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(ctx);
 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         int i;
 761
 762         for (i = 0; i < ctx->nr_user_files; i++) {
 763                 struct file *file = io_file_from_index(&ctx->file_table, i);
 764
 765                 /* skip scm accounted files, they'll be freed by ->ring_sock */
 766                 if (!file || io_file_need_scm(file))
 767                         continue;
 768                 io_file_bitmap_clear(&ctx->file_table, i);
 769                 fput(file);
 770         }
 771
 772 #if defined(CONFIG_UNIX)
 773         if (ctx->ring_sock) {
 774                 struct sock *sock = ctx->ring_sock->sk;
 775                 struct sk_buff *skb;
 776
 777                 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
 778                         kfree_skb(skb);
 779         }
 780 #endif
 781         io_free_file_tables(&ctx->file_table);
 782         io_rsrc_data_free(ctx->file_data);
 783         ctx->file_data = NULL;
 784         ctx->nr_user_files = 0;
 785 }
 786
 787 int io_sqe_files_unregister(struct io_ring_ctx *ctx)
 788 {
 789         unsigned nr = ctx->nr_user_files;
 790         int ret;
 791
 792         if (!ctx->file_data)
 793                 return -ENXIO;
 794
 795         /*
 796          * Quiesce may unlock ->uring_lock, and while it's not held
 797          * prevent new requests using the table.
 798          */
 799         ctx->nr_user_files = 0;
 800         ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
 801         ctx->nr_user_files = nr;
 802         if (!ret)
 803                 __io_sqe_files_unregister(ctx);
 804         return ret;
 805 }
 806
 807 /*
 808  * Ensure the UNIX gc is aware of our file set, so we are certain that
 809  * the io_uring can be safely unregistered on process exit, even if we have
 810  * loops in the file referencing. We account only files that can hold other
 811  * files because otherwise they can't form a loop and so are not interesting
 812  * for GC.
 813  */
 814 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file)
 815 {
 816 #if defined(CONFIG_UNIX)
 817         struct sock *sk = ctx->ring_sock->sk;
 818         struct sk_buff_head *head = &sk->sk_receive_queue;
 819         struct scm_fp_list *fpl;
 820         struct sk_buff *skb;
 821
 822         if (likely(!io_file_need_scm(file)))
 823                 return 0;
 824
 825         /*
 826          * See if we can merge this file into an existing skb SCM_RIGHTS
 827          * file set. If there's no room, fall back to allocating a new skb
 828          * and filling it in.
 829          */
 830         spin_lock_irq(&head->lock);
 831         skb = skb_peek(head);
 832         if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD)
 833                 __skb_unlink(skb, head);
 834         else
 835                 skb = NULL;
 836         spin_unlock_irq(&head->lock);
 837
 838         if (!skb) {
 839                 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
 840                 if (!fpl)
 841                         return -ENOMEM;
 842
 843                 skb = alloc_skb(0, GFP_KERNEL);
 844                 if (!skb) {
 845                         kfree(fpl);
 846                         return -ENOMEM;
 847                 }
 848
 849                 fpl->user = get_uid(current_user());
 850                 fpl->max = SCM_MAX_FD;
 851                 fpl->count = 0;
 852
 853                 UNIXCB(skb).fp = fpl;
 854                 skb->sk = sk;
 855                 skb->scm_io_uring = 1;
 856                 skb->destructor = unix_destruct_scm;
 857                 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
 858         }
 859
 860         fpl = UNIXCB(skb).fp;
 861         fpl->fp[fpl->count++] = get_file(file);
 862         unix_inflight(fpl->user, file);
 863         skb_queue_head(head, skb);
 864         fput(file);
 865 #endif
 866         return 0;
 867 }
 868
 869 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
 870 {
 871         struct file *file = prsrc->file;
 872 #if defined(CONFIG_UNIX)
 873         struct sock *sock = ctx->ring_sock->sk;
 874         struct sk_buff_head list, *head = &sock->sk_receive_queue;
 875         struct sk_buff *skb;
 876         int i;
 877
 878         if (!io_file_need_scm(file)) {
 879                 fput(file);
 880                 return;
 881         }
 882
 883         __skb_queue_head_init(&list);
 884
 885         /*
 886          * Find the skb that holds this file in its SCM_RIGHTS. When found,
 887          * remove this entry and rearrange the file array.
 888          */
 889         skb = skb_dequeue(head);
 890         while (skb) {
 891                 struct scm_fp_list *fp;
 892
 893                 fp = UNIXCB(skb).fp;
 894                 for (i = 0; i < fp->count; i++) {
 895                         int left;
 896
 897                         if (fp->fp[i] != file)
 898                                 continue;
 899
 900                         unix_notinflight(fp->user, fp->fp[i]);
 901                         left = fp->count - 1 - i;
 902                         if (left) {
 903                                 memmove(&fp->fp[i], &fp->fp[i + 1],
 904                                                 left * sizeof(struct file *));
 905                         }
 906                         fp->count--;
 907                         if (!fp->count) {
 908                                 kfree_skb(skb);
 909                                 skb = NULL;
 910                         } else {
 911                                 __skb_queue_tail(&list, skb);
 912                         }
 913                         fput(file);
 914                         file = NULL;
 915                         break;
 916                 }
 917
 918                 if (!file)
 919                         break;
 920
 921                 __skb_queue_tail(&list, skb);
 922
 923                 skb = skb_dequeue(head);
 924         }
 925
 926         if (skb_peek(&list)) {
 927                 spin_lock_irq(&head->lock);
 928                 while ((skb = __skb_dequeue(&list)) != NULL)
 929                         __skb_queue_tail(head, skb);
 930                 spin_unlock_irq(&head->lock);
 931         }
 932 #else
 933         fput(file);
 934 #endif
 935 }
 936
 937 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
 938                           unsigned nr_args, u64 __user *tags)
 939 {
 940         __s32 __user *fds = (__s32 __user *) arg;
 941         struct file *file;
 942         int fd, ret;
 943         unsigned i;
 944
 945         if (ctx->file_data)
 946                 return -EBUSY;
 947         if (!nr_args)
 948                 return -EINVAL;
 949         if (nr_args > IORING_MAX_FIXED_FILES)
 950                 return -EMFILE;
 951         if (nr_args > rlimit(RLIMIT_NOFILE))
 952                 return -EMFILE;
 953         ret = io_rsrc_node_switch_start(ctx);
 954         if (ret)
 955                 return ret;
 956         ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
 957                                  &ctx->file_data);
 958         if (ret)
 959                 return ret;
 960
 961         if (!io_alloc_file_tables(&ctx->file_table, nr_args)) {
 962                 io_rsrc_data_free(ctx->file_data);
 963                 ctx->file_data = NULL;
 964                 return -ENOMEM;
 965         }
 966
 967         for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
 968                 struct io_fixed_file *file_slot;
 969
 970                 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) {
 971                         ret = -EFAULT;
 972                         goto fail;
 973                 }
 974                 /* allow sparse sets */
 975                 if (!fds || fd == -1) {
 976                         ret = -EINVAL;
 977                         if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
 978                                 goto fail;
 979                         continue;
 980                 }
 981
 982                 file = fget(fd);
 983                 ret = -EBADF;
 984                 if (unlikely(!file))
 985                         goto fail;
 986
 987                 /*
 988                  * Don't allow io_uring instances to be registered. If UNIX
 989                  * isn't enabled, then this causes a reference cycle and this
 990                  * instance can never get freed. If UNIX is enabled we'll
 991                  * handle it just fine, but there's still no point in allowing
 992                  * a ring fd as it doesn't support regular read/write anyway.
 993                  */
 994                 if (io_is_uring_fops(file)) {
 995                         fput(file);
 996                         goto fail;
 997                 }
 998                 ret = io_scm_file_account(ctx, file);
 999                 if (ret) {
1000                         fput(file);
1001                         goto fail;
1002                 }
1003                 file_slot = io_fixed_file_slot(&ctx->file_table, i);
1004                 io_fixed_file_set(file_slot, file);
1005                 io_file_bitmap_set(&ctx->file_table, i);
1006         }
1007
1008         /* default it to the whole table */
1009         io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files);
1010         io_rsrc_node_switch(ctx, NULL);
1011         return 0;
1012 fail:
1013         __io_sqe_files_unregister(ctx);
1014         return ret;
1015 }
1016
1017 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
1018 {
1019         io_buffer_unmap(ctx, &prsrc->buf);
1020         prsrc->buf = NULL;
1021 }
1022
1023 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1024 {
1025         unsigned int i;
1026
1027         for (i = 0; i < ctx->nr_user_bufs; i++)
1028                 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
1029         kfree(ctx->user_bufs);
1030         io_rsrc_data_free(ctx->buf_data);
1031         ctx->user_bufs = NULL;
1032         ctx->buf_data = NULL;
1033         ctx->nr_user_bufs = 0;
1034 }
1035
1036 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
1037 {
1038         unsigned nr = ctx->nr_user_bufs;
1039         int ret;
1040
1041         if (!ctx->buf_data)
1042                 return -ENXIO;
1043
1044         /*
1045          * Quiesce may unlock ->uring_lock, and while it's not held
1046          * prevent new requests using the table.
1047          */
1048         ctx->nr_user_bufs = 0;
1049         ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
1050         ctx->nr_user_bufs = nr;
1051         if (!ret)
1052                 __io_sqe_buffers_unregister(ctx);
1053         return ret;
1054 }
1055
1056 /*
1057  * Not super efficient, but this is just a registration time. And we do cache
1058  * the last compound head, so generally we'll only do a full search if we don't
1059  * match that one.
1060  *
1061  * We check if the given compound head page has already been accounted, to
1062  * avoid double accounting it. This allows us to account the full size of the
1063  * page, not just the constituent pages of a huge page.
1064  */
1065 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
1066                                   int nr_pages, struct page *hpage)
1067 {
1068         int i, j;
1069
1070         /* check current page array */
1071         for (i = 0; i < nr_pages; i++) {
1072                 if (!PageCompound(pages[i]))
1073                         continue;
1074                 if (compound_head(pages[i]) == hpage)
1075                         return true;
1076         }
1077
1078         /* check previously registered pages */
1079         for (i = 0; i < ctx->nr_user_bufs; i++) {
1080                 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
1081
1082                 for (j = 0; j < imu->nr_bvecs; j++) {
1083                         if (!PageCompound(imu->bvec[j].bv_page))
1084                                 continue;
1085                         if (compound_head(imu->bvec[j].bv_page) == hpage)
1086                                 return true;
1087                 }
1088         }
1089
1090         return false;
1091 }
1092
1093 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
1094                                  int nr_pages, struct io_mapped_ubuf *imu,
1095                                  struct page **last_hpage)
1096 {
1097         int i, ret;
1098
1099         imu->acct_pages = 0;
1100         for (i = 0; i < nr_pages; i++) {
1101                 if (!PageCompound(pages[i])) {
1102                         imu->acct_pages++;
1103                 } else {
1104                         struct page *hpage;
1105
1106                         hpage = compound_head(pages[i]);
1107                         if (hpage == *last_hpage)
1108                                 continue;
1109                         *last_hpage = hpage;
1110                         if (headpage_already_acct(ctx, pages, i, hpage))
1111                                 continue;
1112                         imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
1113                 }
1114         }
1115
1116         if (!imu->acct_pages)
1117                 return 0;
1118
1119         ret = io_account_mem(ctx, imu->acct_pages);
1120         if (ret)
1121                 imu->acct_pages = 0;
1122         return ret;
1123 }
1124
1125 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages)
1126 {
1127         unsigned long start, end, nr_pages;
1128         struct vm_area_struct **vmas = NULL;
1129         struct page **pages = NULL;
1130         int i, pret, ret = -ENOMEM;
1131
1132         end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1133         start = ubuf >> PAGE_SHIFT;
1134         nr_pages = end - start;
1135
1136         pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
1137         if (!pages)
1138                 goto done;
1139
1140         vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
1141                               GFP_KERNEL);
1142         if (!vmas)
1143                 goto done;
1144
1145         ret = 0;
1146         mmap_read_lock(current->mm);
1147         pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
1148                               pages, vmas);
1149         if (pret == nr_pages) {
1150                 /* don't support file backed memory */
1151                 for (i = 0; i < nr_pages; i++) {
1152                         struct vm_area_struct *vma = vmas[i];
1153
1154                         if (vma_is_shmem(vma))
1155                                 continue;
1156                         if (vma->vm_file &&
1157                             !is_file_hugepages(vma->vm_file)) {
1158                                 ret = -EOPNOTSUPP;
1159                                 break;
1160                         }
1161                 }
1162                 *npages = nr_pages;
1163         } else {
1164                 ret = pret < 0 ? pret : -EFAULT;
1165         }
1166         mmap_read_unlock(current->mm);
1167         if (ret) {
1168                 /*
1169                  * if we did partial map, or found file backed vmas,
1170                  * release any pages we did get
1171                  */
1172                 if (pret > 0)
1173                         unpin_user_pages(pages, pret);
1174                 goto done;
1175         }
1176         ret = 0;
1177 done:
1178         kvfree(vmas);
1179         if (ret < 0) {
1180                 kvfree(pages);
1181                 pages = ERR_PTR(ret);
1182         }
1183         return pages;
1184 }
1185
1186 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
1187                                   struct io_mapped_ubuf **pimu,
1188                                   struct page **last_hpage)
1189 {
1190         struct io_mapped_ubuf *imu = NULL;
1191         struct page **pages = NULL;
1192         unsigned long off;
1193         size_t size;
1194         int ret, nr_pages, i;
1195
1196         *pimu = ctx->dummy_ubuf;
1197         if (!iov->iov_base)
1198                 return 0;
1199
1200         ret = -ENOMEM;
1201         pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
1202                                 &nr_pages);
1203         if (IS_ERR(pages)) {
1204                 ret = PTR_ERR(pages);
1205                 pages = NULL;
1206                 goto done;
1207         }
1208
1209         imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
1210         if (!imu)
1211                 goto done;
1212
1213         ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
1214         if (ret) {
1215                 unpin_user_pages(pages, nr_pages);
1216                 goto done;
1217         }
1218
1219         off = (unsigned long) iov->iov_base & ~PAGE_MASK;
1220         size = iov->iov_len;
1221         for (i = 0; i < nr_pages; i++) {
1222                 size_t vec_len;
1223
1224                 vec_len = min_t(size_t, size, PAGE_SIZE - off);
1225                 imu->bvec[i].bv_page = pages[i];
1226                 imu->bvec[i].bv_len = vec_len;
1227                 imu->bvec[i].bv_offset = off;
1228                 off = 0;
1229                 size -= vec_len;
1230         }
1231         /* store original address for later verification */
1232         imu->ubuf = (unsigned long) iov->iov_base;
1233         imu->ubuf_end = imu->ubuf + iov->iov_len;
1234         imu->nr_bvecs = nr_pages;
1235         *pimu = imu;
1236         ret = 0;
1237 done:
1238         if (ret)
1239                 kvfree(imu);
1240         kvfree(pages);
1241         return ret;
1242 }
1243
1244 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
1245 {
1246         ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
1247         return ctx->user_bufs ? 0 : -ENOMEM;
1248 }
1249
1250 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
1251                             unsigned int nr_args, u64 __user *tags)
1252 {
1253         struct page *last_hpage = NULL;
1254         struct io_rsrc_data *data;
1255         int i, ret;
1256         struct iovec iov;
1257
1258         BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
1259
1260         if (ctx->user_bufs)
1261                 return -EBUSY;
1262         if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
1263                 return -EINVAL;
1264         ret = io_rsrc_node_switch_start(ctx);
1265         if (ret)
1266                 return ret;
1267         ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
1268         if (ret)
1269                 return ret;
1270         ret = io_buffers_map_alloc(ctx, nr_args);
1271         if (ret) {
1272                 io_rsrc_data_free(data);
1273                 return ret;
1274         }
1275
1276         for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
1277                 if (arg) {
1278                         ret = io_copy_iov(ctx, &iov, arg, i);
1279                         if (ret)
1280                                 break;
1281                         ret = io_buffer_validate(&iov);
1282                         if (ret)
1283                                 break;
1284                 } else {
1285                         memset(&iov, 0, sizeof(iov));
1286                 }
1287
1288                 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
1289                         ret = -EINVAL;
1290                         break;
1291                 }
1292
1293                 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
1294                                              &last_hpage);
1295                 if (ret)
1296                         break;
1297         }
1298
1299         WARN_ON_ONCE(ctx->buf_data);
1300
1301         ctx->buf_data = data;
1302         if (ret)
1303                 __io_sqe_buffers_unregister(ctx);
1304         else
1305                 io_rsrc_node_switch(ctx, NULL);
1306         return ret;
1307 }
1308
1309 int io_import_fixed(int ddir, struct iov_iter *iter,
1310                            struct io_mapped_ubuf *imu,
1311                            u64 buf_addr, size_t len)
1312 {
1313         u64 buf_end;
1314         size_t offset;
1315
1316         if (WARN_ON_ONCE(!imu))
1317                 return -EFAULT;
1318         if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
1319                 return -EFAULT;
1320         /* not inside the mapped region */
1321         if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
1322                 return -EFAULT;
1323
1324         /*
1325          * May not be a start of buffer, set size appropriately
1326          * and advance us to the beginning.
1327          */
1328         offset = buf_addr - imu->ubuf;
1329         iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
1330
1331         if (offset) {
1332                 /*
1333                  * Don't use iov_iter_advance() here, as it's really slow for
1334                  * using the latter parts of a big fixed buffer - it iterates
1335                  * over each segment manually. We can cheat a bit here, because
1336                  * we know that:
1337                  *
1338                  * 1) it's a BVEC iter, we set it up
1339                  * 2) all bvecs are PAGE_SIZE in size, except potentially the
1340                  *    first and last bvec
1341                  *
1342                  * So just find our index, and adjust the iterator afterwards.
1343                  * If the offset is within the first bvec (or the whole first
1344                  * bvec, just use iov_iter_advance(). This makes it easier
1345                  * since we can just skip the first segment, which may not
1346                  * be PAGE_SIZE aligned.
1347                  */
1348                 const struct bio_vec *bvec = imu->bvec;
1349
1350                 if (offset <= bvec->bv_len) {
1351                         iov_iter_advance(iter, offset);
1352                 } else {
1353                         unsigned long seg_skip;
1354
1355                         /* skip first vec */
1356                         offset -= bvec->bv_len;
1357                         seg_skip = 1 + (offset >> PAGE_SHIFT);
1358
1359                         iter->bvec = bvec + seg_skip;
1360                         iter->nr_segs -= seg_skip;
1361                         iter->count -= bvec->bv_len + offset;
1362                         iter->iov_offset = offset & ~PAGE_MASK;
1363                 }
1364         }
1365
1366         return 0;
1367 }