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