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/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34 #include <linux/compat.h>
  35 #include "internal.h"
  36
  37 /*
  38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  39  * a vm helper function, it's already simplified quite a bit by the
  40  * addition of remove_mapping(). If success is returned, the caller may
  41  * attempt to reuse this page for another destination.
  42  */
  43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  44                                      struct pipe_buffer *buf)
  45 {
  46         struct page *page = buf->page;
  47         struct address_space *mapping;
  48
  49         lock_page(page);
  50
  51         mapping = page_mapping(page);
  52         if (mapping) {
  53                 WARN_ON(!PageUptodate(page));
  54
  55                 /*
  56                  * At least for ext2 with nobh option, we need to wait on
  57                  * writeback completing on this page, since we'll remove it
  58                  * from the pagecache.  Otherwise truncate wont wait on the
  59                  * page, allowing the disk blocks to be reused by someone else
  60                  * before we actually wrote our data to them. fs corruption
  61                  * ensues.
  62                  */
  63                 wait_on_page_writeback(page);
  64
  65                 if (page_has_private(page) &&
  66                     !try_to_release_page(page, GFP_KERNEL))
  67                         goto out_unlock;
  68
  69                 /*
  70                  * If we succeeded in removing the mapping, set LRU flag
  71                  * and return good.
  72                  */
  73                 if (remove_mapping(mapping, page)) {
  74                         buf->flags |= PIPE_BUF_FLAG_LRU;
  75                         return 0;
  76                 }
  77         }
  78
  79         /*
  80          * Raced with truncate or failed to remove page from current
  81          * address space, unlock and return failure.
  82          */
  83 out_unlock:
  84         unlock_page(page);
  85         return 1;
  86 }
  87
  88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  89                                         struct pipe_buffer *buf)
  90 {
  91         put_page(buf->page);
  92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  93 }
  94
  95 /*
  96  * Check whether the contents of buf is OK to access. Since the content
  97  * is a page cache page, IO may be in flight.
  98  */
  99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 100                                        struct pipe_buffer *buf)
 101 {
 102         struct page *page = buf->page;
 103         int err;
 104
 105         if (!PageUptodate(page)) {
 106                 lock_page(page);
 107
 108                 /*
 109                  * Page got truncated/unhashed. This will cause a 0-byte
 110                  * splice, if this is the first page.
 111                  */
 112                 if (!page->mapping) {
 113                         err = -ENODATA;
 114                         goto error;
 115                 }
 116
 117                 /*
 118                  * Uh oh, read-error from disk.
 119                  */
 120                 if (!PageUptodate(page)) {
 121                         err = -EIO;
 122                         goto error;
 123                 }
 124
 125                 /*
 126                  * Page is ok afterall, we are done.
 127                  */
 128                 unlock_page(page);
 129         }
 130
 131         return 0;
 132 error:
 133         unlock_page(page);
 134         return err;
 135 }
 136
 137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .confirm = generic_pipe_buf_confirm,
 158         .release = page_cache_pipe_buf_release,
 159         .steal = user_page_pipe_buf_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->wait))
 167                 wake_up_interruptible(&pipe->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         int ret = 0, page_nr = 0;
 187
 188         if (!spd_pages)
 189                 return 0;
 190
 191         if (unlikely(!pipe->readers)) {
 192                 send_sig(SIGPIPE, current, 0);
 193                 ret = -EPIPE;
 194                 goto out;
 195         }
 196
 197         while (pipe->nrbufs < pipe->buffers) {
 198                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 199                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 200
 201                 buf->page = spd->pages[page_nr];
 202                 buf->offset = spd->partial[page_nr].offset;
 203                 buf->len = spd->partial[page_nr].len;
 204                 buf->private = spd->partial[page_nr].private;
 205                 buf->ops = spd->ops;
 206
 207                 pipe->nrbufs++;
 208                 page_nr++;
 209                 ret += buf->len;
 210
 211                 if (!--spd->nr_pages)
 212                         break;
 213         }
 214
 215         if (!ret)
 216                 ret = -EAGAIN;
 217
 218 out:
 219         while (page_nr < spd_pages)
 220                 spd->spd_release(spd, page_nr++);
 221
 222         return ret;
 223 }
 224 EXPORT_SYMBOL_GPL(splice_to_pipe);
 225
 226 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 227 {
 228         int ret;
 229
 230         if (unlikely(!pipe->readers)) {
 231                 send_sig(SIGPIPE, current, 0);
 232                 ret = -EPIPE;
 233         } else if (pipe->nrbufs == pipe->buffers) {
 234                 ret = -EAGAIN;
 235         } else {
 236                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 237                 pipe->bufs[newbuf] = *buf;
 238                 pipe->nrbufs++;
 239                 return buf->len;
 240         }
 241         pipe_buf_release(pipe, buf);
 242         return ret;
 243 }
 244 EXPORT_SYMBOL(add_to_pipe);
 245
 246 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 247 {
 248         put_page(spd->pages[i]);
 249 }
 250
 251 /*
 252  * Check if we need to grow the arrays holding pages and partial page
 253  * descriptions.
 254  */
 255 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 256 {
 257         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 258
 259         spd->nr_pages_max = buffers;
 260         if (buffers <= PIPE_DEF_BUFFERS)
 261                 return 0;
 262
 263         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 264         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 265
 266         if (spd->pages && spd->partial)
 267                 return 0;
 268
 269         kfree(spd->pages);
 270         kfree(spd->partial);
 271         return -ENOMEM;
 272 }
 273
 274 void splice_shrink_spd(struct splice_pipe_desc *spd)
 275 {
 276         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 277                 return;
 278
 279         kfree(spd->pages);
 280         kfree(spd->partial);
 281 }
 282
 283 /**
 284  * generic_file_splice_read - splice data from file to a pipe
 285  * @in:         file to splice from
 286  * @ppos:       position in @in
 287  * @pipe:       pipe to splice to
 288  * @len:        number of bytes to splice
 289  * @flags:      splice modifier flags
 290  *
 291  * Description:
 292  *    Will read pages from given file and fill them into a pipe. Can be
 293  *    used as long as it has more or less sane ->read_iter().
 294  *
 295  */
 296 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 297                                  struct pipe_inode_info *pipe, size_t len,
 298                                  unsigned int flags)
 299 {
 300         struct iov_iter to;
 301         struct kiocb kiocb;
 302         int idx, ret;
 303
 304         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 305         idx = to.idx;
 306         init_sync_kiocb(&kiocb, in);
 307         kiocb.ki_pos = *ppos;
 308         ret = in->f_op->read_iter(&kiocb, &to);
 309         if (ret > 0) {
 310                 *ppos = kiocb.ki_pos;
 311                 file_accessed(in);
 312         } else if (ret < 0) {
 313                 to.idx = idx;
 314                 to.iov_offset = 0;
 315                 iov_iter_advance(&to, 0); /* to free what was emitted */
 316                 /*
 317                  * callers of ->splice_read() expect -EAGAIN on
 318                  * "can't put anything in there", rather than -EFAULT.
 319                  */
 320                 if (ret == -EFAULT)
 321                         ret = -EAGAIN;
 322         }
 323
 324         return ret;
 325 }
 326 EXPORT_SYMBOL(generic_file_splice_read);
 327
 328 const struct pipe_buf_operations default_pipe_buf_ops = {
 329         .can_merge = 0,
 330         .confirm = generic_pipe_buf_confirm,
 331         .release = generic_pipe_buf_release,
 332         .steal = generic_pipe_buf_steal,
 333         .get = generic_pipe_buf_get,
 334 };
 335
 336 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 337                                     struct pipe_buffer *buf)
 338 {
 339         return 1;
 340 }
 341
 342 /* Pipe buffer operations for a socket and similar. */
 343 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 344         .can_merge = 0,
 345         .confirm = generic_pipe_buf_confirm,
 346         .release = generic_pipe_buf_release,
 347         .steal = generic_pipe_buf_nosteal,
 348         .get = generic_pipe_buf_get,
 349 };
 350 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 351
 352 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 353                             unsigned long vlen, loff_t offset)
 354 {
 355         mm_segment_t old_fs;
 356         loff_t pos = offset;
 357         ssize_t res;
 358
 359         old_fs = get_fs();
 360         set_fs(get_ds());
 361         /* The cast to a user pointer is valid due to the set_fs() */
 362         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 363         set_fs(old_fs);
 364
 365         return res;
 366 }
 367
 368 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 369                             loff_t pos)
 370 {
 371         mm_segment_t old_fs;
 372         ssize_t res;
 373
 374         old_fs = get_fs();
 375         set_fs(get_ds());
 376         /* The cast to a user pointer is valid due to the set_fs() */
 377         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 378         set_fs(old_fs);
 379
 380         return res;
 381 }
 382 EXPORT_SYMBOL(kernel_write);
 383
 384 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 385                                  struct pipe_inode_info *pipe, size_t len,
 386                                  unsigned int flags)
 387 {
 388         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 389         struct iov_iter to;
 390         struct page **pages;
 391         unsigned int nr_pages;
 392         size_t offset, dummy, copied = 0;
 393         ssize_t res;
 394         int i;
 395
 396         if (pipe->nrbufs == pipe->buffers)
 397                 return -EAGAIN;
 398
 399         /*
 400          * Try to keep page boundaries matching to source pagecache ones -
 401          * it probably won't be much help, but...
 402          */
 403         offset = *ppos & ~PAGE_MASK;
 404
 405         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 406
 407         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
 408         if (res <= 0)
 409                 return -ENOMEM;
 410
 411         nr_pages = res / PAGE_SIZE;
 412
 413         vec = __vec;
 414         if (nr_pages > PIPE_DEF_BUFFERS) {
 415                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 416                 if (unlikely(!vec)) {
 417                         res = -ENOMEM;
 418                         goto out;
 419                 }
 420         }
 421
 422         pipe->bufs[to.idx].offset = offset;
 423         pipe->bufs[to.idx].len -= offset;
 424
 425         for (i = 0; i < nr_pages; i++) {
 426                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 427                 vec[i].iov_base = page_address(pages[i]) + offset;
 428                 vec[i].iov_len = this_len;
 429                 len -= this_len;
 430                 offset = 0;
 431         }
 432
 433         res = kernel_readv(in, vec, nr_pages, *ppos);
 434         if (res > 0) {
 435                 copied = res;
 436                 *ppos += res;
 437         }
 438
 439         if (vec != __vec)
 440                 kfree(vec);
 441 out:
 442         for (i = 0; i < nr_pages; i++)
 443                 put_page(pages[i]);
 444         kvfree(pages);
 445         iov_iter_advance(&to, copied);  /* truncates and discards */
 446         return res;
 447 }
 448
 449 /*
 450  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 451  * using sendpage(). Return the number of bytes sent.
 452  */
 453 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 454                             struct pipe_buffer *buf, struct splice_desc *sd)
 455 {
 456         struct file *file = sd->u.file;
 457         loff_t pos = sd->pos;
 458         int more;
 459
 460         if (!likely(file->f_op->sendpage))
 461                 return -EINVAL;
 462
 463         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 464
 465         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 466                 more |= MSG_SENDPAGE_NOTLAST;
 467
 468         return file->f_op->sendpage(file, buf->page, buf->offset,
 469                                     sd->len, &pos, more);
 470 }
 471
 472 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 473 {
 474         smp_mb();
 475         if (waitqueue_active(&pipe->wait))
 476                 wake_up_interruptible(&pipe->wait);
 477         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 478 }
 479
 480 /**
 481  * splice_from_pipe_feed - feed available data from a pipe to a file
 482  * @pipe:       pipe to splice from
 483  * @sd:         information to @actor
 484  * @actor:      handler that splices the data
 485  *
 486  * Description:
 487  *    This function loops over the pipe and calls @actor to do the
 488  *    actual moving of a single struct pipe_buffer to the desired
 489  *    destination.  It returns when there's no more buffers left in
 490  *    the pipe or if the requested number of bytes (@sd->total_len)
 491  *    have been copied.  It returns a positive number (one) if the
 492  *    pipe needs to be filled with more data, zero if the required
 493  *    number of bytes have been copied and -errno on error.
 494  *
 495  *    This, together with splice_from_pipe_{begin,end,next}, may be
 496  *    used to implement the functionality of __splice_from_pipe() when
 497  *    locking is required around copying the pipe buffers to the
 498  *    destination.
 499  */
 500 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 501                           splice_actor *actor)
 502 {
 503         int ret;
 504
 505         while (pipe->nrbufs) {
 506                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 507
 508                 sd->len = buf->len;
 509                 if (sd->len > sd->total_len)
 510                         sd->len = sd->total_len;
 511
 512                 ret = pipe_buf_confirm(pipe, buf);
 513                 if (unlikely(ret)) {
 514                         if (ret == -ENODATA)
 515                                 ret = 0;
 516                         return ret;
 517                 }
 518
 519                 ret = actor(pipe, buf, sd);
 520                 if (ret <= 0)
 521                         return ret;
 522
 523                 buf->offset += ret;
 524                 buf->len -= ret;
 525
 526                 sd->num_spliced += ret;
 527                 sd->len -= ret;
 528                 sd->pos += ret;
 529                 sd->total_len -= ret;
 530
 531                 if (!buf->len) {
 532                         pipe_buf_release(pipe, buf);
 533                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 534                         pipe->nrbufs--;
 535                         if (pipe->files)
 536                                 sd->need_wakeup = true;
 537                 }
 538
 539                 if (!sd->total_len)
 540                         return 0;
 541         }
 542
 543         return 1;
 544 }
 545
 546 /**
 547  * splice_from_pipe_next - wait for some data to splice from
 548  * @pipe:       pipe to splice from
 549  * @sd:         information about the splice operation
 550  *
 551  * Description:
 552  *    This function will wait for some data and return a positive
 553  *    value (one) if pipe buffers are available.  It will return zero
 554  *    or -errno if no more data needs to be spliced.
 555  */
 556 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 557 {
 558         /*
 559          * Check for signal early to make process killable when there are
 560          * always buffers available
 561          */
 562         if (signal_pending(current))
 563                 return -ERESTARTSYS;
 564
 565         while (!pipe->nrbufs) {
 566                 if (!pipe->writers)
 567                         return 0;
 568
 569                 if (!pipe->waiting_writers && sd->num_spliced)
 570                         return 0;
 571
 572                 if (sd->flags & SPLICE_F_NONBLOCK)
 573                         return -EAGAIN;
 574
 575                 if (signal_pending(current))
 576                         return -ERESTARTSYS;
 577
 578                 if (sd->need_wakeup) {
 579                         wakeup_pipe_writers(pipe);
 580                         sd->need_wakeup = false;
 581                 }
 582
 583                 pipe_wait(pipe);
 584         }
 585
 586         return 1;
 587 }
 588
 589 /**
 590  * splice_from_pipe_begin - start splicing from pipe
 591  * @sd:         information about the splice operation
 592  *
 593  * Description:
 594  *    This function should be called before a loop containing
 595  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 596  *    initialize the necessary fields of @sd.
 597  */
 598 static void splice_from_pipe_begin(struct splice_desc *sd)
 599 {
 600         sd->num_spliced = 0;
 601         sd->need_wakeup = false;
 602 }
 603
 604 /**
 605  * splice_from_pipe_end - finish splicing from pipe
 606  * @pipe:       pipe to splice from
 607  * @sd:         information about the splice operation
 608  *
 609  * Description:
 610  *    This function will wake up pipe writers if necessary.  It should
 611  *    be called after a loop containing splice_from_pipe_next() and
 612  *    splice_from_pipe_feed().
 613  */
 614 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 615 {
 616         if (sd->need_wakeup)
 617                 wakeup_pipe_writers(pipe);
 618 }
 619
 620 /**
 621  * __splice_from_pipe - splice data from a pipe to given actor
 622  * @pipe:       pipe to splice from
 623  * @sd:         information to @actor
 624  * @actor:      handler that splices the data
 625  *
 626  * Description:
 627  *    This function does little more than loop over the pipe and call
 628  *    @actor to do the actual moving of a single struct pipe_buffer to
 629  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 630  *    pipe_to_user.
 631  *
 632  */
 633 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 634                            splice_actor *actor)
 635 {
 636         int ret;
 637
 638         splice_from_pipe_begin(sd);
 639         do {
 640                 cond_resched();
 641                 ret = splice_from_pipe_next(pipe, sd);
 642                 if (ret > 0)
 643                         ret = splice_from_pipe_feed(pipe, sd, actor);
 644         } while (ret > 0);
 645         splice_from_pipe_end(pipe, sd);
 646
 647         return sd->num_spliced ? sd->num_spliced : ret;
 648 }
 649 EXPORT_SYMBOL(__splice_from_pipe);
 650
 651 /**
 652  * splice_from_pipe - splice data from a pipe to a file
 653  * @pipe:       pipe to splice from
 654  * @out:        file to splice to
 655  * @ppos:       position in @out
 656  * @len:        how many bytes to splice
 657  * @flags:      splice modifier flags
 658  * @actor:      handler that splices the data
 659  *
 660  * Description:
 661  *    See __splice_from_pipe. This function locks the pipe inode,
 662  *    otherwise it's identical to __splice_from_pipe().
 663  *
 664  */
 665 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 666                          loff_t *ppos, size_t len, unsigned int flags,
 667                          splice_actor *actor)
 668 {
 669         ssize_t ret;
 670         struct splice_desc sd = {
 671                 .total_len = len,
 672                 .flags = flags,
 673                 .pos = *ppos,
 674                 .u.file = out,
 675         };
 676
 677         pipe_lock(pipe);
 678         ret = __splice_from_pipe(pipe, &sd, actor);
 679         pipe_unlock(pipe);
 680
 681         return ret;
 682 }
 683
 684 /**
 685  * iter_file_splice_write - splice data from a pipe to a file
 686  * @pipe:       pipe info
 687  * @out:        file to write to
 688  * @ppos:       position in @out
 689  * @len:        number of bytes to splice
 690  * @flags:      splice modifier flags
 691  *
 692  * Description:
 693  *    Will either move or copy pages (determined by @flags options) from
 694  *    the given pipe inode to the given file.
 695  *    This one is ->write_iter-based.
 696  *
 697  */
 698 ssize_t
 699 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 700                           loff_t *ppos, size_t len, unsigned int flags)
 701 {
 702         struct splice_desc sd = {
 703                 .total_len = len,
 704                 .flags = flags,
 705                 .pos = *ppos,
 706                 .u.file = out,
 707         };
 708         int nbufs = pipe->buffers;
 709         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 710                                         GFP_KERNEL);
 711         ssize_t ret;
 712
 713         if (unlikely(!array))
 714                 return -ENOMEM;
 715
 716         pipe_lock(pipe);
 717
 718         splice_from_pipe_begin(&sd);
 719         while (sd.total_len) {
 720                 struct iov_iter from;
 721                 size_t left;
 722                 int n, idx;
 723
 724                 ret = splice_from_pipe_next(pipe, &sd);
 725                 if (ret <= 0)
 726                         break;
 727
 728                 if (unlikely(nbufs < pipe->buffers)) {
 729                         kfree(array);
 730                         nbufs = pipe->buffers;
 731                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 732                                         GFP_KERNEL);
 733                         if (!array) {
 734                                 ret = -ENOMEM;
 735                                 break;
 736                         }
 737                 }
 738
 739                 /* build the vector */
 740                 left = sd.total_len;
 741                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 742                         struct pipe_buffer *buf = pipe->bufs + idx;
 743                         size_t this_len = buf->len;
 744
 745                         if (this_len > left)
 746                                 this_len = left;
 747
 748                         if (idx == pipe->buffers - 1)
 749                                 idx = -1;
 750
 751                         ret = pipe_buf_confirm(pipe, buf);
 752                         if (unlikely(ret)) {
 753                                 if (ret == -ENODATA)
 754                                         ret = 0;
 755                                 goto done;
 756                         }
 757
 758                         array[n].bv_page = buf->page;
 759                         array[n].bv_len = this_len;
 760                         array[n].bv_offset = buf->offset;
 761                         left -= this_len;
 762                 }
 763
 764                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 765                               sd.total_len - left);
 766                 ret = vfs_iter_write(out, &from, &sd.pos);
 767                 if (ret <= 0)
 768                         break;
 769
 770                 sd.num_spliced += ret;
 771                 sd.total_len -= ret;
 772                 *ppos = sd.pos;
 773
 774                 /* dismiss the fully eaten buffers, adjust the partial one */
 775                 while (ret) {
 776                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 777                         if (ret >= buf->len) {
 778                                 ret -= buf->len;
 779                                 buf->len = 0;
 780                                 pipe_buf_release(pipe, buf);
 781                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 782                                 pipe->nrbufs--;
 783                                 if (pipe->files)
 784                                         sd.need_wakeup = true;
 785                         } else {
 786                                 buf->offset += ret;
 787                                 buf->len -= ret;
 788                                 ret = 0;
 789                         }
 790                 }
 791         }
 792 done:
 793         kfree(array);
 794         splice_from_pipe_end(pipe, &sd);
 795
 796         pipe_unlock(pipe);
 797
 798         if (sd.num_spliced)
 799                 ret = sd.num_spliced;
 800
 801         return ret;
 802 }
 803
 804 EXPORT_SYMBOL(iter_file_splice_write);
 805
 806 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 807                           struct splice_desc *sd)
 808 {
 809         int ret;
 810         void *data;
 811         loff_t tmp = sd->pos;
 812
 813         data = kmap(buf->page);
 814         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 815         kunmap(buf->page);
 816
 817         return ret;
 818 }
 819
 820 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 821                                          struct file *out, loff_t *ppos,
 822                                          size_t len, unsigned int flags)
 823 {
 824         ssize_t ret;
 825
 826         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 827         if (ret > 0)
 828                 *ppos += ret;
 829
 830         return ret;
 831 }
 832
 833 /**
 834  * generic_splice_sendpage - splice data from a pipe to a socket
 835  * @pipe:       pipe to splice from
 836  * @out:        socket to write to
 837  * @ppos:       position in @out
 838  * @len:        number of bytes to splice
 839  * @flags:      splice modifier flags
 840  *
 841  * Description:
 842  *    Will send @len bytes from the pipe to a network socket. No data copying
 843  *    is involved.
 844  *
 845  */
 846 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 847                                 loff_t *ppos, size_t len, unsigned int flags)
 848 {
 849         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 850 }
 851
 852 EXPORT_SYMBOL(generic_splice_sendpage);
 853
 854 /*
 855  * Attempt to initiate a splice from pipe to file.
 856  */
 857 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 858                            loff_t *ppos, size_t len, unsigned int flags)
 859 {
 860         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 861                                 loff_t *, size_t, unsigned int);
 862
 863         if (out->f_op->splice_write)
 864                 splice_write = out->f_op->splice_write;
 865         else
 866                 splice_write = default_file_splice_write;
 867
 868         return splice_write(pipe, out, ppos, len, flags);
 869 }
 870
 871 /*
 872  * Attempt to initiate a splice from a file to a pipe.
 873  */
 874 static long do_splice_to(struct file *in, loff_t *ppos,
 875                          struct pipe_inode_info *pipe, size_t len,
 876                          unsigned int flags)
 877 {
 878         ssize_t (*splice_read)(struct file *, loff_t *,
 879                                struct pipe_inode_info *, size_t, unsigned int);
 880         int ret;
 881
 882         if (unlikely(!(in->f_mode & FMODE_READ)))
 883                 return -EBADF;
 884
 885         ret = rw_verify_area(READ, in, ppos, len);
 886         if (unlikely(ret < 0))
 887                 return ret;
 888
 889         if (unlikely(len > MAX_RW_COUNT))
 890                 len = MAX_RW_COUNT;
 891
 892         if (in->f_op->splice_read)
 893                 splice_read = in->f_op->splice_read;
 894         else
 895                 splice_read = default_file_splice_read;
 896
 897         return splice_read(in, ppos, pipe, len, flags);
 898 }
 899
 900 /**
 901  * splice_direct_to_actor - splices data directly between two non-pipes
 902  * @in:         file to splice from
 903  * @sd:         actor information on where to splice to
 904  * @actor:      handles the data splicing
 905  *
 906  * Description:
 907  *    This is a special case helper to splice directly between two
 908  *    points, without requiring an explicit pipe. Internally an allocated
 909  *    pipe is cached in the process, and reused during the lifetime of
 910  *    that process.
 911  *
 912  */
 913 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 914                                splice_direct_actor *actor)
 915 {
 916         struct pipe_inode_info *pipe;
 917         long ret, bytes;
 918         umode_t i_mode;
 919         size_t len;
 920         int i, flags, more;
 921
 922         /*
 923          * We require the input being a regular file, as we don't want to
 924          * randomly drop data for eg socket -> socket splicing. Use the
 925          * piped splicing for that!
 926          */
 927         i_mode = file_inode(in)->i_mode;
 928         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 929                 return -EINVAL;
 930
 931         /*
 932          * neither in nor out is a pipe, setup an internal pipe attached to
 933          * 'out' and transfer the wanted data from 'in' to 'out' through that
 934          */
 935         pipe = current->splice_pipe;
 936         if (unlikely(!pipe)) {
 937                 pipe = alloc_pipe_info();
 938                 if (!pipe)
 939                         return -ENOMEM;
 940
 941                 /*
 942                  * We don't have an immediate reader, but we'll read the stuff
 943                  * out of the pipe right after the splice_to_pipe(). So set
 944                  * PIPE_READERS appropriately.
 945                  */
 946                 pipe->readers = 1;
 947
 948                 current->splice_pipe = pipe;
 949         }
 950
 951         /*
 952          * Do the splice.
 953          */
 954         ret = 0;
 955         bytes = 0;
 956         len = sd->total_len;
 957         flags = sd->flags;
 958
 959         /*
 960          * Don't block on output, we have to drain the direct pipe.
 961          */
 962         sd->flags &= ~SPLICE_F_NONBLOCK;
 963         more = sd->flags & SPLICE_F_MORE;
 964
 965         while (len) {
 966                 size_t read_len;
 967                 loff_t pos = sd->pos, prev_pos = pos;
 968
 969                 ret = do_splice_to(in, &pos, pipe, len, flags);
 970                 if (unlikely(ret <= 0))
 971                         goto out_release;
 972
 973                 read_len = ret;
 974                 sd->total_len = read_len;
 975
 976                 /*
 977                  * If more data is pending, set SPLICE_F_MORE
 978                  * If this is the last data and SPLICE_F_MORE was not set
 979                  * initially, clears it.
 980                  */
 981                 if (read_len < len)
 982                         sd->flags |= SPLICE_F_MORE;
 983                 else if (!more)
 984                         sd->flags &= ~SPLICE_F_MORE;
 985                 /*
 986                  * NOTE: nonblocking mode only applies to the input. We
 987                  * must not do the output in nonblocking mode as then we
 988                  * could get stuck data in the internal pipe:
 989                  */
 990                 ret = actor(pipe, sd);
 991                 if (unlikely(ret <= 0)) {
 992                         sd->pos = prev_pos;
 993                         goto out_release;
 994                 }
 995
 996                 bytes += ret;
 997                 len -= ret;
 998                 sd->pos = pos;
 999
1000                 if (ret < read_len) {
1001                         sd->pos = prev_pos + ret;
1002                         goto out_release;
1003                 }
1004         }
1005
1006 done:
1007         pipe->nrbufs = pipe->curbuf = 0;
1008         file_accessed(in);
1009         return bytes;
1010
1011 out_release:
1012         /*
1013          * If we did an incomplete transfer we must release
1014          * the pipe buffers in question:
1015          */
1016         for (i = 0; i < pipe->buffers; i++) {
1017                 struct pipe_buffer *buf = pipe->bufs + i;
1018
1019                 if (buf->ops)
1020                         pipe_buf_release(pipe, buf);
1021         }
1022
1023         if (!bytes)
1024                 bytes = ret;
1025
1026         goto done;
1027 }
1028 EXPORT_SYMBOL(splice_direct_to_actor);
1029
1030 static int direct_splice_actor(struct pipe_inode_info *pipe,
1031                                struct splice_desc *sd)
1032 {
1033         struct file *file = sd->u.file;
1034
1035         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1036                               sd->flags);
1037 }
1038
1039 /**
1040  * do_splice_direct - splices data directly between two files
1041  * @in:         file to splice from
1042  * @ppos:       input file offset
1043  * @out:        file to splice to
1044  * @opos:       output file offset
1045  * @len:        number of bytes to splice
1046  * @flags:      splice modifier flags
1047  *
1048  * Description:
1049  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1050  *    doing it in the application would incur an extra system call
1051  *    (splice in + splice out, as compared to just sendfile()). So this helper
1052  *    can splice directly through a process-private pipe.
1053  *
1054  */
1055 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1056                       loff_t *opos, size_t len, unsigned int flags)
1057 {
1058         struct splice_desc sd = {
1059                 .len            = len,
1060                 .total_len      = len,
1061                 .flags          = flags,
1062                 .pos            = *ppos,
1063                 .u.file         = out,
1064                 .opos           = opos,
1065         };
1066         long ret;
1067
1068         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1069                 return -EBADF;
1070
1071         if (unlikely(out->f_flags & O_APPEND))
1072                 return -EINVAL;
1073
1074         ret = rw_verify_area(WRITE, out, opos, len);
1075         if (unlikely(ret < 0))
1076                 return ret;
1077
1078         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1079         if (ret > 0)
1080                 *ppos = sd.pos;
1081
1082         return ret;
1083 }
1084 EXPORT_SYMBOL(do_splice_direct);
1085
1086 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1087 {
1088         while (pipe->nrbufs == pipe->buffers) {
1089                 if (flags & SPLICE_F_NONBLOCK)
1090                         return -EAGAIN;
1091                 if (signal_pending(current))
1092                         return -ERESTARTSYS;
1093                 pipe->waiting_writers++;
1094                 pipe_wait(pipe);
1095                 pipe->waiting_writers--;
1096         }
1097         return 0;
1098 }
1099
1100 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1101                                struct pipe_inode_info *opipe,
1102                                size_t len, unsigned int flags);
1103
1104 /*
1105  * Determine where to splice to/from.
1106  */
1107 static long do_splice(struct file *in, loff_t __user *off_in,
1108                       struct file *out, loff_t __user *off_out,
1109                       size_t len, unsigned int flags)
1110 {
1111         struct pipe_inode_info *ipipe;
1112         struct pipe_inode_info *opipe;
1113         loff_t offset;
1114         long ret;
1115
1116         ipipe = get_pipe_info(in);
1117         opipe = get_pipe_info(out);
1118
1119         if (ipipe && opipe) {
1120                 if (off_in || off_out)
1121                         return -ESPIPE;
1122
1123                 if (!(in->f_mode & FMODE_READ))
1124                         return -EBADF;
1125
1126                 if (!(out->f_mode & FMODE_WRITE))
1127                         return -EBADF;
1128
1129                 /* Splicing to self would be fun, but... */
1130                 if (ipipe == opipe)
1131                         return -EINVAL;
1132
1133                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1134         }
1135
1136         if (ipipe) {
1137                 if (off_in)
1138                         return -ESPIPE;
1139                 if (off_out) {
1140                         if (!(out->f_mode & FMODE_PWRITE))
1141                                 return -EINVAL;
1142                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1143                                 return -EFAULT;
1144                 } else {
1145                         offset = out->f_pos;
1146                 }
1147
1148                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1149                         return -EBADF;
1150
1151                 if (unlikely(out->f_flags & O_APPEND))
1152                         return -EINVAL;
1153
1154                 ret = rw_verify_area(WRITE, out, &offset, len);
1155                 if (unlikely(ret < 0))
1156                         return ret;
1157
1158                 file_start_write(out);
1159                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1160                 file_end_write(out);
1161
1162                 if (!off_out)
1163                         out->f_pos = offset;
1164                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1165                         ret = -EFAULT;
1166
1167                 return ret;
1168         }
1169
1170         if (opipe) {
1171                 if (off_out)
1172                         return -ESPIPE;
1173                 if (off_in) {
1174                         if (!(in->f_mode & FMODE_PREAD))
1175                                 return -EINVAL;
1176                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1177                                 return -EFAULT;
1178                 } else {
1179                         offset = in->f_pos;
1180                 }
1181
1182                 pipe_lock(opipe);
1183                 ret = wait_for_space(opipe, flags);
1184                 if (!ret)
1185                         ret = do_splice_to(in, &offset, opipe, len, flags);
1186                 pipe_unlock(opipe);
1187                 if (ret > 0)
1188                         wakeup_pipe_readers(opipe);
1189                 if (!off_in)
1190                         in->f_pos = offset;
1191                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1192                         ret = -EFAULT;
1193
1194                 return ret;
1195         }
1196
1197         return -EINVAL;
1198 }
1199
1200 static int iter_to_pipe(struct iov_iter *from,
1201                         struct pipe_inode_info *pipe,
1202                         unsigned flags)
1203 {
1204         struct pipe_buffer buf = {
1205                 .ops = &user_page_pipe_buf_ops,
1206                 .flags = flags
1207         };
1208         size_t total = 0;
1209         int ret = 0;
1210         bool failed = false;
1211
1212         while (iov_iter_count(from) && !failed) {
1213                 struct page *pages[16];
1214                 ssize_t copied;
1215                 size_t start;
1216                 int n;
1217
1218                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1219                 if (copied <= 0) {
1220                         ret = copied;
1221                         break;
1222                 }
1223
1224                 for (n = 0; copied; n++, start = 0) {
1225                         int size = min_t(int, copied, PAGE_SIZE - start);
1226                         if (!failed) {
1227                                 buf.page = pages[n];
1228                                 buf.offset = start;
1229                                 buf.len = size;
1230                                 ret = add_to_pipe(pipe, &buf);
1231                                 if (unlikely(ret < 0)) {
1232                                         failed = true;
1233                                 } else {
1234                                         iov_iter_advance(from, ret);
1235                                         total += ret;
1236                                 }
1237                         } else {
1238                                 put_page(pages[n]);
1239                         }
1240                         copied -= size;
1241                 }
1242         }
1243         return total ? total : ret;
1244 }
1245
1246 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1247                         struct splice_desc *sd)
1248 {
1249         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1250         return n == sd->len ? n : -EFAULT;
1251 }
1252
1253 /*
1254  * For lack of a better implementation, implement vmsplice() to userspace
1255  * as a simple copy of the pipes pages to the user iov.
1256  */
1257 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1258                              unsigned long nr_segs, unsigned int flags)
1259 {
1260         struct pipe_inode_info *pipe;
1261         struct splice_desc sd;
1262         long ret;
1263         struct iovec iovstack[UIO_FASTIOV];
1264         struct iovec *iov = iovstack;
1265         struct iov_iter iter;
1266
1267         pipe = get_pipe_info(file);
1268         if (!pipe)
1269                 return -EBADF;
1270
1271         ret = import_iovec(READ, uiov, nr_segs,
1272                            ARRAY_SIZE(iovstack), &iov, &iter);
1273         if (ret < 0)
1274                 return ret;
1275
1276         sd.total_len = iov_iter_count(&iter);
1277         sd.len = 0;
1278         sd.flags = flags;
1279         sd.u.data = &iter;
1280         sd.pos = 0;
1281
1282         if (sd.total_len) {
1283                 pipe_lock(pipe);
1284                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1285                 pipe_unlock(pipe);
1286         }
1287
1288         kfree(iov);
1289         return ret;
1290 }
1291
1292 /*
1293  * vmsplice splices a user address range into a pipe. It can be thought of
1294  * as splice-from-memory, where the regular splice is splice-from-file (or
1295  * to file). In both cases the output is a pipe, naturally.
1296  */
1297 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1298                              unsigned long nr_segs, unsigned int flags)
1299 {
1300         struct pipe_inode_info *pipe;
1301         struct iovec iovstack[UIO_FASTIOV];
1302         struct iovec *iov = iovstack;
1303         struct iov_iter from;
1304         long ret;
1305         unsigned buf_flag = 0;
1306
1307         if (flags & SPLICE_F_GIFT)
1308                 buf_flag = PIPE_BUF_FLAG_GIFT;
1309
1310         pipe = get_pipe_info(file);
1311         if (!pipe)
1312                 return -EBADF;
1313
1314         ret = import_iovec(WRITE, uiov, nr_segs,
1315                            ARRAY_SIZE(iovstack), &iov, &from);
1316         if (ret < 0)
1317                 return ret;
1318
1319         pipe_lock(pipe);
1320         ret = wait_for_space(pipe, flags);
1321         if (!ret)
1322                 ret = iter_to_pipe(&from, pipe, buf_flag);
1323         pipe_unlock(pipe);
1324         if (ret > 0)
1325                 wakeup_pipe_readers(pipe);
1326         kfree(iov);
1327         return ret;
1328 }
1329
1330 /*
1331  * Note that vmsplice only really supports true splicing _from_ user memory
1332  * to a pipe, not the other way around. Splicing from user memory is a simple
1333  * operation that can be supported without any funky alignment restrictions
1334  * or nasty vm tricks. We simply map in the user memory and fill them into
1335  * a pipe. The reverse isn't quite as easy, though. There are two possible
1336  * solutions for that:
1337  *
1338  *      - memcpy() the data internally, at which point we might as well just
1339  *        do a regular read() on the buffer anyway.
1340  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1341  *        has restriction limitations on both ends of the pipe).
1342  *
1343  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1344  *
1345  */
1346 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1347                 unsigned long, nr_segs, unsigned int, flags)
1348 {
1349         struct fd f;
1350         long error;
1351
1352         if (unlikely(nr_segs > UIO_MAXIOV))
1353                 return -EINVAL;
1354         else if (unlikely(!nr_segs))
1355                 return 0;
1356
1357         error = -EBADF;
1358         f = fdget(fd);
1359         if (f.file) {
1360                 if (f.file->f_mode & FMODE_WRITE)
1361                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1362                 else if (f.file->f_mode & FMODE_READ)
1363                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1364
1365                 fdput(f);
1366         }
1367
1368         return error;
1369 }
1370
1371 #ifdef CONFIG_COMPAT
1372 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1373                     unsigned int, nr_segs, unsigned int, flags)
1374 {
1375         unsigned i;
1376         struct iovec __user *iov;
1377         if (nr_segs > UIO_MAXIOV)
1378                 return -EINVAL;
1379         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1380         for (i = 0; i < nr_segs; i++) {
1381                 struct compat_iovec v;
1382                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1383                     get_user(v.iov_len, &iov32[i].iov_len) ||
1384                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1385                     put_user(v.iov_len, &iov[i].iov_len))
1386                         return -EFAULT;
1387         }
1388         return sys_vmsplice(fd, iov, nr_segs, flags);
1389 }
1390 #endif
1391
1392 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1393                 int, fd_out, loff_t __user *, off_out,
1394                 size_t, len, unsigned int, flags)
1395 {
1396         struct fd in, out;
1397         long error;
1398
1399         if (unlikely(!len))
1400                 return 0;
1401
1402         error = -EBADF;
1403         in = fdget(fd_in);
1404         if (in.file) {
1405                 if (in.file->f_mode & FMODE_READ) {
1406                         out = fdget(fd_out);
1407                         if (out.file) {
1408                                 if (out.file->f_mode & FMODE_WRITE)
1409                                         error = do_splice(in.file, off_in,
1410                                                           out.file, off_out,
1411                                                           len, flags);
1412                                 fdput(out);
1413                         }
1414                 }
1415                 fdput(in);
1416         }
1417         return error;
1418 }
1419
1420 /*
1421  * Make sure there's data to read. Wait for input if we can, otherwise
1422  * return an appropriate error.
1423  */
1424 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1425 {
1426         int ret;
1427
1428         /*
1429          * Check ->nrbufs without the inode lock first. This function
1430          * is speculative anyways, so missing one is ok.
1431          */
1432         if (pipe->nrbufs)
1433                 return 0;
1434
1435         ret = 0;
1436         pipe_lock(pipe);
1437
1438         while (!pipe->nrbufs) {
1439                 if (signal_pending(current)) {
1440                         ret = -ERESTARTSYS;
1441                         break;
1442                 }
1443                 if (!pipe->writers)
1444                         break;
1445                 if (!pipe->waiting_writers) {
1446                         if (flags & SPLICE_F_NONBLOCK) {
1447                                 ret = -EAGAIN;
1448                                 break;
1449                         }
1450                 }
1451                 pipe_wait(pipe);
1452         }
1453
1454         pipe_unlock(pipe);
1455         return ret;
1456 }
1457
1458 /*
1459  * Make sure there's writeable room. Wait for room if we can, otherwise
1460  * return an appropriate error.
1461  */
1462 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1463 {
1464         int ret;
1465
1466         /*
1467          * Check ->nrbufs without the inode lock first. This function
1468          * is speculative anyways, so missing one is ok.
1469          */
1470         if (pipe->nrbufs < pipe->buffers)
1471                 return 0;
1472
1473         ret = 0;
1474         pipe_lock(pipe);
1475
1476         while (pipe->nrbufs >= pipe->buffers) {
1477                 if (!pipe->readers) {
1478                         send_sig(SIGPIPE, current, 0);
1479                         ret = -EPIPE;
1480                         break;
1481                 }
1482                 if (flags & SPLICE_F_NONBLOCK) {
1483                         ret = -EAGAIN;
1484                         break;
1485                 }
1486                 if (signal_pending(current)) {
1487                         ret = -ERESTARTSYS;
1488                         break;
1489                 }
1490                 pipe->waiting_writers++;
1491                 pipe_wait(pipe);
1492                 pipe->waiting_writers--;
1493         }
1494
1495         pipe_unlock(pipe);
1496         return ret;
1497 }
1498
1499 /*
1500  * Splice contents of ipipe to opipe.
1501  */
1502 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1503                                struct pipe_inode_info *opipe,
1504                                size_t len, unsigned int flags)
1505 {
1506         struct pipe_buffer *ibuf, *obuf;
1507         int ret = 0, nbuf;
1508         bool input_wakeup = false;
1509
1510
1511 retry:
1512         ret = ipipe_prep(ipipe, flags);
1513         if (ret)
1514                 return ret;
1515
1516         ret = opipe_prep(opipe, flags);
1517         if (ret)
1518                 return ret;
1519
1520         /*
1521          * Potential ABBA deadlock, work around it by ordering lock
1522          * grabbing by pipe info address. Otherwise two different processes
1523          * could deadlock (one doing tee from A -> B, the other from B -> A).
1524          */
1525         pipe_double_lock(ipipe, opipe);
1526
1527         do {
1528                 if (!opipe->readers) {
1529                         send_sig(SIGPIPE, current, 0);
1530                         if (!ret)
1531                                 ret = -EPIPE;
1532                         break;
1533                 }
1534
1535                 if (!ipipe->nrbufs && !ipipe->writers)
1536                         break;
1537
1538                 /*
1539                  * Cannot make any progress, because either the input
1540                  * pipe is empty or the output pipe is full.
1541                  */
1542                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1543                         /* Already processed some buffers, break */
1544                         if (ret)
1545                                 break;
1546
1547                         if (flags & SPLICE_F_NONBLOCK) {
1548                                 ret = -EAGAIN;
1549                                 break;
1550                         }
1551
1552                         /*
1553                          * We raced with another reader/writer and haven't
1554                          * managed to process any buffers.  A zero return
1555                          * value means EOF, so retry instead.
1556                          */
1557                         pipe_unlock(ipipe);
1558                         pipe_unlock(opipe);
1559                         goto retry;
1560                 }
1561
1562                 ibuf = ipipe->bufs + ipipe->curbuf;
1563                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1564                 obuf = opipe->bufs + nbuf;
1565
1566                 if (len >= ibuf->len) {
1567                         /*
1568                          * Simply move the whole buffer from ipipe to opipe
1569                          */
1570                         *obuf = *ibuf;
1571                         ibuf->ops = NULL;
1572                         opipe->nrbufs++;
1573                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1574                         ipipe->nrbufs--;
1575                         input_wakeup = true;
1576                 } else {
1577                         /*
1578                          * Get a reference to this pipe buffer,
1579                          * so we can copy the contents over.
1580                          */
1581                         pipe_buf_get(ipipe, ibuf);
1582                         *obuf = *ibuf;
1583
1584                         /*
1585                          * Don't inherit the gift flag, we need to
1586                          * prevent multiple steals of this page.
1587                          */
1588                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1589
1590                         obuf->len = len;
1591                         opipe->nrbufs++;
1592                         ibuf->offset += obuf->len;
1593                         ibuf->len -= obuf->len;
1594                 }
1595                 ret += obuf->len;
1596                 len -= obuf->len;
1597         } while (len);
1598
1599         pipe_unlock(ipipe);
1600         pipe_unlock(opipe);
1601
1602         /*
1603          * If we put data in the output pipe, wakeup any potential readers.
1604          */
1605         if (ret > 0)
1606                 wakeup_pipe_readers(opipe);
1607
1608         if (input_wakeup)
1609                 wakeup_pipe_writers(ipipe);
1610
1611         return ret;
1612 }
1613
1614 /*
1615  * Link contents of ipipe to opipe.
1616  */
1617 static int link_pipe(struct pipe_inode_info *ipipe,
1618                      struct pipe_inode_info *opipe,
1619                      size_t len, unsigned int flags)
1620 {
1621         struct pipe_buffer *ibuf, *obuf;
1622         int ret = 0, i = 0, nbuf;
1623
1624         /*
1625          * Potential ABBA deadlock, work around it by ordering lock
1626          * grabbing by pipe info address. Otherwise two different processes
1627          * could deadlock (one doing tee from A -> B, the other from B -> A).
1628          */
1629         pipe_double_lock(ipipe, opipe);
1630
1631         do {
1632                 if (!opipe->readers) {
1633                         send_sig(SIGPIPE, current, 0);
1634                         if (!ret)
1635                                 ret = -EPIPE;
1636                         break;
1637                 }
1638
1639                 /*
1640                  * If we have iterated all input buffers or ran out of
1641                  * output room, break.
1642                  */
1643                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1644                         break;
1645
1646                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1647                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1648
1649                 /*
1650                  * Get a reference to this pipe buffer,
1651                  * so we can copy the contents over.
1652                  */
1653                 pipe_buf_get(ipipe, ibuf);
1654
1655                 obuf = opipe->bufs + nbuf;
1656                 *obuf = *ibuf;
1657
1658                 /*
1659                  * Don't inherit the gift flag, we need to
1660                  * prevent multiple steals of this page.
1661                  */
1662                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1663
1664                 if (obuf->len > len)
1665                         obuf->len = len;
1666
1667                 opipe->nrbufs++;
1668                 ret += obuf->len;
1669                 len -= obuf->len;
1670                 i++;
1671         } while (len);
1672
1673         /*
1674          * return EAGAIN if we have the potential of some data in the
1675          * future, otherwise just return 0
1676          */
1677         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1678                 ret = -EAGAIN;
1679
1680         pipe_unlock(ipipe);
1681         pipe_unlock(opipe);
1682
1683         /*
1684          * If we put data in the output pipe, wakeup any potential readers.
1685          */
1686         if (ret > 0)
1687                 wakeup_pipe_readers(opipe);
1688
1689         return ret;
1690 }
1691
1692 /*
1693  * This is a tee(1) implementation that works on pipes. It doesn't copy
1694  * any data, it simply references the 'in' pages on the 'out' pipe.
1695  * The 'flags' used are the SPLICE_F_* variants, currently the only
1696  * applicable one is SPLICE_F_NONBLOCK.
1697  */
1698 static long do_tee(struct file *in, struct file *out, size_t len,
1699                    unsigned int flags)
1700 {
1701         struct pipe_inode_info *ipipe = get_pipe_info(in);
1702         struct pipe_inode_info *opipe = get_pipe_info(out);
1703         int ret = -EINVAL;
1704
1705         /*
1706          * Duplicate the contents of ipipe to opipe without actually
1707          * copying the data.
1708          */
1709         if (ipipe && opipe && ipipe != opipe) {
1710                 /*
1711                  * Keep going, unless we encounter an error. The ipipe/opipe
1712                  * ordering doesn't really matter.
1713                  */
1714                 ret = ipipe_prep(ipipe, flags);
1715                 if (!ret) {
1716                         ret = opipe_prep(opipe, flags);
1717                         if (!ret)
1718                                 ret = link_pipe(ipipe, opipe, len, flags);
1719                 }
1720         }
1721
1722         return ret;
1723 }
1724
1725 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1726 {
1727         struct fd in;
1728         int error;
1729
1730         if (unlikely(!len))
1731                 return 0;
1732
1733         error = -EBADF;
1734         in = fdget(fdin);
1735         if (in.file) {
1736                 if (in.file->f_mode & FMODE_READ) {
1737                         struct fd out = fdget(fdout);
1738                         if (out.file) {
1739                                 if (out.file->f_mode & FMODE_WRITE)
1740                                         error = do_tee(in.file, out.file,
1741                                                         len, flags);
1742                                 fdput(out);
1743                         }
1744                 }
1745                 fdput(in);
1746         }
1747
1748         return error;
1749 }