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 <linux/kref.h>
  47 #include <uapi/linux/ublk_cmd.h>
  48
  49 #define UBLK_MINORS             (1U << MINORBITS)
  50
  51 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  52 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  53                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  54                 | UBLK_F_NEED_GET_DATA \
  55                 | UBLK_F_USER_RECOVERY \
  56                 | UBLK_F_USER_RECOVERY_REISSUE \
  57                 | UBLK_F_UNPRIVILEGED_DEV \
  58                 | UBLK_F_CMD_IOCTL_ENCODE)
  59
  60 /* All UBLK_PARAM_TYPE_* should be included here */
  61 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | \
  62                 UBLK_PARAM_TYPE_DISCARD | UBLK_PARAM_TYPE_DEVT)
  63
  64 struct ublk_rq_data {
  65         struct llist_node node;
  66
  67         struct kref ref;
  68 };
  69
  70 struct ublk_uring_cmd_pdu {
  71         struct ublk_queue *ubq;
  72 };
  73
  74 /*
  75  * io command is active: sqe cmd is received, and its cqe isn't done
  76  *
  77  * If the flag is set, the io command is owned by ublk driver, and waited
  78  * for incoming blk-mq request from the ublk block device.
  79  *
  80  * If the flag is cleared, the io command will be completed, and owned by
  81  * ublk server.
  82  */
  83 #define UBLK_IO_FLAG_ACTIVE     0x01
  84
  85 /*
  86  * IO command is completed via cqe, and it is being handled by ublksrv, and
  87  * not committed yet
  88  *
  89  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  90  * cross verification
  91  */
  92 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  93
  94 /*
  95  * IO command is aborted, so this flag is set in case of
  96  * !UBLK_IO_FLAG_ACTIVE.
  97  *
  98  * After this flag is observed, any pending or new incoming request
  99  * associated with this io command will be failed immediately
 100  */
 101 #define UBLK_IO_FLAG_ABORTED 0x04
 102
 103 /*
 104  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
 105  * get data buffer address from ublksrv.
 106  *
 107  * Then, bio data could be copied into this data buffer for a WRITE request
 108  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 109  */
 110 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 111
 112 struct ublk_io {
 113         /* userspace buffer address from io cmd */
 114         __u64   addr;
 115         unsigned int flags;
 116         int res;
 117
 118         struct io_uring_cmd *cmd;
 119 };
 120
 121 struct ublk_queue {
 122         int q_id;
 123         int q_depth;
 124
 125         unsigned long flags;
 126         struct task_struct      *ubq_daemon;
 127         char *io_cmd_buf;
 128
 129         struct llist_head       io_cmds;
 130
 131         unsigned long io_addr;  /* mapped vm address */
 132         unsigned int max_io_sz;
 133         bool force_abort;
 134         bool timeout;
 135         unsigned short nr_io_ready;     /* how many ios setup */
 136         struct ublk_device *dev;
 137         struct ublk_io ios[];
 138 };
 139
 140 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 141
 142 struct ublk_device {
 143         struct gendisk          *ub_disk;
 144
 145         char    *__queues;
 146
 147         unsigned int    queue_size;
 148         struct ublksrv_ctrl_dev_info    dev_info;
 149
 150         struct blk_mq_tag_set   tag_set;
 151
 152         struct cdev             cdev;
 153         struct device           cdev_dev;
 154
 155 #define UB_STATE_OPEN           0
 156 #define UB_STATE_USED           1
 157 #define UB_STATE_DELETED        2
 158         unsigned long           state;
 159         int                     ub_number;
 160
 161         struct mutex            mutex;
 162
 163         spinlock_t              mm_lock;
 164         struct mm_struct        *mm;
 165
 166         struct ublk_params      params;
 167
 168         struct completion       completion;
 169         unsigned int            nr_queues_ready;
 170         unsigned int            nr_privileged_daemon;
 171
 172         /*
 173          * Our ubq->daemon may be killed without any notification, so
 174          * monitor each queue's daemon periodically
 175          */
 176         struct delayed_work     monitor_work;
 177         struct work_struct      quiesce_work;
 178         struct work_struct      stop_work;
 179 };
 180
 181 /* header of ublk_params */
 182 struct ublk_params_header {
 183         __u32   len;
 184         __u32   types;
 185 };
 186
 187 static inline void __ublk_complete_rq(struct request *req);
 188 static void ublk_complete_rq(struct kref *ref);
 189
 190 static dev_t ublk_chr_devt;
 191 static struct class *ublk_chr_class;
 192
 193 static DEFINE_IDR(ublk_index_idr);
 194 static DEFINE_SPINLOCK(ublk_idr_lock);
 195 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 196
 197 static DEFINE_MUTEX(ublk_ctl_mutex);
 198
 199 /*
 200  * Max ublk devices allowed to add
 201  *
 202  * It can be extended to one per-user limit in future or even controlled
 203  * by cgroup.
 204  */
 205 static unsigned int ublks_max = 64;
 206 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 207
 208 static struct miscdevice ublk_misc;
 209
 210 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 211 {
 212         struct request_queue *q = ub->ub_disk->queue;
 213         const struct ublk_param_basic *p = &ub->params.basic;
 214
 215         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 216         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 217         blk_queue_io_min(q, 1 << p->io_min_shift);
 218         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 219
 220         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 221                         p->attrs & UBLK_ATTR_FUA);
 222         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 223                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 224         else
 225                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 226
 227         blk_queue_max_hw_sectors(q, p->max_sectors);
 228         blk_queue_chunk_sectors(q, p->chunk_sectors);
 229         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 230
 231         if (p->attrs & UBLK_ATTR_READ_ONLY)
 232                 set_disk_ro(ub->ub_disk, true);
 233
 234         set_capacity(ub->ub_disk, p->dev_sectors);
 235 }
 236
 237 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 238 {
 239         struct request_queue *q = ub->ub_disk->queue;
 240         const struct ublk_param_discard *p = &ub->params.discard;
 241
 242         q->limits.discard_alignment = p->discard_alignment;
 243         q->limits.discard_granularity = p->discard_granularity;
 244         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 245         blk_queue_max_write_zeroes_sectors(q,
 246                         p->max_write_zeroes_sectors);
 247         blk_queue_max_discard_segments(q, p->max_discard_segments);
 248 }
 249
 250 static int ublk_validate_params(const struct ublk_device *ub)
 251 {
 252         /* basic param is the only one which must be set */
 253         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 254                 const struct ublk_param_basic *p = &ub->params.basic;
 255
 256                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 257                         return -EINVAL;
 258
 259                 if (p->logical_bs_shift > p->physical_bs_shift)
 260                         return -EINVAL;
 261
 262                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 263                         return -EINVAL;
 264         } else
 265                 return -EINVAL;
 266
 267         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 268                 const struct ublk_param_discard *p = &ub->params.discard;
 269
 270                 /* So far, only support single segment discard */
 271                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 272                         return -EINVAL;
 273
 274                 if (!p->discard_granularity)
 275                         return -EINVAL;
 276         }
 277
 278         /* dev_t is read-only */
 279         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 280                 return -EINVAL;
 281
 282         return 0;
 283 }
 284
 285 static int ublk_apply_params(struct ublk_device *ub)
 286 {
 287         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 288                 return -EINVAL;
 289
 290         ublk_dev_param_basic_apply(ub);
 291
 292         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 293                 ublk_dev_param_discard_apply(ub);
 294
 295         return 0;
 296 }
 297
 298 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 299 {
 300         return false;
 301 }
 302
 303 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 304                 struct request *req)
 305 {
 306         if (ublk_need_req_ref(ubq)) {
 307                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 308
 309                 kref_init(&data->ref);
 310         }
 311 }
 312
 313 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 314                 struct request *req)
 315 {
 316         if (ublk_need_req_ref(ubq)) {
 317                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 318
 319                 return kref_get_unless_zero(&data->ref);
 320         }
 321
 322         return true;
 323 }
 324
 325 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 326                 struct request *req)
 327 {
 328         if (ublk_need_req_ref(ubq)) {
 329                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 330
 331                 kref_put(&data->ref, ublk_complete_rq);
 332         } else {
 333                 __ublk_complete_rq(req);
 334         }
 335 }
 336
 337 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 338 {
 339         return ubq->flags & UBLK_F_NEED_GET_DATA;
 340 }
 341
 342 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 343 {
 344         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 345                 return ub;
 346         return NULL;
 347 }
 348
 349 static void ublk_put_device(struct ublk_device *ub)
 350 {
 351         put_device(&ub->cdev_dev);
 352 }
 353
 354 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 355                 int qid)
 356 {
 357        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 358 }
 359
 360 static inline bool ublk_rq_has_data(const struct request *rq)
 361 {
 362         return bio_has_data(rq->bio);
 363 }
 364
 365 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 366                 int tag)
 367 {
 368         return (struct ublksrv_io_desc *)
 369                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 370 }
 371
 372 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 373 {
 374         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 375 }
 376
 377 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 378 {
 379         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 380
 381         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 382                         PAGE_SIZE);
 383 }
 384
 385 static inline bool ublk_queue_can_use_recovery_reissue(
 386                 struct ublk_queue *ubq)
 387 {
 388         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 389                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 390 }
 391
 392 static inline bool ublk_queue_can_use_recovery(
 393                 struct ublk_queue *ubq)
 394 {
 395         return ubq->flags & UBLK_F_USER_RECOVERY;
 396 }
 397
 398 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 399 {
 400         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 401 }
 402
 403 static void ublk_free_disk(struct gendisk *disk)
 404 {
 405         struct ublk_device *ub = disk->private_data;
 406
 407         clear_bit(UB_STATE_USED, &ub->state);
 408         put_device(&ub->cdev_dev);
 409 }
 410
 411 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 412                 unsigned int *owner_gid)
 413 {
 414         kuid_t uid;
 415         kgid_t gid;
 416
 417         current_uid_gid(&uid, &gid);
 418
 419         *owner_uid = from_kuid(&init_user_ns, uid);
 420         *owner_gid = from_kgid(&init_user_ns, gid);
 421 }
 422
 423 static int ublk_open(struct block_device *bdev, fmode_t mode)
 424 {
 425         struct ublk_device *ub = bdev->bd_disk->private_data;
 426
 427         if (capable(CAP_SYS_ADMIN))
 428                 return 0;
 429
 430         /*
 431          * If it is one unprivileged device, only owner can open
 432          * the disk. Otherwise it could be one trap made by one
 433          * evil user who grants this disk's privileges to other
 434          * users deliberately.
 435          *
 436          * This way is reasonable too given anyone can create
 437          * unprivileged device, and no need other's grant.
 438          */
 439         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 440                 unsigned int curr_uid, curr_gid;
 441
 442                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 443
 444                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 445                                 ub->dev_info.owner_gid)
 446                         return -EPERM;
 447         }
 448
 449         return 0;
 450 }
 451
 452 static const struct block_device_operations ub_fops = {
 453         .owner =        THIS_MODULE,
 454         .open =         ublk_open,
 455         .free_disk =    ublk_free_disk,
 456 };
 457
 458 #define UBLK_MAX_PIN_PAGES      32
 459
 460 struct ublk_io_iter {
 461         struct page *pages[UBLK_MAX_PIN_PAGES];
 462         struct bio *bio;
 463         struct bvec_iter iter;
 464 };
 465
 466 /* return how many pages are copied */
 467 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 468                 size_t total, size_t pg_off, int dir)
 469 {
 470         unsigned done = 0;
 471         unsigned pg_idx = 0;
 472
 473         while (done < total) {
 474                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 475                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 476                                 (unsigned)(PAGE_SIZE - pg_off));
 477                 void *bv_buf = bvec_kmap_local(&bv);
 478                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 479
 480                 if (dir == ITER_DEST)
 481                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 482                 else
 483                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 484
 485                 kunmap_local(pg_buf);
 486                 kunmap_local(bv_buf);
 487
 488                 /* advance page array */
 489                 pg_off += bytes;
 490                 if (pg_off == PAGE_SIZE) {
 491                         pg_idx += 1;
 492                         pg_off = 0;
 493                 }
 494
 495                 done += bytes;
 496
 497                 /* advance bio */
 498                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 499                 if (!data->iter.bi_size) {
 500                         data->bio = data->bio->bi_next;
 501                         if (data->bio == NULL)
 502                                 break;
 503                         data->iter = data->bio->bi_iter;
 504                 }
 505         }
 506 }
 507
 508 /*
 509  * Copy data between request pages and io_iter, and 'offset'
 510  * is the start point of linear offset of request.
 511  */
 512 static size_t ublk_copy_user_pages(const struct request *req,
 513                 struct iov_iter *uiter, int dir)
 514 {
 515         struct ublk_io_iter iter = {
 516                 .bio    = req->bio,
 517                 .iter   = req->bio->bi_iter,
 518         };
 519         size_t done = 0;
 520
 521         while (iov_iter_count(uiter) && iter.bio) {
 522                 unsigned nr_pages;
 523                 size_t len, off;
 524                 int i;
 525
 526                 len = iov_iter_get_pages2(uiter, iter.pages,
 527                                 iov_iter_count(uiter),
 528                                 UBLK_MAX_PIN_PAGES, &off);
 529                 if (len <= 0)
 530                         return done;
 531
 532                 ublk_copy_io_pages(&iter, len, off, dir);
 533                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 534                 for (i = 0; i < nr_pages; i++) {
 535                         if (dir == ITER_DEST)
 536                                 set_page_dirty(iter.pages[i]);
 537                         put_page(iter.pages[i]);
 538                 }
 539                 done += len;
 540         }
 541
 542         return done;
 543 }
 544
 545 static inline bool ublk_need_map_req(const struct request *req)
 546 {
 547         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 548 }
 549
 550 static inline bool ublk_need_unmap_req(const struct request *req)
 551 {
 552         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_READ;
 553 }
 554
 555 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 556                 struct ublk_io *io)
 557 {
 558         const unsigned int rq_bytes = blk_rq_bytes(req);
 559
 560         /*
 561          * no zero copy, we delay copy WRITE request data into ublksrv
 562          * context and the big benefit is that pinning pages in current
 563          * context is pretty fast, see ublk_pin_user_pages
 564          */
 565         if (ublk_need_map_req(req)) {
 566                 struct iov_iter iter;
 567                 struct iovec iov;
 568                 const int dir = ITER_DEST;
 569
 570                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 571                                 &iov, &iter);
 572
 573                 return ublk_copy_user_pages(req, &iter, dir);
 574         }
 575         return rq_bytes;
 576 }
 577
 578 static int ublk_unmap_io(const struct ublk_queue *ubq,
 579                 const struct request *req,
 580                 struct ublk_io *io)
 581 {
 582         const unsigned int rq_bytes = blk_rq_bytes(req);
 583
 584         if (ublk_need_unmap_req(req)) {
 585                 struct iov_iter iter;
 586                 struct iovec iov;
 587                 const int dir = ITER_SOURCE;
 588
 589                 WARN_ON_ONCE(io->res > rq_bytes);
 590
 591                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 592                                 &iov, &iter);
 593                 return ublk_copy_user_pages(req, &iter, dir);
 594         }
 595         return rq_bytes;
 596 }
 597
 598 static inline unsigned int ublk_req_build_flags(struct request *req)
 599 {
 600         unsigned flags = 0;
 601
 602         if (req->cmd_flags & REQ_FAILFAST_DEV)
 603                 flags |= UBLK_IO_F_FAILFAST_DEV;
 604
 605         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 606                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 607
 608         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 609                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 610
 611         if (req->cmd_flags & REQ_META)
 612                 flags |= UBLK_IO_F_META;
 613
 614         if (req->cmd_flags & REQ_FUA)
 615                 flags |= UBLK_IO_F_FUA;
 616
 617         if (req->cmd_flags & REQ_NOUNMAP)
 618                 flags |= UBLK_IO_F_NOUNMAP;
 619
 620         if (req->cmd_flags & REQ_SWAP)
 621                 flags |= UBLK_IO_F_SWAP;
 622
 623         return flags;
 624 }
 625
 626 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 627 {
 628         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 629         struct ublk_io *io = &ubq->ios[req->tag];
 630         u32 ublk_op;
 631
 632         switch (req_op(req)) {
 633         case REQ_OP_READ:
 634                 ublk_op = UBLK_IO_OP_READ;
 635                 break;
 636         case REQ_OP_WRITE:
 637                 ublk_op = UBLK_IO_OP_WRITE;
 638                 break;
 639         case REQ_OP_FLUSH:
 640                 ublk_op = UBLK_IO_OP_FLUSH;
 641                 break;
 642         case REQ_OP_DISCARD:
 643                 ublk_op = UBLK_IO_OP_DISCARD;
 644                 break;
 645         case REQ_OP_WRITE_ZEROES:
 646                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 647                 break;
 648         default:
 649                 return BLK_STS_IOERR;
 650         }
 651
 652         /* need to translate since kernel may change */
 653         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 654         iod->nr_sectors = blk_rq_sectors(req);
 655         iod->start_sector = blk_rq_pos(req);
 656         iod->addr = io->addr;
 657
 658         return BLK_STS_OK;
 659 }
 660
 661 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 662                 struct io_uring_cmd *ioucmd)
 663 {
 664         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 665 }
 666
 667 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 668 {
 669         return ubq->ubq_daemon->flags & PF_EXITING;
 670 }
 671
 672 /* todo: handle partial completion */
 673 static inline void __ublk_complete_rq(struct request *req)
 674 {
 675         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 676         struct ublk_io *io = &ubq->ios[req->tag];
 677         unsigned int unmapped_bytes;
 678         blk_status_t res = BLK_STS_OK;
 679
 680         /* called from ublk_abort_queue() code path */
 681         if (io->flags & UBLK_IO_FLAG_ABORTED) {
 682                 res = BLK_STS_IOERR;
 683                 goto exit;
 684         }
 685
 686         /* failed read IO if nothing is read */
 687         if (!io->res && req_op(req) == REQ_OP_READ)
 688                 io->res = -EIO;
 689
 690         if (io->res < 0) {
 691                 res = errno_to_blk_status(io->res);
 692                 goto exit;
 693         }
 694
 695         /*
 696          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
 697          * directly.
 698          *
 699          * Both the two needn't unmap.
 700          */
 701         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE)
 702                 goto exit;
 703
 704         /* for READ request, writing data in iod->addr to rq buffers */
 705         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 706
 707         /*
 708          * Extremely impossible since we got data filled in just before
 709          *
 710          * Re-read simply for this unlikely case.
 711          */
 712         if (unlikely(unmapped_bytes < io->res))
 713                 io->res = unmapped_bytes;
 714
 715         if (blk_update_request(req, BLK_STS_OK, io->res))
 716                 blk_mq_requeue_request(req, true);
 717         else
 718                 __blk_mq_end_request(req, BLK_STS_OK);
 719
 720         return;
 721 exit:
 722         blk_mq_end_request(req, res);
 723 }
 724
 725 static void ublk_complete_rq(struct kref *ref)
 726 {
 727         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
 728                         ref);
 729         struct request *req = blk_mq_rq_from_pdu(data);
 730
 731         __ublk_complete_rq(req);
 732 }
 733
 734 /*
 735  * Since __ublk_rq_task_work always fails requests immediately during
 736  * exiting, __ublk_fail_req() is only called from abort context during
 737  * exiting. So lock is unnecessary.
 738  *
 739  * Also aborting may not be started yet, keep in mind that one failed
 740  * request may be issued by block layer again.
 741  */
 742 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
 743                 struct request *req)
 744 {
 745         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 746
 747         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 748                 io->flags |= UBLK_IO_FLAG_ABORTED;
 749                 if (ublk_queue_can_use_recovery_reissue(ubq))
 750                         blk_mq_requeue_request(req, false);
 751                 else
 752                         ublk_put_req_ref(ubq, req);
 753         }
 754 }
 755
 756 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
 757                                 unsigned issue_flags)
 758 {
 759         /* mark this cmd owned by ublksrv */
 760         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 761
 762         /*
 763          * clear ACTIVE since we are done with this sqe/cmd slot
 764          * We can only accept io cmd in case of being not active.
 765          */
 766         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 767
 768         /* tell ublksrv one io request is coming */
 769         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
 770 }
 771
 772 #define UBLK_REQUEUE_DELAY_MS   3
 773
 774 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 775                 struct request *rq)
 776 {
 777         /* We cannot process this rq so just requeue it. */
 778         if (ublk_queue_can_use_recovery(ubq))
 779                 blk_mq_requeue_request(rq, false);
 780         else
 781                 blk_mq_end_request(rq, BLK_STS_IOERR);
 782
 783         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 784 }
 785
 786 static inline void __ublk_rq_task_work(struct request *req,
 787                                        unsigned issue_flags)
 788 {
 789         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 790         int tag = req->tag;
 791         struct ublk_io *io = &ubq->ios[tag];
 792         unsigned int mapped_bytes;
 793
 794         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 795                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 796                         ublk_get_iod(ubq, req->tag)->addr);
 797
 798         /*
 799          * Task is exiting if either:
 800          *
 801          * (1) current != ubq_daemon.
 802          * io_uring_cmd_complete_in_task() tries to run task_work
 803          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 804          *
 805          * (2) current->flags & PF_EXITING.
 806          */
 807         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 808                 __ublk_abort_rq(ubq, req);
 809                 return;
 810         }
 811
 812         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
 813                 /*
 814                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 815                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 816                  * and notify it.
 817                  */
 818                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 819                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 820                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 821                                         __func__, io->cmd->cmd_op, ubq->q_id,
 822                                         req->tag, io->flags);
 823                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
 824                         return;
 825                 }
 826                 /*
 827                  * We have handled UBLK_IO_NEED_GET_DATA command,
 828                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 829                  * do the copy work.
 830                  */
 831                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 832                 /* update iod->addr because ublksrv may have passed a new io buffer */
 833                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 834                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 835                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 836                                 ublk_get_iod(ubq, req->tag)->addr);
 837         }
 838
 839         mapped_bytes = ublk_map_io(ubq, req, io);
 840
 841         /* partially mapped, update io descriptor */
 842         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 843                 /*
 844                  * Nothing mapped, retry until we succeed.
 845                  *
 846                  * We may never succeed in mapping any bytes here because
 847                  * of OOM. TODO: reserve one buffer with single page pinned
 848                  * for providing forward progress guarantee.
 849                  */
 850                 if (unlikely(!mapped_bytes)) {
 851                         blk_mq_requeue_request(req, false);
 852                         blk_mq_delay_kick_requeue_list(req->q,
 853                                         UBLK_REQUEUE_DELAY_MS);
 854                         return;
 855                 }
 856
 857                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 858                         mapped_bytes >> 9;
 859         }
 860
 861         ublk_init_req_ref(ubq, req);
 862         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
 863 }
 864
 865 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
 866                                         unsigned issue_flags)
 867 {
 868         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 869         struct ublk_rq_data *data, *tmp;
 870
 871         io_cmds = llist_reverse_order(io_cmds);
 872         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 873                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
 874 }
 875
 876 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 877 {
 878         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 879         struct ublk_rq_data *data, *tmp;
 880
 881         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 882                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 883 }
 884
 885 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
 886 {
 887         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 888         struct ublk_queue *ubq = pdu->ubq;
 889
 890         ublk_forward_io_cmds(ubq, issue_flags);
 891 }
 892
 893 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 894 {
 895         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 896         struct ublk_io *io;
 897
 898         if (!llist_add(&data->node, &ubq->io_cmds))
 899                 return;
 900
 901         io = &ubq->ios[rq->tag];
 902         /*
 903          * If the check pass, we know that this is a re-issued request aborted
 904          * previously in monitor_work because the ubq_daemon(cmd's task) is
 905          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 906          * because this ioucmd's io_uring context may be freed now if no inflight
 907          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 908          *
 909          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 910          * the tag). Then the request is re-started(allocating the tag) and we are here.
 911          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 912          * guarantees that here is a re-issued request aborted previously.
 913          */
 914         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 915                 ublk_abort_io_cmds(ubq);
 916         } else {
 917                 struct io_uring_cmd *cmd = io->cmd;
 918                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 919
 920                 pdu->ubq = ubq;
 921                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 922         }
 923 }
 924
 925 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
 926 {
 927         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
 928
 929         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
 930                 if (!ubq->timeout) {
 931                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
 932                         ubq->timeout = true;
 933                 }
 934
 935                 return BLK_EH_DONE;
 936         }
 937
 938         return BLK_EH_RESET_TIMER;
 939 }
 940
 941 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 942                 const struct blk_mq_queue_data *bd)
 943 {
 944         struct ublk_queue *ubq = hctx->driver_data;
 945         struct request *rq = bd->rq;
 946         blk_status_t res;
 947
 948         /* fill iod to slot in io cmd buffer */
 949         res = ublk_setup_iod(ubq, rq);
 950         if (unlikely(res != BLK_STS_OK))
 951                 return BLK_STS_IOERR;
 952
 953         /* With recovery feature enabled, force_abort is set in
 954          * ublk_stop_dev() before calling del_gendisk(). We have to
 955          * abort all requeued and new rqs here to let del_gendisk()
 956          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 957          * to avoid UAF on io_uring ctx.
 958          *
 959          * Note: force_abort is guaranteed to be seen because it is set
 960          * before request queue is unqiuesced.
 961          */
 962         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 963                 return BLK_STS_IOERR;
 964
 965         blk_mq_start_request(bd->rq);
 966
 967         if (unlikely(ubq_daemon_is_dying(ubq))) {
 968                 __ublk_abort_rq(ubq, rq);
 969                 return BLK_STS_OK;
 970         }
 971
 972         ublk_queue_cmd(ubq, rq);
 973
 974         return BLK_STS_OK;
 975 }
 976
 977 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 978                 unsigned int hctx_idx)
 979 {
 980         struct ublk_device *ub = driver_data;
 981         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 982
 983         hctx->driver_data = ubq;
 984         return 0;
 985 }
 986
 987 static const struct blk_mq_ops ublk_mq_ops = {
 988         .queue_rq       = ublk_queue_rq,
 989         .init_hctx      = ublk_init_hctx,
 990         .timeout        = ublk_timeout,
 991 };
 992
 993 static int ublk_ch_open(struct inode *inode, struct file *filp)
 994 {
 995         struct ublk_device *ub = container_of(inode->i_cdev,
 996                         struct ublk_device, cdev);
 997
 998         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 999                 return -EBUSY;
1000         filp->private_data = ub;
1001         return 0;
1002 }
1003
1004 static int ublk_ch_release(struct inode *inode, struct file *filp)
1005 {
1006         struct ublk_device *ub = filp->private_data;
1007
1008         clear_bit(UB_STATE_OPEN, &ub->state);
1009         return 0;
1010 }
1011
1012 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1013 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1014 {
1015         struct ublk_device *ub = filp->private_data;
1016         size_t sz = vma->vm_end - vma->vm_start;
1017         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1018         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1019         int q_id, ret = 0;
1020
1021         spin_lock(&ub->mm_lock);
1022         if (!ub->mm)
1023                 ub->mm = current->mm;
1024         if (current->mm != ub->mm)
1025                 ret = -EINVAL;
1026         spin_unlock(&ub->mm_lock);
1027
1028         if (ret)
1029                 return ret;
1030
1031         if (vma->vm_flags & VM_WRITE)
1032                 return -EPERM;
1033
1034         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1035         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1036                 return -EINVAL;
1037
1038         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1039         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1040                         __func__, q_id, current->pid, vma->vm_start,
1041                         phys_off, (unsigned long)sz);
1042
1043         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1044                 return -EINVAL;
1045
1046         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1047         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1048 }
1049
1050 static void ublk_commit_completion(struct ublk_device *ub,
1051                 const struct ublksrv_io_cmd *ub_cmd)
1052 {
1053         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1054         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1055         struct ublk_io *io = &ubq->ios[tag];
1056         struct request *req;
1057
1058         /* now this cmd slot is owned by nbd driver */
1059         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1060         io->res = ub_cmd->result;
1061
1062         /* find the io request and complete */
1063         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1064
1065         if (req && likely(!blk_should_fake_timeout(req->q)))
1066                 ublk_put_req_ref(ubq, req);
1067 }
1068
1069 /*
1070  * When ->ubq_daemon is exiting, either new request is ended immediately,
1071  * or any queued io command is drained, so it is safe to abort queue
1072  * lockless
1073  */
1074 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1075 {
1076         int i;
1077
1078         if (!ublk_get_device(ub))
1079                 return;
1080
1081         for (i = 0; i < ubq->q_depth; i++) {
1082                 struct ublk_io *io = &ubq->ios[i];
1083
1084                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1085                         struct request *rq;
1086
1087                         /*
1088                          * Either we fail the request or ublk_rq_task_work_fn
1089                          * will do it
1090                          */
1091                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1092                         if (rq)
1093                                 __ublk_fail_req(ubq, io, rq);
1094                 }
1095         }
1096         ublk_put_device(ub);
1097 }
1098
1099 static void ublk_daemon_monitor_work(struct work_struct *work)
1100 {
1101         struct ublk_device *ub =
1102                 container_of(work, struct ublk_device, monitor_work.work);
1103         int i;
1104
1105         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1106                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1107
1108                 if (ubq_daemon_is_dying(ubq)) {
1109                         if (ublk_queue_can_use_recovery(ubq))
1110                                 schedule_work(&ub->quiesce_work);
1111                         else
1112                                 schedule_work(&ub->stop_work);
1113
1114                         /* abort queue is for making forward progress */
1115                         ublk_abort_queue(ub, ubq);
1116                 }
1117         }
1118
1119         /*
1120          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1121          * after ublk_remove() or __ublk_quiesce_dev() is started.
1122          *
1123          * No need ub->mutex, monitor work are canceled after state is marked
1124          * as not LIVE, so new state is observed reliably.
1125          */
1126         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1127                 schedule_delayed_work(&ub->monitor_work,
1128                                 UBLK_DAEMON_MONITOR_PERIOD);
1129 }
1130
1131 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1132 {
1133         return ubq->nr_io_ready == ubq->q_depth;
1134 }
1135
1136 static void ublk_cancel_queue(struct ublk_queue *ubq)
1137 {
1138         int i;
1139
1140         if (!ublk_queue_ready(ubq))
1141                 return;
1142
1143         for (i = 0; i < ubq->q_depth; i++) {
1144                 struct ublk_io *io = &ubq->ios[i];
1145
1146                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1147                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1148                                                 IO_URING_F_UNLOCKED);
1149         }
1150
1151         /* all io commands are canceled */
1152         ubq->nr_io_ready = 0;
1153 }
1154
1155 /* Cancel all pending commands, must be called after del_gendisk() returns */
1156 static void ublk_cancel_dev(struct ublk_device *ub)
1157 {
1158         int i;
1159
1160         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1161                 ublk_cancel_queue(ublk_get_queue(ub, i));
1162 }
1163
1164 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1165 {
1166         bool *idle = data;
1167
1168         if (blk_mq_request_started(rq)) {
1169                 *idle = false;
1170                 return false;
1171         }
1172         return true;
1173 }
1174
1175 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1176 {
1177         bool idle;
1178
1179         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1180         while (true) {
1181                 idle = true;
1182                 blk_mq_tagset_busy_iter(&ub->tag_set,
1183                                 ublk_check_inflight_rq, &idle);
1184                 if (idle)
1185                         break;
1186                 msleep(UBLK_REQUEUE_DELAY_MS);
1187         }
1188 }
1189
1190 static void __ublk_quiesce_dev(struct ublk_device *ub)
1191 {
1192         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1193                         __func__, ub->dev_info.dev_id,
1194                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1195                         "LIVE" : "QUIESCED");
1196         blk_mq_quiesce_queue(ub->ub_disk->queue);
1197         ublk_wait_tagset_rqs_idle(ub);
1198         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1199         ublk_cancel_dev(ub);
1200         /* we are going to release task_struct of ubq_daemon and resets
1201          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1202          * Besides, monitor_work is not necessary in QUIESCED state since we have
1203          * already scheduled quiesce_work and quiesced all ubqs.
1204          *
1205          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1206          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1207          */
1208         cancel_delayed_work_sync(&ub->monitor_work);
1209 }
1210
1211 static void ublk_quiesce_work_fn(struct work_struct *work)
1212 {
1213         struct ublk_device *ub =
1214                 container_of(work, struct ublk_device, quiesce_work);
1215
1216         mutex_lock(&ub->mutex);
1217         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1218                 goto unlock;
1219         __ublk_quiesce_dev(ub);
1220  unlock:
1221         mutex_unlock(&ub->mutex);
1222 }
1223
1224 static void ublk_unquiesce_dev(struct ublk_device *ub)
1225 {
1226         int i;
1227
1228         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1229                         __func__, ub->dev_info.dev_id,
1230                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1231                         "LIVE" : "QUIESCED");
1232         /* quiesce_work has run. We let requeued rqs be aborted
1233          * before running fallback_wq. "force_abort" must be seen
1234          * after request queue is unqiuesced. Then del_gendisk()
1235          * can move on.
1236          */
1237         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1238                 ublk_get_queue(ub, i)->force_abort = true;
1239
1240         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1241         /* We may have requeued some rqs in ublk_quiesce_queue() */
1242         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1243 }
1244
1245 static void ublk_stop_dev(struct ublk_device *ub)
1246 {
1247         mutex_lock(&ub->mutex);
1248         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1249                 goto unlock;
1250         if (ublk_can_use_recovery(ub)) {
1251                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1252                         __ublk_quiesce_dev(ub);
1253                 ublk_unquiesce_dev(ub);
1254         }
1255         del_gendisk(ub->ub_disk);
1256         ub->dev_info.state = UBLK_S_DEV_DEAD;
1257         ub->dev_info.ublksrv_pid = -1;
1258         put_disk(ub->ub_disk);
1259         ub->ub_disk = NULL;
1260  unlock:
1261         ublk_cancel_dev(ub);
1262         mutex_unlock(&ub->mutex);
1263         cancel_delayed_work_sync(&ub->monitor_work);
1264 }
1265
1266 /* device can only be started after all IOs are ready */
1267 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1268 {
1269         mutex_lock(&ub->mutex);
1270         ubq->nr_io_ready++;
1271         if (ublk_queue_ready(ubq)) {
1272                 ubq->ubq_daemon = current;
1273                 get_task_struct(ubq->ubq_daemon);
1274                 ub->nr_queues_ready++;
1275
1276                 if (capable(CAP_SYS_ADMIN))
1277                         ub->nr_privileged_daemon++;
1278         }
1279         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1280                 complete_all(&ub->completion);
1281         mutex_unlock(&ub->mutex);
1282 }
1283
1284 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1285                 int tag)
1286 {
1287         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1288         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1289
1290         ublk_queue_cmd(ubq, req);
1291 }
1292
1293 static inline int ublk_check_cmd_op(u32 cmd_op)
1294 {
1295         u32 ioc_type = _IOC_TYPE(cmd_op);
1296
1297         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1298                 return -EOPNOTSUPP;
1299
1300         if (ioc_type != 'u' && ioc_type != 0)
1301                 return -EOPNOTSUPP;
1302
1303         return 0;
1304 }
1305
1306 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1307                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1308 {
1309         io->cmd = cmd;
1310         io->flags |= UBLK_IO_FLAG_ACTIVE;
1311         io->addr = buf_addr;
1312 }
1313
1314 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1315                                unsigned int issue_flags,
1316                                const struct ublksrv_io_cmd *ub_cmd)
1317 {
1318         struct ublk_device *ub = cmd->file->private_data;
1319         struct ublk_queue *ubq;
1320         struct ublk_io *io;
1321         u32 cmd_op = cmd->cmd_op;
1322         unsigned tag = ub_cmd->tag;
1323         int ret = -EINVAL;
1324         struct request *req;
1325
1326         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1327                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1328                         ub_cmd->result);
1329
1330         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1331                 goto out;
1332
1333         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1334         if (!ubq || ub_cmd->q_id != ubq->q_id)
1335                 goto out;
1336
1337         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1338                 goto out;
1339
1340         if (tag >= ubq->q_depth)
1341                 goto out;
1342
1343         io = &ubq->ios[tag];
1344
1345         /* there is pending io cmd, something must be wrong */
1346         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1347                 ret = -EBUSY;
1348                 goto out;
1349         }
1350
1351         /*
1352          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1353          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1354          */
1355         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1356                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1357                 goto out;
1358
1359         ret = ublk_check_cmd_op(cmd_op);
1360         if (ret)
1361                 goto out;
1362
1363         ret = -EINVAL;
1364         switch (_IOC_NR(cmd_op)) {
1365         case UBLK_IO_FETCH_REQ:
1366                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1367                 if (ublk_queue_ready(ubq)) {
1368                         ret = -EBUSY;
1369                         goto out;
1370                 }
1371                 /*
1372                  * The io is being handled by server, so COMMIT_RQ is expected
1373                  * instead of FETCH_REQ
1374                  */
1375                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1376                         goto out;
1377                 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1378                 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1379                         goto out;
1380
1381                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1382                 ublk_mark_io_ready(ub, ubq);
1383                 break;
1384         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1385                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1386                 /*
1387                  * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1388                  * not enabled or it is Read IO.
1389                  */
1390                 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1391                         goto out;
1392                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1393                         goto out;
1394                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1395                 ublk_commit_completion(ub, ub_cmd);
1396                 break;
1397         case UBLK_IO_NEED_GET_DATA:
1398                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1399                         goto out;
1400                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1401                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1402                 break;
1403         default:
1404                 goto out;
1405         }
1406         return -EIOCBQUEUED;
1407
1408  out:
1409         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1410         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1411                         __func__, cmd_op, tag, ret, io->flags);
1412         return -EIOCBQUEUED;
1413 }
1414
1415 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1416 {
1417         /*
1418          * Not necessary for async retry, but let's keep it simple and always
1419          * copy the values to avoid any potential reuse.
1420          */
1421         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1422         const struct ublksrv_io_cmd ub_cmd = {
1423                 .q_id = READ_ONCE(ub_src->q_id),
1424                 .tag = READ_ONCE(ub_src->tag),
1425                 .result = READ_ONCE(ub_src->result),
1426                 .addr = READ_ONCE(ub_src->addr)
1427         };
1428
1429         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1430 }
1431
1432 static const struct file_operations ublk_ch_fops = {
1433         .owner = THIS_MODULE,
1434         .open = ublk_ch_open,
1435         .release = ublk_ch_release,
1436         .llseek = no_llseek,
1437         .uring_cmd = ublk_ch_uring_cmd,
1438         .mmap = ublk_ch_mmap,
1439 };
1440
1441 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1442 {
1443         int size = ublk_queue_cmd_buf_size(ub, q_id);
1444         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1445
1446         if (ubq->ubq_daemon)
1447                 put_task_struct(ubq->ubq_daemon);
1448         if (ubq->io_cmd_buf)
1449                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1450 }
1451
1452 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1453 {
1454         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1455         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1456         void *ptr;
1457         int size;
1458
1459         ubq->flags = ub->dev_info.flags;
1460         ubq->q_id = q_id;
1461         ubq->q_depth = ub->dev_info.queue_depth;
1462         size = ublk_queue_cmd_buf_size(ub, q_id);
1463
1464         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1465         if (!ptr)
1466                 return -ENOMEM;
1467
1468         ubq->io_cmd_buf = ptr;
1469         ubq->dev = ub;
1470         return 0;
1471 }
1472
1473 static void ublk_deinit_queues(struct ublk_device *ub)
1474 {
1475         int nr_queues = ub->dev_info.nr_hw_queues;
1476         int i;
1477
1478         if (!ub->__queues)
1479                 return;
1480
1481         for (i = 0; i < nr_queues; i++)
1482                 ublk_deinit_queue(ub, i);
1483         kfree(ub->__queues);
1484 }
1485
1486 static int ublk_init_queues(struct ublk_device *ub)
1487 {
1488         int nr_queues = ub->dev_info.nr_hw_queues;
1489         int depth = ub->dev_info.queue_depth;
1490         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1491         int i, ret = -ENOMEM;
1492
1493         ub->queue_size = ubq_size;
1494         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1495         if (!ub->__queues)
1496                 return ret;
1497
1498         for (i = 0; i < nr_queues; i++) {
1499                 if (ublk_init_queue(ub, i))
1500                         goto fail;
1501         }
1502
1503         init_completion(&ub->completion);
1504         return 0;
1505
1506  fail:
1507         ublk_deinit_queues(ub);
1508         return ret;
1509 }
1510
1511 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1512 {
1513         int i = idx;
1514         int err;
1515
1516         spin_lock(&ublk_idr_lock);
1517         /* allocate id, if @id >= 0, we're requesting that specific id */
1518         if (i >= 0) {
1519                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1520                 if (err == -ENOSPC)
1521                         err = -EEXIST;
1522         } else {
1523                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1524         }
1525         spin_unlock(&ublk_idr_lock);
1526
1527         if (err >= 0)
1528                 ub->ub_number = err;
1529
1530         return err;
1531 }
1532
1533 static void ublk_free_dev_number(struct ublk_device *ub)
1534 {
1535         spin_lock(&ublk_idr_lock);
1536         idr_remove(&ublk_index_idr, ub->ub_number);
1537         wake_up_all(&ublk_idr_wq);
1538         spin_unlock(&ublk_idr_lock);
1539 }
1540
1541 static void ublk_cdev_rel(struct device *dev)
1542 {
1543         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1544
1545         blk_mq_free_tag_set(&ub->tag_set);
1546         ublk_deinit_queues(ub);
1547         ublk_free_dev_number(ub);
1548         mutex_destroy(&ub->mutex);
1549         kfree(ub);
1550 }
1551
1552 static int ublk_add_chdev(struct ublk_device *ub)
1553 {
1554         struct device *dev = &ub->cdev_dev;
1555         int minor = ub->ub_number;
1556         int ret;
1557
1558         dev->parent = ublk_misc.this_device;
1559         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1560         dev->class = ublk_chr_class;
1561         dev->release = ublk_cdev_rel;
1562         device_initialize(dev);
1563
1564         ret = dev_set_name(dev, "ublkc%d", minor);
1565         if (ret)
1566                 goto fail;
1567
1568         cdev_init(&ub->cdev, &ublk_ch_fops);
1569         ret = cdev_device_add(&ub->cdev, dev);
1570         if (ret)
1571                 goto fail;
1572
1573         ublks_added++;
1574         return 0;
1575  fail:
1576         put_device(dev);
1577         return ret;
1578 }
1579
1580 static void ublk_stop_work_fn(struct work_struct *work)
1581 {
1582         struct ublk_device *ub =
1583                 container_of(work, struct ublk_device, stop_work);
1584
1585         ublk_stop_dev(ub);
1586 }
1587
1588 /* align max io buffer size with PAGE_SIZE */
1589 static void ublk_align_max_io_size(struct ublk_device *ub)
1590 {
1591         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1592
1593         ub->dev_info.max_io_buf_bytes =
1594                 round_down(max_io_bytes, PAGE_SIZE);
1595 }
1596
1597 static int ublk_add_tag_set(struct ublk_device *ub)
1598 {
1599         ub->tag_set.ops = &ublk_mq_ops;
1600         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1601         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1602         ub->tag_set.numa_node = NUMA_NO_NODE;
1603         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1604         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1605         ub->tag_set.driver_data = ub;
1606         return blk_mq_alloc_tag_set(&ub->tag_set);
1607 }
1608
1609 static void ublk_remove(struct ublk_device *ub)
1610 {
1611         ublk_stop_dev(ub);
1612         cancel_work_sync(&ub->stop_work);
1613         cancel_work_sync(&ub->quiesce_work);
1614         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1615         put_device(&ub->cdev_dev);
1616         ublks_added--;
1617 }
1618
1619 static struct ublk_device *ublk_get_device_from_id(int idx)
1620 {
1621         struct ublk_device *ub = NULL;
1622
1623         if (idx < 0)
1624                 return NULL;
1625
1626         spin_lock(&ublk_idr_lock);
1627         ub = idr_find(&ublk_index_idr, idx);
1628         if (ub)
1629                 ub = ublk_get_device(ub);
1630         spin_unlock(&ublk_idr_lock);
1631
1632         return ub;
1633 }
1634
1635 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
1636 {
1637         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1638         int ublksrv_pid = (int)header->data[0];
1639         struct gendisk *disk;
1640         int ret = -EINVAL;
1641
1642         if (ublksrv_pid <= 0)
1643                 return -EINVAL;
1644
1645         wait_for_completion_interruptible(&ub->completion);
1646
1647         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1648
1649         mutex_lock(&ub->mutex);
1650         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1651             test_bit(UB_STATE_USED, &ub->state)) {
1652                 ret = -EEXIST;
1653                 goto out_unlock;
1654         }
1655
1656         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
1657         if (IS_ERR(disk)) {
1658                 ret = PTR_ERR(disk);
1659                 goto out_unlock;
1660         }
1661         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1662         disk->fops = &ub_fops;
1663         disk->private_data = ub;
1664
1665         ub->dev_info.ublksrv_pid = ublksrv_pid;
1666         ub->ub_disk = disk;
1667
1668         ret = ublk_apply_params(ub);
1669         if (ret)
1670                 goto out_put_disk;
1671
1672         /* don't probe partitions if any one ubq daemon is un-trusted */
1673         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1674                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1675
1676         get_device(&ub->cdev_dev);
1677         ub->dev_info.state = UBLK_S_DEV_LIVE;
1678         ret = add_disk(disk);
1679         if (ret) {
1680                 /*
1681                  * Has to drop the reference since ->free_disk won't be
1682                  * called in case of add_disk failure.
1683                  */
1684                 ub->dev_info.state = UBLK_S_DEV_DEAD;
1685                 ublk_put_device(ub);
1686                 goto out_put_disk;
1687         }
1688         set_bit(UB_STATE_USED, &ub->state);
1689 out_put_disk:
1690         if (ret)
1691                 put_disk(disk);
1692 out_unlock:
1693         mutex_unlock(&ub->mutex);
1694         return ret;
1695 }
1696
1697 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
1698                 struct io_uring_cmd *cmd)
1699 {
1700         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1701         void __user *argp = (void __user *)(unsigned long)header->addr;
1702         cpumask_var_t cpumask;
1703         unsigned long queue;
1704         unsigned int retlen;
1705         unsigned int i;
1706         int ret;
1707
1708         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1709                 return -EINVAL;
1710         if (header->len & (sizeof(unsigned long)-1))
1711                 return -EINVAL;
1712         if (!header->addr)
1713                 return -EINVAL;
1714
1715         queue = header->data[0];
1716         if (queue >= ub->dev_info.nr_hw_queues)
1717                 return -EINVAL;
1718
1719         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1720                 return -ENOMEM;
1721
1722         for_each_possible_cpu(i) {
1723                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1724                         cpumask_set_cpu(i, cpumask);
1725         }
1726
1727         ret = -EFAULT;
1728         retlen = min_t(unsigned short, header->len, cpumask_size());
1729         if (copy_to_user(argp, cpumask, retlen))
1730                 goto out_free_cpumask;
1731         if (retlen != header->len &&
1732             clear_user(argp + retlen, header->len - retlen))
1733                 goto out_free_cpumask;
1734
1735         ret = 0;
1736 out_free_cpumask:
1737         free_cpumask_var(cpumask);
1738         return ret;
1739 }
1740
1741 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1742 {
1743         pr_devel("%s: dev id %d flags %llx\n", __func__,
1744                         info->dev_id, info->flags);
1745         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1746                         info->nr_hw_queues, info->queue_depth);
1747 }
1748
1749 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1750 {
1751         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1752         void __user *argp = (void __user *)(unsigned long)header->addr;
1753         struct ublksrv_ctrl_dev_info info;
1754         struct ublk_device *ub;
1755         int ret = -EINVAL;
1756
1757         if (header->len < sizeof(info) || !header->addr)
1758                 return -EINVAL;
1759         if (header->queue_id != (u16)-1) {
1760                 pr_warn("%s: queue_id is wrong %x\n",
1761                         __func__, header->queue_id);
1762                 return -EINVAL;
1763         }
1764
1765         if (copy_from_user(&info, argp, sizeof(info)))
1766                 return -EFAULT;
1767
1768         if (capable(CAP_SYS_ADMIN))
1769                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
1770         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
1771                 return -EPERM;
1772
1773         /*
1774          * unprivileged device can't be trusted, but RECOVERY and
1775          * RECOVERY_REISSUE still may hang error handling, so can't
1776          * support recovery features for unprivileged ublk now
1777          *
1778          * TODO: provide forward progress for RECOVERY handler, so that
1779          * unprivileged device can benefit from it
1780          */
1781         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
1782                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
1783                                 UBLK_F_USER_RECOVERY);
1784
1785         /* the created device is always owned by current user */
1786         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
1787
1788         if (header->dev_id != info.dev_id) {
1789                 pr_warn("%s: dev id not match %u %u\n",
1790                         __func__, header->dev_id, info.dev_id);
1791                 return -EINVAL;
1792         }
1793
1794         ublk_dump_dev_info(&info);
1795
1796         ret = mutex_lock_killable(&ublk_ctl_mutex);
1797         if (ret)
1798                 return ret;
1799
1800         ret = -EACCES;
1801         if (ublks_added >= ublks_max)
1802                 goto out_unlock;
1803
1804         ret = -ENOMEM;
1805         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1806         if (!ub)
1807                 goto out_unlock;
1808         mutex_init(&ub->mutex);
1809         spin_lock_init(&ub->mm_lock);
1810         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1811         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1812         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1813
1814         ret = ublk_alloc_dev_number(ub, header->dev_id);
1815         if (ret < 0)
1816                 goto out_free_ub;
1817
1818         memcpy(&ub->dev_info, &info, sizeof(info));
1819
1820         /* update device id */
1821         ub->dev_info.dev_id = ub->ub_number;
1822
1823         /*
1824          * 64bit flags will be copied back to userspace as feature
1825          * negotiation result, so have to clear flags which driver
1826          * doesn't support yet, then userspace can get correct flags
1827          * (features) to handle.
1828          */
1829         ub->dev_info.flags &= UBLK_F_ALL;
1830
1831         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
1832                 UBLK_F_URING_CMD_COMP_IN_TASK;
1833
1834         /* We are not ready to support zero copy */
1835         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1836
1837         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1838                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1839         ublk_align_max_io_size(ub);
1840
1841         ret = ublk_init_queues(ub);
1842         if (ret)
1843                 goto out_free_dev_number;
1844
1845         ret = ublk_add_tag_set(ub);
1846         if (ret)
1847                 goto out_deinit_queues;
1848
1849         ret = -EFAULT;
1850         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1851                 goto out_free_tag_set;
1852
1853         /*
1854          * Add the char dev so that ublksrv daemon can be setup.
1855          * ublk_add_chdev() will cleanup everything if it fails.
1856          */
1857         ret = ublk_add_chdev(ub);
1858         goto out_unlock;
1859
1860 out_free_tag_set:
1861         blk_mq_free_tag_set(&ub->tag_set);
1862 out_deinit_queues:
1863         ublk_deinit_queues(ub);
1864 out_free_dev_number:
1865         ublk_free_dev_number(ub);
1866 out_free_ub:
1867         mutex_destroy(&ub->mutex);
1868         kfree(ub);
1869 out_unlock:
1870         mutex_unlock(&ublk_ctl_mutex);
1871         return ret;
1872 }
1873
1874 static inline bool ublk_idr_freed(int id)
1875 {
1876         void *ptr;
1877
1878         spin_lock(&ublk_idr_lock);
1879         ptr = idr_find(&ublk_index_idr, id);
1880         spin_unlock(&ublk_idr_lock);
1881
1882         return ptr == NULL;
1883 }
1884
1885 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
1886 {
1887         struct ublk_device *ub = *p_ub;
1888         int idx = ub->ub_number;
1889         int ret;
1890
1891         ret = mutex_lock_killable(&ublk_ctl_mutex);
1892         if (ret)
1893                 return ret;
1894
1895         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
1896                 ublk_remove(ub);
1897                 set_bit(UB_STATE_DELETED, &ub->state);
1898         }
1899
1900         /* Mark the reference as consumed */
1901         *p_ub = NULL;
1902         ublk_put_device(ub);
1903         mutex_unlock(&ublk_ctl_mutex);
1904
1905         /*
1906          * Wait until the idr is removed, then it can be reused after
1907          * DEL_DEV command is returned.
1908          *
1909          * If we returns because of user interrupt, future delete command
1910          * may come:
1911          *
1912          * - the device number isn't freed, this device won't or needn't
1913          *   be deleted again, since UB_STATE_DELETED is set, and device
1914          *   will be released after the last reference is dropped
1915          *
1916          * - the device number is freed already, we will not find this
1917          *   device via ublk_get_device_from_id()
1918          */
1919         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
1920
1921         return 0;
1922 }
1923
1924 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1925 {
1926         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1927
1928         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1929                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1930                         header->data[0], header->addr, header->len);
1931 }
1932
1933 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
1934 {
1935         ublk_stop_dev(ub);
1936         cancel_work_sync(&ub->stop_work);
1937         cancel_work_sync(&ub->quiesce_work);
1938
1939         return 0;
1940 }
1941
1942 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
1943                 struct io_uring_cmd *cmd)
1944 {
1945         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1946         void __user *argp = (void __user *)(unsigned long)header->addr;
1947
1948         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1949                 return -EINVAL;
1950
1951         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1952                 return -EFAULT;
1953
1954         return 0;
1955 }
1956
1957 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
1958 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
1959 {
1960         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
1961         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
1962
1963         if (ub->ub_disk) {
1964                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
1965                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
1966         } else {
1967                 ub->params.devt.disk_major = 0;
1968                 ub->params.devt.disk_minor = 0;
1969         }
1970         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
1971 }
1972
1973 static int ublk_ctrl_get_params(struct ublk_device *ub,
1974                 struct io_uring_cmd *cmd)
1975 {
1976         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
1977         void __user *argp = (void __user *)(unsigned long)header->addr;
1978         struct ublk_params_header ph;
1979         int ret;
1980
1981         if (header->len <= sizeof(ph) || !header->addr)
1982                 return -EINVAL;
1983
1984         if (copy_from_user(&ph, argp, sizeof(ph)))
1985                 return -EFAULT;
1986
1987         if (ph.len > header->len || !ph.len)
1988                 return -EINVAL;
1989
1990         if (ph.len > sizeof(struct ublk_params))
1991                 ph.len = sizeof(struct ublk_params);
1992
1993         mutex_lock(&ub->mutex);
1994         ublk_ctrl_fill_params_devt(ub);
1995         if (copy_to_user(argp, &ub->params, ph.len))
1996                 ret = -EFAULT;
1997         else
1998                 ret = 0;
1999         mutex_unlock(&ub->mutex);
2000
2001         return ret;
2002 }
2003
2004 static int ublk_ctrl_set_params(struct ublk_device *ub,
2005                 struct io_uring_cmd *cmd)
2006 {
2007         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2008         void __user *argp = (void __user *)(unsigned long)header->addr;
2009         struct ublk_params_header ph;
2010         int ret = -EFAULT;
2011
2012         if (header->len <= sizeof(ph) || !header->addr)
2013                 return -EINVAL;
2014
2015         if (copy_from_user(&ph, argp, sizeof(ph)))
2016                 return -EFAULT;
2017
2018         if (ph.len > header->len || !ph.len || !ph.types)
2019                 return -EINVAL;
2020
2021         if (ph.len > sizeof(struct ublk_params))
2022                 ph.len = sizeof(struct ublk_params);
2023
2024         /* parameters can only be changed when device isn't live */
2025         mutex_lock(&ub->mutex);
2026         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2027                 ret = -EACCES;
2028         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2029                 ret = -EFAULT;
2030         } else {
2031                 /* clear all we don't support yet */
2032                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2033                 ret = ublk_validate_params(ub);
2034                 if (ret)
2035                         ub->params.types = 0;
2036         }
2037         mutex_unlock(&ub->mutex);
2038
2039         return ret;
2040 }
2041
2042 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2043 {
2044         int i;
2045
2046         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2047         /* All old ioucmds have to be completed */
2048         WARN_ON_ONCE(ubq->nr_io_ready);
2049         /* old daemon is PF_EXITING, put it now */
2050         put_task_struct(ubq->ubq_daemon);
2051         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2052         ubq->ubq_daemon = NULL;
2053         ubq->timeout = false;
2054
2055         for (i = 0; i < ubq->q_depth; i++) {
2056                 struct ublk_io *io = &ubq->ios[i];
2057
2058                 /* forget everything now and be ready for new FETCH_REQ */
2059                 io->flags = 0;
2060                 io->cmd = NULL;
2061                 io->addr = 0;
2062         }
2063 }
2064
2065 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2066                 struct io_uring_cmd *cmd)
2067 {
2068         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2069         int ret = -EINVAL;
2070         int i;
2071
2072         mutex_lock(&ub->mutex);
2073         if (!ublk_can_use_recovery(ub))
2074                 goto out_unlock;
2075         /*
2076          * START_RECOVERY is only allowd after:
2077          *
2078          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2079          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2080          *     released.
2081          *
2082          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2083          *     (a)has quiesced request queue
2084          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2085          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2086          *     (d)has completed/camceled all ioucmds owned by ther dying process
2087          */
2088         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2089                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2090                 ret = -EBUSY;
2091                 goto out_unlock;
2092         }
2093         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2094         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2095                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2096         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2097         ub->mm = NULL;
2098         ub->nr_queues_ready = 0;
2099         ub->nr_privileged_daemon = 0;
2100         init_completion(&ub->completion);
2101         ret = 0;
2102  out_unlock:
2103         mutex_unlock(&ub->mutex);
2104         return ret;
2105 }
2106
2107 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2108                 struct io_uring_cmd *cmd)
2109 {
2110         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2111         int ublksrv_pid = (int)header->data[0];
2112         int ret = -EINVAL;
2113
2114         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2115                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2116         /* wait until new ubq_daemon sending all FETCH_REQ */
2117         wait_for_completion_interruptible(&ub->completion);
2118         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2119                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2120
2121         mutex_lock(&ub->mutex);
2122         if (!ublk_can_use_recovery(ub))
2123                 goto out_unlock;
2124
2125         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2126                 ret = -EBUSY;
2127                 goto out_unlock;
2128         }
2129         ub->dev_info.ublksrv_pid = ublksrv_pid;
2130         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2131                         __func__, ublksrv_pid, header->dev_id);
2132         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2133         pr_devel("%s: queue unquiesced, dev id %d.\n",
2134                         __func__, header->dev_id);
2135         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2136         ub->dev_info.state = UBLK_S_DEV_LIVE;
2137         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2138         ret = 0;
2139  out_unlock:
2140         mutex_unlock(&ub->mutex);
2141         return ret;
2142 }
2143
2144 /*
2145  * All control commands are sent via /dev/ublk-control, so we have to check
2146  * the destination device's permission
2147  */
2148 static int ublk_char_dev_permission(struct ublk_device *ub,
2149                 const char *dev_path, int mask)
2150 {
2151         int err;
2152         struct path path;
2153         struct kstat stat;
2154
2155         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2156         if (err)
2157                 return err;
2158
2159         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2160         if (err)
2161                 goto exit;
2162
2163         err = -EPERM;
2164         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2165                 goto exit;
2166
2167         err = inode_permission(&nop_mnt_idmap,
2168                         d_backing_inode(path.dentry), mask);
2169 exit:
2170         path_put(&path);
2171         return err;
2172 }
2173
2174 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2175                 struct io_uring_cmd *cmd)
2176 {
2177         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2178         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2179         void __user *argp = (void __user *)(unsigned long)header->addr;
2180         char *dev_path = NULL;
2181         int ret = 0;
2182         int mask;
2183
2184         if (!unprivileged) {
2185                 if (!capable(CAP_SYS_ADMIN))
2186                         return -EPERM;
2187                 /*
2188                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2189                  * char_dev_path in payload too, since userspace may not
2190                  * know if the specified device is created as unprivileged
2191                  * mode.
2192                  */
2193                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2194                         return 0;
2195         }
2196
2197         /*
2198          * User has to provide the char device path for unprivileged ublk
2199          *
2200          * header->addr always points to the dev path buffer, and
2201          * header->dev_path_len records length of dev path buffer.
2202          */
2203         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2204                 return -EINVAL;
2205
2206         if (header->len < header->dev_path_len)
2207                 return -EINVAL;
2208
2209         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2210         if (!dev_path)
2211                 return -ENOMEM;
2212
2213         ret = -EFAULT;
2214         if (copy_from_user(dev_path, argp, header->dev_path_len))
2215                 goto exit;
2216         dev_path[header->dev_path_len] = 0;
2217
2218         ret = -EINVAL;
2219         switch (_IOC_NR(cmd->cmd_op)) {
2220         case UBLK_CMD_GET_DEV_INFO:
2221         case UBLK_CMD_GET_DEV_INFO2:
2222         case UBLK_CMD_GET_QUEUE_AFFINITY:
2223         case UBLK_CMD_GET_PARAMS:
2224                 mask = MAY_READ;
2225                 break;
2226         case UBLK_CMD_START_DEV:
2227         case UBLK_CMD_STOP_DEV:
2228         case UBLK_CMD_ADD_DEV:
2229         case UBLK_CMD_DEL_DEV:
2230         case UBLK_CMD_SET_PARAMS:
2231         case UBLK_CMD_START_USER_RECOVERY:
2232         case UBLK_CMD_END_USER_RECOVERY:
2233                 mask = MAY_READ | MAY_WRITE;
2234                 break;
2235         default:
2236                 goto exit;
2237         }
2238
2239         ret = ublk_char_dev_permission(ub, dev_path, mask);
2240         if (!ret) {
2241                 header->len -= header->dev_path_len;
2242                 header->addr += header->dev_path_len;
2243         }
2244         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2245                         __func__, ub->ub_number, cmd->cmd_op,
2246                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2247                         dev_path, ret);
2248 exit:
2249         kfree(dev_path);
2250         return ret;
2251 }
2252
2253 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2254                 unsigned int issue_flags)
2255 {
2256         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2257         struct ublk_device *ub = NULL;
2258         u32 cmd_op = cmd->cmd_op;
2259         int ret = -EINVAL;
2260
2261         if (issue_flags & IO_URING_F_NONBLOCK)
2262                 return -EAGAIN;
2263
2264         ublk_ctrl_cmd_dump(cmd);
2265
2266         if (!(issue_flags & IO_URING_F_SQE128))
2267                 goto out;
2268
2269         ret = ublk_check_cmd_op(cmd_op);
2270         if (ret)
2271                 goto out;
2272
2273         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2274                 ret = -ENODEV;
2275                 ub = ublk_get_device_from_id(header->dev_id);
2276                 if (!ub)
2277                         goto out;
2278
2279                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2280                 if (ret)
2281                         goto put_dev;
2282         }
2283
2284         switch (_IOC_NR(cmd_op)) {
2285         case UBLK_CMD_START_DEV:
2286                 ret = ublk_ctrl_start_dev(ub, cmd);
2287                 break;
2288         case UBLK_CMD_STOP_DEV:
2289                 ret = ublk_ctrl_stop_dev(ub);
2290                 break;
2291         case UBLK_CMD_GET_DEV_INFO:
2292         case UBLK_CMD_GET_DEV_INFO2:
2293                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2294                 break;
2295         case UBLK_CMD_ADD_DEV:
2296                 ret = ublk_ctrl_add_dev(cmd);
2297                 break;
2298         case UBLK_CMD_DEL_DEV:
2299                 ret = ublk_ctrl_del_dev(&ub);
2300                 break;
2301         case UBLK_CMD_GET_QUEUE_AFFINITY:
2302                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2303                 break;
2304         case UBLK_CMD_GET_PARAMS:
2305                 ret = ublk_ctrl_get_params(ub, cmd);
2306                 break;
2307         case UBLK_CMD_SET_PARAMS:
2308                 ret = ublk_ctrl_set_params(ub, cmd);
2309                 break;
2310         case UBLK_CMD_START_USER_RECOVERY:
2311                 ret = ublk_ctrl_start_recovery(ub, cmd);
2312                 break;
2313         case UBLK_CMD_END_USER_RECOVERY:
2314                 ret = ublk_ctrl_end_recovery(ub, cmd);
2315                 break;
2316         default:
2317                 ret = -ENOTSUPP;
2318                 break;
2319         }
2320
2321  put_dev:
2322         if (ub)
2323                 ublk_put_device(ub);
2324  out:
2325         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2326         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2327                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2328         return -EIOCBQUEUED;
2329 }
2330
2331 static const struct file_operations ublk_ctl_fops = {
2332         .open           = nonseekable_open,
2333         .uring_cmd      = ublk_ctrl_uring_cmd,
2334         .owner          = THIS_MODULE,
2335         .llseek         = noop_llseek,
2336 };
2337
2338 static struct miscdevice ublk_misc = {
2339         .minor          = MISC_DYNAMIC_MINOR,
2340         .name           = "ublk-control",
2341         .fops           = &ublk_ctl_fops,
2342 };
2343
2344 static int __init ublk_init(void)
2345 {
2346         int ret;
2347
2348         init_waitqueue_head(&ublk_idr_wq);
2349
2350         ret = misc_register(&ublk_misc);
2351         if (ret)
2352                 return ret;
2353
2354         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2355         if (ret)
2356                 goto unregister_mis;
2357
2358         ublk_chr_class = class_create("ublk-char");
2359         if (IS_ERR(ublk_chr_class)) {
2360                 ret = PTR_ERR(ublk_chr_class);
2361                 goto free_chrdev_region;
2362         }
2363         return 0;
2364
2365 free_chrdev_region:
2366         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2367 unregister_mis:
2368         misc_deregister(&ublk_misc);
2369         return ret;
2370 }
2371
2372 static void __exit ublk_exit(void)
2373 {
2374         struct ublk_device *ub;
2375         int id;
2376
2377         idr_for_each_entry(&ublk_index_idr, ub, id)
2378                 ublk_remove(ub);
2379
2380         class_destroy(ublk_chr_class);
2381         misc_deregister(&ublk_misc);
2382
2383         idr_destroy(&ublk_index_idr);
2384         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2385 }
2386
2387 module_init(ublk_init);
2388 module_exit(ublk_exit);
2389
2390 module_param(ublks_max, int, 0444);
2391 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2392
2393 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2394 MODULE_LICENSE("GPL");