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