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