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 <uapi/linux/ublk_cmd.h>
  46
  47 #define UBLK_MINORS             (1U << MINORBITS)
  48
  49 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  50 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  51                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  52                 | UBLK_F_NEED_GET_DATA \
  53                 | UBLK_F_USER_RECOVERY \
  54                 | UBLK_F_USER_RECOVERY_REISSUE)
  55
  56 /* All UBLK_PARAM_TYPE_* should be included here */
  57 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD)
  58
  59 struct ublk_rq_data {
  60         union {
  61                 struct callback_head work;
  62                 struct llist_node node;
  63         };
  64 };
  65
  66 struct ublk_uring_cmd_pdu {
  67         struct ublk_queue *ubq;
  68 };
  69
  70 /*
  71  * io command is active: sqe cmd is received, and its cqe isn't done
  72  *
  73  * If the flag is set, the io command is owned by ublk driver, and waited
  74  * for incoming blk-mq request from the ublk block device.
  75  *
  76  * If the flag is cleared, the io command will be completed, and owned by
  77  * ublk server.
  78  */
  79 #define UBLK_IO_FLAG_ACTIVE     0x01
  80
  81 /*
  82  * IO command is completed via cqe, and it is being handled by ublksrv, and
  83  * not committed yet
  84  *
  85  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  86  * cross verification
  87  */
  88 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  89
  90 /*
  91  * IO command is aborted, so this flag is set in case of
  92  * !UBLK_IO_FLAG_ACTIVE.
  93  *
  94  * After this flag is observed, any pending or new incoming request
  95  * associated with this io command will be failed immediately
  96  */
  97 #define UBLK_IO_FLAG_ABORTED 0x04
  98
  99 /*
 100  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
 101  * get data buffer address from ublksrv.
 102  *
 103  * Then, bio data could be copied into this data buffer for a WRITE request
 104  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 105  */
 106 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 107
 108 struct ublk_io {
 109         /* userspace buffer address from io cmd */
 110         __u64   addr;
 111         unsigned int flags;
 112         int res;
 113
 114         struct io_uring_cmd *cmd;
 115 };
 116
 117 struct ublk_queue {
 118         int q_id;
 119         int q_depth;
 120
 121         unsigned long flags;
 122         struct task_struct      *ubq_daemon;
 123         char *io_cmd_buf;
 124
 125         struct llist_head       io_cmds;
 126
 127         unsigned long io_addr;  /* mapped vm address */
 128         unsigned int max_io_sz;
 129         bool force_abort;
 130         unsigned short nr_io_ready;     /* how many ios setup */
 131         struct ublk_device *dev;
 132         struct ublk_io ios[];
 133 };
 134
 135 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 136
 137 struct ublk_device {
 138         struct gendisk          *ub_disk;
 139
 140         char    *__queues;
 141
 142         unsigned short  queue_size;
 143         struct ublksrv_ctrl_dev_info    dev_info;
 144
 145         struct blk_mq_tag_set   tag_set;
 146
 147         struct cdev             cdev;
 148         struct device           cdev_dev;
 149
 150 #define UB_STATE_OPEN           0
 151 #define UB_STATE_USED           1
 152         unsigned long           state;
 153         int                     ub_number;
 154
 155         struct mutex            mutex;
 156
 157         spinlock_t              mm_lock;
 158         struct mm_struct        *mm;
 159
 160         struct ublk_params      params;
 161
 162         struct completion       completion;
 163         unsigned int            nr_queues_ready;
 164         atomic_t                nr_aborted_queues;
 165
 166         /*
 167          * Our ubq->daemon may be killed without any notification, so
 168          * monitor each queue's daemon periodically
 169          */
 170         struct delayed_work     monitor_work;
 171         struct work_struct      quiesce_work;
 172         struct work_struct      stop_work;
 173 };
 174
 175 /* header of ublk_params */
 176 struct ublk_params_header {
 177         __u32   len;
 178         __u32   types;
 179 };
 180
 181 static dev_t ublk_chr_devt;
 182 static struct class *ublk_chr_class;
 183
 184 static DEFINE_IDR(ublk_index_idr);
 185 static DEFINE_SPINLOCK(ublk_idr_lock);
 186 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 187
 188 static DEFINE_MUTEX(ublk_ctl_mutex);
 189
 190 static struct miscdevice ublk_misc;
 191
 192 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 193 {
 194         struct request_queue *q = ub->ub_disk->queue;
 195         const struct ublk_param_basic *p = &ub->params.basic;
 196
 197         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 198         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 199         blk_queue_io_min(q, 1 << p->io_min_shift);
 200         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 201
 202         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 203                         p->attrs & UBLK_ATTR_FUA);
 204         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 205                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 206         else
 207                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 208
 209         blk_queue_max_hw_sectors(q, p->max_sectors);
 210         blk_queue_chunk_sectors(q, p->chunk_sectors);
 211         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 212
 213         if (p->attrs & UBLK_ATTR_READ_ONLY)
 214                 set_disk_ro(ub->ub_disk, true);
 215
 216         set_capacity(ub->ub_disk, p->dev_sectors);
 217 }
 218
 219 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 220 {
 221         struct request_queue *q = ub->ub_disk->queue;
 222         const struct ublk_param_discard *p = &ub->params.discard;
 223
 224         q->limits.discard_alignment = p->discard_alignment;
 225         q->limits.discard_granularity = p->discard_granularity;
 226         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 227         blk_queue_max_write_zeroes_sectors(q,
 228                         p->max_write_zeroes_sectors);
 229         blk_queue_max_discard_segments(q, p->max_discard_segments);
 230 }
 231
 232 static int ublk_validate_params(const struct ublk_device *ub)
 233 {
 234         /* basic param is the only one which must be set */
 235         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 236                 const struct ublk_param_basic *p = &ub->params.basic;
 237
 238                 if (p->logical_bs_shift > PAGE_SHIFT)
 239                         return -EINVAL;
 240
 241                 if (p->logical_bs_shift > p->physical_bs_shift)
 242                         return -EINVAL;
 243
 244                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 245                         return -EINVAL;
 246         } else
 247                 return -EINVAL;
 248
 249         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 250                 const struct ublk_param_discard *p = &ub->params.discard;
 251
 252                 /* So far, only support single segment discard */
 253                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 254                         return -EINVAL;
 255
 256                 if (!p->discard_granularity)
 257                         return -EINVAL;
 258         }
 259
 260         return 0;
 261 }
 262
 263 static int ublk_apply_params(struct ublk_device *ub)
 264 {
 265         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 266                 return -EINVAL;
 267
 268         ublk_dev_param_basic_apply(ub);
 269
 270         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 271                 ublk_dev_param_discard_apply(ub);
 272
 273         return 0;
 274 }
 275
 276 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
 277 {
 278         if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
 279                         !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
 280                 return true;
 281         return false;
 282 }
 283
 284 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 285 {
 286         if (ubq->flags & UBLK_F_NEED_GET_DATA)
 287                 return true;
 288         return false;
 289 }
 290
 291 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 292 {
 293         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 294                 return ub;
 295         return NULL;
 296 }
 297
 298 static void ublk_put_device(struct ublk_device *ub)
 299 {
 300         put_device(&ub->cdev_dev);
 301 }
 302
 303 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 304                 int qid)
 305 {
 306        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 307 }
 308
 309 static inline bool ublk_rq_has_data(const struct request *rq)
 310 {
 311         return rq->bio && bio_has_data(rq->bio);
 312 }
 313
 314 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 315                 int tag)
 316 {
 317         return (struct ublksrv_io_desc *)
 318                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 319 }
 320
 321 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 322 {
 323         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 324 }
 325
 326 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 327 {
 328         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 329
 330         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 331                         PAGE_SIZE);
 332 }
 333
 334 static inline bool ublk_queue_can_use_recovery_reissue(
 335                 struct ublk_queue *ubq)
 336 {
 337         if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
 338                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
 339                 return true;
 340         return false;
 341 }
 342
 343 static inline bool ublk_queue_can_use_recovery(
 344                 struct ublk_queue *ubq)
 345 {
 346         if (ubq->flags & UBLK_F_USER_RECOVERY)
 347                 return true;
 348         return false;
 349 }
 350
 351 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 352 {
 353         if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
 354                 return true;
 355         return false;
 356 }
 357
 358 static void ublk_free_disk(struct gendisk *disk)
 359 {
 360         struct ublk_device *ub = disk->private_data;
 361
 362         clear_bit(UB_STATE_USED, &ub->state);
 363         put_device(&ub->cdev_dev);
 364 }
 365
 366 static const struct block_device_operations ub_fops = {
 367         .owner =        THIS_MODULE,
 368         .free_disk =    ublk_free_disk,
 369 };
 370
 371 #define UBLK_MAX_PIN_PAGES      32
 372
 373 struct ublk_map_data {
 374         const struct ublk_queue *ubq;
 375         const struct request *rq;
 376         const struct ublk_io *io;
 377         unsigned max_bytes;
 378 };
 379
 380 struct ublk_io_iter {
 381         struct page *pages[UBLK_MAX_PIN_PAGES];
 382         unsigned pg_off;        /* offset in the 1st page in pages */
 383         int nr_pages;           /* how many page pointers in pages */
 384         struct bio *bio;
 385         struct bvec_iter iter;
 386 };
 387
 388 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
 389                 unsigned max_bytes, bool to_vm)
 390 {
 391         const unsigned total = min_t(unsigned, max_bytes,
 392                         PAGE_SIZE - data->pg_off +
 393                         ((data->nr_pages - 1) << PAGE_SHIFT));
 394         unsigned done = 0;
 395         unsigned pg_idx = 0;
 396
 397         while (done < total) {
 398                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 399                 const unsigned int bytes = min3(bv.bv_len, total - done,
 400                                 (unsigned)(PAGE_SIZE - data->pg_off));
 401                 void *bv_buf = bvec_kmap_local(&bv);
 402                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 403
 404                 if (to_vm)
 405                         memcpy(pg_buf + data->pg_off, bv_buf, bytes);
 406                 else
 407                         memcpy(bv_buf, pg_buf + data->pg_off, bytes);
 408
 409                 kunmap_local(pg_buf);
 410                 kunmap_local(bv_buf);
 411
 412                 /* advance page array */
 413                 data->pg_off += bytes;
 414                 if (data->pg_off == PAGE_SIZE) {
 415                         pg_idx += 1;
 416                         data->pg_off = 0;
 417                 }
 418
 419                 done += bytes;
 420
 421                 /* advance bio */
 422                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 423                 if (!data->iter.bi_size) {
 424                         data->bio = data->bio->bi_next;
 425                         if (data->bio == NULL)
 426                                 break;
 427                         data->iter = data->bio->bi_iter;
 428                 }
 429         }
 430
 431         return done;
 432 }
 433
 434 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
 435                 bool to_vm)
 436 {
 437         const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
 438         const unsigned long start_vm = data->io->addr;
 439         unsigned int done = 0;
 440         struct ublk_io_iter iter = {
 441                 .pg_off = start_vm & (PAGE_SIZE - 1),
 442                 .bio    = data->rq->bio,
 443                 .iter   = data->rq->bio->bi_iter,
 444         };
 445         const unsigned int nr_pages = round_up(data->max_bytes +
 446                         (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
 447
 448         while (done < nr_pages) {
 449                 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
 450                                 nr_pages - done);
 451                 unsigned i, len;
 452
 453                 iter.nr_pages = get_user_pages_fast(start_vm +
 454                                 (done << PAGE_SHIFT), to_pin, gup_flags,
 455                                 iter.pages);
 456                 if (iter.nr_pages <= 0)
 457                         return done == 0 ? iter.nr_pages : done;
 458                 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
 459                 for (i = 0; i < iter.nr_pages; i++) {
 460                         if (to_vm)
 461                                 set_page_dirty(iter.pages[i]);
 462                         put_page(iter.pages[i]);
 463                 }
 464                 data->max_bytes -= len;
 465                 done += iter.nr_pages;
 466         }
 467
 468         return done;
 469 }
 470
 471 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 472                 struct ublk_io *io)
 473 {
 474         const unsigned int rq_bytes = blk_rq_bytes(req);
 475         /*
 476          * no zero copy, we delay copy WRITE request data into ublksrv
 477          * context and the big benefit is that pinning pages in current
 478          * context is pretty fast, see ublk_pin_user_pages
 479          */
 480         if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
 481                 return rq_bytes;
 482
 483         if (ublk_rq_has_data(req)) {
 484                 struct ublk_map_data data = {
 485                         .ubq    =       ubq,
 486                         .rq     =       req,
 487                         .io     =       io,
 488                         .max_bytes =    rq_bytes,
 489                 };
 490
 491                 ublk_copy_user_pages(&data, true);
 492
 493                 return rq_bytes - data.max_bytes;
 494         }
 495         return rq_bytes;
 496 }
 497
 498 static int ublk_unmap_io(const struct ublk_queue *ubq,
 499                 const struct request *req,
 500                 struct ublk_io *io)
 501 {
 502         const unsigned int rq_bytes = blk_rq_bytes(req);
 503
 504         if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
 505                 struct ublk_map_data data = {
 506                         .ubq    =       ubq,
 507                         .rq     =       req,
 508                         .io     =       io,
 509                         .max_bytes =    io->res,
 510                 };
 511
 512                 WARN_ON_ONCE(io->res > rq_bytes);
 513
 514                 ublk_copy_user_pages(&data, false);
 515
 516                 return io->res - data.max_bytes;
 517         }
 518         return rq_bytes;
 519 }
 520
 521 static inline unsigned int ublk_req_build_flags(struct request *req)
 522 {
 523         unsigned flags = 0;
 524
 525         if (req->cmd_flags & REQ_FAILFAST_DEV)
 526                 flags |= UBLK_IO_F_FAILFAST_DEV;
 527
 528         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 529                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 530
 531         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 532                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 533
 534         if (req->cmd_flags & REQ_META)
 535                 flags |= UBLK_IO_F_META;
 536
 537         if (req->cmd_flags & REQ_FUA)
 538                 flags |= UBLK_IO_F_FUA;
 539
 540         if (req->cmd_flags & REQ_NOUNMAP)
 541                 flags |= UBLK_IO_F_NOUNMAP;
 542
 543         if (req->cmd_flags & REQ_SWAP)
 544                 flags |= UBLK_IO_F_SWAP;
 545
 546         return flags;
 547 }
 548
 549 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 550 {
 551         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 552         struct ublk_io *io = &ubq->ios[req->tag];
 553         u32 ublk_op;
 554
 555         switch (req_op(req)) {
 556         case REQ_OP_READ:
 557                 ublk_op = UBLK_IO_OP_READ;
 558                 break;
 559         case REQ_OP_WRITE:
 560                 ublk_op = UBLK_IO_OP_WRITE;
 561                 break;
 562         case REQ_OP_FLUSH:
 563                 ublk_op = UBLK_IO_OP_FLUSH;
 564                 break;
 565         case REQ_OP_DISCARD:
 566                 ublk_op = UBLK_IO_OP_DISCARD;
 567                 break;
 568         case REQ_OP_WRITE_ZEROES:
 569                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 570                 break;
 571         default:
 572                 return BLK_STS_IOERR;
 573         }
 574
 575         /* need to translate since kernel may change */
 576         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 577         iod->nr_sectors = blk_rq_sectors(req);
 578         iod->start_sector = blk_rq_pos(req);
 579         iod->addr = io->addr;
 580
 581         return BLK_STS_OK;
 582 }
 583
 584 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 585                 struct io_uring_cmd *ioucmd)
 586 {
 587         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 588 }
 589
 590 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 591 {
 592         return ubq->ubq_daemon->flags & PF_EXITING;
 593 }
 594
 595 /* todo: handle partial completion */
 596 static void ublk_complete_rq(struct request *req)
 597 {
 598         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 599         struct ublk_io *io = &ubq->ios[req->tag];
 600         unsigned int unmapped_bytes;
 601
 602         /* failed read IO if nothing is read */
 603         if (!io->res && req_op(req) == REQ_OP_READ)
 604                 io->res = -EIO;
 605
 606         if (io->res < 0) {
 607                 blk_mq_end_request(req, errno_to_blk_status(io->res));
 608                 return;
 609         }
 610
 611         /*
 612          * FLUSH or DISCARD usually won't return bytes returned, so end them
 613          * directly.
 614          *
 615          * Both the two needn't unmap.
 616          */
 617         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
 618                 blk_mq_end_request(req, BLK_STS_OK);
 619                 return;
 620         }
 621
 622         /* for READ request, writing data in iod->addr to rq buffers */
 623         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 624
 625         /*
 626          * Extremely impossible since we got data filled in just before
 627          *
 628          * Re-read simply for this unlikely case.
 629          */
 630         if (unlikely(unmapped_bytes < io->res))
 631                 io->res = unmapped_bytes;
 632
 633         if (blk_update_request(req, BLK_STS_OK, io->res))
 634                 blk_mq_requeue_request(req, true);
 635         else
 636                 __blk_mq_end_request(req, BLK_STS_OK);
 637 }
 638
 639 /*
 640  * Since __ublk_rq_task_work always fails requests immediately during
 641  * exiting, __ublk_fail_req() is only called from abort context during
 642  * exiting. So lock is unnecessary.
 643  *
 644  * Also aborting may not be started yet, keep in mind that one failed
 645  * request may be issued by block layer again.
 646  */
 647 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
 648                 struct request *req)
 649 {
 650         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 651
 652         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 653                 io->flags |= UBLK_IO_FLAG_ABORTED;
 654                 if (ublk_queue_can_use_recovery_reissue(ubq))
 655                         blk_mq_requeue_request(req, false);
 656                 else
 657                         blk_mq_end_request(req, BLK_STS_IOERR);
 658         }
 659 }
 660
 661 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
 662 {
 663         /* mark this cmd owned by ublksrv */
 664         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 665
 666         /*
 667          * clear ACTIVE since we are done with this sqe/cmd slot
 668          * We can only accept io cmd in case of being not active.
 669          */
 670         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 671
 672         /* tell ublksrv one io request is coming */
 673         io_uring_cmd_done(io->cmd, res, 0);
 674 }
 675
 676 #define UBLK_REQUEUE_DELAY_MS   3
 677
 678 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 679                 struct request *rq)
 680 {
 681         /* We cannot process this rq so just requeue it. */
 682         if (ublk_queue_can_use_recovery(ubq))
 683                 blk_mq_requeue_request(rq, false);
 684         else
 685                 blk_mq_end_request(rq, BLK_STS_IOERR);
 686
 687         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 688 }
 689
 690 static inline void __ublk_rq_task_work(struct request *req)
 691 {
 692         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 693         int tag = req->tag;
 694         struct ublk_io *io = &ubq->ios[tag];
 695         unsigned int mapped_bytes;
 696
 697         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 698                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 699                         ublk_get_iod(ubq, req->tag)->addr);
 700
 701         /*
 702          * Task is exiting if either:
 703          *
 704          * (1) current != ubq_daemon.
 705          * io_uring_cmd_complete_in_task() tries to run task_work
 706          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 707          *
 708          * (2) current->flags & PF_EXITING.
 709          */
 710         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 711                 __ublk_abort_rq(ubq, req);
 712                 return;
 713         }
 714
 715         if (ublk_need_get_data(ubq) &&
 716                         (req_op(req) == REQ_OP_WRITE ||
 717                         req_op(req) == REQ_OP_FLUSH)) {
 718                 /*
 719                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 720                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 721                  * and notify it.
 722                  */
 723                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 724                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 725                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 726                                         __func__, io->cmd->cmd_op, ubq->q_id,
 727                                         req->tag, io->flags);
 728                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
 729                         return;
 730                 }
 731                 /*
 732                  * We have handled UBLK_IO_NEED_GET_DATA command,
 733                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 734                  * do the copy work.
 735                  */
 736                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 737                 /* update iod->addr because ublksrv may have passed a new io buffer */
 738                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 739                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 740                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 741                                 ublk_get_iod(ubq, req->tag)->addr);
 742         }
 743
 744         mapped_bytes = ublk_map_io(ubq, req, io);
 745
 746         /* partially mapped, update io descriptor */
 747         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 748                 /*
 749                  * Nothing mapped, retry until we succeed.
 750                  *
 751                  * We may never succeed in mapping any bytes here because
 752                  * of OOM. TODO: reserve one buffer with single page pinned
 753                  * for providing forward progress guarantee.
 754                  */
 755                 if (unlikely(!mapped_bytes)) {
 756                         blk_mq_requeue_request(req, false);
 757                         blk_mq_delay_kick_requeue_list(req->q,
 758                                         UBLK_REQUEUE_DELAY_MS);
 759                         return;
 760                 }
 761
 762                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 763                         mapped_bytes >> 9;
 764         }
 765
 766         ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
 767 }
 768
 769 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
 770 {
 771         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 772         struct ublk_queue *ubq = pdu->ubq;
 773         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 774         struct ublk_rq_data *data;
 775
 776         llist_for_each_entry(data, io_cmds, node)
 777                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
 778 }
 779
 780 static void ublk_rq_task_work_fn(struct callback_head *work)
 781 {
 782         struct ublk_rq_data *data = container_of(work,
 783                         struct ublk_rq_data, work);
 784         struct request *req = blk_mq_rq_from_pdu(data);
 785
 786         __ublk_rq_task_work(req);
 787 }
 788
 789 static void ublk_submit_cmd(struct ublk_queue *ubq, const struct request *rq)
 790 {
 791         struct ublk_io *io = &ubq->ios[rq->tag];
 792
 793         /*
 794          * If the check pass, we know that this is a re-issued request aborted
 795          * previously in monitor_work because the ubq_daemon(cmd's task) is
 796          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 797          * because this ioucmd's io_uring context may be freed now if no inflight
 798          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 799          *
 800          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 801          * the tag). Then the request is re-started(allocating the tag) and we are here.
 802          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 803          * guarantees that here is a re-issued request aborted previously.
 804          */
 805         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 806                 struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 807                 struct ublk_rq_data *data;
 808
 809                 llist_for_each_entry(data, io_cmds, node)
 810                         __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 811         } else {
 812                 struct io_uring_cmd *cmd = io->cmd;
 813                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 814
 815                 pdu->ubq = ubq;
 816                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 817         }
 818 }
 819
 820 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq,
 821                 bool last)
 822 {
 823         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 824
 825         if (ublk_can_use_task_work(ubq)) {
 826                 enum task_work_notify_mode notify_mode = last ?
 827                         TWA_SIGNAL_NO_IPI : TWA_NONE;
 828
 829                 if (task_work_add(ubq->ubq_daemon, &data->work, notify_mode))
 830                         __ublk_abort_rq(ubq, rq);
 831         } else {
 832                 if (llist_add(&data->node, &ubq->io_cmds))
 833                         ublk_submit_cmd(ubq, rq);
 834         }
 835 }
 836
 837 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 838                 const struct blk_mq_queue_data *bd)
 839 {
 840         struct ublk_queue *ubq = hctx->driver_data;
 841         struct request *rq = bd->rq;
 842         blk_status_t res;
 843
 844         /* fill iod to slot in io cmd buffer */
 845         res = ublk_setup_iod(ubq, rq);
 846         if (unlikely(res != BLK_STS_OK))
 847                 return BLK_STS_IOERR;
 848
 849         /* With recovery feature enabled, force_abort is set in
 850          * ublk_stop_dev() before calling del_gendisk(). We have to
 851          * abort all requeued and new rqs here to let del_gendisk()
 852          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 853          * to avoid UAF on io_uring ctx.
 854          *
 855          * Note: force_abort is guaranteed to be seen because it is set
 856          * before request queue is unqiuesced.
 857          */
 858         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 859                 return BLK_STS_IOERR;
 860
 861         blk_mq_start_request(bd->rq);
 862
 863         if (unlikely(ubq_daemon_is_dying(ubq))) {
 864                 __ublk_abort_rq(ubq, rq);
 865                 return BLK_STS_OK;
 866         }
 867
 868         ublk_queue_cmd(ubq, rq, bd->last);
 869
 870         return BLK_STS_OK;
 871 }
 872
 873 static void ublk_commit_rqs(struct blk_mq_hw_ctx *hctx)
 874 {
 875         struct ublk_queue *ubq = hctx->driver_data;
 876
 877         if (ublk_can_use_task_work(ubq))
 878                 __set_notify_signal(ubq->ubq_daemon);
 879 }
 880
 881 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 882                 unsigned int hctx_idx)
 883 {
 884         struct ublk_device *ub = driver_data;
 885         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 886
 887         hctx->driver_data = ubq;
 888         return 0;
 889 }
 890
 891 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 892                 unsigned int hctx_idx, unsigned int numa_node)
 893 {
 894         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 895
 896         init_task_work(&data->work, ublk_rq_task_work_fn);
 897         return 0;
 898 }
 899
 900 static const struct blk_mq_ops ublk_mq_ops = {
 901         .queue_rq       = ublk_queue_rq,
 902         .commit_rqs     = ublk_commit_rqs,
 903         .init_hctx      = ublk_init_hctx,
 904         .init_request   = ublk_init_rq,
 905 };
 906
 907 static int ublk_ch_open(struct inode *inode, struct file *filp)
 908 {
 909         struct ublk_device *ub = container_of(inode->i_cdev,
 910                         struct ublk_device, cdev);
 911
 912         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 913                 return -EBUSY;
 914         filp->private_data = ub;
 915         return 0;
 916 }
 917
 918 static int ublk_ch_release(struct inode *inode, struct file *filp)
 919 {
 920         struct ublk_device *ub = filp->private_data;
 921
 922         clear_bit(UB_STATE_OPEN, &ub->state);
 923         return 0;
 924 }
 925
 926 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
 927 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
 928 {
 929         struct ublk_device *ub = filp->private_data;
 930         size_t sz = vma->vm_end - vma->vm_start;
 931         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
 932         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
 933         int q_id, ret = 0;
 934
 935         spin_lock(&ub->mm_lock);
 936         if (!ub->mm)
 937                 ub->mm = current->mm;
 938         if (current->mm != ub->mm)
 939                 ret = -EINVAL;
 940         spin_unlock(&ub->mm_lock);
 941
 942         if (ret)
 943                 return ret;
 944
 945         if (vma->vm_flags & VM_WRITE)
 946                 return -EPERM;
 947
 948         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
 949         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
 950                 return -EINVAL;
 951
 952         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
 953         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
 954                         __func__, q_id, current->pid, vma->vm_start,
 955                         phys_off, (unsigned long)sz);
 956
 957         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
 958                 return -EINVAL;
 959
 960         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
 961         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
 962 }
 963
 964 static void ublk_commit_completion(struct ublk_device *ub,
 965                 struct ublksrv_io_cmd *ub_cmd)
 966 {
 967         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
 968         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
 969         struct ublk_io *io = &ubq->ios[tag];
 970         struct request *req;
 971
 972         /* now this cmd slot is owned by nbd driver */
 973         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
 974         io->res = ub_cmd->result;
 975
 976         /* find the io request and complete */
 977         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
 978
 979         if (req && likely(!blk_should_fake_timeout(req->q)))
 980                 ublk_complete_rq(req);
 981 }
 982
 983 /*
 984  * When ->ubq_daemon is exiting, either new request is ended immediately,
 985  * or any queued io command is drained, so it is safe to abort queue
 986  * lockless
 987  */
 988 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
 989 {
 990         int i;
 991
 992         if (!ublk_get_device(ub))
 993                 return;
 994
 995         for (i = 0; i < ubq->q_depth; i++) {
 996                 struct ublk_io *io = &ubq->ios[i];
 997
 998                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
 999                         struct request *rq;
1000
1001                         /*
1002                          * Either we fail the request or ublk_rq_task_work_fn
1003                          * will do it
1004                          */
1005                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1006                         if (rq)
1007                                 __ublk_fail_req(ubq, io, rq);
1008                 }
1009         }
1010         ublk_put_device(ub);
1011 }
1012
1013 static void ublk_daemon_monitor_work(struct work_struct *work)
1014 {
1015         struct ublk_device *ub =
1016                 container_of(work, struct ublk_device, monitor_work.work);
1017         int i;
1018
1019         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1020                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1021
1022                 if (ubq_daemon_is_dying(ubq)) {
1023                         if (ublk_queue_can_use_recovery(ubq))
1024                                 schedule_work(&ub->quiesce_work);
1025                         else
1026                                 schedule_work(&ub->stop_work);
1027
1028                         /* abort queue is for making forward progress */
1029                         ublk_abort_queue(ub, ubq);
1030                 }
1031         }
1032
1033         /*
1034          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1035          * after ublk_remove() or __ublk_quiesce_dev() is started.
1036          *
1037          * No need ub->mutex, monitor work are canceled after state is marked
1038          * as not LIVE, so new state is observed reliably.
1039          */
1040         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1041                 schedule_delayed_work(&ub->monitor_work,
1042                                 UBLK_DAEMON_MONITOR_PERIOD);
1043 }
1044
1045 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1046 {
1047         return ubq->nr_io_ready == ubq->q_depth;
1048 }
1049
1050 static void ublk_cancel_queue(struct ublk_queue *ubq)
1051 {
1052         int i;
1053
1054         if (!ublk_queue_ready(ubq))
1055                 return;
1056
1057         for (i = 0; i < ubq->q_depth; i++) {
1058                 struct ublk_io *io = &ubq->ios[i];
1059
1060                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1061                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1062         }
1063
1064         /* all io commands are canceled */
1065         ubq->nr_io_ready = 0;
1066 }
1067
1068 /* Cancel all pending commands, must be called after del_gendisk() returns */
1069 static void ublk_cancel_dev(struct ublk_device *ub)
1070 {
1071         int i;
1072
1073         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1074                 ublk_cancel_queue(ublk_get_queue(ub, i));
1075 }
1076
1077 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1078 {
1079         bool *idle = data;
1080
1081         if (blk_mq_request_started(rq)) {
1082                 *idle = false;
1083                 return false;
1084         }
1085         return true;
1086 }
1087
1088 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1089 {
1090         bool idle;
1091
1092         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1093         while (true) {
1094                 idle = true;
1095                 blk_mq_tagset_busy_iter(&ub->tag_set,
1096                                 ublk_check_inflight_rq, &idle);
1097                 if (idle)
1098                         break;
1099                 msleep(UBLK_REQUEUE_DELAY_MS);
1100         }
1101 }
1102
1103 static void __ublk_quiesce_dev(struct ublk_device *ub)
1104 {
1105         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1106                         __func__, ub->dev_info.dev_id,
1107                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1108                         "LIVE" : "QUIESCED");
1109         blk_mq_quiesce_queue(ub->ub_disk->queue);
1110         ublk_wait_tagset_rqs_idle(ub);
1111         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1112         ublk_cancel_dev(ub);
1113         /* we are going to release task_struct of ubq_daemon and resets
1114          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1115          * Besides, monitor_work is not necessary in QUIESCED state since we have
1116          * already scheduled quiesce_work and quiesced all ubqs.
1117          *
1118          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1119          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1120          */
1121         cancel_delayed_work_sync(&ub->monitor_work);
1122 }
1123
1124 static void ublk_quiesce_work_fn(struct work_struct *work)
1125 {
1126         struct ublk_device *ub =
1127                 container_of(work, struct ublk_device, quiesce_work);
1128
1129         mutex_lock(&ub->mutex);
1130         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1131                 goto unlock;
1132         __ublk_quiesce_dev(ub);
1133  unlock:
1134         mutex_unlock(&ub->mutex);
1135 }
1136
1137 static void ublk_unquiesce_dev(struct ublk_device *ub)
1138 {
1139         int i;
1140
1141         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1142                         __func__, ub->dev_info.dev_id,
1143                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1144                         "LIVE" : "QUIESCED");
1145         /* quiesce_work has run. We let requeued rqs be aborted
1146          * before running fallback_wq. "force_abort" must be seen
1147          * after request queue is unqiuesced. Then del_gendisk()
1148          * can move on.
1149          */
1150         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1151                 ublk_get_queue(ub, i)->force_abort = true;
1152
1153         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1154         /* We may have requeued some rqs in ublk_quiesce_queue() */
1155         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1156 }
1157
1158 static void ublk_stop_dev(struct ublk_device *ub)
1159 {
1160         mutex_lock(&ub->mutex);
1161         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1162                 goto unlock;
1163         if (ublk_can_use_recovery(ub)) {
1164                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1165                         __ublk_quiesce_dev(ub);
1166                 ublk_unquiesce_dev(ub);
1167         }
1168         del_gendisk(ub->ub_disk);
1169         ub->dev_info.state = UBLK_S_DEV_DEAD;
1170         ub->dev_info.ublksrv_pid = -1;
1171         put_disk(ub->ub_disk);
1172         ub->ub_disk = NULL;
1173  unlock:
1174         ublk_cancel_dev(ub);
1175         mutex_unlock(&ub->mutex);
1176         cancel_delayed_work_sync(&ub->monitor_work);
1177 }
1178
1179 /* device can only be started after all IOs are ready */
1180 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1181 {
1182         mutex_lock(&ub->mutex);
1183         ubq->nr_io_ready++;
1184         if (ublk_queue_ready(ubq)) {
1185                 ubq->ubq_daemon = current;
1186                 get_task_struct(ubq->ubq_daemon);
1187                 ub->nr_queues_ready++;
1188         }
1189         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1190                 complete_all(&ub->completion);
1191         mutex_unlock(&ub->mutex);
1192 }
1193
1194 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1195                 int tag)
1196 {
1197         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1198         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1199
1200         ublk_queue_cmd(ubq, req, true);
1201 }
1202
1203 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1204 {
1205         struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1206         struct ublk_device *ub = cmd->file->private_data;
1207         struct ublk_queue *ubq;
1208         struct ublk_io *io;
1209         u32 cmd_op = cmd->cmd_op;
1210         unsigned tag = ub_cmd->tag;
1211         int ret = -EINVAL;
1212
1213         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1214                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1215                         ub_cmd->result);
1216
1217         if (!(issue_flags & IO_URING_F_SQE128))
1218                 goto out;
1219
1220         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1221                 goto out;
1222
1223         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1224         if (!ubq || ub_cmd->q_id != ubq->q_id)
1225                 goto out;
1226
1227         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1228                 goto out;
1229
1230         if (tag >= ubq->q_depth)
1231                 goto out;
1232
1233         io = &ubq->ios[tag];
1234
1235         /* there is pending io cmd, something must be wrong */
1236         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1237                 ret = -EBUSY;
1238                 goto out;
1239         }
1240
1241         /*
1242          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1243          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1244          */
1245         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1246                         ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1247                 goto out;
1248
1249         switch (cmd_op) {
1250         case UBLK_IO_FETCH_REQ:
1251                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1252                 if (ublk_queue_ready(ubq)) {
1253                         ret = -EBUSY;
1254                         goto out;
1255                 }
1256                 /*
1257                  * The io is being handled by server, so COMMIT_RQ is expected
1258                  * instead of FETCH_REQ
1259                  */
1260                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1261                         goto out;
1262                 /* FETCH_RQ has to provide IO buffer */
1263                 if (!ub_cmd->addr)
1264                         goto out;
1265                 io->cmd = cmd;
1266                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1267                 io->addr = ub_cmd->addr;
1268
1269                 ublk_mark_io_ready(ub, ubq);
1270                 break;
1271         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1272                 /* FETCH_RQ has to provide IO buffer */
1273                 if (!ub_cmd->addr)
1274                         goto out;
1275                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1276                         goto out;
1277                 io->addr = ub_cmd->addr;
1278                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1279                 io->cmd = cmd;
1280                 ublk_commit_completion(ub, ub_cmd);
1281                 break;
1282         case UBLK_IO_NEED_GET_DATA:
1283                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1284                         goto out;
1285                 io->addr = ub_cmd->addr;
1286                 io->cmd = cmd;
1287                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1288                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1289                 break;
1290         default:
1291                 goto out;
1292         }
1293         return -EIOCBQUEUED;
1294
1295  out:
1296         io_uring_cmd_done(cmd, ret, 0);
1297         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1298                         __func__, cmd_op, tag, ret, io->flags);
1299         return -EIOCBQUEUED;
1300 }
1301
1302 static const struct file_operations ublk_ch_fops = {
1303         .owner = THIS_MODULE,
1304         .open = ublk_ch_open,
1305         .release = ublk_ch_release,
1306         .llseek = no_llseek,
1307         .uring_cmd = ublk_ch_uring_cmd,
1308         .mmap = ublk_ch_mmap,
1309 };
1310
1311 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1312 {
1313         int size = ublk_queue_cmd_buf_size(ub, q_id);
1314         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1315
1316         if (ubq->ubq_daemon)
1317                 put_task_struct(ubq->ubq_daemon);
1318         if (ubq->io_cmd_buf)
1319                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1320 }
1321
1322 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1323 {
1324         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1325         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1326         void *ptr;
1327         int size;
1328
1329         ubq->flags = ub->dev_info.flags;
1330         ubq->q_id = q_id;
1331         ubq->q_depth = ub->dev_info.queue_depth;
1332         size = ublk_queue_cmd_buf_size(ub, q_id);
1333
1334         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1335         if (!ptr)
1336                 return -ENOMEM;
1337
1338         ubq->io_cmd_buf = ptr;
1339         ubq->dev = ub;
1340         return 0;
1341 }
1342
1343 static void ublk_deinit_queues(struct ublk_device *ub)
1344 {
1345         int nr_queues = ub->dev_info.nr_hw_queues;
1346         int i;
1347
1348         if (!ub->__queues)
1349                 return;
1350
1351         for (i = 0; i < nr_queues; i++)
1352                 ublk_deinit_queue(ub, i);
1353         kfree(ub->__queues);
1354 }
1355
1356 static int ublk_init_queues(struct ublk_device *ub)
1357 {
1358         int nr_queues = ub->dev_info.nr_hw_queues;
1359         int depth = ub->dev_info.queue_depth;
1360         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1361         int i, ret = -ENOMEM;
1362
1363         ub->queue_size = ubq_size;
1364         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1365         if (!ub->__queues)
1366                 return ret;
1367
1368         for (i = 0; i < nr_queues; i++) {
1369                 if (ublk_init_queue(ub, i))
1370                         goto fail;
1371         }
1372
1373         init_completion(&ub->completion);
1374         return 0;
1375
1376  fail:
1377         ublk_deinit_queues(ub);
1378         return ret;
1379 }
1380
1381 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1382 {
1383         int i = idx;
1384         int err;
1385
1386         spin_lock(&ublk_idr_lock);
1387         /* allocate id, if @id >= 0, we're requesting that specific id */
1388         if (i >= 0) {
1389                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1390                 if (err == -ENOSPC)
1391                         err = -EEXIST;
1392         } else {
1393                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1394         }
1395         spin_unlock(&ublk_idr_lock);
1396
1397         if (err >= 0)
1398                 ub->ub_number = err;
1399
1400         return err;
1401 }
1402
1403 static void ublk_free_dev_number(struct ublk_device *ub)
1404 {
1405         spin_lock(&ublk_idr_lock);
1406         idr_remove(&ublk_index_idr, ub->ub_number);
1407         wake_up_all(&ublk_idr_wq);
1408         spin_unlock(&ublk_idr_lock);
1409 }
1410
1411 static void ublk_cdev_rel(struct device *dev)
1412 {
1413         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1414
1415         blk_mq_free_tag_set(&ub->tag_set);
1416         ublk_deinit_queues(ub);
1417         ublk_free_dev_number(ub);
1418         mutex_destroy(&ub->mutex);
1419         kfree(ub);
1420 }
1421
1422 static int ublk_add_chdev(struct ublk_device *ub)
1423 {
1424         struct device *dev = &ub->cdev_dev;
1425         int minor = ub->ub_number;
1426         int ret;
1427
1428         dev->parent = ublk_misc.this_device;
1429         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1430         dev->class = ublk_chr_class;
1431         dev->release = ublk_cdev_rel;
1432         device_initialize(dev);
1433
1434         ret = dev_set_name(dev, "ublkc%d", minor);
1435         if (ret)
1436                 goto fail;
1437
1438         cdev_init(&ub->cdev, &ublk_ch_fops);
1439         ret = cdev_device_add(&ub->cdev, dev);
1440         if (ret)
1441                 goto fail;
1442         return 0;
1443  fail:
1444         put_device(dev);
1445         return ret;
1446 }
1447
1448 static void ublk_stop_work_fn(struct work_struct *work)
1449 {
1450         struct ublk_device *ub =
1451                 container_of(work, struct ublk_device, stop_work);
1452
1453         ublk_stop_dev(ub);
1454 }
1455
1456 /* align max io buffer size with PAGE_SIZE */
1457 static void ublk_align_max_io_size(struct ublk_device *ub)
1458 {
1459         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1460
1461         ub->dev_info.max_io_buf_bytes =
1462                 round_down(max_io_bytes, PAGE_SIZE);
1463 }
1464
1465 static int ublk_add_tag_set(struct ublk_device *ub)
1466 {
1467         ub->tag_set.ops = &ublk_mq_ops;
1468         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1469         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1470         ub->tag_set.numa_node = NUMA_NO_NODE;
1471         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1472         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1473         ub->tag_set.driver_data = ub;
1474         return blk_mq_alloc_tag_set(&ub->tag_set);
1475 }
1476
1477 static void ublk_remove(struct ublk_device *ub)
1478 {
1479         ublk_stop_dev(ub);
1480         cancel_work_sync(&ub->stop_work);
1481         cancel_work_sync(&ub->quiesce_work);
1482         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1483         put_device(&ub->cdev_dev);
1484 }
1485
1486 static struct ublk_device *ublk_get_device_from_id(int idx)
1487 {
1488         struct ublk_device *ub = NULL;
1489
1490         if (idx < 0)
1491                 return NULL;
1492
1493         spin_lock(&ublk_idr_lock);
1494         ub = idr_find(&ublk_index_idr, idx);
1495         if (ub)
1496                 ub = ublk_get_device(ub);
1497         spin_unlock(&ublk_idr_lock);
1498
1499         return ub;
1500 }
1501
1502 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1503 {
1504         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1505         int ublksrv_pid = (int)header->data[0];
1506         struct ublk_device *ub;
1507         struct gendisk *disk;
1508         int ret = -EINVAL;
1509
1510         if (ublksrv_pid <= 0)
1511                 return -EINVAL;
1512
1513         ub = ublk_get_device_from_id(header->dev_id);
1514         if (!ub)
1515                 return -EINVAL;
1516
1517         wait_for_completion_interruptible(&ub->completion);
1518
1519         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1520
1521         mutex_lock(&ub->mutex);
1522         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1523             test_bit(UB_STATE_USED, &ub->state)) {
1524                 ret = -EEXIST;
1525                 goto out_unlock;
1526         }
1527
1528         disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1529         if (IS_ERR(disk)) {
1530                 ret = PTR_ERR(disk);
1531                 goto out_unlock;
1532         }
1533         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1534         disk->fops = &ub_fops;
1535         disk->private_data = ub;
1536
1537         ub->dev_info.ublksrv_pid = ublksrv_pid;
1538         ub->ub_disk = disk;
1539
1540         ret = ublk_apply_params(ub);
1541         if (ret)
1542                 goto out_put_disk;
1543
1544         get_device(&ub->cdev_dev);
1545         ret = add_disk(disk);
1546         if (ret) {
1547                 /*
1548                  * Has to drop the reference since ->free_disk won't be
1549                  * called in case of add_disk failure.
1550                  */
1551                 ublk_put_device(ub);
1552                 goto out_put_disk;
1553         }
1554         set_bit(UB_STATE_USED, &ub->state);
1555         ub->dev_info.state = UBLK_S_DEV_LIVE;
1556 out_put_disk:
1557         if (ret)
1558                 put_disk(disk);
1559 out_unlock:
1560         mutex_unlock(&ub->mutex);
1561         ublk_put_device(ub);
1562         return ret;
1563 }
1564
1565 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1566 {
1567         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1568         void __user *argp = (void __user *)(unsigned long)header->addr;
1569         struct ublk_device *ub;
1570         cpumask_var_t cpumask;
1571         unsigned long queue;
1572         unsigned int retlen;
1573         unsigned int i;
1574         int ret = -EINVAL;
1575
1576         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1577                 return -EINVAL;
1578         if (header->len & (sizeof(unsigned long)-1))
1579                 return -EINVAL;
1580         if (!header->addr)
1581                 return -EINVAL;
1582
1583         ub = ublk_get_device_from_id(header->dev_id);
1584         if (!ub)
1585                 return -EINVAL;
1586
1587         queue = header->data[0];
1588         if (queue >= ub->dev_info.nr_hw_queues)
1589                 goto out_put_device;
1590
1591         ret = -ENOMEM;
1592         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1593                 goto out_put_device;
1594
1595         for_each_possible_cpu(i) {
1596                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1597                         cpumask_set_cpu(i, cpumask);
1598         }
1599
1600         ret = -EFAULT;
1601         retlen = min_t(unsigned short, header->len, cpumask_size());
1602         if (copy_to_user(argp, cpumask, retlen))
1603                 goto out_free_cpumask;
1604         if (retlen != header->len &&
1605             clear_user(argp + retlen, header->len - retlen))
1606                 goto out_free_cpumask;
1607
1608         ret = 0;
1609 out_free_cpumask:
1610         free_cpumask_var(cpumask);
1611 out_put_device:
1612         ublk_put_device(ub);
1613         return ret;
1614 }
1615
1616 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1617 {
1618         pr_devel("%s: dev id %d flags %llx\n", __func__,
1619                         info->dev_id, info->flags);
1620         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1621                         info->nr_hw_queues, info->queue_depth);
1622 }
1623
1624 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1625 {
1626         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1627         void __user *argp = (void __user *)(unsigned long)header->addr;
1628         struct ublksrv_ctrl_dev_info info;
1629         struct ublk_device *ub;
1630         int ret = -EINVAL;
1631
1632         if (header->len < sizeof(info) || !header->addr)
1633                 return -EINVAL;
1634         if (header->queue_id != (u16)-1) {
1635                 pr_warn("%s: queue_id is wrong %x\n",
1636                         __func__, header->queue_id);
1637                 return -EINVAL;
1638         }
1639         if (copy_from_user(&info, argp, sizeof(info)))
1640                 return -EFAULT;
1641         ublk_dump_dev_info(&info);
1642         if (header->dev_id != info.dev_id) {
1643                 pr_warn("%s: dev id not match %u %u\n",
1644                         __func__, header->dev_id, info.dev_id);
1645                 return -EINVAL;
1646         }
1647
1648         ret = mutex_lock_killable(&ublk_ctl_mutex);
1649         if (ret)
1650                 return ret;
1651
1652         ret = -ENOMEM;
1653         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1654         if (!ub)
1655                 goto out_unlock;
1656         mutex_init(&ub->mutex);
1657         spin_lock_init(&ub->mm_lock);
1658         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1659         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1660         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1661
1662         ret = ublk_alloc_dev_number(ub, header->dev_id);
1663         if (ret < 0)
1664                 goto out_free_ub;
1665
1666         memcpy(&ub->dev_info, &info, sizeof(info));
1667
1668         /* update device id */
1669         ub->dev_info.dev_id = ub->ub_number;
1670
1671         /*
1672          * 64bit flags will be copied back to userspace as feature
1673          * negotiation result, so have to clear flags which driver
1674          * doesn't support yet, then userspace can get correct flags
1675          * (features) to handle.
1676          */
1677         ub->dev_info.flags &= UBLK_F_ALL;
1678
1679         if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1680                 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1681
1682         /* We are not ready to support zero copy */
1683         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1684
1685         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1686                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1687         ublk_align_max_io_size(ub);
1688
1689         ret = ublk_init_queues(ub);
1690         if (ret)
1691                 goto out_free_dev_number;
1692
1693         ret = ublk_add_tag_set(ub);
1694         if (ret)
1695                 goto out_deinit_queues;
1696
1697         ret = -EFAULT;
1698         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1699                 goto out_free_tag_set;
1700
1701         /*
1702          * Add the char dev so that ublksrv daemon can be setup.
1703          * ublk_add_chdev() will cleanup everything if it fails.
1704          */
1705         ret = ublk_add_chdev(ub);
1706         goto out_unlock;
1707
1708 out_free_tag_set:
1709         blk_mq_free_tag_set(&ub->tag_set);
1710 out_deinit_queues:
1711         ublk_deinit_queues(ub);
1712 out_free_dev_number:
1713         ublk_free_dev_number(ub);
1714 out_free_ub:
1715         mutex_destroy(&ub->mutex);
1716         kfree(ub);
1717 out_unlock:
1718         mutex_unlock(&ublk_ctl_mutex);
1719         return ret;
1720 }
1721
1722 static inline bool ublk_idr_freed(int id)
1723 {
1724         void *ptr;
1725
1726         spin_lock(&ublk_idr_lock);
1727         ptr = idr_find(&ublk_index_idr, id);
1728         spin_unlock(&ublk_idr_lock);
1729
1730         return ptr == NULL;
1731 }
1732
1733 static int ublk_ctrl_del_dev(int idx)
1734 {
1735         struct ublk_device *ub;
1736         int ret;
1737
1738         ret = mutex_lock_killable(&ublk_ctl_mutex);
1739         if (ret)
1740                 return ret;
1741
1742         ub = ublk_get_device_from_id(idx);
1743         if (ub) {
1744                 ublk_remove(ub);
1745                 ublk_put_device(ub);
1746                 ret = 0;
1747         } else {
1748                 ret = -ENODEV;
1749         }
1750
1751         /*
1752          * Wait until the idr is removed, then it can be reused after
1753          * DEL_DEV command is returned.
1754          */
1755         if (!ret)
1756                 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1757         mutex_unlock(&ublk_ctl_mutex);
1758
1759         return ret;
1760 }
1761
1762 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1763 {
1764         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1765
1766         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1767                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1768                         header->data[0], header->addr, header->len);
1769 }
1770
1771 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1772 {
1773         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1774         struct ublk_device *ub;
1775
1776         ub = ublk_get_device_from_id(header->dev_id);
1777         if (!ub)
1778                 return -EINVAL;
1779
1780         ublk_stop_dev(ub);
1781         cancel_work_sync(&ub->stop_work);
1782         cancel_work_sync(&ub->quiesce_work);
1783
1784         ublk_put_device(ub);
1785         return 0;
1786 }
1787
1788 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1789 {
1790         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1791         void __user *argp = (void __user *)(unsigned long)header->addr;
1792         struct ublk_device *ub;
1793         int ret = 0;
1794
1795         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1796                 return -EINVAL;
1797
1798         ub = ublk_get_device_from_id(header->dev_id);
1799         if (!ub)
1800                 return -EINVAL;
1801
1802         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1803                 ret = -EFAULT;
1804         ublk_put_device(ub);
1805
1806         return ret;
1807 }
1808
1809 static int ublk_ctrl_get_params(struct io_uring_cmd *cmd)
1810 {
1811         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1812         void __user *argp = (void __user *)(unsigned long)header->addr;
1813         struct ublk_params_header ph;
1814         struct ublk_device *ub;
1815         int ret;
1816
1817         if (header->len <= sizeof(ph) || !header->addr)
1818                 return -EINVAL;
1819
1820         if (copy_from_user(&ph, argp, sizeof(ph)))
1821                 return -EFAULT;
1822
1823         if (ph.len > header->len || !ph.len)
1824                 return -EINVAL;
1825
1826         if (ph.len > sizeof(struct ublk_params))
1827                 ph.len = sizeof(struct ublk_params);
1828
1829         ub = ublk_get_device_from_id(header->dev_id);
1830         if (!ub)
1831                 return -EINVAL;
1832
1833         mutex_lock(&ub->mutex);
1834         if (copy_to_user(argp, &ub->params, ph.len))
1835                 ret = -EFAULT;
1836         else
1837                 ret = 0;
1838         mutex_unlock(&ub->mutex);
1839
1840         ublk_put_device(ub);
1841         return ret;
1842 }
1843
1844 static int ublk_ctrl_set_params(struct io_uring_cmd *cmd)
1845 {
1846         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1847         void __user *argp = (void __user *)(unsigned long)header->addr;
1848         struct ublk_params_header ph;
1849         struct ublk_device *ub;
1850         int ret = -EFAULT;
1851
1852         if (header->len <= sizeof(ph) || !header->addr)
1853                 return -EINVAL;
1854
1855         if (copy_from_user(&ph, argp, sizeof(ph)))
1856                 return -EFAULT;
1857
1858         if (ph.len > header->len || !ph.len || !ph.types)
1859                 return -EINVAL;
1860
1861         if (ph.len > sizeof(struct ublk_params))
1862                 ph.len = sizeof(struct ublk_params);
1863
1864         ub = ublk_get_device_from_id(header->dev_id);
1865         if (!ub)
1866                 return -EINVAL;
1867
1868         /* parameters can only be changed when device isn't live */
1869         mutex_lock(&ub->mutex);
1870         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1871                 ret = -EACCES;
1872         } else if (copy_from_user(&ub->params, argp, ph.len)) {
1873                 ret = -EFAULT;
1874         } else {
1875                 /* clear all we don't support yet */
1876                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1877                 ret = ublk_validate_params(ub);
1878         }
1879         mutex_unlock(&ub->mutex);
1880         ublk_put_device(ub);
1881
1882         return ret;
1883 }
1884
1885 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1886 {
1887         int i;
1888
1889         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1890         /* All old ioucmds have to be completed */
1891         WARN_ON_ONCE(ubq->nr_io_ready);
1892         /* old daemon is PF_EXITING, put it now */
1893         put_task_struct(ubq->ubq_daemon);
1894         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1895         ubq->ubq_daemon = NULL;
1896
1897         for (i = 0; i < ubq->q_depth; i++) {
1898                 struct ublk_io *io = &ubq->ios[i];
1899
1900                 /* forget everything now and be ready for new FETCH_REQ */
1901                 io->flags = 0;
1902                 io->cmd = NULL;
1903                 io->addr = 0;
1904         }
1905 }
1906
1907 static int ublk_ctrl_start_recovery(struct io_uring_cmd *cmd)
1908 {
1909         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1910         struct ublk_device *ub;
1911         int ret = -EINVAL;
1912         int i;
1913
1914         ub = ublk_get_device_from_id(header->dev_id);
1915         if (!ub)
1916                 return ret;
1917
1918         mutex_lock(&ub->mutex);
1919         if (!ublk_can_use_recovery(ub))
1920                 goto out_unlock;
1921         /*
1922          * START_RECOVERY is only allowd after:
1923          *
1924          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1925          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
1926          *     released.
1927          *
1928          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1929          *     (a)has quiesced request queue
1930          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
1931          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1932          *     (d)has completed/camceled all ioucmds owned by ther dying process
1933          */
1934         if (test_bit(UB_STATE_OPEN, &ub->state) ||
1935                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1936                 ret = -EBUSY;
1937                 goto out_unlock;
1938         }
1939         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1940         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1941                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1942         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1943         ub->mm = NULL;
1944         ub->nr_queues_ready = 0;
1945         init_completion(&ub->completion);
1946         ret = 0;
1947  out_unlock:
1948         mutex_unlock(&ub->mutex);
1949         ublk_put_device(ub);
1950         return ret;
1951 }
1952
1953 static int ublk_ctrl_end_recovery(struct io_uring_cmd *cmd)
1954 {
1955         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1956         int ublksrv_pid = (int)header->data[0];
1957         struct ublk_device *ub;
1958         int ret = -EINVAL;
1959
1960         ub = ublk_get_device_from_id(header->dev_id);
1961         if (!ub)
1962                 return ret;
1963
1964         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1965                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1966         /* wait until new ubq_daemon sending all FETCH_REQ */
1967         wait_for_completion_interruptible(&ub->completion);
1968         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
1969                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1970
1971         mutex_lock(&ub->mutex);
1972         if (!ublk_can_use_recovery(ub))
1973                 goto out_unlock;
1974
1975         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1976                 ret = -EBUSY;
1977                 goto out_unlock;
1978         }
1979         ub->dev_info.ublksrv_pid = ublksrv_pid;
1980         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
1981                         __func__, ublksrv_pid, header->dev_id);
1982         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1983         pr_devel("%s: queue unquiesced, dev id %d.\n",
1984                         __func__, header->dev_id);
1985         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1986         ub->dev_info.state = UBLK_S_DEV_LIVE;
1987         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1988         ret = 0;
1989  out_unlock:
1990         mutex_unlock(&ub->mutex);
1991         ublk_put_device(ub);
1992         return ret;
1993 }
1994
1995 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1996                 unsigned int issue_flags)
1997 {
1998         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1999         int ret = -EINVAL;
2000
2001         ublk_ctrl_cmd_dump(cmd);
2002
2003         if (!(issue_flags & IO_URING_F_SQE128))
2004                 goto out;
2005
2006         ret = -EPERM;
2007         if (!capable(CAP_SYS_ADMIN))
2008                 goto out;
2009
2010         ret = -ENODEV;
2011         switch (cmd->cmd_op) {
2012         case UBLK_CMD_START_DEV:
2013                 ret = ublk_ctrl_start_dev(cmd);
2014                 break;
2015         case UBLK_CMD_STOP_DEV:
2016                 ret = ublk_ctrl_stop_dev(cmd);
2017                 break;
2018         case UBLK_CMD_GET_DEV_INFO:
2019                 ret = ublk_ctrl_get_dev_info(cmd);
2020                 break;
2021         case UBLK_CMD_ADD_DEV:
2022                 ret = ublk_ctrl_add_dev(cmd);
2023                 break;
2024         case UBLK_CMD_DEL_DEV:
2025                 ret = ublk_ctrl_del_dev(header->dev_id);
2026                 break;
2027         case UBLK_CMD_GET_QUEUE_AFFINITY:
2028                 ret = ublk_ctrl_get_queue_affinity(cmd);
2029                 break;
2030         case UBLK_CMD_GET_PARAMS:
2031                 ret = ublk_ctrl_get_params(cmd);
2032                 break;
2033         case UBLK_CMD_SET_PARAMS:
2034                 ret = ublk_ctrl_set_params(cmd);
2035                 break;
2036         case UBLK_CMD_START_USER_RECOVERY:
2037                 ret = ublk_ctrl_start_recovery(cmd);
2038                 break;
2039         case UBLK_CMD_END_USER_RECOVERY:
2040                 ret = ublk_ctrl_end_recovery(cmd);
2041                 break;
2042         default:
2043                 break;
2044         }
2045  out:
2046         io_uring_cmd_done(cmd, ret, 0);
2047         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2048                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2049         return -EIOCBQUEUED;
2050 }
2051
2052 static const struct file_operations ublk_ctl_fops = {
2053         .open           = nonseekable_open,
2054         .uring_cmd      = ublk_ctrl_uring_cmd,
2055         .owner          = THIS_MODULE,
2056         .llseek         = noop_llseek,
2057 };
2058
2059 static struct miscdevice ublk_misc = {
2060         .minor          = MISC_DYNAMIC_MINOR,
2061         .name           = "ublk-control",
2062         .fops           = &ublk_ctl_fops,
2063 };
2064
2065 static int __init ublk_init(void)
2066 {
2067         int ret;
2068
2069         init_waitqueue_head(&ublk_idr_wq);
2070
2071         ret = misc_register(&ublk_misc);
2072         if (ret)
2073                 return ret;
2074
2075         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2076         if (ret)
2077                 goto unregister_mis;
2078
2079         ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2080         if (IS_ERR(ublk_chr_class)) {
2081                 ret = PTR_ERR(ublk_chr_class);
2082                 goto free_chrdev_region;
2083         }
2084         return 0;
2085
2086 free_chrdev_region:
2087         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2088 unregister_mis:
2089         misc_deregister(&ublk_misc);
2090         return ret;
2091 }
2092
2093 static void __exit ublk_exit(void)
2094 {
2095         struct ublk_device *ub;
2096         int id;
2097
2098         class_destroy(ublk_chr_class);
2099
2100         misc_deregister(&ublk_misc);
2101
2102         idr_for_each_entry(&ublk_index_idr, ub, id)
2103                 ublk_remove(ub);
2104
2105         idr_destroy(&ublk_index_idr);
2106         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2107 }
2108
2109 module_init(ublk_init);
2110 module_exit(ublk_exit);
2111
2112 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2113 MODULE_LICENSE("GPL");