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
 880                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 881                               len - remain);
 882                 ret = sock_sendmsg(sock, &msg);
 883                 if (ret <= 0)
 884                         break;
 885
 886                 spliced += ret;
 887                 len -= ret;
 888                 tail = pipe->tail;
 889                 while (ret > 0) {
 890                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 891                         size_t seg = min_t(size_t, ret, buf->len);
 892
 893                         buf->offset += seg;
 894                         buf->len -= seg;
 895                         ret -= seg;
 896
 897                         if (!buf->len) {
 898                                 pipe_buf_release(pipe, buf);
 899                                 tail++;
 900                         }
 901                 }
 902
 903                 if (tail != pipe->tail) {
 904                         pipe->tail = tail;
 905                         if (pipe->files)
 906                                 need_wakeup = true;
 907                 }
 908         }
 909
 910 out:
 911         pipe_unlock(pipe);
 912         if (need_wakeup)
 913                 wakeup_pipe_writers(pipe);
 914         return spliced ?: ret;
 915 }
 916 #endif
 917
 918 static int warn_unsupported(struct file *file, const char *op)
 919 {
 920         pr_debug_ratelimited(
 921                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 922                 op, file, current->pid, current->comm);
 923         return -EINVAL;
 924 }
 925
 926 /*
 927  * Attempt to initiate a splice from pipe to file.
 928  */
 929 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 930                            loff_t *ppos, size_t len, unsigned int flags)
 931 {
 932         if (unlikely(!out->f_op->splice_write))
 933                 return warn_unsupported(out, "write");
 934         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 935 }
 936
 937 /*
 938  * Indicate to the caller that there was a premature EOF when reading from the
 939  * source and the caller didn't indicate they would be sending more data after
 940  * this.
 941  */
 942 static void do_splice_eof(struct splice_desc *sd)
 943 {
 944         if (sd->splice_eof)
 945                 sd->splice_eof(sd);
 946 }
 947
 948 /**
 949  * vfs_splice_read - Read data from a file and splice it into a pipe
 950  * @in:         File to splice from
 951  * @ppos:       Input file offset
 952  * @pipe:       Pipe to splice to
 953  * @len:        Number of bytes to splice
 954  * @flags:      Splice modifier flags (SPLICE_F_*)
 955  *
 956  * Splice the requested amount of data from the input file to the pipe.  This
 957  * is synchronous as the caller must hold the pipe lock across the entire
 958  * operation.
 959  *
 960  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 961  * a hole and a negative error code otherwise.
 962  */
 963 long vfs_splice_read(struct file *in, loff_t *ppos,
 964                      struct pipe_inode_info *pipe, size_t len,
 965                      unsigned int flags)
 966 {
 967         unsigned int p_space;
 968         int ret;
 969
 970         if (unlikely(!(in->f_mode & FMODE_READ)))
 971                 return -EBADF;
 972         if (!len)
 973                 return 0;
 974
 975         /* Don't try to read more the pipe has space for. */
 976         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 977         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 978
 979         ret = rw_verify_area(READ, in, ppos, len);
 980         if (unlikely(ret < 0))
 981                 return ret;
 982
 983         if (unlikely(len > MAX_RW_COUNT))
 984                 len = MAX_RW_COUNT;
 985
 986         if (unlikely(!in->f_op->splice_read))
 987                 return warn_unsupported(in, "read");
 988         /*
 989          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 990          * buffer, copy into it and splice that into the pipe.
 991          */
 992         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 993                 return copy_splice_read(in, ppos, pipe, len, flags);
 994         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 995 }
 996 EXPORT_SYMBOL_GPL(vfs_splice_read);
 997
 998 /**
 999  * splice_direct_to_actor - splices data directly between two non-pipes
1000  * @in:         file to splice from
1001  * @sd:         actor information on where to splice to
1002  * @actor:      handles the data splicing
1003  *
1004  * Description:
1005  *    This is a special case helper to splice directly between two
1006  *    points, without requiring an explicit pipe. Internally an allocated
1007  *    pipe is cached in the process, and reused during the lifetime of
1008  *    that process.
1009  *
1010  */
1011 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1012                                splice_direct_actor *actor)
1013 {
1014         struct pipe_inode_info *pipe;
1015         long ret, bytes;
1016         size_t len;
1017         int i, flags, more;
1018
1019         /*
1020          * We require the input to be seekable, as we don't want to randomly
1021          * drop data for eg socket -> socket splicing. Use the piped splicing
1022          * for that!
1023          */
1024         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1025                 return -EINVAL;
1026
1027         /*
1028          * neither in nor out is a pipe, setup an internal pipe attached to
1029          * 'out' and transfer the wanted data from 'in' to 'out' through that
1030          */
1031         pipe = current->splice_pipe;
1032         if (unlikely(!pipe)) {
1033                 pipe = alloc_pipe_info();
1034                 if (!pipe)
1035                         return -ENOMEM;
1036
1037                 /*
1038                  * We don't have an immediate reader, but we'll read the stuff
1039                  * out of the pipe right after the splice_to_pipe(). So set
1040                  * PIPE_READERS appropriately.
1041                  */
1042                 pipe->readers = 1;
1043
1044                 current->splice_pipe = pipe;
1045         }
1046
1047         /*
1048          * Do the splice.
1049          */
1050         bytes = 0;
1051         len = sd->total_len;
1052
1053         /* Don't block on output, we have to drain the direct pipe. */
1054         flags = sd->flags;
1055         sd->flags &= ~SPLICE_F_NONBLOCK;
1056
1057         /*
1058          * We signal MORE until we've read sufficient data to fulfill the
1059          * request and we keep signalling it if the caller set it.
1060          */
1061         more = sd->flags & SPLICE_F_MORE;
1062         sd->flags |= SPLICE_F_MORE;
1063
1064         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1065
1066         while (len) {
1067                 size_t read_len;
1068                 loff_t pos = sd->pos, prev_pos = pos;
1069
1070                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1071                 if (unlikely(ret <= 0))
1072                         goto read_failure;
1073
1074                 read_len = ret;
1075                 sd->total_len = read_len;
1076
1077                 /*
1078                  * If we now have sufficient data to fulfill the request then
1079                  * we clear SPLICE_F_MORE if it was not set initially.
1080                  */
1081                 if (read_len >= len && !more)
1082                         sd->flags &= ~SPLICE_F_MORE;
1083
1084                 /*
1085                  * NOTE: nonblocking mode only applies to the input. We
1086                  * must not do the output in nonblocking mode as then we
1087                  * could get stuck data in the internal pipe:
1088                  */
1089                 ret = actor(pipe, sd);
1090                 if (unlikely(ret <= 0)) {
1091                         sd->pos = prev_pos;
1092                         goto out_release;
1093                 }
1094
1095                 bytes += ret;
1096                 len -= ret;
1097                 sd->pos = pos;
1098
1099                 if (ret < read_len) {
1100                         sd->pos = prev_pos + ret;
1101                         goto out_release;
1102                 }
1103         }
1104
1105 done:
1106         pipe->tail = pipe->head = 0;
1107         file_accessed(in);
1108         return bytes;
1109
1110 read_failure:
1111         /*
1112          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1113          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1114          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1115          * least 1 byte *then* we will also do the ->splice_eof() call.
1116          */
1117         if (ret == 0 && !more && len > 0 && bytes)
1118                 do_splice_eof(sd);
1119 out_release:
1120         /*
1121          * If we did an incomplete transfer we must release
1122          * the pipe buffers in question:
1123          */
1124         for (i = 0; i < pipe->ring_size; i++) {
1125                 struct pipe_buffer *buf = &pipe->bufs[i];
1126
1127                 if (buf->ops)
1128                         pipe_buf_release(pipe, buf);
1129         }
1130
1131         if (!bytes)
1132                 bytes = ret;
1133
1134         goto done;
1135 }
1136 EXPORT_SYMBOL(splice_direct_to_actor);
1137
1138 static int direct_splice_actor(struct pipe_inode_info *pipe,
1139                                struct splice_desc *sd)
1140 {
1141         struct file *file = sd->u.file;
1142
1143         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1144                               sd->flags);
1145 }
1146
1147 static void direct_file_splice_eof(struct splice_desc *sd)
1148 {
1149         struct file *file = sd->u.file;
1150
1151         if (file->f_op->splice_eof)
1152                 file->f_op->splice_eof(file);
1153 }
1154
1155 /**
1156  * do_splice_direct - splices data directly between two files
1157  * @in:         file to splice from
1158  * @ppos:       input file offset
1159  * @out:        file to splice to
1160  * @opos:       output file offset
1161  * @len:        number of bytes to splice
1162  * @flags:      splice modifier flags
1163  *
1164  * Description:
1165  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1166  *    doing it in the application would incur an extra system call
1167  *    (splice in + splice out, as compared to just sendfile()). So this helper
1168  *    can splice directly through a process-private pipe.
1169  *
1170  */
1171 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1172                       loff_t *opos, size_t len, unsigned int flags)
1173 {
1174         struct splice_desc sd = {
1175                 .len            = len,
1176                 .total_len      = len,
1177                 .flags          = flags,
1178                 .pos            = *ppos,
1179                 .u.file         = out,
1180                 .splice_eof     = direct_file_splice_eof,
1181                 .opos           = opos,
1182         };
1183         long ret;
1184
1185         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1186                 return -EBADF;
1187
1188         if (unlikely(out->f_flags & O_APPEND))
1189                 return -EINVAL;
1190
1191         ret = rw_verify_area(WRITE, out, opos, len);
1192         if (unlikely(ret < 0))
1193                 return ret;
1194
1195         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1196         if (ret > 0)
1197                 *ppos = sd.pos;
1198
1199         return ret;
1200 }
1201 EXPORT_SYMBOL(do_splice_direct);
1202
1203 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1204 {
1205         for (;;) {
1206                 if (unlikely(!pipe->readers)) {
1207                         send_sig(SIGPIPE, current, 0);
1208                         return -EPIPE;
1209                 }
1210                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1211                         return 0;
1212                 if (flags & SPLICE_F_NONBLOCK)
1213                         return -EAGAIN;
1214                 if (signal_pending(current))
1215                         return -ERESTARTSYS;
1216                 pipe_wait_writable(pipe);
1217         }
1218 }
1219
1220 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1221                                struct pipe_inode_info *opipe,
1222                                size_t len, unsigned int flags);
1223
1224 long splice_file_to_pipe(struct file *in,
1225                          struct pipe_inode_info *opipe,
1226                          loff_t *offset,
1227                          size_t len, unsigned int flags)
1228 {
1229         long ret;
1230
1231         pipe_lock(opipe);
1232         ret = wait_for_space(opipe, flags);
1233         if (!ret)
1234                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1235         pipe_unlock(opipe);
1236         if (ret > 0)
1237                 wakeup_pipe_readers(opipe);
1238         return ret;
1239 }
1240
1241 /*
1242  * Determine where to splice to/from.
1243  */
1244 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1245                loff_t *off_out, size_t len, unsigned int flags)
1246 {
1247         struct pipe_inode_info *ipipe;
1248         struct pipe_inode_info *opipe;
1249         loff_t offset;
1250         long ret;
1251
1252         if (unlikely(!(in->f_mode & FMODE_READ) ||
1253                      !(out->f_mode & FMODE_WRITE)))
1254                 return -EBADF;
1255
1256         ipipe = get_pipe_info(in, true);
1257         opipe = get_pipe_info(out, true);
1258
1259         if (ipipe && opipe) {
1260                 if (off_in || off_out)
1261                         return -ESPIPE;
1262
1263                 /* Splicing to self would be fun, but... */
1264                 if (ipipe == opipe)
1265                         return -EINVAL;
1266
1267                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1268                         flags |= SPLICE_F_NONBLOCK;
1269
1270                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1271         }
1272
1273         if (ipipe) {
1274                 if (off_in)
1275                         return -ESPIPE;
1276                 if (off_out) {
1277                         if (!(out->f_mode & FMODE_PWRITE))
1278                                 return -EINVAL;
1279                         offset = *off_out;
1280                 } else {
1281                         offset = out->f_pos;
1282                 }
1283
1284                 if (unlikely(out->f_flags & O_APPEND))
1285                         return -EINVAL;
1286
1287                 ret = rw_verify_area(WRITE, out, &offset, len);
1288                 if (unlikely(ret < 0))
1289                         return ret;
1290
1291                 if (in->f_flags & O_NONBLOCK)
1292                         flags |= SPLICE_F_NONBLOCK;
1293
1294                 file_start_write(out);
1295                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1296                 file_end_write(out);
1297
1298                 if (ret > 0)
1299                         fsnotify_modify(out);
1300
1301                 if (!off_out)
1302                         out->f_pos = offset;
1303                 else
1304                         *off_out = offset;
1305
1306                 return ret;
1307         }
1308
1309         if (opipe) {
1310                 if (off_out)
1311                         return -ESPIPE;
1312                 if (off_in) {
1313                         if (!(in->f_mode & FMODE_PREAD))
1314                                 return -EINVAL;
1315                         offset = *off_in;
1316                 } else {
1317                         offset = in->f_pos;
1318                 }
1319
1320                 if (out->f_flags & O_NONBLOCK)
1321                         flags |= SPLICE_F_NONBLOCK;
1322
1323                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1324
1325                 if (ret > 0)
1326                         fsnotify_access(in);
1327
1328                 if (!off_in)
1329                         in->f_pos = offset;
1330                 else
1331                         *off_in = offset;
1332
1333                 return ret;
1334         }
1335
1336         return -EINVAL;
1337 }
1338
1339 static long __do_splice(struct file *in, loff_t __user *off_in,
1340                         struct file *out, loff_t __user *off_out,
1341                         size_t len, unsigned int flags)
1342 {
1343         struct pipe_inode_info *ipipe;
1344         struct pipe_inode_info *opipe;
1345         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1346         long ret;
1347
1348         ipipe = get_pipe_info(in, true);
1349         opipe = get_pipe_info(out, true);
1350
1351         if (ipipe) {
1352                 if (off_in)
1353                         return -ESPIPE;
1354                 pipe_clear_nowait(in);
1355         }
1356         if (opipe) {
1357                 if (off_out)
1358                         return -ESPIPE;
1359                 pipe_clear_nowait(out);
1360         }
1361
1362         if (off_out) {
1363                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1364                         return -EFAULT;
1365                 __off_out = &offset;
1366         }
1367         if (off_in) {
1368                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1369                         return -EFAULT;
1370                 __off_in = &offset;
1371         }
1372
1373         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1374         if (ret < 0)
1375                 return ret;
1376
1377         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1378                 return -EFAULT;
1379         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1380                 return -EFAULT;
1381
1382         return ret;
1383 }
1384
1385 static int iter_to_pipe(struct iov_iter *from,
1386                         struct pipe_inode_info *pipe,
1387                         unsigned flags)
1388 {
1389         struct pipe_buffer buf = {
1390                 .ops = &user_page_pipe_buf_ops,
1391                 .flags = flags
1392         };
1393         size_t total = 0;
1394         int ret = 0;
1395
1396         while (iov_iter_count(from)) {
1397                 struct page *pages[16];
1398                 ssize_t left;
1399                 size_t start;
1400                 int i, n;
1401
1402                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1403                 if (left <= 0) {
1404                         ret = left;
1405                         break;
1406                 }
1407
1408                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1409                 for (i = 0; i < n; i++) {
1410                         int size = min_t(int, left, PAGE_SIZE - start);
1411
1412                         buf.page = pages[i];
1413                         buf.offset = start;
1414                         buf.len = size;
1415                         ret = add_to_pipe(pipe, &buf);
1416                         if (unlikely(ret < 0)) {
1417                                 iov_iter_revert(from, left);
1418                                 // this one got dropped by add_to_pipe()
1419                                 while (++i < n)
1420                                         put_page(pages[i]);
1421                                 goto out;
1422                         }
1423                         total += ret;
1424                         left -= size;
1425                         start = 0;
1426                 }
1427         }
1428 out:
1429         return total ? total : ret;
1430 }
1431
1432 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1433                         struct splice_desc *sd)
1434 {
1435         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1436         return n == sd->len ? n : -EFAULT;
1437 }
1438
1439 /*
1440  * For lack of a better implementation, implement vmsplice() to userspace
1441  * as a simple copy of the pipes pages to the user iov.
1442  */
1443 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1444                              unsigned int flags)
1445 {
1446         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1447         struct splice_desc sd = {
1448                 .total_len = iov_iter_count(iter),
1449                 .flags = flags,
1450                 .u.data = iter
1451         };
1452         long ret = 0;
1453
1454         if (!pipe)
1455                 return -EBADF;
1456
1457         pipe_clear_nowait(file);
1458
1459         if (sd.total_len) {
1460                 pipe_lock(pipe);
1461                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1462                 pipe_unlock(pipe);
1463         }
1464
1465         return ret;
1466 }
1467
1468 /*
1469  * vmsplice splices a user address range into a pipe. It can be thought of
1470  * as splice-from-memory, where the regular splice is splice-from-file (or
1471  * to file). In both cases the output is a pipe, naturally.
1472  */
1473 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1474                              unsigned int flags)
1475 {
1476         struct pipe_inode_info *pipe;
1477         long ret = 0;
1478         unsigned buf_flag = 0;
1479
1480         if (flags & SPLICE_F_GIFT)
1481                 buf_flag = PIPE_BUF_FLAG_GIFT;
1482
1483         pipe = get_pipe_info(file, true);
1484         if (!pipe)
1485                 return -EBADF;
1486
1487         pipe_clear_nowait(file);
1488
1489         pipe_lock(pipe);
1490         ret = wait_for_space(pipe, flags);
1491         if (!ret)
1492                 ret = iter_to_pipe(iter, pipe, buf_flag);
1493         pipe_unlock(pipe);
1494         if (ret > 0)
1495                 wakeup_pipe_readers(pipe);
1496         return ret;
1497 }
1498
1499 static int vmsplice_type(struct fd f, int *type)
1500 {
1501         if (!f.file)
1502                 return -EBADF;
1503         if (f.file->f_mode & FMODE_WRITE) {
1504                 *type = ITER_SOURCE;
1505         } else if (f.file->f_mode & FMODE_READ) {
1506                 *type = ITER_DEST;
1507         } else {
1508                 fdput(f);
1509                 return -EBADF;
1510         }
1511         return 0;
1512 }
1513
1514 /*
1515  * Note that vmsplice only really supports true splicing _from_ user memory
1516  * to a pipe, not the other way around. Splicing from user memory is a simple
1517  * operation that can be supported without any funky alignment restrictions
1518  * or nasty vm tricks. We simply map in the user memory and fill them into
1519  * a pipe. The reverse isn't quite as easy, though. There are two possible
1520  * solutions for that:
1521  *
1522  *      - memcpy() the data internally, at which point we might as well just
1523  *        do a regular read() on the buffer anyway.
1524  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1525  *        has restriction limitations on both ends of the pipe).
1526  *
1527  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1528  *
1529  */
1530 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1531                 unsigned long, nr_segs, unsigned int, flags)
1532 {
1533         struct iovec iovstack[UIO_FASTIOV];
1534         struct iovec *iov = iovstack;
1535         struct iov_iter iter;
1536         ssize_t error;
1537         struct fd f;
1538         int type;
1539
1540         if (unlikely(flags & ~SPLICE_F_ALL))
1541                 return -EINVAL;
1542
1543         f = fdget(fd);
1544         error = vmsplice_type(f, &type);
1545         if (error)
1546                 return error;
1547
1548         error = import_iovec(type, uiov, nr_segs,
1549                              ARRAY_SIZE(iovstack), &iov, &iter);
1550         if (error < 0)
1551                 goto out_fdput;
1552
1553         if (!iov_iter_count(&iter))
1554                 error = 0;
1555         else if (type == ITER_SOURCE)
1556                 error = vmsplice_to_pipe(f.file, &iter, flags);
1557         else
1558                 error = vmsplice_to_user(f.file, &iter, flags);
1559
1560         kfree(iov);
1561 out_fdput:
1562         fdput(f);
1563         return error;
1564 }
1565
1566 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1567                 int, fd_out, loff_t __user *, off_out,
1568                 size_t, len, unsigned int, flags)
1569 {
1570         struct fd in, out;
1571         long error;
1572
1573         if (unlikely(!len))
1574                 return 0;
1575
1576         if (unlikely(flags & ~SPLICE_F_ALL))
1577                 return -EINVAL;
1578
1579         error = -EBADF;
1580         in = fdget(fd_in);
1581         if (in.file) {
1582                 out = fdget(fd_out);
1583                 if (out.file) {
1584                         error = __do_splice(in.file, off_in, out.file, off_out,
1585                                                 len, flags);
1586                         fdput(out);
1587                 }
1588                 fdput(in);
1589         }
1590         return error;
1591 }
1592
1593 /*
1594  * Make sure there's data to read. Wait for input if we can, otherwise
1595  * return an appropriate error.
1596  */
1597 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1598 {
1599         int ret;
1600
1601         /*
1602          * Check the pipe occupancy without the inode lock first. This function
1603          * is speculative anyways, so missing one is ok.
1604          */
1605         if (!pipe_empty(pipe->head, pipe->tail))
1606                 return 0;
1607
1608         ret = 0;
1609         pipe_lock(pipe);
1610
1611         while (pipe_empty(pipe->head, pipe->tail)) {
1612                 if (signal_pending(current)) {
1613                         ret = -ERESTARTSYS;
1614                         break;
1615                 }
1616                 if (!pipe->writers)
1617                         break;
1618                 if (flags & SPLICE_F_NONBLOCK) {
1619                         ret = -EAGAIN;
1620                         break;
1621                 }
1622                 pipe_wait_readable(pipe);
1623         }
1624
1625         pipe_unlock(pipe);
1626         return ret;
1627 }
1628
1629 /*
1630  * Make sure there's writeable room. Wait for room if we can, otherwise
1631  * return an appropriate error.
1632  */
1633 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1634 {
1635         int ret;
1636
1637         /*
1638          * Check pipe occupancy without the inode lock first. This function
1639          * is speculative anyways, so missing one is ok.
1640          */
1641         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1642                 return 0;
1643
1644         ret = 0;
1645         pipe_lock(pipe);
1646
1647         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1648                 if (!pipe->readers) {
1649                         send_sig(SIGPIPE, current, 0);
1650                         ret = -EPIPE;
1651                         break;
1652                 }
1653                 if (flags & SPLICE_F_NONBLOCK) {
1654                         ret = -EAGAIN;
1655                         break;
1656                 }
1657                 if (signal_pending(current)) {
1658                         ret = -ERESTARTSYS;
1659                         break;
1660                 }
1661                 pipe_wait_writable(pipe);
1662         }
1663
1664         pipe_unlock(pipe);
1665         return ret;
1666 }
1667
1668 /*
1669  * Splice contents of ipipe to opipe.
1670  */
1671 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1672                                struct pipe_inode_info *opipe,
1673                                size_t len, unsigned int flags)
1674 {
1675         struct pipe_buffer *ibuf, *obuf;
1676         unsigned int i_head, o_head;
1677         unsigned int i_tail, o_tail;
1678         unsigned int i_mask, o_mask;
1679         int ret = 0;
1680         bool input_wakeup = false;
1681
1682
1683 retry:
1684         ret = ipipe_prep(ipipe, flags);
1685         if (ret)
1686                 return ret;
1687
1688         ret = opipe_prep(opipe, flags);
1689         if (ret)
1690                 return ret;
1691
1692         /*
1693          * Potential ABBA deadlock, work around it by ordering lock
1694          * grabbing by pipe info address. Otherwise two different processes
1695          * could deadlock (one doing tee from A -> B, the other from B -> A).
1696          */
1697         pipe_double_lock(ipipe, opipe);
1698
1699         i_tail = ipipe->tail;
1700         i_mask = ipipe->ring_size - 1;
1701         o_head = opipe->head;
1702         o_mask = opipe->ring_size - 1;
1703
1704         do {
1705                 size_t o_len;
1706
1707                 if (!opipe->readers) {
1708                         send_sig(SIGPIPE, current, 0);
1709                         if (!ret)
1710                                 ret = -EPIPE;
1711                         break;
1712                 }
1713
1714                 i_head = ipipe->head;
1715                 o_tail = opipe->tail;
1716
1717                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1718                         break;
1719
1720                 /*
1721                  * Cannot make any progress, because either the input
1722                  * pipe is empty or the output pipe is full.
1723                  */
1724                 if (pipe_empty(i_head, i_tail) ||
1725                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1726                         /* Already processed some buffers, break */
1727                         if (ret)
1728                                 break;
1729
1730                         if (flags & SPLICE_F_NONBLOCK) {
1731                                 ret = -EAGAIN;
1732                                 break;
1733                         }
1734
1735                         /*
1736                          * We raced with another reader/writer and haven't
1737                          * managed to process any buffers.  A zero return
1738                          * value means EOF, so retry instead.
1739                          */
1740                         pipe_unlock(ipipe);
1741                         pipe_unlock(opipe);
1742                         goto retry;
1743                 }
1744
1745                 ibuf = &ipipe->bufs[i_tail & i_mask];
1746                 obuf = &opipe->bufs[o_head & o_mask];
1747
1748                 if (len >= ibuf->len) {
1749                         /*
1750                          * Simply move the whole buffer from ipipe to opipe
1751                          */
1752                         *obuf = *ibuf;
1753                         ibuf->ops = NULL;
1754                         i_tail++;
1755                         ipipe->tail = i_tail;
1756                         input_wakeup = true;
1757                         o_len = obuf->len;
1758                         o_head++;
1759                         opipe->head = o_head;
1760                 } else {
1761                         /*
1762                          * Get a reference to this pipe buffer,
1763                          * so we can copy the contents over.
1764                          */
1765                         if (!pipe_buf_get(ipipe, ibuf)) {
1766                                 if (ret == 0)
1767                                         ret = -EFAULT;
1768                                 break;
1769                         }
1770                         *obuf = *ibuf;
1771
1772                         /*
1773                          * Don't inherit the gift and merge flags, we need to
1774                          * prevent multiple steals of this page.
1775                          */
1776                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1777                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1778
1779                         obuf->len = len;
1780                         ibuf->offset += len;
1781                         ibuf->len -= len;
1782                         o_len = len;
1783                         o_head++;
1784                         opipe->head = o_head;
1785                 }
1786                 ret += o_len;
1787                 len -= o_len;
1788         } while (len);
1789
1790         pipe_unlock(ipipe);
1791         pipe_unlock(opipe);
1792
1793         /*
1794          * If we put data in the output pipe, wakeup any potential readers.
1795          */
1796         if (ret > 0)
1797                 wakeup_pipe_readers(opipe);
1798
1799         if (input_wakeup)
1800                 wakeup_pipe_writers(ipipe);
1801
1802         return ret;
1803 }
1804
1805 /*
1806  * Link contents of ipipe to opipe.
1807  */
1808 static int link_pipe(struct pipe_inode_info *ipipe,
1809                      struct pipe_inode_info *opipe,
1810                      size_t len, unsigned int flags)
1811 {
1812         struct pipe_buffer *ibuf, *obuf;
1813         unsigned int i_head, o_head;
1814         unsigned int i_tail, o_tail;
1815         unsigned int i_mask, o_mask;
1816         int ret = 0;
1817
1818         /*
1819          * Potential ABBA deadlock, work around it by ordering lock
1820          * grabbing by pipe info address. Otherwise two different processes
1821          * could deadlock (one doing tee from A -> B, the other from B -> A).
1822          */
1823         pipe_double_lock(ipipe, opipe);
1824
1825         i_tail = ipipe->tail;
1826         i_mask = ipipe->ring_size - 1;
1827         o_head = opipe->head;
1828         o_mask = opipe->ring_size - 1;
1829
1830         do {
1831                 if (!opipe->readers) {
1832                         send_sig(SIGPIPE, current, 0);
1833                         if (!ret)
1834                                 ret = -EPIPE;
1835                         break;
1836                 }
1837
1838                 i_head = ipipe->head;
1839                 o_tail = opipe->tail;
1840
1841                 /*
1842                  * If we have iterated all input buffers or run out of
1843                  * output room, break.
1844                  */
1845                 if (pipe_empty(i_head, i_tail) ||
1846                     pipe_full(o_head, o_tail, opipe->max_usage))
1847                         break;
1848
1849                 ibuf = &ipipe->bufs[i_tail & i_mask];
1850                 obuf = &opipe->bufs[o_head & o_mask];
1851
1852                 /*
1853                  * Get a reference to this pipe buffer,
1854                  * so we can copy the contents over.
1855                  */
1856                 if (!pipe_buf_get(ipipe, ibuf)) {
1857                         if (ret == 0)
1858                                 ret = -EFAULT;
1859                         break;
1860                 }
1861
1862                 *obuf = *ibuf;
1863
1864                 /*
1865                  * Don't inherit the gift and merge flag, we need to prevent
1866                  * multiple steals of this page.
1867                  */
1868                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1869                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1870
1871                 if (obuf->len > len)
1872                         obuf->len = len;
1873                 ret += obuf->len;
1874                 len -= obuf->len;
1875
1876                 o_head++;
1877                 opipe->head = o_head;
1878                 i_tail++;
1879         } while (len);
1880
1881         pipe_unlock(ipipe);
1882         pipe_unlock(opipe);
1883
1884         /*
1885          * If we put data in the output pipe, wakeup any potential readers.
1886          */
1887         if (ret > 0)
1888                 wakeup_pipe_readers(opipe);
1889
1890         return ret;
1891 }
1892
1893 /*
1894  * This is a tee(1) implementation that works on pipes. It doesn't copy
1895  * any data, it simply references the 'in' pages on the 'out' pipe.
1896  * The 'flags' used are the SPLICE_F_* variants, currently the only
1897  * applicable one is SPLICE_F_NONBLOCK.
1898  */
1899 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1900 {
1901         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1902         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1903         int ret = -EINVAL;
1904
1905         if (unlikely(!(in->f_mode & FMODE_READ) ||
1906                      !(out->f_mode & FMODE_WRITE)))
1907                 return -EBADF;
1908
1909         /*
1910          * Duplicate the contents of ipipe to opipe without actually
1911          * copying the data.
1912          */
1913         if (ipipe && opipe && ipipe != opipe) {
1914                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1915                         flags |= SPLICE_F_NONBLOCK;
1916
1917                 /*
1918                  * Keep going, unless we encounter an error. The ipipe/opipe
1919                  * ordering doesn't really matter.
1920                  */
1921                 ret = ipipe_prep(ipipe, flags);
1922                 if (!ret) {
1923                         ret = opipe_prep(opipe, flags);
1924                         if (!ret)
1925                                 ret = link_pipe(ipipe, opipe, len, flags);
1926                 }
1927         }
1928
1929         return ret;
1930 }
1931
1932 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1933 {
1934         struct fd in, out;
1935         int error;
1936
1937         if (unlikely(flags & ~SPLICE_F_ALL))
1938                 return -EINVAL;
1939
1940         if (unlikely(!len))
1941                 return 0;
1942
1943         error = -EBADF;
1944         in = fdget(fdin);
1945         if (in.file) {
1946                 out = fdget(fdout);
1947                 if (out.file) {
1948                         error = do_tee(in.file, out.file, len, flags);
1949                         fdput(out);
1950                 }
1951                 fdput(in);
1952         }
1953
1954         return error;
1955 }