fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <[email protected]>
  17  * Copyright (C) 2005-2006 Linus Torvalds <[email protected]>
  18  * Copyright (C) 2006 Ingo Molnar <[email protected]>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/fsnotify.h>
  34 #include <linux/security.h>
  35 #include <linux/gfp.h>
  36 #include <linux/net.h>
  37 #include <linux/socket.h>
  38 #include <linux/sched/signal.h>
  39
  40 #include "internal.h"
  41
  42 /*
  43  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  44  * indicate they support non-blocking reads or writes, we must clear it
  45  * here if set to avoid blocking other users of this pipe if splice is
  46  * being done on it.
  47  */
  48 static noinline void noinline pipe_clear_nowait(struct file *file)
  49 {
  50         fmode_t fmode = READ_ONCE(file->f_mode);
  51
  52         do {
  53                 if (!(fmode & FMODE_NOWAIT))
  54                         break;
  55         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  56 }
  57
  58 /*
  59  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  60  * a vm helper function, it's already simplified quite a bit by the
  61  * addition of remove_mapping(). If success is returned, the caller may
  62  * attempt to reuse this page for another destination.
  63  */
  64 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  65                 struct pipe_buffer *buf)
  66 {
  67         struct folio *folio = page_folio(buf->page);
  68         struct address_space *mapping;
  69
  70         folio_lock(folio);
  71
  72         mapping = folio_mapping(folio);
  73         if (mapping) {
  74                 WARN_ON(!folio_test_uptodate(folio));
  75
  76                 /*
  77                  * At least for ext2 with nobh option, we need to wait on
  78                  * writeback completing on this folio, since we'll remove it
  79                  * from the pagecache.  Otherwise truncate wont wait on the
  80                  * folio, allowing the disk blocks to be reused by someone else
  81                  * before we actually wrote our data to them. fs corruption
  82                  * ensues.
  83                  */
  84                 folio_wait_writeback(folio);
  85
  86                 if (!filemap_release_folio(folio, GFP_KERNEL))
  87                         goto out_unlock;
  88
  89                 /*
  90                  * If we succeeded in removing the mapping, set LRU flag
  91                  * and return good.
  92                  */
  93                 if (remove_mapping(mapping, folio)) {
  94                         buf->flags |= PIPE_BUF_FLAG_LRU;
  95                         return true;
  96                 }
  97         }
  98
  99         /*
 100          * Raced with truncate or failed to remove folio from current
 101          * address space, unlock and return failure.
 102          */
 103 out_unlock:
 104         folio_unlock(folio);
 105         return false;
 106 }
 107
 108 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 109                                         struct pipe_buffer *buf)
 110 {
 111         put_page(buf->page);
 112         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 113 }
 114
 115 /*
 116  * Check whether the contents of buf is OK to access. Since the content
 117  * is a page cache page, IO may be in flight.
 118  */
 119 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 120                                        struct pipe_buffer *buf)
 121 {
 122         struct page *page = buf->page;
 123         int err;
 124
 125         if (!PageUptodate(page)) {
 126                 lock_page(page);
 127
 128                 /*
 129                  * Page got truncated/unhashed. This will cause a 0-byte
 130                  * splice, if this is the first page.
 131                  */
 132                 if (!page->mapping) {
 133                         err = -ENODATA;
 134                         goto error;
 135                 }
 136
 137                 /*
 138                  * Uh oh, read-error from disk.
 139                  */
 140                 if (!PageUptodate(page)) {
 141                         err = -EIO;
 142                         goto error;
 143                 }
 144
 145                 /*
 146                  * Page is ok afterall, we are done.
 147                  */
 148                 unlock_page(page);
 149         }
 150
 151         return 0;
 152 error:
 153         unlock_page(page);
 154         return err;
 155 }
 156
 157 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 158         .confirm        = page_cache_pipe_buf_confirm,
 159         .release        = page_cache_pipe_buf_release,
 160         .try_steal      = page_cache_pipe_buf_try_steal,
 161         .get            = generic_pipe_buf_get,
 162 };
 163
 164 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 165                 struct pipe_buffer *buf)
 166 {
 167         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 168                 return false;
 169
 170         buf->flags |= PIPE_BUF_FLAG_LRU;
 171         return generic_pipe_buf_try_steal(pipe, buf);
 172 }
 173
 174 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 175         .release        = page_cache_pipe_buf_release,
 176         .try_steal      = user_page_pipe_buf_try_steal,
 177         .get            = generic_pipe_buf_get,
 178 };
 179
 180 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 181 {
 182         smp_mb();
 183         if (waitqueue_active(&pipe->rd_wait))
 184                 wake_up_interruptible(&pipe->rd_wait);
 185         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 186 }
 187
 188 /**
 189  * splice_to_pipe - fill passed data into a pipe
 190  * @pipe:       pipe to fill
 191  * @spd:        data to fill
 192  *
 193  * Description:
 194  *    @spd contains a map of pages and len/offset tuples, along with
 195  *    the struct pipe_buf_operations associated with these pages. This
 196  *    function will link that data to the pipe.
 197  *
 198  */
 199 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 200                        struct splice_pipe_desc *spd)
 201 {
 202         unsigned int spd_pages = spd->nr_pages;
 203         unsigned int tail = pipe->tail;
 204         unsigned int head = pipe->head;
 205         unsigned int mask = pipe->ring_size - 1;
 206         int ret = 0, page_nr = 0;
 207
 208         if (!spd_pages)
 209                 return 0;
 210
 211         if (unlikely(!pipe->readers)) {
 212                 send_sig(SIGPIPE, current, 0);
 213                 ret = -EPIPE;
 214                 goto out;
 215         }
 216
 217         while (!pipe_full(head, tail, pipe->max_usage)) {
 218                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 219
 220                 buf->page = spd->pages[page_nr];
 221                 buf->offset = spd->partial[page_nr].offset;
 222                 buf->len = spd->partial[page_nr].len;
 223                 buf->private = spd->partial[page_nr].private;
 224                 buf->ops = spd->ops;
 225                 buf->flags = 0;
 226
 227                 head++;
 228                 pipe->head = head;
 229                 page_nr++;
 230                 ret += buf->len;
 231
 232                 if (!--spd->nr_pages)
 233                         break;
 234         }
 235
 236         if (!ret)
 237                 ret = -EAGAIN;
 238
 239 out:
 240         while (page_nr < spd_pages)
 241                 spd->spd_release(spd, page_nr++);
 242
 243         return ret;
 244 }
 245 EXPORT_SYMBOL_GPL(splice_to_pipe);
 246
 247 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 248 {
 249         unsigned int head = pipe->head;
 250         unsigned int tail = pipe->tail;
 251         unsigned int mask = pipe->ring_size - 1;
 252         int ret;
 253
 254         if (unlikely(!pipe->readers)) {
 255                 send_sig(SIGPIPE, current, 0);
 256                 ret = -EPIPE;
 257         } else if (pipe_full(head, tail, pipe->max_usage)) {
 258                 ret = -EAGAIN;
 259         } else {
 260                 pipe->bufs[head & mask] = *buf;
 261                 pipe->head = head + 1;
 262                 return buf->len;
 263         }
 264         pipe_buf_release(pipe, buf);
 265         return ret;
 266 }
 267 EXPORT_SYMBOL(add_to_pipe);
 268
 269 /*
 270  * Check if we need to grow the arrays holding pages and partial page
 271  * descriptions.
 272  */
 273 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 274 {
 275         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 276
 277         spd->nr_pages_max = max_usage;
 278         if (max_usage <= PIPE_DEF_BUFFERS)
 279                 return 0;
 280
 281         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 282         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 283                                      GFP_KERNEL);
 284
 285         if (spd->pages && spd->partial)
 286                 return 0;
 287
 288         kfree(spd->pages);
 289         kfree(spd->partial);
 290         return -ENOMEM;
 291 }
 292
 293 void splice_shrink_spd(struct splice_pipe_desc *spd)
 294 {
 295         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 296                 return;
 297
 298         kfree(spd->pages);
 299         kfree(spd->partial);
 300 }
 301
 302 /**
 303  * copy_splice_read -  Copy data from a file and splice the copy into a pipe
 304  * @in: The file to read from
 305  * @ppos: Pointer to the file position to read from
 306  * @pipe: The pipe to splice into
 307  * @len: The amount to splice
 308  * @flags: The SPLICE_F_* flags
 309  *
 310  * This function allocates a bunch of pages sufficient to hold the requested
 311  * amount of data (but limited by the remaining pipe capacity), passes it to
 312  * the file's ->read_iter() to read into and then splices the used pages into
 313  * the pipe.
 314  *
 315  * Return: On success, the number of bytes read will be returned and *@ppos
 316  * will be updated if appropriate; 0 will be returned if there is no more data
 317  * to be read; -EAGAIN will be returned if the pipe had no space, and some
 318  * other negative error code will be returned on error.  A short read may occur
 319  * if the pipe has insufficient space, we reach the end of the data or we hit a
 320  * hole.
 321  */
 322 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 323                          struct pipe_inode_info *pipe,
 324                          size_t len, unsigned int flags)
 325 {
 326         struct iov_iter to;
 327         struct bio_vec *bv;
 328         struct kiocb kiocb;
 329         struct page **pages;
 330         ssize_t ret;
 331         size_t used, npages, chunk, remain, keep = 0;
 332         int i;
 333
 334         /* Work out how much data we can actually add into the pipe */
 335         used = pipe_occupancy(pipe->head, pipe->tail);
 336         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 337         len = min_t(size_t, len, npages * PAGE_SIZE);
 338         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 339
 340         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 341                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 342         if (!bv)
 343                 return -ENOMEM;
 344
 345         pages = (struct page **)(bv + npages);
 346         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 347         if (!npages) {
 348                 kfree(bv);
 349                 return -ENOMEM;
 350         }
 351
 352         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 353
 354         for (i = 0; i < npages; i++) {
 355                 chunk = min_t(size_t, PAGE_SIZE, remain);
 356                 bv[i].bv_page = pages[i];
 357                 bv[i].bv_offset = 0;
 358                 bv[i].bv_len = chunk;
 359                 remain -= chunk;
 360         }
 361
 362         /* Do the I/O */
 363         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 364         init_sync_kiocb(&kiocb, in);
 365         kiocb.ki_pos = *ppos;
 366         ret = call_read_iter(in, &kiocb, &to);
 367
 368         if (ret > 0) {
 369                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 370                 *ppos = kiocb.ki_pos;
 371         }
 372
 373         /*
 374          * Callers of ->splice_read() expect -EAGAIN on "can't put anything in
 375          * there", rather than -EFAULT.
 376          */
 377         if (ret == -EFAULT)
 378                 ret = -EAGAIN;
 379
 380         /* Free any pages that didn't get touched at all. */
 381         if (keep < npages)
 382                 release_pages(pages + keep, npages - keep);
 383
 384         /* Push the remaining pages into the pipe. */
 385         remain = ret;
 386         for (i = 0; i < keep; i++) {
 387                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 388
 389                 chunk = min_t(size_t, remain, PAGE_SIZE);
 390                 *buf = (struct pipe_buffer) {
 391                         .ops    = &default_pipe_buf_ops,
 392                         .page   = bv[i].bv_page,
 393                         .offset = 0,
 394                         .len    = chunk,
 395                 };
 396                 pipe->head++;
 397                 remain -= chunk;
 398         }
 399
 400         kfree(bv);
 401         return ret;
 402 }
 403 EXPORT_SYMBOL(copy_splice_read);
 404
 405 const struct pipe_buf_operations default_pipe_buf_ops = {
 406         .release        = generic_pipe_buf_release,
 407         .try_steal      = generic_pipe_buf_try_steal,
 408         .get            = generic_pipe_buf_get,
 409 };
 410
 411 /* Pipe buffer operations for a socket and similar. */
 412 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 413         .release        = generic_pipe_buf_release,
 414         .get            = generic_pipe_buf_get,
 415 };
 416 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 417
 418 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 419 {
 420         smp_mb();
 421         if (waitqueue_active(&pipe->wr_wait))
 422                 wake_up_interruptible(&pipe->wr_wait);
 423         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 424 }
 425
 426 /**
 427  * splice_from_pipe_feed - feed available data from a pipe to a file
 428  * @pipe:       pipe to splice from
 429  * @sd:         information to @actor
 430  * @actor:      handler that splices the data
 431  *
 432  * Description:
 433  *    This function loops over the pipe and calls @actor to do the
 434  *    actual moving of a single struct pipe_buffer to the desired
 435  *    destination.  It returns when there's no more buffers left in
 436  *    the pipe or if the requested number of bytes (@sd->total_len)
 437  *    have been copied.  It returns a positive number (one) if the
 438  *    pipe needs to be filled with more data, zero if the required
 439  *    number of bytes have been copied and -errno on error.
 440  *
 441  *    This, together with splice_from_pipe_{begin,end,next}, may be
 442  *    used to implement the functionality of __splice_from_pipe() when
 443  *    locking is required around copying the pipe buffers to the
 444  *    destination.
 445  */
 446 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 447                           splice_actor *actor)
 448 {
 449         unsigned int head = pipe->head;
 450         unsigned int tail = pipe->tail;
 451         unsigned int mask = pipe->ring_size - 1;
 452         int ret;
 453
 454         while (!pipe_empty(head, tail)) {
 455                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 456
 457                 sd->len = buf->len;
 458                 if (sd->len > sd->total_len)
 459                         sd->len = sd->total_len;
 460
 461                 ret = pipe_buf_confirm(pipe, buf);
 462                 if (unlikely(ret)) {
 463                         if (ret == -ENODATA)
 464                                 ret = 0;
 465                         return ret;
 466                 }
 467
 468                 ret = actor(pipe, buf, sd);
 469                 if (ret <= 0)
 470                         return ret;
 471
 472                 buf->offset += ret;
 473                 buf->len -= ret;
 474
 475                 sd->num_spliced += ret;
 476                 sd->len -= ret;
 477                 sd->pos += ret;
 478                 sd->total_len -= ret;
 479
 480                 if (!buf->len) {
 481                         pipe_buf_release(pipe, buf);
 482                         tail++;
 483                         pipe->tail = tail;
 484                         if (pipe->files)
 485                                 sd->need_wakeup = true;
 486                 }
 487
 488                 if (!sd->total_len)
 489                         return 0;
 490         }
 491
 492         return 1;
 493 }
 494
 495 /* We know we have a pipe buffer, but maybe it's empty? */
 496 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 497 {
 498         unsigned int tail = pipe->tail;
 499         unsigned int mask = pipe->ring_size - 1;
 500         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 501
 502         if (unlikely(!buf->len)) {
 503                 pipe_buf_release(pipe, buf);
 504                 pipe->tail = tail+1;
 505                 return true;
 506         }
 507
 508         return false;
 509 }
 510
 511 /**
 512  * splice_from_pipe_next - wait for some data to splice from
 513  * @pipe:       pipe to splice from
 514  * @sd:         information about the splice operation
 515  *
 516  * Description:
 517  *    This function will wait for some data and return a positive
 518  *    value (one) if pipe buffers are available.  It will return zero
 519  *    or -errno if no more data needs to be spliced.
 520  */
 521 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 522 {
 523         /*
 524          * Check for signal early to make process killable when there are
 525          * always buffers available
 526          */
 527         if (signal_pending(current))
 528                 return -ERESTARTSYS;
 529
 530 repeat:
 531         while (pipe_empty(pipe->head, pipe->tail)) {
 532                 if (!pipe->writers)
 533                         return 0;
 534
 535                 if (sd->num_spliced)
 536                         return 0;
 537
 538                 if (sd->flags & SPLICE_F_NONBLOCK)
 539                         return -EAGAIN;
 540
 541                 if (signal_pending(current))
 542                         return -ERESTARTSYS;
 543
 544                 if (sd->need_wakeup) {
 545                         wakeup_pipe_writers(pipe);
 546                         sd->need_wakeup = false;
 547                 }
 548
 549                 pipe_wait_readable(pipe);
 550         }
 551
 552         if (eat_empty_buffer(pipe))
 553                 goto repeat;
 554
 555         return 1;
 556 }
 557
 558 /**
 559  * splice_from_pipe_begin - start splicing from pipe
 560  * @sd:         information about the splice operation
 561  *
 562  * Description:
 563  *    This function should be called before a loop containing
 564  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 565  *    initialize the necessary fields of @sd.
 566  */
 567 static void splice_from_pipe_begin(struct splice_desc *sd)
 568 {
 569         sd->num_spliced = 0;
 570         sd->need_wakeup = false;
 571 }
 572
 573 /**
 574  * splice_from_pipe_end - finish splicing from pipe
 575  * @pipe:       pipe to splice from
 576  * @sd:         information about the splice operation
 577  *
 578  * Description:
 579  *    This function will wake up pipe writers if necessary.  It should
 580  *    be called after a loop containing splice_from_pipe_next() and
 581  *    splice_from_pipe_feed().
 582  */
 583 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 584 {
 585         if (sd->need_wakeup)
 586                 wakeup_pipe_writers(pipe);
 587 }
 588
 589 /**
 590  * __splice_from_pipe - splice data from a pipe to given actor
 591  * @pipe:       pipe to splice from
 592  * @sd:         information to @actor
 593  * @actor:      handler that splices the data
 594  *
 595  * Description:
 596  *    This function does little more than loop over the pipe and call
 597  *    @actor to do the actual moving of a single struct pipe_buffer to
 598  *    the desired destination. See pipe_to_file, pipe_to_sendmsg, or
 599  *    pipe_to_user.
 600  *
 601  */
 602 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 603                            splice_actor *actor)
 604 {
 605         int ret;
 606
 607         splice_from_pipe_begin(sd);
 608         do {
 609                 cond_resched();
 610                 ret = splice_from_pipe_next(pipe, sd);
 611                 if (ret > 0)
 612                         ret = splice_from_pipe_feed(pipe, sd, actor);
 613         } while (ret > 0);
 614         splice_from_pipe_end(pipe, sd);
 615
 616         return sd->num_spliced ? sd->num_spliced : ret;
 617 }
 618 EXPORT_SYMBOL(__splice_from_pipe);
 619
 620 /**
 621  * splice_from_pipe - splice data from a pipe to a file
 622  * @pipe:       pipe to splice from
 623  * @out:        file to splice to
 624  * @ppos:       position in @out
 625  * @len:        how many bytes to splice
 626  * @flags:      splice modifier flags
 627  * @actor:      handler that splices the data
 628  *
 629  * Description:
 630  *    See __splice_from_pipe. This function locks the pipe inode,
 631  *    otherwise it's identical to __splice_from_pipe().
 632  *
 633  */
 634 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 635                          loff_t *ppos, size_t len, unsigned int flags,
 636                          splice_actor *actor)
 637 {
 638         ssize_t ret;
 639         struct splice_desc sd = {
 640                 .total_len = len,
 641                 .flags = flags,
 642                 .pos = *ppos,
 643                 .u.file = out,
 644         };
 645
 646         pipe_lock(pipe);
 647         ret = __splice_from_pipe(pipe, &sd, actor);
 648         pipe_unlock(pipe);
 649
 650         return ret;
 651 }
 652
 653 /**
 654  * iter_file_splice_write - splice data from a pipe to a file
 655  * @pipe:       pipe info
 656  * @out:        file to write to
 657  * @ppos:       position in @out
 658  * @len:        number of bytes to splice
 659  * @flags:      splice modifier flags
 660  *
 661  * Description:
 662  *    Will either move or copy pages (determined by @flags options) from
 663  *    the given pipe inode to the given file.
 664  *    This one is ->write_iter-based.
 665  *
 666  */
 667 ssize_t
 668 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 669                           loff_t *ppos, size_t len, unsigned int flags)
 670 {
 671         struct splice_desc sd = {
 672                 .total_len = len,
 673                 .flags = flags,
 674                 .pos = *ppos,
 675                 .u.file = out,
 676         };
 677         int nbufs = pipe->max_usage;
 678         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 679                                         GFP_KERNEL);
 680         ssize_t ret;
 681
 682         if (unlikely(!array))
 683                 return -ENOMEM;
 684
 685         pipe_lock(pipe);
 686
 687         splice_from_pipe_begin(&sd);
 688         while (sd.total_len) {
 689                 struct iov_iter from;
 690                 unsigned int head, tail, mask;
 691                 size_t left;
 692                 int n;
 693
 694                 ret = splice_from_pipe_next(pipe, &sd);
 695                 if (ret <= 0)
 696                         break;
 697
 698                 if (unlikely(nbufs < pipe->max_usage)) {
 699                         kfree(array);
 700                         nbufs = pipe->max_usage;
 701                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 702                                         GFP_KERNEL);
 703                         if (!array) {
 704                                 ret = -ENOMEM;
 705                                 break;
 706                         }
 707                 }
 708
 709                 head = pipe->head;
 710                 tail = pipe->tail;
 711                 mask = pipe->ring_size - 1;
 712
 713                 /* build the vector */
 714                 left = sd.total_len;
 715                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 716                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 717                         size_t this_len = buf->len;
 718
 719                         /* zero-length bvecs are not supported, skip them */
 720                         if (!this_len)
 721                                 continue;
 722                         this_len = min(this_len, left);
 723
 724                         ret = pipe_buf_confirm(pipe, buf);
 725                         if (unlikely(ret)) {
 726                                 if (ret == -ENODATA)
 727                                         ret = 0;
 728                                 goto done;
 729                         }
 730
 731                         bvec_set_page(&array[n], buf->page, this_len,
 732                                       buf->offset);
 733                         left -= this_len;
 734                         n++;
 735                 }
 736
 737                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 738                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 739                 if (ret <= 0)
 740                         break;
 741
 742                 sd.num_spliced += ret;
 743                 sd.total_len -= ret;
 744                 *ppos = sd.pos;
 745
 746                 /* dismiss the fully eaten buffers, adjust the partial one */
 747                 tail = pipe->tail;
 748                 while (ret) {
 749                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 750                         if (ret >= buf->len) {
 751                                 ret -= buf->len;
 752                                 buf->len = 0;
 753                                 pipe_buf_release(pipe, buf);
 754                                 tail++;
 755                                 pipe->tail = tail;
 756                                 if (pipe->files)
 757                                         sd.need_wakeup = true;
 758                         } else {
 759                                 buf->offset += ret;
 760                                 buf->len -= ret;
 761                                 ret = 0;
 762                         }
 763                 }
 764         }
 765 done:
 766         kfree(array);
 767         splice_from_pipe_end(pipe, &sd);
 768
 769         pipe_unlock(pipe);
 770
 771         if (sd.num_spliced)
 772                 ret = sd.num_spliced;
 773
 774         return ret;
 775 }
 776
 777 EXPORT_SYMBOL(iter_file_splice_write);
 778
 779 #ifdef CONFIG_NET
 780 /**
 781  * splice_to_socket - splice data from a pipe to a socket
 782  * @pipe:       pipe to splice from
 783  * @out:        socket to write to
 784  * @ppos:       position in @out
 785  * @len:        number of bytes to splice
 786  * @flags:      splice modifier flags
 787  *
 788  * Description:
 789  *    Will send @len bytes from the pipe to a network socket. No data copying
 790  *    is involved.
 791  *
 792  */
 793 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 794                          loff_t *ppos, size_t len, unsigned int flags)
 795 {
 796         struct socket *sock = sock_from_file(out);
 797         struct bio_vec bvec[16];
 798         struct msghdr msg = {};
 799         ssize_t ret = 0;
 800         size_t spliced = 0;
 801         bool need_wakeup = false;
 802
 803         pipe_lock(pipe);
 804
 805         while (len > 0) {
 806                 unsigned int head, tail, mask, bc = 0;
 807                 size_t remain = len;
 808
 809                 /*
 810                  * Check for signal early to make process killable when there
 811                  * are always buffers available
 812                  */
 813                 ret = -ERESTARTSYS;
 814                 if (signal_pending(current))
 815                         break;
 816
 817                 while (pipe_empty(pipe->head, pipe->tail)) {
 818                         ret = 0;
 819                         if (!pipe->writers)
 820                                 goto out;
 821
 822                         if (spliced)
 823                                 goto out;
 824
 825                         ret = -EAGAIN;
 826                         if (flags & SPLICE_F_NONBLOCK)
 827                                 goto out;
 828
 829                         ret = -ERESTARTSYS;
 830                         if (signal_pending(current))
 831                                 goto out;
 832
 833                         if (need_wakeup) {
 834                                 wakeup_pipe_writers(pipe);
 835                                 need_wakeup = false;
 836                         }
 837
 838                         pipe_wait_readable(pipe);
 839                 }
 840
 841                 head = pipe->head;
 842                 tail = pipe->tail;
 843                 mask = pipe->ring_size - 1;
 844
 845                 while (!pipe_empty(head, tail)) {
 846                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 847                         size_t seg;
 848
 849                         if (!buf->len) {
 850                                 tail++;
 851                                 continue;
 852                         }
 853
 854                         seg = min_t(size_t, remain, buf->len);
 855
 856                         ret = pipe_buf_confirm(pipe, buf);
 857                         if (unlikely(ret)) {
 858                                 if (ret == -ENODATA)
 859                                         ret = 0;
 860                                 break;
 861                         }
 862
 863                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 864                         remain -= seg;
 865                         if (remain == 0 || bc >= ARRAY_SIZE(bvec))
 866                                 break;
 867                         tail++;
 868                 }
 869
 870                 if (!bc)
 871                         break;
 872
 873                 msg.msg_flags = MSG_SPLICE_PAGES;
 874                 if (flags & SPLICE_F_MORE)
 875                         msg.msg_flags |= MSG_MORE;
 876                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 877                         msg.msg_flags |= MSG_MORE;
 878                 if (out->f_flags & O_NONBLOCK)
 879                         msg.msg_flags |= MSG_DONTWAIT;
 880
 881                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 882                               len - remain);
 883                 ret = sock_sendmsg(sock, &msg);
 884                 if (ret <= 0)
 885                         break;
 886
 887                 spliced += ret;
 888                 len -= ret;
 889                 tail = pipe->tail;
 890                 while (ret > 0) {
 891                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 892                         size_t seg = min_t(size_t, ret, buf->len);
 893
 894                         buf->offset += seg;
 895                         buf->len -= seg;
 896                         ret -= seg;
 897
 898                         if (!buf->len) {
 899                                 pipe_buf_release(pipe, buf);
 900                                 tail++;
 901                         }
 902                 }
 903
 904                 if (tail != pipe->tail) {
 905                         pipe->tail = tail;
 906                         if (pipe->files)
 907                                 need_wakeup = true;
 908                 }
 909         }
 910
 911 out:
 912         pipe_unlock(pipe);
 913         if (need_wakeup)
 914                 wakeup_pipe_writers(pipe);
 915         return spliced ?: ret;
 916 }
 917 #endif
 918
 919 static int warn_unsupported(struct file *file, const char *op)
 920 {
 921         pr_debug_ratelimited(
 922                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 923                 op, file, current->pid, current->comm);
 924         return -EINVAL;
 925 }
 926
 927 /*
 928  * Attempt to initiate a splice from pipe to file.
 929  */
 930 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 931                            loff_t *ppos, size_t len, unsigned int flags)
 932 {
 933         if (unlikely(!out->f_op->splice_write))
 934                 return warn_unsupported(out, "write");
 935         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 936 }
 937
 938 /*
 939  * Indicate to the caller that there was a premature EOF when reading from the
 940  * source and the caller didn't indicate they would be sending more data after
 941  * this.
 942  */
 943 static void do_splice_eof(struct splice_desc *sd)
 944 {
 945         if (sd->splice_eof)
 946                 sd->splice_eof(sd);
 947 }
 948
 949 /**
 950  * vfs_splice_read - Read data from a file and splice it into a pipe
 951  * @in:         File to splice from
 952  * @ppos:       Input file offset
 953  * @pipe:       Pipe to splice to
 954  * @len:        Number of bytes to splice
 955  * @flags:      Splice modifier flags (SPLICE_F_*)
 956  *
 957  * Splice the requested amount of data from the input file to the pipe.  This
 958  * is synchronous as the caller must hold the pipe lock across the entire
 959  * operation.
 960  *
 961  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 962  * a hole and a negative error code otherwise.
 963  */
 964 long vfs_splice_read(struct file *in, loff_t *ppos,
 965                      struct pipe_inode_info *pipe, size_t len,
 966                      unsigned int flags)
 967 {
 968         unsigned int p_space;
 969         int ret;
 970
 971         if (unlikely(!(in->f_mode & FMODE_READ)))
 972                 return -EBADF;
 973         if (!len)
 974                 return 0;
 975
 976         /* Don't try to read more the pipe has space for. */
 977         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 978         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 979
 980         ret = rw_verify_area(READ, in, ppos, len);
 981         if (unlikely(ret < 0))
 982                 return ret;
 983
 984         if (unlikely(len > MAX_RW_COUNT))
 985                 len = MAX_RW_COUNT;
 986
 987         if (unlikely(!in->f_op->splice_read))
 988                 return warn_unsupported(in, "read");
 989         /*
 990          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 991          * buffer, copy into it and splice that into the pipe.
 992          */
 993         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 994                 return copy_splice_read(in, ppos, pipe, len, flags);
 995         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 996 }
 997 EXPORT_SYMBOL_GPL(vfs_splice_read);
 998
 999 /**
1000  * splice_direct_to_actor - splices data directly between two non-pipes
1001  * @in:         file to splice from
1002  * @sd:         actor information on where to splice to
1003  * @actor:      handles the data splicing
1004  *
1005  * Description:
1006  *    This is a special case helper to splice directly between two
1007  *    points, without requiring an explicit pipe. Internally an allocated
1008  *    pipe is cached in the process, and reused during the lifetime of
1009  *    that process.
1010  *
1011  */
1012 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1013                                splice_direct_actor *actor)
1014 {
1015         struct pipe_inode_info *pipe;
1016         long ret, bytes;
1017         size_t len;
1018         int i, flags, more;
1019
1020         /*
1021          * We require the input to be seekable, as we don't want to randomly
1022          * drop data for eg socket -> socket splicing. Use the piped splicing
1023          * for that!
1024          */
1025         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1026                 return -EINVAL;
1027
1028         /*
1029          * neither in nor out is a pipe, setup an internal pipe attached to
1030          * 'out' and transfer the wanted data from 'in' to 'out' through that
1031          */
1032         pipe = current->splice_pipe;
1033         if (unlikely(!pipe)) {
1034                 pipe = alloc_pipe_info();
1035                 if (!pipe)
1036                         return -ENOMEM;
1037
1038                 /*
1039                  * We don't have an immediate reader, but we'll read the stuff
1040                  * out of the pipe right after the splice_to_pipe(). So set
1041                  * PIPE_READERS appropriately.
1042                  */
1043                 pipe->readers = 1;
1044
1045                 current->splice_pipe = pipe;
1046         }
1047
1048         /*
1049          * Do the splice.
1050          */
1051         bytes = 0;
1052         len = sd->total_len;
1053
1054         /* Don't block on output, we have to drain the direct pipe. */
1055         flags = sd->flags;
1056         sd->flags &= ~SPLICE_F_NONBLOCK;
1057
1058         /*
1059          * We signal MORE until we've read sufficient data to fulfill the
1060          * request and we keep signalling it if the caller set it.
1061          */
1062         more = sd->flags & SPLICE_F_MORE;
1063         sd->flags |= SPLICE_F_MORE;
1064
1065         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1066
1067         while (len) {
1068                 size_t read_len;
1069                 loff_t pos = sd->pos, prev_pos = pos;
1070
1071                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1072                 if (unlikely(ret <= 0))
1073                         goto read_failure;
1074
1075                 read_len = ret;
1076                 sd->total_len = read_len;
1077
1078                 /*
1079                  * If we now have sufficient data to fulfill the request then
1080                  * we clear SPLICE_F_MORE if it was not set initially.
1081                  */
1082                 if (read_len >= len && !more)
1083                         sd->flags &= ~SPLICE_F_MORE;
1084
1085                 /*
1086                  * NOTE: nonblocking mode only applies to the input. We
1087                  * must not do the output in nonblocking mode as then we
1088                  * could get stuck data in the internal pipe:
1089                  */
1090                 ret = actor(pipe, sd);
1091                 if (unlikely(ret <= 0)) {
1092                         sd->pos = prev_pos;
1093                         goto out_release;
1094                 }
1095
1096                 bytes += ret;
1097                 len -= ret;
1098                 sd->pos = pos;
1099
1100                 if (ret < read_len) {
1101                         sd->pos = prev_pos + ret;
1102                         goto out_release;
1103                 }
1104         }
1105
1106 done:
1107         pipe->tail = pipe->head = 0;
1108         file_accessed(in);
1109         return bytes;
1110
1111 read_failure:
1112         /*
1113          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1114          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1115          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1116          * least 1 byte *then* we will also do the ->splice_eof() call.
1117          */
1118         if (ret == 0 && !more && len > 0 && bytes)
1119                 do_splice_eof(sd);
1120 out_release:
1121         /*
1122          * If we did an incomplete transfer we must release
1123          * the pipe buffers in question:
1124          */
1125         for (i = 0; i < pipe->ring_size; i++) {
1126                 struct pipe_buffer *buf = &pipe->bufs[i];
1127
1128                 if (buf->ops)
1129                         pipe_buf_release(pipe, buf);
1130         }
1131
1132         if (!bytes)
1133                 bytes = ret;
1134
1135         goto done;
1136 }
1137 EXPORT_SYMBOL(splice_direct_to_actor);
1138
1139 static int direct_splice_actor(struct pipe_inode_info *pipe,
1140                                struct splice_desc *sd)
1141 {
1142         struct file *file = sd->u.file;
1143
1144         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1145                               sd->flags);
1146 }
1147
1148 static void direct_file_splice_eof(struct splice_desc *sd)
1149 {
1150         struct file *file = sd->u.file;
1151
1152         if (file->f_op->splice_eof)
1153                 file->f_op->splice_eof(file);
1154 }
1155
1156 /**
1157  * do_splice_direct - splices data directly between two files
1158  * @in:         file to splice from
1159  * @ppos:       input file offset
1160  * @out:        file to splice to
1161  * @opos:       output file offset
1162  * @len:        number of bytes to splice
1163  * @flags:      splice modifier flags
1164  *
1165  * Description:
1166  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1167  *    doing it in the application would incur an extra system call
1168  *    (splice in + splice out, as compared to just sendfile()). So this helper
1169  *    can splice directly through a process-private pipe.
1170  *
1171  */
1172 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1173                       loff_t *opos, size_t len, unsigned int flags)
1174 {
1175         struct splice_desc sd = {
1176                 .len            = len,
1177                 .total_len      = len,
1178                 .flags          = flags,
1179                 .pos            = *ppos,
1180                 .u.file         = out,
1181                 .splice_eof     = direct_file_splice_eof,
1182                 .opos           = opos,
1183         };
1184         long ret;
1185
1186         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1187                 return -EBADF;
1188
1189         if (unlikely(out->f_flags & O_APPEND))
1190                 return -EINVAL;
1191
1192         ret = rw_verify_area(WRITE, out, opos, len);
1193         if (unlikely(ret < 0))
1194                 return ret;
1195
1196         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1197         if (ret > 0)
1198                 *ppos = sd.pos;
1199
1200         return ret;
1201 }
1202 EXPORT_SYMBOL(do_splice_direct);
1203
1204 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1205 {
1206         for (;;) {
1207                 if (unlikely(!pipe->readers)) {
1208                         send_sig(SIGPIPE, current, 0);
1209                         return -EPIPE;
1210                 }
1211                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1212                         return 0;
1213                 if (flags & SPLICE_F_NONBLOCK)
1214                         return -EAGAIN;
1215                 if (signal_pending(current))
1216                         return -ERESTARTSYS;
1217                 pipe_wait_writable(pipe);
1218         }
1219 }
1220
1221 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1222                                struct pipe_inode_info *opipe,
1223                                size_t len, unsigned int flags);
1224
1225 long splice_file_to_pipe(struct file *in,
1226                          struct pipe_inode_info *opipe,
1227                          loff_t *offset,
1228                          size_t len, unsigned int flags)
1229 {
1230         long ret;
1231
1232         pipe_lock(opipe);
1233         ret = wait_for_space(opipe, flags);
1234         if (!ret)
1235                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1236         pipe_unlock(opipe);
1237         if (ret > 0)
1238                 wakeup_pipe_readers(opipe);
1239         return ret;
1240 }
1241
1242 /*
1243  * Determine where to splice to/from.
1244  */
1245 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1246                loff_t *off_out, size_t len, unsigned int flags)
1247 {
1248         struct pipe_inode_info *ipipe;
1249         struct pipe_inode_info *opipe;
1250         loff_t offset;
1251         long ret;
1252
1253         if (unlikely(!(in->f_mode & FMODE_READ) ||
1254                      !(out->f_mode & FMODE_WRITE)))
1255                 return -EBADF;
1256
1257         ipipe = get_pipe_info(in, true);
1258         opipe = get_pipe_info(out, true);
1259
1260         if (ipipe && opipe) {
1261                 if (off_in || off_out)
1262                         return -ESPIPE;
1263
1264                 /* Splicing to self would be fun, but... */
1265                 if (ipipe == opipe)
1266                         return -EINVAL;
1267
1268                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1269                         flags |= SPLICE_F_NONBLOCK;
1270
1271                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1272         }
1273
1274         if (ipipe) {
1275                 if (off_in)
1276                         return -ESPIPE;
1277                 if (off_out) {
1278                         if (!(out->f_mode & FMODE_PWRITE))
1279                                 return -EINVAL;
1280                         offset = *off_out;
1281                 } else {
1282                         offset = out->f_pos;
1283                 }
1284
1285                 if (unlikely(out->f_flags & O_APPEND))
1286                         return -EINVAL;
1287
1288                 ret = rw_verify_area(WRITE, out, &offset, len);
1289                 if (unlikely(ret < 0))
1290                         return ret;
1291
1292                 if (in->f_flags & O_NONBLOCK)
1293                         flags |= SPLICE_F_NONBLOCK;
1294
1295                 file_start_write(out);
1296                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1297                 file_end_write(out);
1298
1299                 if (ret > 0)
1300                         fsnotify_modify(out);
1301
1302                 if (!off_out)
1303                         out->f_pos = offset;
1304                 else
1305                         *off_out = offset;
1306
1307                 return ret;
1308         }
1309
1310         if (opipe) {
1311                 if (off_out)
1312                         return -ESPIPE;
1313                 if (off_in) {
1314                         if (!(in->f_mode & FMODE_PREAD))
1315                                 return -EINVAL;
1316                         offset = *off_in;
1317                 } else {
1318                         offset = in->f_pos;
1319                 }
1320
1321                 if (out->f_flags & O_NONBLOCK)
1322                         flags |= SPLICE_F_NONBLOCK;
1323
1324                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1325
1326                 if (ret > 0)
1327                         fsnotify_access(in);
1328
1329                 if (!off_in)
1330                         in->f_pos = offset;
1331                 else
1332                         *off_in = offset;
1333
1334                 return ret;
1335         }
1336
1337         return -EINVAL;
1338 }
1339
1340 static long __do_splice(struct file *in, loff_t __user *off_in,
1341                         struct file *out, loff_t __user *off_out,
1342                         size_t len, unsigned int flags)
1343 {
1344         struct pipe_inode_info *ipipe;
1345         struct pipe_inode_info *opipe;
1346         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1347         long ret;
1348
1349         ipipe = get_pipe_info(in, true);
1350         opipe = get_pipe_info(out, true);
1351
1352         if (ipipe) {
1353                 if (off_in)
1354                         return -ESPIPE;
1355                 pipe_clear_nowait(in);
1356         }
1357         if (opipe) {
1358                 if (off_out)
1359                         return -ESPIPE;
1360                 pipe_clear_nowait(out);
1361         }
1362
1363         if (off_out) {
1364                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1365                         return -EFAULT;
1366                 __off_out = &offset;
1367         }
1368         if (off_in) {
1369                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1370                         return -EFAULT;
1371                 __off_in = &offset;
1372         }
1373
1374         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1375         if (ret < 0)
1376                 return ret;
1377
1378         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1379                 return -EFAULT;
1380         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1381                 return -EFAULT;
1382
1383         return ret;
1384 }
1385
1386 static int iter_to_pipe(struct iov_iter *from,
1387                         struct pipe_inode_info *pipe,
1388                         unsigned flags)
1389 {
1390         struct pipe_buffer buf = {
1391                 .ops = &user_page_pipe_buf_ops,
1392                 .flags = flags
1393         };
1394         size_t total = 0;
1395         int ret = 0;
1396
1397         while (iov_iter_count(from)) {
1398                 struct page *pages[16];
1399                 ssize_t left;
1400                 size_t start;
1401                 int i, n;
1402
1403                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1404                 if (left <= 0) {
1405                         ret = left;
1406                         break;
1407                 }
1408
1409                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1410                 for (i = 0; i < n; i++) {
1411                         int size = min_t(int, left, PAGE_SIZE - start);
1412
1413                         buf.page = pages[i];
1414                         buf.offset = start;
1415                         buf.len = size;
1416                         ret = add_to_pipe(pipe, &buf);
1417                         if (unlikely(ret < 0)) {
1418                                 iov_iter_revert(from, left);
1419                                 // this one got dropped by add_to_pipe()
1420                                 while (++i < n)
1421                                         put_page(pages[i]);
1422                                 goto out;
1423                         }
1424                         total += ret;
1425                         left -= size;
1426                         start = 0;
1427                 }
1428         }
1429 out:
1430         return total ? total : ret;
1431 }
1432
1433 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1434                         struct splice_desc *sd)
1435 {
1436         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1437         return n == sd->len ? n : -EFAULT;
1438 }
1439
1440 /*
1441  * For lack of a better implementation, implement vmsplice() to userspace
1442  * as a simple copy of the pipes pages to the user iov.
1443  */
1444 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1445                              unsigned int flags)
1446 {
1447         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1448         struct splice_desc sd = {
1449                 .total_len = iov_iter_count(iter),
1450                 .flags = flags,
1451                 .u.data = iter
1452         };
1453         long ret = 0;
1454
1455         if (!pipe)
1456                 return -EBADF;
1457
1458         pipe_clear_nowait(file);
1459
1460         if (sd.total_len) {
1461                 pipe_lock(pipe);
1462                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1463                 pipe_unlock(pipe);
1464         }
1465
1466         return ret;
1467 }
1468
1469 /*
1470  * vmsplice splices a user address range into a pipe. It can be thought of
1471  * as splice-from-memory, where the regular splice is splice-from-file (or
1472  * to file). In both cases the output is a pipe, naturally.
1473  */
1474 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1475                              unsigned int flags)
1476 {
1477         struct pipe_inode_info *pipe;
1478         long ret = 0;
1479         unsigned buf_flag = 0;
1480
1481         if (flags & SPLICE_F_GIFT)
1482                 buf_flag = PIPE_BUF_FLAG_GIFT;
1483
1484         pipe = get_pipe_info(file, true);
1485         if (!pipe)
1486                 return -EBADF;
1487
1488         pipe_clear_nowait(file);
1489
1490         pipe_lock(pipe);
1491         ret = wait_for_space(pipe, flags);
1492         if (!ret)
1493                 ret = iter_to_pipe(iter, pipe, buf_flag);
1494         pipe_unlock(pipe);
1495         if (ret > 0)
1496                 wakeup_pipe_readers(pipe);
1497         return ret;
1498 }
1499
1500 static int vmsplice_type(struct fd f, int *type)
1501 {
1502         if (!f.file)
1503                 return -EBADF;
1504         if (f.file->f_mode & FMODE_WRITE) {
1505                 *type = ITER_SOURCE;
1506         } else if (f.file->f_mode & FMODE_READ) {
1507                 *type = ITER_DEST;
1508         } else {
1509                 fdput(f);
1510                 return -EBADF;
1511         }
1512         return 0;
1513 }
1514
1515 /*
1516  * Note that vmsplice only really supports true splicing _from_ user memory
1517  * to a pipe, not the other way around. Splicing from user memory is a simple
1518  * operation that can be supported without any funky alignment restrictions
1519  * or nasty vm tricks. We simply map in the user memory and fill them into
1520  * a pipe. The reverse isn't quite as easy, though. There are two possible
1521  * solutions for that:
1522  *
1523  *      - memcpy() the data internally, at which point we might as well just
1524  *        do a regular read() on the buffer anyway.
1525  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1526  *        has restriction limitations on both ends of the pipe).
1527  *
1528  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1529  *
1530  */
1531 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1532                 unsigned long, nr_segs, unsigned int, flags)
1533 {
1534         struct iovec iovstack[UIO_FASTIOV];
1535         struct iovec *iov = iovstack;
1536         struct iov_iter iter;
1537         ssize_t error;
1538         struct fd f;
1539         int type;
1540
1541         if (unlikely(flags & ~SPLICE_F_ALL))
1542                 return -EINVAL;
1543
1544         f = fdget(fd);
1545         error = vmsplice_type(f, &type);
1546         if (error)
1547                 return error;
1548
1549         error = import_iovec(type, uiov, nr_segs,
1550                              ARRAY_SIZE(iovstack), &iov, &iter);
1551         if (error < 0)
1552                 goto out_fdput;
1553
1554         if (!iov_iter_count(&iter))
1555                 error = 0;
1556         else if (type == ITER_SOURCE)
1557                 error = vmsplice_to_pipe(f.file, &iter, flags);
1558         else
1559                 error = vmsplice_to_user(f.file, &iter, flags);
1560
1561         kfree(iov);
1562 out_fdput:
1563         fdput(f);
1564         return error;
1565 }
1566
1567 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1568                 int, fd_out, loff_t __user *, off_out,
1569                 size_t, len, unsigned int, flags)
1570 {
1571         struct fd in, out;
1572         long error;
1573
1574         if (unlikely(!len))
1575                 return 0;
1576
1577         if (unlikely(flags & ~SPLICE_F_ALL))
1578                 return -EINVAL;
1579
1580         error = -EBADF;
1581         in = fdget(fd_in);
1582         if (in.file) {
1583                 out = fdget(fd_out);
1584                 if (out.file) {
1585                         error = __do_splice(in.file, off_in, out.file, off_out,
1586                                                 len, flags);
1587                         fdput(out);
1588                 }
1589                 fdput(in);
1590         }
1591         return error;
1592 }
1593
1594 /*
1595  * Make sure there's data to read. Wait for input if we can, otherwise
1596  * return an appropriate error.
1597  */
1598 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1599 {
1600         int ret;
1601
1602         /*
1603          * Check the pipe occupancy without the inode lock first. This function
1604          * is speculative anyways, so missing one is ok.
1605          */
1606         if (!pipe_empty(pipe->head, pipe->tail))
1607                 return 0;
1608
1609         ret = 0;
1610         pipe_lock(pipe);
1611
1612         while (pipe_empty(pipe->head, pipe->tail)) {
1613                 if (signal_pending(current)) {
1614                         ret = -ERESTARTSYS;
1615                         break;
1616                 }
1617                 if (!pipe->writers)
1618                         break;
1619                 if (flags & SPLICE_F_NONBLOCK) {
1620                         ret = -EAGAIN;
1621                         break;
1622                 }
1623                 pipe_wait_readable(pipe);
1624         }
1625
1626         pipe_unlock(pipe);
1627         return ret;
1628 }
1629
1630 /*
1631  * Make sure there's writeable room. Wait for room if we can, otherwise
1632  * return an appropriate error.
1633  */
1634 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1635 {
1636         int ret;
1637
1638         /*
1639          * Check pipe occupancy without the inode lock first. This function
1640          * is speculative anyways, so missing one is ok.
1641          */
1642         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1643                 return 0;
1644
1645         ret = 0;
1646         pipe_lock(pipe);
1647
1648         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1649                 if (!pipe->readers) {
1650                         send_sig(SIGPIPE, current, 0);
1651                         ret = -EPIPE;
1652                         break;
1653                 }
1654                 if (flags & SPLICE_F_NONBLOCK) {
1655                         ret = -EAGAIN;
1656                         break;
1657                 }
1658                 if (signal_pending(current)) {
1659                         ret = -ERESTARTSYS;
1660                         break;
1661                 }
1662                 pipe_wait_writable(pipe);
1663         }
1664
1665         pipe_unlock(pipe);
1666         return ret;
1667 }
1668
1669 /*
1670  * Splice contents of ipipe to opipe.
1671  */
1672 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1673                                struct pipe_inode_info *opipe,
1674                                size_t len, unsigned int flags)
1675 {
1676         struct pipe_buffer *ibuf, *obuf;
1677         unsigned int i_head, o_head;
1678         unsigned int i_tail, o_tail;
1679         unsigned int i_mask, o_mask;
1680         int ret = 0;
1681         bool input_wakeup = false;
1682
1683
1684 retry:
1685         ret = ipipe_prep(ipipe, flags);
1686         if (ret)
1687                 return ret;
1688
1689         ret = opipe_prep(opipe, flags);
1690         if (ret)
1691                 return ret;
1692
1693         /*
1694          * Potential ABBA deadlock, work around it by ordering lock
1695          * grabbing by pipe info address. Otherwise two different processes
1696          * could deadlock (one doing tee from A -> B, the other from B -> A).
1697          */
1698         pipe_double_lock(ipipe, opipe);
1699
1700         i_tail = ipipe->tail;
1701         i_mask = ipipe->ring_size - 1;
1702         o_head = opipe->head;
1703         o_mask = opipe->ring_size - 1;
1704
1705         do {
1706                 size_t o_len;
1707
1708                 if (!opipe->readers) {
1709                         send_sig(SIGPIPE, current, 0);
1710                         if (!ret)
1711                                 ret = -EPIPE;
1712                         break;
1713                 }
1714
1715                 i_head = ipipe->head;
1716                 o_tail = opipe->tail;
1717
1718                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1719                         break;
1720
1721                 /*
1722                  * Cannot make any progress, because either the input
1723                  * pipe is empty or the output pipe is full.
1724                  */
1725                 if (pipe_empty(i_head, i_tail) ||
1726                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1727                         /* Already processed some buffers, break */
1728                         if (ret)
1729                                 break;
1730
1731                         if (flags & SPLICE_F_NONBLOCK) {
1732                                 ret = -EAGAIN;
1733                                 break;
1734                         }
1735
1736                         /*
1737                          * We raced with another reader/writer and haven't
1738                          * managed to process any buffers.  A zero return
1739                          * value means EOF, so retry instead.
1740                          */
1741                         pipe_unlock(ipipe);
1742                         pipe_unlock(opipe);
1743                         goto retry;
1744                 }
1745
1746                 ibuf = &ipipe->bufs[i_tail & i_mask];
1747                 obuf = &opipe->bufs[o_head & o_mask];
1748
1749                 if (len >= ibuf->len) {
1750                         /*
1751                          * Simply move the whole buffer from ipipe to opipe
1752                          */
1753                         *obuf = *ibuf;
1754                         ibuf->ops = NULL;
1755                         i_tail++;
1756                         ipipe->tail = i_tail;
1757                         input_wakeup = true;
1758                         o_len = obuf->len;
1759                         o_head++;
1760                         opipe->head = o_head;
1761                 } else {
1762                         /*
1763                          * Get a reference to this pipe buffer,
1764                          * so we can copy the contents over.
1765                          */
1766                         if (!pipe_buf_get(ipipe, ibuf)) {
1767                                 if (ret == 0)
1768                                         ret = -EFAULT;
1769                                 break;
1770                         }
1771                         *obuf = *ibuf;
1772
1773                         /*
1774                          * Don't inherit the gift and merge flags, we need to
1775                          * prevent multiple steals of this page.
1776                          */
1777                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1778                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1779
1780                         obuf->len = len;
1781                         ibuf->offset += len;
1782                         ibuf->len -= len;
1783                         o_len = len;
1784                         o_head++;
1785                         opipe->head = o_head;
1786                 }
1787                 ret += o_len;
1788                 len -= o_len;
1789         } while (len);
1790
1791         pipe_unlock(ipipe);
1792         pipe_unlock(opipe);
1793
1794         /*
1795          * If we put data in the output pipe, wakeup any potential readers.
1796          */
1797         if (ret > 0)
1798                 wakeup_pipe_readers(opipe);
1799
1800         if (input_wakeup)
1801                 wakeup_pipe_writers(ipipe);
1802
1803         return ret;
1804 }
1805
1806 /*
1807  * Link contents of ipipe to opipe.
1808  */
1809 static int link_pipe(struct pipe_inode_info *ipipe,
1810                      struct pipe_inode_info *opipe,
1811                      size_t len, unsigned int flags)
1812 {
1813         struct pipe_buffer *ibuf, *obuf;
1814         unsigned int i_head, o_head;
1815         unsigned int i_tail, o_tail;
1816         unsigned int i_mask, o_mask;
1817         int ret = 0;
1818
1819         /*
1820          * Potential ABBA deadlock, work around it by ordering lock
1821          * grabbing by pipe info address. Otherwise two different processes
1822          * could deadlock (one doing tee from A -> B, the other from B -> A).
1823          */
1824         pipe_double_lock(ipipe, opipe);
1825
1826         i_tail = ipipe->tail;
1827         i_mask = ipipe->ring_size - 1;
1828         o_head = opipe->head;
1829         o_mask = opipe->ring_size - 1;
1830
1831         do {
1832                 if (!opipe->readers) {
1833                         send_sig(SIGPIPE, current, 0);
1834                         if (!ret)
1835                                 ret = -EPIPE;
1836                         break;
1837                 }
1838
1839                 i_head = ipipe->head;
1840                 o_tail = opipe->tail;
1841
1842                 /*
1843                  * If we have iterated all input buffers or run out of
1844                  * output room, break.
1845                  */
1846                 if (pipe_empty(i_head, i_tail) ||
1847                     pipe_full(o_head, o_tail, opipe->max_usage))
1848                         break;
1849
1850                 ibuf = &ipipe->bufs[i_tail & i_mask];
1851                 obuf = &opipe->bufs[o_head & o_mask];
1852
1853                 /*
1854                  * Get a reference to this pipe buffer,
1855                  * so we can copy the contents over.
1856                  */
1857                 if (!pipe_buf_get(ipipe, ibuf)) {
1858                         if (ret == 0)
1859                                 ret = -EFAULT;
1860                         break;
1861                 }
1862
1863                 *obuf = *ibuf;
1864
1865                 /*
1866                  * Don't inherit the gift and merge flag, we need to prevent
1867                  * multiple steals of this page.
1868                  */
1869                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1870                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1871
1872                 if (obuf->len > len)
1873                         obuf->len = len;
1874                 ret += obuf->len;
1875                 len -= obuf->len;
1876
1877                 o_head++;
1878                 opipe->head = o_head;
1879                 i_tail++;
1880         } while (len);
1881
1882         pipe_unlock(ipipe);
1883         pipe_unlock(opipe);
1884
1885         /*
1886          * If we put data in the output pipe, wakeup any potential readers.
1887          */
1888         if (ret > 0)
1889                 wakeup_pipe_readers(opipe);
1890
1891         return ret;
1892 }
1893
1894 /*
1895  * This is a tee(1) implementation that works on pipes. It doesn't copy
1896  * any data, it simply references the 'in' pages on the 'out' pipe.
1897  * The 'flags' used are the SPLICE_F_* variants, currently the only
1898  * applicable one is SPLICE_F_NONBLOCK.
1899  */
1900 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1901 {
1902         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1903         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1904         int ret = -EINVAL;
1905
1906         if (unlikely(!(in->f_mode & FMODE_READ) ||
1907                      !(out->f_mode & FMODE_WRITE)))
1908                 return -EBADF;
1909
1910         /*
1911          * Duplicate the contents of ipipe to opipe without actually
1912          * copying the data.
1913          */
1914         if (ipipe && opipe && ipipe != opipe) {
1915                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1916                         flags |= SPLICE_F_NONBLOCK;
1917
1918                 /*
1919                  * Keep going, unless we encounter an error. The ipipe/opipe
1920                  * ordering doesn't really matter.
1921                  */
1922                 ret = ipipe_prep(ipipe, flags);
1923                 if (!ret) {
1924                         ret = opipe_prep(opipe, flags);
1925                         if (!ret)
1926                                 ret = link_pipe(ipipe, opipe, len, flags);
1927                 }
1928         }
1929
1930         return ret;
1931 }
1932
1933 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1934 {
1935         struct fd in, out;
1936         int error;
1937
1938         if (unlikely(flags & ~SPLICE_F_ALL))
1939                 return -EINVAL;
1940
1941         if (unlikely(!len))
1942                 return 0;
1943
1944         error = -EBADF;
1945         in = fdget(fdin);
1946         if (in.file) {
1947                 out = fdget(fdout);
1948                 if (out.file) {
1949                         error = do_tee(in.file, out.file, len, flags);
1950                         fdput(out);
1951                 }
1952                 fdput(in);
1953         }
1954
1955         return error;
1956 }