fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <[email protected]>
  16  * Copyright (C) 2005-2006 Linus Torvalds <[email protected]>
  17  * Copyright (C) 2006 Ingo Molnar <[email protected]>
  18  *
  19  */
  20 #include <linux/bvec.h>
  21 #include <linux/fs.h>
  22 #include <linux/file.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/splice.h>
  25 #include <linux/memcontrol.h>
  26 #include <linux/mm_inline.h>
  27 #include <linux/swap.h>
  28 #include <linux/writeback.h>
  29 #include <linux/export.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33 #include <linux/gfp.h>
  34 #include <linux/socket.h>
  35 #include <linux/compat.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 int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  47                                      struct pipe_buffer *buf)
  48 {
  49         struct page *page = buf->page;
  50         struct address_space *mapping;
  51
  52         lock_page(page);
  53
  54         mapping = page_mapping(page);
  55         if (mapping) {
  56                 WARN_ON(!PageUptodate(page));
  57
  58                 /*
  59                  * At least for ext2 with nobh option, we need to wait on
  60                  * writeback completing on this page, since we'll remove it
  61                  * from the pagecache.  Otherwise truncate wont wait on the
  62                  * page, 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                 wait_on_page_writeback(page);
  67
  68                 if (page_has_private(page) &&
  69                     !try_to_release_page(page, 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, page)) {
  77                         buf->flags |= PIPE_BUF_FLAG_LRU;
  78                         return 0;
  79                 }
  80         }
  81
  82         /*
  83          * Raced with truncate or failed to remove page from current
  84          * address space, unlock and return failure.
  85          */
  86 out_unlock:
  87         unlock_page(page);
  88         return 1;
  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         .can_merge = 0,
 142         .confirm = page_cache_pipe_buf_confirm,
 143         .release = page_cache_pipe_buf_release,
 144         .steal = page_cache_pipe_buf_steal,
 145         .get = generic_pipe_buf_get,
 146 };
 147
 148 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 149                                     struct pipe_buffer *buf)
 150 {
 151         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 152                 return 1;
 153
 154         buf->flags |= PIPE_BUF_FLAG_LRU;
 155         return generic_pipe_buf_steal(pipe, buf);
 156 }
 157
 158 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 159         .can_merge = 0,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 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->nrbufs < pipe->buffers) {
 201                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 202                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 203
 204                 buf->page = spd->pages[page_nr];
 205                 buf->offset = spd->partial[page_nr].offset;
 206                 buf->len = spd->partial[page_nr].len;
 207                 buf->private = spd->partial[page_nr].private;
 208                 buf->ops = spd->ops;
 209                 buf->flags = 0;
 210
 211                 pipe->nrbufs++;
 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         int ret;
 233
 234         if (unlikely(!pipe->readers)) {
 235                 send_sig(SIGPIPE, current, 0);
 236                 ret = -EPIPE;
 237         } else if (pipe->nrbufs == pipe->buffers) {
 238                 ret = -EAGAIN;
 239         } else {
 240                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 241                 pipe->bufs[newbuf] = *buf;
 242                 pipe->nrbufs++;
 243                 return buf->len;
 244         }
 245         pipe_buf_release(pipe, buf);
 246         return ret;
 247 }
 248 EXPORT_SYMBOL(add_to_pipe);
 249
 250 /*
 251  * Check if we need to grow the arrays holding pages and partial page
 252  * descriptions.
 253  */
 254 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 255 {
 256         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 257
 258         spd->nr_pages_max = buffers;
 259         if (buffers <= PIPE_DEF_BUFFERS)
 260                 return 0;
 261
 262         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 263         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 264
 265         if (spd->pages && spd->partial)
 266                 return 0;
 267
 268         kfree(spd->pages);
 269         kfree(spd->partial);
 270         return -ENOMEM;
 271 }
 272
 273 void splice_shrink_spd(struct splice_pipe_desc *spd)
 274 {
 275         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 276                 return;
 277
 278         kfree(spd->pages);
 279         kfree(spd->partial);
 280 }
 281
 282 /**
 283  * generic_file_splice_read - splice data from file to a pipe
 284  * @in:         file to splice from
 285  * @ppos:       position in @in
 286  * @pipe:       pipe to splice to
 287  * @len:        number of bytes to splice
 288  * @flags:      splice modifier flags
 289  *
 290  * Description:
 291  *    Will read pages from given file and fill them into a pipe. Can be
 292  *    used as long as it has more or less sane ->read_iter().
 293  *
 294  */
 295 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 296                                  struct pipe_inode_info *pipe, size_t len,
 297                                  unsigned int flags)
 298 {
 299         struct iov_iter to;
 300         struct kiocb kiocb;
 301         int idx, ret;
 302
 303         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 304         idx = to.idx;
 305         init_sync_kiocb(&kiocb, in);
 306         kiocb.ki_pos = *ppos;
 307         ret = call_read_iter(in, &kiocb, &to);
 308         if (ret > 0) {
 309                 *ppos = kiocb.ki_pos;
 310                 file_accessed(in);
 311         } else if (ret < 0) {
 312                 to.idx = idx;
 313                 to.iov_offset = 0;
 314                 iov_iter_advance(&to, 0); /* to free what was emitted */
 315                 /*
 316                  * callers of ->splice_read() expect -EAGAIN on
 317                  * "can't put anything in there", rather than -EFAULT.
 318                  */
 319                 if (ret == -EFAULT)
 320                         ret = -EAGAIN;
 321         }
 322
 323         return ret;
 324 }
 325 EXPORT_SYMBOL(generic_file_splice_read);
 326
 327 const struct pipe_buf_operations default_pipe_buf_ops = {
 328         .can_merge = 0,
 329         .confirm = generic_pipe_buf_confirm,
 330         .release = generic_pipe_buf_release,
 331         .steal = generic_pipe_buf_steal,
 332         .get = generic_pipe_buf_get,
 333 };
 334
 335 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 336                                     struct pipe_buffer *buf)
 337 {
 338         return 1;
 339 }
 340
 341 /* Pipe buffer operations for a socket and similar. */
 342 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 343         .can_merge = 0,
 344         .confirm = generic_pipe_buf_confirm,
 345         .release = generic_pipe_buf_release,
 346         .steal = generic_pipe_buf_nosteal,
 347         .get = generic_pipe_buf_get,
 348 };
 349 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 350
 351 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 352                             unsigned long vlen, loff_t offset)
 353 {
 354         mm_segment_t old_fs;
 355         loff_t pos = offset;
 356         ssize_t res;
 357
 358         old_fs = get_fs();
 359         set_fs(get_ds());
 360         /* The cast to a user pointer is valid due to the set_fs() */
 361         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 362         set_fs(old_fs);
 363
 364         return res;
 365 }
 366
 367 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 368                             loff_t pos)
 369 {
 370         mm_segment_t old_fs;
 371         ssize_t res;
 372
 373         old_fs = get_fs();
 374         set_fs(get_ds());
 375         /* The cast to a user pointer is valid due to the set_fs() */
 376         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 377         set_fs(old_fs);
 378
 379         return res;
 380 }
 381 EXPORT_SYMBOL(kernel_write);
 382
 383 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 384                                  struct pipe_inode_info *pipe, size_t len,
 385                                  unsigned int flags)
 386 {
 387         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 388         struct iov_iter to;
 389         struct page **pages;
 390         unsigned int nr_pages;
 391         size_t offset, base, copied = 0;
 392         ssize_t res;
 393         int i;
 394
 395         if (pipe->nrbufs == pipe->buffers)
 396                 return -EAGAIN;
 397
 398         /*
 399          * Try to keep page boundaries matching to source pagecache ones -
 400          * it probably won't be much help, but...
 401          */
 402         offset = *ppos & ~PAGE_MASK;
 403
 404         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 405
 406         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 407         if (res <= 0)
 408                 return -ENOMEM;
 409
 410         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 411
 412         vec = __vec;
 413         if (nr_pages > PIPE_DEF_BUFFERS) {
 414                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 415                 if (unlikely(!vec)) {
 416                         res = -ENOMEM;
 417                         goto out;
 418                 }
 419         }
 420
 421         pipe->bufs[to.idx].offset = offset;
 422         pipe->bufs[to.idx].len -= offset;
 423
 424         for (i = 0; i < nr_pages; i++) {
 425                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 426                 vec[i].iov_base = page_address(pages[i]) + offset;
 427                 vec[i].iov_len = this_len;
 428                 len -= this_len;
 429                 offset = 0;
 430         }
 431
 432         res = kernel_readv(in, vec, nr_pages, *ppos);
 433         if (res > 0) {
 434                 copied = res;
 435                 *ppos += res;
 436         }
 437
 438         if (vec != __vec)
 439                 kfree(vec);
 440 out:
 441         for (i = 0; i < nr_pages; i++)
 442                 put_page(pages[i]);
 443         kvfree(pages);
 444         iov_iter_advance(&to, copied);  /* truncates and discards */
 445         return res;
 446 }
 447
 448 /*
 449  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 450  * using sendpage(). Return the number of bytes sent.
 451  */
 452 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 453                             struct pipe_buffer *buf, struct splice_desc *sd)
 454 {
 455         struct file *file = sd->u.file;
 456         loff_t pos = sd->pos;
 457         int more;
 458
 459         if (!likely(file->f_op->sendpage))
 460                 return -EINVAL;
 461
 462         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 463
 464         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 465                 more |= MSG_SENDPAGE_NOTLAST;
 466
 467         return file->f_op->sendpage(file, buf->page, buf->offset,
 468                                     sd->len, &pos, more);
 469 }
 470
 471 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 472 {
 473         smp_mb();
 474         if (waitqueue_active(&pipe->wait))
 475                 wake_up_interruptible(&pipe->wait);
 476         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 477 }
 478
 479 /**
 480  * splice_from_pipe_feed - feed available data from a pipe to a file
 481  * @pipe:       pipe to splice from
 482  * @sd:         information to @actor
 483  * @actor:      handler that splices the data
 484  *
 485  * Description:
 486  *    This function loops over the pipe and calls @actor to do the
 487  *    actual moving of a single struct pipe_buffer to the desired
 488  *    destination.  It returns when there's no more buffers left in
 489  *    the pipe or if the requested number of bytes (@sd->total_len)
 490  *    have been copied.  It returns a positive number (one) if the
 491  *    pipe needs to be filled with more data, zero if the required
 492  *    number of bytes have been copied and -errno on error.
 493  *
 494  *    This, together with splice_from_pipe_{begin,end,next}, may be
 495  *    used to implement the functionality of __splice_from_pipe() when
 496  *    locking is required around copying the pipe buffers to the
 497  *    destination.
 498  */
 499 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 500                           splice_actor *actor)
 501 {
 502         int ret;
 503
 504         while (pipe->nrbufs) {
 505                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 506
 507                 sd->len = buf->len;
 508                 if (sd->len > sd->total_len)
 509                         sd->len = sd->total_len;
 510
 511                 ret = pipe_buf_confirm(pipe, buf);
 512                 if (unlikely(ret)) {
 513                         if (ret == -ENODATA)
 514                                 ret = 0;
 515                         return ret;
 516                 }
 517
 518                 ret = actor(pipe, buf, sd);
 519                 if (ret <= 0)
 520                         return ret;
 521
 522                 buf->offset += ret;
 523                 buf->len -= ret;
 524
 525                 sd->num_spliced += ret;
 526                 sd->len -= ret;
 527                 sd->pos += ret;
 528                 sd->total_len -= ret;
 529
 530                 if (!buf->len) {
 531                         pipe_buf_release(pipe, buf);
 532                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 533                         pipe->nrbufs--;
 534                         if (pipe->files)
 535                                 sd->need_wakeup = true;
 536                 }
 537
 538                 if (!sd->total_len)
 539                         return 0;
 540         }
 541
 542         return 1;
 543 }
 544
 545 /**
 546  * splice_from_pipe_next - wait for some data to splice from
 547  * @pipe:       pipe to splice from
 548  * @sd:         information about the splice operation
 549  *
 550  * Description:
 551  *    This function will wait for some data and return a positive
 552  *    value (one) if pipe buffers are available.  It will return zero
 553  *    or -errno if no more data needs to be spliced.
 554  */
 555 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 556 {
 557         /*
 558          * Check for signal early to make process killable when there are
 559          * always buffers available
 560          */
 561         if (signal_pending(current))
 562                 return -ERESTARTSYS;
 563
 564         while (!pipe->nrbufs) {
 565                 if (!pipe->writers)
 566                         return 0;
 567
 568                 if (!pipe->waiting_writers && sd->num_spliced)
 569                         return 0;
 570
 571                 if (sd->flags & SPLICE_F_NONBLOCK)
 572                         return -EAGAIN;
 573
 574                 if (signal_pending(current))
 575                         return -ERESTARTSYS;
 576
 577                 if (sd->need_wakeup) {
 578                         wakeup_pipe_writers(pipe);
 579                         sd->need_wakeup = false;
 580                 }
 581
 582                 pipe_wait(pipe);
 583         }
 584
 585         return 1;
 586 }
 587
 588 /**
 589  * splice_from_pipe_begin - start splicing from pipe
 590  * @sd:         information about the splice operation
 591  *
 592  * Description:
 593  *    This function should be called before a loop containing
 594  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 595  *    initialize the necessary fields of @sd.
 596  */
 597 static void splice_from_pipe_begin(struct splice_desc *sd)
 598 {
 599         sd->num_spliced = 0;
 600         sd->need_wakeup = false;
 601 }
 602
 603 /**
 604  * splice_from_pipe_end - finish splicing from pipe
 605  * @pipe:       pipe to splice from
 606  * @sd:         information about the splice operation
 607  *
 608  * Description:
 609  *    This function will wake up pipe writers if necessary.  It should
 610  *    be called after a loop containing splice_from_pipe_next() and
 611  *    splice_from_pipe_feed().
 612  */
 613 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 614 {
 615         if (sd->need_wakeup)
 616                 wakeup_pipe_writers(pipe);
 617 }
 618
 619 /**
 620  * __splice_from_pipe - splice data from a pipe to given actor
 621  * @pipe:       pipe to splice from
 622  * @sd:         information to @actor
 623  * @actor:      handler that splices the data
 624  *
 625  * Description:
 626  *    This function does little more than loop over the pipe and call
 627  *    @actor to do the actual moving of a single struct pipe_buffer to
 628  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 629  *    pipe_to_user.
 630  *
 631  */
 632 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 633                            splice_actor *actor)
 634 {
 635         int ret;
 636
 637         splice_from_pipe_begin(sd);
 638         do {
 639                 cond_resched();
 640                 ret = splice_from_pipe_next(pipe, sd);
 641                 if (ret > 0)
 642                         ret = splice_from_pipe_feed(pipe, sd, actor);
 643         } while (ret > 0);
 644         splice_from_pipe_end(pipe, sd);
 645
 646         return sd->num_spliced ? sd->num_spliced : ret;
 647 }
 648 EXPORT_SYMBOL(__splice_from_pipe);
 649
 650 /**
 651  * splice_from_pipe - splice data from a pipe to a file
 652  * @pipe:       pipe to splice from
 653  * @out:        file to splice to
 654  * @ppos:       position in @out
 655  * @len:        how many bytes to splice
 656  * @flags:      splice modifier flags
 657  * @actor:      handler that splices the data
 658  *
 659  * Description:
 660  *    See __splice_from_pipe. This function locks the pipe inode,
 661  *    otherwise it's identical to __splice_from_pipe().
 662  *
 663  */
 664 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 665                          loff_t *ppos, size_t len, unsigned int flags,
 666                          splice_actor *actor)
 667 {
 668         ssize_t ret;
 669         struct splice_desc sd = {
 670                 .total_len = len,
 671                 .flags = flags,
 672                 .pos = *ppos,
 673                 .u.file = out,
 674         };
 675
 676         pipe_lock(pipe);
 677         ret = __splice_from_pipe(pipe, &sd, actor);
 678         pipe_unlock(pipe);
 679
 680         return ret;
 681 }
 682
 683 /**
 684  * iter_file_splice_write - splice data from a pipe to a file
 685  * @pipe:       pipe info
 686  * @out:        file to write to
 687  * @ppos:       position in @out
 688  * @len:        number of bytes to splice
 689  * @flags:      splice modifier flags
 690  *
 691  * Description:
 692  *    Will either move or copy pages (determined by @flags options) from
 693  *    the given pipe inode to the given file.
 694  *    This one is ->write_iter-based.
 695  *
 696  */
 697 ssize_t
 698 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 699                           loff_t *ppos, size_t len, unsigned int flags)
 700 {
 701         struct splice_desc sd = {
 702                 .total_len = len,
 703                 .flags = flags,
 704                 .pos = *ppos,
 705                 .u.file = out,
 706         };
 707         int nbufs = pipe->buffers;
 708         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 709                                         GFP_KERNEL);
 710         ssize_t ret;
 711
 712         if (unlikely(!array))
 713                 return -ENOMEM;
 714
 715         pipe_lock(pipe);
 716
 717         splice_from_pipe_begin(&sd);
 718         while (sd.total_len) {
 719                 struct iov_iter from;
 720                 size_t left;
 721                 int n, idx;
 722
 723                 ret = splice_from_pipe_next(pipe, &sd);
 724                 if (ret <= 0)
 725                         break;
 726
 727                 if (unlikely(nbufs < pipe->buffers)) {
 728                         kfree(array);
 729                         nbufs = pipe->buffers;
 730                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 731                                         GFP_KERNEL);
 732                         if (!array) {
 733                                 ret = -ENOMEM;
 734                                 break;
 735                         }
 736                 }
 737
 738                 /* build the vector */
 739                 left = sd.total_len;
 740                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 741                         struct pipe_buffer *buf = pipe->bufs + idx;
 742                         size_t this_len = buf->len;
 743
 744                         if (this_len > left)
 745                                 this_len = left;
 746
 747                         if (idx == pipe->buffers - 1)
 748                                 idx = -1;
 749
 750                         ret = pipe_buf_confirm(pipe, buf);
 751                         if (unlikely(ret)) {
 752                                 if (ret == -ENODATA)
 753                                         ret = 0;
 754                                 goto done;
 755                         }
 756
 757                         array[n].bv_page = buf->page;
 758                         array[n].bv_len = this_len;
 759                         array[n].bv_offset = buf->offset;
 760                         left -= this_len;
 761                 }
 762
 763                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 764                               sd.total_len - left);
 765                 ret = vfs_iter_write(out, &from, &sd.pos);
 766                 if (ret <= 0)
 767                         break;
 768
 769                 sd.num_spliced += ret;
 770                 sd.total_len -= ret;
 771                 *ppos = sd.pos;
 772
 773                 /* dismiss the fully eaten buffers, adjust the partial one */
 774                 while (ret) {
 775                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 776                         if (ret >= buf->len) {
 777                                 ret -= buf->len;
 778                                 buf->len = 0;
 779                                 pipe_buf_release(pipe, buf);
 780                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 781                                 pipe->nrbufs--;
 782                                 if (pipe->files)
 783                                         sd.need_wakeup = true;
 784                         } else {
 785                                 buf->offset += ret;
 786                                 buf->len -= ret;
 787                                 ret = 0;
 788                         }
 789                 }
 790         }
 791 done:
 792         kfree(array);
 793         splice_from_pipe_end(pipe, &sd);
 794
 795         pipe_unlock(pipe);
 796
 797         if (sd.num_spliced)
 798                 ret = sd.num_spliced;
 799
 800         return ret;
 801 }
 802
 803 EXPORT_SYMBOL(iter_file_splice_write);
 804
 805 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 806                           struct splice_desc *sd)
 807 {
 808         int ret;
 809         void *data;
 810         loff_t tmp = sd->pos;
 811
 812         data = kmap(buf->page);
 813         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 814         kunmap(buf->page);
 815
 816         return ret;
 817 }
 818
 819 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 820                                          struct file *out, loff_t *ppos,
 821                                          size_t len, unsigned int flags)
 822 {
 823         ssize_t ret;
 824
 825         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 826         if (ret > 0)
 827                 *ppos += ret;
 828
 829         return ret;
 830 }
 831
 832 /**
 833  * generic_splice_sendpage - splice data from a pipe to a socket
 834  * @pipe:       pipe to splice from
 835  * @out:        socket to write to
 836  * @ppos:       position in @out
 837  * @len:        number of bytes to splice
 838  * @flags:      splice modifier flags
 839  *
 840  * Description:
 841  *    Will send @len bytes from the pipe to a network socket. No data copying
 842  *    is involved.
 843  *
 844  */
 845 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 846                                 loff_t *ppos, size_t len, unsigned int flags)
 847 {
 848         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 849 }
 850
 851 EXPORT_SYMBOL(generic_splice_sendpage);
 852
 853 /*
 854  * Attempt to initiate a splice from pipe to file.
 855  */
 856 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 857                            loff_t *ppos, size_t len, unsigned int flags)
 858 {
 859         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 860                                 loff_t *, size_t, unsigned int);
 861
 862         if (out->f_op->splice_write)
 863                 splice_write = out->f_op->splice_write;
 864         else
 865                 splice_write = default_file_splice_write;
 866
 867         return splice_write(pipe, out, ppos, len, flags);
 868 }
 869
 870 /*
 871  * Attempt to initiate a splice from a file to a pipe.
 872  */
 873 static long do_splice_to(struct file *in, loff_t *ppos,
 874                          struct pipe_inode_info *pipe, size_t len,
 875                          unsigned int flags)
 876 {
 877         ssize_t (*splice_read)(struct file *, loff_t *,
 878                                struct pipe_inode_info *, size_t, unsigned int);
 879         int ret;
 880
 881         if (unlikely(!(in->f_mode & FMODE_READ)))
 882                 return -EBADF;
 883
 884         ret = rw_verify_area(READ, in, ppos, len);
 885         if (unlikely(ret < 0))
 886                 return ret;
 887
 888         if (unlikely(len > MAX_RW_COUNT))
 889                 len = MAX_RW_COUNT;
 890
 891         if (in->f_op->splice_read)
 892                 splice_read = in->f_op->splice_read;
 893         else
 894                 splice_read = default_file_splice_read;
 895
 896         return splice_read(in, ppos, pipe, len, flags);
 897 }
 898
 899 /**
 900  * splice_direct_to_actor - splices data directly between two non-pipes
 901  * @in:         file to splice from
 902  * @sd:         actor information on where to splice to
 903  * @actor:      handles the data splicing
 904  *
 905  * Description:
 906  *    This is a special case helper to splice directly between two
 907  *    points, without requiring an explicit pipe. Internally an allocated
 908  *    pipe is cached in the process, and reused during the lifetime of
 909  *    that process.
 910  *
 911  */
 912 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 913                                splice_direct_actor *actor)
 914 {
 915         struct pipe_inode_info *pipe;
 916         long ret, bytes;
 917         umode_t i_mode;
 918         size_t len;
 919         int i, flags, more;
 920
 921         /*
 922          * We require the input being a regular file, as we don't want to
 923          * randomly drop data for eg socket -> socket splicing. Use the
 924          * piped splicing for that!
 925          */
 926         i_mode = file_inode(in)->i_mode;
 927         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 928                 return -EINVAL;
 929
 930         /*
 931          * neither in nor out is a pipe, setup an internal pipe attached to
 932          * 'out' and transfer the wanted data from 'in' to 'out' through that
 933          */
 934         pipe = current->splice_pipe;
 935         if (unlikely(!pipe)) {
 936                 pipe = alloc_pipe_info();
 937                 if (!pipe)
 938                         return -ENOMEM;
 939
 940                 /*
 941                  * We don't have an immediate reader, but we'll read the stuff
 942                  * out of the pipe right after the splice_to_pipe(). So set
 943                  * PIPE_READERS appropriately.
 944                  */
 945                 pipe->readers = 1;
 946
 947                 current->splice_pipe = pipe;
 948         }
 949
 950         /*
 951          * Do the splice.
 952          */
 953         ret = 0;
 954         bytes = 0;
 955         len = sd->total_len;
 956         flags = sd->flags;
 957
 958         /*
 959          * Don't block on output, we have to drain the direct pipe.
 960          */
 961         sd->flags &= ~SPLICE_F_NONBLOCK;
 962         more = sd->flags & SPLICE_F_MORE;
 963
 964         while (len) {
 965                 size_t read_len;
 966                 loff_t pos = sd->pos, prev_pos = pos;
 967
 968                 ret = do_splice_to(in, &pos, pipe, len, flags);
 969                 if (unlikely(ret <= 0))
 970                         goto out_release;
 971
 972                 read_len = ret;
 973                 sd->total_len = read_len;
 974
 975                 /*
 976                  * If more data is pending, set SPLICE_F_MORE
 977                  * If this is the last data and SPLICE_F_MORE was not set
 978                  * initially, clears it.
 979                  */
 980                 if (read_len < len)
 981                         sd->flags |= SPLICE_F_MORE;
 982                 else if (!more)
 983                         sd->flags &= ~SPLICE_F_MORE;
 984                 /*
 985                  * NOTE: nonblocking mode only applies to the input. We
 986                  * must not do the output in nonblocking mode as then we
 987                  * could get stuck data in the internal pipe:
 988                  */
 989                 ret = actor(pipe, sd);
 990                 if (unlikely(ret <= 0)) {
 991                         sd->pos = prev_pos;
 992                         goto out_release;
 993                 }
 994
 995                 bytes += ret;
 996                 len -= ret;
 997                 sd->pos = pos;
 998
 999                 if (ret < read_len) {
1000                         sd->pos = prev_pos + ret;
1001                         goto out_release;
1002                 }
1003         }
1004
1005 done:
1006         pipe->nrbufs = pipe->curbuf = 0;
1007         file_accessed(in);
1008         return bytes;
1009
1010 out_release:
1011         /*
1012          * If we did an incomplete transfer we must release
1013          * the pipe buffers in question:
1014          */
1015         for (i = 0; i < pipe->buffers; i++) {
1016                 struct pipe_buffer *buf = pipe->bufs + i;
1017
1018                 if (buf->ops)
1019                         pipe_buf_release(pipe, buf);
1020         }
1021
1022         if (!bytes)
1023                 bytes = ret;
1024
1025         goto done;
1026 }
1027 EXPORT_SYMBOL(splice_direct_to_actor);
1028
1029 static int direct_splice_actor(struct pipe_inode_info *pipe,
1030                                struct splice_desc *sd)
1031 {
1032         struct file *file = sd->u.file;
1033
1034         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1035                               sd->flags);
1036 }
1037
1038 /**
1039  * do_splice_direct - splices data directly between two files
1040  * @in:         file to splice from
1041  * @ppos:       input file offset
1042  * @out:        file to splice to
1043  * @opos:       output file offset
1044  * @len:        number of bytes to splice
1045  * @flags:      splice modifier flags
1046  *
1047  * Description:
1048  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1049  *    doing it in the application would incur an extra system call
1050  *    (splice in + splice out, as compared to just sendfile()). So this helper
1051  *    can splice directly through a process-private pipe.
1052  *
1053  */
1054 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1055                       loff_t *opos, size_t len, unsigned int flags)
1056 {
1057         struct splice_desc sd = {
1058                 .len            = len,
1059                 .total_len      = len,
1060                 .flags          = flags,
1061                 .pos            = *ppos,
1062                 .u.file         = out,
1063                 .opos           = opos,
1064         };
1065         long ret;
1066
1067         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1068                 return -EBADF;
1069
1070         if (unlikely(out->f_flags & O_APPEND))
1071                 return -EINVAL;
1072
1073         ret = rw_verify_area(WRITE, out, opos, len);
1074         if (unlikely(ret < 0))
1075                 return ret;
1076
1077         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1078         if (ret > 0)
1079                 *ppos = sd.pos;
1080
1081         return ret;
1082 }
1083 EXPORT_SYMBOL(do_splice_direct);
1084
1085 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1086 {
1087         for (;;) {
1088                 if (unlikely(!pipe->readers)) {
1089                         send_sig(SIGPIPE, current, 0);
1090                         return -EPIPE;
1091                 }
1092                 if (pipe->nrbufs != pipe->buffers)
1093                         return 0;
1094                 if (flags & SPLICE_F_NONBLOCK)
1095                         return -EAGAIN;
1096                 if (signal_pending(current))
1097                         return -ERESTARTSYS;
1098                 pipe->waiting_writers++;
1099                 pipe_wait(pipe);
1100                 pipe->waiting_writers--;
1101         }
1102 }
1103
1104 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1105                                struct pipe_inode_info *opipe,
1106                                size_t len, unsigned int flags);
1107
1108 /*
1109  * Determine where to splice to/from.
1110  */
1111 static long do_splice(struct file *in, loff_t __user *off_in,
1112                       struct file *out, loff_t __user *off_out,
1113                       size_t len, unsigned int flags)
1114 {
1115         struct pipe_inode_info *ipipe;
1116         struct pipe_inode_info *opipe;
1117         loff_t offset;
1118         long ret;
1119
1120         ipipe = get_pipe_info(in);
1121         opipe = get_pipe_info(out);
1122
1123         if (ipipe && opipe) {
1124                 if (off_in || off_out)
1125                         return -ESPIPE;
1126
1127                 if (!(in->f_mode & FMODE_READ))
1128                         return -EBADF;
1129
1130                 if (!(out->f_mode & FMODE_WRITE))
1131                         return -EBADF;
1132
1133                 /* Splicing to self would be fun, but... */
1134                 if (ipipe == opipe)
1135                         return -EINVAL;
1136
1137                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1138         }
1139
1140         if (ipipe) {
1141                 if (off_in)
1142                         return -ESPIPE;
1143                 if (off_out) {
1144                         if (!(out->f_mode & FMODE_PWRITE))
1145                                 return -EINVAL;
1146                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1147                                 return -EFAULT;
1148                 } else {
1149                         offset = out->f_pos;
1150                 }
1151
1152                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1153                         return -EBADF;
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                 file_start_write(out);
1163                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1164                 file_end_write(out);
1165
1166                 if (!off_out)
1167                         out->f_pos = offset;
1168                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1169                         ret = -EFAULT;
1170
1171                 return ret;
1172         }
1173
1174         if (opipe) {
1175                 if (off_out)
1176                         return -ESPIPE;
1177                 if (off_in) {
1178                         if (!(in->f_mode & FMODE_PREAD))
1179                                 return -EINVAL;
1180                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1181                                 return -EFAULT;
1182                 } else {
1183                         offset = in->f_pos;
1184                 }
1185
1186                 pipe_lock(opipe);
1187                 ret = wait_for_space(opipe, flags);
1188                 if (!ret)
1189                         ret = do_splice_to(in, &offset, opipe, len, flags);
1190                 pipe_unlock(opipe);
1191                 if (ret > 0)
1192                         wakeup_pipe_readers(opipe);
1193                 if (!off_in)
1194                         in->f_pos = offset;
1195                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1196                         ret = -EFAULT;
1197
1198                 return ret;
1199         }
1200
1201         return -EINVAL;
1202 }
1203
1204 static int iter_to_pipe(struct iov_iter *from,
1205                         struct pipe_inode_info *pipe,
1206                         unsigned flags)
1207 {
1208         struct pipe_buffer buf = {
1209                 .ops = &user_page_pipe_buf_ops,
1210                 .flags = flags
1211         };
1212         size_t total = 0;
1213         int ret = 0;
1214         bool failed = false;
1215
1216         while (iov_iter_count(from) && !failed) {
1217                 struct page *pages[16];
1218                 ssize_t copied;
1219                 size_t start;
1220                 int n;
1221
1222                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1223                 if (copied <= 0) {
1224                         ret = copied;
1225                         break;
1226                 }
1227
1228                 for (n = 0; copied; n++, start = 0) {
1229                         int size = min_t(int, copied, PAGE_SIZE - start);
1230                         if (!failed) {
1231                                 buf.page = pages[n];
1232                                 buf.offset = start;
1233                                 buf.len = size;
1234                                 ret = add_to_pipe(pipe, &buf);
1235                                 if (unlikely(ret < 0)) {
1236                                         failed = true;
1237                                 } else {
1238                                         iov_iter_advance(from, ret);
1239                                         total += ret;
1240                                 }
1241                         } else {
1242                                 put_page(pages[n]);
1243                         }
1244                         copied -= size;
1245                 }
1246         }
1247         return total ? total : ret;
1248 }
1249
1250 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1251                         struct splice_desc *sd)
1252 {
1253         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1254         return n == sd->len ? n : -EFAULT;
1255 }
1256
1257 /*
1258  * For lack of a better implementation, implement vmsplice() to userspace
1259  * as a simple copy of the pipes pages to the user iov.
1260  */
1261 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1262                              unsigned long nr_segs, unsigned int flags)
1263 {
1264         struct pipe_inode_info *pipe;
1265         struct splice_desc sd;
1266         long ret;
1267         struct iovec iovstack[UIO_FASTIOV];
1268         struct iovec *iov = iovstack;
1269         struct iov_iter iter;
1270
1271         pipe = get_pipe_info(file);
1272         if (!pipe)
1273                 return -EBADF;
1274
1275         ret = import_iovec(READ, uiov, nr_segs,
1276                            ARRAY_SIZE(iovstack), &iov, &iter);
1277         if (ret < 0)
1278                 return ret;
1279
1280         sd.total_len = iov_iter_count(&iter);
1281         sd.len = 0;
1282         sd.flags = flags;
1283         sd.u.data = &iter;
1284         sd.pos = 0;
1285
1286         if (sd.total_len) {
1287                 pipe_lock(pipe);
1288                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1289                 pipe_unlock(pipe);
1290         }
1291
1292         kfree(iov);
1293         return ret;
1294 }
1295
1296 /*
1297  * vmsplice splices a user address range into a pipe. It can be thought of
1298  * as splice-from-memory, where the regular splice is splice-from-file (or
1299  * to file). In both cases the output is a pipe, naturally.
1300  */
1301 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1302                              unsigned long nr_segs, unsigned int flags)
1303 {
1304         struct pipe_inode_info *pipe;
1305         struct iovec iovstack[UIO_FASTIOV];
1306         struct iovec *iov = iovstack;
1307         struct iov_iter from;
1308         long ret;
1309         unsigned buf_flag = 0;
1310
1311         if (flags & SPLICE_F_GIFT)
1312                 buf_flag = PIPE_BUF_FLAG_GIFT;
1313
1314         pipe = get_pipe_info(file);
1315         if (!pipe)
1316                 return -EBADF;
1317
1318         ret = import_iovec(WRITE, uiov, nr_segs,
1319                            ARRAY_SIZE(iovstack), &iov, &from);
1320         if (ret < 0)
1321                 return ret;
1322
1323         pipe_lock(pipe);
1324         ret = wait_for_space(pipe, flags);
1325         if (!ret)
1326                 ret = iter_to_pipe(&from, pipe, buf_flag);
1327         pipe_unlock(pipe);
1328         if (ret > 0)
1329                 wakeup_pipe_readers(pipe);
1330         kfree(iov);
1331         return ret;
1332 }
1333
1334 /*
1335  * Note that vmsplice only really supports true splicing _from_ user memory
1336  * to a pipe, not the other way around. Splicing from user memory is a simple
1337  * operation that can be supported without any funky alignment restrictions
1338  * or nasty vm tricks. We simply map in the user memory and fill them into
1339  * a pipe. The reverse isn't quite as easy, though. There are two possible
1340  * solutions for that:
1341  *
1342  *      - memcpy() the data internally, at which point we might as well just
1343  *        do a regular read() on the buffer anyway.
1344  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1345  *        has restriction limitations on both ends of the pipe).
1346  *
1347  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1348  *
1349  */
1350 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1351                 unsigned long, nr_segs, unsigned int, flags)
1352 {
1353         struct fd f;
1354         long error;
1355
1356         if (unlikely(flags & ~SPLICE_F_ALL))
1357                 return -EINVAL;
1358         if (unlikely(nr_segs > UIO_MAXIOV))
1359                 return -EINVAL;
1360         else if (unlikely(!nr_segs))
1361                 return 0;
1362
1363         error = -EBADF;
1364         f = fdget(fd);
1365         if (f.file) {
1366                 if (f.file->f_mode & FMODE_WRITE)
1367                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1368                 else if (f.file->f_mode & FMODE_READ)
1369                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1370
1371                 fdput(f);
1372         }
1373
1374         return error;
1375 }
1376
1377 #ifdef CONFIG_COMPAT
1378 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1379                     unsigned int, nr_segs, unsigned int, flags)
1380 {
1381         unsigned i;
1382         struct iovec __user *iov;
1383         if (nr_segs > UIO_MAXIOV)
1384                 return -EINVAL;
1385         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1386         for (i = 0; i < nr_segs; i++) {
1387                 struct compat_iovec v;
1388                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1389                     get_user(v.iov_len, &iov32[i].iov_len) ||
1390                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1391                     put_user(v.iov_len, &iov[i].iov_len))
1392                         return -EFAULT;
1393         }
1394         return sys_vmsplice(fd, iov, nr_segs, flags);
1395 }
1396 #endif
1397
1398 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1399                 int, fd_out, loff_t __user *, off_out,
1400                 size_t, len, unsigned int, flags)
1401 {
1402         struct fd in, out;
1403         long error;
1404
1405         if (unlikely(!len))
1406                 return 0;
1407
1408         if (unlikely(flags & ~SPLICE_F_ALL))
1409                 return -EINVAL;
1410
1411         error = -EBADF;
1412         in = fdget(fd_in);
1413         if (in.file) {
1414                 if (in.file->f_mode & FMODE_READ) {
1415                         out = fdget(fd_out);
1416                         if (out.file) {
1417                                 if (out.file->f_mode & FMODE_WRITE)
1418                                         error = do_splice(in.file, off_in,
1419                                                           out.file, off_out,
1420                                                           len, flags);
1421                                 fdput(out);
1422                         }
1423                 }
1424                 fdput(in);
1425         }
1426         return error;
1427 }
1428
1429 /*
1430  * Make sure there's data to read. Wait for input if we can, otherwise
1431  * return an appropriate error.
1432  */
1433 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1434 {
1435         int ret;
1436
1437         /*
1438          * Check ->nrbufs without the inode lock first. This function
1439          * is speculative anyways, so missing one is ok.
1440          */
1441         if (pipe->nrbufs)
1442                 return 0;
1443
1444         ret = 0;
1445         pipe_lock(pipe);
1446
1447         while (!pipe->nrbufs) {
1448                 if (signal_pending(current)) {
1449                         ret = -ERESTARTSYS;
1450                         break;
1451                 }
1452                 if (!pipe->writers)
1453                         break;
1454                 if (!pipe->waiting_writers) {
1455                         if (flags & SPLICE_F_NONBLOCK) {
1456                                 ret = -EAGAIN;
1457                                 break;
1458                         }
1459                 }
1460                 pipe_wait(pipe);
1461         }
1462
1463         pipe_unlock(pipe);
1464         return ret;
1465 }
1466
1467 /*
1468  * Make sure there's writeable room. Wait for room if we can, otherwise
1469  * return an appropriate error.
1470  */
1471 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1472 {
1473         int ret;
1474
1475         /*
1476          * Check ->nrbufs without the inode lock first. This function
1477          * is speculative anyways, so missing one is ok.
1478          */
1479         if (pipe->nrbufs < pipe->buffers)
1480                 return 0;
1481
1482         ret = 0;
1483         pipe_lock(pipe);
1484
1485         while (pipe->nrbufs >= pipe->buffers) {
1486                 if (!pipe->readers) {
1487                         send_sig(SIGPIPE, current, 0);
1488                         ret = -EPIPE;
1489                         break;
1490                 }
1491                 if (flags & SPLICE_F_NONBLOCK) {
1492                         ret = -EAGAIN;
1493                         break;
1494                 }
1495                 if (signal_pending(current)) {
1496                         ret = -ERESTARTSYS;
1497                         break;
1498                 }
1499                 pipe->waiting_writers++;
1500                 pipe_wait(pipe);
1501                 pipe->waiting_writers--;
1502         }
1503
1504         pipe_unlock(pipe);
1505         return ret;
1506 }
1507
1508 /*
1509  * Splice contents of ipipe to opipe.
1510  */
1511 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1512                                struct pipe_inode_info *opipe,
1513                                size_t len, unsigned int flags)
1514 {
1515         struct pipe_buffer *ibuf, *obuf;
1516         int ret = 0, nbuf;
1517         bool input_wakeup = false;
1518
1519
1520 retry:
1521         ret = ipipe_prep(ipipe, flags);
1522         if (ret)
1523                 return ret;
1524
1525         ret = opipe_prep(opipe, flags);
1526         if (ret)
1527                 return ret;
1528
1529         /*
1530          * Potential ABBA deadlock, work around it by ordering lock
1531          * grabbing by pipe info address. Otherwise two different processes
1532          * could deadlock (one doing tee from A -> B, the other from B -> A).
1533          */
1534         pipe_double_lock(ipipe, opipe);
1535
1536         do {
1537                 if (!opipe->readers) {
1538                         send_sig(SIGPIPE, current, 0);
1539                         if (!ret)
1540                                 ret = -EPIPE;
1541                         break;
1542                 }
1543
1544                 if (!ipipe->nrbufs && !ipipe->writers)
1545                         break;
1546
1547                 /*
1548                  * Cannot make any progress, because either the input
1549                  * pipe is empty or the output pipe is full.
1550                  */
1551                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1552                         /* Already processed some buffers, break */
1553                         if (ret)
1554                                 break;
1555
1556                         if (flags & SPLICE_F_NONBLOCK) {
1557                                 ret = -EAGAIN;
1558                                 break;
1559                         }
1560
1561                         /*
1562                          * We raced with another reader/writer and haven't
1563                          * managed to process any buffers.  A zero return
1564                          * value means EOF, so retry instead.
1565                          */
1566                         pipe_unlock(ipipe);
1567                         pipe_unlock(opipe);
1568                         goto retry;
1569                 }
1570
1571                 ibuf = ipipe->bufs + ipipe->curbuf;
1572                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1573                 obuf = opipe->bufs + nbuf;
1574
1575                 if (len >= ibuf->len) {
1576                         /*
1577                          * Simply move the whole buffer from ipipe to opipe
1578                          */
1579                         *obuf = *ibuf;
1580                         ibuf->ops = NULL;
1581                         opipe->nrbufs++;
1582                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1583                         ipipe->nrbufs--;
1584                         input_wakeup = true;
1585                 } else {
1586                         /*
1587                          * Get a reference to this pipe buffer,
1588                          * so we can copy the contents over.
1589                          */
1590                         pipe_buf_get(ipipe, ibuf);
1591                         *obuf = *ibuf;
1592
1593                         /*
1594                          * Don't inherit the gift flag, we need to
1595                          * prevent multiple steals of this page.
1596                          */
1597                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1598
1599                         obuf->len = len;
1600                         opipe->nrbufs++;
1601                         ibuf->offset += obuf->len;
1602                         ibuf->len -= obuf->len;
1603                 }
1604                 ret += obuf->len;
1605                 len -= obuf->len;
1606         } while (len);
1607
1608         pipe_unlock(ipipe);
1609         pipe_unlock(opipe);
1610
1611         /*
1612          * If we put data in the output pipe, wakeup any potential readers.
1613          */
1614         if (ret > 0)
1615                 wakeup_pipe_readers(opipe);
1616
1617         if (input_wakeup)
1618                 wakeup_pipe_writers(ipipe);
1619
1620         return ret;
1621 }
1622
1623 /*
1624  * Link contents of ipipe to opipe.
1625  */
1626 static int link_pipe(struct pipe_inode_info *ipipe,
1627                      struct pipe_inode_info *opipe,
1628                      size_t len, unsigned int flags)
1629 {
1630         struct pipe_buffer *ibuf, *obuf;
1631         int ret = 0, i = 0, nbuf;
1632
1633         /*
1634          * Potential ABBA deadlock, work around it by ordering lock
1635          * grabbing by pipe info address. Otherwise two different processes
1636          * could deadlock (one doing tee from A -> B, the other from B -> A).
1637          */
1638         pipe_double_lock(ipipe, opipe);
1639
1640         do {
1641                 if (!opipe->readers) {
1642                         send_sig(SIGPIPE, current, 0);
1643                         if (!ret)
1644                                 ret = -EPIPE;
1645                         break;
1646                 }
1647
1648                 /*
1649                  * If we have iterated all input buffers or ran out of
1650                  * output room, break.
1651                  */
1652                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1653                         break;
1654
1655                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1656                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1657
1658                 /*
1659                  * Get a reference to this pipe buffer,
1660                  * so we can copy the contents over.
1661                  */
1662                 pipe_buf_get(ipipe, ibuf);
1663
1664                 obuf = opipe->bufs + nbuf;
1665                 *obuf = *ibuf;
1666
1667                 /*
1668                  * Don't inherit the gift flag, we need to
1669                  * prevent multiple steals of this page.
1670                  */
1671                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1672
1673                 if (obuf->len > len)
1674                         obuf->len = len;
1675
1676                 opipe->nrbufs++;
1677                 ret += obuf->len;
1678                 len -= obuf->len;
1679                 i++;
1680         } while (len);
1681
1682         /*
1683          * return EAGAIN if we have the potential of some data in the
1684          * future, otherwise just return 0
1685          */
1686         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1687                 ret = -EAGAIN;
1688
1689         pipe_unlock(ipipe);
1690         pipe_unlock(opipe);
1691
1692         /*
1693          * If we put data in the output pipe, wakeup any potential readers.
1694          */
1695         if (ret > 0)
1696                 wakeup_pipe_readers(opipe);
1697
1698         return ret;
1699 }
1700
1701 /*
1702  * This is a tee(1) implementation that works on pipes. It doesn't copy
1703  * any data, it simply references the 'in' pages on the 'out' pipe.
1704  * The 'flags' used are the SPLICE_F_* variants, currently the only
1705  * applicable one is SPLICE_F_NONBLOCK.
1706  */
1707 static long do_tee(struct file *in, struct file *out, size_t len,
1708                    unsigned int flags)
1709 {
1710         struct pipe_inode_info *ipipe = get_pipe_info(in);
1711         struct pipe_inode_info *opipe = get_pipe_info(out);
1712         int ret = -EINVAL;
1713
1714         /*
1715          * Duplicate the contents of ipipe to opipe without actually
1716          * copying the data.
1717          */
1718         if (ipipe && opipe && ipipe != opipe) {
1719                 /*
1720                  * Keep going, unless we encounter an error. The ipipe/opipe
1721                  * ordering doesn't really matter.
1722                  */
1723                 ret = ipipe_prep(ipipe, flags);
1724                 if (!ret) {
1725                         ret = opipe_prep(opipe, flags);
1726                         if (!ret)
1727                                 ret = link_pipe(ipipe, opipe, len, flags);
1728                 }
1729         }
1730
1731         return ret;
1732 }
1733
1734 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1735 {
1736         struct fd in;
1737         int error;
1738
1739         if (unlikely(flags & ~SPLICE_F_ALL))
1740                 return -EINVAL;
1741
1742         if (unlikely(!len))
1743                 return 0;
1744
1745         error = -EBADF;
1746         in = fdget(fdin);
1747         if (in.file) {
1748                 if (in.file->f_mode & FMODE_READ) {
1749                         struct fd out = fdget(fdout);
1750                         if (out.file) {
1751                                 if (out.file->f_mode & FMODE_WRITE)
1752                                         error = do_tee(in.file, out.file,
1753                                                         len, flags);
1754                                 fdput(out);
1755                         }
1756                 }
1757                 fdput(in);
1758         }
1759
1760         return error;
1761 }