fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  17  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/security.h>
  34 #include <linux/gfp.h>
  35 #include <linux/socket.h>
  36 #include <linux/compat.h>
  37 #include <linux/sched/signal.h>
  38
  39 #include "internal.h"
  40
  41 /*
  42  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  43  * a vm helper function, it's already simplified quite a bit by the
  44  * addition of remove_mapping(). If success is returned, the caller may
  45  * attempt to reuse this page for another destination.
  46  */
  47 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  48                                      struct pipe_buffer *buf)
  49 {
  50         struct page *page = buf->page;
  51         struct address_space *mapping;
  52
  53         lock_page(page);
  54
  55         mapping = page_mapping(page);
  56         if (mapping) {
  57                 WARN_ON(!PageUptodate(page));
  58
  59                 /*
  60                  * At least for ext2 with nobh option, we need to wait on
  61                  * writeback completing on this page, since we'll remove it
  62                  * from the pagecache.  Otherwise truncate wont wait on the
  63                  * page, allowing the disk blocks to be reused by someone else
  64                  * before we actually wrote our data to them. fs corruption
  65                  * ensues.
  66                  */
  67                 wait_on_page_writeback(page);
  68
  69                 if (page_has_private(page) &&
  70                     !try_to_release_page(page, GFP_KERNEL))
  71                         goto out_unlock;
  72
  73                 /*
  74                  * If we succeeded in removing the mapping, set LRU flag
  75                  * and return good.
  76                  */
  77                 if (remove_mapping(mapping, page)) {
  78                         buf->flags |= PIPE_BUF_FLAG_LRU;
  79                         return 0;
  80                 }
  81         }
  82
  83         /*
  84          * Raced with truncate or failed to remove page from current
  85          * address space, unlock and return failure.
  86          */
  87 out_unlock:
  88         unlock_page(page);
  89         return 1;
  90 }
  91
  92 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  93                                         struct pipe_buffer *buf)
  94 {
  95         put_page(buf->page);
  96         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  97 }
  98
  99 /*
 100  * Check whether the contents of buf is OK to access. Since the content
 101  * is a page cache page, IO may be in flight.
 102  */
 103 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 104                                        struct pipe_buffer *buf)
 105 {
 106         struct page *page = buf->page;
 107         int err;
 108
 109         if (!PageUptodate(page)) {
 110                 lock_page(page);
 111
 112                 /*
 113                  * Page got truncated/unhashed. This will cause a 0-byte
 114                  * splice, if this is the first page.
 115                  */
 116                 if (!page->mapping) {
 117                         err = -ENODATA;
 118                         goto error;
 119                 }
 120
 121                 /*
 122                  * Uh oh, read-error from disk.
 123                  */
 124                 if (!PageUptodate(page)) {
 125                         err = -EIO;
 126                         goto error;
 127                 }
 128
 129                 /*
 130                  * Page is ok afterall, we are done.
 131                  */
 132                 unlock_page(page);
 133         }
 134
 135         return 0;
 136 error:
 137         unlock_page(page);
 138         return err;
 139 }
 140
 141 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 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         .confirm = generic_pipe_buf_confirm,
 160         .release = page_cache_pipe_buf_release,
 161         .steal = user_page_pipe_buf_steal,
 162         .get = generic_pipe_buf_get,
 163 };
 164
 165 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 166 {
 167         smp_mb();
 168         if (waitqueue_active(&pipe->wait))
 169                 wake_up_interruptible(&pipe->wait);
 170         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 171 }
 172
 173 /**
 174  * splice_to_pipe - fill passed data into a pipe
 175  * @pipe:       pipe to fill
 176  * @spd:        data to fill
 177  *
 178  * Description:
 179  *    @spd contains a map of pages and len/offset tuples, along with
 180  *    the struct pipe_buf_operations associated with these pages. This
 181  *    function will link that data to the pipe.
 182  *
 183  */
 184 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 185                        struct splice_pipe_desc *spd)
 186 {
 187         unsigned int spd_pages = spd->nr_pages;
 188         unsigned int tail = pipe->tail;
 189         unsigned int head = pipe->head;
 190         unsigned int mask = pipe->ring_size - 1;
 191         int ret = 0, page_nr = 0;
 192
 193         if (!spd_pages)
 194                 return 0;
 195
 196         if (unlikely(!pipe->readers)) {
 197                 send_sig(SIGPIPE, current, 0);
 198                 ret = -EPIPE;
 199                 goto out;
 200         }
 201
 202         while (!pipe_full(head, tail, pipe->max_usage)) {
 203                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 204
 205                 buf->page = spd->pages[page_nr];
 206                 buf->offset = spd->partial[page_nr].offset;
 207                 buf->len = spd->partial[page_nr].len;
 208                 buf->private = spd->partial[page_nr].private;
 209                 buf->ops = spd->ops;
 210                 buf->flags = 0;
 211
 212                 head++;
 213                 pipe->head = head;
 214                 page_nr++;
 215                 ret += buf->len;
 216
 217                 if (!--spd->nr_pages)
 218                         break;
 219         }
 220
 221         if (!ret)
 222                 ret = -EAGAIN;
 223
 224 out:
 225         while (page_nr < spd_pages)
 226                 spd->spd_release(spd, page_nr++);
 227
 228         return ret;
 229 }
 230 EXPORT_SYMBOL_GPL(splice_to_pipe);
 231
 232 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 233 {
 234         unsigned int head = pipe->head;
 235         unsigned int tail = pipe->tail;
 236         unsigned int mask = pipe->ring_size - 1;
 237         int ret;
 238
 239         if (unlikely(!pipe->readers)) {
 240                 send_sig(SIGPIPE, current, 0);
 241                 ret = -EPIPE;
 242         } else if (pipe_full(head, tail, pipe->max_usage)) {
 243                 ret = -EAGAIN;
 244         } else {
 245                 pipe->bufs[head & mask] = *buf;
 246                 pipe->head = head + 1;
 247                 return buf->len;
 248         }
 249         pipe_buf_release(pipe, buf);
 250         return ret;
 251 }
 252 EXPORT_SYMBOL(add_to_pipe);
 253
 254 /*
 255  * Check if we need to grow the arrays holding pages and partial page
 256  * descriptions.
 257  */
 258 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 259 {
 260         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 261
 262         spd->nr_pages_max = max_usage;
 263         if (max_usage <= PIPE_DEF_BUFFERS)
 264                 return 0;
 265
 266         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 267         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 268                                      GFP_KERNEL);
 269
 270         if (spd->pages && spd->partial)
 271                 return 0;
 272
 273         kfree(spd->pages);
 274         kfree(spd->partial);
 275         return -ENOMEM;
 276 }
 277
 278 void splice_shrink_spd(struct splice_pipe_desc *spd)
 279 {
 280         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 281                 return;
 282
 283         kfree(spd->pages);
 284         kfree(spd->partial);
 285 }
 286
 287 /**
 288  * generic_file_splice_read - splice data from file to a pipe
 289  * @in:         file to splice from
 290  * @ppos:       position in @in
 291  * @pipe:       pipe to splice to
 292  * @len:        number of bytes to splice
 293  * @flags:      splice modifier flags
 294  *
 295  * Description:
 296  *    Will read pages from given file and fill them into a pipe. Can be
 297  *    used as long as it has more or less sane ->read_iter().
 298  *
 299  */
 300 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 301                                  struct pipe_inode_info *pipe, size_t len,
 302                                  unsigned int flags)
 303 {
 304         struct iov_iter to;
 305         struct kiocb kiocb;
 306         unsigned int i_head;
 307         int ret;
 308
 309         iov_iter_pipe(&to, READ, pipe, len);
 310         i_head = to.head;
 311         init_sync_kiocb(&kiocb, in);
 312         kiocb.ki_pos = *ppos;
 313         ret = call_read_iter(in, &kiocb, &to);
 314         if (ret > 0) {
 315                 *ppos = kiocb.ki_pos;
 316                 file_accessed(in);
 317         } else if (ret < 0) {
 318                 to.head = i_head;
 319                 to.iov_offset = 0;
 320                 iov_iter_advance(&to, 0); /* to free what was emitted */
 321                 /*
 322                  * callers of ->splice_read() expect -EAGAIN on
 323                  * "can't put anything in there", rather than -EFAULT.
 324                  */
 325                 if (ret == -EFAULT)
 326                         ret = -EAGAIN;
 327         }
 328
 329         return ret;
 330 }
 331 EXPORT_SYMBOL(generic_file_splice_read);
 332
 333 const struct pipe_buf_operations default_pipe_buf_ops = {
 334         .confirm = generic_pipe_buf_confirm,
 335         .release = generic_pipe_buf_release,
 336         .steal = generic_pipe_buf_steal,
 337         .get = generic_pipe_buf_get,
 338 };
 339
 340 int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 341                              struct pipe_buffer *buf)
 342 {
 343         return 1;
 344 }
 345
 346 /* Pipe buffer operations for a socket and similar. */
 347 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 348         .confirm = generic_pipe_buf_confirm,
 349         .release = generic_pipe_buf_release,
 350         .steal = generic_pipe_buf_nosteal,
 351         .get = generic_pipe_buf_get,
 352 };
 353 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 354
 355 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 356                             unsigned long vlen, loff_t offset)
 357 {
 358         mm_segment_t old_fs;
 359         loff_t pos = offset;
 360         ssize_t res;
 361
 362         old_fs = get_fs();
 363         set_fs(KERNEL_DS);
 364         /* The cast to a user pointer is valid due to the set_fs() */
 365         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 366         set_fs(old_fs);
 367
 368         return res;
 369 }
 370
 371 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 372                                  struct pipe_inode_info *pipe, size_t len,
 373                                  unsigned int flags)
 374 {
 375         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 376         struct iov_iter to;
 377         struct page **pages;
 378         unsigned int nr_pages;
 379         unsigned int mask;
 380         size_t offset, base, copied = 0;
 381         ssize_t res;
 382         int i;
 383
 384         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 385                 return -EAGAIN;
 386
 387         /*
 388          * Try to keep page boundaries matching to source pagecache ones -
 389          * it probably won't be much help, but...
 390          */
 391         offset = *ppos & ~PAGE_MASK;
 392
 393         iov_iter_pipe(&to, READ, pipe, len + offset);
 394
 395         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 396         if (res <= 0)
 397                 return -ENOMEM;
 398
 399         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 400
 401         vec = __vec;
 402         if (nr_pages > PIPE_DEF_BUFFERS) {
 403                 vec = kmalloc_array(nr_pages, sizeof(struct kvec), GFP_KERNEL);
 404                 if (unlikely(!vec)) {
 405                         res = -ENOMEM;
 406                         goto out;
 407                 }
 408         }
 409
 410         mask = pipe->ring_size - 1;
 411         pipe->bufs[to.head & mask].offset = offset;
 412         pipe->bufs[to.head & mask].len -= offset;
 413
 414         for (i = 0; i < nr_pages; i++) {
 415                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 416                 vec[i].iov_base = page_address(pages[i]) + offset;
 417                 vec[i].iov_len = this_len;
 418                 len -= this_len;
 419                 offset = 0;
 420         }
 421
 422         res = kernel_readv(in, vec, nr_pages, *ppos);
 423         if (res > 0) {
 424                 copied = res;
 425                 *ppos += res;
 426         }
 427
 428         if (vec != __vec)
 429                 kfree(vec);
 430 out:
 431         for (i = 0; i < nr_pages; i++)
 432                 put_page(pages[i]);
 433         kvfree(pages);
 434         iov_iter_advance(&to, copied);  /* truncates and discards */
 435         return res;
 436 }
 437
 438 /*
 439  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 440  * using sendpage(). Return the number of bytes sent.
 441  */
 442 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 443                             struct pipe_buffer *buf, struct splice_desc *sd)
 444 {
 445         struct file *file = sd->u.file;
 446         loff_t pos = sd->pos;
 447         int more;
 448
 449         if (!likely(file->f_op->sendpage))
 450                 return -EINVAL;
 451
 452         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 453
 454         if (sd->len < sd->total_len &&
 455             pipe_occupancy(pipe->head, pipe->tail) > 1)
 456                 more |= MSG_SENDPAGE_NOTLAST;
 457
 458         return file->f_op->sendpage(file, buf->page, buf->offset,
 459                                     sd->len, &pos, more);
 460 }
 461
 462 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 463 {
 464         smp_mb();
 465         if (waitqueue_active(&pipe->wait))
 466                 wake_up_interruptible(&pipe->wait);
 467         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 468 }
 469
 470 /**
 471  * splice_from_pipe_feed - feed available data from a pipe to a file
 472  * @pipe:       pipe to splice from
 473  * @sd:         information to @actor
 474  * @actor:      handler that splices the data
 475  *
 476  * Description:
 477  *    This function loops over the pipe and calls @actor to do the
 478  *    actual moving of a single struct pipe_buffer to the desired
 479  *    destination.  It returns when there's no more buffers left in
 480  *    the pipe or if the requested number of bytes (@sd->total_len)
 481  *    have been copied.  It returns a positive number (one) if the
 482  *    pipe needs to be filled with more data, zero if the required
 483  *    number of bytes have been copied and -errno on error.
 484  *
 485  *    This, together with splice_from_pipe_{begin,end,next}, may be
 486  *    used to implement the functionality of __splice_from_pipe() when
 487  *    locking is required around copying the pipe buffers to the
 488  *    destination.
 489  */
 490 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 491                           splice_actor *actor)
 492 {
 493         unsigned int head = pipe->head;
 494         unsigned int tail = pipe->tail;
 495         unsigned int mask = pipe->ring_size - 1;
 496         int ret;
 497
 498         while (!pipe_empty(head, tail)) {
 499                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 500
 501                 sd->len = buf->len;
 502                 if (sd->len > sd->total_len)
 503                         sd->len = sd->total_len;
 504
 505                 ret = pipe_buf_confirm(pipe, buf);
 506                 if (unlikely(ret)) {
 507                         if (ret == -ENODATA)
 508                                 ret = 0;
 509                         return ret;
 510                 }
 511
 512                 ret = actor(pipe, buf, sd);
 513                 if (ret <= 0)
 514                         return ret;
 515
 516                 buf->offset += ret;
 517                 buf->len -= ret;
 518
 519                 sd->num_spliced += ret;
 520                 sd->len -= ret;
 521                 sd->pos += ret;
 522                 sd->total_len -= ret;
 523
 524                 if (!buf->len) {
 525                         pipe_buf_release(pipe, buf);
 526                         tail++;
 527                         pipe->tail = tail;
 528                         if (pipe->files)
 529                                 sd->need_wakeup = true;
 530                 }
 531
 532                 if (!sd->total_len)
 533                         return 0;
 534         }
 535
 536         return 1;
 537 }
 538
 539 /**
 540  * splice_from_pipe_next - wait for some data to splice from
 541  * @pipe:       pipe to splice from
 542  * @sd:         information about the splice operation
 543  *
 544  * Description:
 545  *    This function will wait for some data and return a positive
 546  *    value (one) if pipe buffers are available.  It will return zero
 547  *    or -errno if no more data needs to be spliced.
 548  */
 549 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 550 {
 551         /*
 552          * Check for signal early to make process killable when there are
 553          * always buffers available
 554          */
 555         if (signal_pending(current))
 556                 return -ERESTARTSYS;
 557
 558         while (pipe_empty(pipe->head, pipe->tail)) {
 559                 if (!pipe->writers)
 560                         return 0;
 561
 562                 if (sd->num_spliced)
 563                         return 0;
 564
 565                 if (sd->flags & SPLICE_F_NONBLOCK)
 566                         return -EAGAIN;
 567
 568                 if (signal_pending(current))
 569                         return -ERESTARTSYS;
 570
 571                 if (sd->need_wakeup) {
 572                         wakeup_pipe_writers(pipe);
 573                         sd->need_wakeup = false;
 574                 }
 575
 576                 pipe_wait(pipe);
 577         }
 578
 579         return 1;
 580 }
 581
 582 /**
 583  * splice_from_pipe_begin - start splicing from pipe
 584  * @sd:         information about the splice operation
 585  *
 586  * Description:
 587  *    This function should be called before a loop containing
 588  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 589  *    initialize the necessary fields of @sd.
 590  */
 591 static void splice_from_pipe_begin(struct splice_desc *sd)
 592 {
 593         sd->num_spliced = 0;
 594         sd->need_wakeup = false;
 595 }
 596
 597 /**
 598  * splice_from_pipe_end - finish splicing from pipe
 599  * @pipe:       pipe to splice from
 600  * @sd:         information about the splice operation
 601  *
 602  * Description:
 603  *    This function will wake up pipe writers if necessary.  It should
 604  *    be called after a loop containing splice_from_pipe_next() and
 605  *    splice_from_pipe_feed().
 606  */
 607 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 608 {
 609         if (sd->need_wakeup)
 610                 wakeup_pipe_writers(pipe);
 611 }
 612
 613 /**
 614  * __splice_from_pipe - splice data from a pipe to given actor
 615  * @pipe:       pipe to splice from
 616  * @sd:         information to @actor
 617  * @actor:      handler that splices the data
 618  *
 619  * Description:
 620  *    This function does little more than loop over the pipe and call
 621  *    @actor to do the actual moving of a single struct pipe_buffer to
 622  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 623  *    pipe_to_user.
 624  *
 625  */
 626 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 627                            splice_actor *actor)
 628 {
 629         int ret;
 630
 631         splice_from_pipe_begin(sd);
 632         do {
 633                 cond_resched();
 634                 ret = splice_from_pipe_next(pipe, sd);
 635                 if (ret > 0)
 636                         ret = splice_from_pipe_feed(pipe, sd, actor);
 637         } while (ret > 0);
 638         splice_from_pipe_end(pipe, sd);
 639
 640         return sd->num_spliced ? sd->num_spliced : ret;
 641 }
 642 EXPORT_SYMBOL(__splice_from_pipe);
 643
 644 /**
 645  * splice_from_pipe - splice data from a pipe to a file
 646  * @pipe:       pipe to splice from
 647  * @out:        file to splice to
 648  * @ppos:       position in @out
 649  * @len:        how many bytes to splice
 650  * @flags:      splice modifier flags
 651  * @actor:      handler that splices the data
 652  *
 653  * Description:
 654  *    See __splice_from_pipe. This function locks the pipe inode,
 655  *    otherwise it's identical to __splice_from_pipe().
 656  *
 657  */
 658 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 659                          loff_t *ppos, size_t len, unsigned int flags,
 660                          splice_actor *actor)
 661 {
 662         ssize_t ret;
 663         struct splice_desc sd = {
 664                 .total_len = len,
 665                 .flags = flags,
 666                 .pos = *ppos,
 667                 .u.file = out,
 668         };
 669
 670         pipe_lock(pipe);
 671         ret = __splice_from_pipe(pipe, &sd, actor);
 672         pipe_unlock(pipe);
 673
 674         return ret;
 675 }
 676
 677 /**
 678  * iter_file_splice_write - splice data from a pipe to a file
 679  * @pipe:       pipe info
 680  * @out:        file to write to
 681  * @ppos:       position in @out
 682  * @len:        number of bytes to splice
 683  * @flags:      splice modifier flags
 684  *
 685  * Description:
 686  *    Will either move or copy pages (determined by @flags options) from
 687  *    the given pipe inode to the given file.
 688  *    This one is ->write_iter-based.
 689  *
 690  */
 691 ssize_t
 692 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 693                           loff_t *ppos, size_t len, unsigned int flags)
 694 {
 695         struct splice_desc sd = {
 696                 .total_len = len,
 697                 .flags = flags,
 698                 .pos = *ppos,
 699                 .u.file = out,
 700         };
 701         int nbufs = pipe->max_usage;
 702         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 703                                         GFP_KERNEL);
 704         ssize_t ret;
 705
 706         if (unlikely(!array))
 707                 return -ENOMEM;
 708
 709         pipe_lock(pipe);
 710
 711         splice_from_pipe_begin(&sd);
 712         while (sd.total_len) {
 713                 struct iov_iter from;
 714                 unsigned int head, tail, mask;
 715                 size_t left;
 716                 int n;
 717
 718                 ret = splice_from_pipe_next(pipe, &sd);
 719                 if (ret <= 0)
 720                         break;
 721
 722                 if (unlikely(nbufs < pipe->max_usage)) {
 723                         kfree(array);
 724                         nbufs = pipe->max_usage;
 725                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 726                                         GFP_KERNEL);
 727                         if (!array) {
 728                                 ret = -ENOMEM;
 729                                 break;
 730                         }
 731                 }
 732
 733                 head = pipe->head;
 734                 tail = pipe->tail;
 735                 mask = pipe->ring_size - 1;
 736
 737                 /* build the vector */
 738                 left = sd.total_len;
 739                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++, n++) {
 740                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 741                         size_t this_len = buf->len;
 742
 743                         if (this_len > left)
 744                                 this_len = left;
 745
 746                         ret = pipe_buf_confirm(pipe, buf);
 747                         if (unlikely(ret)) {
 748                                 if (ret == -ENODATA)
 749                                         ret = 0;
 750                                 goto done;
 751                         }
 752
 753                         array[n].bv_page = buf->page;
 754                         array[n].bv_len = this_len;
 755                         array[n].bv_offset = buf->offset;
 756                         left -= this_len;
 757                 }
 758
 759                 iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
 760                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 761                 if (ret <= 0)
 762                         break;
 763
 764                 sd.num_spliced += ret;
 765                 sd.total_len -= ret;
 766                 *ppos = sd.pos;
 767
 768                 /* dismiss the fully eaten buffers, adjust the partial one */
 769                 tail = pipe->tail;
 770                 while (ret) {
 771                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 772                         if (ret >= buf->len) {
 773                                 ret -= buf->len;
 774                                 buf->len = 0;
 775                                 pipe_buf_release(pipe, buf);
 776                                 tail++;
 777                                 pipe->tail = tail;
 778                                 if (pipe->files)
 779                                         sd.need_wakeup = true;
 780                         } else {
 781                                 buf->offset += ret;
 782                                 buf->len -= ret;
 783                                 ret = 0;
 784                         }
 785                 }
 786         }
 787 done:
 788         kfree(array);
 789         splice_from_pipe_end(pipe, &sd);
 790
 791         pipe_unlock(pipe);
 792
 793         if (sd.num_spliced)
 794                 ret = sd.num_spliced;
 795
 796         return ret;
 797 }
 798
 799 EXPORT_SYMBOL(iter_file_splice_write);
 800
 801 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 802                           struct splice_desc *sd)
 803 {
 804         int ret;
 805         void *data;
 806         loff_t tmp = sd->pos;
 807
 808         data = kmap(buf->page);
 809         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 810         kunmap(buf->page);
 811
 812         return ret;
 813 }
 814
 815 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 816                                          struct file *out, loff_t *ppos,
 817                                          size_t len, unsigned int flags)
 818 {
 819         ssize_t ret;
 820
 821         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 822         if (ret > 0)
 823                 *ppos += ret;
 824
 825         return ret;
 826 }
 827
 828 /**
 829  * generic_splice_sendpage - splice data from a pipe to a socket
 830  * @pipe:       pipe to splice from
 831  * @out:        socket to write to
 832  * @ppos:       position in @out
 833  * @len:        number of bytes to splice
 834  * @flags:      splice modifier flags
 835  *
 836  * Description:
 837  *    Will send @len bytes from the pipe to a network socket. No data copying
 838  *    is involved.
 839  *
 840  */
 841 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 842                                 loff_t *ppos, size_t len, unsigned int flags)
 843 {
 844         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 845 }
 846
 847 EXPORT_SYMBOL(generic_splice_sendpage);
 848
 849 /*
 850  * Attempt to initiate a splice from pipe to file.
 851  */
 852 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 853                            loff_t *ppos, size_t len, unsigned int flags)
 854 {
 855         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 856                                 loff_t *, size_t, unsigned int);
 857
 858         if (out->f_op->splice_write)
 859                 splice_write = out->f_op->splice_write;
 860         else
 861                 splice_write = default_file_splice_write;
 862
 863         return splice_write(pipe, out, ppos, len, flags);
 864 }
 865
 866 /*
 867  * Attempt to initiate a splice from a file to a pipe.
 868  */
 869 static long do_splice_to(struct file *in, loff_t *ppos,
 870                          struct pipe_inode_info *pipe, size_t len,
 871                          unsigned int flags)
 872 {
 873         ssize_t (*splice_read)(struct file *, loff_t *,
 874                                struct pipe_inode_info *, size_t, unsigned int);
 875         int ret;
 876
 877         if (unlikely(!(in->f_mode & FMODE_READ)))
 878                 return -EBADF;
 879
 880         ret = rw_verify_area(READ, in, ppos, len);
 881         if (unlikely(ret < 0))
 882                 return ret;
 883
 884         if (unlikely(len > MAX_RW_COUNT))
 885                 len = MAX_RW_COUNT;
 886
 887         if (in->f_op->splice_read)
 888                 splice_read = in->f_op->splice_read;
 889         else
 890                 splice_read = default_file_splice_read;
 891
 892         return splice_read(in, ppos, pipe, len, flags);
 893 }
 894
 895 /**
 896  * splice_direct_to_actor - splices data directly between two non-pipes
 897  * @in:         file to splice from
 898  * @sd:         actor information on where to splice to
 899  * @actor:      handles the data splicing
 900  *
 901  * Description:
 902  *    This is a special case helper to splice directly between two
 903  *    points, without requiring an explicit pipe. Internally an allocated
 904  *    pipe is cached in the process, and reused during the lifetime of
 905  *    that process.
 906  *
 907  */
 908 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 909                                splice_direct_actor *actor)
 910 {
 911         struct pipe_inode_info *pipe;
 912         long ret, bytes;
 913         umode_t i_mode;
 914         size_t len;
 915         int i, flags, more;
 916
 917         /*
 918          * We require the input being a regular file, as we don't want to
 919          * randomly drop data for eg socket -> socket splicing. Use the
 920          * piped splicing for that!
 921          */
 922         i_mode = file_inode(in)->i_mode;
 923         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 924                 return -EINVAL;
 925
 926         /*
 927          * neither in nor out is a pipe, setup an internal pipe attached to
 928          * 'out' and transfer the wanted data from 'in' to 'out' through that
 929          */
 930         pipe = current->splice_pipe;
 931         if (unlikely(!pipe)) {
 932                 pipe = alloc_pipe_info();
 933                 if (!pipe)
 934                         return -ENOMEM;
 935
 936                 /*
 937                  * We don't have an immediate reader, but we'll read the stuff
 938                  * out of the pipe right after the splice_to_pipe(). So set
 939                  * PIPE_READERS appropriately.
 940                  */
 941                 pipe->readers = 1;
 942
 943                 current->splice_pipe = pipe;
 944         }
 945
 946         /*
 947          * Do the splice.
 948          */
 949         ret = 0;
 950         bytes = 0;
 951         len = sd->total_len;
 952         flags = sd->flags;
 953
 954         /*
 955          * Don't block on output, we have to drain the direct pipe.
 956          */
 957         sd->flags &= ~SPLICE_F_NONBLOCK;
 958         more = sd->flags & SPLICE_F_MORE;
 959
 960         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 961
 962         while (len) {
 963                 unsigned int p_space;
 964                 size_t read_len;
 965                 loff_t pos = sd->pos, prev_pos = pos;
 966
 967                 /* Don't try to read more the pipe has space for. */
 968                 p_space = pipe->max_usage -
 969                         pipe_occupancy(pipe->head, pipe->tail);
 970                 read_len = min_t(size_t, len, p_space << PAGE_SHIFT);
 971                 ret = do_splice_to(in, &pos, pipe, read_len, flags);
 972                 if (unlikely(ret <= 0))
 973                         goto out_release;
 974
 975                 read_len = ret;
 976                 sd->total_len = read_len;
 977
 978                 /*
 979                  * If more data is pending, set SPLICE_F_MORE
 980                  * If this is the last data and SPLICE_F_MORE was not set
 981                  * initially, clears it.
 982                  */
 983                 if (read_len < len)
 984                         sd->flags |= SPLICE_F_MORE;
 985                 else if (!more)
 986                         sd->flags &= ~SPLICE_F_MORE;
 987                 /*
 988                  * NOTE: nonblocking mode only applies to the input. We
 989                  * must not do the output in nonblocking mode as then we
 990                  * could get stuck data in the internal pipe:
 991                  */
 992                 ret = actor(pipe, sd);
 993                 if (unlikely(ret <= 0)) {
 994                         sd->pos = prev_pos;
 995                         goto out_release;
 996                 }
 997
 998                 bytes += ret;
 999                 len -= ret;
1000                 sd->pos = pos;
1001
1002                 if (ret < read_len) {
1003                         sd->pos = prev_pos + ret;
1004                         goto out_release;
1005                 }
1006         }
1007
1008 done:
1009         pipe->tail = pipe->head = 0;
1010         file_accessed(in);
1011         return bytes;
1012
1013 out_release:
1014         /*
1015          * If we did an incomplete transfer we must release
1016          * the pipe buffers in question:
1017          */
1018         for (i = 0; i < pipe->ring_size; i++) {
1019                 struct pipe_buffer *buf = &pipe->bufs[i];
1020
1021                 if (buf->ops)
1022                         pipe_buf_release(pipe, buf);
1023         }
1024
1025         if (!bytes)
1026                 bytes = ret;
1027
1028         goto done;
1029 }
1030 EXPORT_SYMBOL(splice_direct_to_actor);
1031
1032 static int direct_splice_actor(struct pipe_inode_info *pipe,
1033                                struct splice_desc *sd)
1034 {
1035         struct file *file = sd->u.file;
1036
1037         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1038                               sd->flags);
1039 }
1040
1041 /**
1042  * do_splice_direct - splices data directly between two files
1043  * @in:         file to splice from
1044  * @ppos:       input file offset
1045  * @out:        file to splice to
1046  * @opos:       output file offset
1047  * @len:        number of bytes to splice
1048  * @flags:      splice modifier flags
1049  *
1050  * Description:
1051  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1052  *    doing it in the application would incur an extra system call
1053  *    (splice in + splice out, as compared to just sendfile()). So this helper
1054  *    can splice directly through a process-private pipe.
1055  *
1056  */
1057 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1058                       loff_t *opos, size_t len, unsigned int flags)
1059 {
1060         struct splice_desc sd = {
1061                 .len            = len,
1062                 .total_len      = len,
1063                 .flags          = flags,
1064                 .pos            = *ppos,
1065                 .u.file         = out,
1066                 .opos           = opos,
1067         };
1068         long ret;
1069
1070         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1071                 return -EBADF;
1072
1073         if (unlikely(out->f_flags & O_APPEND))
1074                 return -EINVAL;
1075
1076         ret = rw_verify_area(WRITE, out, opos, len);
1077         if (unlikely(ret < 0))
1078                 return ret;
1079
1080         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1081         if (ret > 0)
1082                 *ppos = sd.pos;
1083
1084         return ret;
1085 }
1086 EXPORT_SYMBOL(do_splice_direct);
1087
1088 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1089 {
1090         for (;;) {
1091                 if (unlikely(!pipe->readers)) {
1092                         send_sig(SIGPIPE, current, 0);
1093                         return -EPIPE;
1094                 }
1095                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1096                         return 0;
1097                 if (flags & SPLICE_F_NONBLOCK)
1098                         return -EAGAIN;
1099                 if (signal_pending(current))
1100                         return -ERESTARTSYS;
1101                 pipe_wait(pipe);
1102         }
1103 }
1104
1105 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1106                                struct pipe_inode_info *opipe,
1107                                size_t len, unsigned int flags);
1108
1109 /*
1110  * Determine where to splice to/from.
1111  */
1112 static long do_splice(struct file *in, loff_t __user *off_in,
1113                       struct file *out, loff_t __user *off_out,
1114                       size_t len, unsigned int flags)
1115 {
1116         struct pipe_inode_info *ipipe;
1117         struct pipe_inode_info *opipe;
1118         loff_t offset;
1119         long ret;
1120
1121         ipipe = get_pipe_info(in);
1122         opipe = get_pipe_info(out);
1123
1124         if (ipipe && opipe) {
1125                 if (off_in || off_out)
1126                         return -ESPIPE;
1127
1128                 if (!(in->f_mode & FMODE_READ))
1129                         return -EBADF;
1130
1131                 if (!(out->f_mode & FMODE_WRITE))
1132                         return -EBADF;
1133
1134                 /* Splicing to self would be fun, but... */
1135                 if (ipipe == opipe)
1136                         return -EINVAL;
1137
1138                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1139                         flags |= SPLICE_F_NONBLOCK;
1140
1141                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1142         }
1143
1144         if (ipipe) {
1145                 if (off_in)
1146                         return -ESPIPE;
1147                 if (off_out) {
1148                         if (!(out->f_mode & FMODE_PWRITE))
1149                                 return -EINVAL;
1150                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1151                                 return -EFAULT;
1152                 } else {
1153                         offset = out->f_pos;
1154                 }
1155
1156                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1157                         return -EBADF;
1158
1159                 if (unlikely(out->f_flags & O_APPEND))
1160                         return -EINVAL;
1161
1162                 ret = rw_verify_area(WRITE, out, &offset, len);
1163                 if (unlikely(ret < 0))
1164                         return ret;
1165
1166                 if (in->f_flags & O_NONBLOCK)
1167                         flags |= SPLICE_F_NONBLOCK;
1168
1169                 file_start_write(out);
1170                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1171                 file_end_write(out);
1172
1173                 if (!off_out)
1174                         out->f_pos = offset;
1175                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1176                         ret = -EFAULT;
1177
1178                 return ret;
1179         }
1180
1181         if (opipe) {
1182                 if (off_out)
1183                         return -ESPIPE;
1184                 if (off_in) {
1185                         if (!(in->f_mode & FMODE_PREAD))
1186                                 return -EINVAL;
1187                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1188                                 return -EFAULT;
1189                 } else {
1190                         offset = in->f_pos;
1191                 }
1192
1193                 if (out->f_flags & O_NONBLOCK)
1194                         flags |= SPLICE_F_NONBLOCK;
1195
1196                 pipe_lock(opipe);
1197                 ret = wait_for_space(opipe, flags);
1198                 if (!ret) {
1199                         unsigned int p_space;
1200
1201                         /* Don't try to read more the pipe has space for. */
1202                         p_space = opipe->max_usage - pipe_occupancy(opipe->head, opipe->tail);
1203                         len = min_t(size_t, len, p_space << PAGE_SHIFT);
1204
1205                         ret = do_splice_to(in, &offset, opipe, len, flags);
1206                 }
1207                 pipe_unlock(opipe);
1208                 if (ret > 0)
1209                         wakeup_pipe_readers(opipe);
1210                 if (!off_in)
1211                         in->f_pos = offset;
1212                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1213                         ret = -EFAULT;
1214
1215                 return ret;
1216         }
1217
1218         return -EINVAL;
1219 }
1220
1221 static int iter_to_pipe(struct iov_iter *from,
1222                         struct pipe_inode_info *pipe,
1223                         unsigned flags)
1224 {
1225         struct pipe_buffer buf = {
1226                 .ops = &user_page_pipe_buf_ops,
1227                 .flags = flags
1228         };
1229         size_t total = 0;
1230         int ret = 0;
1231         bool failed = false;
1232
1233         while (iov_iter_count(from) && !failed) {
1234                 struct page *pages[16];
1235                 ssize_t copied;
1236                 size_t start;
1237                 int n;
1238
1239                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1240                 if (copied <= 0) {
1241                         ret = copied;
1242                         break;
1243                 }
1244
1245                 for (n = 0; copied; n++, start = 0) {
1246                         int size = min_t(int, copied, PAGE_SIZE - start);
1247                         if (!failed) {
1248                                 buf.page = pages[n];
1249                                 buf.offset = start;
1250                                 buf.len = size;
1251                                 ret = add_to_pipe(pipe, &buf);
1252                                 if (unlikely(ret < 0)) {
1253                                         failed = true;
1254                                 } else {
1255                                         iov_iter_advance(from, ret);
1256                                         total += ret;
1257                                 }
1258                         } else {
1259                                 put_page(pages[n]);
1260                         }
1261                         copied -= size;
1262                 }
1263         }
1264         return total ? total : ret;
1265 }
1266
1267 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1268                         struct splice_desc *sd)
1269 {
1270         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1271         return n == sd->len ? n : -EFAULT;
1272 }
1273
1274 /*
1275  * For lack of a better implementation, implement vmsplice() to userspace
1276  * as a simple copy of the pipes pages to the user iov.
1277  */
1278 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1279                              unsigned int flags)
1280 {
1281         struct pipe_inode_info *pipe = get_pipe_info(file);
1282         struct splice_desc sd = {
1283                 .total_len = iov_iter_count(iter),
1284                 .flags = flags,
1285                 .u.data = iter
1286         };
1287         long ret = 0;
1288
1289         if (!pipe)
1290                 return -EBADF;
1291
1292         if (sd.total_len) {
1293                 pipe_lock(pipe);
1294                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1295                 pipe_unlock(pipe);
1296         }
1297
1298         return ret;
1299 }
1300
1301 /*
1302  * vmsplice splices a user address range into a pipe. It can be thought of
1303  * as splice-from-memory, where the regular splice is splice-from-file (or
1304  * to file). In both cases the output is a pipe, naturally.
1305  */
1306 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1307                              unsigned int flags)
1308 {
1309         struct pipe_inode_info *pipe;
1310         long ret = 0;
1311         unsigned buf_flag = 0;
1312
1313         if (flags & SPLICE_F_GIFT)
1314                 buf_flag = PIPE_BUF_FLAG_GIFT;
1315
1316         pipe = get_pipe_info(file);
1317         if (!pipe)
1318                 return -EBADF;
1319
1320         pipe_lock(pipe);
1321         ret = wait_for_space(pipe, flags);
1322         if (!ret)
1323                 ret = iter_to_pipe(iter, pipe, buf_flag);
1324         pipe_unlock(pipe);
1325         if (ret > 0)
1326                 wakeup_pipe_readers(pipe);
1327         return ret;
1328 }
1329
1330 static int vmsplice_type(struct fd f, int *type)
1331 {
1332         if (!f.file)
1333                 return -EBADF;
1334         if (f.file->f_mode & FMODE_WRITE) {
1335                 *type = WRITE;
1336         } else if (f.file->f_mode & FMODE_READ) {
1337                 *type = READ;
1338         } else {
1339                 fdput(f);
1340                 return -EBADF;
1341         }
1342         return 0;
1343 }
1344
1345 /*
1346  * Note that vmsplice only really supports true splicing _from_ user memory
1347  * to a pipe, not the other way around. Splicing from user memory is a simple
1348  * operation that can be supported without any funky alignment restrictions
1349  * or nasty vm tricks. We simply map in the user memory and fill them into
1350  * a pipe. The reverse isn't quite as easy, though. There are two possible
1351  * solutions for that:
1352  *
1353  *      - memcpy() the data internally, at which point we might as well just
1354  *        do a regular read() on the buffer anyway.
1355  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1356  *        has restriction limitations on both ends of the pipe).
1357  *
1358  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1359  *
1360  */
1361 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1362 {
1363         if (unlikely(flags & ~SPLICE_F_ALL))
1364                 return -EINVAL;
1365
1366         if (!iov_iter_count(iter))
1367                 return 0;
1368
1369         if (iov_iter_rw(iter) == WRITE)
1370                 return vmsplice_to_pipe(f, iter, flags);
1371         else
1372                 return vmsplice_to_user(f, iter, flags);
1373 }
1374
1375 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1376                 unsigned long, nr_segs, unsigned int, flags)
1377 {
1378         struct iovec iovstack[UIO_FASTIOV];
1379         struct iovec *iov = iovstack;
1380         struct iov_iter iter;
1381         ssize_t error;
1382         struct fd f;
1383         int type;
1384
1385         f = fdget(fd);
1386         error = vmsplice_type(f, &type);
1387         if (error)
1388                 return error;
1389
1390         error = import_iovec(type, uiov, nr_segs,
1391                              ARRAY_SIZE(iovstack), &iov, &iter);
1392         if (error >= 0) {
1393                 error = do_vmsplice(f.file, &iter, flags);
1394                 kfree(iov);
1395         }
1396         fdput(f);
1397         return error;
1398 }
1399
1400 #ifdef CONFIG_COMPAT
1401 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1402                     unsigned int, nr_segs, unsigned int, flags)
1403 {
1404         struct iovec iovstack[UIO_FASTIOV];
1405         struct iovec *iov = iovstack;
1406         struct iov_iter iter;
1407         ssize_t error;
1408         struct fd f;
1409         int type;
1410
1411         f = fdget(fd);
1412         error = vmsplice_type(f, &type);
1413         if (error)
1414                 return error;
1415
1416         error = compat_import_iovec(type, iov32, nr_segs,
1417                              ARRAY_SIZE(iovstack), &iov, &iter);
1418         if (error >= 0) {
1419                 error = do_vmsplice(f.file, &iter, flags);
1420                 kfree(iov);
1421         }
1422         fdput(f);
1423         return error;
1424 }
1425 #endif
1426
1427 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1428                 int, fd_out, loff_t __user *, off_out,
1429                 size_t, len, unsigned int, flags)
1430 {
1431         struct fd in, out;
1432         long error;
1433
1434         if (unlikely(!len))
1435                 return 0;
1436
1437         if (unlikely(flags & ~SPLICE_F_ALL))
1438                 return -EINVAL;
1439
1440         error = -EBADF;
1441         in = fdget(fd_in);
1442         if (in.file) {
1443                 if (in.file->f_mode & FMODE_READ) {
1444                         out = fdget(fd_out);
1445                         if (out.file) {
1446                                 if (out.file->f_mode & FMODE_WRITE)
1447                                         error = do_splice(in.file, off_in,
1448                                                           out.file, off_out,
1449                                                           len, flags);
1450                                 fdput(out);
1451                         }
1452                 }
1453                 fdput(in);
1454         }
1455         return error;
1456 }
1457
1458 /*
1459  * Make sure there's data to read. Wait for input if we can, otherwise
1460  * return an appropriate error.
1461  */
1462 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1463 {
1464         int ret;
1465
1466         /*
1467          * Check the pipe occupancy without the inode lock first. This function
1468          * is speculative anyways, so missing one is ok.
1469          */
1470         if (!pipe_empty(pipe->head, pipe->tail))
1471                 return 0;
1472
1473         ret = 0;
1474         pipe_lock(pipe);
1475
1476         while (pipe_empty(pipe->head, pipe->tail)) {
1477                 if (signal_pending(current)) {
1478                         ret = -ERESTARTSYS;
1479                         break;
1480                 }
1481                 if (!pipe->writers)
1482                         break;
1483                 if (flags & SPLICE_F_NONBLOCK) {
1484                         ret = -EAGAIN;
1485                         break;
1486                 }
1487                 pipe_wait(pipe);
1488         }
1489
1490         pipe_unlock(pipe);
1491         return ret;
1492 }
1493
1494 /*
1495  * Make sure there's writeable room. Wait for room if we can, otherwise
1496  * return an appropriate error.
1497  */
1498 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1499 {
1500         int ret;
1501
1502         /*
1503          * Check pipe occupancy without the inode lock first. This function
1504          * is speculative anyways, so missing one is ok.
1505          */
1506         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1507                 return 0;
1508
1509         ret = 0;
1510         pipe_lock(pipe);
1511
1512         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1513                 if (!pipe->readers) {
1514                         send_sig(SIGPIPE, current, 0);
1515                         ret = -EPIPE;
1516                         break;
1517                 }
1518                 if (flags & SPLICE_F_NONBLOCK) {
1519                         ret = -EAGAIN;
1520                         break;
1521                 }
1522                 if (signal_pending(current)) {
1523                         ret = -ERESTARTSYS;
1524                         break;
1525                 }
1526                 pipe_wait(pipe);
1527         }
1528
1529         pipe_unlock(pipe);
1530         return ret;
1531 }
1532
1533 /*
1534  * Splice contents of ipipe to opipe.
1535  */
1536 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1537                                struct pipe_inode_info *opipe,
1538                                size_t len, unsigned int flags)
1539 {
1540         struct pipe_buffer *ibuf, *obuf;
1541         unsigned int i_head, o_head;
1542         unsigned int i_tail, o_tail;
1543         unsigned int i_mask, o_mask;
1544         int ret = 0;
1545         bool input_wakeup = false;
1546
1547
1548 retry:
1549         ret = ipipe_prep(ipipe, flags);
1550         if (ret)
1551                 return ret;
1552
1553         ret = opipe_prep(opipe, flags);
1554         if (ret)
1555                 return ret;
1556
1557         /*
1558          * Potential ABBA deadlock, work around it by ordering lock
1559          * grabbing by pipe info address. Otherwise two different processes
1560          * could deadlock (one doing tee from A -> B, the other from B -> A).
1561          */
1562         pipe_double_lock(ipipe, opipe);
1563
1564         i_tail = ipipe->tail;
1565         i_mask = ipipe->ring_size - 1;
1566         o_head = opipe->head;
1567         o_mask = opipe->ring_size - 1;
1568
1569         do {
1570                 size_t o_len;
1571
1572                 if (!opipe->readers) {
1573                         send_sig(SIGPIPE, current, 0);
1574                         if (!ret)
1575                                 ret = -EPIPE;
1576                         break;
1577                 }
1578
1579                 i_head = ipipe->head;
1580                 o_tail = opipe->tail;
1581
1582                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1583                         break;
1584
1585                 /*
1586                  * Cannot make any progress, because either the input
1587                  * pipe is empty or the output pipe is full.
1588                  */
1589                 if (pipe_empty(i_head, i_tail) ||
1590                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1591                         /* Already processed some buffers, break */
1592                         if (ret)
1593                                 break;
1594
1595                         if (flags & SPLICE_F_NONBLOCK) {
1596                                 ret = -EAGAIN;
1597                                 break;
1598                         }
1599
1600                         /*
1601                          * We raced with another reader/writer and haven't
1602                          * managed to process any buffers.  A zero return
1603                          * value means EOF, so retry instead.
1604                          */
1605                         pipe_unlock(ipipe);
1606                         pipe_unlock(opipe);
1607                         goto retry;
1608                 }
1609
1610                 ibuf = &ipipe->bufs[i_tail & i_mask];
1611                 obuf = &opipe->bufs[o_head & o_mask];
1612
1613                 if (len >= ibuf->len) {
1614                         /*
1615                          * Simply move the whole buffer from ipipe to opipe
1616                          */
1617                         *obuf = *ibuf;
1618                         ibuf->ops = NULL;
1619                         i_tail++;
1620                         ipipe->tail = i_tail;
1621                         input_wakeup = true;
1622                         o_len = obuf->len;
1623                         o_head++;
1624                         opipe->head = o_head;
1625                 } else {
1626                         /*
1627                          * Get a reference to this pipe buffer,
1628                          * so we can copy the contents over.
1629                          */
1630                         if (!pipe_buf_get(ipipe, ibuf)) {
1631                                 if (ret == 0)
1632                                         ret = -EFAULT;
1633                                 break;
1634                         }
1635                         *obuf = *ibuf;
1636
1637                         /*
1638                          * Don't inherit the gift flag, we need to
1639                          * prevent multiple steals of this page.
1640                          */
1641                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1642
1643                         pipe_buf_mark_unmergeable(obuf);
1644
1645                         obuf->len = len;
1646                         ibuf->offset += len;
1647                         ibuf->len -= len;
1648                         o_len = len;
1649                         o_head++;
1650                         opipe->head = o_head;
1651                 }
1652                 ret += o_len;
1653                 len -= o_len;
1654         } while (len);
1655
1656         pipe_unlock(ipipe);
1657         pipe_unlock(opipe);
1658
1659         /*
1660          * If we put data in the output pipe, wakeup any potential readers.
1661          */
1662         if (ret > 0)
1663                 wakeup_pipe_readers(opipe);
1664
1665         if (input_wakeup)
1666                 wakeup_pipe_writers(ipipe);
1667
1668         return ret;
1669 }
1670
1671 /*
1672  * Link contents of ipipe to opipe.
1673  */
1674 static int link_pipe(struct pipe_inode_info *ipipe,
1675                      struct pipe_inode_info *opipe,
1676                      size_t len, unsigned int flags)
1677 {
1678         struct pipe_buffer *ibuf, *obuf;
1679         unsigned int i_head, o_head;
1680         unsigned int i_tail, o_tail;
1681         unsigned int i_mask, o_mask;
1682         int ret = 0;
1683
1684         /*
1685          * Potential ABBA deadlock, work around it by ordering lock
1686          * grabbing by pipe info address. Otherwise two different processes
1687          * could deadlock (one doing tee from A -> B, the other from B -> A).
1688          */
1689         pipe_double_lock(ipipe, opipe);
1690
1691         i_tail = ipipe->tail;
1692         i_mask = ipipe->ring_size - 1;
1693         o_head = opipe->head;
1694         o_mask = opipe->ring_size - 1;
1695
1696         do {
1697                 if (!opipe->readers) {
1698                         send_sig(SIGPIPE, current, 0);
1699                         if (!ret)
1700                                 ret = -EPIPE;
1701                         break;
1702                 }
1703
1704                 i_head = ipipe->head;
1705                 o_tail = opipe->tail;
1706
1707                 /*
1708                  * If we have iterated all input buffers or run out of
1709                  * output room, break.
1710                  */
1711                 if (pipe_empty(i_head, i_tail) ||
1712                     pipe_full(o_head, o_tail, opipe->max_usage))
1713                         break;
1714
1715                 ibuf = &ipipe->bufs[i_tail & i_mask];
1716                 obuf = &opipe->bufs[o_head & o_mask];
1717
1718                 /*
1719                  * Get a reference to this pipe buffer,
1720                  * so we can copy the contents over.
1721                  */
1722                 if (!pipe_buf_get(ipipe, ibuf)) {
1723                         if (ret == 0)
1724                                 ret = -EFAULT;
1725                         break;
1726                 }
1727
1728                 *obuf = *ibuf;
1729
1730                 /*
1731                  * Don't inherit the gift flag, we need to
1732                  * prevent multiple steals of this page.
1733                  */
1734                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1735
1736                 pipe_buf_mark_unmergeable(obuf);
1737
1738                 if (obuf->len > len)
1739                         obuf->len = len;
1740                 ret += obuf->len;
1741                 len -= obuf->len;
1742
1743                 o_head++;
1744                 opipe->head = o_head;
1745                 i_tail++;
1746         } while (len);
1747
1748         pipe_unlock(ipipe);
1749         pipe_unlock(opipe);
1750
1751         /*
1752          * If we put data in the output pipe, wakeup any potential readers.
1753          */
1754         if (ret > 0)
1755                 wakeup_pipe_readers(opipe);
1756
1757         return ret;
1758 }
1759
1760 /*
1761  * This is a tee(1) implementation that works on pipes. It doesn't copy
1762  * any data, it simply references the 'in' pages on the 'out' pipe.
1763  * The 'flags' used are the SPLICE_F_* variants, currently the only
1764  * applicable one is SPLICE_F_NONBLOCK.
1765  */
1766 static long do_tee(struct file *in, struct file *out, size_t len,
1767                    unsigned int flags)
1768 {
1769         struct pipe_inode_info *ipipe = get_pipe_info(in);
1770         struct pipe_inode_info *opipe = get_pipe_info(out);
1771         int ret = -EINVAL;
1772
1773         /*
1774          * Duplicate the contents of ipipe to opipe without actually
1775          * copying the data.
1776          */
1777         if (ipipe && opipe && ipipe != opipe) {
1778                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1779                         flags |= SPLICE_F_NONBLOCK;
1780
1781                 /*
1782                  * Keep going, unless we encounter an error. The ipipe/opipe
1783                  * ordering doesn't really matter.
1784                  */
1785                 ret = ipipe_prep(ipipe, flags);
1786                 if (!ret) {
1787                         ret = opipe_prep(opipe, flags);
1788                         if (!ret)
1789                                 ret = link_pipe(ipipe, opipe, len, flags);
1790                 }
1791         }
1792
1793         return ret;
1794 }
1795
1796 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1797 {
1798         struct fd in;
1799         int error;
1800
1801         if (unlikely(flags & ~SPLICE_F_ALL))
1802                 return -EINVAL;
1803
1804         if (unlikely(!len))
1805                 return 0;
1806
1807         error = -EBADF;
1808         in = fdget(fdin);
1809         if (in.file) {
1810                 if (in.file->f_mode & FMODE_READ) {
1811                         struct fd out = fdget(fdout);
1812                         if (out.file) {
1813                                 if (out.file->f_mode & FMODE_WRITE)
1814                                         error = do_tee(in.file, out.file,
1815                                                         len, flags);
1816                                 fdput(out);
1817                         }
1818                 }
1819                 fdput(in);
1820         }
1821
1822         return error;
1823 }