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 <uapi/linux/ublk_cmd.h>
  46
  47 #define UBLK_MINORS             (1U << MINORBITS)
  48
  49 /* All UBLK_F_* have to be included into UBLK_F_ALL */
  50 #define UBLK_F_ALL (UBLK_F_SUPPORT_ZERO_COPY \
  51                 | UBLK_F_URING_CMD_COMP_IN_TASK \
  52                 | UBLK_F_NEED_GET_DATA \
  53                 | UBLK_F_USER_RECOVERY \
  54                 | UBLK_F_USER_RECOVERY_REISSUE)
  55
  56 /* All UBLK_PARAM_TYPE_* should be included here */
  57 #define UBLK_PARAM_TYPE_ALL (UBLK_PARAM_TYPE_BASIC | UBLK_PARAM_TYPE_DISCARD)
  58
  59 struct ublk_rq_data {
  60         struct llist_node node;
  61         struct callback_head work;
  62 };
  63
  64 struct ublk_uring_cmd_pdu {
  65         struct ublk_queue *ubq;
  66 };
  67
  68 /*
  69  * io command is active: sqe cmd is received, and its cqe isn't done
  70  *
  71  * If the flag is set, the io command is owned by ublk driver, and waited
  72  * for incoming blk-mq request from the ublk block device.
  73  *
  74  * If the flag is cleared, the io command will be completed, and owned by
  75  * ublk server.
  76  */
  77 #define UBLK_IO_FLAG_ACTIVE     0x01
  78
  79 /*
  80  * IO command is completed via cqe, and it is being handled by ublksrv, and
  81  * not committed yet
  82  *
  83  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  84  * cross verification
  85  */
  86 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  87
  88 /*
  89  * IO command is aborted, so this flag is set in case of
  90  * !UBLK_IO_FLAG_ACTIVE.
  91  *
  92  * After this flag is observed, any pending or new incoming request
  93  * associated with this io command will be failed immediately
  94  */
  95 #define UBLK_IO_FLAG_ABORTED 0x04
  96
  97 /*
  98  * UBLK_IO_FLAG_NEED_GET_DATA is set because IO command requires
  99  * get data buffer address from ublksrv.
 100  *
 101  * Then, bio data could be copied into this data buffer for a WRITE request
 102  * after the IO command is issued again and UBLK_IO_FLAG_NEED_GET_DATA is unset.
 103  */
 104 #define UBLK_IO_FLAG_NEED_GET_DATA 0x08
 105
 106 struct ublk_io {
 107         /* userspace buffer address from io cmd */
 108         __u64   addr;
 109         unsigned int flags;
 110         int res;
 111
 112         struct io_uring_cmd *cmd;
 113 };
 114
 115 struct ublk_queue {
 116         int q_id;
 117         int q_depth;
 118
 119         unsigned long flags;
 120         struct task_struct      *ubq_daemon;
 121         char *io_cmd_buf;
 122
 123         struct llist_head       io_cmds;
 124
 125         unsigned long io_addr;  /* mapped vm address */
 126         unsigned int max_io_sz;
 127         bool force_abort;
 128         unsigned short nr_io_ready;     /* how many ios setup */
 129         struct ublk_device *dev;
 130         struct ublk_io ios[];
 131 };
 132
 133 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 134
 135 struct ublk_device {
 136         struct gendisk          *ub_disk;
 137
 138         char    *__queues;
 139
 140         unsigned short  queue_size;
 141         struct ublksrv_ctrl_dev_info    dev_info;
 142
 143         struct blk_mq_tag_set   tag_set;
 144
 145         struct cdev             cdev;
 146         struct device           cdev_dev;
 147
 148 #define UB_STATE_OPEN           0
 149 #define UB_STATE_USED           1
 150         unsigned long           state;
 151         int                     ub_number;
 152
 153         struct mutex            mutex;
 154
 155         spinlock_t              mm_lock;
 156         struct mm_struct        *mm;
 157
 158         struct ublk_params      params;
 159
 160         struct completion       completion;
 161         unsigned int            nr_queues_ready;
 162         unsigned int            nr_privileged_daemon;
 163
 164         /*
 165          * Our ubq->daemon may be killed without any notification, so
 166          * monitor each queue's daemon periodically
 167          */
 168         struct delayed_work     monitor_work;
 169         struct work_struct      quiesce_work;
 170         struct work_struct      stop_work;
 171 };
 172
 173 /* header of ublk_params */
 174 struct ublk_params_header {
 175         __u32   len;
 176         __u32   types;
 177 };
 178
 179 static dev_t ublk_chr_devt;
 180 static struct class *ublk_chr_class;
 181
 182 static DEFINE_IDR(ublk_index_idr);
 183 static DEFINE_SPINLOCK(ublk_idr_lock);
 184 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 185
 186 static DEFINE_MUTEX(ublk_ctl_mutex);
 187
 188 static struct miscdevice ublk_misc;
 189
 190 static void ublk_dev_param_basic_apply(struct ublk_device *ub)
 191 {
 192         struct request_queue *q = ub->ub_disk->queue;
 193         const struct ublk_param_basic *p = &ub->params.basic;
 194
 195         blk_queue_logical_block_size(q, 1 << p->logical_bs_shift);
 196         blk_queue_physical_block_size(q, 1 << p->physical_bs_shift);
 197         blk_queue_io_min(q, 1 << p->io_min_shift);
 198         blk_queue_io_opt(q, 1 << p->io_opt_shift);
 199
 200         blk_queue_write_cache(q, p->attrs & UBLK_ATTR_VOLATILE_CACHE,
 201                         p->attrs & UBLK_ATTR_FUA);
 202         if (p->attrs & UBLK_ATTR_ROTATIONAL)
 203                 blk_queue_flag_clear(QUEUE_FLAG_NONROT, q);
 204         else
 205                 blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
 206
 207         blk_queue_max_hw_sectors(q, p->max_sectors);
 208         blk_queue_chunk_sectors(q, p->chunk_sectors);
 209         blk_queue_virt_boundary(q, p->virt_boundary_mask);
 210
 211         if (p->attrs & UBLK_ATTR_READ_ONLY)
 212                 set_disk_ro(ub->ub_disk, true);
 213
 214         set_capacity(ub->ub_disk, p->dev_sectors);
 215 }
 216
 217 static void ublk_dev_param_discard_apply(struct ublk_device *ub)
 218 {
 219         struct request_queue *q = ub->ub_disk->queue;
 220         const struct ublk_param_discard *p = &ub->params.discard;
 221
 222         q->limits.discard_alignment = p->discard_alignment;
 223         q->limits.discard_granularity = p->discard_granularity;
 224         blk_queue_max_discard_sectors(q, p->max_discard_sectors);
 225         blk_queue_max_write_zeroes_sectors(q,
 226                         p->max_write_zeroes_sectors);
 227         blk_queue_max_discard_segments(q, p->max_discard_segments);
 228 }
 229
 230 static int ublk_validate_params(const struct ublk_device *ub)
 231 {
 232         /* basic param is the only one which must be set */
 233         if (ub->params.types & UBLK_PARAM_TYPE_BASIC) {
 234                 const struct ublk_param_basic *p = &ub->params.basic;
 235
 236                 if (p->logical_bs_shift > PAGE_SHIFT)
 237                         return -EINVAL;
 238
 239                 if (p->logical_bs_shift > p->physical_bs_shift)
 240                         return -EINVAL;
 241
 242                 if (p->max_sectors > (ub->dev_info.max_io_buf_bytes >> 9))
 243                         return -EINVAL;
 244         } else
 245                 return -EINVAL;
 246
 247         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD) {
 248                 const struct ublk_param_discard *p = &ub->params.discard;
 249
 250                 /* So far, only support single segment discard */
 251                 if (p->max_discard_sectors && p->max_discard_segments != 1)
 252                         return -EINVAL;
 253
 254                 if (!p->discard_granularity)
 255                         return -EINVAL;
 256         }
 257
 258         return 0;
 259 }
 260
 261 static int ublk_apply_params(struct ublk_device *ub)
 262 {
 263         if (!(ub->params.types & UBLK_PARAM_TYPE_BASIC))
 264                 return -EINVAL;
 265
 266         ublk_dev_param_basic_apply(ub);
 267
 268         if (ub->params.types & UBLK_PARAM_TYPE_DISCARD)
 269                 ublk_dev_param_discard_apply(ub);
 270
 271         return 0;
 272 }
 273
 274 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
 275 {
 276         if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
 277                         !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
 278                 return true;
 279         return false;
 280 }
 281
 282 static inline bool ublk_need_get_data(const struct ublk_queue *ubq)
 283 {
 284         if (ubq->flags & UBLK_F_NEED_GET_DATA)
 285                 return true;
 286         return false;
 287 }
 288
 289 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 290 {
 291         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 292                 return ub;
 293         return NULL;
 294 }
 295
 296 static void ublk_put_device(struct ublk_device *ub)
 297 {
 298         put_device(&ub->cdev_dev);
 299 }
 300
 301 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 302                 int qid)
 303 {
 304        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 305 }
 306
 307 static inline bool ublk_rq_has_data(const struct request *rq)
 308 {
 309         return rq->bio && bio_has_data(rq->bio);
 310 }
 311
 312 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 313                 int tag)
 314 {
 315         return (struct ublksrv_io_desc *)
 316                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 317 }
 318
 319 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 320 {
 321         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 322 }
 323
 324 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 325 {
 326         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 327
 328         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 329                         PAGE_SIZE);
 330 }
 331
 332 static inline bool ublk_queue_can_use_recovery_reissue(
 333                 struct ublk_queue *ubq)
 334 {
 335         if ((ubq->flags & UBLK_F_USER_RECOVERY) &&
 336                         (ubq->flags & UBLK_F_USER_RECOVERY_REISSUE))
 337                 return true;
 338         return false;
 339 }
 340
 341 static inline bool ublk_queue_can_use_recovery(
 342                 struct ublk_queue *ubq)
 343 {
 344         if (ubq->flags & UBLK_F_USER_RECOVERY)
 345                 return true;
 346         return false;
 347 }
 348
 349 static inline bool ublk_can_use_recovery(struct ublk_device *ub)
 350 {
 351         if (ub->dev_info.flags & UBLK_F_USER_RECOVERY)
 352                 return true;
 353         return false;
 354 }
 355
 356 static void ublk_free_disk(struct gendisk *disk)
 357 {
 358         struct ublk_device *ub = disk->private_data;
 359
 360         clear_bit(UB_STATE_USED, &ub->state);
 361         put_device(&ub->cdev_dev);
 362 }
 363
 364 static const struct block_device_operations ub_fops = {
 365         .owner =        THIS_MODULE,
 366         .free_disk =    ublk_free_disk,
 367 };
 368
 369 #define UBLK_MAX_PIN_PAGES      32
 370
 371 struct ublk_map_data {
 372         const struct ublk_queue *ubq;
 373         const struct request *rq;
 374         const struct ublk_io *io;
 375         unsigned max_bytes;
 376 };
 377
 378 struct ublk_io_iter {
 379         struct page *pages[UBLK_MAX_PIN_PAGES];
 380         unsigned pg_off;        /* offset in the 1st page in pages */
 381         int nr_pages;           /* how many page pointers in pages */
 382         struct bio *bio;
 383         struct bvec_iter iter;
 384 };
 385
 386 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
 387                 unsigned max_bytes, bool to_vm)
 388 {
 389         const unsigned total = min_t(unsigned, max_bytes,
 390                         PAGE_SIZE - data->pg_off +
 391                         ((data->nr_pages - 1) << PAGE_SHIFT));
 392         unsigned done = 0;
 393         unsigned pg_idx = 0;
 394
 395         while (done < total) {
 396                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 397                 const unsigned int bytes = min3(bv.bv_len, total - done,
 398                                 (unsigned)(PAGE_SIZE - data->pg_off));
 399                 void *bv_buf = bvec_kmap_local(&bv);
 400                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 401
 402                 if (to_vm)
 403                         memcpy(pg_buf + data->pg_off, bv_buf, bytes);
 404                 else
 405                         memcpy(bv_buf, pg_buf + data->pg_off, bytes);
 406
 407                 kunmap_local(pg_buf);
 408                 kunmap_local(bv_buf);
 409
 410                 /* advance page array */
 411                 data->pg_off += bytes;
 412                 if (data->pg_off == PAGE_SIZE) {
 413                         pg_idx += 1;
 414                         data->pg_off = 0;
 415                 }
 416
 417                 done += bytes;
 418
 419                 /* advance bio */
 420                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 421                 if (!data->iter.bi_size) {
 422                         data->bio = data->bio->bi_next;
 423                         if (data->bio == NULL)
 424                                 break;
 425                         data->iter = data->bio->bi_iter;
 426                 }
 427         }
 428
 429         return done;
 430 }
 431
 432 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
 433                 bool to_vm)
 434 {
 435         const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
 436         const unsigned long start_vm = data->io->addr;
 437         unsigned int done = 0;
 438         struct ublk_io_iter iter = {
 439                 .pg_off = start_vm & (PAGE_SIZE - 1),
 440                 .bio    = data->rq->bio,
 441                 .iter   = data->rq->bio->bi_iter,
 442         };
 443         const unsigned int nr_pages = round_up(data->max_bytes +
 444                         (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
 445
 446         while (done < nr_pages) {
 447                 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
 448                                 nr_pages - done);
 449                 unsigned i, len;
 450
 451                 iter.nr_pages = get_user_pages_fast(start_vm +
 452                                 (done << PAGE_SHIFT), to_pin, gup_flags,
 453                                 iter.pages);
 454                 if (iter.nr_pages <= 0)
 455                         return done == 0 ? iter.nr_pages : done;
 456                 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
 457                 for (i = 0; i < iter.nr_pages; i++) {
 458                         if (to_vm)
 459                                 set_page_dirty(iter.pages[i]);
 460                         put_page(iter.pages[i]);
 461                 }
 462                 data->max_bytes -= len;
 463                 done += iter.nr_pages;
 464         }
 465
 466         return done;
 467 }
 468
 469 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 470                 struct ublk_io *io)
 471 {
 472         const unsigned int rq_bytes = blk_rq_bytes(req);
 473         /*
 474          * no zero copy, we delay copy WRITE request data into ublksrv
 475          * context and the big benefit is that pinning pages in current
 476          * context is pretty fast, see ublk_pin_user_pages
 477          */
 478         if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
 479                 return rq_bytes;
 480
 481         if (ublk_rq_has_data(req)) {
 482                 struct ublk_map_data data = {
 483                         .ubq    =       ubq,
 484                         .rq     =       req,
 485                         .io     =       io,
 486                         .max_bytes =    rq_bytes,
 487                 };
 488
 489                 ublk_copy_user_pages(&data, true);
 490
 491                 return rq_bytes - data.max_bytes;
 492         }
 493         return rq_bytes;
 494 }
 495
 496 static int ublk_unmap_io(const struct ublk_queue *ubq,
 497                 const struct request *req,
 498                 struct ublk_io *io)
 499 {
 500         const unsigned int rq_bytes = blk_rq_bytes(req);
 501
 502         if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
 503                 struct ublk_map_data data = {
 504                         .ubq    =       ubq,
 505                         .rq     =       req,
 506                         .io     =       io,
 507                         .max_bytes =    io->res,
 508                 };
 509
 510                 WARN_ON_ONCE(io->res > rq_bytes);
 511
 512                 ublk_copy_user_pages(&data, false);
 513
 514                 return io->res - data.max_bytes;
 515         }
 516         return rq_bytes;
 517 }
 518
 519 static inline unsigned int ublk_req_build_flags(struct request *req)
 520 {
 521         unsigned flags = 0;
 522
 523         if (req->cmd_flags & REQ_FAILFAST_DEV)
 524                 flags |= UBLK_IO_F_FAILFAST_DEV;
 525
 526         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 527                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 528
 529         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 530                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 531
 532         if (req->cmd_flags & REQ_META)
 533                 flags |= UBLK_IO_F_META;
 534
 535         if (req->cmd_flags & REQ_FUA)
 536                 flags |= UBLK_IO_F_FUA;
 537
 538         if (req->cmd_flags & REQ_NOUNMAP)
 539                 flags |= UBLK_IO_F_NOUNMAP;
 540
 541         if (req->cmd_flags & REQ_SWAP)
 542                 flags |= UBLK_IO_F_SWAP;
 543
 544         return flags;
 545 }
 546
 547 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 548 {
 549         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 550         struct ublk_io *io = &ubq->ios[req->tag];
 551         u32 ublk_op;
 552
 553         switch (req_op(req)) {
 554         case REQ_OP_READ:
 555                 ublk_op = UBLK_IO_OP_READ;
 556                 break;
 557         case REQ_OP_WRITE:
 558                 ublk_op = UBLK_IO_OP_WRITE;
 559                 break;
 560         case REQ_OP_FLUSH:
 561                 ublk_op = UBLK_IO_OP_FLUSH;
 562                 break;
 563         case REQ_OP_DISCARD:
 564                 ublk_op = UBLK_IO_OP_DISCARD;
 565                 break;
 566         case REQ_OP_WRITE_ZEROES:
 567                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 568                 break;
 569         default:
 570                 return BLK_STS_IOERR;
 571         }
 572
 573         /* need to translate since kernel may change */
 574         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 575         iod->nr_sectors = blk_rq_sectors(req);
 576         iod->start_sector = blk_rq_pos(req);
 577         iod->addr = io->addr;
 578
 579         return BLK_STS_OK;
 580 }
 581
 582 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 583                 struct io_uring_cmd *ioucmd)
 584 {
 585         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 586 }
 587
 588 static inline bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 589 {
 590         return ubq->ubq_daemon->flags & PF_EXITING;
 591 }
 592
 593 /* todo: handle partial completion */
 594 static void ublk_complete_rq(struct request *req)
 595 {
 596         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 597         struct ublk_io *io = &ubq->ios[req->tag];
 598         unsigned int unmapped_bytes;
 599
 600         /* failed read IO if nothing is read */
 601         if (!io->res && req_op(req) == REQ_OP_READ)
 602                 io->res = -EIO;
 603
 604         if (io->res < 0) {
 605                 blk_mq_end_request(req, errno_to_blk_status(io->res));
 606                 return;
 607         }
 608
 609         /*
 610          * FLUSH or DISCARD usually won't return bytes returned, so end them
 611          * directly.
 612          *
 613          * Both the two needn't unmap.
 614          */
 615         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
 616                 blk_mq_end_request(req, BLK_STS_OK);
 617                 return;
 618         }
 619
 620         /* for READ request, writing data in iod->addr to rq buffers */
 621         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 622
 623         /*
 624          * Extremely impossible since we got data filled in just before
 625          *
 626          * Re-read simply for this unlikely case.
 627          */
 628         if (unlikely(unmapped_bytes < io->res))
 629                 io->res = unmapped_bytes;
 630
 631         if (blk_update_request(req, BLK_STS_OK, io->res))
 632                 blk_mq_requeue_request(req, true);
 633         else
 634                 __blk_mq_end_request(req, BLK_STS_OK);
 635 }
 636
 637 /*
 638  * Since __ublk_rq_task_work always fails requests immediately during
 639  * exiting, __ublk_fail_req() is only called from abort context during
 640  * exiting. So lock is unnecessary.
 641  *
 642  * Also aborting may not be started yet, keep in mind that one failed
 643  * request may be issued by block layer again.
 644  */
 645 static void __ublk_fail_req(struct ublk_queue *ubq, struct ublk_io *io,
 646                 struct request *req)
 647 {
 648         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 649
 650         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 651                 io->flags |= UBLK_IO_FLAG_ABORTED;
 652                 if (ublk_queue_can_use_recovery_reissue(ubq))
 653                         blk_mq_requeue_request(req, false);
 654                 else
 655                         blk_mq_end_request(req, BLK_STS_IOERR);
 656         }
 657 }
 658
 659 static void ubq_complete_io_cmd(struct ublk_io *io, int res)
 660 {
 661         /* mark this cmd owned by ublksrv */
 662         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 663
 664         /*
 665          * clear ACTIVE since we are done with this sqe/cmd slot
 666          * We can only accept io cmd in case of being not active.
 667          */
 668         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 669
 670         /* tell ublksrv one io request is coming */
 671         io_uring_cmd_done(io->cmd, res, 0);
 672 }
 673
 674 #define UBLK_REQUEUE_DELAY_MS   3
 675
 676 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 677                 struct request *rq)
 678 {
 679         /* We cannot process this rq so just requeue it. */
 680         if (ublk_queue_can_use_recovery(ubq))
 681                 blk_mq_requeue_request(rq, false);
 682         else
 683                 blk_mq_end_request(rq, BLK_STS_IOERR);
 684
 685         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 686 }
 687
 688 static inline void __ublk_rq_task_work(struct request *req)
 689 {
 690         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 691         int tag = req->tag;
 692         struct ublk_io *io = &ubq->ios[tag];
 693         unsigned int mapped_bytes;
 694
 695         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 696                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 697                         ublk_get_iod(ubq, req->tag)->addr);
 698
 699         /*
 700          * Task is exiting if either:
 701          *
 702          * (1) current != ubq_daemon.
 703          * io_uring_cmd_complete_in_task() tries to run task_work
 704          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 705          *
 706          * (2) current->flags & PF_EXITING.
 707          */
 708         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 709                 __ublk_abort_rq(ubq, req);
 710                 return;
 711         }
 712
 713         if (ublk_need_get_data(ubq) &&
 714                         (req_op(req) == REQ_OP_WRITE ||
 715                         req_op(req) == REQ_OP_FLUSH)) {
 716                 /*
 717                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 718                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 719                  * and notify it.
 720                  */
 721                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 722                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 723                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 724                                         __func__, io->cmd->cmd_op, ubq->q_id,
 725                                         req->tag, io->flags);
 726                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA);
 727                         return;
 728                 }
 729                 /*
 730                  * We have handled UBLK_IO_NEED_GET_DATA command,
 731                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 732                  * do the copy work.
 733                  */
 734                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 735                 /* update iod->addr because ublksrv may have passed a new io buffer */
 736                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 737                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 738                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 739                                 ublk_get_iod(ubq, req->tag)->addr);
 740         }
 741
 742         mapped_bytes = ublk_map_io(ubq, req, io);
 743
 744         /* partially mapped, update io descriptor */
 745         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 746                 /*
 747                  * Nothing mapped, retry until we succeed.
 748                  *
 749                  * We may never succeed in mapping any bytes here because
 750                  * of OOM. TODO: reserve one buffer with single page pinned
 751                  * for providing forward progress guarantee.
 752                  */
 753                 if (unlikely(!mapped_bytes)) {
 754                         blk_mq_requeue_request(req, false);
 755                         blk_mq_delay_kick_requeue_list(req->q,
 756                                         UBLK_REQUEUE_DELAY_MS);
 757                         return;
 758                 }
 759
 760                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 761                         mapped_bytes >> 9;
 762         }
 763
 764         ubq_complete_io_cmd(io, UBLK_IO_RES_OK);
 765 }
 766
 767 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq)
 768 {
 769         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 770         struct ublk_rq_data *data, *tmp;
 771
 772         io_cmds = llist_reverse_order(io_cmds);
 773         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 774                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data));
 775 }
 776
 777 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 778 {
 779         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 780         struct ublk_rq_data *data, *tmp;
 781
 782         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 783                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 784 }
 785
 786 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
 787 {
 788         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 789         struct ublk_queue *ubq = pdu->ubq;
 790
 791         ublk_forward_io_cmds(ubq);
 792 }
 793
 794 static void ublk_rq_task_work_fn(struct callback_head *work)
 795 {
 796         struct ublk_rq_data *data = container_of(work,
 797                         struct ublk_rq_data, work);
 798         struct request *req = blk_mq_rq_from_pdu(data);
 799         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 800
 801         ublk_forward_io_cmds(ubq);
 802 }
 803
 804 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 805 {
 806         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 807         struct ublk_io *io;
 808
 809         if (!llist_add(&data->node, &ubq->io_cmds))
 810                 return;
 811
 812         io = &ubq->ios[rq->tag];
 813         /*
 814          * If the check pass, we know that this is a re-issued request aborted
 815          * previously in monitor_work because the ubq_daemon(cmd's task) is
 816          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 817          * because this ioucmd's io_uring context may be freed now if no inflight
 818          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 819          *
 820          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 821          * the tag). Then the request is re-started(allocating the tag) and we are here.
 822          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 823          * guarantees that here is a re-issued request aborted previously.
 824          */
 825         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 826                 ublk_abort_io_cmds(ubq);
 827         } else if (ublk_can_use_task_work(ubq)) {
 828                 if (task_work_add(ubq->ubq_daemon, &data->work,
 829                                         TWA_SIGNAL_NO_IPI))
 830                         ublk_abort_io_cmds(ubq);
 831         } else {
 832                 struct io_uring_cmd *cmd = io->cmd;
 833                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 834
 835                 pdu->ubq = ubq;
 836                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 837         }
 838 }
 839
 840 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 841                 const struct blk_mq_queue_data *bd)
 842 {
 843         struct ublk_queue *ubq = hctx->driver_data;
 844         struct request *rq = bd->rq;
 845         blk_status_t res;
 846
 847         /* fill iod to slot in io cmd buffer */
 848         res = ublk_setup_iod(ubq, rq);
 849         if (unlikely(res != BLK_STS_OK))
 850                 return BLK_STS_IOERR;
 851
 852         /* With recovery feature enabled, force_abort is set in
 853          * ublk_stop_dev() before calling del_gendisk(). We have to
 854          * abort all requeued and new rqs here to let del_gendisk()
 855          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 856          * to avoid UAF on io_uring ctx.
 857          *
 858          * Note: force_abort is guaranteed to be seen because it is set
 859          * before request queue is unqiuesced.
 860          */
 861         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 862                 return BLK_STS_IOERR;
 863
 864         blk_mq_start_request(bd->rq);
 865
 866         if (unlikely(ubq_daemon_is_dying(ubq))) {
 867                 __ublk_abort_rq(ubq, rq);
 868                 return BLK_STS_OK;
 869         }
 870
 871         ublk_queue_cmd(ubq, rq);
 872
 873         return BLK_STS_OK;
 874 }
 875
 876 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 877                 unsigned int hctx_idx)
 878 {
 879         struct ublk_device *ub = driver_data;
 880         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 881
 882         hctx->driver_data = ubq;
 883         return 0;
 884 }
 885
 886 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 887                 unsigned int hctx_idx, unsigned int numa_node)
 888 {
 889         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 890
 891         init_task_work(&data->work, ublk_rq_task_work_fn);
 892         return 0;
 893 }
 894
 895 static const struct blk_mq_ops ublk_mq_ops = {
 896         .queue_rq       = ublk_queue_rq,
 897         .init_hctx      = ublk_init_hctx,
 898         .init_request   = ublk_init_rq,
 899 };
 900
 901 static int ublk_ch_open(struct inode *inode, struct file *filp)
 902 {
 903         struct ublk_device *ub = container_of(inode->i_cdev,
 904                         struct ublk_device, cdev);
 905
 906         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 907                 return -EBUSY;
 908         filp->private_data = ub;
 909         return 0;
 910 }
 911
 912 static int ublk_ch_release(struct inode *inode, struct file *filp)
 913 {
 914         struct ublk_device *ub = filp->private_data;
 915
 916         clear_bit(UB_STATE_OPEN, &ub->state);
 917         return 0;
 918 }
 919
 920 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
 921 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
 922 {
 923         struct ublk_device *ub = filp->private_data;
 924         size_t sz = vma->vm_end - vma->vm_start;
 925         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
 926         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
 927         int q_id, ret = 0;
 928
 929         spin_lock(&ub->mm_lock);
 930         if (!ub->mm)
 931                 ub->mm = current->mm;
 932         if (current->mm != ub->mm)
 933                 ret = -EINVAL;
 934         spin_unlock(&ub->mm_lock);
 935
 936         if (ret)
 937                 return ret;
 938
 939         if (vma->vm_flags & VM_WRITE)
 940                 return -EPERM;
 941
 942         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
 943         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
 944                 return -EINVAL;
 945
 946         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
 947         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
 948                         __func__, q_id, current->pid, vma->vm_start,
 949                         phys_off, (unsigned long)sz);
 950
 951         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
 952                 return -EINVAL;
 953
 954         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
 955         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
 956 }
 957
 958 static void ublk_commit_completion(struct ublk_device *ub,
 959                 struct ublksrv_io_cmd *ub_cmd)
 960 {
 961         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
 962         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
 963         struct ublk_io *io = &ubq->ios[tag];
 964         struct request *req;
 965
 966         /* now this cmd slot is owned by nbd driver */
 967         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
 968         io->res = ub_cmd->result;
 969
 970         /* find the io request and complete */
 971         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
 972
 973         if (req && likely(!blk_should_fake_timeout(req->q)))
 974                 ublk_complete_rq(req);
 975 }
 976
 977 /*
 978  * When ->ubq_daemon is exiting, either new request is ended immediately,
 979  * or any queued io command is drained, so it is safe to abort queue
 980  * lockless
 981  */
 982 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
 983 {
 984         int i;
 985
 986         if (!ublk_get_device(ub))
 987                 return;
 988
 989         for (i = 0; i < ubq->q_depth; i++) {
 990                 struct ublk_io *io = &ubq->ios[i];
 991
 992                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
 993                         struct request *rq;
 994
 995                         /*
 996                          * Either we fail the request or ublk_rq_task_work_fn
 997                          * will do it
 998                          */
 999                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1000                         if (rq)
1001                                 __ublk_fail_req(ubq, io, rq);
1002                 }
1003         }
1004         ublk_put_device(ub);
1005 }
1006
1007 static void ublk_daemon_monitor_work(struct work_struct *work)
1008 {
1009         struct ublk_device *ub =
1010                 container_of(work, struct ublk_device, monitor_work.work);
1011         int i;
1012
1013         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1014                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1015
1016                 if (ubq_daemon_is_dying(ubq)) {
1017                         if (ublk_queue_can_use_recovery(ubq))
1018                                 schedule_work(&ub->quiesce_work);
1019                         else
1020                                 schedule_work(&ub->stop_work);
1021
1022                         /* abort queue is for making forward progress */
1023                         ublk_abort_queue(ub, ubq);
1024                 }
1025         }
1026
1027         /*
1028          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1029          * after ublk_remove() or __ublk_quiesce_dev() is started.
1030          *
1031          * No need ub->mutex, monitor work are canceled after state is marked
1032          * as not LIVE, so new state is observed reliably.
1033          */
1034         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1035                 schedule_delayed_work(&ub->monitor_work,
1036                                 UBLK_DAEMON_MONITOR_PERIOD);
1037 }
1038
1039 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1040 {
1041         return ubq->nr_io_ready == ubq->q_depth;
1042 }
1043
1044 static void ublk_cancel_queue(struct ublk_queue *ubq)
1045 {
1046         int i;
1047
1048         if (!ublk_queue_ready(ubq))
1049                 return;
1050
1051         for (i = 0; i < ubq->q_depth; i++) {
1052                 struct ublk_io *io = &ubq->ios[i];
1053
1054                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1055                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
1056         }
1057
1058         /* all io commands are canceled */
1059         ubq->nr_io_ready = 0;
1060 }
1061
1062 /* Cancel all pending commands, must be called after del_gendisk() returns */
1063 static void ublk_cancel_dev(struct ublk_device *ub)
1064 {
1065         int i;
1066
1067         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1068                 ublk_cancel_queue(ublk_get_queue(ub, i));
1069 }
1070
1071 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1072 {
1073         bool *idle = data;
1074
1075         if (blk_mq_request_started(rq)) {
1076                 *idle = false;
1077                 return false;
1078         }
1079         return true;
1080 }
1081
1082 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1083 {
1084         bool idle;
1085
1086         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1087         while (true) {
1088                 idle = true;
1089                 blk_mq_tagset_busy_iter(&ub->tag_set,
1090                                 ublk_check_inflight_rq, &idle);
1091                 if (idle)
1092                         break;
1093                 msleep(UBLK_REQUEUE_DELAY_MS);
1094         }
1095 }
1096
1097 static void __ublk_quiesce_dev(struct ublk_device *ub)
1098 {
1099         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1100                         __func__, ub->dev_info.dev_id,
1101                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1102                         "LIVE" : "QUIESCED");
1103         blk_mq_quiesce_queue(ub->ub_disk->queue);
1104         ublk_wait_tagset_rqs_idle(ub);
1105         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1106         ublk_cancel_dev(ub);
1107         /* we are going to release task_struct of ubq_daemon and resets
1108          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1109          * Besides, monitor_work is not necessary in QUIESCED state since we have
1110          * already scheduled quiesce_work and quiesced all ubqs.
1111          *
1112          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1113          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1114          */
1115         cancel_delayed_work_sync(&ub->monitor_work);
1116 }
1117
1118 static void ublk_quiesce_work_fn(struct work_struct *work)
1119 {
1120         struct ublk_device *ub =
1121                 container_of(work, struct ublk_device, quiesce_work);
1122
1123         mutex_lock(&ub->mutex);
1124         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1125                 goto unlock;
1126         __ublk_quiesce_dev(ub);
1127  unlock:
1128         mutex_unlock(&ub->mutex);
1129 }
1130
1131 static void ublk_unquiesce_dev(struct ublk_device *ub)
1132 {
1133         int i;
1134
1135         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1136                         __func__, ub->dev_info.dev_id,
1137                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1138                         "LIVE" : "QUIESCED");
1139         /* quiesce_work has run. We let requeued rqs be aborted
1140          * before running fallback_wq. "force_abort" must be seen
1141          * after request queue is unqiuesced. Then del_gendisk()
1142          * can move on.
1143          */
1144         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1145                 ublk_get_queue(ub, i)->force_abort = true;
1146
1147         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1148         /* We may have requeued some rqs in ublk_quiesce_queue() */
1149         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1150 }
1151
1152 static void ublk_stop_dev(struct ublk_device *ub)
1153 {
1154         mutex_lock(&ub->mutex);
1155         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1156                 goto unlock;
1157         if (ublk_can_use_recovery(ub)) {
1158                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1159                         __ublk_quiesce_dev(ub);
1160                 ublk_unquiesce_dev(ub);
1161         }
1162         del_gendisk(ub->ub_disk);
1163         ub->dev_info.state = UBLK_S_DEV_DEAD;
1164         ub->dev_info.ublksrv_pid = -1;
1165         put_disk(ub->ub_disk);
1166         ub->ub_disk = NULL;
1167  unlock:
1168         ublk_cancel_dev(ub);
1169         mutex_unlock(&ub->mutex);
1170         cancel_delayed_work_sync(&ub->monitor_work);
1171 }
1172
1173 /* device can only be started after all IOs are ready */
1174 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1175 {
1176         mutex_lock(&ub->mutex);
1177         ubq->nr_io_ready++;
1178         if (ublk_queue_ready(ubq)) {
1179                 ubq->ubq_daemon = current;
1180                 get_task_struct(ubq->ubq_daemon);
1181                 ub->nr_queues_ready++;
1182
1183                 if (capable(CAP_SYS_ADMIN))
1184                         ub->nr_privileged_daemon++;
1185         }
1186         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1187                 complete_all(&ub->completion);
1188         mutex_unlock(&ub->mutex);
1189 }
1190
1191 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1192                 int tag)
1193 {
1194         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1195         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1196
1197         ublk_queue_cmd(ubq, req);
1198 }
1199
1200 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1201 {
1202         struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
1203         struct ublk_device *ub = cmd->file->private_data;
1204         struct ublk_queue *ubq;
1205         struct ublk_io *io;
1206         u32 cmd_op = cmd->cmd_op;
1207         unsigned tag = ub_cmd->tag;
1208         int ret = -EINVAL;
1209
1210         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1211                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1212                         ub_cmd->result);
1213
1214         if (!(issue_flags & IO_URING_F_SQE128))
1215                 goto out;
1216
1217         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1218                 goto out;
1219
1220         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1221         if (!ubq || ub_cmd->q_id != ubq->q_id)
1222                 goto out;
1223
1224         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1225                 goto out;
1226
1227         if (tag >= ubq->q_depth)
1228                 goto out;
1229
1230         io = &ubq->ios[tag];
1231
1232         /* there is pending io cmd, something must be wrong */
1233         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1234                 ret = -EBUSY;
1235                 goto out;
1236         }
1237
1238         /*
1239          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1240          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1241          */
1242         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1243                         ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1244                 goto out;
1245
1246         switch (cmd_op) {
1247         case UBLK_IO_FETCH_REQ:
1248                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1249                 if (ublk_queue_ready(ubq)) {
1250                         ret = -EBUSY;
1251                         goto out;
1252                 }
1253                 /*
1254                  * The io is being handled by server, so COMMIT_RQ is expected
1255                  * instead of FETCH_REQ
1256                  */
1257                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1258                         goto out;
1259                 /* FETCH_RQ has to provide IO buffer */
1260                 if (!ub_cmd->addr)
1261                         goto out;
1262                 io->cmd = cmd;
1263                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1264                 io->addr = ub_cmd->addr;
1265
1266                 ublk_mark_io_ready(ub, ubq);
1267                 break;
1268         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1269                 /* FETCH_RQ has to provide IO buffer */
1270                 if (!ub_cmd->addr)
1271                         goto out;
1272                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1273                         goto out;
1274                 io->addr = ub_cmd->addr;
1275                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1276                 io->cmd = cmd;
1277                 ublk_commit_completion(ub, ub_cmd);
1278                 break;
1279         case UBLK_IO_NEED_GET_DATA:
1280                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1281                         goto out;
1282                 io->addr = ub_cmd->addr;
1283                 io->cmd = cmd;
1284                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1285                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1286                 break;
1287         default:
1288                 goto out;
1289         }
1290         return -EIOCBQUEUED;
1291
1292  out:
1293         io_uring_cmd_done(cmd, ret, 0);
1294         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1295                         __func__, cmd_op, tag, ret, io->flags);
1296         return -EIOCBQUEUED;
1297 }
1298
1299 static const struct file_operations ublk_ch_fops = {
1300         .owner = THIS_MODULE,
1301         .open = ublk_ch_open,
1302         .release = ublk_ch_release,
1303         .llseek = no_llseek,
1304         .uring_cmd = ublk_ch_uring_cmd,
1305         .mmap = ublk_ch_mmap,
1306 };
1307
1308 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1309 {
1310         int size = ublk_queue_cmd_buf_size(ub, q_id);
1311         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1312
1313         if (ubq->ubq_daemon)
1314                 put_task_struct(ubq->ubq_daemon);
1315         if (ubq->io_cmd_buf)
1316                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1317 }
1318
1319 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1320 {
1321         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1322         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1323         void *ptr;
1324         int size;
1325
1326         ubq->flags = ub->dev_info.flags;
1327         ubq->q_id = q_id;
1328         ubq->q_depth = ub->dev_info.queue_depth;
1329         size = ublk_queue_cmd_buf_size(ub, q_id);
1330
1331         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1332         if (!ptr)
1333                 return -ENOMEM;
1334
1335         ubq->io_cmd_buf = ptr;
1336         ubq->dev = ub;
1337         return 0;
1338 }
1339
1340 static void ublk_deinit_queues(struct ublk_device *ub)
1341 {
1342         int nr_queues = ub->dev_info.nr_hw_queues;
1343         int i;
1344
1345         if (!ub->__queues)
1346                 return;
1347
1348         for (i = 0; i < nr_queues; i++)
1349                 ublk_deinit_queue(ub, i);
1350         kfree(ub->__queues);
1351 }
1352
1353 static int ublk_init_queues(struct ublk_device *ub)
1354 {
1355         int nr_queues = ub->dev_info.nr_hw_queues;
1356         int depth = ub->dev_info.queue_depth;
1357         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1358         int i, ret = -ENOMEM;
1359
1360         ub->queue_size = ubq_size;
1361         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1362         if (!ub->__queues)
1363                 return ret;
1364
1365         for (i = 0; i < nr_queues; i++) {
1366                 if (ublk_init_queue(ub, i))
1367                         goto fail;
1368         }
1369
1370         init_completion(&ub->completion);
1371         return 0;
1372
1373  fail:
1374         ublk_deinit_queues(ub);
1375         return ret;
1376 }
1377
1378 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1379 {
1380         int i = idx;
1381         int err;
1382
1383         spin_lock(&ublk_idr_lock);
1384         /* allocate id, if @id >= 0, we're requesting that specific id */
1385         if (i >= 0) {
1386                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1387                 if (err == -ENOSPC)
1388                         err = -EEXIST;
1389         } else {
1390                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1391         }
1392         spin_unlock(&ublk_idr_lock);
1393
1394         if (err >= 0)
1395                 ub->ub_number = err;
1396
1397         return err;
1398 }
1399
1400 static void ublk_free_dev_number(struct ublk_device *ub)
1401 {
1402         spin_lock(&ublk_idr_lock);
1403         idr_remove(&ublk_index_idr, ub->ub_number);
1404         wake_up_all(&ublk_idr_wq);
1405         spin_unlock(&ublk_idr_lock);
1406 }
1407
1408 static void ublk_cdev_rel(struct device *dev)
1409 {
1410         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1411
1412         blk_mq_free_tag_set(&ub->tag_set);
1413         ublk_deinit_queues(ub);
1414         ublk_free_dev_number(ub);
1415         mutex_destroy(&ub->mutex);
1416         kfree(ub);
1417 }
1418
1419 static int ublk_add_chdev(struct ublk_device *ub)
1420 {
1421         struct device *dev = &ub->cdev_dev;
1422         int minor = ub->ub_number;
1423         int ret;
1424
1425         dev->parent = ublk_misc.this_device;
1426         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1427         dev->class = ublk_chr_class;
1428         dev->release = ublk_cdev_rel;
1429         device_initialize(dev);
1430
1431         ret = dev_set_name(dev, "ublkc%d", minor);
1432         if (ret)
1433                 goto fail;
1434
1435         cdev_init(&ub->cdev, &ublk_ch_fops);
1436         ret = cdev_device_add(&ub->cdev, dev);
1437         if (ret)
1438                 goto fail;
1439         return 0;
1440  fail:
1441         put_device(dev);
1442         return ret;
1443 }
1444
1445 static void ublk_stop_work_fn(struct work_struct *work)
1446 {
1447         struct ublk_device *ub =
1448                 container_of(work, struct ublk_device, stop_work);
1449
1450         ublk_stop_dev(ub);
1451 }
1452
1453 /* align max io buffer size with PAGE_SIZE */
1454 static void ublk_align_max_io_size(struct ublk_device *ub)
1455 {
1456         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1457
1458         ub->dev_info.max_io_buf_bytes =
1459                 round_down(max_io_bytes, PAGE_SIZE);
1460 }
1461
1462 static int ublk_add_tag_set(struct ublk_device *ub)
1463 {
1464         ub->tag_set.ops = &ublk_mq_ops;
1465         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1466         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1467         ub->tag_set.numa_node = NUMA_NO_NODE;
1468         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1469         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1470         ub->tag_set.driver_data = ub;
1471         return blk_mq_alloc_tag_set(&ub->tag_set);
1472 }
1473
1474 static void ublk_remove(struct ublk_device *ub)
1475 {
1476         ublk_stop_dev(ub);
1477         cancel_work_sync(&ub->stop_work);
1478         cancel_work_sync(&ub->quiesce_work);
1479         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1480         put_device(&ub->cdev_dev);
1481 }
1482
1483 static struct ublk_device *ublk_get_device_from_id(int idx)
1484 {
1485         struct ublk_device *ub = NULL;
1486
1487         if (idx < 0)
1488                 return NULL;
1489
1490         spin_lock(&ublk_idr_lock);
1491         ub = idr_find(&ublk_index_idr, idx);
1492         if (ub)
1493                 ub = ublk_get_device(ub);
1494         spin_unlock(&ublk_idr_lock);
1495
1496         return ub;
1497 }
1498
1499 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1500 {
1501         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1502         int ublksrv_pid = (int)header->data[0];
1503         struct ublk_device *ub;
1504         struct gendisk *disk;
1505         int ret = -EINVAL;
1506
1507         if (ublksrv_pid <= 0)
1508                 return -EINVAL;
1509
1510         ub = ublk_get_device_from_id(header->dev_id);
1511         if (!ub)
1512                 return -EINVAL;
1513
1514         wait_for_completion_interruptible(&ub->completion);
1515
1516         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1517
1518         mutex_lock(&ub->mutex);
1519         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1520             test_bit(UB_STATE_USED, &ub->state)) {
1521                 ret = -EEXIST;
1522                 goto out_unlock;
1523         }
1524
1525         disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1526         if (IS_ERR(disk)) {
1527                 ret = PTR_ERR(disk);
1528                 goto out_unlock;
1529         }
1530         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1531         disk->fops = &ub_fops;
1532         disk->private_data = ub;
1533
1534         ub->dev_info.ublksrv_pid = ublksrv_pid;
1535         ub->ub_disk = disk;
1536
1537         ret = ublk_apply_params(ub);
1538         if (ret)
1539                 goto out_put_disk;
1540
1541         /* don't probe partitions if any one ubq daemon is un-trusted */
1542         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1543                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1544
1545         get_device(&ub->cdev_dev);
1546         ret = add_disk(disk);
1547         if (ret) {
1548                 /*
1549                  * Has to drop the reference since ->free_disk won't be
1550                  * called in case of add_disk failure.
1551                  */
1552                 ublk_put_device(ub);
1553                 goto out_put_disk;
1554         }
1555         set_bit(UB_STATE_USED, &ub->state);
1556         ub->dev_info.state = UBLK_S_DEV_LIVE;
1557 out_put_disk:
1558         if (ret)
1559                 put_disk(disk);
1560 out_unlock:
1561         mutex_unlock(&ub->mutex);
1562         ublk_put_device(ub);
1563         return ret;
1564 }
1565
1566 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1567 {
1568         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1569         void __user *argp = (void __user *)(unsigned long)header->addr;
1570         struct ublk_device *ub;
1571         cpumask_var_t cpumask;
1572         unsigned long queue;
1573         unsigned int retlen;
1574         unsigned int i;
1575         int ret = -EINVAL;
1576
1577         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1578                 return -EINVAL;
1579         if (header->len & (sizeof(unsigned long)-1))
1580                 return -EINVAL;
1581         if (!header->addr)
1582                 return -EINVAL;
1583
1584         ub = ublk_get_device_from_id(header->dev_id);
1585         if (!ub)
1586                 return -EINVAL;
1587
1588         queue = header->data[0];
1589         if (queue >= ub->dev_info.nr_hw_queues)
1590                 goto out_put_device;
1591
1592         ret = -ENOMEM;
1593         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1594                 goto out_put_device;
1595
1596         for_each_possible_cpu(i) {
1597                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1598                         cpumask_set_cpu(i, cpumask);
1599         }
1600
1601         ret = -EFAULT;
1602         retlen = min_t(unsigned short, header->len, cpumask_size());
1603         if (copy_to_user(argp, cpumask, retlen))
1604                 goto out_free_cpumask;
1605         if (retlen != header->len &&
1606             clear_user(argp + retlen, header->len - retlen))
1607                 goto out_free_cpumask;
1608
1609         ret = 0;
1610 out_free_cpumask:
1611         free_cpumask_var(cpumask);
1612 out_put_device:
1613         ublk_put_device(ub);
1614         return ret;
1615 }
1616
1617 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1618 {
1619         pr_devel("%s: dev id %d flags %llx\n", __func__,
1620                         info->dev_id, info->flags);
1621         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1622                         info->nr_hw_queues, info->queue_depth);
1623 }
1624
1625 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1626 {
1627         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1628         void __user *argp = (void __user *)(unsigned long)header->addr;
1629         struct ublksrv_ctrl_dev_info info;
1630         struct ublk_device *ub;
1631         int ret = -EINVAL;
1632
1633         if (header->len < sizeof(info) || !header->addr)
1634                 return -EINVAL;
1635         if (header->queue_id != (u16)-1) {
1636                 pr_warn("%s: queue_id is wrong %x\n",
1637                         __func__, header->queue_id);
1638                 return -EINVAL;
1639         }
1640         if (copy_from_user(&info, argp, sizeof(info)))
1641                 return -EFAULT;
1642         ublk_dump_dev_info(&info);
1643         if (header->dev_id != info.dev_id) {
1644                 pr_warn("%s: dev id not match %u %u\n",
1645                         __func__, header->dev_id, info.dev_id);
1646                 return -EINVAL;
1647         }
1648
1649         ret = mutex_lock_killable(&ublk_ctl_mutex);
1650         if (ret)
1651                 return ret;
1652
1653         ret = -ENOMEM;
1654         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1655         if (!ub)
1656                 goto out_unlock;
1657         mutex_init(&ub->mutex);
1658         spin_lock_init(&ub->mm_lock);
1659         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1660         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1661         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1662
1663         ret = ublk_alloc_dev_number(ub, header->dev_id);
1664         if (ret < 0)
1665                 goto out_free_ub;
1666
1667         memcpy(&ub->dev_info, &info, sizeof(info));
1668
1669         /* update device id */
1670         ub->dev_info.dev_id = ub->ub_number;
1671
1672         /*
1673          * 64bit flags will be copied back to userspace as feature
1674          * negotiation result, so have to clear flags which driver
1675          * doesn't support yet, then userspace can get correct flags
1676          * (features) to handle.
1677          */
1678         ub->dev_info.flags &= UBLK_F_ALL;
1679
1680         if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1681                 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1682
1683         /* We are not ready to support zero copy */
1684         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1685
1686         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1687                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1688         ublk_align_max_io_size(ub);
1689
1690         ret = ublk_init_queues(ub);
1691         if (ret)
1692                 goto out_free_dev_number;
1693
1694         ret = ublk_add_tag_set(ub);
1695         if (ret)
1696                 goto out_deinit_queues;
1697
1698         ret = -EFAULT;
1699         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1700                 goto out_free_tag_set;
1701
1702         /*
1703          * Add the char dev so that ublksrv daemon can be setup.
1704          * ublk_add_chdev() will cleanup everything if it fails.
1705          */
1706         ret = ublk_add_chdev(ub);
1707         goto out_unlock;
1708
1709 out_free_tag_set:
1710         blk_mq_free_tag_set(&ub->tag_set);
1711 out_deinit_queues:
1712         ublk_deinit_queues(ub);
1713 out_free_dev_number:
1714         ublk_free_dev_number(ub);
1715 out_free_ub:
1716         mutex_destroy(&ub->mutex);
1717         kfree(ub);
1718 out_unlock:
1719         mutex_unlock(&ublk_ctl_mutex);
1720         return ret;
1721 }
1722
1723 static inline bool ublk_idr_freed(int id)
1724 {
1725         void *ptr;
1726
1727         spin_lock(&ublk_idr_lock);
1728         ptr = idr_find(&ublk_index_idr, id);
1729         spin_unlock(&ublk_idr_lock);
1730
1731         return ptr == NULL;
1732 }
1733
1734 static int ublk_ctrl_del_dev(int idx)
1735 {
1736         struct ublk_device *ub;
1737         int ret;
1738
1739         ret = mutex_lock_killable(&ublk_ctl_mutex);
1740         if (ret)
1741                 return ret;
1742
1743         ub = ublk_get_device_from_id(idx);
1744         if (ub) {
1745                 ublk_remove(ub);
1746                 ublk_put_device(ub);
1747                 ret = 0;
1748         } else {
1749                 ret = -ENODEV;
1750         }
1751
1752         /*
1753          * Wait until the idr is removed, then it can be reused after
1754          * DEL_DEV command is returned.
1755          */
1756         if (!ret)
1757                 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1758         mutex_unlock(&ublk_ctl_mutex);
1759
1760         return ret;
1761 }
1762
1763 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1764 {
1765         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1766
1767         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1768                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1769                         header->data[0], header->addr, header->len);
1770 }
1771
1772 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1773 {
1774         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1775         struct ublk_device *ub;
1776
1777         ub = ublk_get_device_from_id(header->dev_id);
1778         if (!ub)
1779                 return -EINVAL;
1780
1781         ublk_stop_dev(ub);
1782         cancel_work_sync(&ub->stop_work);
1783         cancel_work_sync(&ub->quiesce_work);
1784
1785         ublk_put_device(ub);
1786         return 0;
1787 }
1788
1789 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1790 {
1791         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1792         void __user *argp = (void __user *)(unsigned long)header->addr;
1793         struct ublk_device *ub;
1794         int ret = 0;
1795
1796         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1797                 return -EINVAL;
1798
1799         ub = ublk_get_device_from_id(header->dev_id);
1800         if (!ub)
1801                 return -EINVAL;
1802
1803         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1804                 ret = -EFAULT;
1805         ublk_put_device(ub);
1806
1807         return ret;
1808 }
1809
1810 static int ublk_ctrl_get_params(struct io_uring_cmd *cmd)
1811 {
1812         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1813         void __user *argp = (void __user *)(unsigned long)header->addr;
1814         struct ublk_params_header ph;
1815         struct ublk_device *ub;
1816         int ret;
1817
1818         if (header->len <= sizeof(ph) || !header->addr)
1819                 return -EINVAL;
1820
1821         if (copy_from_user(&ph, argp, sizeof(ph)))
1822                 return -EFAULT;
1823
1824         if (ph.len > header->len || !ph.len)
1825                 return -EINVAL;
1826
1827         if (ph.len > sizeof(struct ublk_params))
1828                 ph.len = sizeof(struct ublk_params);
1829
1830         ub = ublk_get_device_from_id(header->dev_id);
1831         if (!ub)
1832                 return -EINVAL;
1833
1834         mutex_lock(&ub->mutex);
1835         if (copy_to_user(argp, &ub->params, ph.len))
1836                 ret = -EFAULT;
1837         else
1838                 ret = 0;
1839         mutex_unlock(&ub->mutex);
1840
1841         ublk_put_device(ub);
1842         return ret;
1843 }
1844
1845 static int ublk_ctrl_set_params(struct io_uring_cmd *cmd)
1846 {
1847         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1848         void __user *argp = (void __user *)(unsigned long)header->addr;
1849         struct ublk_params_header ph;
1850         struct ublk_device *ub;
1851         int ret = -EFAULT;
1852
1853         if (header->len <= sizeof(ph) || !header->addr)
1854                 return -EINVAL;
1855
1856         if (copy_from_user(&ph, argp, sizeof(ph)))
1857                 return -EFAULT;
1858
1859         if (ph.len > header->len || !ph.len || !ph.types)
1860                 return -EINVAL;
1861
1862         if (ph.len > sizeof(struct ublk_params))
1863                 ph.len = sizeof(struct ublk_params);
1864
1865         ub = ublk_get_device_from_id(header->dev_id);
1866         if (!ub)
1867                 return -EINVAL;
1868
1869         /* parameters can only be changed when device isn't live */
1870         mutex_lock(&ub->mutex);
1871         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1872                 ret = -EACCES;
1873         } else if (copy_from_user(&ub->params, argp, ph.len)) {
1874                 ret = -EFAULT;
1875         } else {
1876                 /* clear all we don't support yet */
1877                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1878                 ret = ublk_validate_params(ub);
1879         }
1880         mutex_unlock(&ub->mutex);
1881         ublk_put_device(ub);
1882
1883         return ret;
1884 }
1885
1886 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1887 {
1888         int i;
1889
1890         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1891         /* All old ioucmds have to be completed */
1892         WARN_ON_ONCE(ubq->nr_io_ready);
1893         /* old daemon is PF_EXITING, put it now */
1894         put_task_struct(ubq->ubq_daemon);
1895         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1896         ubq->ubq_daemon = NULL;
1897
1898         for (i = 0; i < ubq->q_depth; i++) {
1899                 struct ublk_io *io = &ubq->ios[i];
1900
1901                 /* forget everything now and be ready for new FETCH_REQ */
1902                 io->flags = 0;
1903                 io->cmd = NULL;
1904                 io->addr = 0;
1905         }
1906 }
1907
1908 static int ublk_ctrl_start_recovery(struct io_uring_cmd *cmd)
1909 {
1910         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1911         struct ublk_device *ub;
1912         int ret = -EINVAL;
1913         int i;
1914
1915         ub = ublk_get_device_from_id(header->dev_id);
1916         if (!ub)
1917                 return ret;
1918
1919         mutex_lock(&ub->mutex);
1920         if (!ublk_can_use_recovery(ub))
1921                 goto out_unlock;
1922         /*
1923          * START_RECOVERY is only allowd after:
1924          *
1925          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1926          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
1927          *     released.
1928          *
1929          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1930          *     (a)has quiesced request queue
1931          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
1932          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1933          *     (d)has completed/camceled all ioucmds owned by ther dying process
1934          */
1935         if (test_bit(UB_STATE_OPEN, &ub->state) ||
1936                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1937                 ret = -EBUSY;
1938                 goto out_unlock;
1939         }
1940         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1941         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1942                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1943         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1944         ub->mm = NULL;
1945         ub->nr_queues_ready = 0;
1946         ub->nr_privileged_daemon = 0;
1947         init_completion(&ub->completion);
1948         ret = 0;
1949  out_unlock:
1950         mutex_unlock(&ub->mutex);
1951         ublk_put_device(ub);
1952         return ret;
1953 }
1954
1955 static int ublk_ctrl_end_recovery(struct io_uring_cmd *cmd)
1956 {
1957         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1958         int ublksrv_pid = (int)header->data[0];
1959         struct ublk_device *ub;
1960         int ret = -EINVAL;
1961
1962         ub = ublk_get_device_from_id(header->dev_id);
1963         if (!ub)
1964                 return ret;
1965
1966         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1967                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1968         /* wait until new ubq_daemon sending all FETCH_REQ */
1969         wait_for_completion_interruptible(&ub->completion);
1970         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
1971                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1972
1973         mutex_lock(&ub->mutex);
1974         if (!ublk_can_use_recovery(ub))
1975                 goto out_unlock;
1976
1977         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1978                 ret = -EBUSY;
1979                 goto out_unlock;
1980         }
1981         ub->dev_info.ublksrv_pid = ublksrv_pid;
1982         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
1983                         __func__, ublksrv_pid, header->dev_id);
1984         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1985         pr_devel("%s: queue unquiesced, dev id %d.\n",
1986                         __func__, header->dev_id);
1987         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1988         ub->dev_info.state = UBLK_S_DEV_LIVE;
1989         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1990         ret = 0;
1991  out_unlock:
1992         mutex_unlock(&ub->mutex);
1993         ublk_put_device(ub);
1994         return ret;
1995 }
1996
1997 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1998                 unsigned int issue_flags)
1999 {
2000         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2001         int ret = -EINVAL;
2002
2003         if (issue_flags & IO_URING_F_NONBLOCK)
2004                 return -EAGAIN;
2005
2006         ublk_ctrl_cmd_dump(cmd);
2007
2008         if (!(issue_flags & IO_URING_F_SQE128))
2009                 goto out;
2010
2011         ret = -EPERM;
2012         if (!capable(CAP_SYS_ADMIN))
2013                 goto out;
2014
2015         ret = -ENODEV;
2016         switch (cmd->cmd_op) {
2017         case UBLK_CMD_START_DEV:
2018                 ret = ublk_ctrl_start_dev(cmd);
2019                 break;
2020         case UBLK_CMD_STOP_DEV:
2021                 ret = ublk_ctrl_stop_dev(cmd);
2022                 break;
2023         case UBLK_CMD_GET_DEV_INFO:
2024                 ret = ublk_ctrl_get_dev_info(cmd);
2025                 break;
2026         case UBLK_CMD_ADD_DEV:
2027                 ret = ublk_ctrl_add_dev(cmd);
2028                 break;
2029         case UBLK_CMD_DEL_DEV:
2030                 ret = ublk_ctrl_del_dev(header->dev_id);
2031                 break;
2032         case UBLK_CMD_GET_QUEUE_AFFINITY:
2033                 ret = ublk_ctrl_get_queue_affinity(cmd);
2034                 break;
2035         case UBLK_CMD_GET_PARAMS:
2036                 ret = ublk_ctrl_get_params(cmd);
2037                 break;
2038         case UBLK_CMD_SET_PARAMS:
2039                 ret = ublk_ctrl_set_params(cmd);
2040                 break;
2041         case UBLK_CMD_START_USER_RECOVERY:
2042                 ret = ublk_ctrl_start_recovery(cmd);
2043                 break;
2044         case UBLK_CMD_END_USER_RECOVERY:
2045                 ret = ublk_ctrl_end_recovery(cmd);
2046                 break;
2047         default:
2048                 break;
2049         }
2050  out:
2051         io_uring_cmd_done(cmd, ret, 0);
2052         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2053                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2054         return -EIOCBQUEUED;
2055 }
2056
2057 static const struct file_operations ublk_ctl_fops = {
2058         .open           = nonseekable_open,
2059         .uring_cmd      = ublk_ctrl_uring_cmd,
2060         .owner          = THIS_MODULE,
2061         .llseek         = noop_llseek,
2062 };
2063
2064 static struct miscdevice ublk_misc = {
2065         .minor          = MISC_DYNAMIC_MINOR,
2066         .name           = "ublk-control",
2067         .fops           = &ublk_ctl_fops,
2068 };
2069
2070 static int __init ublk_init(void)
2071 {
2072         int ret;
2073
2074         init_waitqueue_head(&ublk_idr_wq);
2075
2076         ret = misc_register(&ublk_misc);
2077         if (ret)
2078                 return ret;
2079
2080         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2081         if (ret)
2082                 goto unregister_mis;
2083
2084         ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2085         if (IS_ERR(ublk_chr_class)) {
2086                 ret = PTR_ERR(ublk_chr_class);
2087                 goto free_chrdev_region;
2088         }
2089         return 0;
2090
2091 free_chrdev_region:
2092         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2093 unregister_mis:
2094         misc_deregister(&ublk_misc);
2095         return ret;
2096 }
2097
2098 static void __exit ublk_exit(void)
2099 {
2100         struct ublk_device *ub;
2101         int id;
2102
2103         idr_for_each_entry(&ublk_index_idr, ub, id)
2104                 ublk_remove(ub);
2105
2106         class_destroy(ublk_chr_class);
2107         misc_deregister(&ublk_misc);
2108
2109         idr_destroy(&ublk_index_idr);
2110         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2111 }
2112
2113 module_init(ublk_init);
2114 module_exit(ublk_exit);
2115
2116 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2117 MODULE_LICENSE("GPL");