fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/bvec.h>
  21 #include <linux/fs.h>
  22 #include <linux/file.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/splice.h>
  25 #include <linux/memcontrol.h>
  26 #include <linux/mm_inline.h>
  27 #include <linux/swap.h>
  28 #include <linux/writeback.h>
  29 #include <linux/export.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33 #include <linux/gfp.h>
  34 #include <linux/socket.h>
  35 #include <linux/compat.h>
  36 #include <linux/sched/signal.h>
  37
  38 #include "internal.h"
  39
  40 /*
  41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42  * a vm helper function, it's already simplified quite a bit by the
  43  * addition of remove_mapping(). If success is returned, the caller may
  44  * attempt to reuse this page for another destination.
  45  */
  46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  47                                      struct pipe_buffer *buf)
  48 {
  49         struct page *page = buf->page;
  50         struct address_space *mapping;
  51
  52         lock_page(page);
  53
  54         mapping = page_mapping(page);
  55         if (mapping) {
  56                 WARN_ON(!PageUptodate(page));
  57
  58                 /*
  59                  * At least for ext2 with nobh option, we need to wait on
  60                  * writeback completing on this page, since we'll remove it
  61                  * from the pagecache.  Otherwise truncate wont wait on the
  62                  * page, allowing the disk blocks to be reused by someone else
  63                  * before we actually wrote our data to them. fs corruption
  64                  * ensues.
  65                  */
  66                 wait_on_page_writeback(page);
  67
  68                 if (page_has_private(page) &&
  69                     !try_to_release_page(page, GFP_KERNEL))
  70                         goto out_unlock;
  71
  72                 /*
  73                  * If we succeeded in removing the mapping, set LRU flag
  74                  * and return good.
  75                  */
  76                 if (remove_mapping(mapping, page)) {
  77                         buf->flags |= PIPE_BUF_FLAG_LRU;
  78                         return 0;
  79                 }
  80         }
  81
  82         /*
  83          * Raced with truncate or failed to remove page from current
  84          * address space, unlock and return failure.
  85          */
  86 out_unlock:
  87         unlock_page(page);
  88         return 1;
  89 }
  90
  91 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  92                                         struct pipe_buffer *buf)
  93 {
  94         put_page(buf->page);
  95         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  96 }
  97
  98 /*
  99  * Check whether the contents of buf is OK to access. Since the content
 100  * is a page cache page, IO may be in flight.
 101  */
 102 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 103                                        struct pipe_buffer *buf)
 104 {
 105         struct page *page = buf->page;
 106         int err;
 107
 108         if (!PageUptodate(page)) {
 109                 lock_page(page);
 110
 111                 /*
 112                  * Page got truncated/unhashed. This will cause a 0-byte
 113                  * splice, if this is the first page.
 114                  */
 115                 if (!page->mapping) {
 116                         err = -ENODATA;
 117                         goto error;
 118                 }
 119
 120                 /*
 121                  * Uh oh, read-error from disk.
 122                  */
 123                 if (!PageUptodate(page)) {
 124                         err = -EIO;
 125                         goto error;
 126                 }
 127
 128                 /*
 129                  * Page is ok afterall, we are done.
 130                  */
 131                 unlock_page(page);
 132         }
 133
 134         return 0;
 135 error:
 136         unlock_page(page);
 137         return err;
 138 }
 139
 140 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 141         .can_merge = 0,
 142         .confirm = page_cache_pipe_buf_confirm,
 143         .release = page_cache_pipe_buf_release,
 144         .steal = page_cache_pipe_buf_steal,
 145         .get = generic_pipe_buf_get,
 146 };
 147
 148 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 149                                     struct pipe_buffer *buf)
 150 {
 151         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 152                 return 1;
 153
 154         buf->flags |= PIPE_BUF_FLAG_LRU;
 155         return generic_pipe_buf_steal(pipe, buf);
 156 }
 157
 158 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 159         .can_merge = 0,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 189         int ret = 0, page_nr = 0;
 190
 191         if (!spd_pages)
 192                 return 0;
 193
 194         if (unlikely(!pipe->readers)) {
 195                 send_sig(SIGPIPE, current, 0);
 196                 ret = -EPIPE;
 197                 goto out;
 198         }
 199
 200         while (pipe->nrbufs < pipe->buffers) {
 201                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 202                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 203
 204                 buf->page = spd->pages[page_nr];
 205                 buf->offset = spd->partial[page_nr].offset;
 206                 buf->len = spd->partial[page_nr].len;
 207                 buf->private = spd->partial[page_nr].private;
 208                 buf->ops = spd->ops;
 209                 buf->flags = 0;
 210
 211                 pipe->nrbufs++;
 212                 page_nr++;
 213                 ret += buf->len;
 214
 215                 if (!--spd->nr_pages)
 216                         break;
 217         }
 218
 219         if (!ret)
 220                 ret = -EAGAIN;
 221
 222 out:
 223         while (page_nr < spd_pages)
 224                 spd->spd_release(spd, page_nr++);
 225
 226         return ret;
 227 }
 228 EXPORT_SYMBOL_GPL(splice_to_pipe);
 229
 230 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 231 {
 232         int ret;
 233
 234         if (unlikely(!pipe->readers)) {
 235                 send_sig(SIGPIPE, current, 0);
 236                 ret = -EPIPE;
 237         } else if (pipe->nrbufs == pipe->buffers) {
 238                 ret = -EAGAIN;
 239         } else {
 240                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 241                 pipe->bufs[newbuf] = *buf;
 242                 pipe->nrbufs++;
 243                 return buf->len;
 244         }
 245         pipe_buf_release(pipe, buf);
 246         return ret;
 247 }
 248 EXPORT_SYMBOL(add_to_pipe);
 249
 250 /*
 251  * Check if we need to grow the arrays holding pages and partial page
 252  * descriptions.
 253  */
 254 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 255 {
 256         unsigned int buffers = READ_ONCE(pipe->buffers);
 257
 258         spd->nr_pages_max = buffers;
 259         if (buffers <= PIPE_DEF_BUFFERS)
 260                 return 0;
 261
 262         spd->pages = kmalloc_array(buffers, sizeof(struct page *), GFP_KERNEL);
 263         spd->partial = kmalloc_array(buffers, sizeof(struct partial_page),
 264                                      GFP_KERNEL);
 265
 266         if (spd->pages && spd->partial)
 267                 return 0;
 268
 269         kfree(spd->pages);
 270         kfree(spd->partial);
 271         return -ENOMEM;
 272 }
 273
 274 void splice_shrink_spd(struct splice_pipe_desc *spd)
 275 {
 276         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 277                 return;
 278
 279         kfree(spd->pages);
 280         kfree(spd->partial);
 281 }
 282
 283 /**
 284  * generic_file_splice_read - splice data from file to a pipe
 285  * @in:         file to splice from
 286  * @ppos:       position in @in
 287  * @pipe:       pipe to splice to
 288  * @len:        number of bytes to splice
 289  * @flags:      splice modifier flags
 290  *
 291  * Description:
 292  *    Will read pages from given file and fill them into a pipe. Can be
 293  *    used as long as it has more or less sane ->read_iter().
 294  *
 295  */
 296 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 297                                  struct pipe_inode_info *pipe, size_t len,
 298                                  unsigned int flags)
 299 {
 300         struct iov_iter to;
 301         struct kiocb kiocb;
 302         int idx, ret;
 303
 304         iov_iter_pipe(&to, READ, pipe, len);
 305         idx = to.idx;
 306         init_sync_kiocb(&kiocb, in);
 307         kiocb.ki_pos = *ppos;
 308         ret = call_read_iter(in, &kiocb, &to);
 309         if (ret > 0) {
 310                 *ppos = kiocb.ki_pos;
 311                 file_accessed(in);
 312         } else if (ret < 0) {
 313                 to.idx = idx;
 314                 to.iov_offset = 0;
 315                 iov_iter_advance(&to, 0); /* to free what was emitted */
 316                 /*
 317                  * callers of ->splice_read() expect -EAGAIN on
 318                  * "can't put anything in there", rather than -EFAULT.
 319                  */
 320                 if (ret == -EFAULT)
 321                         ret = -EAGAIN;
 322         }
 323
 324         return ret;
 325 }
 326 EXPORT_SYMBOL(generic_file_splice_read);
 327
 328 const struct pipe_buf_operations default_pipe_buf_ops = {
 329         .can_merge = 0,
 330         .confirm = generic_pipe_buf_confirm,
 331         .release = generic_pipe_buf_release,
 332         .steal = generic_pipe_buf_steal,
 333         .get = generic_pipe_buf_get,
 334 };
 335
 336 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 337                                     struct pipe_buffer *buf)
 338 {
 339         return 1;
 340 }
 341
 342 /* Pipe buffer operations for a socket and similar. */
 343 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 344         .can_merge = 0,
 345         .confirm = generic_pipe_buf_confirm,
 346         .release = generic_pipe_buf_release,
 347         .steal = generic_pipe_buf_nosteal,
 348         .get = generic_pipe_buf_get,
 349 };
 350 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 351
 352 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 353                             unsigned long vlen, loff_t offset)
 354 {
 355         mm_segment_t old_fs;
 356         loff_t pos = offset;
 357         ssize_t res;
 358
 359         old_fs = get_fs();
 360         set_fs(get_ds());
 361         /* The cast to a user pointer is valid due to the set_fs() */
 362         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 363         set_fs(old_fs);
 364
 365         return res;
 366 }
 367
 368 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 369                                  struct pipe_inode_info *pipe, size_t len,
 370                                  unsigned int flags)
 371 {
 372         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 373         struct iov_iter to;
 374         struct page **pages;
 375         unsigned int nr_pages;
 376         size_t offset, base, copied = 0;
 377         ssize_t res;
 378         int i;
 379
 380         if (pipe->nrbufs == pipe->buffers)
 381                 return -EAGAIN;
 382
 383         /*
 384          * Try to keep page boundaries matching to source pagecache ones -
 385          * it probably won't be much help, but...
 386          */
 387         offset = *ppos & ~PAGE_MASK;
 388
 389         iov_iter_pipe(&to, READ, pipe, len + offset);
 390
 391         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 392         if (res <= 0)
 393                 return -ENOMEM;
 394
 395         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 396
 397         vec = __vec;
 398         if (nr_pages > PIPE_DEF_BUFFERS) {
 399                 vec = kmalloc_array(nr_pages, sizeof(struct kvec), GFP_KERNEL);
 400                 if (unlikely(!vec)) {
 401                         res = -ENOMEM;
 402                         goto out;
 403                 }
 404         }
 405
 406         pipe->bufs[to.idx].offset = offset;
 407         pipe->bufs[to.idx].len -= offset;
 408
 409         for (i = 0; i < nr_pages; i++) {
 410                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 411                 vec[i].iov_base = page_address(pages[i]) + offset;
 412                 vec[i].iov_len = this_len;
 413                 len -= this_len;
 414                 offset = 0;
 415         }
 416
 417         res = kernel_readv(in, vec, nr_pages, *ppos);
 418         if (res > 0) {
 419                 copied = res;
 420                 *ppos += res;
 421         }
 422
 423         if (vec != __vec)
 424                 kfree(vec);
 425 out:
 426         for (i = 0; i < nr_pages; i++)
 427                 put_page(pages[i]);
 428         kvfree(pages);
 429         iov_iter_advance(&to, copied);  /* truncates and discards */
 430         return res;
 431 }
 432
 433 /*
 434  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 435  * using sendpage(). Return the number of bytes sent.
 436  */
 437 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 438                             struct pipe_buffer *buf, struct splice_desc *sd)
 439 {
 440         struct file *file = sd->u.file;
 441         loff_t pos = sd->pos;
 442         int more;
 443
 444         if (!likely(file->f_op->sendpage))
 445                 return -EINVAL;
 446
 447         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 448
 449         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 450                 more |= MSG_SENDPAGE_NOTLAST;
 451
 452         return file->f_op->sendpage(file, buf->page, buf->offset,
 453                                     sd->len, &pos, more);
 454 }
 455
 456 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 457 {
 458         smp_mb();
 459         if (waitqueue_active(&pipe->wait))
 460                 wake_up_interruptible(&pipe->wait);
 461         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 462 }
 463
 464 /**
 465  * splice_from_pipe_feed - feed available data from a pipe to a file
 466  * @pipe:       pipe to splice from
 467  * @sd:         information to @actor
 468  * @actor:      handler that splices the data
 469  *
 470  * Description:
 471  *    This function loops over the pipe and calls @actor to do the
 472  *    actual moving of a single struct pipe_buffer to the desired
 473  *    destination.  It returns when there's no more buffers left in
 474  *    the pipe or if the requested number of bytes (@sd->total_len)
 475  *    have been copied.  It returns a positive number (one) if the
 476  *    pipe needs to be filled with more data, zero if the required
 477  *    number of bytes have been copied and -errno on error.
 478  *
 479  *    This, together with splice_from_pipe_{begin,end,next}, may be
 480  *    used to implement the functionality of __splice_from_pipe() when
 481  *    locking is required around copying the pipe buffers to the
 482  *    destination.
 483  */
 484 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 485                           splice_actor *actor)
 486 {
 487         int ret;
 488
 489         while (pipe->nrbufs) {
 490                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 491
 492                 sd->len = buf->len;
 493                 if (sd->len > sd->total_len)
 494                         sd->len = sd->total_len;
 495
 496                 ret = pipe_buf_confirm(pipe, buf);
 497                 if (unlikely(ret)) {
 498                         if (ret == -ENODATA)
 499                                 ret = 0;
 500                         return ret;
 501                 }
 502
 503                 ret = actor(pipe, buf, sd);
 504                 if (ret <= 0)
 505                         return ret;
 506
 507                 buf->offset += ret;
 508                 buf->len -= ret;
 509
 510                 sd->num_spliced += ret;
 511                 sd->len -= ret;
 512                 sd->pos += ret;
 513                 sd->total_len -= ret;
 514
 515                 if (!buf->len) {
 516                         pipe_buf_release(pipe, buf);
 517                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 518                         pipe->nrbufs--;
 519                         if (pipe->files)
 520                                 sd->need_wakeup = true;
 521                 }
 522
 523                 if (!sd->total_len)
 524                         return 0;
 525         }
 526
 527         return 1;
 528 }
 529
 530 /**
 531  * splice_from_pipe_next - wait for some data to splice from
 532  * @pipe:       pipe to splice from
 533  * @sd:         information about the splice operation
 534  *
 535  * Description:
 536  *    This function will wait for some data and return a positive
 537  *    value (one) if pipe buffers are available.  It will return zero
 538  *    or -errno if no more data needs to be spliced.
 539  */
 540 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 541 {
 542         /*
 543          * Check for signal early to make process killable when there are
 544          * always buffers available
 545          */
 546         if (signal_pending(current))
 547                 return -ERESTARTSYS;
 548
 549         while (!pipe->nrbufs) {
 550                 if (!pipe->writers)
 551                         return 0;
 552
 553                 if (!pipe->waiting_writers && sd->num_spliced)
 554                         return 0;
 555
 556                 if (sd->flags & SPLICE_F_NONBLOCK)
 557                         return -EAGAIN;
 558
 559                 if (signal_pending(current))
 560                         return -ERESTARTSYS;
 561
 562                 if (sd->need_wakeup) {
 563                         wakeup_pipe_writers(pipe);
 564                         sd->need_wakeup = false;
 565                 }
 566
 567                 pipe_wait(pipe);
 568         }
 569
 570         return 1;
 571 }
 572
 573 /**
 574  * splice_from_pipe_begin - start splicing from pipe
 575  * @sd:         information about the splice operation
 576  *
 577  * Description:
 578  *    This function should be called before a loop containing
 579  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 580  *    initialize the necessary fields of @sd.
 581  */
 582 static void splice_from_pipe_begin(struct splice_desc *sd)
 583 {
 584         sd->num_spliced = 0;
 585         sd->need_wakeup = false;
 586 }
 587
 588 /**
 589  * splice_from_pipe_end - finish splicing from pipe
 590  * @pipe:       pipe to splice from
 591  * @sd:         information about the splice operation
 592  *
 593  * Description:
 594  *    This function will wake up pipe writers if necessary.  It should
 595  *    be called after a loop containing splice_from_pipe_next() and
 596  *    splice_from_pipe_feed().
 597  */
 598 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 599 {
 600         if (sd->need_wakeup)
 601                 wakeup_pipe_writers(pipe);
 602 }
 603
 604 /**
 605  * __splice_from_pipe - splice data from a pipe to given actor
 606  * @pipe:       pipe to splice from
 607  * @sd:         information to @actor
 608  * @actor:      handler that splices the data
 609  *
 610  * Description:
 611  *    This function does little more than loop over the pipe and call
 612  *    @actor to do the actual moving of a single struct pipe_buffer to
 613  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 614  *    pipe_to_user.
 615  *
 616  */
 617 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 618                            splice_actor *actor)
 619 {
 620         int ret;
 621
 622         splice_from_pipe_begin(sd);
 623         do {
 624                 cond_resched();
 625                 ret = splice_from_pipe_next(pipe, sd);
 626                 if (ret > 0)
 627                         ret = splice_from_pipe_feed(pipe, sd, actor);
 628         } while (ret > 0);
 629         splice_from_pipe_end(pipe, sd);
 630
 631         return sd->num_spliced ? sd->num_spliced : ret;
 632 }
 633 EXPORT_SYMBOL(__splice_from_pipe);
 634
 635 /**
 636  * splice_from_pipe - splice data from a pipe to a file
 637  * @pipe:       pipe to splice from
 638  * @out:        file to splice to
 639  * @ppos:       position in @out
 640  * @len:        how many bytes to splice
 641  * @flags:      splice modifier flags
 642  * @actor:      handler that splices the data
 643  *
 644  * Description:
 645  *    See __splice_from_pipe. This function locks the pipe inode,
 646  *    otherwise it's identical to __splice_from_pipe().
 647  *
 648  */
 649 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 650                          loff_t *ppos, size_t len, unsigned int flags,
 651                          splice_actor *actor)
 652 {
 653         ssize_t ret;
 654         struct splice_desc sd = {
 655                 .total_len = len,
 656                 .flags = flags,
 657                 .pos = *ppos,
 658                 .u.file = out,
 659         };
 660
 661         pipe_lock(pipe);
 662         ret = __splice_from_pipe(pipe, &sd, actor);
 663         pipe_unlock(pipe);
 664
 665         return ret;
 666 }
 667
 668 /**
 669  * iter_file_splice_write - splice data from a pipe to a file
 670  * @pipe:       pipe info
 671  * @out:        file to write to
 672  * @ppos:       position in @out
 673  * @len:        number of bytes to splice
 674  * @flags:      splice modifier flags
 675  *
 676  * Description:
 677  *    Will either move or copy pages (determined by @flags options) from
 678  *    the given pipe inode to the given file.
 679  *    This one is ->write_iter-based.
 680  *
 681  */
 682 ssize_t
 683 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 684                           loff_t *ppos, size_t len, unsigned int flags)
 685 {
 686         struct splice_desc sd = {
 687                 .total_len = len,
 688                 .flags = flags,
 689                 .pos = *ppos,
 690                 .u.file = out,
 691         };
 692         int nbufs = pipe->buffers;
 693         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 694                                         GFP_KERNEL);
 695         ssize_t ret;
 696
 697         if (unlikely(!array))
 698                 return -ENOMEM;
 699
 700         pipe_lock(pipe);
 701
 702         splice_from_pipe_begin(&sd);
 703         while (sd.total_len) {
 704                 struct iov_iter from;
 705                 size_t left;
 706                 int n, idx;
 707
 708                 ret = splice_from_pipe_next(pipe, &sd);
 709                 if (ret <= 0)
 710                         break;
 711
 712                 if (unlikely(nbufs < pipe->buffers)) {
 713                         kfree(array);
 714                         nbufs = pipe->buffers;
 715                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 716                                         GFP_KERNEL);
 717                         if (!array) {
 718                                 ret = -ENOMEM;
 719                                 break;
 720                         }
 721                 }
 722
 723                 /* build the vector */
 724                 left = sd.total_len;
 725                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 726                         struct pipe_buffer *buf = pipe->bufs + idx;
 727                         size_t this_len = buf->len;
 728
 729                         if (this_len > left)
 730                                 this_len = left;
 731
 732                         if (idx == pipe->buffers - 1)
 733                                 idx = -1;
 734
 735                         ret = pipe_buf_confirm(pipe, buf);
 736                         if (unlikely(ret)) {
 737                                 if (ret == -ENODATA)
 738                                         ret = 0;
 739                                 goto done;
 740                         }
 741
 742                         array[n].bv_page = buf->page;
 743                         array[n].bv_len = this_len;
 744                         array[n].bv_offset = buf->offset;
 745                         left -= this_len;
 746                 }
 747
 748                 iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
 749                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 750                 if (ret <= 0)
 751                         break;
 752
 753                 sd.num_spliced += ret;
 754                 sd.total_len -= ret;
 755                 *ppos = sd.pos;
 756
 757                 /* dismiss the fully eaten buffers, adjust the partial one */
 758                 while (ret) {
 759                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 760                         if (ret >= buf->len) {
 761                                 ret -= buf->len;
 762                                 buf->len = 0;
 763                                 pipe_buf_release(pipe, buf);
 764                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 765                                 pipe->nrbufs--;
 766                                 if (pipe->files)
 767                                         sd.need_wakeup = true;
 768                         } else {
 769                                 buf->offset += ret;
 770                                 buf->len -= ret;
 771                                 ret = 0;
 772                         }
 773                 }
 774         }
 775 done:
 776         kfree(array);
 777         splice_from_pipe_end(pipe, &sd);
 778
 779         pipe_unlock(pipe);
 780
 781         if (sd.num_spliced)
 782                 ret = sd.num_spliced;
 783
 784         return ret;
 785 }
 786
 787 EXPORT_SYMBOL(iter_file_splice_write);
 788
 789 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 790                           struct splice_desc *sd)
 791 {
 792         int ret;
 793         void *data;
 794         loff_t tmp = sd->pos;
 795
 796         data = kmap(buf->page);
 797         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 798         kunmap(buf->page);
 799
 800         return ret;
 801 }
 802
 803 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 804                                          struct file *out, loff_t *ppos,
 805                                          size_t len, unsigned int flags)
 806 {
 807         ssize_t ret;
 808
 809         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 810         if (ret > 0)
 811                 *ppos += ret;
 812
 813         return ret;
 814 }
 815
 816 /**
 817  * generic_splice_sendpage - splice data from a pipe to a socket
 818  * @pipe:       pipe to splice from
 819  * @out:        socket to write to
 820  * @ppos:       position in @out
 821  * @len:        number of bytes to splice
 822  * @flags:      splice modifier flags
 823  *
 824  * Description:
 825  *    Will send @len bytes from the pipe to a network socket. No data copying
 826  *    is involved.
 827  *
 828  */
 829 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 830                                 loff_t *ppos, size_t len, unsigned int flags)
 831 {
 832         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 833 }
 834
 835 EXPORT_SYMBOL(generic_splice_sendpage);
 836
 837 /*
 838  * Attempt to initiate a splice from pipe to file.
 839  */
 840 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 841                            loff_t *ppos, size_t len, unsigned int flags)
 842 {
 843         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 844                                 loff_t *, size_t, unsigned int);
 845
 846         if (out->f_op->splice_write)
 847                 splice_write = out->f_op->splice_write;
 848         else
 849                 splice_write = default_file_splice_write;
 850
 851         return splice_write(pipe, out, ppos, len, flags);
 852 }
 853
 854 /*
 855  * Attempt to initiate a splice from a file to a pipe.
 856  */
 857 static long do_splice_to(struct file *in, loff_t *ppos,
 858                          struct pipe_inode_info *pipe, size_t len,
 859                          unsigned int flags)
 860 {
 861         ssize_t (*splice_read)(struct file *, loff_t *,
 862                                struct pipe_inode_info *, size_t, unsigned int);
 863         int ret;
 864
 865         if (unlikely(!(in->f_mode & FMODE_READ)))
 866                 return -EBADF;
 867
 868         ret = rw_verify_area(READ, in, ppos, len);
 869         if (unlikely(ret < 0))
 870                 return ret;
 871
 872         if (unlikely(len > MAX_RW_COUNT))
 873                 len = MAX_RW_COUNT;
 874
 875         if (in->f_op->splice_read)
 876                 splice_read = in->f_op->splice_read;
 877         else
 878                 splice_read = default_file_splice_read;
 879
 880         return splice_read(in, ppos, pipe, len, flags);
 881 }
 882
 883 /**
 884  * splice_direct_to_actor - splices data directly between two non-pipes
 885  * @in:         file to splice from
 886  * @sd:         actor information on where to splice to
 887  * @actor:      handles the data splicing
 888  *
 889  * Description:
 890  *    This is a special case helper to splice directly between two
 891  *    points, without requiring an explicit pipe. Internally an allocated
 892  *    pipe is cached in the process, and reused during the lifetime of
 893  *    that process.
 894  *
 895  */
 896 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 897                                splice_direct_actor *actor)
 898 {
 899         struct pipe_inode_info *pipe;
 900         long ret, bytes;
 901         umode_t i_mode;
 902         size_t len;
 903         int i, flags, more;
 904
 905         /*
 906          * We require the input being a regular file, as we don't want to
 907          * randomly drop data for eg socket -> socket splicing. Use the
 908          * piped splicing for that!
 909          */
 910         i_mode = file_inode(in)->i_mode;
 911         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 912                 return -EINVAL;
 913
 914         /*
 915          * neither in nor out is a pipe, setup an internal pipe attached to
 916          * 'out' and transfer the wanted data from 'in' to 'out' through that
 917          */
 918         pipe = current->splice_pipe;
 919         if (unlikely(!pipe)) {
 920                 pipe = alloc_pipe_info();
 921                 if (!pipe)
 922                         return -ENOMEM;
 923
 924                 /*
 925                  * We don't have an immediate reader, but we'll read the stuff
 926                  * out of the pipe right after the splice_to_pipe(). So set
 927                  * PIPE_READERS appropriately.
 928                  */
 929                 pipe->readers = 1;
 930
 931                 current->splice_pipe = pipe;
 932         }
 933
 934         /*
 935          * Do the splice.
 936          */
 937         ret = 0;
 938         bytes = 0;
 939         len = sd->total_len;
 940         flags = sd->flags;
 941
 942         /*
 943          * Don't block on output, we have to drain the direct pipe.
 944          */
 945         sd->flags &= ~SPLICE_F_NONBLOCK;
 946         more = sd->flags & SPLICE_F_MORE;
 947
 948         while (len) {
 949                 size_t read_len;
 950                 loff_t pos = sd->pos, prev_pos = pos;
 951
 952                 ret = do_splice_to(in, &pos, pipe, len, flags);
 953                 if (unlikely(ret <= 0))
 954                         goto out_release;
 955
 956                 read_len = ret;
 957                 sd->total_len = read_len;
 958
 959                 /*
 960                  * If more data is pending, set SPLICE_F_MORE
 961                  * If this is the last data and SPLICE_F_MORE was not set
 962                  * initially, clears it.
 963                  */
 964                 if (read_len < len)
 965                         sd->flags |= SPLICE_F_MORE;
 966                 else if (!more)
 967                         sd->flags &= ~SPLICE_F_MORE;
 968                 /*
 969                  * NOTE: nonblocking mode only applies to the input. We
 970                  * must not do the output in nonblocking mode as then we
 971                  * could get stuck data in the internal pipe:
 972                  */
 973                 ret = actor(pipe, sd);
 974                 if (unlikely(ret <= 0)) {
 975                         sd->pos = prev_pos;
 976                         goto out_release;
 977                 }
 978
 979                 bytes += ret;
 980                 len -= ret;
 981                 sd->pos = pos;
 982
 983                 if (ret < read_len) {
 984                         sd->pos = prev_pos + ret;
 985                         goto out_release;
 986                 }
 987         }
 988
 989 done:
 990         pipe->nrbufs = pipe->curbuf = 0;
 991         file_accessed(in);
 992         return bytes;
 993
 994 out_release:
 995         /*
 996          * If we did an incomplete transfer we must release
 997          * the pipe buffers in question:
 998          */
 999         for (i = 0; i < pipe->buffers; i++) {
1000                 struct pipe_buffer *buf = pipe->bufs + i;
1001
1002                 if (buf->ops)
1003                         pipe_buf_release(pipe, buf);
1004         }
1005
1006         if (!bytes)
1007                 bytes = ret;
1008
1009         goto done;
1010 }
1011 EXPORT_SYMBOL(splice_direct_to_actor);
1012
1013 static int direct_splice_actor(struct pipe_inode_info *pipe,
1014                                struct splice_desc *sd)
1015 {
1016         struct file *file = sd->u.file;
1017
1018         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1019                               sd->flags);
1020 }
1021
1022 /**
1023  * do_splice_direct - splices data directly between two files
1024  * @in:         file to splice from
1025  * @ppos:       input file offset
1026  * @out:        file to splice to
1027  * @opos:       output file offset
1028  * @len:        number of bytes to splice
1029  * @flags:      splice modifier flags
1030  *
1031  * Description:
1032  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1033  *    doing it in the application would incur an extra system call
1034  *    (splice in + splice out, as compared to just sendfile()). So this helper
1035  *    can splice directly through a process-private pipe.
1036  *
1037  */
1038 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1039                       loff_t *opos, size_t len, unsigned int flags)
1040 {
1041         struct splice_desc sd = {
1042                 .len            = len,
1043                 .total_len      = len,
1044                 .flags          = flags,
1045                 .pos            = *ppos,
1046                 .u.file         = out,
1047                 .opos           = opos,
1048         };
1049         long ret;
1050
1051         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1052                 return -EBADF;
1053
1054         if (unlikely(out->f_flags & O_APPEND))
1055                 return -EINVAL;
1056
1057         ret = rw_verify_area(WRITE, out, opos, len);
1058         if (unlikely(ret < 0))
1059                 return ret;
1060
1061         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1062         if (ret > 0)
1063                 *ppos = sd.pos;
1064
1065         return ret;
1066 }
1067 EXPORT_SYMBOL(do_splice_direct);
1068
1069 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1070 {
1071         for (;;) {
1072                 if (unlikely(!pipe->readers)) {
1073                         send_sig(SIGPIPE, current, 0);
1074                         return -EPIPE;
1075                 }
1076                 if (pipe->nrbufs != pipe->buffers)
1077                         return 0;
1078                 if (flags & SPLICE_F_NONBLOCK)
1079                         return -EAGAIN;
1080                 if (signal_pending(current))
1081                         return -ERESTARTSYS;
1082                 pipe->waiting_writers++;
1083                 pipe_wait(pipe);
1084                 pipe->waiting_writers--;
1085         }
1086 }
1087
1088 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1089                                struct pipe_inode_info *opipe,
1090                                size_t len, unsigned int flags);
1091
1092 /*
1093  * Determine where to splice to/from.
1094  */
1095 static long do_splice(struct file *in, loff_t __user *off_in,
1096                       struct file *out, loff_t __user *off_out,
1097                       size_t len, unsigned int flags)
1098 {
1099         struct pipe_inode_info *ipipe;
1100         struct pipe_inode_info *opipe;
1101         loff_t offset;
1102         long ret;
1103
1104         ipipe = get_pipe_info(in);
1105         opipe = get_pipe_info(out);
1106
1107         if (ipipe && opipe) {
1108                 if (off_in || off_out)
1109                         return -ESPIPE;
1110
1111                 if (!(in->f_mode & FMODE_READ))
1112                         return -EBADF;
1113
1114                 if (!(out->f_mode & FMODE_WRITE))
1115                         return -EBADF;
1116
1117                 /* Splicing to self would be fun, but... */
1118                 if (ipipe == opipe)
1119                         return -EINVAL;
1120
1121                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1122         }
1123
1124         if (ipipe) {
1125                 if (off_in)
1126                         return -ESPIPE;
1127                 if (off_out) {
1128                         if (!(out->f_mode & FMODE_PWRITE))
1129                                 return -EINVAL;
1130                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1131                                 return -EFAULT;
1132                 } else {
1133                         offset = out->f_pos;
1134                 }
1135
1136                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1137                         return -EBADF;
1138
1139                 if (unlikely(out->f_flags & O_APPEND))
1140                         return -EINVAL;
1141
1142                 ret = rw_verify_area(WRITE, out, &offset, len);
1143                 if (unlikely(ret < 0))
1144                         return ret;
1145
1146                 file_start_write(out);
1147                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1148                 file_end_write(out);
1149
1150                 if (!off_out)
1151                         out->f_pos = offset;
1152                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1153                         ret = -EFAULT;
1154
1155                 return ret;
1156         }
1157
1158         if (opipe) {
1159                 if (off_out)
1160                         return -ESPIPE;
1161                 if (off_in) {
1162                         if (!(in->f_mode & FMODE_PREAD))
1163                                 return -EINVAL;
1164                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1165                                 return -EFAULT;
1166                 } else {
1167                         offset = in->f_pos;
1168                 }
1169
1170                 pipe_lock(opipe);
1171                 ret = wait_for_space(opipe, flags);
1172                 if (!ret)
1173                         ret = do_splice_to(in, &offset, opipe, len, flags);
1174                 pipe_unlock(opipe);
1175                 if (ret > 0)
1176                         wakeup_pipe_readers(opipe);
1177                 if (!off_in)
1178                         in->f_pos = offset;
1179                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1180                         ret = -EFAULT;
1181
1182                 return ret;
1183         }
1184
1185         return -EINVAL;
1186 }
1187
1188 static int iter_to_pipe(struct iov_iter *from,
1189                         struct pipe_inode_info *pipe,
1190                         unsigned flags)
1191 {
1192         struct pipe_buffer buf = {
1193                 .ops = &user_page_pipe_buf_ops,
1194                 .flags = flags
1195         };
1196         size_t total = 0;
1197         int ret = 0;
1198         bool failed = false;
1199
1200         while (iov_iter_count(from) && !failed) {
1201                 struct page *pages[16];
1202                 ssize_t copied;
1203                 size_t start;
1204                 int n;
1205
1206                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1207                 if (copied <= 0) {
1208                         ret = copied;
1209                         break;
1210                 }
1211
1212                 for (n = 0; copied; n++, start = 0) {
1213                         int size = min_t(int, copied, PAGE_SIZE - start);
1214                         if (!failed) {
1215                                 buf.page = pages[n];
1216                                 buf.offset = start;
1217                                 buf.len = size;
1218                                 ret = add_to_pipe(pipe, &buf);
1219                                 if (unlikely(ret < 0)) {
1220                                         failed = true;
1221                                 } else {
1222                                         iov_iter_advance(from, ret);
1223                                         total += ret;
1224                                 }
1225                         } else {
1226                                 put_page(pages[n]);
1227                         }
1228                         copied -= size;
1229                 }
1230         }
1231         return total ? total : ret;
1232 }
1233
1234 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1235                         struct splice_desc *sd)
1236 {
1237         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1238         return n == sd->len ? n : -EFAULT;
1239 }
1240
1241 /*
1242  * For lack of a better implementation, implement vmsplice() to userspace
1243  * as a simple copy of the pipes pages to the user iov.
1244  */
1245 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1246                              unsigned int flags)
1247 {
1248         struct pipe_inode_info *pipe = get_pipe_info(file);
1249         struct splice_desc sd = {
1250                 .total_len = iov_iter_count(iter),
1251                 .flags = flags,
1252                 .u.data = iter
1253         };
1254         long ret = 0;
1255
1256         if (!pipe)
1257                 return -EBADF;
1258
1259         if (sd.total_len) {
1260                 pipe_lock(pipe);
1261                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1262                 pipe_unlock(pipe);
1263         }
1264
1265         return ret;
1266 }
1267
1268 /*
1269  * vmsplice splices a user address range into a pipe. It can be thought of
1270  * as splice-from-memory, where the regular splice is splice-from-file (or
1271  * to file). In both cases the output is a pipe, naturally.
1272  */
1273 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1274                              unsigned int flags)
1275 {
1276         struct pipe_inode_info *pipe;
1277         long ret = 0;
1278         unsigned buf_flag = 0;
1279
1280         if (flags & SPLICE_F_GIFT)
1281                 buf_flag = PIPE_BUF_FLAG_GIFT;
1282
1283         pipe = get_pipe_info(file);
1284         if (!pipe)
1285                 return -EBADF;
1286
1287         pipe_lock(pipe);
1288         ret = wait_for_space(pipe, flags);
1289         if (!ret)
1290                 ret = iter_to_pipe(iter, pipe, buf_flag);
1291         pipe_unlock(pipe);
1292         if (ret > 0)
1293                 wakeup_pipe_readers(pipe);
1294         return ret;
1295 }
1296
1297 static int vmsplice_type(struct fd f, int *type)
1298 {
1299         if (!f.file)
1300                 return -EBADF;
1301         if (f.file->f_mode & FMODE_WRITE) {
1302                 *type = WRITE;
1303         } else if (f.file->f_mode & FMODE_READ) {
1304                 *type = READ;
1305         } else {
1306                 fdput(f);
1307                 return -EBADF;
1308         }
1309         return 0;
1310 }
1311
1312 /*
1313  * Note that vmsplice only really supports true splicing _from_ user memory
1314  * to a pipe, not the other way around. Splicing from user memory is a simple
1315  * operation that can be supported without any funky alignment restrictions
1316  * or nasty vm tricks. We simply map in the user memory and fill them into
1317  * a pipe. The reverse isn't quite as easy, though. There are two possible
1318  * solutions for that:
1319  *
1320  *      - memcpy() the data internally, at which point we might as well just
1321  *        do a regular read() on the buffer anyway.
1322  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1323  *        has restriction limitations on both ends of the pipe).
1324  *
1325  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1326  *
1327  */
1328 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1329 {
1330         if (unlikely(flags & ~SPLICE_F_ALL))
1331                 return -EINVAL;
1332
1333         if (!iov_iter_count(iter))
1334                 return 0;
1335
1336         if (iov_iter_rw(iter) == WRITE)
1337                 return vmsplice_to_pipe(f, iter, flags);
1338         else
1339                 return vmsplice_to_user(f, iter, flags);
1340 }
1341
1342 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1343                 unsigned long, nr_segs, unsigned int, flags)
1344 {
1345         struct iovec iovstack[UIO_FASTIOV];
1346         struct iovec *iov = iovstack;
1347         struct iov_iter iter;
1348         long error;
1349         struct fd f;
1350         int type;
1351
1352         f = fdget(fd);
1353         error = vmsplice_type(f, &type);
1354         if (error)
1355                 return error;
1356
1357         error = import_iovec(type, uiov, nr_segs,
1358                              ARRAY_SIZE(iovstack), &iov, &iter);
1359         if (!error) {
1360                 error = do_vmsplice(f.file, &iter, flags);
1361                 kfree(iov);
1362         }
1363         fdput(f);
1364         return error;
1365 }
1366
1367 #ifdef CONFIG_COMPAT
1368 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1369                     unsigned int, nr_segs, unsigned int, flags)
1370 {
1371         struct iovec iovstack[UIO_FASTIOV];
1372         struct iovec *iov = iovstack;
1373         struct iov_iter iter;
1374         long error;
1375         struct fd f;
1376         int type;
1377
1378         f = fdget(fd);
1379         error = vmsplice_type(f, &type);
1380         if (error)
1381                 return error;
1382
1383         error = compat_import_iovec(type, iov32, nr_segs,
1384                              ARRAY_SIZE(iovstack), &iov, &iter);
1385         if (!error) {
1386                 error = do_vmsplice(f.file, &iter, flags);
1387                 kfree(iov);
1388         }
1389         fdput(f);
1390         return error;
1391 }
1392 #endif
1393
1394 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1395                 int, fd_out, loff_t __user *, off_out,
1396                 size_t, len, unsigned int, flags)
1397 {
1398         struct fd in, out;
1399         long error;
1400
1401         if (unlikely(!len))
1402                 return 0;
1403
1404         if (unlikely(flags & ~SPLICE_F_ALL))
1405                 return -EINVAL;
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(flags & ~SPLICE_F_ALL))
1736                 return -EINVAL;
1737
1738         if (unlikely(!len))
1739                 return 0;
1740
1741         error = -EBADF;
1742         in = fdget(fdin);
1743         if (in.file) {
1744                 if (in.file->f_mode & FMODE_READ) {
1745                         struct fd out = fdget(fdout);
1746                         if (out.file) {
1747                                 if (out.file->f_mode & FMODE_WRITE)
1748                                         error = do_tee(in.file, out.file,
1749                                                         len, flags);
1750                                 fdput(out);
1751                         }
1752                 }
1753                 fdput(in);
1754         }
1755
1756         return error;
1757 }