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_flag_set(QUEUE_FLAG_ZONE_RESETALL, ub->ub_disk->queue);
 255         blk_queue_required_elevator_features(ub->ub_disk->queue,
 256                                              ELEVATOR_F_ZBD_SEQ_WRITE);
 257         disk_set_max_active_zones(ub->ub_disk, p->max_active_zones);
 258         disk_set_max_open_zones(ub->ub_disk, p->max_open_zones);
 259         blk_queue_max_zone_append_sectors(ub->ub_disk->queue, p->max_zone_append_sectors);
 260
 261         ub->ub_disk->nr_zones = ublk_get_nr_zones(ub);
 262
 263         return 0;
 264 }
 265
 266 /* Based on virtblk_alloc_report_buffer */
 267 static void *ublk_alloc_report_buffer(struct ublk_device *ublk,
 268                                       unsigned int nr_zones, size_t *buflen)
 269 {
 270         struct request_queue *q = ublk->ub_disk->queue;
 271         size_t bufsize;
 272         void *buf;
 273
 274         nr_zones = min_t(unsigned int, nr_zones,
 275                          ublk->ub_disk->nr_zones);
 276
 277         bufsize = nr_zones * sizeof(struct blk_zone);
 278         bufsize =
 279                 min_t(size_t, bufsize, queue_max_hw_sectors(q) << SECTOR_SHIFT);
 280
 281         while (bufsize >= sizeof(struct blk_zone)) {
 282                 buf = kvmalloc(bufsize, GFP_KERNEL | __GFP_NORETRY);
 283                 if (buf) {
 284                         *buflen = bufsize;
 285                         return buf;
 286                 }
 287                 bufsize >>= 1;
 288         }
 289
 290         *buflen = 0;
 291         return NULL;
 292 }
 293
 294 static int ublk_report_zones(struct gendisk *disk, sector_t sector,
 295                       unsigned int nr_zones, report_zones_cb cb, void *data)
 296 {
 297         struct ublk_device *ub = disk->private_data;
 298         unsigned int zone_size_sectors = disk->queue->limits.chunk_sectors;
 299         unsigned int first_zone = sector >> ilog2(zone_size_sectors);
 300         unsigned int done_zones = 0;
 301         unsigned int max_zones_per_request;
 302         int ret;
 303         struct blk_zone *buffer;
 304         size_t buffer_length;
 305
 306         nr_zones = min_t(unsigned int, ub->ub_disk->nr_zones - first_zone,
 307                          nr_zones);
 308
 309         buffer = ublk_alloc_report_buffer(ub, nr_zones, &buffer_length);
 310         if (!buffer)
 311                 return -ENOMEM;
 312
 313         max_zones_per_request = buffer_length / sizeof(struct blk_zone);
 314
 315         while (done_zones < nr_zones) {
 316                 unsigned int remaining_zones = nr_zones - done_zones;
 317                 unsigned int zones_in_request =
 318                         min_t(unsigned int, remaining_zones, max_zones_per_request);
 319                 struct request *req;
 320                 struct ublk_rq_data *pdu;
 321                 blk_status_t status;
 322
 323                 memset(buffer, 0, buffer_length);
 324
 325                 req = blk_mq_alloc_request(disk->queue, REQ_OP_DRV_IN, 0);
 326                 if (IS_ERR(req)) {
 327                         ret = PTR_ERR(req);
 328                         goto out;
 329                 }
 330
 331                 pdu = blk_mq_rq_to_pdu(req);
 332                 pdu->operation = UBLK_IO_OP_REPORT_ZONES;
 333                 pdu->sector = sector;
 334                 pdu->nr_zones = zones_in_request;
 335
 336                 ret = blk_rq_map_kern(disk->queue, req, buffer, buffer_length,
 337                                         GFP_KERNEL);
 338                 if (ret) {
 339                         blk_mq_free_request(req);
 340                         goto out;
 341                 }
 342
 343                 status = blk_execute_rq(req, 0);
 344                 ret = blk_status_to_errno(status);
 345                 blk_mq_free_request(req);
 346                 if (ret)
 347                         goto out;
 348
 349                 for (unsigned int i = 0; i < zones_in_request; i++) {
 350                         struct blk_zone *zone = buffer + i;
 351
 352                         /* A zero length zone means no more zones in this response */
 353                         if (!zone->len)
 354                                 break;
 355
 356                         ret = cb(zone, i, data);
 357                         if (ret)
 358                                 goto out;
 359
 360                         done_zones++;
 361                         sector += zone_size_sectors;
 362
 363                 }
 364         }
 365
 366         ret = done_zones;
 367
 368 out:
 369         kvfree(buffer);
 370         return ret;
 371 }
 372
 373 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 374                                          struct request *req)
 375 {
 376         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 377         struct ublk_io *io = &ubq->ios[req->tag];
 378         struct ublk_rq_data *pdu = blk_mq_rq_to_pdu(req);
 379         u32 ublk_op;
 380
 381         switch (req_op(req)) {
 382         case REQ_OP_ZONE_OPEN:
 383                 ublk_op = UBLK_IO_OP_ZONE_OPEN;
 384                 break;
 385         case REQ_OP_ZONE_CLOSE:
 386                 ublk_op = UBLK_IO_OP_ZONE_CLOSE;
 387                 break;
 388         case REQ_OP_ZONE_FINISH:
 389                 ublk_op = UBLK_IO_OP_ZONE_FINISH;
 390                 break;
 391         case REQ_OP_ZONE_RESET:
 392                 ublk_op = UBLK_IO_OP_ZONE_RESET;
 393                 break;
 394         case REQ_OP_ZONE_APPEND:
 395                 ublk_op = UBLK_IO_OP_ZONE_APPEND;
 396                 break;
 397         case REQ_OP_ZONE_RESET_ALL:
 398                 ublk_op = UBLK_IO_OP_ZONE_RESET_ALL;
 399                 break;
 400         case REQ_OP_DRV_IN:
 401                 ublk_op = pdu->operation;
 402                 switch (ublk_op) {
 403                 case UBLK_IO_OP_REPORT_ZONES:
 404                         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 405                         iod->nr_zones = pdu->nr_zones;
 406                         iod->start_sector = pdu->sector;
 407                         return BLK_STS_OK;
 408                 default:
 409                         return BLK_STS_IOERR;
 410                 }
 411         case REQ_OP_DRV_OUT:
 412                 /* We do not support drv_out */
 413                 return BLK_STS_NOTSUPP;
 414         default:
 415                 return BLK_STS_IOERR;
 416         }
 417
 418         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 419         iod->nr_sectors = blk_rq_sectors(req);
 420         iod->start_sector = blk_rq_pos(req);
 421         iod->addr = io->addr;
 422
 423         return BLK_STS_OK;
 424 }
 425
 426 #else
 427
 428 #define ublk_report_zones (NULL)
 429
 430 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 431 {
 432         return -EOPNOTSUPP;
 433 }
 434
 435 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
 436 {
 437         return -EOPNOTSUPP;
 438 }
 439
 440 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 441 {
 442         return 0;
 443 }
 444
 445 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 446                                          struct request *req)
 447 {
 448         return BLK_STS_NOTSUPP;
 449 }
 450
 451 #endif
 452
 453 static inline void __ublk_complete_rq(struct request *req);
 454 static void ublk_complete_rq(struct kref *ref);
 455
 456 static dev_t ublk_chr_devt;
 457 static const struct class ublk_chr_class = {
 458         .name = "ublk-char",
 459 };
 460
 461 static DEFINE_IDR(ublk_index_idr);
 462 static DEFINE_SPINLOCK(ublk_idr_lock);
 463 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 464
 465 static DEFINE_MUTEX(ublk_ctl_mutex);
 466
 467 /*
 468  * Max ublk devices allowed to add
 469  *
 470  * It can be extended to one per-user limit in future or even controlled
 471  * by cgroup.
 472  */
 473 #define UBLK_MAX_UBLKS UBLK_MINORS
 474 static unsigned int ublks_max = 64;
 475 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 476
 477 static struct miscdevice ublk_misc;
 478
 479 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 480 {
 481         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 482                 UBLK_QID_BITS_MASK;
 483 }
 484
 485 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 486 {
 487         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 488 }
 489
 490 static inline unsigned ublk_pos_to_tag(loff_t pos)
 491 {
 492         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 493                 UBLK_TAG_BITS_MASK;
 494 }
 495
 496 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 497 {
 498         struct request_queue *q = ub->ub_disk->queue;
 499         const struct ublk_param_basic *p = &ub->params.basic;
 500
 501         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 502         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 503         blk_queue_io_min(q, 1 << p->io_min_shift);
 504         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 505
 506         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 507                         p->attrs & UBLK_ATTR_FUA);
 508         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 509                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 510         else
 511                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 512
 513         blk_queue_max_hw_sectors(q, p->max_sectors);
 514         blk_queue_chunk_sectors(q, p->chunk_sectors);
 515         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 516
 517         if (p->attrs & UBLK_ATTR_READ_ONLY)
 518                 set_disk_ro(ub->ub_disk, true);
 519
 520         set_capacity(ub->ub_disk, p->dev_sectors);
 521 }
 522
 523 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 524 {
 525         struct request_queue *q = ub->ub_disk->queue;
 526         const struct ublk_param_discard *p = &ub->params.discard;
 527
 528         q->limits.discard_alignment = p->discard_alignment;
 529         q->limits.discard_granularity = p->discard_granularity;
 530         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 531         blk_queue_max_write_zeroes_sectors(q,
 532                         p->max_write_zeroes_sectors);
 533         blk_queue_max_discard_segments(q, p->max_discard_segments);
 534 }
 535
 536 static int ublk_validate_params(const struct ublk_device *ub)
 537 {
 538         /* basic param is the only one which must be set */
 539         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 540                 const struct ublk_param_basic *p = &ub->params.basic;
 541
 542                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 543                         return -EINVAL;
 544
 545                 if (p->logical_bs_shift > p->physical_bs_shift)
 546                         return -EINVAL;
 547
 548                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 549                         return -EINVAL;
 550
 551                 if (ublk_dev_is_zoned(ub) && !p->chunk_sectors)
 552                         return -EINVAL;
 553         } else
 554                 return -EINVAL;
 555
 556         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 557                 const struct ublk_param_discard *p = &ub->params.discard;
 558
 559                 /* So far, only support single segment discard */
 560                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 561                         return -EINVAL;
 562
 563                 if (!p->discard_granularity)
 564                         return -EINVAL;
 565         }
 566
 567         /* dev_t is read-only */
 568         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 569                 return -EINVAL;
 570
 571         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 572                 return ublk_dev_param_zoned_validate(ub);
 573         else if (ublk_dev_is_zoned(ub))
 574                 return -EINVAL;
 575
 576         return 0;
 577 }
 578
 579 static int ublk_apply_params(struct ublk_device *ub)
 580 {
 581         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 582                 return -EINVAL;
 583
 584         ublk_dev_param_basic_apply(ub);
 585
 586         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 587                 ublk_dev_param_discard_apply(ub);
 588
 589         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 590                 return ublk_dev_param_zoned_apply(ub);
 591
 592         return 0;
 593 }
 594
 595 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 596 {
 597         return ubq->flags & UBLK_F_USER_COPY;
 598 }
 599
 600 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 601 {
 602         /*
 603          * read()/write() is involved in user copy, so request reference
 604          * has to be grabbed
 605          */
 606         return ublk_support_user_copy(ubq);
 607 }
 608
 609 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 610                 struct request *req)
 611 {
 612         if (ublk_need_req_ref(ubq)) {
 613                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 614
 615                 kref_init(&data->ref);
 616         }
 617 }
 618
 619 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 620                 struct request *req)
 621 {
 622         if (ublk_need_req_ref(ubq)) {
 623                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 624
 625                 return kref_get_unless_zero(&data->ref);
 626         }
 627
 628         return true;
 629 }
 630
 631 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 632                 struct request *req)
 633 {
 634         if (ublk_need_req_ref(ubq)) {
 635                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 636
 637                 kref_put(&data->ref, ublk_complete_rq);
 638         } else {
 639                 __ublk_complete_rq(req);
 640         }
 641 }
 642
 643 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 644 {
 645         return ubq->flags & UBLK_F_NEED_GET_DATA;
 646 }
 647
 648 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 649 {
 650         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 651                 return ub;
 652         return NULL;
 653 }
 654
 655 static void ublk_put_device(struct ublk_device *ub)
 656 {
 657         put_device(&ub->cdev_dev);
 658 }
 659
 660 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 661                 int qid)
 662 {
 663        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 664 }
 665
 666 static inline bool ublk_rq_has_data(const struct request *rq)
 667 {
 668         return bio_has_data(rq->bio);
 669 }
 670
 671 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 672                 int tag)
 673 {
 674         return (struct ublksrv_io_desc *)
 675                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 676 }
 677
 678 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 679 {
 680         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 681 }
 682
 683 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 684 {
 685         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 686
 687         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 688                         PAGE_SIZE);
 689 }
 690
 691 static inline bool ublk_queue_can_use_recovery_reissue(
 692                 struct ublk_queue *ubq)
 693 {
 694         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 695                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 696 }
 697
 698 static inline bool ublk_queue_can_use_recovery(
 699                 struct ublk_queue *ubq)
 700 {
 701         return ubq->flags & UBLK_F_USER_RECOVERY;
 702 }
 703
 704 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 705 {
 706         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 707 }
 708
 709 static void ublk_free_disk(struct gendisk *disk)
 710 {
 711         struct ublk_device *ub = disk->private_data;
 712
 713         clear_bit(UB_STATE_USED, &ub->state);
 714         put_device(&ub->cdev_dev);
 715 }
 716
 717 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 718                 unsigned int *owner_gid)
 719 {
 720         kuid_t uid;
 721         kgid_t gid;
 722
 723         current_uid_gid(&uid, &gid);
 724
 725         *owner_uid = from_kuid(&init_user_ns, uid);
 726         *owner_gid = from_kgid(&init_user_ns, gid);
 727 }
 728
 729 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 730 {
 731         struct ublk_device *ub = disk->private_data;
 732
 733         if (capable(CAP_SYS_ADMIN))
 734                 return 0;
 735
 736         /*
 737          * If it is one unprivileged device, only owner can open
 738          * the disk. Otherwise it could be one trap made by one
 739          * evil user who grants this disk's privileges to other
 740          * users deliberately.
 741          *
 742          * This way is reasonable too given anyone can create
 743          * unprivileged device, and no need other's grant.
 744          */
 745         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 746                 unsigned int curr_uid, curr_gid;
 747
 748                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 749
 750                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 751                                 ub->dev_info.owner_gid)
 752                         return -EPERM;
 753         }
 754
 755         return 0;
 756 }
 757
 758 static const struct block_device_operations ub_fops = {
 759         .owner =        THIS_MODULE,
 760         .open =         ublk_open,
 761         .free_disk =    ublk_free_disk,
 762         .report_zones = ublk_report_zones,
 763 };
 764
 765 #define UBLK_MAX_PIN_PAGES      32
 766
 767 struct ublk_io_iter {
 768         struct page *pages[UBLK_MAX_PIN_PAGES];
 769         struct bio *bio;
 770         struct bvec_iter iter;
 771 };
 772
 773 /* return how many pages are copied */
 774 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 775                 size_t total, size_t pg_off, int dir)
 776 {
 777         unsigned done = 0;
 778         unsigned pg_idx = 0;
 779
 780         while (done < total) {
 781                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 782                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 783                                 (unsigned)(PAGE_SIZE - pg_off));
 784                 void *bv_buf = bvec_kmap_local(&bv);
 785                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 786
 787                 if (dir == ITER_DEST)
 788                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 789                 else
 790                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 791
 792                 kunmap_local(pg_buf);
 793                 kunmap_local(bv_buf);
 794
 795                 /* advance page array */
 796                 pg_off += bytes;
 797                 if (pg_off == PAGE_SIZE) {
 798                         pg_idx += 1;
 799                         pg_off = 0;
 800                 }
 801
 802                 done += bytes;
 803
 804                 /* advance bio */
 805                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 806                 if (!data->iter.bi_size) {
 807                         data->bio = data->bio->bi_next;
 808                         if (data->bio == NULL)
 809                                 break;
 810                         data->iter = data->bio->bi_iter;
 811                 }
 812         }
 813 }
 814
 815 static bool ublk_advance_io_iter(const struct request *req,
 816                 struct ublk_io_iter *iter, unsigned int offset)
 817 {
 818         struct bio *bio = req->bio;
 819
 820         for_each_bio(bio) {
 821                 if (bio->bi_iter.bi_size > offset) {
 822                         iter->bio = bio;
 823                         iter->iter = bio->bi_iter;
 824                         bio_advance_iter(iter->bio, &iter->iter, offset);
 825                         return true;
 826                 }
 827                 offset -= bio->bi_iter.bi_size;
 828         }
 829         return false;
 830 }
 831
 832 /*
 833  * Copy data between request pages and io_iter, and 'offset'
 834  * is the start point of linear offset of request.
 835  */
 836 static size_t ublk_copy_user_pages(const struct request *req,
 837                 unsigned offset, struct iov_iter *uiter, int dir)
 838 {
 839         struct ublk_io_iter iter;
 840         size_t done = 0;
 841
 842         if (!ublk_advance_io_iter(req, &iter, offset))
 843                 return 0;
 844
 845         while (iov_iter_count(uiter) && iter.bio) {
 846                 unsigned nr_pages;
 847                 ssize_t len;
 848                 size_t off;
 849                 int i;
 850
 851                 len = iov_iter_get_pages2(uiter, iter.pages,
 852                                 iov_iter_count(uiter),
 853                                 UBLK_MAX_PIN_PAGES, &off);
 854                 if (len <= 0)
 855                         return done;
 856
 857                 ublk_copy_io_pages(&iter, len, off, dir);
 858                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 859                 for (i = 0; i < nr_pages; i++) {
 860                         if (dir == ITER_DEST)
 861                                 set_page_dirty(iter.pages[i]);
 862                         put_page(iter.pages[i]);
 863                 }
 864                 done += len;
 865         }
 866
 867         return done;
 868 }
 869
 870 static inline bool ublk_need_map_req(const struct request *req)
 871 {
 872         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 873 }
 874
 875 static inline bool ublk_need_unmap_req(const struct request *req)
 876 {
 877         return ublk_rq_has_data(req) &&
 878                (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN);
 879 }
 880
 881 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 882                 struct ublk_io *io)
 883 {
 884         const unsigned int rq_bytes = blk_rq_bytes(req);
 885
 886         if (ublk_support_user_copy(ubq))
 887                 return rq_bytes;
 888
 889         /*
 890          * no zero copy, we delay copy WRITE request data into ublksrv
 891          * context and the big benefit is that pinning pages in current
 892          * context is pretty fast, see ublk_pin_user_pages
 893          */
 894         if (ublk_need_map_req(req)) {
 895                 struct iov_iter iter;
 896                 struct iovec iov;
 897                 const int dir = ITER_DEST;
 898
 899                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 900                                 &iov, &iter);
 901
 902                 return ublk_copy_user_pages(req, 0, &iter, dir);
 903         }
 904         return rq_bytes;
 905 }
 906
 907 static int ublk_unmap_io(const struct ublk_queue *ubq,
 908                 const struct request *req,
 909                 struct ublk_io *io)
 910 {
 911         const unsigned int rq_bytes = blk_rq_bytes(req);
 912
 913         if (ublk_support_user_copy(ubq))
 914                 return rq_bytes;
 915
 916         if (ublk_need_unmap_req(req)) {
 917                 struct iov_iter iter;
 918                 struct iovec iov;
 919                 const int dir = ITER_SOURCE;
 920
 921                 WARN_ON_ONCE(io->res > rq_bytes);
 922
 923                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 924                                 &iov, &iter);
 925                 return ublk_copy_user_pages(req, 0, &iter, dir);
 926         }
 927         return rq_bytes;
 928 }
 929
 930 static inline unsigned int ublk_req_build_flags(struct request *req)
 931 {
 932         unsigned flags = 0;
 933
 934         if (req->cmd_flags & REQ_FAILFAST_DEV)
 935                 flags |= UBLK_IO_F_FAILFAST_DEV;
 936
 937         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 938                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 939
 940         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 941                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 942
 943         if (req->cmd_flags & REQ_META)
 944                 flags |= UBLK_IO_F_META;
 945
 946         if (req->cmd_flags & REQ_FUA)
 947                 flags |= UBLK_IO_F_FUA;
 948
 949         if (req->cmd_flags & REQ_NOUNMAP)
 950                 flags |= UBLK_IO_F_NOUNMAP;
 951
 952         if (req->cmd_flags & REQ_SWAP)
 953                 flags |= UBLK_IO_F_SWAP;
 954
 955         return flags;
 956 }
 957
 958 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 959 {
 960         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 961         struct ublk_io *io = &ubq->ios[req->tag];
 962         enum req_op op = req_op(req);
 963         u32 ublk_op;
 964
 965         if (!ublk_queue_is_zoned(ubq) &&
 966             (op_is_zone_mgmt(op) || op == REQ_OP_ZONE_APPEND))
 967                 return BLK_STS_IOERR;
 968
 969         switch (req_op(req)) {
 970         case REQ_OP_READ:
 971                 ublk_op = UBLK_IO_OP_READ;
 972                 break;
 973         case REQ_OP_WRITE:
 974                 ublk_op = UBLK_IO_OP_WRITE;
 975                 break;
 976         case REQ_OP_FLUSH:
 977                 ublk_op = UBLK_IO_OP_FLUSH;
 978                 break;
 979         case REQ_OP_DISCARD:
 980                 ublk_op = UBLK_IO_OP_DISCARD;
 981                 break;
 982         case REQ_OP_WRITE_ZEROES:
 983                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 984                 break;
 985         default:
 986                 if (ublk_queue_is_zoned(ubq))
 987                         return ublk_setup_iod_zoned(ubq, req);
 988                 return BLK_STS_IOERR;
 989         }
 990
 991         /* need to translate since kernel may change */
 992         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 993         iod->nr_sectors = blk_rq_sectors(req);
 994         iod->start_sector = blk_rq_pos(req);
 995         iod->addr = io->addr;
 996
 997         return BLK_STS_OK;
 998 }
 999
1000 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
1001                 struct io_uring_cmd *ioucmd)
1002 {
1003         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
1004 }
1005
1006 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
1007 {
1008         return ubq->ubq_daemon->flags & PF_EXITING;
1009 }
1010
1011 /* todo: handle partial completion */
1012 static inline void __ublk_complete_rq(struct request *req)
1013 {
1014         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1015         struct ublk_io *io = &ubq->ios[req->tag];
1016         unsigned int unmapped_bytes;
1017         blk_status_t res = BLK_STS_OK;
1018
1019         /* called from ublk_abort_queue() code path */
1020         if (io->flags & UBLK_IO_FLAG_ABORTED) {
1021                 res = BLK_STS_IOERR;
1022                 goto exit;
1023         }
1024
1025         /* failed read IO if nothing is read */
1026         if (!io->res && req_op(req) == REQ_OP_READ)
1027                 io->res = -EIO;
1028
1029         if (io->res < 0) {
1030                 res = errno_to_blk_status(io->res);
1031                 goto exit;
1032         }
1033
1034         /*
1035          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
1036          * directly.
1037          *
1038          * Both the two needn't unmap.
1039          */
1040         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE &&
1041             req_op(req) != REQ_OP_DRV_IN)
1042                 goto exit;
1043
1044         /* for READ request, writing data in iod->addr to rq buffers */
1045         unmapped_bytes = ublk_unmap_io(ubq, req, io);
1046
1047         /*
1048          * Extremely impossible since we got data filled in just before
1049          *
1050          * Re-read simply for this unlikely case.
1051          */
1052         if (unlikely(unmapped_bytes < io->res))
1053                 io->res = unmapped_bytes;
1054
1055         if (blk_update_request(req, BLK_STS_OK, io->res))
1056                 blk_mq_requeue_request(req, true);
1057         else
1058                 __blk_mq_end_request(req, BLK_STS_OK);
1059
1060         return;
1061 exit:
1062         blk_mq_end_request(req, res);
1063 }
1064
1065 static void ublk_complete_rq(struct kref *ref)
1066 {
1067         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
1068                         ref);
1069         struct request *req = blk_mq_rq_from_pdu(data);
1070
1071         __ublk_complete_rq(req);
1072 }
1073
1074 /*
1075  * Since __ublk_rq_task_work always fails requests immediately during
1076  * exiting, __ublk_fail_req() is only called from abort context during
1077  * exiting. So lock is unnecessary.
1078  *
1079  * Also aborting may not be started yet, keep in mind that one failed
1080  * request may be issued by block layer again.
1081  */
1082 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
1083                 struct request *req)
1084 {
1085         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
1086
1087         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
1088                 io->flags |= UBLK_IO_FLAG_ABORTED;
1089                 if (ublk_queue_can_use_recovery_reissue(ubq))
1090                         blk_mq_requeue_request(req, false);
1091                 else
1092                         ublk_put_req_ref(ubq, req);
1093         }
1094 }
1095
1096 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
1097                                 unsigned issue_flags)
1098 {
1099         /* mark this cmd owned by ublksrv */
1100         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
1101
1102         /*
1103          * clear ACTIVE since we are done with this sqe/cmd slot
1104          * We can only accept io cmd in case of being not active.
1105          */
1106         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
1107
1108         /* tell ublksrv one io request is coming */
1109         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
1110 }
1111
1112 #define UBLK_REQUEUE_DELAY_MS   3
1113
1114 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
1115                 struct request *rq)
1116 {
1117         /* We cannot process this rq so just requeue it. */
1118         if (ublk_queue_can_use_recovery(ubq))
1119                 blk_mq_requeue_request(rq, false);
1120         else
1121                 blk_mq_end_request(rq, BLK_STS_IOERR);
1122
1123         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
1124 }
1125
1126 static inline void __ublk_rq_task_work(struct request *req,
1127                                        unsigned issue_flags)
1128 {
1129         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1130         int tag = req->tag;
1131         struct ublk_io *io = &ubq->ios[tag];
1132         unsigned int mapped_bytes;
1133
1134         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
1135                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1136                         ublk_get_iod(ubq, req->tag)->addr);
1137
1138         /*
1139          * Task is exiting if either:
1140          *
1141          * (1) current != ubq_daemon.
1142          * io_uring_cmd_complete_in_task() tries to run task_work
1143          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
1144          *
1145          * (2) current->flags & PF_EXITING.
1146          */
1147         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
1148                 __ublk_abort_rq(ubq, req);
1149                 return;
1150         }
1151
1152         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
1153                 /*
1154                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
1155                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
1156                  * and notify it.
1157                  */
1158                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
1159                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
1160                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
1161                                         __func__, io->cmd->cmd_op, ubq->q_id,
1162                                         req->tag, io->flags);
1163                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
1164                         return;
1165                 }
1166                 /*
1167                  * We have handled UBLK_IO_NEED_GET_DATA command,
1168                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
1169                  * do the copy work.
1170                  */
1171                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
1172                 /* update iod->addr because ublksrv may have passed a new io buffer */
1173                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
1174                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
1175                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1176                                 ublk_get_iod(ubq, req->tag)->addr);
1177         }
1178
1179         mapped_bytes = ublk_map_io(ubq, req, io);
1180
1181         /* partially mapped, update io descriptor */
1182         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
1183                 /*
1184                  * Nothing mapped, retry until we succeed.
1185                  *
1186                  * We may never succeed in mapping any bytes here because
1187                  * of OOM. TODO: reserve one buffer with single page pinned
1188                  * for providing forward progress guarantee.
1189                  */
1190                 if (unlikely(!mapped_bytes)) {
1191                         blk_mq_requeue_request(req, false);
1192                         blk_mq_delay_kick_requeue_list(req->q,
1193                                         UBLK_REQUEUE_DELAY_MS);
1194                         return;
1195                 }
1196
1197                 ublk_get_iod(ubq, req->tag)->nr_sectors =
1198                         mapped_bytes >> 9;
1199         }
1200
1201         ublk_init_req_ref(ubq, req);
1202         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
1203 }
1204
1205 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
1206                                         unsigned issue_flags)
1207 {
1208         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1209         struct ublk_rq_data *data, *tmp;
1210
1211         io_cmds = llist_reverse_order(io_cmds);
1212         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1213                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
1214 }
1215
1216 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
1217 {
1218         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1219         struct ublk_rq_data *data, *tmp;
1220
1221         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1222                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
1223 }
1224
1225 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1226 {
1227         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1228         struct ublk_queue *ubq = pdu->ubq;
1229
1230         ublk_forward_io_cmds(ubq, issue_flags);
1231 }
1232
1233 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
1234 {
1235         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
1236         struct ublk_io *io;
1237
1238         if (!llist_add(&data->node, &ubq->io_cmds))
1239                 return;
1240
1241         io = &ubq->ios[rq->tag];
1242         /*
1243          * If the check pass, we know that this is a re-issued request aborted
1244          * previously in monitor_work because the ubq_daemon(cmd's task) is
1245          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
1246          * because this ioucmd's io_uring context may be freed now if no inflight
1247          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
1248          *
1249          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
1250          * the tag). Then the request is re-started(allocating the tag) and we are here.
1251          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
1252          * guarantees that here is a re-issued request aborted previously.
1253          */
1254         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
1255                 ublk_abort_io_cmds(ubq);
1256         } else {
1257                 struct io_uring_cmd *cmd = io->cmd;
1258                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1259
1260                 pdu->ubq = ubq;
1261                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
1262         }
1263 }
1264
1265 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
1266 {
1267         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
1268
1269         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
1270                 if (!ubq->timeout) {
1271                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
1272                         ubq->timeout = true;
1273                 }
1274
1275                 return BLK_EH_DONE;
1276         }
1277
1278         return BLK_EH_RESET_TIMER;
1279 }
1280
1281 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
1282                 const struct blk_mq_queue_data *bd)
1283 {
1284         struct ublk_queue *ubq = hctx->driver_data;
1285         struct request *rq = bd->rq;
1286         blk_status_t res;
1287
1288         /* fill iod to slot in io cmd buffer */
1289         res = ublk_setup_iod(ubq, rq);
1290         if (unlikely(res != BLK_STS_OK))
1291                 return BLK_STS_IOERR;
1292
1293         /* With recovery feature enabled, force_abort is set in
1294          * ublk_stop_dev() before calling del_gendisk(). We have to
1295          * abort all requeued and new rqs here to let del_gendisk()
1296          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1297          * to avoid UAF on io_uring ctx.
1298          *
1299          * Note: force_abort is guaranteed to be seen because it is set
1300          * before request queue is unqiuesced.
1301          */
1302         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1303                 return BLK_STS_IOERR;
1304
1305         blk_mq_start_request(bd->rq);
1306
1307         if (unlikely(ubq_daemon_is_dying(ubq))) {
1308                 __ublk_abort_rq(ubq, rq);
1309                 return BLK_STS_OK;
1310         }
1311
1312         ublk_queue_cmd(ubq, rq);
1313
1314         return BLK_STS_OK;
1315 }
1316
1317 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1318                 unsigned int hctx_idx)
1319 {
1320         struct ublk_device *ub = driver_data;
1321         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1322
1323         hctx->driver_data = ubq;
1324         return 0;
1325 }
1326
1327 static const struct blk_mq_ops ublk_mq_ops = {
1328         .queue_rq       = ublk_queue_rq,
1329         .init_hctx      = ublk_init_hctx,
1330         .timeout        = ublk_timeout,
1331 };
1332
1333 static int ublk_ch_open(struct inode *inode, struct file *filp)
1334 {
1335         struct ublk_device *ub = container_of(inode->i_cdev,
1336                         struct ublk_device, cdev);
1337
1338         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1339                 return -EBUSY;
1340         filp->private_data = ub;
1341         return 0;
1342 }
1343
1344 static int ublk_ch_release(struct inode *inode, struct file *filp)
1345 {
1346         struct ublk_device *ub = filp->private_data;
1347
1348         clear_bit(UB_STATE_OPEN, &ub->state);
1349         return 0;
1350 }
1351
1352 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1353 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1354 {
1355         struct ublk_device *ub = filp->private_data;
1356         size_t sz = vma->vm_end - vma->vm_start;
1357         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1358         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1359         int q_id, ret = 0;
1360
1361         spin_lock(&ub->mm_lock);
1362         if (!ub->mm)
1363                 ub->mm = current->mm;
1364         if (current->mm != ub->mm)
1365                 ret = -EINVAL;
1366         spin_unlock(&ub->mm_lock);
1367
1368         if (ret)
1369                 return ret;
1370
1371         if (vma->vm_flags & VM_WRITE)
1372                 return -EPERM;
1373
1374         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1375         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1376                 return -EINVAL;
1377
1378         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1379         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1380                         __func__, q_id, current->pid, vma->vm_start,
1381                         phys_off, (unsigned long)sz);
1382
1383         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1384                 return -EINVAL;
1385
1386         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1387         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1388 }
1389
1390 static void ublk_commit_completion(struct ublk_device *ub,
1391                 const struct ublksrv_io_cmd *ub_cmd)
1392 {
1393         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1394         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1395         struct ublk_io *io = &ubq->ios[tag];
1396         struct request *req;
1397
1398         /* now this cmd slot is owned by nbd driver */
1399         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1400         io->res = ub_cmd->result;
1401
1402         /* find the io request and complete */
1403         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1404         if (WARN_ON_ONCE(unlikely(!req)))
1405                 return;
1406
1407         if (req_op(req) == REQ_OP_ZONE_APPEND)
1408                 req->__sector = ub_cmd->zone_append_lba;
1409
1410         if (likely(!blk_should_fake_timeout(req->q)))
1411                 ublk_put_req_ref(ubq, req);
1412 }
1413
1414 /*
1415  * When ->ubq_daemon is exiting, either new request is ended immediately,
1416  * or any queued io command is drained, so it is safe to abort queue
1417  * lockless
1418  */
1419 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1420 {
1421         int i;
1422
1423         if (!ublk_get_device(ub))
1424                 return;
1425
1426         for (i = 0; i < ubq->q_depth; i++) {
1427                 struct ublk_io *io = &ubq->ios[i];
1428
1429                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1430                         struct request *rq;
1431
1432                         /*
1433                          * Either we fail the request or ublk_rq_task_work_fn
1434                          * will do it
1435                          */
1436                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1437                         if (rq)
1438                                 __ublk_fail_req(ubq, io, rq);
1439                 }
1440         }
1441         ublk_put_device(ub);
1442 }
1443
1444 static void ublk_daemon_monitor_work(struct work_struct *work)
1445 {
1446         struct ublk_device *ub =
1447                 container_of(work, struct ublk_device, monitor_work.work);
1448         int i;
1449
1450         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1451                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1452
1453                 if (ubq_daemon_is_dying(ubq)) {
1454                         if (ublk_queue_can_use_recovery(ubq))
1455                                 schedule_work(&ub->quiesce_work);
1456                         else
1457                                 schedule_work(&ub->stop_work);
1458
1459                         /* abort queue is for making forward progress */
1460                         ublk_abort_queue(ub, ubq);
1461                 }
1462         }
1463
1464         /*
1465          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1466          * after ublk_remove() or __ublk_quiesce_dev() is started.
1467          *
1468          * No need ub->mutex, monitor work are canceled after state is marked
1469          * as not LIVE, so new state is observed reliably.
1470          */
1471         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1472                 schedule_delayed_work(&ub->monitor_work,
1473                                 UBLK_DAEMON_MONITOR_PERIOD);
1474 }
1475
1476 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1477 {
1478         return ubq->nr_io_ready == ubq->q_depth;
1479 }
1480
1481 static void ublk_cmd_cancel_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1482 {
1483         io_uring_cmd_done(cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1484 }
1485
1486 static void ublk_cancel_queue(struct ublk_queue *ubq)
1487 {
1488         int i;
1489
1490         if (!ublk_queue_ready(ubq))
1491                 return;
1492
1493         for (i = 0; i < ubq->q_depth; i++) {
1494                 struct ublk_io *io = &ubq->ios[i];
1495
1496                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1497                         io_uring_cmd_complete_in_task(io->cmd,
1498                                                       ublk_cmd_cancel_cb);
1499         }
1500
1501         /* all io commands are canceled */
1502         ubq->nr_io_ready = 0;
1503 }
1504
1505 /* Cancel all pending commands, must be called after del_gendisk() returns */
1506 static void ublk_cancel_dev(struct ublk_device *ub)
1507 {
1508         int i;
1509
1510         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1511                 ublk_cancel_queue(ublk_get_queue(ub, i));
1512 }
1513
1514 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1515 {
1516         bool *idle = data;
1517
1518         if (blk_mq_request_started(rq)) {
1519                 *idle = false;
1520                 return false;
1521         }
1522         return true;
1523 }
1524
1525 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1526 {
1527         bool idle;
1528
1529         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1530         while (true) {
1531                 idle = true;
1532                 blk_mq_tagset_busy_iter(&ub->tag_set,
1533                                 ublk_check_inflight_rq, &idle);
1534                 if (idle)
1535                         break;
1536                 msleep(UBLK_REQUEUE_DELAY_MS);
1537         }
1538 }
1539
1540 static void __ublk_quiesce_dev(struct ublk_device *ub)
1541 {
1542         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1543                         __func__, ub->dev_info.dev_id,
1544                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1545                         "LIVE" : "QUIESCED");
1546         blk_mq_quiesce_queue(ub->ub_disk->queue);
1547         ublk_wait_tagset_rqs_idle(ub);
1548         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1549         ublk_cancel_dev(ub);
1550         /* we are going to release task_struct of ubq_daemon and resets
1551          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1552          * Besides, monitor_work is not necessary in QUIESCED state since we have
1553          * already scheduled quiesce_work and quiesced all ubqs.
1554          *
1555          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1556          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1557          */
1558         cancel_delayed_work_sync(&ub->monitor_work);
1559 }
1560
1561 static void ublk_quiesce_work_fn(struct work_struct *work)
1562 {
1563         struct ublk_device *ub =
1564                 container_of(work, struct ublk_device, quiesce_work);
1565
1566         mutex_lock(&ub->mutex);
1567         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1568                 goto unlock;
1569         __ublk_quiesce_dev(ub);
1570  unlock:
1571         mutex_unlock(&ub->mutex);
1572 }
1573
1574 static void ublk_unquiesce_dev(struct ublk_device *ub)
1575 {
1576         int i;
1577
1578         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1579                         __func__, ub->dev_info.dev_id,
1580                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1581                         "LIVE" : "QUIESCED");
1582         /* quiesce_work has run. We let requeued rqs be aborted
1583          * before running fallback_wq. "force_abort" must be seen
1584          * after request queue is unqiuesced. Then del_gendisk()
1585          * can move on.
1586          */
1587         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1588                 ublk_get_queue(ub, i)->force_abort = true;
1589
1590         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1591         /* We may have requeued some rqs in ublk_quiesce_queue() */
1592         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1593 }
1594
1595 static void ublk_stop_dev(struct ublk_device *ub)
1596 {
1597         mutex_lock(&ub->mutex);
1598         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1599                 goto unlock;
1600         if (ublk_can_use_recovery(ub)) {
1601                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1602                         __ublk_quiesce_dev(ub);
1603                 ublk_unquiesce_dev(ub);
1604         }
1605         del_gendisk(ub->ub_disk);
1606         ub->dev_info.state = UBLK_S_DEV_DEAD;
1607         ub->dev_info.ublksrv_pid = -1;
1608         put_disk(ub->ub_disk);
1609         ub->ub_disk = NULL;
1610  unlock:
1611         ublk_cancel_dev(ub);
1612         mutex_unlock(&ub->mutex);
1613         cancel_delayed_work_sync(&ub->monitor_work);
1614 }
1615
1616 /* device can only be started after all IOs are ready */
1617 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1618 {
1619         mutex_lock(&ub->mutex);
1620         ubq->nr_io_ready++;
1621         if (ublk_queue_ready(ubq)) {
1622                 ubq->ubq_daemon = current;
1623                 get_task_struct(ubq->ubq_daemon);
1624                 ub->nr_queues_ready++;
1625
1626                 if (capable(CAP_SYS_ADMIN))
1627                         ub->nr_privileged_daemon++;
1628         }
1629         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1630                 complete_all(&ub->completion);
1631         mutex_unlock(&ub->mutex);
1632 }
1633
1634 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1635                 int tag)
1636 {
1637         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1638         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1639
1640         ublk_queue_cmd(ubq, req);
1641 }
1642
1643 static inline int ublk_check_cmd_op(u32 cmd_op)
1644 {
1645         u32 ioc_type = _IOC_TYPE(cmd_op);
1646
1647         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1648                 return -EOPNOTSUPP;
1649
1650         if (ioc_type != 'u' && ioc_type != 0)
1651                 return -EOPNOTSUPP;
1652
1653         return 0;
1654 }
1655
1656 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1657                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1658 {
1659         io->cmd = cmd;
1660         io->flags |= UBLK_IO_FLAG_ACTIVE;
1661         io->addr = buf_addr;
1662 }
1663
1664 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1665                                unsigned int issue_flags,
1666                                const struct ublksrv_io_cmd *ub_cmd)
1667 {
1668         struct ublk_device *ub = cmd->file->private_data;
1669         struct ublk_queue *ubq;
1670         struct ublk_io *io;
1671         u32 cmd_op = cmd->cmd_op;
1672         unsigned tag = ub_cmd->tag;
1673         int ret = -EINVAL;
1674         struct request *req;
1675
1676         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1677                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1678                         ub_cmd->result);
1679
1680         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1681                 goto out;
1682
1683         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1684         if (!ubq || ub_cmd->q_id != ubq->q_id)
1685                 goto out;
1686
1687         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1688                 goto out;
1689
1690         if (tag >= ubq->q_depth)
1691                 goto out;
1692
1693         io = &ubq->ios[tag];
1694
1695         /* there is pending io cmd, something must be wrong */
1696         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1697                 ret = -EBUSY;
1698                 goto out;
1699         }
1700
1701         /*
1702          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1703          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1704          */
1705         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1706                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1707                 goto out;
1708
1709         ret = ublk_check_cmd_op(cmd_op);
1710         if (ret)
1711                 goto out;
1712
1713         ret = -EINVAL;
1714         switch (_IOC_NR(cmd_op)) {
1715         case UBLK_IO_FETCH_REQ:
1716                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1717                 if (ublk_queue_ready(ubq)) {
1718                         ret = -EBUSY;
1719                         goto out;
1720                 }
1721                 /*
1722                  * The io is being handled by server, so COMMIT_RQ is expected
1723                  * instead of FETCH_REQ
1724                  */
1725                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1726                         goto out;
1727
1728                 if (!ublk_support_user_copy(ubq)) {
1729                         /*
1730                          * FETCH_RQ has to provide IO buffer if NEED GET
1731                          * DATA is not enabled
1732                          */
1733                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1734                                 goto out;
1735                 } else if (ub_cmd->addr) {
1736                         /* User copy requires addr to be unset */
1737                         ret = -EINVAL;
1738                         goto out;
1739                 }
1740
1741                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1742                 ublk_mark_io_ready(ub, ubq);
1743                 break;
1744         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1745                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1746
1747                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1748                         goto out;
1749
1750                 if (!ublk_support_user_copy(ubq)) {
1751                         /*
1752                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1753                          * NEED GET DATA is not enabled or it is Read IO.
1754                          */
1755                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1756                                                 req_op(req) == REQ_OP_READ))
1757                                 goto out;
1758                 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1759                         /*
1760                          * User copy requires addr to be unset when command is
1761                          * not zone append
1762                          */
1763                         ret = -EINVAL;
1764                         goto out;
1765                 }
1766
1767                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1768                 ublk_commit_completion(ub, ub_cmd);
1769                 break;
1770         case UBLK_IO_NEED_GET_DATA:
1771                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1772                         goto out;
1773                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1774                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1775                 break;
1776         default:
1777                 goto out;
1778         }
1779         return -EIOCBQUEUED;
1780
1781  out:
1782         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1783         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1784                         __func__, cmd_op, tag, ret, io->flags);
1785         return -EIOCBQUEUED;
1786 }
1787
1788 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1789                 struct ublk_queue *ubq, int tag, size_t offset)
1790 {
1791         struct request *req;
1792
1793         if (!ublk_need_req_ref(ubq))
1794                 return NULL;
1795
1796         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1797         if (!req)
1798                 return NULL;
1799
1800         if (!ublk_get_req_ref(ubq, req))
1801                 return NULL;
1802
1803         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1804                 goto fail_put;
1805
1806         if (!ublk_rq_has_data(req))
1807                 goto fail_put;
1808
1809         if (offset > blk_rq_bytes(req))
1810                 goto fail_put;
1811
1812         return req;
1813 fail_put:
1814         ublk_put_req_ref(ubq, req);
1815         return NULL;
1816 }
1817
1818 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1819 {
1820         /*
1821          * Not necessary for async retry, but let's keep it simple and always
1822          * copy the values to avoid any potential reuse.
1823          */
1824         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1825         const struct ublksrv_io_cmd ub_cmd = {
1826                 .q_id = READ_ONCE(ub_src->q_id),
1827                 .tag = READ_ONCE(ub_src->tag),
1828                 .result = READ_ONCE(ub_src->result),
1829                 .addr = READ_ONCE(ub_src->addr)
1830         };
1831
1832         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1833 }
1834
1835 static inline bool ublk_check_ubuf_dir(const struct request *req,
1836                 int ubuf_dir)
1837 {
1838         /* copy ubuf to request pages */
1839         if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1840             ubuf_dir == ITER_SOURCE)
1841                 return true;
1842
1843         /* copy request pages to ubuf */
1844         if ((req_op(req) == REQ_OP_WRITE ||
1845              req_op(req) == REQ_OP_ZONE_APPEND) &&
1846             ubuf_dir == ITER_DEST)
1847                 return true;
1848
1849         return false;
1850 }
1851
1852 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1853                 struct iov_iter *iter, size_t *off, int dir)
1854 {
1855         struct ublk_device *ub = iocb->ki_filp->private_data;
1856         struct ublk_queue *ubq;
1857         struct request *req;
1858         size_t buf_off;
1859         u16 tag, q_id;
1860
1861         if (!ub)
1862                 return ERR_PTR(-EACCES);
1863
1864         if (!user_backed_iter(iter))
1865                 return ERR_PTR(-EACCES);
1866
1867         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1868                 return ERR_PTR(-EACCES);
1869
1870         tag = ublk_pos_to_tag(iocb->ki_pos);
1871         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1872         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1873
1874         if (q_id >= ub->dev_info.nr_hw_queues)
1875                 return ERR_PTR(-EINVAL);
1876
1877         ubq = ublk_get_queue(ub, q_id);
1878         if (!ubq)
1879                 return ERR_PTR(-EINVAL);
1880
1881         if (tag >= ubq->q_depth)
1882                 return ERR_PTR(-EINVAL);
1883
1884         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1885         if (!req)
1886                 return ERR_PTR(-EINVAL);
1887
1888         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1889                 goto fail;
1890
1891         if (!ublk_check_ubuf_dir(req, dir))
1892                 goto fail;
1893
1894         *off = buf_off;
1895         return req;
1896 fail:
1897         ublk_put_req_ref(ubq, req);
1898         return ERR_PTR(-EACCES);
1899 }
1900
1901 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1902 {
1903         struct ublk_queue *ubq;
1904         struct request *req;
1905         size_t buf_off;
1906         size_t ret;
1907
1908         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1909         if (IS_ERR(req))
1910                 return PTR_ERR(req);
1911
1912         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1913         ubq = req->mq_hctx->driver_data;
1914         ublk_put_req_ref(ubq, req);
1915
1916         return ret;
1917 }
1918
1919 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1920 {
1921         struct ublk_queue *ubq;
1922         struct request *req;
1923         size_t buf_off;
1924         size_t ret;
1925
1926         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1927         if (IS_ERR(req))
1928                 return PTR_ERR(req);
1929
1930         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1931         ubq = req->mq_hctx->driver_data;
1932         ublk_put_req_ref(ubq, req);
1933
1934         return ret;
1935 }
1936
1937 static const struct file_operations ublk_ch_fops = {
1938         .owner = THIS_MODULE,
1939         .open = ublk_ch_open,
1940         .release = ublk_ch_release,
1941         .llseek = no_llseek,
1942         .read_iter = ublk_ch_read_iter,
1943         .write_iter = ublk_ch_write_iter,
1944         .uring_cmd = ublk_ch_uring_cmd,
1945         .mmap = ublk_ch_mmap,
1946 };
1947
1948 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1949 {
1950         int size = ublk_queue_cmd_buf_size(ub, q_id);
1951         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1952
1953         if (ubq->ubq_daemon)
1954                 put_task_struct(ubq->ubq_daemon);
1955         if (ubq->io_cmd_buf)
1956                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1957 }
1958
1959 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1960 {
1961         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1962         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1963         void *ptr;
1964         int size;
1965
1966         ubq->flags = ub->dev_info.flags;
1967         ubq->q_id = q_id;
1968         ubq->q_depth = ub->dev_info.queue_depth;
1969         size = ublk_queue_cmd_buf_size(ub, q_id);
1970
1971         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1972         if (!ptr)
1973                 return -ENOMEM;
1974
1975         ubq->io_cmd_buf = ptr;
1976         ubq->dev = ub;
1977         return 0;
1978 }
1979
1980 static void ublk_deinit_queues(struct ublk_device *ub)
1981 {
1982         int nr_queues = ub->dev_info.nr_hw_queues;
1983         int i;
1984
1985         if (!ub->__queues)
1986                 return;
1987
1988         for (i = 0; i < nr_queues; i++)
1989                 ublk_deinit_queue(ub, i);
1990         kfree(ub->__queues);
1991 }
1992
1993 static int ublk_init_queues(struct ublk_device *ub)
1994 {
1995         int nr_queues = ub->dev_info.nr_hw_queues;
1996         int depth = ub->dev_info.queue_depth;
1997         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1998         int i, ret = -ENOMEM;
1999
2000         ub->queue_size = ubq_size;
2001         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
2002         if (!ub->__queues)
2003                 return ret;
2004
2005         for (i = 0; i < nr_queues; i++) {
2006                 if (ublk_init_queue(ub, i))
2007                         goto fail;
2008         }
2009
2010         init_completion(&ub->completion);
2011         return 0;
2012
2013  fail:
2014         ublk_deinit_queues(ub);
2015         return ret;
2016 }
2017
2018 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2019 {
2020         int i = idx;
2021         int err;
2022
2023         spin_lock(&ublk_idr_lock);
2024         /* allocate id, if @id >= 0, we're requesting that specific id */
2025         if (i >= 0) {
2026                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2027                 if (err == -ENOSPC)
2028                         err = -EEXIST;
2029         } else {
2030                 err = idr_alloc(&ublk_index_idr, ub, 0, UBLK_MAX_UBLKS,
2031                                 GFP_NOWAIT);
2032         }
2033         spin_unlock(&ublk_idr_lock);
2034
2035         if (err >= 0)
2036                 ub->ub_number = err;
2037
2038         return err;
2039 }
2040
2041 static void ublk_free_dev_number(struct ublk_device *ub)
2042 {
2043         spin_lock(&ublk_idr_lock);
2044         idr_remove(&ublk_index_idr, ub->ub_number);
2045         wake_up_all(&ublk_idr_wq);
2046         spin_unlock(&ublk_idr_lock);
2047 }
2048
2049 static void ublk_cdev_rel(struct device *dev)
2050 {
2051         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2052
2053         blk_mq_free_tag_set(&ub->tag_set);
2054         ublk_deinit_queues(ub);
2055         ublk_free_dev_number(ub);
2056         mutex_destroy(&ub->mutex);
2057         kfree(ub);
2058 }
2059
2060 static int ublk_add_chdev(struct ublk_device *ub)
2061 {
2062         struct device *dev = &ub->cdev_dev;
2063         int minor = ub->ub_number;
2064         int ret;
2065
2066         dev->parent = ublk_misc.this_device;
2067         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2068         dev->class = &ublk_chr_class;
2069         dev->release = ublk_cdev_rel;
2070         device_initialize(dev);
2071
2072         ret = dev_set_name(dev, "ublkc%d", minor);
2073         if (ret)
2074                 goto fail;
2075
2076         cdev_init(&ub->cdev, &ublk_ch_fops);
2077         ret = cdev_device_add(&ub->cdev, dev);
2078         if (ret)
2079                 goto fail;
2080
2081         ublks_added++;
2082         return 0;
2083  fail:
2084         put_device(dev);
2085         return ret;
2086 }
2087
2088 static void ublk_stop_work_fn(struct work_struct *work)
2089 {
2090         struct ublk_device *ub =
2091                 container_of(work, struct ublk_device, stop_work);
2092
2093         ublk_stop_dev(ub);
2094 }
2095
2096 /* align max io buffer size with PAGE_SIZE */
2097 static void ublk_align_max_io_size(struct ublk_device *ub)
2098 {
2099         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2100
2101         ub->dev_info.max_io_buf_bytes =
2102                 round_down(max_io_bytes, PAGE_SIZE);
2103 }
2104
2105 static int ublk_add_tag_set(struct ublk_device *ub)
2106 {
2107         ub->tag_set.ops = &ublk_mq_ops;
2108         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2109         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2110         ub->tag_set.numa_node = NUMA_NO_NODE;
2111         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2112         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2113         ub->tag_set.driver_data = ub;
2114         return blk_mq_alloc_tag_set(&ub->tag_set);
2115 }
2116
2117 static void ublk_remove(struct ublk_device *ub)
2118 {
2119         ublk_stop_dev(ub);
2120         cancel_work_sync(&ub->stop_work);
2121         cancel_work_sync(&ub->quiesce_work);
2122         cdev_device_del(&ub->cdev, &ub->cdev_dev);
2123         put_device(&ub->cdev_dev);
2124         ublks_added--;
2125 }
2126
2127 static struct ublk_device *ublk_get_device_from_id(int idx)
2128 {
2129         struct ublk_device *ub = NULL;
2130
2131         if (idx < 0)
2132                 return NULL;
2133
2134         spin_lock(&ublk_idr_lock);
2135         ub = idr_find(&ublk_index_idr, idx);
2136         if (ub)
2137                 ub = ublk_get_device(ub);
2138         spin_unlock(&ublk_idr_lock);
2139
2140         return ub;
2141 }
2142
2143 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2144 {
2145         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2146         int ublksrv_pid = (int)header->data[0];
2147         struct gendisk *disk;
2148         int ret = -EINVAL;
2149
2150         if (ublksrv_pid <= 0)
2151                 return -EINVAL;
2152
2153         if (wait_for_completion_interruptible(&ub->completion) != 0)
2154                 return -EINTR;
2155
2156         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2157
2158         mutex_lock(&ub->mutex);
2159         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2160             test_bit(UB_STATE_USED, &ub->state)) {
2161                 ret = -EEXIST;
2162                 goto out_unlock;
2163         }
2164
2165         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
2166         if (IS_ERR(disk)) {
2167                 ret = PTR_ERR(disk);
2168                 goto out_unlock;
2169         }
2170         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2171         disk->fops = &ub_fops;
2172         disk->private_data = ub;
2173
2174         ub->dev_info.ublksrv_pid = ublksrv_pid;
2175         ub->ub_disk = disk;
2176
2177         ret = ublk_apply_params(ub);
2178         if (ret)
2179                 goto out_put_disk;
2180
2181         /* don't probe partitions if any one ubq daemon is un-trusted */
2182         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2183                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2184
2185         get_device(&ub->cdev_dev);
2186         ub->dev_info.state = UBLK_S_DEV_LIVE;
2187
2188         if (ublk_dev_is_zoned(ub)) {
2189                 ret = ublk_revalidate_disk_zones(ub);
2190                 if (ret)
2191                         goto out_put_cdev;
2192         }
2193
2194         ret = add_disk(disk);
2195         if (ret)
2196                 goto out_put_cdev;
2197
2198         set_bit(UB_STATE_USED, &ub->state);
2199
2200 out_put_cdev:
2201         if (ret) {
2202                 ub->dev_info.state = UBLK_S_DEV_DEAD;
2203                 ublk_put_device(ub);
2204         }
2205 out_put_disk:
2206         if (ret)
2207                 put_disk(disk);
2208 out_unlock:
2209         mutex_unlock(&ub->mutex);
2210         return ret;
2211 }
2212
2213 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2214                 struct io_uring_cmd *cmd)
2215 {
2216         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2217         void __user *argp = (void __user *)(unsigned long)header->addr;
2218         cpumask_var_t cpumask;
2219         unsigned long queue;
2220         unsigned int retlen;
2221         unsigned int i;
2222         int ret;
2223
2224         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2225                 return -EINVAL;
2226         if (header->len & (sizeof(unsigned long)-1))
2227                 return -EINVAL;
2228         if (!header->addr)
2229                 return -EINVAL;
2230
2231         queue = header->data[0];
2232         if (queue >= ub->dev_info.nr_hw_queues)
2233                 return -EINVAL;
2234
2235         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2236                 return -ENOMEM;
2237
2238         for_each_possible_cpu(i) {
2239                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2240                         cpumask_set_cpu(i, cpumask);
2241         }
2242
2243         ret = -EFAULT;
2244         retlen = min_t(unsigned short, header->len, cpumask_size());
2245         if (copy_to_user(argp, cpumask, retlen))
2246                 goto out_free_cpumask;
2247         if (retlen != header->len &&
2248             clear_user(argp + retlen, header->len - retlen))
2249                 goto out_free_cpumask;
2250
2251         ret = 0;
2252 out_free_cpumask:
2253         free_cpumask_var(cpumask);
2254         return ret;
2255 }
2256
2257 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2258 {
2259         pr_devel("%s: dev id %d flags %llx\n", __func__,
2260                         info->dev_id, info->flags);
2261         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2262                         info->nr_hw_queues, info->queue_depth);
2263 }
2264
2265 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2266 {
2267         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2268         void __user *argp = (void __user *)(unsigned long)header->addr;
2269         struct ublksrv_ctrl_dev_info info;
2270         struct ublk_device *ub;
2271         int ret = -EINVAL;
2272
2273         if (header->len < sizeof(info) || !header->addr)
2274                 return -EINVAL;
2275         if (header->queue_id != (u16)-1) {
2276                 pr_warn("%s: queue_id is wrong %x\n",
2277                         __func__, header->queue_id);
2278                 return -EINVAL;
2279         }
2280
2281         if (copy_from_user(&info, argp, sizeof(info)))
2282                 return -EFAULT;
2283
2284         if (capable(CAP_SYS_ADMIN))
2285                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2286         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2287                 return -EPERM;
2288
2289         /*
2290          * unprivileged device can't be trusted, but RECOVERY and
2291          * RECOVERY_REISSUE still may hang error handling, so can't
2292          * support recovery features for unprivileged ublk now
2293          *
2294          * TODO: provide forward progress for RECOVERY handler, so that
2295          * unprivileged device can benefit from it
2296          */
2297         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2298                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2299                                 UBLK_F_USER_RECOVERY);
2300
2301         /* the created device is always owned by current user */
2302         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2303
2304         if (header->dev_id != info.dev_id) {
2305                 pr_warn("%s: dev id not match %u %u\n",
2306                         __func__, header->dev_id, info.dev_id);
2307                 return -EINVAL;
2308         }
2309
2310         if (header->dev_id != U32_MAX && header->dev_id >= UBLK_MAX_UBLKS) {
2311                 pr_warn("%s: dev id is too large. Max supported is %d\n",
2312                         __func__, UBLK_MAX_UBLKS - 1);
2313                 return -EINVAL;
2314         }
2315
2316         ublk_dump_dev_info(&info);
2317
2318         ret = mutex_lock_killable(&ublk_ctl_mutex);
2319         if (ret)
2320                 return ret;
2321
2322         ret = -EACCES;
2323         if (ublks_added >= ublks_max)
2324                 goto out_unlock;
2325
2326         ret = -ENOMEM;
2327         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2328         if (!ub)
2329                 goto out_unlock;
2330         mutex_init(&ub->mutex);
2331         spin_lock_init(&ub->mm_lock);
2332         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2333         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2334         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
2335
2336         ret = ublk_alloc_dev_number(ub, header->dev_id);
2337         if (ret < 0)
2338                 goto out_free_ub;
2339
2340         memcpy(&ub->dev_info, &info, sizeof(info));
2341
2342         /* update device id */
2343         ub->dev_info.dev_id = ub->ub_number;
2344
2345         /*
2346          * 64bit flags will be copied back to userspace as feature
2347          * negotiation result, so have to clear flags which driver
2348          * doesn't support yet, then userspace can get correct flags
2349          * (features) to handle.
2350          */
2351         ub->dev_info.flags &= UBLK_F_ALL;
2352
2353         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2354                 UBLK_F_URING_CMD_COMP_IN_TASK;
2355
2356         /* GET_DATA isn't needed any more with USER_COPY */
2357         if (ublk_dev_is_user_copy(ub))
2358                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2359
2360         /* Zoned storage support requires user copy feature */
2361         if (ublk_dev_is_zoned(ub) &&
2362             (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2363                 ret = -EINVAL;
2364                 goto out_free_dev_number;
2365         }
2366
2367         /* We are not ready to support zero copy */
2368         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2369
2370         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2371                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2372         ublk_align_max_io_size(ub);
2373
2374         ret = ublk_init_queues(ub);
2375         if (ret)
2376                 goto out_free_dev_number;
2377
2378         ret = ublk_add_tag_set(ub);
2379         if (ret)
2380                 goto out_deinit_queues;
2381
2382         ret = -EFAULT;
2383         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2384                 goto out_free_tag_set;
2385
2386         /*
2387          * Add the char dev so that ublksrv daemon can be setup.
2388          * ublk_add_chdev() will cleanup everything if it fails.
2389          */
2390         ret = ublk_add_chdev(ub);
2391         goto out_unlock;
2392
2393 out_free_tag_set:
2394         blk_mq_free_tag_set(&ub->tag_set);
2395 out_deinit_queues:
2396         ublk_deinit_queues(ub);
2397 out_free_dev_number:
2398         ublk_free_dev_number(ub);
2399 out_free_ub:
2400         mutex_destroy(&ub->mutex);
2401         kfree(ub);
2402 out_unlock:
2403         mutex_unlock(&ublk_ctl_mutex);
2404         return ret;
2405 }
2406
2407 static inline bool ublk_idr_freed(int id)
2408 {
2409         void *ptr;
2410
2411         spin_lock(&ublk_idr_lock);
2412         ptr = idr_find(&ublk_index_idr, id);
2413         spin_unlock(&ublk_idr_lock);
2414
2415         return ptr == NULL;
2416 }
2417
2418 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2419 {
2420         struct ublk_device *ub = *p_ub;
2421         int idx = ub->ub_number;
2422         int ret;
2423
2424         ret = mutex_lock_killable(&ublk_ctl_mutex);
2425         if (ret)
2426                 return ret;
2427
2428         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2429                 ublk_remove(ub);
2430                 set_bit(UB_STATE_DELETED, &ub->state);
2431         }
2432
2433         /* Mark the reference as consumed */
2434         *p_ub = NULL;
2435         ublk_put_device(ub);
2436         mutex_unlock(&ublk_ctl_mutex);
2437
2438         /*
2439          * Wait until the idr is removed, then it can be reused after
2440          * DEL_DEV command is returned.
2441          *
2442          * If we returns because of user interrupt, future delete command
2443          * may come:
2444          *
2445          * - the device number isn't freed, this device won't or needn't
2446          *   be deleted again, since UB_STATE_DELETED is set, and device
2447          *   will be released after the last reference is dropped
2448          *
2449          * - the device number is freed already, we will not find this
2450          *   device via ublk_get_device_from_id()
2451          */
2452         if (wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx)))
2453                 return -EINTR;
2454         return 0;
2455 }
2456
2457 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2458 {
2459         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2460
2461         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2462                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2463                         header->data[0], header->addr, header->len);
2464 }
2465
2466 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2467 {
2468         ublk_stop_dev(ub);
2469         cancel_work_sync(&ub->stop_work);
2470         cancel_work_sync(&ub->quiesce_work);
2471
2472         return 0;
2473 }
2474
2475 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2476                 struct io_uring_cmd *cmd)
2477 {
2478         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2479         void __user *argp = (void __user *)(unsigned long)header->addr;
2480
2481         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2482                 return -EINVAL;
2483
2484         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2485                 return -EFAULT;
2486
2487         return 0;
2488 }
2489
2490 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2491 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2492 {
2493         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2494         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2495
2496         if (ub->ub_disk) {
2497                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2498                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2499         } else {
2500                 ub->params.devt.disk_major = 0;
2501                 ub->params.devt.disk_minor = 0;
2502         }
2503         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2504 }
2505
2506 static int ublk_ctrl_get_params(struct ublk_device *ub,
2507                 struct io_uring_cmd *cmd)
2508 {
2509         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2510         void __user *argp = (void __user *)(unsigned long)header->addr;
2511         struct ublk_params_header ph;
2512         int ret;
2513
2514         if (header->len <= sizeof(ph) || !header->addr)
2515                 return -EINVAL;
2516
2517         if (copy_from_user(&ph, argp, sizeof(ph)))
2518                 return -EFAULT;
2519
2520         if (ph.len > header->len || !ph.len)
2521                 return -EINVAL;
2522
2523         if (ph.len > sizeof(struct ublk_params))
2524                 ph.len = sizeof(struct ublk_params);
2525
2526         mutex_lock(&ub->mutex);
2527         ublk_ctrl_fill_params_devt(ub);
2528         if (copy_to_user(argp, &ub->params, ph.len))
2529                 ret = -EFAULT;
2530         else
2531                 ret = 0;
2532         mutex_unlock(&ub->mutex);
2533
2534         return ret;
2535 }
2536
2537 static int ublk_ctrl_set_params(struct ublk_device *ub,
2538                 struct io_uring_cmd *cmd)
2539 {
2540         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2541         void __user *argp = (void __user *)(unsigned long)header->addr;
2542         struct ublk_params_header ph;
2543         int ret = -EFAULT;
2544
2545         if (header->len <= sizeof(ph) || !header->addr)
2546                 return -EINVAL;
2547
2548         if (copy_from_user(&ph, argp, sizeof(ph)))
2549                 return -EFAULT;
2550
2551         if (ph.len > header->len || !ph.len || !ph.types)
2552                 return -EINVAL;
2553
2554         if (ph.len > sizeof(struct ublk_params))
2555                 ph.len = sizeof(struct ublk_params);
2556
2557         /* parameters can only be changed when device isn't live */
2558         mutex_lock(&ub->mutex);
2559         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2560                 ret = -EACCES;
2561         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2562                 ret = -EFAULT;
2563         } else {
2564                 /* clear all we don't support yet */
2565                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2566                 ret = ublk_validate_params(ub);
2567                 if (ret)
2568                         ub->params.types = 0;
2569         }
2570         mutex_unlock(&ub->mutex);
2571
2572         return ret;
2573 }
2574
2575 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2576 {
2577         int i;
2578
2579         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2580         /* All old ioucmds have to be completed */
2581         WARN_ON_ONCE(ubq->nr_io_ready);
2582         /* old daemon is PF_EXITING, put it now */
2583         put_task_struct(ubq->ubq_daemon);
2584         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2585         ubq->ubq_daemon = NULL;
2586         ubq->timeout = false;
2587
2588         for (i = 0; i < ubq->q_depth; i++) {
2589                 struct ublk_io *io = &ubq->ios[i];
2590
2591                 /* forget everything now and be ready for new FETCH_REQ */
2592                 io->flags = 0;
2593                 io->cmd = NULL;
2594                 io->addr = 0;
2595         }
2596 }
2597
2598 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2599                 struct io_uring_cmd *cmd)
2600 {
2601         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2602         int ret = -EINVAL;
2603         int i;
2604
2605         mutex_lock(&ub->mutex);
2606         if (!ublk_can_use_recovery(ub))
2607                 goto out_unlock;
2608         /*
2609          * START_RECOVERY is only allowd after:
2610          *
2611          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2612          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2613          *     released.
2614          *
2615          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2616          *     (a)has quiesced request queue
2617          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2618          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2619          *     (d)has completed/camceled all ioucmds owned by ther dying process
2620          */
2621         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2622                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2623                 ret = -EBUSY;
2624                 goto out_unlock;
2625         }
2626         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2627         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2628                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2629         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2630         ub->mm = NULL;
2631         ub->nr_queues_ready = 0;
2632         ub->nr_privileged_daemon = 0;
2633         init_completion(&ub->completion);
2634         ret = 0;
2635  out_unlock:
2636         mutex_unlock(&ub->mutex);
2637         return ret;
2638 }
2639
2640 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2641                 struct io_uring_cmd *cmd)
2642 {
2643         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2644         int ublksrv_pid = (int)header->data[0];
2645         int ret = -EINVAL;
2646
2647         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2648                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2649         /* wait until new ubq_daemon sending all FETCH_REQ */
2650         if (wait_for_completion_interruptible(&ub->completion))
2651                 return -EINTR;
2652
2653         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2654                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2655
2656         mutex_lock(&ub->mutex);
2657         if (!ublk_can_use_recovery(ub))
2658                 goto out_unlock;
2659
2660         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2661                 ret = -EBUSY;
2662                 goto out_unlock;
2663         }
2664         ub->dev_info.ublksrv_pid = ublksrv_pid;
2665         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2666                         __func__, ublksrv_pid, header->dev_id);
2667         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2668         pr_devel("%s: queue unquiesced, dev id %d.\n",
2669                         __func__, header->dev_id);
2670         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2671         ub->dev_info.state = UBLK_S_DEV_LIVE;
2672         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2673         ret = 0;
2674  out_unlock:
2675         mutex_unlock(&ub->mutex);
2676         return ret;
2677 }
2678
2679 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2680 {
2681         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2682         void __user *argp = (void __user *)(unsigned long)header->addr;
2683         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2684
2685         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2686                 return -EINVAL;
2687
2688         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2689                 return -EFAULT;
2690
2691         return 0;
2692 }
2693
2694 /*
2695  * All control commands are sent via /dev/ublk-control, so we have to check
2696  * the destination device's permission
2697  */
2698 static int ublk_char_dev_permission(struct ublk_device *ub,
2699                 const char *dev_path, int mask)
2700 {
2701         int err;
2702         struct path path;
2703         struct kstat stat;
2704
2705         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2706         if (err)
2707                 return err;
2708
2709         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2710         if (err)
2711                 goto exit;
2712
2713         err = -EPERM;
2714         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2715                 goto exit;
2716
2717         err = inode_permission(&nop_mnt_idmap,
2718                         d_backing_inode(path.dentry), mask);
2719 exit:
2720         path_put(&path);
2721         return err;
2722 }
2723
2724 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2725                 struct io_uring_cmd *cmd)
2726 {
2727         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2728         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2729         void __user *argp = (void __user *)(unsigned long)header->addr;
2730         char *dev_path = NULL;
2731         int ret = 0;
2732         int mask;
2733
2734         if (!unprivileged) {
2735                 if (!capable(CAP_SYS_ADMIN))
2736                         return -EPERM;
2737                 /*
2738                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2739                  * char_dev_path in payload too, since userspace may not
2740                  * know if the specified device is created as unprivileged
2741                  * mode.
2742                  */
2743                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2744                         return 0;
2745         }
2746
2747         /*
2748          * User has to provide the char device path for unprivileged ublk
2749          *
2750          * header->addr always points to the dev path buffer, and
2751          * header->dev_path_len records length of dev path buffer.
2752          */
2753         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2754                 return -EINVAL;
2755
2756         if (header->len < header->dev_path_len)
2757                 return -EINVAL;
2758
2759         dev_path = memdup_user_nul(argp, header->dev_path_len);
2760         if (IS_ERR(dev_path))
2761                 return PTR_ERR(dev_path);
2762
2763         ret = -EINVAL;
2764         switch (_IOC_NR(cmd->cmd_op)) {
2765         case UBLK_CMD_GET_DEV_INFO:
2766         case UBLK_CMD_GET_DEV_INFO2:
2767         case UBLK_CMD_GET_QUEUE_AFFINITY:
2768         case UBLK_CMD_GET_PARAMS:
2769         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2770                 mask = MAY_READ;
2771                 break;
2772         case UBLK_CMD_START_DEV:
2773         case UBLK_CMD_STOP_DEV:
2774         case UBLK_CMD_ADD_DEV:
2775         case UBLK_CMD_DEL_DEV:
2776         case UBLK_CMD_SET_PARAMS:
2777         case UBLK_CMD_START_USER_RECOVERY:
2778         case UBLK_CMD_END_USER_RECOVERY:
2779                 mask = MAY_READ | MAY_WRITE;
2780                 break;
2781         default:
2782                 goto exit;
2783         }
2784
2785         ret = ublk_char_dev_permission(ub, dev_path, mask);
2786         if (!ret) {
2787                 header->len -= header->dev_path_len;
2788                 header->addr += header->dev_path_len;
2789         }
2790         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2791                         __func__, ub->ub_number, cmd->cmd_op,
2792                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2793                         dev_path, ret);
2794 exit:
2795         kfree(dev_path);
2796         return ret;
2797 }
2798
2799 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2800                 unsigned int issue_flags)
2801 {
2802         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2803         struct ublk_device *ub = NULL;
2804         u32 cmd_op = cmd->cmd_op;
2805         int ret = -EINVAL;
2806
2807         if (issue_flags & IO_URING_F_NONBLOCK)
2808                 return -EAGAIN;
2809
2810         ublk_ctrl_cmd_dump(cmd);
2811
2812         if (!(issue_flags & IO_URING_F_SQE128))
2813                 goto out;
2814
2815         ret = ublk_check_cmd_op(cmd_op);
2816         if (ret)
2817                 goto out;
2818
2819         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2820                 ret = ublk_ctrl_get_features(cmd);
2821                 goto out;
2822         }
2823
2824         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2825                 ret = -ENODEV;
2826                 ub = ublk_get_device_from_id(header->dev_id);
2827                 if (!ub)
2828                         goto out;
2829
2830                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2831                 if (ret)
2832                         goto put_dev;
2833         }
2834
2835         switch (_IOC_NR(cmd_op)) {
2836         case UBLK_CMD_START_DEV:
2837                 ret = ublk_ctrl_start_dev(ub, cmd);
2838                 break;
2839         case UBLK_CMD_STOP_DEV:
2840                 ret = ublk_ctrl_stop_dev(ub);
2841                 break;
2842         case UBLK_CMD_GET_DEV_INFO:
2843         case UBLK_CMD_GET_DEV_INFO2:
2844                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2845                 break;
2846         case UBLK_CMD_ADD_DEV:
2847                 ret = ublk_ctrl_add_dev(cmd);
2848                 break;
2849         case UBLK_CMD_DEL_DEV:
2850                 ret = ublk_ctrl_del_dev(&ub);
2851                 break;
2852         case UBLK_CMD_GET_QUEUE_AFFINITY:
2853                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2854                 break;
2855         case UBLK_CMD_GET_PARAMS:
2856                 ret = ublk_ctrl_get_params(ub, cmd);
2857                 break;
2858         case UBLK_CMD_SET_PARAMS:
2859                 ret = ublk_ctrl_set_params(ub, cmd);
2860                 break;
2861         case UBLK_CMD_START_USER_RECOVERY:
2862                 ret = ublk_ctrl_start_recovery(ub, cmd);
2863                 break;
2864         case UBLK_CMD_END_USER_RECOVERY:
2865                 ret = ublk_ctrl_end_recovery(ub, cmd);
2866                 break;
2867         default:
2868                 ret = -ENOTSUPP;
2869                 break;
2870         }
2871
2872  put_dev:
2873         if (ub)
2874                 ublk_put_device(ub);
2875  out:
2876         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2877         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2878                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2879         return -EIOCBQUEUED;
2880 }
2881
2882 static const struct file_operations ublk_ctl_fops = {
2883         .open           = nonseekable_open,
2884         .uring_cmd      = ublk_ctrl_uring_cmd,
2885         .owner          = THIS_MODULE,
2886         .llseek         = noop_llseek,
2887 };
2888
2889 static struct miscdevice ublk_misc = {
2890         .minor          = MISC_DYNAMIC_MINOR,
2891         .name           = "ublk-control",
2892         .fops           = &ublk_ctl_fops,
2893 };
2894
2895 static int __init ublk_init(void)
2896 {
2897         int ret;
2898
2899         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2900                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2901
2902         init_waitqueue_head(&ublk_idr_wq);
2903
2904         ret = misc_register(&ublk_misc);
2905         if (ret)
2906                 return ret;
2907
2908         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2909         if (ret)
2910                 goto unregister_mis;
2911
2912         ret = class_register(&ublk_chr_class);
2913         if (ret)
2914                 goto free_chrdev_region;
2915
2916         return 0;
2917
2918 free_chrdev_region:
2919         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2920 unregister_mis:
2921         misc_deregister(&ublk_misc);
2922         return ret;
2923 }
2924
2925 static void __exit ublk_exit(void)
2926 {
2927         struct ublk_device *ub;
2928         int id;
2929
2930         idr_for_each_entry(&ublk_index_idr, ub, id)
2931                 ublk_remove(ub);
2932
2933         class_unregister(&ublk_chr_class);
2934         misc_deregister(&ublk_misc);
2935
2936         idr_destroy(&ublk_index_idr);
2937         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2938 }
2939
2940 module_init(ublk_init);
2941 module_exit(ublk_exit);
2942
2943 module_param(ublks_max, int, 0444);
2944 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2945
2946 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2947 MODULE_LICENSE("GPL");