fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34 #include <linux/compat.h>
  35 #include "internal.h"
  36
  37 /*
  38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  39  * a vm helper function, it's already simplified quite a bit by the
  40  * addition of remove_mapping(). If success is returned, the caller may
  41  * attempt to reuse this page for another destination.
  42  */
  43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  44                                      struct pipe_buffer *buf)
  45 {
  46         struct page *page = buf->page;
  47         struct address_space *mapping;
  48
  49         lock_page(page);
  50
  51         mapping = page_mapping(page);
  52         if (mapping) {
  53                 WARN_ON(!PageUptodate(page));
  54
  55                 /*
  56                  * At least for ext2 with nobh option, we need to wait on
  57                  * writeback completing on this page, since we'll remove it
  58                  * from the pagecache.  Otherwise truncate wont wait on the
  59                  * page, allowing the disk blocks to be reused by someone else
  60                  * before we actually wrote our data to them. fs corruption
  61                  * ensues.
  62                  */
  63                 wait_on_page_writeback(page);
  64
  65                 if (page_has_private(page) &&
  66                     !try_to_release_page(page, GFP_KERNEL))
  67                         goto out_unlock;
  68
  69                 /*
  70                  * If we succeeded in removing the mapping, set LRU flag
  71                  * and return good.
  72                  */
  73                 if (remove_mapping(mapping, page)) {
  74                         buf->flags |= PIPE_BUF_FLAG_LRU;
  75                         return 0;
  76                 }
  77         }
  78
  79         /*
  80          * Raced with truncate or failed to remove page from current
  81          * address space, unlock and return failure.
  82          */
  83 out_unlock:
  84         unlock_page(page);
  85         return 1;
  86 }
  87
  88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  89                                         struct pipe_buffer *buf)
  90 {
  91         put_page(buf->page);
  92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  93 }
  94
  95 /*
  96  * Check whether the contents of buf is OK to access. Since the content
  97  * is a page cache page, IO may be in flight.
  98  */
  99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 100                                        struct pipe_buffer *buf)
 101 {
 102         struct page *page = buf->page;
 103         int err;
 104
 105         if (!PageUptodate(page)) {
 106                 lock_page(page);
 107
 108                 /*
 109                  * Page got truncated/unhashed. This will cause a 0-byte
 110                  * splice, if this is the first page.
 111                  */
 112                 if (!page->mapping) {
 113                         err = -ENODATA;
 114                         goto error;
 115                 }
 116
 117                 /*
 118                  * Uh oh, read-error from disk.
 119                  */
 120                 if (!PageUptodate(page)) {
 121                         err = -EIO;
 122                         goto error;
 123                 }
 124
 125                 /*
 126                  * Page is ok afterall, we are done.
 127                  */
 128                 unlock_page(page);
 129         }
 130
 131         return 0;
 132 error:
 133         unlock_page(page);
 134         return err;
 135 }
 136
 137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .confirm = page_cache_pipe_buf_confirm,
 140         .release = page_cache_pipe_buf_release,
 141         .steal = page_cache_pipe_buf_steal,
 142         .get = generic_pipe_buf_get,
 143 };
 144
 145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                     struct pipe_buffer *buf)
 147 {
 148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                 return 1;
 150
 151         buf->flags |= PIPE_BUF_FLAG_LRU;
 152         return generic_pipe_buf_steal(pipe, buf);
 153 }
 154
 155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156         .can_merge = 0,
 157         .confirm = generic_pipe_buf_confirm,
 158         .release = page_cache_pipe_buf_release,
 159         .steal = user_page_pipe_buf_steal,
 160         .get = generic_pipe_buf_get,
 161 };
 162
 163 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 164 {
 165         smp_mb();
 166         if (waitqueue_active(&pipe->wait))
 167                 wake_up_interruptible(&pipe->wait);
 168         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 169 }
 170
 171 /**
 172  * splice_to_pipe - fill passed data into a pipe
 173  * @pipe:       pipe to fill
 174  * @spd:        data to fill
 175  *
 176  * Description:
 177  *    @spd contains a map of pages and len/offset tuples, along with
 178  *    the struct pipe_buf_operations associated with these pages. This
 179  *    function will link that data to the pipe.
 180  *
 181  */
 182 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 183                        struct splice_pipe_desc *spd)
 184 {
 185         unsigned int spd_pages = spd->nr_pages;
 186         int ret, do_wakeup, page_nr;
 187
 188         if (!spd_pages)
 189                 return 0;
 190
 191         ret = 0;
 192         do_wakeup = 0;
 193         page_nr = 0;
 194
 195         pipe_lock(pipe);
 196
 197         for (;;) {
 198                 if (!pipe->readers) {
 199                         send_sig(SIGPIPE, current, 0);
 200                         if (!ret)
 201                                 ret = -EPIPE;
 202                         break;
 203                 }
 204
 205                 if (pipe->nrbufs < pipe->buffers) {
 206                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 207                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 208
 209                         buf->page = spd->pages[page_nr];
 210                         buf->offset = spd->partial[page_nr].offset;
 211                         buf->len = spd->partial[page_nr].len;
 212                         buf->private = spd->partial[page_nr].private;
 213                         buf->ops = spd->ops;
 214                         if (spd->flags & SPLICE_F_GIFT)
 215                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 216
 217                         pipe->nrbufs++;
 218                         page_nr++;
 219                         ret += buf->len;
 220
 221                         if (pipe->files)
 222                                 do_wakeup = 1;
 223
 224                         if (!--spd->nr_pages)
 225                                 break;
 226                         if (pipe->nrbufs < pipe->buffers)
 227                                 continue;
 228
 229                         break;
 230                 }
 231
 232                 if (spd->flags & SPLICE_F_NONBLOCK) {
 233                         if (!ret)
 234                                 ret = -EAGAIN;
 235                         break;
 236                 }
 237
 238                 if (signal_pending(current)) {
 239                         if (!ret)
 240                                 ret = -ERESTARTSYS;
 241                         break;
 242                 }
 243
 244                 if (do_wakeup) {
 245                         smp_mb();
 246                         if (waitqueue_active(&pipe->wait))
 247                                 wake_up_interruptible_sync(&pipe->wait);
 248                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 249                         do_wakeup = 0;
 250                 }
 251
 252                 pipe->waiting_writers++;
 253                 pipe_wait(pipe);
 254                 pipe->waiting_writers--;
 255         }
 256
 257         pipe_unlock(pipe);
 258
 259         if (do_wakeup)
 260                 wakeup_pipe_readers(pipe);
 261
 262         while (page_nr < spd_pages)
 263                 spd->spd_release(spd, page_nr++);
 264
 265         return ret;
 266 }
 267 EXPORT_SYMBOL_GPL(splice_to_pipe);
 268
 269 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 270 {
 271         put_page(spd->pages[i]);
 272 }
 273
 274 /*
 275  * Check if we need to grow the arrays holding pages and partial page
 276  * descriptions.
 277  */
 278 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 279 {
 280         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 281
 282         spd->nr_pages_max = buffers;
 283         if (buffers <= PIPE_DEF_BUFFERS)
 284                 return 0;
 285
 286         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 287         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 288
 289         if (spd->pages && spd->partial)
 290                 return 0;
 291
 292         kfree(spd->pages);
 293         kfree(spd->partial);
 294         return -ENOMEM;
 295 }
 296
 297 void splice_shrink_spd(struct splice_pipe_desc *spd)
 298 {
 299         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 300                 return;
 301
 302         kfree(spd->pages);
 303         kfree(spd->partial);
 304 }
 305
 306 static int
 307 __generic_file_splice_read(struct file *in, loff_t *ppos,
 308                            struct pipe_inode_info *pipe, size_t len,
 309                            unsigned int flags)
 310 {
 311         struct address_space *mapping = in->f_mapping;
 312         unsigned int loff, nr_pages, req_pages;
 313         struct page *pages[PIPE_DEF_BUFFERS];
 314         struct partial_page partial[PIPE_DEF_BUFFERS];
 315         struct page *page;
 316         pgoff_t index, end_index;
 317         loff_t isize;
 318         int error, page_nr;
 319         struct splice_pipe_desc spd = {
 320                 .pages = pages,
 321                 .partial = partial,
 322                 .nr_pages_max = PIPE_DEF_BUFFERS,
 323                 .flags = flags,
 324                 .ops = &page_cache_pipe_buf_ops,
 325                 .spd_release = spd_release_page,
 326         };
 327
 328         if (splice_grow_spd(pipe, &spd))
 329                 return -ENOMEM;
 330
 331         index = *ppos >> PAGE_SHIFT;
 332         loff = *ppos & ~PAGE_MASK;
 333         req_pages = (len + loff + PAGE_SIZE - 1) >> PAGE_SHIFT;
 334         nr_pages = min(req_pages, spd.nr_pages_max);
 335
 336         /*
 337          * Lookup the (hopefully) full range of pages we need.
 338          */
 339         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 340         index += spd.nr_pages;
 341
 342         /*
 343          * If find_get_pages_contig() returned fewer pages than we needed,
 344          * readahead/allocate the rest and fill in the holes.
 345          */
 346         if (spd.nr_pages < nr_pages)
 347                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 348                                 index, req_pages - spd.nr_pages);
 349
 350         error = 0;
 351         while (spd.nr_pages < nr_pages) {
 352                 /*
 353                  * Page could be there, find_get_pages_contig() breaks on
 354                  * the first hole.
 355                  */
 356                 page = find_get_page(mapping, index);
 357                 if (!page) {
 358                         /*
 359                          * page didn't exist, allocate one.
 360                          */
 361                         page = page_cache_alloc_cold(mapping);
 362                         if (!page)
 363                                 break;
 364
 365                         error = add_to_page_cache_lru(page, mapping, index,
 366                                    mapping_gfp_constraint(mapping, GFP_KERNEL));
 367                         if (unlikely(error)) {
 368                                 put_page(page);
 369                                 if (error == -EEXIST)
 370                                         continue;
 371                                 break;
 372                         }
 373                         /*
 374                          * add_to_page_cache() locks the page, unlock it
 375                          * to avoid convoluting the logic below even more.
 376                          */
 377                         unlock_page(page);
 378                 }
 379
 380                 spd.pages[spd.nr_pages++] = page;
 381                 index++;
 382         }
 383
 384         /*
 385          * Now loop over the map and see if we need to start IO on any
 386          * pages, fill in the partial map, etc.
 387          */
 388         index = *ppos >> PAGE_SHIFT;
 389         nr_pages = spd.nr_pages;
 390         spd.nr_pages = 0;
 391         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 392                 unsigned int this_len;
 393
 394                 if (!len)
 395                         break;
 396
 397                 /*
 398                  * this_len is the max we'll use from this page
 399                  */
 400                 this_len = min_t(unsigned long, len, PAGE_SIZE - loff);
 401                 page = spd.pages[page_nr];
 402
 403                 if (PageReadahead(page))
 404                         page_cache_async_readahead(mapping, &in->f_ra, in,
 405                                         page, index, req_pages - page_nr);
 406
 407                 /*
 408                  * If the page isn't uptodate, we may need to start io on it
 409                  */
 410                 if (!PageUptodate(page)) {
 411                         lock_page(page);
 412
 413                         /*
 414                          * Page was truncated, or invalidated by the
 415                          * filesystem.  Redo the find/create, but this time the
 416                          * page is kept locked, so there's no chance of another
 417                          * race with truncate/invalidate.
 418                          */
 419                         if (!page->mapping) {
 420                                 unlock_page(page);
 421 retry_lookup:
 422                                 page = find_or_create_page(mapping, index,
 423                                                 mapping_gfp_mask(mapping));
 424
 425                                 if (!page) {
 426                                         error = -ENOMEM;
 427                                         break;
 428                                 }
 429                                 put_page(spd.pages[page_nr]);
 430                                 spd.pages[page_nr] = page;
 431                         }
 432                         /*
 433                          * page was already under io and is now done, great
 434                          */
 435                         if (PageUptodate(page)) {
 436                                 unlock_page(page);
 437                                 goto fill_it;
 438                         }
 439
 440                         /*
 441                          * need to read in the page
 442                          */
 443                         error = mapping->a_ops->readpage(in, page);
 444                         if (unlikely(error)) {
 445                                 /*
 446                                  * Re-lookup the page
 447                                  */
 448                                 if (error == AOP_TRUNCATED_PAGE)
 449                                         goto retry_lookup;
 450
 451                                 break;
 452                         }
 453                 }
 454 fill_it:
 455                 /*
 456                  * i_size must be checked after PageUptodate.
 457                  */
 458                 isize = i_size_read(mapping->host);
 459                 end_index = (isize - 1) >> PAGE_SHIFT;
 460                 if (unlikely(!isize || index > end_index))
 461                         break;
 462
 463                 /*
 464                  * if this is the last page, see if we need to shrink
 465                  * the length and stop
 466                  */
 467                 if (end_index == index) {
 468                         unsigned int plen;
 469
 470                         /*
 471                          * max good bytes in this page
 472                          */
 473                         plen = ((isize - 1) & ~PAGE_MASK) + 1;
 474                         if (plen <= loff)
 475                                 break;
 476
 477                         /*
 478                          * force quit after adding this page
 479                          */
 480                         this_len = min(this_len, plen - loff);
 481                         len = this_len;
 482                 }
 483
 484                 spd.partial[page_nr].offset = loff;
 485                 spd.partial[page_nr].len = this_len;
 486                 len -= this_len;
 487                 loff = 0;
 488                 spd.nr_pages++;
 489                 index++;
 490         }
 491
 492         /*
 493          * Release any pages at the end, if we quit early. 'page_nr' is how far
 494          * we got, 'nr_pages' is how many pages are in the map.
 495          */
 496         while (page_nr < nr_pages)
 497                 put_page(spd.pages[page_nr++]);
 498         in->f_ra.prev_pos = (loff_t)index << PAGE_SHIFT;
 499
 500         if (spd.nr_pages)
 501                 error = splice_to_pipe(pipe, &spd);
 502
 503         splice_shrink_spd(&spd);
 504         return error;
 505 }
 506
 507 /**
 508  * generic_file_splice_read - splice data from file to a pipe
 509  * @in:         file to splice from
 510  * @ppos:       position in @in
 511  * @pipe:       pipe to splice to
 512  * @len:        number of bytes to splice
 513  * @flags:      splice modifier flags
 514  *
 515  * Description:
 516  *    Will read pages from given file and fill them into a pipe. Can be
 517  *    used as long as the address_space operations for the source implements
 518  *    a readpage() hook.
 519  *
 520  */
 521 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 522                                  struct pipe_inode_info *pipe, size_t len,
 523                                  unsigned int flags)
 524 {
 525         loff_t isize, left;
 526         int ret;
 527
 528         if (IS_DAX(in->f_mapping->host))
 529                 return default_file_splice_read(in, ppos, pipe, len, flags);
 530
 531         isize = i_size_read(in->f_mapping->host);
 532         if (unlikely(*ppos >= isize))
 533                 return 0;
 534
 535         left = isize - *ppos;
 536         if (unlikely(left < len))
 537                 len = left;
 538
 539         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 540         if (ret > 0) {
 541                 *ppos += ret;
 542                 file_accessed(in);
 543         }
 544
 545         return ret;
 546 }
 547 EXPORT_SYMBOL(generic_file_splice_read);
 548
 549 static const struct pipe_buf_operations default_pipe_buf_ops = {
 550         .can_merge = 0,
 551         .confirm = generic_pipe_buf_confirm,
 552         .release = generic_pipe_buf_release,
 553         .steal = generic_pipe_buf_steal,
 554         .get = generic_pipe_buf_get,
 555 };
 556
 557 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 558                                     struct pipe_buffer *buf)
 559 {
 560         return 1;
 561 }
 562
 563 /* Pipe buffer operations for a socket and similar. */
 564 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 565         .can_merge = 0,
 566         .confirm = generic_pipe_buf_confirm,
 567         .release = generic_pipe_buf_release,
 568         .steal = generic_pipe_buf_nosteal,
 569         .get = generic_pipe_buf_get,
 570 };
 571 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 572
 573 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 574                             unsigned long vlen, loff_t offset)
 575 {
 576         mm_segment_t old_fs;
 577         loff_t pos = offset;
 578         ssize_t res;
 579
 580         old_fs = get_fs();
 581         set_fs(get_ds());
 582         /* The cast to a user pointer is valid due to the set_fs() */
 583         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
 584         set_fs(old_fs);
 585
 586         return res;
 587 }
 588
 589 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 590                             loff_t pos)
 591 {
 592         mm_segment_t old_fs;
 593         ssize_t res;
 594
 595         old_fs = get_fs();
 596         set_fs(get_ds());
 597         /* The cast to a user pointer is valid due to the set_fs() */
 598         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 599         set_fs(old_fs);
 600
 601         return res;
 602 }
 603 EXPORT_SYMBOL(kernel_write);
 604
 605 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 606                                  struct pipe_inode_info *pipe, size_t len,
 607                                  unsigned int flags)
 608 {
 609         unsigned int nr_pages;
 610         unsigned int nr_freed;
 611         size_t offset;
 612         struct page *pages[PIPE_DEF_BUFFERS];
 613         struct partial_page partial[PIPE_DEF_BUFFERS];
 614         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 615         ssize_t res;
 616         size_t this_len;
 617         int error;
 618         int i;
 619         struct splice_pipe_desc spd = {
 620                 .pages = pages,
 621                 .partial = partial,
 622                 .nr_pages_max = PIPE_DEF_BUFFERS,
 623                 .flags = flags,
 624                 .ops = &default_pipe_buf_ops,
 625                 .spd_release = spd_release_page,
 626         };
 627
 628         if (splice_grow_spd(pipe, &spd))
 629                 return -ENOMEM;
 630
 631         res = -ENOMEM;
 632         vec = __vec;
 633         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 634                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 635                 if (!vec)
 636                         goto shrink_ret;
 637         }
 638
 639         offset = *ppos & ~PAGE_MASK;
 640         nr_pages = (len + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 641
 642         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 643                 struct page *page;
 644
 645                 page = alloc_page(GFP_USER);
 646                 error = -ENOMEM;
 647                 if (!page)
 648                         goto err;
 649
 650                 this_len = min_t(size_t, len, PAGE_SIZE - offset);
 651                 vec[i].iov_base = (void __user *) page_address(page);
 652                 vec[i].iov_len = this_len;
 653                 spd.pages[i] = page;
 654                 spd.nr_pages++;
 655                 len -= this_len;
 656                 offset = 0;
 657         }
 658
 659         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 660         if (res < 0) {
 661                 error = res;
 662                 goto err;
 663         }
 664
 665         error = 0;
 666         if (!res)
 667                 goto err;
 668
 669         nr_freed = 0;
 670         for (i = 0; i < spd.nr_pages; i++) {
 671                 this_len = min_t(size_t, vec[i].iov_len, res);
 672                 spd.partial[i].offset = 0;
 673                 spd.partial[i].len = this_len;
 674                 if (!this_len) {
 675                         __free_page(spd.pages[i]);
 676                         spd.pages[i] = NULL;
 677                         nr_freed++;
 678                 }
 679                 res -= this_len;
 680         }
 681         spd.nr_pages -= nr_freed;
 682
 683         res = splice_to_pipe(pipe, &spd);
 684         if (res > 0)
 685                 *ppos += res;
 686
 687 shrink_ret:
 688         if (vec != __vec)
 689                 kfree(vec);
 690         splice_shrink_spd(&spd);
 691         return res;
 692
 693 err:
 694         for (i = 0; i < spd.nr_pages; i++)
 695                 __free_page(spd.pages[i]);
 696
 697         res = error;
 698         goto shrink_ret;
 699 }
 700 EXPORT_SYMBOL(default_file_splice_read);
 701
 702 /*
 703  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 704  * using sendpage(). Return the number of bytes sent.
 705  */
 706 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 707                             struct pipe_buffer *buf, struct splice_desc *sd)
 708 {
 709         struct file *file = sd->u.file;
 710         loff_t pos = sd->pos;
 711         int more;
 712
 713         if (!likely(file->f_op->sendpage))
 714                 return -EINVAL;
 715
 716         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 717
 718         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 719                 more |= MSG_SENDPAGE_NOTLAST;
 720
 721         return file->f_op->sendpage(file, buf->page, buf->offset,
 722                                     sd->len, &pos, more);
 723 }
 724
 725 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 726 {
 727         smp_mb();
 728         if (waitqueue_active(&pipe->wait))
 729                 wake_up_interruptible(&pipe->wait);
 730         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 731 }
 732
 733 /**
 734  * splice_from_pipe_feed - feed available data from a pipe to a file
 735  * @pipe:       pipe to splice from
 736  * @sd:         information to @actor
 737  * @actor:      handler that splices the data
 738  *
 739  * Description:
 740  *    This function loops over the pipe and calls @actor to do the
 741  *    actual moving of a single struct pipe_buffer to the desired
 742  *    destination.  It returns when there's no more buffers left in
 743  *    the pipe or if the requested number of bytes (@sd->total_len)
 744  *    have been copied.  It returns a positive number (one) if the
 745  *    pipe needs to be filled with more data, zero if the required
 746  *    number of bytes have been copied and -errno on error.
 747  *
 748  *    This, together with splice_from_pipe_{begin,end,next}, may be
 749  *    used to implement the functionality of __splice_from_pipe() when
 750  *    locking is required around copying the pipe buffers to the
 751  *    destination.
 752  */
 753 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 754                           splice_actor *actor)
 755 {
 756         int ret;
 757
 758         while (pipe->nrbufs) {
 759                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 760                 const struct pipe_buf_operations *ops = buf->ops;
 761
 762                 sd->len = buf->len;
 763                 if (sd->len > sd->total_len)
 764                         sd->len = sd->total_len;
 765
 766                 ret = buf->ops->confirm(pipe, buf);
 767                 if (unlikely(ret)) {
 768                         if (ret == -ENODATA)
 769                                 ret = 0;
 770                         return ret;
 771                 }
 772
 773                 ret = actor(pipe, buf, sd);
 774                 if (ret <= 0)
 775                         return ret;
 776
 777                 buf->offset += ret;
 778                 buf->len -= ret;
 779
 780                 sd->num_spliced += ret;
 781                 sd->len -= ret;
 782                 sd->pos += ret;
 783                 sd->total_len -= ret;
 784
 785                 if (!buf->len) {
 786                         buf->ops = NULL;
 787                         ops->release(pipe, buf);
 788                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 789                         pipe->nrbufs--;
 790                         if (pipe->files)
 791                                 sd->need_wakeup = true;
 792                 }
 793
 794                 if (!sd->total_len)
 795                         return 0;
 796         }
 797
 798         return 1;
 799 }
 800
 801 /**
 802  * splice_from_pipe_next - wait for some data to splice from
 803  * @pipe:       pipe to splice from
 804  * @sd:         information about the splice operation
 805  *
 806  * Description:
 807  *    This function will wait for some data and return a positive
 808  *    value (one) if pipe buffers are available.  It will return zero
 809  *    or -errno if no more data needs to be spliced.
 810  */
 811 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 812 {
 813         /*
 814          * Check for signal early to make process killable when there are
 815          * always buffers available
 816          */
 817         if (signal_pending(current))
 818                 return -ERESTARTSYS;
 819
 820         while (!pipe->nrbufs) {
 821                 if (!pipe->writers)
 822                         return 0;
 823
 824                 if (!pipe->waiting_writers && sd->num_spliced)
 825                         return 0;
 826
 827                 if (sd->flags & SPLICE_F_NONBLOCK)
 828                         return -EAGAIN;
 829
 830                 if (signal_pending(current))
 831                         return -ERESTARTSYS;
 832
 833                 if (sd->need_wakeup) {
 834                         wakeup_pipe_writers(pipe);
 835                         sd->need_wakeup = false;
 836                 }
 837
 838                 pipe_wait(pipe);
 839         }
 840
 841         return 1;
 842 }
 843
 844 /**
 845  * splice_from_pipe_begin - start splicing from pipe
 846  * @sd:         information about the splice operation
 847  *
 848  * Description:
 849  *    This function should be called before a loop containing
 850  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 851  *    initialize the necessary fields of @sd.
 852  */
 853 static void splice_from_pipe_begin(struct splice_desc *sd)
 854 {
 855         sd->num_spliced = 0;
 856         sd->need_wakeup = false;
 857 }
 858
 859 /**
 860  * splice_from_pipe_end - finish splicing from pipe
 861  * @pipe:       pipe to splice from
 862  * @sd:         information about the splice operation
 863  *
 864  * Description:
 865  *    This function will wake up pipe writers if necessary.  It should
 866  *    be called after a loop containing splice_from_pipe_next() and
 867  *    splice_from_pipe_feed().
 868  */
 869 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 870 {
 871         if (sd->need_wakeup)
 872                 wakeup_pipe_writers(pipe);
 873 }
 874
 875 /**
 876  * __splice_from_pipe - splice data from a pipe to given actor
 877  * @pipe:       pipe to splice from
 878  * @sd:         information to @actor
 879  * @actor:      handler that splices the data
 880  *
 881  * Description:
 882  *    This function does little more than loop over the pipe and call
 883  *    @actor to do the actual moving of a single struct pipe_buffer to
 884  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 885  *    pipe_to_user.
 886  *
 887  */
 888 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 889                            splice_actor *actor)
 890 {
 891         int ret;
 892
 893         splice_from_pipe_begin(sd);
 894         do {
 895                 cond_resched();
 896                 ret = splice_from_pipe_next(pipe, sd);
 897                 if (ret > 0)
 898                         ret = splice_from_pipe_feed(pipe, sd, actor);
 899         } while (ret > 0);
 900         splice_from_pipe_end(pipe, sd);
 901
 902         return sd->num_spliced ? sd->num_spliced : ret;
 903 }
 904 EXPORT_SYMBOL(__splice_from_pipe);
 905
 906 /**
 907  * splice_from_pipe - splice data from a pipe to a file
 908  * @pipe:       pipe to splice from
 909  * @out:        file to splice to
 910  * @ppos:       position in @out
 911  * @len:        how many bytes to splice
 912  * @flags:      splice modifier flags
 913  * @actor:      handler that splices the data
 914  *
 915  * Description:
 916  *    See __splice_from_pipe. This function locks the pipe inode,
 917  *    otherwise it's identical to __splice_from_pipe().
 918  *
 919  */
 920 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 921                          loff_t *ppos, size_t len, unsigned int flags,
 922                          splice_actor *actor)
 923 {
 924         ssize_t ret;
 925         struct splice_desc sd = {
 926                 .total_len = len,
 927                 .flags = flags,
 928                 .pos = *ppos,
 929                 .u.file = out,
 930         };
 931
 932         pipe_lock(pipe);
 933         ret = __splice_from_pipe(pipe, &sd, actor);
 934         pipe_unlock(pipe);
 935
 936         return ret;
 937 }
 938
 939 /**
 940  * iter_file_splice_write - splice data from a pipe to a file
 941  * @pipe:       pipe info
 942  * @out:        file to write to
 943  * @ppos:       position in @out
 944  * @len:        number of bytes to splice
 945  * @flags:      splice modifier flags
 946  *
 947  * Description:
 948  *    Will either move or copy pages (determined by @flags options) from
 949  *    the given pipe inode to the given file.
 950  *    This one is ->write_iter-based.
 951  *
 952  */
 953 ssize_t
 954 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 955                           loff_t *ppos, size_t len, unsigned int flags)
 956 {
 957         struct splice_desc sd = {
 958                 .total_len = len,
 959                 .flags = flags,
 960                 .pos = *ppos,
 961                 .u.file = out,
 962         };
 963         int nbufs = pipe->buffers;
 964         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 965                                         GFP_KERNEL);
 966         ssize_t ret;
 967
 968         if (unlikely(!array))
 969                 return -ENOMEM;
 970
 971         pipe_lock(pipe);
 972
 973         splice_from_pipe_begin(&sd);
 974         while (sd.total_len) {
 975                 struct iov_iter from;
 976                 size_t left;
 977                 int n, idx;
 978
 979                 ret = splice_from_pipe_next(pipe, &sd);
 980                 if (ret <= 0)
 981                         break;
 982
 983                 if (unlikely(nbufs < pipe->buffers)) {
 984                         kfree(array);
 985                         nbufs = pipe->buffers;
 986                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 987                                         GFP_KERNEL);
 988                         if (!array) {
 989                                 ret = -ENOMEM;
 990                                 break;
 991                         }
 992                 }
 993
 994                 /* build the vector */
 995                 left = sd.total_len;
 996                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 997                         struct pipe_buffer *buf = pipe->bufs + idx;
 998                         size_t this_len = buf->len;
 999
1000                         if (this_len > left)
1001                                 this_len = left;
1002
1003                         if (idx == pipe->buffers - 1)
1004                                 idx = -1;
1005
1006                         ret = buf->ops->confirm(pipe, buf);
1007                         if (unlikely(ret)) {
1008                                 if (ret == -ENODATA)
1009                                         ret = 0;
1010                                 goto done;
1011                         }
1012
1013                         array[n].bv_page = buf->page;
1014                         array[n].bv_len = this_len;
1015                         array[n].bv_offset = buf->offset;
1016                         left -= this_len;
1017                 }
1018
1019                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
1020                               sd.total_len - left);
1021                 ret = vfs_iter_write(out, &from, &sd.pos);
1022                 if (ret <= 0)
1023                         break;
1024
1025                 sd.num_spliced += ret;
1026                 sd.total_len -= ret;
1027                 *ppos = sd.pos;
1028
1029                 /* dismiss the fully eaten buffers, adjust the partial one */
1030                 while (ret) {
1031                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1032                         if (ret >= buf->len) {
1033                                 const struct pipe_buf_operations *ops = buf->ops;
1034                                 ret -= buf->len;
1035                                 buf->len = 0;
1036                                 buf->ops = NULL;
1037                                 ops->release(pipe, buf);
1038                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1039                                 pipe->nrbufs--;
1040                                 if (pipe->files)
1041                                         sd.need_wakeup = true;
1042                         } else {
1043                                 buf->offset += ret;
1044                                 buf->len -= ret;
1045                                 ret = 0;
1046                         }
1047                 }
1048         }
1049 done:
1050         kfree(array);
1051         splice_from_pipe_end(pipe, &sd);
1052
1053         pipe_unlock(pipe);
1054
1055         if (sd.num_spliced)
1056                 ret = sd.num_spliced;
1057
1058         return ret;
1059 }
1060
1061 EXPORT_SYMBOL(iter_file_splice_write);
1062
1063 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1064                           struct splice_desc *sd)
1065 {
1066         int ret;
1067         void *data;
1068         loff_t tmp = sd->pos;
1069
1070         data = kmap(buf->page);
1071         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1072         kunmap(buf->page);
1073
1074         return ret;
1075 }
1076
1077 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1078                                          struct file *out, loff_t *ppos,
1079                                          size_t len, unsigned int flags)
1080 {
1081         ssize_t ret;
1082
1083         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1084         if (ret > 0)
1085                 *ppos += ret;
1086
1087         return ret;
1088 }
1089
1090 /**
1091  * generic_splice_sendpage - splice data from a pipe to a socket
1092  * @pipe:       pipe to splice from
1093  * @out:        socket to write to
1094  * @ppos:       position in @out
1095  * @len:        number of bytes to splice
1096  * @flags:      splice modifier flags
1097  *
1098  * Description:
1099  *    Will send @len bytes from the pipe to a network socket. No data copying
1100  *    is involved.
1101  *
1102  */
1103 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1104                                 loff_t *ppos, size_t len, unsigned int flags)
1105 {
1106         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1107 }
1108
1109 EXPORT_SYMBOL(generic_splice_sendpage);
1110
1111 /*
1112  * Attempt to initiate a splice from pipe to file.
1113  */
1114 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1115                            loff_t *ppos, size_t len, unsigned int flags)
1116 {
1117         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1118                                 loff_t *, size_t, unsigned int);
1119
1120         if (out->f_op->splice_write)
1121                 splice_write = out->f_op->splice_write;
1122         else
1123                 splice_write = default_file_splice_write;
1124
1125         return splice_write(pipe, out, ppos, len, flags);
1126 }
1127
1128 /*
1129  * Attempt to initiate a splice from a file to a pipe.
1130  */
1131 static long do_splice_to(struct file *in, loff_t *ppos,
1132                          struct pipe_inode_info *pipe, size_t len,
1133                          unsigned int flags)
1134 {
1135         ssize_t (*splice_read)(struct file *, loff_t *,
1136                                struct pipe_inode_info *, size_t, unsigned int);
1137         int ret;
1138
1139         if (unlikely(!(in->f_mode & FMODE_READ)))
1140                 return -EBADF;
1141
1142         ret = rw_verify_area(READ, in, ppos, len);
1143         if (unlikely(ret < 0))
1144                 return ret;
1145
1146         if (in->f_op->splice_read)
1147                 splice_read = in->f_op->splice_read;
1148         else
1149                 splice_read = default_file_splice_read;
1150
1151         return splice_read(in, ppos, pipe, len, flags);
1152 }
1153
1154 /**
1155  * splice_direct_to_actor - splices data directly between two non-pipes
1156  * @in:         file to splice from
1157  * @sd:         actor information on where to splice to
1158  * @actor:      handles the data splicing
1159  *
1160  * Description:
1161  *    This is a special case helper to splice directly between two
1162  *    points, without requiring an explicit pipe. Internally an allocated
1163  *    pipe is cached in the process, and reused during the lifetime of
1164  *    that process.
1165  *
1166  */
1167 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1168                                splice_direct_actor *actor)
1169 {
1170         struct pipe_inode_info *pipe;
1171         long ret, bytes;
1172         umode_t i_mode;
1173         size_t len;
1174         int i, flags, more;
1175
1176         /*
1177          * We require the input being a regular file, as we don't want to
1178          * randomly drop data for eg socket -> socket splicing. Use the
1179          * piped splicing for that!
1180          */
1181         i_mode = file_inode(in)->i_mode;
1182         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1183                 return -EINVAL;
1184
1185         /*
1186          * neither in nor out is a pipe, setup an internal pipe attached to
1187          * 'out' and transfer the wanted data from 'in' to 'out' through that
1188          */
1189         pipe = current->splice_pipe;
1190         if (unlikely(!pipe)) {
1191                 pipe = alloc_pipe_info();
1192                 if (!pipe)
1193                         return -ENOMEM;
1194
1195                 /*
1196                  * We don't have an immediate reader, but we'll read the stuff
1197                  * out of the pipe right after the splice_to_pipe(). So set
1198                  * PIPE_READERS appropriately.
1199                  */
1200                 pipe->readers = 1;
1201
1202                 current->splice_pipe = pipe;
1203         }
1204
1205         /*
1206          * Do the splice.
1207          */
1208         ret = 0;
1209         bytes = 0;
1210         len = sd->total_len;
1211         flags = sd->flags;
1212
1213         /*
1214          * Don't block on output, we have to drain the direct pipe.
1215          */
1216         sd->flags &= ~SPLICE_F_NONBLOCK;
1217         more = sd->flags & SPLICE_F_MORE;
1218
1219         while (len) {
1220                 size_t read_len;
1221                 loff_t pos = sd->pos, prev_pos = pos;
1222
1223                 ret = do_splice_to(in, &pos, pipe, len, flags);
1224                 if (unlikely(ret <= 0))
1225                         goto out_release;
1226
1227                 read_len = ret;
1228                 sd->total_len = read_len;
1229
1230                 /*
1231                  * If more data is pending, set SPLICE_F_MORE
1232                  * If this is the last data and SPLICE_F_MORE was not set
1233                  * initially, clears it.
1234                  */
1235                 if (read_len < len)
1236                         sd->flags |= SPLICE_F_MORE;
1237                 else if (!more)
1238                         sd->flags &= ~SPLICE_F_MORE;
1239                 /*
1240                  * NOTE: nonblocking mode only applies to the input. We
1241                  * must not do the output in nonblocking mode as then we
1242                  * could get stuck data in the internal pipe:
1243                  */
1244                 ret = actor(pipe, sd);
1245                 if (unlikely(ret <= 0)) {
1246                         sd->pos = prev_pos;
1247                         goto out_release;
1248                 }
1249
1250                 bytes += ret;
1251                 len -= ret;
1252                 sd->pos = pos;
1253
1254                 if (ret < read_len) {
1255                         sd->pos = prev_pos + ret;
1256                         goto out_release;
1257                 }
1258         }
1259
1260 done:
1261         pipe->nrbufs = pipe->curbuf = 0;
1262         file_accessed(in);
1263         return bytes;
1264
1265 out_release:
1266         /*
1267          * If we did an incomplete transfer we must release
1268          * the pipe buffers in question:
1269          */
1270         for (i = 0; i < pipe->buffers; i++) {
1271                 struct pipe_buffer *buf = pipe->bufs + i;
1272
1273                 if (buf->ops) {
1274                         buf->ops->release(pipe, buf);
1275                         buf->ops = NULL;
1276                 }
1277         }
1278
1279         if (!bytes)
1280                 bytes = ret;
1281
1282         goto done;
1283 }
1284 EXPORT_SYMBOL(splice_direct_to_actor);
1285
1286 static int direct_splice_actor(struct pipe_inode_info *pipe,
1287                                struct splice_desc *sd)
1288 {
1289         struct file *file = sd->u.file;
1290
1291         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1292                               sd->flags);
1293 }
1294
1295 /**
1296  * do_splice_direct - splices data directly between two files
1297  * @in:         file to splice from
1298  * @ppos:       input file offset
1299  * @out:        file to splice to
1300  * @opos:       output file offset
1301  * @len:        number of bytes to splice
1302  * @flags:      splice modifier flags
1303  *
1304  * Description:
1305  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1306  *    doing it in the application would incur an extra system call
1307  *    (splice in + splice out, as compared to just sendfile()). So this helper
1308  *    can splice directly through a process-private pipe.
1309  *
1310  */
1311 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1312                       loff_t *opos, size_t len, unsigned int flags)
1313 {
1314         struct splice_desc sd = {
1315                 .len            = len,
1316                 .total_len      = len,
1317                 .flags          = flags,
1318                 .pos            = *ppos,
1319                 .u.file         = out,
1320                 .opos           = opos,
1321         };
1322         long ret;
1323
1324         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1325                 return -EBADF;
1326
1327         if (unlikely(out->f_flags & O_APPEND))
1328                 return -EINVAL;
1329
1330         ret = rw_verify_area(WRITE, out, opos, len);
1331         if (unlikely(ret < 0))
1332                 return ret;
1333
1334         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1335         if (ret > 0)
1336                 *ppos = sd.pos;
1337
1338         return ret;
1339 }
1340 EXPORT_SYMBOL(do_splice_direct);
1341
1342 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1343                                struct pipe_inode_info *opipe,
1344                                size_t len, unsigned int flags);
1345
1346 /*
1347  * Determine where to splice to/from.
1348  */
1349 static long do_splice(struct file *in, loff_t __user *off_in,
1350                       struct file *out, loff_t __user *off_out,
1351                       size_t len, unsigned int flags)
1352 {
1353         struct pipe_inode_info *ipipe;
1354         struct pipe_inode_info *opipe;
1355         loff_t offset;
1356         long ret;
1357
1358         ipipe = get_pipe_info(in);
1359         opipe = get_pipe_info(out);
1360
1361         if (ipipe && opipe) {
1362                 if (off_in || off_out)
1363                         return -ESPIPE;
1364
1365                 if (!(in->f_mode & FMODE_READ))
1366                         return -EBADF;
1367
1368                 if (!(out->f_mode & FMODE_WRITE))
1369                         return -EBADF;
1370
1371                 /* Splicing to self would be fun, but... */
1372                 if (ipipe == opipe)
1373                         return -EINVAL;
1374
1375                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1376         }
1377
1378         if (ipipe) {
1379                 if (off_in)
1380                         return -ESPIPE;
1381                 if (off_out) {
1382                         if (!(out->f_mode & FMODE_PWRITE))
1383                                 return -EINVAL;
1384                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1385                                 return -EFAULT;
1386                 } else {
1387                         offset = out->f_pos;
1388                 }
1389
1390                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1391                         return -EBADF;
1392
1393                 if (unlikely(out->f_flags & O_APPEND))
1394                         return -EINVAL;
1395
1396                 ret = rw_verify_area(WRITE, out, &offset, len);
1397                 if (unlikely(ret < 0))
1398                         return ret;
1399
1400                 file_start_write(out);
1401                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1402                 file_end_write(out);
1403
1404                 if (!off_out)
1405                         out->f_pos = offset;
1406                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1407                         ret = -EFAULT;
1408
1409                 return ret;
1410         }
1411
1412         if (opipe) {
1413                 if (off_out)
1414                         return -ESPIPE;
1415                 if (off_in) {
1416                         if (!(in->f_mode & FMODE_PREAD))
1417                                 return -EINVAL;
1418                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1419                                 return -EFAULT;
1420                 } else {
1421                         offset = in->f_pos;
1422                 }
1423
1424                 ret = do_splice_to(in, &offset, opipe, len, flags);
1425
1426                 if (!off_in)
1427                         in->f_pos = offset;
1428                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1429                         ret = -EFAULT;
1430
1431                 return ret;
1432         }
1433
1434         return -EINVAL;
1435 }
1436
1437 /*
1438  * Map an iov into an array of pages and offset/length tupples. With the
1439  * partial_page structure, we can map several non-contiguous ranges into
1440  * our ones pages[] map instead of splitting that operation into pieces.
1441  * Could easily be exported as a generic helper for other users, in which
1442  * case one would probably want to add a 'max_nr_pages' parameter as well.
1443  */
1444 static int get_iovec_page_array(const struct iovec __user *iov,
1445                                 unsigned int nr_vecs, struct page **pages,
1446                                 struct partial_page *partial, bool aligned,
1447                                 unsigned int pipe_buffers)
1448 {
1449         int buffers = 0, error = 0;
1450
1451         while (nr_vecs) {
1452                 unsigned long off, npages;
1453                 struct iovec entry;
1454                 void __user *base;
1455                 size_t len;
1456                 int i;
1457
1458                 error = -EFAULT;
1459                 if (copy_from_user(&entry, iov, sizeof(entry)))
1460                         break;
1461
1462                 base = entry.iov_base;
1463                 len = entry.iov_len;
1464
1465                 /*
1466                  * Sanity check this iovec. 0 read succeeds.
1467                  */
1468                 error = 0;
1469                 if (unlikely(!len))
1470                         break;
1471                 error = -EFAULT;
1472                 if (!access_ok(VERIFY_READ, base, len))
1473                         break;
1474
1475                 /*
1476                  * Get this base offset and number of pages, then map
1477                  * in the user pages.
1478                  */
1479                 off = (unsigned long) base & ~PAGE_MASK;
1480
1481                 /*
1482                  * If asked for alignment, the offset must be zero and the
1483                  * length a multiple of the PAGE_SIZE.
1484                  */
1485                 error = -EINVAL;
1486                 if (aligned && (off || len & ~PAGE_MASK))
1487                         break;
1488
1489                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1490                 if (npages > pipe_buffers - buffers)
1491                         npages = pipe_buffers - buffers;
1492
1493                 error = get_user_pages_fast((unsigned long)base, npages,
1494                                         0, &pages[buffers]);
1495
1496                 if (unlikely(error <= 0))
1497                         break;
1498
1499                 /*
1500                  * Fill this contiguous range into the partial page map.
1501                  */
1502                 for (i = 0; i < error; i++) {
1503                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1504
1505                         partial[buffers].offset = off;
1506                         partial[buffers].len = plen;
1507
1508                         off = 0;
1509                         len -= plen;
1510                         buffers++;
1511                 }
1512
1513                 /*
1514                  * We didn't complete this iov, stop here since it probably
1515                  * means we have to move some of this into a pipe to
1516                  * be able to continue.
1517                  */
1518                 if (len)
1519                         break;
1520
1521                 /*
1522                  * Don't continue if we mapped fewer pages than we asked for,
1523                  * or if we mapped the max number of pages that we have
1524                  * room for.
1525                  */
1526                 if (error < npages || buffers == pipe_buffers)
1527                         break;
1528
1529                 nr_vecs--;
1530                 iov++;
1531         }
1532
1533         if (buffers)
1534                 return buffers;
1535
1536         return error;
1537 }
1538
1539 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1540                         struct splice_desc *sd)
1541 {
1542         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1543         return n == sd->len ? n : -EFAULT;
1544 }
1545
1546 /*
1547  * For lack of a better implementation, implement vmsplice() to userspace
1548  * as a simple copy of the pipes pages to the user iov.
1549  */
1550 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1551                              unsigned long nr_segs, unsigned int flags)
1552 {
1553         struct pipe_inode_info *pipe;
1554         struct splice_desc sd;
1555         long ret;
1556         struct iovec iovstack[UIO_FASTIOV];
1557         struct iovec *iov = iovstack;
1558         struct iov_iter iter;
1559
1560         pipe = get_pipe_info(file);
1561         if (!pipe)
1562                 return -EBADF;
1563
1564         ret = import_iovec(READ, uiov, nr_segs,
1565                            ARRAY_SIZE(iovstack), &iov, &iter);
1566         if (ret < 0)
1567                 return ret;
1568
1569         sd.total_len = iov_iter_count(&iter);
1570         sd.len = 0;
1571         sd.flags = flags;
1572         sd.u.data = &iter;
1573         sd.pos = 0;
1574
1575         if (sd.total_len) {
1576                 pipe_lock(pipe);
1577                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1578                 pipe_unlock(pipe);
1579         }
1580
1581         kfree(iov);
1582         return ret;
1583 }
1584
1585 /*
1586  * vmsplice splices a user address range into a pipe. It can be thought of
1587  * as splice-from-memory, where the regular splice is splice-from-file (or
1588  * to file). In both cases the output is a pipe, naturally.
1589  */
1590 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1591                              unsigned long nr_segs, unsigned int flags)
1592 {
1593         struct pipe_inode_info *pipe;
1594         struct page *pages[PIPE_DEF_BUFFERS];
1595         struct partial_page partial[PIPE_DEF_BUFFERS];
1596         struct splice_pipe_desc spd = {
1597                 .pages = pages,
1598                 .partial = partial,
1599                 .nr_pages_max = PIPE_DEF_BUFFERS,
1600                 .flags = flags,
1601                 .ops = &user_page_pipe_buf_ops,
1602                 .spd_release = spd_release_page,
1603         };
1604         long ret;
1605
1606         pipe = get_pipe_info(file);
1607         if (!pipe)
1608                 return -EBADF;
1609
1610         if (splice_grow_spd(pipe, &spd))
1611                 return -ENOMEM;
1612
1613         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1614                                             spd.partial, false,
1615                                             spd.nr_pages_max);
1616         if (spd.nr_pages <= 0)
1617                 ret = spd.nr_pages;
1618         else
1619                 ret = splice_to_pipe(pipe, &spd);
1620
1621         splice_shrink_spd(&spd);
1622         return ret;
1623 }
1624
1625 /*
1626  * Note that vmsplice only really supports true splicing _from_ user memory
1627  * to a pipe, not the other way around. Splicing from user memory is a simple
1628  * operation that can be supported without any funky alignment restrictions
1629  * or nasty vm tricks. We simply map in the user memory and fill them into
1630  * a pipe. The reverse isn't quite as easy, though. There are two possible
1631  * solutions for that:
1632  *
1633  *      - memcpy() the data internally, at which point we might as well just
1634  *        do a regular read() on the buffer anyway.
1635  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1636  *        has restriction limitations on both ends of the pipe).
1637  *
1638  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1639  *
1640  */
1641 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1642                 unsigned long, nr_segs, unsigned int, flags)
1643 {
1644         struct fd f;
1645         long error;
1646
1647         if (unlikely(nr_segs > UIO_MAXIOV))
1648                 return -EINVAL;
1649         else if (unlikely(!nr_segs))
1650                 return 0;
1651
1652         error = -EBADF;
1653         f = fdget(fd);
1654         if (f.file) {
1655                 if (f.file->f_mode & FMODE_WRITE)
1656                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1657                 else if (f.file->f_mode & FMODE_READ)
1658                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1659
1660                 fdput(f);
1661         }
1662
1663         return error;
1664 }
1665
1666 #ifdef CONFIG_COMPAT
1667 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1668                     unsigned int, nr_segs, unsigned int, flags)
1669 {
1670         unsigned i;
1671         struct iovec __user *iov;
1672         if (nr_segs > UIO_MAXIOV)
1673                 return -EINVAL;
1674         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1675         for (i = 0; i < nr_segs; i++) {
1676                 struct compat_iovec v;
1677                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1678                     get_user(v.iov_len, &iov32[i].iov_len) ||
1679                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1680                     put_user(v.iov_len, &iov[i].iov_len))
1681                         return -EFAULT;
1682         }
1683         return sys_vmsplice(fd, iov, nr_segs, flags);
1684 }
1685 #endif
1686
1687 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1688                 int, fd_out, loff_t __user *, off_out,
1689                 size_t, len, unsigned int, flags)
1690 {
1691         struct fd in, out;
1692         long error;
1693
1694         if (unlikely(!len))
1695                 return 0;
1696
1697         error = -EBADF;
1698         in = fdget(fd_in);
1699         if (in.file) {
1700                 if (in.file->f_mode & FMODE_READ) {
1701                         out = fdget(fd_out);
1702                         if (out.file) {
1703                                 if (out.file->f_mode & FMODE_WRITE)
1704                                         error = do_splice(in.file, off_in,
1705                                                           out.file, off_out,
1706                                                           len, flags);
1707                                 fdput(out);
1708                         }
1709                 }
1710                 fdput(in);
1711         }
1712         return error;
1713 }
1714
1715 /*
1716  * Make sure there's data to read. Wait for input if we can, otherwise
1717  * return an appropriate error.
1718  */
1719 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1720 {
1721         int ret;
1722
1723         /*
1724          * Check ->nrbufs without the inode lock first. This function
1725          * is speculative anyways, so missing one is ok.
1726          */
1727         if (pipe->nrbufs)
1728                 return 0;
1729
1730         ret = 0;
1731         pipe_lock(pipe);
1732
1733         while (!pipe->nrbufs) {
1734                 if (signal_pending(current)) {
1735                         ret = -ERESTARTSYS;
1736                         break;
1737                 }
1738                 if (!pipe->writers)
1739                         break;
1740                 if (!pipe->waiting_writers) {
1741                         if (flags & SPLICE_F_NONBLOCK) {
1742                                 ret = -EAGAIN;
1743                                 break;
1744                         }
1745                 }
1746                 pipe_wait(pipe);
1747         }
1748
1749         pipe_unlock(pipe);
1750         return ret;
1751 }
1752
1753 /*
1754  * Make sure there's writeable room. Wait for room if we can, otherwise
1755  * return an appropriate error.
1756  */
1757 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1758 {
1759         int ret;
1760
1761         /*
1762          * Check ->nrbufs without the inode lock first. This function
1763          * is speculative anyways, so missing one is ok.
1764          */
1765         if (pipe->nrbufs < pipe->buffers)
1766                 return 0;
1767
1768         ret = 0;
1769         pipe_lock(pipe);
1770
1771         while (pipe->nrbufs >= pipe->buffers) {
1772                 if (!pipe->readers) {
1773                         send_sig(SIGPIPE, current, 0);
1774                         ret = -EPIPE;
1775                         break;
1776                 }
1777                 if (flags & SPLICE_F_NONBLOCK) {
1778                         ret = -EAGAIN;
1779                         break;
1780                 }
1781                 if (signal_pending(current)) {
1782                         ret = -ERESTARTSYS;
1783                         break;
1784                 }
1785                 pipe->waiting_writers++;
1786                 pipe_wait(pipe);
1787                 pipe->waiting_writers--;
1788         }
1789
1790         pipe_unlock(pipe);
1791         return ret;
1792 }
1793
1794 /*
1795  * Splice contents of ipipe to opipe.
1796  */
1797 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1798                                struct pipe_inode_info *opipe,
1799                                size_t len, unsigned int flags)
1800 {
1801         struct pipe_buffer *ibuf, *obuf;
1802         int ret = 0, nbuf;
1803         bool input_wakeup = false;
1804
1805
1806 retry:
1807         ret = ipipe_prep(ipipe, flags);
1808         if (ret)
1809                 return ret;
1810
1811         ret = opipe_prep(opipe, flags);
1812         if (ret)
1813                 return ret;
1814
1815         /*
1816          * Potential ABBA deadlock, work around it by ordering lock
1817          * grabbing by pipe info address. Otherwise two different processes
1818          * could deadlock (one doing tee from A -> B, the other from B -> A).
1819          */
1820         pipe_double_lock(ipipe, opipe);
1821
1822         do {
1823                 if (!opipe->readers) {
1824                         send_sig(SIGPIPE, current, 0);
1825                         if (!ret)
1826                                 ret = -EPIPE;
1827                         break;
1828                 }
1829
1830                 if (!ipipe->nrbufs && !ipipe->writers)
1831                         break;
1832
1833                 /*
1834                  * Cannot make any progress, because either the input
1835                  * pipe is empty or the output pipe is full.
1836                  */
1837                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1838                         /* Already processed some buffers, break */
1839                         if (ret)
1840                                 break;
1841
1842                         if (flags & SPLICE_F_NONBLOCK) {
1843                                 ret = -EAGAIN;
1844                                 break;
1845                         }
1846
1847                         /*
1848                          * We raced with another reader/writer and haven't
1849                          * managed to process any buffers.  A zero return
1850                          * value means EOF, so retry instead.
1851                          */
1852                         pipe_unlock(ipipe);
1853                         pipe_unlock(opipe);
1854                         goto retry;
1855                 }
1856
1857                 ibuf = ipipe->bufs + ipipe->curbuf;
1858                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1859                 obuf = opipe->bufs + nbuf;
1860
1861                 if (len >= ibuf->len) {
1862                         /*
1863                          * Simply move the whole buffer from ipipe to opipe
1864                          */
1865                         *obuf = *ibuf;
1866                         ibuf->ops = NULL;
1867                         opipe->nrbufs++;
1868                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1869                         ipipe->nrbufs--;
1870                         input_wakeup = true;
1871                 } else {
1872                         /*
1873                          * Get a reference to this pipe buffer,
1874                          * so we can copy the contents over.
1875                          */
1876                         ibuf->ops->get(ipipe, ibuf);
1877                         *obuf = *ibuf;
1878
1879                         /*
1880                          * Don't inherit the gift flag, we need to
1881                          * prevent multiple steals of this page.
1882                          */
1883                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1884
1885                         obuf->len = len;
1886                         opipe->nrbufs++;
1887                         ibuf->offset += obuf->len;
1888                         ibuf->len -= obuf->len;
1889                 }
1890                 ret += obuf->len;
1891                 len -= obuf->len;
1892         } while (len);
1893
1894         pipe_unlock(ipipe);
1895         pipe_unlock(opipe);
1896
1897         /*
1898          * If we put data in the output pipe, wakeup any potential readers.
1899          */
1900         if (ret > 0)
1901                 wakeup_pipe_readers(opipe);
1902
1903         if (input_wakeup)
1904                 wakeup_pipe_writers(ipipe);
1905
1906         return ret;
1907 }
1908
1909 /*
1910  * Link contents of ipipe to opipe.
1911  */
1912 static int link_pipe(struct pipe_inode_info *ipipe,
1913                      struct pipe_inode_info *opipe,
1914                      size_t len, unsigned int flags)
1915 {
1916         struct pipe_buffer *ibuf, *obuf;
1917         int ret = 0, i = 0, nbuf;
1918
1919         /*
1920          * Potential ABBA deadlock, work around it by ordering lock
1921          * grabbing by pipe info address. Otherwise two different processes
1922          * could deadlock (one doing tee from A -> B, the other from B -> A).
1923          */
1924         pipe_double_lock(ipipe, opipe);
1925
1926         do {
1927                 if (!opipe->readers) {
1928                         send_sig(SIGPIPE, current, 0);
1929                         if (!ret)
1930                                 ret = -EPIPE;
1931                         break;
1932                 }
1933
1934                 /*
1935                  * If we have iterated all input buffers or ran out of
1936                  * output room, break.
1937                  */
1938                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1939                         break;
1940
1941                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1942                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1943
1944                 /*
1945                  * Get a reference to this pipe buffer,
1946                  * so we can copy the contents over.
1947                  */
1948                 ibuf->ops->get(ipipe, ibuf);
1949
1950                 obuf = opipe->bufs + nbuf;
1951                 *obuf = *ibuf;
1952
1953                 /*
1954                  * Don't inherit the gift flag, we need to
1955                  * prevent multiple steals of this page.
1956                  */
1957                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1958
1959                 if (obuf->len > len)
1960                         obuf->len = len;
1961
1962                 opipe->nrbufs++;
1963                 ret += obuf->len;
1964                 len -= obuf->len;
1965                 i++;
1966         } while (len);
1967
1968         /*
1969          * return EAGAIN if we have the potential of some data in the
1970          * future, otherwise just return 0
1971          */
1972         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1973                 ret = -EAGAIN;
1974
1975         pipe_unlock(ipipe);
1976         pipe_unlock(opipe);
1977
1978         /*
1979          * If we put data in the output pipe, wakeup any potential readers.
1980          */
1981         if (ret > 0)
1982                 wakeup_pipe_readers(opipe);
1983
1984         return ret;
1985 }
1986
1987 /*
1988  * This is a tee(1) implementation that works on pipes. It doesn't copy
1989  * any data, it simply references the 'in' pages on the 'out' pipe.
1990  * The 'flags' used are the SPLICE_F_* variants, currently the only
1991  * applicable one is SPLICE_F_NONBLOCK.
1992  */
1993 static long do_tee(struct file *in, struct file *out, size_t len,
1994                    unsigned int flags)
1995 {
1996         struct pipe_inode_info *ipipe = get_pipe_info(in);
1997         struct pipe_inode_info *opipe = get_pipe_info(out);
1998         int ret = -EINVAL;
1999
2000         /*
2001          * Duplicate the contents of ipipe to opipe without actually
2002          * copying the data.
2003          */
2004         if (ipipe && opipe && ipipe != opipe) {
2005                 /*
2006                  * Keep going, unless we encounter an error. The ipipe/opipe
2007                  * ordering doesn't really matter.
2008                  */
2009                 ret = ipipe_prep(ipipe, flags);
2010                 if (!ret) {
2011                         ret = opipe_prep(opipe, flags);
2012                         if (!ret)
2013                                 ret = link_pipe(ipipe, opipe, len, flags);
2014                 }
2015         }
2016
2017         return ret;
2018 }
2019
2020 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2021 {
2022         struct fd in;
2023         int error;
2024
2025         if (unlikely(!len))
2026                 return 0;
2027
2028         error = -EBADF;
2029         in = fdget(fdin);
2030         if (in.file) {
2031                 if (in.file->f_mode & FMODE_READ) {
2032                         struct fd out = fdget(fdout);
2033                         if (out.file) {
2034                                 if (out.file->f_mode & FMODE_WRITE)
2035                                         error = do_tee(in.file, out.file,
2036                                                         len, flags);
2037                                 fdput(out);
2038                         }
2039                 }
2040                 fdput(in);
2041         }
2042
2043         return error;
2044 }