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