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