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/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34 #include <linux/compat.h>
  35 #include "internal.h"
  36
  37 /*
  38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  39  * a vm helper function, it's already simplified quite a bit by the
  40  * addition of remove_mapping(). If success is returned, the caller may
  41  * attempt to reuse this page for another destination.
  42  */
  43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  44                                      struct pipe_buffer *buf)
  45 {
  46         struct page *page = buf->page;
  47         struct address_space *mapping;
  48
  49         lock_page(page);
  50
  51         mapping = page_mapping(page);
  52         if (mapping) {
  53                 WARN_ON(!PageUptodate(page));
  54
  55                 /*
  56                  * At least for ext2 with nobh option, we need to wait on
  57                  * writeback completing on this page, since we'll remove it
  58                  * from the pagecache.  Otherwise truncate wont wait on the
  59                  * page, allowing the disk blocks to be reused by someone else
  60                  * before we actually wrote our data to them. fs corruption
  61                  * ensues.
  62                  */
  63                 wait_on_page_writeback(page);
  64
  65                 if (page_has_private(page) &&
  66                     !try_to_release_page(page, GFP_KERNEL))
  67                         goto out_unlock;
  68
  69                 /*
  70                  * If we succeeded in removing the mapping, set LRU flag
  71                  * and return good.
  72                  */
  73                 if (remove_mapping(mapping, page)) {
  74                         buf->flags |= PIPE_BUF_FLAG_LRU;
  75                         return 0;
  76                 }
  77         }
  78
  79         /*
  80          * Raced with truncate or failed to remove page from current
  81          * address space, unlock and return failure.
  82          */
  83 out_unlock:
  84         unlock_page(page);
  85         return 1;
  86 }
  87
  88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  89                                         struct pipe_buffer *buf)
  90 {
  91         put_page(buf->page);
  92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  93 }
  94
  95 /*
  96  * Check whether the contents of buf is OK to access. Since the content
  97  * is a page cache page, IO may be in flight.
  98  */
  99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 100                                        struct pipe_buffer *buf)
 101 {
 102         struct page *page = buf->page;
 103         int err;
 104
 105         if (!PageUptodate(page)) {
 106                 lock_page(page);
 107
 108                 /*
 109                  * Page got truncated/unhashed. This will cause a 0-byte
 110                  * splice, if this is the first page.
 111                  */
 112                 if (!page->mapping) {
 113                         err = -ENODATA;
 114                         goto error;
 115                 }
 116
 117                 /*
 118                  * Uh oh, read-error from disk.
 119                  */
 120                 if (!PageUptodate(page)) {
 121                         err = -EIO;
 122                         goto error;
 123                 }
 124
 125                 /*
 126                  * Page is ok afterall, we are done.
 127                  */
 128                 unlock_page(page);
 129         }
 130
 131         return 0;
 132 error:
 133         unlock_page(page);
 134         return err;
 135 }
 136
 137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .confirm = generic_pipe_buf_confirm,
 158         .release = page_cache_pipe_buf_release,
 159         .steal = user_page_pipe_buf_steal,
 160         .get = generic_pipe_buf_get,
 161 };
 162
 163 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 164 {
 165         smp_mb();
 166         if (waitqueue_active(&pipe->wait))
 167                 wake_up_interruptible(&pipe->wait);
 168         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 169 }
 170
 171 /**
 172  * splice_to_pipe - fill passed data into a pipe
 173  * @pipe:       pipe to fill
 174  * @spd:        data to fill
 175  *
 176  * Description:
 177  *    @spd contains a map of pages and len/offset tuples, along with
 178  *    the struct pipe_buf_operations associated with these pages. This
 179  *    function will link that data to the pipe.
 180  *
 181  */
 182 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 183                        struct splice_pipe_desc *spd)
 184 {
 185         unsigned int spd_pages = spd->nr_pages;
 186         int ret = 0, page_nr = 0;
 187
 188         if (!spd_pages)
 189                 return 0;
 190
 191         if (unlikely(!pipe->readers)) {
 192                 send_sig(SIGPIPE, current, 0);
 193                 ret = -EPIPE;
 194                 goto out;
 195         }
 196
 197         while (pipe->nrbufs < pipe->buffers) {
 198                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 199                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 200
 201                 buf->page = spd->pages[page_nr];
 202                 buf->offset = spd->partial[page_nr].offset;
 203                 buf->len = spd->partial[page_nr].len;
 204                 buf->private = spd->partial[page_nr].private;
 205                 buf->ops = spd->ops;
 206
 207                 pipe->nrbufs++;
 208                 page_nr++;
 209                 ret += buf->len;
 210
 211                 if (!--spd->nr_pages)
 212                         break;
 213         }
 214
 215         if (!ret)
 216                 ret = -EAGAIN;
 217
 218 out:
 219         while (page_nr < spd_pages)
 220                 spd->spd_release(spd, page_nr++);
 221
 222         return ret;
 223 }
 224 EXPORT_SYMBOL_GPL(splice_to_pipe);
 225
 226 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 227 {
 228         int ret;
 229
 230         if (unlikely(!pipe->readers)) {
 231                 send_sig(SIGPIPE, current, 0);
 232                 ret = -EPIPE;
 233         } else if (pipe->nrbufs == pipe->buffers) {
 234                 ret = -EAGAIN;
 235         } else {
 236                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 237                 pipe->bufs[newbuf] = *buf;
 238                 pipe->nrbufs++;
 239                 return buf->len;
 240         }
 241         pipe_buf_release(pipe, buf);
 242         return ret;
 243 }
 244 EXPORT_SYMBOL(add_to_pipe);
 245
 246 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 247 {
 248         put_page(spd->pages[i]);
 249 }
 250
 251 /*
 252  * Check if we need to grow the arrays holding pages and partial page
 253  * descriptions.
 254  */
 255 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 256 {
 257         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 258
 259         spd->nr_pages_max = buffers;
 260         if (buffers <= PIPE_DEF_BUFFERS)
 261                 return 0;
 262
 263         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 264         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 265
 266         if (spd->pages && spd->partial)
 267                 return 0;
 268
 269         kfree(spd->pages);
 270         kfree(spd->partial);
 271         return -ENOMEM;
 272 }
 273
 274 void splice_shrink_spd(struct splice_pipe_desc *spd)
 275 {
 276         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 277                 return;
 278
 279         kfree(spd->pages);
 280         kfree(spd->partial);
 281 }
 282
 283 /**
 284  * generic_file_splice_read - splice data from file to a pipe
 285  * @in:         file to splice from
 286  * @ppos:       position in @in
 287  * @pipe:       pipe to splice to
 288  * @len:        number of bytes to splice
 289  * @flags:      splice modifier flags
 290  *
 291  * Description:
 292  *    Will read pages from given file and fill them into a pipe. Can be
 293  *    used as long as it has more or less sane ->read_iter().
 294  *
 295  */
 296 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 297                                  struct pipe_inode_info *pipe, size_t len,
 298                                  unsigned int flags)
 299 {
 300         struct iov_iter to;
 301         struct kiocb kiocb;
 302         int idx, ret;
 303
 304         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 305         idx = to.idx;
 306         init_sync_kiocb(&kiocb, in);
 307         kiocb.ki_pos = *ppos;
 308         ret = in->f_op->read_iter(&kiocb, &to);
 309         if (ret > 0) {
 310                 *ppos = kiocb.ki_pos;
 311                 file_accessed(in);
 312         } else if (ret < 0) {
 313                 to.idx = idx;
 314                 to.iov_offset = 0;
 315                 iov_iter_advance(&to, 0); /* to free what was emitted */
 316                 /*
 317                  * callers of ->splice_read() expect -EAGAIN on
 318                  * "can't put anything in there", rather than -EFAULT.
 319                  */
 320                 if (ret == -EFAULT)
 321                         ret = -EAGAIN;
 322         }
 323
 324         return ret;
 325 }
 326 EXPORT_SYMBOL(generic_file_splice_read);
 327
 328 const struct pipe_buf_operations default_pipe_buf_ops = {
 329         .can_merge = 0,
 330         .confirm = generic_pipe_buf_confirm,
 331         .release = generic_pipe_buf_release,
 332         .steal = generic_pipe_buf_steal,
 333         .get = generic_pipe_buf_get,
 334 };
 335
 336 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 337                                     struct pipe_buffer *buf)
 338 {
 339         return 1;
 340 }
 341
 342 /* Pipe buffer operations for a socket and similar. */
 343 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 344         .can_merge = 0,
 345         .confirm = generic_pipe_buf_confirm,
 346         .release = generic_pipe_buf_release,
 347         .steal = generic_pipe_buf_nosteal,
 348         .get = generic_pipe_buf_get,
 349 };
 350 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 351
 352 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 353                             unsigned long vlen, loff_t offset)
 354 {
 355         mm_segment_t old_fs;
 356         loff_t pos = offset;
 357         ssize_t res;
 358
 359         old_fs = get_fs();
 360         set_fs(get_ds());
 361         /* The cast to a user pointer is valid due to the set_fs() */
 362         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 363         set_fs(old_fs);
 364
 365         return res;
 366 }
 367
 368 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 369                             loff_t pos)
 370 {
 371         mm_segment_t old_fs;
 372         ssize_t res;
 373
 374         old_fs = get_fs();
 375         set_fs(get_ds());
 376         /* The cast to a user pointer is valid due to the set_fs() */
 377         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 378         set_fs(old_fs);
 379
 380         return res;
 381 }
 382 EXPORT_SYMBOL(kernel_write);
 383
 384 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 385                                  struct pipe_inode_info *pipe, size_t len,
 386                                  unsigned int flags)
 387 {
 388         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 389         struct iov_iter to;
 390         struct page **pages;
 391         unsigned int nr_pages;
 392         size_t offset, dummy, copied = 0;
 393         ssize_t res;
 394         int i;
 395
 396         if (pipe->nrbufs == pipe->buffers)
 397                 return -EAGAIN;
 398
 399         /*
 400          * Try to keep page boundaries matching to source pagecache ones -
 401          * it probably won't be much help, but...
 402          */
 403         offset = *ppos & ~PAGE_MASK;
 404
 405         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 406
 407         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
 408         if (res <= 0)
 409                 return -ENOMEM;
 410
 411         BUG_ON(dummy);
 412         nr_pages = DIV_ROUND_UP(res, PAGE_SIZE);
 413
 414         vec = __vec;
 415         if (nr_pages > PIPE_DEF_BUFFERS) {
 416                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 417                 if (unlikely(!vec)) {
 418                         res = -ENOMEM;
 419                         goto out;
 420                 }
 421         }
 422
 423         pipe->bufs[to.idx].offset = offset;
 424         pipe->bufs[to.idx].len -= offset;
 425
 426         for (i = 0; i < nr_pages; i++) {
 427                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 428                 vec[i].iov_base = page_address(pages[i]) + offset;
 429                 vec[i].iov_len = this_len;
 430                 len -= this_len;
 431                 offset = 0;
 432         }
 433
 434         res = kernel_readv(in, vec, nr_pages, *ppos);
 435         if (res > 0) {
 436                 copied = res;
 437                 *ppos += res;
 438         }
 439
 440         if (vec != __vec)
 441                 kfree(vec);
 442 out:
 443         for (i = 0; i < nr_pages; i++)
 444                 put_page(pages[i]);
 445         kvfree(pages);
 446         iov_iter_advance(&to, copied);  /* truncates and discards */
 447         return res;
 448 }
 449
 450 /*
 451  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 452  * using sendpage(). Return the number of bytes sent.
 453  */
 454 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 455                             struct pipe_buffer *buf, struct splice_desc *sd)
 456 {
 457         struct file *file = sd->u.file;
 458         loff_t pos = sd->pos;
 459         int more;
 460
 461         if (!likely(file->f_op->sendpage))
 462                 return -EINVAL;
 463
 464         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 465
 466         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 467                 more |= MSG_SENDPAGE_NOTLAST;
 468
 469         return file->f_op->sendpage(file, buf->page, buf->offset,
 470                                     sd->len, &pos, more);
 471 }
 472
 473 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 474 {
 475         smp_mb();
 476         if (waitqueue_active(&pipe->wait))
 477                 wake_up_interruptible(&pipe->wait);
 478         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 479 }
 480
 481 /**
 482  * splice_from_pipe_feed - feed available data from a pipe to a file
 483  * @pipe:       pipe to splice from
 484  * @sd:         information to @actor
 485  * @actor:      handler that splices the data
 486  *
 487  * Description:
 488  *    This function loops over the pipe and calls @actor to do the
 489  *    actual moving of a single struct pipe_buffer to the desired
 490  *    destination.  It returns when there's no more buffers left in
 491  *    the pipe or if the requested number of bytes (@sd->total_len)
 492  *    have been copied.  It returns a positive number (one) if the
 493  *    pipe needs to be filled with more data, zero if the required
 494  *    number of bytes have been copied and -errno on error.
 495  *
 496  *    This, together with splice_from_pipe_{begin,end,next}, may be
 497  *    used to implement the functionality of __splice_from_pipe() when
 498  *    locking is required around copying the pipe buffers to the
 499  *    destination.
 500  */
 501 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 502                           splice_actor *actor)
 503 {
 504         int ret;
 505
 506         while (pipe->nrbufs) {
 507                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 508
 509                 sd->len = buf->len;
 510                 if (sd->len > sd->total_len)
 511                         sd->len = sd->total_len;
 512
 513                 ret = pipe_buf_confirm(pipe, buf);
 514                 if (unlikely(ret)) {
 515                         if (ret == -ENODATA)
 516                                 ret = 0;
 517                         return ret;
 518                 }
 519
 520                 ret = actor(pipe, buf, sd);
 521                 if (ret <= 0)
 522                         return ret;
 523
 524                 buf->offset += ret;
 525                 buf->len -= ret;
 526
 527                 sd->num_spliced += ret;
 528                 sd->len -= ret;
 529                 sd->pos += ret;
 530                 sd->total_len -= ret;
 531
 532                 if (!buf->len) {
 533                         pipe_buf_release(pipe, buf);
 534                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 535                         pipe->nrbufs--;
 536                         if (pipe->files)
 537                                 sd->need_wakeup = true;
 538                 }
 539
 540                 if (!sd->total_len)
 541                         return 0;
 542         }
 543
 544         return 1;
 545 }
 546
 547 /**
 548  * splice_from_pipe_next - wait for some data to splice from
 549  * @pipe:       pipe to splice from
 550  * @sd:         information about the splice operation
 551  *
 552  * Description:
 553  *    This function will wait for some data and return a positive
 554  *    value (one) if pipe buffers are available.  It will return zero
 555  *    or -errno if no more data needs to be spliced.
 556  */
 557 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 558 {
 559         /*
 560          * Check for signal early to make process killable when there are
 561          * always buffers available
 562          */
 563         if (signal_pending(current))
 564                 return -ERESTARTSYS;
 565
 566         while (!pipe->nrbufs) {
 567                 if (!pipe->writers)
 568                         return 0;
 569
 570                 if (!pipe->waiting_writers && sd->num_spliced)
 571                         return 0;
 572
 573                 if (sd->flags & SPLICE_F_NONBLOCK)
 574                         return -EAGAIN;
 575
 576                 if (signal_pending(current))
 577                         return -ERESTARTSYS;
 578
 579                 if (sd->need_wakeup) {
 580                         wakeup_pipe_writers(pipe);
 581                         sd->need_wakeup = false;
 582                 }
 583
 584                 pipe_wait(pipe);
 585         }
 586
 587         return 1;
 588 }
 589
 590 /**
 591  * splice_from_pipe_begin - start splicing from pipe
 592  * @sd:         information about the splice operation
 593  *
 594  * Description:
 595  *    This function should be called before a loop containing
 596  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 597  *    initialize the necessary fields of @sd.
 598  */
 599 static void splice_from_pipe_begin(struct splice_desc *sd)
 600 {
 601         sd->num_spliced = 0;
 602         sd->need_wakeup = false;
 603 }
 604
 605 /**
 606  * splice_from_pipe_end - finish splicing from pipe
 607  * @pipe:       pipe to splice from
 608  * @sd:         information about the splice operation
 609  *
 610  * Description:
 611  *    This function will wake up pipe writers if necessary.  It should
 612  *    be called after a loop containing splice_from_pipe_next() and
 613  *    splice_from_pipe_feed().
 614  */
 615 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 616 {
 617         if (sd->need_wakeup)
 618                 wakeup_pipe_writers(pipe);
 619 }
 620
 621 /**
 622  * __splice_from_pipe - splice data from a pipe to given actor
 623  * @pipe:       pipe to splice from
 624  * @sd:         information to @actor
 625  * @actor:      handler that splices the data
 626  *
 627  * Description:
 628  *    This function does little more than loop over the pipe and call
 629  *    @actor to do the actual moving of a single struct pipe_buffer to
 630  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 631  *    pipe_to_user.
 632  *
 633  */
 634 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 635                            splice_actor *actor)
 636 {
 637         int ret;
 638
 639         splice_from_pipe_begin(sd);
 640         do {
 641                 cond_resched();
 642                 ret = splice_from_pipe_next(pipe, sd);
 643                 if (ret > 0)
 644                         ret = splice_from_pipe_feed(pipe, sd, actor);
 645         } while (ret > 0);
 646         splice_from_pipe_end(pipe, sd);
 647
 648         return sd->num_spliced ? sd->num_spliced : ret;
 649 }
 650 EXPORT_SYMBOL(__splice_from_pipe);
 651
 652 /**
 653  * splice_from_pipe - splice data from a pipe to a file
 654  * @pipe:       pipe to splice from
 655  * @out:        file to splice to
 656  * @ppos:       position in @out
 657  * @len:        how many bytes to splice
 658  * @flags:      splice modifier flags
 659  * @actor:      handler that splices the data
 660  *
 661  * Description:
 662  *    See __splice_from_pipe. This function locks the pipe inode,
 663  *    otherwise it's identical to __splice_from_pipe().
 664  *
 665  */
 666 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 667                          loff_t *ppos, size_t len, unsigned int flags,
 668                          splice_actor *actor)
 669 {
 670         ssize_t ret;
 671         struct splice_desc sd = {
 672                 .total_len = len,
 673                 .flags = flags,
 674                 .pos = *ppos,
 675                 .u.file = out,
 676         };
 677
 678         pipe_lock(pipe);
 679         ret = __splice_from_pipe(pipe, &sd, actor);
 680         pipe_unlock(pipe);
 681
 682         return ret;
 683 }
 684
 685 /**
 686  * iter_file_splice_write - splice data from a pipe to a file
 687  * @pipe:       pipe info
 688  * @out:        file to write to
 689  * @ppos:       position in @out
 690  * @len:        number of bytes to splice
 691  * @flags:      splice modifier flags
 692  *
 693  * Description:
 694  *    Will either move or copy pages (determined by @flags options) from
 695  *    the given pipe inode to the given file.
 696  *    This one is ->write_iter-based.
 697  *
 698  */
 699 ssize_t
 700 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 701                           loff_t *ppos, size_t len, unsigned int flags)
 702 {
 703         struct splice_desc sd = {
 704                 .total_len = len,
 705                 .flags = flags,
 706                 .pos = *ppos,
 707                 .u.file = out,
 708         };
 709         int nbufs = pipe->buffers;
 710         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 711                                         GFP_KERNEL);
 712         ssize_t ret;
 713
 714         if (unlikely(!array))
 715                 return -ENOMEM;
 716
 717         pipe_lock(pipe);
 718
 719         splice_from_pipe_begin(&sd);
 720         while (sd.total_len) {
 721                 struct iov_iter from;
 722                 size_t left;
 723                 int n, idx;
 724
 725                 ret = splice_from_pipe_next(pipe, &sd);
 726                 if (ret <= 0)
 727                         break;
 728
 729                 if (unlikely(nbufs < pipe->buffers)) {
 730                         kfree(array);
 731                         nbufs = pipe->buffers;
 732                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 733                                         GFP_KERNEL);
 734                         if (!array) {
 735                                 ret = -ENOMEM;
 736                                 break;
 737                         }
 738                 }
 739
 740                 /* build the vector */
 741                 left = sd.total_len;
 742                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 743                         struct pipe_buffer *buf = pipe->bufs + idx;
 744                         size_t this_len = buf->len;
 745
 746                         if (this_len > left)
 747                                 this_len = left;
 748
 749                         if (idx == pipe->buffers - 1)
 750                                 idx = -1;
 751
 752                         ret = pipe_buf_confirm(pipe, buf);
 753                         if (unlikely(ret)) {
 754                                 if (ret == -ENODATA)
 755                                         ret = 0;
 756                                 goto done;
 757                         }
 758
 759                         array[n].bv_page = buf->page;
 760                         array[n].bv_len = this_len;
 761                         array[n].bv_offset = buf->offset;
 762                         left -= this_len;
 763                 }
 764
 765                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 766                               sd.total_len - left);
 767                 ret = vfs_iter_write(out, &from, &sd.pos);
 768                 if (ret <= 0)
 769                         break;
 770
 771                 sd.num_spliced += ret;
 772                 sd.total_len -= ret;
 773                 *ppos = sd.pos;
 774
 775                 /* dismiss the fully eaten buffers, adjust the partial one */
 776                 while (ret) {
 777                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 778                         if (ret >= buf->len) {
 779                                 ret -= buf->len;
 780                                 buf->len = 0;
 781                                 pipe_buf_release(pipe, buf);
 782                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 783                                 pipe->nrbufs--;
 784                                 if (pipe->files)
 785                                         sd.need_wakeup = true;
 786                         } else {
 787                                 buf->offset += ret;
 788                                 buf->len -= ret;
 789                                 ret = 0;
 790                         }
 791                 }
 792         }
 793 done:
 794         kfree(array);
 795         splice_from_pipe_end(pipe, &sd);
 796
 797         pipe_unlock(pipe);
 798
 799         if (sd.num_spliced)
 800                 ret = sd.num_spliced;
 801
 802         return ret;
 803 }
 804
 805 EXPORT_SYMBOL(iter_file_splice_write);
 806
 807 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 808                           struct splice_desc *sd)
 809 {
 810         int ret;
 811         void *data;
 812         loff_t tmp = sd->pos;
 813
 814         data = kmap(buf->page);
 815         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 816         kunmap(buf->page);
 817
 818         return ret;
 819 }
 820
 821 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 822                                          struct file *out, loff_t *ppos,
 823                                          size_t len, unsigned int flags)
 824 {
 825         ssize_t ret;
 826
 827         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 828         if (ret > 0)
 829                 *ppos += ret;
 830
 831         return ret;
 832 }
 833
 834 /**
 835  * generic_splice_sendpage - splice data from a pipe to a socket
 836  * @pipe:       pipe to splice from
 837  * @out:        socket to write to
 838  * @ppos:       position in @out
 839  * @len:        number of bytes to splice
 840  * @flags:      splice modifier flags
 841  *
 842  * Description:
 843  *    Will send @len bytes from the pipe to a network socket. No data copying
 844  *    is involved.
 845  *
 846  */
 847 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 848                                 loff_t *ppos, size_t len, unsigned int flags)
 849 {
 850         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 851 }
 852
 853 EXPORT_SYMBOL(generic_splice_sendpage);
 854
 855 /*
 856  * Attempt to initiate a splice from pipe to file.
 857  */
 858 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 859                            loff_t *ppos, size_t len, unsigned int flags)
 860 {
 861         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 862                                 loff_t *, size_t, unsigned int);
 863
 864         if (out->f_op->splice_write)
 865                 splice_write = out->f_op->splice_write;
 866         else
 867                 splice_write = default_file_splice_write;
 868
 869         return splice_write(pipe, out, ppos, len, flags);
 870 }
 871
 872 /*
 873  * Attempt to initiate a splice from a file to a pipe.
 874  */
 875 static long do_splice_to(struct file *in, loff_t *ppos,
 876                          struct pipe_inode_info *pipe, size_t len,
 877                          unsigned int flags)
 878 {
 879         ssize_t (*splice_read)(struct file *, loff_t *,
 880                                struct pipe_inode_info *, size_t, unsigned int);
 881         int ret;
 882
 883         if (unlikely(!(in->f_mode & FMODE_READ)))
 884                 return -EBADF;
 885
 886         ret = rw_verify_area(READ, in, ppos, len);
 887         if (unlikely(ret < 0))
 888                 return ret;
 889
 890         if (unlikely(len > MAX_RW_COUNT))
 891                 len = MAX_RW_COUNT;
 892
 893         if (in->f_op->splice_read)
 894                 splice_read = in->f_op->splice_read;
 895         else
 896                 splice_read = default_file_splice_read;
 897
 898         return splice_read(in, ppos, pipe, len, flags);
 899 }
 900
 901 /**
 902  * splice_direct_to_actor - splices data directly between two non-pipes
 903  * @in:         file to splice from
 904  * @sd:         actor information on where to splice to
 905  * @actor:      handles the data splicing
 906  *
 907  * Description:
 908  *    This is a special case helper to splice directly between two
 909  *    points, without requiring an explicit pipe. Internally an allocated
 910  *    pipe is cached in the process, and reused during the lifetime of
 911  *    that process.
 912  *
 913  */
 914 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 915                                splice_direct_actor *actor)
 916 {
 917         struct pipe_inode_info *pipe;
 918         long ret, bytes;
 919         umode_t i_mode;
 920         size_t len;
 921         int i, flags, more;
 922
 923         /*
 924          * We require the input being a regular file, as we don't want to
 925          * randomly drop data for eg socket -> socket splicing. Use the
 926          * piped splicing for that!
 927          */
 928         i_mode = file_inode(in)->i_mode;
 929         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 930                 return -EINVAL;
 931
 932         /*
 933          * neither in nor out is a pipe, setup an internal pipe attached to
 934          * 'out' and transfer the wanted data from 'in' to 'out' through that
 935          */
 936         pipe = current->splice_pipe;
 937         if (unlikely(!pipe)) {
 938                 pipe = alloc_pipe_info();
 939                 if (!pipe)
 940                         return -ENOMEM;
 941
 942                 /*
 943                  * We don't have an immediate reader, but we'll read the stuff
 944                  * out of the pipe right after the splice_to_pipe(). So set
 945                  * PIPE_READERS appropriately.
 946                  */
 947                 pipe->readers = 1;
 948
 949                 current->splice_pipe = pipe;
 950         }
 951
 952         /*
 953          * Do the splice.
 954          */
 955         ret = 0;
 956         bytes = 0;
 957         len = sd->total_len;
 958         flags = sd->flags;
 959
 960         /*
 961          * Don't block on output, we have to drain the direct pipe.
 962          */
 963         sd->flags &= ~SPLICE_F_NONBLOCK;
 964         more = sd->flags & SPLICE_F_MORE;
 965
 966         while (len) {
 967                 size_t read_len;
 968                 loff_t pos = sd->pos, prev_pos = pos;
 969
 970                 ret = do_splice_to(in, &pos, pipe, len, flags);
 971                 if (unlikely(ret <= 0))
 972                         goto out_release;
 973
 974                 read_len = ret;
 975                 sd->total_len = read_len;
 976
 977                 /*
 978                  * If more data is pending, set SPLICE_F_MORE
 979                  * If this is the last data and SPLICE_F_MORE was not set
 980                  * initially, clears it.
 981                  */
 982                 if (read_len < len)
 983                         sd->flags |= SPLICE_F_MORE;
 984                 else if (!more)
 985                         sd->flags &= ~SPLICE_F_MORE;
 986                 /*
 987                  * NOTE: nonblocking mode only applies to the input. We
 988                  * must not do the output in nonblocking mode as then we
 989                  * could get stuck data in the internal pipe:
 990                  */
 991                 ret = actor(pipe, sd);
 992                 if (unlikely(ret <= 0)) {
 993                         sd->pos = prev_pos;
 994                         goto out_release;
 995                 }
 996
 997                 bytes += ret;
 998                 len -= ret;
 999                 sd->pos = pos;
1000
1001                 if (ret < read_len) {
1002                         sd->pos = prev_pos + ret;
1003                         goto out_release;
1004                 }
1005         }
1006
1007 done:
1008         pipe->nrbufs = pipe->curbuf = 0;
1009         file_accessed(in);
1010         return bytes;
1011
1012 out_release:
1013         /*
1014          * If we did an incomplete transfer we must release
1015          * the pipe buffers in question:
1016          */
1017         for (i = 0; i < pipe->buffers; i++) {
1018                 struct pipe_buffer *buf = pipe->bufs + i;
1019
1020                 if (buf->ops)
1021                         pipe_buf_release(pipe, buf);
1022         }
1023
1024         if (!bytes)
1025                 bytes = ret;
1026
1027         goto done;
1028 }
1029 EXPORT_SYMBOL(splice_direct_to_actor);
1030
1031 static int direct_splice_actor(struct pipe_inode_info *pipe,
1032                                struct splice_desc *sd)
1033 {
1034         struct file *file = sd->u.file;
1035
1036         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1037                               sd->flags);
1038 }
1039
1040 /**
1041  * do_splice_direct - splices data directly between two files
1042  * @in:         file to splice from
1043  * @ppos:       input file offset
1044  * @out:        file to splice to
1045  * @opos:       output file offset
1046  * @len:        number of bytes to splice
1047  * @flags:      splice modifier flags
1048  *
1049  * Description:
1050  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1051  *    doing it in the application would incur an extra system call
1052  *    (splice in + splice out, as compared to just sendfile()). So this helper
1053  *    can splice directly through a process-private pipe.
1054  *
1055  */
1056 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1057                       loff_t *opos, size_t len, unsigned int flags)
1058 {
1059         struct splice_desc sd = {
1060                 .len            = len,
1061                 .total_len      = len,
1062                 .flags          = flags,
1063                 .pos            = *ppos,
1064                 .u.file         = out,
1065                 .opos           = opos,
1066         };
1067         long ret;
1068
1069         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1070                 return -EBADF;
1071
1072         if (unlikely(out->f_flags & O_APPEND))
1073                 return -EINVAL;
1074
1075         ret = rw_verify_area(WRITE, out, opos, len);
1076         if (unlikely(ret < 0))
1077                 return ret;
1078
1079         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1080         if (ret > 0)
1081                 *ppos = sd.pos;
1082
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(do_splice_direct);
1086
1087 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1088 {
1089         while (pipe->nrbufs == pipe->buffers) {
1090                 if (flags & SPLICE_F_NONBLOCK)
1091                         return -EAGAIN;
1092                 if (signal_pending(current))
1093                         return -ERESTARTSYS;
1094                 pipe->waiting_writers++;
1095                 pipe_wait(pipe);
1096                 pipe->waiting_writers--;
1097         }
1098         return 0;
1099 }
1100
1101 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1102                                struct pipe_inode_info *opipe,
1103                                size_t len, unsigned int flags);
1104
1105 /*
1106  * Determine where to splice to/from.
1107  */
1108 static long do_splice(struct file *in, loff_t __user *off_in,
1109                       struct file *out, loff_t __user *off_out,
1110                       size_t len, unsigned int flags)
1111 {
1112         struct pipe_inode_info *ipipe;
1113         struct pipe_inode_info *opipe;
1114         loff_t offset;
1115         long ret;
1116
1117         ipipe = get_pipe_info(in);
1118         opipe = get_pipe_info(out);
1119
1120         if (ipipe && opipe) {
1121                 if (off_in || off_out)
1122                         return -ESPIPE;
1123
1124                 if (!(in->f_mode & FMODE_READ))
1125                         return -EBADF;
1126
1127                 if (!(out->f_mode & FMODE_WRITE))
1128                         return -EBADF;
1129
1130                 /* Splicing to self would be fun, but... */
1131                 if (ipipe == opipe)
1132                         return -EINVAL;
1133
1134                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1135         }
1136
1137         if (ipipe) {
1138                 if (off_in)
1139                         return -ESPIPE;
1140                 if (off_out) {
1141                         if (!(out->f_mode & FMODE_PWRITE))
1142                                 return -EINVAL;
1143                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1144                                 return -EFAULT;
1145                 } else {
1146                         offset = out->f_pos;
1147                 }
1148
1149                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1150                         return -EBADF;
1151
1152                 if (unlikely(out->f_flags & O_APPEND))
1153                         return -EINVAL;
1154
1155                 ret = rw_verify_area(WRITE, out, &offset, len);
1156                 if (unlikely(ret < 0))
1157                         return ret;
1158
1159                 file_start_write(out);
1160                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1161                 file_end_write(out);
1162
1163                 if (!off_out)
1164                         out->f_pos = offset;
1165                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1166                         ret = -EFAULT;
1167
1168                 return ret;
1169         }
1170
1171         if (opipe) {
1172                 if (off_out)
1173                         return -ESPIPE;
1174                 if (off_in) {
1175                         if (!(in->f_mode & FMODE_PREAD))
1176                                 return -EINVAL;
1177                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1178                                 return -EFAULT;
1179                 } else {
1180                         offset = in->f_pos;
1181                 }
1182
1183                 pipe_lock(opipe);
1184                 ret = wait_for_space(opipe, flags);
1185                 if (!ret)
1186                         ret = do_splice_to(in, &offset, opipe, len, flags);
1187                 pipe_unlock(opipe);
1188                 if (ret > 0)
1189                         wakeup_pipe_readers(opipe);
1190                 if (!off_in)
1191                         in->f_pos = offset;
1192                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1193                         ret = -EFAULT;
1194
1195                 return ret;
1196         }
1197
1198         return -EINVAL;
1199 }
1200
1201 static int iter_to_pipe(struct iov_iter *from,
1202                         struct pipe_inode_info *pipe,
1203                         unsigned flags)
1204 {
1205         struct pipe_buffer buf = {
1206                 .ops = &user_page_pipe_buf_ops,
1207                 .flags = flags
1208         };
1209         size_t total = 0;
1210         int ret = 0;
1211         bool failed = false;
1212
1213         while (iov_iter_count(from) && !failed) {
1214                 struct page *pages[16];
1215                 ssize_t copied;
1216                 size_t start;
1217                 int n;
1218
1219                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1220                 if (copied <= 0) {
1221                         ret = copied;
1222                         break;
1223                 }
1224
1225                 for (n = 0; copied; n++, start = 0) {
1226                         int size = min_t(int, copied, PAGE_SIZE - start);
1227                         if (!failed) {
1228                                 buf.page = pages[n];
1229                                 buf.offset = start;
1230                                 buf.len = size;
1231                                 ret = add_to_pipe(pipe, &buf);
1232                                 if (unlikely(ret < 0)) {
1233                                         failed = true;
1234                                 } else {
1235                                         iov_iter_advance(from, ret);
1236                                         total += ret;
1237                                 }
1238                         } else {
1239                                 put_page(pages[n]);
1240                         }
1241                         copied -= size;
1242                 }
1243         }
1244         return total ? total : ret;
1245 }
1246
1247 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1248                         struct splice_desc *sd)
1249 {
1250         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1251         return n == sd->len ? n : -EFAULT;
1252 }
1253
1254 /*
1255  * For lack of a better implementation, implement vmsplice() to userspace
1256  * as a simple copy of the pipes pages to the user iov.
1257  */
1258 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1259                              unsigned long nr_segs, unsigned int flags)
1260 {
1261         struct pipe_inode_info *pipe;
1262         struct splice_desc sd;
1263         long ret;
1264         struct iovec iovstack[UIO_FASTIOV];
1265         struct iovec *iov = iovstack;
1266         struct iov_iter iter;
1267
1268         pipe = get_pipe_info(file);
1269         if (!pipe)
1270                 return -EBADF;
1271
1272         ret = import_iovec(READ, uiov, nr_segs,
1273                            ARRAY_SIZE(iovstack), &iov, &iter);
1274         if (ret < 0)
1275                 return ret;
1276
1277         sd.total_len = iov_iter_count(&iter);
1278         sd.len = 0;
1279         sd.flags = flags;
1280         sd.u.data = &iter;
1281         sd.pos = 0;
1282
1283         if (sd.total_len) {
1284                 pipe_lock(pipe);
1285                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1286                 pipe_unlock(pipe);
1287         }
1288
1289         kfree(iov);
1290         return ret;
1291 }
1292
1293 /*
1294  * vmsplice splices a user address range into a pipe. It can be thought of
1295  * as splice-from-memory, where the regular splice is splice-from-file (or
1296  * to file). In both cases the output is a pipe, naturally.
1297  */
1298 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1299                              unsigned long nr_segs, unsigned int flags)
1300 {
1301         struct pipe_inode_info *pipe;
1302         struct iovec iovstack[UIO_FASTIOV];
1303         struct iovec *iov = iovstack;
1304         struct iov_iter from;
1305         long ret;
1306         unsigned buf_flag = 0;
1307
1308         if (flags & SPLICE_F_GIFT)
1309                 buf_flag = PIPE_BUF_FLAG_GIFT;
1310
1311         pipe = get_pipe_info(file);
1312         if (!pipe)
1313                 return -EBADF;
1314
1315         ret = import_iovec(WRITE, uiov, nr_segs,
1316                            ARRAY_SIZE(iovstack), &iov, &from);
1317         if (ret < 0)
1318                 return ret;
1319
1320         pipe_lock(pipe);
1321         ret = wait_for_space(pipe, flags);
1322         if (!ret)
1323                 ret = iter_to_pipe(&from, pipe, buf_flag);
1324         pipe_unlock(pipe);
1325         if (ret > 0)
1326                 wakeup_pipe_readers(pipe);
1327         kfree(iov);
1328         return ret;
1329 }
1330
1331 /*
1332  * Note that vmsplice only really supports true splicing _from_ user memory
1333  * to a pipe, not the other way around. Splicing from user memory is a simple
1334  * operation that can be supported without any funky alignment restrictions
1335  * or nasty vm tricks. We simply map in the user memory and fill them into
1336  * a pipe. The reverse isn't quite as easy, though. There are two possible
1337  * solutions for that:
1338  *
1339  *      - memcpy() the data internally, at which point we might as well just
1340  *        do a regular read() on the buffer anyway.
1341  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1342  *        has restriction limitations on both ends of the pipe).
1343  *
1344  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1345  *
1346  */
1347 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1348                 unsigned long, nr_segs, unsigned int, flags)
1349 {
1350         struct fd f;
1351         long error;
1352
1353         if (unlikely(nr_segs > UIO_MAXIOV))
1354                 return -EINVAL;
1355         else if (unlikely(!nr_segs))
1356                 return 0;
1357
1358         error = -EBADF;
1359         f = fdget(fd);
1360         if (f.file) {
1361                 if (f.file->f_mode & FMODE_WRITE)
1362                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1363                 else if (f.file->f_mode & FMODE_READ)
1364                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1365
1366                 fdput(f);
1367         }
1368
1369         return error;
1370 }
1371
1372 #ifdef CONFIG_COMPAT
1373 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1374                     unsigned int, nr_segs, unsigned int, flags)
1375 {
1376         unsigned i;
1377         struct iovec __user *iov;
1378         if (nr_segs > UIO_MAXIOV)
1379                 return -EINVAL;
1380         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1381         for (i = 0; i < nr_segs; i++) {
1382                 struct compat_iovec v;
1383                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1384                     get_user(v.iov_len, &iov32[i].iov_len) ||
1385                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1386                     put_user(v.iov_len, &iov[i].iov_len))
1387                         return -EFAULT;
1388         }
1389         return sys_vmsplice(fd, iov, nr_segs, flags);
1390 }
1391 #endif
1392
1393 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1394                 int, fd_out, loff_t __user *, off_out,
1395                 size_t, len, unsigned int, flags)
1396 {
1397         struct fd in, out;
1398         long error;
1399
1400         if (unlikely(!len))
1401                 return 0;
1402
1403         error = -EBADF;
1404         in = fdget(fd_in);
1405         if (in.file) {
1406                 if (in.file->f_mode & FMODE_READ) {
1407                         out = fdget(fd_out);
1408                         if (out.file) {
1409                                 if (out.file->f_mode & FMODE_WRITE)
1410                                         error = do_splice(in.file, off_in,
1411                                                           out.file, off_out,
1412                                                           len, flags);
1413                                 fdput(out);
1414                         }
1415                 }
1416                 fdput(in);
1417         }
1418         return error;
1419 }
1420
1421 /*
1422  * Make sure there's data to read. Wait for input if we can, otherwise
1423  * return an appropriate error.
1424  */
1425 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1426 {
1427         int ret;
1428
1429         /*
1430          * Check ->nrbufs without the inode lock first. This function
1431          * is speculative anyways, so missing one is ok.
1432          */
1433         if (pipe->nrbufs)
1434                 return 0;
1435
1436         ret = 0;
1437         pipe_lock(pipe);
1438
1439         while (!pipe->nrbufs) {
1440                 if (signal_pending(current)) {
1441                         ret = -ERESTARTSYS;
1442                         break;
1443                 }
1444                 if (!pipe->writers)
1445                         break;
1446                 if (!pipe->waiting_writers) {
1447                         if (flags & SPLICE_F_NONBLOCK) {
1448                                 ret = -EAGAIN;
1449                                 break;
1450                         }
1451                 }
1452                 pipe_wait(pipe);
1453         }
1454
1455         pipe_unlock(pipe);
1456         return ret;
1457 }
1458
1459 /*
1460  * Make sure there's writeable room. Wait for room if we can, otherwise
1461  * return an appropriate error.
1462  */
1463 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1464 {
1465         int ret;
1466
1467         /*
1468          * Check ->nrbufs without the inode lock first. This function
1469          * is speculative anyways, so missing one is ok.
1470          */
1471         if (pipe->nrbufs < pipe->buffers)
1472                 return 0;
1473
1474         ret = 0;
1475         pipe_lock(pipe);
1476
1477         while (pipe->nrbufs >= pipe->buffers) {
1478                 if (!pipe->readers) {
1479                         send_sig(SIGPIPE, current, 0);
1480                         ret = -EPIPE;
1481                         break;
1482                 }
1483                 if (flags & SPLICE_F_NONBLOCK) {
1484                         ret = -EAGAIN;
1485                         break;
1486                 }
1487                 if (signal_pending(current)) {
1488                         ret = -ERESTARTSYS;
1489                         break;
1490                 }
1491                 pipe->waiting_writers++;
1492                 pipe_wait(pipe);
1493                 pipe->waiting_writers--;
1494         }
1495
1496         pipe_unlock(pipe);
1497         return ret;
1498 }
1499
1500 /*
1501  * Splice contents of ipipe to opipe.
1502  */
1503 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1504                                struct pipe_inode_info *opipe,
1505                                size_t len, unsigned int flags)
1506 {
1507         struct pipe_buffer *ibuf, *obuf;
1508         int ret = 0, nbuf;
1509         bool input_wakeup = false;
1510
1511
1512 retry:
1513         ret = ipipe_prep(ipipe, flags);
1514         if (ret)
1515                 return ret;
1516
1517         ret = opipe_prep(opipe, flags);
1518         if (ret)
1519                 return ret;
1520
1521         /*
1522          * Potential ABBA deadlock, work around it by ordering lock
1523          * grabbing by pipe info address. Otherwise two different processes
1524          * could deadlock (one doing tee from A -> B, the other from B -> A).
1525          */
1526         pipe_double_lock(ipipe, opipe);
1527
1528         do {
1529                 if (!opipe->readers) {
1530                         send_sig(SIGPIPE, current, 0);
1531                         if (!ret)
1532                                 ret = -EPIPE;
1533                         break;
1534                 }
1535
1536                 if (!ipipe->nrbufs && !ipipe->writers)
1537                         break;
1538
1539                 /*
1540                  * Cannot make any progress, because either the input
1541                  * pipe is empty or the output pipe is full.
1542                  */
1543                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1544                         /* Already processed some buffers, break */
1545                         if (ret)
1546                                 break;
1547
1548                         if (flags & SPLICE_F_NONBLOCK) {
1549                                 ret = -EAGAIN;
1550                                 break;
1551                         }
1552
1553                         /*
1554                          * We raced with another reader/writer and haven't
1555                          * managed to process any buffers.  A zero return
1556                          * value means EOF, so retry instead.
1557                          */
1558                         pipe_unlock(ipipe);
1559                         pipe_unlock(opipe);
1560                         goto retry;
1561                 }
1562
1563                 ibuf = ipipe->bufs + ipipe->curbuf;
1564                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1565                 obuf = opipe->bufs + nbuf;
1566
1567                 if (len >= ibuf->len) {
1568                         /*
1569                          * Simply move the whole buffer from ipipe to opipe
1570                          */
1571                         *obuf = *ibuf;
1572                         ibuf->ops = NULL;
1573                         opipe->nrbufs++;
1574                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1575                         ipipe->nrbufs--;
1576                         input_wakeup = true;
1577                 } else {
1578                         /*
1579                          * Get a reference to this pipe buffer,
1580                          * so we can copy the contents over.
1581                          */
1582                         pipe_buf_get(ipipe, ibuf);
1583                         *obuf = *ibuf;
1584
1585                         /*
1586                          * Don't inherit the gift flag, we need to
1587                          * prevent multiple steals of this page.
1588                          */
1589                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1590
1591                         obuf->len = len;
1592                         opipe->nrbufs++;
1593                         ibuf->offset += obuf->len;
1594                         ibuf->len -= obuf->len;
1595                 }
1596                 ret += obuf->len;
1597                 len -= obuf->len;
1598         } while (len);
1599
1600         pipe_unlock(ipipe);
1601         pipe_unlock(opipe);
1602
1603         /*
1604          * If we put data in the output pipe, wakeup any potential readers.
1605          */
1606         if (ret > 0)
1607                 wakeup_pipe_readers(opipe);
1608
1609         if (input_wakeup)
1610                 wakeup_pipe_writers(ipipe);
1611
1612         return ret;
1613 }
1614
1615 /*
1616  * Link contents of ipipe to opipe.
1617  */
1618 static int link_pipe(struct pipe_inode_info *ipipe,
1619                      struct pipe_inode_info *opipe,
1620                      size_t len, unsigned int flags)
1621 {
1622         struct pipe_buffer *ibuf, *obuf;
1623         int ret = 0, i = 0, nbuf;
1624
1625         /*
1626          * Potential ABBA deadlock, work around it by ordering lock
1627          * grabbing by pipe info address. Otherwise two different processes
1628          * could deadlock (one doing tee from A -> B, the other from B -> A).
1629          */
1630         pipe_double_lock(ipipe, opipe);
1631
1632         do {
1633                 if (!opipe->readers) {
1634                         send_sig(SIGPIPE, current, 0);
1635                         if (!ret)
1636                                 ret = -EPIPE;
1637                         break;
1638                 }
1639
1640                 /*
1641                  * If we have iterated all input buffers or ran out of
1642                  * output room, break.
1643                  */
1644                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1645                         break;
1646
1647                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1648                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1649
1650                 /*
1651                  * Get a reference to this pipe buffer,
1652                  * so we can copy the contents over.
1653                  */
1654                 pipe_buf_get(ipipe, ibuf);
1655
1656                 obuf = opipe->bufs + nbuf;
1657                 *obuf = *ibuf;
1658
1659                 /*
1660                  * Don't inherit the gift flag, we need to
1661                  * prevent multiple steals of this page.
1662                  */
1663                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1664
1665                 if (obuf->len > len)
1666                         obuf->len = len;
1667
1668                 opipe->nrbufs++;
1669                 ret += obuf->len;
1670                 len -= obuf->len;
1671                 i++;
1672         } while (len);
1673
1674         /*
1675          * return EAGAIN if we have the potential of some data in the
1676          * future, otherwise just return 0
1677          */
1678         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1679                 ret = -EAGAIN;
1680
1681         pipe_unlock(ipipe);
1682         pipe_unlock(opipe);
1683
1684         /*
1685          * If we put data in the output pipe, wakeup any potential readers.
1686          */
1687         if (ret > 0)
1688                 wakeup_pipe_readers(opipe);
1689
1690         return ret;
1691 }
1692
1693 /*
1694  * This is a tee(1) implementation that works on pipes. It doesn't copy
1695  * any data, it simply references the 'in' pages on the 'out' pipe.
1696  * The 'flags' used are the SPLICE_F_* variants, currently the only
1697  * applicable one is SPLICE_F_NONBLOCK.
1698  */
1699 static long do_tee(struct file *in, struct file *out, size_t len,
1700                    unsigned int flags)
1701 {
1702         struct pipe_inode_info *ipipe = get_pipe_info(in);
1703         struct pipe_inode_info *opipe = get_pipe_info(out);
1704         int ret = -EINVAL;
1705
1706         /*
1707          * Duplicate the contents of ipipe to opipe without actually
1708          * copying the data.
1709          */
1710         if (ipipe && opipe && ipipe != opipe) {
1711                 /*
1712                  * Keep going, unless we encounter an error. The ipipe/opipe
1713                  * ordering doesn't really matter.
1714                  */
1715                 ret = ipipe_prep(ipipe, flags);
1716                 if (!ret) {
1717                         ret = opipe_prep(opipe, flags);
1718                         if (!ret)
1719                                 ret = link_pipe(ipipe, opipe, len, flags);
1720                 }
1721         }
1722
1723         return ret;
1724 }
1725
1726 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1727 {
1728         struct fd in;
1729         int error;
1730
1731         if (unlikely(!len))
1732                 return 0;
1733
1734         error = -EBADF;
1735         in = fdget(fdin);
1736         if (in.file) {
1737                 if (in.file->f_mode & FMODE_READ) {
1738                         struct fd out = fdget(fdout);
1739                         if (out.file) {
1740                                 if (out.file->f_mode & FMODE_WRITE)
1741                                         error = do_tee(in.file, out.file,
1742                                                         len, flags);
1743                                 fdput(out);
1744                         }
1745                 }
1746                 fdput(in);
1747         }
1748
1749         return error;
1750 }