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