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 int    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 || p->logical_bs_shift < 9)
 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                                 unsigned issue_flags)
 661 {
 662         /* mark this cmd owned by ublksrv */
 663         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 664
 665         /*
 666          * clear ACTIVE since we are done with this sqe/cmd slot
 667          * We can only accept io cmd in case of being not active.
 668          */
 669         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 670
 671         /* tell ublksrv one io request is coming */
 672         io_uring_cmd_done(io->cmd, res, 0, issue_flags);
 673 }
 674
 675 #define UBLK_REQUEUE_DELAY_MS   3
 676
 677 static inline void __ublk_abort_rq(struct ublk_queue *ubq,
 678                 struct request *rq)
 679 {
 680         /* We cannot process this rq so just requeue it. */
 681         if (ublk_queue_can_use_recovery(ubq))
 682                 blk_mq_requeue_request(rq, false);
 683         else
 684                 blk_mq_end_request(rq, BLK_STS_IOERR);
 685
 686         mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 687 }
 688
 689 static inline void __ublk_rq_task_work(struct request *req,
 690                                        unsigned issue_flags)
 691 {
 692         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 693         int tag = req->tag;
 694         struct ublk_io *io = &ubq->ios[tag];
 695         unsigned int mapped_bytes;
 696
 697         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 698                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 699                         ublk_get_iod(ubq, req->tag)->addr);
 700
 701         /*
 702          * Task is exiting if either:
 703          *
 704          * (1) current != ubq_daemon.
 705          * io_uring_cmd_complete_in_task() tries to run task_work
 706          * in a workqueue if ubq_daemon(cmd's task) is PF_EXITING.
 707          *
 708          * (2) current->flags & PF_EXITING.
 709          */
 710         if (unlikely(current != ubq->ubq_daemon || current->flags & PF_EXITING)) {
 711                 __ublk_abort_rq(ubq, req);
 712                 return;
 713         }
 714
 715         if (ublk_need_get_data(ubq) &&
 716                         (req_op(req) == REQ_OP_WRITE ||
 717                         req_op(req) == REQ_OP_FLUSH)) {
 718                 /*
 719                  * We have not handled UBLK_IO_NEED_GET_DATA command yet,
 720                  * so immepdately pass UBLK_IO_RES_NEED_GET_DATA to ublksrv
 721                  * and notify it.
 722                  */
 723                 if (!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA)) {
 724                         io->flags |= UBLK_IO_FLAG_NEED_GET_DATA;
 725                         pr_devel("%s: need get data. op %d, qid %d tag %d io_flags %x\n",
 726                                         __func__, io->cmd->cmd_op, ubq->q_id,
 727                                         req->tag, io->flags);
 728                         ubq_complete_io_cmd(io, UBLK_IO_RES_NEED_GET_DATA, issue_flags);
 729                         return;
 730                 }
 731                 /*
 732                  * We have handled UBLK_IO_NEED_GET_DATA command,
 733                  * so clear UBLK_IO_FLAG_NEED_GET_DATA now and just
 734                  * do the copy work.
 735                  */
 736                 io->flags &= ~UBLK_IO_FLAG_NEED_GET_DATA;
 737                 /* update iod->addr because ublksrv may have passed a new io buffer */
 738                 ublk_get_iod(ubq, req->tag)->addr = io->addr;
 739                 pr_devel("%s: update iod->addr: op %d, qid %d tag %d io_flags %x addr %llx\n",
 740                                 __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 741                                 ublk_get_iod(ubq, req->tag)->addr);
 742         }
 743
 744         mapped_bytes = ublk_map_io(ubq, req, io);
 745
 746         /* partially mapped, update io descriptor */
 747         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 748                 /*
 749                  * Nothing mapped, retry until we succeed.
 750                  *
 751                  * We may never succeed in mapping any bytes here because
 752                  * of OOM. TODO: reserve one buffer with single page pinned
 753                  * for providing forward progress guarantee.
 754                  */
 755                 if (unlikely(!mapped_bytes)) {
 756                         blk_mq_requeue_request(req, false);
 757                         blk_mq_delay_kick_requeue_list(req->q,
 758                                         UBLK_REQUEUE_DELAY_MS);
 759                         return;
 760                 }
 761
 762                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 763                         mapped_bytes >> 9;
 764         }
 765
 766         ubq_complete_io_cmd(io, UBLK_IO_RES_OK, issue_flags);
 767 }
 768
 769 static inline void ublk_forward_io_cmds(struct ublk_queue *ubq,
 770                                         unsigned issue_flags)
 771 {
 772         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 773         struct ublk_rq_data *data, *tmp;
 774
 775         io_cmds = llist_reverse_order(io_cmds);
 776         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 777                 __ublk_rq_task_work(blk_mq_rq_from_pdu(data), issue_flags);
 778 }
 779
 780 static inline void ublk_abort_io_cmds(struct ublk_queue *ubq)
 781 {
 782         struct llist_node *io_cmds = llist_del_all(&ubq->io_cmds);
 783         struct ublk_rq_data *data, *tmp;
 784
 785         llist_for_each_entry_safe(data, tmp, io_cmds, node)
 786                 __ublk_abort_rq(ubq, blk_mq_rq_from_pdu(data));
 787 }
 788
 789 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd, unsigned issue_flags)
 790 {
 791         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 792         struct ublk_queue *ubq = pdu->ubq;
 793
 794         ublk_forward_io_cmds(ubq, issue_flags);
 795 }
 796
 797 static void ublk_rq_task_work_fn(struct callback_head *work)
 798 {
 799         struct ublk_rq_data *data = container_of(work,
 800                         struct ublk_rq_data, work);
 801         struct request *req = blk_mq_rq_from_pdu(data);
 802         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 803         unsigned issue_flags = IO_URING_F_UNLOCKED;
 804
 805         ublk_forward_io_cmds(ubq, issue_flags);
 806 }
 807
 808 static void ublk_queue_cmd(struct ublk_queue *ubq, struct request *rq)
 809 {
 810         struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 811         struct ublk_io *io;
 812
 813         if (!llist_add(&data->node, &ubq->io_cmds))
 814                 return;
 815
 816         io = &ubq->ios[rq->tag];
 817         /*
 818          * If the check pass, we know that this is a re-issued request aborted
 819          * previously in monitor_work because the ubq_daemon(cmd's task) is
 820          * PF_EXITING. We cannot call io_uring_cmd_complete_in_task() anymore
 821          * because this ioucmd's io_uring context may be freed now if no inflight
 822          * ioucmd exists. Otherwise we may cause null-deref in ctx->fallback_work.
 823          *
 824          * Note: monitor_work sets UBLK_IO_FLAG_ABORTED and ends this request(releasing
 825          * the tag). Then the request is re-started(allocating the tag) and we are here.
 826          * Since releasing/allocating a tag implies smp_mb(), finding UBLK_IO_FLAG_ABORTED
 827          * guarantees that here is a re-issued request aborted previously.
 828          */
 829         if (unlikely(io->flags & UBLK_IO_FLAG_ABORTED)) {
 830                 ublk_abort_io_cmds(ubq);
 831         } else if (ublk_can_use_task_work(ubq)) {
 832                 if (task_work_add(ubq->ubq_daemon, &data->work,
 833                                         TWA_SIGNAL_NO_IPI))
 834                         ublk_abort_io_cmds(ubq);
 835         } else {
 836                 struct io_uring_cmd *cmd = io->cmd;
 837                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 838
 839                 pdu->ubq = ubq;
 840                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 841         }
 842 }
 843
 844 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 845                 const struct blk_mq_queue_data *bd)
 846 {
 847         struct ublk_queue *ubq = hctx->driver_data;
 848         struct request *rq = bd->rq;
 849         blk_status_t res;
 850
 851         /* fill iod to slot in io cmd buffer */
 852         res = ublk_setup_iod(ubq, rq);
 853         if (unlikely(res != BLK_STS_OK))
 854                 return BLK_STS_IOERR;
 855
 856         /* With recovery feature enabled, force_abort is set in
 857          * ublk_stop_dev() before calling del_gendisk(). We have to
 858          * abort all requeued and new rqs here to let del_gendisk()
 859          * move on. Besides, we cannot not call io_uring_cmd_complete_in_task()
 860          * to avoid UAF on io_uring ctx.
 861          *
 862          * Note: force_abort is guaranteed to be seen because it is set
 863          * before request queue is unqiuesced.
 864          */
 865         if (ublk_queue_can_use_recovery(ubq) && unlikely(ubq->force_abort))
 866                 return BLK_STS_IOERR;
 867
 868         blk_mq_start_request(bd->rq);
 869
 870         if (unlikely(ubq_daemon_is_dying(ubq))) {
 871                 __ublk_abort_rq(ubq, rq);
 872                 return BLK_STS_OK;
 873         }
 874
 875         ublk_queue_cmd(ubq, rq);
 876
 877         return BLK_STS_OK;
 878 }
 879
 880 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 881                 unsigned int hctx_idx)
 882 {
 883         struct ublk_device *ub = driver_data;
 884         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 885
 886         hctx->driver_data = ubq;
 887         return 0;
 888 }
 889
 890 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 891                 unsigned int hctx_idx, unsigned int numa_node)
 892 {
 893         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 894
 895         init_task_work(&data->work, ublk_rq_task_work_fn);
 896         return 0;
 897 }
 898
 899 static const struct blk_mq_ops ublk_mq_ops = {
 900         .queue_rq       = ublk_queue_rq,
 901         .init_hctx      = ublk_init_hctx,
 902         .init_request   = ublk_init_rq,
 903 };
 904
 905 static int ublk_ch_open(struct inode *inode, struct file *filp)
 906 {
 907         struct ublk_device *ub = container_of(inode->i_cdev,
 908                         struct ublk_device, cdev);
 909
 910         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 911                 return -EBUSY;
 912         filp->private_data = ub;
 913         return 0;
 914 }
 915
 916 static int ublk_ch_release(struct inode *inode, struct file *filp)
 917 {
 918         struct ublk_device *ub = filp->private_data;
 919
 920         clear_bit(UB_STATE_OPEN, &ub->state);
 921         return 0;
 922 }
 923
 924 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
 925 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
 926 {
 927         struct ublk_device *ub = filp->private_data;
 928         size_t sz = vma->vm_end - vma->vm_start;
 929         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
 930         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
 931         int q_id, ret = 0;
 932
 933         spin_lock(&ub->mm_lock);
 934         if (!ub->mm)
 935                 ub->mm = current->mm;
 936         if (current->mm != ub->mm)
 937                 ret = -EINVAL;
 938         spin_unlock(&ub->mm_lock);
 939
 940         if (ret)
 941                 return ret;
 942
 943         if (vma->vm_flags & VM_WRITE)
 944                 return -EPERM;
 945
 946         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
 947         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
 948                 return -EINVAL;
 949
 950         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
 951         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
 952                         __func__, q_id, current->pid, vma->vm_start,
 953                         phys_off, (unsigned long)sz);
 954
 955         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
 956                 return -EINVAL;
 957
 958         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
 959         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
 960 }
 961
 962 static void ublk_commit_completion(struct ublk_device *ub,
 963                 struct ublksrv_io_cmd *ub_cmd)
 964 {
 965         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
 966         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
 967         struct ublk_io *io = &ubq->ios[tag];
 968         struct request *req;
 969
 970         /* now this cmd slot is owned by nbd driver */
 971         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
 972         io->res = ub_cmd->result;
 973
 974         /* find the io request and complete */
 975         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
 976
 977         if (req && likely(!blk_should_fake_timeout(req->q)))
 978                 ublk_complete_rq(req);
 979 }
 980
 981 /*
 982  * When ->ubq_daemon is exiting, either new request is ended immediately,
 983  * or any queued io command is drained, so it is safe to abort queue
 984  * lockless
 985  */
 986 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
 987 {
 988         int i;
 989
 990         if (!ublk_get_device(ub))
 991                 return;
 992
 993         for (i = 0; i < ubq->q_depth; i++) {
 994                 struct ublk_io *io = &ubq->ios[i];
 995
 996                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
 997                         struct request *rq;
 998
 999                         /*
1000                          * Either we fail the request or ublk_rq_task_work_fn
1001                          * will do it
1002                          */
1003                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
1004                         if (rq)
1005                                 __ublk_fail_req(ubq, io, rq);
1006                 }
1007         }
1008         ublk_put_device(ub);
1009 }
1010
1011 static void ublk_daemon_monitor_work(struct work_struct *work)
1012 {
1013         struct ublk_device *ub =
1014                 container_of(work, struct ublk_device, monitor_work.work);
1015         int i;
1016
1017         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
1018                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
1019
1020                 if (ubq_daemon_is_dying(ubq)) {
1021                         if (ublk_queue_can_use_recovery(ubq))
1022                                 schedule_work(&ub->quiesce_work);
1023                         else
1024                                 schedule_work(&ub->stop_work);
1025
1026                         /* abort queue is for making forward progress */
1027                         ublk_abort_queue(ub, ubq);
1028                 }
1029         }
1030
1031         /*
1032          * We can't schedule monitor work after ub's state is not UBLK_S_DEV_LIVE.
1033          * after ublk_remove() or __ublk_quiesce_dev() is started.
1034          *
1035          * No need ub->mutex, monitor work are canceled after state is marked
1036          * as not LIVE, so new state is observed reliably.
1037          */
1038         if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1039                 schedule_delayed_work(&ub->monitor_work,
1040                                 UBLK_DAEMON_MONITOR_PERIOD);
1041 }
1042
1043 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
1044 {
1045         return ubq->nr_io_ready == ubq->q_depth;
1046 }
1047
1048 static void ublk_cancel_queue(struct ublk_queue *ubq)
1049 {
1050         int i;
1051
1052         if (!ublk_queue_ready(ubq))
1053                 return;
1054
1055         for (i = 0; i < ubq->q_depth; i++) {
1056                 struct ublk_io *io = &ubq->ios[i];
1057
1058                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
1059                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0,
1060                                                 IO_URING_F_UNLOCKED);
1061         }
1062
1063         /* all io commands are canceled */
1064         ubq->nr_io_ready = 0;
1065 }
1066
1067 /* Cancel all pending commands, must be called after del_gendisk() returns */
1068 static void ublk_cancel_dev(struct ublk_device *ub)
1069 {
1070         int i;
1071
1072         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1073                 ublk_cancel_queue(ublk_get_queue(ub, i));
1074 }
1075
1076 static bool ublk_check_inflight_rq(struct request *rq, void *data)
1077 {
1078         bool *idle = data;
1079
1080         if (blk_mq_request_started(rq)) {
1081                 *idle = false;
1082                 return false;
1083         }
1084         return true;
1085 }
1086
1087 static void ublk_wait_tagset_rqs_idle(struct ublk_device *ub)
1088 {
1089         bool idle;
1090
1091         WARN_ON_ONCE(!blk_queue_quiesced(ub->ub_disk->queue));
1092         while (true) {
1093                 idle = true;
1094                 blk_mq_tagset_busy_iter(&ub->tag_set,
1095                                 ublk_check_inflight_rq, &idle);
1096                 if (idle)
1097                         break;
1098                 msleep(UBLK_REQUEUE_DELAY_MS);
1099         }
1100 }
1101
1102 static void __ublk_quiesce_dev(struct ublk_device *ub)
1103 {
1104         pr_devel("%s: quiesce ub: dev_id %d state %s\n",
1105                         __func__, ub->dev_info.dev_id,
1106                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1107                         "LIVE" : "QUIESCED");
1108         blk_mq_quiesce_queue(ub->ub_disk->queue);
1109         ublk_wait_tagset_rqs_idle(ub);
1110         ub->dev_info.state = UBLK_S_DEV_QUIESCED;
1111         ublk_cancel_dev(ub);
1112         /* we are going to release task_struct of ubq_daemon and resets
1113          * ->ubq_daemon to NULL. So in monitor_work, check on ubq_daemon causes UAF.
1114          * Besides, monitor_work is not necessary in QUIESCED state since we have
1115          * already scheduled quiesce_work and quiesced all ubqs.
1116          *
1117          * Do not let monitor_work schedule itself if state it QUIESCED. And we cancel
1118          * it here and re-schedule it in END_USER_RECOVERY to avoid UAF.
1119          */
1120         cancel_delayed_work_sync(&ub->monitor_work);
1121 }
1122
1123 static void ublk_quiesce_work_fn(struct work_struct *work)
1124 {
1125         struct ublk_device *ub =
1126                 container_of(work, struct ublk_device, quiesce_work);
1127
1128         mutex_lock(&ub->mutex);
1129         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
1130                 goto unlock;
1131         __ublk_quiesce_dev(ub);
1132  unlock:
1133         mutex_unlock(&ub->mutex);
1134 }
1135
1136 static void ublk_unquiesce_dev(struct ublk_device *ub)
1137 {
1138         int i;
1139
1140         pr_devel("%s: unquiesce ub: dev_id %d state %s\n",
1141                         __func__, ub->dev_info.dev_id,
1142                         ub->dev_info.state == UBLK_S_DEV_LIVE ?
1143                         "LIVE" : "QUIESCED");
1144         /* quiesce_work has run. We let requeued rqs be aborted
1145          * before running fallback_wq. "force_abort" must be seen
1146          * after request queue is unqiuesced. Then del_gendisk()
1147          * can move on.
1148          */
1149         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1150                 ublk_get_queue(ub, i)->force_abort = true;
1151
1152         blk_mq_unquiesce_queue(ub->ub_disk->queue);
1153         /* We may have requeued some rqs in ublk_quiesce_queue() */
1154         blk_mq_kick_requeue_list(ub->ub_disk->queue);
1155 }
1156
1157 static void ublk_stop_dev(struct ublk_device *ub)
1158 {
1159         mutex_lock(&ub->mutex);
1160         if (ub->dev_info.state == UBLK_S_DEV_DEAD)
1161                 goto unlock;
1162         if (ublk_can_use_recovery(ub)) {
1163                 if (ub->dev_info.state == UBLK_S_DEV_LIVE)
1164                         __ublk_quiesce_dev(ub);
1165                 ublk_unquiesce_dev(ub);
1166         }
1167         del_gendisk(ub->ub_disk);
1168         ub->dev_info.state = UBLK_S_DEV_DEAD;
1169         ub->dev_info.ublksrv_pid = -1;
1170         put_disk(ub->ub_disk);
1171         ub->ub_disk = NULL;
1172  unlock:
1173         ublk_cancel_dev(ub);
1174         mutex_unlock(&ub->mutex);
1175         cancel_delayed_work_sync(&ub->monitor_work);
1176 }
1177
1178 /* device can only be started after all IOs are ready */
1179 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
1180 {
1181         mutex_lock(&ub->mutex);
1182         ubq->nr_io_ready++;
1183         if (ublk_queue_ready(ubq)) {
1184                 ubq->ubq_daemon = current;
1185                 get_task_struct(ubq->ubq_daemon);
1186                 ub->nr_queues_ready++;
1187
1188                 if (capable(CAP_SYS_ADMIN))
1189                         ub->nr_privileged_daemon++;
1190         }
1191         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
1192                 complete_all(&ub->completion);
1193         mutex_unlock(&ub->mutex);
1194 }
1195
1196 static void ublk_handle_need_get_data(struct ublk_device *ub, int q_id,
1197                 int tag)
1198 {
1199         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1200         struct request *req = blk_mq_tag_to_rq(ub->tag_set.tags[q_id], tag);
1201
1202         ublk_queue_cmd(ubq, req);
1203 }
1204
1205 static int __ublk_ch_uring_cmd(struct io_uring_cmd *cmd,
1206                                unsigned int issue_flags,
1207                                struct ublksrv_io_cmd *ub_cmd)
1208 {
1209         struct ublk_device *ub = cmd->file->private_data;
1210         struct ublk_queue *ubq;
1211         struct ublk_io *io;
1212         u32 cmd_op = cmd->cmd_op;
1213         unsigned tag = ub_cmd->tag;
1214         int ret = -EINVAL;
1215         struct request *req;
1216
1217         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
1218                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
1219                         ub_cmd->result);
1220
1221         if (!(issue_flags & IO_URING_F_SQE128))
1222                 goto out;
1223
1224         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
1225                 goto out;
1226
1227         ubq = ublk_get_queue(ub, ub_cmd->q_id);
1228         if (!ubq || ub_cmd->q_id != ubq->q_id)
1229                 goto out;
1230
1231         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
1232                 goto out;
1233
1234         if (tag >= ubq->q_depth)
1235                 goto out;
1236
1237         io = &ubq->ios[tag];
1238
1239         /* there is pending io cmd, something must be wrong */
1240         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
1241                 ret = -EBUSY;
1242                 goto out;
1243         }
1244
1245         /*
1246          * ensure that the user issues UBLK_IO_NEED_GET_DATA
1247          * iff the driver have set the UBLK_IO_FLAG_NEED_GET_DATA.
1248          */
1249         if ((!!(io->flags & UBLK_IO_FLAG_NEED_GET_DATA))
1250                         ^ (cmd_op == UBLK_IO_NEED_GET_DATA))
1251                 goto out;
1252
1253         switch (cmd_op) {
1254         case UBLK_IO_FETCH_REQ:
1255                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
1256                 if (ublk_queue_ready(ubq)) {
1257                         ret = -EBUSY;
1258                         goto out;
1259                 }
1260                 /*
1261                  * The io is being handled by server, so COMMIT_RQ is expected
1262                  * instead of FETCH_REQ
1263                  */
1264                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
1265                         goto out;
1266                 /* FETCH_RQ has to provide IO buffer if NEED GET DATA is not enabled */
1267                 if (!ub_cmd->addr && !ublk_need_get_data(ubq))
1268                         goto out;
1269                 io->cmd = cmd;
1270                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1271                 io->addr = ub_cmd->addr;
1272
1273                 ublk_mark_io_ready(ub, ubq);
1274                 break;
1275         case UBLK_IO_COMMIT_AND_FETCH_REQ:
1276                 req = blk_mq_tag_to_rq(ub->tag_set.tags[ub_cmd->q_id], tag);
1277                 /*
1278                  * COMMIT_AND_FETCH_REQ has to provide IO buffer if NEED GET DATA is
1279                  * not enabled or it is Read IO.
1280                  */
1281                 if (!ub_cmd->addr && (!ublk_need_get_data(ubq) || req_op(req) == REQ_OP_READ))
1282                         goto out;
1283                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1284                         goto out;
1285                 io->addr = ub_cmd->addr;
1286                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1287                 io->cmd = cmd;
1288                 ublk_commit_completion(ub, ub_cmd);
1289                 break;
1290         case UBLK_IO_NEED_GET_DATA:
1291                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
1292                         goto out;
1293                 io->addr = ub_cmd->addr;
1294                 io->cmd = cmd;
1295                 io->flags |= UBLK_IO_FLAG_ACTIVE;
1296                 ublk_handle_need_get_data(ub, ub_cmd->q_id, ub_cmd->tag);
1297                 break;
1298         default:
1299                 goto out;
1300         }
1301         return -EIOCBQUEUED;
1302
1303  out:
1304         io_uring_cmd_done(cmd, ret, 0, issue_flags);
1305         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
1306                         __func__, cmd_op, tag, ret, io->flags);
1307         return -EIOCBQUEUED;
1308 }
1309
1310 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
1311 {
1312         struct ublksrv_io_cmd *ub_src = (struct ublksrv_io_cmd *) cmd->cmd;
1313         struct ublksrv_io_cmd ub_cmd;
1314
1315         /*
1316          * Not necessary for async retry, but let's keep it simple and always
1317          * copy the values to avoid any potential reuse.
1318          */
1319         ub_cmd.q_id = READ_ONCE(ub_src->q_id);
1320         ub_cmd.tag = READ_ONCE(ub_src->tag);
1321         ub_cmd.result = READ_ONCE(ub_src->result);
1322         ub_cmd.addr = READ_ONCE(ub_src->addr);
1323
1324         return __ublk_ch_uring_cmd(cmd, issue_flags, &ub_cmd);
1325 }
1326
1327 static const struct file_operations ublk_ch_fops = {
1328         .owner = THIS_MODULE,
1329         .open = ublk_ch_open,
1330         .release = ublk_ch_release,
1331         .llseek = no_llseek,
1332         .uring_cmd = ublk_ch_uring_cmd,
1333         .mmap = ublk_ch_mmap,
1334 };
1335
1336 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
1337 {
1338         int size = ublk_queue_cmd_buf_size(ub, q_id);
1339         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1340
1341         if (ubq->ubq_daemon)
1342                 put_task_struct(ubq->ubq_daemon);
1343         if (ubq->io_cmd_buf)
1344                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
1345 }
1346
1347 static int ublk_init_queue(struct ublk_device *ub, int q_id)
1348 {
1349         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
1350         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
1351         void *ptr;
1352         int size;
1353
1354         ubq->flags = ub->dev_info.flags;
1355         ubq->q_id = q_id;
1356         ubq->q_depth = ub->dev_info.queue_depth;
1357         size = ublk_queue_cmd_buf_size(ub, q_id);
1358
1359         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
1360         if (!ptr)
1361                 return -ENOMEM;
1362
1363         ubq->io_cmd_buf = ptr;
1364         ubq->dev = ub;
1365         return 0;
1366 }
1367
1368 static void ublk_deinit_queues(struct ublk_device *ub)
1369 {
1370         int nr_queues = ub->dev_info.nr_hw_queues;
1371         int i;
1372
1373         if (!ub->__queues)
1374                 return;
1375
1376         for (i = 0; i < nr_queues; i++)
1377                 ublk_deinit_queue(ub, i);
1378         kfree(ub->__queues);
1379 }
1380
1381 static int ublk_init_queues(struct ublk_device *ub)
1382 {
1383         int nr_queues = ub->dev_info.nr_hw_queues;
1384         int depth = ub->dev_info.queue_depth;
1385         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
1386         int i, ret = -ENOMEM;
1387
1388         ub->queue_size = ubq_size;
1389         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
1390         if (!ub->__queues)
1391                 return ret;
1392
1393         for (i = 0; i < nr_queues; i++) {
1394                 if (ublk_init_queue(ub, i))
1395                         goto fail;
1396         }
1397
1398         init_completion(&ub->completion);
1399         return 0;
1400
1401  fail:
1402         ublk_deinit_queues(ub);
1403         return ret;
1404 }
1405
1406 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1407 {
1408         int i = idx;
1409         int err;
1410
1411         spin_lock(&ublk_idr_lock);
1412         /* allocate id, if @id >= 0, we're requesting that specific id */
1413         if (i >= 0) {
1414                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1415                 if (err == -ENOSPC)
1416                         err = -EEXIST;
1417         } else {
1418                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1419         }
1420         spin_unlock(&ublk_idr_lock);
1421
1422         if (err >= 0)
1423                 ub->ub_number = err;
1424
1425         return err;
1426 }
1427
1428 static void ublk_free_dev_number(struct ublk_device *ub)
1429 {
1430         spin_lock(&ublk_idr_lock);
1431         idr_remove(&ublk_index_idr, ub->ub_number);
1432         wake_up_all(&ublk_idr_wq);
1433         spin_unlock(&ublk_idr_lock);
1434 }
1435
1436 static void ublk_cdev_rel(struct device *dev)
1437 {
1438         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1439
1440         blk_mq_free_tag_set(&ub->tag_set);
1441         ublk_deinit_queues(ub);
1442         ublk_free_dev_number(ub);
1443         mutex_destroy(&ub->mutex);
1444         kfree(ub);
1445 }
1446
1447 static int ublk_add_chdev(struct ublk_device *ub)
1448 {
1449         struct device *dev = &ub->cdev_dev;
1450         int minor = ub->ub_number;
1451         int ret;
1452
1453         dev->parent = ublk_misc.this_device;
1454         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1455         dev->class = ublk_chr_class;
1456         dev->release = ublk_cdev_rel;
1457         device_initialize(dev);
1458
1459         ret = dev_set_name(dev, "ublkc%d", minor);
1460         if (ret)
1461                 goto fail;
1462
1463         cdev_init(&ub->cdev, &ublk_ch_fops);
1464         ret = cdev_device_add(&ub->cdev, dev);
1465         if (ret)
1466                 goto fail;
1467         return 0;
1468  fail:
1469         put_device(dev);
1470         return ret;
1471 }
1472
1473 static void ublk_stop_work_fn(struct work_struct *work)
1474 {
1475         struct ublk_device *ub =
1476                 container_of(work, struct ublk_device, stop_work);
1477
1478         ublk_stop_dev(ub);
1479 }
1480
1481 /* align max io buffer size with PAGE_SIZE */
1482 static void ublk_align_max_io_size(struct ublk_device *ub)
1483 {
1484         unsigned int max_io_bytes = ub->dev_info.max_io_buf_bytes;
1485
1486         ub->dev_info.max_io_buf_bytes =
1487                 round_down(max_io_bytes, PAGE_SIZE);
1488 }
1489
1490 static int ublk_add_tag_set(struct ublk_device *ub)
1491 {
1492         ub->tag_set.ops = &ublk_mq_ops;
1493         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1494         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1495         ub->tag_set.numa_node = NUMA_NO_NODE;
1496         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1497         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1498         ub->tag_set.driver_data = ub;
1499         return blk_mq_alloc_tag_set(&ub->tag_set);
1500 }
1501
1502 static void ublk_remove(struct ublk_device *ub)
1503 {
1504         ublk_stop_dev(ub);
1505         cancel_work_sync(&ub->stop_work);
1506         cancel_work_sync(&ub->quiesce_work);
1507         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1508         put_device(&ub->cdev_dev);
1509 }
1510
1511 static struct ublk_device *ublk_get_device_from_id(int idx)
1512 {
1513         struct ublk_device *ub = NULL;
1514
1515         if (idx < 0)
1516                 return NULL;
1517
1518         spin_lock(&ublk_idr_lock);
1519         ub = idr_find(&ublk_index_idr, idx);
1520         if (ub)
1521                 ub = ublk_get_device(ub);
1522         spin_unlock(&ublk_idr_lock);
1523
1524         return ub;
1525 }
1526
1527 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1528 {
1529         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1530         int ublksrv_pid = (int)header->data[0];
1531         struct ublk_device *ub;
1532         struct gendisk *disk;
1533         int ret = -EINVAL;
1534
1535         if (ublksrv_pid <= 0)
1536                 return -EINVAL;
1537
1538         ub = ublk_get_device_from_id(header->dev_id);
1539         if (!ub)
1540                 return -EINVAL;
1541
1542         wait_for_completion_interruptible(&ub->completion);
1543
1544         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1545
1546         mutex_lock(&ub->mutex);
1547         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1548             test_bit(UB_STATE_USED, &ub->state)) {
1549                 ret = -EEXIST;
1550                 goto out_unlock;
1551         }
1552
1553         disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1554         if (IS_ERR(disk)) {
1555                 ret = PTR_ERR(disk);
1556                 goto out_unlock;
1557         }
1558         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1559         disk->fops = &ub_fops;
1560         disk->private_data = ub;
1561
1562         ub->dev_info.ublksrv_pid = ublksrv_pid;
1563         ub->ub_disk = disk;
1564
1565         ret = ublk_apply_params(ub);
1566         if (ret)
1567                 goto out_put_disk;
1568
1569         /* don't probe partitions if any one ubq daemon is un-trusted */
1570         if (ub->nr_privileged_daemon != ub->nr_queues_ready)
1571                 set_bit(GD_SUPPRESS_PART_SCAN, &disk->state);
1572
1573         get_device(&ub->cdev_dev);
1574         ret = add_disk(disk);
1575         if (ret) {
1576                 /*
1577                  * Has to drop the reference since ->free_disk won't be
1578                  * called in case of add_disk failure.
1579                  */
1580                 ublk_put_device(ub);
1581                 goto out_put_disk;
1582         }
1583         set_bit(UB_STATE_USED, &ub->state);
1584         ub->dev_info.state = UBLK_S_DEV_LIVE;
1585 out_put_disk:
1586         if (ret)
1587                 put_disk(disk);
1588 out_unlock:
1589         mutex_unlock(&ub->mutex);
1590         ublk_put_device(ub);
1591         return ret;
1592 }
1593
1594 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1595 {
1596         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1597         void __user *argp = (void __user *)(unsigned long)header->addr;
1598         struct ublk_device *ub;
1599         cpumask_var_t cpumask;
1600         unsigned long queue;
1601         unsigned int retlen;
1602         unsigned int i;
1603         int ret = -EINVAL;
1604
1605         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1606                 return -EINVAL;
1607         if (header->len & (sizeof(unsigned long)-1))
1608                 return -EINVAL;
1609         if (!header->addr)
1610                 return -EINVAL;
1611
1612         ub = ublk_get_device_from_id(header->dev_id);
1613         if (!ub)
1614                 return -EINVAL;
1615
1616         queue = header->data[0];
1617         if (queue >= ub->dev_info.nr_hw_queues)
1618                 goto out_put_device;
1619
1620         ret = -ENOMEM;
1621         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1622                 goto out_put_device;
1623
1624         for_each_possible_cpu(i) {
1625                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1626                         cpumask_set_cpu(i, cpumask);
1627         }
1628
1629         ret = -EFAULT;
1630         retlen = min_t(unsigned short, header->len, cpumask_size());
1631         if (copy_to_user(argp, cpumask, retlen))
1632                 goto out_free_cpumask;
1633         if (retlen != header->len &&
1634             clear_user(argp + retlen, header->len - retlen))
1635                 goto out_free_cpumask;
1636
1637         ret = 0;
1638 out_free_cpumask:
1639         free_cpumask_var(cpumask);
1640 out_put_device:
1641         ublk_put_device(ub);
1642         return ret;
1643 }
1644
1645 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1646 {
1647         pr_devel("%s: dev id %d flags %llx\n", __func__,
1648                         info->dev_id, info->flags);
1649         pr_devel("\t nr_hw_queues %d queue_depth %d\n",
1650                         info->nr_hw_queues, info->queue_depth);
1651 }
1652
1653 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1654 {
1655         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1656         void __user *argp = (void __user *)(unsigned long)header->addr;
1657         struct ublksrv_ctrl_dev_info info;
1658         struct ublk_device *ub;
1659         int ret = -EINVAL;
1660
1661         if (header->len < sizeof(info) || !header->addr)
1662                 return -EINVAL;
1663         if (header->queue_id != (u16)-1) {
1664                 pr_warn("%s: queue_id is wrong %x\n",
1665                         __func__, header->queue_id);
1666                 return -EINVAL;
1667         }
1668         if (copy_from_user(&info, argp, sizeof(info)))
1669                 return -EFAULT;
1670         ublk_dump_dev_info(&info);
1671         if (header->dev_id != info.dev_id) {
1672                 pr_warn("%s: dev id not match %u %u\n",
1673                         __func__, header->dev_id, info.dev_id);
1674                 return -EINVAL;
1675         }
1676
1677         ret = mutex_lock_killable(&ublk_ctl_mutex);
1678         if (ret)
1679                 return ret;
1680
1681         ret = -ENOMEM;
1682         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1683         if (!ub)
1684                 goto out_unlock;
1685         mutex_init(&ub->mutex);
1686         spin_lock_init(&ub->mm_lock);
1687         INIT_WORK(&ub->quiesce_work, ublk_quiesce_work_fn);
1688         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1689         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1690
1691         ret = ublk_alloc_dev_number(ub, header->dev_id);
1692         if (ret < 0)
1693                 goto out_free_ub;
1694
1695         memcpy(&ub->dev_info, &info, sizeof(info));
1696
1697         /* update device id */
1698         ub->dev_info.dev_id = ub->ub_number;
1699
1700         /*
1701          * 64bit flags will be copied back to userspace as feature
1702          * negotiation result, so have to clear flags which driver
1703          * doesn't support yet, then userspace can get correct flags
1704          * (features) to handle.
1705          */
1706         ub->dev_info.flags &= UBLK_F_ALL;
1707
1708         if (!IS_BUILTIN(CONFIG_BLK_DEV_UBLK))
1709                 ub->dev_info.flags |= UBLK_F_URING_CMD_COMP_IN_TASK;
1710
1711         /* We are not ready to support zero copy */
1712         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1713
1714         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1715                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1716         ublk_align_max_io_size(ub);
1717
1718         ret = ublk_init_queues(ub);
1719         if (ret)
1720                 goto out_free_dev_number;
1721
1722         ret = ublk_add_tag_set(ub);
1723         if (ret)
1724                 goto out_deinit_queues;
1725
1726         ret = -EFAULT;
1727         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1728                 goto out_free_tag_set;
1729
1730         /*
1731          * Add the char dev so that ublksrv daemon can be setup.
1732          * ublk_add_chdev() will cleanup everything if it fails.
1733          */
1734         ret = ublk_add_chdev(ub);
1735         goto out_unlock;
1736
1737 out_free_tag_set:
1738         blk_mq_free_tag_set(&ub->tag_set);
1739 out_deinit_queues:
1740         ublk_deinit_queues(ub);
1741 out_free_dev_number:
1742         ublk_free_dev_number(ub);
1743 out_free_ub:
1744         mutex_destroy(&ub->mutex);
1745         kfree(ub);
1746 out_unlock:
1747         mutex_unlock(&ublk_ctl_mutex);
1748         return ret;
1749 }
1750
1751 static inline bool ublk_idr_freed(int id)
1752 {
1753         void *ptr;
1754
1755         spin_lock(&ublk_idr_lock);
1756         ptr = idr_find(&ublk_index_idr, id);
1757         spin_unlock(&ublk_idr_lock);
1758
1759         return ptr == NULL;
1760 }
1761
1762 static int ublk_ctrl_del_dev(int idx)
1763 {
1764         struct ublk_device *ub;
1765         int ret;
1766
1767         ret = mutex_lock_killable(&ublk_ctl_mutex);
1768         if (ret)
1769                 return ret;
1770
1771         ub = ublk_get_device_from_id(idx);
1772         if (ub) {
1773                 ublk_remove(ub);
1774                 ublk_put_device(ub);
1775                 ret = 0;
1776         } else {
1777                 ret = -ENODEV;
1778         }
1779
1780         /*
1781          * Wait until the idr is removed, then it can be reused after
1782          * DEL_DEV command is returned.
1783          */
1784         if (!ret)
1785                 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1786         mutex_unlock(&ublk_ctl_mutex);
1787
1788         return ret;
1789 }
1790
1791 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1792 {
1793         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1794
1795         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1796                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1797                         header->data[0], header->addr, header->len);
1798 }
1799
1800 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1801 {
1802         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1803         struct ublk_device *ub;
1804
1805         ub = ublk_get_device_from_id(header->dev_id);
1806         if (!ub)
1807                 return -EINVAL;
1808
1809         ublk_stop_dev(ub);
1810         cancel_work_sync(&ub->stop_work);
1811         cancel_work_sync(&ub->quiesce_work);
1812
1813         ublk_put_device(ub);
1814         return 0;
1815 }
1816
1817 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1818 {
1819         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1820         void __user *argp = (void __user *)(unsigned long)header->addr;
1821         struct ublk_device *ub;
1822         int ret = 0;
1823
1824         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1825                 return -EINVAL;
1826
1827         ub = ublk_get_device_from_id(header->dev_id);
1828         if (!ub)
1829                 return -EINVAL;
1830
1831         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1832                 ret = -EFAULT;
1833         ublk_put_device(ub);
1834
1835         return ret;
1836 }
1837
1838 static int ublk_ctrl_get_params(struct io_uring_cmd *cmd)
1839 {
1840         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1841         void __user *argp = (void __user *)(unsigned long)header->addr;
1842         struct ublk_params_header ph;
1843         struct ublk_device *ub;
1844         int ret;
1845
1846         if (header->len <= sizeof(ph) || !header->addr)
1847                 return -EINVAL;
1848
1849         if (copy_from_user(&ph, argp, sizeof(ph)))
1850                 return -EFAULT;
1851
1852         if (ph.len > header->len || !ph.len)
1853                 return -EINVAL;
1854
1855         if (ph.len > sizeof(struct ublk_params))
1856                 ph.len = sizeof(struct ublk_params);
1857
1858         ub = ublk_get_device_from_id(header->dev_id);
1859         if (!ub)
1860                 return -EINVAL;
1861
1862         mutex_lock(&ub->mutex);
1863         if (copy_to_user(argp, &ub->params, ph.len))
1864                 ret = -EFAULT;
1865         else
1866                 ret = 0;
1867         mutex_unlock(&ub->mutex);
1868
1869         ublk_put_device(ub);
1870         return ret;
1871 }
1872
1873 static int ublk_ctrl_set_params(struct io_uring_cmd *cmd)
1874 {
1875         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1876         void __user *argp = (void __user *)(unsigned long)header->addr;
1877         struct ublk_params_header ph;
1878         struct ublk_device *ub;
1879         int ret = -EFAULT;
1880
1881         if (header->len <= sizeof(ph) || !header->addr)
1882                 return -EINVAL;
1883
1884         if (copy_from_user(&ph, argp, sizeof(ph)))
1885                 return -EFAULT;
1886
1887         if (ph.len > header->len || !ph.len || !ph.types)
1888                 return -EINVAL;
1889
1890         if (ph.len > sizeof(struct ublk_params))
1891                 ph.len = sizeof(struct ublk_params);
1892
1893         ub = ublk_get_device_from_id(header->dev_id);
1894         if (!ub)
1895                 return -EINVAL;
1896
1897         /* parameters can only be changed when device isn't live */
1898         mutex_lock(&ub->mutex);
1899         if (ub->dev_info.state == UBLK_S_DEV_LIVE) {
1900                 ret = -EACCES;
1901         } else if (copy_from_user(&ub->params, argp, ph.len)) {
1902                 ret = -EFAULT;
1903         } else {
1904                 /* clear all we don't support yet */
1905                 ub->params.types &= UBLK_PARAM_TYPE_ALL;
1906                 ret = ublk_validate_params(ub);
1907                 if (ret)
1908                         ub->params.types = 0;
1909         }
1910         mutex_unlock(&ub->mutex);
1911         ublk_put_device(ub);
1912
1913         return ret;
1914 }
1915
1916 static void ublk_queue_reinit(struct ublk_device *ub, struct ublk_queue *ubq)
1917 {
1918         int i;
1919
1920         WARN_ON_ONCE(!(ubq->ubq_daemon && ubq_daemon_is_dying(ubq)));
1921         /* All old ioucmds have to be completed */
1922         WARN_ON_ONCE(ubq->nr_io_ready);
1923         /* old daemon is PF_EXITING, put it now */
1924         put_task_struct(ubq->ubq_daemon);
1925         /* We have to reset it to NULL, otherwise ub won't accept new FETCH_REQ */
1926         ubq->ubq_daemon = NULL;
1927
1928         for (i = 0; i < ubq->q_depth; i++) {
1929                 struct ublk_io *io = &ubq->ios[i];
1930
1931                 /* forget everything now and be ready for new FETCH_REQ */
1932                 io->flags = 0;
1933                 io->cmd = NULL;
1934                 io->addr = 0;
1935         }
1936 }
1937
1938 static int ublk_ctrl_start_recovery(struct io_uring_cmd *cmd)
1939 {
1940         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1941         struct ublk_device *ub;
1942         int ret = -EINVAL;
1943         int i;
1944
1945         ub = ublk_get_device_from_id(header->dev_id);
1946         if (!ub)
1947                 return ret;
1948
1949         mutex_lock(&ub->mutex);
1950         if (!ublk_can_use_recovery(ub))
1951                 goto out_unlock;
1952         /*
1953          * START_RECOVERY is only allowd after:
1954          *
1955          * (1) UB_STATE_OPEN is not set, which means the dying process is exited
1956          *     and related io_uring ctx is freed so file struct of /dev/ublkcX is
1957          *     released.
1958          *
1959          * (2) UBLK_S_DEV_QUIESCED is set, which means the quiesce_work:
1960          *     (a)has quiesced request queue
1961          *     (b)has requeued every inflight rqs whose io_flags is ACTIVE
1962          *     (c)has requeued/aborted every inflight rqs whose io_flags is NOT ACTIVE
1963          *     (d)has completed/camceled all ioucmds owned by ther dying process
1964          */
1965         if (test_bit(UB_STATE_OPEN, &ub->state) ||
1966                         ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
1967                 ret = -EBUSY;
1968                 goto out_unlock;
1969         }
1970         pr_devel("%s: start recovery for dev id %d.\n", __func__, header->dev_id);
1971         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
1972                 ublk_queue_reinit(ub, ublk_get_queue(ub, i));
1973         /* set to NULL, otherwise new ubq_daemon cannot mmap the io_cmd_buf */
1974         ub->mm = NULL;
1975         ub->nr_queues_ready = 0;
1976         ub->nr_privileged_daemon = 0;
1977         init_completion(&ub->completion);
1978         ret = 0;
1979  out_unlock:
1980         mutex_unlock(&ub->mutex);
1981         ublk_put_device(ub);
1982         return ret;
1983 }
1984
1985 static int ublk_ctrl_end_recovery(struct io_uring_cmd *cmd)
1986 {
1987         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1988         int ublksrv_pid = (int)header->data[0];
1989         struct ublk_device *ub;
1990         int ret = -EINVAL;
1991
1992         ub = ublk_get_device_from_id(header->dev_id);
1993         if (!ub)
1994                 return ret;
1995
1996         pr_devel("%s: Waiting for new ubq_daemons(nr: %d) are ready, dev id %d...\n",
1997                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
1998         /* wait until new ubq_daemon sending all FETCH_REQ */
1999         wait_for_completion_interruptible(&ub->completion);
2000         pr_devel("%s: All new ubq_daemons(nr: %d) are ready, dev id %d\n",
2001                         __func__, ub->dev_info.nr_hw_queues, header->dev_id);
2002
2003         mutex_lock(&ub->mutex);
2004         if (!ublk_can_use_recovery(ub))
2005                 goto out_unlock;
2006
2007         if (ub->dev_info.state != UBLK_S_DEV_QUIESCED) {
2008                 ret = -EBUSY;
2009                 goto out_unlock;
2010         }
2011         ub->dev_info.ublksrv_pid = ublksrv_pid;
2012         pr_devel("%s: new ublksrv_pid %d, dev id %d\n",
2013                         __func__, ublksrv_pid, header->dev_id);
2014         blk_mq_unquiesce_queue(ub->ub_disk->queue);
2015         pr_devel("%s: queue unquiesced, dev id %d.\n",
2016                         __func__, header->dev_id);
2017         blk_mq_kick_requeue_list(ub->ub_disk->queue);
2018         ub->dev_info.state = UBLK_S_DEV_LIVE;
2019         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
2020         ret = 0;
2021  out_unlock:
2022         mutex_unlock(&ub->mutex);
2023         ublk_put_device(ub);
2024         return ret;
2025 }
2026
2027 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
2028                 unsigned int issue_flags)
2029 {
2030         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
2031         int ret = -EINVAL;
2032
2033         if (issue_flags & IO_URING_F_NONBLOCK)
2034                 return -EAGAIN;
2035
2036         ublk_ctrl_cmd_dump(cmd);
2037
2038         if (!(issue_flags & IO_URING_F_SQE128))
2039                 goto out;
2040
2041         ret = -EPERM;
2042         if (!capable(CAP_SYS_ADMIN))
2043                 goto out;
2044
2045         ret = -ENODEV;
2046         switch (cmd->cmd_op) {
2047         case UBLK_CMD_START_DEV:
2048                 ret = ublk_ctrl_start_dev(cmd);
2049                 break;
2050         case UBLK_CMD_STOP_DEV:
2051                 ret = ublk_ctrl_stop_dev(cmd);
2052                 break;
2053         case UBLK_CMD_GET_DEV_INFO:
2054                 ret = ublk_ctrl_get_dev_info(cmd);
2055                 break;
2056         case UBLK_CMD_ADD_DEV:
2057                 ret = ublk_ctrl_add_dev(cmd);
2058                 break;
2059         case UBLK_CMD_DEL_DEV:
2060                 ret = ublk_ctrl_del_dev(header->dev_id);
2061                 break;
2062         case UBLK_CMD_GET_QUEUE_AFFINITY:
2063                 ret = ublk_ctrl_get_queue_affinity(cmd);
2064                 break;
2065         case UBLK_CMD_GET_PARAMS:
2066                 ret = ublk_ctrl_get_params(cmd);
2067                 break;
2068         case UBLK_CMD_SET_PARAMS:
2069                 ret = ublk_ctrl_set_params(cmd);
2070                 break;
2071         case UBLK_CMD_START_USER_RECOVERY:
2072                 ret = ublk_ctrl_start_recovery(cmd);
2073                 break;
2074         case UBLK_CMD_END_USER_RECOVERY:
2075                 ret = ublk_ctrl_end_recovery(cmd);
2076                 break;
2077         default:
2078                 break;
2079         }
2080  out:
2081         io_uring_cmd_done(cmd, ret, 0, issue_flags);
2082         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
2083                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
2084         return -EIOCBQUEUED;
2085 }
2086
2087 static const struct file_operations ublk_ctl_fops = {
2088         .open           = nonseekable_open,
2089         .uring_cmd      = ublk_ctrl_uring_cmd,
2090         .owner          = THIS_MODULE,
2091         .llseek         = noop_llseek,
2092 };
2093
2094 static struct miscdevice ublk_misc = {
2095         .minor          = MISC_DYNAMIC_MINOR,
2096         .name           = "ublk-control",
2097         .fops           = &ublk_ctl_fops,
2098 };
2099
2100 static int __init ublk_init(void)
2101 {
2102         int ret;
2103
2104         init_waitqueue_head(&ublk_idr_wq);
2105
2106         ret = misc_register(&ublk_misc);
2107         if (ret)
2108                 return ret;
2109
2110         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
2111         if (ret)
2112                 goto unregister_mis;
2113
2114         ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
2115         if (IS_ERR(ublk_chr_class)) {
2116                 ret = PTR_ERR(ublk_chr_class);
2117                 goto free_chrdev_region;
2118         }
2119         return 0;
2120
2121 free_chrdev_region:
2122         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2123 unregister_mis:
2124         misc_deregister(&ublk_misc);
2125         return ret;
2126 }
2127
2128 static void __exit ublk_exit(void)
2129 {
2130         struct ublk_device *ub;
2131         int id;
2132
2133         idr_for_each_entry(&ublk_index_idr, ub, id)
2134                 ublk_remove(ub);
2135
2136         class_destroy(ublk_chr_class);
2137         misc_deregister(&ublk_misc);
2138
2139         idr_destroy(&ublk_index_idr);
2140         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
2141 }
2142
2143 module_init(ublk_init);
2144 module_exit(ublk_exit);
2145
2146 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
2147 MODULE_LICENSE("GPL");