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 | UBLK_F_URING_CMD_COMP_IN_TASK)
  51
  52 struct ublk_rq_data {
  53         struct callback_head work;
  54 };
  55
  56 struct ublk_uring_cmd_pdu {
  57         struct request *req;
  58 };
  59
  60 /*
  61  * io command is active: sqe cmd is received, and its cqe isn't done
  62  *
  63  * If the flag is set, the io command is owned by ublk driver, and waited
  64  * for incoming blk-mq request from the ublk block device.
  65  *
  66  * If the flag is cleared, the io command will be completed, and owned by
  67  * ublk server.
  68  */
  69 #define UBLK_IO_FLAG_ACTIVE     0x01
  70
  71 /*
  72  * IO command is completed via cqe, and it is being handled by ublksrv, and
  73  * not committed yet
  74  *
  75  * Basically exclusively with UBLK_IO_FLAG_ACTIVE, so can be served for
  76  * cross verification
  77  */
  78 #define UBLK_IO_FLAG_OWNED_BY_SRV 0x02
  79
  80 /*
  81  * IO command is aborted, so this flag is set in case of
  82  * !UBLK_IO_FLAG_ACTIVE.
  83  *
  84  * After this flag is observed, any pending or new incoming request
  85  * associated with this io command will be failed immediately
  86  */
  87 #define UBLK_IO_FLAG_ABORTED 0x04
  88
  89 struct ublk_io {
  90         /* userspace buffer address from io cmd */
  91         __u64   addr;
  92         unsigned int flags;
  93         int res;
  94
  95         struct io_uring_cmd *cmd;
  96 };
  97
  98 struct ublk_queue {
  99         int q_id;
 100         int q_depth;
 101
 102         unsigned long flags;
 103         struct task_struct      *ubq_daemon;
 104         char *io_cmd_buf;
 105
 106         unsigned long io_addr;  /* mapped vm address */
 107         unsigned int max_io_sz;
 108         bool abort_work_pending;
 109         unsigned short nr_io_ready;     /* how many ios setup */
 110         struct ublk_device *dev;
 111         struct ublk_io ios[0];
 112 };
 113
 114 #define UBLK_DAEMON_MONITOR_PERIOD      (5 * HZ)
 115
 116 struct ublk_device {
 117         struct gendisk          *ub_disk;
 118
 119         char    *__queues;
 120
 121         unsigned short  queue_size;
 122         unsigned short  bs_shift;
 123         struct ublksrv_ctrl_dev_info    dev_info;
 124
 125         struct blk_mq_tag_set   tag_set;
 126
 127         struct cdev             cdev;
 128         struct device           cdev_dev;
 129
 130 #define UB_STATE_OPEN           0
 131 #define UB_STATE_USED           1
 132         unsigned long           state;
 133         int                     ub_number;
 134
 135         struct mutex            mutex;
 136
 137         spinlock_t              mm_lock;
 138         struct mm_struct        *mm;
 139
 140         struct completion       completion;
 141         unsigned int            nr_queues_ready;
 142         atomic_t                nr_aborted_queues;
 143
 144         /*
 145          * Our ubq->daemon may be killed without any notification, so
 146          * monitor each queue's daemon periodically
 147          */
 148         struct delayed_work     monitor_work;
 149         struct work_struct      stop_work;
 150 };
 151
 152 static dev_t ublk_chr_devt;
 153 static struct class *ublk_chr_class;
 154
 155 static DEFINE_IDR(ublk_index_idr);
 156 static DEFINE_SPINLOCK(ublk_idr_lock);
 157 static wait_queue_head_t ublk_idr_wq;   /* wait until one idr is freed */
 158
 159 static DEFINE_MUTEX(ublk_ctl_mutex);
 160
 161 static struct miscdevice ublk_misc;
 162
 163 static inline bool ublk_can_use_task_work(const struct ublk_queue *ubq)
 164 {
 165         if (IS_BUILTIN(CONFIG_BLK_DEV_UBLK) &&
 166                         !(ubq->flags & UBLK_F_URING_CMD_COMP_IN_TASK))
 167                 return true;
 168         return false;
 169 }
 170
 171 static struct ublk_device *ublk_get_device(struct ublk_device *ub)
 172 {
 173         if (kobject_get_unless_zero(&ub->cdev_dev.kobj))
 174                 return ub;
 175         return NULL;
 176 }
 177
 178 static void ublk_put_device(struct ublk_device *ub)
 179 {
 180         put_device(&ub->cdev_dev);
 181 }
 182
 183 static inline struct ublk_queue *ublk_get_queue(struct ublk_device *dev,
 184                 int qid)
 185 {
 186        return (struct ublk_queue *)&(dev->__queues[qid * dev->queue_size]);
 187 }
 188
 189 static inline bool ublk_rq_has_data(const struct request *rq)
 190 {
 191         return rq->bio && bio_has_data(rq->bio);
 192 }
 193
 194 static inline struct ublksrv_io_desc *ublk_get_iod(struct ublk_queue *ubq,
 195                 int tag)
 196 {
 197         return (struct ublksrv_io_desc *)
 198                 &(ubq->io_cmd_buf[tag * sizeof(struct ublksrv_io_desc)]);
 199 }
 200
 201 static inline char *ublk_queue_cmd_buf(struct ublk_device *ub, int q_id)
 202 {
 203         return ublk_get_queue(ub, q_id)->io_cmd_buf;
 204 }
 205
 206 static inline int ublk_queue_cmd_buf_size(struct ublk_device *ub, int q_id)
 207 {
 208         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 209
 210         return round_up(ubq->q_depth * sizeof(struct ublksrv_io_desc),
 211                         PAGE_SIZE);
 212 }
 213
 214 static void ublk_free_disk(struct gendisk *disk)
 215 {
 216         struct ublk_device *ub = disk->private_data;
 217
 218         clear_bit(UB_STATE_USED, &ub->state);
 219         put_device(&ub->cdev_dev);
 220 }
 221
 222 static const struct block_device_operations ub_fops = {
 223         .owner =        THIS_MODULE,
 224         .free_disk =    ublk_free_disk,
 225 };
 226
 227 #define UBLK_MAX_PIN_PAGES      32
 228
 229 struct ublk_map_data {
 230         const struct ublk_queue *ubq;
 231         const struct request *rq;
 232         const struct ublk_io *io;
 233         unsigned max_bytes;
 234 };
 235
 236 struct ublk_io_iter {
 237         struct page *pages[UBLK_MAX_PIN_PAGES];
 238         unsigned pg_off;        /* offset in the 1st page in pages */
 239         int nr_pages;           /* how many page pointers in pages */
 240         struct bio *bio;
 241         struct bvec_iter iter;
 242 };
 243
 244 static inline unsigned ublk_copy_io_pages(struct ublk_io_iter *data,
 245                 unsigned max_bytes, bool to_vm)
 246 {
 247         const unsigned total = min_t(unsigned, max_bytes,
 248                         PAGE_SIZE - data->pg_off +
 249                         ((data->nr_pages - 1) << PAGE_SHIFT));
 250         unsigned done = 0;
 251         unsigned pg_idx = 0;
 252
 253         while (done < total) {
 254                 struct bio_vec bv = bio_iter_iovec(data->bio, data->iter);
 255                 const unsigned int bytes = min3(bv.bv_len, total - done,
 256                                 (unsigned)(PAGE_SIZE - data->pg_off));
 257                 void *bv_buf = bvec_kmap_local(&bv);
 258                 void *pg_buf = kmap_local_page(data->pages[pg_idx]);
 259
 260                 if (to_vm)
 261                         memcpy(pg_buf + data->pg_off, bv_buf, bytes);
 262                 else
 263                         memcpy(bv_buf, pg_buf + data->pg_off, bytes);
 264
 265                 kunmap_local(pg_buf);
 266                 kunmap_local(bv_buf);
 267
 268                 /* advance page array */
 269                 data->pg_off += bytes;
 270                 if (data->pg_off == PAGE_SIZE) {
 271                         pg_idx += 1;
 272                         data->pg_off = 0;
 273                 }
 274
 275                 done += bytes;
 276
 277                 /* advance bio */
 278                 bio_advance_iter_single(data->bio, &data->iter, bytes);
 279                 if (!data->iter.bi_size) {
 280                         data->bio = data->bio->bi_next;
 281                         if (data->bio == NULL)
 282                                 break;
 283                         data->iter = data->bio->bi_iter;
 284                 }
 285         }
 286
 287         return done;
 288 }
 289
 290 static inline int ublk_copy_user_pages(struct ublk_map_data *data,
 291                 bool to_vm)
 292 {
 293         const unsigned int gup_flags = to_vm ? FOLL_WRITE : 0;
 294         const unsigned long start_vm = data->io->addr;
 295         unsigned int done = 0;
 296         struct ublk_io_iter iter = {
 297                 .pg_off = start_vm & (PAGE_SIZE - 1),
 298                 .bio    = data->rq->bio,
 299                 .iter   = data->rq->bio->bi_iter,
 300         };
 301         const unsigned int nr_pages = round_up(data->max_bytes +
 302                         (start_vm & (PAGE_SIZE - 1)), PAGE_SIZE) >> PAGE_SHIFT;
 303
 304         while (done < nr_pages) {
 305                 const unsigned to_pin = min_t(unsigned, UBLK_MAX_PIN_PAGES,
 306                                 nr_pages - done);
 307                 unsigned i, len;
 308
 309                 iter.nr_pages = get_user_pages_fast(start_vm +
 310                                 (done << PAGE_SHIFT), to_pin, gup_flags,
 311                                 iter.pages);
 312                 if (iter.nr_pages <= 0)
 313                         return done == 0 ? iter.nr_pages : done;
 314                 len = ublk_copy_io_pages(&iter, data->max_bytes, to_vm);
 315                 for (i = 0; i < iter.nr_pages; i++) {
 316                         if (to_vm)
 317                                 set_page_dirty(iter.pages[i]);
 318                         put_page(iter.pages[i]);
 319                 }
 320                 data->max_bytes -= len;
 321                 done += iter.nr_pages;
 322         }
 323
 324         return done;
 325 }
 326
 327 static int ublk_map_io(const struct ublk_queue *ubq, const struct request *req,
 328                 struct ublk_io *io)
 329 {
 330         const unsigned int rq_bytes = blk_rq_bytes(req);
 331         /*
 332          * no zero copy, we delay copy WRITE request data into ublksrv
 333          * context and the big benefit is that pinning pages in current
 334          * context is pretty fast, see ublk_pin_user_pages
 335          */
 336         if (req_op(req) != REQ_OP_WRITE && req_op(req) != REQ_OP_FLUSH)
 337                 return rq_bytes;
 338
 339         if (ublk_rq_has_data(req)) {
 340                 struct ublk_map_data data = {
 341                         .ubq    =       ubq,
 342                         .rq     =       req,
 343                         .io     =       io,
 344                         .max_bytes =    rq_bytes,
 345                 };
 346
 347                 ublk_copy_user_pages(&data, true);
 348
 349                 return rq_bytes - data.max_bytes;
 350         }
 351         return rq_bytes;
 352 }
 353
 354 static int ublk_unmap_io(const struct ublk_queue *ubq,
 355                 const struct request *req,
 356                 struct ublk_io *io)
 357 {
 358         const unsigned int rq_bytes = blk_rq_bytes(req);
 359
 360         if (req_op(req) == REQ_OP_READ && ublk_rq_has_data(req)) {
 361                 struct ublk_map_data data = {
 362                         .ubq    =       ubq,
 363                         .rq     =       req,
 364                         .io     =       io,
 365                         .max_bytes =    io->res,
 366                 };
 367
 368                 WARN_ON_ONCE(io->res > rq_bytes);
 369
 370                 ublk_copy_user_pages(&data, false);
 371
 372                 return io->res - data.max_bytes;
 373         }
 374         return rq_bytes;
 375 }
 376
 377 static inline unsigned int ublk_req_build_flags(struct request *req)
 378 {
 379         unsigned flags = 0;
 380
 381         if (req->cmd_flags & REQ_FAILFAST_DEV)
 382                 flags |= UBLK_IO_F_FAILFAST_DEV;
 383
 384         if (req->cmd_flags & REQ_FAILFAST_TRANSPORT)
 385                 flags |= UBLK_IO_F_FAILFAST_TRANSPORT;
 386
 387         if (req->cmd_flags & REQ_FAILFAST_DRIVER)
 388                 flags |= UBLK_IO_F_FAILFAST_DRIVER;
 389
 390         if (req->cmd_flags & REQ_META)
 391                 flags |= UBLK_IO_F_META;
 392
 393         if (req->cmd_flags & REQ_FUA)
 394                 flags |= UBLK_IO_F_FUA;
 395
 396         if (req->cmd_flags & REQ_NOUNMAP)
 397                 flags |= UBLK_IO_F_NOUNMAP;
 398
 399         if (req->cmd_flags & REQ_SWAP)
 400                 flags |= UBLK_IO_F_SWAP;
 401
 402         return flags;
 403 }
 404
 405 static blk_status_t ublk_setup_iod(struct ublk_queue *ubq, struct request *req)
 406 {
 407         struct ublksrv_io_desc *iod = ublk_get_iod(ubq, req->tag);
 408         struct ublk_io *io = &ubq->ios[req->tag];
 409         u32 ublk_op;
 410
 411         switch (req_op(req)) {
 412         case REQ_OP_READ:
 413                 ublk_op = UBLK_IO_OP_READ;
 414                 break;
 415         case REQ_OP_WRITE:
 416                 ublk_op = UBLK_IO_OP_WRITE;
 417                 break;
 418         case REQ_OP_FLUSH:
 419                 ublk_op = UBLK_IO_OP_FLUSH;
 420                 break;
 421         case REQ_OP_DISCARD:
 422                 ublk_op = UBLK_IO_OP_DISCARD;
 423                 break;
 424         case REQ_OP_WRITE_ZEROES:
 425                 ublk_op = UBLK_IO_OP_WRITE_ZEROES;
 426                 break;
 427         default:
 428                 return BLK_STS_IOERR;
 429         }
 430
 431         /* need to translate since kernel may change */
 432         iod->op_flags = ublk_op | ublk_req_build_flags(req);
 433         iod->nr_sectors = blk_rq_sectors(req);
 434         iod->start_sector = blk_rq_pos(req);
 435         iod->addr = io->addr;
 436
 437         return BLK_STS_OK;
 438 }
 439
 440 static inline struct ublk_uring_cmd_pdu *ublk_get_uring_cmd_pdu(
 441                 struct io_uring_cmd *ioucmd)
 442 {
 443         return (struct ublk_uring_cmd_pdu *)&ioucmd->pdu;
 444 }
 445
 446 static bool ubq_daemon_is_dying(struct ublk_queue *ubq)
 447 {
 448         return ubq->ubq_daemon->flags & PF_EXITING;
 449 }
 450
 451 /* todo: handle partial completion */
 452 static void ublk_complete_rq(struct request *req)
 453 {
 454         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 455         struct ublk_io *io = &ubq->ios[req->tag];
 456         unsigned int unmapped_bytes;
 457
 458         /* failed read IO if nothing is read */
 459         if (!io->res && req_op(req) == REQ_OP_READ)
 460                 io->res = -EIO;
 461
 462         if (io->res < 0) {
 463                 blk_mq_end_request(req, errno_to_blk_status(io->res));
 464                 return;
 465         }
 466
 467         /*
 468          * FLUSH or DISCARD usually won't return bytes returned, so end them
 469          * directly.
 470          *
 471          * Both the two needn't unmap.
 472          */
 473         if (req_op(req) != REQ_OP_READ && req_op(req) != REQ_OP_WRITE) {
 474                 blk_mq_end_request(req, BLK_STS_OK);
 475                 return;
 476         }
 477
 478         /* for READ request, writing data in iod->addr to rq buffers */
 479         unmapped_bytes = ublk_unmap_io(ubq, req, io);
 480
 481         /*
 482          * Extremely impossible since we got data filled in just before
 483          *
 484          * Re-read simply for this unlikely case.
 485          */
 486         if (unlikely(unmapped_bytes < io->res))
 487                 io->res = unmapped_bytes;
 488
 489         if (blk_update_request(req, BLK_STS_OK, io->res))
 490                 blk_mq_requeue_request(req, true);
 491         else
 492                 __blk_mq_end_request(req, BLK_STS_OK);
 493 }
 494
 495 /*
 496  * __ublk_fail_req() may be called from abort context or ->ubq_daemon
 497  * context during exiting, so lock is required.
 498  *
 499  * Also aborting may not be started yet, keep in mind that one failed
 500  * request may be issued by block layer again.
 501  */
 502 static void __ublk_fail_req(struct ublk_io *io, struct request *req)
 503 {
 504         WARN_ON_ONCE(io->flags & UBLK_IO_FLAG_ACTIVE);
 505
 506         if (!(io->flags & UBLK_IO_FLAG_ABORTED)) {
 507                 io->flags |= UBLK_IO_FLAG_ABORTED;
 508                 blk_mq_end_request(req, BLK_STS_IOERR);
 509         }
 510 }
 511
 512 #define UBLK_REQUEUE_DELAY_MS   3
 513
 514 static inline void __ublk_rq_task_work(struct request *req)
 515 {
 516         struct ublk_queue *ubq = req->mq_hctx->driver_data;
 517         struct ublk_device *ub = ubq->dev;
 518         int tag = req->tag;
 519         struct ublk_io *io = &ubq->ios[tag];
 520         bool task_exiting = current != ubq->ubq_daemon ||
 521                 (current->flags & PF_EXITING);
 522         unsigned int mapped_bytes;
 523
 524         pr_devel("%s: complete: op %d, qid %d tag %d io_flags %x addr %llx\n",
 525                         __func__, io->cmd->cmd_op, ubq->q_id, req->tag, io->flags,
 526                         ublk_get_iod(ubq, req->tag)->addr);
 527
 528         if (unlikely(task_exiting)) {
 529                 blk_mq_end_request(req, BLK_STS_IOERR);
 530                 mod_delayed_work(system_wq, &ub->monitor_work, 0);
 531                 return;
 532         }
 533
 534         mapped_bytes = ublk_map_io(ubq, req, io);
 535
 536         /* partially mapped, update io descriptor */
 537         if (unlikely(mapped_bytes != blk_rq_bytes(req))) {
 538                 /*
 539                  * Nothing mapped, retry until we succeed.
 540                  *
 541                  * We may never succeed in mapping any bytes here because
 542                  * of OOM. TODO: reserve one buffer with single page pinned
 543                  * for providing forward progress guarantee.
 544                  */
 545                 if (unlikely(!mapped_bytes)) {
 546                         blk_mq_requeue_request(req, false);
 547                         blk_mq_delay_kick_requeue_list(req->q,
 548                                         UBLK_REQUEUE_DELAY_MS);
 549                         return;
 550                 }
 551
 552                 ublk_get_iod(ubq, req->tag)->nr_sectors =
 553                         mapped_bytes >> 9;
 554         }
 555
 556         /* mark this cmd owned by ublksrv */
 557         io->flags |= UBLK_IO_FLAG_OWNED_BY_SRV;
 558
 559         /*
 560          * clear ACTIVE since we are done with this sqe/cmd slot
 561          * We can only accept io cmd in case of being not active.
 562          */
 563         io->flags &= ~UBLK_IO_FLAG_ACTIVE;
 564
 565         /* tell ublksrv one io request is coming */
 566         io_uring_cmd_done(io->cmd, UBLK_IO_RES_OK, 0);
 567 }
 568
 569 static void ublk_rq_task_work_cb(struct io_uring_cmd *cmd)
 570 {
 571         struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 572
 573         __ublk_rq_task_work(pdu->req);
 574 }
 575
 576 static void ublk_rq_task_work_fn(struct callback_head *work)
 577 {
 578         struct ublk_rq_data *data = container_of(work,
 579                         struct ublk_rq_data, work);
 580         struct request *req = blk_mq_rq_from_pdu(data);
 581
 582         __ublk_rq_task_work(req);
 583 }
 584
 585 static blk_status_t ublk_queue_rq(struct blk_mq_hw_ctx *hctx,
 586                 const struct blk_mq_queue_data *bd)
 587 {
 588         struct ublk_queue *ubq = hctx->driver_data;
 589         struct request *rq = bd->rq;
 590         blk_status_t res;
 591
 592         /* fill iod to slot in io cmd buffer */
 593         res = ublk_setup_iod(ubq, rq);
 594         if (unlikely(res != BLK_STS_OK))
 595                 return BLK_STS_IOERR;
 596
 597         blk_mq_start_request(bd->rq);
 598
 599         if (unlikely(ubq_daemon_is_dying(ubq))) {
 600  fail:
 601                 mod_delayed_work(system_wq, &ubq->dev->monitor_work, 0);
 602                 return BLK_STS_IOERR;
 603         }
 604
 605         if (ublk_can_use_task_work(ubq)) {
 606                 struct ublk_rq_data *data = blk_mq_rq_to_pdu(rq);
 607                 enum task_work_notify_mode notify_mode = bd->last ?
 608                         TWA_SIGNAL_NO_IPI : TWA_NONE;
 609
 610                 if (task_work_add(ubq->ubq_daemon, &data->work, notify_mode))
 611                         goto fail;
 612         } else {
 613                 struct io_uring_cmd *cmd = ubq->ios[rq->tag].cmd;
 614                 struct ublk_uring_cmd_pdu *pdu = ublk_get_uring_cmd_pdu(cmd);
 615
 616                 pdu->req = rq;
 617                 io_uring_cmd_complete_in_task(cmd, ublk_rq_task_work_cb);
 618         }
 619
 620         return BLK_STS_OK;
 621 }
 622
 623 static void ublk_commit_rqs(struct blk_mq_hw_ctx *hctx)
 624 {
 625         struct ublk_queue *ubq = hctx->driver_data;
 626
 627         if (ublk_can_use_task_work(ubq))
 628                 __set_notify_signal(ubq->ubq_daemon);
 629 }
 630
 631 static int ublk_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 632                 unsigned int hctx_idx)
 633 {
 634         struct ublk_device *ub = driver_data;
 635         struct ublk_queue *ubq = ublk_get_queue(ub, hctx->queue_num);
 636
 637         hctx->driver_data = ubq;
 638         return 0;
 639 }
 640
 641 static int ublk_init_rq(struct blk_mq_tag_set *set, struct request *req,
 642                 unsigned int hctx_idx, unsigned int numa_node)
 643 {
 644         struct ublk_rq_data *data = blk_mq_rq_to_pdu(req);
 645
 646         init_task_work(&data->work, ublk_rq_task_work_fn);
 647         return 0;
 648 }
 649
 650 static const struct blk_mq_ops ublk_mq_ops = {
 651         .queue_rq       = ublk_queue_rq,
 652         .commit_rqs     = ublk_commit_rqs,
 653         .init_hctx      = ublk_init_hctx,
 654         .init_request   = ublk_init_rq,
 655 };
 656
 657 static int ublk_ch_open(struct inode *inode, struct file *filp)
 658 {
 659         struct ublk_device *ub = container_of(inode->i_cdev,
 660                         struct ublk_device, cdev);
 661
 662         if (test_and_set_bit(UB_STATE_OPEN, &ub->state))
 663                 return -EBUSY;
 664         filp->private_data = ub;
 665         return 0;
 666 }
 667
 668 static int ublk_ch_release(struct inode *inode, struct file *filp)
 669 {
 670         struct ublk_device *ub = filp->private_data;
 671
 672         clear_bit(UB_STATE_OPEN, &ub->state);
 673         return 0;
 674 }
 675
 676 /* map pre-allocated per-queue cmd buffer to ublksrv daemon */
 677 static int ublk_ch_mmap(struct file *filp, struct vm_area_struct *vma)
 678 {
 679         struct ublk_device *ub = filp->private_data;
 680         size_t sz = vma->vm_end - vma->vm_start;
 681         unsigned max_sz = UBLK_MAX_QUEUE_DEPTH * sizeof(struct ublksrv_io_desc);
 682         unsigned long pfn, end, phys_off = vma->vm_pgoff << PAGE_SHIFT;
 683         int q_id, ret = 0;
 684
 685         spin_lock(&ub->mm_lock);
 686         if (!ub->mm)
 687                 ub->mm = current->mm;
 688         if (current->mm != ub->mm)
 689                 ret = -EINVAL;
 690         spin_unlock(&ub->mm_lock);
 691
 692         if (ret)
 693                 return ret;
 694
 695         if (vma->vm_flags & VM_WRITE)
 696                 return -EPERM;
 697
 698         end = UBLKSRV_CMD_BUF_OFFSET + ub->dev_info.nr_hw_queues * max_sz;
 699         if (phys_off < UBLKSRV_CMD_BUF_OFFSET || phys_off >= end)
 700                 return -EINVAL;
 701
 702         q_id = (phys_off - UBLKSRV_CMD_BUF_OFFSET) / max_sz;
 703         pr_devel("%s: qid %d, pid %d, addr %lx pg_off %lx sz %lu\n",
 704                         __func__, q_id, current->pid, vma->vm_start,
 705                         phys_off, (unsigned long)sz);
 706
 707         if (sz != ublk_queue_cmd_buf_size(ub, q_id))
 708                 return -EINVAL;
 709
 710         pfn = virt_to_phys(ublk_queue_cmd_buf(ub, q_id)) >> PAGE_SHIFT;
 711         return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
 712 }
 713
 714 static void ublk_commit_completion(struct ublk_device *ub,
 715                 struct ublksrv_io_cmd *ub_cmd)
 716 {
 717         u32 qid = ub_cmd->q_id, tag = ub_cmd->tag;
 718         struct ublk_queue *ubq = ublk_get_queue(ub, qid);
 719         struct ublk_io *io = &ubq->ios[tag];
 720         struct request *req;
 721
 722         /* now this cmd slot is owned by nbd driver */
 723         io->flags &= ~UBLK_IO_FLAG_OWNED_BY_SRV;
 724         io->res = ub_cmd->result;
 725
 726         /* find the io request and complete */
 727         req = blk_mq_tag_to_rq(ub->tag_set.tags[qid], tag);
 728
 729         if (req && likely(!blk_should_fake_timeout(req->q)))
 730                 ublk_complete_rq(req);
 731 }
 732
 733 /*
 734  * When ->ubq_daemon is exiting, either new request is ended immediately,
 735  * or any queued io command is drained, so it is safe to abort queue
 736  * lockless
 737  */
 738 static void ublk_abort_queue(struct ublk_device *ub, struct ublk_queue *ubq)
 739 {
 740         int i;
 741
 742         if (!ublk_get_device(ub))
 743                 return;
 744
 745         for (i = 0; i < ubq->q_depth; i++) {
 746                 struct ublk_io *io = &ubq->ios[i];
 747
 748                 if (!(io->flags & UBLK_IO_FLAG_ACTIVE)) {
 749                         struct request *rq;
 750
 751                         /*
 752                          * Either we fail the request or ublk_rq_task_work_fn
 753                          * will do it
 754                          */
 755                         rq = blk_mq_tag_to_rq(ub->tag_set.tags[ubq->q_id], i);
 756                         if (rq)
 757                                 __ublk_fail_req(io, rq);
 758                 }
 759         }
 760         ublk_put_device(ub);
 761 }
 762
 763 static void ublk_daemon_monitor_work(struct work_struct *work)
 764 {
 765         struct ublk_device *ub =
 766                 container_of(work, struct ublk_device, monitor_work.work);
 767         int i;
 768
 769         for (i = 0; i < ub->dev_info.nr_hw_queues; i++) {
 770                 struct ublk_queue *ubq = ublk_get_queue(ub, i);
 771
 772                 if (ubq_daemon_is_dying(ubq)) {
 773                         schedule_work(&ub->stop_work);
 774
 775                         /* abort queue is for making forward progress */
 776                         ublk_abort_queue(ub, ubq);
 777                 }
 778         }
 779
 780         /*
 781          * We can't schedule monitor work after ublk_remove() is started.
 782          *
 783          * No need ub->mutex, monitor work are canceled after state is marked
 784          * as DEAD, so DEAD state is observed reliably.
 785          */
 786         if (ub->dev_info.state != UBLK_S_DEV_DEAD)
 787                 schedule_delayed_work(&ub->monitor_work,
 788                                 UBLK_DAEMON_MONITOR_PERIOD);
 789 }
 790
 791 static void ublk_cancel_queue(struct ublk_queue *ubq)
 792 {
 793         int i;
 794
 795         for (i = 0; i < ubq->q_depth; i++) {
 796                 struct ublk_io *io = &ubq->ios[i];
 797
 798                 if (io->flags & UBLK_IO_FLAG_ACTIVE)
 799                         io_uring_cmd_done(io->cmd, UBLK_IO_RES_ABORT, 0);
 800         }
 801 }
 802
 803 /* Cancel all pending commands, must be called after del_gendisk() returns */
 804 static void ublk_cancel_dev(struct ublk_device *ub)
 805 {
 806         int i;
 807
 808         for (i = 0; i < ub->dev_info.nr_hw_queues; i++)
 809                 ublk_cancel_queue(ublk_get_queue(ub, i));
 810 }
 811
 812 static void ublk_stop_dev(struct ublk_device *ub)
 813 {
 814         mutex_lock(&ub->mutex);
 815         if (ub->dev_info.state != UBLK_S_DEV_LIVE)
 816                 goto unlock;
 817
 818         del_gendisk(ub->ub_disk);
 819         ub->dev_info.state = UBLK_S_DEV_DEAD;
 820         ub->dev_info.ublksrv_pid = -1;
 821         ublk_cancel_dev(ub);
 822         put_disk(ub->ub_disk);
 823         ub->ub_disk = NULL;
 824  unlock:
 825         mutex_unlock(&ub->mutex);
 826         cancel_delayed_work_sync(&ub->monitor_work);
 827 }
 828
 829 static inline bool ublk_queue_ready(struct ublk_queue *ubq)
 830 {
 831         return ubq->nr_io_ready == ubq->q_depth;
 832 }
 833
 834 /* device can only be started after all IOs are ready */
 835 static void ublk_mark_io_ready(struct ublk_device *ub, struct ublk_queue *ubq)
 836 {
 837         mutex_lock(&ub->mutex);
 838         ubq->nr_io_ready++;
 839         if (ublk_queue_ready(ubq)) {
 840                 ubq->ubq_daemon = current;
 841                 get_task_struct(ubq->ubq_daemon);
 842                 ub->nr_queues_ready++;
 843         }
 844         if (ub->nr_queues_ready == ub->dev_info.nr_hw_queues)
 845                 complete_all(&ub->completion);
 846         mutex_unlock(&ub->mutex);
 847 }
 848
 849 static int ublk_ch_uring_cmd(struct io_uring_cmd *cmd, unsigned int issue_flags)
 850 {
 851         struct ublksrv_io_cmd *ub_cmd = (struct ublksrv_io_cmd *)cmd->cmd;
 852         struct ublk_device *ub = cmd->file->private_data;
 853         struct ublk_queue *ubq;
 854         struct ublk_io *io;
 855         u32 cmd_op = cmd->cmd_op;
 856         unsigned tag = ub_cmd->tag;
 857         int ret = -EINVAL;
 858
 859         pr_devel("%s: received: cmd op %d queue %d tag %d result %d\n",
 860                         __func__, cmd->cmd_op, ub_cmd->q_id, tag,
 861                         ub_cmd->result);
 862
 863         if (!(issue_flags & IO_URING_F_SQE128))
 864                 goto out;
 865
 866         if (ub_cmd->q_id >= ub->dev_info.nr_hw_queues)
 867                 goto out;
 868
 869         ubq = ublk_get_queue(ub, ub_cmd->q_id);
 870         if (!ubq || ub_cmd->q_id != ubq->q_id)
 871                 goto out;
 872
 873         if (ubq->ubq_daemon && ubq->ubq_daemon != current)
 874                 goto out;
 875
 876         if (tag >= ubq->q_depth)
 877                 goto out;
 878
 879         io = &ubq->ios[tag];
 880
 881         /* there is pending io cmd, something must be wrong */
 882         if (io->flags & UBLK_IO_FLAG_ACTIVE) {
 883                 ret = -EBUSY;
 884                 goto out;
 885         }
 886
 887         switch (cmd_op) {
 888         case UBLK_IO_FETCH_REQ:
 889                 /* UBLK_IO_FETCH_REQ is only allowed before queue is setup */
 890                 if (ublk_queue_ready(ubq)) {
 891                         ret = -EBUSY;
 892                         goto out;
 893                 }
 894                 /*
 895                  * The io is being handled by server, so COMMIT_RQ is expected
 896                  * instead of FETCH_REQ
 897                  */
 898                 if (io->flags & UBLK_IO_FLAG_OWNED_BY_SRV)
 899                         goto out;
 900                 /* FETCH_RQ has to provide IO buffer */
 901                 if (!ub_cmd->addr)
 902                         goto out;
 903                 io->cmd = cmd;
 904                 io->flags |= UBLK_IO_FLAG_ACTIVE;
 905                 io->addr = ub_cmd->addr;
 906
 907                 ublk_mark_io_ready(ub, ubq);
 908                 break;
 909         case UBLK_IO_COMMIT_AND_FETCH_REQ:
 910                 /* FETCH_RQ has to provide IO buffer */
 911                 if (!ub_cmd->addr)
 912                         goto out;
 913                 if (!(io->flags & UBLK_IO_FLAG_OWNED_BY_SRV))
 914                         goto out;
 915                 io->addr = ub_cmd->addr;
 916                 io->flags |= UBLK_IO_FLAG_ACTIVE;
 917                 io->cmd = cmd;
 918                 ublk_commit_completion(ub, ub_cmd);
 919                 break;
 920         default:
 921                 goto out;
 922         }
 923         return -EIOCBQUEUED;
 924
 925  out:
 926         io_uring_cmd_done(cmd, ret, 0);
 927         pr_devel("%s: complete: cmd op %d, tag %d ret %x io_flags %x\n",
 928                         __func__, cmd_op, tag, ret, io->flags);
 929         return -EIOCBQUEUED;
 930 }
 931
 932 static const struct file_operations ublk_ch_fops = {
 933         .owner = THIS_MODULE,
 934         .open = ublk_ch_open,
 935         .release = ublk_ch_release,
 936         .llseek = no_llseek,
 937         .uring_cmd = ublk_ch_uring_cmd,
 938         .mmap = ublk_ch_mmap,
 939 };
 940
 941 static void ublk_deinit_queue(struct ublk_device *ub, int q_id)
 942 {
 943         int size = ublk_queue_cmd_buf_size(ub, q_id);
 944         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 945
 946         if (ubq->ubq_daemon)
 947                 put_task_struct(ubq->ubq_daemon);
 948         if (ubq->io_cmd_buf)
 949                 free_pages((unsigned long)ubq->io_cmd_buf, get_order(size));
 950 }
 951
 952 static int ublk_init_queue(struct ublk_device *ub, int q_id)
 953 {
 954         struct ublk_queue *ubq = ublk_get_queue(ub, q_id);
 955         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO;
 956         void *ptr;
 957         int size;
 958
 959         ubq->flags = ub->dev_info.flags;
 960         ubq->q_id = q_id;
 961         ubq->q_depth = ub->dev_info.queue_depth;
 962         size = ublk_queue_cmd_buf_size(ub, q_id);
 963
 964         ptr = (void *) __get_free_pages(gfp_flags, get_order(size));
 965         if (!ptr)
 966                 return -ENOMEM;
 967
 968         ubq->io_cmd_buf = ptr;
 969         ubq->dev = ub;
 970         return 0;
 971 }
 972
 973 static void ublk_deinit_queues(struct ublk_device *ub)
 974 {
 975         int nr_queues = ub->dev_info.nr_hw_queues;
 976         int i;
 977
 978         if (!ub->__queues)
 979                 return;
 980
 981         for (i = 0; i < nr_queues; i++)
 982                 ublk_deinit_queue(ub, i);
 983         kfree(ub->__queues);
 984 }
 985
 986 static int ublk_init_queues(struct ublk_device *ub)
 987 {
 988         int nr_queues = ub->dev_info.nr_hw_queues;
 989         int depth = ub->dev_info.queue_depth;
 990         int ubq_size = sizeof(struct ublk_queue) + depth * sizeof(struct ublk_io);
 991         int i, ret = -ENOMEM;
 992
 993         ub->queue_size = ubq_size;
 994         ub->__queues = kcalloc(nr_queues, ubq_size, GFP_KERNEL);
 995         if (!ub->__queues)
 996                 return ret;
 997
 998         for (i = 0; i < nr_queues; i++) {
 999                 if (ublk_init_queue(ub, i))
1000                         goto fail;
1001         }
1002
1003         init_completion(&ub->completion);
1004         return 0;
1005
1006  fail:
1007         ublk_deinit_queues(ub);
1008         return ret;
1009 }
1010
1011 static int ublk_alloc_dev_number(struct ublk_device *ub, int idx)
1012 {
1013         int i = idx;
1014         int err;
1015
1016         spin_lock(&ublk_idr_lock);
1017         /* allocate id, if @id >= 0, we're requesting that specific id */
1018         if (i >= 0) {
1019                 err = idr_alloc(&ublk_index_idr, ub, i, i + 1, GFP_NOWAIT);
1020                 if (err == -ENOSPC)
1021                         err = -EEXIST;
1022         } else {
1023                 err = idr_alloc(&ublk_index_idr, ub, 0, 0, GFP_NOWAIT);
1024         }
1025         spin_unlock(&ublk_idr_lock);
1026
1027         if (err >= 0)
1028                 ub->ub_number = err;
1029
1030         return err;
1031 }
1032
1033 static void ublk_free_dev_number(struct ublk_device *ub)
1034 {
1035         spin_lock(&ublk_idr_lock);
1036         idr_remove(&ublk_index_idr, ub->ub_number);
1037         wake_up_all(&ublk_idr_wq);
1038         spin_unlock(&ublk_idr_lock);
1039 }
1040
1041 static void ublk_cdev_rel(struct device *dev)
1042 {
1043         struct ublk_device *ub = container_of(dev, struct ublk_device, cdev_dev);
1044
1045         blk_mq_free_tag_set(&ub->tag_set);
1046         ublk_deinit_queues(ub);
1047         ublk_free_dev_number(ub);
1048         mutex_destroy(&ub->mutex);
1049         kfree(ub);
1050 }
1051
1052 static int ublk_add_chdev(struct ublk_device *ub)
1053 {
1054         struct device *dev = &ub->cdev_dev;
1055         int minor = ub->ub_number;
1056         int ret;
1057
1058         dev->parent = ublk_misc.this_device;
1059         dev->devt = MKDEV(MAJOR(ublk_chr_devt), minor);
1060         dev->class = ublk_chr_class;
1061         dev->release = ublk_cdev_rel;
1062         device_initialize(dev);
1063
1064         ret = dev_set_name(dev, "ublkc%d", minor);
1065         if (ret)
1066                 goto fail;
1067
1068         cdev_init(&ub->cdev, &ublk_ch_fops);
1069         ret = cdev_device_add(&ub->cdev, dev);
1070         if (ret)
1071                 goto fail;
1072         return 0;
1073  fail:
1074         put_device(dev);
1075         return ret;
1076 }
1077
1078 static void ublk_stop_work_fn(struct work_struct *work)
1079 {
1080         struct ublk_device *ub =
1081                 container_of(work, struct ublk_device, stop_work);
1082
1083         ublk_stop_dev(ub);
1084 }
1085
1086 /* align maximum I/O size to PAGE_SIZE */
1087 static void ublk_align_max_io_size(struct ublk_device *ub)
1088 {
1089         unsigned int max_rq_bytes = ub->dev_info.rq_max_blocks << ub->bs_shift;
1090
1091         ub->dev_info.rq_max_blocks =
1092                 round_down(max_rq_bytes, PAGE_SIZE) >> ub->bs_shift;
1093 }
1094
1095 static int ublk_add_tag_set(struct ublk_device *ub)
1096 {
1097         ub->tag_set.ops = &ublk_mq_ops;
1098         ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
1099         ub->tag_set.queue_depth = ub->dev_info.queue_depth;
1100         ub->tag_set.numa_node = NUMA_NO_NODE;
1101         ub->tag_set.cmd_size = sizeof(struct ublk_rq_data);
1102         ub->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1103         ub->tag_set.driver_data = ub;
1104         return blk_mq_alloc_tag_set(&ub->tag_set);
1105 }
1106
1107 static void ublk_remove(struct ublk_device *ub)
1108 {
1109         ublk_stop_dev(ub);
1110         cancel_work_sync(&ub->stop_work);
1111         cdev_device_del(&ub->cdev, &ub->cdev_dev);
1112         put_device(&ub->cdev_dev);
1113 }
1114
1115 static struct ublk_device *ublk_get_device_from_id(int idx)
1116 {
1117         struct ublk_device *ub = NULL;
1118
1119         if (idx < 0)
1120                 return NULL;
1121
1122         spin_lock(&ublk_idr_lock);
1123         ub = idr_find(&ublk_index_idr, idx);
1124         if (ub)
1125                 ub = ublk_get_device(ub);
1126         spin_unlock(&ublk_idr_lock);
1127
1128         return ub;
1129 }
1130
1131 static int ublk_ctrl_start_dev(struct io_uring_cmd *cmd)
1132 {
1133         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1134         int ublksrv_pid = (int)header->data[0];
1135         unsigned long dev_blocks = header->data[1];
1136         struct ublk_device *ub;
1137         struct gendisk *disk;
1138         int ret = -EINVAL;
1139
1140         if (ublksrv_pid <= 0)
1141                 return -EINVAL;
1142
1143         ub = ublk_get_device_from_id(header->dev_id);
1144         if (!ub)
1145                 return -EINVAL;
1146
1147         wait_for_completion_interruptible(&ub->completion);
1148
1149         schedule_delayed_work(&ub->monitor_work, UBLK_DAEMON_MONITOR_PERIOD);
1150
1151         mutex_lock(&ub->mutex);
1152         if (ub->dev_info.state == UBLK_S_DEV_LIVE ||
1153             test_bit(UB_STATE_USED, &ub->state)) {
1154                 ret = -EEXIST;
1155                 goto out_unlock;
1156         }
1157
1158         /* We may get disk size updated */
1159         if (dev_blocks)
1160                 ub->dev_info.dev_blocks = dev_blocks;
1161
1162         disk = blk_mq_alloc_disk(&ub->tag_set, ub);
1163         if (IS_ERR(disk)) {
1164                 ret = PTR_ERR(disk);
1165                 goto out_unlock;
1166         }
1167         sprintf(disk->disk_name, "ublkb%d", ub->ub_number);
1168         disk->fops = &ub_fops;
1169         disk->private_data = ub;
1170
1171         blk_queue_logical_block_size(disk->queue, ub->dev_info.block_size);
1172         blk_queue_physical_block_size(disk->queue, ub->dev_info.block_size);
1173         blk_queue_io_min(disk->queue, ub->dev_info.block_size);
1174         blk_queue_max_hw_sectors(disk->queue,
1175                 ub->dev_info.rq_max_blocks << (ub->bs_shift - 9));
1176         disk->queue->limits.discard_granularity = PAGE_SIZE;
1177         blk_queue_max_discard_sectors(disk->queue, UINT_MAX >> 9);
1178         blk_queue_max_write_zeroes_sectors(disk->queue, UINT_MAX >> 9);
1179
1180         set_capacity(disk, ub->dev_info.dev_blocks << (ub->bs_shift - 9));
1181
1182         ub->dev_info.ublksrv_pid = ublksrv_pid;
1183         ub->ub_disk = disk;
1184         get_device(&ub->cdev_dev);
1185         ret = add_disk(disk);
1186         if (ret) {
1187                 put_disk(disk);
1188                 goto out_unlock;
1189         }
1190         set_bit(UB_STATE_USED, &ub->state);
1191         ub->dev_info.state = UBLK_S_DEV_LIVE;
1192 out_unlock:
1193         mutex_unlock(&ub->mutex);
1194         ublk_put_device(ub);
1195         return ret;
1196 }
1197
1198 static int ublk_ctrl_get_queue_affinity(struct io_uring_cmd *cmd)
1199 {
1200         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1201         void __user *argp = (void __user *)(unsigned long)header->addr;
1202         struct ublk_device *ub;
1203         cpumask_var_t cpumask;
1204         unsigned long queue;
1205         unsigned int retlen;
1206         unsigned int i;
1207         int ret = -EINVAL;
1208
1209         if (header->len * BITS_PER_BYTE < nr_cpu_ids)
1210                 return -EINVAL;
1211         if (header->len & (sizeof(unsigned long)-1))
1212                 return -EINVAL;
1213         if (!header->addr)
1214                 return -EINVAL;
1215
1216         ub = ublk_get_device_from_id(header->dev_id);
1217         if (!ub)
1218                 return -EINVAL;
1219
1220         queue = header->data[0];
1221         if (queue >= ub->dev_info.nr_hw_queues)
1222                 goto out_put_device;
1223
1224         ret = -ENOMEM;
1225         if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
1226                 goto out_put_device;
1227
1228         for_each_possible_cpu(i) {
1229                 if (ub->tag_set.map[HCTX_TYPE_DEFAULT].mq_map[i] == queue)
1230                         cpumask_set_cpu(i, cpumask);
1231         }
1232
1233         ret = -EFAULT;
1234         retlen = min_t(unsigned short, header->len, cpumask_size());
1235         if (copy_to_user(argp, cpumask, retlen))
1236                 goto out_free_cpumask;
1237         if (retlen != header->len &&
1238             clear_user(argp + retlen, header->len - retlen))
1239                 goto out_free_cpumask;
1240
1241         ret = 0;
1242 out_free_cpumask:
1243         free_cpumask_var(cpumask);
1244 out_put_device:
1245         ublk_put_device(ub);
1246         return ret;
1247 }
1248
1249 static inline void ublk_dump_dev_info(struct ublksrv_ctrl_dev_info *info)
1250 {
1251         pr_devel("%s: dev id %d flags %llx\n", __func__,
1252                         info->dev_id, info->flags);
1253         pr_devel("\t nr_hw_queues %d queue_depth %d block size %d dev_capacity %lld\n",
1254                         info->nr_hw_queues, info->queue_depth,
1255                         info->block_size, info->dev_blocks);
1256 }
1257
1258 static int ublk_ctrl_add_dev(struct io_uring_cmd *cmd)
1259 {
1260         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1261         void __user *argp = (void __user *)(unsigned long)header->addr;
1262         struct ublksrv_ctrl_dev_info info;
1263         struct ublk_device *ub;
1264         int ret = -EINVAL;
1265
1266         if (header->len < sizeof(info) || !header->addr)
1267                 return -EINVAL;
1268         if (header->queue_id != (u16)-1) {
1269                 pr_warn("%s: queue_id is wrong %x\n",
1270                         __func__, header->queue_id);
1271                 return -EINVAL;
1272         }
1273         if (copy_from_user(&info, argp, sizeof(info)))
1274                 return -EFAULT;
1275         ublk_dump_dev_info(&info);
1276         if (header->dev_id != info.dev_id) {
1277                 pr_warn("%s: dev id not match %u %u\n",
1278                         __func__, header->dev_id, info.dev_id);
1279                 return -EINVAL;
1280         }
1281
1282         ret = mutex_lock_killable(&ublk_ctl_mutex);
1283         if (ret)
1284                 return ret;
1285
1286         ret = -ENOMEM;
1287         ub = kzalloc(sizeof(*ub), GFP_KERNEL);
1288         if (!ub)
1289                 goto out_unlock;
1290         mutex_init(&ub->mutex);
1291         spin_lock_init(&ub->mm_lock);
1292         INIT_WORK(&ub->stop_work, ublk_stop_work_fn);
1293         INIT_DELAYED_WORK(&ub->monitor_work, ublk_daemon_monitor_work);
1294
1295         ret = ublk_alloc_dev_number(ub, header->dev_id);
1296         if (ret < 0)
1297                 goto out_free_ub;
1298
1299         memcpy(&ub->dev_info, &info, sizeof(info));
1300
1301         /* update device id */
1302         ub->dev_info.dev_id = ub->ub_number;
1303
1304         /*
1305          * 64bit flags will be copied back to userspace as feature
1306          * negotiation result, so have to clear flags which driver
1307          * doesn't support yet, then userspace can get correct flags
1308          * (features) to handle.
1309          */
1310         ub->dev_info.flags &= UBLK_F_ALL;
1311
1312         /* We are not ready to support zero copy */
1313         ub->dev_info.flags &= ~UBLK_F_SUPPORT_ZERO_COPY;
1314
1315         ub->bs_shift = ilog2(ub->dev_info.block_size);
1316         ub->dev_info.nr_hw_queues = min_t(unsigned int,
1317                         ub->dev_info.nr_hw_queues, nr_cpu_ids);
1318         ublk_align_max_io_size(ub);
1319
1320         ret = ublk_init_queues(ub);
1321         if (ret)
1322                 goto out_free_dev_number;
1323
1324         ret = ublk_add_tag_set(ub);
1325         if (ret)
1326                 goto out_deinit_queues;
1327
1328         ret = -EFAULT;
1329         if (copy_to_user(argp, &ub->dev_info, sizeof(info)))
1330                 goto out_free_tag_set;
1331
1332         /*
1333          * Add the char dev so that ublksrv daemon can be setup.
1334          * ublk_add_chdev() will cleanup everything if it fails.
1335          */
1336         ret = ublk_add_chdev(ub);
1337         goto out_unlock;
1338
1339 out_free_tag_set:
1340         blk_mq_free_tag_set(&ub->tag_set);
1341 out_deinit_queues:
1342         ublk_deinit_queues(ub);
1343 out_free_dev_number:
1344         ublk_free_dev_number(ub);
1345 out_free_ub:
1346         mutex_destroy(&ub->mutex);
1347         kfree(ub);
1348 out_unlock:
1349         mutex_unlock(&ublk_ctl_mutex);
1350         return ret;
1351 }
1352
1353 static inline bool ublk_idr_freed(int id)
1354 {
1355         void *ptr;
1356
1357         spin_lock(&ublk_idr_lock);
1358         ptr = idr_find(&ublk_index_idr, id);
1359         spin_unlock(&ublk_idr_lock);
1360
1361         return ptr == NULL;
1362 }
1363
1364 static int ublk_ctrl_del_dev(int idx)
1365 {
1366         struct ublk_device *ub;
1367         int ret;
1368
1369         ret = mutex_lock_killable(&ublk_ctl_mutex);
1370         if (ret)
1371                 return ret;
1372
1373         ub = ublk_get_device_from_id(idx);
1374         if (ub) {
1375                 ublk_remove(ub);
1376                 ublk_put_device(ub);
1377                 ret = 0;
1378         } else {
1379                 ret = -ENODEV;
1380         }
1381
1382         /*
1383          * Wait until the idr is removed, then it can be reused after
1384          * DEL_DEV command is returned.
1385          */
1386         if (!ret)
1387                 wait_event(ublk_idr_wq, ublk_idr_freed(idx));
1388         mutex_unlock(&ublk_ctl_mutex);
1389
1390         return ret;
1391 }
1392
1393 static inline void ublk_ctrl_cmd_dump(struct io_uring_cmd *cmd)
1394 {
1395         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1396
1397         pr_devel("%s: cmd_op %x, dev id %d qid %d data %llx buf %llx len %u\n",
1398                         __func__, cmd->cmd_op, header->dev_id, header->queue_id,
1399                         header->data[0], header->addr, header->len);
1400 }
1401
1402 static int ublk_ctrl_stop_dev(struct io_uring_cmd *cmd)
1403 {
1404         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1405         struct ublk_device *ub;
1406
1407         ub = ublk_get_device_from_id(header->dev_id);
1408         if (!ub)
1409                 return -EINVAL;
1410
1411         ublk_stop_dev(ub);
1412         cancel_work_sync(&ub->stop_work);
1413
1414         ublk_put_device(ub);
1415         return 0;
1416 }
1417
1418 static int ublk_ctrl_get_dev_info(struct io_uring_cmd *cmd)
1419 {
1420         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1421         void __user *argp = (void __user *)(unsigned long)header->addr;
1422         struct ublk_device *ub;
1423         int ret = 0;
1424
1425         if (header->len < sizeof(struct ublksrv_ctrl_dev_info) || !header->addr)
1426                 return -EINVAL;
1427
1428         ub = ublk_get_device_from_id(header->dev_id);
1429         if (!ub)
1430                 return -EINVAL;
1431
1432         if (copy_to_user(argp, &ub->dev_info, sizeof(ub->dev_info)))
1433                 ret = -EFAULT;
1434         ublk_put_device(ub);
1435
1436         return ret;
1437 }
1438
1439 static int ublk_ctrl_uring_cmd(struct io_uring_cmd *cmd,
1440                 unsigned int issue_flags)
1441 {
1442         struct ublksrv_ctrl_cmd *header = (struct ublksrv_ctrl_cmd *)cmd->cmd;
1443         int ret = -EINVAL;
1444
1445         ublk_ctrl_cmd_dump(cmd);
1446
1447         if (!(issue_flags & IO_URING_F_SQE128))
1448                 goto out;
1449
1450         ret = -EPERM;
1451         if (!capable(CAP_SYS_ADMIN))
1452                 goto out;
1453
1454         ret = -ENODEV;
1455         switch (cmd->cmd_op) {
1456         case UBLK_CMD_START_DEV:
1457                 ret = ublk_ctrl_start_dev(cmd);
1458                 break;
1459         case UBLK_CMD_STOP_DEV:
1460                 ret = ublk_ctrl_stop_dev(cmd);
1461                 break;
1462         case UBLK_CMD_GET_DEV_INFO:
1463                 ret = ublk_ctrl_get_dev_info(cmd);
1464                 break;
1465         case UBLK_CMD_ADD_DEV:
1466                 ret = ublk_ctrl_add_dev(cmd);
1467                 break;
1468         case UBLK_CMD_DEL_DEV:
1469                 ret = ublk_ctrl_del_dev(header->dev_id);
1470                 break;
1471         case UBLK_CMD_GET_QUEUE_AFFINITY:
1472                 ret = ublk_ctrl_get_queue_affinity(cmd);
1473                 break;
1474         default:
1475                 break;
1476         }
1477  out:
1478         io_uring_cmd_done(cmd, ret, 0);
1479         pr_devel("%s: cmd done ret %d cmd_op %x, dev id %d qid %d\n",
1480                         __func__, ret, cmd->cmd_op, header->dev_id, header->queue_id);
1481         return -EIOCBQUEUED;
1482 }
1483
1484 static const struct file_operations ublk_ctl_fops = {
1485         .open           = nonseekable_open,
1486         .uring_cmd      = ublk_ctrl_uring_cmd,
1487         .owner          = THIS_MODULE,
1488         .llseek         = noop_llseek,
1489 };
1490
1491 static struct miscdevice ublk_misc = {
1492         .minor          = MISC_DYNAMIC_MINOR,
1493         .name           = "ublk-control",
1494         .fops           = &ublk_ctl_fops,
1495 };
1496
1497 static int __init ublk_init(void)
1498 {
1499         int ret;
1500
1501         init_waitqueue_head(&ublk_idr_wq);
1502
1503         ret = misc_register(&ublk_misc);
1504         if (ret)
1505                 return ret;
1506
1507         ret = alloc_chrdev_region(&ublk_chr_devt, 0, UBLK_MINORS, "ublk-char");
1508         if (ret)
1509                 goto unregister_mis;
1510
1511         ublk_chr_class = class_create(THIS_MODULE, "ublk-char");
1512         if (IS_ERR(ublk_chr_class)) {
1513                 ret = PTR_ERR(ublk_chr_class);
1514                 goto free_chrdev_region;
1515         }
1516         return 0;
1517
1518 free_chrdev_region:
1519         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
1520 unregister_mis:
1521         misc_deregister(&ublk_misc);
1522         return ret;
1523 }
1524
1525 static void __exit ublk_exit(void)
1526 {
1527         struct ublk_device *ub;
1528         int id;
1529
1530         class_destroy(ublk_chr_class);
1531
1532         misc_deregister(&ublk_misc);
1533
1534         idr_for_each_entry(&ublk_index_idr, ub, id)
1535                 ublk_remove(ub);
1536
1537         idr_destroy(&ublk_index_idr);
1538         unregister_chrdev_region(ublk_chr_devt, UBLK_MINORS);
1539 }
1540
1541 module_init(ublk_init);
1542 module_exit(ublk_exit);
1543
1544 MODULE_AUTHOR("Ming Lei <ming.lei@redhat.com>");
1545 MODULE_LICENSE("GPL");