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 "internal.h"
  37
  38 /*
  39  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  40  * a vm helper function, it's already simplified quite a bit by the
  41  * addition of remove_mapping(). If success is returned, the caller may
  42  * attempt to reuse this page for another destination.
  43  */
  44 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  45                                      struct pipe_buffer *buf)
  46 {
  47         struct page *page = buf->page;
  48         struct address_space *mapping;
  49
  50         lock_page(page);
  51
  52         mapping = page_mapping(page);
  53         if (mapping) {
  54                 WARN_ON(!PageUptodate(page));
  55
  56                 /*
  57                  * At least for ext2 with nobh option, we need to wait on
  58                  * writeback completing on this page, since we'll remove it
  59                  * from the pagecache.  Otherwise truncate wont wait on the
  60                  * page, allowing the disk blocks to be reused by someone else
  61                  * before we actually wrote our data to them. fs corruption
  62                  * ensues.
  63                  */
  64                 wait_on_page_writeback(page);
  65
  66                 if (page_has_private(page) &&
  67                     !try_to_release_page(page, GFP_KERNEL))
  68                         goto out_unlock;
  69
  70                 /*
  71                  * If we succeeded in removing the mapping, set LRU flag
  72                  * and return good.
  73                  */
  74                 if (remove_mapping(mapping, page)) {
  75                         buf->flags |= PIPE_BUF_FLAG_LRU;
  76                         return 0;
  77                 }
  78         }
  79
  80         /*
  81          * Raced with truncate or failed to remove page from current
  82          * address space, unlock and return failure.
  83          */
  84 out_unlock:
  85         unlock_page(page);
  86         return 1;
  87 }
  88
  89 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  90                                         struct pipe_buffer *buf)
  91 {
  92         put_page(buf->page);
  93         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  94 }
  95
  96 /*
  97  * Check whether the contents of buf is OK to access. Since the content
  98  * is a page cache page, IO may be in flight.
  99  */
 100 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 101                                        struct pipe_buffer *buf)
 102 {
 103         struct page *page = buf->page;
 104         int err;
 105
 106         if (!PageUptodate(page)) {
 107                 lock_page(page);
 108
 109                 /*
 110                  * Page got truncated/unhashed. This will cause a 0-byte
 111                  * splice, if this is the first page.
 112                  */
 113                 if (!page->mapping) {
 114                         err = -ENODATA;
 115                         goto error;
 116                 }
 117
 118                 /*
 119                  * Uh oh, read-error from disk.
 120                  */
 121                 if (!PageUptodate(page)) {
 122                         err = -EIO;
 123                         goto error;
 124                 }
 125
 126                 /*
 127                  * Page is ok afterall, we are done.
 128                  */
 129                 unlock_page(page);
 130         }
 131
 132         return 0;
 133 error:
 134         unlock_page(page);
 135         return err;
 136 }
 137
 138 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 139         .can_merge = 0,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 158         .confirm = generic_pipe_buf_confirm,
 159         .release = page_cache_pipe_buf_release,
 160         .steal = user_page_pipe_buf_steal,
 161         .get = generic_pipe_buf_get,
 162 };
 163
 164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 165 {
 166         smp_mb();
 167         if (waitqueue_active(&pipe->wait))
 168                 wake_up_interruptible(&pipe->wait);
 169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 170 }
 171
 172 /**
 173  * splice_to_pipe - fill passed data into a pipe
 174  * @pipe:       pipe to fill
 175  * @spd:        data to fill
 176  *
 177  * Description:
 178  *    @spd contains a map of pages and len/offset tuples, along with
 179  *    the struct pipe_buf_operations associated with these pages. This
 180  *    function will link that data to the pipe.
 181  *
 182  */
 183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 184                        struct splice_pipe_desc *spd)
 185 {
 186         unsigned int spd_pages = spd->nr_pages;
 187         int ret = 0, page_nr = 0;
 188
 189         if (!spd_pages)
 190                 return 0;
 191
 192         if (unlikely(!pipe->readers)) {
 193                 send_sig(SIGPIPE, current, 0);
 194                 ret = -EPIPE;
 195                 goto out;
 196         }
 197
 198         while (pipe->nrbufs < pipe->buffers) {
 199                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 200                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 201
 202                 buf->page = spd->pages[page_nr];
 203                 buf->offset = spd->partial[page_nr].offset;
 204                 buf->len = spd->partial[page_nr].len;
 205                 buf->private = spd->partial[page_nr].private;
 206                 buf->ops = spd->ops;
 207
 208                 pipe->nrbufs++;
 209                 page_nr++;
 210                 ret += buf->len;
 211
 212                 if (!--spd->nr_pages)
 213                         break;
 214         }
 215
 216         if (!ret)
 217                 ret = -EAGAIN;
 218
 219 out:
 220         while (page_nr < spd_pages)
 221                 spd->spd_release(spd, page_nr++);
 222
 223         return ret;
 224 }
 225 EXPORT_SYMBOL_GPL(splice_to_pipe);
 226
 227 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 228 {
 229         int ret;
 230
 231         if (unlikely(!pipe->readers)) {
 232                 send_sig(SIGPIPE, current, 0);
 233                 ret = -EPIPE;
 234         } else if (pipe->nrbufs == pipe->buffers) {
 235                 ret = -EAGAIN;
 236         } else {
 237                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 238                 pipe->bufs[newbuf] = *buf;
 239                 pipe->nrbufs++;
 240                 return buf->len;
 241         }
 242         pipe_buf_release(pipe, buf);
 243         return ret;
 244 }
 245 EXPORT_SYMBOL(add_to_pipe);
 246
 247 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 248 {
 249         put_page(spd->pages[i]);
 250 }
 251
 252 /*
 253  * Check if we need to grow the arrays holding pages and partial page
 254  * descriptions.
 255  */
 256 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 257 {
 258         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 259
 260         spd->nr_pages_max = buffers;
 261         if (buffers <= PIPE_DEF_BUFFERS)
 262                 return 0;
 263
 264         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 265         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 266
 267         if (spd->pages && spd->partial)
 268                 return 0;
 269
 270         kfree(spd->pages);
 271         kfree(spd->partial);
 272         return -ENOMEM;
 273 }
 274
 275 void splice_shrink_spd(struct splice_pipe_desc *spd)
 276 {
 277         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 278                 return;
 279
 280         kfree(spd->pages);
 281         kfree(spd->partial);
 282 }
 283
 284 /**
 285  * generic_file_splice_read - splice data from file to a pipe
 286  * @in:         file to splice from
 287  * @ppos:       position in @in
 288  * @pipe:       pipe to splice to
 289  * @len:        number of bytes to splice
 290  * @flags:      splice modifier flags
 291  *
 292  * Description:
 293  *    Will read pages from given file and fill them into a pipe. Can be
 294  *    used as long as it has more or less sane ->read_iter().
 295  *
 296  */
 297 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 298                                  struct pipe_inode_info *pipe, size_t len,
 299                                  unsigned int flags)
 300 {
 301         struct iov_iter to;
 302         struct kiocb kiocb;
 303         loff_t isize;
 304         int idx, ret;
 305
 306         isize = i_size_read(in->f_mapping->host);
 307         if (unlikely(*ppos >= isize))
 308                 return 0;
 309
 310         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
 311         idx = to.idx;
 312         init_sync_kiocb(&kiocb, in);
 313         kiocb.ki_pos = *ppos;
 314         ret = in->f_op->read_iter(&kiocb, &to);
 315         if (ret > 0) {
 316                 *ppos = kiocb.ki_pos;
 317                 file_accessed(in);
 318         } else if (ret < 0) {
 319                 to.idx = idx;
 320                 to.iov_offset = 0;
 321                 iov_iter_advance(&to, 0); /* to free what was emitted */
 322                 /*
 323                  * callers of ->splice_read() expect -EAGAIN on
 324                  * "can't put anything in there", rather than -EFAULT.
 325                  */
 326                 if (ret == -EFAULT)
 327                         ret = -EAGAIN;
 328         }
 329
 330         return ret;
 331 }
 332 EXPORT_SYMBOL(generic_file_splice_read);
 333
 334 const struct pipe_buf_operations default_pipe_buf_ops = {
 335         .can_merge = 0,
 336         .confirm = generic_pipe_buf_confirm,
 337         .release = generic_pipe_buf_release,
 338         .steal = generic_pipe_buf_steal,
 339         .get = generic_pipe_buf_get,
 340 };
 341
 342 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 343                                     struct pipe_buffer *buf)
 344 {
 345         return 1;
 346 }
 347
 348 /* Pipe buffer operations for a socket and similar. */
 349 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 350         .can_merge = 0,
 351         .confirm = generic_pipe_buf_confirm,
 352         .release = generic_pipe_buf_release,
 353         .steal = generic_pipe_buf_nosteal,
 354         .get = generic_pipe_buf_get,
 355 };
 356 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 357
 358 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 359                             unsigned long vlen, loff_t offset)
 360 {
 361         mm_segment_t old_fs;
 362         loff_t pos = offset;
 363         ssize_t res;
 364
 365         old_fs = get_fs();
 366         set_fs(get_ds());
 367         /* The cast to a user pointer is valid due to the set_fs() */
 368         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 369         set_fs(old_fs);
 370
 371         return res;
 372 }
 373
 374 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 375                             loff_t pos)
 376 {
 377         mm_segment_t old_fs;
 378         ssize_t res;
 379
 380         old_fs = get_fs();
 381         set_fs(get_ds());
 382         /* The cast to a user pointer is valid due to the set_fs() */
 383         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 384         set_fs(old_fs);
 385
 386         return res;
 387 }
 388 EXPORT_SYMBOL(kernel_write);
 389
 390 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 391                                  struct pipe_inode_info *pipe, size_t len,
 392                                  unsigned int flags)
 393 {
 394         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 395         struct iov_iter to;
 396         struct page **pages;
 397         unsigned int nr_pages;
 398         size_t offset, dummy, copied = 0;
 399         ssize_t res;
 400         int i;
 401
 402         if (pipe->nrbufs == pipe->buffers)
 403                 return -EAGAIN;
 404
 405         /*
 406          * Try to keep page boundaries matching to source pagecache ones -
 407          * it probably won't be much help, but...
 408          */
 409         offset = *ppos & ~PAGE_MASK;
 410
 411         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 412
 413         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
 414         if (res <= 0)
 415                 return -ENOMEM;
 416
 417         nr_pages = res / PAGE_SIZE;
 418
 419         vec = __vec;
 420         if (nr_pages > PIPE_DEF_BUFFERS) {
 421                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 422                 if (unlikely(!vec)) {
 423                         res = -ENOMEM;
 424                         goto out;
 425                 }
 426         }
 427
 428         pipe->bufs[to.idx].offset = offset;
 429         pipe->bufs[to.idx].len -= offset;
 430
 431         for (i = 0; i < nr_pages; i++) {
 432                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 433                 vec[i].iov_base = page_address(pages[i]) + offset;
 434                 vec[i].iov_len = this_len;
 435                 len -= this_len;
 436                 offset = 0;
 437         }
 438
 439         res = kernel_readv(in, vec, nr_pages, *ppos);
 440         if (res > 0) {
 441                 copied = res;
 442                 *ppos += res;
 443         }
 444
 445         if (vec != __vec)
 446                 kfree(vec);
 447 out:
 448         for (i = 0; i < nr_pages; i++)
 449                 put_page(pages[i]);
 450         kvfree(pages);
 451         iov_iter_advance(&to, copied);  /* truncates and discards */
 452         return res;
 453 }
 454
 455 /*
 456  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 457  * using sendpage(). Return the number of bytes sent.
 458  */
 459 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 460                             struct pipe_buffer *buf, struct splice_desc *sd)
 461 {
 462         struct file *file = sd->u.file;
 463         loff_t pos = sd->pos;
 464         int more;
 465
 466         if (!likely(file->f_op->sendpage))
 467                 return -EINVAL;
 468
 469         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 470
 471         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 472                 more |= MSG_SENDPAGE_NOTLAST;
 473
 474         return file->f_op->sendpage(file, buf->page, buf->offset,
 475                                     sd->len, &pos, more);
 476 }
 477
 478 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 479 {
 480         smp_mb();
 481         if (waitqueue_active(&pipe->wait))
 482                 wake_up_interruptible(&pipe->wait);
 483         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 484 }
 485
 486 /**
 487  * splice_from_pipe_feed - feed available data from a pipe to a file
 488  * @pipe:       pipe to splice from
 489  * @sd:         information to @actor
 490  * @actor:      handler that splices the data
 491  *
 492  * Description:
 493  *    This function loops over the pipe and calls @actor to do the
 494  *    actual moving of a single struct pipe_buffer to the desired
 495  *    destination.  It returns when there's no more buffers left in
 496  *    the pipe or if the requested number of bytes (@sd->total_len)
 497  *    have been copied.  It returns a positive number (one) if the
 498  *    pipe needs to be filled with more data, zero if the required
 499  *    number of bytes have been copied and -errno on error.
 500  *
 501  *    This, together with splice_from_pipe_{begin,end,next}, may be
 502  *    used to implement the functionality of __splice_from_pipe() when
 503  *    locking is required around copying the pipe buffers to the
 504  *    destination.
 505  */
 506 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 507                           splice_actor *actor)
 508 {
 509         int ret;
 510
 511         while (pipe->nrbufs) {
 512                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 513
 514                 sd->len = buf->len;
 515                 if (sd->len > sd->total_len)
 516                         sd->len = sd->total_len;
 517
 518                 ret = pipe_buf_confirm(pipe, buf);
 519                 if (unlikely(ret)) {
 520                         if (ret == -ENODATA)
 521                                 ret = 0;
 522                         return ret;
 523                 }
 524
 525                 ret = actor(pipe, buf, sd);
 526                 if (ret <= 0)
 527                         return ret;
 528
 529                 buf->offset += ret;
 530                 buf->len -= ret;
 531
 532                 sd->num_spliced += ret;
 533                 sd->len -= ret;
 534                 sd->pos += ret;
 535                 sd->total_len -= ret;
 536
 537                 if (!buf->len) {
 538                         pipe_buf_release(pipe, buf);
 539                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 540                         pipe->nrbufs--;
 541                         if (pipe->files)
 542                                 sd->need_wakeup = true;
 543                 }
 544
 545                 if (!sd->total_len)
 546                         return 0;
 547         }
 548
 549         return 1;
 550 }
 551
 552 /**
 553  * splice_from_pipe_next - wait for some data to splice from
 554  * @pipe:       pipe to splice from
 555  * @sd:         information about the splice operation
 556  *
 557  * Description:
 558  *    This function will wait for some data and return a positive
 559  *    value (one) if pipe buffers are available.  It will return zero
 560  *    or -errno if no more data needs to be spliced.
 561  */
 562 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 563 {
 564         /*
 565          * Check for signal early to make process killable when there are
 566          * always buffers available
 567          */
 568         if (signal_pending(current))
 569                 return -ERESTARTSYS;
 570
 571         while (!pipe->nrbufs) {
 572                 if (!pipe->writers)
 573                         return 0;
 574
 575                 if (!pipe->waiting_writers && sd->num_spliced)
 576                         return 0;
 577
 578                 if (sd->flags & SPLICE_F_NONBLOCK)
 579                         return -EAGAIN;
 580
 581                 if (signal_pending(current))
 582                         return -ERESTARTSYS;
 583
 584                 if (sd->need_wakeup) {
 585                         wakeup_pipe_writers(pipe);
 586                         sd->need_wakeup = false;
 587                 }
 588
 589                 pipe_wait(pipe);
 590         }
 591
 592         return 1;
 593 }
 594
 595 /**
 596  * splice_from_pipe_begin - start splicing from pipe
 597  * @sd:         information about the splice operation
 598  *
 599  * Description:
 600  *    This function should be called before a loop containing
 601  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 602  *    initialize the necessary fields of @sd.
 603  */
 604 static void splice_from_pipe_begin(struct splice_desc *sd)
 605 {
 606         sd->num_spliced = 0;
 607         sd->need_wakeup = false;
 608 }
 609
 610 /**
 611  * splice_from_pipe_end - finish splicing from pipe
 612  * @pipe:       pipe to splice from
 613  * @sd:         information about the splice operation
 614  *
 615  * Description:
 616  *    This function will wake up pipe writers if necessary.  It should
 617  *    be called after a loop containing splice_from_pipe_next() and
 618  *    splice_from_pipe_feed().
 619  */
 620 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 621 {
 622         if (sd->need_wakeup)
 623                 wakeup_pipe_writers(pipe);
 624 }
 625
 626 /**
 627  * __splice_from_pipe - splice data from a pipe to given actor
 628  * @pipe:       pipe to splice from
 629  * @sd:         information to @actor
 630  * @actor:      handler that splices the data
 631  *
 632  * Description:
 633  *    This function does little more than loop over the pipe and call
 634  *    @actor to do the actual moving of a single struct pipe_buffer to
 635  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 636  *    pipe_to_user.
 637  *
 638  */
 639 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 640                            splice_actor *actor)
 641 {
 642         int ret;
 643
 644         splice_from_pipe_begin(sd);
 645         do {
 646                 cond_resched();
 647                 ret = splice_from_pipe_next(pipe, sd);
 648                 if (ret > 0)
 649                         ret = splice_from_pipe_feed(pipe, sd, actor);
 650         } while (ret > 0);
 651         splice_from_pipe_end(pipe, sd);
 652
 653         return sd->num_spliced ? sd->num_spliced : ret;
 654 }
 655 EXPORT_SYMBOL(__splice_from_pipe);
 656
 657 /**
 658  * splice_from_pipe - splice data from a pipe to a file
 659  * @pipe:       pipe to splice from
 660  * @out:        file to splice to
 661  * @ppos:       position in @out
 662  * @len:        how many bytes to splice
 663  * @flags:      splice modifier flags
 664  * @actor:      handler that splices the data
 665  *
 666  * Description:
 667  *    See __splice_from_pipe. This function locks the pipe inode,
 668  *    otherwise it's identical to __splice_from_pipe().
 669  *
 670  */
 671 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 672                          loff_t *ppos, size_t len, unsigned int flags,
 673                          splice_actor *actor)
 674 {
 675         ssize_t ret;
 676         struct splice_desc sd = {
 677                 .total_len = len,
 678                 .flags = flags,
 679                 .pos = *ppos,
 680                 .u.file = out,
 681         };
 682
 683         pipe_lock(pipe);
 684         ret = __splice_from_pipe(pipe, &sd, actor);
 685         pipe_unlock(pipe);
 686
 687         return ret;
 688 }
 689
 690 /**
 691  * iter_file_splice_write - splice data from a pipe to a file
 692  * @pipe:       pipe info
 693  * @out:        file to write to
 694  * @ppos:       position in @out
 695  * @len:        number of bytes to splice
 696  * @flags:      splice modifier flags
 697  *
 698  * Description:
 699  *    Will either move or copy pages (determined by @flags options) from
 700  *    the given pipe inode to the given file.
 701  *    This one is ->write_iter-based.
 702  *
 703  */
 704 ssize_t
 705 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 706                           loff_t *ppos, size_t len, unsigned int flags)
 707 {
 708         struct splice_desc sd = {
 709                 .total_len = len,
 710                 .flags = flags,
 711                 .pos = *ppos,
 712                 .u.file = out,
 713         };
 714         int nbufs = pipe->buffers;
 715         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 716                                         GFP_KERNEL);
 717         ssize_t ret;
 718
 719         if (unlikely(!array))
 720                 return -ENOMEM;
 721
 722         pipe_lock(pipe);
 723
 724         splice_from_pipe_begin(&sd);
 725         while (sd.total_len) {
 726                 struct iov_iter from;
 727                 size_t left;
 728                 int n, idx;
 729
 730                 ret = splice_from_pipe_next(pipe, &sd);
 731                 if (ret <= 0)
 732                         break;
 733
 734                 if (unlikely(nbufs < pipe->buffers)) {
 735                         kfree(array);
 736                         nbufs = pipe->buffers;
 737                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 738                                         GFP_KERNEL);
 739                         if (!array) {
 740                                 ret = -ENOMEM;
 741                                 break;
 742                         }
 743                 }
 744
 745                 /* build the vector */
 746                 left = sd.total_len;
 747                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 748                         struct pipe_buffer *buf = pipe->bufs + idx;
 749                         size_t this_len = buf->len;
 750
 751                         if (this_len > left)
 752                                 this_len = left;
 753
 754                         if (idx == pipe->buffers - 1)
 755                                 idx = -1;
 756
 757                         ret = pipe_buf_confirm(pipe, buf);
 758                         if (unlikely(ret)) {
 759                                 if (ret == -ENODATA)
 760                                         ret = 0;
 761                                 goto done;
 762                         }
 763
 764                         array[n].bv_page = buf->page;
 765                         array[n].bv_len = this_len;
 766                         array[n].bv_offset = buf->offset;
 767                         left -= this_len;
 768                 }
 769
 770                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 771                               sd.total_len - left);
 772                 ret = vfs_iter_write(out, &from, &sd.pos);
 773                 if (ret <= 0)
 774                         break;
 775
 776                 sd.num_spliced += ret;
 777                 sd.total_len -= ret;
 778                 *ppos = sd.pos;
 779
 780                 /* dismiss the fully eaten buffers, adjust the partial one */
 781                 while (ret) {
 782                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 783                         if (ret >= buf->len) {
 784                                 ret -= buf->len;
 785                                 buf->len = 0;
 786                                 pipe_buf_release(pipe, buf);
 787                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 788                                 pipe->nrbufs--;
 789                                 if (pipe->files)
 790                                         sd.need_wakeup = true;
 791                         } else {
 792                                 buf->offset += ret;
 793                                 buf->len -= ret;
 794                                 ret = 0;
 795                         }
 796                 }
 797         }
 798 done:
 799         kfree(array);
 800         splice_from_pipe_end(pipe, &sd);
 801
 802         pipe_unlock(pipe);
 803
 804         if (sd.num_spliced)
 805                 ret = sd.num_spliced;
 806
 807         return ret;
 808 }
 809
 810 EXPORT_SYMBOL(iter_file_splice_write);
 811
 812 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 813                           struct splice_desc *sd)
 814 {
 815         int ret;
 816         void *data;
 817         loff_t tmp = sd->pos;
 818
 819         data = kmap(buf->page);
 820         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 821         kunmap(buf->page);
 822
 823         return ret;
 824 }
 825
 826 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 827                                          struct file *out, loff_t *ppos,
 828                                          size_t len, unsigned int flags)
 829 {
 830         ssize_t ret;
 831
 832         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 833         if (ret > 0)
 834                 *ppos += ret;
 835
 836         return ret;
 837 }
 838
 839 /**
 840  * generic_splice_sendpage - splice data from a pipe to a socket
 841  * @pipe:       pipe to splice from
 842  * @out:        socket to write to
 843  * @ppos:       position in @out
 844  * @len:        number of bytes to splice
 845  * @flags:      splice modifier flags
 846  *
 847  * Description:
 848  *    Will send @len bytes from the pipe to a network socket. No data copying
 849  *    is involved.
 850  *
 851  */
 852 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 853                                 loff_t *ppos, size_t len, unsigned int flags)
 854 {
 855         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 856 }
 857
 858 EXPORT_SYMBOL(generic_splice_sendpage);
 859
 860 /*
 861  * Attempt to initiate a splice from pipe to file.
 862  */
 863 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 864                            loff_t *ppos, size_t len, unsigned int flags)
 865 {
 866         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 867                                 loff_t *, size_t, unsigned int);
 868
 869         if (out->f_op->splice_write)
 870                 splice_write = out->f_op->splice_write;
 871         else
 872                 splice_write = default_file_splice_write;
 873
 874         return splice_write(pipe, out, ppos, len, flags);
 875 }
 876
 877 /*
 878  * Attempt to initiate a splice from a file to a pipe.
 879  */
 880 static long do_splice_to(struct file *in, loff_t *ppos,
 881                          struct pipe_inode_info *pipe, size_t len,
 882                          unsigned int flags)
 883 {
 884         ssize_t (*splice_read)(struct file *, loff_t *,
 885                                struct pipe_inode_info *, size_t, unsigned int);
 886         int ret;
 887
 888         if (unlikely(!(in->f_mode & FMODE_READ)))
 889                 return -EBADF;
 890
 891         ret = rw_verify_area(READ, in, ppos, len);
 892         if (unlikely(ret < 0))
 893                 return ret;
 894
 895         if (unlikely(len > MAX_RW_COUNT))
 896                 len = MAX_RW_COUNT;
 897
 898         if (in->f_op->splice_read)
 899                 splice_read = in->f_op->splice_read;
 900         else
 901                 splice_read = default_file_splice_read;
 902
 903         return splice_read(in, ppos, pipe, len, flags);
 904 }
 905
 906 /**
 907  * splice_direct_to_actor - splices data directly between two non-pipes
 908  * @in:         file to splice from
 909  * @sd:         actor information on where to splice to
 910  * @actor:      handles the data splicing
 911  *
 912  * Description:
 913  *    This is a special case helper to splice directly between two
 914  *    points, without requiring an explicit pipe. Internally an allocated
 915  *    pipe is cached in the process, and reused during the lifetime of
 916  *    that process.
 917  *
 918  */
 919 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 920                                splice_direct_actor *actor)
 921 {
 922         struct pipe_inode_info *pipe;
 923         long ret, bytes;
 924         umode_t i_mode;
 925         size_t len;
 926         int i, flags, more;
 927
 928         /*
 929          * We require the input being a regular file, as we don't want to
 930          * randomly drop data for eg socket -> socket splicing. Use the
 931          * piped splicing for that!
 932          */
 933         i_mode = file_inode(in)->i_mode;
 934         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 935                 return -EINVAL;
 936
 937         /*
 938          * neither in nor out is a pipe, setup an internal pipe attached to
 939          * 'out' and transfer the wanted data from 'in' to 'out' through that
 940          */
 941         pipe = current->splice_pipe;
 942         if (unlikely(!pipe)) {
 943                 pipe = alloc_pipe_info();
 944                 if (!pipe)
 945                         return -ENOMEM;
 946
 947                 /*
 948                  * We don't have an immediate reader, but we'll read the stuff
 949                  * out of the pipe right after the splice_to_pipe(). So set
 950                  * PIPE_READERS appropriately.
 951                  */
 952                 pipe->readers = 1;
 953
 954                 current->splice_pipe = pipe;
 955         }
 956
 957         /*
 958          * Do the splice.
 959          */
 960         ret = 0;
 961         bytes = 0;
 962         len = sd->total_len;
 963         flags = sd->flags;
 964
 965         /*
 966          * Don't block on output, we have to drain the direct pipe.
 967          */
 968         sd->flags &= ~SPLICE_F_NONBLOCK;
 969         more = sd->flags & SPLICE_F_MORE;
 970
 971         while (len) {
 972                 size_t read_len;
 973                 loff_t pos = sd->pos, prev_pos = pos;
 974
 975                 ret = do_splice_to(in, &pos, pipe, len, flags);
 976                 if (unlikely(ret <= 0))
 977                         goto out_release;
 978
 979                 read_len = ret;
 980                 sd->total_len = read_len;
 981
 982                 /*
 983                  * If more data is pending, set SPLICE_F_MORE
 984                  * If this is the last data and SPLICE_F_MORE was not set
 985                  * initially, clears it.
 986                  */
 987                 if (read_len < len)
 988                         sd->flags |= SPLICE_F_MORE;
 989                 else if (!more)
 990                         sd->flags &= ~SPLICE_F_MORE;
 991                 /*
 992                  * NOTE: nonblocking mode only applies to the input. We
 993                  * must not do the output in nonblocking mode as then we
 994                  * could get stuck data in the internal pipe:
 995                  */
 996                 ret = actor(pipe, sd);
 997                 if (unlikely(ret <= 0)) {
 998                         sd->pos = prev_pos;
 999                         goto out_release;
1000                 }
1001
1002                 bytes += ret;
1003                 len -= ret;
1004                 sd->pos = pos;
1005
1006                 if (ret < read_len) {
1007                         sd->pos = prev_pos + ret;
1008                         goto out_release;
1009                 }
1010         }
1011
1012 done:
1013         pipe->nrbufs = pipe->curbuf = 0;
1014         file_accessed(in);
1015         return bytes;
1016
1017 out_release:
1018         /*
1019          * If we did an incomplete transfer we must release
1020          * the pipe buffers in question:
1021          */
1022         for (i = 0; i < pipe->buffers; i++) {
1023                 struct pipe_buffer *buf = pipe->bufs + i;
1024
1025                 if (buf->ops)
1026                         pipe_buf_release(pipe, buf);
1027         }
1028
1029         if (!bytes)
1030                 bytes = ret;
1031
1032         goto done;
1033 }
1034 EXPORT_SYMBOL(splice_direct_to_actor);
1035
1036 static int direct_splice_actor(struct pipe_inode_info *pipe,
1037                                struct splice_desc *sd)
1038 {
1039         struct file *file = sd->u.file;
1040
1041         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1042                               sd->flags);
1043 }
1044
1045 /**
1046  * do_splice_direct - splices data directly between two files
1047  * @in:         file to splice from
1048  * @ppos:       input file offset
1049  * @out:        file to splice to
1050  * @opos:       output file offset
1051  * @len:        number of bytes to splice
1052  * @flags:      splice modifier flags
1053  *
1054  * Description:
1055  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1056  *    doing it in the application would incur an extra system call
1057  *    (splice in + splice out, as compared to just sendfile()). So this helper
1058  *    can splice directly through a process-private pipe.
1059  *
1060  */
1061 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1062                       loff_t *opos, size_t len, unsigned int flags)
1063 {
1064         struct splice_desc sd = {
1065                 .len            = len,
1066                 .total_len      = len,
1067                 .flags          = flags,
1068                 .pos            = *ppos,
1069                 .u.file         = out,
1070                 .opos           = opos,
1071         };
1072         long ret;
1073
1074         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1075                 return -EBADF;
1076
1077         if (unlikely(out->f_flags & O_APPEND))
1078                 return -EINVAL;
1079
1080         ret = rw_verify_area(WRITE, out, opos, len);
1081         if (unlikely(ret < 0))
1082                 return ret;
1083
1084         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1085         if (ret > 0)
1086                 *ppos = sd.pos;
1087
1088         return ret;
1089 }
1090 EXPORT_SYMBOL(do_splice_direct);
1091
1092 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1093 {
1094         while (pipe->nrbufs == pipe->buffers) {
1095                 if (flags & SPLICE_F_NONBLOCK)
1096                         return -EAGAIN;
1097                 if (signal_pending(current))
1098                         return -ERESTARTSYS;
1099                 pipe->waiting_writers++;
1100                 pipe_wait(pipe);
1101                 pipe->waiting_writers--;
1102         }
1103         return 0;
1104 }
1105
1106 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1107                                struct pipe_inode_info *opipe,
1108                                size_t len, unsigned int flags);
1109
1110 /*
1111  * Determine where to splice to/from.
1112  */
1113 static long do_splice(struct file *in, loff_t __user *off_in,
1114                       struct file *out, loff_t __user *off_out,
1115                       size_t len, unsigned int flags)
1116 {
1117         struct pipe_inode_info *ipipe;
1118         struct pipe_inode_info *opipe;
1119         loff_t offset;
1120         long ret;
1121
1122         ipipe = get_pipe_info(in);
1123         opipe = get_pipe_info(out);
1124
1125         if (ipipe && opipe) {
1126                 if (off_in || off_out)
1127                         return -ESPIPE;
1128
1129                 if (!(in->f_mode & FMODE_READ))
1130                         return -EBADF;
1131
1132                 if (!(out->f_mode & FMODE_WRITE))
1133                         return -EBADF;
1134
1135                 /* Splicing to self would be fun, but... */
1136                 if (ipipe == opipe)
1137                         return -EINVAL;
1138
1139                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1140         }
1141
1142         if (ipipe) {
1143                 if (off_in)
1144                         return -ESPIPE;
1145                 if (off_out) {
1146                         if (!(out->f_mode & FMODE_PWRITE))
1147                                 return -EINVAL;
1148                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1149                                 return -EFAULT;
1150                 } else {
1151                         offset = out->f_pos;
1152                 }
1153
1154                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1155                         return -EBADF;
1156
1157                 if (unlikely(out->f_flags & O_APPEND))
1158                         return -EINVAL;
1159
1160                 ret = rw_verify_area(WRITE, out, &offset, len);
1161                 if (unlikely(ret < 0))
1162                         return ret;
1163
1164                 file_start_write(out);
1165                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1166                 file_end_write(out);
1167
1168                 if (!off_out)
1169                         out->f_pos = offset;
1170                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1171                         ret = -EFAULT;
1172
1173                 return ret;
1174         }
1175
1176         if (opipe) {
1177                 if (off_out)
1178                         return -ESPIPE;
1179                 if (off_in) {
1180                         if (!(in->f_mode & FMODE_PREAD))
1181                                 return -EINVAL;
1182                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1183                                 return -EFAULT;
1184                 } else {
1185                         offset = in->f_pos;
1186                 }
1187
1188                 pipe_lock(opipe);
1189                 ret = wait_for_space(opipe, flags);
1190                 if (!ret)
1191                         ret = do_splice_to(in, &offset, opipe, len, flags);
1192                 pipe_unlock(opipe);
1193                 if (ret > 0)
1194                         wakeup_pipe_readers(opipe);
1195                 if (!off_in)
1196                         in->f_pos = offset;
1197                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1198                         ret = -EFAULT;
1199
1200                 return ret;
1201         }
1202
1203         return -EINVAL;
1204 }
1205
1206 static int iter_to_pipe(struct iov_iter *from,
1207                         struct pipe_inode_info *pipe,
1208                         unsigned flags)
1209 {
1210         struct pipe_buffer buf = {
1211                 .ops = &user_page_pipe_buf_ops,
1212                 .flags = flags
1213         };
1214         size_t total = 0;
1215         int ret = 0;
1216         bool failed = false;
1217
1218         while (iov_iter_count(from) && !failed) {
1219                 struct page *pages[16];
1220                 ssize_t copied;
1221                 size_t start;
1222                 int n;
1223
1224                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1225                 if (copied <= 0) {
1226                         ret = copied;
1227                         break;
1228                 }
1229
1230                 for (n = 0; copied; n++, start = 0) {
1231                         int size = min_t(int, copied, PAGE_SIZE - start);
1232                         if (!failed) {
1233                                 buf.page = pages[n];
1234                                 buf.offset = start;
1235                                 buf.len = size;
1236                                 ret = add_to_pipe(pipe, &buf);
1237                                 if (unlikely(ret < 0)) {
1238                                         failed = true;
1239                                 } else {
1240                                         iov_iter_advance(from, ret);
1241                                         total += ret;
1242                                 }
1243                         } else {
1244                                 put_page(pages[n]);
1245                         }
1246                         copied -= size;
1247                 }
1248         }
1249         return total ? total : ret;
1250 }
1251
1252 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1253                         struct splice_desc *sd)
1254 {
1255         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1256         return n == sd->len ? n : -EFAULT;
1257 }
1258
1259 /*
1260  * For lack of a better implementation, implement vmsplice() to userspace
1261  * as a simple copy of the pipes pages to the user iov.
1262  */
1263 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1264                              unsigned long nr_segs, unsigned int flags)
1265 {
1266         struct pipe_inode_info *pipe;
1267         struct splice_desc sd;
1268         long ret;
1269         struct iovec iovstack[UIO_FASTIOV];
1270         struct iovec *iov = iovstack;
1271         struct iov_iter iter;
1272
1273         pipe = get_pipe_info(file);
1274         if (!pipe)
1275                 return -EBADF;
1276
1277         ret = import_iovec(READ, uiov, nr_segs,
1278                            ARRAY_SIZE(iovstack), &iov, &iter);
1279         if (ret < 0)
1280                 return ret;
1281
1282         sd.total_len = iov_iter_count(&iter);
1283         sd.len = 0;
1284         sd.flags = flags;
1285         sd.u.data = &iter;
1286         sd.pos = 0;
1287
1288         if (sd.total_len) {
1289                 pipe_lock(pipe);
1290                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1291                 pipe_unlock(pipe);
1292         }
1293
1294         kfree(iov);
1295         return ret;
1296 }
1297
1298 /*
1299  * vmsplice splices a user address range into a pipe. It can be thought of
1300  * as splice-from-memory, where the regular splice is splice-from-file (or
1301  * to file). In both cases the output is a pipe, naturally.
1302  */
1303 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1304                              unsigned long nr_segs, unsigned int flags)
1305 {
1306         struct pipe_inode_info *pipe;
1307         struct iovec iovstack[UIO_FASTIOV];
1308         struct iovec *iov = iovstack;
1309         struct iov_iter from;
1310         long ret;
1311         unsigned buf_flag = 0;
1312
1313         if (flags & SPLICE_F_GIFT)
1314                 buf_flag = PIPE_BUF_FLAG_GIFT;
1315
1316         pipe = get_pipe_info(file);
1317         if (!pipe)
1318                 return -EBADF;
1319
1320         ret = import_iovec(WRITE, uiov, nr_segs,
1321                            ARRAY_SIZE(iovstack), &iov, &from);
1322         if (ret < 0)
1323                 return ret;
1324
1325         pipe_lock(pipe);
1326         ret = wait_for_space(pipe, flags);
1327         if (!ret)
1328                 ret = iter_to_pipe(&from, pipe, buf_flag);
1329         pipe_unlock(pipe);
1330         if (ret > 0)
1331                 wakeup_pipe_readers(pipe);
1332         kfree(iov);
1333         return ret;
1334 }
1335
1336 /*
1337  * Note that vmsplice only really supports true splicing _from_ user memory
1338  * to a pipe, not the other way around. Splicing from user memory is a simple
1339  * operation that can be supported without any funky alignment restrictions
1340  * or nasty vm tricks. We simply map in the user memory and fill them into
1341  * a pipe. The reverse isn't quite as easy, though. There are two possible
1342  * solutions for that:
1343  *
1344  *      - memcpy() the data internally, at which point we might as well just
1345  *        do a regular read() on the buffer anyway.
1346  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1347  *        has restriction limitations on both ends of the pipe).
1348  *
1349  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1350  *
1351  */
1352 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1353                 unsigned long, nr_segs, unsigned int, flags)
1354 {
1355         struct fd f;
1356         long error;
1357
1358         if (unlikely(nr_segs > UIO_MAXIOV))
1359                 return -EINVAL;
1360         else if (unlikely(!nr_segs))
1361                 return 0;
1362
1363         error = -EBADF;
1364         f = fdget(fd);
1365         if (f.file) {
1366                 if (f.file->f_mode & FMODE_WRITE)
1367                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1368                 else if (f.file->f_mode & FMODE_READ)
1369                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1370
1371                 fdput(f);
1372         }
1373
1374         return error;
1375 }
1376
1377 #ifdef CONFIG_COMPAT
1378 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1379                     unsigned int, nr_segs, unsigned int, flags)
1380 {
1381         unsigned i;
1382         struct iovec __user *iov;
1383         if (nr_segs > UIO_MAXIOV)
1384                 return -EINVAL;
1385         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1386         for (i = 0; i < nr_segs; i++) {
1387                 struct compat_iovec v;
1388                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1389                     get_user(v.iov_len, &iov32[i].iov_len) ||
1390                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1391                     put_user(v.iov_len, &iov[i].iov_len))
1392                         return -EFAULT;
1393         }
1394         return sys_vmsplice(fd, iov, nr_segs, flags);
1395 }
1396 #endif
1397
1398 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1399                 int, fd_out, loff_t __user *, off_out,
1400                 size_t, len, unsigned int, flags)
1401 {
1402         struct fd in, out;
1403         long error;
1404
1405         if (unlikely(!len))
1406                 return 0;
1407
1408         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                         pipe_buf_get(ipipe, ibuf);
1588                         *obuf = *ibuf;
1589
1590                         /*
1591                          * Don't inherit the gift flag, we need to
1592                          * prevent multiple steals of this page.
1593                          */
1594                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1595
1596                         obuf->len = len;
1597                         opipe->nrbufs++;
1598                         ibuf->offset += obuf->len;
1599                         ibuf->len -= obuf->len;
1600                 }
1601                 ret += obuf->len;
1602                 len -= obuf->len;
1603         } while (len);
1604
1605         pipe_unlock(ipipe);
1606         pipe_unlock(opipe);
1607
1608         /*
1609          * If we put data in the output pipe, wakeup any potential readers.
1610          */
1611         if (ret > 0)
1612                 wakeup_pipe_readers(opipe);
1613
1614         if (input_wakeup)
1615                 wakeup_pipe_writers(ipipe);
1616
1617         return ret;
1618 }
1619
1620 /*
1621  * Link contents of ipipe to opipe.
1622  */
1623 static int link_pipe(struct pipe_inode_info *ipipe,
1624                      struct pipe_inode_info *opipe,
1625                      size_t len, unsigned int flags)
1626 {
1627         struct pipe_buffer *ibuf, *obuf;
1628         int ret = 0, i = 0, nbuf;
1629
1630         /*
1631          * Potential ABBA deadlock, work around it by ordering lock
1632          * grabbing by pipe info address. Otherwise two different processes
1633          * could deadlock (one doing tee from A -> B, the other from B -> A).
1634          */
1635         pipe_double_lock(ipipe, opipe);
1636
1637         do {
1638                 if (!opipe->readers) {
1639                         send_sig(SIGPIPE, current, 0);
1640                         if (!ret)
1641                                 ret = -EPIPE;
1642                         break;
1643                 }
1644
1645                 /*
1646                  * If we have iterated all input buffers or ran out of
1647                  * output room, break.
1648                  */
1649                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1650                         break;
1651
1652                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1653                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1654
1655                 /*
1656                  * Get a reference to this pipe buffer,
1657                  * so we can copy the contents over.
1658                  */
1659                 pipe_buf_get(ipipe, ibuf);
1660
1661                 obuf = opipe->bufs + nbuf;
1662                 *obuf = *ibuf;
1663
1664                 /*
1665                  * Don't inherit the gift flag, we need to
1666                  * prevent multiple steals of this page.
1667                  */
1668                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1669
1670                 if (obuf->len > len)
1671                         obuf->len = len;
1672
1673                 opipe->nrbufs++;
1674                 ret += obuf->len;
1675                 len -= obuf->len;
1676                 i++;
1677         } while (len);
1678
1679         /*
1680          * return EAGAIN if we have the potential of some data in the
1681          * future, otherwise just return 0
1682          */
1683         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1684                 ret = -EAGAIN;
1685
1686         pipe_unlock(ipipe);
1687         pipe_unlock(opipe);
1688
1689         /*
1690          * If we put data in the output pipe, wakeup any potential readers.
1691          */
1692         if (ret > 0)
1693                 wakeup_pipe_readers(opipe);
1694
1695         return ret;
1696 }
1697
1698 /*
1699  * This is a tee(1) implementation that works on pipes. It doesn't copy
1700  * any data, it simply references the 'in' pages on the 'out' pipe.
1701  * The 'flags' used are the SPLICE_F_* variants, currently the only
1702  * applicable one is SPLICE_F_NONBLOCK.
1703  */
1704 static long do_tee(struct file *in, struct file *out, size_t len,
1705                    unsigned int flags)
1706 {
1707         struct pipe_inode_info *ipipe = get_pipe_info(in);
1708         struct pipe_inode_info *opipe = get_pipe_info(out);
1709         int ret = -EINVAL;
1710
1711         /*
1712          * Duplicate the contents of ipipe to opipe without actually
1713          * copying the data.
1714          */
1715         if (ipipe && opipe && ipipe != opipe) {
1716                 /*
1717                  * Keep going, unless we encounter an error. The ipipe/opipe
1718                  * ordering doesn't really matter.
1719                  */
1720                 ret = ipipe_prep(ipipe, flags);
1721                 if (!ret) {
1722                         ret = opipe_prep(opipe, flags);
1723                         if (!ret)
1724                                 ret = link_pipe(ipipe, opipe, len, flags);
1725                 }
1726         }
1727
1728         return ret;
1729 }
1730
1731 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1732 {
1733         struct fd in;
1734         int error;
1735
1736         if (unlikely(!len))
1737                 return 0;
1738
1739         error = -EBADF;
1740         in = fdget(fdin);
1741         if (in.file) {
1742                 if (in.file->f_mode & FMODE_READ) {
1743                         struct fd out = fdget(fdout);
1744                         if (out.file) {
1745                                 if (out.file->f_mode & FMODE_WRITE)
1746                                         error = do_tee(in.file, out.file,
1747                                                         len, flags);
1748                                 fdput(out);
1749                         }
1750                 }
1751                 fdput(in);
1752         }
1753
1754         return error;
1755 }