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                 | UBLK_F_USER_COPY \
  60                 | UBLK_F_ZONED)
  61
  62 /* All UBLK_PARAM_TYPE_* should be included here */
  63 #define UBLK_PARAM_TYPE_ALL                                \
  64         (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD | \
  65          UBLK_PARAM_TYPE_DEVT | UBLK_PARAM_TYPE_ZONED)
  66
  67 struct ublk_rq_data {
  68         struct llist_node node;
  69
  70         struct kref ref;
  71         __u64 sector;
  72         __u32 operation;
  73         __u32 nr_zones;
  74 };
  75
  76 struct ublk_uring_cmd_pdu {
  77         struct ublk_queue *ubq;
  78 };
  79
  80 /*
  81  * io command is active: sqe cmd is received, and its cqe isn't done
  82  *
  83  * If the flag is set, the io command is owned by ublk driver, and waited
  84  * for incoming blk-mq request from the ublk block device.
  85  *
  86  * If the flag is cleared, the io command will be completed, and owned by
  87  * ublk server.
  88  */
  89 #define UBLK_IO_FLAG_ACTIVE     0x01
  90
  91 /*
  92  * IO command is completed via cqe, and it is being handled by ublksrv, and
  93  * not committed yet
  94  *
  95  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  96  * cross verification
  97  */
  98 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  99
 100 /*
 101  * IO command is aborted, so this flag is set in case of
 102  * !UBLK_IO_FLAG_ACTIVE.
 103  *
 104  * After this flag is observed, any pending or new incoming request
 105  * associated with this io command will be failed immediately
 106  */
 107 #define UBLK_IO_FLAG_ABORTED 0x04
 108
 109 /*
 110  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
 111  * get data buffer address from ublksrv.
 112  *
 113  * Then, bio data could be copied into this data buffer for a WRITE request
 114  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 115  */
 116 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 117
 118 struct ublk_io {
 119         /* userspace buffer address from io cmd */
 120         __u64   addr;
 121         unsigned int flags;
 122         int res;
 123
 124         struct io_uring_cmd *cmd;
 125 };
 126
 127 struct ublk_queue {
 128         int q_id;
 129         int q_depth;
 130
 131         unsigned long flags;
 132         struct task_struct      *ubq_daemon;
 133         char *io_cmd_buf;
 134
 135         struct llist_head       io_cmds;
 136
 137         unsigned long io_addr;  /* mapped vm address */
 138         unsigned int max_io_sz;
 139         bool force_abort;
 140         bool timeout;
 141         unsigned short nr_io_ready;     /* how many ios setup */
 142         struct ublk_device *dev;
 143         struct ublk_io ios[];
 144 };
 145
 146 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 147
 148 struct ublk_device {
 149         struct gendisk          *ub_disk;
 150
 151         char    *__queues;
 152
 153         unsigned int    queue_size;
 154         struct ublksrv_ctrl_dev_info    dev_info;
 155
 156         struct blk_mq_tag_set   tag_set;
 157
 158         struct cdev             cdev;
 159         struct device           cdev_dev;
 160
 161 #define UB_STATE_OPEN           0
 162 #define UB_STATE_USED           1
 163 #define UB_STATE_DELETED        2
 164         unsigned long           state;
 165         int                     ub_number;
 166
 167         struct mutex            mutex;
 168
 169         spinlock_t              mm_lock;
 170         struct mm_struct        *mm;
 171
 172         struct ublk_params      params;
 173
 174         struct completion       completion;
 175         unsigned int            nr_queues_ready;
 176         unsigned int            nr_privileged_daemon;
 177
 178         /*
 179          * Our ubq->daemon may be killed without any notification, so
 180          * monitor each queue's daemon periodically
 181          */
 182         struct delayed_work     monitor_work;
 183         struct work_struct      quiesce_work;
 184         struct work_struct      stop_work;
 185 };
 186
 187 /* header of ublk_params */
 188 struct ublk_params_header {
 189         __u32   len;
 190         __u32   types;
 191 };
 192
 193 static inline unsigned int ublk_req_build_flags(struct request *req);
 194 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 195                                                    int tag);
 196
 197 static inline bool ublk_dev_is_user_copy(const struct ublk_device *ub)
 198 {
 199         return ub->dev_info.flags & UBLK_F_USER_COPY;
 200 }
 201
 202 static inline bool ublk_dev_is_zoned(const struct ublk_device *ub)
 203 {
 204         return ub->dev_info.flags & UBLK_F_ZONED;
 205 }
 206
 207 static inline bool ublk_queue_is_zoned(struct ublk_queue *ubq)
 208 {
 209         return ubq->flags & UBLK_F_ZONED;
 210 }
 211
 212 #ifdef CONFIG_BLK_DEV_ZONED
 213
 214 static int ublk_get_nr_zones(const struct ublk_device *ub)
 215 {
 216         const struct ublk_param_basic *p = &ub->params.basic;
 217
 218         /* Zone size is a power of 2 */
 219         return p->dev_sectors >> ilog2(p->chunk_sectors);
 220 }
 221
 222 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 223 {
 224         return blk_revalidate_disk_zones(ub->ub_disk, NULL);
 225 }
 226
 227 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 228 {
 229         const struct ublk_param_zoned *p = &ub->params.zoned;
 230         int nr_zones;
 231
 232         if (!ublk_dev_is_zoned(ub))
 233                 return -EINVAL;
 234
 235         if (!p->max_zone_append_sectors)
 236                 return -EINVAL;
 237
 238         nr_zones = ublk_get_nr_zones(ub);
 239
 240         if (p->max_active_zones > nr_zones)
 241                 return -EINVAL;
 242
 243         if (p->max_open_zones > nr_zones)
 244                 return -EINVAL;
 245
 246         return 0;
 247 }
 248
 249 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
 250 {
 251         const struct ublk_param_zoned *p = &ub->params.zoned;
 252
 253         disk_set_zoned(ub->ub_disk, BLK_ZONED_HM);
 254         blk_queue_required_elevator_features(ub->ub_disk->queue,
 255                                              ELEVATOR_F_ZBD_SEQ_WRITE);
 256         disk_set_max_active_zones(ub->ub_disk, p->max_active_zones);
 257         disk_set_max_open_zones(ub->ub_disk, p->max_open_zones);
 258         blk_queue_max_zone_append_sectors(ub->ub_disk->queue, p->max_zone_append_sectors);
 259
 260         ub->ub_disk->nr_zones = ublk_get_nr_zones(ub);
 261
 262         return 0;
 263 }
 264
 265 /* Based on virtblk_alloc_report_buffer */
 266 static void *ublk_alloc_report_buffer(struct ublk_device *ublk,
 267                                       unsigned int nr_zones, size_t *buflen)
 268 {
 269         struct request_queue *q = ublk->ub_disk->queue;
 270         size_t bufsize;
 271         void *buf;
 272
 273         nr_zones = min_t(unsigned int, nr_zones,
 274                          ublk->ub_disk->nr_zones);
 275
 276         bufsize = nr_zones * sizeof(struct blk_zone);
 277         bufsize =
 278                 min_t(size_t, bufsize, queue_max_hw_sectors(q) << SECTOR_SHIFT);
 279
 280         while (bufsize >= sizeof(struct blk_zone)) {
 281                 buf = kvmalloc(bufsize, GFP_KERNEL | __GFP_NORETRY);
 282                 if (buf) {
 283                         *buflen = bufsize;
 284                         return buf;
 285                 }
 286                 bufsize >>= 1;
 287         }
 288
 289         *buflen = 0;
 290         return NULL;
 291 }
 292
 293 static int ublk_report_zones(struct gendisk *disk, sector_t sector,
 294                       unsigned int nr_zones, report_zones_cb cb, void *data)
 295 {
 296         struct ublk_device *ub = disk->private_data;
 297         unsigned int zone_size_sectors = disk->queue->limits.chunk_sectors;
 298         unsigned int first_zone = sector >> ilog2(zone_size_sectors);
 299         unsigned int done_zones = 0;
 300         unsigned int max_zones_per_request;
 301         int ret;
 302         struct blk_zone *buffer;
 303         size_t buffer_length;
 304
 305         nr_zones = min_t(unsigned int, ub->ub_disk->nr_zones - first_zone,
 306                          nr_zones);
 307
 308         buffer = ublk_alloc_report_buffer(ub, nr_zones, &buffer_length);
 309         if (!buffer)
 310                 return -ENOMEM;
 311
 312         max_zones_per_request = buffer_length / sizeof(struct blk_zone);
 313
 314         while (done_zones < nr_zones) {
 315                 unsigned int remaining_zones = nr_zones - done_zones;
 316                 unsigned int zones_in_request =
 317                         min_t(unsigned int, remaining_zones, max_zones_per_request);
 318                 struct request *req;
 319                 struct ublk_rq_data *pdu;
 320                 blk_status_t status;
 321
 322                 memset(buffer, 0, buffer_length);
 323
 324                 req = blk_mq_alloc_request(disk->queue, REQ_OP_DRV_IN, 0);
 325                 if (IS_ERR(req)) {
 326                         ret = PTR_ERR(req);
 327                         goto out;
 328                 }
 329
 330                 pdu = blk_mq_rq_to_pdu(req);
 331                 pdu->operation = UBLK_IO_OP_REPORT_ZONES;
 332                 pdu->sector = sector;
 333                 pdu->nr_zones = zones_in_request;
 334
 335                 ret = blk_rq_map_kern(disk->queue, req, buffer, buffer_length,
 336                                         GFP_KERNEL);
 337                 if (ret) {
 338                         blk_mq_free_request(req);
 339                         goto out;
 340                 }
 341
 342                 status = blk_execute_rq(req, 0);
 343                 ret = blk_status_to_errno(status);
 344                 blk_mq_free_request(req);
 345                 if (ret)
 346                         goto out;
 347
 348                 for (unsigned int i = 0; i < zones_in_request; i++) {
 349                         struct blk_zone *zone = buffer + i;
 350
 351                         /* A zero length zone means no more zones in this response */
 352                         if (!zone->len)
 353                                 break;
 354
 355                         ret = cb(zone, i, data);
 356                         if (ret)
 357                                 goto out;
 358
 359                         done_zones++;
 360                         sector += zone_size_sectors;
 361
 362                 }
 363         }
 364
 365         ret = done_zones;
 366
 367 out:
 368         kvfree(buffer);
 369         return ret;
 370 }
 371
 372 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 373                                          struct request *req)
 374 {
 375         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 376         struct ublk_io *io = &ubq->ios[req->tag];
 377         struct ublk_rq_data *pdu = blk_mq_rq_to_pdu(req);
 378         u32 ublk_op;
 379
 380         switch (req_op(req)) {
 381         case REQ_OP_ZONE_OPEN:
 382                 ublk_op = UBLK_IO_OP_ZONE_OPEN;
 383                 break;
 384         case REQ_OP_ZONE_CLOSE:
 385                 ublk_op = UBLK_IO_OP_ZONE_CLOSE;
 386                 break;
 387         case REQ_OP_ZONE_FINISH:
 388                 ublk_op = UBLK_IO_OP_ZONE_FINISH;
 389                 break;
 390         case REQ_OP_ZONE_RESET:
 391                 ublk_op = UBLK_IO_OP_ZONE_RESET;
 392                 break;
 393         case REQ_OP_ZONE_APPEND:
 394                 ublk_op = UBLK_IO_OP_ZONE_APPEND;
 395                 break;
 396         case REQ_OP_DRV_IN:
 397                 ublk_op = pdu->operation;
 398                 switch (ublk_op) {
 399                 case UBLK_IO_OP_REPORT_ZONES:
 400                         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 401                         iod->nr_zones = pdu->nr_zones;
 402                         iod->start_sector = pdu->sector;
 403                         return BLK_STS_OK;
 404                 default:
 405                         return BLK_STS_IOERR;
 406                 }
 407         case REQ_OP_ZONE_RESET_ALL:
 408         case REQ_OP_DRV_OUT:
 409                 /* We do not support reset_all and drv_out */
 410                 return BLK_STS_NOTSUPP;
 411         default:
 412                 return BLK_STS_IOERR;
 413         }
 414
 415         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 416         iod->nr_sectors = blk_rq_sectors(req);
 417         iod->start_sector = blk_rq_pos(req);
 418         iod->addr = io->addr;
 419
 420         return BLK_STS_OK;
 421 }
 422
 423 #else
 424
 425 #define ublk_report_zones (NULL)
 426
 427 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 428 {
 429         return -EOPNOTSUPP;
 430 }
 431
 432 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
 433 {
 434         return -EOPNOTSUPP;
 435 }
 436
 437 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 438 {
 439         return 0;
 440 }
 441
 442 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 443                                          struct request *req)
 444 {
 445         return BLK_STS_NOTSUPP;
 446 }
 447
 448 #endif
 449
 450 static inline void __ublk_complete_rq(struct request *req);
 451 static void ublk_complete_rq(struct kref *ref);
 452
 453 static dev_t ublk_chr_devt;
 454 static const struct class ublk_chr_class = {
 455         .name = "ublk-char",
 456 };
 457
 458 static DEFINE_IDR(ublk_index_idr);
 459 static DEFINE_SPINLOCK(ublk_idr_lock);
 460 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 461
 462 static DEFINE_MUTEX(ublk_ctl_mutex);
 463
 464 /*
 465  * Max ublk devices allowed to add
 466  *
 467  * It can be extended to one per-user limit in future or even controlled
 468  * by cgroup.
 469  */
 470 static unsigned int ublks_max = 64;
 471 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 472
 473 static struct miscdevice ublk_misc;
 474
 475 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 476 {
 477         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 478                 UBLK_QID_BITS_MASK;
 479 }
 480
 481 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 482 {
 483         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 484 }
 485
 486 static inline unsigned ublk_pos_to_tag(loff_t pos)
 487 {
 488         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 489                 UBLK_TAG_BITS_MASK;
 490 }
 491
 492 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 493 {
 494         struct request_queue *q = ub->ub_disk->queue;
 495         const struct ublk_param_basic *p = &ub->params.basic;
 496
 497         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 498         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 499         blk_queue_io_min(q, 1 << p->io_min_shift);
 500         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 501
 502         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 503                         p->attrs & UBLK_ATTR_FUA);
 504         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 505                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 506         else
 507                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 508
 509         blk_queue_max_hw_sectors(q, p->max_sectors);
 510         blk_queue_chunk_sectors(q, p->chunk_sectors);
 511         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 512
 513         if (p->attrs & UBLK_ATTR_READ_ONLY)
 514                 set_disk_ro(ub->ub_disk, true);
 515
 516         set_capacity(ub->ub_disk, p->dev_sectors);
 517 }
 518
 519 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 520 {
 521         struct request_queue *q = ub->ub_disk->queue;
 522         const struct ublk_param_discard *p = &ub->params.discard;
 523
 524         q->limits.discard_alignment = p->discard_alignment;
 525         q->limits.discard_granularity = p->discard_granularity;
 526         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 527         blk_queue_max_write_zeroes_sectors(q,
 528                         p->max_write_zeroes_sectors);
 529         blk_queue_max_discard_segments(q, p->max_discard_segments);
 530 }
 531
 532 static int ublk_validate_params(const struct ublk_device *ub)
 533 {
 534         /* basic param is the only one which must be set */
 535         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 536                 const struct ublk_param_basic *p = &ub->params.basic;
 537
 538                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 539                         return -EINVAL;
 540
 541                 if (p->logical_bs_shift > p->physical_bs_shift)
 542                         return -EINVAL;
 543
 544                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 545                         return -EINVAL;
 546
 547                 if (ublk_dev_is_zoned(ub) && !p->chunk_sectors)
 548                         return -EINVAL;
 549         } else
 550                 return -EINVAL;
 551
 552         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 553                 const struct ublk_param_discard *p = &ub->params.discard;
 554
 555                 /* So far, only support single segment discard */
 556                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 557                         return -EINVAL;
 558
 559                 if (!p->discard_granularity)
 560                         return -EINVAL;
 561         }
 562
 563         /* dev_t is read-only */
 564         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 565                 return -EINVAL;
 566
 567         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 568                 return ublk_dev_param_zoned_validate(ub);
 569         else if (ublk_dev_is_zoned(ub))
 570                 return -EINVAL;
 571
 572         return 0;
 573 }
 574
 575 static int ublk_apply_params(struct ublk_device *ub)
 576 {
 577         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 578                 return -EINVAL;
 579
 580         ublk_dev_param_basic_apply(ub);
 581
 582         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 583                 ublk_dev_param_discard_apply(ub);
 584
 585         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 586                 return ublk_dev_param_zoned_apply(ub);
 587
 588         return 0;
 589 }
 590
 591 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 592 {
 593         return ubq->flags & UBLK_F_USER_COPY;
 594 }
 595
 596 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 597 {
 598         /*
 599          * read()/write() is involved in user copy, so request reference
 600          * has to be grabbed
 601          */
 602         return ublk_support_user_copy(ubq);
 603 }
 604
 605 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 606                 struct request *req)
 607 {
 608         if (ublk_need_req_ref(ubq)) {
 609                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 610
 611                 kref_init(&data->ref);
 612         }
 613 }
 614
 615 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 616                 struct request *req)
 617 {
 618         if (ublk_need_req_ref(ubq)) {
 619                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 620
 621                 return kref_get_unless_zero(&data->ref);
 622         }
 623
 624         return true;
 625 }
 626
 627 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 628                 struct request *req)
 629 {
 630         if (ublk_need_req_ref(ubq)) {
 631                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 632
 633                 kref_put(&data->ref, ublk_complete_rq);
 634         } else {
 635                 __ublk_complete_rq(req);
 636         }
 637 }
 638
 639 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 640 {
 641         return ubq->flags & UBLK_F_NEED_GET_DATA;
 642 }
 643
 644 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 645 {
 646         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 647                 return ub;
 648         return NULL;
 649 }
 650
 651 static void ublk_put_device(struct ublk_device *ub)
 652 {
 653         put_device(&ub->cdev_dev);
 654 }
 655
 656 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 657                 int qid)
 658 {
 659        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 660 }
 661
 662 static inline bool ublk_rq_has_data(const struct request *rq)
 663 {
 664         return bio_has_data(rq->bio);
 665 }
 666
 667 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 668                 int tag)
 669 {
 670         return (struct ublksrv_io_desc *)
 671                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 672 }
 673
 674 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 675 {
 676         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 677 }
 678
 679 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 680 {
 681         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 682
 683         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 684                         PAGE_SIZE);
 685 }
 686
 687 static inline bool ublk_queue_can_use_recovery_reissue(
 688                 struct ublk_queue *ubq)
 689 {
 690         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 691                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 692 }
 693
 694 static inline bool ublk_queue_can_use_recovery(
 695                 struct ublk_queue *ubq)
 696 {
 697         return ubq->flags & UBLK_F_USER_RECOVERY;
 698 }
 699
 700 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 701 {
 702         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 703 }
 704
 705 static void ublk_free_disk(struct gendisk *disk)
 706 {
 707         struct ublk_device *ub = disk->private_data;
 708
 709         clear_bit(UB_STATE_USED, &ub->state);
 710         put_device(&ub->cdev_dev);
 711 }
 712
 713 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 714                 unsigned int *owner_gid)
 715 {
 716         kuid_t uid;
 717         kgid_t gid;
 718
 719         current_uid_gid(&uid, &gid);
 720
 721         *owner_uid = from_kuid(&init_user_ns, uid);
 722         *owner_gid = from_kgid(&init_user_ns, gid);
 723 }
 724
 725 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 726 {
 727         struct ublk_device *ub = disk->private_data;
 728
 729         if (capable(CAP_SYS_ADMIN))
 730                 return 0;
 731
 732         /*
 733          * If it is one unprivileged device, only owner can open
 734          * the disk. Otherwise it could be one trap made by one
 735          * evil user who grants this disk's privileges to other
 736          * users deliberately.
 737          *
 738          * This way is reasonable too given anyone can create
 739          * unprivileged device, and no need other's grant.
 740          */
 741         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 742                 unsigned int curr_uid, curr_gid;
 743
 744                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 745
 746                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 747                                 ub->dev_info.owner_gid)
 748                         return -EPERM;
 749         }
 750
 751         return 0;
 752 }
 753
 754 static const struct block_device_operations ub_fops = {
 755         .owner =        THIS_MODULE,
 756         .open =         ublk_open,
 757         .free_disk =    ublk_free_disk,
 758         .report_zones = ublk_report_zones,
 759 };
 760
 761 #define UBLK_MAX_PIN_PAGES      32
 762
 763 struct ublk_io_iter {
 764         struct page *pages[UBLK_MAX_PIN_PAGES];
 765         struct bio *bio;
 766         struct bvec_iter iter;
 767 };
 768
 769 /* return how many pages are copied */
 770 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 771                 size_t total, size_t pg_off, int dir)
 772 {
 773         unsigned done = 0;
 774         unsigned pg_idx = 0;
 775
 776         while (done < total) {
 777                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 778                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 779                                 (unsigned)(PAGE_SIZE - pg_off));
 780                 void *bv_buf = bvec_kmap_local(&bv);
 781                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 782
 783                 if (dir == ITER_DEST)
 784                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 785                 else
 786                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 787
 788                 kunmap_local(pg_buf);
 789                 kunmap_local(bv_buf);
 790
 791                 /* advance page array */
 792                 pg_off += bytes;
 793                 if (pg_off == PAGE_SIZE) {
 794                         pg_idx += 1;
 795                         pg_off = 0;
 796                 }
 797
 798                 done += bytes;
 799
 800                 /* advance bio */
 801                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 802                 if (!data->iter.bi_size) {
 803                         data->bio = data->bio->bi_next;
 804                         if (data->bio == NULL)
 805                                 break;
 806                         data->iter = data->bio->bi_iter;
 807                 }
 808         }
 809 }
 810
 811 static bool ublk_advance_io_iter(const struct request *req,
 812                 struct ublk_io_iter *iter, unsigned int offset)
 813 {
 814         struct bio *bio = req->bio;
 815
 816         for_each_bio(bio) {
 817                 if (bio->bi_iter.bi_size > offset) {
 818                         iter->bio = bio;
 819                         iter->iter = bio->bi_iter;
 820                         bio_advance_iter(iter->bio, &iter->iter, offset);
 821                         return true;
 822                 }
 823                 offset -= bio->bi_iter.bi_size;
 824         }
 825         return false;
 826 }
 827
 828 /*
 829  * Copy data between request pages and io_iter, and 'offset'
 830  * is the start point of linear offset of request.
 831  */
 832 static size_t ublk_copy_user_pages(const struct request *req,
 833                 unsigned offset, struct iov_iter *uiter, int dir)
 834 {
 835         struct ublk_io_iter iter;
 836         size_t done = 0;
 837
 838         if (!ublk_advance_io_iter(req, &iter, offset))
 839                 return 0;
 840
 841         while (iov_iter_count(uiter) && iter.bio) {
 842                 unsigned nr_pages;
 843                 ssize_t len;
 844                 size_t off;
 845                 int i;
 846
 847                 len = iov_iter_get_pages2(uiter, iter.pages,
 848                                 iov_iter_count(uiter),
 849                                 UBLK_MAX_PIN_PAGES, &off);
 850                 if (len <= 0)
 851                         return done;
 852
 853                 ublk_copy_io_pages(&iter, len, off, dir);
 854                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 855                 for (i = 0; i < nr_pages; i++) {
 856                         if (dir == ITER_DEST)
 857                                 set_page_dirty(iter.pages[i]);
 858                         put_page(iter.pages[i]);
 859                 }
 860                 done += len;
 861         }
 862
 863         return done;
 864 }
 865
 866 static inline bool ublk_need_map_req(const struct request *req)
 867 {
 868         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 869 }
 870
 871 static inline bool ublk_need_unmap_req(const struct request *req)
 872 {
 873         return ublk_rq_has_data(req) &&
 874                (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN);
 875 }
 876
 877 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 878                 struct ublk_io *io)
 879 {
 880         const unsigned int rq_bytes = blk_rq_bytes(req);
 881
 882         if (ublk_support_user_copy(ubq))
 883                 return rq_bytes;
 884
 885         /*
 886          * no zero copy, we delay copy WRITE request data into ublksrv
 887          * context and the big benefit is that pinning pages in current
 888          * context is pretty fast, see ublk_pin_user_pages
 889          */
 890         if (ublk_need_map_req(req)) {
 891                 struct iov_iter iter;
 892                 struct iovec iov;
 893                 const int dir = ITER_DEST;
 894
 895                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 896                                 &iov, &iter);
 897
 898                 return ublk_copy_user_pages(req, 0, &iter, dir);
 899         }
 900         return rq_bytes;
 901 }
 902
 903 static int ublk_unmap_io(const struct ublk_queue *ubq,
 904                 const struct request *req,
 905                 struct ublk_io *io)
 906 {
 907         const unsigned int rq_bytes = blk_rq_bytes(req);
 908
 909         if (ublk_support_user_copy(ubq))
 910                 return rq_bytes;
 911
 912         if (ublk_need_unmap_req(req)) {
 913                 struct iov_iter iter;
 914                 struct iovec iov;
 915                 const int dir = ITER_SOURCE;
 916
 917                 WARN_ON_ONCE(io->res > rq_bytes);
 918
 919                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 920                                 &iov, &iter);
 921                 return ublk_copy_user_pages(req, 0, &iter, dir);
 922         }
 923         return rq_bytes;
 924 }
 925
 926 static inline unsigned int ublk_req_build_flags(struct request *req)
 927 {
 928         unsigned flags = 0;
 929
 930         if (req->cmd_flags & REQ_FAILFAST_DEV)
 931                 flags |= UBLK_IO_F_FAILFAST_DEV;
 932
 933         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 934                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 935
 936         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 937                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 938
 939         if (req->cmd_flags & REQ_META)
 940                 flags |= UBLK_IO_F_META;
 941
 942         if (req->cmd_flags & REQ_FUA)
 943                 flags |= UBLK_IO_F_FUA;
 944
 945         if (req->cmd_flags & REQ_NOUNMAP)
 946                 flags |= UBLK_IO_F_NOUNMAP;
 947
 948         if (req->cmd_flags & REQ_SWAP)
 949                 flags |= UBLK_IO_F_SWAP;
 950
 951         return flags;
 952 }
 953
 954 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 955 {
 956         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 957         struct ublk_io *io = &ubq->ios[req->tag];
 958         enum req_op op = req_op(req);
 959         u32 ublk_op;
 960
 961         if (!ublk_queue_is_zoned(ubq) &&
 962             (op_is_zone_mgmt(op) || op == REQ_OP_ZONE_APPEND))
 963                 return BLK_STS_IOERR;
 964
 965         switch (req_op(req)) {
 966         case REQ_OP_READ:
 967                 ublk_op = UBLK_IO_OP_READ;
 968                 break;
 969         case REQ_OP_WRITE:
 970                 ublk_op = UBLK_IO_OP_WRITE;
 971                 break;
 972         case REQ_OP_FLUSH:
 973                 ublk_op = UBLK_IO_OP_FLUSH;
 974                 break;
 975         case REQ_OP_DISCARD:
 976                 ublk_op = UBLK_IO_OP_DISCARD;
 977                 break;
 978         case REQ_OP_WRITE_ZEROES:
 979                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 980                 break;
 981         default:
 982                 if (ublk_queue_is_zoned(ubq))
 983                         return ublk_setup_iod_zoned(ubq, req);
 984                 return BLK_STS_IOERR;
 985         }
 986
 987         /* need to translate since kernel may change */
 988         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 989         iod->nr_sectors = blk_rq_sectors(req);
 990         iod->start_sector = blk_rq_pos(req);
 991         iod->addr = io->addr;
 992
 993         return BLK_STS_OK;
 994 }
 995
 996 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 997                 struct io_uring_cmd *ioucmd)
 998 {
 999         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
1000 }
1001
1002 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
1003 {
1004         return ubq->ubq_daemon->flags & PF_EXITING;
1005 }
1006
1007 /* todo: handle partial completion */
1008 static inline void __ublk_complete_rq(struct request *req)
1009 {
1010         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1011         struct ublk_io *io = &ubq->ios[req->tag];
1012         unsigned int unmapped_bytes;
1013         blk_status_t res = BLK_STS_OK;
1014
1015         /* called from ublk_abort_queue() code path */
1016         if (io->flags & UBLK_IO_FLAG_ABORTED) {
1017                 res = BLK_STS_IOERR;
1018                 goto exit;
1019         }
1020
1021         /* failed read IO if nothing is read */
1022         if (!io->res && req_op(req) == REQ_OP_READ)
1023                 io->res = -EIO;
1024
1025         if (io->res < 0) {
1026                 res = errno_to_blk_status(io->res);
1027                 goto exit;
1028         }
1029
1030         /*
1031          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
1032          * directly.
1033          *
1034          * Both the two needn't unmap.
1035          */
1036         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE &&
1037             req_op(req) != REQ_OP_DRV_IN)
1038                 goto exit;
1039
1040         /* for READ request, writing data in iod->addr to rq buffers */
1041         unmapped_bytes = ublk_unmap_io(ubq, req, io);
1042
1043         /*
1044          * Extremely impossible since we got data filled in just before
1045          *
1046          * Re-read simply for this unlikely case.
1047          */
1048         if (unlikely(unmapped_bytes < io->res))
1049                 io->res = unmapped_bytes;
1050
1051         if (blk_update_request(req, BLK_STS_OK, io->res))
1052                 blk_mq_requeue_request(req, true);
1053         else
1054                 __blk_mq_end_request(req, BLK_STS_OK);
1055
1056         return;
1057 exit:
1058         blk_mq_end_request(req, res);
1059 }
1060
1061 static void ublk_complete_rq(struct kref *ref)
1062 {
1063         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
1064                         ref);
1065         struct request *req = blk_mq_rq_from_pdu(data);
1066
1067         __ublk_complete_rq(req);
1068 }
1069
1070 /*
1071  * Since __ublk_rq_task_work always fails requests immediately during
1072  * exiting, __ublk_fail_req() is only called from abort context during
1073  * exiting. So lock is unnecessary.
1074  *
1075  * Also aborting may not be started yet, keep in mind that one failed
1076  * request may be issued by block layer again.
1077  */
1078 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
1079                 struct request *req)
1080 {
1081         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
1082
1083         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
1084                 io->flags |= UBLK_IO_FLAG_ABORTED;
1085                 if (ublk_queue_can_use_recovery_reissue(ubq))
1086                         blk_mq_requeue_request(req, false);
1087                 else
1088                         ublk_put_req_ref(ubq, req);
1089         }
1090 }
1091
1092 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
1093                                 unsigned issue_flags)
1094 {
1095         /* mark this cmd owned by ublksrv */
1096         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
1097
1098         /*
1099          * clear ACTIVE since we are done with this sqe/cmd slot
1100          * We can only accept io cmd in case of being not active.
1101          */
1102         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
1103
1104         /* tell ublksrv one io request is coming */
1105         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
1106 }
1107
1108 #define UBLK_REQUEUE_DELAY_MS   3
1109
1110 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
1111                 struct request *rq)
1112 {
1113         /* We cannot process this rq so just requeue it. */
1114         if (ublk_queue_can_use_recovery(ubq))
1115                 blk_mq_requeue_request(rq, false);
1116         else
1117                 blk_mq_end_request(rq, BLK_STS_IOERR);
1118
1119         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
1120 }
1121
1122 static inline void __ublk_rq_task_work(struct request *req,
1123                                        unsigned issue_flags)
1124 {
1125         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1126         int tag = req->tag;
1127         struct ublk_io *io = &ubq->ios[tag];
1128         unsigned int mapped_bytes;
1129
1130         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
1131                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1132                         ublk_get_iod(ubq, req->tag)->addr);
1133
1134         /*
1135          * Task is exiting if either:
1136          *
1137          * (1) current != ubq_daemon.
1138          * io_uring_cmd_complete_in_task() tries to run task_work
1139          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
1140          *
1141          * (2) current->flags & PF_EXITING.
1142          */
1143         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
1144                 __ublk_abort_rq(ubq, req);
1145                 return;
1146         }
1147
1148         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
1149                 /*
1150                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
1151                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
1152                  * and notify it.
1153                  */
1154                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
1155                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
1156                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
1157                                         __func__, io->cmd->cmd_op, ubq->q_id,
1158                                         req->tag, io->flags);
1159                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
1160                         return;
1161                 }
1162                 /*
1163                  * We have handled UBLK_IO_NEED_GET_DATA command,
1164                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
1165                  * do the copy work.
1166                  */
1167                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
1168                 /* update iod->addr because ublksrv may have passed a new io buffer */
1169                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
1170                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
1171                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1172                                 ublk_get_iod(ubq, req->tag)->addr);
1173         }
1174
1175         mapped_bytes = ublk_map_io(ubq, req, io);
1176
1177         /* partially mapped, update io descriptor */
1178         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
1179                 /*
1180                  * Nothing mapped, retry until we succeed.
1181                  *
1182                  * We may never succeed in mapping any bytes here because
1183                  * of OOM. TODO: reserve one buffer with single page pinned
1184                  * for providing forward progress guarantee.
1185                  */
1186                 if (unlikely(!mapped_bytes)) {
1187                         blk_mq_requeue_request(req, false);
1188                         blk_mq_delay_kick_requeue_list(req->q,
1189                                         UBLK_REQUEUE_DELAY_MS);
1190                         return;
1191                 }
1192
1193                 ublk_get_iod(ubq, req->tag)->nr_sectors =
1194                         mapped_bytes >> 9;
1195         }
1196
1197         ublk_init_req_ref(ubq, req);
1198         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
1199 }
1200
1201 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
1202                                         unsigned issue_flags)
1203 {
1204         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1205         struct ublk_rq_data *data, *tmp;
1206
1207         io_cmds = llist_reverse_order(io_cmds);
1208         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1209                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
1210 }
1211
1212 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
1213 {
1214         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1215         struct ublk_rq_data *data, *tmp;
1216
1217         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1218                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
1219 }
1220
1221 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1222 {
1223         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1224         struct ublk_queue *ubq = pdu->ubq;
1225
1226         ublk_forward_io_cmds(ubq, issue_flags);
1227 }
1228
1229 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
1230 {
1231         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
1232         struct ublk_io *io;
1233
1234         if (!llist_add(&data->node, &ubq->io_cmds))
1235                 return;
1236
1237         io = &ubq->ios[rq->tag];
1238         /*
1239          * If the check pass, we know that this is a re-issued request aborted
1240          * previously in monitor_work because the ubq_daemon(cmd's task) is
1241          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
1242          * because this ioucmd's io_uring context may be freed now if no inflight
1243          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
1244          *
1245          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
1246          * the tag). Then the request is re-started(allocating the tag) and we are here.
1247          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
1248          * guarantees that here is a re-issued request aborted previously.
1249          */
1250         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
1251                 ublk_abort_io_cmds(ubq);
1252         } else {
1253                 struct io_uring_cmd *cmd = io->cmd;
1254                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1255
1256                 pdu->ubq = ubq;
1257                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
1258         }
1259 }
1260
1261 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
1262 {
1263         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
1264
1265         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
1266                 if (!ubq->timeout) {
1267                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
1268                         ubq->timeout = true;
1269                 }
1270
1271                 return BLK_EH_DONE;
1272         }
1273
1274         return BLK_EH_RESET_TIMER;
1275 }
1276
1277 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
1278                 const struct blk_mq_queue_data *bd)
1279 {
1280         struct ublk_queue *ubq = hctx->driver_data;
1281         struct request *rq = bd->rq;
1282         blk_status_t res;
1283
1284         /* fill iod to slot in io cmd buffer */
1285         res = ublk_setup_iod(ubq, rq);
1286         if (unlikely(res != BLK_STS_OK))
1287                 return BLK_STS_IOERR;
1288
1289         /* With recovery feature enabled, force_abort is set in
1290          * ublk_stop_dev() before calling del_gendisk(). We have to
1291          * abort all requeued and new rqs here to let del_gendisk()
1292          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1293          * to avoid UAF on io_uring ctx.
1294          *
1295          * Note: force_abort is guaranteed to be seen because it is set
1296          * before request queue is unqiuesced.
1297          */
1298         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1299                 return BLK_STS_IOERR;
1300
1301         blk_mq_start_request(bd->rq);
1302
1303         if (unlikely(ubq_daemon_is_dying(ubq))) {
1304                 __ublk_abort_rq(ubq, rq);
1305                 return BLK_STS_OK;
1306         }
1307
1308         ublk_queue_cmd(ubq, rq);
1309
1310         return BLK_STS_OK;
1311 }
1312
1313 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1314                 unsigned int hctx_idx)
1315 {
1316         struct ublk_device *ub = driver_data;
1317         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1318
1319         hctx->driver_data = ubq;
1320         return 0;
1321 }
1322
1323 static const struct blk_mq_ops ublk_mq_ops = {
1324         .queue_rq       = ublk_queue_rq,
1325         .init_hctx      = ublk_init_hctx,
1326         .timeout        = ublk_timeout,
1327 };
1328
1329 static int ublk_ch_open(struct inode *inode, struct file *filp)
1330 {
1331         struct ublk_device *ub = container_of(inode->i_cdev,
1332                         struct ublk_device, cdev);
1333
1334         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1335                 return -EBUSY;
1336         filp->private_data = ub;
1337         return 0;
1338 }
1339
1340 static int ublk_ch_release(struct inode *inode, struct file *filp)
1341 {
1342         struct ublk_device *ub = filp->private_data;
1343
1344         clear_bit(UB_STATE_OPEN, &ub->state);
1345         return 0;
1346 }
1347
1348 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1349 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1350 {
1351         struct ublk_device *ub = filp->private_data;
1352         size_t sz = vma->vm_end - vma->vm_start;
1353         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1354         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1355         int q_id, ret = 0;
1356
1357         spin_lock(&ub->mm_lock);
1358         if (!ub->mm)
1359                 ub->mm = current->mm;
1360         if (current->mm != ub->mm)
1361                 ret = -EINVAL;
1362         spin_unlock(&ub->mm_lock);
1363
1364         if (ret)
1365                 return ret;
1366
1367         if (vma->vm_flags & VM_WRITE)
1368                 return -EPERM;
1369
1370         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1371         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1372                 return -EINVAL;
1373
1374         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1375         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1376                         __func__, q_id, current->pid, vma->vm_start,
1377                         phys_off, (unsigned long)sz);
1378
1379         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1380                 return -EINVAL;
1381
1382         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1383         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1384 }
1385
1386 static void ublk_commit_completion(struct ublk_device *ub,
1387                 const struct ublksrv_io_cmd *ub_cmd)
1388 {
1389         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1390         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1391         struct ublk_io *io = &ubq->ios[tag];
1392         struct request *req;
1393
1394         /* now this cmd slot is owned by nbd driver */
1395         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1396         io->res = ub_cmd->result;
1397
1398         /* find the io request and complete */
1399         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1400         if (WARN_ON_ONCE(unlikely(!req)))
1401                 return;
1402
1403         if (req_op(req) == REQ_OP_ZONE_APPEND)
1404                 req->__sector = ub_cmd->zone_append_lba;
1405
1406         if (likely(!blk_should_fake_timeout(req->q)))
1407                 ublk_put_req_ref(ubq, req);
1408 }
1409
1410 /*
1411  * When ->ubq_daemon is exiting, either new request is ended immediately,
1412  * or any queued io command is drained, so it is safe to abort queue
1413  * lockless
1414  */
1415 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1416 {
1417         int i;
1418
1419         if (!ublk_get_device(ub))
1420                 return;
1421
1422         for (i = 0; i < ubq->q_depth; i++) {
1423                 struct ublk_io *io = &ubq->ios[i];
1424
1425                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1426                         struct request *rq;
1427
1428                         /*
1429                          * Either we fail the request or ublk_rq_task_work_fn
1430                          * will do it
1431                          */
1432                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1433                         if (rq)
1434                                 __ublk_fail_req(ubq, io, rq);
1435                 }
1436         }
1437         ublk_put_device(ub);
1438 }
1439
1440 static void ublk_daemon_monitor_work(struct work_struct *work)
1441 {
1442         struct ublk_device *ub =
1443                 container_of(work, struct ublk_device, monitor_work.work);
1444         int i;
1445
1446         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1447                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1448
1449                 if (ubq_daemon_is_dying(ubq)) {
1450                         if (ublk_queue_can_use_recovery(ubq))
1451                                 schedule_work(&ub->quiesce_work);
1452                         else
1453                                 schedule_work(&ub->stop_work);
1454
1455                         /* abort queue is for making forward progress */
1456                         ublk_abort_queue(ub, ubq);
1457                 }
1458         }
1459
1460         /*
1461          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1462          * after ublk_remove() or __ublk_quiesce_dev() is started.
1463          *
1464          * No need ub->mutex, monitor work are canceled after state is marked
1465          * as not LIVE, so new state is observed reliably.
1466          */
1467         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1468                 schedule_delayed_work(&ub->monitor_work,
1469                                 UBLK_DAEMON_MONITOR_PERIOD);
1470 }
1471
1472 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1473 {
1474         return ubq->nr_io_ready == ubq->q_depth;
1475 }
1476
1477 static void ublk_cmd_cancel_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1478 {
1479         io_uring_cmd_done(cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1480 }
1481
1482 static void ublk_cancel_queue(struct ublk_queue *ubq)
1483 {
1484         int i;
1485
1486         if (!ublk_queue_ready(ubq))
1487                 return;
1488
1489         for (i = 0; i < ubq->q_depth; i++) {
1490                 struct ublk_io *io = &ubq->ios[i];
1491
1492                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1493                         io_uring_cmd_complete_in_task(io->cmd,
1494                                                       ublk_cmd_cancel_cb);
1495         }
1496
1497         /* all io commands are canceled */
1498         ubq->nr_io_ready = 0;
1499 }
1500
1501 /* Cancel all pending commands, must be called after del_gendisk() returns */
1502 static void ublk_cancel_dev(struct ublk_device *ub)
1503 {
1504         int i;
1505
1506         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1507                 ublk_cancel_queue(ublk_get_queue(ub, i));
1508 }
1509
1510 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1511 {
1512         bool *idle = data;
1513
1514         if (blk_mq_request_started(rq)) {
1515                 *idle = false;
1516                 return false;
1517         }
1518         return true;
1519 }
1520
1521 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1522 {
1523         bool idle;
1524
1525         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1526         while (true) {
1527                 idle = true;
1528                 blk_mq_tagset_busy_iter(&ub->tag_set,
1529                                 ublk_check_inflight_rq, &idle);
1530                 if (idle)
1531                         break;
1532                 msleep(UBLK_REQUEUE_DELAY_MS);
1533         }
1534 }
1535
1536 static void __ublk_quiesce_dev(struct ublk_device *ub)
1537 {
1538         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1539                         __func__, ub->dev_info.dev_id,
1540                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1541                         "LIVE" : "QUIESCED");
1542         blk_mq_quiesce_queue(ub->ub_disk->queue);
1543         ublk_wait_tagset_rqs_idle(ub);
1544         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1545         ublk_cancel_dev(ub);
1546         /* we are going to release task_struct of ubq_daemon and resets
1547          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1548          * Besides, monitor_work is not necessary in QUIESCED state since we have
1549          * already scheduled quiesce_work and quiesced all ubqs.
1550          *
1551          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1552          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1553          */
1554         cancel_delayed_work_sync(&ub->monitor_work);
1555 }
1556
1557 static void ublk_quiesce_work_fn(struct work_struct *work)
1558 {
1559         struct ublk_device *ub =
1560                 container_of(work, struct ublk_device, quiesce_work);
1561
1562         mutex_lock(&ub->mutex);
1563         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1564                 goto unlock;
1565         __ublk_quiesce_dev(ub);
1566  unlock:
1567         mutex_unlock(&ub->mutex);
1568 }
1569
1570 static void ublk_unquiesce_dev(struct ublk_device *ub)
1571 {
1572         int i;
1573
1574         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1575                         __func__, ub->dev_info.dev_id,
1576                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1577                         "LIVE" : "QUIESCED");
1578         /* quiesce_work has run. We let requeued rqs be aborted
1579          * before running fallback_wq. "force_abort" must be seen
1580          * after request queue is unqiuesced. Then del_gendisk()
1581          * can move on.
1582          */
1583         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1584                 ublk_get_queue(ub, i)->force_abort = true;
1585
1586         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1587         /* We may have requeued some rqs in ublk_quiesce_queue() */
1588         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1589 }
1590
1591 static void ublk_stop_dev(struct ublk_device *ub)
1592 {
1593         mutex_lock(&ub->mutex);
1594         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1595                 goto unlock;
1596         if (ublk_can_use_recovery(ub)) {
1597                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1598                         __ublk_quiesce_dev(ub);
1599                 ublk_unquiesce_dev(ub);
1600         }
1601         del_gendisk(ub->ub_disk);
1602         ub->dev_info.state = UBLK_S_DEV_DEAD;
1603         ub->dev_info.ublksrv_pid = -1;
1604         put_disk(ub->ub_disk);
1605         ub->ub_disk = NULL;
1606  unlock:
1607         ublk_cancel_dev(ub);
1608         mutex_unlock(&ub->mutex);
1609         cancel_delayed_work_sync(&ub->monitor_work);
1610 }
1611
1612 /* device can only be started after all IOs are ready */
1613 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1614 {
1615         mutex_lock(&ub->mutex);
1616         ubq->nr_io_ready++;
1617         if (ublk_queue_ready(ubq)) {
1618                 ubq->ubq_daemon = current;
1619                 get_task_struct(ubq->ubq_daemon);
1620                 ub->nr_queues_ready++;
1621
1622                 if (capable(CAP_SYS_ADMIN))
1623                         ub->nr_privileged_daemon++;
1624         }
1625         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1626                 complete_all(&ub->completion);
1627         mutex_unlock(&ub->mutex);
1628 }
1629
1630 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1631                 int tag)
1632 {
1633         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1634         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1635
1636         ublk_queue_cmd(ubq, req);
1637 }
1638
1639 static inline int ublk_check_cmd_op(u32 cmd_op)
1640 {
1641         u32 ioc_type = _IOC_TYPE(cmd_op);
1642
1643         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1644                 return -EOPNOTSUPP;
1645
1646         if (ioc_type != 'u' && ioc_type != 0)
1647                 return -EOPNOTSUPP;
1648
1649         return 0;
1650 }
1651
1652 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1653                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1654 {
1655         io->cmd = cmd;
1656         io->flags |= UBLK_IO_FLAG_ACTIVE;
1657         io->addr = buf_addr;
1658 }
1659
1660 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1661                                unsigned int issue_flags,
1662                                const struct ublksrv_io_cmd *ub_cmd)
1663 {
1664         struct ublk_device *ub = cmd->file->private_data;
1665         struct ublk_queue *ubq;
1666         struct ublk_io *io;
1667         u32 cmd_op = cmd->cmd_op;
1668         unsigned tag = ub_cmd->tag;
1669         int ret = -EINVAL;
1670         struct request *req;
1671
1672         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1673                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1674                         ub_cmd->result);
1675
1676         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1677                 goto out;
1678
1679         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1680         if (!ubq || ub_cmd->q_id != ubq->q_id)
1681                 goto out;
1682
1683         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1684                 goto out;
1685
1686         if (tag >= ubq->q_depth)
1687                 goto out;
1688
1689         io = &ubq->ios[tag];
1690
1691         /* there is pending io cmd, something must be wrong */
1692         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1693                 ret = -EBUSY;
1694                 goto out;
1695         }
1696
1697         /*
1698          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1699          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1700          */
1701         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1702                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1703                 goto out;
1704
1705         ret = ublk_check_cmd_op(cmd_op);
1706         if (ret)
1707                 goto out;
1708
1709         ret = -EINVAL;
1710         switch (_IOC_NR(cmd_op)) {
1711         case UBLK_IO_FETCH_REQ:
1712                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1713                 if (ublk_queue_ready(ubq)) {
1714                         ret = -EBUSY;
1715                         goto out;
1716                 }
1717                 /*
1718                  * The io is being handled by server, so COMMIT_RQ is expected
1719                  * instead of FETCH_REQ
1720                  */
1721                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1722                         goto out;
1723
1724                 if (!ublk_support_user_copy(ubq)) {
1725                         /*
1726                          * FETCH_RQ has to provide IO buffer if NEED GET
1727                          * DATA is not enabled
1728                          */
1729                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1730                                 goto out;
1731                 } else if (ub_cmd->addr) {
1732                         /* User copy requires addr to be unset */
1733                         ret = -EINVAL;
1734                         goto out;
1735                 }
1736
1737                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1738                 ublk_mark_io_ready(ub, ubq);
1739                 break;
1740         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1741                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1742
1743                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1744                         goto out;
1745
1746                 if (!ublk_support_user_copy(ubq)) {
1747                         /*
1748                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1749                          * NEED GET DATA is not enabled or it is Read IO.
1750                          */
1751                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1752                                                 req_op(req) == REQ_OP_READ))
1753                                 goto out;
1754                 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1755                         /*
1756                          * User copy requires addr to be unset when command is
1757                          * not zone append
1758                          */
1759                         ret = -EINVAL;
1760                         goto out;
1761                 }
1762
1763                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1764                 ublk_commit_completion(ub, ub_cmd);
1765                 break;
1766         case UBLK_IO_NEED_GET_DATA:
1767                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1768                         goto out;
1769                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1770                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1771                 break;
1772         default:
1773                 goto out;
1774         }
1775         return -EIOCBQUEUED;
1776
1777  out:
1778         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1779         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1780                         __func__, cmd_op, tag, ret, io->flags);
1781         return -EIOCBQUEUED;
1782 }
1783
1784 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1785                 struct ublk_queue *ubq, int tag, size_t offset)
1786 {
1787         struct request *req;
1788
1789         if (!ublk_need_req_ref(ubq))
1790                 return NULL;
1791
1792         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1793         if (!req)
1794                 return NULL;
1795
1796         if (!ublk_get_req_ref(ubq, req))
1797                 return NULL;
1798
1799         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1800                 goto fail_put;
1801
1802         if (!ublk_rq_has_data(req))
1803                 goto fail_put;
1804
1805         if (offset > blk_rq_bytes(req))
1806                 goto fail_put;
1807
1808         return req;
1809 fail_put:
1810         ublk_put_req_ref(ubq, req);
1811         return NULL;
1812 }
1813
1814 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1815 {
1816         /*
1817          * Not necessary for async retry, but let's keep it simple and always
1818          * copy the values to avoid any potential reuse.
1819          */
1820         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1821         const struct ublksrv_io_cmd ub_cmd = {
1822                 .q_id = READ_ONCE(ub_src->q_id),
1823                 .tag = READ_ONCE(ub_src->tag),
1824                 .result = READ_ONCE(ub_src->result),
1825                 .addr = READ_ONCE(ub_src->addr)
1826         };
1827
1828         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1829 }
1830
1831 static inline bool ublk_check_ubuf_dir(const struct request *req,
1832                 int ubuf_dir)
1833 {
1834         /* copy ubuf to request pages */
1835         if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1836             ubuf_dir == ITER_SOURCE)
1837                 return true;
1838
1839         /* copy request pages to ubuf */
1840         if ((req_op(req) == REQ_OP_WRITE ||
1841              req_op(req) == REQ_OP_ZONE_APPEND) &&
1842             ubuf_dir == ITER_DEST)
1843                 return true;
1844
1845         return false;
1846 }
1847
1848 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1849                 struct iov_iter *iter, size_t *off, int dir)
1850 {
1851         struct ublk_device *ub = iocb->ki_filp->private_data;
1852         struct ublk_queue *ubq;
1853         struct request *req;
1854         size_t buf_off;
1855         u16 tag, q_id;
1856
1857         if (!ub)
1858                 return ERR_PTR(-EACCES);
1859
1860         if (!user_backed_iter(iter))
1861                 return ERR_PTR(-EACCES);
1862
1863         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1864                 return ERR_PTR(-EACCES);
1865
1866         tag = ublk_pos_to_tag(iocb->ki_pos);
1867         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1868         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1869
1870         if (q_id >= ub->dev_info.nr_hw_queues)
1871                 return ERR_PTR(-EINVAL);
1872
1873         ubq = ublk_get_queue(ub, q_id);
1874         if (!ubq)
1875                 return ERR_PTR(-EINVAL);
1876
1877         if (tag >= ubq->q_depth)
1878                 return ERR_PTR(-EINVAL);
1879
1880         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1881         if (!req)
1882                 return ERR_PTR(-EINVAL);
1883
1884         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1885                 goto fail;
1886
1887         if (!ublk_check_ubuf_dir(req, dir))
1888                 goto fail;
1889
1890         *off = buf_off;
1891         return req;
1892 fail:
1893         ublk_put_req_ref(ubq, req);
1894         return ERR_PTR(-EACCES);
1895 }
1896
1897 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1898 {
1899         struct ublk_queue *ubq;
1900         struct request *req;
1901         size_t buf_off;
1902         size_t ret;
1903
1904         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1905         if (IS_ERR(req))
1906                 return PTR_ERR(req);
1907
1908         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1909         ubq = req->mq_hctx->driver_data;
1910         ublk_put_req_ref(ubq, req);
1911
1912         return ret;
1913 }
1914
1915 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1916 {
1917         struct ublk_queue *ubq;
1918         struct request *req;
1919         size_t buf_off;
1920         size_t ret;
1921
1922         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1923         if (IS_ERR(req))
1924                 return PTR_ERR(req);
1925
1926         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1927         ubq = req->mq_hctx->driver_data;
1928         ublk_put_req_ref(ubq, req);
1929
1930         return ret;
1931 }
1932
1933 static const struct file_operations ublk_ch_fops = {
1934         .owner = THIS_MODULE,
1935         .open = ublk_ch_open,
1936         .release = ublk_ch_release,
1937         .llseek = no_llseek,
1938         .read_iter = ublk_ch_read_iter,
1939         .write_iter = ublk_ch_write_iter,
1940         .uring_cmd = ublk_ch_uring_cmd,
1941         .mmap = ublk_ch_mmap,
1942 };
1943
1944 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1945 {
1946         int size = ublk_queue_cmd_buf_size(ub, q_id);
1947         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1948
1949         if (ubq->ubq_daemon)
1950                 put_task_struct(ubq->ubq_daemon);
1951         if (ubq->io_cmd_buf)
1952                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1953 }
1954
1955 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1956 {
1957         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1958         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1959         void *ptr;
1960         int size;
1961
1962         ubq->flags = ub->dev_info.flags;
1963         ubq->q_id = q_id;
1964         ubq->q_depth = ub->dev_info.queue_depth;
1965         size = ublk_queue_cmd_buf_size(ub, q_id);
1966
1967         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1968         if (!ptr)
1969                 return -ENOMEM;
1970
1971         ubq->io_cmd_buf = ptr;
1972         ubq->dev = ub;
1973         return 0;
1974 }
1975
1976 static void ublk_deinit_queues(struct ublk_device *ub)
1977 {
1978         int nr_queues = ub->dev_info.nr_hw_queues;
1979         int i;
1980
1981         if (!ub->__queues)
1982                 return;
1983
1984         for (i = 0; i < nr_queues; i++)
1985                 ublk_deinit_queue(ub, i);
1986         kfree(ub->__queues);
1987 }
1988
1989 static int ublk_init_queues(struct ublk_device *ub)
1990 {
1991         int nr_queues = ub->dev_info.nr_hw_queues;
1992         int depth = ub->dev_info.queue_depth;
1993         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1994         int i, ret = -ENOMEM;
1995
1996         ub->queue_size = ubq_size;
1997         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1998         if (!ub->__queues)
1999                 return ret;
2000
2001         for (i = 0; i < nr_queues; i++) {
2002                 if (ublk_init_queue(ub, i))
2003                         goto fail;
2004         }
2005
2006         init_completion(&ub->completion);
2007         return 0;
2008
2009  fail:
2010         ublk_deinit_queues(ub);
2011         return ret;
2012 }
2013
2014 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2015 {
2016         int i = idx;
2017         int err;
2018
2019         spin_lock(&ublk_idr_lock);
2020         /* allocate id, if @id >= 0, we're requesting that specific id */
2021         if (i >= 0) {
2022                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2023                 if (err == -ENOSPC)
2024                         err = -EEXIST;
2025         } else {
2026                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
2027         }
2028         spin_unlock(&ublk_idr_lock);
2029
2030         if (err >= 0)
2031                 ub->ub_number = err;
2032
2033         return err;
2034 }
2035
2036 static void ublk_free_dev_number(struct ublk_device *ub)
2037 {
2038         spin_lock(&ublk_idr_lock);
2039         idr_remove(&ublk_index_idr, ub->ub_number);
2040         wake_up_all(&ublk_idr_wq);
2041         spin_unlock(&ublk_idr_lock);
2042 }
2043
2044 static void ublk_cdev_rel(struct device *dev)
2045 {
2046         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2047
2048         blk_mq_free_tag_set(&ub->tag_set);
2049         ublk_deinit_queues(ub);
2050         ublk_free_dev_number(ub);
2051         mutex_destroy(&ub->mutex);
2052         kfree(ub);
2053 }
2054
2055 static int ublk_add_chdev(struct ublk_device *ub)
2056 {
2057         struct device *dev = &ub->cdev_dev;
2058         int minor = ub->ub_number;
2059         int ret;
2060
2061         dev->parent = ublk_misc.this_device;
2062         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2063         dev->class = &ublk_chr_class;
2064         dev->release = ublk_cdev_rel;
2065         device_initialize(dev);
2066
2067         ret = dev_set_name(dev, "ublkc%d", minor);
2068         if (ret)
2069                 goto fail;
2070
2071         cdev_init(&ub->cdev, &ublk_ch_fops);
2072         ret = cdev_device_add(&ub->cdev, dev);
2073         if (ret)
2074                 goto fail;
2075
2076         ublks_added++;
2077         return 0;
2078  fail:
2079         put_device(dev);
2080         return ret;
2081 }
2082
2083 static void ublk_stop_work_fn(struct work_struct *work)
2084 {
2085         struct ublk_device *ub =
2086                 container_of(work, struct ublk_device, stop_work);
2087
2088         ublk_stop_dev(ub);
2089 }
2090
2091 /* align max io buffer size with PAGE_SIZE */
2092 static void ublk_align_max_io_size(struct ublk_device *ub)
2093 {
2094         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2095
2096         ub->dev_info.max_io_buf_bytes =
2097                 round_down(max_io_bytes, PAGE_SIZE);
2098 }
2099
2100 static int ublk_add_tag_set(struct ublk_device *ub)
2101 {
2102         ub->tag_set.ops = &ublk_mq_ops;
2103         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2104         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2105         ub->tag_set.numa_node = NUMA_NO_NODE;
2106         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2107         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2108         ub->tag_set.driver_data = ub;
2109         return blk_mq_alloc_tag_set(&ub->tag_set);
2110 }
2111
2112 static void ublk_remove(struct ublk_device *ub)
2113 {
2114         ublk_stop_dev(ub);
2115         cancel_work_sync(&ub->stop_work);
2116         cancel_work_sync(&ub->quiesce_work);
2117         cdev_device_del(&ub->cdev, &ub->cdev_dev);
2118         put_device(&ub->cdev_dev);
2119         ublks_added--;
2120 }
2121
2122 static struct ublk_device *ublk_get_device_from_id(int idx)
2123 {
2124         struct ublk_device *ub = NULL;
2125
2126         if (idx < 0)
2127                 return NULL;
2128
2129         spin_lock(&ublk_idr_lock);
2130         ub = idr_find(&ublk_index_idr, idx);
2131         if (ub)
2132                 ub = ublk_get_device(ub);
2133         spin_unlock(&ublk_idr_lock);
2134
2135         return ub;
2136 }
2137
2138 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2139 {
2140         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2141         int ublksrv_pid = (int)header->data[0];
2142         struct gendisk *disk;
2143         int ret = -EINVAL;
2144
2145         if (ublksrv_pid <= 0)
2146                 return -EINVAL;
2147
2148         wait_for_completion_interruptible(&ub->completion);
2149
2150         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2151
2152         mutex_lock(&ub->mutex);
2153         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2154             test_bit(UB_STATE_USED, &ub->state)) {
2155                 ret = -EEXIST;
2156                 goto out_unlock;
2157         }
2158
2159         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
2160         if (IS_ERR(disk)) {
2161                 ret = PTR_ERR(disk);
2162                 goto out_unlock;
2163         }
2164         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2165         disk->fops = &ub_fops;
2166         disk->private_data = ub;
2167
2168         ub->dev_info.ublksrv_pid = ublksrv_pid;
2169         ub->ub_disk = disk;
2170
2171         ret = ublk_apply_params(ub);
2172         if (ret)
2173                 goto out_put_disk;
2174
2175         /* don't probe partitions if any one ubq daemon is un-trusted */
2176         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2177                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2178
2179         get_device(&ub->cdev_dev);
2180         ub->dev_info.state = UBLK_S_DEV_LIVE;
2181
2182         if (ublk_dev_is_zoned(ub)) {
2183                 ret = ublk_revalidate_disk_zones(ub);
2184                 if (ret)
2185                         goto out_put_cdev;
2186         }
2187
2188         ret = add_disk(disk);
2189         if (ret)
2190                 goto out_put_cdev;
2191
2192         set_bit(UB_STATE_USED, &ub->state);
2193
2194 out_put_cdev:
2195         if (ret) {
2196                 ub->dev_info.state = UBLK_S_DEV_DEAD;
2197                 ublk_put_device(ub);
2198         }
2199 out_put_disk:
2200         if (ret)
2201                 put_disk(disk);
2202 out_unlock:
2203         mutex_unlock(&ub->mutex);
2204         return ret;
2205 }
2206
2207 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2208                 struct io_uring_cmd *cmd)
2209 {
2210         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2211         void __user *argp = (void __user *)(unsigned long)header->addr;
2212         cpumask_var_t cpumask;
2213         unsigned long queue;
2214         unsigned int retlen;
2215         unsigned int i;
2216         int ret;
2217
2218         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2219                 return -EINVAL;
2220         if (header->len & (sizeof(unsigned long)-1))
2221                 return -EINVAL;
2222         if (!header->addr)
2223                 return -EINVAL;
2224
2225         queue = header->data[0];
2226         if (queue >= ub->dev_info.nr_hw_queues)
2227                 return -EINVAL;
2228
2229         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2230                 return -ENOMEM;
2231
2232         for_each_possible_cpu(i) {
2233                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2234                         cpumask_set_cpu(i, cpumask);
2235         }
2236
2237         ret = -EFAULT;
2238         retlen = min_t(unsigned short, header->len, cpumask_size());
2239         if (copy_to_user(argp, cpumask, retlen))
2240                 goto out_free_cpumask;
2241         if (retlen != header->len &&
2242             clear_user(argp + retlen, header->len - retlen))
2243                 goto out_free_cpumask;
2244
2245         ret = 0;
2246 out_free_cpumask:
2247         free_cpumask_var(cpumask);
2248         return ret;
2249 }
2250
2251 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2252 {
2253         pr_devel("%s: dev id %d flags %llx\n", __func__,
2254                         info->dev_id, info->flags);
2255         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2256                         info->nr_hw_queues, info->queue_depth);
2257 }
2258
2259 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2260 {
2261         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2262         void __user *argp = (void __user *)(unsigned long)header->addr;
2263         struct ublksrv_ctrl_dev_info info;
2264         struct ublk_device *ub;
2265         int ret = -EINVAL;
2266
2267         if (header->len < sizeof(info) || !header->addr)
2268                 return -EINVAL;
2269         if (header->queue_id != (u16)-1) {
2270                 pr_warn("%s: queue_id is wrong %x\n",
2271                         __func__, header->queue_id);
2272                 return -EINVAL;
2273         }
2274
2275         if (copy_from_user(&info, argp, sizeof(info)))
2276                 return -EFAULT;
2277
2278         if (capable(CAP_SYS_ADMIN))
2279                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2280         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2281                 return -EPERM;
2282
2283         /*
2284          * unprivileged device can't be trusted, but RECOVERY and
2285          * RECOVERY_REISSUE still may hang error handling, so can't
2286          * support recovery features for unprivileged ublk now
2287          *
2288          * TODO: provide forward progress for RECOVERY handler, so that
2289          * unprivileged device can benefit from it
2290          */
2291         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2292                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2293                                 UBLK_F_USER_RECOVERY);
2294
2295         /* the created device is always owned by current user */
2296         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2297
2298         if (header->dev_id != info.dev_id) {
2299                 pr_warn("%s: dev id not match %u %u\n",
2300                         __func__, header->dev_id, info.dev_id);
2301                 return -EINVAL;
2302         }
2303
2304         ublk_dump_dev_info(&info);
2305
2306         ret = mutex_lock_killable(&ublk_ctl_mutex);
2307         if (ret)
2308                 return ret;
2309
2310         ret = -EACCES;
2311         if (ublks_added >= ublks_max)
2312                 goto out_unlock;
2313
2314         ret = -ENOMEM;
2315         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2316         if (!ub)
2317                 goto out_unlock;
2318         mutex_init(&ub->mutex);
2319         spin_lock_init(&ub->mm_lock);
2320         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2321         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2322         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
2323
2324         ret = ublk_alloc_dev_number(ub, header->dev_id);
2325         if (ret < 0)
2326                 goto out_free_ub;
2327
2328         memcpy(&ub->dev_info, &info, sizeof(info));
2329
2330         /* update device id */
2331         ub->dev_info.dev_id = ub->ub_number;
2332
2333         /*
2334          * 64bit flags will be copied back to userspace as feature
2335          * negotiation result, so have to clear flags which driver
2336          * doesn't support yet, then userspace can get correct flags
2337          * (features) to handle.
2338          */
2339         ub->dev_info.flags &= UBLK_F_ALL;
2340
2341         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2342                 UBLK_F_URING_CMD_COMP_IN_TASK;
2343
2344         /* GET_DATA isn't needed any more with USER_COPY */
2345         if (ublk_dev_is_user_copy(ub))
2346                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2347
2348         /* Zoned storage support requires user copy feature */
2349         if (ublk_dev_is_zoned(ub) &&
2350             (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2351                 ret = -EINVAL;
2352                 goto out_free_dev_number;
2353         }
2354
2355         /* We are not ready to support zero copy */
2356         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2357
2358         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2359                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2360         ublk_align_max_io_size(ub);
2361
2362         ret = ublk_init_queues(ub);
2363         if (ret)
2364                 goto out_free_dev_number;
2365
2366         ret = ublk_add_tag_set(ub);
2367         if (ret)
2368                 goto out_deinit_queues;
2369
2370         ret = -EFAULT;
2371         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2372                 goto out_free_tag_set;
2373
2374         /*
2375          * Add the char dev so that ublksrv daemon can be setup.
2376          * ublk_add_chdev() will cleanup everything if it fails.
2377          */
2378         ret = ublk_add_chdev(ub);
2379         goto out_unlock;
2380
2381 out_free_tag_set:
2382         blk_mq_free_tag_set(&ub->tag_set);
2383 out_deinit_queues:
2384         ublk_deinit_queues(ub);
2385 out_free_dev_number:
2386         ublk_free_dev_number(ub);
2387 out_free_ub:
2388         mutex_destroy(&ub->mutex);
2389         kfree(ub);
2390 out_unlock:
2391         mutex_unlock(&ublk_ctl_mutex);
2392         return ret;
2393 }
2394
2395 static inline bool ublk_idr_freed(int id)
2396 {
2397         void *ptr;
2398
2399         spin_lock(&ublk_idr_lock);
2400         ptr = idr_find(&ublk_index_idr, id);
2401         spin_unlock(&ublk_idr_lock);
2402
2403         return ptr == NULL;
2404 }
2405
2406 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2407 {
2408         struct ublk_device *ub = *p_ub;
2409         int idx = ub->ub_number;
2410         int ret;
2411
2412         ret = mutex_lock_killable(&ublk_ctl_mutex);
2413         if (ret)
2414                 return ret;
2415
2416         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2417                 ublk_remove(ub);
2418                 set_bit(UB_STATE_DELETED, &ub->state);
2419         }
2420
2421         /* Mark the reference as consumed */
2422         *p_ub = NULL;
2423         ublk_put_device(ub);
2424         mutex_unlock(&ublk_ctl_mutex);
2425
2426         /*
2427          * Wait until the idr is removed, then it can be reused after
2428          * DEL_DEV command is returned.
2429          *
2430          * If we returns because of user interrupt, future delete command
2431          * may come:
2432          *
2433          * - the device number isn't freed, this device won't or needn't
2434          *   be deleted again, since UB_STATE_DELETED is set, and device
2435          *   will be released after the last reference is dropped
2436          *
2437          * - the device number is freed already, we will not find this
2438          *   device via ublk_get_device_from_id()
2439          */
2440         wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx));
2441
2442         return 0;
2443 }
2444
2445 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2446 {
2447         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2448
2449         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2450                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2451                         header->data[0], header->addr, header->len);
2452 }
2453
2454 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2455 {
2456         ublk_stop_dev(ub);
2457         cancel_work_sync(&ub->stop_work);
2458         cancel_work_sync(&ub->quiesce_work);
2459
2460         return 0;
2461 }
2462
2463 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2464                 struct io_uring_cmd *cmd)
2465 {
2466         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2467         void __user *argp = (void __user *)(unsigned long)header->addr;
2468
2469         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2470                 return -EINVAL;
2471
2472         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2473                 return -EFAULT;
2474
2475         return 0;
2476 }
2477
2478 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2479 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2480 {
2481         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2482         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2483
2484         if (ub->ub_disk) {
2485                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2486                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2487         } else {
2488                 ub->params.devt.disk_major = 0;
2489                 ub->params.devt.disk_minor = 0;
2490         }
2491         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2492 }
2493
2494 static int ublk_ctrl_get_params(struct ublk_device *ub,
2495                 struct io_uring_cmd *cmd)
2496 {
2497         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2498         void __user *argp = (void __user *)(unsigned long)header->addr;
2499         struct ublk_params_header ph;
2500         int ret;
2501
2502         if (header->len <= sizeof(ph) || !header->addr)
2503                 return -EINVAL;
2504
2505         if (copy_from_user(&ph, argp, sizeof(ph)))
2506                 return -EFAULT;
2507
2508         if (ph.len > header->len || !ph.len)
2509                 return -EINVAL;
2510
2511         if (ph.len > sizeof(struct ublk_params))
2512                 ph.len = sizeof(struct ublk_params);
2513
2514         mutex_lock(&ub->mutex);
2515         ublk_ctrl_fill_params_devt(ub);
2516         if (copy_to_user(argp, &ub->params, ph.len))
2517                 ret = -EFAULT;
2518         else
2519                 ret = 0;
2520         mutex_unlock(&ub->mutex);
2521
2522         return ret;
2523 }
2524
2525 static int ublk_ctrl_set_params(struct ublk_device *ub,
2526                 struct io_uring_cmd *cmd)
2527 {
2528         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2529         void __user *argp = (void __user *)(unsigned long)header->addr;
2530         struct ublk_params_header ph;
2531         int ret = -EFAULT;
2532
2533         if (header->len <= sizeof(ph) || !header->addr)
2534                 return -EINVAL;
2535
2536         if (copy_from_user(&ph, argp, sizeof(ph)))
2537                 return -EFAULT;
2538
2539         if (ph.len > header->len || !ph.len || !ph.types)
2540                 return -EINVAL;
2541
2542         if (ph.len > sizeof(struct ublk_params))
2543                 ph.len = sizeof(struct ublk_params);
2544
2545         /* parameters can only be changed when device isn't live */
2546         mutex_lock(&ub->mutex);
2547         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2548                 ret = -EACCES;
2549         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2550                 ret = -EFAULT;
2551         } else {
2552                 /* clear all we don't support yet */
2553                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2554                 ret = ublk_validate_params(ub);
2555                 if (ret)
2556                         ub->params.types = 0;
2557         }
2558         mutex_unlock(&ub->mutex);
2559
2560         return ret;
2561 }
2562
2563 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2564 {
2565         int i;
2566
2567         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2568         /* All old ioucmds have to be completed */
2569         WARN_ON_ONCE(ubq->nr_io_ready);
2570         /* old daemon is PF_EXITING, put it now */
2571         put_task_struct(ubq->ubq_daemon);
2572         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2573         ubq->ubq_daemon = NULL;
2574         ubq->timeout = false;
2575
2576         for (i = 0; i < ubq->q_depth; i++) {
2577                 struct ublk_io *io = &ubq->ios[i];
2578
2579                 /* forget everything now and be ready for new FETCH_REQ */
2580                 io->flags = 0;
2581                 io->cmd = NULL;
2582                 io->addr = 0;
2583         }
2584 }
2585
2586 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2587                 struct io_uring_cmd *cmd)
2588 {
2589         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2590         int ret = -EINVAL;
2591         int i;
2592
2593         mutex_lock(&ub->mutex);
2594         if (!ublk_can_use_recovery(ub))
2595                 goto out_unlock;
2596         /*
2597          * START_RECOVERY is only allowd after:
2598          *
2599          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2600          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2601          *     released.
2602          *
2603          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2604          *     (a)has quiesced request queue
2605          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2606          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2607          *     (d)has completed/camceled all ioucmds owned by ther dying process
2608          */
2609         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2610                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2611                 ret = -EBUSY;
2612                 goto out_unlock;
2613         }
2614         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2615         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2616                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2617         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2618         ub->mm = NULL;
2619         ub->nr_queues_ready = 0;
2620         ub->nr_privileged_daemon = 0;
2621         init_completion(&ub->completion);
2622         ret = 0;
2623  out_unlock:
2624         mutex_unlock(&ub->mutex);
2625         return ret;
2626 }
2627
2628 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2629                 struct io_uring_cmd *cmd)
2630 {
2631         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2632         int ublksrv_pid = (int)header->data[0];
2633         int ret = -EINVAL;
2634
2635         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2636                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2637         /* wait until new ubq_daemon sending all FETCH_REQ */
2638         wait_for_completion_interruptible(&ub->completion);
2639         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2640                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2641
2642         mutex_lock(&ub->mutex);
2643         if (!ublk_can_use_recovery(ub))
2644                 goto out_unlock;
2645
2646         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2647                 ret = -EBUSY;
2648                 goto out_unlock;
2649         }
2650         ub->dev_info.ublksrv_pid = ublksrv_pid;
2651         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2652                         __func__, ublksrv_pid, header->dev_id);
2653         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2654         pr_devel("%s: queue unquiesced, dev id %d.\n",
2655                         __func__, header->dev_id);
2656         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2657         ub->dev_info.state = UBLK_S_DEV_LIVE;
2658         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2659         ret = 0;
2660  out_unlock:
2661         mutex_unlock(&ub->mutex);
2662         return ret;
2663 }
2664
2665 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2666 {
2667         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2668         void __user *argp = (void __user *)(unsigned long)header->addr;
2669         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2670
2671         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2672                 return -EINVAL;
2673
2674         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2675                 return -EFAULT;
2676
2677         return 0;
2678 }
2679
2680 /*
2681  * All control commands are sent via /dev/ublk-control, so we have to check
2682  * the destination device's permission
2683  */
2684 static int ublk_char_dev_permission(struct ublk_device *ub,
2685                 const char *dev_path, int mask)
2686 {
2687         int err;
2688         struct path path;
2689         struct kstat stat;
2690
2691         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2692         if (err)
2693                 return err;
2694
2695         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2696         if (err)
2697                 goto exit;
2698
2699         err = -EPERM;
2700         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2701                 goto exit;
2702
2703         err = inode_permission(&nop_mnt_idmap,
2704                         d_backing_inode(path.dentry), mask);
2705 exit:
2706         path_put(&path);
2707         return err;
2708 }
2709
2710 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2711                 struct io_uring_cmd *cmd)
2712 {
2713         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2714         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2715         void __user *argp = (void __user *)(unsigned long)header->addr;
2716         char *dev_path = NULL;
2717         int ret = 0;
2718         int mask;
2719
2720         if (!unprivileged) {
2721                 if (!capable(CAP_SYS_ADMIN))
2722                         return -EPERM;
2723                 /*
2724                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2725                  * char_dev_path in payload too, since userspace may not
2726                  * know if the specified device is created as unprivileged
2727                  * mode.
2728                  */
2729                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2730                         return 0;
2731         }
2732
2733         /*
2734          * User has to provide the char device path for unprivileged ublk
2735          *
2736          * header->addr always points to the dev path buffer, and
2737          * header->dev_path_len records length of dev path buffer.
2738          */
2739         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2740                 return -EINVAL;
2741
2742         if (header->len < header->dev_path_len)
2743                 return -EINVAL;
2744
2745         dev_path = kmalloc(header->dev_path_len + 1, GFP_KERNEL);
2746         if (!dev_path)
2747                 return -ENOMEM;
2748
2749         ret = -EFAULT;
2750         if (copy_from_user(dev_path, argp, header->dev_path_len))
2751                 goto exit;
2752         dev_path[header->dev_path_len] = 0;
2753
2754         ret = -EINVAL;
2755         switch (_IOC_NR(cmd->cmd_op)) {
2756         case UBLK_CMD_GET_DEV_INFO:
2757         case UBLK_CMD_GET_DEV_INFO2:
2758         case UBLK_CMD_GET_QUEUE_AFFINITY:
2759         case UBLK_CMD_GET_PARAMS:
2760         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2761                 mask = MAY_READ;
2762                 break;
2763         case UBLK_CMD_START_DEV:
2764         case UBLK_CMD_STOP_DEV:
2765         case UBLK_CMD_ADD_DEV:
2766         case UBLK_CMD_DEL_DEV:
2767         case UBLK_CMD_SET_PARAMS:
2768         case UBLK_CMD_START_USER_RECOVERY:
2769         case UBLK_CMD_END_USER_RECOVERY:
2770                 mask = MAY_READ | MAY_WRITE;
2771                 break;
2772         default:
2773                 goto exit;
2774         }
2775
2776         ret = ublk_char_dev_permission(ub, dev_path, mask);
2777         if (!ret) {
2778                 header->len -= header->dev_path_len;
2779                 header->addr += header->dev_path_len;
2780         }
2781         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2782                         __func__, ub->ub_number, cmd->cmd_op,
2783                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2784                         dev_path, ret);
2785 exit:
2786         kfree(dev_path);
2787         return ret;
2788 }
2789
2790 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2791                 unsigned int issue_flags)
2792 {
2793         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2794         struct ublk_device *ub = NULL;
2795         u32 cmd_op = cmd->cmd_op;
2796         int ret = -EINVAL;
2797
2798         if (issue_flags & IO_URING_F_NONBLOCK)
2799                 return -EAGAIN;
2800
2801         ublk_ctrl_cmd_dump(cmd);
2802
2803         if (!(issue_flags & IO_URING_F_SQE128))
2804                 goto out;
2805
2806         ret = ublk_check_cmd_op(cmd_op);
2807         if (ret)
2808                 goto out;
2809
2810         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2811                 ret = ublk_ctrl_get_features(cmd);
2812                 goto out;
2813         }
2814
2815         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2816                 ret = -ENODEV;
2817                 ub = ublk_get_device_from_id(header->dev_id);
2818                 if (!ub)
2819                         goto out;
2820
2821                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2822                 if (ret)
2823                         goto put_dev;
2824         }
2825
2826         switch (_IOC_NR(cmd_op)) {
2827         case UBLK_CMD_START_DEV:
2828                 ret = ublk_ctrl_start_dev(ub, cmd);
2829                 break;
2830         case UBLK_CMD_STOP_DEV:
2831                 ret = ublk_ctrl_stop_dev(ub);
2832                 break;
2833         case UBLK_CMD_GET_DEV_INFO:
2834         case UBLK_CMD_GET_DEV_INFO2:
2835                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2836                 break;
2837         case UBLK_CMD_ADD_DEV:
2838                 ret = ublk_ctrl_add_dev(cmd);
2839                 break;
2840         case UBLK_CMD_DEL_DEV:
2841                 ret = ublk_ctrl_del_dev(&ub);
2842                 break;
2843         case UBLK_CMD_GET_QUEUE_AFFINITY:
2844                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2845                 break;
2846         case UBLK_CMD_GET_PARAMS:
2847                 ret = ublk_ctrl_get_params(ub, cmd);
2848                 break;
2849         case UBLK_CMD_SET_PARAMS:
2850                 ret = ublk_ctrl_set_params(ub, cmd);
2851                 break;
2852         case UBLK_CMD_START_USER_RECOVERY:
2853                 ret = ublk_ctrl_start_recovery(ub, cmd);
2854                 break;
2855         case UBLK_CMD_END_USER_RECOVERY:
2856                 ret = ublk_ctrl_end_recovery(ub, cmd);
2857                 break;
2858         default:
2859                 ret = -ENOTSUPP;
2860                 break;
2861         }
2862
2863  put_dev:
2864         if (ub)
2865                 ublk_put_device(ub);
2866  out:
2867         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2868         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2869                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2870         return -EIOCBQUEUED;
2871 }
2872
2873 static const struct file_operations ublk_ctl_fops = {
2874         .open           = nonseekable_open,
2875         .uring_cmd      = ublk_ctrl_uring_cmd,
2876         .owner          = THIS_MODULE,
2877         .llseek         = noop_llseek,
2878 };
2879
2880 static struct miscdevice ublk_misc = {
2881         .minor          = MISC_DYNAMIC_MINOR,
2882         .name           = "ublk-control",
2883         .fops           = &ublk_ctl_fops,
2884 };
2885
2886 static int __init ublk_init(void)
2887 {
2888         int ret;
2889
2890         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2891                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2892
2893         init_waitqueue_head(&ublk_idr_wq);
2894
2895         ret = misc_register(&ublk_misc);
2896         if (ret)
2897                 return ret;
2898
2899         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2900         if (ret)
2901                 goto unregister_mis;
2902
2903         ret = class_register(&ublk_chr_class);
2904         if (ret)
2905                 goto free_chrdev_region;
2906
2907         return 0;
2908
2909 free_chrdev_region:
2910         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2911 unregister_mis:
2912         misc_deregister(&ublk_misc);
2913         return ret;
2914 }
2915
2916 static void __exit ublk_exit(void)
2917 {
2918         struct ublk_device *ub;
2919         int id;
2920
2921         idr_for_each_entry(&ublk_index_idr, ub, id)
2922                 ublk_remove(ub);
2923
2924         class_unregister(&ublk_chr_class);
2925         misc_deregister(&ublk_misc);
2926
2927         idr_destroy(&ublk_index_idr);
2928         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2929 }
2930
2931 module_init(ublk_init);
2932 module_exit(ublk_exit);
2933
2934 module_param(ublks_max, int, 0444);
2935 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2936
2937 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2938 MODULE_LICENSE("GPL");