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 <[email protected]>
   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/cmd.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 /* private ioctl command mirror */
  52 #define UBLK_CMD_DEL_DEV_ASYNC  _IOC_NR(UBLK_U_CMD_DEL_DEV_ASYNC)
  53
  54 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  55 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  56                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  57                 | UBLK_F_NEED_GET_DATA \
  58                 | UBLK_F_USER_RECOVERY \
  59                 | UBLK_F_USER_RECOVERY_REISSUE \
  60                 | UBLK_F_UNPRIVILEGED_DEV \
  61                 | UBLK_F_CMD_IOCTL_ENCODE \
  62                 | UBLK_F_USER_COPY \
  63                 | UBLK_F_ZONED)
  64
  65 /* All UBLK_PARAM_TYPE_* should be included here */
  66 #define UBLK_PARAM_TYPE_ALL                                \
  67         (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD | \
  68          UBLK_PARAM_TYPE_DEVT | UBLK_PARAM_TYPE_ZONED)
  69
  70 struct ublk_rq_data {
  71         struct llist_node node;
  72
  73         struct kref ref;
  74 };
  75
  76 struct ublk_uring_cmd_pdu {
  77         struct ublk_queue *ubq;
  78         u16 tag;
  79 };
  80
  81 /*
  82  * io command is active: sqe cmd is received, and its cqe isn't done
  83  *
  84  * If the flag is set, the io command is owned by ublk driver, and waited
  85  * for incoming blk-mq request from the ublk block device.
  86  *
  87  * If the flag is cleared, the io command will be completed, and owned by
  88  * ublk server.
  89  */
  90 #define UBLK_IO_FLAG_ACTIVE     0x01
  91
  92 /*
  93  * IO command is completed via cqe, and it is being handled by ublksrv, and
  94  * not committed yet
  95  *
  96  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  97  * cross verification
  98  */
  99 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
 100
 101 /*
 102  * IO command is aborted, so this flag is set in case of
 103  * !UBLK_IO_FLAG_ACTIVE.
 104  *
 105  * After this flag is observed, any pending or new incoming request
 106  * associated with this io command will be failed immediately
 107  */
 108 #define UBLK_IO_FLAG_ABORTED 0x04
 109
 110 /*
 111  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
 112  * get data buffer address from ublksrv.
 113  *
 114  * Then, bio data could be copied into this data buffer for a WRITE request
 115  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 116  */
 117 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 118
 119 /* atomic RW with ubq->cancel_lock */
 120 #define UBLK_IO_FLAG_CANCELED   0x80000000
 121
 122 struct ublk_io {
 123         /* userspace buffer address from io cmd */
 124         __u64   addr;
 125         unsigned int flags;
 126         int res;
 127
 128         struct io_uring_cmd *cmd;
 129 };
 130
 131 struct ublk_queue {
 132         int q_id;
 133         int q_depth;
 134
 135         unsigned long flags;
 136         struct task_struct      *ubq_daemon;
 137         char *io_cmd_buf;
 138
 139         struct llist_head       io_cmds;
 140
 141         unsigned long io_addr;  /* mapped vm address */
 142         unsigned int max_io_sz;
 143         bool force_abort;
 144         bool timeout;
 145         bool canceling;
 146         unsigned short nr_io_ready;     /* how many ios setup */
 147         spinlock_t              cancel_lock;
 148         struct ublk_device *dev;
 149         struct ublk_io ios[];
 150 };
 151
 152 struct ublk_device {
 153         struct gendisk          *ub_disk;
 154
 155         char    *__queues;
 156
 157         unsigned int    queue_size;
 158         struct ublksrv_ctrl_dev_info    dev_info;
 159
 160         struct blk_mq_tag_set   tag_set;
 161
 162         struct cdev             cdev;
 163         struct device           cdev_dev;
 164
 165 #define UB_STATE_OPEN           0
 166 #define UB_STATE_USED           1
 167 #define UB_STATE_DELETED        2
 168         unsigned long           state;
 169         int                     ub_number;
 170
 171         struct mutex            mutex;
 172
 173         spinlock_t              lock;
 174         struct mm_struct        *mm;
 175
 176         struct ublk_params      params;
 177
 178         struct completion       completion;
 179         unsigned int            nr_queues_ready;
 180         unsigned int            nr_privileged_daemon;
 181
 182         struct work_struct      quiesce_work;
 183         struct work_struct      stop_work;
 184 };
 185
 186 /* header of ublk_params */
 187 struct ublk_params_header {
 188         __u32   len;
 189         __u32   types;
 190 };
 191
 192 static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq);
 193
 194 static inline unsigned int ublk_req_build_flags(struct request *req);
 195 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 196                                                    int tag);
 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 struct ublk_zoned_report_desc {
 215         __u64 sector;
 216         __u32 operation;
 217         __u32 nr_zones;
 218 };
 219
 220 static DEFINE_XARRAY(ublk_zoned_report_descs);
 221
 222 static int ublk_zoned_insert_report_desc(const struct request *req,
 223                 struct ublk_zoned_report_desc *desc)
 224 {
 225         return xa_insert(&ublk_zoned_report_descs, (unsigned long)req,
 226                             desc, GFP_KERNEL);
 227 }
 228
 229 static struct ublk_zoned_report_desc *ublk_zoned_erase_report_desc(
 230                 const struct request *req)
 231 {
 232         return xa_erase(&ublk_zoned_report_descs, (unsigned long)req);
 233 }
 234
 235 static struct ublk_zoned_report_desc *ublk_zoned_get_report_desc(
 236                 const struct request *req)
 237 {
 238         return xa_load(&ublk_zoned_report_descs, (unsigned long)req);
 239 }
 240
 241 static int ublk_get_nr_zones(const struct ublk_device *ub)
 242 {
 243         const struct ublk_param_basic *p = &ub->params.basic;
 244
 245         /* Zone size is a power of 2 */
 246         return p->dev_sectors >> ilog2(p->chunk_sectors);
 247 }
 248
 249 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 250 {
 251         return blk_revalidate_disk_zones(ub->ub_disk);
 252 }
 253
 254 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 255 {
 256         const struct ublk_param_zoned *p = &ub->params.zoned;
 257         int nr_zones;
 258
 259         if (!ublk_dev_is_zoned(ub))
 260                 return -EINVAL;
 261
 262         if (!p->max_zone_append_sectors)
 263                 return -EINVAL;
 264
 265         nr_zones = ublk_get_nr_zones(ub);
 266
 267         if (p->max_active_zones > nr_zones)
 268                 return -EINVAL;
 269
 270         if (p->max_open_zones > nr_zones)
 271                 return -EINVAL;
 272
 273         return 0;
 274 }
 275
 276 static void ublk_dev_param_zoned_apply(struct ublk_device *ub)
 277 {
 278         ub->ub_disk->nr_zones = ublk_get_nr_zones(ub);
 279 }
 280
 281 /* Based on virtblk_alloc_report_buffer */
 282 static void *ublk_alloc_report_buffer(struct ublk_device *ublk,
 283                                       unsigned int nr_zones, size_t *buflen)
 284 {
 285         struct request_queue *q = ublk->ub_disk->queue;
 286         size_t bufsize;
 287         void *buf;
 288
 289         nr_zones = min_t(unsigned int, nr_zones,
 290                          ublk->ub_disk->nr_zones);
 291
 292         bufsize = nr_zones * sizeof(struct blk_zone);
 293         bufsize =
 294                 min_t(size_t, bufsize, queue_max_hw_sectors(q) << SECTOR_SHIFT);
 295
 296         while (bufsize >= sizeof(struct blk_zone)) {
 297                 buf = kvmalloc(bufsize, GFP_KERNEL | __GFP_NORETRY);
 298                 if (buf) {
 299                         *buflen = bufsize;
 300                         return buf;
 301                 }
 302                 bufsize >>= 1;
 303         }
 304
 305         *buflen = 0;
 306         return NULL;
 307 }
 308
 309 static int ublk_report_zones(struct gendisk *disk, sector_t sector,
 310                       unsigned int nr_zones, report_zones_cb cb, void *data)
 311 {
 312         struct ublk_device *ub = disk->private_data;
 313         unsigned int zone_size_sectors = disk->queue->limits.chunk_sectors;
 314         unsigned int first_zone = sector >> ilog2(zone_size_sectors);
 315         unsigned int done_zones = 0;
 316         unsigned int max_zones_per_request;
 317         int ret;
 318         struct blk_zone *buffer;
 319         size_t buffer_length;
 320
 321         nr_zones = min_t(unsigned int, ub->ub_disk->nr_zones - first_zone,
 322                          nr_zones);
 323
 324         buffer = ublk_alloc_report_buffer(ub, nr_zones, &buffer_length);
 325         if (!buffer)
 326                 return -ENOMEM;
 327
 328         max_zones_per_request = buffer_length / sizeof(struct blk_zone);
 329
 330         while (done_zones < nr_zones) {
 331                 unsigned int remaining_zones = nr_zones - done_zones;
 332                 unsigned int zones_in_request =
 333                         min_t(unsigned int, remaining_zones, max_zones_per_request);
 334                 struct request *req;
 335                 struct ublk_zoned_report_desc desc;
 336                 blk_status_t status;
 337
 338                 memset(buffer, 0, buffer_length);
 339
 340                 req = blk_mq_alloc_request(disk->queue, REQ_OP_DRV_IN, 0);
 341                 if (IS_ERR(req)) {
 342                         ret = PTR_ERR(req);
 343                         goto out;
 344                 }
 345
 346                 desc.operation = UBLK_IO_OP_REPORT_ZONES;
 347                 desc.sector = sector;
 348                 desc.nr_zones = zones_in_request;
 349                 ret = ublk_zoned_insert_report_desc(req, &desc);
 350                 if (ret)
 351                         goto free_req;
 352
 353                 ret = blk_rq_map_kern(disk->queue, req, buffer, buffer_length,
 354                                         GFP_KERNEL);
 355                 if (ret)
 356                         goto erase_desc;
 357
 358                 status = blk_execute_rq(req, 0);
 359                 ret = blk_status_to_errno(status);
 360 erase_desc:
 361                 ublk_zoned_erase_report_desc(req);
 362 free_req:
 363                 blk_mq_free_request(req);
 364                 if (ret)
 365                         goto out;
 366
 367                 for (unsigned int i = 0; i < zones_in_request; i++) {
 368                         struct blk_zone *zone = buffer + i;
 369
 370                         /* A zero length zone means no more zones in this response */
 371                         if (!zone->len)
 372                                 break;
 373
 374                         ret = cb(zone, i, data);
 375                         if (ret)
 376                                 goto out;
 377
 378                         done_zones++;
 379                         sector += zone_size_sectors;
 380
 381                 }
 382         }
 383
 384         ret = done_zones;
 385
 386 out:
 387         kvfree(buffer);
 388         return ret;
 389 }
 390
 391 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 392                                          struct request *req)
 393 {
 394         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 395         struct ublk_io *io = &ubq->ios[req->tag];
 396         struct ublk_zoned_report_desc *desc;
 397         u32 ublk_op;
 398
 399         switch (req_op(req)) {
 400         case REQ_OP_ZONE_OPEN:
 401                 ublk_op = UBLK_IO_OP_ZONE_OPEN;
 402                 break;
 403         case REQ_OP_ZONE_CLOSE:
 404                 ublk_op = UBLK_IO_OP_ZONE_CLOSE;
 405                 break;
 406         case REQ_OP_ZONE_FINISH:
 407                 ublk_op = UBLK_IO_OP_ZONE_FINISH;
 408                 break;
 409         case REQ_OP_ZONE_RESET:
 410                 ublk_op = UBLK_IO_OP_ZONE_RESET;
 411                 break;
 412         case REQ_OP_ZONE_APPEND:
 413                 ublk_op = UBLK_IO_OP_ZONE_APPEND;
 414                 break;
 415         case REQ_OP_ZONE_RESET_ALL:
 416                 ublk_op = UBLK_IO_OP_ZONE_RESET_ALL;
 417                 break;
 418         case REQ_OP_DRV_IN:
 419                 desc = ublk_zoned_get_report_desc(req);
 420                 if (!desc)
 421                         return BLK_STS_IOERR;
 422                 ublk_op = desc->operation;
 423                 switch (ublk_op) {
 424                 case UBLK_IO_OP_REPORT_ZONES:
 425                         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 426                         iod->nr_zones = desc->nr_zones;
 427                         iod->start_sector = desc->sector;
 428                         return BLK_STS_OK;
 429                 default:
 430                         return BLK_STS_IOERR;
 431                 }
 432         case REQ_OP_DRV_OUT:
 433                 /* We do not support drv_out */
 434                 return BLK_STS_NOTSUPP;
 435         default:
 436                 return BLK_STS_IOERR;
 437         }
 438
 439         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 440         iod->nr_sectors = blk_rq_sectors(req);
 441         iod->start_sector = blk_rq_pos(req);
 442         iod->addr = io->addr;
 443
 444         return BLK_STS_OK;
 445 }
 446
 447 #else
 448
 449 #define ublk_report_zones (NULL)
 450
 451 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 452 {
 453         return -EOPNOTSUPP;
 454 }
 455
 456 static void ublk_dev_param_zoned_apply(struct ublk_device *ub)
 457 {
 458 }
 459
 460 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 461 {
 462         return 0;
 463 }
 464
 465 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 466                                          struct request *req)
 467 {
 468         return BLK_STS_NOTSUPP;
 469 }
 470
 471 #endif
 472
 473 static inline void __ublk_complete_rq(struct request *req);
 474 static void ublk_complete_rq(struct kref *ref);
 475
 476 static dev_t ublk_chr_devt;
 477 static const struct class ublk_chr_class = {
 478         .name = "ublk-char",
 479 };
 480
 481 static DEFINE_IDR(ublk_index_idr);
 482 static DEFINE_SPINLOCK(ublk_idr_lock);
 483 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 484
 485 static DEFINE_MUTEX(ublk_ctl_mutex);
 486
 487 /*
 488  * Max ublk devices allowed to add
 489  *
 490  * It can be extended to one per-user limit in future or even controlled
 491  * by cgroup.
 492  */
 493 #define UBLK_MAX_UBLKS UBLK_MINORS
 494 static unsigned int ublks_max = 64;
 495 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 496
 497 static struct miscdevice ublk_misc;
 498
 499 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 500 {
 501         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 502                 UBLK_QID_BITS_MASK;
 503 }
 504
 505 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 506 {
 507         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 508 }
 509
 510 static inline unsigned ublk_pos_to_tag(loff_t pos)
 511 {
 512         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 513                 UBLK_TAG_BITS_MASK;
 514 }
 515
 516 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 517 {
 518         const struct ublk_param_basic *p = &ub->params.basic;
 519
 520         if (p->attrs & UBLK_ATTR_READ_ONLY)
 521                 set_disk_ro(ub->ub_disk, true);
 522
 523         set_capacity(ub->ub_disk, p->dev_sectors);
 524 }
 525
 526 static int ublk_validate_params(const struct ublk_device *ub)
 527 {
 528         /* basic param is the only one which must be set */
 529         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 530                 const struct ublk_param_basic *p = &ub->params.basic;
 531
 532                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 533                         return -EINVAL;
 534
 535                 if (p->logical_bs_shift > p->physical_bs_shift)
 536                         return -EINVAL;
 537
 538                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 539                         return -EINVAL;
 540
 541                 if (ublk_dev_is_zoned(ub) && !p->chunk_sectors)
 542                         return -EINVAL;
 543         } else
 544                 return -EINVAL;
 545
 546         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 547                 const struct ublk_param_discard *p = &ub->params.discard;
 548
 549                 /* So far, only support single segment discard */
 550                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 551                         return -EINVAL;
 552
 553                 if (!p->discard_granularity)
 554                         return -EINVAL;
 555         }
 556
 557         /* dev_t is read-only */
 558         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 559                 return -EINVAL;
 560
 561         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 562                 return ublk_dev_param_zoned_validate(ub);
 563         else if (ublk_dev_is_zoned(ub))
 564                 return -EINVAL;
 565
 566         return 0;
 567 }
 568
 569 static void ublk_apply_params(struct ublk_device *ub)
 570 {
 571         ublk_dev_param_basic_apply(ub);
 572
 573         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 574                 ublk_dev_param_zoned_apply(ub);
 575 }
 576
 577 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 578 {
 579         return ubq->flags & UBLK_F_USER_COPY;
 580 }
 581
 582 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 583 {
 584         /*
 585          * read()/write() is involved in user copy, so request reference
 586          * has to be grabbed
 587          */
 588         return ublk_support_user_copy(ubq);
 589 }
 590
 591 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 592                 struct request *req)
 593 {
 594         if (ublk_need_req_ref(ubq)) {
 595                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 596
 597                 kref_init(&data->ref);
 598         }
 599 }
 600
 601 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 602                 struct request *req)
 603 {
 604         if (ublk_need_req_ref(ubq)) {
 605                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 606
 607                 return kref_get_unless_zero(&data->ref);
 608         }
 609
 610         return true;
 611 }
 612
 613 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 614                 struct request *req)
 615 {
 616         if (ublk_need_req_ref(ubq)) {
 617                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 618
 619                 kref_put(&data->ref, ublk_complete_rq);
 620         } else {
 621                 __ublk_complete_rq(req);
 622         }
 623 }
 624
 625 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 626 {
 627         return ubq->flags & UBLK_F_NEED_GET_DATA;
 628 }
 629
 630 /* Called in slow path only, keep it noinline for trace purpose */
 631 static noinline struct ublk_device *ublk_get_device(struct ublk_device *ub)
 632 {
 633         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 634                 return ub;
 635         return NULL;
 636 }
 637
 638 /* Called in slow path only, keep it noinline for trace purpose */
 639 static noinline void ublk_put_device(struct ublk_device *ub)
 640 {
 641         put_device(&ub->cdev_dev);
 642 }
 643
 644 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 645                 int qid)
 646 {
 647        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 648 }
 649
 650 static inline bool ublk_rq_has_data(const struct request *rq)
 651 {
 652         return bio_has_data(rq->bio);
 653 }
 654
 655 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 656                 int tag)
 657 {
 658         return (struct ublksrv_io_desc *)
 659                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 660 }
 661
 662 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 663 {
 664         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 665 }
 666
 667 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 668 {
 669         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 670
 671         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 672                         PAGE_SIZE);
 673 }
 674
 675 static inline bool ublk_queue_can_use_recovery_reissue(
 676                 struct ublk_queue *ubq)
 677 {
 678         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 679                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 680 }
 681
 682 static inline bool ublk_queue_can_use_recovery(
 683                 struct ublk_queue *ubq)
 684 {
 685         return ubq->flags & UBLK_F_USER_RECOVERY;
 686 }
 687
 688 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 689 {
 690         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 691 }
 692
 693 static void ublk_free_disk(struct gendisk *disk)
 694 {
 695         struct ublk_device *ub = disk->private_data;
 696
 697         clear_bit(UB_STATE_USED, &ub->state);
 698         ublk_put_device(ub);
 699 }
 700
 701 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 702                 unsigned int *owner_gid)
 703 {
 704         kuid_t uid;
 705         kgid_t gid;
 706
 707         current_uid_gid(&uid, &gid);
 708
 709         *owner_uid = from_kuid(&init_user_ns, uid);
 710         *owner_gid = from_kgid(&init_user_ns, gid);
 711 }
 712
 713 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 714 {
 715         struct ublk_device *ub = disk->private_data;
 716
 717         if (capable(CAP_SYS_ADMIN))
 718                 return 0;
 719
 720         /*
 721          * If it is one unprivileged device, only owner can open
 722          * the disk. Otherwise it could be one trap made by one
 723          * evil user who grants this disk's privileges to other
 724          * users deliberately.
 725          *
 726          * This way is reasonable too given anyone can create
 727          * unprivileged device, and no need other's grant.
 728          */
 729         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 730                 unsigned int curr_uid, curr_gid;
 731
 732                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 733
 734                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 735                                 ub->dev_info.owner_gid)
 736                         return -EPERM;
 737         }
 738
 739         return 0;
 740 }
 741
 742 static const struct block_device_operations ub_fops = {
 743         .owner =        THIS_MODULE,
 744         .open =         ublk_open,
 745         .free_disk =    ublk_free_disk,
 746         .report_zones = ublk_report_zones,
 747 };
 748
 749 #define UBLK_MAX_PIN_PAGES      32
 750
 751 struct ublk_io_iter {
 752         struct page *pages[UBLK_MAX_PIN_PAGES];
 753         struct bio *bio;
 754         struct bvec_iter iter;
 755 };
 756
 757 /* return how many pages are copied */
 758 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 759                 size_t total, size_t pg_off, int dir)
 760 {
 761         unsigned done = 0;
 762         unsigned pg_idx = 0;
 763
 764         while (done < total) {
 765                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 766                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 767                                 (unsigned)(PAGE_SIZE - pg_off));
 768                 void *bv_buf = bvec_kmap_local(&bv);
 769                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 770
 771                 if (dir == ITER_DEST)
 772                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 773                 else
 774                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 775
 776                 kunmap_local(pg_buf);
 777                 kunmap_local(bv_buf);
 778
 779                 /* advance page array */
 780                 pg_off += bytes;
 781                 if (pg_off == PAGE_SIZE) {
 782                         pg_idx += 1;
 783                         pg_off = 0;
 784                 }
 785
 786                 done += bytes;
 787
 788                 /* advance bio */
 789                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 790                 if (!data->iter.bi_size) {
 791                         data->bio = data->bio->bi_next;
 792                         if (data->bio == NULL)
 793                                 break;
 794                         data->iter = data->bio->bi_iter;
 795                 }
 796         }
 797 }
 798
 799 static bool ublk_advance_io_iter(const struct request *req,
 800                 struct ublk_io_iter *iter, unsigned int offset)
 801 {
 802         struct bio *bio = req->bio;
 803
 804         for_each_bio(bio) {
 805                 if (bio->bi_iter.bi_size > offset) {
 806                         iter->bio = bio;
 807                         iter->iter = bio->bi_iter;
 808                         bio_advance_iter(iter->bio, &iter->iter, offset);
 809                         return true;
 810                 }
 811                 offset -= bio->bi_iter.bi_size;
 812         }
 813         return false;
 814 }
 815
 816 /*
 817  * Copy data between request pages and io_iter, and 'offset'
 818  * is the start point of linear offset of request.
 819  */
 820 static size_t ublk_copy_user_pages(const struct request *req,
 821                 unsigned offset, struct iov_iter *uiter, int dir)
 822 {
 823         struct ublk_io_iter iter;
 824         size_t done = 0;
 825
 826         if (!ublk_advance_io_iter(req, &iter, offset))
 827                 return 0;
 828
 829         while (iov_iter_count(uiter) && iter.bio) {
 830                 unsigned nr_pages;
 831                 ssize_t len;
 832                 size_t off;
 833                 int i;
 834
 835                 len = iov_iter_get_pages2(uiter, iter.pages,
 836                                 iov_iter_count(uiter),
 837                                 UBLK_MAX_PIN_PAGES, &off);
 838                 if (len <= 0)
 839                         return done;
 840
 841                 ublk_copy_io_pages(&iter, len, off, dir);
 842                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 843                 for (i = 0; i < nr_pages; i++) {
 844                         if (dir == ITER_DEST)
 845                                 set_page_dirty(iter.pages[i]);
 846                         put_page(iter.pages[i]);
 847                 }
 848                 done += len;
 849         }
 850
 851         return done;
 852 }
 853
 854 static inline bool ublk_need_map_req(const struct request *req)
 855 {
 856         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 857 }
 858
 859 static inline bool ublk_need_unmap_req(const struct request *req)
 860 {
 861         return ublk_rq_has_data(req) &&
 862                (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN);
 863 }
 864
 865 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 866                 struct ublk_io *io)
 867 {
 868         const unsigned int rq_bytes = blk_rq_bytes(req);
 869
 870         if (ublk_support_user_copy(ubq))
 871                 return rq_bytes;
 872
 873         /*
 874          * no zero copy, we delay copy WRITE request data into ublksrv
 875          * context and the big benefit is that pinning pages in current
 876          * context is pretty fast, see ublk_pin_user_pages
 877          */
 878         if (ublk_need_map_req(req)) {
 879                 struct iov_iter iter;
 880                 const int dir = ITER_DEST;
 881
 882                 import_ubuf(dir, u64_to_user_ptr(io->addr), rq_bytes, &iter);
 883                 return ublk_copy_user_pages(req, 0, &iter, dir);
 884         }
 885         return rq_bytes;
 886 }
 887
 888 static int ublk_unmap_io(const struct ublk_queue *ubq,
 889                 const struct request *req,
 890                 struct ublk_io *io)
 891 {
 892         const unsigned int rq_bytes = blk_rq_bytes(req);
 893
 894         if (ublk_support_user_copy(ubq))
 895                 return rq_bytes;
 896
 897         if (ublk_need_unmap_req(req)) {
 898                 struct iov_iter iter;
 899                 const int dir = ITER_SOURCE;
 900
 901                 WARN_ON_ONCE(io->res > rq_bytes);
 902
 903                 import_ubuf(dir, u64_to_user_ptr(io->addr), io->res, &iter);
 904                 return ublk_copy_user_pages(req, 0, &iter, dir);
 905         }
 906         return rq_bytes;
 907 }
 908
 909 static inline unsigned int ublk_req_build_flags(struct request *req)
 910 {
 911         unsigned flags = 0;
 912
 913         if (req->cmd_flags & REQ_FAILFAST_DEV)
 914                 flags |= UBLK_IO_F_FAILFAST_DEV;
 915
 916         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 917                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 918
 919         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 920                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 921
 922         if (req->cmd_flags & REQ_META)
 923                 flags |= UBLK_IO_F_META;
 924
 925         if (req->cmd_flags & REQ_FUA)
 926                 flags |= UBLK_IO_F_FUA;
 927
 928         if (req->cmd_flags & REQ_NOUNMAP)
 929                 flags |= UBLK_IO_F_NOUNMAP;
 930
 931         if (req->cmd_flags & REQ_SWAP)
 932                 flags |= UBLK_IO_F_SWAP;
 933
 934         return flags;
 935 }
 936
 937 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 938 {
 939         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 940         struct ublk_io *io = &ubq->ios[req->tag];
 941         enum req_op op = req_op(req);
 942         u32 ublk_op;
 943
 944         if (!ublk_queue_is_zoned(ubq) &&
 945             (op_is_zone_mgmt(op) || op == REQ_OP_ZONE_APPEND))
 946                 return BLK_STS_IOERR;
 947
 948         switch (req_op(req)) {
 949         case REQ_OP_READ:
 950                 ublk_op = UBLK_IO_OP_READ;
 951                 break;
 952         case REQ_OP_WRITE:
 953                 ublk_op = UBLK_IO_OP_WRITE;
 954                 break;
 955         case REQ_OP_FLUSH:
 956                 ublk_op = UBLK_IO_OP_FLUSH;
 957                 break;
 958         case REQ_OP_DISCARD:
 959                 ublk_op = UBLK_IO_OP_DISCARD;
 960                 break;
 961         case REQ_OP_WRITE_ZEROES:
 962                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 963                 break;
 964         default:
 965                 if (ublk_queue_is_zoned(ubq))
 966                         return ublk_setup_iod_zoned(ubq, req);
 967                 return BLK_STS_IOERR;
 968         }
 969
 970         /* need to translate since kernel may change */
 971         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 972         iod->nr_sectors = blk_rq_sectors(req);
 973         iod->start_sector = blk_rq_pos(req);
 974         iod->addr = io->addr;
 975
 976         return BLK_STS_OK;
 977 }
 978
 979 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 980                 struct io_uring_cmd *ioucmd)
 981 {
 982         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 983 }
 984
 985 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 986 {
 987         return ubq->ubq_daemon->flags & PF_EXITING;
 988 }
 989
 990 /* todo: handle partial completion */
 991 static inline void __ublk_complete_rq(struct request *req)
 992 {
 993         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 994         struct ublk_io *io = &ubq->ios[req->tag];
 995         unsigned int unmapped_bytes;
 996         blk_status_t res = BLK_STS_OK;
 997
 998         /* called from ublk_abort_queue() code path */
 999         if (io->flags & UBLK_IO_FLAG_ABORTED) {
1000                 res = BLK_STS_IOERR;
1001                 goto exit;
1002         }
1003
1004         /* failed read IO if nothing is read */
1005         if (!io->res && req_op(req) == REQ_OP_READ)
1006                 io->res = -EIO;
1007
1008         if (io->res < 0) {
1009                 res = errno_to_blk_status(io->res);
1010                 goto exit;
1011         }
1012
1013         /*
1014          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
1015          * directly.
1016          *
1017          * Both the two needn't unmap.
1018          */
1019         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE &&
1020             req_op(req) != REQ_OP_DRV_IN)
1021                 goto exit;
1022
1023         /* for READ request, writing data in iod->addr to rq buffers */
1024         unmapped_bytes = ublk_unmap_io(ubq, req, io);
1025
1026         /*
1027          * Extremely impossible since we got data filled in just before
1028          *
1029          * Re-read simply for this unlikely case.
1030          */
1031         if (unlikely(unmapped_bytes < io->res))
1032                 io->res = unmapped_bytes;
1033
1034         if (blk_update_request(req, BLK_STS_OK, io->res))
1035                 blk_mq_requeue_request(req, true);
1036         else
1037                 __blk_mq_end_request(req, BLK_STS_OK);
1038
1039         return;
1040 exit:
1041         blk_mq_end_request(req, res);
1042 }
1043
1044 static void ublk_complete_rq(struct kref *ref)
1045 {
1046         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
1047                         ref);
1048         struct request *req = blk_mq_rq_from_pdu(data);
1049
1050         __ublk_complete_rq(req);
1051 }
1052
1053 /*
1054  * Since __ublk_rq_task_work always fails requests immediately during
1055  * exiting, __ublk_fail_req() is only called from abort context during
1056  * exiting. So lock is unnecessary.
1057  *
1058  * Also aborting may not be started yet, keep in mind that one failed
1059  * request may be issued by block layer again.
1060  */
1061 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
1062                 struct request *req)
1063 {
1064         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
1065
1066         if (ublk_queue_can_use_recovery_reissue(ubq))
1067                 blk_mq_requeue_request(req, false);
1068         else
1069                 ublk_put_req_ref(ubq, req);
1070 }
1071
1072 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
1073                                 unsigned issue_flags)
1074 {
1075         /* mark this cmd owned by ublksrv */
1076         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
1077
1078         /*
1079          * clear ACTIVE since we are done with this sqe/cmd slot
1080          * We can only accept io cmd in case of being not active.
1081          */
1082         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
1083
1084         /* tell ublksrv one io request is coming */
1085         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
1086 }
1087
1088 #define UBLK_REQUEUE_DELAY_MS   3
1089
1090 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
1091                 struct request *rq)
1092 {
1093         /* We cannot process this rq so just requeue it. */
1094         if (ublk_queue_can_use_recovery(ubq))
1095                 blk_mq_requeue_request(rq, false);
1096         else
1097                 blk_mq_end_request(rq, BLK_STS_IOERR);
1098 }
1099
1100 static inline void __ublk_rq_task_work(struct request *req,
1101                                        unsigned issue_flags)
1102 {
1103         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1104         int tag = req->tag;
1105         struct ublk_io *io = &ubq->ios[tag];
1106         unsigned int mapped_bytes;
1107
1108         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
1109                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1110                         ublk_get_iod(ubq, req->tag)->addr);
1111
1112         /*
1113          * Task is exiting if either:
1114          *
1115          * (1) current != ubq_daemon.
1116          * io_uring_cmd_complete_in_task() tries to run task_work
1117          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
1118          *
1119          * (2) current->flags & PF_EXITING.
1120          */
1121         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
1122                 __ublk_abort_rq(ubq, req);
1123                 return;
1124         }
1125
1126         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
1127                 /*
1128                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
1129                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
1130                  * and notify it.
1131                  */
1132                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
1133                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
1134                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
1135                                         __func__, io->cmd->cmd_op, ubq->q_id,
1136                                         req->tag, io->flags);
1137                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
1138                         return;
1139                 }
1140                 /*
1141                  * We have handled UBLK_IO_NEED_GET_DATA command,
1142                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
1143                  * do the copy work.
1144                  */
1145                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
1146                 /* update iod->addr because ublksrv may have passed a new io buffer */
1147                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
1148                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
1149                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1150                                 ublk_get_iod(ubq, req->tag)->addr);
1151         }
1152
1153         mapped_bytes = ublk_map_io(ubq, req, io);
1154
1155         /* partially mapped, update io descriptor */
1156         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
1157                 /*
1158                  * Nothing mapped, retry until we succeed.
1159                  *
1160                  * We may never succeed in mapping any bytes here because
1161                  * of OOM. TODO: reserve one buffer with single page pinned
1162                  * for providing forward progress guarantee.
1163                  */
1164                 if (unlikely(!mapped_bytes)) {
1165                         blk_mq_requeue_request(req, false);
1166                         blk_mq_delay_kick_requeue_list(req->q,
1167                                         UBLK_REQUEUE_DELAY_MS);
1168                         return;
1169                 }
1170
1171                 ublk_get_iod(ubq, req->tag)->nr_sectors =
1172                         mapped_bytes >> 9;
1173         }
1174
1175         ublk_init_req_ref(ubq, req);
1176         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
1177 }
1178
1179 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
1180                                         unsigned issue_flags)
1181 {
1182         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1183         struct ublk_rq_data *data, *tmp;
1184
1185         io_cmds = llist_reverse_order(io_cmds);
1186         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1187                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
1188 }
1189
1190 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1191 {
1192         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1193         struct ublk_queue *ubq = pdu->ubq;
1194
1195         ublk_forward_io_cmds(ubq, issue_flags);
1196 }
1197
1198 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
1199 {
1200         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
1201
1202         if (llist_add(&data->node, &ubq->io_cmds)) {
1203                 struct ublk_io *io = &ubq->ios[rq->tag];
1204
1205                 io_uring_cmd_complete_in_task(io->cmd, ublk_rq_task_work_cb);
1206         }
1207 }
1208
1209 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
1210 {
1211         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
1212         unsigned int nr_inflight = 0;
1213         int i;
1214
1215         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
1216                 if (!ubq->timeout) {
1217                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
1218                         ubq->timeout = true;
1219                 }
1220
1221                 return BLK_EH_DONE;
1222         }
1223
1224         if (!ubq_daemon_is_dying(ubq))
1225                 return BLK_EH_RESET_TIMER;
1226
1227         for (i = 0; i < ubq->q_depth; i++) {
1228                 struct ublk_io *io = &ubq->ios[i];
1229
1230                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1231                         nr_inflight++;
1232         }
1233
1234         /* cancelable uring_cmd can't help us if all commands are in-flight */
1235         if (nr_inflight == ubq->q_depth) {
1236                 struct ublk_device *ub = ubq->dev;
1237
1238                 if (ublk_abort_requests(ub, ubq)) {
1239                         if (ublk_can_use_recovery(ub))
1240                                 schedule_work(&ub->quiesce_work);
1241                         else
1242                                 schedule_work(&ub->stop_work);
1243                 }
1244                 return BLK_EH_DONE;
1245         }
1246
1247         return BLK_EH_RESET_TIMER;
1248 }
1249
1250 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
1251                 const struct blk_mq_queue_data *bd)
1252 {
1253         struct ublk_queue *ubq = hctx->driver_data;
1254         struct request *rq = bd->rq;
1255         blk_status_t res;
1256
1257         /* fill iod to slot in io cmd buffer */
1258         res = ublk_setup_iod(ubq, rq);
1259         if (unlikely(res != BLK_STS_OK))
1260                 return BLK_STS_IOERR;
1261
1262         /* With recovery feature enabled, force_abort is set in
1263          * ublk_stop_dev() before calling del_gendisk(). We have to
1264          * abort all requeued and new rqs here to let del_gendisk()
1265          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1266          * to avoid UAF on io_uring ctx.
1267          *
1268          * Note: force_abort is guaranteed to be seen because it is set
1269          * before request queue is unqiuesced.
1270          */
1271         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1272                 return BLK_STS_IOERR;
1273
1274         if (unlikely(ubq->canceling)) {
1275                 __ublk_abort_rq(ubq, rq);
1276                 return BLK_STS_OK;
1277         }
1278
1279         blk_mq_start_request(bd->rq);
1280         ublk_queue_cmd(ubq, rq);
1281
1282         return BLK_STS_OK;
1283 }
1284
1285 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1286                 unsigned int hctx_idx)
1287 {
1288         struct ublk_device *ub = driver_data;
1289         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1290
1291         hctx->driver_data = ubq;
1292         return 0;
1293 }
1294
1295 static const struct blk_mq_ops ublk_mq_ops = {
1296         .queue_rq       = ublk_queue_rq,
1297         .init_hctx      = ublk_init_hctx,
1298         .timeout        = ublk_timeout,
1299 };
1300
1301 static int ublk_ch_open(struct inode *inode, struct file *filp)
1302 {
1303         struct ublk_device *ub = container_of(inode->i_cdev,
1304                         struct ublk_device, cdev);
1305
1306         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1307                 return -EBUSY;
1308         filp->private_data = ub;
1309         return 0;
1310 }
1311
1312 static int ublk_ch_release(struct inode *inode, struct file *filp)
1313 {
1314         struct ublk_device *ub = filp->private_data;
1315
1316         clear_bit(UB_STATE_OPEN, &ub->state);
1317         return 0;
1318 }
1319
1320 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1321 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1322 {
1323         struct ublk_device *ub = filp->private_data;
1324         size_t sz = vma->vm_end - vma->vm_start;
1325         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1326         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1327         int q_id, ret = 0;
1328
1329         spin_lock(&ub->lock);
1330         if (!ub->mm)
1331                 ub->mm = current->mm;
1332         if (current->mm != ub->mm)
1333                 ret = -EINVAL;
1334         spin_unlock(&ub->lock);
1335
1336         if (ret)
1337                 return ret;
1338
1339         if (vma->vm_flags & VM_WRITE)
1340                 return -EPERM;
1341
1342         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1343         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1344                 return -EINVAL;
1345
1346         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1347         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1348                         __func__, q_id, current->pid, vma->vm_start,
1349                         phys_off, (unsigned long)sz);
1350
1351         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1352                 return -EINVAL;
1353
1354         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1355         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1356 }
1357
1358 static void ublk_commit_completion(struct ublk_device *ub,
1359                 const struct ublksrv_io_cmd *ub_cmd)
1360 {
1361         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1362         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1363         struct ublk_io *io = &ubq->ios[tag];
1364         struct request *req;
1365
1366         /* now this cmd slot is owned by nbd driver */
1367         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1368         io->res = ub_cmd->result;
1369
1370         /* find the io request and complete */
1371         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1372         if (WARN_ON_ONCE(unlikely(!req)))
1373                 return;
1374
1375         if (req_op(req) == REQ_OP_ZONE_APPEND)
1376                 req->__sector = ub_cmd->zone_append_lba;
1377
1378         if (likely(!blk_should_fake_timeout(req->q)))
1379                 ublk_put_req_ref(ubq, req);
1380 }
1381
1382 /*
1383  * Called from ubq_daemon context via cancel fn, meantime quiesce ublk
1384  * blk-mq queue, so we are called exclusively with blk-mq and ubq_daemon
1385  * context, so everything is serialized.
1386  */
1387 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1388 {
1389         int i;
1390
1391         for (i = 0; i < ubq->q_depth; i++) {
1392                 struct ublk_io *io = &ubq->ios[i];
1393
1394                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1395                         struct request *rq;
1396
1397                         /*
1398                          * Either we fail the request or ublk_rq_task_work_fn
1399                          * will do it
1400                          */
1401                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1402                         if (rq && blk_mq_request_started(rq)) {
1403                                 io->flags |= UBLK_IO_FLAG_ABORTED;
1404                                 __ublk_fail_req(ubq, io, rq);
1405                         }
1406                 }
1407         }
1408 }
1409
1410 static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq)
1411 {
1412         struct gendisk *disk;
1413
1414         spin_lock(&ubq->cancel_lock);
1415         if (ubq->canceling) {
1416                 spin_unlock(&ubq->cancel_lock);
1417                 return false;
1418         }
1419         ubq->canceling = true;
1420         spin_unlock(&ubq->cancel_lock);
1421
1422         spin_lock(&ub->lock);
1423         disk = ub->ub_disk;
1424         if (disk)
1425                 get_device(disk_to_dev(disk));
1426         spin_unlock(&ub->lock);
1427
1428         /* Our disk has been dead */
1429         if (!disk)
1430                 return false;
1431
1432         /* Now we are serialized with ublk_queue_rq() */
1433         blk_mq_quiesce_queue(disk->queue);
1434         /* abort queue is for making forward progress */
1435         ublk_abort_queue(ub, ubq);
1436         blk_mq_unquiesce_queue(disk->queue);
1437         put_device(disk_to_dev(disk));
1438
1439         return true;
1440 }
1441
1442 static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io,
1443                 unsigned int issue_flags)
1444 {
1445         bool done;
1446
1447         if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1448                 return;
1449
1450         spin_lock(&ubq->cancel_lock);
1451         done = !!(io->flags & UBLK_IO_FLAG_CANCELED);
1452         if (!done)
1453                 io->flags |= UBLK_IO_FLAG_CANCELED;
1454         spin_unlock(&ubq->cancel_lock);
1455
1456         if (!done)
1457                 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1458 }
1459
1460 /*
1461  * The ublk char device won't be closed when calling cancel fn, so both
1462  * ublk device and queue are guaranteed to be live
1463  */
1464 static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd,
1465                 unsigned int issue_flags)
1466 {
1467         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1468         struct ublk_queue *ubq = pdu->ubq;
1469         struct task_struct *task;
1470         struct ublk_device *ub;
1471         bool need_schedule;
1472         struct ublk_io *io;
1473
1474         if (WARN_ON_ONCE(!ubq))
1475                 return;
1476
1477         if (WARN_ON_ONCE(pdu->tag >= ubq->q_depth))
1478                 return;
1479
1480         task = io_uring_cmd_get_task(cmd);
1481         if (WARN_ON_ONCE(task && task != ubq->ubq_daemon))
1482                 return;
1483
1484         ub = ubq->dev;
1485         need_schedule = ublk_abort_requests(ub, ubq);
1486
1487         io = &ubq->ios[pdu->tag];
1488         WARN_ON_ONCE(io->cmd != cmd);
1489         ublk_cancel_cmd(ubq, io, issue_flags);
1490
1491         if (need_schedule) {
1492                 if (ublk_can_use_recovery(ub))
1493                         schedule_work(&ub->quiesce_work);
1494                 else
1495                         schedule_work(&ub->stop_work);
1496         }
1497 }
1498
1499 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1500 {
1501         return ubq->nr_io_ready == ubq->q_depth;
1502 }
1503
1504 static void ublk_cancel_queue(struct ublk_queue *ubq)
1505 {
1506         int i;
1507
1508         for (i = 0; i < ubq->q_depth; i++)
1509                 ublk_cancel_cmd(ubq, &ubq->ios[i], IO_URING_F_UNLOCKED);
1510 }
1511
1512 /* Cancel all pending commands, must be called after del_gendisk() returns */
1513 static void ublk_cancel_dev(struct ublk_device *ub)
1514 {
1515         int i;
1516
1517         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1518                 ublk_cancel_queue(ublk_get_queue(ub, i));
1519 }
1520
1521 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1522 {
1523         bool *idle = data;
1524
1525         if (blk_mq_request_started(rq)) {
1526                 *idle = false;
1527                 return false;
1528         }
1529         return true;
1530 }
1531
1532 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1533 {
1534         bool idle;
1535
1536         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1537         while (true) {
1538                 idle = true;
1539                 blk_mq_tagset_busy_iter(&ub->tag_set,
1540                                 ublk_check_inflight_rq, &idle);
1541                 if (idle)
1542                         break;
1543                 msleep(UBLK_REQUEUE_DELAY_MS);
1544         }
1545 }
1546
1547 static void __ublk_quiesce_dev(struct ublk_device *ub)
1548 {
1549         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1550                         __func__, ub->dev_info.dev_id,
1551                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1552                         "LIVE" : "QUIESCED");
1553         blk_mq_quiesce_queue(ub->ub_disk->queue);
1554         ublk_wait_tagset_rqs_idle(ub);
1555         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1556 }
1557
1558 static void ublk_quiesce_work_fn(struct work_struct *work)
1559 {
1560         struct ublk_device *ub =
1561                 container_of(work, struct ublk_device, quiesce_work);
1562
1563         mutex_lock(&ub->mutex);
1564         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1565                 goto unlock;
1566         __ublk_quiesce_dev(ub);
1567  unlock:
1568         mutex_unlock(&ub->mutex);
1569         ublk_cancel_dev(ub);
1570 }
1571
1572 static void ublk_unquiesce_dev(struct ublk_device *ub)
1573 {
1574         int i;
1575
1576         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1577                         __func__, ub->dev_info.dev_id,
1578                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1579                         "LIVE" : "QUIESCED");
1580         /* quiesce_work has run. We let requeued rqs be aborted
1581          * before running fallback_wq. "force_abort" must be seen
1582          * after request queue is unqiuesced. Then del_gendisk()
1583          * can move on.
1584          */
1585         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1586                 ublk_get_queue(ub, i)->force_abort = true;
1587
1588         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1589         /* We may have requeued some rqs in ublk_quiesce_queue() */
1590         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1591 }
1592
1593 static void ublk_stop_dev(struct ublk_device *ub)
1594 {
1595         struct gendisk *disk;
1596
1597         mutex_lock(&ub->mutex);
1598         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1599                 goto unlock;
1600         if (ublk_can_use_recovery(ub)) {
1601                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1602                         __ublk_quiesce_dev(ub);
1603                 ublk_unquiesce_dev(ub);
1604         }
1605         del_gendisk(ub->ub_disk);
1606
1607         /* Sync with ublk_abort_queue() by holding the lock */
1608         spin_lock(&ub->lock);
1609         disk = ub->ub_disk;
1610         ub->dev_info.state = UBLK_S_DEV_DEAD;
1611         ub->dev_info.ublksrv_pid = -1;
1612         ub->ub_disk = NULL;
1613         spin_unlock(&ub->lock);
1614         put_disk(disk);
1615  unlock:
1616         mutex_unlock(&ub->mutex);
1617         ublk_cancel_dev(ub);
1618 }
1619
1620 /* device can only be started after all IOs are ready */
1621 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1622 {
1623         mutex_lock(&ub->mutex);
1624         ubq->nr_io_ready++;
1625         if (ublk_queue_ready(ubq)) {
1626                 ubq->ubq_daemon = current;
1627                 get_task_struct(ubq->ubq_daemon);
1628                 ub->nr_queues_ready++;
1629
1630                 if (capable(CAP_SYS_ADMIN))
1631                         ub->nr_privileged_daemon++;
1632         }
1633         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1634                 complete_all(&ub->completion);
1635         mutex_unlock(&ub->mutex);
1636 }
1637
1638 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1639                 int tag)
1640 {
1641         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1642         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1643
1644         ublk_queue_cmd(ubq, req);
1645 }
1646
1647 static inline int ublk_check_cmd_op(u32 cmd_op)
1648 {
1649         u32 ioc_type = _IOC_TYPE(cmd_op);
1650
1651         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1652                 return -EOPNOTSUPP;
1653
1654         if (ioc_type != 'u' && ioc_type != 0)
1655                 return -EOPNOTSUPP;
1656
1657         return 0;
1658 }
1659
1660 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1661                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1662 {
1663         io->cmd = cmd;
1664         io->flags |= UBLK_IO_FLAG_ACTIVE;
1665         io->addr = buf_addr;
1666 }
1667
1668 static inline void ublk_prep_cancel(struct io_uring_cmd *cmd,
1669                                     unsigned int issue_flags,
1670                                     struct ublk_queue *ubq, unsigned int tag)
1671 {
1672         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1673
1674         /*
1675          * Safe to refer to @ubq since ublk_queue won't be died until its
1676          * commands are completed
1677          */
1678         pdu->ubq = ubq;
1679         pdu->tag = tag;
1680         io_uring_cmd_mark_cancelable(cmd, issue_flags);
1681 }
1682
1683 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1684                                unsigned int issue_flags,
1685                                const struct ublksrv_io_cmd *ub_cmd)
1686 {
1687         struct ublk_device *ub = cmd->file->private_data;
1688         struct ublk_queue *ubq;
1689         struct ublk_io *io;
1690         u32 cmd_op = cmd->cmd_op;
1691         unsigned tag = ub_cmd->tag;
1692         int ret = -EINVAL;
1693         struct request *req;
1694
1695         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1696                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1697                         ub_cmd->result);
1698
1699         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1700                 goto out;
1701
1702         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1703         if (!ubq || ub_cmd->q_id != ubq->q_id)
1704                 goto out;
1705
1706         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1707                 goto out;
1708
1709         if (tag >= ubq->q_depth)
1710                 goto out;
1711
1712         io = &ubq->ios[tag];
1713
1714         /* there is pending io cmd, something must be wrong */
1715         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1716                 ret = -EBUSY;
1717                 goto out;
1718         }
1719
1720         /*
1721          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1722          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1723          */
1724         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1725                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1726                 goto out;
1727
1728         ret = ublk_check_cmd_op(cmd_op);
1729         if (ret)
1730                 goto out;
1731
1732         ret = -EINVAL;
1733         switch (_IOC_NR(cmd_op)) {
1734         case UBLK_IO_FETCH_REQ:
1735                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1736                 if (ublk_queue_ready(ubq)) {
1737                         ret = -EBUSY;
1738                         goto out;
1739                 }
1740                 /*
1741                  * The io is being handled by server, so COMMIT_RQ is expected
1742                  * instead of FETCH_REQ
1743                  */
1744                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1745                         goto out;
1746
1747                 if (!ublk_support_user_copy(ubq)) {
1748                         /*
1749                          * FETCH_RQ has to provide IO buffer if NEED GET
1750                          * DATA is not enabled
1751                          */
1752                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1753                                 goto out;
1754                 } else if (ub_cmd->addr) {
1755                         /* User copy requires addr to be unset */
1756                         ret = -EINVAL;
1757                         goto out;
1758                 }
1759
1760                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1761                 ublk_mark_io_ready(ub, ubq);
1762                 break;
1763         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1764                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1765
1766                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1767                         goto out;
1768
1769                 if (!ublk_support_user_copy(ubq)) {
1770                         /*
1771                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1772                          * NEED GET DATA is not enabled or it is Read IO.
1773                          */
1774                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1775                                                 req_op(req) == REQ_OP_READ))
1776                                 goto out;
1777                 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1778                         /*
1779                          * User copy requires addr to be unset when command is
1780                          * not zone append
1781                          */
1782                         ret = -EINVAL;
1783                         goto out;
1784                 }
1785
1786                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1787                 ublk_commit_completion(ub, ub_cmd);
1788                 break;
1789         case UBLK_IO_NEED_GET_DATA:
1790                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1791                         goto out;
1792                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1793                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1794                 break;
1795         default:
1796                 goto out;
1797         }
1798         ublk_prep_cancel(cmd, issue_flags, ubq, tag);
1799         return -EIOCBQUEUED;
1800
1801  out:
1802         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1803         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1804                         __func__, cmd_op, tag, ret, io->flags);
1805         return -EIOCBQUEUED;
1806 }
1807
1808 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1809                 struct ublk_queue *ubq, int tag, size_t offset)
1810 {
1811         struct request *req;
1812
1813         if (!ublk_need_req_ref(ubq))
1814                 return NULL;
1815
1816         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1817         if (!req)
1818                 return NULL;
1819
1820         if (!ublk_get_req_ref(ubq, req))
1821                 return NULL;
1822
1823         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1824                 goto fail_put;
1825
1826         if (!ublk_rq_has_data(req))
1827                 goto fail_put;
1828
1829         if (offset > blk_rq_bytes(req))
1830                 goto fail_put;
1831
1832         return req;
1833 fail_put:
1834         ublk_put_req_ref(ubq, req);
1835         return NULL;
1836 }
1837
1838 static inline int ublk_ch_uring_cmd_local(struct io_uring_cmd *cmd,
1839                 unsigned int issue_flags)
1840 {
1841         /*
1842          * Not necessary for async retry, but let's keep it simple and always
1843          * copy the values to avoid any potential reuse.
1844          */
1845         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1846         const struct ublksrv_io_cmd ub_cmd = {
1847                 .q_id = READ_ONCE(ub_src->q_id),
1848                 .tag = READ_ONCE(ub_src->tag),
1849                 .result = READ_ONCE(ub_src->result),
1850                 .addr = READ_ONCE(ub_src->addr)
1851         };
1852
1853         WARN_ON_ONCE(issue_flags & IO_URING_F_UNLOCKED);
1854
1855         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1856 }
1857
1858 static void ublk_ch_uring_cmd_cb(struct io_uring_cmd *cmd,
1859                 unsigned int issue_flags)
1860 {
1861         ublk_ch_uring_cmd_local(cmd, issue_flags);
1862 }
1863
1864 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1865 {
1866         if (unlikely(issue_flags & IO_URING_F_CANCEL)) {
1867                 ublk_uring_cmd_cancel_fn(cmd, issue_flags);
1868                 return 0;
1869         }
1870
1871         /* well-implemented server won't run into unlocked */
1872         if (unlikely(issue_flags & IO_URING_F_UNLOCKED)) {
1873                 io_uring_cmd_complete_in_task(cmd, ublk_ch_uring_cmd_cb);
1874                 return -EIOCBQUEUED;
1875         }
1876
1877         return ublk_ch_uring_cmd_local(cmd, issue_flags);
1878 }
1879
1880 static inline bool ublk_check_ubuf_dir(const struct request *req,
1881                 int ubuf_dir)
1882 {
1883         /* copy ubuf to request pages */
1884         if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1885             ubuf_dir == ITER_SOURCE)
1886                 return true;
1887
1888         /* copy request pages to ubuf */
1889         if ((req_op(req) == REQ_OP_WRITE ||
1890              req_op(req) == REQ_OP_ZONE_APPEND) &&
1891             ubuf_dir == ITER_DEST)
1892                 return true;
1893
1894         return false;
1895 }
1896
1897 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1898                 struct iov_iter *iter, size_t *off, int dir)
1899 {
1900         struct ublk_device *ub = iocb->ki_filp->private_data;
1901         struct ublk_queue *ubq;
1902         struct request *req;
1903         size_t buf_off;
1904         u16 tag, q_id;
1905
1906         if (!ub)
1907                 return ERR_PTR(-EACCES);
1908
1909         if (!user_backed_iter(iter))
1910                 return ERR_PTR(-EACCES);
1911
1912         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1913                 return ERR_PTR(-EACCES);
1914
1915         tag = ublk_pos_to_tag(iocb->ki_pos);
1916         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1917         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1918
1919         if (q_id >= ub->dev_info.nr_hw_queues)
1920                 return ERR_PTR(-EINVAL);
1921
1922         ubq = ublk_get_queue(ub, q_id);
1923         if (!ubq)
1924                 return ERR_PTR(-EINVAL);
1925
1926         if (tag >= ubq->q_depth)
1927                 return ERR_PTR(-EINVAL);
1928
1929         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1930         if (!req)
1931                 return ERR_PTR(-EINVAL);
1932
1933         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1934                 goto fail;
1935
1936         if (!ublk_check_ubuf_dir(req, dir))
1937                 goto fail;
1938
1939         *off = buf_off;
1940         return req;
1941 fail:
1942         ublk_put_req_ref(ubq, req);
1943         return ERR_PTR(-EACCES);
1944 }
1945
1946 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1947 {
1948         struct ublk_queue *ubq;
1949         struct request *req;
1950         size_t buf_off;
1951         size_t ret;
1952
1953         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1954         if (IS_ERR(req))
1955                 return PTR_ERR(req);
1956
1957         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1958         ubq = req->mq_hctx->driver_data;
1959         ublk_put_req_ref(ubq, req);
1960
1961         return ret;
1962 }
1963
1964 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1965 {
1966         struct ublk_queue *ubq;
1967         struct request *req;
1968         size_t buf_off;
1969         size_t ret;
1970
1971         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1972         if (IS_ERR(req))
1973                 return PTR_ERR(req);
1974
1975         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1976         ubq = req->mq_hctx->driver_data;
1977         ublk_put_req_ref(ubq, req);
1978
1979         return ret;
1980 }
1981
1982 static const struct file_operations ublk_ch_fops = {
1983         .owner = THIS_MODULE,
1984         .open = ublk_ch_open,
1985         .release = ublk_ch_release,
1986         .read_iter = ublk_ch_read_iter,
1987         .write_iter = ublk_ch_write_iter,
1988         .uring_cmd = ublk_ch_uring_cmd,
1989         .mmap = ublk_ch_mmap,
1990 };
1991
1992 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1993 {
1994         int size = ublk_queue_cmd_buf_size(ub, q_id);
1995         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1996
1997         if (ubq->ubq_daemon)
1998                 put_task_struct(ubq->ubq_daemon);
1999         if (ubq->io_cmd_buf)
2000                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
2001 }
2002
2003 static int ublk_init_queue(struct ublk_device *ub, int q_id)
2004 {
2005         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
2006         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
2007         void *ptr;
2008         int size;
2009
2010         spin_lock_init(&ubq->cancel_lock);
2011         ubq->flags = ub->dev_info.flags;
2012         ubq->q_id = q_id;
2013         ubq->q_depth = ub->dev_info.queue_depth;
2014         size = ublk_queue_cmd_buf_size(ub, q_id);
2015
2016         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
2017         if (!ptr)
2018                 return -ENOMEM;
2019
2020         ubq->io_cmd_buf = ptr;
2021         ubq->dev = ub;
2022         return 0;
2023 }
2024
2025 static void ublk_deinit_queues(struct ublk_device *ub)
2026 {
2027         int nr_queues = ub->dev_info.nr_hw_queues;
2028         int i;
2029
2030         if (!ub->__queues)
2031                 return;
2032
2033         for (i = 0; i < nr_queues; i++)
2034                 ublk_deinit_queue(ub, i);
2035         kfree(ub->__queues);
2036 }
2037
2038 static int ublk_init_queues(struct ublk_device *ub)
2039 {
2040         int nr_queues = ub->dev_info.nr_hw_queues;
2041         int depth = ub->dev_info.queue_depth;
2042         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
2043         int i, ret = -ENOMEM;
2044
2045         ub->queue_size = ubq_size;
2046         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
2047         if (!ub->__queues)
2048                 return ret;
2049
2050         for (i = 0; i < nr_queues; i++) {
2051                 if (ublk_init_queue(ub, i))
2052                         goto fail;
2053         }
2054
2055         init_completion(&ub->completion);
2056         return 0;
2057
2058  fail:
2059         ublk_deinit_queues(ub);
2060         return ret;
2061 }
2062
2063 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2064 {
2065         int i = idx;
2066         int err;
2067
2068         spin_lock(&ublk_idr_lock);
2069         /* allocate id, if @id >= 0, we're requesting that specific id */
2070         if (i >= 0) {
2071                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2072                 if (err == -ENOSPC)
2073                         err = -EEXIST;
2074         } else {
2075                 err = idr_alloc(&ublk_index_idr, ub, 0, UBLK_MAX_UBLKS,
2076                                 GFP_NOWAIT);
2077         }
2078         spin_unlock(&ublk_idr_lock);
2079
2080         if (err >= 0)
2081                 ub->ub_number = err;
2082
2083         return err;
2084 }
2085
2086 static void ublk_free_dev_number(struct ublk_device *ub)
2087 {
2088         spin_lock(&ublk_idr_lock);
2089         idr_remove(&ublk_index_idr, ub->ub_number);
2090         wake_up_all(&ublk_idr_wq);
2091         spin_unlock(&ublk_idr_lock);
2092 }
2093
2094 static void ublk_cdev_rel(struct device *dev)
2095 {
2096         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2097
2098         blk_mq_free_tag_set(&ub->tag_set);
2099         ublk_deinit_queues(ub);
2100         ublk_free_dev_number(ub);
2101         mutex_destroy(&ub->mutex);
2102         kfree(ub);
2103 }
2104
2105 static int ublk_add_chdev(struct ublk_device *ub)
2106 {
2107         struct device *dev = &ub->cdev_dev;
2108         int minor = ub->ub_number;
2109         int ret;
2110
2111         dev->parent = ublk_misc.this_device;
2112         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2113         dev->class = &ublk_chr_class;
2114         dev->release = ublk_cdev_rel;
2115         device_initialize(dev);
2116
2117         ret = dev_set_name(dev, "ublkc%d", minor);
2118         if (ret)
2119                 goto fail;
2120
2121         cdev_init(&ub->cdev, &ublk_ch_fops);
2122         ret = cdev_device_add(&ub->cdev, dev);
2123         if (ret)
2124                 goto fail;
2125
2126         ublks_added++;
2127         return 0;
2128  fail:
2129         put_device(dev);
2130         return ret;
2131 }
2132
2133 static void ublk_stop_work_fn(struct work_struct *work)
2134 {
2135         struct ublk_device *ub =
2136                 container_of(work, struct ublk_device, stop_work);
2137
2138         ublk_stop_dev(ub);
2139 }
2140
2141 /* align max io buffer size with PAGE_SIZE */
2142 static void ublk_align_max_io_size(struct ublk_device *ub)
2143 {
2144         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2145
2146         ub->dev_info.max_io_buf_bytes =
2147                 round_down(max_io_bytes, PAGE_SIZE);
2148 }
2149
2150 static int ublk_add_tag_set(struct ublk_device *ub)
2151 {
2152         ub->tag_set.ops = &ublk_mq_ops;
2153         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2154         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2155         ub->tag_set.numa_node = NUMA_NO_NODE;
2156         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2157         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2158         ub->tag_set.driver_data = ub;
2159         return blk_mq_alloc_tag_set(&ub->tag_set);
2160 }
2161
2162 static void ublk_remove(struct ublk_device *ub)
2163 {
2164         ublk_stop_dev(ub);
2165         cancel_work_sync(&ub->stop_work);
2166         cancel_work_sync(&ub->quiesce_work);
2167         cdev_device_del(&ub->cdev, &ub->cdev_dev);
2168         ublk_put_device(ub);
2169         ublks_added--;
2170 }
2171
2172 static struct ublk_device *ublk_get_device_from_id(int idx)
2173 {
2174         struct ublk_device *ub = NULL;
2175
2176         if (idx < 0)
2177                 return NULL;
2178
2179         spin_lock(&ublk_idr_lock);
2180         ub = idr_find(&ublk_index_idr, idx);
2181         if (ub)
2182                 ub = ublk_get_device(ub);
2183         spin_unlock(&ublk_idr_lock);
2184
2185         return ub;
2186 }
2187
2188 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2189 {
2190         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2191         const struct ublk_param_basic *p = &ub->params.basic;
2192         int ublksrv_pid = (int)header->data[0];
2193         struct queue_limits lim = {
2194                 .logical_block_size     = 1 << p->logical_bs_shift,
2195                 .physical_block_size    = 1 << p->physical_bs_shift,
2196                 .io_min                 = 1 << p->io_min_shift,
2197                 .io_opt                 = 1 << p->io_opt_shift,
2198                 .max_hw_sectors         = p->max_sectors,
2199                 .chunk_sectors          = p->chunk_sectors,
2200                 .virt_boundary_mask     = p->virt_boundary_mask,
2201                 .max_segments           = USHRT_MAX,
2202                 .max_segment_size       = UINT_MAX,
2203                 .dma_alignment          = 3,
2204         };
2205         struct gendisk *disk;
2206         int ret = -EINVAL;
2207
2208         if (ublksrv_pid <= 0)
2209                 return -EINVAL;
2210         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
2211                 return -EINVAL;
2212
2213         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
2214                 const struct ublk_param_discard *pd = &ub->params.discard;
2215
2216                 lim.discard_alignment = pd->discard_alignment;
2217                 lim.discard_granularity = pd->discard_granularity;
2218                 lim.max_hw_discard_sectors = pd->max_discard_sectors;
2219                 lim.max_write_zeroes_sectors = pd->max_write_zeroes_sectors;
2220                 lim.max_discard_segments = pd->max_discard_segments;
2221         }
2222
2223         if (ub->params.types & UBLK_PARAM_TYPE_ZONED) {
2224                 const struct ublk_param_zoned *p = &ub->params.zoned;
2225
2226                 if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED))
2227                         return -EOPNOTSUPP;
2228
2229                 lim.features |= BLK_FEAT_ZONED;
2230                 lim.max_active_zones = p->max_active_zones;
2231                 lim.max_open_zones =  p->max_open_zones;
2232                 lim.max_zone_append_sectors = p->max_zone_append_sectors;
2233         }
2234
2235         if (ub->params.basic.attrs & UBLK_ATTR_VOLATILE_CACHE) {
2236                 lim.features |= BLK_FEAT_WRITE_CACHE;
2237                 if (ub->params.basic.attrs & UBLK_ATTR_FUA)
2238                         lim.features |= BLK_FEAT_FUA;
2239         }
2240
2241         if (ub->params.basic.attrs & UBLK_ATTR_ROTATIONAL)
2242                 lim.features |= BLK_FEAT_ROTATIONAL;
2243
2244         if (wait_for_completion_interruptible(&ub->completion) != 0)
2245                 return -EINTR;
2246
2247         mutex_lock(&ub->mutex);
2248         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2249             test_bit(UB_STATE_USED, &ub->state)) {
2250                 ret = -EEXIST;
2251                 goto out_unlock;
2252         }
2253
2254         disk = blk_mq_alloc_disk(&ub->tag_set, &lim, NULL);
2255         if (IS_ERR(disk)) {
2256                 ret = PTR_ERR(disk);
2257                 goto out_unlock;
2258         }
2259         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2260         disk->fops = &ub_fops;
2261         disk->private_data = ub;
2262
2263         ub->dev_info.ublksrv_pid = ublksrv_pid;
2264         ub->ub_disk = disk;
2265
2266         ublk_apply_params(ub);
2267
2268         /* don't probe partitions if any one ubq daemon is un-trusted */
2269         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2270                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2271
2272         ublk_get_device(ub);
2273         ub->dev_info.state = UBLK_S_DEV_LIVE;
2274
2275         if (ublk_dev_is_zoned(ub)) {
2276                 ret = ublk_revalidate_disk_zones(ub);
2277                 if (ret)
2278                         goto out_put_cdev;
2279         }
2280
2281         ret = add_disk(disk);
2282         if (ret)
2283                 goto out_put_cdev;
2284
2285         set_bit(UB_STATE_USED, &ub->state);
2286
2287 out_put_cdev:
2288         if (ret) {
2289                 ub->dev_info.state = UBLK_S_DEV_DEAD;
2290                 ublk_put_device(ub);
2291         }
2292         if (ret)
2293                 put_disk(disk);
2294 out_unlock:
2295         mutex_unlock(&ub->mutex);
2296         return ret;
2297 }
2298
2299 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2300                 struct io_uring_cmd *cmd)
2301 {
2302         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2303         void __user *argp = (void __user *)(unsigned long)header->addr;
2304         cpumask_var_t cpumask;
2305         unsigned long queue;
2306         unsigned int retlen;
2307         unsigned int i;
2308         int ret;
2309
2310         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2311                 return -EINVAL;
2312         if (header->len & (sizeof(unsigned long)-1))
2313                 return -EINVAL;
2314         if (!header->addr)
2315                 return -EINVAL;
2316
2317         queue = header->data[0];
2318         if (queue >= ub->dev_info.nr_hw_queues)
2319                 return -EINVAL;
2320
2321         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2322                 return -ENOMEM;
2323
2324         for_each_possible_cpu(i) {
2325                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2326                         cpumask_set_cpu(i, cpumask);
2327         }
2328
2329         ret = -EFAULT;
2330         retlen = min_t(unsigned short, header->len, cpumask_size());
2331         if (copy_to_user(argp, cpumask, retlen))
2332                 goto out_free_cpumask;
2333         if (retlen != header->len &&
2334             clear_user(argp + retlen, header->len - retlen))
2335                 goto out_free_cpumask;
2336
2337         ret = 0;
2338 out_free_cpumask:
2339         free_cpumask_var(cpumask);
2340         return ret;
2341 }
2342
2343 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2344 {
2345         pr_devel("%s: dev id %d flags %llx\n", __func__,
2346                         info->dev_id, info->flags);
2347         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2348                         info->nr_hw_queues, info->queue_depth);
2349 }
2350
2351 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2352 {
2353         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2354         void __user *argp = (void __user *)(unsigned long)header->addr;
2355         struct ublksrv_ctrl_dev_info info;
2356         struct ublk_device *ub;
2357         int ret = -EINVAL;
2358
2359         if (header->len < sizeof(info) || !header->addr)
2360                 return -EINVAL;
2361         if (header->queue_id != (u16)-1) {
2362                 pr_warn("%s: queue_id is wrong %x\n",
2363                         __func__, header->queue_id);
2364                 return -EINVAL;
2365         }
2366
2367         if (copy_from_user(&info, argp, sizeof(info)))
2368                 return -EFAULT;
2369
2370         if (capable(CAP_SYS_ADMIN))
2371                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2372         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2373                 return -EPERM;
2374
2375         /*
2376          * unprivileged device can't be trusted, but RECOVERY and
2377          * RECOVERY_REISSUE still may hang error handling, so can't
2378          * support recovery features for unprivileged ublk now
2379          *
2380          * TODO: provide forward progress for RECOVERY handler, so that
2381          * unprivileged device can benefit from it
2382          */
2383         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2384                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2385                                 UBLK_F_USER_RECOVERY);
2386
2387         /* the created device is always owned by current user */
2388         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2389
2390         if (header->dev_id != info.dev_id) {
2391                 pr_warn("%s: dev id not match %u %u\n",
2392                         __func__, header->dev_id, info.dev_id);
2393                 return -EINVAL;
2394         }
2395
2396         if (header->dev_id != U32_MAX && header->dev_id >= UBLK_MAX_UBLKS) {
2397                 pr_warn("%s: dev id is too large. Max supported is %d\n",
2398                         __func__, UBLK_MAX_UBLKS - 1);
2399                 return -EINVAL;
2400         }
2401
2402         ublk_dump_dev_info(&info);
2403
2404         ret = mutex_lock_killable(&ublk_ctl_mutex);
2405         if (ret)
2406                 return ret;
2407
2408         ret = -EACCES;
2409         if (ublks_added >= ublks_max)
2410                 goto out_unlock;
2411
2412         ret = -ENOMEM;
2413         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2414         if (!ub)
2415                 goto out_unlock;
2416         mutex_init(&ub->mutex);
2417         spin_lock_init(&ub->lock);
2418         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2419         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2420
2421         ret = ublk_alloc_dev_number(ub, header->dev_id);
2422         if (ret < 0)
2423                 goto out_free_ub;
2424
2425         memcpy(&ub->dev_info, &info, sizeof(info));
2426
2427         /* update device id */
2428         ub->dev_info.dev_id = ub->ub_number;
2429
2430         /*
2431          * 64bit flags will be copied back to userspace as feature
2432          * negotiation result, so have to clear flags which driver
2433          * doesn't support yet, then userspace can get correct flags
2434          * (features) to handle.
2435          */
2436         ub->dev_info.flags &= UBLK_F_ALL;
2437
2438         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2439                 UBLK_F_URING_CMD_COMP_IN_TASK;
2440
2441         /* GET_DATA isn't needed any more with USER_COPY */
2442         if (ublk_dev_is_user_copy(ub))
2443                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2444
2445         /* Zoned storage support requires user copy feature */
2446         if (ublk_dev_is_zoned(ub) &&
2447             (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2448                 ret = -EINVAL;
2449                 goto out_free_dev_number;
2450         }
2451
2452         /* We are not ready to support zero copy */
2453         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2454
2455         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2456                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2457         ublk_align_max_io_size(ub);
2458
2459         ret = ublk_init_queues(ub);
2460         if (ret)
2461                 goto out_free_dev_number;
2462
2463         ret = ublk_add_tag_set(ub);
2464         if (ret)
2465                 goto out_deinit_queues;
2466
2467         ret = -EFAULT;
2468         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2469                 goto out_free_tag_set;
2470
2471         /*
2472          * Add the char dev so that ublksrv daemon can be setup.
2473          * ublk_add_chdev() will cleanup everything if it fails.
2474          */
2475         ret = ublk_add_chdev(ub);
2476         goto out_unlock;
2477
2478 out_free_tag_set:
2479         blk_mq_free_tag_set(&ub->tag_set);
2480 out_deinit_queues:
2481         ublk_deinit_queues(ub);
2482 out_free_dev_number:
2483         ublk_free_dev_number(ub);
2484 out_free_ub:
2485         mutex_destroy(&ub->mutex);
2486         kfree(ub);
2487 out_unlock:
2488         mutex_unlock(&ublk_ctl_mutex);
2489         return ret;
2490 }
2491
2492 static inline bool ublk_idr_freed(int id)
2493 {
2494         void *ptr;
2495
2496         spin_lock(&ublk_idr_lock);
2497         ptr = idr_find(&ublk_index_idr, id);
2498         spin_unlock(&ublk_idr_lock);
2499
2500         return ptr == NULL;
2501 }
2502
2503 static int ublk_ctrl_del_dev(struct ublk_device **p_ub, bool wait)
2504 {
2505         struct ublk_device *ub = *p_ub;
2506         int idx = ub->ub_number;
2507         int ret;
2508
2509         ret = mutex_lock_killable(&ublk_ctl_mutex);
2510         if (ret)
2511                 return ret;
2512
2513         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2514                 ublk_remove(ub);
2515                 set_bit(UB_STATE_DELETED, &ub->state);
2516         }
2517
2518         /* Mark the reference as consumed */
2519         *p_ub = NULL;
2520         ublk_put_device(ub);
2521         mutex_unlock(&ublk_ctl_mutex);
2522
2523         /*
2524          * Wait until the idr is removed, then it can be reused after
2525          * DEL_DEV command is returned.
2526          *
2527          * If we returns because of user interrupt, future delete command
2528          * may come:
2529          *
2530          * - the device number isn't freed, this device won't or needn't
2531          *   be deleted again, since UB_STATE_DELETED is set, and device
2532          *   will be released after the last reference is dropped
2533          *
2534          * - the device number is freed already, we will not find this
2535          *   device via ublk_get_device_from_id()
2536          */
2537         if (wait && wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx)))
2538                 return -EINTR;
2539         return 0;
2540 }
2541
2542 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2543 {
2544         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2545
2546         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2547                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2548                         header->data[0], header->addr, header->len);
2549 }
2550
2551 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2552 {
2553         ublk_stop_dev(ub);
2554         cancel_work_sync(&ub->stop_work);
2555         cancel_work_sync(&ub->quiesce_work);
2556
2557         return 0;
2558 }
2559
2560 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2561                 struct io_uring_cmd *cmd)
2562 {
2563         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2564         void __user *argp = (void __user *)(unsigned long)header->addr;
2565
2566         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2567                 return -EINVAL;
2568
2569         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2570                 return -EFAULT;
2571
2572         return 0;
2573 }
2574
2575 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2576 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2577 {
2578         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2579         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2580
2581         if (ub->ub_disk) {
2582                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2583                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2584         } else {
2585                 ub->params.devt.disk_major = 0;
2586                 ub->params.devt.disk_minor = 0;
2587         }
2588         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2589 }
2590
2591 static int ublk_ctrl_get_params(struct ublk_device *ub,
2592                 struct io_uring_cmd *cmd)
2593 {
2594         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2595         void __user *argp = (void __user *)(unsigned long)header->addr;
2596         struct ublk_params_header ph;
2597         int ret;
2598
2599         if (header->len <= sizeof(ph) || !header->addr)
2600                 return -EINVAL;
2601
2602         if (copy_from_user(&ph, argp, sizeof(ph)))
2603                 return -EFAULT;
2604
2605         if (ph.len > header->len || !ph.len)
2606                 return -EINVAL;
2607
2608         if (ph.len > sizeof(struct ublk_params))
2609                 ph.len = sizeof(struct ublk_params);
2610
2611         mutex_lock(&ub->mutex);
2612         ublk_ctrl_fill_params_devt(ub);
2613         if (copy_to_user(argp, &ub->params, ph.len))
2614                 ret = -EFAULT;
2615         else
2616                 ret = 0;
2617         mutex_unlock(&ub->mutex);
2618
2619         return ret;
2620 }
2621
2622 static int ublk_ctrl_set_params(struct ublk_device *ub,
2623                 struct io_uring_cmd *cmd)
2624 {
2625         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2626         void __user *argp = (void __user *)(unsigned long)header->addr;
2627         struct ublk_params_header ph;
2628         int ret = -EFAULT;
2629
2630         if (header->len <= sizeof(ph) || !header->addr)
2631                 return -EINVAL;
2632
2633         if (copy_from_user(&ph, argp, sizeof(ph)))
2634                 return -EFAULT;
2635
2636         if (ph.len > header->len || !ph.len || !ph.types)
2637                 return -EINVAL;
2638
2639         if (ph.len > sizeof(struct ublk_params))
2640                 ph.len = sizeof(struct ublk_params);
2641
2642         /* parameters can only be changed when device isn't live */
2643         mutex_lock(&ub->mutex);
2644         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2645                 ret = -EACCES;
2646         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2647                 ret = -EFAULT;
2648         } else {
2649                 /* clear all we don't support yet */
2650                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2651                 ret = ublk_validate_params(ub);
2652                 if (ret)
2653                         ub->params.types = 0;
2654         }
2655         mutex_unlock(&ub->mutex);
2656
2657         return ret;
2658 }
2659
2660 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2661 {
2662         int i;
2663
2664         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2665
2666         /* All old ioucmds have to be completed */
2667         ubq->nr_io_ready = 0;
2668         /* old daemon is PF_EXITING, put it now */
2669         put_task_struct(ubq->ubq_daemon);
2670         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2671         ubq->ubq_daemon = NULL;
2672         ubq->timeout = false;
2673         ubq->canceling = false;
2674
2675         for (i = 0; i < ubq->q_depth; i++) {
2676                 struct ublk_io *io = &ubq->ios[i];
2677
2678                 /* forget everything now and be ready for new FETCH_REQ */
2679                 io->flags = 0;
2680                 io->cmd = NULL;
2681                 io->addr = 0;
2682         }
2683 }
2684
2685 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2686                 struct io_uring_cmd *cmd)
2687 {
2688         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2689         int ret = -EINVAL;
2690         int i;
2691
2692         mutex_lock(&ub->mutex);
2693         if (!ublk_can_use_recovery(ub))
2694                 goto out_unlock;
2695         if (!ub->nr_queues_ready)
2696                 goto out_unlock;
2697         /*
2698          * START_RECOVERY is only allowd after:
2699          *
2700          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2701          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2702          *     released.
2703          *
2704          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2705          *     (a)has quiesced request queue
2706          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2707          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2708          *     (d)has completed/camceled all ioucmds owned by ther dying process
2709          */
2710         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2711                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2712                 ret = -EBUSY;
2713                 goto out_unlock;
2714         }
2715         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2716         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2717                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2718         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2719         ub->mm = NULL;
2720         ub->nr_queues_ready = 0;
2721         ub->nr_privileged_daemon = 0;
2722         init_completion(&ub->completion);
2723         ret = 0;
2724  out_unlock:
2725         mutex_unlock(&ub->mutex);
2726         return ret;
2727 }
2728
2729 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2730                 struct io_uring_cmd *cmd)
2731 {
2732         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2733         int ublksrv_pid = (int)header->data[0];
2734         int ret = -EINVAL;
2735
2736         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2737                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2738         /* wait until new ubq_daemon sending all FETCH_REQ */
2739         if (wait_for_completion_interruptible(&ub->completion))
2740                 return -EINTR;
2741
2742         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2743                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2744
2745         mutex_lock(&ub->mutex);
2746         if (!ublk_can_use_recovery(ub))
2747                 goto out_unlock;
2748
2749         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2750                 ret = -EBUSY;
2751                 goto out_unlock;
2752         }
2753         ub->dev_info.ublksrv_pid = ublksrv_pid;
2754         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2755                         __func__, ublksrv_pid, header->dev_id);
2756         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2757         pr_devel("%s: queue unquiesced, dev id %d.\n",
2758                         __func__, header->dev_id);
2759         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2760         ub->dev_info.state = UBLK_S_DEV_LIVE;
2761         ret = 0;
2762  out_unlock:
2763         mutex_unlock(&ub->mutex);
2764         return ret;
2765 }
2766
2767 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2768 {
2769         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2770         void __user *argp = (void __user *)(unsigned long)header->addr;
2771         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2772
2773         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2774                 return -EINVAL;
2775
2776         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2777                 return -EFAULT;
2778
2779         return 0;
2780 }
2781
2782 /*
2783  * All control commands are sent via /dev/ublk-control, so we have to check
2784  * the destination device's permission
2785  */
2786 static int ublk_char_dev_permission(struct ublk_device *ub,
2787                 const char *dev_path, int mask)
2788 {
2789         int err;
2790         struct path path;
2791         struct kstat stat;
2792
2793         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2794         if (err)
2795                 return err;
2796
2797         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2798         if (err)
2799                 goto exit;
2800
2801         err = -EPERM;
2802         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2803                 goto exit;
2804
2805         err = inode_permission(&nop_mnt_idmap,
2806                         d_backing_inode(path.dentry), mask);
2807 exit:
2808         path_put(&path);
2809         return err;
2810 }
2811
2812 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2813                 struct io_uring_cmd *cmd)
2814 {
2815         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2816         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2817         void __user *argp = (void __user *)(unsigned long)header->addr;
2818         char *dev_path = NULL;
2819         int ret = 0;
2820         int mask;
2821
2822         if (!unprivileged) {
2823                 if (!capable(CAP_SYS_ADMIN))
2824                         return -EPERM;
2825                 /*
2826                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2827                  * char_dev_path in payload too, since userspace may not
2828                  * know if the specified device is created as unprivileged
2829                  * mode.
2830                  */
2831                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2832                         return 0;
2833         }
2834
2835         /*
2836          * User has to provide the char device path for unprivileged ublk
2837          *
2838          * header->addr always points to the dev path buffer, and
2839          * header->dev_path_len records length of dev path buffer.
2840          */
2841         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2842                 return -EINVAL;
2843
2844         if (header->len < header->dev_path_len)
2845                 return -EINVAL;
2846
2847         dev_path = memdup_user_nul(argp, header->dev_path_len);
2848         if (IS_ERR(dev_path))
2849                 return PTR_ERR(dev_path);
2850
2851         ret = -EINVAL;
2852         switch (_IOC_NR(cmd->cmd_op)) {
2853         case UBLK_CMD_GET_DEV_INFO:
2854         case UBLK_CMD_GET_DEV_INFO2:
2855         case UBLK_CMD_GET_QUEUE_AFFINITY:
2856         case UBLK_CMD_GET_PARAMS:
2857         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2858                 mask = MAY_READ;
2859                 break;
2860         case UBLK_CMD_START_DEV:
2861         case UBLK_CMD_STOP_DEV:
2862         case UBLK_CMD_ADD_DEV:
2863         case UBLK_CMD_DEL_DEV:
2864         case UBLK_CMD_SET_PARAMS:
2865         case UBLK_CMD_START_USER_RECOVERY:
2866         case UBLK_CMD_END_USER_RECOVERY:
2867                 mask = MAY_READ | MAY_WRITE;
2868                 break;
2869         default:
2870                 goto exit;
2871         }
2872
2873         ret = ublk_char_dev_permission(ub, dev_path, mask);
2874         if (!ret) {
2875                 header->len -= header->dev_path_len;
2876                 header->addr += header->dev_path_len;
2877         }
2878         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2879                         __func__, ub->ub_number, cmd->cmd_op,
2880                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2881                         dev_path, ret);
2882 exit:
2883         kfree(dev_path);
2884         return ret;
2885 }
2886
2887 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2888                 unsigned int issue_flags)
2889 {
2890         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2891         struct ublk_device *ub = NULL;
2892         u32 cmd_op = cmd->cmd_op;
2893         int ret = -EINVAL;
2894
2895         if (issue_flags & IO_URING_F_NONBLOCK)
2896                 return -EAGAIN;
2897
2898         ublk_ctrl_cmd_dump(cmd);
2899
2900         if (!(issue_flags & IO_URING_F_SQE128))
2901                 goto out;
2902
2903         ret = ublk_check_cmd_op(cmd_op);
2904         if (ret)
2905                 goto out;
2906
2907         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2908                 ret = ublk_ctrl_get_features(cmd);
2909                 goto out;
2910         }
2911
2912         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2913                 ret = -ENODEV;
2914                 ub = ublk_get_device_from_id(header->dev_id);
2915                 if (!ub)
2916                         goto out;
2917
2918                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2919                 if (ret)
2920                         goto put_dev;
2921         }
2922
2923         switch (_IOC_NR(cmd_op)) {
2924         case UBLK_CMD_START_DEV:
2925                 ret = ublk_ctrl_start_dev(ub, cmd);
2926                 break;
2927         case UBLK_CMD_STOP_DEV:
2928                 ret = ublk_ctrl_stop_dev(ub);
2929                 break;
2930         case UBLK_CMD_GET_DEV_INFO:
2931         case UBLK_CMD_GET_DEV_INFO2:
2932                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2933                 break;
2934         case UBLK_CMD_ADD_DEV:
2935                 ret = ublk_ctrl_add_dev(cmd);
2936                 break;
2937         case UBLK_CMD_DEL_DEV:
2938                 ret = ublk_ctrl_del_dev(&ub, true);
2939                 break;
2940         case UBLK_CMD_DEL_DEV_ASYNC:
2941                 ret = ublk_ctrl_del_dev(&ub, false);
2942                 break;
2943         case UBLK_CMD_GET_QUEUE_AFFINITY:
2944                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2945                 break;
2946         case UBLK_CMD_GET_PARAMS:
2947                 ret = ublk_ctrl_get_params(ub, cmd);
2948                 break;
2949         case UBLK_CMD_SET_PARAMS:
2950                 ret = ublk_ctrl_set_params(ub, cmd);
2951                 break;
2952         case UBLK_CMD_START_USER_RECOVERY:
2953                 ret = ublk_ctrl_start_recovery(ub, cmd);
2954                 break;
2955         case UBLK_CMD_END_USER_RECOVERY:
2956                 ret = ublk_ctrl_end_recovery(ub, cmd);
2957                 break;
2958         default:
2959                 ret = -ENOTSUPP;
2960                 break;
2961         }
2962
2963  put_dev:
2964         if (ub)
2965                 ublk_put_device(ub);
2966  out:
2967         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2968         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2969                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2970         return -EIOCBQUEUED;
2971 }
2972
2973 static const struct file_operations ublk_ctl_fops = {
2974         .open           = nonseekable_open,
2975         .uring_cmd      = ublk_ctrl_uring_cmd,
2976         .owner          = THIS_MODULE,
2977         .llseek         = noop_llseek,
2978 };
2979
2980 static struct miscdevice ublk_misc = {
2981         .minor          = MISC_DYNAMIC_MINOR,
2982         .name           = "ublk-control",
2983         .fops           = &ublk_ctl_fops,
2984 };
2985
2986 static int __init ublk_init(void)
2987 {
2988         int ret;
2989
2990         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2991                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2992
2993         init_waitqueue_head(&ublk_idr_wq);
2994
2995         ret = misc_register(&ublk_misc);
2996         if (ret)
2997                 return ret;
2998
2999         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
3000         if (ret)
3001                 goto unregister_mis;
3002
3003         ret = class_register(&ublk_chr_class);
3004         if (ret)
3005                 goto free_chrdev_region;
3006
3007         return 0;
3008
3009 free_chrdev_region:
3010         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
3011 unregister_mis:
3012         misc_deregister(&ublk_misc);
3013         return ret;
3014 }
3015
3016 static void __exit ublk_exit(void)
3017 {
3018         struct ublk_device *ub;
3019         int id;
3020
3021         idr_for_each_entry(&ublk_index_idr, ub, id)
3022                 ublk_remove(ub);
3023
3024         class_unregister(&ublk_chr_class);
3025         misc_deregister(&ublk_misc);
3026
3027         idr_destroy(&ublk_index_idr);
3028         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
3029 }
3030
3031 module_init(ublk_init);
3032 module_exit(ublk_exit);
3033
3034 static int ublk_set_max_ublks(const char *buf, const struct kernel_param *kp)
3035 {
3036         return param_set_uint_minmax(buf, kp, 0, UBLK_MAX_UBLKS);
3037 }
3038
3039 static int ublk_get_max_ublks(char *buf, const struct kernel_param *kp)
3040 {
3041         return sysfs_emit(buf, "%u\n", ublks_max);
3042 }
3043
3044 static const struct kernel_param_ops ublk_max_ublks_ops = {
3045         .set = ublk_set_max_ublks,
3046         .get = ublk_get_max_ublks,
3047 };
3048
3049 module_param_cb(ublks_max, &ublk_max_ublks_ops, &ublks_max, 0644);
3050 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
3051
3052 MODULE_AUTHOR("Ming Lei <[email protected]>");
3053 MODULE_DESCRIPTION("Userspace block device");
3054 MODULE_LICENSE("GPL");