UB_STATE_DELETED is set, and device
1906          *   will be released after the last reference is dropped
1907          *
1908          * - the device number is freed already, we will not find this
1909          *   device via ublk_get_device_from_id()
1910          */
1911         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1912
1913         return 0;
1914 }
1915
1916 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1917 {
1918         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1919
1920         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1921                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1922                         header->data[0], header->addr, header->len);
1923 }
1924
1925 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1926 {
1927         ublk_stop_dev(ub);
1928         cancel_work_sync(&ub->stop_work);
1929         cancel_work_sync(&ub->quiesce_work);
1930
1931         return 0;
1932 }
1933
1934 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1935                 struct io_uring_cmd *cmd)
1936 {
1937         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1938         void __user *argp = (void __user *)(unsigned long)header->addr;
1939
1940         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1941                 return -EINVAL;
1942
1943         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1944                 return -EFAULT;
1945
1946         return 0;
1947 }
1948
1949 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1950 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1951 {
1952         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1953         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1954
1955         if (ub->ub_disk) {
1956                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1957                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1958         } else {
1959                 ub->params.devt.disk_major = 0;
1960                 ub->params.devt.disk_minor = 0;
1961         }
1962         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1963 }
1964
1965 static int ublk_ctrl_get_params(struct ublk_device *ub,
1966                 struct io_uring_cmd *cmd)
1967 {
1968         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1969         void __user *argp = (void __user *)(unsigned long)header->addr;
1970         struct ublk_params_header ph;
1971         int ret;
1972
1973         if (header->len <= sizeof(ph) || !header->addr)
1974                 return -EINVAL;
1975
1976         if (copy_from_user(&ph, argp, sizeof(ph)))
1977                 return -EFAULT;
1978
1979         if (ph.len > header->len || !ph.len)
1980                 return -EINVAL;
1981
1982         if (ph.len > sizeof(struct ublk_params))
1983                 ph.len = sizeof(struct ublk_params);
1984
1985         mutex_lock(&ub->mutex);
1986         ublk_ctrl_fill_params_devt(ub);
1987         if (copy_to_user(argp, &ub->params, ph.len))
1988                 ret = -EFAULT;
1989         else
1990                 ret = 0;
1991         mutex_unlock(&ub->mutex);
1992
1993         return ret;
1994 }
1995
1996 static int ublk_ctrl_set_params(struct ublk_device *ub,
1997                 struct io_uring_cmd *cmd)
1998 {
1999         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2000         void __user *argp = (void __user *)(unsigned long)header->addr;
2001         struct ublk_params_header ph;
2002         int ret = -EFAULT;
2003
2004         if (header->len <= sizeof(ph) || !header->addr)
2005                 return -EINVAL;
2006
2007         if (copy_from_user(&ph, argp, sizeof(ph)))
2008                 return -EFAULT;
2009
2010         if (ph.len > header->len || !ph.len || !ph.types)
2011                 return -EINVAL;
2012
2013         if (ph.len > sizeof(struct ublk_params))
2014                 ph.len = sizeof(struct ublk_params);
2015
2016         /* parameters can only be changed when device isn't live */
2017         mutex_lock(&ub->mutex);
2018         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2019                 ret = -EACCES;
2020         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2021                 ret = -EFAULT;
2022         } else {
2023                 /* clear all we don't support yet */
2024                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2025                 ret = ublk_validate_params(ub);
2026                 if (ret)
2027                         ub->params.types = 0;
2028         }
2029         mutex_unlock(&ub->mutex);
2030
2031         return ret;
2032 }
2033
2034 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2035 {
2036         int i;
2037
2038         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2039         /* All old ioucmds have to be completed */
2040         WARN_ON_ONCE(ubq->nr_io_ready);
2041         /* old daemon is PF_EXITING, put it now */
2042         put_task_struct(ubq->ubq_daemon);
2043         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2044         ubq->ubq_daemon = NULL;
2045         ubq->timeout = false;
2046
2047         for (i = 0; i < ubq->q_depth; i++) {
2048                 struct ublk_io *io = &ubq->ios[i];
2049
2050                 /* forget everything now and be ready for new FETCH_REQ */
2051                 io->flags = 0;
2052                 io->cmd = NULL;
2053                 io->addr = 0;
2054         }
2055 }
2056
2057 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2058                 struct io_uring_cmd *cmd)
2059 {
2060         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2061         int ret = -EINVAL;
2062         int i;
2063
2064         mutex_lock(&ub->mutex);
2065         if (!ublk_can_use_recovery(ub))
2066                 goto out_unlock;
2067         /*
2068          * START_RECOVERY is only allowd after:
2069          *
2070          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2071          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2072          *     released.
2073          *
2074          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2075          *     (a)has quiesced request queue
2076          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2077          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2078          *     (d)has completed/camceled all ioucmds owned by ther dying process
2079          */
2080         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2081                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2082                 ret = -EBUSY;
2083                 goto out_unlock;
2084         }
2085         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2086         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2087                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2088         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2089         ub->mm = NULL;
2090         ub->nr_queues_ready = 0;
2091         ub->nr_privileged_daemon = 0;
2092         init_completion(&ub->completion);
2093         ret = 0;
2094  out_unlock:
2095         mutex_unlock(&ub->mutex);
2096         return ret;
2097 }
2098
2099 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2100                 struct io_uring_cmd *cmd)
2101 {
2102         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2103         int ublksrv_pid = (int)header->data[0];
2104         int ret = -EINVAL;
2105
2106         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2107                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2108         /* wait until new ubq_daemon sending all FETCH_REQ */
2109         wait_for_completion_interruptible(&ub->completion);
2110         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2111                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2112
2113         mutex_lock(&ub->mutex);
2114         if (!ublk_can_use_recovery(ub))
2115                 goto out_unlock;
2116
2117         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2118                 ret = -EBUSY;
2119                 goto out_unlock;
2120         }
2121         ub->dev_info.ublksrv_pid = ublksrv_pid;
2122         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2123                         __func__, ublksrv_pid, header->dev_id);
2124         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2125         pr_devel("%s: queue unquiesced, dev id %d.\n",
2126                         __func__, header->dev_id);
2127         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2128         ub->dev_info.state = UBLK_S_DEV_LIVE;
2129         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2130         ret = 0;
2131  out_unlock:
2132         mutex_unlock(&ub->mutex);
2133         return ret;
2134 }
2135
2136 /*
2137  * All control commands are sent via /dev/ublk-control, so we have to check
2138  * the destination device's permission
2139  */
2140 static int ublk_char_dev_permission(struct ublk_device *ub,
2141                 const char *dev_path, int mask)
2142 {
2143         int err;
2144         struct path path;
2145         struct kstat stat;
2146
2147         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2148         if (err)
2149                 return err;
2150
2151         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2152         if (err)
2153                 goto exit;
2154
2155         err = -EPERM;
2156         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2157                 goto exit;
2158
2159         err = inode_permission(&nop_mnt_idmap,
2160                         d_backing_inode(path.dentry), mask);
2161 exit:
2162         path_put(&path);
2163         return err;
2164 }
2165
2166 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2167                 struct io_uring_cmd *cmd)
2168 {
2169         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2170         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2171         void __user *argp = (void __user *)(unsigned long)header->addr;
2172         char *dev_path = NULL;
2173         int ret = 0;
2174         int mask;
2175
2176         if (!unprivileged) {
2177                 if (!capable(CAP_SYS_ADMIN))
2178                         return -EPERM;
2179                 /*
2180                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2181                  * char_dev_path in payload too, since userspace may not
2182                  * know if the specified device is created as unprivileged
2183                  * mode.
2184                  */
2185                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2186                         return 0;
2187         }
2188
2189         /*
2190          * User has to provide the char device path for unprivileged ublk
2191          *
2192          * header->addr always points to the dev path buffer, and
2193          * header->dev_path_len records length of dev path buffer.
2194          */
2195         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2196                 return -EINVAL;
2197
2198         if (header->len < header->dev_path_len)
2199                 return -EINVAL;
2200
2201         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2202         if (!dev_path)
2203                 return -ENOMEM;
2204
2205         ret = -EFAULT;
2206         if (copy_from_user(dev_path, argp, header->dev_path_len))
2207                 goto exit;
2208         dev_path[header->dev_path_len] = 0;
2209
2210         ret = -EINVAL;
2211         switch (_IOC_NR(cmd->cmd_op)) {
2212         case UBLK_CMD_GET_DEV_INFO:
2213         case UBLK_CMD_GET_DEV_INFO2:
2214         case UBLK_CMD_GET_QUEUE_AFFINITY:
2215         case UBLK_CMD_GET_PARAMS:
2216                 mask = MAY_READ;
2217                 break;
2218         case UBLK_CMD_START_DEV:
2219         case UBLK_CMD_STOP_DEV:
2220         case UBLK_CMD_ADD_DEV:
2221         case UBLK_CMD_DEL_DEV:
2222         case UBLK_CMD_SET_PARAMS:
2223         case UBLK_CMD_START_USER_RECOVERY:
2224         case UBLK_CMD_END_USER_RECOVERY:
2225                 mask = MAY_READ | MAY_WRITE;
2226                 break;
2227         default:
2228                 goto exit;
2229         }
2230
2231         ret = ublk_char_dev_permission(ub, dev_path, mask);
2232         if (!ret) {
2233                 header->len -= header->dev_path_len;
2234                 header->addr += header->dev_path_len;
2235         }
2236         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2237                         __func__, ub->ub_number, cmd->cmd_op,
2238                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2239                         dev_path, ret);
2240 exit:
2241         kfree(dev_path);
2242         return ret;
2243 }
2244
2245 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2246                 unsigned int issue_flags)
2247 {
2248         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2249         struct ublk_device *ub = NULL;
2250         u32 cmd_op = cmd->cmd_op;
2251         int ret = -EINVAL;
2252
2253         if (issue_flags & IO_URING_F_NONBLOCK)
2254                 return -EAGAIN;
2255
2256         ublk_ctrl_cmd_dump(cmd);
2257
2258         if (!(issue_flags & IO_URING_F_SQE128))
2259                 goto out;
2260
2261         ret = ublk_check_cmd_op(cmd_op);
2262         if (ret)
2263                 goto out;
2264
2265         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2266                 ret = -ENODEV;
2267                 ub = ublk_get_device_from_id(header->dev_id);
2268                 if (!ub)
2269                         goto out;
2270
2271                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2272                 if (ret)
2273                         goto put_dev;
2274         }
2275
2276         switch (_IOC_NR(cmd_op)) {
2277         case UBLK_CMD_START_DEV:
2278                 ret = ublk_ctrl_start_dev(ub, cmd);
2279                 break;
2280         case UBLK_CMD_STOP_DEV:
2281                 ret = ublk_ctrl_stop_dev(ub);
2282                 break;
2283         case UBLK_CMD_GET_DEV_INFO:
2284         case UBLK_CMD_GET_DEV_INFO2:
2285                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2286                 break;
2287         case UBLK_CMD_ADD_DEV:
2288                 ret = ublk_ctrl_add_dev(cmd);
2289                 break;
2290         case UBLK_CMD_DEL_DEV:
2291                 ret = ublk_ctrl_del_dev(&ub);
2292                 break;
2293         case UBLK_CMD_GET_QUEUE_AFFINITY:
2294                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2295                 break;
2296         case UBLK_CMD_GET_PARAMS:
2297                 ret = ublk_ctrl_get_params(ub, cmd);
2298                 break;
2299         case UBLK_CMD_SET_PARAMS:
2300                 ret = ublk_ctrl_set_params(ub, cmd);
2301                 break;
2302         case UBLK_CMD_START_USER_RECOVERY:
2303                 ret = ublk_ctrl_start_recovery(ub, cmd);
2304                 break;
2305         case UBLK_CMD_END_USER_RECOVERY:
2306                 ret = ublk_ctrl_end_recovery(ub, cmd);
2307                 break;
2308         default:
2309                 ret = -ENOTSUPP;
2310                 break;
2311         }
2312
2313  put_dev:
2314         if (ub)
2315                 ublk_put_device(ub);
2316  out:
2317         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2318         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2319                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2320         return -EIOCBQUEUED;
2321 }
2322
2323 static const struct file_operations ublk_ctl_fops = {
2324         .open           = nonseekable_open,
2325         .uring_cmd      = ublk_ctrl_uring_cmd,
2326         .owner          = THIS_MODULE,
2327         .llseek         = noop_llseek,
2328 };
2329
2330 static struct miscdevice ublk_misc = {
2331         .minor          = MISC_DYNAMIC_MINOR,
2332         .name           = "ublk-control",
2333         .fops           = &ublk_ctl_fops,
2334 };
2335
2336 static int __init ublk_init(void)
2337 {
2338         int ret;
2339
2340         init_waitqueue_head(&ublk_idr_wq);
2341
2342         ret = misc_register(&ublk_misc);
2343         if (ret)
2344                 return ret;
2345
2346         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2347         if (ret)
2348                 goto unregister_mis;
2349
2350         ublk_chr_class = class_create("ublk-char");
2351         if (IS_ERR(ublk_chr_class)) {
2352                 ret = PTR_ERR(ublk_chr_class);
2353                 goto free_chrdev_region;
2354         }
2355         return 0;
2356
2357 free_chrdev_region:
2358         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2359 unregister_mis:
2360         misc_deregister(&ublk_misc);
2361         return ret;
2362 }
2363
2364 static void __exit ublk_exit(void)
2365 {
2366         struct ublk_device *ub;
2367         int id;
2368
2369         idr_for_each_entry(&ublk_index_idr, ub, id)
2370                 ublk_remove(ub);
2371
2372         class_destroy(ublk_chr_class);
2373         misc_deregister(&ublk_misc);
2374
2375         idr_destroy(&ublk_index_idr);
2376         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2377 }
2378
2379 module_init(ublk_init);
2380 module_exit(ublk_exit);
2381
2382 module_param(ublks_max, int, 0444);
2383 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2384
2385 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2386 MODULE_LICENSE("GPL");