io_uring/rsrc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/kernel.h>
   3 #include <linux/errno.h>
   4 #include <linux/fs.h>
   5 #include <linux/file.h>
   6 #include <linux/mm.h>
   7 #include <linux/slab.h>
   8 #include <linux/nospec.h>
   9 #include <linux/hugetlb.h>
  10 #include <linux/compat.h>
  11 #include <linux/io_uring.h>
  12
  13 #include <uapi/linux/io_uring.h>
  14
  15 #include "io_uring.h"
  16 #include "openclose.h"
  17 #include "rsrc.h"
  18
  19 struct io_rsrc_update {
  20         struct file                     *file;
  21         u64                             arg;
  22         u32                             nr_args;
  23         u32                             offset;
  24 };
  25
  26 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
  27                                   struct io_mapped_ubuf **pimu,
  28                                   struct page **last_hpage);
  29
  30 #define IO_RSRC_REF_BATCH       100
  31
  32 /* only define max */
  33 #define IORING_MAX_FIXED_FILES  (1U << 20)
  34 #define IORING_MAX_REG_BUFFERS  (1U << 14)
  35
  36 void io_rsrc_refs_drop(struct io_ring_ctx *ctx)
  37         __must_hold(&ctx->uring_lock)
  38 {
  39         if (ctx->rsrc_cached_refs) {
  40                 io_rsrc_put_node(ctx->rsrc_node, ctx->rsrc_cached_refs);
  41                 ctx->rsrc_cached_refs = 0;
  42         }
  43 }
  44
  45 int __io_account_mem(struct user_struct *user, unsigned long nr_pages)
  46 {
  47         unsigned long page_limit, cur_pages, new_pages;
  48
  49         if (!nr_pages)
  50                 return 0;
  51
  52         /* Don't allow more pages than we can safely lock */
  53         page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
  54
  55         cur_pages = atomic_long_read(&user->locked_vm);
  56         do {
  57                 new_pages = cur_pages + nr_pages;
  58                 if (new_pages > page_limit)
  59                         return -ENOMEM;
  60         } while (!atomic_long_try_cmpxchg(&user->locked_vm,
  61                                           &cur_pages, new_pages));
  62         return 0;
  63 }
  64
  65 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
  66 {
  67         if (ctx->user)
  68                 __io_unaccount_mem(ctx->user, nr_pages);
  69
  70         if (ctx->mm_account)
  71                 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
  72 }
  73
  74 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
  75 {
  76         int ret;
  77
  78         if (ctx->user) {
  79                 ret = __io_account_mem(ctx->user, nr_pages);
  80                 if (ret)
  81                         return ret;
  82         }
  83
  84         if (ctx->mm_account)
  85                 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
  86
  87         return 0;
  88 }
  89
  90 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
  91                        void __user *arg, unsigned index)
  92 {
  93         struct iovec __user *src;
  94
  95 #ifdef CONFIG_COMPAT
  96         if (ctx->compat) {
  97                 struct compat_iovec __user *ciovs;
  98                 struct compat_iovec ciov;
  99
 100                 ciovs = (struct compat_iovec __user *) arg;
 101                 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
 102                         return -EFAULT;
 103
 104                 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
 105                 dst->iov_len = ciov.iov_len;
 106                 return 0;
 107         }
 108 #endif
 109         src = (struct iovec __user *) arg;
 110         if (copy_from_user(dst, &src[index], sizeof(*dst)))
 111                 return -EFAULT;
 112         return 0;
 113 }
 114
 115 static int io_buffer_validate(struct iovec *iov)
 116 {
 117         unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
 118
 119         /*
 120          * Don't impose further limits on the size and buffer
 121          * constraints here, we'll -EINVAL later when IO is
 122          * submitted if they are wrong.
 123          */
 124         if (!iov->iov_base)
 125                 return iov->iov_len ? -EFAULT : 0;
 126         if (!iov->iov_len)
 127                 return -EFAULT;
 128
 129         /* arbitrary limit, but we need something */
 130         if (iov->iov_len > SZ_1G)
 131                 return -EFAULT;
 132
 133         if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
 134                 return -EOVERFLOW;
 135
 136         return 0;
 137 }
 138
 139 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
 140 {
 141         struct io_mapped_ubuf *imu = *slot;
 142         unsigned int i;
 143
 144         if (imu != ctx->dummy_ubuf) {
 145                 for (i = 0; i < imu->nr_bvecs; i++)
 146                         unpin_user_page(imu->bvec[i].bv_page);
 147                 if (imu->acct_pages)
 148                         io_unaccount_mem(ctx, imu->acct_pages);
 149                 kvfree(imu);
 150         }
 151         *slot = NULL;
 152 }
 153
 154 void io_rsrc_refs_refill(struct io_ring_ctx *ctx)
 155         __must_hold(&ctx->uring_lock)
 156 {
 157         ctx->rsrc_cached_refs += IO_RSRC_REF_BATCH;
 158         percpu_ref_get_many(&ctx->rsrc_node->refs, IO_RSRC_REF_BATCH);
 159 }
 160
 161 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
 162 {
 163         struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
 164         struct io_ring_ctx *ctx = rsrc_data->ctx;
 165         struct io_rsrc_put *prsrc, *tmp;
 166
 167         list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
 168                 list_del(&prsrc->list);
 169
 170                 if (prsrc->tag) {
 171                         if (ctx->flags & IORING_SETUP_IOPOLL) {
 172                                 mutex_lock(&ctx->uring_lock);
 173                                 io_post_aux_cqe(ctx, prsrc->tag, 0, 0, true);
 174                                 mutex_unlock(&ctx->uring_lock);
 175                         } else {
 176                                 io_post_aux_cqe(ctx, prsrc->tag, 0, 0, true);
 177                         }
 178                 }
 179
 180                 rsrc_data->do_put(ctx, prsrc);
 181                 kfree(prsrc);
 182         }
 183
 184         io_rsrc_node_destroy(ref_node);
 185         if (atomic_dec_and_test(&rsrc_data->refs))
 186                 complete(&rsrc_data->done);
 187 }
 188
 189 void io_rsrc_put_work(struct work_struct *work)
 190 {
 191         struct io_ring_ctx *ctx;
 192         struct llist_node *node;
 193
 194         ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
 195         node = llist_del_all(&ctx->rsrc_put_llist);
 196
 197         while (node) {
 198                 struct io_rsrc_node *ref_node;
 199                 struct llist_node *next = node->next;
 200
 201                 ref_node = llist_entry(node, struct io_rsrc_node, llist);
 202                 __io_rsrc_put_work(ref_node);
 203                 node = next;
 204         }
 205 }
 206
 207 void io_wait_rsrc_data(struct io_rsrc_data *data)
 208 {
 209         if (data && !atomic_dec_and_test(&data->refs))
 210                 wait_for_completion(&data->done);
 211 }
 212
 213 void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
 214 {
 215         percpu_ref_exit(&ref_node->refs);
 216         kfree(ref_node);
 217 }
 218
 219 static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
 220 {
 221         struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
 222         struct io_ring_ctx *ctx = node->rsrc_data->ctx;
 223         unsigned long flags;
 224         bool first_add = false;
 225         unsigned long delay = HZ;
 226
 227         spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
 228         node->done = true;
 229
 230         /* if we are mid-quiesce then do not delay */
 231         if (node->rsrc_data->quiesce)
 232                 delay = 0;
 233
 234         while (!list_empty(&ctx->rsrc_ref_list)) {
 235                 node = list_first_entry(&ctx->rsrc_ref_list,
 236                                             struct io_rsrc_node, node);
 237                 /* recycle ref nodes in order */
 238                 if (!node->done)
 239                         break;
 240                 list_del(&node->node);
 241                 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
 242         }
 243         spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
 244
 245         if (first_add)
 246                 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
 247 }
 248
 249 static struct io_rsrc_node *io_rsrc_node_alloc(void)
 250 {
 251         struct io_rsrc_node *ref_node;
 252
 253         ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
 254         if (!ref_node)
 255                 return NULL;
 256
 257         if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
 258                             0, GFP_KERNEL)) {
 259                 kfree(ref_node);
 260                 return NULL;
 261         }
 262         INIT_LIST_HEAD(&ref_node->node);
 263         INIT_LIST_HEAD(&ref_node->rsrc_list);
 264         ref_node->done = false;
 265         return ref_node;
 266 }
 267
 268 void io_rsrc_node_switch(struct io_ring_ctx *ctx,
 269                          struct io_rsrc_data *data_to_kill)
 270         __must_hold(&ctx->uring_lock)
 271 {
 272         WARN_ON_ONCE(!ctx->rsrc_backup_node);
 273         WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
 274
 275         io_rsrc_refs_drop(ctx);
 276
 277         if (data_to_kill) {
 278                 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
 279
 280                 rsrc_node->rsrc_data = data_to_kill;
 281                 spin_lock_irq(&ctx->rsrc_ref_lock);
 282                 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
 283                 spin_unlock_irq(&ctx->rsrc_ref_lock);
 284
 285                 atomic_inc(&data_to_kill->refs);
 286                 percpu_ref_kill(&rsrc_node->refs);
 287                 ctx->rsrc_node = NULL;
 288         }
 289
 290         if (!ctx->rsrc_node) {
 291                 ctx->rsrc_node = ctx->rsrc_backup_node;
 292                 ctx->rsrc_backup_node = NULL;
 293         }
 294 }
 295
 296 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
 297 {
 298         if (ctx->rsrc_backup_node)
 299                 return 0;
 300         ctx->rsrc_backup_node = io_rsrc_node_alloc();
 301         return ctx->rsrc_backup_node ? 0 : -ENOMEM;
 302 }
 303
 304 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
 305                                       struct io_ring_ctx *ctx)
 306 {
 307         int ret;
 308
 309         /* As we may drop ->uring_lock, other task may have started quiesce */
 310         if (data->quiesce)
 311                 return -ENXIO;
 312
 313         data->quiesce = true;
 314         do {
 315                 ret = io_rsrc_node_switch_start(ctx);
 316                 if (ret)
 317                         break;
 318                 io_rsrc_node_switch(ctx, data);
 319
 320                 /* kill initial ref, already quiesced if zero */
 321                 if (atomic_dec_and_test(&data->refs))
 322                         break;
 323                 mutex_unlock(&ctx->uring_lock);
 324                 flush_delayed_work(&ctx->rsrc_put_work);
 325                 ret = wait_for_completion_interruptible(&data->done);
 326                 if (!ret) {
 327                         mutex_lock(&ctx->uring_lock);
 328                         if (atomic_read(&data->refs) > 0) {
 329                                 /*
 330                                  * it has been revived by another thread while
 331                                  * we were unlocked
 332                                  */
 333                                 mutex_unlock(&ctx->uring_lock);
 334                         } else {
 335                                 break;
 336                         }
 337                 }
 338
 339                 atomic_inc(&data->refs);
 340                 /* wait for all works potentially completing data->done */
 341                 flush_delayed_work(&ctx->rsrc_put_work);
 342                 reinit_completion(&data->done);
 343
 344                 ret = io_run_task_work_sig();
 345                 mutex_lock(&ctx->uring_lock);
 346         } while (ret >= 0);
 347         data->quiesce = false;
 348
 349         return ret;
 350 }
 351
 352 static void io_free_page_table(void **table, size_t size)
 353 {
 354         unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
 355
 356         for (i = 0; i < nr_tables; i++)
 357                 kfree(table[i]);
 358         kfree(table);
 359 }
 360
 361 static void io_rsrc_data_free(struct io_rsrc_data *data)
 362 {
 363         size_t size = data->nr * sizeof(data->tags[0][0]);
 364
 365         if (data->tags)
 366                 io_free_page_table((void **)data->tags, size);
 367         kfree(data);
 368 }
 369
 370 static __cold void **io_alloc_page_table(size_t size)
 371 {
 372         unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
 373         size_t init_size = size;
 374         void **table;
 375
 376         table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
 377         if (!table)
 378                 return NULL;
 379
 380         for (i = 0; i < nr_tables; i++) {
 381                 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
 382
 383                 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
 384                 if (!table[i]) {
 385                         io_free_page_table(table, init_size);
 386                         return NULL;
 387                 }
 388                 size -= this_size;
 389         }
 390         return table;
 391 }
 392
 393 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx,
 394                                      rsrc_put_fn *do_put, u64 __user *utags,
 395                                      unsigned nr, struct io_rsrc_data **pdata)
 396 {
 397         struct io_rsrc_data *data;
 398         int ret = -ENOMEM;
 399         unsigned i;
 400
 401         data = kzalloc(sizeof(*data), GFP_KERNEL);
 402         if (!data)
 403                 return -ENOMEM;
 404         data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
 405         if (!data->tags) {
 406                 kfree(data);
 407                 return -ENOMEM;
 408         }
 409
 410         data->nr = nr;
 411         data->ctx = ctx;
 412         data->do_put = do_put;
 413         if (utags) {
 414                 ret = -EFAULT;
 415                 for (i = 0; i < nr; i++) {
 416                         u64 *tag_slot = io_get_tag_slot(data, i);
 417
 418                         if (copy_from_user(tag_slot, &utags[i],
 419                                            sizeof(*tag_slot)))
 420                                 goto fail;
 421                 }
 422         }
 423
 424         atomic_set(&data->refs, 1);
 425         init_completion(&data->done);
 426         *pdata = data;
 427         return 0;
 428 fail:
 429         io_rsrc_data_free(data);
 430         return ret;
 431 }
 432
 433 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
 434                                  struct io_uring_rsrc_update2 *up,
 435                                  unsigned nr_args)
 436 {
 437         u64 __user *tags = u64_to_user_ptr(up->tags);
 438         __s32 __user *fds = u64_to_user_ptr(up->data);
 439         struct io_rsrc_data *data = ctx->file_data;
 440         struct io_fixed_file *file_slot;
 441         struct file *file;
 442         int fd, i, err = 0;
 443         unsigned int done;
 444         bool needs_switch = false;
 445
 446         if (!ctx->file_data)
 447                 return -ENXIO;
 448         if (up->offset + nr_args > ctx->nr_user_files)
 449                 return -EINVAL;
 450
 451         for (done = 0; done < nr_args; done++) {
 452                 u64 tag = 0;
 453
 454                 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
 455                     copy_from_user(&fd, &fds[done], sizeof(fd))) {
 456                         err = -EFAULT;
 457                         break;
 458                 }
 459                 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
 460                         err = -EINVAL;
 461                         break;
 462                 }
 463                 if (fd == IORING_REGISTER_FILES_SKIP)
 464                         continue;
 465
 466                 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
 467                 file_slot = io_fixed_file_slot(&ctx->file_table, i);
 468
 469                 if (file_slot->file_ptr) {
 470                         file = (struct file *)(file_slot->file_ptr & FFS_MASK);
 471                         err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file);
 472                         if (err)
 473                                 break;
 474                         file_slot->file_ptr = 0;
 475                         io_file_bitmap_clear(&ctx->file_table, i);
 476                         needs_switch = true;
 477                 }
 478                 if (fd != -1) {
 479                         file = fget(fd);
 480                         if (!file) {
 481                                 err = -EBADF;
 482                                 break;
 483                         }
 484                         /*
 485                          * Don't allow io_uring instances to be registered. If
 486                          * UNIX isn't enabled, then this causes a reference
 487                          * cycle and this instance can never get freed. If UNIX
 488                          * is enabled we'll handle it just fine, but there's
 489                          * still no point in allowing a ring fd as it doesn't
 490                          * support regular read/write anyway.
 491                          */
 492                         if (io_is_uring_fops(file)) {
 493                                 fput(file);
 494                                 err = -EBADF;
 495                                 break;
 496                         }
 497                         err = io_scm_file_account(ctx, file);
 498                         if (err) {
 499                                 fput(file);
 500                                 break;
 501                         }
 502                         *io_get_tag_slot(data, i) = tag;
 503                         io_fixed_file_set(file_slot, file);
 504                         io_file_bitmap_set(&ctx->file_table, i);
 505                 }
 506         }
 507
 508         if (needs_switch)
 509                 io_rsrc_node_switch(ctx, data);
 510         return done ? done : err;
 511 }
 512
 513 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
 514                                    struct io_uring_rsrc_update2 *up,
 515                                    unsigned int nr_args)
 516 {
 517         u64 __user *tags = u64_to_user_ptr(up->tags);
 518         struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
 519         struct page *last_hpage = NULL;
 520         bool needs_switch = false;
 521         __u32 done;
 522         int i, err;
 523
 524         if (!ctx->buf_data)
 525                 return -ENXIO;
 526         if (up->offset + nr_args > ctx->nr_user_bufs)
 527                 return -EINVAL;
 528
 529         for (done = 0; done < nr_args; done++) {
 530                 struct io_mapped_ubuf *imu;
 531                 int offset = up->offset + done;
 532                 u64 tag = 0;
 533
 534                 err = io_copy_iov(ctx, &iov, iovs, done);
 535                 if (err)
 536                         break;
 537                 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
 538                         err = -EFAULT;
 539                         break;
 540                 }
 541                 err = io_buffer_validate(&iov);
 542                 if (err)
 543                         break;
 544                 if (!iov.iov_base && tag) {
 545                         err = -EINVAL;
 546                         break;
 547                 }
 548                 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
 549                 if (err)
 550                         break;
 551
 552                 i = array_index_nospec(offset, ctx->nr_user_bufs);
 553                 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
 554                         err = io_queue_rsrc_removal(ctx->buf_data, i,
 555                                                     ctx->rsrc_node, ctx->user_bufs[i]);
 556                         if (unlikely(err)) {
 557                                 io_buffer_unmap(ctx, &imu);
 558                                 break;
 559                         }
 560                         ctx->user_bufs[i] = ctx->dummy_ubuf;
 561                         needs_switch = true;
 562                 }
 563
 564                 ctx->user_bufs[i] = imu;
 565                 *io_get_tag_slot(ctx->buf_data, offset) = tag;
 566         }
 567
 568         if (needs_switch)
 569                 io_rsrc_node_switch(ctx, ctx->buf_data);
 570         return done ? done : err;
 571 }
 572
 573 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
 574                                      struct io_uring_rsrc_update2 *up,
 575                                      unsigned nr_args)
 576 {
 577         __u32 tmp;
 578         int err;
 579
 580         if (check_add_overflow(up->offset, nr_args, &tmp))
 581                 return -EOVERFLOW;
 582         err = io_rsrc_node_switch_start(ctx);
 583         if (err)
 584                 return err;
 585
 586         switch (type) {
 587         case IORING_RSRC_FILE:
 588                 return __io_sqe_files_update(ctx, up, nr_args);
 589         case IORING_RSRC_BUFFER:
 590                 return __io_sqe_buffers_update(ctx, up, nr_args);
 591         }
 592         return -EINVAL;
 593 }
 594
 595 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
 596                              unsigned nr_args)
 597 {
 598         struct io_uring_rsrc_update2 up;
 599
 600         if (!nr_args)
 601                 return -EINVAL;
 602         memset(&up, 0, sizeof(up));
 603         if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
 604                 return -EFAULT;
 605         if (up.resv || up.resv2)
 606                 return -EINVAL;
 607         return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
 608 }
 609
 610 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
 611                             unsigned size, unsigned type)
 612 {
 613         struct io_uring_rsrc_update2 up;
 614
 615         if (size != sizeof(up))
 616                 return -EINVAL;
 617         if (copy_from_user(&up, arg, sizeof(up)))
 618                 return -EFAULT;
 619         if (!up.nr || up.resv || up.resv2)
 620                 return -EINVAL;
 621         return __io_register_rsrc_update(ctx, type, &up, up.nr);
 622 }
 623
 624 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
 625                             unsigned int size, unsigned int type)
 626 {
 627         struct io_uring_rsrc_register rr;
 628
 629         /* keep it extendible */
 630         if (size != sizeof(rr))
 631                 return -EINVAL;
 632
 633         memset(&rr, 0, sizeof(rr));
 634         if (copy_from_user(&rr, arg, size))
 635                 return -EFAULT;
 636         if (!rr.nr || rr.resv2)
 637                 return -EINVAL;
 638         if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE)
 639                 return -EINVAL;
 640
 641         switch (type) {
 642         case IORING_RSRC_FILE:
 643                 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
 644                         break;
 645                 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
 646                                              rr.nr, u64_to_user_ptr(rr.tags));
 647         case IORING_RSRC_BUFFER:
 648                 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data)
 649                         break;
 650                 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
 651                                                rr.nr, u64_to_user_ptr(rr.tags));
 652         }
 653         return -EINVAL;
 654 }
 655
 656 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 657 {
 658         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 659
 660         if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
 661                 return -EINVAL;
 662         if (sqe->rw_flags || sqe->splice_fd_in)
 663                 return -EINVAL;
 664
 665         up->offset = READ_ONCE(sqe->off);
 666         up->nr_args = READ_ONCE(sqe->len);
 667         if (!up->nr_args)
 668                 return -EINVAL;
 669         up->arg = READ_ONCE(sqe->addr);
 670         return 0;
 671 }
 672
 673 static int io_files_update_with_index_alloc(struct io_kiocb *req,
 674                                             unsigned int issue_flags)
 675 {
 676         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 677         __s32 __user *fds = u64_to_user_ptr(up->arg);
 678         unsigned int done;
 679         struct file *file;
 680         int ret, fd;
 681
 682         if (!req->ctx->file_data)
 683                 return -ENXIO;
 684
 685         for (done = 0; done < up->nr_args; done++) {
 686                 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
 687                         ret = -EFAULT;
 688                         break;
 689                 }
 690
 691                 file = fget(fd);
 692                 if (!file) {
 693                         ret = -EBADF;
 694                         break;
 695                 }
 696                 ret = io_fixed_fd_install(req, issue_flags, file,
 697                                           IORING_FILE_INDEX_ALLOC);
 698                 if (ret < 0)
 699                         break;
 700                 if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
 701                         __io_close_fixed(req->ctx, issue_flags, ret);
 702                         ret = -EFAULT;
 703                         break;
 704                 }
 705         }
 706
 707         if (done)
 708                 return done;
 709         return ret;
 710 }
 711
 712 int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
 713 {
 714         struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update);
 715         struct io_ring_ctx *ctx = req->ctx;
 716         struct io_uring_rsrc_update2 up2;
 717         int ret;
 718
 719         up2.offset = up->offset;
 720         up2.data = up->arg;
 721         up2.nr = 0;
 722         up2.tags = 0;
 723         up2.resv = 0;
 724         up2.resv2 = 0;
 725
 726         if (up->offset == IORING_FILE_INDEX_ALLOC) {
 727                 ret = io_files_update_with_index_alloc(req, issue_flags);
 728         } else {
 729                 io_ring_submit_lock(ctx, issue_flags);
 730                 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
 731                                                 &up2, up->nr_args);
 732                 io_ring_submit_unlock(ctx, issue_flags);
 733         }
 734
 735         if (ret < 0)
 736                 req_set_fail(req);
 737         io_req_set_res(req, ret, 0);
 738         return IOU_OK;
 739 }
 740
 741 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
 742                           struct io_rsrc_node *node, void *rsrc)
 743 {
 744         u64 *tag_slot = io_get_tag_slot(data, idx);
 745         struct io_rsrc_put *prsrc;
 746
 747         prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
 748         if (!prsrc)
 749                 return -ENOMEM;
 750
 751         prsrc->tag = *tag_slot;
 752         *tag_slot = 0;
 753         prsrc->rsrc = rsrc;
 754         list_add(&prsrc->list, &node->rsrc_list);
 755         return 0;
 756 }
 757
 758 void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
 759 {
 760 #if !defined(IO_URING_SCM_ALL)
 761         int i;
 762
 763         for (i = 0; i < ctx->nr_user_files; i++) {
 764                 struct file *file = io_file_from_index(&ctx->file_table, i);
 765
 766                 if (!file)
 767                         continue;
 768                 if (io_fixed_file_slot(&ctx->file_table, i)->file_ptr & FFS_SCM)
 769                         continue;
 770                 io_file_bitmap_clear(&ctx->file_table, i);
 771                 fput(file);
 772         }
 773 #endif
 774
 775 #if defined(CONFIG_UNIX)
 776         if (ctx->ring_sock) {
 777                 struct sock *sock = ctx->ring_sock->sk;
 778                 struct sk_buff *skb;
 779
 780                 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
 781                         kfree_skb(skb);
 782         }
 783 #endif
 784         io_free_file_tables(&ctx->file_table);
 785         io_rsrc_data_free(ctx->file_data);
 786         ctx->file_data = NULL;
 787         ctx->nr_user_files = 0;
 788 }
 789
 790 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
 791 {
 792         unsigned nr = ctx->nr_user_files;
 793         int ret;
 794
 795         if (!ctx->file_data)
 796                 return -ENXIO;
 797
 798         /*
 799          * Quiesce may unlock ->uring_lock, and while it's not held
 800          * prevent new requests using the table.
 801          */
 802         ctx->nr_user_files = 0;
 803         ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
 804         ctx->nr_user_files = nr;
 805         if (!ret)
 806                 __io_sqe_files_unregister(ctx);
 807         return ret;
 808 }
 809
 810 /*
 811  * Ensure the UNIX gc is aware of our file set, so we are certain that
 812  * the io_uring can be safely unregistered on process exit, even if we have
 813  * loops in the file referencing. We account only files that can hold other
 814  * files because otherwise they can't form a loop and so are not interesting
 815  * for GC.
 816  */
 817 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
 818 {
 819 #if defined(CONFIG_UNIX)
 820         struct sock *sk = ctx->ring_sock->sk;
 821         struct sk_buff_head *head = &sk->sk_receive_queue;
 822         struct scm_fp_list *fpl;
 823         struct sk_buff *skb;
 824
 825         if (likely(!io_file_need_scm(file)))
 826                 return 0;
 827
 828         /*
 829          * See if we can merge this file into an existing skb SCM_RIGHTS
 830          * file set. If there's no room, fall back to allocating a new skb
 831          * and filling it in.
 832          */
 833         spin_lock_irq(&head->lock);
 834         skb = skb_peek(head);
 835         if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
 836                 __skb_unlink(skb, head);
 837         else
 838                 skb = NULL;
 839         spin_unlock_irq(&head->lock);
 840
 841         if (!skb) {
 842                 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
 843                 if (!fpl)
 844                         return -ENOMEM;
 845
 846                 skb = alloc_skb(0, GFP_KERNEL);
 847                 if (!skb) {
 848                         kfree(fpl);
 849                         return -ENOMEM;
 850                 }
 851
 852                 fpl->user = get_uid(current_user());
 853                 fpl->max = SCM_MAX_FD;
 854                 fpl->count = 0;
 855
 856                 UNIXCB(skb).fp = fpl;
 857                 skb->sk = sk;
 858                 skb->destructor = unix_destruct_scm;
 859                 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
 860         }
 861
 862         fpl = UNIXCB(skb).fp;
 863         fpl->fp[fpl->count++] = get_file(file);
 864         unix_inflight(fpl->user, file);
 865         skb_queue_head(head, skb);
 866         fput(file);
 867 #endif
 868         return 0;
 869 }
 870
 871 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
 872 {
 873         struct file *file = prsrc->file;
 874 #if defined(CONFIG_UNIX)
 875         struct sock *sock = ctx->ring_sock->sk;
 876         struct sk_buff_head list, *head = &sock->sk_receive_queue;
 877         struct sk_buff *skb;
 878         int i;
 879
 880         if (!io_file_need_scm(file)) {
 881                 fput(file);
 882                 return;
 883         }
 884
 885         __skb_queue_head_init(&list);
 886
 887         /*
 888          * Find the skb that holds this file in its SCM_RIGHTS. When found,
 889          * remove this entry and rearrange the file array.
 890          */
 891         skb = skb_dequeue(head);
 892         while (skb) {
 893                 struct scm_fp_list *fp;
 894
 895                 fp = UNIXCB(skb).fp;
 896                 for (i = 0; i < fp->count; i++) {
 897                         int left;
 898
 899                         if (fp->fp[i] != file)
 900                                 continue;
 901
 902                         unix_notinflight(fp->user, fp->fp[i]);
 903                         left = fp->count - 1 - i;
 904                         if (left) {
 905                                 memmove(&fp->fp[i], &fp->fp[i + 1],
 906                                                 left * sizeof(struct file *));
 907                         }
 908                         fp->count--;
 909                         if (!fp->count) {
 910                                 kfree_skb(skb);
 911                                 skb = NULL;
 912                         } else {
 913                                 __skb_queue_tail(&list, skb);
 914                         }
 915                         fput(file);
 916                         file = NULL;
 917                         break;
 918                 }
 919
 920                 if (!file)
 921                         break;
 922
 923                 __skb_queue_tail(&list, skb);
 924
 925                 skb = skb_dequeue(head);
 926         }
 927
 928         if (skb_peek(&list)) {
 929                 spin_lock_irq(&head->lock);
 930                 while ((skb = __skb_dequeue(&list)) != NULL)
 931                         __skb_queue_tail(head, skb);
 932                 spin_unlock_irq(&head->lock);
 933         }
 934 #else
 935         fput(file);
 936 #endif
 937 }
 938
 939 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
 940                           unsigned nr_args, u64 __user *tags)
 941 {
 942         __s32 __user *fds = (__s32 __user *) arg;
 943         struct file *file;
 944         int fd, ret;
 945         unsigned i;
 946
 947         if (ctx->file_data)
 948                 return -EBUSY;
 949         if (!nr_args)
 950                 return -EINVAL;
 951         if (nr_args > IORING_MAX_FIXED_FILES)
 952                 return -EMFILE;
 953         if (nr_args > rlimit(RLIMIT_NOFILE))
 954                 return -EMFILE;
 955         ret = io_rsrc_node_switch_start(ctx);
 956         if (ret)
 957                 return ret;
 958         ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
 959                                  &ctx->file_data);
 960         if (ret)
 961                 return ret;
 962
 963         if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
 964                 io_rsrc_data_free(ctx->file_data);
 965                 ctx->file_data = NULL;
 966                 return -ENOMEM;
 967         }
 968
 969         for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
 970                 struct io_fixed_file *file_slot;
 971
 972                 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
 973                         ret = -EFAULT;
 974                         goto fail;
 975                 }
 976                 /* allow sparse sets */
 977                 if (!fds || fd == -1) {
 978                         ret = -EINVAL;
 979                         if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
 980                                 goto fail;
 981                         continue;
 982                 }
 983
 984                 file = fget(fd);
 985                 ret = -EBADF;
 986                 if (unlikely(!file))
 987                         goto fail;
 988
 989                 /*
 990                  * Don't allow io_uring instances to be registered. If UNIX
 991                  * isn't enabled, then this causes a reference cycle and this
 992                  * instance can never get freed. If UNIX is enabled we'll
 993                  * handle it just fine, but there's still no point in allowing
 994                  * a ring fd as it doesn't support regular read/write anyway.
 995                  */
 996                 if (io_is_uring_fops(file)) {
 997                         fput(file);
 998                         goto fail;
 999                 }
1000                 ret = io_scm_file_account(ctx, file);
1001                 if (ret) {
1002                         fput(file);
1003                         goto fail;
1004                 }
1005                 file_slot = io_fixed_file_slot(&ctx->file_table, i);
1006                 io_fixed_file_set(file_slot, file);
1007                 io_file_bitmap_set(&ctx->file_table, i);
1008         }
1009
1010         /* default it to the whole table */
1011         io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
1012         io_rsrc_node_switch(ctx, NULL);
1013         return 0;
1014 fail:
1015         __io_sqe_files_unregister(ctx);
1016         return ret;
1017 }
1018
1019 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
1020 {
1021         io_buffer_unmap(ctx, &prsrc->buf);
1022         prsrc->buf = NULL;
1023 }
1024
1025 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1026 {
1027         unsigned int i;
1028
1029         for (i = 0; i < ctx->nr_user_bufs; i++)
1030                 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
1031         kfree(ctx->user_bufs);
1032         io_rsrc_data_free(ctx->buf_data);
1033         ctx->user_bufs = NULL;
1034         ctx->buf_data = NULL;
1035         ctx->nr_user_bufs = 0;
1036 }
1037
1038 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1039 {
1040         unsigned nr = ctx->nr_user_bufs;
1041         int ret;
1042
1043         if (!ctx->buf_data)
1044                 return -ENXIO;
1045
1046         /*
1047          * Quiesce may unlock ->uring_lock, and while it's not held
1048          * prevent new requests using the table.
1049          */
1050         ctx->nr_user_bufs = 0;
1051         ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1052         ctx->nr_user_bufs = nr;
1053         if (!ret)
1054                 __io_sqe_buffers_unregister(ctx);
1055         return ret;
1056 }
1057
1058 /*
1059  * Not super efficient, but this is just a registration time. And we do cache
1060  * the last compound head, so generally we'll only do a full search if we don't
1061  * match that one.
1062  *
1063  * We check if the given compound head page has already been accounted, to
1064  * avoid double accounting it. This allows us to account the full size of the
1065  * page, not just the constituent pages of a huge page.
1066  */
1067 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1068                                   int nr_pages, struct page *hpage)
1069 {
1070         int i, j;
1071
1072         /* check current page array */
1073         for (i = 0; i < nr_pages; i++) {
1074                 if (!PageCompound(pages[i]))
1075                         continue;
1076                 if (compound_head(pages[i]) == hpage)
1077                         return true;
1078         }
1079
1080         /* check previously registered pages */
1081         for (i = 0; i < ctx->nr_user_bufs; i++) {
1082                 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1083
1084                 for (j = 0; j < imu->nr_bvecs; j++) {
1085                         if (!PageCompound(imu->bvec[j].bv_page))
1086                                 continue;
1087                         if (compound_head(imu->bvec[j].bv_page) == hpage)
1088                                 return true;
1089                 }
1090         }
1091
1092         return false;
1093 }
1094
1095 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1096                                  int nr_pages, struct io_mapped_ubuf *imu,
1097                                  struct page **last_hpage)
1098 {
1099         int i, ret;
1100
1101         imu->acct_pages = 0;
1102         for (i = 0; i < nr_pages; i++) {
1103                 if (!PageCompound(pages[i])) {
1104                         imu->acct_pages++;
1105                 } else {
1106                         struct page *hpage;
1107
1108                         hpage = compound_head(pages[i]);
1109                         if (hpage == *last_hpage)
1110                                 continue;
1111                         *last_hpage = hpage;
1112                         if (headpage_already_acct(ctx, pages, i, hpage))
1113                                 continue;
1114                         imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1115                 }
1116         }
1117
1118         if (!imu->acct_pages)
1119                 return 0;
1120
1121         ret = io_account_mem(ctx, imu->acct_pages);
1122         if (ret)
1123                 imu->acct_pages = 0;
1124         return ret;
1125 }
1126
1127 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1128 {
1129         unsigned long start, end, nr_pages;
1130         struct vm_area_struct **vmas = NULL;
1131         struct page **pages = NULL;
1132         int i, pret, ret = -ENOMEM;
1133
1134         end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1135         start = ubuf >> PAGE_SHIFT;
1136         nr_pages = end - start;
1137
1138         pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1139         if (!pages)
1140                 goto done;
1141
1142         vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1143                               GFP_KERNEL);
1144         if (!vmas)
1145                 goto done;
1146
1147         ret = 0;
1148         mmap_read_lock(current->mm);
1149         pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1150                               pages, vmas);
1151         if (pret == nr_pages) {
1152                 /* don't support file backed memory */
1153                 for (i = 0; i < nr_pages; i++) {
1154                         struct vm_area_struct *vma = vmas[i];
1155
1156                         if (vma_is_shmem(vma))
1157                                 continue;
1158                         if (vma->vm_file &&
1159                             !is_file_hugepages(vma->vm_file)) {
1160                                 ret = -EOPNOTSUPP;
1161                                 break;
1162                         }
1163                 }
1164                 *npages = nr_pages;
1165         } else {
1166                 ret = pret < 0 ? pret : -EFAULT;
1167         }
1168         mmap_read_unlock(current->mm);
1169         if (ret) {
1170                 /*
1171                  * if we did partial map, or found file backed vmas,
1172                  * release any pages we did get
1173                  */
1174                 if (pret > 0)
1175                         unpin_user_pages(pages, pret);
1176                 goto done;
1177         }
1178         ret = 0;
1179 done:
1180         kvfree(vmas);
1181         if (ret < 0) {
1182                 kvfree(pages);
1183                 pages = ERR_PTR(ret);
1184         }
1185         return pages;
1186 }
1187
1188 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1189                                   struct io_mapped_ubuf **pimu,
1190                                   struct page **last_hpage)
1191 {
1192         struct io_mapped_ubuf *imu = NULL;
1193         struct page **pages = NULL;
1194         unsigned long off;
1195         size_t size;
1196         int ret, nr_pages, i;
1197
1198         *pimu = ctx->dummy_ubuf;
1199         if (!iov->iov_base)
1200                 return 0;
1201
1202         ret = -ENOMEM;
1203         pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1204                                 &nr_pages);
1205         if (IS_ERR(pages)) {
1206                 ret = PTR_ERR(pages);
1207                 pages = NULL;
1208                 goto done;
1209         }
1210
1211         imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1212         if (!imu)
1213                 goto done;
1214
1215         ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1216         if (ret) {
1217                 unpin_user_pages(pages, nr_pages);
1218                 goto done;
1219         }
1220
1221         off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1222         size = iov->iov_len;
1223         for (i = 0; i < nr_pages; i++) {
1224                 size_t vec_len;
1225
1226                 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1227                 imu->bvec[i].bv_page = pages[i];
1228                 imu->bvec[i].bv_len = vec_len;
1229                 imu->bvec[i].bv_offset = off;
1230                 off = 0;
1231                 size -= vec_len;
1232         }
1233         /* store original address for later verification */
1234         imu->ubuf = (unsigned long) iov->iov_base;
1235         imu->ubuf_end = imu->ubuf + iov->iov_len;
1236         imu->nr_bvecs = nr_pages;
1237         *pimu = imu;
1238         ret = 0;
1239 done:
1240         if (ret)
1241                 kvfree(imu);
1242         kvfree(pages);
1243         return ret;
1244 }
1245
1246 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1247 {
1248         ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1249         return ctx->user_bufs ? 0 : -ENOMEM;
1250 }
1251
1252 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1253                             unsigned int nr_args, u64 __user *tags)
1254 {
1255         struct page *last_hpage = NULL;
1256         struct io_rsrc_data *data;
1257         int i, ret;
1258         struct iovec iov;
1259
1260         BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1261
1262         if (ctx->user_bufs)
1263                 return -EBUSY;
1264         if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1265                 return -EINVAL;
1266         ret = io_rsrc_node_switch_start(ctx);
1267         if (ret)
1268                 return ret;
1269         ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1270         if (ret)
1271                 return ret;
1272         ret = io_buffers_map_alloc(ctx, nr_args);
1273         if (ret) {
1274                 io_rsrc_data_free(data);
1275                 return ret;
1276         }
1277
1278         for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1279                 if (arg) {
1280                         ret = io_copy_iov(ctx, &iov, arg, i);
1281                         if (ret)
1282                                 break;
1283                         ret = io_buffer_validate(&iov);
1284                         if (ret)
1285                                 break;
1286                 } else {
1287                         memset(&iov, 0, sizeof(iov));
1288                 }
1289
1290                 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1291                         ret = -EINVAL;
1292                         break;
1293                 }
1294
1295                 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1296                                              &last_hpage);
1297                 if (ret)
1298                         break;
1299         }
1300
1301         WARN_ON_ONCE(ctx->buf_data);
1302
1303         ctx->buf_data = data;
1304         if (ret)
1305                 __io_sqe_buffers_unregister(ctx);
1306         else
1307                 io_rsrc_node_switch(ctx, NULL);
1308         return ret;
1309 }
1310
1311 int io_import_fixed(int ddir, struct iov_iter *iter,
1312                            struct io_mapped_ubuf *imu,
1313                            u64 buf_addr, size_t len)
1314 {
1315         u64 buf_end;
1316         size_t offset;
1317
1318         if (WARN_ON_ONCE(!imu))
1319                 return -EFAULT;
1320         if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1321                 return -EFAULT;
1322         /* not inside the mapped region */
1323         if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1324                 return -EFAULT;
1325
1326         /*
1327          * May not be a start of buffer, set size appropriately
1328          * and advance us to the beginning.
1329          */
1330         offset = buf_addr - imu->ubuf;
1331         iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1332
1333         if (offset) {
1334                 /*
1335                  * Don't use iov_iter_advance() here, as it's really slow for
1336                  * using the latter parts of a big fixed buffer - it iterates
1337                  * over each segment manually. We can cheat a bit here, because
1338                  * we know that:
1339                  *
1340                  * 1) it's a BVEC iter, we set it up
1341                  * 2) all bvecs are PAGE_SIZE in size, except potentially the
1342                  *    first and last bvec
1343                  *
1344                  * So just find our index, and adjust the iterator afterwards.
1345                  * If the offset is within the first bvec (or the whole first
1346                  * bvec, just use iov_iter_advance(). This makes it easier
1347                  * since we can just skip the first segment, which may not
1348                  * be PAGE_SIZE aligned.
1349                  */
1350                 const struct bio_vec *bvec = imu->bvec;
1351
1352                 if (offset <= bvec->bv_len) {
1353                         iov_iter_advance(iter, offset);
1354                 } else {
1355                         unsigned long seg_skip;
1356
1357                         /* skip first vec */
1358                         offset -= bvec->bv_len;
1359                         seg_skip = 1 + (offset >> PAGE_SHIFT);
1360
1361                         iter->bvec = bvec + seg_skip;
1362                         iter->nr_segs -= seg_skip;
1363                         iter->count -= bvec->bv_len + offset;
1364                         iter->iov_offset = offset & ~PAGE_MASK;
1365                 }
1366         }
1367
1368         return 0;
1369 }