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         buf->ops->release(pipe, buf);
 242         buf->ops = NULL;
 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                 if (WARN_ON(to.idx != idx || to.iov_offset)) {
 320                         /*
 321                          * a bogus ->read_iter() has copied something and still
 322                          * returned an error instead of a short read.
 323                          */
 324                         to.idx = idx;
 325                         to.iov_offset = 0;
 326                         iov_iter_advance(&to, 0); /* to free what was emitted */
 327                 }
 328                 /*
 329                  * callers of ->splice_read() expect -EAGAIN on
 330                  * "can't put anything in there", rather than -EFAULT.
 331                  */
 332                 if (ret == -EFAULT)
 333                         ret = -EAGAIN;
 334         }
 335
 336         return ret;
 337 }
 338 EXPORT_SYMBOL(generic_file_splice_read);
 339
 340 const struct pipe_buf_operations default_pipe_buf_ops = {
 341         .can_merge = 0,
 342         .confirm = generic_pipe_buf_confirm,
 343         .release = generic_pipe_buf_release,
 344         .steal = generic_pipe_buf_steal,
 345         .get = generic_pipe_buf_get,
 346 };
 347
 348 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 349                                     struct pipe_buffer *buf)
 350 {
 351         return 1;
 352 }
 353
 354 /* Pipe buffer operations for a socket and similar. */
 355 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 356         .can_merge = 0,
 357         .confirm = generic_pipe_buf_confirm,
 358         .release = generic_pipe_buf_release,
 359         .steal = generic_pipe_buf_nosteal,
 360         .get = generic_pipe_buf_get,
 361 };
 362 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 363
 364 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 365                             unsigned long vlen, loff_t offset)
 366 {
 367         mm_segment_t old_fs;
 368         loff_t pos = offset;
 369         ssize_t res;
 370
 371         old_fs = get_fs();
 372         set_fs(get_ds());
 373         /* The cast to a user pointer is valid due to the set_fs() */
 374         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 375         set_fs(old_fs);
 376
 377         return res;
 378 }
 379
 380 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 381                             loff_t pos)
 382 {
 383         mm_segment_t old_fs;
 384         ssize_t res;
 385
 386         old_fs = get_fs();
 387         set_fs(get_ds());
 388         /* The cast to a user pointer is valid due to the set_fs() */
 389         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 390         set_fs(old_fs);
 391
 392         return res;
 393 }
 394 EXPORT_SYMBOL(kernel_write);
 395
 396 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 397                                  struct pipe_inode_info *pipe, size_t len,
 398                                  unsigned int flags)
 399 {
 400         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 401         struct iov_iter to;
 402         struct page **pages;
 403         unsigned int nr_pages;
 404         size_t offset, dummy, copied = 0;
 405         ssize_t res;
 406         int i;
 407
 408         if (pipe->nrbufs == pipe->buffers)
 409                 return -EAGAIN;
 410
 411         /*
 412          * Try to keep page boundaries matching to source pagecache ones -
 413          * it probably won't be much help, but...
 414          */
 415         offset = *ppos & ~PAGE_MASK;
 416
 417         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 418
 419         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
 420         if (res <= 0)
 421                 return -ENOMEM;
 422
 423         nr_pages = res / PAGE_SIZE;
 424
 425         vec = __vec;
 426         if (nr_pages > PIPE_DEF_BUFFERS) {
 427                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 428                 if (unlikely(!vec)) {
 429                         res = -ENOMEM;
 430                         goto out;
 431                 }
 432         }
 433
 434         pipe->bufs[to.idx].offset = offset;
 435         pipe->bufs[to.idx].len -= offset;
 436
 437         for (i = 0; i < nr_pages; i++) {
 438                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 439                 vec[i].iov_base = page_address(pages[i]) + offset;
 440                 vec[i].iov_len = this_len;
 441                 len -= this_len;
 442                 offset = 0;
 443         }
 444
 445         res = kernel_readv(in, vec, nr_pages, *ppos);
 446         if (res > 0) {
 447                 copied = res;
 448                 *ppos += res;
 449         }
 450
 451         if (vec != __vec)
 452                 kfree(vec);
 453 out:
 454         for (i = 0; i < nr_pages; i++)
 455                 put_page(pages[i]);
 456         kvfree(pages);
 457         iov_iter_advance(&to, copied);  /* truncates and discards */
 458         return res;
 459 }
 460
 461 /*
 462  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 463  * using sendpage(). Return the number of bytes sent.
 464  */
 465 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 466                             struct pipe_buffer *buf, struct splice_desc *sd)
 467 {
 468         struct file *file = sd->u.file;
 469         loff_t pos = sd->pos;
 470         int more;
 471
 472         if (!likely(file->f_op->sendpage))
 473                 return -EINVAL;
 474
 475         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 476
 477         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 478                 more |= MSG_SENDPAGE_NOTLAST;
 479
 480         return file->f_op->sendpage(file, buf->page, buf->offset,
 481                                     sd->len, &pos, more);
 482 }
 483
 484 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 485 {
 486         smp_mb();
 487         if (waitqueue_active(&pipe->wait))
 488                 wake_up_interruptible(&pipe->wait);
 489         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 490 }
 491
 492 /**
 493  * splice_from_pipe_feed - feed available data from a pipe to a file
 494  * @pipe:       pipe to splice from
 495  * @sd:         information to @actor
 496  * @actor:      handler that splices the data
 497  *
 498  * Description:
 499  *    This function loops over the pipe and calls @actor to do the
 500  *    actual moving of a single struct pipe_buffer to the desired
 501  *    destination.  It returns when there's no more buffers left in
 502  *    the pipe or if the requested number of bytes (@sd->total_len)
 503  *    have been copied.  It returns a positive number (one) if the
 504  *    pipe needs to be filled with more data, zero if the required
 505  *    number of bytes have been copied and -errno on error.
 506  *
 507  *    This, together with splice_from_pipe_{begin,end,next}, may be
 508  *    used to implement the functionality of __splice_from_pipe() when
 509  *    locking is required around copying the pipe buffers to the
 510  *    destination.
 511  */
 512 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 513                           splice_actor *actor)
 514 {
 515         int ret;
 516
 517         while (pipe->nrbufs) {
 518                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 519                 const struct pipe_buf_operations *ops = buf->ops;
 520
 521                 sd->len = buf->len;
 522                 if (sd->len > sd->total_len)
 523                         sd->len = sd->total_len;
 524
 525                 ret = buf->ops->confirm(pipe, buf);
 526                 if (unlikely(ret)) {
 527                         if (ret == -ENODATA)
 528                                 ret = 0;
 529                         return ret;
 530                 }
 531
 532                 ret = actor(pipe, buf, sd);
 533                 if (ret <= 0)
 534                         return ret;
 535
 536                 buf->offset += ret;
 537                 buf->len -= ret;
 538
 539                 sd->num_spliced += ret;
 540                 sd->len -= ret;
 541                 sd->pos += ret;
 542                 sd->total_len -= ret;
 543
 544                 if (!buf->len) {
 545                         buf->ops = NULL;
 546                         ops->release(pipe, buf);
 547                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 548                         pipe->nrbufs--;
 549                         if (pipe->files)
 550                                 sd->need_wakeup = true;
 551                 }
 552
 553                 if (!sd->total_len)
 554                         return 0;
 555         }
 556
 557         return 1;
 558 }
 559
 560 /**
 561  * splice_from_pipe_next - wait for some data to splice from
 562  * @pipe:       pipe to splice from
 563  * @sd:         information about the splice operation
 564  *
 565  * Description:
 566  *    This function will wait for some data and return a positive
 567  *    value (one) if pipe buffers are available.  It will return zero
 568  *    or -errno if no more data needs to be spliced.
 569  */
 570 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 571 {
 572         /*
 573          * Check for signal early to make process killable when there are
 574          * always buffers available
 575          */
 576         if (signal_pending(current))
 577                 return -ERESTARTSYS;
 578
 579         while (!pipe->nrbufs) {
 580                 if (!pipe->writers)
 581                         return 0;
 582
 583                 if (!pipe->waiting_writers && sd->num_spliced)
 584                         return 0;
 585
 586                 if (sd->flags & SPLICE_F_NONBLOCK)
 587                         return -EAGAIN;
 588
 589                 if (signal_pending(current))
 590                         return -ERESTARTSYS;
 591
 592                 if (sd->need_wakeup) {
 593                         wakeup_pipe_writers(pipe);
 594                         sd->need_wakeup = false;
 595                 }
 596
 597                 pipe_wait(pipe);
 598         }
 599
 600         return 1;
 601 }
 602
 603 /**
 604  * splice_from_pipe_begin - start splicing from pipe
 605  * @sd:         information about the splice operation
 606  *
 607  * Description:
 608  *    This function should be called before a loop containing
 609  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 610  *    initialize the necessary fields of @sd.
 611  */
 612 static void splice_from_pipe_begin(struct splice_desc *sd)
 613 {
 614         sd->num_spliced = 0;
 615         sd->need_wakeup = false;
 616 }
 617
 618 /**
 619  * splice_from_pipe_end - finish splicing from pipe
 620  * @pipe:       pipe to splice from
 621  * @sd:         information about the splice operation
 622  *
 623  * Description:
 624  *    This function will wake up pipe writers if necessary.  It should
 625  *    be called after a loop containing splice_from_pipe_next() and
 626  *    splice_from_pipe_feed().
 627  */
 628 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 629 {
 630         if (sd->need_wakeup)
 631                 wakeup_pipe_writers(pipe);
 632 }
 633
 634 /**
 635  * __splice_from_pipe - splice data from a pipe to given actor
 636  * @pipe:       pipe to splice from
 637  * @sd:         information to @actor
 638  * @actor:      handler that splices the data
 639  *
 640  * Description:
 641  *    This function does little more than loop over the pipe and call
 642  *    @actor to do the actual moving of a single struct pipe_buffer to
 643  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 644  *    pipe_to_user.
 645  *
 646  */
 647 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 648                            splice_actor *actor)
 649 {
 650         int ret;
 651
 652         splice_from_pipe_begin(sd);
 653         do {
 654                 cond_resched();
 655                 ret = splice_from_pipe_next(pipe, sd);
 656                 if (ret > 0)
 657                         ret = splice_from_pipe_feed(pipe, sd, actor);
 658         } while (ret > 0);
 659         splice_from_pipe_end(pipe, sd);
 660
 661         return sd->num_spliced ? sd->num_spliced : ret;
 662 }
 663 EXPORT_SYMBOL(__splice_from_pipe);
 664
 665 /**
 666  * splice_from_pipe - splice data from a pipe to a file
 667  * @pipe:       pipe to splice from
 668  * @out:        file to splice to
 669  * @ppos:       position in @out
 670  * @len:        how many bytes to splice
 671  * @flags:      splice modifier flags
 672  * @actor:      handler that splices the data
 673  *
 674  * Description:
 675  *    See __splice_from_pipe. This function locks the pipe inode,
 676  *    otherwise it's identical to __splice_from_pipe().
 677  *
 678  */
 679 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 680                          loff_t *ppos, size_t len, unsigned int flags,
 681                          splice_actor *actor)
 682 {
 683         ssize_t ret;
 684         struct splice_desc sd = {
 685                 .total_len = len,
 686                 .flags = flags,
 687                 .pos = *ppos,
 688                 .u.file = out,
 689         };
 690
 691         pipe_lock(pipe);
 692         ret = __splice_from_pipe(pipe, &sd, actor);
 693         pipe_unlock(pipe);
 694
 695         return ret;
 696 }
 697
 698 /**
 699  * iter_file_splice_write - splice data from a pipe to a file
 700  * @pipe:       pipe info
 701  * @out:        file to write to
 702  * @ppos:       position in @out
 703  * @len:        number of bytes to splice
 704  * @flags:      splice modifier flags
 705  *
 706  * Description:
 707  *    Will either move or copy pages (determined by @flags options) from
 708  *    the given pipe inode to the given file.
 709  *    This one is ->write_iter-based.
 710  *
 711  */
 712 ssize_t
 713 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 714                           loff_t *ppos, size_t len, unsigned int flags)
 715 {
 716         struct splice_desc sd = {
 717                 .total_len = len,
 718                 .flags = flags,
 719                 .pos = *ppos,
 720                 .u.file = out,
 721         };
 722         int nbufs = pipe->buffers;
 723         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 724                                         GFP_KERNEL);
 725         ssize_t ret;
 726
 727         if (unlikely(!array))
 728                 return -ENOMEM;
 729
 730         pipe_lock(pipe);
 731
 732         splice_from_pipe_begin(&sd);
 733         while (sd.total_len) {
 734                 struct iov_iter from;
 735                 size_t left;
 736                 int n, idx;
 737
 738                 ret = splice_from_pipe_next(pipe, &sd);
 739                 if (ret <= 0)
 740                         break;
 741
 742                 if (unlikely(nbufs < pipe->buffers)) {
 743                         kfree(array);
 744                         nbufs = pipe->buffers;
 745                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 746                                         GFP_KERNEL);
 747                         if (!array) {
 748                                 ret = -ENOMEM;
 749                                 break;
 750                         }
 751                 }
 752
 753                 /* build the vector */
 754                 left = sd.total_len;
 755                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 756                         struct pipe_buffer *buf = pipe->bufs + idx;
 757                         size_t this_len = buf->len;
 758
 759                         if (this_len > left)
 760                                 this_len = left;
 761
 762                         if (idx == pipe->buffers - 1)
 763                                 idx = -1;
 764
 765                         ret = buf->ops->confirm(pipe, buf);
 766                         if (unlikely(ret)) {
 767                                 if (ret == -ENODATA)
 768                                         ret = 0;
 769                                 goto done;
 770                         }
 771
 772                         array[n].bv_page = buf->page;
 773                         array[n].bv_len = this_len;
 774                         array[n].bv_offset = buf->offset;
 775                         left -= this_len;
 776                 }
 777
 778                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 779                               sd.total_len - left);
 780                 ret = vfs_iter_write(out, &from, &sd.pos);
 781                 if (ret <= 0)
 782                         break;
 783
 784                 sd.num_spliced += ret;
 785                 sd.total_len -= ret;
 786                 *ppos = sd.pos;
 787
 788                 /* dismiss the fully eaten buffers, adjust the partial one */
 789                 while (ret) {
 790                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 791                         if (ret >= buf->len) {
 792                                 const struct pipe_buf_operations *ops = buf->ops;
 793                                 ret -= buf->len;
 794                                 buf->len = 0;
 795                                 buf->ops = NULL;
 796                                 ops->release(pipe, buf);
 797                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 798                                 pipe->nrbufs--;
 799                                 if (pipe->files)
 800                                         sd.need_wakeup = true;
 801                         } else {
 802                                 buf->offset += ret;
 803                                 buf->len -= ret;
 804                                 ret = 0;
 805                         }
 806                 }
 807         }
 808 done:
 809         kfree(array);
 810         splice_from_pipe_end(pipe, &sd);
 811
 812         pipe_unlock(pipe);
 813
 814         if (sd.num_spliced)
 815                 ret = sd.num_spliced;
 816
 817         return ret;
 818 }
 819
 820 EXPORT_SYMBOL(iter_file_splice_write);
 821
 822 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 823                           struct splice_desc *sd)
 824 {
 825         int ret;
 826         void *data;
 827         loff_t tmp = sd->pos;
 828
 829         data = kmap(buf->page);
 830         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 831         kunmap(buf->page);
 832
 833         return ret;
 834 }
 835
 836 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 837                                          struct file *out, loff_t *ppos,
 838                                          size_t len, unsigned int flags)
 839 {
 840         ssize_t ret;
 841
 842         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 843         if (ret > 0)
 844                 *ppos += ret;
 845
 846         return ret;
 847 }
 848
 849 /**
 850  * generic_splice_sendpage - splice data from a pipe to a socket
 851  * @pipe:       pipe to splice from
 852  * @out:        socket to write to
 853  * @ppos:       position in @out
 854  * @len:        number of bytes to splice
 855  * @flags:      splice modifier flags
 856  *
 857  * Description:
 858  *    Will send @len bytes from the pipe to a network socket. No data copying
 859  *    is involved.
 860  *
 861  */
 862 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 863                                 loff_t *ppos, size_t len, unsigned int flags)
 864 {
 865         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 866 }
 867
 868 EXPORT_SYMBOL(generic_splice_sendpage);
 869
 870 /*
 871  * Attempt to initiate a splice from pipe to file.
 872  */
 873 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 874                            loff_t *ppos, size_t len, unsigned int flags)
 875 {
 876         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 877                                 loff_t *, size_t, unsigned int);
 878
 879         if (out->f_op->splice_write)
 880                 splice_write = out->f_op->splice_write;
 881         else
 882                 splice_write = default_file_splice_write;
 883
 884         return splice_write(pipe, out, ppos, len, flags);
 885 }
 886
 887 /*
 888  * Attempt to initiate a splice from a file to a pipe.
 889  */
 890 static long do_splice_to(struct file *in, loff_t *ppos,
 891                          struct pipe_inode_info *pipe, size_t len,
 892                          unsigned int flags)
 893 {
 894         ssize_t (*splice_read)(struct file *, loff_t *,
 895                                struct pipe_inode_info *, size_t, unsigned int);
 896         int ret;
 897
 898         if (unlikely(!(in->f_mode & FMODE_READ)))
 899                 return -EBADF;
 900
 901         ret = rw_verify_area(READ, in, ppos, len);
 902         if (unlikely(ret < 0))
 903                 return ret;
 904
 905         if (unlikely(len > MAX_RW_COUNT))
 906                 len = MAX_RW_COUNT;
 907
 908         if (in->f_op->splice_read)
 909                 splice_read = in->f_op->splice_read;
 910         else
 911                 splice_read = default_file_splice_read;
 912
 913         return splice_read(in, ppos, pipe, len, flags);
 914 }
 915
 916 /**
 917  * splice_direct_to_actor - splices data directly between two non-pipes
 918  * @in:         file to splice from
 919  * @sd:         actor information on where to splice to
 920  * @actor:      handles the data splicing
 921  *
 922  * Description:
 923  *    This is a special case helper to splice directly between two
 924  *    points, without requiring an explicit pipe. Internally an allocated
 925  *    pipe is cached in the process, and reused during the lifetime of
 926  *    that process.
 927  *
 928  */
 929 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 930                                splice_direct_actor *actor)
 931 {
 932         struct pipe_inode_info *pipe;
 933         long ret, bytes;
 934         umode_t i_mode;
 935         size_t len;
 936         int i, flags, more;
 937
 938         /*
 939          * We require the input being a regular file, as we don't want to
 940          * randomly drop data for eg socket -> socket splicing. Use the
 941          * piped splicing for that!
 942          */
 943         i_mode = file_inode(in)->i_mode;
 944         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 945                 return -EINVAL;
 946
 947         /*
 948          * neither in nor out is a pipe, setup an internal pipe attached to
 949          * 'out' and transfer the wanted data from 'in' to 'out' through that
 950          */
 951         pipe = current->splice_pipe;
 952         if (unlikely(!pipe)) {
 953                 pipe = alloc_pipe_info();
 954                 if (!pipe)
 955                         return -ENOMEM;
 956
 957                 /*
 958                  * We don't have an immediate reader, but we'll read the stuff
 959                  * out of the pipe right after the splice_to_pipe(). So set
 960                  * PIPE_READERS appropriately.
 961                  */
 962                 pipe->readers = 1;
 963
 964                 current->splice_pipe = pipe;
 965         }
 966
 967         /*
 968          * Do the splice.
 969          */
 970         ret = 0;
 971         bytes = 0;
 972         len = sd->total_len;
 973         flags = sd->flags;
 974
 975         /*
 976          * Don't block on output, we have to drain the direct pipe.
 977          */
 978         sd->flags &= ~SPLICE_F_NONBLOCK;
 979         more = sd->flags & SPLICE_F_MORE;
 980
 981         while (len) {
 982                 size_t read_len;
 983                 loff_t pos = sd->pos, prev_pos = pos;
 984
 985                 ret = do_splice_to(in, &pos, pipe, len, flags);
 986                 if (unlikely(ret <= 0))
 987                         goto out_release;
 988
 989                 read_len = ret;
 990                 sd->total_len = read_len;
 991
 992                 /*
 993                  * If more data is pending, set SPLICE_F_MORE
 994                  * If this is the last data and SPLICE_F_MORE was not set
 995                  * initially, clears it.
 996                  */
 997                 if (read_len < len)
 998                         sd->flags |= SPLICE_F_MORE;
 999                 else if (!more)
1000                         sd->flags &= ~SPLICE_F_MORE;
1001                 /*
1002                  * NOTE: nonblocking mode only applies to the input. We
1003                  * must not do the output in nonblocking mode as then we
1004                  * could get stuck data in the internal pipe:
1005                  */
1006                 ret = actor(pipe, sd);
1007                 if (unlikely(ret <= 0)) {
1008                         sd->pos = prev_pos;
1009                         goto out_release;
1010                 }
1011
1012                 bytes += ret;
1013                 len -= ret;
1014                 sd->pos = pos;
1015
1016                 if (ret < read_len) {
1017                         sd->pos = prev_pos + ret;
1018                         goto out_release;
1019                 }
1020         }
1021
1022 done:
1023         pipe->nrbufs = pipe->curbuf = 0;
1024         file_accessed(in);
1025         return bytes;
1026
1027 out_release:
1028         /*
1029          * If we did an incomplete transfer we must release
1030          * the pipe buffers in question:
1031          */
1032         for (i = 0; i < pipe->buffers; i++) {
1033                 struct pipe_buffer *buf = pipe->bufs + i;
1034
1035                 if (buf->ops) {
1036                         buf->ops->release(pipe, buf);
1037                         buf->ops = NULL;
1038                 }
1039         }
1040
1041         if (!bytes)
1042                 bytes = ret;
1043
1044         goto done;
1045 }
1046 EXPORT_SYMBOL(splice_direct_to_actor);
1047
1048 static int direct_splice_actor(struct pipe_inode_info *pipe,
1049                                struct splice_desc *sd)
1050 {
1051         struct file *file = sd->u.file;
1052
1053         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1054                               sd->flags);
1055 }
1056
1057 /**
1058  * do_splice_direct - splices data directly between two files
1059  * @in:         file to splice from
1060  * @ppos:       input file offset
1061  * @out:        file to splice to
1062  * @opos:       output file offset
1063  * @len:        number of bytes to splice
1064  * @flags:      splice modifier flags
1065  *
1066  * Description:
1067  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1068  *    doing it in the application would incur an extra system call
1069  *    (splice in + splice out, as compared to just sendfile()). So this helper
1070  *    can splice directly through a process-private pipe.
1071  *
1072  */
1073 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1074                       loff_t *opos, size_t len, unsigned int flags)
1075 {
1076         struct splice_desc sd = {
1077                 .len            = len,
1078                 .total_len      = len,
1079                 .flags          = flags,
1080                 .pos            = *ppos,
1081                 .u.file         = out,
1082                 .opos           = opos,
1083         };
1084         long ret;
1085
1086         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1087                 return -EBADF;
1088
1089         if (unlikely(out->f_flags & O_APPEND))
1090                 return -EINVAL;
1091
1092         ret = rw_verify_area(WRITE, out, opos, len);
1093         if (unlikely(ret < 0))
1094                 return ret;
1095
1096         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1097         if (ret > 0)
1098                 *ppos = sd.pos;
1099
1100         return ret;
1101 }
1102 EXPORT_SYMBOL(do_splice_direct);
1103
1104 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1105 {
1106         while (pipe->nrbufs == pipe->buffers) {
1107                 if (flags & SPLICE_F_NONBLOCK)
1108                         return -EAGAIN;
1109                 if (signal_pending(current))
1110                         return -ERESTARTSYS;
1111                 pipe->waiting_writers++;
1112                 pipe_wait(pipe);
1113                 pipe->waiting_writers--;
1114         }
1115         return 0;
1116 }
1117
1118 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1119                                struct pipe_inode_info *opipe,
1120                                size_t len, unsigned int flags);
1121
1122 /*
1123  * Determine where to splice to/from.
1124  */
1125 static long do_splice(struct file *in, loff_t __user *off_in,
1126                       struct file *out, loff_t __user *off_out,
1127                       size_t len, unsigned int flags)
1128 {
1129         struct pipe_inode_info *ipipe;
1130         struct pipe_inode_info *opipe;
1131         loff_t offset;
1132         long ret;
1133
1134         ipipe = get_pipe_info(in);
1135         opipe = get_pipe_info(out);
1136
1137         if (ipipe && opipe) {
1138                 if (off_in || off_out)
1139                         return -ESPIPE;
1140
1141                 if (!(in->f_mode & FMODE_READ))
1142                         return -EBADF;
1143
1144                 if (!(out->f_mode & FMODE_WRITE))
1145                         return -EBADF;
1146
1147                 /* Splicing to self would be fun, but... */
1148                 if (ipipe == opipe)
1149                         return -EINVAL;
1150
1151                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1152         }
1153
1154         if (ipipe) {
1155                 if (off_in)
1156                         return -ESPIPE;
1157                 if (off_out) {
1158                         if (!(out->f_mode & FMODE_PWRITE))
1159                                 return -EINVAL;
1160                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1161                                 return -EFAULT;
1162                 } else {
1163                         offset = out->f_pos;
1164                 }
1165
1166                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1167                         return -EBADF;
1168
1169                 if (unlikely(out->f_flags & O_APPEND))
1170                         return -EINVAL;
1171
1172                 ret = rw_verify_area(WRITE, out, &offset, len);
1173                 if (unlikely(ret < 0))
1174                         return ret;
1175
1176                 file_start_write(out);
1177                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1178                 file_end_write(out);
1179
1180                 if (!off_out)
1181                         out->f_pos = offset;
1182                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1183                         ret = -EFAULT;
1184
1185                 return ret;
1186         }
1187
1188         if (opipe) {
1189                 if (off_out)
1190                         return -ESPIPE;
1191                 if (off_in) {
1192                         if (!(in->f_mode & FMODE_PREAD))
1193                                 return -EINVAL;
1194                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1195                                 return -EFAULT;
1196                 } else {
1197                         offset = in->f_pos;
1198                 }
1199
1200                 pipe_lock(opipe);
1201                 ret = wait_for_space(opipe, flags);
1202                 if (!ret)
1203                         ret = do_splice_to(in, &offset, opipe, len, flags);
1204                 pipe_unlock(opipe);
1205                 if (ret > 0)
1206                         wakeup_pipe_readers(opipe);
1207                 if (!off_in)
1208                         in->f_pos = offset;
1209                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1210                         ret = -EFAULT;
1211
1212                 return ret;
1213         }
1214
1215         return -EINVAL;
1216 }
1217
1218 static int iter_to_pipe(struct iov_iter *from,
1219                         struct pipe_inode_info *pipe,
1220                         unsigned flags)
1221 {
1222         struct pipe_buffer buf = {
1223                 .ops = &user_page_pipe_buf_ops,
1224                 .flags = flags
1225         };
1226         size_t total = 0;
1227         int ret = 0;
1228         bool failed = false;
1229
1230         while (iov_iter_count(from) && !failed) {
1231                 struct page *pages[16];
1232                 ssize_t copied;
1233                 size_t start;
1234                 int n;
1235
1236                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1237                 if (copied <= 0) {
1238                         ret = copied;
1239                         break;
1240                 }
1241
1242                 for (n = 0; copied; n++, start = 0) {
1243                         int size = min_t(int, copied, PAGE_SIZE - start);
1244                         if (!failed) {
1245                                 buf.page = pages[n];
1246                                 buf.offset = start;
1247                                 buf.len = size;
1248                                 ret = add_to_pipe(pipe, &buf);
1249                                 if (unlikely(ret < 0)) {
1250                                         failed = true;
1251                                 } else {
1252                                         iov_iter_advance(from, ret);
1253                                         total += ret;
1254                                 }
1255                         } else {
1256                                 put_page(pages[n]);
1257                         }
1258                         copied -= size;
1259                 }
1260         }
1261         return total ? total : ret;
1262 }
1263
1264 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1265                         struct splice_desc *sd)
1266 {
1267         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1268         return n == sd->len ? n : -EFAULT;
1269 }
1270
1271 /*
1272  * For lack of a better implementation, implement vmsplice() to userspace
1273  * as a simple copy of the pipes pages to the user iov.
1274  */
1275 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1276                              unsigned long nr_segs, unsigned int flags)
1277 {
1278         struct pipe_inode_info *pipe;
1279         struct splice_desc sd;
1280         long ret;
1281         struct iovec iovstack[UIO_FASTIOV];
1282         struct iovec *iov = iovstack;
1283         struct iov_iter iter;
1284
1285         pipe = get_pipe_info(file);
1286         if (!pipe)
1287                 return -EBADF;
1288
1289         ret = import_iovec(READ, uiov, nr_segs,
1290                            ARRAY_SIZE(iovstack), &iov, &iter);
1291         if (ret < 0)
1292                 return ret;
1293
1294         sd.total_len = iov_iter_count(&iter);
1295         sd.len = 0;
1296         sd.flags = flags;
1297         sd.u.data = &iter;
1298         sd.pos = 0;
1299
1300         if (sd.total_len) {
1301                 pipe_lock(pipe);
1302                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1303                 pipe_unlock(pipe);
1304         }
1305
1306         kfree(iov);
1307         return ret;
1308 }
1309
1310 /*
1311  * vmsplice splices a user address range into a pipe. It can be thought of
1312  * as splice-from-memory, where the regular splice is splice-from-file (or
1313  * to file). In both cases the output is a pipe, naturally.
1314  */
1315 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1316                              unsigned long nr_segs, unsigned int flags)
1317 {
1318         struct pipe_inode_info *pipe;
1319         struct iovec iovstack[UIO_FASTIOV];
1320         struct iovec *iov = iovstack;
1321         struct iov_iter from;
1322         long ret;
1323         unsigned buf_flag = 0;
1324
1325         if (flags & SPLICE_F_GIFT)
1326                 buf_flag = PIPE_BUF_FLAG_GIFT;
1327
1328         pipe = get_pipe_info(file);
1329         if (!pipe)
1330                 return -EBADF;
1331
1332         ret = import_iovec(WRITE, uiov, nr_segs,
1333                            ARRAY_SIZE(iovstack), &iov, &from);
1334         if (ret < 0)
1335                 return ret;
1336
1337         pipe_lock(pipe);
1338         ret = wait_for_space(pipe, flags);
1339         if (!ret)
1340                 ret = iter_to_pipe(&from, pipe, buf_flag);
1341         pipe_unlock(pipe);
1342         if (ret > 0)
1343                 wakeup_pipe_readers(pipe);
1344         kfree(iov);
1345         return ret;
1346 }
1347
1348 /*
1349  * Note that vmsplice only really supports true splicing _from_ user memory
1350  * to a pipe, not the other way around. Splicing from user memory is a simple
1351  * operation that can be supported without any funky alignment restrictions
1352  * or nasty vm tricks. We simply map in the user memory and fill them into
1353  * a pipe. The reverse isn't quite as easy, though. There are two possible
1354  * solutions for that:
1355  *
1356  *      - memcpy() the data internally, at which point we might as well just
1357  *        do a regular read() on the buffer anyway.
1358  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1359  *        has restriction limitations on both ends of the pipe).
1360  *
1361  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1362  *
1363  */
1364 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1365                 unsigned long, nr_segs, unsigned int, flags)
1366 {
1367         struct fd f;
1368         long error;
1369
1370         if (unlikely(nr_segs > UIO_MAXIOV))
1371                 return -EINVAL;
1372         else if (unlikely(!nr_segs))
1373                 return 0;
1374
1375         error = -EBADF;
1376         f = fdget(fd);
1377         if (f.file) {
1378                 if (f.file->f_mode & FMODE_WRITE)
1379                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1380                 else if (f.file->f_mode & FMODE_READ)
1381                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1382
1383                 fdput(f);
1384         }
1385
1386         return error;
1387 }
1388
1389 #ifdef CONFIG_COMPAT
1390 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1391                     unsigned int, nr_segs, unsigned int, flags)
1392 {
1393         unsigned i;
1394         struct iovec __user *iov;
1395         if (nr_segs > UIO_MAXIOV)
1396                 return -EINVAL;
1397         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1398         for (i = 0; i < nr_segs; i++) {
1399                 struct compat_iovec v;
1400                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1401                     get_user(v.iov_len, &iov32[i].iov_len) ||
1402                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1403                     put_user(v.iov_len, &iov[i].iov_len))
1404                         return -EFAULT;
1405         }
1406         return sys_vmsplice(fd, iov, nr_segs, flags);
1407 }
1408 #endif
1409
1410 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1411                 int, fd_out, loff_t __user *, off_out,
1412                 size_t, len, unsigned int, flags)
1413 {
1414         struct fd in, out;
1415         long error;
1416
1417         if (unlikely(!len))
1418                 return 0;
1419
1420         error = -EBADF;
1421         in = fdget(fd_in);
1422         if (in.file) {
1423                 if (in.file->f_mode & FMODE_READ) {
1424                         out = fdget(fd_out);
1425                         if (out.file) {
1426                                 if (out.file->f_mode & FMODE_WRITE)
1427                                         error = do_splice(in.file, off_in,
1428                                                           out.file, off_out,
1429                                                           len, flags);
1430                                 fdput(out);
1431                         }
1432                 }
1433                 fdput(in);
1434         }
1435         return error;
1436 }
1437
1438 /*
1439  * Make sure there's data to read. Wait for input if we can, otherwise
1440  * return an appropriate error.
1441  */
1442 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1443 {
1444         int ret;
1445
1446         /*
1447          * Check ->nrbufs without the inode lock first. This function
1448          * is speculative anyways, so missing one is ok.
1449          */
1450         if (pipe->nrbufs)
1451                 return 0;
1452
1453         ret = 0;
1454         pipe_lock(pipe);
1455
1456         while (!pipe->nrbufs) {
1457                 if (signal_pending(current)) {
1458                         ret = -ERESTARTSYS;
1459                         break;
1460                 }
1461                 if (!pipe->writers)
1462                         break;
1463                 if (!pipe->waiting_writers) {
1464                         if (flags & SPLICE_F_NONBLOCK) {
1465                                 ret = -EAGAIN;
1466                                 break;
1467                         }
1468                 }
1469                 pipe_wait(pipe);
1470         }
1471
1472         pipe_unlock(pipe);
1473         return ret;
1474 }
1475
1476 /*
1477  * Make sure there's writeable room. Wait for room if we can, otherwise
1478  * return an appropriate error.
1479  */
1480 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1481 {
1482         int ret;
1483
1484         /*
1485          * Check ->nrbufs without the inode lock first. This function
1486          * is speculative anyways, so missing one is ok.
1487          */
1488         if (pipe->nrbufs < pipe->buffers)
1489                 return 0;
1490
1491         ret = 0;
1492         pipe_lock(pipe);
1493
1494         while (pipe->nrbufs >= pipe->buffers) {
1495                 if (!pipe->readers) {
1496                         send_sig(SIGPIPE, current, 0);
1497                         ret = -EPIPE;
1498                         break;
1499                 }
1500                 if (flags & SPLICE_F_NONBLOCK) {
1501                         ret = -EAGAIN;
1502                         break;
1503                 }
1504                 if (signal_pending(current)) {
1505                         ret = -ERESTARTSYS;
1506                         break;
1507                 }
1508                 pipe->waiting_writers++;
1509                 pipe_wait(pipe);
1510                 pipe->waiting_writers--;
1511         }
1512
1513         pipe_unlock(pipe);
1514         return ret;
1515 }
1516
1517 /*
1518  * Splice contents of ipipe to opipe.
1519  */
1520 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1521                                struct pipe_inode_info *opipe,
1522                                size_t len, unsigned int flags)
1523 {
1524         struct pipe_buffer *ibuf, *obuf;
1525         int ret = 0, nbuf;
1526         bool input_wakeup = false;
1527
1528
1529 retry:
1530         ret = ipipe_prep(ipipe, flags);
1531         if (ret)
1532                 return ret;
1533
1534         ret = opipe_prep(opipe, flags);
1535         if (ret)
1536                 return ret;
1537
1538         /*
1539          * Potential ABBA deadlock, work around it by ordering lock
1540          * grabbing by pipe info address. Otherwise two different processes
1541          * could deadlock (one doing tee from A -> B, the other from B -> A).
1542          */
1543         pipe_double_lock(ipipe, opipe);
1544
1545         do {
1546                 if (!opipe->readers) {
1547                         send_sig(SIGPIPE, current, 0);
1548                         if (!ret)
1549                                 ret = -EPIPE;
1550                         break;
1551                 }
1552
1553                 if (!ipipe->nrbufs && !ipipe->writers)
1554                         break;
1555
1556                 /*
1557                  * Cannot make any progress, because either the input
1558                  * pipe is empty or the output pipe is full.
1559                  */
1560                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1561                         /* Already processed some buffers, break */
1562                         if (ret)
1563                                 break;
1564
1565                         if (flags & SPLICE_F_NONBLOCK) {
1566                                 ret = -EAGAIN;
1567                                 break;
1568                         }
1569
1570                         /*
1571                          * We raced with another reader/writer and haven't
1572                          * managed to process any buffers.  A zero return
1573                          * value means EOF, so retry instead.
1574                          */
1575                         pipe_unlock(ipipe);
1576                         pipe_unlock(opipe);
1577                         goto retry;
1578                 }
1579
1580                 ibuf = ipipe->bufs + ipipe->curbuf;
1581                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1582                 obuf = opipe->bufs + nbuf;
1583
1584                 if (len >= ibuf->len) {
1585                         /*
1586                          * Simply move the whole buffer from ipipe to opipe
1587                          */
1588                         *obuf = *ibuf;
1589                         ibuf->ops = NULL;
1590                         opipe->nrbufs++;
1591                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1592                         ipipe->nrbufs--;
1593                         input_wakeup = true;
1594                 } else {
1595                         /*
1596                          * Get a reference to this pipe buffer,
1597                          * so we can copy the contents over.
1598                          */
1599                         pipe_buf_get(ipipe, ibuf);
1600                         *obuf = *ibuf;
1601
1602                         /*
1603                          * Don't inherit the gift flag, we need to
1604                          * prevent multiple steals of this page.
1605                          */
1606                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1607
1608                         obuf->len = len;
1609                         opipe->nrbufs++;
1610                         ibuf->offset += obuf->len;
1611                         ibuf->len -= obuf->len;
1612                 }
1613                 ret += obuf->len;
1614                 len -= obuf->len;
1615         } while (len);
1616
1617         pipe_unlock(ipipe);
1618         pipe_unlock(opipe);
1619
1620         /*
1621          * If we put data in the output pipe, wakeup any potential readers.
1622          */
1623         if (ret > 0)
1624                 wakeup_pipe_readers(opipe);
1625
1626         if (input_wakeup)
1627                 wakeup_pipe_writers(ipipe);
1628
1629         return ret;
1630 }
1631
1632 /*
1633  * Link contents of ipipe to opipe.
1634  */
1635 static int link_pipe(struct pipe_inode_info *ipipe,
1636                      struct pipe_inode_info *opipe,
1637                      size_t len, unsigned int flags)
1638 {
1639         struct pipe_buffer *ibuf, *obuf;
1640         int ret = 0, i = 0, nbuf;
1641
1642         /*
1643          * Potential ABBA deadlock, work around it by ordering lock
1644          * grabbing by pipe info address. Otherwise two different processes
1645          * could deadlock (one doing tee from A -> B, the other from B -> A).
1646          */
1647         pipe_double_lock(ipipe, opipe);
1648
1649         do {
1650                 if (!opipe->readers) {
1651                         send_sig(SIGPIPE, current, 0);
1652                         if (!ret)
1653                                 ret = -EPIPE;
1654                         break;
1655                 }
1656
1657                 /*
1658                  * If we have iterated all input buffers or ran out of
1659                  * output room, break.
1660                  */
1661                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1662                         break;
1663
1664                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1665                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1666
1667                 /*
1668                  * Get a reference to this pipe buffer,
1669                  * so we can copy the contents over.
1670                  */
1671                 pipe_buf_get(ipipe, ibuf);
1672
1673                 obuf = opipe->bufs + nbuf;
1674                 *obuf = *ibuf;
1675
1676                 /*
1677                  * Don't inherit the gift flag, we need to
1678                  * prevent multiple steals of this page.
1679                  */
1680                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1681
1682                 if (obuf->len > len)
1683                         obuf->len = len;
1684
1685                 opipe->nrbufs++;
1686                 ret += obuf->len;
1687                 len -= obuf->len;
1688                 i++;
1689         } while (len);
1690
1691         /*
1692          * return EAGAIN if we have the potential of some data in the
1693          * future, otherwise just return 0
1694          */
1695         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1696                 ret = -EAGAIN;
1697
1698         pipe_unlock(ipipe);
1699         pipe_unlock(opipe);
1700
1701         /*
1702          * If we put data in the output pipe, wakeup any potential readers.
1703          */
1704         if (ret > 0)
1705                 wakeup_pipe_readers(opipe);
1706
1707         return ret;
1708 }
1709
1710 /*
1711  * This is a tee(1) implementation that works on pipes. It doesn't copy
1712  * any data, it simply references the 'in' pages on the 'out' pipe.
1713  * The 'flags' used are the SPLICE_F_* variants, currently the only
1714  * applicable one is SPLICE_F_NONBLOCK.
1715  */
1716 static long do_tee(struct file *in, struct file *out, size_t len,
1717                    unsigned int flags)
1718 {
1719         struct pipe_inode_info *ipipe = get_pipe_info(in);
1720         struct pipe_inode_info *opipe = get_pipe_info(out);
1721         int ret = -EINVAL;
1722
1723         /*
1724          * Duplicate the contents of ipipe to opipe without actually
1725          * copying the data.
1726          */
1727         if (ipipe && opipe && ipipe != opipe) {
1728                 /*
1729                  * Keep going, unless we encounter an error. The ipipe/opipe
1730                  * ordering doesn't really matter.
1731                  */
1732                 ret = ipipe_prep(ipipe, flags);
1733                 if (!ret) {
1734                         ret = opipe_prep(opipe, flags);
1735                         if (!ret)
1736                                 ret = link_pipe(ipipe, opipe, len, flags);
1737                 }
1738         }
1739
1740         return ret;
1741 }
1742
1743 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1744 {
1745         struct fd in;
1746         int error;
1747
1748         if (unlikely(!len))
1749                 return 0;
1750
1751         error = -EBADF;
1752         in = fdget(fdin);
1753         if (in.file) {
1754                 if (in.file->f_mode & FMODE_READ) {
1755                         struct fd out = fdget(fdout);
1756                         if (out.file) {
1757                                 if (out.file->f_mode & FMODE_WRITE)
1758                                         error = do_tee(in.file, out.file,
1759                                                         len, flags);
1760                                 fdput(out);
1761                         }
1762                 }
1763                 fdput(in);
1764         }
1765
1766         return error;
1767 }