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         for (i = 0; i < ubq->q_depth; i++) {
1424                 struct ublk_io *io = &ubq->ios[i];
1425
1426                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1427                         struct request *rq;
1428
1429                         /*
1430                          * Either we fail the request or ublk_rq_task_work_fn
1431                          * will do it
1432                          */
1433                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1434                         if (rq)
1435                                 __ublk_fail_req(ubq, io, rq);
1436                 }
1437         }
1438 }
1439
1440 static void ublk_daemon_monitor_work(struct work_struct *work)
1441 {
1442         struct ublk_device *ub =
1443                 container_of(work, struct ublk_device, monitor_work.work);
1444         int i;
1445
1446         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1447                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1448
1449                 if (ubq_daemon_is_dying(ubq)) {
1450                         if (ublk_queue_can_use_recovery(ubq))
1451                                 schedule_work(&ub->quiesce_work);
1452                         else
1453                                 schedule_work(&ub->stop_work);
1454
1455                         /* abort queue is for making forward progress */
1456                         ublk_abort_queue(ub, ubq);
1457                 }
1458         }
1459
1460         /*
1461          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1462          * after ublk_remove() or __ublk_quiesce_dev() is started.
1463          *
1464          * No need ub->mutex, monitor work are canceled after state is marked
1465          * as not LIVE, so new state is observed reliably.
1466          */
1467         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1468                 schedule_delayed_work(&ub->monitor_work,
1469                                 UBLK_DAEMON_MONITOR_PERIOD);
1470 }
1471
1472 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1473 {
1474         return ubq->nr_io_ready == ubq->q_depth;
1475 }
1476
1477 static void ublk_cmd_cancel_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1478 {
1479         io_uring_cmd_done(cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1480 }
1481
1482 static void ublk_cancel_queue(struct ublk_queue *ubq)
1483 {
1484         int i;
1485
1486         if (!ublk_queue_ready(ubq))
1487                 return;
1488
1489         for (i = 0; i < ubq->q_depth; i++) {
1490                 struct ublk_io *io = &ubq->ios[i];
1491
1492                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1493                         io_uring_cmd_complete_in_task(io->cmd,
1494                                                       ublk_cmd_cancel_cb);
1495         }
1496
1497         /* all io commands are canceled */
1498         ubq->nr_io_ready = 0;
1499 }
1500
1501 /* Cancel all pending commands, must be called after del_gendisk() returns */
1502 static void ublk_cancel_dev(struct ublk_device *ub)
1503 {
1504         int i;
1505
1506         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1507                 ublk_cancel_queue(ublk_get_queue(ub, i));
1508 }
1509
1510 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1511 {
1512         bool *idle = data;
1513
1514         if (blk_mq_request_started(rq)) {
1515                 *idle = false;
1516                 return false;
1517         }
1518         return true;
1519 }
1520
1521 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1522 {
1523         bool idle;
1524
1525         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1526         while (true) {
1527                 idle = true;
1528                 blk_mq_tagset_busy_iter(&ub->tag_set,
1529                                 ublk_check_inflight_rq, &idle);
1530                 if (idle)
1531                         break;
1532                 msleep(UBLK_REQUEUE_DELAY_MS);
1533         }
1534 }
1535
1536 static void __ublk_quiesce_dev(struct ublk_device *ub)
1537 {
1538         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1539                         __func__, ub->dev_info.dev_id,
1540                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1541                         "LIVE" : "QUIESCED");
1542         blk_mq_quiesce_queue(ub->ub_disk->queue);
1543         ublk_wait_tagset_rqs_idle(ub);
1544         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1545         ublk_cancel_dev(ub);
1546         /* we are going to release task_struct of ubq_daemon and resets
1547          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1548          * Besides, monitor_work is not necessary in QUIESCED state since we have
1549          * already scheduled quiesce_work and quiesced all ubqs.
1550          *
1551          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1552          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1553          */
1554         cancel_delayed_work_sync(&ub->monitor_work);
1555 }
1556
1557 static void ublk_quiesce_work_fn(struct work_struct *work)
1558 {
1559         struct ublk_device *ub =
1560                 container_of(work, struct ublk_device, quiesce_work);
1561
1562         mutex_lock(&ub->mutex);
1563         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1564                 goto unlock;
1565         __ublk_quiesce_dev(ub);
1566  unlock:
1567         mutex_unlock(&ub->mutex);
1568 }
1569
1570 static void ublk_unquiesce_dev(struct ublk_device *ub)
1571 {
1572         int i;
1573
1574         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1575                         __func__, ub->dev_info.dev_id,
1576                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1577                         "LIVE" : "QUIESCED");
1578         /* quiesce_work has run. We let requeued rqs be aborted
1579          * before running fallback_wq. "force_abort" must be seen
1580          * after request queue is unqiuesced. Then del_gendisk()
1581          * can move on.
1582          */
1583         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1584                 ublk_get_queue(ub, i)->force_abort = true;
1585
1586         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1587         /* We may have requeued some rqs in ublk_quiesce_queue() */
1588         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1589 }
1590
1591 static void ublk_stop_dev(struct ublk_device *ub)
1592 {
1593         mutex_lock(&ub->mutex);
1594         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1595                 goto unlock;
1596         if (ublk_can_use_recovery(ub)) {
1597                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1598                         __ublk_quiesce_dev(ub);
1599                 ublk_unquiesce_dev(ub);
1600         }
1601         del_gendisk(ub->ub_disk);
1602         ub->dev_info.state = UBLK_S_DEV_DEAD;
1603         ub->dev_info.ublksrv_pid = -1;
1604         put_disk(ub->ub_disk);
1605         ub->ub_disk = NULL;
1606  unlock:
1607         ublk_cancel_dev(ub);
1608         mutex_unlock(&ub->mutex);
1609         cancel_delayed_work_sync(&ub->monitor_work);
1610 }
1611
1612 /* device can only be started after all IOs are ready */
1613 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1614 {
1615         mutex_lock(&ub->mutex);
1616         ubq->nr_io_ready++;
1617         if (ublk_queue_ready(ubq)) {
1618                 ubq->ubq_daemon = current;
1619                 get_task_struct(ubq->ubq_daemon);
1620                 ub->nr_queues_ready++;
1621
1622                 if (capable(CAP_SYS_ADMIN))
1623                         ub->nr_privileged_daemon++;
1624         }
1625         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1626                 complete_all(&ub->completion);
1627         mutex_unlock(&ub->mutex);
1628 }
1629
1630 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1631                 int tag)
1632 {
1633         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1634         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1635
1636         ublk_queue_cmd(ubq, req);
1637 }
1638
1639 static inline int ublk_check_cmd_op(u32 cmd_op)
1640 {
1641         u32 ioc_type = _IOC_TYPE(cmd_op);
1642
1643         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1644                 return -EOPNOTSUPP;
1645
1646         if (ioc_type != 'u' && ioc_type != 0)
1647                 return -EOPNOTSUPP;
1648
1649         return 0;
1650 }
1651
1652 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1653                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1654 {
1655         io->cmd = cmd;
1656         io->flags |= UBLK_IO_FLAG_ACTIVE;
1657         io->addr = buf_addr;
1658 }
1659
1660 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1661                                unsigned int issue_flags,
1662                                const struct ublksrv_io_cmd *ub_cmd)
1663 {
1664         struct ublk_device *ub = cmd->file->private_data;
1665         struct ublk_queue *ubq;
1666         struct ublk_io *io;
1667         u32 cmd_op = cmd->cmd_op;
1668         unsigned tag = ub_cmd->tag;
1669         int ret = -EINVAL;
1670         struct request *req;
1671
1672         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1673                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1674                         ub_cmd->result);
1675
1676         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1677                 goto out;
1678
1679         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1680         if (!ubq || ub_cmd->q_id != ubq->q_id)
1681                 goto out;
1682
1683         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1684                 goto out;
1685
1686         if (tag >= ubq->q_depth)
1687                 goto out;
1688
1689         io = &ubq->ios[tag];
1690
1691         /* there is pending io cmd, something must be wrong */
1692         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1693                 ret = -EBUSY;
1694                 goto out;
1695         }
1696
1697         /*
1698          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1699          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1700          */
1701         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1702                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1703                 goto out;
1704
1705         ret = ublk_check_cmd_op(cmd_op);
1706         if (ret)
1707                 goto out;
1708
1709         ret = -EINVAL;
1710         switch (_IOC_NR(cmd_op)) {
1711         case UBLK_IO_FETCH_REQ:
1712                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1713                 if (ublk_queue_ready(ubq)) {
1714                         ret = -EBUSY;
1715                         goto out;
1716                 }
1717                 /*
1718                  * The io is being handled by server, so COMMIT_RQ is expected
1719                  * instead of FETCH_REQ
1720                  */
1721                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1722                         goto out;
1723
1724                 if (!ublk_support_user_copy(ubq)) {
1725                         /*
1726                          * FETCH_RQ has to provide IO buffer if NEED GET
1727                          * DATA is not enabled
1728                          */
1729                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1730                                 goto out;
1731                 } else if (ub_cmd->addr) {
1732                         /* User copy requires addr to be unset */
1733                         ret = -EINVAL;
1734                         goto out;
1735                 }
1736
1737                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1738                 ublk_mark_io_ready(ub, ubq);
1739                 break;
1740         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1741                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1742
1743                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1744                         goto out;
1745
1746                 if (!ublk_support_user_copy(ubq)) {
1747                         /*
1748                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1749                          * NEED GET DATA is not enabled or it is Read IO.
1750                          */
1751                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1752                                                 req_op(req) == REQ_OP_READ))
1753                                 goto out;
1754                 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1755                         /*
1756                          * User copy requires addr to be unset when command is
1757                          * not zone append
1758                          */
1759                         ret = -EINVAL;
1760                         goto out;
1761                 }
1762
1763                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1764                 ublk_commit_completion(ub, ub_cmd);
1765                 break;
1766         case UBLK_IO_NEED_GET_DATA:
1767                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1768                         goto out;
1769                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1770                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1771                 break;
1772         default:
1773                 goto out;
1774         }
1775         return -EIOCBQUEUED;
1776
1777  out:
1778         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1779         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1780                         __func__, cmd_op, tag, ret, io->flags);
1781         return -EIOCBQUEUED;
1782 }
1783
1784 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1785                 struct ublk_queue *ubq, int tag, size_t offset)
1786 {
1787         struct request *req;
1788
1789         if (!ublk_need_req_ref(ubq))
1790                 return NULL;
1791
1792         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1793         if (!req)
1794                 return NULL;
1795
1796         if (!ublk_get_req_ref(ubq, req))
1797                 return NULL;
1798
1799         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1800                 goto fail_put;
1801
1802         if (!ublk_rq_has_data(req))
1803                 goto fail_put;
1804
1805         if (offset > blk_rq_bytes(req))
1806                 goto fail_put;
1807
1808         return req;
1809 fail_put:
1810         ublk_put_req_ref(ubq, req);
1811         return NULL;
1812 }
1813
1814 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1815 {
1816         /*
1817          * Not necessary for async retry, but let's keep it simple and always
1818          * copy the values to avoid any potential reuse.
1819          */
1820         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1821         const struct ublksrv_io_cmd ub_cmd = {
1822                 .q_id = READ_ONCE(ub_src->q_id),
1823                 .tag = READ_ONCE(ub_src->tag),
1824                 .result = READ_ONCE(ub_src->result),
1825                 .addr = READ_ONCE(ub_src->addr)
1826         };
1827
1828         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1829 }
1830
1831 static inline bool ublk_check_ubuf_dir(const struct request *req,
1832                 int ubuf_dir)
1833 {
1834         /* copy ubuf to request pages */
1835         if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1836             ubuf_dir == ITER_SOURCE)
1837                 return true;
1838
1839         /* copy request pages to ubuf */
1840         if ((req_op(req) == REQ_OP_WRITE ||
1841              req_op(req) == REQ_OP_ZONE_APPEND) &&
1842             ubuf_dir == ITER_DEST)
1843                 return true;
1844
1845         return false;
1846 }
1847
1848 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1849                 struct iov_iter *iter, size_t *off, int dir)
1850 {
1851         struct ublk_device *ub = iocb->ki_filp->private_data;
1852         struct ublk_queue *ubq;
1853         struct request *req;
1854         size_t buf_off;
1855         u16 tag, q_id;
1856
1857         if (!ub)
1858                 return ERR_PTR(-EACCES);
1859
1860         if (!user_backed_iter(iter))
1861                 return ERR_PTR(-EACCES);
1862
1863         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1864                 return ERR_PTR(-EACCES);
1865
1866         tag = ublk_pos_to_tag(iocb->ki_pos);
1867         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1868         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1869
1870         if (q_id >= ub->dev_info.nr_hw_queues)
1871                 return ERR_PTR(-EINVAL);
1872
1873         ubq = ublk_get_queue(ub, q_id);
1874         if (!ubq)
1875                 return ERR_PTR(-EINVAL);
1876
1877         if (tag >= ubq->q_depth)
1878                 return ERR_PTR(-EINVAL);
1879
1880         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1881         if (!req)
1882                 return ERR_PTR(-EINVAL);
1883
1884         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1885                 goto fail;
1886
1887         if (!ublk_check_ubuf_dir(req, dir))
1888                 goto fail;
1889
1890         *off = buf_off;
1891         return req;
1892 fail:
1893         ublk_put_req_ref(ubq, req);
1894         return ERR_PTR(-EACCES);
1895 }
1896
1897 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1898 {
1899         struct ublk_queue *ubq;
1900         struct request *req;
1901         size_t buf_off;
1902         size_t ret;
1903
1904         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1905         if (IS_ERR(req))
1906                 return PTR_ERR(req);
1907
1908         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1909         ubq = req->mq_hctx->driver_data;
1910         ublk_put_req_ref(ubq, req);
1911
1912         return ret;
1913 }
1914
1915 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1916 {
1917         struct ublk_queue *ubq;
1918         struct request *req;
1919         size_t buf_off;
1920         size_t ret;
1921
1922         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1923         if (IS_ERR(req))
1924                 return PTR_ERR(req);
1925
1926         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1927         ubq = req->mq_hctx->driver_data;
1928         ublk_put_req_ref(ubq, req);
1929
1930         return ret;
1931 }
1932
1933 static const struct file_operations ublk_ch_fops = {
1934         .owner = THIS_MODULE,
1935         .open = ublk_ch_open,
1936         .release = ublk_ch_release,
1937         .llseek = no_llseek,
1938         .read_iter = ublk_ch_read_iter,
1939         .write_iter = ublk_ch_write_iter,
1940         .uring_cmd = ublk_ch_uring_cmd,
1941         .mmap = ublk_ch_mmap,
1942 };
1943
1944 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1945 {
1946         int size = ublk_queue_cmd_buf_size(ub, q_id);
1947         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1948
1949         if (ubq->ubq_daemon)
1950                 put_task_struct(ubq->ubq_daemon);
1951         if (ubq->io_cmd_buf)
1952                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1953 }
1954
1955 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1956 {
1957         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1958         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1959         void *ptr;
1960         int size;
1961
1962         ubq->flags = ub->dev_info.flags;
1963         ubq->q_id = q_id;
1964         ubq->q_depth = ub->dev_info.queue_depth;
1965         size = ublk_queue_cmd_buf_size(ub, q_id);
1966
1967         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1968         if (!ptr)
1969                 return -ENOMEM;
1970
1971         ubq->io_cmd_buf = ptr;
1972         ubq->dev = ub;
1973         return 0;
1974 }
1975
1976 static void ublk_deinit_queues(struct ublk_device *ub)
1977 {
1978         int nr_queues = ub->dev_info.nr_hw_queues;
1979         int i;
1980
1981         if (!ub->__queues)
1982                 return;
1983
1984         for (i = 0; i < nr_queues; i++)
1985                 ublk_deinit_queue(ub, i);
1986         kfree(ub->__queues);
1987 }
1988
1989 static int ublk_init_queues(struct ublk_device *ub)
1990 {
1991         int nr_queues = ub->dev_info.nr_hw_queues;
1992         int depth = ub->dev_info.queue_depth;
1993         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1994         int i, ret = -ENOMEM;
1995
1996         ub->queue_size = ubq_size;
1997         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1998         if (!ub->__queues)
1999                 return ret;
2000
2001         for (i = 0; i < nr_queues; i++) {
2002                 if (ublk_init_queue(ub, i))
2003                         goto fail;
2004         }
2005
2006         init_completion(&ub->completion);
2007         return 0;
2008
2009  fail:
2010         ublk_deinit_queues(ub);
2011         return ret;
2012 }
2013
2014 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2015 {
2016         int i = idx;
2017         int err;
2018
2019         spin_lock(&ublk_idr_lock);
2020         /* allocate id, if @id >= 0, we're requesting that specific id */
2021         if (i >= 0) {
2022                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2023                 if (err == -ENOSPC)
2024                         err = -EEXIST;
2025         } else {
2026                 err = idr_alloc(&ublk_index_idr, ub, 0, UBLK_MAX_UBLKS,
2027                                 GFP_NOWAIT);
2028         }
2029         spin_unlock(&ublk_idr_lock);
2030
2031         if (err >= 0)
2032                 ub->ub_number = err;
2033
2034         return err;
2035 }
2036
2037 static void ublk_free_dev_number(struct ublk_device *ub)
2038 {
2039         spin_lock(&ublk_idr_lock);
2040         idr_remove(&ublk_index_idr, ub->ub_number);
2041         wake_up_all(&ublk_idr_wq);
2042         spin_unlock(&ublk_idr_lock);
2043 }
2044
2045 static void ublk_cdev_rel(struct device *dev)
2046 {
2047         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2048
2049         blk_mq_free_tag_set(&ub->tag_set);
2050         ublk_deinit_queues(ub);
2051         ublk_free_dev_number(ub);
2052         mutex_destroy(&ub->mutex);
2053         kfree(ub);
2054 }
2055
2056 static int ublk_add_chdev(struct ublk_device *ub)
2057 {
2058         struct device *dev = &ub->cdev_dev;
2059         int minor = ub->ub_number;
2060         int ret;
2061
2062         dev->parent = ublk_misc.this_device;
2063         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2064         dev->class = &ublk_chr_class;
2065         dev->release = ublk_cdev_rel;
2066         device_initialize(dev);
2067
2068         ret = dev_set_name(dev, "ublkc%d", minor);
2069         if (ret)
2070                 goto fail;
2071
2072         cdev_init(&ub->cdev, &ublk_ch_fops);
2073         ret = cdev_device_add(&ub->cdev, dev);
2074         if (ret)
2075                 goto fail;
2076
2077         ublks_added++;
2078         return 0;
2079  fail:
2080         put_device(dev);
2081         return ret;
2082 }
2083
2084 static void ublk_stop_work_fn(struct work_struct *work)
2085 {
2086         struct ublk_device *ub =
2087                 container_of(work, struct ublk_device, stop_work);
2088
2089         ublk_stop_dev(ub);
2090 }
2091
2092 /* align max io buffer size with PAGE_SIZE */
2093 static void ublk_align_max_io_size(struct ublk_device *ub)
2094 {
2095         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2096
2097         ub->dev_info.max_io_buf_bytes =
2098                 round_down(max_io_bytes, PAGE_SIZE);
2099 }
2100
2101 static int ublk_add_tag_set(struct ublk_device *ub)
2102 {
2103         ub->tag_set.ops = &ublk_mq_ops;
2104         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2105         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2106         ub->tag_set.numa_node = NUMA_NO_NODE;
2107         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2108         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2109         ub->tag_set.driver_data = ub;
2110         return blk_mq_alloc_tag_set(&ub->tag_set);
2111 }
2112
2113 static void ublk_remove(struct ublk_device *ub)
2114 {
2115         ublk_stop_dev(ub);
2116         cancel_work_sync(&ub->stop_work);
2117         cancel_work_sync(&ub->quiesce_work);
2118         cdev_device_del(&ub->cdev, &ub->cdev_dev);
2119         put_device(&ub->cdev_dev);
2120         ublks_added--;
2121 }
2122
2123 static struct ublk_device *ublk_get_device_from_id(int idx)
2124 {
2125         struct ublk_device *ub = NULL;
2126
2127         if (idx < 0)
2128                 return NULL;
2129
2130         spin_lock(&ublk_idr_lock);
2131         ub = idr_find(&ublk_index_idr, idx);
2132         if (ub)
2133                 ub = ublk_get_device(ub);
2134         spin_unlock(&ublk_idr_lock);
2135
2136         return ub;
2137 }
2138
2139 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2140 {
2141         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2142         int ublksrv_pid = (int)header->data[0];
2143         struct gendisk *disk;
2144         int ret = -EINVAL;
2145
2146         if (ublksrv_pid <= 0)
2147                 return -EINVAL;
2148
2149         if (wait_for_completion_interruptible(&ub->completion) != 0)
2150                 return -EINTR;
2151
2152         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2153
2154         mutex_lock(&ub->mutex);
2155         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2156             test_bit(UB_STATE_USED, &ub->state)) {
2157                 ret = -EEXIST;
2158                 goto out_unlock;
2159         }
2160
2161         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
2162         if (IS_ERR(disk)) {
2163                 ret = PTR_ERR(disk);
2164                 goto out_unlock;
2165         }
2166         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2167         disk->fops = &ub_fops;
2168         disk->private_data = ub;
2169
2170         ub->dev_info.ublksrv_pid = ublksrv_pid;
2171         ub->ub_disk = disk;
2172
2173         ret = ublk_apply_params(ub);
2174         if (ret)
2175                 goto out_put_disk;
2176
2177         /* don't probe partitions if any one ubq daemon is un-trusted */
2178         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2179                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2180
2181         get_device(&ub->cdev_dev);
2182         ub->dev_info.state = UBLK_S_DEV_LIVE;
2183
2184         if (ublk_dev_is_zoned(ub)) {
2185                 ret = ublk_revalidate_disk_zones(ub);
2186                 if (ret)
2187                         goto out_put_cdev;
2188         }
2189
2190         ret = add_disk(disk);
2191         if (ret)
2192                 goto out_put_cdev;
2193
2194         set_bit(UB_STATE_USED, &ub->state);
2195
2196 out_put_cdev:
2197         if (ret) {
2198                 ub->dev_info.state = UBLK_S_DEV_DEAD;
2199                 ublk_put_device(ub);
2200         }
2201 out_put_disk:
2202         if (ret)
2203                 put_disk(disk);
2204 out_unlock:
2205         mutex_unlock(&ub->mutex);
2206         return ret;
2207 }
2208
2209 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2210                 struct io_uring_cmd *cmd)
2211 {
2212         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2213         void __user *argp = (void __user *)(unsigned long)header->addr;
2214         cpumask_var_t cpumask;
2215         unsigned long queue;
2216         unsigned int retlen;
2217         unsigned int i;
2218         int ret;
2219
2220         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2221                 return -EINVAL;
2222         if (header->len & (sizeof(unsigned long)-1))
2223                 return -EINVAL;
2224         if (!header->addr)
2225                 return -EINVAL;
2226
2227         queue = header->data[0];
2228         if (queue >= ub->dev_info.nr_hw_queues)
2229                 return -EINVAL;
2230
2231         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2232                 return -ENOMEM;
2233
2234         for_each_possible_cpu(i) {
2235                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2236                         cpumask_set_cpu(i, cpumask);
2237         }
2238
2239         ret = -EFAULT;
2240         retlen = min_t(unsigned short, header->len, cpumask_size());
2241         if (copy_to_user(argp, cpumask, retlen))
2242                 goto out_free_cpumask;
2243         if (retlen != header->len &&
2244             clear_user(argp + retlen, header->len - retlen))
2245                 goto out_free_cpumask;
2246
2247         ret = 0;
2248 out_free_cpumask:
2249         free_cpumask_var(cpumask);
2250         return ret;
2251 }
2252
2253 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2254 {
2255         pr_devel("%s: dev id %d flags %llx\n", __func__,
2256                         info->dev_id, info->flags);
2257         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2258                         info->nr_hw_queues, info->queue_depth);
2259 }
2260
2261 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2262 {
2263         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2264         void __user *argp = (void __user *)(unsigned long)header->addr;
2265         struct ublksrv_ctrl_dev_info info;
2266         struct ublk_device *ub;
2267         int ret = -EINVAL;
2268
2269         if (header->len < sizeof(info) || !header->addr)
2270                 return -EINVAL;
2271         if (header->queue_id != (u16)-1) {
2272                 pr_warn("%s: queue_id is wrong %x\n",
2273                         __func__, header->queue_id);
2274                 return -EINVAL;
2275         }
2276
2277         if (copy_from_user(&info, argp, sizeof(info)))
2278                 return -EFAULT;
2279
2280         if (capable(CAP_SYS_ADMIN))
2281                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2282         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2283                 return -EPERM;
2284
2285         /*
2286          * unprivileged device can't be trusted, but RECOVERY and
2287          * RECOVERY_REISSUE still may hang error handling, so can't
2288          * support recovery features for unprivileged ublk now
2289          *
2290          * TODO: provide forward progress for RECOVERY handler, so that
2291          * unprivileged device can benefit from it
2292          */
2293         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2294                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2295                                 UBLK_F_USER_RECOVERY);
2296
2297         /* the created device is always owned by current user */
2298         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2299
2300         if (header->dev_id != info.dev_id) {
2301                 pr_warn("%s: dev id not match %u %u\n",
2302                         __func__, header->dev_id, info.dev_id);
2303                 return -EINVAL;
2304         }
2305
2306         if (header->dev_id != U32_MAX && header->dev_id >= UBLK_MAX_UBLKS) {
2307                 pr_warn("%s: dev id is too large. Max supported is %d\n",
2308                         __func__, UBLK_MAX_UBLKS - 1);
2309                 return -EINVAL;
2310         }
2311
2312         ublk_dump_dev_info(&info);
2313
2314         ret = mutex_lock_killable(&ublk_ctl_mutex);
2315         if (ret)
2316                 return ret;
2317
2318         ret = -EACCES;
2319         if (ublks_added >= ublks_max)
2320                 goto out_unlock;
2321
2322         ret = -ENOMEM;
2323         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2324         if (!ub)
2325                 goto out_unlock;
2326         mutex_init(&ub->mutex);
2327         spin_lock_init(&ub->mm_lock);
2328         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2329         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2330         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
2331
2332         ret = ublk_alloc_dev_number(ub, header->dev_id);
2333         if (ret < 0)
2334                 goto out_free_ub;
2335
2336         memcpy(&ub->dev_info, &info, sizeof(info));
2337
2338         /* update device id */
2339         ub->dev_info.dev_id = ub->ub_number;
2340
2341         /*
2342          * 64bit flags will be copied back to userspace as feature
2343          * negotiation result, so have to clear flags which driver
2344          * doesn't support yet, then userspace can get correct flags
2345          * (features) to handle.
2346          */
2347         ub->dev_info.flags &= UBLK_F_ALL;
2348
2349         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2350                 UBLK_F_URING_CMD_COMP_IN_TASK;
2351
2352         /* GET_DATA isn't needed any more with USER_COPY */
2353         if (ublk_dev_is_user_copy(ub))
2354                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2355
2356         /* Zoned storage support requires user copy feature */
2357         if (ublk_dev_is_zoned(ub) &&
2358             (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2359                 ret = -EINVAL;
2360                 goto out_free_dev_number;
2361         }
2362
2363         /* We are not ready to support zero copy */
2364         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2365
2366         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2367                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2368         ublk_align_max_io_size(ub);
2369
2370         ret = ublk_init_queues(ub);
2371         if (ret)
2372                 goto out_free_dev_number;
2373
2374         ret = ublk_add_tag_set(ub);
2375         if (ret)
2376                 goto out_deinit_queues;
2377
2378         ret = -EFAULT;
2379         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2380                 goto out_free_tag_set;
2381
2382         /*
2383          * Add the char dev so that ublksrv daemon can be setup.
2384          * ublk_add_chdev() will cleanup everything if it fails.
2385          */
2386         ret = ublk_add_chdev(ub);
2387         goto out_unlock;
2388
2389 out_free_tag_set:
2390         blk_mq_free_tag_set(&ub->tag_set);
2391 out_deinit_queues:
2392         ublk_deinit_queues(ub);
2393 out_free_dev_number:
2394         ublk_free_dev_number(ub);
2395 out_free_ub:
2396         mutex_destroy(&ub->mutex);
2397         kfree(ub);
2398 out_unlock:
2399         mutex_unlock(&ublk_ctl_mutex);
2400         return ret;
2401 }
2402
2403 static inline bool ublk_idr_freed(int id)
2404 {
2405         void *ptr;
2406
2407         spin_lock(&ublk_idr_lock);
2408         ptr = idr_find(&ublk_index_idr, id);
2409         spin_unlock(&ublk_idr_lock);
2410
2411         return ptr == NULL;
2412 }
2413
2414 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2415 {
2416         struct ublk_device *ub = *p_ub;
2417         int idx = ub->ub_number;
2418         int ret;
2419
2420         ret = mutex_lock_killable(&ublk_ctl_mutex);
2421         if (ret)
2422                 return ret;
2423
2424         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2425                 ublk_remove(ub);
2426                 set_bit(UB_STATE_DELETED, &ub->state);
2427         }
2428
2429         /* Mark the reference as consumed */
2430         *p_ub = NULL;
2431         ublk_put_device(ub);
2432         mutex_unlock(&ublk_ctl_mutex);
2433
2434         /*
2435          * Wait until the idr is removed, then it can be reused after
2436          * DEL_DEV command is returned.
2437          *
2438          * If we returns because of user interrupt, future delete command
2439          * may come:
2440          *
2441          * - the device number isn't freed, this device won't or needn't
2442          *   be deleted again, since UB_STATE_DELETED is set, and device
2443          *   will be released after the last reference is dropped
2444          *
2445          * - the device number is freed already, we will not find this
2446          *   device via ublk_get_device_from_id()
2447          */
2448         if (wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx)))
2449                 return -EINTR;
2450         return 0;
2451 }
2452
2453 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2454 {
2455         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2456
2457         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2458                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2459                         header->data[0], header->addr, header->len);
2460 }
2461
2462 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2463 {
2464         ublk_stop_dev(ub);
2465         cancel_work_sync(&ub->stop_work);
2466         cancel_work_sync(&ub->quiesce_work);
2467
2468         return 0;
2469 }
2470
2471 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2472                 struct io_uring_cmd *cmd)
2473 {
2474         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2475         void __user *argp = (void __user *)(unsigned long)header->addr;
2476
2477         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2478                 return -EINVAL;
2479
2480         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2481                 return -EFAULT;
2482
2483         return 0;
2484 }
2485
2486 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2487 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2488 {
2489         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2490         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2491
2492         if (ub->ub_disk) {
2493                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2494                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2495         } else {
2496                 ub->params.devt.disk_major = 0;
2497                 ub->params.devt.disk_minor = 0;
2498         }
2499         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2500 }
2501
2502 static int ublk_ctrl_get_params(struct ublk_device *ub,
2503                 struct io_uring_cmd *cmd)
2504 {
2505         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2506         void __user *argp = (void __user *)(unsigned long)header->addr;
2507         struct ublk_params_header ph;
2508         int ret;
2509
2510         if (header->len <= sizeof(ph) || !header->addr)
2511                 return -EINVAL;
2512
2513         if (copy_from_user(&ph, argp, sizeof(ph)))
2514                 return -EFAULT;
2515
2516         if (ph.len > header->len || !ph.len)
2517                 return -EINVAL;
2518
2519         if (ph.len > sizeof(struct ublk_params))
2520                 ph.len = sizeof(struct ublk_params);
2521
2522         mutex_lock(&ub->mutex);
2523         ublk_ctrl_fill_params_devt(ub);
2524         if (copy_to_user(argp, &ub->params, ph.len))
2525                 ret = -EFAULT;
2526         else
2527                 ret = 0;
2528         mutex_unlock(&ub->mutex);
2529
2530         return ret;
2531 }
2532
2533 static int ublk_ctrl_set_params(struct ublk_device *ub,
2534                 struct io_uring_cmd *cmd)
2535 {
2536         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2537         void __user *argp = (void __user *)(unsigned long)header->addr;
2538         struct ublk_params_header ph;
2539         int ret = -EFAULT;
2540
2541         if (header->len <= sizeof(ph) || !header->addr)
2542                 return -EINVAL;
2543
2544         if (copy_from_user(&ph, argp, sizeof(ph)))
2545                 return -EFAULT;
2546
2547         if (ph.len > header->len || !ph.len || !ph.types)
2548                 return -EINVAL;
2549
2550         if (ph.len > sizeof(struct ublk_params))
2551                 ph.len = sizeof(struct ublk_params);
2552
2553         /* parameters can only be changed when device isn't live */
2554         mutex_lock(&ub->mutex);
2555         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2556                 ret = -EACCES;
2557         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2558                 ret = -EFAULT;
2559         } else {
2560                 /* clear all we don't support yet */
2561                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2562                 ret = ublk_validate_params(ub);
2563                 if (ret)
2564                         ub->params.types = 0;
2565         }
2566         mutex_unlock(&ub->mutex);
2567
2568         return ret;
2569 }
2570
2571 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2572 {
2573         int i;
2574
2575         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2576         /* All old ioucmds have to be completed */
2577         WARN_ON_ONCE(ubq->nr_io_ready);
2578         /* old daemon is PF_EXITING, put it now */
2579         put_task_struct(ubq->ubq_daemon);
2580         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2581         ubq->ubq_daemon = NULL;
2582         ubq->timeout = false;
2583
2584         for (i = 0; i < ubq->q_depth; i++) {
2585                 struct ublk_io *io = &ubq->ios[i];
2586
2587                 /* forget everything now and be ready for new FETCH_REQ */
2588                 io->flags = 0;
2589                 io->cmd = NULL;
2590                 io->addr = 0;
2591         }
2592 }
2593
2594 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2595                 struct io_uring_cmd *cmd)
2596 {
2597         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2598         int ret = -EINVAL;
2599         int i;
2600
2601         mutex_lock(&ub->mutex);
2602         if (!ublk_can_use_recovery(ub))
2603                 goto out_unlock;
2604         /*
2605          * START_RECOVERY is only allowd after:
2606          *
2607          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2608          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2609          *     released.
2610          *
2611          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2612          *     (a)has quiesced request queue
2613          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2614          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2615          *     (d)has completed/camceled all ioucmds owned by ther dying process
2616          */
2617         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2618                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2619                 ret = -EBUSY;
2620                 goto out_unlock;
2621         }
2622         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2623         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2624                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2625         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2626         ub->mm = NULL;
2627         ub->nr_queues_ready = 0;
2628         ub->nr_privileged_daemon = 0;
2629         init_completion(&ub->completion);
2630         ret = 0;
2631  out_unlock:
2632         mutex_unlock(&ub->mutex);
2633         return ret;
2634 }
2635
2636 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2637                 struct io_uring_cmd *cmd)
2638 {
2639         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2640         int ublksrv_pid = (int)header->data[0];
2641         int ret = -EINVAL;
2642
2643         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2644                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2645         /* wait until new ubq_daemon sending all FETCH_REQ */
2646         if (wait_for_completion_interruptible(&ub->completion))
2647                 return -EINTR;
2648
2649         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2650                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2651
2652         mutex_lock(&ub->mutex);
2653         if (!ublk_can_use_recovery(ub))
2654                 goto out_unlock;
2655
2656         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2657                 ret = -EBUSY;
2658                 goto out_unlock;
2659         }
2660         ub->dev_info.ublksrv_pid = ublksrv_pid;
2661         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2662                         __func__, ublksrv_pid, header->dev_id);
2663         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2664         pr_devel("%s: queue unquiesced, dev id %d.\n",
2665                         __func__, header->dev_id);
2666         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2667         ub->dev_info.state = UBLK_S_DEV_LIVE;
2668         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2669         ret = 0;
2670  out_unlock:
2671         mutex_unlock(&ub->mutex);
2672         return ret;
2673 }
2674
2675 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2676 {
2677         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2678         void __user *argp = (void __user *)(unsigned long)header->addr;
2679         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2680
2681         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2682                 return -EINVAL;
2683
2684         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2685                 return -EFAULT;
2686
2687         return 0;
2688 }
2689
2690 /*
2691  * All control commands are sent via /dev/ublk-control, so we have to check
2692  * the destination device's permission
2693  */
2694 static int ublk_char_dev_permission(struct ublk_device *ub,
2695                 const char *dev_path, int mask)
2696 {
2697         int err;
2698         struct path path;
2699         struct kstat stat;
2700
2701         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2702         if (err)
2703                 return err;
2704
2705         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2706         if (err)
2707                 goto exit;
2708
2709         err = -EPERM;
2710         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2711                 goto exit;
2712
2713         err = inode_permission(&nop_mnt_idmap,
2714                         d_backing_inode(path.dentry), mask);
2715 exit:
2716         path_put(&path);
2717         return err;
2718 }
2719
2720 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2721                 struct io_uring_cmd *cmd)
2722 {
2723         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2724         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2725         void __user *argp = (void __user *)(unsigned long)header->addr;
2726         char *dev_path = NULL;
2727         int ret = 0;
2728         int mask;
2729
2730         if (!unprivileged) {
2731                 if (!capable(CAP_SYS_ADMIN))
2732                         return -EPERM;
2733                 /*
2734                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2735                  * char_dev_path in payload too, since userspace may not
2736                  * know if the specified device is created as unprivileged
2737                  * mode.
2738                  */
2739                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2740                         return 0;
2741         }
2742
2743         /*
2744          * User has to provide the char device path for unprivileged ublk
2745          *
2746          * header->addr always points to the dev path buffer, and
2747          * header->dev_path_len records length of dev path buffer.
2748          */
2749         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2750                 return -EINVAL;
2751
2752         if (header->len < header->dev_path_len)
2753                 return -EINVAL;
2754
2755         dev_path = memdup_user_nul(argp, header->dev_path_len);
2756         if (IS_ERR(dev_path))
2757                 return PTR_ERR(dev_path);
2758
2759         ret = -EINVAL;
2760         switch (_IOC_NR(cmd->cmd_op)) {
2761         case UBLK_CMD_GET_DEV_INFO:
2762         case UBLK_CMD_GET_DEV_INFO2:
2763         case UBLK_CMD_GET_QUEUE_AFFINITY:
2764         case UBLK_CMD_GET_PARAMS:
2765         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2766                 mask = MAY_READ;
2767                 break;
2768         case UBLK_CMD_START_DEV:
2769         case UBLK_CMD_STOP_DEV:
2770         case UBLK_CMD_ADD_DEV:
2771         case UBLK_CMD_DEL_DEV:
2772         case UBLK_CMD_SET_PARAMS:
2773         case UBLK_CMD_START_USER_RECOVERY:
2774         case UBLK_CMD_END_USER_RECOVERY:
2775                 mask = MAY_READ | MAY_WRITE;
2776                 break;
2777         default:
2778                 goto exit;
2779         }
2780
2781         ret = ublk_char_dev_permission(ub, dev_path, mask);
2782         if (!ret) {
2783                 header->len -= header->dev_path_len;
2784                 header->addr += header->dev_path_len;
2785         }
2786         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2787                         __func__, ub->ub_number, cmd->cmd_op,
2788                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2789                         dev_path, ret);
2790 exit:
2791         kfree(dev_path);
2792         return ret;
2793 }
2794
2795 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2796                 unsigned int issue_flags)
2797 {
2798         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2799         struct ublk_device *ub = NULL;
2800         u32 cmd_op = cmd->cmd_op;
2801         int ret = -EINVAL;
2802
2803         if (issue_flags & IO_URING_F_NONBLOCK)
2804                 return -EAGAIN;
2805
2806         ublk_ctrl_cmd_dump(cmd);
2807
2808         if (!(issue_flags & IO_URING_F_SQE128))
2809                 goto out;
2810
2811         ret = ublk_check_cmd_op(cmd_op);
2812         if (ret)
2813                 goto out;
2814
2815         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2816                 ret = ublk_ctrl_get_features(cmd);
2817                 goto out;
2818         }
2819
2820         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2821                 ret = -ENODEV;
2822                 ub = ublk_get_device_from_id(header->dev_id);
2823                 if (!ub)
2824                         goto out;
2825
2826                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2827                 if (ret)
2828                         goto put_dev;
2829         }
2830
2831         switch (_IOC_NR(cmd_op)) {
2832         case UBLK_CMD_START_DEV:
2833                 ret = ublk_ctrl_start_dev(ub, cmd);
2834                 break;
2835         case UBLK_CMD_STOP_DEV:
2836                 ret = ublk_ctrl_stop_dev(ub);
2837                 break;
2838         case UBLK_CMD_GET_DEV_INFO:
2839         case UBLK_CMD_GET_DEV_INFO2:
2840                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2841                 break;
2842         case UBLK_CMD_ADD_DEV:
2843                 ret = ublk_ctrl_add_dev(cmd);
2844                 break;
2845         case UBLK_CMD_DEL_DEV:
2846                 ret = ublk_ctrl_del_dev(&ub);
2847                 break;
2848         case UBLK_CMD_GET_QUEUE_AFFINITY:
2849                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2850                 break;
2851         case UBLK_CMD_GET_PARAMS:
2852                 ret = ublk_ctrl_get_params(ub, cmd);
2853                 break;
2854         case UBLK_CMD_SET_PARAMS:
2855                 ret = ublk_ctrl_set_params(ub, cmd);
2856                 break;
2857         case UBLK_CMD_START_USER_RECOVERY:
2858                 ret = ublk_ctrl_start_recovery(ub, cmd);
2859                 break;
2860         case UBLK_CMD_END_USER_RECOVERY:
2861                 ret = ublk_ctrl_end_recovery(ub, cmd);
2862                 break;
2863         default:
2864                 ret = -ENOTSUPP;
2865                 break;
2866         }
2867
2868  put_dev:
2869         if (ub)
2870                 ublk_put_device(ub);
2871  out:
2872         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2873         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2874                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2875         return -EIOCBQUEUED;
2876 }
2877
2878 static const struct file_operations ublk_ctl_fops = {
2879         .open           = nonseekable_open,
2880         .uring_cmd      = ublk_ctrl_uring_cmd,
2881         .owner          = THIS_MODULE,
2882         .llseek         = noop_llseek,
2883 };
2884
2885 static struct miscdevice ublk_misc = {
2886         .minor          = MISC_DYNAMIC_MINOR,
2887         .name           = "ublk-control",
2888         .fops           = &ublk_ctl_fops,
2889 };
2890
2891 static int __init ublk_init(void)
2892 {
2893         int ret;
2894
2895         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2896                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2897
2898         init_waitqueue_head(&ublk_idr_wq);
2899
2900         ret = misc_register(&ublk_misc);
2901         if (ret)
2902                 return ret;
2903
2904         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2905         if (ret)
2906                 goto unregister_mis;
2907
2908         ret = class_register(&ublk_chr_class);
2909         if (ret)
2910                 goto free_chrdev_region;
2911
2912         return 0;
2913
2914 free_chrdev_region:
2915         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2916 unregister_mis:
2917         misc_deregister(&ublk_misc);
2918         return ret;
2919 }
2920
2921 static void __exit ublk_exit(void)
2922 {
2923         struct ublk_device *ub;
2924         int id;
2925
2926         idr_for_each_entry(&ublk_index_idr, ub, id)
2927                 ublk_remove(ub);
2928
2929         class_unregister(&ublk_chr_class);
2930         misc_deregister(&ublk_misc);
2931
2932         idr_destroy(&ublk_index_idr);
2933         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2934 }
2935
2936 module_init(ublk_init);
2937 module_exit(ublk_exit);
2938
2939 static int ublk_set_max_ublks(const char *buf, const struct kernel_param *kp)
2940 {
2941         return param_set_uint_minmax(buf, kp, 0, UBLK_MAX_UBLKS);
2942 }
2943
2944 static int ublk_get_max_ublks(char *buf, const struct kernel_param *kp)
2945 {
2946         return sysfs_emit(buf, "%u\n", ublks_max);
2947 }
2948
2949 static const struct kernel_param_ops ublk_max_ublks_ops = {
2950         .set = ublk_set_max_ublks,
2951         .get = ublk_get_max_ublks,
2952 };
2953
2954 module_param_cb(ublks_max, &ublk_max_ublks_ops, &ublks_max, 0644);
2955 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
2956
2957 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2958 MODULE_LICENSE("GPL");