fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/bvec.h>
  21 #include <linux/fs.h>
  22 #include <linux/file.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/splice.h>
  25 #include <linux/memcontrol.h>
  26 #include <linux/mm_inline.h>
  27 #include <linux/swap.h>
  28 #include <linux/writeback.h>
  29 #include <linux/export.h>
  30 #include <linux/syscalls.h>
  31 #include <linux/uio.h>
  32 #include <linux/security.h>
  33 #include <linux/gfp.h>
  34 #include <linux/socket.h>
  35 #include <linux/compat.h>
  36 #include <linux/sched/signal.h>
  37
  38 #include "internal.h"
  39
  40 /*
  41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42  * a vm helper function, it's already simplified quite a bit by the
  43  * addition of remove_mapping(). If success is returned, the caller may
  44  * attempt to reuse this page for another destination.
  45  */
  46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  47                                      struct pipe_buffer *buf)
  48 {
  49         struct page *page = buf->page;
  50         struct address_space *mapping;
  51
  52         lock_page(page);
  53
  54         mapping = page_mapping(page);
  55         if (mapping) {
  56                 WARN_ON(!PageUptodate(page));
  57
  58                 /*
  59                  * At least for ext2 with nobh option, we need to wait on
  60                  * writeback completing on this page, since we'll remove it
  61                  * from the pagecache.  Otherwise truncate wont wait on the
  62                  * page, allowing the disk blocks to be reused by someone else
  63                  * before we actually wrote our data to them. fs corruption
  64                  * ensues.
  65                  */
  66                 wait_on_page_writeback(page);
  67
  68                 if (page_has_private(page) &&
  69                     !try_to_release_page(page, GFP_KERNEL))
  70                         goto out_unlock;
  71
  72                 /*
  73                  * If we succeeded in removing the mapping, set LRU flag
  74                  * and return good.
  75                  */
  76                 if (remove_mapping(mapping, page)) {
  77                         buf->flags |= PIPE_BUF_FLAG_LRU;
  78                         return 0;
  79                 }
  80         }
  81
  82         /*
  83          * Raced with truncate or failed to remove page from current
  84          * address space, unlock and return failure.
  85          */
  86 out_unlock:
  87         unlock_page(page);
  88         return 1;
  89 }
  90
  91 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  92                                         struct pipe_buffer *buf)
  93 {
  94         put_page(buf->page);
  95         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  96 }
  97
  98 /*
  99  * Check whether the contents of buf is OK to access. Since the content
 100  * is a page cache page, IO may be in flight.
 101  */
 102 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 103                                        struct pipe_buffer *buf)
 104 {
 105         struct page *page = buf->page;
 106         int err;
 107
 108         if (!PageUptodate(page)) {
 109                 lock_page(page);
 110
 111                 /*
 112                  * Page got truncated/unhashed. This will cause a 0-byte
 113                  * splice, if this is the first page.
 114                  */
 115                 if (!page->mapping) {
 116                         err = -ENODATA;
 117                         goto error;
 118                 }
 119
 120                 /*
 121                  * Uh oh, read-error from disk.
 122                  */
 123                 if (!PageUptodate(page)) {
 124                         err = -EIO;
 125                         goto error;
 126                 }
 127
 128                 /*
 129                  * Page is ok afterall, we are done.
 130                  */
 131                 unlock_page(page);
 132         }
 133
 134         return 0;
 135 error:
 136         unlock_page(page);
 137         return err;
 138 }
 139
 140 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 141         .can_merge = 0,
 142         .confirm = page_cache_pipe_buf_confirm,
 143         .release = page_cache_pipe_buf_release,
 144         .steal = page_cache_pipe_buf_steal,
 145         .get = generic_pipe_buf_get,
 146 };
 147
 148 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 149                                     struct pipe_buffer *buf)
 150 {
 151         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 152                 return 1;
 153
 154         buf->flags |= PIPE_BUF_FLAG_LRU;
 155         return generic_pipe_buf_steal(pipe, buf);
 156 }
 157
 158 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 159         .can_merge = 0,
 160         .confirm = generic_pipe_buf_confirm,
 161         .release = page_cache_pipe_buf_release,
 162         .steal = user_page_pipe_buf_steal,
 163         .get = generic_pipe_buf_get,
 164 };
 165
 166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 167 {
 168         smp_mb();
 169         if (waitqueue_active(&pipe->wait))
 170                 wake_up_interruptible(&pipe->wait);
 171         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 172 }
 173
 174 /**
 175  * splice_to_pipe - fill passed data into a pipe
 176  * @pipe:       pipe to fill
 177  * @spd:        data to fill
 178  *
 179  * Description:
 180  *    @spd contains a map of pages and len/offset tuples, along with
 181  *    the struct pipe_buf_operations associated with these pages. This
 182  *    function will link that data to the pipe.
 183  *
 184  */
 185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 186                        struct splice_pipe_desc *spd)
 187 {
 188         unsigned int spd_pages = spd->nr_pages;
 189         int ret = 0, page_nr = 0;
 190
 191         if (!spd_pages)
 192                 return 0;
 193
 194         if (unlikely(!pipe->readers)) {
 195                 send_sig(SIGPIPE, current, 0);
 196                 ret = -EPIPE;
 197                 goto out;
 198         }
 199
 200         while (pipe->nrbufs < pipe->buffers) {
 201                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 202                 struct pipe_buffer *buf = pipe->bufs + newbuf;
 203
 204                 buf->page = spd->pages[page_nr];
 205                 buf->offset = spd->partial[page_nr].offset;
 206                 buf->len = spd->partial[page_nr].len;
 207                 buf->private = spd->partial[page_nr].private;
 208                 buf->ops = spd->ops;
 209                 buf->flags = 0;
 210
 211                 pipe->nrbufs++;
 212                 page_nr++;
 213                 ret += buf->len;
 214
 215                 if (!--spd->nr_pages)
 216                         break;
 217         }
 218
 219         if (!ret)
 220                 ret = -EAGAIN;
 221
 222 out:
 223         while (page_nr < spd_pages)
 224                 spd->spd_release(spd, page_nr++);
 225
 226         return ret;
 227 }
 228 EXPORT_SYMBOL_GPL(splice_to_pipe);
 229
 230 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 231 {
 232         int ret;
 233
 234         if (unlikely(!pipe->readers)) {
 235                 send_sig(SIGPIPE, current, 0);
 236                 ret = -EPIPE;
 237         } else if (pipe->nrbufs == pipe->buffers) {
 238                 ret = -EAGAIN;
 239         } else {
 240                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 241                 pipe->bufs[newbuf] = *buf;
 242                 pipe->nrbufs++;
 243                 return buf->len;
 244         }
 245         pipe_buf_release(pipe, buf);
 246         return ret;
 247 }
 248 EXPORT_SYMBOL(add_to_pipe);
 249
 250 /*
 251  * Check if we need to grow the arrays holding pages and partial page
 252  * descriptions.
 253  */
 254 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 255 {
 256         unsigned int buffers = READ_ONCE(pipe->buffers);
 257
 258         spd->nr_pages_max = buffers;
 259         if (buffers <= PIPE_DEF_BUFFERS)
 260                 return 0;
 261
 262         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 263         spd->partial = kmalloc(buffers * sizeof(struct partial_page), 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, ITER_PIPE | 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         .can_merge = 0,
 329         .confirm = generic_pipe_buf_confirm,
 330         .release = generic_pipe_buf_release,
 331         .steal = generic_pipe_buf_steal,
 332         .get = generic_pipe_buf_get,
 333 };
 334
 335 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 336                                     struct pipe_buffer *buf)
 337 {
 338         return 1;
 339 }
 340
 341 /* Pipe buffer operations for a socket and similar. */
 342 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 343         .can_merge = 0,
 344         .confirm = generic_pipe_buf_confirm,
 345         .release = generic_pipe_buf_release,
 346         .steal = generic_pipe_buf_nosteal,
 347         .get = generic_pipe_buf_get,
 348 };
 349 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 350
 351 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
 352                             unsigned long vlen, loff_t offset)
 353 {
 354         mm_segment_t old_fs;
 355         loff_t pos = offset;
 356         ssize_t res;
 357
 358         old_fs = get_fs();
 359         set_fs(get_ds());
 360         /* The cast to a user pointer is valid due to the set_fs() */
 361         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 362         set_fs(old_fs);
 363
 364         return res;
 365 }
 366
 367 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 368                                  struct pipe_inode_info *pipe, size_t len,
 369                                  unsigned int flags)
 370 {
 371         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
 372         struct iov_iter to;
 373         struct page **pages;
 374         unsigned int nr_pages;
 375         size_t offset, base, copied = 0;
 376         ssize_t res;
 377         int i;
 378
 379         if (pipe->nrbufs == pipe->buffers)
 380                 return -EAGAIN;
 381
 382         /*
 383          * Try to keep page boundaries matching to source pagecache ones -
 384          * it probably won't be much help, but...
 385          */
 386         offset = *ppos & ~PAGE_MASK;
 387
 388         iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
 389
 390         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
 391         if (res <= 0)
 392                 return -ENOMEM;
 393
 394         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
 395
 396         vec = __vec;
 397         if (nr_pages > PIPE_DEF_BUFFERS) {
 398                 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
 399                 if (unlikely(!vec)) {
 400                         res = -ENOMEM;
 401                         goto out;
 402                 }
 403         }
 404
 405         pipe->bufs[to.idx].offset = offset;
 406         pipe->bufs[to.idx].len -= offset;
 407
 408         for (i = 0; i < nr_pages; i++) {
 409                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
 410                 vec[i].iov_base = page_address(pages[i]) + offset;
 411                 vec[i].iov_len = this_len;
 412                 len -= this_len;
 413                 offset = 0;
 414         }
 415
 416         res = kernel_readv(in, vec, nr_pages, *ppos);
 417         if (res > 0) {
 418                 copied = res;
 419                 *ppos += res;
 420         }
 421
 422         if (vec != __vec)
 423                 kfree(vec);
 424 out:
 425         for (i = 0; i < nr_pages; i++)
 426                 put_page(pages[i]);
 427         kvfree(pages);
 428         iov_iter_advance(&to, copied);  /* truncates and discards */
 429         return res;
 430 }
 431
 432 /*
 433  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 434  * using sendpage(). Return the number of bytes sent.
 435  */
 436 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 437                             struct pipe_buffer *buf, struct splice_desc *sd)
 438 {
 439         struct file *file = sd->u.file;
 440         loff_t pos = sd->pos;
 441         int more;
 442
 443         if (!likely(file->f_op->sendpage))
 444                 return -EINVAL;
 445
 446         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 447
 448         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 449                 more |= MSG_SENDPAGE_NOTLAST;
 450
 451         return file->f_op->sendpage(file, buf->page, buf->offset,
 452                                     sd->len, &pos, more);
 453 }
 454
 455 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 456 {
 457         smp_mb();
 458         if (waitqueue_active(&pipe->wait))
 459                 wake_up_interruptible(&pipe->wait);
 460         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 461 }
 462
 463 /**
 464  * splice_from_pipe_feed - feed available data from a pipe to a file
 465  * @pipe:       pipe to splice from
 466  * @sd:         information to @actor
 467  * @actor:      handler that splices the data
 468  *
 469  * Description:
 470  *    This function loops over the pipe and calls @actor to do the
 471  *    actual moving of a single struct pipe_buffer to the desired
 472  *    destination.  It returns when there's no more buffers left in
 473  *    the pipe or if the requested number of bytes (@sd->total_len)
 474  *    have been copied.  It returns a positive number (one) if the
 475  *    pipe needs to be filled with more data, zero if the required
 476  *    number of bytes have been copied and -errno on error.
 477  *
 478  *    This, together with splice_from_pipe_{begin,end,next}, may be
 479  *    used to implement the functionality of __splice_from_pipe() when
 480  *    locking is required around copying the pipe buffers to the
 481  *    destination.
 482  */
 483 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 484                           splice_actor *actor)
 485 {
 486         int ret;
 487
 488         while (pipe->nrbufs) {
 489                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 490
 491                 sd->len = buf->len;
 492                 if (sd->len > sd->total_len)
 493                         sd->len = sd->total_len;
 494
 495                 ret = pipe_buf_confirm(pipe, buf);
 496                 if (unlikely(ret)) {
 497                         if (ret == -ENODATA)
 498                                 ret = 0;
 499                         return ret;
 500                 }
 501
 502                 ret = actor(pipe, buf, sd);
 503                 if (ret <= 0)
 504                         return ret;
 505
 506                 buf->offset += ret;
 507                 buf->len -= ret;
 508
 509                 sd->num_spliced += ret;
 510                 sd->len -= ret;
 511                 sd->pos += ret;
 512                 sd->total_len -= ret;
 513
 514                 if (!buf->len) {
 515                         pipe_buf_release(pipe, buf);
 516                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 517                         pipe->nrbufs--;
 518                         if (pipe->files)
 519                                 sd->need_wakeup = true;
 520                 }
 521
 522                 if (!sd->total_len)
 523                         return 0;
 524         }
 525
 526         return 1;
 527 }
 528
 529 /**
 530  * splice_from_pipe_next - wait for some data to splice from
 531  * @pipe:       pipe to splice from
 532  * @sd:         information about the splice operation
 533  *
 534  * Description:
 535  *    This function will wait for some data and return a positive
 536  *    value (one) if pipe buffers are available.  It will return zero
 537  *    or -errno if no more data needs to be spliced.
 538  */
 539 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 540 {
 541         /*
 542          * Check for signal early to make process killable when there are
 543          * always buffers available
 544          */
 545         if (signal_pending(current))
 546                 return -ERESTARTSYS;
 547
 548         while (!pipe->nrbufs) {
 549                 if (!pipe->writers)
 550                         return 0;
 551
 552                 if (!pipe->waiting_writers && sd->num_spliced)
 553                         return 0;
 554
 555                 if (sd->flags & SPLICE_F_NONBLOCK)
 556                         return -EAGAIN;
 557
 558                 if (signal_pending(current))
 559                         return -ERESTARTSYS;
 560
 561                 if (sd->need_wakeup) {
 562                         wakeup_pipe_writers(pipe);
 563                         sd->need_wakeup = false;
 564                 }
 565
 566                 pipe_wait(pipe);
 567         }
 568
 569         return 1;
 570 }
 571
 572 /**
 573  * splice_from_pipe_begin - start splicing from pipe
 574  * @sd:         information about the splice operation
 575  *
 576  * Description:
 577  *    This function should be called before a loop containing
 578  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 579  *    initialize the necessary fields of @sd.
 580  */
 581 static void splice_from_pipe_begin(struct splice_desc *sd)
 582 {
 583         sd->num_spliced = 0;
 584         sd->need_wakeup = false;
 585 }
 586
 587 /**
 588  * splice_from_pipe_end - finish splicing from pipe
 589  * @pipe:       pipe to splice from
 590  * @sd:         information about the splice operation
 591  *
 592  * Description:
 593  *    This function will wake up pipe writers if necessary.  It should
 594  *    be called after a loop containing splice_from_pipe_next() and
 595  *    splice_from_pipe_feed().
 596  */
 597 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 598 {
 599         if (sd->need_wakeup)
 600                 wakeup_pipe_writers(pipe);
 601 }
 602
 603 /**
 604  * __splice_from_pipe - splice data from a pipe to given actor
 605  * @pipe:       pipe to splice from
 606  * @sd:         information to @actor
 607  * @actor:      handler that splices the data
 608  *
 609  * Description:
 610  *    This function does little more than loop over the pipe and call
 611  *    @actor to do the actual moving of a single struct pipe_buffer to
 612  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 613  *    pipe_to_user.
 614  *
 615  */
 616 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 617                            splice_actor *actor)
 618 {
 619         int ret;
 620
 621         splice_from_pipe_begin(sd);
 622         do {
 623                 cond_resched();
 624                 ret = splice_from_pipe_next(pipe, sd);
 625                 if (ret > 0)
 626                         ret = splice_from_pipe_feed(pipe, sd, actor);
 627         } while (ret > 0);
 628         splice_from_pipe_end(pipe, sd);
 629
 630         return sd->num_spliced ? sd->num_spliced : ret;
 631 }
 632 EXPORT_SYMBOL(__splice_from_pipe);
 633
 634 /**
 635  * splice_from_pipe - splice data from a pipe to a file
 636  * @pipe:       pipe to splice from
 637  * @out:        file to splice to
 638  * @ppos:       position in @out
 639  * @len:        how many bytes to splice
 640  * @flags:      splice modifier flags
 641  * @actor:      handler that splices the data
 642  *
 643  * Description:
 644  *    See __splice_from_pipe. This function locks the pipe inode,
 645  *    otherwise it's identical to __splice_from_pipe().
 646  *
 647  */
 648 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 649                          loff_t *ppos, size_t len, unsigned int flags,
 650                          splice_actor *actor)
 651 {
 652         ssize_t ret;
 653         struct splice_desc sd = {
 654                 .total_len = len,
 655                 .flags = flags,
 656                 .pos = *ppos,
 657                 .u.file = out,
 658         };
 659
 660         pipe_lock(pipe);
 661         ret = __splice_from_pipe(pipe, &sd, actor);
 662         pipe_unlock(pipe);
 663
 664         return ret;
 665 }
 666
 667 /**
 668  * iter_file_splice_write - splice data from a pipe to a file
 669  * @pipe:       pipe info
 670  * @out:        file to write to
 671  * @ppos:       position in @out
 672  * @len:        number of bytes to splice
 673  * @flags:      splice modifier flags
 674  *
 675  * Description:
 676  *    Will either move or copy pages (determined by @flags options) from
 677  *    the given pipe inode to the given file.
 678  *    This one is ->write_iter-based.
 679  *
 680  */
 681 ssize_t
 682 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 683                           loff_t *ppos, size_t len, unsigned int flags)
 684 {
 685         struct splice_desc sd = {
 686                 .total_len = len,
 687                 .flags = flags,
 688                 .pos = *ppos,
 689                 .u.file = out,
 690         };
 691         int nbufs = pipe->buffers;
 692         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 693                                         GFP_KERNEL);
 694         ssize_t ret;
 695
 696         if (unlikely(!array))
 697                 return -ENOMEM;
 698
 699         pipe_lock(pipe);
 700
 701         splice_from_pipe_begin(&sd);
 702         while (sd.total_len) {
 703                 struct iov_iter from;
 704                 size_t left;
 705                 int n, idx;
 706
 707                 ret = splice_from_pipe_next(pipe, &sd);
 708                 if (ret <= 0)
 709                         break;
 710
 711                 if (unlikely(nbufs < pipe->buffers)) {
 712                         kfree(array);
 713                         nbufs = pipe->buffers;
 714                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 715                                         GFP_KERNEL);
 716                         if (!array) {
 717                                 ret = -ENOMEM;
 718                                 break;
 719                         }
 720                 }
 721
 722                 /* build the vector */
 723                 left = sd.total_len;
 724                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 725                         struct pipe_buffer *buf = pipe->bufs + idx;
 726                         size_t this_len = buf->len;
 727
 728                         if (this_len > left)
 729                                 this_len = left;
 730
 731                         if (idx == pipe->buffers - 1)
 732                                 idx = -1;
 733
 734                         ret = pipe_buf_confirm(pipe, buf);
 735                         if (unlikely(ret)) {
 736                                 if (ret == -ENODATA)
 737                                         ret = 0;
 738                                 goto done;
 739                         }
 740
 741                         array[n].bv_page = buf->page;
 742                         array[n].bv_len = this_len;
 743                         array[n].bv_offset = buf->offset;
 744                         left -= this_len;
 745                 }
 746
 747                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
 748                               sd.total_len - left);
 749                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 750                 if (ret <= 0)
 751                         break;
 752
 753                 sd.num_spliced += ret;
 754                 sd.total_len -= ret;
 755                 *ppos = sd.pos;
 756
 757                 /* dismiss the fully eaten buffers, adjust the partial one */
 758                 while (ret) {
 759                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 760                         if (ret >= buf->len) {
 761                                 ret -= buf->len;
 762                                 buf->len = 0;
 763                                 pipe_buf_release(pipe, buf);
 764                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 765                                 pipe->nrbufs--;
 766                                 if (pipe->files)
 767                                         sd.need_wakeup = true;
 768                         } else {
 769                                 buf->offset += ret;
 770                                 buf->len -= ret;
 771                                 ret = 0;
 772                         }
 773                 }
 774         }
 775 done:
 776         kfree(array);
 777         splice_from_pipe_end(pipe, &sd);
 778
 779         pipe_unlock(pipe);
 780
 781         if (sd.num_spliced)
 782                 ret = sd.num_spliced;
 783
 784         return ret;
 785 }
 786
 787 EXPORT_SYMBOL(iter_file_splice_write);
 788
 789 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 790                           struct splice_desc *sd)
 791 {
 792         int ret;
 793         void *data;
 794         loff_t tmp = sd->pos;
 795
 796         data = kmap(buf->page);
 797         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
 798         kunmap(buf->page);
 799
 800         return ret;
 801 }
 802
 803 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
 804                                          struct file *out, loff_t *ppos,
 805                                          size_t len, unsigned int flags)
 806 {
 807         ssize_t ret;
 808
 809         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
 810         if (ret > 0)
 811                 *ppos += ret;
 812
 813         return ret;
 814 }
 815
 816 /**
 817  * generic_splice_sendpage - splice data from a pipe to a socket
 818  * @pipe:       pipe to splice from
 819  * @out:        socket to write to
 820  * @ppos:       position in @out
 821  * @len:        number of bytes to splice
 822  * @flags:      splice modifier flags
 823  *
 824  * Description:
 825  *    Will send @len bytes from the pipe to a network socket. No data copying
 826  *    is involved.
 827  *
 828  */
 829 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 830                                 loff_t *ppos, size_t len, unsigned int flags)
 831 {
 832         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 833 }
 834
 835 EXPORT_SYMBOL(generic_splice_sendpage);
 836
 837 /*
 838  * Attempt to initiate a splice from pipe to file.
 839  */
 840 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 841                            loff_t *ppos, size_t len, unsigned int flags)
 842 {
 843         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
 844                                 loff_t *, size_t, unsigned int);
 845
 846         if (out->f_op->splice_write)
 847                 splice_write = out->f_op->splice_write;
 848         else
 849                 splice_write = default_file_splice_write;
 850
 851         return splice_write(pipe, out, ppos, len, flags);
 852 }
 853
 854 /*
 855  * Attempt to initiate a splice from a file to a pipe.
 856  */
 857 static long do_splice_to(struct file *in, loff_t *ppos,
 858                          struct pipe_inode_info *pipe, size_t len,
 859                          unsigned int flags)
 860 {
 861         ssize_t (*splice_read)(struct file *, loff_t *,
 862                                struct pipe_inode_info *, size_t, unsigned int);
 863         int ret;
 864
 865         if (unlikely(!(in->f_mode & FMODE_READ)))
 866                 return -EBADF;
 867
 868         ret = rw_verify_area(READ, in, ppos, len);
 869         if (unlikely(ret < 0))
 870                 return ret;
 871
 872         if (unlikely(len > MAX_RW_COUNT))
 873                 len = MAX_RW_COUNT;
 874
 875         if (in->f_op->splice_read)
 876                 splice_read = in->f_op->splice_read;
 877         else
 878                 splice_read = default_file_splice_read;
 879
 880         return splice_read(in, ppos, pipe, len, flags);
 881 }
 882
 883 /**
 884  * splice_direct_to_actor - splices data directly between two non-pipes
 885  * @in:         file to splice from
 886  * @sd:         actor information on where to splice to
 887  * @actor:      handles the data splicing
 888  *
 889  * Description:
 890  *    This is a special case helper to splice directly between two
 891  *    points, without requiring an explicit pipe. Internally an allocated
 892  *    pipe is cached in the process, and reused during the lifetime of
 893  *    that process.
 894  *
 895  */
 896 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 897                                splice_direct_actor *actor)
 898 {
 899         struct pipe_inode_info *pipe;
 900         long ret, bytes;
 901         umode_t i_mode;
 902         size_t len;
 903         int i, flags, more;
 904
 905         /*
 906          * We require the input being a regular file, as we don't want to
 907          * randomly drop data for eg socket -> socket splicing. Use the
 908          * piped splicing for that!
 909          */
 910         i_mode = file_inode(in)->i_mode;
 911         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
 912                 return -EINVAL;
 913
 914         /*
 915          * neither in nor out is a pipe, setup an internal pipe attached to
 916          * 'out' and transfer the wanted data from 'in' to 'out' through that
 917          */
 918         pipe = current->splice_pipe;
 919         if (unlikely(!pipe)) {
 920                 pipe = alloc_pipe_info();
 921                 if (!pipe)
 922                         return -ENOMEM;
 923
 924                 /*
 925                  * We don't have an immediate reader, but we'll read the stuff
 926                  * out of the pipe right after the splice_to_pipe(). So set
 927                  * PIPE_READERS appropriately.
 928                  */
 929                 pipe->readers = 1;
 930
 931                 current->splice_pipe = pipe;
 932         }
 933
 934         /*
 935          * Do the splice.
 936          */
 937         ret = 0;
 938         bytes = 0;
 939         len = sd->total_len;
 940         flags = sd->flags;
 941
 942         /*
 943          * Don't block on output, we have to drain the direct pipe.
 944          */
 945         sd->flags &= ~SPLICE_F_NONBLOCK;
 946         more = sd->flags & SPLICE_F_MORE;
 947
 948         while (len) {
 949                 size_t read_len;
 950                 loff_t pos = sd->pos, prev_pos = pos;
 951
 952                 ret = do_splice_to(in, &pos, pipe, len, flags);
 953                 if (unlikely(ret <= 0))
 954                         goto out_release;
 955
 956                 read_len = ret;
 957                 sd->total_len = read_len;
 958
 959                 /*
 960                  * If more data is pending, set SPLICE_F_MORE
 961                  * If this is the last data and SPLICE_F_MORE was not set
 962                  * initially, clears it.
 963                  */
 964                 if (read_len < len)
 965                         sd->flags |= SPLICE_F_MORE;
 966                 else if (!more)
 967                         sd->flags &= ~SPLICE_F_MORE;
 968                 /*
 969                  * NOTE: nonblocking mode only applies to the input. We
 970                  * must not do the output in nonblocking mode as then we
 971                  * could get stuck data in the internal pipe:
 972                  */
 973                 ret = actor(pipe, sd);
 974                 if (unlikely(ret <= 0)) {
 975                         sd->pos = prev_pos;
 976                         goto out_release;
 977                 }
 978
 979                 bytes += ret;
 980                 len -= ret;
 981                 sd->pos = pos;
 982
 983                 if (ret < read_len) {
 984                         sd->pos = prev_pos + ret;
 985                         goto out_release;
 986                 }
 987         }
 988
 989 done:
 990         pipe->nrbufs = pipe->curbuf = 0;
 991         file_accessed(in);
 992         return bytes;
 993
 994 out_release:
 995         /*
 996          * If we did an incomplete transfer we must release
 997          * the pipe buffers in question:
 998          */
 999         for (i = 0; i < pipe->buffers; i++) {
1000                 struct pipe_buffer *buf = pipe->bufs + i;
1001
1002                 if (buf->ops)
1003                         pipe_buf_release(pipe, buf);
1004         }
1005
1006         if (!bytes)
1007                 bytes = ret;
1008
1009         goto done;
1010 }
1011 EXPORT_SYMBOL(splice_direct_to_actor);
1012
1013 static int direct_splice_actor(struct pipe_inode_info *pipe,
1014                                struct splice_desc *sd)
1015 {
1016         struct file *file = sd->u.file;
1017
1018         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1019                               sd->flags);
1020 }
1021
1022 /**
1023  * do_splice_direct - splices data directly between two files
1024  * @in:         file to splice from
1025  * @ppos:       input file offset
1026  * @out:        file to splice to
1027  * @opos:       output file offset
1028  * @len:        number of bytes to splice
1029  * @flags:      splice modifier flags
1030  *
1031  * Description:
1032  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1033  *    doing it in the application would incur an extra system call
1034  *    (splice in + splice out, as compared to just sendfile()). So this helper
1035  *    can splice directly through a process-private pipe.
1036  *
1037  */
1038 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1039                       loff_t *opos, size_t len, unsigned int flags)
1040 {
1041         struct splice_desc sd = {
1042                 .len            = len,
1043                 .total_len      = len,
1044                 .flags          = flags,
1045                 .pos            = *ppos,
1046                 .u.file         = out,
1047                 .opos           = opos,
1048         };
1049         long ret;
1050
1051         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1052                 return -EBADF;
1053
1054         if (unlikely(out->f_flags & O_APPEND))
1055                 return -EINVAL;
1056
1057         ret = rw_verify_area(WRITE, out, opos, len);
1058         if (unlikely(ret < 0))
1059                 return ret;
1060
1061         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1062         if (ret > 0)
1063                 *ppos = sd.pos;
1064
1065         return ret;
1066 }
1067 EXPORT_SYMBOL(do_splice_direct);
1068
1069 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1070 {
1071         for (;;) {
1072                 if (unlikely(!pipe->readers)) {
1073                         send_sig(SIGPIPE, current, 0);
1074                         return -EPIPE;
1075                 }
1076                 if (pipe->nrbufs != pipe->buffers)
1077                         return 0;
1078                 if (flags & SPLICE_F_NONBLOCK)
1079                         return -EAGAIN;
1080                 if (signal_pending(current))
1081                         return -ERESTARTSYS;
1082                 pipe->waiting_writers++;
1083                 pipe_wait(pipe);
1084                 pipe->waiting_writers--;
1085         }
1086 }
1087
1088 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1089                                struct pipe_inode_info *opipe,
1090                                size_t len, unsigned int flags);
1091
1092 /*
1093  * Determine where to splice to/from.
1094  */
1095 static long do_splice(struct file *in, loff_t __user *off_in,
1096                       struct file *out, loff_t __user *off_out,
1097                       size_t len, unsigned int flags)
1098 {
1099         struct pipe_inode_info *ipipe;
1100         struct pipe_inode_info *opipe;
1101         loff_t offset;
1102         long ret;
1103
1104         ipipe = get_pipe_info(in);
1105         opipe = get_pipe_info(out);
1106
1107         if (ipipe && opipe) {
1108                 if (off_in || off_out)
1109                         return -ESPIPE;
1110
1111                 if (!(in->f_mode & FMODE_READ))
1112                         return -EBADF;
1113
1114                 if (!(out->f_mode & FMODE_WRITE))
1115                         return -EBADF;
1116
1117                 /* Splicing to self would be fun, but... */
1118                 if (ipipe == opipe)
1119                         return -EINVAL;
1120
1121                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1122         }
1123
1124         if (ipipe) {
1125                 if (off_in)
1126                         return -ESPIPE;
1127                 if (off_out) {
1128                         if (!(out->f_mode & FMODE_PWRITE))
1129                                 return -EINVAL;
1130                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1131                                 return -EFAULT;
1132                 } else {
1133                         offset = out->f_pos;
1134                 }
1135
1136                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1137                         return -EBADF;
1138
1139                 if (unlikely(out->f_flags & O_APPEND))
1140                         return -EINVAL;
1141
1142                 ret = rw_verify_area(WRITE, out, &offset, len);
1143                 if (unlikely(ret < 0))
1144                         return ret;
1145
1146                 file_start_write(out);
1147                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1148                 file_end_write(out);
1149
1150                 if (!off_out)
1151                         out->f_pos = offset;
1152                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1153                         ret = -EFAULT;
1154
1155                 return ret;
1156         }
1157
1158         if (opipe) {
1159                 if (off_out)
1160                         return -ESPIPE;
1161                 if (off_in) {
1162                         if (!(in->f_mode & FMODE_PREAD))
1163                                 return -EINVAL;
1164                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1165                                 return -EFAULT;
1166                 } else {
1167                         offset = in->f_pos;
1168                 }
1169
1170                 pipe_lock(opipe);
1171                 ret = wait_for_space(opipe, flags);
1172                 if (!ret)
1173                         ret = do_splice_to(in, &offset, opipe, len, flags);
1174                 pipe_unlock(opipe);
1175                 if (ret > 0)
1176                         wakeup_pipe_readers(opipe);
1177                 if (!off_in)
1178                         in->f_pos = offset;
1179                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1180                         ret = -EFAULT;
1181
1182                 return ret;
1183         }
1184
1185         return -EINVAL;
1186 }
1187
1188 static int iter_to_pipe(struct iov_iter *from,
1189                         struct pipe_inode_info *pipe,
1190                         unsigned flags)
1191 {
1192         struct pipe_buffer buf = {
1193                 .ops = &user_page_pipe_buf_ops,
1194                 .flags = flags
1195         };
1196         size_t total = 0;
1197         int ret = 0;
1198         bool failed = false;
1199
1200         while (iov_iter_count(from) && !failed) {
1201                 struct page *pages[16];
1202                 ssize_t copied;
1203                 size_t start;
1204                 int n;
1205
1206                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1207                 if (copied <= 0) {
1208                         ret = copied;
1209                         break;
1210                 }
1211
1212                 for (n = 0; copied; n++, start = 0) {
1213                         int size = min_t(int, copied, PAGE_SIZE - start);
1214                         if (!failed) {
1215                                 buf.page = pages[n];
1216                                 buf.offset = start;
1217                                 buf.len = size;
1218                                 ret = add_to_pipe(pipe, &buf);
1219                                 if (unlikely(ret < 0)) {
1220                                         failed = true;
1221                                 } else {
1222                                         iov_iter_advance(from, ret);
1223                                         total += ret;
1224                                 }
1225                         } else {
1226                                 put_page(pages[n]);
1227                         }
1228                         copied -= size;
1229                 }
1230         }
1231         return total ? total : ret;
1232 }
1233
1234 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1235                         struct splice_desc *sd)
1236 {
1237         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1238         return n == sd->len ? n : -EFAULT;
1239 }
1240
1241 /*
1242  * For lack of a better implementation, implement vmsplice() to userspace
1243  * as a simple copy of the pipes pages to the user iov.
1244  */
1245 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1246                              unsigned long nr_segs, unsigned int flags)
1247 {
1248         struct pipe_inode_info *pipe;
1249         struct splice_desc sd;
1250         long ret;
1251         struct iovec iovstack[UIO_FASTIOV];
1252         struct iovec *iov = iovstack;
1253         struct iov_iter iter;
1254
1255         pipe = get_pipe_info(file);
1256         if (!pipe)
1257                 return -EBADF;
1258
1259         ret = import_iovec(READ, uiov, nr_segs,
1260                            ARRAY_SIZE(iovstack), &iov, &iter);
1261         if (ret < 0)
1262                 return ret;
1263
1264         sd.total_len = iov_iter_count(&iter);
1265         sd.len = 0;
1266         sd.flags = flags;
1267         sd.u.data = &iter;
1268         sd.pos = 0;
1269
1270         if (sd.total_len) {
1271                 pipe_lock(pipe);
1272                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1273                 pipe_unlock(pipe);
1274         }
1275
1276         kfree(iov);
1277         return ret;
1278 }
1279
1280 /*
1281  * vmsplice splices a user address range into a pipe. It can be thought of
1282  * as splice-from-memory, where the regular splice is splice-from-file (or
1283  * to file). In both cases the output is a pipe, naturally.
1284  */
1285 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1286                              unsigned long nr_segs, unsigned int flags)
1287 {
1288         struct pipe_inode_info *pipe;
1289         struct iovec iovstack[UIO_FASTIOV];
1290         struct iovec *iov = iovstack;
1291         struct iov_iter from;
1292         long ret;
1293         unsigned buf_flag = 0;
1294
1295         if (flags & SPLICE_F_GIFT)
1296                 buf_flag = PIPE_BUF_FLAG_GIFT;
1297
1298         pipe = get_pipe_info(file);
1299         if (!pipe)
1300                 return -EBADF;
1301
1302         ret = import_iovec(WRITE, uiov, nr_segs,
1303                            ARRAY_SIZE(iovstack), &iov, &from);
1304         if (ret < 0)
1305                 return ret;
1306
1307         pipe_lock(pipe);
1308         ret = wait_for_space(pipe, flags);
1309         if (!ret)
1310                 ret = iter_to_pipe(&from, pipe, buf_flag);
1311         pipe_unlock(pipe);
1312         if (ret > 0)
1313                 wakeup_pipe_readers(pipe);
1314         kfree(iov);
1315         return ret;
1316 }
1317
1318 /*
1319  * Note that vmsplice only really supports true splicing _from_ user memory
1320  * to a pipe, not the other way around. Splicing from user memory is a simple
1321  * operation that can be supported without any funky alignment restrictions
1322  * or nasty vm tricks. We simply map in the user memory and fill them into
1323  * a pipe. The reverse isn't quite as easy, though. There are two possible
1324  * solutions for that:
1325  *
1326  *      - memcpy() the data internally, at which point we might as well just
1327  *        do a regular read() on the buffer anyway.
1328  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1329  *        has restriction limitations on both ends of the pipe).
1330  *
1331  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1332  *
1333  */
1334 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1335                 unsigned long, nr_segs, unsigned int, flags)
1336 {
1337         struct fd f;
1338         long error;
1339
1340         if (unlikely(flags & ~SPLICE_F_ALL))
1341                 return -EINVAL;
1342         if (unlikely(nr_segs > UIO_MAXIOV))
1343                 return -EINVAL;
1344         else if (unlikely(!nr_segs))
1345                 return 0;
1346
1347         error = -EBADF;
1348         f = fdget(fd);
1349         if (f.file) {
1350                 if (f.file->f_mode & FMODE_WRITE)
1351                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1352                 else if (f.file->f_mode & FMODE_READ)
1353                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1354
1355                 fdput(f);
1356         }
1357
1358         return error;
1359 }
1360
1361 #ifdef CONFIG_COMPAT
1362 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1363                     unsigned int, nr_segs, unsigned int, flags)
1364 {
1365         unsigned i;
1366         struct iovec __user *iov;
1367         if (nr_segs > UIO_MAXIOV)
1368                 return -EINVAL;
1369         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1370         for (i = 0; i < nr_segs; i++) {
1371                 struct compat_iovec v;
1372                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1373                     get_user(v.iov_len, &iov32[i].iov_len) ||
1374                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1375                     put_user(v.iov_len, &iov[i].iov_len))
1376                         return -EFAULT;
1377         }
1378         return sys_vmsplice(fd, iov, nr_segs, flags);
1379 }
1380 #endif
1381
1382 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1383                 int, fd_out, loff_t __user *, off_out,
1384                 size_t, len, unsigned int, flags)
1385 {
1386         struct fd in, out;
1387         long error;
1388
1389         if (unlikely(!len))
1390                 return 0;
1391
1392         if (unlikely(flags & ~SPLICE_F_ALL))
1393                 return -EINVAL;
1394
1395         error = -EBADF;
1396         in = fdget(fd_in);
1397         if (in.file) {
1398                 if (in.file->f_mode & FMODE_READ) {
1399                         out = fdget(fd_out);
1400                         if (out.file) {
1401                                 if (out.file->f_mode & FMODE_WRITE)
1402                                         error = do_splice(in.file, off_in,
1403                                                           out.file, off_out,
1404                                                           len, flags);
1405                                 fdput(out);
1406                         }
1407                 }
1408                 fdput(in);
1409         }
1410         return error;
1411 }
1412
1413 /*
1414  * Make sure there's data to read. Wait for input if we can, otherwise
1415  * return an appropriate error.
1416  */
1417 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1418 {
1419         int ret;
1420
1421         /*
1422          * Check ->nrbufs without the inode lock first. This function
1423          * is speculative anyways, so missing one is ok.
1424          */
1425         if (pipe->nrbufs)
1426                 return 0;
1427
1428         ret = 0;
1429         pipe_lock(pipe);
1430
1431         while (!pipe->nrbufs) {
1432                 if (signal_pending(current)) {
1433                         ret = -ERESTARTSYS;
1434                         break;
1435                 }
1436                 if (!pipe->writers)
1437                         break;
1438                 if (!pipe->waiting_writers) {
1439                         if (flags & SPLICE_F_NONBLOCK) {
1440                                 ret = -EAGAIN;
1441                                 break;
1442                         }
1443                 }
1444                 pipe_wait(pipe);
1445         }
1446
1447         pipe_unlock(pipe);
1448         return ret;
1449 }
1450
1451 /*
1452  * Make sure there's writeable room. Wait for room if we can, otherwise
1453  * return an appropriate error.
1454  */
1455 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1456 {
1457         int ret;
1458
1459         /*
1460          * Check ->nrbufs without the inode lock first. This function
1461          * is speculative anyways, so missing one is ok.
1462          */
1463         if (pipe->nrbufs < pipe->buffers)
1464                 return 0;
1465
1466         ret = 0;
1467         pipe_lock(pipe);
1468
1469         while (pipe->nrbufs >= pipe->buffers) {
1470                 if (!pipe->readers) {
1471                         send_sig(SIGPIPE, current, 0);
1472                         ret = -EPIPE;
1473                         break;
1474                 }
1475                 if (flags & SPLICE_F_NONBLOCK) {
1476                         ret = -EAGAIN;
1477                         break;
1478                 }
1479                 if (signal_pending(current)) {
1480                         ret = -ERESTARTSYS;
1481                         break;
1482                 }
1483                 pipe->waiting_writers++;
1484                 pipe_wait(pipe);
1485                 pipe->waiting_writers--;
1486         }
1487
1488         pipe_unlock(pipe);
1489         return ret;
1490 }
1491
1492 /*
1493  * Splice contents of ipipe to opipe.
1494  */
1495 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1496                                struct pipe_inode_info *opipe,
1497                                size_t len, unsigned int flags)
1498 {
1499         struct pipe_buffer *ibuf, *obuf;
1500         int ret = 0, nbuf;
1501         bool input_wakeup = false;
1502
1503
1504 retry:
1505         ret = ipipe_prep(ipipe, flags);
1506         if (ret)
1507                 return ret;
1508
1509         ret = opipe_prep(opipe, flags);
1510         if (ret)
1511                 return ret;
1512
1513         /*
1514          * Potential ABBA deadlock, work around it by ordering lock
1515          * grabbing by pipe info address. Otherwise two different processes
1516          * could deadlock (one doing tee from A -> B, the other from B -> A).
1517          */
1518         pipe_double_lock(ipipe, opipe);
1519
1520         do {
1521                 if (!opipe->readers) {
1522                         send_sig(SIGPIPE, current, 0);
1523                         if (!ret)
1524                                 ret = -EPIPE;
1525                         break;
1526                 }
1527
1528                 if (!ipipe->nrbufs && !ipipe->writers)
1529                         break;
1530
1531                 /*
1532                  * Cannot make any progress, because either the input
1533                  * pipe is empty or the output pipe is full.
1534                  */
1535                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1536                         /* Already processed some buffers, break */
1537                         if (ret)
1538                                 break;
1539
1540                         if (flags & SPLICE_F_NONBLOCK) {
1541                                 ret = -EAGAIN;
1542                                 break;
1543                         }
1544
1545                         /*
1546                          * We raced with another reader/writer and haven't
1547                          * managed to process any buffers.  A zero return
1548                          * value means EOF, so retry instead.
1549                          */
1550                         pipe_unlock(ipipe);
1551                         pipe_unlock(opipe);
1552                         goto retry;
1553                 }
1554
1555                 ibuf = ipipe->bufs + ipipe->curbuf;
1556                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1557                 obuf = opipe->bufs + nbuf;
1558
1559                 if (len >= ibuf->len) {
1560                         /*
1561                          * Simply move the whole buffer from ipipe to opipe
1562                          */
1563                         *obuf = *ibuf;
1564                         ibuf->ops = NULL;
1565                         opipe->nrbufs++;
1566                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1567                         ipipe->nrbufs--;
1568                         input_wakeup = true;
1569                 } else {
1570                         /*
1571                          * Get a reference to this pipe buffer,
1572                          * so we can copy the contents over.
1573                          */
1574                         pipe_buf_get(ipipe, ibuf);
1575                         *obuf = *ibuf;
1576
1577                         /*
1578                          * Don't inherit the gift flag, we need to
1579                          * prevent multiple steals of this page.
1580                          */
1581                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1582
1583                         obuf->len = len;
1584                         opipe->nrbufs++;
1585                         ibuf->offset += obuf->len;
1586                         ibuf->len -= obuf->len;
1587                 }
1588                 ret += obuf->len;
1589                 len -= obuf->len;
1590         } while (len);
1591
1592         pipe_unlock(ipipe);
1593         pipe_unlock(opipe);
1594
1595         /*
1596          * If we put data in the output pipe, wakeup any potential readers.
1597          */
1598         if (ret > 0)
1599                 wakeup_pipe_readers(opipe);
1600
1601         if (input_wakeup)
1602                 wakeup_pipe_writers(ipipe);
1603
1604         return ret;
1605 }
1606
1607 /*
1608  * Link contents of ipipe to opipe.
1609  */
1610 static int link_pipe(struct pipe_inode_info *ipipe,
1611                      struct pipe_inode_info *opipe,
1612                      size_t len, unsigned int flags)
1613 {
1614         struct pipe_buffer *ibuf, *obuf;
1615         int ret = 0, i = 0, nbuf;
1616
1617         /*
1618          * Potential ABBA deadlock, work around it by ordering lock
1619          * grabbing by pipe info address. Otherwise two different processes
1620          * could deadlock (one doing tee from A -> B, the other from B -> A).
1621          */
1622         pipe_double_lock(ipipe, opipe);
1623
1624         do {
1625                 if (!opipe->readers) {
1626                         send_sig(SIGPIPE, current, 0);
1627                         if (!ret)
1628                                 ret = -EPIPE;
1629                         break;
1630                 }
1631
1632                 /*
1633                  * If we have iterated all input buffers or ran out of
1634                  * output room, break.
1635                  */
1636                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1637                         break;
1638
1639                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1640                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1641
1642                 /*
1643                  * Get a reference to this pipe buffer,
1644                  * so we can copy the contents over.
1645                  */
1646                 pipe_buf_get(ipipe, ibuf);
1647
1648                 obuf = opipe->bufs + nbuf;
1649                 *obuf = *ibuf;
1650
1651                 /*
1652                  * Don't inherit the gift flag, we need to
1653                  * prevent multiple steals of this page.
1654                  */
1655                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1656
1657                 if (obuf->len > len)
1658                         obuf->len = len;
1659
1660                 opipe->nrbufs++;
1661                 ret += obuf->len;
1662                 len -= obuf->len;
1663                 i++;
1664         } while (len);
1665
1666         /*
1667          * return EAGAIN if we have the potential of some data in the
1668          * future, otherwise just return 0
1669          */
1670         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1671                 ret = -EAGAIN;
1672
1673         pipe_unlock(ipipe);
1674         pipe_unlock(opipe);
1675
1676         /*
1677          * If we put data in the output pipe, wakeup any potential readers.
1678          */
1679         if (ret > 0)
1680                 wakeup_pipe_readers(opipe);
1681
1682         return ret;
1683 }
1684
1685 /*
1686  * This is a tee(1) implementation that works on pipes. It doesn't copy
1687  * any data, it simply references the 'in' pages on the 'out' pipe.
1688  * The 'flags' used are the SPLICE_F_* variants, currently the only
1689  * applicable one is SPLICE_F_NONBLOCK.
1690  */
1691 static long do_tee(struct file *in, struct file *out, size_t len,
1692                    unsigned int flags)
1693 {
1694         struct pipe_inode_info *ipipe = get_pipe_info(in);
1695         struct pipe_inode_info *opipe = get_pipe_info(out);
1696         int ret = -EINVAL;
1697
1698         /*
1699          * Duplicate the contents of ipipe to opipe without actually
1700          * copying the data.
1701          */
1702         if (ipipe && opipe && ipipe != opipe) {
1703                 /*
1704                  * Keep going, unless we encounter an error. The ipipe/opipe
1705                  * ordering doesn't really matter.
1706                  */
1707                 ret = ipipe_prep(ipipe, flags);
1708                 if (!ret) {
1709                         ret = opipe_prep(opipe, flags);
1710                         if (!ret)
1711                                 ret = link_pipe(ipipe, opipe, len, flags);
1712                 }
1713         }
1714
1715         return ret;
1716 }
1717
1718 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1719 {
1720         struct fd in;
1721         int error;
1722
1723         if (unlikely(flags & ~SPLICE_F_ALL))
1724                 return -EINVAL;
1725
1726         if (unlikely(!len))
1727                 return 0;
1728
1729         error = -EBADF;
1730         in = fdget(fdin);
1731         if (in.file) {
1732                 if (in.file->f_mode & FMODE_READ) {
1733                         struct fd out = fdget(fdout);
1734                         if (out.file) {
1735                                 if (out.file->f_mode & FMODE_WRITE)
1736                                         error = do_tee(in.file, out.file,
1737                                                         len, flags);
1738                                 fdput(out);
1739                         }
1740                 }
1741                 fdput(in);
1742         }
1743
1744         return error;
1745 }