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 <axboe@kernel.dk>
  17  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  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 (folio_has_private(folio) &&
  87                     !filemap_release_folio(folio, GFP_KERNEL))
  88                         goto out_unlock;
  89
  90                 /*
  91                  * If we succeeded in removing the mapping, set LRU flag
  92                  * and return good.
  93                  */
  94                 if (remove_mapping(mapping, folio)) {
  95                         buf->flags |= PIPE_BUF_FLAG_LRU;
  96                         return true;
  97                 }
  98         }
  99
 100         /*
 101          * Raced with truncate or failed to remove folio from current
 102          * address space, unlock and return failure.
 103          */
 104 out_unlock:
 105         folio_unlock(folio);
 106         return false;
 107 }
 108
 109 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 110                                         struct pipe_buffer *buf)
 111 {
 112         put_page(buf->page);
 113         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 114 }
 115
 116 /*
 117  * Check whether the contents of buf is OK to access. Since the content
 118  * is a page cache page, IO may be in flight.
 119  */
 120 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 121                                        struct pipe_buffer *buf)
 122 {
 123         struct page *page = buf->page;
 124         int err;
 125
 126         if (!PageUptodate(page)) {
 127                 lock_page(page);
 128
 129                 /*
 130                  * Page got truncated/unhashed. This will cause a 0-byte
 131                  * splice, if this is the first page.
 132                  */
 133                 if (!page->mapping) {
 134                         err = -ENODATA;
 135                         goto error;
 136                 }
 137
 138                 /*
 139                  * Uh oh, read-error from disk.
 140                  */
 141                 if (!PageUptodate(page)) {
 142                         err = -EIO;
 143                         goto error;
 144                 }
 145
 146                 /*
 147                  * Page is ok afterall, we are done.
 148                  */
 149                 unlock_page(page);
 150         }
 151
 152         return 0;
 153 error:
 154         unlock_page(page);
 155         return err;
 156 }
 157
 158 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 159         .confirm        = page_cache_pipe_buf_confirm,
 160         .release        = page_cache_pipe_buf_release,
 161         .try_steal      = page_cache_pipe_buf_try_steal,
 162         .get            = generic_pipe_buf_get,
 163 };
 164
 165 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 166                 struct pipe_buffer *buf)
 167 {
 168         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 169                 return false;
 170
 171         buf->flags |= PIPE_BUF_FLAG_LRU;
 172         return generic_pipe_buf_try_steal(pipe, buf);
 173 }
 174
 175 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 176         .release        = page_cache_pipe_buf_release,
 177         .try_steal      = user_page_pipe_buf_try_steal,
 178         .get            = generic_pipe_buf_get,
 179 };
 180
 181 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 182 {
 183         smp_mb();
 184         if (waitqueue_active(&pipe->rd_wait))
 185                 wake_up_interruptible(&pipe->rd_wait);
 186         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 187 }
 188
 189 /**
 190  * splice_to_pipe - fill passed data into a pipe
 191  * @pipe:       pipe to fill
 192  * @spd:        data to fill
 193  *
 194  * Description:
 195  *    @spd contains a map of pages and len/offset tuples, along with
 196  *    the struct pipe_buf_operations associated with these pages. This
 197  *    function will link that data to the pipe.
 198  *
 199  */
 200 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 201                        struct splice_pipe_desc *spd)
 202 {
 203         unsigned int spd_pages = spd->nr_pages;
 204         unsigned int tail = pipe->tail;
 205         unsigned int head = pipe->head;
 206         unsigned int mask = pipe->ring_size - 1;
 207         int ret = 0, page_nr = 0;
 208
 209         if (!spd_pages)
 210                 return 0;
 211
 212         if (unlikely(!pipe->readers)) {
 213                 send_sig(SIGPIPE, current, 0);
 214                 ret = -EPIPE;
 215                 goto out;
 216         }
 217
 218         while (!pipe_full(head, tail, pipe->max_usage)) {
 219                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 220
 221                 buf->page = spd->pages[page_nr];
 222                 buf->offset = spd->partial[page_nr].offset;
 223                 buf->len = spd->partial[page_nr].len;
 224                 buf->private = spd->partial[page_nr].private;
 225                 buf->ops = spd->ops;
 226                 buf->flags = 0;
 227
 228                 head++;
 229                 pipe->head = head;
 230                 page_nr++;
 231                 ret += buf->len;
 232
 233                 if (!--spd->nr_pages)
 234                         break;
 235         }
 236
 237         if (!ret)
 238                 ret = -EAGAIN;
 239
 240 out:
 241         while (page_nr < spd_pages)
 242                 spd->spd_release(spd, page_nr++);
 243
 244         return ret;
 245 }
 246 EXPORT_SYMBOL_GPL(splice_to_pipe);
 247
 248 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 249 {
 250         unsigned int head = pipe->head;
 251         unsigned int tail = pipe->tail;
 252         unsigned int mask = pipe->ring_size - 1;
 253         int ret;
 254
 255         if (unlikely(!pipe->readers)) {
 256                 send_sig(SIGPIPE, current, 0);
 257                 ret = -EPIPE;
 258         } else if (pipe_full(head, tail, pipe->max_usage)) {
 259                 ret = -EAGAIN;
 260         } else {
 261                 pipe->bufs[head & mask] = *buf;
 262                 pipe->head = head + 1;
 263                 return buf->len;
 264         }
 265         pipe_buf_release(pipe, buf);
 266         return ret;
 267 }
 268 EXPORT_SYMBOL(add_to_pipe);
 269
 270 /*
 271  * Check if we need to grow the arrays holding pages and partial page
 272  * descriptions.
 273  */
 274 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 275 {
 276         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 277
 278         spd->nr_pages_max = max_usage;
 279         if (max_usage <= PIPE_DEF_BUFFERS)
 280                 return 0;
 281
 282         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 283         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 284                                      GFP_KERNEL);
 285
 286         if (spd->pages && spd->partial)
 287                 return 0;
 288
 289         kfree(spd->pages);
 290         kfree(spd->partial);
 291         return -ENOMEM;
 292 }
 293
 294 void splice_shrink_spd(struct splice_pipe_desc *spd)
 295 {
 296         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 297                 return;
 298
 299         kfree(spd->pages);
 300         kfree(spd->partial);
 301 }
 302
 303 /**
 304  * copy_splice_read -  Copy data from a file and splice the copy into a pipe
 305  * @in: The file to read from
 306  * @ppos: Pointer to the file position to read from
 307  * @pipe: The pipe to splice into
 308  * @len: The amount to splice
 309  * @flags: The SPLICE_F_* flags
 310  *
 311  * This function allocates a bunch of pages sufficient to hold the requested
 312  * amount of data (but limited by the remaining pipe capacity), passes it to
 313  * the file's ->read_iter() to read into and then splices the used pages into
 314  * the pipe.
 315  *
 316  * Return: On success, the number of bytes read will be returned and *@ppos
 317  * will be updated if appropriate; 0 will be returned if there is no more data
 318  * to be read; -EAGAIN will be returned if the pipe had no space, and some
 319  * other negative error code will be returned on error.  A short read may occur
 320  * if the pipe has insufficient space, we reach the end of the data or we hit a
 321  * hole.
 322  */
 323 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 324                          struct pipe_inode_info *pipe,
 325                          size_t len, unsigned int flags)
 326 {
 327         struct iov_iter to;
 328         struct bio_vec *bv;
 329         struct kiocb kiocb;
 330         struct page **pages;
 331         ssize_t ret;
 332         size_t used, npages, chunk, remain, keep = 0;
 333         int i;
 334
 335         /* Work out how much data we can actually add into the pipe */
 336         used = pipe_occupancy(pipe->head, pipe->tail);
 337         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 338         len = min_t(size_t, len, npages * PAGE_SIZE);
 339         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 340
 341         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 342                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 343         if (!bv)
 344                 return -ENOMEM;
 345
 346         pages = (struct page **)(bv + npages);
 347         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 348         if (!npages) {
 349                 kfree(bv);
 350                 return -ENOMEM;
 351         }
 352
 353         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 354
 355         for (i = 0; i < npages; i++) {
 356                 chunk = min_t(size_t, PAGE_SIZE, remain);
 357                 bv[i].bv_page = pages[i];
 358                 bv[i].bv_offset = 0;
 359                 bv[i].bv_len = chunk;
 360                 remain -= chunk;
 361         }
 362
 363         /* Do the I/O */
 364         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 365         init_sync_kiocb(&kiocb, in);
 366         kiocb.ki_pos = *ppos;
 367         ret = call_read_iter(in, &kiocb, &to);
 368
 369         if (ret > 0) {
 370                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 371                 *ppos = kiocb.ki_pos;
 372         }
 373
 374         /*
 375          * Callers of ->splice_read() expect -EAGAIN on "can't put anything in
 376          * there", rather than -EFAULT.
 377          */
 378         if (ret == -EFAULT)
 379                 ret = -EAGAIN;
 380
 381         /* Free any pages that didn't get touched at all. */
 382         if (keep < npages)
 383                 release_pages(pages + keep, npages - keep);
 384
 385         /* Push the remaining pages into the pipe. */
 386         remain = ret;
 387         for (i = 0; i < keep; i++) {
 388                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 389
 390                 chunk = min_t(size_t, remain, PAGE_SIZE);
 391                 *buf = (struct pipe_buffer) {
 392                         .ops    = &default_pipe_buf_ops,
 393                         .page   = bv[i].bv_page,
 394                         .offset = 0,
 395                         .len    = chunk,
 396                 };
 397                 pipe->head++;
 398                 remain -= chunk;
 399         }
 400
 401         kfree(bv);
 402         return ret;
 403 }
 404 EXPORT_SYMBOL(copy_splice_read);
 405
 406 const struct pipe_buf_operations default_pipe_buf_ops = {
 407         .release        = generic_pipe_buf_release,
 408         .try_steal      = generic_pipe_buf_try_steal,
 409         .get            = generic_pipe_buf_get,
 410 };
 411
 412 /* Pipe buffer operations for a socket and similar. */
 413 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 414         .release        = generic_pipe_buf_release,
 415         .get            = generic_pipe_buf_get,
 416 };
 417 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 418
 419 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 420 {
 421         smp_mb();
 422         if (waitqueue_active(&pipe->wr_wait))
 423                 wake_up_interruptible(&pipe->wr_wait);
 424         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 425 }
 426
 427 /**
 428  * splice_from_pipe_feed - feed available data from a pipe to a file
 429  * @pipe:       pipe to splice from
 430  * @sd:         information to @actor
 431  * @actor:      handler that splices the data
 432  *
 433  * Description:
 434  *    This function loops over the pipe and calls @actor to do the
 435  *    actual moving of a single struct pipe_buffer to the desired
 436  *    destination.  It returns when there's no more buffers left in
 437  *    the pipe or if the requested number of bytes (@sd->total_len)
 438  *    have been copied.  It returns a positive number (one) if the
 439  *    pipe needs to be filled with more data, zero if the required
 440  *    number of bytes have been copied and -errno on error.
 441  *
 442  *    This, together with splice_from_pipe_{begin,end,next}, may be
 443  *    used to implement the functionality of __splice_from_pipe() when
 444  *    locking is required around copying the pipe buffers to the
 445  *    destination.
 446  */
 447 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 448                           splice_actor *actor)
 449 {
 450         unsigned int head = pipe->head;
 451         unsigned int tail = pipe->tail;
 452         unsigned int mask = pipe->ring_size - 1;
 453         int ret;
 454
 455         while (!pipe_empty(head, tail)) {
 456                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 457
 458                 sd->len = buf->len;
 459                 if (sd->len > sd->total_len)
 460                         sd->len = sd->total_len;
 461
 462                 ret = pipe_buf_confirm(pipe, buf);
 463                 if (unlikely(ret)) {
 464                         if (ret == -ENODATA)
 465                                 ret = 0;
 466                         return ret;
 467                 }
 468
 469                 ret = actor(pipe, buf, sd);
 470                 if (ret <= 0)
 471                         return ret;
 472
 473                 buf->offset += ret;
 474                 buf->len -= ret;
 475
 476                 sd->num_spliced += ret;
 477                 sd->len -= ret;
 478                 sd->pos += ret;
 479                 sd->total_len -= ret;
 480
 481                 if (!buf->len) {
 482                         pipe_buf_release(pipe, buf);
 483                         tail++;
 484                         pipe->tail = tail;
 485                         if (pipe->files)
 486                                 sd->need_wakeup = true;
 487                 }
 488
 489                 if (!sd->total_len)
 490                         return 0;
 491         }
 492
 493         return 1;
 494 }
 495
 496 /* We know we have a pipe buffer, but maybe it's empty? */
 497 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 498 {
 499         unsigned int tail = pipe->tail;
 500         unsigned int mask = pipe->ring_size - 1;
 501         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 502
 503         if (unlikely(!buf->len)) {
 504                 pipe_buf_release(pipe, buf);
 505                 pipe->tail = tail+1;
 506                 return true;
 507         }
 508
 509         return false;
 510 }
 511
 512 /**
 513  * splice_from_pipe_next - wait for some data to splice from
 514  * @pipe:       pipe to splice from
 515  * @sd:         information about the splice operation
 516  *
 517  * Description:
 518  *    This function will wait for some data and return a positive
 519  *    value (one) if pipe buffers are available.  It will return zero
 520  *    or -errno if no more data needs to be spliced.
 521  */
 522 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 523 {
 524         /*
 525          * Check for signal early to make process killable when there are
 526          * always buffers available
 527          */
 528         if (signal_pending(current))
 529                 return -ERESTARTSYS;
 530
 531 repeat:
 532         while (pipe_empty(pipe->head, pipe->tail)) {
 533                 if (!pipe->writers)
 534                         return 0;
 535
 536                 if (sd->num_spliced)
 537                         return 0;
 538
 539                 if (sd->flags & SPLICE_F_NONBLOCK)
 540                         return -EAGAIN;
 541
 542                 if (signal_pending(current))
 543                         return -ERESTARTSYS;
 544
 545                 if (sd->need_wakeup) {
 546                         wakeup_pipe_writers(pipe);
 547                         sd->need_wakeup = false;
 548                 }
 549
 550                 pipe_wait_readable(pipe);
 551         }
 552
 553         if (eat_empty_buffer(pipe))
 554                 goto repeat;
 555
 556         return 1;
 557 }
 558
 559 /**
 560  * splice_from_pipe_begin - start splicing from pipe
 561  * @sd:         information about the splice operation
 562  *
 563  * Description:
 564  *    This function should be called before a loop containing
 565  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 566  *    initialize the necessary fields of @sd.
 567  */
 568 static void splice_from_pipe_begin(struct splice_desc *sd)
 569 {
 570         sd->num_spliced = 0;
 571         sd->need_wakeup = false;
 572 }
 573
 574 /**
 575  * splice_from_pipe_end - finish splicing from pipe
 576  * @pipe:       pipe to splice from
 577  * @sd:         information about the splice operation
 578  *
 579  * Description:
 580  *    This function will wake up pipe writers if necessary.  It should
 581  *    be called after a loop containing splice_from_pipe_next() and
 582  *    splice_from_pipe_feed().
 583  */
 584 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 585 {
 586         if (sd->need_wakeup)
 587                 wakeup_pipe_writers(pipe);
 588 }
 589
 590 /**
 591  * __splice_from_pipe - splice data from a pipe to given actor
 592  * @pipe:       pipe to splice from
 593  * @sd:         information to @actor
 594  * @actor:      handler that splices the data
 595  *
 596  * Description:
 597  *    This function does little more than loop over the pipe and call
 598  *    @actor to do the actual moving of a single struct pipe_buffer to
 599  *    the desired destination. See pipe_to_file, pipe_to_sendmsg, or
 600  *    pipe_to_user.
 601  *
 602  */
 603 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 604                            splice_actor *actor)
 605 {
 606         int ret;
 607
 608         splice_from_pipe_begin(sd);
 609         do {
 610                 cond_resched();
 611                 ret = splice_from_pipe_next(pipe, sd);
 612                 if (ret > 0)
 613                         ret = splice_from_pipe_feed(pipe, sd, actor);
 614         } while (ret > 0);
 615         splice_from_pipe_end(pipe, sd);
 616
 617         return sd->num_spliced ? sd->num_spliced : ret;
 618 }
 619 EXPORT_SYMBOL(__splice_from_pipe);
 620
 621 /**
 622  * splice_from_pipe - splice data from a pipe to a file
 623  * @pipe:       pipe to splice from
 624  * @out:        file to splice to
 625  * @ppos:       position in @out
 626  * @len:        how many bytes to splice
 627  * @flags:      splice modifier flags
 628  * @actor:      handler that splices the data
 629  *
 630  * Description:
 631  *    See __splice_from_pipe. This function locks the pipe inode,
 632  *    otherwise it's identical to __splice_from_pipe().
 633  *
 634  */
 635 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 636                          loff_t *ppos, size_t len, unsigned int flags,
 637                          splice_actor *actor)
 638 {
 639         ssize_t ret;
 640         struct splice_desc sd = {
 641                 .total_len = len,
 642                 .flags = flags,
 643                 .pos = *ppos,
 644                 .u.file = out,
 645         };
 646
 647         pipe_lock(pipe);
 648         ret = __splice_from_pipe(pipe, &sd, actor);
 649         pipe_unlock(pipe);
 650
 651         return ret;
 652 }
 653
 654 /**
 655  * iter_file_splice_write - splice data from a pipe to a file
 656  * @pipe:       pipe info
 657  * @out:        file to write to
 658  * @ppos:       position in @out
 659  * @len:        number of bytes to splice
 660  * @flags:      splice modifier flags
 661  *
 662  * Description:
 663  *    Will either move or copy pages (determined by @flags options) from
 664  *    the given pipe inode to the given file.
 665  *    This one is ->write_iter-based.
 666  *
 667  */
 668 ssize_t
 669 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 670                           loff_t *ppos, size_t len, unsigned int flags)
 671 {
 672         struct splice_desc sd = {
 673                 .total_len = len,
 674                 .flags = flags,
 675                 .pos = *ppos,
 676                 .u.file = out,
 677         };
 678         int nbufs = pipe->max_usage;
 679         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 680                                         GFP_KERNEL);
 681         ssize_t ret;
 682
 683         if (unlikely(!array))
 684                 return -ENOMEM;
 685
 686         pipe_lock(pipe);
 687
 688         splice_from_pipe_begin(&sd);
 689         while (sd.total_len) {
 690                 struct iov_iter from;
 691                 unsigned int head, tail, mask;
 692                 size_t left;
 693                 int n;
 694
 695                 ret = splice_from_pipe_next(pipe, &sd);
 696                 if (ret <= 0)
 697                         break;
 698
 699                 if (unlikely(nbufs < pipe->max_usage)) {
 700                         kfree(array);
 701                         nbufs = pipe->max_usage;
 702                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 703                                         GFP_KERNEL);
 704                         if (!array) {
 705                                 ret = -ENOMEM;
 706                                 break;
 707                         }
 708                 }
 709
 710                 head = pipe->head;
 711                 tail = pipe->tail;
 712                 mask = pipe->ring_size - 1;
 713
 714                 /* build the vector */
 715                 left = sd.total_len;
 716                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 717                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 718                         size_t this_len = buf->len;
 719
 720                         /* zero-length bvecs are not supported, skip them */
 721                         if (!this_len)
 722                                 continue;
 723                         this_len = min(this_len, left);
 724
 725                         ret = pipe_buf_confirm(pipe, buf);
 726                         if (unlikely(ret)) {
 727                                 if (ret == -ENODATA)
 728                                         ret = 0;
 729                                 goto done;
 730                         }
 731
 732                         bvec_set_page(&array[n], buf->page, this_len,
 733                                       buf->offset);
 734                         left -= this_len;
 735                         n++;
 736                 }
 737
 738                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 739                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 740                 if (ret <= 0)
 741                         break;
 742
 743                 sd.num_spliced += ret;
 744                 sd.total_len -= ret;
 745                 *ppos = sd.pos;
 746
 747                 /* dismiss the fully eaten buffers, adjust the partial one */
 748                 tail = pipe->tail;
 749                 while (ret) {
 750                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 751                         if (ret >= buf->len) {
 752                                 ret -= buf->len;
 753                                 buf->len = 0;
 754                                 pipe_buf_release(pipe, buf);
 755                                 tail++;
 756                                 pipe->tail = tail;
 757                                 if (pipe->files)
 758                                         sd.need_wakeup = true;
 759                         } else {
 760                                 buf->offset += ret;
 761                                 buf->len -= ret;
 762                                 ret = 0;
 763                         }
 764                 }
 765         }
 766 done:
 767         kfree(array);
 768         splice_from_pipe_end(pipe, &sd);
 769
 770         pipe_unlock(pipe);
 771
 772         if (sd.num_spliced)
 773                 ret = sd.num_spliced;
 774
 775         return ret;
 776 }
 777
 778 EXPORT_SYMBOL(iter_file_splice_write);
 779
 780 #ifdef CONFIG_NET
 781 /**
 782  * splice_to_socket - splice data from a pipe to a socket
 783  * @pipe:       pipe to splice from
 784  * @out:        socket to write to
 785  * @ppos:       position in @out
 786  * @len:        number of bytes to splice
 787  * @flags:      splice modifier flags
 788  *
 789  * Description:
 790  *    Will send @len bytes from the pipe to a network socket. No data copying
 791  *    is involved.
 792  *
 793  */
 794 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 795                          loff_t *ppos, size_t len, unsigned int flags)
 796 {
 797         struct socket *sock = sock_from_file(out);
 798         struct bio_vec bvec[16];
 799         struct msghdr msg = {};
 800         ssize_t ret = 0;
 801         size_t spliced = 0;
 802         bool need_wakeup = false;
 803
 804         pipe_lock(pipe);
 805
 806         while (len > 0) {
 807                 unsigned int head, tail, mask, bc = 0;
 808                 size_t remain = len;
 809
 810                 /*
 811                  * Check for signal early to make process killable when there
 812                  * are always buffers available
 813                  */
 814                 ret = -ERESTARTSYS;
 815                 if (signal_pending(current))
 816                         break;
 817
 818                 while (pipe_empty(pipe->head, pipe->tail)) {
 819                         ret = 0;
 820                         if (!pipe->writers)
 821                                 goto out;
 822
 823                         if (spliced)
 824                                 goto out;
 825
 826                         ret = -EAGAIN;
 827                         if (flags & SPLICE_F_NONBLOCK)
 828                                 goto out;
 829
 830                         ret = -ERESTARTSYS;
 831                         if (signal_pending(current))
 832                                 goto out;
 833
 834                         if (need_wakeup) {
 835                                 wakeup_pipe_writers(pipe);
 836                                 need_wakeup = false;
 837                         }
 838
 839                         pipe_wait_readable(pipe);
 840                 }
 841
 842                 head = pipe->head;
 843                 tail = pipe->tail;
 844                 mask = pipe->ring_size - 1;
 845
 846                 while (!pipe_empty(head, tail)) {
 847                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 848                         size_t seg;
 849
 850                         if (!buf->len) {
 851                                 tail++;
 852                                 continue;
 853                         }
 854
 855                         seg = min_t(size_t, remain, buf->len);
 856
 857                         ret = pipe_buf_confirm(pipe, buf);
 858                         if (unlikely(ret)) {
 859                                 if (ret == -ENODATA)
 860                                         ret = 0;
 861                                 break;
 862                         }
 863
 864                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 865                         remain -= seg;
 866                         if (remain == 0 || bc >= ARRAY_SIZE(bvec))
 867                                 break;
 868                         tail++;
 869                 }
 870
 871                 if (!bc)
 872                         break;
 873
 874                 msg.msg_flags = MSG_SPLICE_PAGES;
 875                 if (flags & SPLICE_F_MORE)
 876                         msg.msg_flags |= MSG_MORE;
 877                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 878                         msg.msg_flags |= MSG_MORE;
 879                 if (out->f_flags & O_NONBLOCK)
 880                         msg.msg_flags |= MSG_DONTWAIT;
 881
 882                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 883                               len - remain);
 884                 ret = sock_sendmsg(sock, &msg);
 885                 if (ret <= 0)
 886                         break;
 887
 888                 spliced += ret;
 889                 len -= ret;
 890                 tail = pipe->tail;
 891                 while (ret > 0) {
 892                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 893                         size_t seg = min_t(size_t, ret, buf->len);
 894
 895                         buf->offset += seg;
 896                         buf->len -= seg;
 897                         ret -= seg;
 898
 899                         if (!buf->len) {
 900                                 pipe_buf_release(pipe, buf);
 901                                 tail++;
 902                         }
 903                 }
 904
 905                 if (tail != pipe->tail) {
 906                         pipe->tail = tail;
 907                         if (pipe->files)
 908                                 need_wakeup = true;
 909                 }
 910         }
 911
 912 out:
 913         pipe_unlock(pipe);
 914         if (need_wakeup)
 915                 wakeup_pipe_writers(pipe);
 916         return spliced ?: ret;
 917 }
 918 #endif
 919
 920 static int warn_unsupported(struct file *file, const char *op)
 921 {
 922         pr_debug_ratelimited(
 923                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 924                 op, file, current->pid, current->comm);
 925         return -EINVAL;
 926 }
 927
 928 /*
 929  * Attempt to initiate a splice from pipe to file.
 930  */
 931 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 932                            loff_t *ppos, size_t len, unsigned int flags)
 933 {
 934         if (unlikely(!out->f_op->splice_write))
 935                 return warn_unsupported(out, "write");
 936         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 937 }
 938
 939 /*
 940  * Indicate to the caller that there was a premature EOF when reading from the
 941  * source and the caller didn't indicate they would be sending more data after
 942  * this.
 943  */
 944 static void do_splice_eof(struct splice_desc *sd)
 945 {
 946         if (sd->splice_eof)
 947                 sd->splice_eof(sd);
 948 }
 949
 950 /**
 951  * vfs_splice_read - Read data from a file and splice it into a pipe
 952  * @in:         File to splice from
 953  * @ppos:       Input file offset
 954  * @pipe:       Pipe to splice to
 955  * @len:        Number of bytes to splice
 956  * @flags:      Splice modifier flags (SPLICE_F_*)
 957  *
 958  * Splice the requested amount of data from the input file to the pipe.  This
 959  * is synchronous as the caller must hold the pipe lock across the entire
 960  * operation.
 961  *
 962  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 963  * a hole and a negative error code otherwise.
 964  */
 965 long vfs_splice_read(struct file *in, loff_t *ppos,
 966                      struct pipe_inode_info *pipe, size_t len,
 967                      unsigned int flags)
 968 {
 969         unsigned int p_space;
 970         int ret;
 971
 972         if (unlikely(!(in->f_mode & FMODE_READ)))
 973                 return -EBADF;
 974         if (!len)
 975                 return 0;
 976
 977         /* Don't try to read more the pipe has space for. */
 978         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 979         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 980
 981         ret = rw_verify_area(READ, in, ppos, len);
 982         if (unlikely(ret < 0))
 983                 return ret;
 984
 985         if (unlikely(len > MAX_RW_COUNT))
 986                 len = MAX_RW_COUNT;
 987
 988         if (unlikely(!in->f_op->splice_read))
 989                 return warn_unsupported(in, "read");
 990         /*
 991          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 992          * buffer, copy into it and splice that into the pipe.
 993          */
 994         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 995                 return copy_splice_read(in, ppos, pipe, len, flags);
 996         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 997 }
 998 EXPORT_SYMBOL_GPL(vfs_splice_read);
 999
1000 /**
1001  * splice_direct_to_actor - splices data directly between two non-pipes
1002  * @in:         file to splice from
1003  * @sd:         actor information on where to splice to
1004  * @actor:      handles the data splicing
1005  *
1006  * Description:
1007  *    This is a special case helper to splice directly between two
1008  *    points, without requiring an explicit pipe. Internally an allocated
1009  *    pipe is cached in the process, and reused during the lifetime of
1010  *    that process.
1011  *
1012  */
1013 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1014                                splice_direct_actor *actor)
1015 {
1016         struct pipe_inode_info *pipe;
1017         long ret, bytes;
1018         size_t len;
1019         int i, flags, more;
1020
1021         /*
1022          * We require the input to be seekable, as we don't want to randomly
1023          * drop data for eg socket -> socket splicing. Use the piped splicing
1024          * for that!
1025          */
1026         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1027                 return -EINVAL;
1028
1029         /*
1030          * neither in nor out is a pipe, setup an internal pipe attached to
1031          * 'out' and transfer the wanted data from 'in' to 'out' through that
1032          */
1033         pipe = current->splice_pipe;
1034         if (unlikely(!pipe)) {
1035                 pipe = alloc_pipe_info();
1036                 if (!pipe)
1037                         return -ENOMEM;
1038
1039                 /*
1040                  * We don't have an immediate reader, but we'll read the stuff
1041                  * out of the pipe right after the splice_to_pipe(). So set
1042                  * PIPE_READERS appropriately.
1043                  */
1044                 pipe->readers = 1;
1045
1046                 current->splice_pipe = pipe;
1047         }
1048
1049         /*
1050          * Do the splice.
1051          */
1052         bytes = 0;
1053         len = sd->total_len;
1054
1055         /* Don't block on output, we have to drain the direct pipe. */
1056         flags = sd->flags;
1057         sd->flags &= ~SPLICE_F_NONBLOCK;
1058
1059         /*
1060          * We signal MORE until we've read sufficient data to fulfill the
1061          * request and we keep signalling it if the caller set it.
1062          */
1063         more = sd->flags & SPLICE_F_MORE;
1064         sd->flags |= SPLICE_F_MORE;
1065
1066         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1067
1068         while (len) {
1069                 size_t read_len;
1070                 loff_t pos = sd->pos, prev_pos = pos;
1071
1072                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1073                 if (unlikely(ret <= 0))
1074                         goto read_failure;
1075
1076                 read_len = ret;
1077                 sd->total_len = read_len;
1078
1079                 /*
1080                  * If we now have sufficient data to fulfill the request then
1081                  * we clear SPLICE_F_MORE if it was not set initially.
1082                  */
1083                 if (read_len >= len && !more)
1084                         sd->flags &= ~SPLICE_F_MORE;
1085
1086                 /*
1087                  * NOTE: nonblocking mode only applies to the input. We
1088                  * must not do the output in nonblocking mode as then we
1089                  * could get stuck data in the internal pipe:
1090                  */
1091                 ret = actor(pipe, sd);
1092                 if (unlikely(ret <= 0)) {
1093                         sd->pos = prev_pos;
1094                         goto out_release;
1095                 }
1096
1097                 bytes += ret;
1098                 len -= ret;
1099                 sd->pos = pos;
1100
1101                 if (ret < read_len) {
1102                         sd->pos = prev_pos + ret;
1103                         goto out_release;
1104                 }
1105         }
1106
1107 done:
1108         pipe->tail = pipe->head = 0;
1109         file_accessed(in);
1110         return bytes;
1111
1112 read_failure:
1113         /*
1114          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1115          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1116          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1117          * least 1 byte *then* we will also do the ->splice_eof() call.
1118          */
1119         if (ret == 0 && !more && len > 0 && bytes)
1120                 do_splice_eof(sd);
1121 out_release:
1122         /*
1123          * If we did an incomplete transfer we must release
1124          * the pipe buffers in question:
1125          */
1126         for (i = 0; i < pipe->ring_size; i++) {
1127                 struct pipe_buffer *buf = &pipe->bufs[i];
1128
1129                 if (buf->ops)
1130                         pipe_buf_release(pipe, buf);
1131         }
1132
1133         if (!bytes)
1134                 bytes = ret;
1135
1136         goto done;
1137 }
1138 EXPORT_SYMBOL(splice_direct_to_actor);
1139
1140 static int direct_splice_actor(struct pipe_inode_info *pipe,
1141                                struct splice_desc *sd)
1142 {
1143         struct file *file = sd->u.file;
1144
1145         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1146                               sd->flags);
1147 }
1148
1149 static void direct_file_splice_eof(struct splice_desc *sd)
1150 {
1151         struct file *file = sd->u.file;
1152
1153         if (file->f_op->splice_eof)
1154                 file->f_op->splice_eof(file);
1155 }
1156
1157 /**
1158  * do_splice_direct - splices data directly between two files
1159  * @in:         file to splice from
1160  * @ppos:       input file offset
1161  * @out:        file to splice to
1162  * @opos:       output file offset
1163  * @len:        number of bytes to splice
1164  * @flags:      splice modifier flags
1165  *
1166  * Description:
1167  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1168  *    doing it in the application would incur an extra system call
1169  *    (splice in + splice out, as compared to just sendfile()). So this helper
1170  *    can splice directly through a process-private pipe.
1171  *
1172  */
1173 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1174                       loff_t *opos, size_t len, unsigned int flags)
1175 {
1176         struct splice_desc sd = {
1177                 .len            = len,
1178                 .total_len      = len,
1179                 .flags          = flags,
1180                 .pos            = *ppos,
1181                 .u.file         = out,
1182                 .splice_eof     = direct_file_splice_eof,
1183                 .opos           = opos,
1184         };
1185         long ret;
1186
1187         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1188                 return -EBADF;
1189
1190         if (unlikely(out->f_flags & O_APPEND))
1191                 return -EINVAL;
1192
1193         ret = rw_verify_area(WRITE, out, opos, len);
1194         if (unlikely(ret < 0))
1195                 return ret;
1196
1197         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1198         if (ret > 0)
1199                 *ppos = sd.pos;
1200
1201         return ret;
1202 }
1203 EXPORT_SYMBOL(do_splice_direct);
1204
1205 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1206 {
1207         for (;;) {
1208                 if (unlikely(!pipe->readers)) {
1209                         send_sig(SIGPIPE, current, 0);
1210                         return -EPIPE;
1211                 }
1212                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1213                         return 0;
1214                 if (flags & SPLICE_F_NONBLOCK)
1215                         return -EAGAIN;
1216                 if (signal_pending(current))
1217                         return -ERESTARTSYS;
1218                 pipe_wait_writable(pipe);
1219         }
1220 }
1221
1222 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1223                                struct pipe_inode_info *opipe,
1224                                size_t len, unsigned int flags);
1225
1226 long splice_file_to_pipe(struct file *in,
1227                          struct pipe_inode_info *opipe,
1228                          loff_t *offset,
1229                          size_t len, unsigned int flags)
1230 {
1231         long ret;
1232
1233         pipe_lock(opipe);
1234         ret = wait_for_space(opipe, flags);
1235         if (!ret)
1236                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1237         pipe_unlock(opipe);
1238         if (ret > 0)
1239                 wakeup_pipe_readers(opipe);
1240         return ret;
1241 }
1242
1243 /*
1244  * Determine where to splice to/from.
1245  */
1246 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1247                loff_t *off_out, size_t len, unsigned int flags)
1248 {
1249         struct pipe_inode_info *ipipe;
1250         struct pipe_inode_info *opipe;
1251         loff_t offset;
1252         long ret;
1253
1254         if (unlikely(!(in->f_mode & FMODE_READ) ||
1255                      !(out->f_mode & FMODE_WRITE)))
1256                 return -EBADF;
1257
1258         ipipe = get_pipe_info(in, true);
1259         opipe = get_pipe_info(out, true);
1260
1261         if (ipipe && opipe) {
1262                 if (off_in || off_out)
1263                         return -ESPIPE;
1264
1265                 /* Splicing to self would be fun, but... */
1266                 if (ipipe == opipe)
1267                         return -EINVAL;
1268
1269                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1270                         flags |= SPLICE_F_NONBLOCK;
1271
1272                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1273         }
1274
1275         if (ipipe) {
1276                 if (off_in)
1277                         return -ESPIPE;
1278                 if (off_out) {
1279                         if (!(out->f_mode & FMODE_PWRITE))
1280                                 return -EINVAL;
1281                         offset = *off_out;
1282                 } else {
1283                         offset = out->f_pos;
1284                 }
1285
1286                 if (unlikely(out->f_flags & O_APPEND))
1287                         return -EINVAL;
1288
1289                 ret = rw_verify_area(WRITE, out, &offset, len);
1290                 if (unlikely(ret < 0))
1291                         return ret;
1292
1293                 if (in->f_flags & O_NONBLOCK)
1294                         flags |= SPLICE_F_NONBLOCK;
1295
1296                 file_start_write(out);
1297                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1298                 file_end_write(out);
1299
1300                 if (ret > 0)
1301                         fsnotify_modify(out);
1302
1303                 if (!off_out)
1304                         out->f_pos = offset;
1305                 else
1306                         *off_out = offset;
1307
1308                 return ret;
1309         }
1310
1311         if (opipe) {
1312                 if (off_out)
1313                         return -ESPIPE;
1314                 if (off_in) {
1315                         if (!(in->f_mode & FMODE_PREAD))
1316                                 return -EINVAL;
1317                         offset = *off_in;
1318                 } else {
1319                         offset = in->f_pos;
1320                 }
1321
1322                 if (out->f_flags & O_NONBLOCK)
1323                         flags |= SPLICE_F_NONBLOCK;
1324
1325                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1326
1327                 if (ret > 0)
1328                         fsnotify_access(in);
1329
1330                 if (!off_in)
1331                         in->f_pos = offset;
1332                 else
1333                         *off_in = offset;
1334
1335                 return ret;
1336         }
1337
1338         return -EINVAL;
1339 }
1340
1341 static long __do_splice(struct file *in, loff_t __user *off_in,
1342                         struct file *out, loff_t __user *off_out,
1343                         size_t len, unsigned int flags)
1344 {
1345         struct pipe_inode_info *ipipe;
1346         struct pipe_inode_info *opipe;
1347         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1348         long ret;
1349
1350         ipipe = get_pipe_info(in, true);
1351         opipe = get_pipe_info(out, true);
1352
1353         if (ipipe) {
1354                 if (off_in)
1355                         return -ESPIPE;
1356                 pipe_clear_nowait(in);
1357         }
1358         if (opipe) {
1359                 if (off_out)
1360                         return -ESPIPE;
1361                 pipe_clear_nowait(out);
1362         }
1363
1364         if (off_out) {
1365                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1366                         return -EFAULT;
1367                 __off_out = &offset;
1368         }
1369         if (off_in) {
1370                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1371                         return -EFAULT;
1372                 __off_in = &offset;
1373         }
1374
1375         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1376         if (ret < 0)
1377                 return ret;
1378
1379         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1380                 return -EFAULT;
1381         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1382                 return -EFAULT;
1383
1384         return ret;
1385 }
1386
1387 static int iter_to_pipe(struct iov_iter *from,
1388                         struct pipe_inode_info *pipe,
1389                         unsigned flags)
1390 {
1391         struct pipe_buffer buf = {
1392                 .ops = &user_page_pipe_buf_ops,
1393                 .flags = flags
1394         };
1395         size_t total = 0;
1396         int ret = 0;
1397
1398         while (iov_iter_count(from)) {
1399                 struct page *pages[16];
1400                 ssize_t left;
1401                 size_t start;
1402                 int i, n;
1403
1404                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1405                 if (left <= 0) {
1406                         ret = left;
1407                         break;
1408                 }
1409
1410                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1411                 for (i = 0; i < n; i++) {
1412                         int size = min_t(int, left, PAGE_SIZE - start);
1413
1414                         buf.page = pages[i];
1415                         buf.offset = start;
1416                         buf.len = size;
1417                         ret = add_to_pipe(pipe, &buf);
1418                         if (unlikely(ret < 0)) {
1419                                 iov_iter_revert(from, left);
1420                                 // this one got dropped by add_to_pipe()
1421                                 while (++i < n)
1422                                         put_page(pages[i]);
1423                                 goto out;
1424                         }
1425                         total += ret;
1426                         left -= size;
1427                         start = 0;
1428                 }
1429         }
1430 out:
1431         return total ? total : ret;
1432 }
1433
1434 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1435                         struct splice_desc *sd)
1436 {
1437         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1438         return n == sd->len ? n : -EFAULT;
1439 }
1440
1441 /*
1442  * For lack of a better implementation, implement vmsplice() to userspace
1443  * as a simple copy of the pipes pages to the user iov.
1444  */
1445 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1446                              unsigned int flags)
1447 {
1448         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1449         struct splice_desc sd = {
1450                 .total_len = iov_iter_count(iter),
1451                 .flags = flags,
1452                 .u.data = iter
1453         };
1454         long ret = 0;
1455
1456         if (!pipe)
1457                 return -EBADF;
1458
1459         pipe_clear_nowait(file);
1460
1461         if (sd.total_len) {
1462                 pipe_lock(pipe);
1463                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1464                 pipe_unlock(pipe);
1465         }
1466
1467         return ret;
1468 }
1469
1470 /*
1471  * vmsplice splices a user address range into a pipe. It can be thought of
1472  * as splice-from-memory, where the regular splice is splice-from-file (or
1473  * to file). In both cases the output is a pipe, naturally.
1474  */
1475 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1476                              unsigned int flags)
1477 {
1478         struct pipe_inode_info *pipe;
1479         long ret = 0;
1480         unsigned buf_flag = 0;
1481
1482         if (flags & SPLICE_F_GIFT)
1483                 buf_flag = PIPE_BUF_FLAG_GIFT;
1484
1485         pipe = get_pipe_info(file, true);
1486         if (!pipe)
1487                 return -EBADF;
1488
1489         pipe_clear_nowait(file);
1490
1491         pipe_lock(pipe);
1492         ret = wait_for_space(pipe, flags);
1493         if (!ret)
1494                 ret = iter_to_pipe(iter, pipe, buf_flag);
1495         pipe_unlock(pipe);
1496         if (ret > 0)
1497                 wakeup_pipe_readers(pipe);
1498         return ret;
1499 }
1500
1501 static int vmsplice_type(struct fd f, int *type)
1502 {
1503         if (!f.file)
1504                 return -EBADF;
1505         if (f.file->f_mode & FMODE_WRITE) {
1506                 *type = ITER_SOURCE;
1507         } else if (f.file->f_mode & FMODE_READ) {
1508                 *type = ITER_DEST;
1509         } else {
1510                 fdput(f);
1511                 return -EBADF;
1512         }
1513         return 0;
1514 }
1515
1516 /*
1517  * Note that vmsplice only really supports true splicing _from_ user memory
1518  * to a pipe, not the other way around. Splicing from user memory is a simple
1519  * operation that can be supported without any funky alignment restrictions
1520  * or nasty vm tricks. We simply map in the user memory and fill them into
1521  * a pipe. The reverse isn't quite as easy, though. There are two possible
1522  * solutions for that:
1523  *
1524  *      - memcpy() the data internally, at which point we might as well just
1525  *        do a regular read() on the buffer anyway.
1526  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1527  *        has restriction limitations on both ends of the pipe).
1528  *
1529  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1530  *
1531  */
1532 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1533                 unsigned long, nr_segs, unsigned int, flags)
1534 {
1535         struct iovec iovstack[UIO_FASTIOV];
1536         struct iovec *iov = iovstack;
1537         struct iov_iter iter;
1538         ssize_t error;
1539         struct fd f;
1540         int type;
1541
1542         if (unlikely(flags & ~SPLICE_F_ALL))
1543                 return -EINVAL;
1544
1545         f = fdget(fd);
1546         error = vmsplice_type(f, &type);
1547         if (error)
1548                 return error;
1549
1550         error = import_iovec(type, uiov, nr_segs,
1551                              ARRAY_SIZE(iovstack), &iov, &iter);
1552         if (error < 0)
1553                 goto out_fdput;
1554
1555         if (!iov_iter_count(&iter))
1556                 error = 0;
1557         else if (type == ITER_SOURCE)
1558                 error = vmsplice_to_pipe(f.file, &iter, flags);
1559         else
1560                 error = vmsplice_to_user(f.file, &iter, flags);
1561
1562         kfree(iov);
1563 out_fdput:
1564         fdput(f);
1565         return error;
1566 }
1567
1568 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1569                 int, fd_out, loff_t __user *, off_out,
1570                 size_t, len, unsigned int, flags)
1571 {
1572         struct fd in, out;
1573         long error;
1574
1575         if (unlikely(!len))
1576                 return 0;
1577
1578         if (unlikely(flags & ~SPLICE_F_ALL))
1579                 return -EINVAL;
1580
1581         error = -EBADF;
1582         in = fdget(fd_in);
1583         if (in.file) {
1584                 out = fdget(fd_out);
1585                 if (out.file) {
1586                         error = __do_splice(in.file, off_in, out.file, off_out,
1587                                                 len, flags);
1588                         fdput(out);
1589                 }
1590                 fdput(in);
1591         }
1592         return error;
1593 }
1594
1595 /*
1596  * Make sure there's data to read. Wait for input if we can, otherwise
1597  * return an appropriate error.
1598  */
1599 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1600 {
1601         int ret;
1602
1603         /*
1604          * Check the pipe occupancy without the inode lock first. This function
1605          * is speculative anyways, so missing one is ok.
1606          */
1607         if (!pipe_empty(pipe->head, pipe->tail))
1608                 return 0;
1609
1610         ret = 0;
1611         pipe_lock(pipe);
1612
1613         while (pipe_empty(pipe->head, pipe->tail)) {
1614                 if (signal_pending(current)) {
1615                         ret = -ERESTARTSYS;
1616                         break;
1617                 }
1618                 if (!pipe->writers)
1619                         break;
1620                 if (flags & SPLICE_F_NONBLOCK) {
1621                         ret = -EAGAIN;
1622                         break;
1623                 }
1624                 pipe_wait_readable(pipe);
1625         }
1626
1627         pipe_unlock(pipe);
1628         return ret;
1629 }
1630
1631 /*
1632  * Make sure there's writeable room. Wait for room if we can, otherwise
1633  * return an appropriate error.
1634  */
1635 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1636 {
1637         int ret;
1638
1639         /*
1640          * Check pipe occupancy without the inode lock first. This function
1641          * is speculative anyways, so missing one is ok.
1642          */
1643         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1644                 return 0;
1645
1646         ret = 0;
1647         pipe_lock(pipe);
1648
1649         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1650                 if (!pipe->readers) {
1651                         send_sig(SIGPIPE, current, 0);
1652                         ret = -EPIPE;
1653                         break;
1654                 }
1655                 if (flags & SPLICE_F_NONBLOCK) {
1656                         ret = -EAGAIN;
1657                         break;
1658                 }
1659                 if (signal_pending(current)) {
1660                         ret = -ERESTARTSYS;
1661                         break;
1662                 }
1663                 pipe_wait_writable(pipe);
1664         }
1665
1666         pipe_unlock(pipe);
1667         return ret;
1668 }
1669
1670 /*
1671  * Splice contents of ipipe to opipe.
1672  */
1673 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1674                                struct pipe_inode_info *opipe,
1675                                size_t len, unsigned int flags)
1676 {
1677         struct pipe_buffer *ibuf, *obuf;
1678         unsigned int i_head, o_head;
1679         unsigned int i_tail, o_tail;
1680         unsigned int i_mask, o_mask;
1681         int ret = 0;
1682         bool input_wakeup = false;
1683
1684
1685 retry:
1686         ret = ipipe_prep(ipipe, flags);
1687         if (ret)
1688                 return ret;
1689
1690         ret = opipe_prep(opipe, flags);
1691         if (ret)
1692                 return ret;
1693
1694         /*
1695          * Potential ABBA deadlock, work around it by ordering lock
1696          * grabbing by pipe info address. Otherwise two different processes
1697          * could deadlock (one doing tee from A -> B, the other from B -> A).
1698          */
1699         pipe_double_lock(ipipe, opipe);
1700
1701         i_tail = ipipe->tail;
1702         i_mask = ipipe->ring_size - 1;
1703         o_head = opipe->head;
1704         o_mask = opipe->ring_size - 1;
1705
1706         do {
1707                 size_t o_len;
1708
1709                 if (!opipe->readers) {
1710                         send_sig(SIGPIPE, current, 0);
1711                         if (!ret)
1712                                 ret = -EPIPE;
1713                         break;
1714                 }
1715
1716                 i_head = ipipe->head;
1717                 o_tail = opipe->tail;
1718
1719                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1720                         break;
1721
1722                 /*
1723                  * Cannot make any progress, because either the input
1724                  * pipe is empty or the output pipe is full.
1725                  */
1726                 if (pipe_empty(i_head, i_tail) ||
1727                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1728                         /* Already processed some buffers, break */
1729                         if (ret)
1730                                 break;
1731
1732                         if (flags & SPLICE_F_NONBLOCK) {
1733                                 ret = -EAGAIN;
1734                                 break;
1735                         }
1736
1737                         /*
1738                          * We raced with another reader/writer and haven't
1739                          * managed to process any buffers.  A zero return
1740                          * value means EOF, so retry instead.
1741                          */
1742                         pipe_unlock(ipipe);
1743                         pipe_unlock(opipe);
1744                         goto retry;
1745                 }
1746
1747                 ibuf = &ipipe->bufs[i_tail & i_mask];
1748                 obuf = &opipe->bufs[o_head & o_mask];
1749
1750                 if (len >= ibuf->len) {
1751                         /*
1752                          * Simply move the whole buffer from ipipe to opipe
1753                          */
1754                         *obuf = *ibuf;
1755                         ibuf->ops = NULL;
1756                         i_tail++;
1757                         ipipe->tail = i_tail;
1758                         input_wakeup = true;
1759                         o_len = obuf->len;
1760                         o_head++;
1761                         opipe->head = o_head;
1762                 } else {
1763                         /*
1764                          * Get a reference to this pipe buffer,
1765                          * so we can copy the contents over.
1766                          */
1767                         if (!pipe_buf_get(ipipe, ibuf)) {
1768                                 if (ret == 0)
1769                                         ret = -EFAULT;
1770                                 break;
1771                         }
1772                         *obuf = *ibuf;
1773
1774                         /*
1775                          * Don't inherit the gift and merge flags, we need to
1776                          * prevent multiple steals of this page.
1777                          */
1778                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1779                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1780
1781                         obuf->len = len;
1782                         ibuf->offset += len;
1783                         ibuf->len -= len;
1784                         o_len = len;
1785                         o_head++;
1786                         opipe->head = o_head;
1787                 }
1788                 ret += o_len;
1789                 len -= o_len;
1790         } while (len);
1791
1792         pipe_unlock(ipipe);
1793         pipe_unlock(opipe);
1794
1795         /*
1796          * If we put data in the output pipe, wakeup any potential readers.
1797          */
1798         if (ret > 0)
1799                 wakeup_pipe_readers(opipe);
1800
1801         if (input_wakeup)
1802                 wakeup_pipe_writers(ipipe);
1803
1804         return ret;
1805 }
1806
1807 /*
1808  * Link contents of ipipe to opipe.
1809  */
1810 static int link_pipe(struct pipe_inode_info *ipipe,
1811                      struct pipe_inode_info *opipe,
1812                      size_t len, unsigned int flags)
1813 {
1814         struct pipe_buffer *ibuf, *obuf;
1815         unsigned int i_head, o_head;
1816         unsigned int i_tail, o_tail;
1817         unsigned int i_mask, o_mask;
1818         int ret = 0;
1819
1820         /*
1821          * Potential ABBA deadlock, work around it by ordering lock
1822          * grabbing by pipe info address. Otherwise two different processes
1823          * could deadlock (one doing tee from A -> B, the other from B -> A).
1824          */
1825         pipe_double_lock(ipipe, opipe);
1826
1827         i_tail = ipipe->tail;
1828         i_mask = ipipe->ring_size - 1;
1829         o_head = opipe->head;
1830         o_mask = opipe->ring_size - 1;
1831
1832         do {
1833                 if (!opipe->readers) {
1834                         send_sig(SIGPIPE, current, 0);
1835                         if (!ret)
1836                                 ret = -EPIPE;
1837                         break;
1838                 }
1839
1840                 i_head = ipipe->head;
1841                 o_tail = opipe->tail;
1842
1843                 /*
1844                  * If we have iterated all input buffers or run out of
1845                  * output room, break.
1846                  */
1847                 if (pipe_empty(i_head, i_tail) ||
1848                     pipe_full(o_head, o_tail, opipe->max_usage))
1849                         break;
1850
1851                 ibuf = &ipipe->bufs[i_tail & i_mask];
1852                 obuf = &opipe->bufs[o_head & o_mask];
1853
1854                 /*
1855                  * Get a reference to this pipe buffer,
1856                  * so we can copy the contents over.
1857                  */
1858                 if (!pipe_buf_get(ipipe, ibuf)) {
1859                         if (ret == 0)
1860                                 ret = -EFAULT;
1861                         break;
1862                 }
1863
1864                 *obuf = *ibuf;
1865
1866                 /*
1867                  * Don't inherit the gift and merge flag, we need to prevent
1868                  * multiple steals of this page.
1869                  */
1870                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1871                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1872
1873                 if (obuf->len > len)
1874                         obuf->len = len;
1875                 ret += obuf->len;
1876                 len -= obuf->len;
1877
1878                 o_head++;
1879                 opipe->head = o_head;
1880                 i_tail++;
1881         } while (len);
1882
1883         pipe_unlock(ipipe);
1884         pipe_unlock(opipe);
1885
1886         /*
1887          * If we put data in the output pipe, wakeup any potential readers.
1888          */
1889         if (ret > 0)
1890                 wakeup_pipe_readers(opipe);
1891
1892         return ret;
1893 }
1894
1895 /*
1896  * This is a tee(1) implementation that works on pipes. It doesn't copy
1897  * any data, it simply references the 'in' pages on the 'out' pipe.
1898  * The 'flags' used are the SPLICE_F_* variants, currently the only
1899  * applicable one is SPLICE_F_NONBLOCK.
1900  */
1901 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1902 {
1903         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1904         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1905         int ret = -EINVAL;
1906
1907         if (unlikely(!(in->f_mode & FMODE_READ) ||
1908                      !(out->f_mode & FMODE_WRITE)))
1909                 return -EBADF;
1910
1911         /*
1912          * Duplicate the contents of ipipe to opipe without actually
1913          * copying the data.
1914          */
1915         if (ipipe && opipe && ipipe != opipe) {
1916                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1917                         flags |= SPLICE_F_NONBLOCK;
1918
1919                 /*
1920                  * Keep going, unless we encounter an error. The ipipe/opipe
1921                  * ordering doesn't really matter.
1922                  */
1923                 ret = ipipe_prep(ipipe, flags);
1924                 if (!ret) {
1925                         ret = opipe_prep(opipe, flags);
1926                         if (!ret)
1927                                 ret = link_pipe(ipipe, opipe, len, flags);
1928                 }
1929         }
1930
1931         return ret;
1932 }
1933
1934 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1935 {
1936         struct fd in, out;
1937         int error;
1938
1939         if (unlikely(flags & ~SPLICE_F_ALL))
1940                 return -EINVAL;
1941
1942         if (unlikely(!len))
1943                 return 0;
1944
1945         error = -EBADF;
1946         in = fdget(fdin);
1947         if (in.file) {
1948                 out = fdget(fdout);
1949                 if (out.file) {
1950                         error = do_tee(in.file, out.file, len, flags);
1951                         fdput(out);
1952                 }
1953                 fdput(in);
1954         }
1955
1956         return error;
1957 }