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