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