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