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 /*
 251  * Check if we need to grow the arrays holding pages and partial page
 252  * descriptions.
 253  */
 254 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 255 {
 256         unsigned int buffers = READ_ONCE(pipe->buffers);
 257
 258         spd->nr_pages_max = buffers;
 259         if (buffers <= PIPE_DEF_BUFFERS)
 260                 return 0;
 261
 262         spd->pages = kmalloc_array(buffers, sizeof(struct page *), GFP_KERNEL);
 263         spd->partial = kmalloc_array(buffers, sizeof(struct partial_page),
 264                                      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 = call_read_iter(in, &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 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 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 369                                  struct pipe_inode_info *pipe, size_t len,
 370                                  unsigned int flags)
 371 {
 372         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 373         struct iov_iter to;
 374         struct page **pages;
 375         unsigned int nr_pages;
 376         size_t offset, base, copied = 0;
 377         ssize_t res;
 378         int i;
 379
 380         if (pipe->nrbufs == pipe->buffers)
 381                 return -EAGAIN;
 382
 383         /*
 384          * Try to keep page boundaries matching to source pagecache ones -
 385          * it probably won't be much help, but...
 386          */
 387         offset = *ppos & ~PAGE_MASK;
 388
 389         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 390
 391         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 392         if (res <= 0)
 393                 return -ENOMEM;
 394
 395         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 396
 397         vec = __vec;
 398         if (nr_pages > PIPE_DEF_BUFFERS) {
 399                 vec = kmalloc_array(nr_pages, sizeof(struct kvec), GFP_KERNEL);
 400                 if (unlikely(!vec)) {
 401                         res = -ENOMEM;
 402                         goto out;
 403                 }
 404         }
 405
 406         pipe->bufs[to.idx].offset = offset;
 407         pipe->bufs[to.idx].len -= offset;
 408
 409         for (i = 0; i < nr_pages; i++) {
 410                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 411                 vec[i].iov_base = page_address(pages[i]) + offset;
 412                 vec[i].iov_len = this_len;
 413                 len -= this_len;
 414                 offset = 0;
 415         }
 416
 417         res = kernel_readv(in, vec, nr_pages, *ppos);
 418         if (res > 0) {
 419                 copied = res;
 420                 *ppos += res;
 421         }
 422
 423         if (vec != __vec)
 424                 kfree(vec);
 425 out:
 426         for (i = 0; i < nr_pages; i++)
 427                 put_page(pages[i]);
 428         kvfree(pages);
 429         iov_iter_advance(&to, copied);  /* truncates and discards */
 430         return res;
 431 }
 432
 433 /*
 434  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 435  * using sendpage(). Return the number of bytes sent.
 436  */
 437 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 438                             struct pipe_buffer *buf, struct splice_desc *sd)
 439 {
 440         struct file *file = sd->u.file;
 441         loff_t pos = sd->pos;
 442         int more;
 443
 444         if (!likely(file->f_op->sendpage))
 445                 return -EINVAL;
 446
 447         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 448
 449         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 450                 more |= MSG_SENDPAGE_NOTLAST;
 451
 452         return file->f_op->sendpage(file, buf->page, buf->offset,
 453                                     sd->len, &pos, more);
 454 }
 455
 456 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 457 {
 458         smp_mb();
 459         if (waitqueue_active(&pipe->wait))
 460                 wake_up_interruptible(&pipe->wait);
 461         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 462 }
 463
 464 /**
 465  * splice_from_pipe_feed - feed available data from a pipe to a file
 466  * @pipe:       pipe to splice from
 467  * @sd:         information to @actor
 468  * @actor:      handler that splices the data
 469  *
 470  * Description:
 471  *    This function loops over the pipe and calls @actor to do the
 472  *    actual moving of a single struct pipe_buffer to the desired
 473  *    destination.  It returns when there's no more buffers left in
 474  *    the pipe or if the requested number of bytes (@sd->total_len)
 475  *    have been copied.  It returns a positive number (one) if the
 476  *    pipe needs to be filled with more data, zero if the required
 477  *    number of bytes have been copied and -errno on error.
 478  *
 479  *    This, together with splice_from_pipe_{begin,end,next}, may be
 480  *    used to implement the functionality of __splice_from_pipe() when
 481  *    locking is required around copying the pipe buffers to the
 482  *    destination.
 483  */
 484 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 485                           splice_actor *actor)
 486 {
 487         int ret;
 488
 489         while (pipe->nrbufs) {
 490                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 491
 492                 sd->len = buf->len;
 493                 if (sd->len > sd->total_len)
 494                         sd->len = sd->total_len;
 495
 496                 ret = pipe_buf_confirm(pipe, buf);
 497                 if (unlikely(ret)) {
 498                         if (ret == -ENODATA)
 499                                 ret = 0;
 500                         return ret;
 501                 }
 502
 503                 ret = actor(pipe, buf, sd);
 504                 if (ret <= 0)
 505                         return ret;
 506
 507                 buf->offset += ret;
 508                 buf->len -= ret;
 509
 510                 sd->num_spliced += ret;
 511                 sd->len -= ret;
 512                 sd->pos += ret;
 513                 sd->total_len -= ret;
 514
 515                 if (!buf->len) {
 516                         pipe_buf_release(pipe, buf);
 517                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 518                         pipe->nrbufs--;
 519                         if (pipe->files)
 520                                 sd->need_wakeup = true;
 521                 }
 522
 523                 if (!sd->total_len)
 524                         return 0;
 525         }
 526
 527         return 1;
 528 }
 529
 530 /**
 531  * splice_from_pipe_next - wait for some data to splice from
 532  * @pipe:       pipe to splice from
 533  * @sd:         information about the splice operation
 534  *
 535  * Description:
 536  *    This function will wait for some data and return a positive
 537  *    value (one) if pipe buffers are available.  It will return zero
 538  *    or -errno if no more data needs to be spliced.
 539  */
 540 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 541 {
 542         /*
 543          * Check for signal early to make process killable when there are
 544          * always buffers available
 545          */
 546         if (signal_pending(current))
 547                 return -ERESTARTSYS;
 548
 549         while (!pipe->nrbufs) {
 550                 if (!pipe->writers)
 551                         return 0;
 552
 553                 if (!pipe->waiting_writers && sd->num_spliced)
 554                         return 0;
 555
 556                 if (sd->flags & SPLICE_F_NONBLOCK)
 557                         return -EAGAIN;
 558
 559                 if (signal_pending(current))
 560                         return -ERESTARTSYS;
 561
 562                 if (sd->need_wakeup) {
 563                         wakeup_pipe_writers(pipe);
 564                         sd->need_wakeup = false;
 565                 }
 566
 567                 pipe_wait(pipe);
 568         }
 569
 570         return 1;
 571 }
 572
 573 /**
 574  * splice_from_pipe_begin - start splicing from pipe
 575  * @sd:         information about the splice operation
 576  *
 577  * Description:
 578  *    This function should be called before a loop containing
 579  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 580  *    initialize the necessary fields of @sd.
 581  */
 582 static void splice_from_pipe_begin(struct splice_desc *sd)
 583 {
 584         sd->num_spliced = 0;
 585         sd->need_wakeup = false;
 586 }
 587
 588 /**
 589  * splice_from_pipe_end - finish splicing from pipe
 590  * @pipe:       pipe to splice from
 591  * @sd:         information about the splice operation
 592  *
 593  * Description:
 594  *    This function will wake up pipe writers if necessary.  It should
 595  *    be called after a loop containing splice_from_pipe_next() and
 596  *    splice_from_pipe_feed().
 597  */
 598 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 599 {
 600         if (sd->need_wakeup)
 601                 wakeup_pipe_writers(pipe);
 602 }
 603
 604 /**
 605  * __splice_from_pipe - splice data from a pipe to given actor
 606  * @pipe:       pipe to splice from
 607  * @sd:         information to @actor
 608  * @actor:      handler that splices the data
 609  *
 610  * Description:
 611  *    This function does little more than loop over the pipe and call
 612  *    @actor to do the actual moving of a single struct pipe_buffer to
 613  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 614  *    pipe_to_user.
 615  *
 616  */
 617 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 618                            splice_actor *actor)
 619 {
 620         int ret;
 621
 622         splice_from_pipe_begin(sd);
 623         do {
 624                 cond_resched();
 625                 ret = splice_from_pipe_next(pipe, sd);
 626                 if (ret > 0)
 627                         ret = splice_from_pipe_feed(pipe, sd, actor);
 628         } while (ret > 0);
 629         splice_from_pipe_end(pipe, sd);
 630
 631         return sd->num_spliced ? sd->num_spliced : ret;
 632 }
 633 EXPORT_SYMBOL(__splice_from_pipe);
 634
 635 /**
 636  * splice_from_pipe - splice data from a pipe to a file
 637  * @pipe:       pipe to splice from
 638  * @out:        file to splice to
 639  * @ppos:       position in @out
 640  * @len:        how many bytes to splice
 641  * @flags:      splice modifier flags
 642  * @actor:      handler that splices the data
 643  *
 644  * Description:
 645  *    See __splice_from_pipe. This function locks the pipe inode,
 646  *    otherwise it's identical to __splice_from_pipe().
 647  *
 648  */
 649 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 650                          loff_t *ppos, size_t len, unsigned int flags,
 651                          splice_actor *actor)
 652 {
 653         ssize_t ret;
 654         struct splice_desc sd = {
 655                 .total_len = len,
 656                 .flags = flags,
 657                 .pos = *ppos,
 658                 .u.file = out,
 659         };
 660
 661         pipe_lock(pipe);
 662         ret = __splice_from_pipe(pipe, &sd, actor);
 663         pipe_unlock(pipe);
 664
 665         return ret;
 666 }
 667
 668 /**
 669  * iter_file_splice_write - splice data from a pipe to a file
 670  * @pipe:       pipe info
 671  * @out:        file to write to
 672  * @ppos:       position in @out
 673  * @len:        number of bytes to splice
 674  * @flags:      splice modifier flags
 675  *
 676  * Description:
 677  *    Will either move or copy pages (determined by @flags options) from
 678  *    the given pipe inode to the given file.
 679  *    This one is ->write_iter-based.
 680  *
 681  */
 682 ssize_t
 683 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 684                           loff_t *ppos, size_t len, unsigned int flags)
 685 {
 686         struct splice_desc sd = {
 687                 .total_len = len,
 688                 .flags = flags,
 689                 .pos = *ppos,
 690                 .u.file = out,
 691         };
 692         int nbufs = pipe->buffers;
 693         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 694                                         GFP_KERNEL);
 695         ssize_t ret;
 696
 697         if (unlikely(!array))
 698                 return -ENOMEM;
 699
 700         pipe_lock(pipe);
 701
 702         splice_from_pipe_begin(&sd);
 703         while (sd.total_len) {
 704                 struct iov_iter from;
 705                 size_t left;
 706                 int n, idx;
 707
 708                 ret = splice_from_pipe_next(pipe, &sd);
 709                 if (ret <= 0)
 710                         break;
 711
 712                 if (unlikely(nbufs < pipe->buffers)) {
 713                         kfree(array);
 714                         nbufs = pipe->buffers;
 715                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 716                                         GFP_KERNEL);
 717                         if (!array) {
 718                                 ret = -ENOMEM;
 719                                 break;
 720                         }
 721                 }
 722
 723                 /* build the vector */
 724                 left = sd.total_len;
 725                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 726                         struct pipe_buffer *buf = pipe->bufs + idx;
 727                         size_t this_len = buf->len;
 728
 729                         if (this_len > left)
 730                                 this_len = left;
 731
 732                         if (idx == pipe->buffers - 1)
 733                                 idx = -1;
 734
 735                         ret = pipe_buf_confirm(pipe, buf);
 736                         if (unlikely(ret)) {
 737                                 if (ret == -ENODATA)
 738                                         ret = 0;
 739                                 goto done;
 740                         }
 741
 742                         array[n].bv_page = buf->page;
 743                         array[n].bv_len = this_len;
 744                         array[n].bv_offset = buf->offset;
 745                         left -= this_len;
 746                 }
 747
 748                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 749                               sd.total_len - left);
 750                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 751                 if (ret <= 0)
 752                         break;
 753
 754                 sd.num_spliced += ret;
 755                 sd.total_len -= ret;
 756                 *ppos = sd.pos;
 757
 758                 /* dismiss the fully eaten buffers, adjust the partial one */
 759                 while (ret) {
 760                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 761                         if (ret >= buf->len) {
 762                                 ret -= buf->len;
 763                                 buf->len = 0;
 764                                 pipe_buf_release(pipe, buf);
 765                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 766                                 pipe->nrbufs--;
 767                                 if (pipe->files)
 768                                         sd.need_wakeup = true;
 769                         } else {
 770                                 buf->offset += ret;
 771                                 buf->len -= ret;
 772                                 ret = 0;
 773                         }
 774                 }
 775         }
 776 done:
 777         kfree(array);
 778         splice_from_pipe_end(pipe, &sd);
 779
 780         pipe_unlock(pipe);
 781
 782         if (sd.num_spliced)
 783                 ret = sd.num_spliced;
 784
 785         return ret;
 786 }
 787
 788 EXPORT_SYMBOL(iter_file_splice_write);
 789
 790 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 791                           struct splice_desc *sd)
 792 {
 793         int ret;
 794         void *data;
 795         loff_t tmp = sd->pos;
 796
 797         data = kmap(buf->page);
 798         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 799         kunmap(buf->page);
 800
 801         return ret;
 802 }
 803
 804 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 805                                          struct file *out, loff_t *ppos,
 806                                          size_t len, unsigned int flags)
 807 {
 808         ssize_t ret;
 809
 810         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 811         if (ret > 0)
 812                 *ppos += ret;
 813
 814         return ret;
 815 }
 816
 817 /**
 818  * generic_splice_sendpage - splice data from a pipe to a socket
 819  * @pipe:       pipe to splice from
 820  * @out:        socket to write to
 821  * @ppos:       position in @out
 822  * @len:        number of bytes to splice
 823  * @flags:      splice modifier flags
 824  *
 825  * Description:
 826  *    Will send @len bytes from the pipe to a network socket. No data copying
 827  *    is involved.
 828  *
 829  */
 830 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 831                                 loff_t *ppos, size_t len, unsigned int flags)
 832 {
 833         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 834 }
 835
 836 EXPORT_SYMBOL(generic_splice_sendpage);
 837
 838 /*
 839  * Attempt to initiate a splice from pipe to file.
 840  */
 841 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 842                            loff_t *ppos, size_t len, unsigned int flags)
 843 {
 844         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 845                                 loff_t *, size_t, unsigned int);
 846
 847         if (out->f_op->splice_write)
 848                 splice_write = out->f_op->splice_write;
 849         else
 850                 splice_write = default_file_splice_write;
 851
 852         return splice_write(pipe, out, ppos, len, flags);
 853 }
 854
 855 /*
 856  * Attempt to initiate a splice from a file to a pipe.
 857  */
 858 static long do_splice_to(struct file *in, loff_t *ppos,
 859                          struct pipe_inode_info *pipe, size_t len,
 860                          unsigned int flags)
 861 {
 862         ssize_t (*splice_read)(struct file *, loff_t *,
 863                                struct pipe_inode_info *, size_t, unsigned int);
 864         int ret;
 865
 866         if (unlikely(!(in->f_mode & FMODE_READ)))
 867                 return -EBADF;
 868
 869         ret = rw_verify_area(READ, in, ppos, len);
 870         if (unlikely(ret < 0))
 871                 return ret;
 872
 873         if (unlikely(len > MAX_RW_COUNT))
 874                 len = MAX_RW_COUNT;
 875
 876         if (in->f_op->splice_read)
 877                 splice_read = in->f_op->splice_read;
 878         else
 879                 splice_read = default_file_splice_read;
 880
 881         return splice_read(in, ppos, pipe, len, flags);
 882 }
 883
 884 /**
 885  * splice_direct_to_actor - splices data directly between two non-pipes
 886  * @in:         file to splice from
 887  * @sd:         actor information on where to splice to
 888  * @actor:      handles the data splicing
 889  *
 890  * Description:
 891  *    This is a special case helper to splice directly between two
 892  *    points, without requiring an explicit pipe. Internally an allocated
 893  *    pipe is cached in the process, and reused during the lifetime of
 894  *    that process.
 895  *
 896  */
 897 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 898                                splice_direct_actor *actor)
 899 {
 900         struct pipe_inode_info *pipe;
 901         long ret, bytes;
 902         umode_t i_mode;
 903         size_t len;
 904         int i, flags, more;
 905
 906         /*
 907          * We require the input being a regular file, as we don't want to
 908          * randomly drop data for eg socket -> socket splicing. Use the
 909          * piped splicing for that!
 910          */
 911         i_mode = file_inode(in)->i_mode;
 912         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 913                 return -EINVAL;
 914
 915         /*
 916          * neither in nor out is a pipe, setup an internal pipe attached to
 917          * 'out' and transfer the wanted data from 'in' to 'out' through that
 918          */
 919         pipe = current->splice_pipe;
 920         if (unlikely(!pipe)) {
 921                 pipe = alloc_pipe_info();
 922                 if (!pipe)
 923                         return -ENOMEM;
 924
 925                 /*
 926                  * We don't have an immediate reader, but we'll read the stuff
 927                  * out of the pipe right after the splice_to_pipe(). So set
 928                  * PIPE_READERS appropriately.
 929                  */
 930                 pipe->readers = 1;
 931
 932                 current->splice_pipe = pipe;
 933         }
 934
 935         /*
 936          * Do the splice.
 937          */
 938         ret = 0;
 939         bytes = 0;
 940         len = sd->total_len;
 941         flags = sd->flags;
 942
 943         /*
 944          * Don't block on output, we have to drain the direct pipe.
 945          */
 946         sd->flags &= ~SPLICE_F_NONBLOCK;
 947         more = sd->flags & SPLICE_F_MORE;
 948
 949         while (len) {
 950                 size_t read_len;
 951                 loff_t pos = sd->pos, prev_pos = pos;
 952
 953                 ret = do_splice_to(in, &pos, pipe, len, flags);
 954                 if (unlikely(ret <= 0))
 955                         goto out_release;
 956
 957                 read_len = ret;
 958                 sd->total_len = read_len;
 959
 960                 /*
 961                  * If more data is pending, set SPLICE_F_MORE
 962                  * If this is the last data and SPLICE_F_MORE was not set
 963                  * initially, clears it.
 964                  */
 965                 if (read_len < len)
 966                         sd->flags |= SPLICE_F_MORE;
 967                 else if (!more)
 968                         sd->flags &= ~SPLICE_F_MORE;
 969                 /*
 970                  * NOTE: nonblocking mode only applies to the input. We
 971                  * must not do the output in nonblocking mode as then we
 972                  * could get stuck data in the internal pipe:
 973                  */
 974                 ret = actor(pipe, sd);
 975                 if (unlikely(ret <= 0)) {
 976                         sd->pos = prev_pos;
 977                         goto out_release;
 978                 }
 979
 980                 bytes += ret;
 981                 len -= ret;
 982                 sd->pos = pos;
 983
 984                 if (ret < read_len) {
 985                         sd->pos = prev_pos + ret;
 986                         goto out_release;
 987                 }
 988         }
 989
 990 done:
 991         pipe->nrbufs = pipe->curbuf = 0;
 992         file_accessed(in);
 993         return bytes;
 994
 995 out_release:
 996         /*
 997          * If we did an incomplete transfer we must release
 998          * the pipe buffers in question:
 999          */
1000         for (i = 0; i < pipe->buffers; i++) {
1001                 struct pipe_buffer *buf = pipe->bufs + i;
1002
1003                 if (buf->ops)
1004                         pipe_buf_release(pipe, buf);
1005         }
1006
1007         if (!bytes)
1008                 bytes = ret;
1009
1010         goto done;
1011 }
1012 EXPORT_SYMBOL(splice_direct_to_actor);
1013
1014 static int direct_splice_actor(struct pipe_inode_info *pipe,
1015                                struct splice_desc *sd)
1016 {
1017         struct file *file = sd->u.file;
1018
1019         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1020                               sd->flags);
1021 }
1022
1023 /**
1024  * do_splice_direct - splices data directly between two files
1025  * @in:         file to splice from
1026  * @ppos:       input file offset
1027  * @out:        file to splice to
1028  * @opos:       output file offset
1029  * @len:        number of bytes to splice
1030  * @flags:      splice modifier flags
1031  *
1032  * Description:
1033  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1034  *    doing it in the application would incur an extra system call
1035  *    (splice in + splice out, as compared to just sendfile()). So this helper
1036  *    can splice directly through a process-private pipe.
1037  *
1038  */
1039 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1040                       loff_t *opos, size_t len, unsigned int flags)
1041 {
1042         struct splice_desc sd = {
1043                 .len            = len,
1044                 .total_len      = len,
1045                 .flags          = flags,
1046                 .pos            = *ppos,
1047                 .u.file         = out,
1048                 .opos           = opos,
1049         };
1050         long ret;
1051
1052         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1053                 return -EBADF;
1054
1055         if (unlikely(out->f_flags & O_APPEND))
1056                 return -EINVAL;
1057
1058         ret = rw_verify_area(WRITE, out, opos, len);
1059         if (unlikely(ret < 0))
1060                 return ret;
1061
1062         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1063         if (ret > 0)
1064                 *ppos = sd.pos;
1065
1066         return ret;
1067 }
1068 EXPORT_SYMBOL(do_splice_direct);
1069
1070 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1071 {
1072         for (;;) {
1073                 if (unlikely(!pipe->readers)) {
1074                         send_sig(SIGPIPE, current, 0);
1075                         return -EPIPE;
1076                 }
1077                 if (pipe->nrbufs != pipe->buffers)
1078                         return 0;
1079                 if (flags & SPLICE_F_NONBLOCK)
1080                         return -EAGAIN;
1081                 if (signal_pending(current))
1082                         return -ERESTARTSYS;
1083                 pipe->waiting_writers++;
1084                 pipe_wait(pipe);
1085                 pipe->waiting_writers--;
1086         }
1087 }
1088
1089 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1090                                struct pipe_inode_info *opipe,
1091                                size_t len, unsigned int flags);
1092
1093 /*
1094  * Determine where to splice to/from.
1095  */
1096 static long do_splice(struct file *in, loff_t __user *off_in,
1097                       struct file *out, loff_t __user *off_out,
1098                       size_t len, unsigned int flags)
1099 {
1100         struct pipe_inode_info *ipipe;
1101         struct pipe_inode_info *opipe;
1102         loff_t offset;
1103         long ret;
1104
1105         ipipe = get_pipe_info(in);
1106         opipe = get_pipe_info(out);
1107
1108         if (ipipe && opipe) {
1109                 if (off_in || off_out)
1110                         return -ESPIPE;
1111
1112                 if (!(in->f_mode & FMODE_READ))
1113                         return -EBADF;
1114
1115                 if (!(out->f_mode & FMODE_WRITE))
1116                         return -EBADF;
1117
1118                 /* Splicing to self would be fun, but... */
1119                 if (ipipe == opipe)
1120                         return -EINVAL;
1121
1122                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1123         }
1124
1125         if (ipipe) {
1126                 if (off_in)
1127                         return -ESPIPE;
1128                 if (off_out) {
1129                         if (!(out->f_mode & FMODE_PWRITE))
1130                                 return -EINVAL;
1131                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1132                                 return -EFAULT;
1133                 } else {
1134                         offset = out->f_pos;
1135                 }
1136
1137                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1138                         return -EBADF;
1139
1140                 if (unlikely(out->f_flags & O_APPEND))
1141                         return -EINVAL;
1142
1143                 ret = rw_verify_area(WRITE, out, &offset, len);
1144                 if (unlikely(ret < 0))
1145                         return ret;
1146
1147                 file_start_write(out);
1148                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1149                 file_end_write(out);
1150
1151                 if (!off_out)
1152                         out->f_pos = offset;
1153                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1154                         ret = -EFAULT;
1155
1156                 return ret;
1157         }
1158
1159         if (opipe) {
1160                 if (off_out)
1161                         return -ESPIPE;
1162                 if (off_in) {
1163                         if (!(in->f_mode & FMODE_PREAD))
1164                                 return -EINVAL;
1165                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1166                                 return -EFAULT;
1167                 } else {
1168                         offset = in->f_pos;
1169                 }
1170
1171                 pipe_lock(opipe);
1172                 ret = wait_for_space(opipe, flags);
1173                 if (!ret)
1174                         ret = do_splice_to(in, &offset, opipe, len, flags);
1175                 pipe_unlock(opipe);
1176                 if (ret > 0)
1177                         wakeup_pipe_readers(opipe);
1178                 if (!off_in)
1179                         in->f_pos = offset;
1180                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1181                         ret = -EFAULT;
1182
1183                 return ret;
1184         }
1185
1186         return -EINVAL;
1187 }
1188
1189 static int iter_to_pipe(struct iov_iter *from,
1190                         struct pipe_inode_info *pipe,
1191                         unsigned flags)
1192 {
1193         struct pipe_buffer buf = {
1194                 .ops = &user_page_pipe_buf_ops,
1195                 .flags = flags
1196         };
1197         size_t total = 0;
1198         int ret = 0;
1199         bool failed = false;
1200
1201         while (iov_iter_count(from) && !failed) {
1202                 struct page *pages[16];
1203                 ssize_t copied;
1204                 size_t start;
1205                 int n;
1206
1207                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1208                 if (copied <= 0) {
1209                         ret = copied;
1210                         break;
1211                 }
1212
1213                 for (n = 0; copied; n++, start = 0) {
1214                         int size = min_t(int, copied, PAGE_SIZE - start);
1215                         if (!failed) {
1216                                 buf.page = pages[n];
1217                                 buf.offset = start;
1218                                 buf.len = size;
1219                                 ret = add_to_pipe(pipe, &buf);
1220                                 if (unlikely(ret < 0)) {
1221                                         failed = true;
1222                                 } else {
1223                                         iov_iter_advance(from, ret);
1224                                         total += ret;
1225                                 }
1226                         } else {
1227                                 put_page(pages[n]);
1228                         }
1229                         copied -= size;
1230                 }
1231         }
1232         return total ? total : ret;
1233 }
1234
1235 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1236                         struct splice_desc *sd)
1237 {
1238         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1239         return n == sd->len ? n : -EFAULT;
1240 }
1241
1242 /*
1243  * For lack of a better implementation, implement vmsplice() to userspace
1244  * as a simple copy of the pipes pages to the user iov.
1245  */
1246 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1247                              unsigned int flags)
1248 {
1249         struct pipe_inode_info *pipe = get_pipe_info(file);
1250         struct splice_desc sd = {
1251                 .total_len = iov_iter_count(iter),
1252                 .flags = flags,
1253                 .u.data = iter
1254         };
1255         long ret = 0;
1256
1257         if (!pipe)
1258                 return -EBADF;
1259
1260         if (sd.total_len) {
1261                 pipe_lock(pipe);
1262                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1263                 pipe_unlock(pipe);
1264         }
1265
1266         return ret;
1267 }
1268
1269 /*
1270  * vmsplice splices a user address range into a pipe. It can be thought of
1271  * as splice-from-memory, where the regular splice is splice-from-file (or
1272  * to file). In both cases the output is a pipe, naturally.
1273  */
1274 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1275                              unsigned int flags)
1276 {
1277         struct pipe_inode_info *pipe;
1278         long ret = 0;
1279         unsigned buf_flag = 0;
1280
1281         if (flags & SPLICE_F_GIFT)
1282                 buf_flag = PIPE_BUF_FLAG_GIFT;
1283
1284         pipe = get_pipe_info(file);
1285         if (!pipe)
1286                 return -EBADF;
1287
1288         pipe_lock(pipe);
1289         ret = wait_for_space(pipe, flags);
1290         if (!ret)
1291                 ret = iter_to_pipe(iter, pipe, buf_flag);
1292         pipe_unlock(pipe);
1293         if (ret > 0)
1294                 wakeup_pipe_readers(pipe);
1295         return ret;
1296 }
1297
1298 static int vmsplice_type(struct fd f, int *type)
1299 {
1300         if (!f.file)
1301                 return -EBADF;
1302         if (f.file->f_mode & FMODE_WRITE) {
1303                 *type = WRITE;
1304         } else if (f.file->f_mode & FMODE_READ) {
1305                 *type = READ;
1306         } else {
1307                 fdput(f);
1308                 return -EBADF;
1309         }
1310         return 0;
1311 }
1312
1313 /*
1314  * Note that vmsplice only really supports true splicing _from_ user memory
1315  * to a pipe, not the other way around. Splicing from user memory is a simple
1316  * operation that can be supported without any funky alignment restrictions
1317  * or nasty vm tricks. We simply map in the user memory and fill them into
1318  * a pipe. The reverse isn't quite as easy, though. There are two possible
1319  * solutions for that:
1320  *
1321  *      - memcpy() the data internally, at which point we might as well just
1322  *        do a regular read() on the buffer anyway.
1323  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1324  *        has restriction limitations on both ends of the pipe).
1325  *
1326  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1327  *
1328  */
1329 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1330 {
1331         if (unlikely(flags & ~SPLICE_F_ALL))
1332                 return -EINVAL;
1333
1334         if (!iov_iter_count(iter))
1335                 return 0;
1336
1337         if (iov_iter_rw(iter) == WRITE)
1338                 return vmsplice_to_pipe(f, iter, flags);
1339         else
1340                 return vmsplice_to_user(f, iter, flags);
1341 }
1342
1343 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1344                 unsigned long, nr_segs, unsigned int, flags)
1345 {
1346         struct iovec iovstack[UIO_FASTIOV];
1347         struct iovec *iov = iovstack;
1348         struct iov_iter iter;
1349         long error;
1350         struct fd f;
1351         int type;
1352
1353         f = fdget(fd);
1354         error = vmsplice_type(f, &type);
1355         if (error)
1356                 return error;
1357
1358         error = import_iovec(type, uiov, nr_segs,
1359                              ARRAY_SIZE(iovstack), &iov, &iter);
1360         if (!error) {
1361                 error = do_vmsplice(f.file, &iter, flags);
1362                 kfree(iov);
1363         }
1364         fdput(f);
1365         return error;
1366 }
1367
1368 #ifdef CONFIG_COMPAT
1369 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1370                     unsigned int, nr_segs, unsigned int, flags)
1371 {
1372         struct iovec iovstack[UIO_FASTIOV];
1373         struct iovec *iov = iovstack;
1374         struct iov_iter iter;
1375         long error;
1376         struct fd f;
1377         int type;
1378
1379         f = fdget(fd);
1380         error = vmsplice_type(f, &type);
1381         if (error)
1382                 return error;
1383
1384         error = compat_import_iovec(type, iov32, nr_segs,
1385                              ARRAY_SIZE(iovstack), &iov, &iter);
1386         if (!error) {
1387                 error = do_vmsplice(f.file, &iter, flags);
1388                 kfree(iov);
1389         }
1390         fdput(f);
1391         return error;
1392 }
1393 #endif
1394
1395 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1396                 int, fd_out, loff_t __user *, off_out,
1397                 size_t, len, unsigned int, flags)
1398 {
1399         struct fd in, out;
1400         long error;
1401
1402         if (unlikely(!len))
1403                 return 0;
1404
1405         if (unlikely(flags & ~SPLICE_F_ALL))
1406                 return -EINVAL;
1407
1408         error = -EBADF;
1409         in = fdget(fd_in);
1410         if (in.file) {
1411                 if (in.file->f_mode & FMODE_READ) {
1412                         out = fdget(fd_out);
1413                         if (out.file) {
1414                                 if (out.file->f_mode & FMODE_WRITE)
1415                                         error = do_splice(in.file, off_in,
1416                                                           out.file, off_out,
1417                                                           len, flags);
1418                                 fdput(out);
1419                         }
1420                 }
1421                 fdput(in);
1422         }
1423         return error;
1424 }
1425
1426 /*
1427  * Make sure there's data to read. Wait for input if we can, otherwise
1428  * return an appropriate error.
1429  */
1430 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1431 {
1432         int ret;
1433
1434         /*
1435          * Check ->nrbufs without the inode lock first. This function
1436          * is speculative anyways, so missing one is ok.
1437          */
1438         if (pipe->nrbufs)
1439                 return 0;
1440
1441         ret = 0;
1442         pipe_lock(pipe);
1443
1444         while (!pipe->nrbufs) {
1445                 if (signal_pending(current)) {
1446                         ret = -ERESTARTSYS;
1447                         break;
1448                 }
1449                 if (!pipe->writers)
1450                         break;
1451                 if (!pipe->waiting_writers) {
1452                         if (flags & SPLICE_F_NONBLOCK) {
1453                                 ret = -EAGAIN;
1454                                 break;
1455                         }
1456                 }
1457                 pipe_wait(pipe);
1458         }
1459
1460         pipe_unlock(pipe);
1461         return ret;
1462 }
1463
1464 /*
1465  * Make sure there's writeable room. Wait for room if we can, otherwise
1466  * return an appropriate error.
1467  */
1468 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1469 {
1470         int ret;
1471
1472         /*
1473          * Check ->nrbufs without the inode lock first. This function
1474          * is speculative anyways, so missing one is ok.
1475          */
1476         if (pipe->nrbufs < pipe->buffers)
1477                 return 0;
1478
1479         ret = 0;
1480         pipe_lock(pipe);
1481
1482         while (pipe->nrbufs >= pipe->buffers) {
1483                 if (!pipe->readers) {
1484                         send_sig(SIGPIPE, current, 0);
1485                         ret = -EPIPE;
1486                         break;
1487                 }
1488                 if (flags & SPLICE_F_NONBLOCK) {
1489                         ret = -EAGAIN;
1490                         break;
1491                 }
1492                 if (signal_pending(current)) {
1493                         ret = -ERESTARTSYS;
1494                         break;
1495                 }
1496                 pipe->waiting_writers++;
1497                 pipe_wait(pipe);
1498                 pipe->waiting_writers--;
1499         }
1500
1501         pipe_unlock(pipe);
1502         return ret;
1503 }
1504
1505 /*
1506  * Splice contents of ipipe to opipe.
1507  */
1508 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1509                                struct pipe_inode_info *opipe,
1510                                size_t len, unsigned int flags)
1511 {
1512         struct pipe_buffer *ibuf, *obuf;
1513         int ret = 0, nbuf;
1514         bool input_wakeup = false;
1515
1516
1517 retry:
1518         ret = ipipe_prep(ipipe, flags);
1519         if (ret)
1520                 return ret;
1521
1522         ret = opipe_prep(opipe, flags);
1523         if (ret)
1524                 return ret;
1525
1526         /*
1527          * Potential ABBA deadlock, work around it by ordering lock
1528          * grabbing by pipe info address. Otherwise two different processes
1529          * could deadlock (one doing tee from A -> B, the other from B -> A).
1530          */
1531         pipe_double_lock(ipipe, opipe);
1532
1533         do {
1534                 if (!opipe->readers) {
1535                         send_sig(SIGPIPE, current, 0);
1536                         if (!ret)
1537                                 ret = -EPIPE;
1538                         break;
1539                 }
1540
1541                 if (!ipipe->nrbufs && !ipipe->writers)
1542                         break;
1543
1544                 /*
1545                  * Cannot make any progress, because either the input
1546                  * pipe is empty or the output pipe is full.
1547                  */
1548                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1549                         /* Already processed some buffers, break */
1550                         if (ret)
1551                                 break;
1552
1553                         if (flags & SPLICE_F_NONBLOCK) {
1554                                 ret = -EAGAIN;
1555                                 break;
1556                         }
1557
1558                         /*
1559                          * We raced with another reader/writer and haven't
1560                          * managed to process any buffers.  A zero return
1561                          * value means EOF, so retry instead.
1562                          */
1563                         pipe_unlock(ipipe);
1564                         pipe_unlock(opipe);
1565                         goto retry;
1566                 }
1567
1568                 ibuf = ipipe->bufs + ipipe->curbuf;
1569                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1570                 obuf = opipe->bufs + nbuf;
1571
1572                 if (len >= ibuf->len) {
1573                         /*
1574                          * Simply move the whole buffer from ipipe to opipe
1575                          */
1576                         *obuf = *ibuf;
1577                         ibuf->ops = NULL;
1578                         opipe->nrbufs++;
1579                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1580                         ipipe->nrbufs--;
1581                         input_wakeup = true;
1582                 } else {
1583                         /*
1584                          * Get a reference to this pipe buffer,
1585                          * so we can copy the contents over.
1586                          */
1587                         if (!pipe_buf_get(ipipe, ibuf)) {
1588                                 if (ret == 0)
1589                                         ret = -EFAULT;
1590                                 break;
1591                         }
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                         pipe_buf_mark_unmergeable(obuf);
1601
1602                         obuf->len = len;
1603                         opipe->nrbufs++;
1604                         ibuf->offset += obuf->len;
1605                         ibuf->len -= obuf->len;
1606                 }
1607                 ret += obuf->len;
1608                 len -= obuf->len;
1609         } while (len);
1610
1611         pipe_unlock(ipipe);
1612         pipe_unlock(opipe);
1613
1614         /*
1615          * If we put data in the output pipe, wakeup any potential readers.
1616          */
1617         if (ret > 0)
1618                 wakeup_pipe_readers(opipe);
1619
1620         if (input_wakeup)
1621                 wakeup_pipe_writers(ipipe);
1622
1623         return ret;
1624 }
1625
1626 /*
1627  * Link contents of ipipe to opipe.
1628  */
1629 static int link_pipe(struct pipe_inode_info *ipipe,
1630                      struct pipe_inode_info *opipe,
1631                      size_t len, unsigned int flags)
1632 {
1633         struct pipe_buffer *ibuf, *obuf;
1634         int ret = 0, i = 0, nbuf;
1635
1636         /*
1637          * Potential ABBA deadlock, work around it by ordering lock
1638          * grabbing by pipe info address. Otherwise two different processes
1639          * could deadlock (one doing tee from A -> B, the other from B -> A).
1640          */
1641         pipe_double_lock(ipipe, opipe);
1642
1643         do {
1644                 if (!opipe->readers) {
1645                         send_sig(SIGPIPE, current, 0);
1646                         if (!ret)
1647                                 ret = -EPIPE;
1648                         break;
1649                 }
1650
1651                 /*
1652                  * If we have iterated all input buffers or ran out of
1653                  * output room, break.
1654                  */
1655                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1656                         break;
1657
1658                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1659                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1660
1661                 /*
1662                  * Get a reference to this pipe buffer,
1663                  * so we can copy the contents over.
1664                  */
1665                 if (!pipe_buf_get(ipipe, ibuf)) {
1666                         if (ret == 0)
1667                                 ret = -EFAULT;
1668                         break;
1669                 }
1670
1671                 obuf = opipe->bufs + nbuf;
1672                 *obuf = *ibuf;
1673
1674                 /*
1675                  * Don't inherit the gift flag, we need to
1676                  * prevent multiple steals of this page.
1677                  */
1678                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1679
1680                 pipe_buf_mark_unmergeable(obuf);
1681
1682                 if (obuf->len > len)
1683                         obuf->len = len;
1684
1685                 opipe->nrbufs++;
1686                 ret += obuf->len;
1687                 len -= obuf->len;
1688                 i++;
1689         } while (len);
1690
1691         /*
1692          * return EAGAIN if we have the potential of some data in the
1693          * future, otherwise just return 0
1694          */
1695         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1696                 ret = -EAGAIN;
1697
1698         pipe_unlock(ipipe);
1699         pipe_unlock(opipe);
1700
1701         /*
1702          * If we put data in the output pipe, wakeup any potential readers.
1703          */
1704         if (ret > 0)
1705                 wakeup_pipe_readers(opipe);
1706
1707         return ret;
1708 }
1709
1710 /*
1711  * This is a tee(1) implementation that works on pipes. It doesn't copy
1712  * any data, it simply references the 'in' pages on the 'out' pipe.
1713  * The 'flags' used are the SPLICE_F_* variants, currently the only
1714  * applicable one is SPLICE_F_NONBLOCK.
1715  */
1716 static long do_tee(struct file *in, struct file *out, size_t len,
1717                    unsigned int flags)
1718 {
1719         struct pipe_inode_info *ipipe = get_pipe_info(in);
1720         struct pipe_inode_info *opipe = get_pipe_info(out);
1721         int ret = -EINVAL;
1722
1723         /*
1724          * Duplicate the contents of ipipe to opipe without actually
1725          * copying the data.
1726          */
1727         if (ipipe && opipe && ipipe != opipe) {
1728                 /*
1729                  * Keep going, unless we encounter an error. The ipipe/opipe
1730                  * ordering doesn't really matter.
1731                  */
1732                 ret = ipipe_prep(ipipe, flags);
1733                 if (!ret) {
1734                         ret = opipe_prep(opipe, flags);
1735                         if (!ret)
1736                                 ret = link_pipe(ipipe, opipe, len, flags);
1737                 }
1738         }
1739
1740         return ret;
1741 }
1742
1743 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1744 {
1745         struct fd in;
1746         int error;
1747
1748         if (unlikely(flags & ~SPLICE_F_ALL))
1749                 return -EINVAL;
1750
1751         if (unlikely(!len))
1752                 return 0;
1753
1754         error = -EBADF;
1755         in = fdget(fdin);
1756         if (in.file) {
1757                 if (in.file->f_mode & FMODE_READ) {
1758                         struct fd out = fdget(fdout);
1759                         if (out.file) {
1760                                 if (out.file->f_mode & FMODE_WRITE)
1761                                         error = do_tee(in.file, out.file,
1762                                                         len, flags);
1763                                 fdput(out);
1764                         }
1765                 }
1766                 fdput(in);
1767         }
1768
1769         return error;
1770 }