drivers/block/ublk_drv.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Userspace block device - block device which IO is handled from userspace
   4  *
   5  * Take full use of io_uring passthrough command for communicating with
   6  * ublk userspace daemon(ublksrvd) for handling basic IO request.
   7  *
   8  * Copyright 2022 Ming Lei <ming.lei@redhat.com>
   9  *
  10  * (part of code stolen from loop.c)
  11  */
  12 #include <linux/module.h>
  13 #include <linux/moduleparam.h>
  14 #include <linux/sched.h>
  15 #include <linux/fs.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/file.h>
  18 #include <linux/stat.h>
  19 #include <linux/errno.h>
  20 #include <linux/major.h>
  21 #include <linux/wait.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/init.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/compat.h>
  27 #include <linux/mutex.h>
  28 #include <linux/writeback.h>
  29 #include <linux/completion.h>
  30 #include <linux/highmem.h>
  31 #include <linux/sysfs.h>
  32 #include <linux/miscdevice.h>
  33 #include <linux/falloc.h>
  34 #include <linux/uio.h>
  35 #include <linux/ioprio.h>
  36 #include <linux/sched/mm.h>
  37 #include <linux/uaccess.h>
  38 #include <linux/cdev.h>
  39 #include <linux/io_uring.h>
  40 #include <linux/blk-mq.h>
  41 #include <linux/delay.h>
  42 #include <linux/mm.h>
  43 #include <asm/page.h>
  44 #include <linux/task_work.h>
  45 #include <linux/namei.h>
  46 #include <linux/kref.h>
  47 #include <uapi/linux/ublk_cmd.h>
  48
  49 #define UBLK_MINORS             (1U << MINORBITS)
  50
  51 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  52 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  53                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  54                 | UBLK_F_NEED_GET_DATA \
  55                 | UBLK_F_USER_RECOVERY \
  56                 | UBLK_F_USER_RECOVERY_REISSUE \
  57                 | UBLK_F_UNPRIVILEGED_DEV \
  58                 | UBLK_F_CMD_IOCTL_ENCODE \
  59                 | UBLK_F_USER_COPY \
  60                 | UBLK_F_ZONED)
  61
  62 /* All UBLK_PARAM_TYPE_* should be included here */
  63 #define UBLK_PARAM_TYPE_ALL                                \
  64         (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD | \
  65          UBLK_PARAM_TYPE_DEVT | UBLK_PARAM_TYPE_ZONED)
  66
  67 struct ublk_rq_data {
  68         struct llist_node node;
  69
  70         struct kref ref;
  71         __u64 sector;
  72         __u32 operation;
  73         __u32 nr_zones;
  74 };
  75
  76 struct ublk_uring_cmd_pdu {
  77         struct ublk_queue *ubq;
  78         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 static int ublk_get_nr_zones(const struct ublk_device *ub)
 215 {
 216         const struct ublk_param_basic *p = &ub->params.basic;
 217
 218         /* Zone size is a power of 2 */
 219         return p->dev_sectors >> ilog2(p->chunk_sectors);
 220 }
 221
 222 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 223 {
 224         return blk_revalidate_disk_zones(ub->ub_disk, NULL);
 225 }
 226
 227 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 228 {
 229         const struct ublk_param_zoned *p = &ub->params.zoned;
 230         int nr_zones;
 231
 232         if (!ublk_dev_is_zoned(ub))
 233                 return -EINVAL;
 234
 235         if (!p->max_zone_append_sectors)
 236                 return -EINVAL;
 237
 238         nr_zones = ublk_get_nr_zones(ub);
 239
 240         if (p->max_active_zones > nr_zones)
 241                 return -EINVAL;
 242
 243         if (p->max_open_zones > nr_zones)
 244                 return -EINVAL;
 245
 246         return 0;
 247 }
 248
 249 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
 250 {
 251         const struct ublk_param_zoned *p = &ub->params.zoned;
 252
 253         disk_set_zoned(ub->ub_disk, BLK_ZONED_HM);
 254         blk_queue_flag_set(QUEUE_FLAG_ZONE_RESETALL, ub->ub_disk->queue);
 255         blk_queue_required_elevator_features(ub->ub_disk->queue,
 256                                              ELEVATOR_F_ZBD_SEQ_WRITE);
 257         disk_set_max_active_zones(ub->ub_disk, p->max_active_zones);
 258         disk_set_max_open_zones(ub->ub_disk, p->max_open_zones);
 259         blk_queue_max_zone_append_sectors(ub->ub_disk->queue, p->max_zone_append_sectors);
 260
 261         ub->ub_disk->nr_zones = ublk_get_nr_zones(ub);
 262
 263         return 0;
 264 }
 265
 266 /* Based on virtblk_alloc_report_buffer */
 267 static void *ublk_alloc_report_buffer(struct ublk_device *ublk,
 268                                       unsigned int nr_zones, size_t *buflen)
 269 {
 270         struct request_queue *q = ublk->ub_disk->queue;
 271         size_t bufsize;
 272         void *buf;
 273
 274         nr_zones = min_t(unsigned int, nr_zones,
 275                          ublk->ub_disk->nr_zones);
 276
 277         bufsize = nr_zones * sizeof(struct blk_zone);
 278         bufsize =
 279                 min_t(size_t, bufsize, queue_max_hw_sectors(q) << SECTOR_SHIFT);
 280
 281         while (bufsize >= sizeof(struct blk_zone)) {
 282                 buf = kvmalloc(bufsize, GFP_KERNEL | __GFP_NORETRY);
 283                 if (buf) {
 284                         *buflen = bufsize;
 285                         return buf;
 286                 }
 287                 bufsize >>= 1;
 288         }
 289
 290         *buflen = 0;
 291         return NULL;
 292 }
 293
 294 static int ublk_report_zones(struct gendisk *disk, sector_t sector,
 295                       unsigned int nr_zones, report_zones_cb cb, void *data)
 296 {
 297         struct ublk_device *ub = disk->private_data;
 298         unsigned int zone_size_sectors = disk->queue->limits.chunk_sectors;
 299         unsigned int first_zone = sector >> ilog2(zone_size_sectors);
 300         unsigned int done_zones = 0;
 301         unsigned int max_zones_per_request;
 302         int ret;
 303         struct blk_zone *buffer;
 304         size_t buffer_length;
 305
 306         nr_zones = min_t(unsigned int, ub->ub_disk->nr_zones - first_zone,
 307                          nr_zones);
 308
 309         buffer = ublk_alloc_report_buffer(ub, nr_zones, &buffer_length);
 310         if (!buffer)
 311                 return -ENOMEM;
 312
 313         max_zones_per_request = buffer_length / sizeof(struct blk_zone);
 314
 315         while (done_zones < nr_zones) {
 316                 unsigned int remaining_zones = nr_zones - done_zones;
 317                 unsigned int zones_in_request =
 318                         min_t(unsigned int, remaining_zones, max_zones_per_request);
 319                 struct request *req;
 320                 struct ublk_rq_data *pdu;
 321                 blk_status_t status;
 322
 323                 memset(buffer, 0, buffer_length);
 324
 325                 req = blk_mq_alloc_request(disk->queue, REQ_OP_DRV_IN, 0);
 326                 if (IS_ERR(req)) {
 327                         ret = PTR_ERR(req);
 328                         goto out;
 329                 }
 330
 331                 pdu = blk_mq_rq_to_pdu(req);
 332                 pdu->operation = UBLK_IO_OP_REPORT_ZONES;
 333                 pdu->sector = sector;
 334                 pdu->nr_zones = zones_in_request;
 335
 336                 ret = blk_rq_map_kern(disk->queue, req, buffer, buffer_length,
 337                                         GFP_KERNEL);
 338                 if (ret) {
 339                         blk_mq_free_request(req);
 340                         goto out;
 341                 }
 342
 343                 status = blk_execute_rq(req, 0);
 344                 ret = blk_status_to_errno(status);
 345                 blk_mq_free_request(req);
 346                 if (ret)
 347                         goto out;
 348
 349                 for (unsigned int i = 0; i < zones_in_request; i++) {
 350                         struct blk_zone *zone = buffer + i;
 351
 352                         /* A zero length zone means no more zones in this response */
 353                         if (!zone->len)
 354                                 break;
 355
 356                         ret = cb(zone, i, data);
 357                         if (ret)
 358                                 goto out;
 359
 360                         done_zones++;
 361                         sector += zone_size_sectors;
 362
 363                 }
 364         }
 365
 366         ret = done_zones;
 367
 368 out:
 369         kvfree(buffer);
 370         return ret;
 371 }
 372
 373 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 374                                          struct request *req)
 375 {
 376         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 377         struct ublk_io *io = &ubq->ios[req->tag];
 378         struct ublk_rq_data *pdu = blk_mq_rq_to_pdu(req);
 379         u32 ublk_op;
 380
 381         switch (req_op(req)) {
 382         case REQ_OP_ZONE_OPEN:
 383                 ublk_op = UBLK_IO_OP_ZONE_OPEN;
 384                 break;
 385         case REQ_OP_ZONE_CLOSE:
 386                 ublk_op = UBLK_IO_OP_ZONE_CLOSE;
 387                 break;
 388         case REQ_OP_ZONE_FINISH:
 389                 ublk_op = UBLK_IO_OP_ZONE_FINISH;
 390                 break;
 391         case REQ_OP_ZONE_RESET:
 392                 ublk_op = UBLK_IO_OP_ZONE_RESET;
 393                 break;
 394         case REQ_OP_ZONE_APPEND:
 395                 ublk_op = UBLK_IO_OP_ZONE_APPEND;
 396                 break;
 397         case REQ_OP_ZONE_RESET_ALL:
 398                 ublk_op = UBLK_IO_OP_ZONE_RESET_ALL;
 399                 break;
 400         case REQ_OP_DRV_IN:
 401                 ublk_op = pdu->operation;
 402                 switch (ublk_op) {
 403                 case UBLK_IO_OP_REPORT_ZONES:
 404                         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 405                         iod->nr_zones = pdu->nr_zones;
 406                         iod->start_sector = pdu->sector;
 407                         return BLK_STS_OK;
 408                 default:
 409                         return BLK_STS_IOERR;
 410                 }
 411         case REQ_OP_DRV_OUT:
 412                 /* We do not support drv_out */
 413                 return BLK_STS_NOTSUPP;
 414         default:
 415                 return BLK_STS_IOERR;
 416         }
 417
 418         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 419         iod->nr_sectors = blk_rq_sectors(req);
 420         iod->start_sector = blk_rq_pos(req);
 421         iod->addr = io->addr;
 422
 423         return BLK_STS_OK;
 424 }
 425
 426 #else
 427
 428 #define ublk_report_zones (NULL)
 429
 430 static int ublk_dev_param_zoned_validate(const struct ublk_device *ub)
 431 {
 432         return -EOPNOTSUPP;
 433 }
 434
 435 static int ublk_dev_param_zoned_apply(struct ublk_device *ub)
 436 {
 437         return -EOPNOTSUPP;
 438 }
 439
 440 static int ublk_revalidate_disk_zones(struct ublk_device *ub)
 441 {
 442         return 0;
 443 }
 444
 445 static blk_status_t ublk_setup_iod_zoned(struct ublk_queue *ubq,
 446                                          struct request *req)
 447 {
 448         return BLK_STS_NOTSUPP;
 449 }
 450
 451 #endif
 452
 453 static inline void __ublk_complete_rq(struct request *req);
 454 static void ublk_complete_rq(struct kref *ref);
 455
 456 static dev_t ublk_chr_devt;
 457 static const struct class ublk_chr_class = {
 458         .name = "ublk-char",
 459 };
 460
 461 static DEFINE_IDR(ublk_index_idr);
 462 static DEFINE_SPINLOCK(ublk_idr_lock);
 463 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 464
 465 static DEFINE_MUTEX(ublk_ctl_mutex);
 466
 467 /*
 468  * Max ublk devices allowed to add
 469  *
 470  * It can be extended to one per-user limit in future or even controlled
 471  * by cgroup.
 472  */
 473 #define UBLK_MAX_UBLKS UBLK_MINORS
 474 static unsigned int ublks_max = 64;
 475 static unsigned int ublks_added;        /* protected by ublk_ctl_mutex */
 476
 477 static struct miscdevice ublk_misc;
 478
 479 static inline unsigned ublk_pos_to_hwq(loff_t pos)
 480 {
 481         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_QID_OFF) &
 482                 UBLK_QID_BITS_MASK;
 483 }
 484
 485 static inline unsigned ublk_pos_to_buf_off(loff_t pos)
 486 {
 487         return (pos - UBLKSRV_IO_BUF_OFFSET) & UBLK_IO_BUF_BITS_MASK;
 488 }
 489
 490 static inline unsigned ublk_pos_to_tag(loff_t pos)
 491 {
 492         return ((pos - UBLKSRV_IO_BUF_OFFSET) >> UBLK_TAG_OFF) &
 493                 UBLK_TAG_BITS_MASK;
 494 }
 495
 496 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 497 {
 498         struct request_queue *q = ub->ub_disk->queue;
 499         const struct ublk_param_basic *p = &ub->params.basic;
 500
 501         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 502         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 503         blk_queue_io_min(q, 1 << p->io_min_shift);
 504         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 505
 506         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 507                         p->attrs & UBLK_ATTR_FUA);
 508         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 509                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 510         else
 511                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 512
 513         blk_queue_max_hw_sectors(q, p->max_sectors);
 514         blk_queue_chunk_sectors(q, p->chunk_sectors);
 515         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 516
 517         if (p->attrs & UBLK_ATTR_READ_ONLY)
 518                 set_disk_ro(ub->ub_disk, true);
 519
 520         set_capacity(ub->ub_disk, p->dev_sectors);
 521 }
 522
 523 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 524 {
 525         struct request_queue *q = ub->ub_disk->queue;
 526         const struct ublk_param_discard *p = &ub->params.discard;
 527
 528         q->limits.discard_alignment = p->discard_alignment;
 529         q->limits.discard_granularity = p->discard_granularity;
 530         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 531         blk_queue_max_write_zeroes_sectors(q,
 532                         p->max_write_zeroes_sectors);
 533         blk_queue_max_discard_segments(q, p->max_discard_segments);
 534 }
 535
 536 static int ublk_validate_params(const struct ublk_device *ub)
 537 {
 538         /* basic param is the only one which must be set */
 539         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 540                 const struct ublk_param_basic *p = &ub->params.basic;
 541
 542                 if (p->logical_bs_shift > PAGE_SHIFT || p->logical_bs_shift < 9)
 543                         return -EINVAL;
 544
 545                 if (p->logical_bs_shift > p->physical_bs_shift)
 546                         return -EINVAL;
 547
 548                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 549                         return -EINVAL;
 550
 551                 if (ublk_dev_is_zoned(ub) && !p->chunk_sectors)
 552                         return -EINVAL;
 553         } else
 554                 return -EINVAL;
 555
 556         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 557                 const struct ublk_param_discard *p = &ub->params.discard;
 558
 559                 /* So far, only support single segment discard */
 560                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 561                         return -EINVAL;
 562
 563                 if (!p->discard_granularity)
 564                         return -EINVAL;
 565         }
 566
 567         /* dev_t is read-only */
 568         if (ub->params.types & UBLK_PARAM_TYPE_DEVT)
 569                 return -EINVAL;
 570
 571         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 572                 return ublk_dev_param_zoned_validate(ub);
 573         else if (ublk_dev_is_zoned(ub))
 574                 return -EINVAL;
 575
 576         return 0;
 577 }
 578
 579 static int ublk_apply_params(struct ublk_device *ub)
 580 {
 581         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 582                 return -EINVAL;
 583
 584         ublk_dev_param_basic_apply(ub);
 585
 586         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 587                 ublk_dev_param_discard_apply(ub);
 588
 589         if (ub->params.types & UBLK_PARAM_TYPE_ZONED)
 590                 return ublk_dev_param_zoned_apply(ub);
 591
 592         return 0;
 593 }
 594
 595 static inline bool ublk_support_user_copy(const struct ublk_queue *ubq)
 596 {
 597         return ubq->flags & UBLK_F_USER_COPY;
 598 }
 599
 600 static inline bool ublk_need_req_ref(const struct ublk_queue *ubq)
 601 {
 602         /*
 603          * read()/write() is involved in user copy, so request reference
 604          * has to be grabbed
 605          */
 606         return ublk_support_user_copy(ubq);
 607 }
 608
 609 static inline void ublk_init_req_ref(const struct ublk_queue *ubq,
 610                 struct request *req)
 611 {
 612         if (ublk_need_req_ref(ubq)) {
 613                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 614
 615                 kref_init(&data->ref);
 616         }
 617 }
 618
 619 static inline bool ublk_get_req_ref(const struct ublk_queue *ubq,
 620                 struct request *req)
 621 {
 622         if (ublk_need_req_ref(ubq)) {
 623                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 624
 625                 return kref_get_unless_zero(&data->ref);
 626         }
 627
 628         return true;
 629 }
 630
 631 static inline void ublk_put_req_ref(const struct ublk_queue *ubq,
 632                 struct request *req)
 633 {
 634         if (ublk_need_req_ref(ubq)) {
 635                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 636
 637                 kref_put(&data->ref, ublk_complete_rq);
 638         } else {
 639                 __ublk_complete_rq(req);
 640         }
 641 }
 642
 643 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 644 {
 645         return ubq->flags & UBLK_F_NEED_GET_DATA;
 646 }
 647
 648 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 649 {
 650         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 651                 return ub;
 652         return NULL;
 653 }
 654
 655 static void ublk_put_device(struct ublk_device *ub)
 656 {
 657         put_device(&ub->cdev_dev);
 658 }
 659
 660 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 661                 int qid)
 662 {
 663        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 664 }
 665
 666 static inline bool ublk_rq_has_data(const struct request *rq)
 667 {
 668         return bio_has_data(rq->bio);
 669 }
 670
 671 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 672                 int tag)
 673 {
 674         return (struct ublksrv_io_desc *)
 675                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 676 }
 677
 678 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 679 {
 680         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 681 }
 682
 683 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 684 {
 685         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 686
 687         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 688                         PAGE_SIZE);
 689 }
 690
 691 static inline bool ublk_queue_can_use_recovery_reissue(
 692                 struct ublk_queue *ubq)
 693 {
 694         return (ubq->flags & UBLK_F_USER_RECOVERY) &&
 695                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE);
 696 }
 697
 698 static inline bool ublk_queue_can_use_recovery(
 699                 struct ublk_queue *ubq)
 700 {
 701         return ubq->flags & UBLK_F_USER_RECOVERY;
 702 }
 703
 704 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 705 {
 706         return ub->dev_info.flags & UBLK_F_USER_RECOVERY;
 707 }
 708
 709 static void ublk_free_disk(struct gendisk *disk)
 710 {
 711         struct ublk_device *ub = disk->private_data;
 712
 713         clear_bit(UB_STATE_USED, &ub->state);
 714         put_device(&ub->cdev_dev);
 715 }
 716
 717 static void ublk_store_owner_uid_gid(unsigned int *owner_uid,
 718                 unsigned int *owner_gid)
 719 {
 720         kuid_t uid;
 721         kgid_t gid;
 722
 723         current_uid_gid(&uid, &gid);
 724
 725         *owner_uid = from_kuid(&init_user_ns, uid);
 726         *owner_gid = from_kgid(&init_user_ns, gid);
 727 }
 728
 729 static int ublk_open(struct gendisk *disk, blk_mode_t mode)
 730 {
 731         struct ublk_device *ub = disk->private_data;
 732
 733         if (capable(CAP_SYS_ADMIN))
 734                 return 0;
 735
 736         /*
 737          * If it is one unprivileged device, only owner can open
 738          * the disk. Otherwise it could be one trap made by one
 739          * evil user who grants this disk's privileges to other
 740          * users deliberately.
 741          *
 742          * This way is reasonable too given anyone can create
 743          * unprivileged device, and no need other's grant.
 744          */
 745         if (ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV) {
 746                 unsigned int curr_uid, curr_gid;
 747
 748                 ublk_store_owner_uid_gid(&curr_uid, &curr_gid);
 749
 750                 if (curr_uid != ub->dev_info.owner_uid || curr_gid !=
 751                                 ub->dev_info.owner_gid)
 752                         return -EPERM;
 753         }
 754
 755         return 0;
 756 }
 757
 758 static const struct block_device_operations ub_fops = {
 759         .owner =        THIS_MODULE,
 760         .open =         ublk_open,
 761         .free_disk =    ublk_free_disk,
 762         .report_zones = ublk_report_zones,
 763 };
 764
 765 #define UBLK_MAX_PIN_PAGES      32
 766
 767 struct ublk_io_iter {
 768         struct page *pages[UBLK_MAX_PIN_PAGES];
 769         struct bio *bio;
 770         struct bvec_iter iter;
 771 };
 772
 773 /* return how many pages are copied */
 774 static void ublk_copy_io_pages(struct ublk_io_iter *data,
 775                 size_t total, size_t pg_off, int dir)
 776 {
 777         unsigned done = 0;
 778         unsigned pg_idx = 0;
 779
 780         while (done < total) {
 781                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 782                 unsigned int bytes = min3(bv.bv_len, (unsigned)total - done,
 783                                 (unsigned)(PAGE_SIZE - pg_off));
 784                 void *bv_buf = bvec_kmap_local(&bv);
 785                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 786
 787                 if (dir == ITER_DEST)
 788                         memcpy(pg_buf + pg_off, bv_buf, bytes);
 789                 else
 790                         memcpy(bv_buf, pg_buf + pg_off, bytes);
 791
 792                 kunmap_local(pg_buf);
 793                 kunmap_local(bv_buf);
 794
 795                 /* advance page array */
 796                 pg_off += bytes;
 797                 if (pg_off == PAGE_SIZE) {
 798                         pg_idx += 1;
 799                         pg_off = 0;
 800                 }
 801
 802                 done += bytes;
 803
 804                 /* advance bio */
 805                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 806                 if (!data->iter.bi_size) {
 807                         data->bio = data->bio->bi_next;
 808                         if (data->bio == NULL)
 809                                 break;
 810                         data->iter = data->bio->bi_iter;
 811                 }
 812         }
 813 }
 814
 815 static bool ublk_advance_io_iter(const struct request *req,
 816                 struct ublk_io_iter *iter, unsigned int offset)
 817 {
 818         struct bio *bio = req->bio;
 819
 820         for_each_bio(bio) {
 821                 if (bio->bi_iter.bi_size > offset) {
 822                         iter->bio = bio;
 823                         iter->iter = bio->bi_iter;
 824                         bio_advance_iter(iter->bio, &iter->iter, offset);
 825                         return true;
 826                 }
 827                 offset -= bio->bi_iter.bi_size;
 828         }
 829         return false;
 830 }
 831
 832 /*
 833  * Copy data between request pages and io_iter, and 'offset'
 834  * is the start point of linear offset of request.
 835  */
 836 static size_t ublk_copy_user_pages(const struct request *req,
 837                 unsigned offset, struct iov_iter *uiter, int dir)
 838 {
 839         struct ublk_io_iter iter;
 840         size_t done = 0;
 841
 842         if (!ublk_advance_io_iter(req, &iter, offset))
 843                 return 0;
 844
 845         while (iov_iter_count(uiter) && iter.bio) {
 846                 unsigned nr_pages;
 847                 ssize_t len;
 848                 size_t off;
 849                 int i;
 850
 851                 len = iov_iter_get_pages2(uiter, iter.pages,
 852                                 iov_iter_count(uiter),
 853                                 UBLK_MAX_PIN_PAGES, &off);
 854                 if (len <= 0)
 855                         return done;
 856
 857                 ublk_copy_io_pages(&iter, len, off, dir);
 858                 nr_pages = DIV_ROUND_UP(len + off, PAGE_SIZE);
 859                 for (i = 0; i < nr_pages; i++) {
 860                         if (dir == ITER_DEST)
 861                                 set_page_dirty(iter.pages[i]);
 862                         put_page(iter.pages[i]);
 863                 }
 864                 done += len;
 865         }
 866
 867         return done;
 868 }
 869
 870 static inline bool ublk_need_map_req(const struct request *req)
 871 {
 872         return ublk_rq_has_data(req) && req_op(req) == REQ_OP_WRITE;
 873 }
 874
 875 static inline bool ublk_need_unmap_req(const struct request *req)
 876 {
 877         return ublk_rq_has_data(req) &&
 878                (req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN);
 879 }
 880
 881 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 882                 struct ublk_io *io)
 883 {
 884         const unsigned int rq_bytes = blk_rq_bytes(req);
 885
 886         if (ublk_support_user_copy(ubq))
 887                 return rq_bytes;
 888
 889         /*
 890          * no zero copy, we delay copy WRITE request data into ublksrv
 891          * context and the big benefit is that pinning pages in current
 892          * context is pretty fast, see ublk_pin_user_pages
 893          */
 894         if (ublk_need_map_req(req)) {
 895                 struct iov_iter iter;
 896                 struct iovec iov;
 897                 const int dir = ITER_DEST;
 898
 899                 import_single_range(dir, u64_to_user_ptr(io->addr), rq_bytes,
 900                                 &iov, &iter);
 901
 902                 return ublk_copy_user_pages(req, 0, &iter, dir);
 903         }
 904         return rq_bytes;
 905 }
 906
 907 static int ublk_unmap_io(const struct ublk_queue *ubq,
 908                 const struct request *req,
 909                 struct ublk_io *io)
 910 {
 911         const unsigned int rq_bytes = blk_rq_bytes(req);
 912
 913         if (ublk_support_user_copy(ubq))
 914                 return rq_bytes;
 915
 916         if (ublk_need_unmap_req(req)) {
 917                 struct iov_iter iter;
 918                 struct iovec iov;
 919                 const int dir = ITER_SOURCE;
 920
 921                 WARN_ON_ONCE(io->res > rq_bytes);
 922
 923                 import_single_range(dir, u64_to_user_ptr(io->addr), io->res,
 924                                 &iov, &iter);
 925                 return ublk_copy_user_pages(req, 0, &iter, dir);
 926         }
 927         return rq_bytes;
 928 }
 929
 930 static inline unsigned int ublk_req_build_flags(struct request *req)
 931 {
 932         unsigned flags = 0;
 933
 934         if (req->cmd_flags & REQ_FAILFAST_DEV)
 935                 flags |= UBLK_IO_F_FAILFAST_DEV;
 936
 937         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 938                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 939
 940         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 941                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 942
 943         if (req->cmd_flags & REQ_META)
 944                 flags |= UBLK_IO_F_META;
 945
 946         if (req->cmd_flags & REQ_FUA)
 947                 flags |= UBLK_IO_F_FUA;
 948
 949         if (req->cmd_flags & REQ_NOUNMAP)
 950                 flags |= UBLK_IO_F_NOUNMAP;
 951
 952         if (req->cmd_flags & REQ_SWAP)
 953                 flags |= UBLK_IO_F_SWAP;
 954
 955         return flags;
 956 }
 957
 958 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 959 {
 960         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 961         struct ublk_io *io = &ubq->ios[req->tag];
 962         enum req_op op = req_op(req);
 963         u32 ublk_op;
 964
 965         if (!ublk_queue_is_zoned(ubq) &&
 966             (op_is_zone_mgmt(op) || op == REQ_OP_ZONE_APPEND))
 967                 return BLK_STS_IOERR;
 968
 969         switch (req_op(req)) {
 970         case REQ_OP_READ:
 971                 ublk_op = UBLK_IO_OP_READ;
 972                 break;
 973         case REQ_OP_WRITE:
 974                 ublk_op = UBLK_IO_OP_WRITE;
 975                 break;
 976         case REQ_OP_FLUSH:
 977                 ublk_op = UBLK_IO_OP_FLUSH;
 978                 break;
 979         case REQ_OP_DISCARD:
 980                 ublk_op = UBLK_IO_OP_DISCARD;
 981                 break;
 982         case REQ_OP_WRITE_ZEROES:
 983                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 984                 break;
 985         default:
 986                 if (ublk_queue_is_zoned(ubq))
 987                         return ublk_setup_iod_zoned(ubq, req);
 988                 return BLK_STS_IOERR;
 989         }
 990
 991         /* need to translate since kernel may change */
 992         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 993         iod->nr_sectors = blk_rq_sectors(req);
 994         iod->start_sector = blk_rq_pos(req);
 995         iod->addr = io->addr;
 996
 997         return BLK_STS_OK;
 998 }
 999
1000 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
1001                 struct io_uring_cmd *ioucmd)
1002 {
1003         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
1004 }
1005
1006 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
1007 {
1008         return ubq->ubq_daemon->flags & PF_EXITING;
1009 }
1010
1011 /* todo: handle partial completion */
1012 static inline void __ublk_complete_rq(struct request *req)
1013 {
1014         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1015         struct ublk_io *io = &ubq->ios[req->tag];
1016         unsigned int unmapped_bytes;
1017         blk_status_t res = BLK_STS_OK;
1018
1019         /* called from ublk_abort_queue() code path */
1020         if (io->flags & UBLK_IO_FLAG_ABORTED) {
1021                 res = BLK_STS_IOERR;
1022                 goto exit;
1023         }
1024
1025         /* failed read IO if nothing is read */
1026         if (!io->res && req_op(req) == REQ_OP_READ)
1027                 io->res = -EIO;
1028
1029         if (io->res < 0) {
1030                 res = errno_to_blk_status(io->res);
1031                 goto exit;
1032         }
1033
1034         /*
1035          * FLUSH, DISCARD or WRITE_ZEROES usually won't return bytes returned, so end them
1036          * directly.
1037          *
1038          * Both the two needn't unmap.
1039          */
1040         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE &&
1041             req_op(req) != REQ_OP_DRV_IN)
1042                 goto exit;
1043
1044         /* for READ request, writing data in iod->addr to rq buffers */
1045         unmapped_bytes = ublk_unmap_io(ubq, req, io);
1046
1047         /*
1048          * Extremely impossible since we got data filled in just before
1049          *
1050          * Re-read simply for this unlikely case.
1051          */
1052         if (unlikely(unmapped_bytes < io->res))
1053                 io->res = unmapped_bytes;
1054
1055         if (blk_update_request(req, BLK_STS_OK, io->res))
1056                 blk_mq_requeue_request(req, true);
1057         else
1058                 __blk_mq_end_request(req, BLK_STS_OK);
1059
1060         return;
1061 exit:
1062         blk_mq_end_request(req, res);
1063 }
1064
1065 static void ublk_complete_rq(struct kref *ref)
1066 {
1067         struct ublk_rq_data *data = container_of(ref, struct ublk_rq_data,
1068                         ref);
1069         struct request *req = blk_mq_rq_from_pdu(data);
1070
1071         __ublk_complete_rq(req);
1072 }
1073
1074 /*
1075  * Since __ublk_rq_task_work always fails requests immediately during
1076  * exiting, __ublk_fail_req() is only called from abort context during
1077  * exiting. So lock is unnecessary.
1078  *
1079  * Also aborting may not be started yet, keep in mind that one failed
1080  * request may be issued by block layer again.
1081  */
1082 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
1083                 struct request *req)
1084 {
1085         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
1086
1087         if (ublk_queue_can_use_recovery_reissue(ubq))
1088                 blk_mq_requeue_request(req, false);
1089         else
1090                 ublk_put_req_ref(ubq, req);
1091 }
1092
1093 static void ubq_complete_io_cmd(struct ublk_io *io, int res,
1094                                 unsigned issue_flags)
1095 {
1096         /* mark this cmd owned by ublksrv */
1097         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
1098
1099         /*
1100          * clear ACTIVE since we are done with this sqe/cmd slot
1101          * We can only accept io cmd in case of being not active.
1102          */
1103         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
1104
1105         /* tell ublksrv one io request is coming */
1106         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
1107 }
1108
1109 #define UBLK_REQUEUE_DELAY_MS   3
1110
1111 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
1112                 struct request *rq)
1113 {
1114         /* We cannot process this rq so just requeue it. */
1115         if (ublk_queue_can_use_recovery(ubq))
1116                 blk_mq_requeue_request(rq, false);
1117         else
1118                 blk_mq_end_request(rq, BLK_STS_IOERR);
1119 }
1120
1121 static inline void __ublk_rq_task_work(struct request *req,
1122                                        unsigned issue_flags)
1123 {
1124         struct ublk_queue *ubq = req->mq_hctx->driver_data;
1125         int tag = req->tag;
1126         struct ublk_io *io = &ubq->ios[tag];
1127         unsigned int mapped_bytes;
1128
1129         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
1130                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1131                         ublk_get_iod(ubq, req->tag)->addr);
1132
1133         /*
1134          * Task is exiting if either:
1135          *
1136          * (1) current != ubq_daemon.
1137          * io_uring_cmd_complete_in_task() tries to run task_work
1138          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
1139          *
1140          * (2) current->flags & PF_EXITING.
1141          */
1142         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
1143                 __ublk_abort_rq(ubq, req);
1144                 return;
1145         }
1146
1147         if (ublk_need_get_data(ubq) && ublk_need_map_req(req)) {
1148                 /*
1149                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
1150                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
1151                  * and notify it.
1152                  */
1153                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
1154                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
1155                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
1156                                         __func__, io->cmd->cmd_op, ubq->q_id,
1157                                         req->tag, io->flags);
1158                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
1159                         return;
1160                 }
1161                 /*
1162                  * We have handled UBLK_IO_NEED_GET_DATA command,
1163                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
1164                  * do the copy work.
1165                  */
1166                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
1167                 /* update iod->addr because ublksrv may have passed a new io buffer */
1168                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
1169                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
1170                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
1171                                 ublk_get_iod(ubq, req->tag)->addr);
1172         }
1173
1174         mapped_bytes = ublk_map_io(ubq, req, io);
1175
1176         /* partially mapped, update io descriptor */
1177         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
1178                 /*
1179                  * Nothing mapped, retry until we succeed.
1180                  *
1181                  * We may never succeed in mapping any bytes here because
1182                  * of OOM. TODO: reserve one buffer with single page pinned
1183                  * for providing forward progress guarantee.
1184                  */
1185                 if (unlikely(!mapped_bytes)) {
1186                         blk_mq_requeue_request(req, false);
1187                         blk_mq_delay_kick_requeue_list(req->q,
1188                                         UBLK_REQUEUE_DELAY_MS);
1189                         return;
1190                 }
1191
1192                 ublk_get_iod(ubq, req->tag)->nr_sectors =
1193                         mapped_bytes >> 9;
1194         }
1195
1196         ublk_init_req_ref(ubq, req);
1197         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
1198 }
1199
1200 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
1201                                         unsigned issue_flags)
1202 {
1203         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
1204         struct ublk_rq_data *data, *tmp;
1205
1206         io_cmds = llist_reverse_order(io_cmds);
1207         llist_for_each_entry_safe(data, tmp, io_cmds, node)
1208                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
1209 }
1210
1211 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
1212 {
1213         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1214         struct ublk_queue *ubq = pdu->ubq;
1215
1216         ublk_forward_io_cmds(ubq, issue_flags);
1217 }
1218
1219 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
1220 {
1221         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
1222
1223         if (llist_add(&data->node, &ubq->io_cmds)) {
1224                 struct ublk_io *io = &ubq->ios[rq->tag];
1225
1226                 io_uring_cmd_complete_in_task(io->cmd, ublk_rq_task_work_cb);
1227         }
1228 }
1229
1230 static enum blk_eh_timer_return ublk_timeout(struct request *rq)
1231 {
1232         struct ublk_queue *ubq = rq->mq_hctx->driver_data;
1233         unsigned int nr_inflight = 0;
1234         int i;
1235
1236         if (ubq->flags & UBLK_F_UNPRIVILEGED_DEV) {
1237                 if (!ubq->timeout) {
1238                         send_sig(SIGKILL, ubq->ubq_daemon, 0);
1239                         ubq->timeout = true;
1240                 }
1241
1242                 return BLK_EH_DONE;
1243         }
1244
1245         if (!ubq_daemon_is_dying(ubq))
1246                 return BLK_EH_RESET_TIMER;
1247
1248         for (i = 0; i < ubq->q_depth; i++) {
1249                 struct ublk_io *io = &ubq->ios[i];
1250
1251                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1252                         nr_inflight++;
1253         }
1254
1255         /* cancelable uring_cmd can't help us if all commands are in-flight */
1256         if (nr_inflight == ubq->q_depth) {
1257                 struct ublk_device *ub = ubq->dev;
1258
1259                 if (ublk_abort_requests(ub, ubq)) {
1260                         if (ublk_can_use_recovery(ub))
1261                                 schedule_work(&ub->quiesce_work);
1262                         else
1263                                 schedule_work(&ub->stop_work);
1264                 }
1265                 return BLK_EH_DONE;
1266         }
1267
1268         return BLK_EH_RESET_TIMER;
1269 }
1270
1271 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
1272                 const struct blk_mq_queue_data *bd)
1273 {
1274         struct ublk_queue *ubq = hctx->driver_data;
1275         struct request *rq = bd->rq;
1276         blk_status_t res;
1277
1278         /* fill iod to slot in io cmd buffer */
1279         res = ublk_setup_iod(ubq, rq);
1280         if (unlikely(res != BLK_STS_OK))
1281                 return BLK_STS_IOERR;
1282
1283         /* With recovery feature enabled, force_abort is set in
1284          * ublk_stop_dev() before calling del_gendisk(). We have to
1285          * abort all requeued and new rqs here to let del_gendisk()
1286          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
1287          * to avoid UAF on io_uring ctx.
1288          *
1289          * Note: force_abort is guaranteed to be seen because it is set
1290          * before request queue is unqiuesced.
1291          */
1292         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
1293                 return BLK_STS_IOERR;
1294
1295         if (unlikely(ubq->canceling)) {
1296                 __ublk_abort_rq(ubq, rq);
1297                 return BLK_STS_OK;
1298         }
1299
1300         blk_mq_start_request(bd->rq);
1301         ublk_queue_cmd(ubq, rq);
1302
1303         return BLK_STS_OK;
1304 }
1305
1306 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
1307                 unsigned int hctx_idx)
1308 {
1309         struct ublk_device *ub = driver_data;
1310         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
1311
1312         hctx->driver_data = ubq;
1313         return 0;
1314 }
1315
1316 static const struct blk_mq_ops ublk_mq_ops = {
1317         .queue_rq       = ublk_queue_rq,
1318         .init_hctx      = ublk_init_hctx,
1319         .timeout        = ublk_timeout,
1320 };
1321
1322 static int ublk_ch_open(struct inode *inode, struct file *filp)
1323 {
1324         struct ublk_device *ub = container_of(inode->i_cdev,
1325                         struct ublk_device, cdev);
1326
1327         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
1328                 return -EBUSY;
1329         filp->private_data = ub;
1330         return 0;
1331 }
1332
1333 static int ublk_ch_release(struct inode *inode, struct file *filp)
1334 {
1335         struct ublk_device *ub = filp->private_data;
1336
1337         clear_bit(UB_STATE_OPEN, &ub->state);
1338         return 0;
1339 }
1340
1341 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
1342 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
1343 {
1344         struct ublk_device *ub = filp->private_data;
1345         size_t sz = vma->vm_end - vma->vm_start;
1346         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
1347         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
1348         int q_id, ret = 0;
1349
1350         spin_lock(&ub->lock);
1351         if (!ub->mm)
1352                 ub->mm = current->mm;
1353         if (current->mm != ub->mm)
1354                 ret = -EINVAL;
1355         spin_unlock(&ub->lock);
1356
1357         if (ret)
1358                 return ret;
1359
1360         if (vma->vm_flags & VM_WRITE)
1361                 return -EPERM;
1362
1363         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
1364         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
1365                 return -EINVAL;
1366
1367         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
1368         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
1369                         __func__, q_id, current->pid, vma->vm_start,
1370                         phys_off, (unsigned long)sz);
1371
1372         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
1373                 return -EINVAL;
1374
1375         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
1376         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
1377 }
1378
1379 static void ublk_commit_completion(struct ublk_device *ub,
1380                 const struct ublksrv_io_cmd *ub_cmd)
1381 {
1382         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
1383         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
1384         struct ublk_io *io = &ubq->ios[tag];
1385         struct request *req;
1386
1387         /* now this cmd slot is owned by nbd driver */
1388         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
1389         io->res = ub_cmd->result;
1390
1391         /* find the io request and complete */
1392         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
1393         if (WARN_ON_ONCE(unlikely(!req)))
1394                 return;
1395
1396         if (req_op(req) == REQ_OP_ZONE_APPEND)
1397                 req->__sector = ub_cmd->zone_append_lba;
1398
1399         if (likely(!blk_should_fake_timeout(req->q)))
1400                 ublk_put_req_ref(ubq, req);
1401 }
1402
1403 /*
1404  * Called from ubq_daemon context via cancel fn, meantime quiesce ublk
1405  * blk-mq queue, so we are called exclusively with blk-mq and ubq_daemon
1406  * context, so everything is serialized.
1407  */
1408 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
1409 {
1410         int i;
1411
1412         for (i = 0; i < ubq->q_depth; i++) {
1413                 struct ublk_io *io = &ubq->ios[i];
1414
1415                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
1416                         struct request *rq;
1417
1418                         /*
1419                          * Either we fail the request or ublk_rq_task_work_fn
1420                          * will do it
1421                          */
1422                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1423                         if (rq && blk_mq_request_started(rq)) {
1424                                 io->flags |= UBLK_IO_FLAG_ABORTED;
1425                                 __ublk_fail_req(ubq, io, rq);
1426                         }
1427                 }
1428         }
1429 }
1430
1431 static bool ublk_abort_requests(struct ublk_device *ub, struct ublk_queue *ubq)
1432 {
1433         struct gendisk *disk;
1434
1435         spin_lock(&ubq->cancel_lock);
1436         if (ubq->canceling) {
1437                 spin_unlock(&ubq->cancel_lock);
1438                 return false;
1439         }
1440         ubq->canceling = true;
1441         spin_unlock(&ubq->cancel_lock);
1442
1443         spin_lock(&ub->lock);
1444         disk = ub->ub_disk;
1445         if (disk)
1446                 get_device(disk_to_dev(disk));
1447         spin_unlock(&ub->lock);
1448
1449         /* Our disk has been dead */
1450         if (!disk)
1451                 return false;
1452
1453         /* Now we are serialized with ublk_queue_rq() */
1454         blk_mq_quiesce_queue(disk->queue);
1455         /* abort queue is for making forward progress */
1456         ublk_abort_queue(ub, ubq);
1457         blk_mq_unquiesce_queue(disk->queue);
1458         put_device(disk_to_dev(disk));
1459
1460         return true;
1461 }
1462
1463 static void ublk_cancel_cmd(struct ublk_queue *ubq, struct ublk_io *io,
1464                 unsigned int issue_flags)
1465 {
1466         bool done;
1467
1468         if (!(io->flags & UBLK_IO_FLAG_ACTIVE))
1469                 return;
1470
1471         spin_lock(&ubq->cancel_lock);
1472         done = !!(io->flags & UBLK_IO_FLAG_CANCELED);
1473         if (!done)
1474                 io->flags |= UBLK_IO_FLAG_CANCELED;
1475         spin_unlock(&ubq->cancel_lock);
1476
1477         if (!done)
1478                 io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0, issue_flags);
1479 }
1480
1481 /*
1482  * The ublk char device won't be closed when calling cancel fn, so both
1483  * ublk device and queue are guaranteed to be live
1484  */
1485 static void ublk_uring_cmd_cancel_fn(struct io_uring_cmd *cmd,
1486                 unsigned int issue_flags)
1487 {
1488         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1489         struct ublk_queue *ubq = pdu->ubq;
1490         struct task_struct *task;
1491         struct ublk_device *ub;
1492         bool need_schedule;
1493         struct ublk_io *io;
1494
1495         if (WARN_ON_ONCE(!ubq))
1496                 return;
1497
1498         if (WARN_ON_ONCE(pdu->tag >= ubq->q_depth))
1499                 return;
1500
1501         task = io_uring_cmd_get_task(cmd);
1502         if (WARN_ON_ONCE(task && task != ubq->ubq_daemon))
1503                 return;
1504
1505         ub = ubq->dev;
1506         need_schedule = ublk_abort_requests(ub, ubq);
1507
1508         io = &ubq->ios[pdu->tag];
1509         WARN_ON_ONCE(io->cmd != cmd);
1510         ublk_cancel_cmd(ubq, io, issue_flags);
1511
1512         if (need_schedule) {
1513                 if (ublk_can_use_recovery(ub))
1514                         schedule_work(&ub->quiesce_work);
1515                 else
1516                         schedule_work(&ub->stop_work);
1517         }
1518 }
1519
1520 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1521 {
1522         return ubq->nr_io_ready == ubq->q_depth;
1523 }
1524
1525 static void ublk_cancel_queue(struct ublk_queue *ubq)
1526 {
1527         int i;
1528
1529         for (i = 0; i < ubq->q_depth; i++)
1530                 ublk_cancel_cmd(ubq, &ubq->ios[i], IO_URING_F_UNLOCKED);
1531 }
1532
1533 /* Cancel all pending commands, must be called after del_gendisk() returns */
1534 static void ublk_cancel_dev(struct ublk_device *ub)
1535 {
1536         int i;
1537
1538         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1539                 ublk_cancel_queue(ublk_get_queue(ub, i));
1540 }
1541
1542 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1543 {
1544         bool *idle = data;
1545
1546         if (blk_mq_request_started(rq)) {
1547                 *idle = false;
1548                 return false;
1549         }
1550         return true;
1551 }
1552
1553 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1554 {
1555         bool idle;
1556
1557         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1558         while (true) {
1559                 idle = true;
1560                 blk_mq_tagset_busy_iter(&ub->tag_set,
1561                                 ublk_check_inflight_rq, &idle);
1562                 if (idle)
1563                         break;
1564                 msleep(UBLK_REQUEUE_DELAY_MS);
1565         }
1566 }
1567
1568 static void __ublk_quiesce_dev(struct ublk_device *ub)
1569 {
1570         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1571                         __func__, ub->dev_info.dev_id,
1572                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1573                         "LIVE" : "QUIESCED");
1574         blk_mq_quiesce_queue(ub->ub_disk->queue);
1575         ublk_wait_tagset_rqs_idle(ub);
1576         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1577 }
1578
1579 static void ublk_quiesce_work_fn(struct work_struct *work)
1580 {
1581         struct ublk_device *ub =
1582                 container_of(work, struct ublk_device, quiesce_work);
1583
1584         mutex_lock(&ub->mutex);
1585         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1586                 goto unlock;
1587         __ublk_quiesce_dev(ub);
1588  unlock:
1589         mutex_unlock(&ub->mutex);
1590         ublk_cancel_dev(ub);
1591 }
1592
1593 static void ublk_unquiesce_dev(struct ublk_device *ub)
1594 {
1595         int i;
1596
1597         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1598                         __func__, ub->dev_info.dev_id,
1599                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1600                         "LIVE" : "QUIESCED");
1601         /* quiesce_work has run. We let requeued rqs be aborted
1602          * before running fallback_wq. "force_abort" must be seen
1603          * after request queue is unqiuesced. Then del_gendisk()
1604          * can move on.
1605          */
1606         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1607                 ublk_get_queue(ub, i)->force_abort = true;
1608
1609         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1610         /* We may have requeued some rqs in ublk_quiesce_queue() */
1611         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1612 }
1613
1614 static void ublk_stop_dev(struct ublk_device *ub)
1615 {
1616         struct gendisk *disk;
1617
1618         mutex_lock(&ub->mutex);
1619         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1620                 goto unlock;
1621         if (ublk_can_use_recovery(ub)) {
1622                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1623                         __ublk_quiesce_dev(ub);
1624                 ublk_unquiesce_dev(ub);
1625         }
1626         del_gendisk(ub->ub_disk);
1627
1628         /* Sync with ublk_abort_queue() by holding the lock */
1629         spin_lock(&ub->lock);
1630         disk = ub->ub_disk;
1631         ub->dev_info.state = UBLK_S_DEV_DEAD;
1632         ub->dev_info.ublksrv_pid = -1;
1633         ub->ub_disk = NULL;
1634         spin_unlock(&ub->lock);
1635         put_disk(disk);
1636  unlock:
1637         mutex_unlock(&ub->mutex);
1638         ublk_cancel_dev(ub);
1639 }
1640
1641 /* device can only be started after all IOs are ready */
1642 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1643 {
1644         mutex_lock(&ub->mutex);
1645         ubq->nr_io_ready++;
1646         if (ublk_queue_ready(ubq)) {
1647                 ubq->ubq_daemon = current;
1648                 get_task_struct(ubq->ubq_daemon);
1649                 ub->nr_queues_ready++;
1650
1651                 if (capable(CAP_SYS_ADMIN))
1652                         ub->nr_privileged_daemon++;
1653         }
1654         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1655                 complete_all(&ub->completion);
1656         mutex_unlock(&ub->mutex);
1657 }
1658
1659 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1660                 int tag)
1661 {
1662         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1663         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1664
1665         ublk_queue_cmd(ubq, req);
1666 }
1667
1668 static inline int ublk_check_cmd_op(u32 cmd_op)
1669 {
1670         u32 ioc_type = _IOC_TYPE(cmd_op);
1671
1672         if (!IS_ENABLED(CONFIG_BLKDEV_UBLK_LEGACY_OPCODES) && ioc_type != 'u')
1673                 return -EOPNOTSUPP;
1674
1675         if (ioc_type != 'u' && ioc_type != 0)
1676                 return -EOPNOTSUPP;
1677
1678         return 0;
1679 }
1680
1681 static inline void ublk_fill_io_cmd(struct ublk_io *io,
1682                 struct io_uring_cmd *cmd, unsigned long buf_addr)
1683 {
1684         io->cmd = cmd;
1685         io->flags |= UBLK_IO_FLAG_ACTIVE;
1686         io->addr = buf_addr;
1687 }
1688
1689 static inline void ublk_prep_cancel(struct io_uring_cmd *cmd,
1690                                     unsigned int issue_flags,
1691                                     struct ublk_queue *ubq, unsigned int tag)
1692 {
1693         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
1694
1695         /*
1696          * Safe to refer to @ubq since ublk_queue won't be died until its
1697          * commands are completed
1698          */
1699         pdu->ubq = ubq;
1700         pdu->tag = tag;
1701         io_uring_cmd_mark_cancelable(cmd, issue_flags);
1702 }
1703
1704 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1705                                unsigned int issue_flags,
1706                                const struct ublksrv_io_cmd *ub_cmd)
1707 {
1708         struct ublk_device *ub = cmd->file->private_data;
1709         struct ublk_queue *ubq;
1710         struct ublk_io *io;
1711         u32 cmd_op = cmd->cmd_op;
1712         unsigned tag = ub_cmd->tag;
1713         int ret = -EINVAL;
1714         struct request *req;
1715
1716         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1717                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1718                         ub_cmd->result);
1719
1720         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1721                 goto out;
1722
1723         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1724         if (!ubq || ub_cmd->q_id != ubq->q_id)
1725                 goto out;
1726
1727         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1728                 goto out;
1729
1730         if (tag >= ubq->q_depth)
1731                 goto out;
1732
1733         io = &ubq->ios[tag];
1734
1735         /* there is pending io cmd, something must be wrong */
1736         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1737                 ret = -EBUSY;
1738                 goto out;
1739         }
1740
1741         /*
1742          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1743          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1744          */
1745         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1746                         ^ (_IOC_NR(cmd_op) == UBLK_IO_NEED_GET_DATA))
1747                 goto out;
1748
1749         ret = ublk_check_cmd_op(cmd_op);
1750         if (ret)
1751                 goto out;
1752
1753         ret = -EINVAL;
1754         switch (_IOC_NR(cmd_op)) {
1755         case UBLK_IO_FETCH_REQ:
1756                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1757                 if (ublk_queue_ready(ubq)) {
1758                         ret = -EBUSY;
1759                         goto out;
1760                 }
1761                 /*
1762                  * The io is being handled by server, so COMMIT_RQ is expected
1763                  * instead of FETCH_REQ
1764                  */
1765                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1766                         goto out;
1767
1768                 if (!ublk_support_user_copy(ubq)) {
1769                         /*
1770                          * FETCH_RQ has to provide IO buffer if NEED GET
1771                          * DATA is not enabled
1772                          */
1773                         if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1774                                 goto out;
1775                 } else if (ub_cmd->addr) {
1776                         /* User copy requires addr to be unset */
1777                         ret = -EINVAL;
1778                         goto out;
1779                 }
1780
1781                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1782                 ublk_mark_io_ready(ub, ubq);
1783                 break;
1784         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1785                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1786
1787                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1788                         goto out;
1789
1790                 if (!ublk_support_user_copy(ubq)) {
1791                         /*
1792                          * COMMIT_AND_FETCH_REQ has to provide IO buffer if
1793                          * NEED GET DATA is not enabled or it is Read IO.
1794                          */
1795                         if (!ub_cmd->addr && (!ublk_need_get_data(ubq) ||
1796                                                 req_op(req) == REQ_OP_READ))
1797                                 goto out;
1798                 } else if (req_op(req) != REQ_OP_ZONE_APPEND && ub_cmd->addr) {
1799                         /*
1800                          * User copy requires addr to be unset when command is
1801                          * not zone append
1802                          */
1803                         ret = -EINVAL;
1804                         goto out;
1805                 }
1806
1807                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1808                 ublk_commit_completion(ub, ub_cmd);
1809                 break;
1810         case UBLK_IO_NEED_GET_DATA:
1811                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1812                         goto out;
1813                 ublk_fill_io_cmd(io, cmd, ub_cmd->addr);
1814                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1815                 break;
1816         default:
1817                 goto out;
1818         }
1819         ublk_prep_cancel(cmd, issue_flags, ubq, tag);
1820         return -EIOCBQUEUED;
1821
1822  out:
1823         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1824         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1825                         __func__, cmd_op, tag, ret, io->flags);
1826         return -EIOCBQUEUED;
1827 }
1828
1829 static inline struct request *__ublk_check_and_get_req(struct ublk_device *ub,
1830                 struct ublk_queue *ubq, int tag, size_t offset)
1831 {
1832         struct request *req;
1833
1834         if (!ublk_need_req_ref(ubq))
1835                 return NULL;
1836
1837         req = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], tag);
1838         if (!req)
1839                 return NULL;
1840
1841         if (!ublk_get_req_ref(ubq, req))
1842                 return NULL;
1843
1844         if (unlikely(!blk_mq_request_started(req) || req->tag != tag))
1845                 goto fail_put;
1846
1847         if (!ublk_rq_has_data(req))
1848                 goto fail_put;
1849
1850         if (offset > blk_rq_bytes(req))
1851                 goto fail_put;
1852
1853         return req;
1854 fail_put:
1855         ublk_put_req_ref(ubq, req);
1856         return NULL;
1857 }
1858
1859 static inline int ublk_ch_uring_cmd_local(struct io_uring_cmd *cmd,
1860                 unsigned int issue_flags)
1861 {
1862         /*
1863          * Not necessary for async retry, but let's keep it simple and always
1864          * copy the values to avoid any potential reuse.
1865          */
1866         const struct ublksrv_io_cmd *ub_src = io_uring_sqe_cmd(cmd->sqe);
1867         const struct ublksrv_io_cmd ub_cmd = {
1868                 .q_id = READ_ONCE(ub_src->q_id),
1869                 .tag = READ_ONCE(ub_src->tag),
1870                 .result = READ_ONCE(ub_src->result),
1871                 .addr = READ_ONCE(ub_src->addr)
1872         };
1873
1874         WARN_ON_ONCE(issue_flags & IO_URING_F_UNLOCKED);
1875
1876         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1877 }
1878
1879 static void ublk_ch_uring_cmd_cb(struct io_uring_cmd *cmd,
1880                 unsigned int issue_flags)
1881 {
1882         ublk_ch_uring_cmd_local(cmd, issue_flags);
1883 }
1884
1885 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1886 {
1887         if (unlikely(issue_flags & IO_URING_F_CANCEL)) {
1888                 ublk_uring_cmd_cancel_fn(cmd, issue_flags);
1889                 return 0;
1890         }
1891
1892         /* well-implemented server won't run into unlocked */
1893         if (unlikely(issue_flags & IO_URING_F_UNLOCKED)) {
1894                 io_uring_cmd_complete_in_task(cmd, ublk_ch_uring_cmd_cb);
1895                 return -EIOCBQUEUED;
1896         }
1897
1898         return ublk_ch_uring_cmd_local(cmd, issue_flags);
1899 }
1900
1901 static inline bool ublk_check_ubuf_dir(const struct request *req,
1902                 int ubuf_dir)
1903 {
1904         /* copy ubuf to request pages */
1905         if ((req_op(req) == REQ_OP_READ || req_op(req) == REQ_OP_DRV_IN) &&
1906             ubuf_dir == ITER_SOURCE)
1907                 return true;
1908
1909         /* copy request pages to ubuf */
1910         if ((req_op(req) == REQ_OP_WRITE ||
1911              req_op(req) == REQ_OP_ZONE_APPEND) &&
1912             ubuf_dir == ITER_DEST)
1913                 return true;
1914
1915         return false;
1916 }
1917
1918 static struct request *ublk_check_and_get_req(struct kiocb *iocb,
1919                 struct iov_iter *iter, size_t *off, int dir)
1920 {
1921         struct ublk_device *ub = iocb->ki_filp->private_data;
1922         struct ublk_queue *ubq;
1923         struct request *req;
1924         size_t buf_off;
1925         u16 tag, q_id;
1926
1927         if (!ub)
1928                 return ERR_PTR(-EACCES);
1929
1930         if (!user_backed_iter(iter))
1931                 return ERR_PTR(-EACCES);
1932
1933         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1934                 return ERR_PTR(-EACCES);
1935
1936         tag = ublk_pos_to_tag(iocb->ki_pos);
1937         q_id = ublk_pos_to_hwq(iocb->ki_pos);
1938         buf_off = ublk_pos_to_buf_off(iocb->ki_pos);
1939
1940         if (q_id >= ub->dev_info.nr_hw_queues)
1941                 return ERR_PTR(-EINVAL);
1942
1943         ubq = ublk_get_queue(ub, q_id);
1944         if (!ubq)
1945                 return ERR_PTR(-EINVAL);
1946
1947         if (tag >= ubq->q_depth)
1948                 return ERR_PTR(-EINVAL);
1949
1950         req = __ublk_check_and_get_req(ub, ubq, tag, buf_off);
1951         if (!req)
1952                 return ERR_PTR(-EINVAL);
1953
1954         if (!req->mq_hctx || !req->mq_hctx->driver_data)
1955                 goto fail;
1956
1957         if (!ublk_check_ubuf_dir(req, dir))
1958                 goto fail;
1959
1960         *off = buf_off;
1961         return req;
1962 fail:
1963         ublk_put_req_ref(ubq, req);
1964         return ERR_PTR(-EACCES);
1965 }
1966
1967 static ssize_t ublk_ch_read_iter(struct kiocb *iocb, struct iov_iter *to)
1968 {
1969         struct ublk_queue *ubq;
1970         struct request *req;
1971         size_t buf_off;
1972         size_t ret;
1973
1974         req = ublk_check_and_get_req(iocb, to, &buf_off, ITER_DEST);
1975         if (IS_ERR(req))
1976                 return PTR_ERR(req);
1977
1978         ret = ublk_copy_user_pages(req, buf_off, to, ITER_DEST);
1979         ubq = req->mq_hctx->driver_data;
1980         ublk_put_req_ref(ubq, req);
1981
1982         return ret;
1983 }
1984
1985 static ssize_t ublk_ch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1986 {
1987         struct ublk_queue *ubq;
1988         struct request *req;
1989         size_t buf_off;
1990         size_t ret;
1991
1992         req = ublk_check_and_get_req(iocb, from, &buf_off, ITER_SOURCE);
1993         if (IS_ERR(req))
1994                 return PTR_ERR(req);
1995
1996         ret = ublk_copy_user_pages(req, buf_off, from, ITER_SOURCE);
1997         ubq = req->mq_hctx->driver_data;
1998         ublk_put_req_ref(ubq, req);
1999
2000         return ret;
2001 }
2002
2003 static const struct file_operations ublk_ch_fops = {
2004         .owner = THIS_MODULE,
2005         .open = ublk_ch_open,
2006         .release = ublk_ch_release,
2007         .llseek = no_llseek,
2008         .read_iter = ublk_ch_read_iter,
2009         .write_iter = ublk_ch_write_iter,
2010         .uring_cmd = ublk_ch_uring_cmd,
2011         .mmap = ublk_ch_mmap,
2012 };
2013
2014 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
2015 {
2016         int size = ublk_queue_cmd_buf_size(ub, q_id);
2017         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
2018
2019         if (ubq->ubq_daemon)
2020                 put_task_struct(ubq->ubq_daemon);
2021         if (ubq->io_cmd_buf)
2022                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
2023 }
2024
2025 static int ublk_init_queue(struct ublk_device *ub, int q_id)
2026 {
2027         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
2028         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
2029         void *ptr;
2030         int size;
2031
2032         spin_lock_init(&ubq->cancel_lock);
2033         ubq->flags = ub->dev_info.flags;
2034         ubq->q_id = q_id;
2035         ubq->q_depth = ub->dev_info.queue_depth;
2036         size = ublk_queue_cmd_buf_size(ub, q_id);
2037
2038         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
2039         if (!ptr)
2040                 return -ENOMEM;
2041
2042         ubq->io_cmd_buf = ptr;
2043         ubq->dev = ub;
2044         return 0;
2045 }
2046
2047 static void ublk_deinit_queues(struct ublk_device *ub)
2048 {
2049         int nr_queues = ub->dev_info.nr_hw_queues;
2050         int i;
2051
2052         if (!ub->__queues)
2053                 return;
2054
2055         for (i = 0; i < nr_queues; i++)
2056                 ublk_deinit_queue(ub, i);
2057         kfree(ub->__queues);
2058 }
2059
2060 static int ublk_init_queues(struct ublk_device *ub)
2061 {
2062         int nr_queues = ub->dev_info.nr_hw_queues;
2063         int depth = ub->dev_info.queue_depth;
2064         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
2065         int i, ret = -ENOMEM;
2066
2067         ub->queue_size = ubq_size;
2068         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
2069         if (!ub->__queues)
2070                 return ret;
2071
2072         for (i = 0; i < nr_queues; i++) {
2073                 if (ublk_init_queue(ub, i))
2074                         goto fail;
2075         }
2076
2077         init_completion(&ub->completion);
2078         return 0;
2079
2080  fail:
2081         ublk_deinit_queues(ub);
2082         return ret;
2083 }
2084
2085 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
2086 {
2087         int i = idx;
2088         int err;
2089
2090         spin_lock(&ublk_idr_lock);
2091         /* allocate id, if @id >= 0, we're requesting that specific id */
2092         if (i >= 0) {
2093                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
2094                 if (err == -ENOSPC)
2095                         err = -EEXIST;
2096         } else {
2097                 err = idr_alloc(&ublk_index_idr, ub, 0, UBLK_MAX_UBLKS,
2098                                 GFP_NOWAIT);
2099         }
2100         spin_unlock(&ublk_idr_lock);
2101
2102         if (err >= 0)
2103                 ub->ub_number = err;
2104
2105         return err;
2106 }
2107
2108 static void ublk_free_dev_number(struct ublk_device *ub)
2109 {
2110         spin_lock(&ublk_idr_lock);
2111         idr_remove(&ublk_index_idr, ub->ub_number);
2112         wake_up_all(&ublk_idr_wq);
2113         spin_unlock(&ublk_idr_lock);
2114 }
2115
2116 static void ublk_cdev_rel(struct device *dev)
2117 {
2118         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
2119
2120         blk_mq_free_tag_set(&ub->tag_set);
2121         ublk_deinit_queues(ub);
2122         ublk_free_dev_number(ub);
2123         mutex_destroy(&ub->mutex);
2124         kfree(ub);
2125 }
2126
2127 static int ublk_add_chdev(struct ublk_device *ub)
2128 {
2129         struct device *dev = &ub->cdev_dev;
2130         int minor = ub->ub_number;
2131         int ret;
2132
2133         dev->parent = ublk_misc.this_device;
2134         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
2135         dev->class = &ublk_chr_class;
2136         dev->release = ublk_cdev_rel;
2137         device_initialize(dev);
2138
2139         ret = dev_set_name(dev, "ublkc%d", minor);
2140         if (ret)
2141                 goto fail;
2142
2143         cdev_init(&ub->cdev, &ublk_ch_fops);
2144         ret = cdev_device_add(&ub->cdev, dev);
2145         if (ret)
2146                 goto fail;
2147
2148         ublks_added++;
2149         return 0;
2150  fail:
2151         put_device(dev);
2152         return ret;
2153 }
2154
2155 static void ublk_stop_work_fn(struct work_struct *work)
2156 {
2157         struct ublk_device *ub =
2158                 container_of(work, struct ublk_device, stop_work);
2159
2160         ublk_stop_dev(ub);
2161 }
2162
2163 /* align max io buffer size with PAGE_SIZE */
2164 static void ublk_align_max_io_size(struct ublk_device *ub)
2165 {
2166         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
2167
2168         ub->dev_info.max_io_buf_bytes =
2169                 round_down(max_io_bytes, PAGE_SIZE);
2170 }
2171
2172 static int ublk_add_tag_set(struct ublk_device *ub)
2173 {
2174         ub->tag_set.ops = &ublk_mq_ops;
2175         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
2176         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
2177         ub->tag_set.numa_node = NUMA_NO_NODE;
2178         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
2179         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
2180         ub->tag_set.driver_data = ub;
2181         return blk_mq_alloc_tag_set(&ub->tag_set);
2182 }
2183
2184 static void ublk_remove(struct ublk_device *ub)
2185 {
2186         ublk_stop_dev(ub);
2187         cancel_work_sync(&ub->stop_work);
2188         cancel_work_sync(&ub->quiesce_work);
2189         cdev_device_del(&ub->cdev, &ub->cdev_dev);
2190         put_device(&ub->cdev_dev);
2191         ublks_added--;
2192 }
2193
2194 static struct ublk_device *ublk_get_device_from_id(int idx)
2195 {
2196         struct ublk_device *ub = NULL;
2197
2198         if (idx < 0)
2199                 return NULL;
2200
2201         spin_lock(&ublk_idr_lock);
2202         ub = idr_find(&ublk_index_idr, idx);
2203         if (ub)
2204                 ub = ublk_get_device(ub);
2205         spin_unlock(&ublk_idr_lock);
2206
2207         return ub;
2208 }
2209
2210 static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd)
2211 {
2212         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2213         int ublksrv_pid = (int)header->data[0];
2214         struct gendisk *disk;
2215         int ret = -EINVAL;
2216
2217         if (ublksrv_pid <= 0)
2218                 return -EINVAL;
2219
2220         if (wait_for_completion_interruptible(&ub->completion) != 0)
2221                 return -EINTR;
2222
2223         mutex_lock(&ub->mutex);
2224         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
2225             test_bit(UB_STATE_USED, &ub->state)) {
2226                 ret = -EEXIST;
2227                 goto out_unlock;
2228         }
2229
2230         disk = blk_mq_alloc_disk(&ub->tag_set, NULL);
2231         if (IS_ERR(disk)) {
2232                 ret = PTR_ERR(disk);
2233                 goto out_unlock;
2234         }
2235         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
2236         disk->fops = &ub_fops;
2237         disk->private_data = ub;
2238
2239         ub->dev_info.ublksrv_pid = ublksrv_pid;
2240         ub->ub_disk = disk;
2241
2242         ret = ublk_apply_params(ub);
2243         if (ret)
2244                 goto out_put_disk;
2245
2246         /* don't probe partitions if any one ubq daemon is un-trusted */
2247         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
2248                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
2249
2250         get_device(&ub->cdev_dev);
2251         ub->dev_info.state = UBLK_S_DEV_LIVE;
2252
2253         if (ublk_dev_is_zoned(ub)) {
2254                 ret = ublk_revalidate_disk_zones(ub);
2255                 if (ret)
2256                         goto out_put_cdev;
2257         }
2258
2259         ret = add_disk(disk);
2260         if (ret)
2261                 goto out_put_cdev;
2262
2263         set_bit(UB_STATE_USED, &ub->state);
2264
2265 out_put_cdev:
2266         if (ret) {
2267                 ub->dev_info.state = UBLK_S_DEV_DEAD;
2268                 ublk_put_device(ub);
2269         }
2270 out_put_disk:
2271         if (ret)
2272                 put_disk(disk);
2273 out_unlock:
2274         mutex_unlock(&ub->mutex);
2275         return ret;
2276 }
2277
2278 static int ublk_ctrl_get_queue_affinity(struct ublk_device *ub,
2279                 struct io_uring_cmd *cmd)
2280 {
2281         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2282         void __user *argp = (void __user *)(unsigned long)header->addr;
2283         cpumask_var_t cpumask;
2284         unsigned long queue;
2285         unsigned int retlen;
2286         unsigned int i;
2287         int ret;
2288
2289         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
2290                 return -EINVAL;
2291         if (header->len & (sizeof(unsigned long)-1))
2292                 return -EINVAL;
2293         if (!header->addr)
2294                 return -EINVAL;
2295
2296         queue = header->data[0];
2297         if (queue >= ub->dev_info.nr_hw_queues)
2298                 return -EINVAL;
2299
2300         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
2301                 return -ENOMEM;
2302
2303         for_each_possible_cpu(i) {
2304                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
2305                         cpumask_set_cpu(i, cpumask);
2306         }
2307
2308         ret = -EFAULT;
2309         retlen = min_t(unsigned short, header->len, cpumask_size());
2310         if (copy_to_user(argp, cpumask, retlen))
2311                 goto out_free_cpumask;
2312         if (retlen != header->len &&
2313             clear_user(argp + retlen, header->len - retlen))
2314                 goto out_free_cpumask;
2315
2316         ret = 0;
2317 out_free_cpumask:
2318         free_cpumask_var(cpumask);
2319         return ret;
2320 }
2321
2322 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
2323 {
2324         pr_devel("%s: dev id %d flags %llx\n", __func__,
2325                         info->dev_id, info->flags);
2326         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
2327                         info->nr_hw_queues, info->queue_depth);
2328 }
2329
2330 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
2331 {
2332         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2333         void __user *argp = (void __user *)(unsigned long)header->addr;
2334         struct ublksrv_ctrl_dev_info info;
2335         struct ublk_device *ub;
2336         int ret = -EINVAL;
2337
2338         if (header->len < sizeof(info) || !header->addr)
2339                 return -EINVAL;
2340         if (header->queue_id != (u16)-1) {
2341                 pr_warn("%s: queue_id is wrong %x\n",
2342                         __func__, header->queue_id);
2343                 return -EINVAL;
2344         }
2345
2346         if (copy_from_user(&info, argp, sizeof(info)))
2347                 return -EFAULT;
2348
2349         if (capable(CAP_SYS_ADMIN))
2350                 info.flags &= ~UBLK_F_UNPRIVILEGED_DEV;
2351         else if (!(info.flags & UBLK_F_UNPRIVILEGED_DEV))
2352                 return -EPERM;
2353
2354         /*
2355          * unprivileged device can't be trusted, but RECOVERY and
2356          * RECOVERY_REISSUE still may hang error handling, so can't
2357          * support recovery features for unprivileged ublk now
2358          *
2359          * TODO: provide forward progress for RECOVERY handler, so that
2360          * unprivileged device can benefit from it
2361          */
2362         if (info.flags & UBLK_F_UNPRIVILEGED_DEV)
2363                 info.flags &= ~(UBLK_F_USER_RECOVERY_REISSUE |
2364                                 UBLK_F_USER_RECOVERY);
2365
2366         /* the created device is always owned by current user */
2367         ublk_store_owner_uid_gid(&info.owner_uid, &info.owner_gid);
2368
2369         if (header->dev_id != info.dev_id) {
2370                 pr_warn("%s: dev id not match %u %u\n",
2371                         __func__, header->dev_id, info.dev_id);
2372                 return -EINVAL;
2373         }
2374
2375         if (header->dev_id != U32_MAX && header->dev_id >= UBLK_MAX_UBLKS) {
2376                 pr_warn("%s: dev id is too large. Max supported is %d\n",
2377                         __func__, UBLK_MAX_UBLKS - 1);
2378                 return -EINVAL;
2379         }
2380
2381         ublk_dump_dev_info(&info);
2382
2383         ret = mutex_lock_killable(&ublk_ctl_mutex);
2384         if (ret)
2385                 return ret;
2386
2387         ret = -EACCES;
2388         if (ublks_added >= ublks_max)
2389                 goto out_unlock;
2390
2391         ret = -ENOMEM;
2392         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
2393         if (!ub)
2394                 goto out_unlock;
2395         mutex_init(&ub->mutex);
2396         spin_lock_init(&ub->lock);
2397         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
2398         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
2399
2400         ret = ublk_alloc_dev_number(ub, header->dev_id);
2401         if (ret < 0)
2402                 goto out_free_ub;
2403
2404         memcpy(&ub->dev_info, &info, sizeof(info));
2405
2406         /* update device id */
2407         ub->dev_info.dev_id = ub->ub_number;
2408
2409         /*
2410          * 64bit flags will be copied back to userspace as feature
2411          * negotiation result, so have to clear flags which driver
2412          * doesn't support yet, then userspace can get correct flags
2413          * (features) to handle.
2414          */
2415         ub->dev_info.flags &= UBLK_F_ALL;
2416
2417         ub->dev_info.flags |= UBLK_F_CMD_IOCTL_ENCODE |
2418                 UBLK_F_URING_CMD_COMP_IN_TASK;
2419
2420         /* GET_DATA isn't needed any more with USER_COPY */
2421         if (ublk_dev_is_user_copy(ub))
2422                 ub->dev_info.flags &= ~UBLK_F_NEED_GET_DATA;
2423
2424         /* Zoned storage support requires user copy feature */
2425         if (ublk_dev_is_zoned(ub) &&
2426             (!IS_ENABLED(CONFIG_BLK_DEV_ZONED) || !ublk_dev_is_user_copy(ub))) {
2427                 ret = -EINVAL;
2428                 goto out_free_dev_number;
2429         }
2430
2431         /* We are not ready to support zero copy */
2432         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
2433
2434         ub->dev_info.nr_hw_queues = min_t(unsigned int,
2435                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
2436         ublk_align_max_io_size(ub);
2437
2438         ret = ublk_init_queues(ub);
2439         if (ret)
2440                 goto out_free_dev_number;
2441
2442         ret = ublk_add_tag_set(ub);
2443         if (ret)
2444                 goto out_deinit_queues;
2445
2446         ret = -EFAULT;
2447         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
2448                 goto out_free_tag_set;
2449
2450         /*
2451          * Add the char dev so that ublksrv daemon can be setup.
2452          * ublk_add_chdev() will cleanup everything if it fails.
2453          */
2454         ret = ublk_add_chdev(ub);
2455         goto out_unlock;
2456
2457 out_free_tag_set:
2458         blk_mq_free_tag_set(&ub->tag_set);
2459 out_deinit_queues:
2460         ublk_deinit_queues(ub);
2461 out_free_dev_number:
2462         ublk_free_dev_number(ub);
2463 out_free_ub:
2464         mutex_destroy(&ub->mutex);
2465         kfree(ub);
2466 out_unlock:
2467         mutex_unlock(&ublk_ctl_mutex);
2468         return ret;
2469 }
2470
2471 static inline bool ublk_idr_freed(int id)
2472 {
2473         void *ptr;
2474
2475         spin_lock(&ublk_idr_lock);
2476         ptr = idr_find(&ublk_index_idr, id);
2477         spin_unlock(&ublk_idr_lock);
2478
2479         return ptr == NULL;
2480 }
2481
2482 static int ublk_ctrl_del_dev(struct ublk_device **p_ub)
2483 {
2484         struct ublk_device *ub = *p_ub;
2485         int idx = ub->ub_number;
2486         int ret;
2487
2488         ret = mutex_lock_killable(&ublk_ctl_mutex);
2489         if (ret)
2490                 return ret;
2491
2492         if (!test_bit(UB_STATE_DELETED, &ub->state)) {
2493                 ublk_remove(ub);
2494                 set_bit(UB_STATE_DELETED, &ub->state);
2495         }
2496
2497         /* Mark the reference as consumed */
2498         *p_ub = NULL;
2499         ublk_put_device(ub);
2500         mutex_unlock(&ublk_ctl_mutex);
2501
2502         /*
2503          * Wait until the idr is removed, then it can be reused after
2504          * DEL_DEV command is returned.
2505          *
2506          * If we returns because of user interrupt, future delete command
2507          * may come:
2508          *
2509          * - the device number isn't freed, this device won't or needn't
2510          *   be deleted again, since UB_STATE_DELETED is set, and device
2511          *   will be released after the last reference is dropped
2512          *
2513          * - the device number is freed already, we will not find this
2514          *   device via ublk_get_device_from_id()
2515          */
2516         if (wait_event_interruptible(ublk_idr_wq, ublk_idr_freed(idx)))
2517                 return -EINTR;
2518         return 0;
2519 }
2520
2521 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
2522 {
2523         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2524
2525         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
2526                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
2527                         header->data[0], header->addr, header->len);
2528 }
2529
2530 static int ublk_ctrl_stop_dev(struct ublk_device *ub)
2531 {
2532         ublk_stop_dev(ub);
2533         cancel_work_sync(&ub->stop_work);
2534         cancel_work_sync(&ub->quiesce_work);
2535
2536         return 0;
2537 }
2538
2539 static int ublk_ctrl_get_dev_info(struct ublk_device *ub,
2540                 struct io_uring_cmd *cmd)
2541 {
2542         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2543         void __user *argp = (void __user *)(unsigned long)header->addr;
2544
2545         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
2546                 return -EINVAL;
2547
2548         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
2549                 return -EFAULT;
2550
2551         return 0;
2552 }
2553
2554 /* TYPE_DEVT is readonly, so fill it up before returning to userspace */
2555 static void ublk_ctrl_fill_params_devt(struct ublk_device *ub)
2556 {
2557         ub->params.devt.char_major = MAJOR(ub->cdev_dev.devt);
2558         ub->params.devt.char_minor = MINOR(ub->cdev_dev.devt);
2559
2560         if (ub->ub_disk) {
2561                 ub->params.devt.disk_major = MAJOR(disk_devt(ub->ub_disk));
2562                 ub->params.devt.disk_minor = MINOR(disk_devt(ub->ub_disk));
2563         } else {
2564                 ub->params.devt.disk_major = 0;
2565                 ub->params.devt.disk_minor = 0;
2566         }
2567         ub->params.types |= UBLK_PARAM_TYPE_DEVT;
2568 }
2569
2570 static int ublk_ctrl_get_params(struct ublk_device *ub,
2571                 struct io_uring_cmd *cmd)
2572 {
2573         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2574         void __user *argp = (void __user *)(unsigned long)header->addr;
2575         struct ublk_params_header ph;
2576         int ret;
2577
2578         if (header->len <= sizeof(ph) || !header->addr)
2579                 return -EINVAL;
2580
2581         if (copy_from_user(&ph, argp, sizeof(ph)))
2582                 return -EFAULT;
2583
2584         if (ph.len > header->len || !ph.len)
2585                 return -EINVAL;
2586
2587         if (ph.len > sizeof(struct ublk_params))
2588                 ph.len = sizeof(struct ublk_params);
2589
2590         mutex_lock(&ub->mutex);
2591         ublk_ctrl_fill_params_devt(ub);
2592         if (copy_to_user(argp, &ub->params, ph.len))
2593                 ret = -EFAULT;
2594         else
2595                 ret = 0;
2596         mutex_unlock(&ub->mutex);
2597
2598         return ret;
2599 }
2600
2601 static int ublk_ctrl_set_params(struct ublk_device *ub,
2602                 struct io_uring_cmd *cmd)
2603 {
2604         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2605         void __user *argp = (void __user *)(unsigned long)header->addr;
2606         struct ublk_params_header ph;
2607         int ret = -EFAULT;
2608
2609         if (header->len <= sizeof(ph) || !header->addr)
2610                 return -EINVAL;
2611
2612         if (copy_from_user(&ph, argp, sizeof(ph)))
2613                 return -EFAULT;
2614
2615         if (ph.len > header->len || !ph.len || !ph.types)
2616                 return -EINVAL;
2617
2618         if (ph.len > sizeof(struct ublk_params))
2619                 ph.len = sizeof(struct ublk_params);
2620
2621         /* parameters can only be changed when device isn't live */
2622         mutex_lock(&ub->mutex);
2623         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
2624                 ret = -EACCES;
2625         } else if (copy_from_user(&ub->params, argp, ph.len)) {
2626                 ret = -EFAULT;
2627         } else {
2628                 /* clear all we don't support yet */
2629                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
2630                 ret = ublk_validate_params(ub);
2631                 if (ret)
2632                         ub->params.types = 0;
2633         }
2634         mutex_unlock(&ub->mutex);
2635
2636         return ret;
2637 }
2638
2639 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
2640 {
2641         int i;
2642
2643         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
2644
2645         /* All old ioucmds have to be completed */
2646         ubq->nr_io_ready = 0;
2647         /* old daemon is PF_EXITING, put it now */
2648         put_task_struct(ubq->ubq_daemon);
2649         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
2650         ubq->ubq_daemon = NULL;
2651         ubq->timeout = false;
2652         ubq->canceling = false;
2653
2654         for (i = 0; i < ubq->q_depth; i++) {
2655                 struct ublk_io *io = &ubq->ios[i];
2656
2657                 /* forget everything now and be ready for new FETCH_REQ */
2658                 io->flags = 0;
2659                 io->cmd = NULL;
2660                 io->addr = 0;
2661         }
2662 }
2663
2664 static int ublk_ctrl_start_recovery(struct ublk_device *ub,
2665                 struct io_uring_cmd *cmd)
2666 {
2667         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2668         int ret = -EINVAL;
2669         int i;
2670
2671         mutex_lock(&ub->mutex);
2672         if (!ublk_can_use_recovery(ub))
2673                 goto out_unlock;
2674         /*
2675          * START_RECOVERY is only allowd after:
2676          *
2677          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
2678          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
2679          *     released.
2680          *
2681          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
2682          *     (a)has quiesced request queue
2683          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
2684          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
2685          *     (d)has completed/camceled all ioucmds owned by ther dying process
2686          */
2687         if (test_bit(UB_STATE_OPEN, &ub->state) ||
2688                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2689                 ret = -EBUSY;
2690                 goto out_unlock;
2691         }
2692         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
2693         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
2694                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
2695         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
2696         ub->mm = NULL;
2697         ub->nr_queues_ready = 0;
2698         ub->nr_privileged_daemon = 0;
2699         init_completion(&ub->completion);
2700         ret = 0;
2701  out_unlock:
2702         mutex_unlock(&ub->mutex);
2703         return ret;
2704 }
2705
2706 static int ublk_ctrl_end_recovery(struct ublk_device *ub,
2707                 struct io_uring_cmd *cmd)
2708 {
2709         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2710         int ublksrv_pid = (int)header->data[0];
2711         int ret = -EINVAL;
2712
2713         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
2714                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2715         /* wait until new ubq_daemon sending all FETCH_REQ */
2716         if (wait_for_completion_interruptible(&ub->completion))
2717                 return -EINTR;
2718
2719         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2720                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2721
2722         mutex_lock(&ub->mutex);
2723         if (!ublk_can_use_recovery(ub))
2724                 goto out_unlock;
2725
2726         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2727                 ret = -EBUSY;
2728                 goto out_unlock;
2729         }
2730         ub->dev_info.ublksrv_pid = ublksrv_pid;
2731         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2732                         __func__, ublksrv_pid, header->dev_id);
2733         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2734         pr_devel("%s: queue unquiesced, dev id %d.\n",
2735                         __func__, header->dev_id);
2736         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2737         ub->dev_info.state = UBLK_S_DEV_LIVE;
2738         ret = 0;
2739  out_unlock:
2740         mutex_unlock(&ub->mutex);
2741         return ret;
2742 }
2743
2744 static int ublk_ctrl_get_features(struct io_uring_cmd *cmd)
2745 {
2746         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2747         void __user *argp = (void __user *)(unsigned long)header->addr;
2748         u64 features = UBLK_F_ALL & ~UBLK_F_SUPPORT_ZERO_COPY;
2749
2750         if (header->len != UBLK_FEATURES_LEN || !header->addr)
2751                 return -EINVAL;
2752
2753         if (copy_to_user(argp, &features, UBLK_FEATURES_LEN))
2754                 return -EFAULT;
2755
2756         return 0;
2757 }
2758
2759 /*
2760  * All control commands are sent via /dev/ublk-control, so we have to check
2761  * the destination device's permission
2762  */
2763 static int ublk_char_dev_permission(struct ublk_device *ub,
2764                 const char *dev_path, int mask)
2765 {
2766         int err;
2767         struct path path;
2768         struct kstat stat;
2769
2770         err = kern_path(dev_path, LOOKUP_FOLLOW, &path);
2771         if (err)
2772                 return err;
2773
2774         err = vfs_getattr(&path, &stat, STATX_TYPE, AT_STATX_SYNC_AS_STAT);
2775         if (err)
2776                 goto exit;
2777
2778         err = -EPERM;
2779         if (stat.rdev != ub->cdev_dev.devt || !S_ISCHR(stat.mode))
2780                 goto exit;
2781
2782         err = inode_permission(&nop_mnt_idmap,
2783                         d_backing_inode(path.dentry), mask);
2784 exit:
2785         path_put(&path);
2786         return err;
2787 }
2788
2789 static int ublk_ctrl_uring_cmd_permission(struct ublk_device *ub,
2790                 struct io_uring_cmd *cmd)
2791 {
2792         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)io_uring_sqe_cmd(cmd->sqe);
2793         bool unprivileged = ub->dev_info.flags & UBLK_F_UNPRIVILEGED_DEV;
2794         void __user *argp = (void __user *)(unsigned long)header->addr;
2795         char *dev_path = NULL;
2796         int ret = 0;
2797         int mask;
2798
2799         if (!unprivileged) {
2800                 if (!capable(CAP_SYS_ADMIN))
2801                         return -EPERM;
2802                 /*
2803                  * The new added command of UBLK_CMD_GET_DEV_INFO2 includes
2804                  * char_dev_path in payload too, since userspace may not
2805                  * know if the specified device is created as unprivileged
2806                  * mode.
2807                  */
2808                 if (_IOC_NR(cmd->cmd_op) != UBLK_CMD_GET_DEV_INFO2)
2809                         return 0;
2810         }
2811
2812         /*
2813          * User has to provide the char device path for unprivileged ublk
2814          *
2815          * header->addr always points to the dev path buffer, and
2816          * header->dev_path_len records length of dev path buffer.
2817          */
2818         if (!header->dev_path_len || header->dev_path_len > PATH_MAX)
2819                 return -EINVAL;
2820
2821         if (header->len < header->dev_path_len)
2822                 return -EINVAL;
2823
2824         dev_path = memdup_user_nul(argp, header->dev_path_len);
2825         if (IS_ERR(dev_path))
2826                 return PTR_ERR(dev_path);
2827
2828         ret = -EINVAL;
2829         switch (_IOC_NR(cmd->cmd_op)) {
2830         case UBLK_CMD_GET_DEV_INFO:
2831         case UBLK_CMD_GET_DEV_INFO2:
2832         case UBLK_CMD_GET_QUEUE_AFFINITY:
2833         case UBLK_CMD_GET_PARAMS:
2834         case (_IOC_NR(UBLK_U_CMD_GET_FEATURES)):
2835                 mask = MAY_READ;
2836                 break;
2837         case UBLK_CMD_START_DEV:
2838         case UBLK_CMD_STOP_DEV:
2839         case UBLK_CMD_ADD_DEV:
2840         case UBLK_CMD_DEL_DEV:
2841         case UBLK_CMD_SET_PARAMS:
2842         case UBLK_CMD_START_USER_RECOVERY:
2843         case UBLK_CMD_END_USER_RECOVERY:
2844                 mask = MAY_READ | MAY_WRITE;
2845                 break;
2846         default:
2847                 goto exit;
2848         }
2849
2850         ret = ublk_char_dev_permission(ub, dev_path, mask);
2851         if (!ret) {
2852                 header->len -= header->dev_path_len;
2853                 header->addr += header->dev_path_len;
2854         }
2855         pr_devel("%s: dev id %d cmd_op %x uid %d gid %d path %s ret %d\n",
2856                         __func__, ub->ub_number, cmd->cmd_op,
2857                         ub->dev_info.owner_uid, ub->dev_info.owner_gid,
2858                         dev_path, ret);
2859 exit:
2860         kfree(dev_path);
2861         return ret;
2862 }
2863
2864 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2865                 unsigned int issue_flags)
2866 {
2867         const struct ublksrv_ctrl_cmd *header = io_uring_sqe_cmd(cmd->sqe);
2868         struct ublk_device *ub = NULL;
2869         u32 cmd_op = cmd->cmd_op;
2870         int ret = -EINVAL;
2871
2872         if (issue_flags & IO_URING_F_NONBLOCK)
2873                 return -EAGAIN;
2874
2875         ublk_ctrl_cmd_dump(cmd);
2876
2877         if (!(issue_flags & IO_URING_F_SQE128))
2878                 goto out;
2879
2880         ret = ublk_check_cmd_op(cmd_op);
2881         if (ret)
2882                 goto out;
2883
2884         if (cmd_op == UBLK_U_CMD_GET_FEATURES) {
2885                 ret = ublk_ctrl_get_features(cmd);
2886                 goto out;
2887         }
2888
2889         if (_IOC_NR(cmd_op) != UBLK_CMD_ADD_DEV) {
2890                 ret = -ENODEV;
2891                 ub = ublk_get_device_from_id(header->dev_id);
2892                 if (!ub)
2893                         goto out;
2894
2895                 ret = ublk_ctrl_uring_cmd_permission(ub, cmd);
2896                 if (ret)
2897                         goto put_dev;
2898         }
2899
2900         switch (_IOC_NR(cmd_op)) {
2901         case UBLK_CMD_START_DEV:
2902                 ret = ublk_ctrl_start_dev(ub, cmd);
2903                 break;
2904         case UBLK_CMD_STOP_DEV:
2905                 ret = ublk_ctrl_stop_dev(ub);
2906                 break;
2907         case UBLK_CMD_GET_DEV_INFO:
2908         case UBLK_CMD_GET_DEV_INFO2:
2909                 ret = ublk_ctrl_get_dev_info(ub, cmd);
2910                 break;
2911         case UBLK_CMD_ADD_DEV:
2912                 ret = ublk_ctrl_add_dev(cmd);
2913                 break;
2914         case UBLK_CMD_DEL_DEV:
2915                 ret = ublk_ctrl_del_dev(&ub);
2916                 break;
2917         case UBLK_CMD_GET_QUEUE_AFFINITY:
2918                 ret = ublk_ctrl_get_queue_affinity(ub, cmd);
2919                 break;
2920         case UBLK_CMD_GET_PARAMS:
2921                 ret = ublk_ctrl_get_params(ub, cmd);
2922                 break;
2923         case UBLK_CMD_SET_PARAMS:
2924                 ret = ublk_ctrl_set_params(ub, cmd);
2925                 break;
2926         case UBLK_CMD_START_USER_RECOVERY:
2927                 ret = ublk_ctrl_start_recovery(ub, cmd);
2928                 break;
2929         case UBLK_CMD_END_USER_RECOVERY:
2930                 ret = ublk_ctrl_end_recovery(ub, cmd);
2931                 break;
2932         default:
2933                 ret = -ENOTSUPP;
2934                 break;
2935         }
2936
2937  put_dev:
2938         if (ub)
2939                 ublk_put_device(ub);
2940  out:
2941         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2942         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2943                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2944         return -EIOCBQUEUED;
2945 }
2946
2947 static const struct file_operations ublk_ctl_fops = {
2948         .open           = nonseekable_open,
2949         .uring_cmd      = ublk_ctrl_uring_cmd,
2950         .owner          = THIS_MODULE,
2951         .llseek         = noop_llseek,
2952 };
2953
2954 static struct miscdevice ublk_misc = {
2955         .minor          = MISC_DYNAMIC_MINOR,
2956         .name           = "ublk-control",
2957         .fops           = &ublk_ctl_fops,
2958 };
2959
2960 static int __init ublk_init(void)
2961 {
2962         int ret;
2963
2964         BUILD_BUG_ON((u64)UBLKSRV_IO_BUF_OFFSET +
2965                         UBLKSRV_IO_BUF_TOTAL_SIZE < UBLKSRV_IO_BUF_OFFSET);
2966
2967         init_waitqueue_head(&ublk_idr_wq);
2968
2969         ret = misc_register(&ublk_misc);
2970         if (ret)
2971                 return ret;
2972
2973         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2974         if (ret)
2975                 goto unregister_mis;
2976
2977         ret = class_register(&ublk_chr_class);
2978         if (ret)
2979                 goto free_chrdev_region;
2980
2981         return 0;
2982
2983 free_chrdev_region:
2984         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2985 unregister_mis:
2986         misc_deregister(&ublk_misc);
2987         return ret;
2988 }
2989
2990 static void __exit ublk_exit(void)
2991 {
2992         struct ublk_device *ub;
2993         int id;
2994
2995         idr_for_each_entry(&ublk_index_idr, ub, id)
2996                 ublk_remove(ub);
2997
2998         class_unregister(&ublk_chr_class);
2999         misc_deregister(&ublk_misc);
3000
3001         idr_destroy(&ublk_index_idr);
3002         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
3003 }
3004
3005 module_init(ublk_init);
3006 module_exit(ublk_exit);
3007
3008 static int ublk_set_max_ublks(const char *buf, const struct kernel_param *kp)
3009 {
3010         return param_set_uint_minmax(buf, kp, 0, UBLK_MAX_UBLKS);
3011 }
3012
3013 static int ublk_get_max_ublks(char *buf, const struct kernel_param *kp)
3014 {
3015         return sysfs_emit(buf, "%u\n", ublks_max);
3016 }
3017
3018 static const struct kernel_param_ops ublk_max_ublks_ops = {
3019         .set = ublk_set_max_ublks,
3020         .get = ublk_get_max_ublks,
3021 };
3022
3023 module_param_cb(ublks_max, &ublk_max_ublks_ops, &ublks_max, 0644);
3024 MODULE_PARM_DESC(ublks_max, "max number of ublk devices allowed to add(default: 64)");
3025
3026 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
3027 MODULE_LICENSE("GPL");