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 <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  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 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 251 {
 252         put_page(spd->pages[i]);
 253 }
 254
 255 /*
 256  * Check if we need to grow the arrays holding pages and partial page
 257  * descriptions.
 258  */
 259 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 260 {
 261         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 262
 263         spd->nr_pages_max = buffers;
 264         if (buffers <= PIPE_DEF_BUFFERS)
 265                 return 0;
 266
 267         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 268         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 269
 270         if (spd->pages && spd->partial)
 271                 return 0;
 272
 273         kfree(spd->pages);
 274         kfree(spd->partial);
 275         return -ENOMEM;
 276 }
 277
 278 void splice_shrink_spd(struct splice_pipe_desc *spd)
 279 {
 280         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 281                 return;
 282
 283         kfree(spd->pages);
 284         kfree(spd->partial);
 285 }
 286
 287 /**
 288  * generic_file_splice_read - splice data from file to a pipe
 289  * @in:         file to splice from
 290  * @ppos:       position in @in
 291  * @pipe:       pipe to splice to
 292  * @len:        number of bytes to splice
 293  * @flags:      splice modifier flags
 294  *
 295  * Description:
 296  *    Will read pages from given file and fill them into a pipe. Can be
 297  *    used as long as it has more or less sane ->read_iter().
 298  *
 299  */
 300 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 301                                  struct pipe_inode_info *pipe, size_t len,
 302                                  unsigned int flags)
 303 {
 304         struct iov_iter to;
 305         struct kiocb kiocb;
 306         int idx, ret;
 307
 308         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 309         idx = to.idx;
 310         init_sync_kiocb(&kiocb, in);
 311         kiocb.ki_pos = *ppos;
 312         ret = call_read_iter(in, &kiocb, &to);
 313         if (ret > 0) {
 314                 *ppos = kiocb.ki_pos;
 315                 file_accessed(in);
 316         } else if (ret < 0) {
 317                 to.idx = idx;
 318                 to.iov_offset = 0;
 319                 iov_iter_advance(&to, 0); /* to free what was emitted */
 320                 /*
 321                  * callers of ->splice_read() expect -EAGAIN on
 322                  * "can't put anything in there", rather than -EFAULT.
 323                  */
 324                 if (ret == -EFAULT)
 325                         ret = -EAGAIN;
 326         }
 327
 328         return ret;
 329 }
 330 EXPORT_SYMBOL(generic_file_splice_read);
 331
 332 const struct pipe_buf_operations default_pipe_buf_ops = {
 333         .can_merge = 0,
 334         .confirm = generic_pipe_buf_confirm,
 335         .release = generic_pipe_buf_release,
 336         .steal = generic_pipe_buf_steal,
 337         .get = generic_pipe_buf_get,
 338 };
 339
 340 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 341                                     struct pipe_buffer *buf)
 342 {
 343         return 1;
 344 }
 345
 346 /* Pipe buffer operations for a socket and similar. */
 347 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 348         .can_merge = 0,
 349         .confirm = generic_pipe_buf_confirm,
 350         .release = generic_pipe_buf_release,
 351         .steal = generic_pipe_buf_nosteal,
 352         .get = generic_pipe_buf_get,
 353 };
 354 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 355
 356 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 357                             unsigned long vlen, loff_t offset)
 358 {
 359         mm_segment_t old_fs;
 360         loff_t pos = offset;
 361         ssize_t res;
 362
 363         old_fs = get_fs();
 364         set_fs(get_ds());
 365         /* The cast to a user pointer is valid due to the set_fs() */
 366         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 367         set_fs(old_fs);
 368
 369         return res;
 370 }
 371
 372 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 373                             loff_t pos)
 374 {
 375         mm_segment_t old_fs;
 376         ssize_t res;
 377
 378         old_fs = get_fs();
 379         set_fs(get_ds());
 380         /* The cast to a user pointer is valid due to the set_fs() */
 381         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 382         set_fs(old_fs);
 383
 384         return res;
 385 }
 386 EXPORT_SYMBOL(kernel_write);
 387
 388 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 389                                  struct pipe_inode_info *pipe, size_t len,
 390                                  unsigned int flags)
 391 {
 392         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 393         struct iov_iter to;
 394         struct page **pages;
 395         unsigned int nr_pages;
 396         size_t offset, dummy, copied = 0;
 397         ssize_t res;
 398         int i;
 399
 400         if (pipe->nrbufs == pipe->buffers)
 401                 return -EAGAIN;
 402
 403         /*
 404          * Try to keep page boundaries matching to source pagecache ones -
 405          * it probably won't be much help, but...
 406          */
 407         offset = *ppos & ~PAGE_MASK;
 408
 409         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 410
 411         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
 412         if (res <= 0)
 413                 return -ENOMEM;
 414
 415         BUG_ON(dummy);
 416         nr_pages = DIV_ROUND_UP(res, PAGE_SIZE);
 417
 418         vec = __vec;
 419         if (nr_pages > PIPE_DEF_BUFFERS) {
 420                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 421                 if (unlikely(!vec)) {
 422                         res = -ENOMEM;
 423                         goto out;
 424                 }
 425         }
 426
 427         pipe->bufs[to.idx].offset = offset;
 428         pipe->bufs[to.idx].len -= offset;
 429
 430         for (i = 0; i < nr_pages; i++) {
 431                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 432                 vec[i].iov_base = page_address(pages[i]) + offset;
 433                 vec[i].iov_len = this_len;
 434                 len -= this_len;
 435                 offset = 0;
 436         }
 437
 438         res = kernel_readv(in, vec, nr_pages, *ppos);
 439         if (res > 0) {
 440                 copied = res;
 441                 *ppos += res;
 442         }
 443
 444         if (vec != __vec)
 445                 kfree(vec);
 446 out:
 447         for (i = 0; i < nr_pages; i++)
 448                 put_page(pages[i]);
 449         kvfree(pages);
 450         iov_iter_advance(&to, copied);  /* truncates and discards */
 451         return res;
 452 }
 453
 454 /*
 455  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 456  * using sendpage(). Return the number of bytes sent.
 457  */
 458 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 459                             struct pipe_buffer *buf, struct splice_desc *sd)
 460 {
 461         struct file *file = sd->u.file;
 462         loff_t pos = sd->pos;
 463         int more;
 464
 465         if (!likely(file->f_op->sendpage))
 466                 return -EINVAL;
 467
 468         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 469
 470         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 471                 more |= MSG_SENDPAGE_NOTLAST;
 472
 473         return file->f_op->sendpage(file, buf->page, buf->offset,
 474                                     sd->len, &pos, more);
 475 }
 476
 477 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 478 {
 479         smp_mb();
 480         if (waitqueue_active(&pipe->wait))
 481                 wake_up_interruptible(&pipe->wait);
 482         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 483 }
 484
 485 /**
 486  * splice_from_pipe_feed - feed available data from a pipe to a file
 487  * @pipe:       pipe to splice from
 488  * @sd:         information to @actor
 489  * @actor:      handler that splices the data
 490  *
 491  * Description:
 492  *    This function loops over the pipe and calls @actor to do the
 493  *    actual moving of a single struct pipe_buffer to the desired
 494  *    destination.  It returns when there's no more buffers left in
 495  *    the pipe or if the requested number of bytes (@sd->total_len)
 496  *    have been copied.  It returns a positive number (one) if the
 497  *    pipe needs to be filled with more data, zero if the required
 498  *    number of bytes have been copied and -errno on error.
 499  *
 500  *    This, together with splice_from_pipe_{begin,end,next}, may be
 501  *    used to implement the functionality of __splice_from_pipe() when
 502  *    locking is required around copying the pipe buffers to the
 503  *    destination.
 504  */
 505 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 506                           splice_actor *actor)
 507 {
 508         int ret;
 509
 510         while (pipe->nrbufs) {
 511                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 512
 513                 sd->len = buf->len;
 514                 if (sd->len > sd->total_len)
 515                         sd->len = sd->total_len;
 516
 517                 ret = pipe_buf_confirm(pipe, buf);
 518                 if (unlikely(ret)) {
 519                         if (ret == -ENODATA)
 520                                 ret = 0;
 521                         return ret;
 522                 }
 523
 524                 ret = actor(pipe, buf, sd);
 525                 if (ret <= 0)
 526                         return ret;
 527
 528                 buf->offset += ret;
 529                 buf->len -= ret;
 530
 531                 sd->num_spliced += ret;
 532                 sd->len -= ret;
 533                 sd->pos += ret;
 534                 sd->total_len -= ret;
 535
 536                 if (!buf->len) {
 537                         pipe_buf_release(pipe, buf);
 538                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 539                         pipe->nrbufs--;
 540                         if (pipe->files)
 541                                 sd->need_wakeup = true;
 542                 }
 543
 544                 if (!sd->total_len)
 545                         return 0;
 546         }
 547
 548         return 1;
 549 }
 550
 551 /**
 552  * splice_from_pipe_next - wait for some data to splice from
 553  * @pipe:       pipe to splice from
 554  * @sd:         information about the splice operation
 555  *
 556  * Description:
 557  *    This function will wait for some data and return a positive
 558  *    value (one) if pipe buffers are available.  It will return zero
 559  *    or -errno if no more data needs to be spliced.
 560  */
 561 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 562 {
 563         /*
 564          * Check for signal early to make process killable when there are
 565          * always buffers available
 566          */
 567         if (signal_pending(current))
 568                 return -ERESTARTSYS;
 569
 570         while (!pipe->nrbufs) {
 571                 if (!pipe->writers)
 572                         return 0;
 573
 574                 if (!pipe->waiting_writers && sd->num_spliced)
 575                         return 0;
 576
 577                 if (sd->flags & SPLICE_F_NONBLOCK)
 578                         return -EAGAIN;
 579
 580                 if (signal_pending(current))
 581                         return -ERESTARTSYS;
 582
 583                 if (sd->need_wakeup) {
 584                         wakeup_pipe_writers(pipe);
 585                         sd->need_wakeup = false;
 586                 }
 587
 588                 pipe_wait(pipe);
 589         }
 590
 591         return 1;
 592 }
 593
 594 /**
 595  * splice_from_pipe_begin - start splicing from pipe
 596  * @sd:         information about the splice operation
 597  *
 598  * Description:
 599  *    This function should be called before a loop containing
 600  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 601  *    initialize the necessary fields of @sd.
 602  */
 603 static void splice_from_pipe_begin(struct splice_desc *sd)
 604 {
 605         sd->num_spliced = 0;
 606         sd->need_wakeup = false;
 607 }
 608
 609 /**
 610  * splice_from_pipe_end - finish splicing from pipe
 611  * @pipe:       pipe to splice from
 612  * @sd:         information about the splice operation
 613  *
 614  * Description:
 615  *    This function will wake up pipe writers if necessary.  It should
 616  *    be called after a loop containing splice_from_pipe_next() and
 617  *    splice_from_pipe_feed().
 618  */
 619 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 620 {
 621         if (sd->need_wakeup)
 622                 wakeup_pipe_writers(pipe);
 623 }
 624
 625 /**
 626  * __splice_from_pipe - splice data from a pipe to given actor
 627  * @pipe:       pipe to splice from
 628  * @sd:         information to @actor
 629  * @actor:      handler that splices the data
 630  *
 631  * Description:
 632  *    This function does little more than loop over the pipe and call
 633  *    @actor to do the actual moving of a single struct pipe_buffer to
 634  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 635  *    pipe_to_user.
 636  *
 637  */
 638 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 639                            splice_actor *actor)
 640 {
 641         int ret;
 642
 643         splice_from_pipe_begin(sd);
 644         do {
 645                 cond_resched();
 646                 ret = splice_from_pipe_next(pipe, sd);
 647                 if (ret > 0)
 648                         ret = splice_from_pipe_feed(pipe, sd, actor);
 649         } while (ret > 0);
 650         splice_from_pipe_end(pipe, sd);
 651
 652         return sd->num_spliced ? sd->num_spliced : ret;
 653 }
 654 EXPORT_SYMBOL(__splice_from_pipe);
 655
 656 /**
 657  * splice_from_pipe - splice data from a pipe to a file
 658  * @pipe:       pipe to splice from
 659  * @out:        file to splice to
 660  * @ppos:       position in @out
 661  * @len:        how many bytes to splice
 662  * @flags:      splice modifier flags
 663  * @actor:      handler that splices the data
 664  *
 665  * Description:
 666  *    See __splice_from_pipe. This function locks the pipe inode,
 667  *    otherwise it's identical to __splice_from_pipe().
 668  *
 669  */
 670 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 671                          loff_t *ppos, size_t len, unsigned int flags,
 672                          splice_actor *actor)
 673 {
 674         ssize_t ret;
 675         struct splice_desc sd = {
 676                 .total_len = len,
 677                 .flags = flags,
 678                 .pos = *ppos,
 679                 .u.file = out,
 680         };
 681
 682         pipe_lock(pipe);
 683         ret = __splice_from_pipe(pipe, &sd, actor);
 684         pipe_unlock(pipe);
 685
 686         return ret;
 687 }
 688
 689 /**
 690  * iter_file_splice_write - splice data from a pipe to a file
 691  * @pipe:       pipe info
 692  * @out:        file to write to
 693  * @ppos:       position in @out
 694  * @len:        number of bytes to splice
 695  * @flags:      splice modifier flags
 696  *
 697  * Description:
 698  *    Will either move or copy pages (determined by @flags options) from
 699  *    the given pipe inode to the given file.
 700  *    This one is ->write_iter-based.
 701  *
 702  */
 703 ssize_t
 704 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 705                           loff_t *ppos, size_t len, unsigned int flags)
 706 {
 707         struct splice_desc sd = {
 708                 .total_len = len,
 709                 .flags = flags,
 710                 .pos = *ppos,
 711                 .u.file = out,
 712         };
 713         int nbufs = pipe->buffers;
 714         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 715                                         GFP_KERNEL);
 716         ssize_t ret;
 717
 718         if (unlikely(!array))
 719                 return -ENOMEM;
 720
 721         pipe_lock(pipe);
 722
 723         splice_from_pipe_begin(&sd);
 724         while (sd.total_len) {
 725                 struct iov_iter from;
 726                 size_t left;
 727                 int n, idx;
 728
 729                 ret = splice_from_pipe_next(pipe, &sd);
 730                 if (ret <= 0)
 731                         break;
 732
 733                 if (unlikely(nbufs < pipe->buffers)) {
 734                         kfree(array);
 735                         nbufs = pipe->buffers;
 736                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 737                                         GFP_KERNEL);
 738                         if (!array) {
 739                                 ret = -ENOMEM;
 740                                 break;
 741                         }
 742                 }
 743
 744                 /* build the vector */
 745                 left = sd.total_len;
 746                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 747                         struct pipe_buffer *buf = pipe->bufs + idx;
 748                         size_t this_len = buf->len;
 749
 750                         if (this_len > left)
 751                                 this_len = left;
 752
 753                         if (idx == pipe->buffers - 1)
 754                                 idx = -1;
 755
 756                         ret = pipe_buf_confirm(pipe, buf);
 757                         if (unlikely(ret)) {
 758                                 if (ret == -ENODATA)
 759                                         ret = 0;
 760                                 goto done;
 761                         }
 762
 763                         array[n].bv_page = buf->page;
 764                         array[n].bv_len = this_len;
 765                         array[n].bv_offset = buf->offset;
 766                         left -= this_len;
 767                 }
 768
 769                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 770                               sd.total_len - left);
 771                 ret = vfs_iter_write(out, &from, &sd.pos);
 772                 if (ret <= 0)
 773                         break;
 774
 775                 sd.num_spliced += ret;
 776                 sd.total_len -= ret;
 777                 *ppos = sd.pos;
 778
 779                 /* dismiss the fully eaten buffers, adjust the partial one */
 780                 while (ret) {
 781                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 782                         if (ret >= buf->len) {
 783                                 ret -= buf->len;
 784                                 buf->len = 0;
 785                                 pipe_buf_release(pipe, buf);
 786                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 787                                 pipe->nrbufs--;
 788                                 if (pipe->files)
 789                                         sd.need_wakeup = true;
 790                         } else {
 791                                 buf->offset += ret;
 792                                 buf->len -= ret;
 793                                 ret = 0;
 794                         }
 795                 }
 796         }
 797 done:
 798         kfree(array);
 799         splice_from_pipe_end(pipe, &sd);
 800
 801         pipe_unlock(pipe);
 802
 803         if (sd.num_spliced)
 804                 ret = sd.num_spliced;
 805
 806         return ret;
 807 }
 808
 809 EXPORT_SYMBOL(iter_file_splice_write);
 810
 811 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 812                           struct splice_desc *sd)
 813 {
 814         int ret;
 815         void *data;
 816         loff_t tmp = sd->pos;
 817
 818         data = kmap(buf->page);
 819         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 820         kunmap(buf->page);
 821
 822         return ret;
 823 }
 824
 825 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 826                                          struct file *out, loff_t *ppos,
 827                                          size_t len, unsigned int flags)
 828 {
 829         ssize_t ret;
 830
 831         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 832         if (ret > 0)
 833                 *ppos += ret;
 834
 835         return ret;
 836 }
 837
 838 /**
 839  * generic_splice_sendpage - splice data from a pipe to a socket
 840  * @pipe:       pipe to splice from
 841  * @out:        socket to write to
 842  * @ppos:       position in @out
 843  * @len:        number of bytes to splice
 844  * @flags:      splice modifier flags
 845  *
 846  * Description:
 847  *    Will send @len bytes from the pipe to a network socket. No data copying
 848  *    is involved.
 849  *
 850  */
 851 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 852                                 loff_t *ppos, size_t len, unsigned int flags)
 853 {
 854         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 855 }
 856
 857 EXPORT_SYMBOL(generic_splice_sendpage);
 858
 859 /*
 860  * Attempt to initiate a splice from pipe to file.
 861  */
 862 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 863                            loff_t *ppos, size_t len, unsigned int flags)
 864 {
 865         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 866                                 loff_t *, size_t, unsigned int);
 867
 868         if (out->f_op->splice_write)
 869                 splice_write = out->f_op->splice_write;
 870         else
 871                 splice_write = default_file_splice_write;
 872
 873         return splice_write(pipe, out, ppos, len, flags);
 874 }
 875
 876 /*
 877  * Attempt to initiate a splice from a file to a pipe.
 878  */
 879 static long do_splice_to(struct file *in, loff_t *ppos,
 880                          struct pipe_inode_info *pipe, size_t len,
 881                          unsigned int flags)
 882 {
 883         ssize_t (*splice_read)(struct file *, loff_t *,
 884                                struct pipe_inode_info *, size_t, unsigned int);
 885         int ret;
 886
 887         if (unlikely(!(in->f_mode & FMODE_READ)))
 888                 return -EBADF;
 889
 890         ret = rw_verify_area(READ, in, ppos, len);
 891         if (unlikely(ret < 0))
 892                 return ret;
 893
 894         if (unlikely(len > MAX_RW_COUNT))
 895                 len = MAX_RW_COUNT;
 896
 897         if (in->f_op->splice_read)
 898                 splice_read = in->f_op->splice_read;
 899         else
 900                 splice_read = default_file_splice_read;
 901
 902         return splice_read(in, ppos, pipe, len, flags);
 903 }
 904
 905 /**
 906  * splice_direct_to_actor - splices data directly between two non-pipes
 907  * @in:         file to splice from
 908  * @sd:         actor information on where to splice to
 909  * @actor:      handles the data splicing
 910  *
 911  * Description:
 912  *    This is a special case helper to splice directly between two
 913  *    points, without requiring an explicit pipe. Internally an allocated
 914  *    pipe is cached in the process, and reused during the lifetime of
 915  *    that process.
 916  *
 917  */
 918 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 919                                splice_direct_actor *actor)
 920 {
 921         struct pipe_inode_info *pipe;
 922         long ret, bytes;
 923         umode_t i_mode;
 924         size_t len;
 925         int i, flags, more;
 926
 927         /*
 928          * We require the input being a regular file, as we don't want to
 929          * randomly drop data for eg socket -> socket splicing. Use the
 930          * piped splicing for that!
 931          */
 932         i_mode = file_inode(in)->i_mode;
 933         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 934                 return -EINVAL;
 935
 936         /*
 937          * neither in nor out is a pipe, setup an internal pipe attached to
 938          * 'out' and transfer the wanted data from 'in' to 'out' through that
 939          */
 940         pipe = current->splice_pipe;
 941         if (unlikely(!pipe)) {
 942                 pipe = alloc_pipe_info();
 943                 if (!pipe)
 944                         return -ENOMEM;
 945
 946                 /*
 947                  * We don't have an immediate reader, but we'll read the stuff
 948                  * out of the pipe right after the splice_to_pipe(). So set
 949                  * PIPE_READERS appropriately.
 950                  */
 951                 pipe->readers = 1;
 952
 953                 current->splice_pipe = pipe;
 954         }
 955
 956         /*
 957          * Do the splice.
 958          */
 959         ret = 0;
 960         bytes = 0;
 961         len = sd->total_len;
 962         flags = sd->flags;
 963
 964         /*
 965          * Don't block on output, we have to drain the direct pipe.
 966          */
 967         sd->flags &= ~SPLICE_F_NONBLOCK;
 968         more = sd->flags & SPLICE_F_MORE;
 969
 970         while (len) {
 971                 size_t read_len;
 972                 loff_t pos = sd->pos, prev_pos = pos;
 973
 974                 ret = do_splice_to(in, &pos, pipe, len, flags);
 975                 if (unlikely(ret <= 0))
 976                         goto out_release;
 977
 978                 read_len = ret;
 979                 sd->total_len = read_len;
 980
 981                 /*
 982                  * If more data is pending, set SPLICE_F_MORE
 983                  * If this is the last data and SPLICE_F_MORE was not set
 984                  * initially, clears it.
 985                  */
 986                 if (read_len < len)
 987                         sd->flags |= SPLICE_F_MORE;
 988                 else if (!more)
 989                         sd->flags &= ~SPLICE_F_MORE;
 990                 /*
 991                  * NOTE: nonblocking mode only applies to the input. We
 992                  * must not do the output in nonblocking mode as then we
 993                  * could get stuck data in the internal pipe:
 994                  */
 995                 ret = actor(pipe, sd);
 996                 if (unlikely(ret <= 0)) {
 997                         sd->pos = prev_pos;
 998                         goto out_release;
 999                 }
1000
1001                 bytes += ret;
1002                 len -= ret;
1003                 sd->pos = pos;
1004
1005                 if (ret < read_len) {
1006                         sd->pos = prev_pos + ret;
1007                         goto out_release;
1008                 }
1009         }
1010
1011 done:
1012         pipe->nrbufs = pipe->curbuf = 0;
1013         file_accessed(in);
1014         return bytes;
1015
1016 out_release:
1017         /*
1018          * If we did an incomplete transfer we must release
1019          * the pipe buffers in question:
1020          */
1021         for (i = 0; i < pipe->buffers; i++) {
1022                 struct pipe_buffer *buf = pipe->bufs + i;
1023
1024                 if (buf->ops)
1025                         pipe_buf_release(pipe, buf);
1026         }
1027
1028         if (!bytes)
1029                 bytes = ret;
1030
1031         goto done;
1032 }
1033 EXPORT_SYMBOL(splice_direct_to_actor);
1034
1035 static int direct_splice_actor(struct pipe_inode_info *pipe,
1036                                struct splice_desc *sd)
1037 {
1038         struct file *file = sd->u.file;
1039
1040         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1041                               sd->flags);
1042 }
1043
1044 /**
1045  * do_splice_direct - splices data directly between two files
1046  * @in:         file to splice from
1047  * @ppos:       input file offset
1048  * @out:        file to splice to
1049  * @opos:       output file offset
1050  * @len:        number of bytes to splice
1051  * @flags:      splice modifier flags
1052  *
1053  * Description:
1054  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1055  *    doing it in the application would incur an extra system call
1056  *    (splice in + splice out, as compared to just sendfile()). So this helper
1057  *    can splice directly through a process-private pipe.
1058  *
1059  */
1060 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1061                       loff_t *opos, size_t len, unsigned int flags)
1062 {
1063         struct splice_desc sd = {
1064                 .len            = len,
1065                 .total_len      = len,
1066                 .flags          = flags,
1067                 .pos            = *ppos,
1068                 .u.file         = out,
1069                 .opos           = opos,
1070         };
1071         long ret;
1072
1073         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1074                 return -EBADF;
1075
1076         if (unlikely(out->f_flags & O_APPEND))
1077                 return -EINVAL;
1078
1079         ret = rw_verify_area(WRITE, out, opos, len);
1080         if (unlikely(ret < 0))
1081                 return ret;
1082
1083         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1084         if (ret > 0)
1085                 *ppos = sd.pos;
1086
1087         return ret;
1088 }
1089 EXPORT_SYMBOL(do_splice_direct);
1090
1091 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1092 {
1093         for (;;) {
1094                 if (unlikely(!pipe->readers)) {
1095                         send_sig(SIGPIPE, current, 0);
1096                         return -EPIPE;
1097                 }
1098                 if (pipe->nrbufs != pipe->buffers)
1099                         return 0;
1100                 if (flags & SPLICE_F_NONBLOCK)
1101                         return -EAGAIN;
1102                 if (signal_pending(current))
1103                         return -ERESTARTSYS;
1104                 pipe->waiting_writers++;
1105                 pipe_wait(pipe);
1106                 pipe->waiting_writers--;
1107         }
1108 }
1109
1110 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1111                                struct pipe_inode_info *opipe,
1112                                size_t len, unsigned int flags);
1113
1114 /*
1115  * Determine where to splice to/from.
1116  */
1117 static long do_splice(struct file *in, loff_t __user *off_in,
1118                       struct file *out, loff_t __user *off_out,
1119                       size_t len, unsigned int flags)
1120 {
1121         struct pipe_inode_info *ipipe;
1122         struct pipe_inode_info *opipe;
1123         loff_t offset;
1124         long ret;
1125
1126         ipipe = get_pipe_info(in);
1127         opipe = get_pipe_info(out);
1128
1129         if (ipipe && opipe) {
1130                 if (off_in || off_out)
1131                         return -ESPIPE;
1132
1133                 if (!(in->f_mode & FMODE_READ))
1134                         return -EBADF;
1135
1136                 if (!(out->f_mode & FMODE_WRITE))
1137                         return -EBADF;
1138
1139                 /* Splicing to self would be fun, but... */
1140                 if (ipipe == opipe)
1141                         return -EINVAL;
1142
1143                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1144         }
1145
1146         if (ipipe) {
1147                 if (off_in)
1148                         return -ESPIPE;
1149                 if (off_out) {
1150                         if (!(out->f_mode & FMODE_PWRITE))
1151                                 return -EINVAL;
1152                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1153                                 return -EFAULT;
1154                 } else {
1155                         offset = out->f_pos;
1156                 }
1157
1158                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1159                         return -EBADF;
1160
1161                 if (unlikely(out->f_flags & O_APPEND))
1162                         return -EINVAL;
1163
1164                 ret = rw_verify_area(WRITE, out, &offset, len);
1165                 if (unlikely(ret < 0))
1166                         return ret;
1167
1168                 file_start_write(out);
1169                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1170                 file_end_write(out);
1171
1172                 if (!off_out)
1173                         out->f_pos = offset;
1174                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1175                         ret = -EFAULT;
1176
1177                 return ret;
1178         }
1179
1180         if (opipe) {
1181                 if (off_out)
1182                         return -ESPIPE;
1183                 if (off_in) {
1184                         if (!(in->f_mode & FMODE_PREAD))
1185                                 return -EINVAL;
1186                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1187                                 return -EFAULT;
1188                 } else {
1189                         offset = in->f_pos;
1190                 }
1191
1192                 pipe_lock(opipe);
1193                 ret = wait_for_space(opipe, flags);
1194                 if (!ret)
1195                         ret = do_splice_to(in, &offset, opipe, len, flags);
1196                 pipe_unlock(opipe);
1197                 if (ret > 0)
1198                         wakeup_pipe_readers(opipe);
1199                 if (!off_in)
1200                         in->f_pos = offset;
1201                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1202                         ret = -EFAULT;
1203
1204                 return ret;
1205         }
1206
1207         return -EINVAL;
1208 }
1209
1210 static int iter_to_pipe(struct iov_iter *from,
1211                         struct pipe_inode_info *pipe,
1212                         unsigned flags)
1213 {
1214         struct pipe_buffer buf = {
1215                 .ops = &user_page_pipe_buf_ops,
1216                 .flags = flags
1217         };
1218         size_t total = 0;
1219         int ret = 0;
1220         bool failed = false;
1221
1222         while (iov_iter_count(from) && !failed) {
1223                 struct page *pages[16];
1224                 ssize_t copied;
1225                 size_t start;
1226                 int n;
1227
1228                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1229                 if (copied <= 0) {
1230                         ret = copied;
1231                         break;
1232                 }
1233
1234                 for (n = 0; copied; n++, start = 0) {
1235                         int size = min_t(int, copied, PAGE_SIZE - start);
1236                         if (!failed) {
1237                                 buf.page = pages[n];
1238                                 buf.offset = start;
1239                                 buf.len = size;
1240                                 ret = add_to_pipe(pipe, &buf);
1241                                 if (unlikely(ret < 0)) {
1242                                         failed = true;
1243                                 } else {
1244                                         iov_iter_advance(from, ret);
1245                                         total += ret;
1246                                 }
1247                         } else {
1248                                 put_page(pages[n]);
1249                         }
1250                         copied -= size;
1251                 }
1252         }
1253         return total ? total : ret;
1254 }
1255
1256 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1257                         struct splice_desc *sd)
1258 {
1259         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1260         return n == sd->len ? n : -EFAULT;
1261 }
1262
1263 /*
1264  * For lack of a better implementation, implement vmsplice() to userspace
1265  * as a simple copy of the pipes pages to the user iov.
1266  */
1267 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1268                              unsigned long nr_segs, unsigned int flags)
1269 {
1270         struct pipe_inode_info *pipe;
1271         struct splice_desc sd;
1272         long ret;
1273         struct iovec iovstack[UIO_FASTIOV];
1274         struct iovec *iov = iovstack;
1275         struct iov_iter iter;
1276
1277         pipe = get_pipe_info(file);
1278         if (!pipe)
1279                 return -EBADF;
1280
1281         ret = import_iovec(READ, uiov, nr_segs,
1282                            ARRAY_SIZE(iovstack), &iov, &iter);
1283         if (ret < 0)
1284                 return ret;
1285
1286         sd.total_len = iov_iter_count(&iter);
1287         sd.len = 0;
1288         sd.flags = flags;
1289         sd.u.data = &iter;
1290         sd.pos = 0;
1291
1292         if (sd.total_len) {
1293                 pipe_lock(pipe);
1294                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1295                 pipe_unlock(pipe);
1296         }
1297
1298         kfree(iov);
1299         return ret;
1300 }
1301
1302 /*
1303  * vmsplice splices a user address range into a pipe. It can be thought of
1304  * as splice-from-memory, where the regular splice is splice-from-file (or
1305  * to file). In both cases the output is a pipe, naturally.
1306  */
1307 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1308                              unsigned long nr_segs, unsigned int flags)
1309 {
1310         struct pipe_inode_info *pipe;
1311         struct iovec iovstack[UIO_FASTIOV];
1312         struct iovec *iov = iovstack;
1313         struct iov_iter from;
1314         long ret;
1315         unsigned buf_flag = 0;
1316
1317         if (flags & SPLICE_F_GIFT)
1318                 buf_flag = PIPE_BUF_FLAG_GIFT;
1319
1320         pipe = get_pipe_info(file);
1321         if (!pipe)
1322                 return -EBADF;
1323
1324         ret = import_iovec(WRITE, uiov, nr_segs,
1325                            ARRAY_SIZE(iovstack), &iov, &from);
1326         if (ret < 0)
1327                 return ret;
1328
1329         pipe_lock(pipe);
1330         ret = wait_for_space(pipe, flags);
1331         if (!ret)
1332                 ret = iter_to_pipe(&from, pipe, buf_flag);
1333         pipe_unlock(pipe);
1334         if (ret > 0)
1335                 wakeup_pipe_readers(pipe);
1336         kfree(iov);
1337         return ret;
1338 }
1339
1340 /*
1341  * Note that vmsplice only really supports true splicing _from_ user memory
1342  * to a pipe, not the other way around. Splicing from user memory is a simple
1343  * operation that can be supported without any funky alignment restrictions
1344  * or nasty vm tricks. We simply map in the user memory and fill them into
1345  * a pipe. The reverse isn't quite as easy, though. There are two possible
1346  * solutions for that:
1347  *
1348  *      - memcpy() the data internally, at which point we might as well just
1349  *        do a regular read() on the buffer anyway.
1350  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1351  *        has restriction limitations on both ends of the pipe).
1352  *
1353  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1354  *
1355  */
1356 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1357                 unsigned long, nr_segs, unsigned int, flags)
1358 {
1359         struct fd f;
1360         long error;
1361
1362         if (unlikely(nr_segs > UIO_MAXIOV))
1363                 return -EINVAL;
1364         else if (unlikely(!nr_segs))
1365                 return 0;
1366
1367         error = -EBADF;
1368         f = fdget(fd);
1369         if (f.file) {
1370                 if (f.file->f_mode & FMODE_WRITE)
1371                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1372                 else if (f.file->f_mode & FMODE_READ)
1373                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1374
1375                 fdput(f);
1376         }
1377
1378         return error;
1379 }
1380
1381 #ifdef CONFIG_COMPAT
1382 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1383                     unsigned int, nr_segs, unsigned int, flags)
1384 {
1385         unsigned i;
1386         struct iovec __user *iov;
1387         if (nr_segs > UIO_MAXIOV)
1388                 return -EINVAL;
1389         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1390         for (i = 0; i < nr_segs; i++) {
1391                 struct compat_iovec v;
1392                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1393                     get_user(v.iov_len, &iov32[i].iov_len) ||
1394                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1395                     put_user(v.iov_len, &iov[i].iov_len))
1396                         return -EFAULT;
1397         }
1398         return sys_vmsplice(fd, iov, nr_segs, flags);
1399 }
1400 #endif
1401
1402 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1403                 int, fd_out, loff_t __user *, off_out,
1404                 size_t, len, unsigned int, flags)
1405 {
1406         struct fd in, out;
1407         long error;
1408
1409         if (unlikely(!len))
1410                 return 0;
1411
1412         error = -EBADF;
1413         in = fdget(fd_in);
1414         if (in.file) {
1415                 if (in.file->f_mode & FMODE_READ) {
1416                         out = fdget(fd_out);
1417                         if (out.file) {
1418                                 if (out.file->f_mode & FMODE_WRITE)
1419                                         error = do_splice(in.file, off_in,
1420                                                           out.file, off_out,
1421                                                           len, flags);
1422                                 fdput(out);
1423                         }
1424                 }
1425                 fdput(in);
1426         }
1427         return error;
1428 }
1429
1430 /*
1431  * Make sure there's data to read. Wait for input if we can, otherwise
1432  * return an appropriate error.
1433  */
1434 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1435 {
1436         int ret;
1437
1438         /*
1439          * Check ->nrbufs without the inode lock first. This function
1440          * is speculative anyways, so missing one is ok.
1441          */
1442         if (pipe->nrbufs)
1443                 return 0;
1444
1445         ret = 0;
1446         pipe_lock(pipe);
1447
1448         while (!pipe->nrbufs) {
1449                 if (signal_pending(current)) {
1450                         ret = -ERESTARTSYS;
1451                         break;
1452                 }
1453                 if (!pipe->writers)
1454                         break;
1455                 if (!pipe->waiting_writers) {
1456                         if (flags & SPLICE_F_NONBLOCK) {
1457                                 ret = -EAGAIN;
1458                                 break;
1459                         }
1460                 }
1461                 pipe_wait(pipe);
1462         }
1463
1464         pipe_unlock(pipe);
1465         return ret;
1466 }
1467
1468 /*
1469  * Make sure there's writeable room. Wait for room if we can, otherwise
1470  * return an appropriate error.
1471  */
1472 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1473 {
1474         int ret;
1475
1476         /*
1477          * Check ->nrbufs without the inode lock first. This function
1478          * is speculative anyways, so missing one is ok.
1479          */
1480         if (pipe->nrbufs < pipe->buffers)
1481                 return 0;
1482
1483         ret = 0;
1484         pipe_lock(pipe);
1485
1486         while (pipe->nrbufs >= pipe->buffers) {
1487                 if (!pipe->readers) {
1488                         send_sig(SIGPIPE, current, 0);
1489                         ret = -EPIPE;
1490                         break;
1491                 }
1492                 if (flags & SPLICE_F_NONBLOCK) {
1493                         ret = -EAGAIN;
1494                         break;
1495                 }
1496                 if (signal_pending(current)) {
1497                         ret = -ERESTARTSYS;
1498                         break;
1499                 }
1500                 pipe->waiting_writers++;
1501                 pipe_wait(pipe);
1502                 pipe->waiting_writers--;
1503         }
1504
1505         pipe_unlock(pipe);
1506         return ret;
1507 }
1508
1509 /*
1510  * Splice contents of ipipe to opipe.
1511  */
1512 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1513                                struct pipe_inode_info *opipe,
1514                                size_t len, unsigned int flags)
1515 {
1516         struct pipe_buffer *ibuf, *obuf;
1517         int ret = 0, nbuf;
1518         bool input_wakeup = false;
1519
1520
1521 retry:
1522         ret = ipipe_prep(ipipe, flags);
1523         if (ret)
1524                 return ret;
1525
1526         ret = opipe_prep(opipe, flags);
1527         if (ret)
1528                 return ret;
1529
1530         /*
1531          * Potential ABBA deadlock, work around it by ordering lock
1532          * grabbing by pipe info address. Otherwise two different processes
1533          * could deadlock (one doing tee from A -> B, the other from B -> A).
1534          */
1535         pipe_double_lock(ipipe, opipe);
1536
1537         do {
1538                 if (!opipe->readers) {
1539                         send_sig(SIGPIPE, current, 0);
1540                         if (!ret)
1541                                 ret = -EPIPE;
1542                         break;
1543                 }
1544
1545                 if (!ipipe->nrbufs && !ipipe->writers)
1546                         break;
1547
1548                 /*
1549                  * Cannot make any progress, because either the input
1550                  * pipe is empty or the output pipe is full.
1551                  */
1552                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1553                         /* Already processed some buffers, break */
1554                         if (ret)
1555                                 break;
1556
1557                         if (flags & SPLICE_F_NONBLOCK) {
1558                                 ret = -EAGAIN;
1559                                 break;
1560                         }
1561
1562                         /*
1563                          * We raced with another reader/writer and haven't
1564                          * managed to process any buffers.  A zero return
1565                          * value means EOF, so retry instead.
1566                          */
1567                         pipe_unlock(ipipe);
1568                         pipe_unlock(opipe);
1569                         goto retry;
1570                 }
1571
1572                 ibuf = ipipe->bufs + ipipe->curbuf;
1573                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1574                 obuf = opipe->bufs + nbuf;
1575
1576                 if (len >= ibuf->len) {
1577                         /*
1578                          * Simply move the whole buffer from ipipe to opipe
1579                          */
1580                         *obuf = *ibuf;
1581                         ibuf->ops = NULL;
1582                         opipe->nrbufs++;
1583                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1584                         ipipe->nrbufs--;
1585                         input_wakeup = true;
1586                 } else {
1587                         /*
1588                          * Get a reference to this pipe buffer,
1589                          * so we can copy the contents over.
1590                          */
1591                         pipe_buf_get(ipipe, ibuf);
1592                         *obuf = *ibuf;
1593
1594                         /*
1595                          * Don't inherit the gift flag, we need to
1596                          * prevent multiple steals of this page.
1597                          */
1598                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1599
1600                         obuf->len = len;
1601                         opipe->nrbufs++;
1602                         ibuf->offset += obuf->len;
1603                         ibuf->len -= obuf->len;
1604                 }
1605                 ret += obuf->len;
1606                 len -= obuf->len;
1607         } while (len);
1608
1609         pipe_unlock(ipipe);
1610         pipe_unlock(opipe);
1611
1612         /*
1613          * If we put data in the output pipe, wakeup any potential readers.
1614          */
1615         if (ret > 0)
1616                 wakeup_pipe_readers(opipe);
1617
1618         if (input_wakeup)
1619                 wakeup_pipe_writers(ipipe);
1620
1621         return ret;
1622 }
1623
1624 /*
1625  * Link contents of ipipe to opipe.
1626  */
1627 static int link_pipe(struct pipe_inode_info *ipipe,
1628                      struct pipe_inode_info *opipe,
1629                      size_t len, unsigned int flags)
1630 {
1631         struct pipe_buffer *ibuf, *obuf;
1632         int ret = 0, i = 0, nbuf;
1633
1634         /*
1635          * Potential ABBA deadlock, work around it by ordering lock
1636          * grabbing by pipe info address. Otherwise two different processes
1637          * could deadlock (one doing tee from A -> B, the other from B -> A).
1638          */
1639         pipe_double_lock(ipipe, opipe);
1640
1641         do {
1642                 if (!opipe->readers) {
1643                         send_sig(SIGPIPE, current, 0);
1644                         if (!ret)
1645                                 ret = -EPIPE;
1646                         break;
1647                 }
1648
1649                 /*
1650                  * If we have iterated all input buffers or ran out of
1651                  * output room, break.
1652                  */
1653                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1654                         break;
1655
1656                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1657                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1658
1659                 /*
1660                  * Get a reference to this pipe buffer,
1661                  * so we can copy the contents over.
1662                  */
1663                 pipe_buf_get(ipipe, ibuf);
1664
1665                 obuf = opipe->bufs + nbuf;
1666                 *obuf = *ibuf;
1667
1668                 /*
1669                  * Don't inherit the gift flag, we need to
1670                  * prevent multiple steals of this page.
1671                  */
1672                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1673
1674                 if (obuf->len > len)
1675                         obuf->len = len;
1676
1677                 opipe->nrbufs++;
1678                 ret += obuf->len;
1679                 len -= obuf->len;
1680                 i++;
1681         } while (len);
1682
1683         /*
1684          * return EAGAIN if we have the potential of some data in the
1685          * future, otherwise just return 0
1686          */
1687         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1688                 ret = -EAGAIN;
1689
1690         pipe_unlock(ipipe);
1691         pipe_unlock(opipe);
1692
1693         /*
1694          * If we put data in the output pipe, wakeup any potential readers.
1695          */
1696         if (ret > 0)
1697                 wakeup_pipe_readers(opipe);
1698
1699         return ret;
1700 }
1701
1702 /*
1703  * This is a tee(1) implementation that works on pipes. It doesn't copy
1704  * any data, it simply references the 'in' pages on the 'out' pipe.
1705  * The 'flags' used are the SPLICE_F_* variants, currently the only
1706  * applicable one is SPLICE_F_NONBLOCK.
1707  */
1708 static long do_tee(struct file *in, struct file *out, size_t len,
1709                    unsigned int flags)
1710 {
1711         struct pipe_inode_info *ipipe = get_pipe_info(in);
1712         struct pipe_inode_info *opipe = get_pipe_info(out);
1713         int ret = -EINVAL;
1714
1715         /*
1716          * Duplicate the contents of ipipe to opipe without actually
1717          * copying the data.
1718          */
1719         if (ipipe && opipe && ipipe != opipe) {
1720                 /*
1721                  * Keep going, unless we encounter an error. The ipipe/opipe
1722                  * ordering doesn't really matter.
1723                  */
1724                 ret = ipipe_prep(ipipe, flags);
1725                 if (!ret) {
1726                         ret = opipe_prep(opipe, flags);
1727                         if (!ret)
1728                                 ret = link_pipe(ipipe, opipe, len, flags);
1729                 }
1730         }
1731
1732         return ret;
1733 }
1734
1735 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1736 {
1737         struct fd in;
1738         int error;
1739
1740         if (unlikely(!len))
1741                 return 0;
1742
1743         error = -EBADF;
1744         in = fdget(fdin);
1745         if (in.file) {
1746                 if (in.file->f_mode & FMODE_READ) {
1747                         struct fd out = fdget(fdout);
1748                         if (out.file) {
1749                                 if (out.file->f_mode & FMODE_WRITE)
1750                                         error = do_tee(in.file, out.file,
1751                                                         len, flags);
1752                                 fdput(out);
1753                         }
1754                 }
1755                 fdput(in);
1756         }
1757
1758         return error;
1759 }