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         page_cache_release(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         page_cache_release(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_CACHE_SHIFT;
 332         loff = *ppos & ~PAGE_CACHE_MASK;
 333         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_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                                 page_cache_release(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_CACHE_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_CACHE_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                                 page = find_or_create_page(mapping, index,
 422                                                 mapping_gfp_mask(mapping));
 423
 424                                 if (!page) {
 425                                         error = -ENOMEM;
 426                                         break;
 427                                 }
 428                                 page_cache_release(spd.pages[page_nr]);
 429                                 spd.pages[page_nr] = page;
 430                         }
 431                         /*
 432                          * page was already under io and is now done, great
 433                          */
 434                         if (PageUptodate(page)) {
 435                                 unlock_page(page);
 436                                 goto fill_it;
 437                         }
 438
 439                         /*
 440                          * need to read in the page
 441                          */
 442                         error = mapping->a_ops->readpage(in, page);
 443                         if (unlikely(error)) {
 444                                 /*
 445                                  * We really should re-lookup the page here,
 446                                  * but it complicates things a lot. Instead
 447                                  * lets just do what we already stored, and
 448                                  * we'll get it the next time we are called.
 449                                  */
 450                                 if (error == AOP_TRUNCATED_PAGE)
 451                                         error = 0;
 452
 453                                 break;
 454                         }
 455                 }
 456 fill_it:
 457                 /*
 458                  * i_size must be checked after PageUptodate.
 459                  */
 460                 isize = i_size_read(mapping->host);
 461                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 462                 if (unlikely(!isize || index > end_index))
 463                         break;
 464
 465                 /*
 466                  * if this is the last page, see if we need to shrink
 467                  * the length and stop
 468                  */
 469                 if (end_index == index) {
 470                         unsigned int plen;
 471
 472                         /*
 473                          * max good bytes in this page
 474                          */
 475                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 476                         if (plen <= loff)
 477                                 break;
 478
 479                         /*
 480                          * force quit after adding this page
 481                          */
 482                         this_len = min(this_len, plen - loff);
 483                         len = this_len;
 484                 }
 485
 486                 spd.partial[page_nr].offset = loff;
 487                 spd.partial[page_nr].len = this_len;
 488                 len -= this_len;
 489                 loff = 0;
 490                 spd.nr_pages++;
 491                 index++;
 492         }
 493
 494         /*
 495          * Release any pages at the end, if we quit early. 'page_nr' is how far
 496          * we got, 'nr_pages' is how many pages are in the map.
 497          */
 498         while (page_nr < nr_pages)
 499                 page_cache_release(spd.pages[page_nr++]);
 500         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 501
 502         if (spd.nr_pages)
 503                 error = splice_to_pipe(pipe, &spd);
 504
 505         splice_shrink_spd(&spd);
 506         return error;
 507 }
 508
 509 /**
 510  * generic_file_splice_read - splice data from file to a pipe
 511  * @in:         file to splice from
 512  * @ppos:       position in @in
 513  * @pipe:       pipe to splice to
 514  * @len:        number of bytes to splice
 515  * @flags:      splice modifier flags
 516  *
 517  * Description:
 518  *    Will read pages from given file and fill them into a pipe. Can be
 519  *    used as long as the address_space operations for the source implements
 520  *    a readpage() hook.
 521  *
 522  */
 523 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 524                                  struct pipe_inode_info *pipe, size_t len,
 525                                  unsigned int flags)
 526 {
 527         loff_t isize, left;
 528         int ret;
 529
 530         if (IS_DAX(in->f_mapping->host))
 531                 return default_file_splice_read(in, ppos, pipe, len, flags);
 532
 533         isize = i_size_read(in->f_mapping->host);
 534         if (unlikely(*ppos >= isize))
 535                 return 0;
 536
 537         left = isize - *ppos;
 538         if (unlikely(left < len))
 539                 len = left;
 540
 541         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 542         if (ret > 0) {
 543                 *ppos += ret;
 544                 file_accessed(in);
 545         }
 546
 547         return ret;
 548 }
 549 EXPORT_SYMBOL(generic_file_splice_read);
 550
 551 static const struct pipe_buf_operations default_pipe_buf_ops = {
 552         .can_merge = 0,
 553         .confirm = generic_pipe_buf_confirm,
 554         .release = generic_pipe_buf_release,
 555         .steal = generic_pipe_buf_steal,
 556         .get = generic_pipe_buf_get,
 557 };
 558
 559 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 560                                     struct pipe_buffer *buf)
 561 {
 562         return 1;
 563 }
 564
 565 /* Pipe buffer operations for a socket and similar. */
 566 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 567         .can_merge = 0,
 568         .confirm = generic_pipe_buf_confirm,
 569         .release = generic_pipe_buf_release,
 570         .steal = generic_pipe_buf_nosteal,
 571         .get = generic_pipe_buf_get,
 572 };
 573 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 574
 575 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 576                             unsigned long vlen, loff_t offset)
 577 {
 578         mm_segment_t old_fs;
 579         loff_t pos = offset;
 580         ssize_t res;
 581
 582         old_fs = get_fs();
 583         set_fs(get_ds());
 584         /* The cast to a user pointer is valid due to the set_fs() */
 585         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 586         set_fs(old_fs);
 587
 588         return res;
 589 }
 590
 591 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 592                             loff_t pos)
 593 {
 594         mm_segment_t old_fs;
 595         ssize_t res;
 596
 597         old_fs = get_fs();
 598         set_fs(get_ds());
 599         /* The cast to a user pointer is valid due to the set_fs() */
 600         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 601         set_fs(old_fs);
 602
 603         return res;
 604 }
 605 EXPORT_SYMBOL(kernel_write);
 606
 607 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 608                                  struct pipe_inode_info *pipe, size_t len,
 609                                  unsigned int flags)
 610 {
 611         unsigned int nr_pages;
 612         unsigned int nr_freed;
 613         size_t offset;
 614         struct page *pages[PIPE_DEF_BUFFERS];
 615         struct partial_page partial[PIPE_DEF_BUFFERS];
 616         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 617         ssize_t res;
 618         size_t this_len;
 619         int error;
 620         int i;
 621         struct splice_pipe_desc spd = {
 622                 .pages = pages,
 623                 .partial = partial,
 624                 .nr_pages_max = PIPE_DEF_BUFFERS,
 625                 .flags = flags,
 626                 .ops = &default_pipe_buf_ops,
 627                 .spd_release = spd_release_page,
 628         };
 629
 630         if (splice_grow_spd(pipe, &spd))
 631                 return -ENOMEM;
 632
 633         res = -ENOMEM;
 634         vec = __vec;
 635         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 636                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 637                 if (!vec)
 638                         goto shrink_ret;
 639         }
 640
 641         offset = *ppos & ~PAGE_CACHE_MASK;
 642         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 643
 644         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 645                 struct page *page;
 646
 647                 page = alloc_page(GFP_USER);
 648                 error = -ENOMEM;
 649                 if (!page)
 650                         goto err;
 651
 652                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 653                 vec[i].iov_base = (void __user *) page_address(page);
 654                 vec[i].iov_len = this_len;
 655                 spd.pages[i] = page;
 656                 spd.nr_pages++;
 657                 len -= this_len;
 658                 offset = 0;
 659         }
 660
 661         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 662         if (res < 0) {
 663                 error = res;
 664                 goto err;
 665         }
 666
 667         error = 0;
 668         if (!res)
 669                 goto err;
 670
 671         nr_freed = 0;
 672         for (i = 0; i < spd.nr_pages; i++) {
 673                 this_len = min_t(size_t, vec[i].iov_len, res);
 674                 spd.partial[i].offset = 0;
 675                 spd.partial[i].len = this_len;
 676                 if (!this_len) {
 677                         __free_page(spd.pages[i]);
 678                         spd.pages[i] = NULL;
 679                         nr_freed++;
 680                 }
 681                 res -= this_len;
 682         }
 683         spd.nr_pages -= nr_freed;
 684
 685         res = splice_to_pipe(pipe, &spd);
 686         if (res > 0)
 687                 *ppos += res;
 688
 689 shrink_ret:
 690         if (vec != __vec)
 691                 kfree(vec);
 692         splice_shrink_spd(&spd);
 693         return res;
 694
 695 err:
 696         for (i = 0; i < spd.nr_pages; i++)
 697                 __free_page(spd.pages[i]);
 698
 699         res = error;
 700         goto shrink_ret;
 701 }
 702 EXPORT_SYMBOL(default_file_splice_read);
 703
 704 /*
 705  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 706  * using sendpage(). Return the number of bytes sent.
 707  */
 708 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 709                             struct pipe_buffer *buf, struct splice_desc *sd)
 710 {
 711         struct file *file = sd->u.file;
 712         loff_t pos = sd->pos;
 713         int more;
 714
 715         if (!likely(file->f_op->sendpage))
 716                 return -EINVAL;
 717
 718         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 719
 720         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 721                 more |= MSG_SENDPAGE_NOTLAST;
 722
 723         return file->f_op->sendpage(file, buf->page, buf->offset,
 724                                     sd->len, &pos, more);
 725 }
 726
 727 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 728 {
 729         smp_mb();
 730         if (waitqueue_active(&pipe->wait))
 731                 wake_up_interruptible(&pipe->wait);
 732         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 733 }
 734
 735 /**
 736  * splice_from_pipe_feed - feed available data from a pipe to a file
 737  * @pipe:       pipe to splice from
 738  * @sd:         information to @actor
 739  * @actor:      handler that splices the data
 740  *
 741  * Description:
 742  *    This function loops over the pipe and calls @actor to do the
 743  *    actual moving of a single struct pipe_buffer to the desired
 744  *    destination.  It returns when there's no more buffers left in
 745  *    the pipe or if the requested number of bytes (@sd->total_len)
 746  *    have been copied.  It returns a positive number (one) if the
 747  *    pipe needs to be filled with more data, zero if the required
 748  *    number of bytes have been copied and -errno on error.
 749  *
 750  *    This, together with splice_from_pipe_{begin,end,next}, may be
 751  *    used to implement the functionality of __splice_from_pipe() when
 752  *    locking is required around copying the pipe buffers to the
 753  *    destination.
 754  */
 755 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 756                           splice_actor *actor)
 757 {
 758         int ret;
 759
 760         while (pipe->nrbufs) {
 761                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 762                 const struct pipe_buf_operations *ops = buf->ops;
 763
 764                 sd->len = buf->len;
 765                 if (sd->len > sd->total_len)
 766                         sd->len = sd->total_len;
 767
 768                 ret = buf->ops->confirm(pipe, buf);
 769                 if (unlikely(ret)) {
 770                         if (ret == -ENODATA)
 771                                 ret = 0;
 772                         return ret;
 773                 }
 774
 775                 ret = actor(pipe, buf, sd);
 776                 if (ret <= 0)
 777                         return ret;
 778
 779                 buf->offset += ret;
 780                 buf->len -= ret;
 781
 782                 sd->num_spliced += ret;
 783                 sd->len -= ret;
 784                 sd->pos += ret;
 785                 sd->total_len -= ret;
 786
 787                 if (!buf->len) {
 788                         buf->ops = NULL;
 789                         ops->release(pipe, buf);
 790                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 791                         pipe->nrbufs--;
 792                         if (pipe->files)
 793                                 sd->need_wakeup = true;
 794                 }
 795
 796                 if (!sd->total_len)
 797                         return 0;
 798         }
 799
 800         return 1;
 801 }
 802
 803 /**
 804  * splice_from_pipe_next - wait for some data to splice from
 805  * @pipe:       pipe to splice from
 806  * @sd:         information about the splice operation
 807  *
 808  * Description:
 809  *    This function will wait for some data and return a positive
 810  *    value (one) if pipe buffers are available.  It will return zero
 811  *    or -errno if no more data needs to be spliced.
 812  */
 813 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 814 {
 815         /*
 816          * Check for signal early to make process killable when there are
 817          * always buffers available
 818          */
 819         if (signal_pending(current))
 820                 return -ERESTARTSYS;
 821
 822         while (!pipe->nrbufs) {
 823                 if (!pipe->writers)
 824                         return 0;
 825
 826                 if (!pipe->waiting_writers && sd->num_spliced)
 827                         return 0;
 828
 829                 if (sd->flags & SPLICE_F_NONBLOCK)
 830                         return -EAGAIN;
 831
 832                 if (signal_pending(current))
 833                         return -ERESTARTSYS;
 834
 835                 if (sd->need_wakeup) {
 836                         wakeup_pipe_writers(pipe);
 837                         sd->need_wakeup = false;
 838                 }
 839
 840                 pipe_wait(pipe);
 841         }
 842
 843         return 1;
 844 }
 845
 846 /**
 847  * splice_from_pipe_begin - start splicing from pipe
 848  * @sd:         information about the splice operation
 849  *
 850  * Description:
 851  *    This function should be called before a loop containing
 852  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 853  *    initialize the necessary fields of @sd.
 854  */
 855 static void splice_from_pipe_begin(struct splice_desc *sd)
 856 {
 857         sd->num_spliced = 0;
 858         sd->need_wakeup = false;
 859 }
 860
 861 /**
 862  * splice_from_pipe_end - finish splicing from pipe
 863  * @pipe:       pipe to splice from
 864  * @sd:         information about the splice operation
 865  *
 866  * Description:
 867  *    This function will wake up pipe writers if necessary.  It should
 868  *    be called after a loop containing splice_from_pipe_next() and
 869  *    splice_from_pipe_feed().
 870  */
 871 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 872 {
 873         if (sd->need_wakeup)
 874                 wakeup_pipe_writers(pipe);
 875 }
 876
 877 /**
 878  * __splice_from_pipe - splice data from a pipe to given actor
 879  * @pipe:       pipe to splice from
 880  * @sd:         information to @actor
 881  * @actor:      handler that splices the data
 882  *
 883  * Description:
 884  *    This function does little more than loop over the pipe and call
 885  *    @actor to do the actual moving of a single struct pipe_buffer to
 886  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 887  *    pipe_to_user.
 888  *
 889  */
 890 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 891                            splice_actor *actor)
 892 {
 893         int ret;
 894
 895         splice_from_pipe_begin(sd);
 896         do {
 897                 cond_resched();
 898                 ret = splice_from_pipe_next(pipe, sd);
 899                 if (ret > 0)
 900                         ret = splice_from_pipe_feed(pipe, sd, actor);
 901         } while (ret > 0);
 902         splice_from_pipe_end(pipe, sd);
 903
 904         return sd->num_spliced ? sd->num_spliced : ret;
 905 }
 906 EXPORT_SYMBOL(__splice_from_pipe);
 907
 908 /**
 909  * splice_from_pipe - splice data from a pipe to a file
 910  * @pipe:       pipe to splice from
 911  * @out:        file to splice to
 912  * @ppos:       position in @out
 913  * @len:        how many bytes to splice
 914  * @flags:      splice modifier flags
 915  * @actor:      handler that splices the data
 916  *
 917  * Description:
 918  *    See __splice_from_pipe. This function locks the pipe inode,
 919  *    otherwise it's identical to __splice_from_pipe().
 920  *
 921  */
 922 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 923                          loff_t *ppos, size_t len, unsigned int flags,
 924                          splice_actor *actor)
 925 {
 926         ssize_t ret;
 927         struct splice_desc sd = {
 928                 .total_len = len,
 929                 .flags = flags,
 930                 .pos = *ppos,
 931                 .u.file = out,
 932         };
 933
 934         pipe_lock(pipe);
 935         ret = __splice_from_pipe(pipe, &sd, actor);
 936         pipe_unlock(pipe);
 937
 938         return ret;
 939 }
 940
 941 /**
 942  * iter_file_splice_write - splice data from a pipe to a file
 943  * @pipe:       pipe info
 944  * @out:        file to write to
 945  * @ppos:       position in @out
 946  * @len:        number of bytes to splice
 947  * @flags:      splice modifier flags
 948  *
 949  * Description:
 950  *    Will either move or copy pages (determined by @flags options) from
 951  *    the given pipe inode to the given file.
 952  *    This one is ->write_iter-based.
 953  *
 954  */
 955 ssize_t
 956 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 957                           loff_t *ppos, size_t len, unsigned int flags)
 958 {
 959         struct splice_desc sd = {
 960                 .total_len = len,
 961                 .flags = flags,
 962                 .pos = *ppos,
 963                 .u.file = out,
 964         };
 965         int nbufs = pipe->buffers;
 966         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 967                                         GFP_KERNEL);
 968         ssize_t ret;
 969
 970         if (unlikely(!array))
 971                 return -ENOMEM;
 972
 973         pipe_lock(pipe);
 974
 975         splice_from_pipe_begin(&sd);
 976         while (sd.total_len) {
 977                 struct iov_iter from;
 978                 size_t left;
 979                 int n, idx;
 980
 981                 ret = splice_from_pipe_next(pipe, &sd);
 982                 if (ret <= 0)
 983                         break;
 984
 985                 if (unlikely(nbufs < pipe->buffers)) {
 986                         kfree(array);
 987                         nbufs = pipe->buffers;
 988                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 989                                         GFP_KERNEL);
 990                         if (!array) {
 991                                 ret = -ENOMEM;
 992                                 break;
 993                         }
 994                 }
 995
 996                 /* build the vector */
 997                 left = sd.total_len;
 998                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 999                         struct pipe_buffer *buf = pipe->bufs + idx;
1000                         size_t this_len = buf->len;
1001
1002                         if (this_len > left)
1003                                 this_len = left;
1004
1005                         if (idx == pipe->buffers - 1)
1006                                 idx = -1;
1007
1008                         ret = buf->ops->confirm(pipe, buf);
1009                         if (unlikely(ret)) {
1010                                 if (ret == -ENODATA)
1011                                         ret = 0;
1012                                 goto done;
1013                         }
1014
1015                         array[n].bv_page = buf->page;
1016                         array[n].bv_len = this_len;
1017                         array[n].bv_offset = buf->offset;
1018                         left -= this_len;
1019                 }
1020
1021                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
1022                               sd.total_len - left);
1023                 ret = vfs_iter_write(out, &from, &sd.pos);
1024                 if (ret <= 0)
1025                         break;
1026
1027                 sd.num_spliced += ret;
1028                 sd.total_len -= ret;
1029                 *ppos = sd.pos;
1030
1031                 /* dismiss the fully eaten buffers, adjust the partial one */
1032                 while (ret) {
1033                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1034                         if (ret >= buf->len) {
1035                                 const struct pipe_buf_operations *ops = buf->ops;
1036                                 ret -= buf->len;
1037                                 buf->len = 0;
1038                                 buf->ops = NULL;
1039                                 ops->release(pipe, buf);
1040                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1041                                 pipe->nrbufs--;
1042                                 if (pipe->files)
1043                                         sd.need_wakeup = true;
1044                         } else {
1045                                 buf->offset += ret;
1046                                 buf->len -= ret;
1047                                 ret = 0;
1048                         }
1049                 }
1050         }
1051 done:
1052         kfree(array);
1053         splice_from_pipe_end(pipe, &sd);
1054
1055         pipe_unlock(pipe);
1056
1057         if (sd.num_spliced)
1058                 ret = sd.num_spliced;
1059
1060         return ret;
1061 }
1062
1063 EXPORT_SYMBOL(iter_file_splice_write);
1064
1065 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1066                           struct splice_desc *sd)
1067 {
1068         int ret;
1069         void *data;
1070         loff_t tmp = sd->pos;
1071
1072         data = kmap(buf->page);
1073         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1074         kunmap(buf->page);
1075
1076         return ret;
1077 }
1078
1079 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1080                                          struct file *out, loff_t *ppos,
1081                                          size_t len, unsigned int flags)
1082 {
1083         ssize_t ret;
1084
1085         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1086         if (ret > 0)
1087                 *ppos += ret;
1088
1089         return ret;
1090 }
1091
1092 /**
1093  * generic_splice_sendpage - splice data from a pipe to a socket
1094  * @pipe:       pipe to splice from
1095  * @out:        socket to write to
1096  * @ppos:       position in @out
1097  * @len:        number of bytes to splice
1098  * @flags:      splice modifier flags
1099  *
1100  * Description:
1101  *    Will send @len bytes from the pipe to a network socket. No data copying
1102  *    is involved.
1103  *
1104  */
1105 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1106                                 loff_t *ppos, size_t len, unsigned int flags)
1107 {
1108         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1109 }
1110
1111 EXPORT_SYMBOL(generic_splice_sendpage);
1112
1113 /*
1114  * Attempt to initiate a splice from pipe to file.
1115  */
1116 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1117                            loff_t *ppos, size_t len, unsigned int flags)
1118 {
1119         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1120                                 loff_t *, size_t, unsigned int);
1121
1122         if (out->f_op->splice_write)
1123                 splice_write = out->f_op->splice_write;
1124         else
1125                 splice_write = default_file_splice_write;
1126
1127         return splice_write(pipe, out, ppos, len, flags);
1128 }
1129
1130 /*
1131  * Attempt to initiate a splice from a file to a pipe.
1132  */
1133 static long do_splice_to(struct file *in, loff_t *ppos,
1134                          struct pipe_inode_info *pipe, size_t len,
1135                          unsigned int flags)
1136 {
1137         ssize_t (*splice_read)(struct file *, loff_t *,
1138                                struct pipe_inode_info *, size_t, unsigned int);
1139         int ret;
1140
1141         if (unlikely(!(in->f_mode & FMODE_READ)))
1142                 return -EBADF;
1143
1144         ret = rw_verify_area(READ, in, ppos, len);
1145         if (unlikely(ret < 0))
1146                 return ret;
1147
1148         if (in->f_op->splice_read)
1149                 splice_read = in->f_op->splice_read;
1150         else
1151                 splice_read = default_file_splice_read;
1152
1153         return splice_read(in, ppos, pipe, len, flags);
1154 }
1155
1156 /**
1157  * splice_direct_to_actor - splices data directly between two non-pipes
1158  * @in:         file to splice from
1159  * @sd:         actor information on where to splice to
1160  * @actor:      handles the data splicing
1161  *
1162  * Description:
1163  *    This is a special case helper to splice directly between two
1164  *    points, without requiring an explicit pipe. Internally an allocated
1165  *    pipe is cached in the process, and reused during the lifetime of
1166  *    that process.
1167  *
1168  */
1169 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1170                                splice_direct_actor *actor)
1171 {
1172         struct pipe_inode_info *pipe;
1173         long ret, bytes;
1174         umode_t i_mode;
1175         size_t len;
1176         int i, flags, more;
1177
1178         /*
1179          * We require the input being a regular file, as we don't want to
1180          * randomly drop data for eg socket -> socket splicing. Use the
1181          * piped splicing for that!
1182          */
1183         i_mode = file_inode(in)->i_mode;
1184         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1185                 return -EINVAL;
1186
1187         /*
1188          * neither in nor out is a pipe, setup an internal pipe attached to
1189          * 'out' and transfer the wanted data from 'in' to 'out' through that
1190          */
1191         pipe = current->splice_pipe;
1192         if (unlikely(!pipe)) {
1193                 pipe = alloc_pipe_info();
1194                 if (!pipe)
1195                         return -ENOMEM;
1196
1197                 /*
1198                  * We don't have an immediate reader, but we'll read the stuff
1199                  * out of the pipe right after the splice_to_pipe(). So set
1200                  * PIPE_READERS appropriately.
1201                  */
1202                 pipe->readers = 1;
1203
1204                 current->splice_pipe = pipe;
1205         }
1206
1207         /*
1208          * Do the splice.
1209          */
1210         ret = 0;
1211         bytes = 0;
1212         len = sd->total_len;
1213         flags = sd->flags;
1214
1215         /*
1216          * Don't block on output, we have to drain the direct pipe.
1217          */
1218         sd->flags &= ~SPLICE_F_NONBLOCK;
1219         more = sd->flags & SPLICE_F_MORE;
1220
1221         while (len) {
1222                 size_t read_len;
1223                 loff_t pos = sd->pos, prev_pos = pos;
1224
1225                 ret = do_splice_to(in, &pos, pipe, len, flags);
1226                 if (unlikely(ret <= 0))
1227                         goto out_release;
1228
1229                 read_len = ret;
1230                 sd->total_len = read_len;
1231
1232                 /*
1233                  * If more data is pending, set SPLICE_F_MORE
1234                  * If this is the last data and SPLICE_F_MORE was not set
1235                  * initially, clears it.
1236                  */
1237                 if (read_len < len)
1238                         sd->flags |= SPLICE_F_MORE;
1239                 else if (!more)
1240                         sd->flags &= ~SPLICE_F_MORE;
1241                 /*
1242                  * NOTE: nonblocking mode only applies to the input. We
1243                  * must not do the output in nonblocking mode as then we
1244                  * could get stuck data in the internal pipe:
1245                  */
1246                 ret = actor(pipe, sd);
1247                 if (unlikely(ret <= 0)) {
1248                         sd->pos = prev_pos;
1249                         goto out_release;
1250                 }
1251
1252                 bytes += ret;
1253                 len -= ret;
1254                 sd->pos = pos;
1255
1256                 if (ret < read_len) {
1257                         sd->pos = prev_pos + ret;
1258                         goto out_release;
1259                 }
1260         }
1261
1262 done:
1263         pipe->nrbufs = pipe->curbuf = 0;
1264         file_accessed(in);
1265         return bytes;
1266
1267 out_release:
1268         /*
1269          * If we did an incomplete transfer we must release
1270          * the pipe buffers in question:
1271          */
1272         for (i = 0; i < pipe->buffers; i++) {
1273                 struct pipe_buffer *buf = pipe->bufs + i;
1274
1275                 if (buf->ops) {
1276                         buf->ops->release(pipe, buf);
1277                         buf->ops = NULL;
1278                 }
1279         }
1280
1281         if (!bytes)
1282                 bytes = ret;
1283
1284         goto done;
1285 }
1286 EXPORT_SYMBOL(splice_direct_to_actor);
1287
1288 static int direct_splice_actor(struct pipe_inode_info *pipe,
1289                                struct splice_desc *sd)
1290 {
1291         struct file *file = sd->u.file;
1292
1293         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1294                               sd->flags);
1295 }
1296
1297 /**
1298  * do_splice_direct - splices data directly between two files
1299  * @in:         file to splice from
1300  * @ppos:       input file offset
1301  * @out:        file to splice to
1302  * @opos:       output file offset
1303  * @len:        number of bytes to splice
1304  * @flags:      splice modifier flags
1305  *
1306  * Description:
1307  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1308  *    doing it in the application would incur an extra system call
1309  *    (splice in + splice out, as compared to just sendfile()). So this helper
1310  *    can splice directly through a process-private pipe.
1311  *
1312  */
1313 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1314                       loff_t *opos, size_t len, unsigned int flags)
1315 {
1316         struct splice_desc sd = {
1317                 .len            = len,
1318                 .total_len      = len,
1319                 .flags          = flags,
1320                 .pos            = *ppos,
1321                 .u.file         = out,
1322                 .opos           = opos,
1323         };
1324         long ret;
1325
1326         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1327                 return -EBADF;
1328
1329         if (unlikely(out->f_flags & O_APPEND))
1330                 return -EINVAL;
1331
1332         ret = rw_verify_area(WRITE, out, opos, len);
1333         if (unlikely(ret < 0))
1334                 return ret;
1335
1336         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1337         if (ret > 0)
1338                 *ppos = sd.pos;
1339
1340         return ret;
1341 }
1342 EXPORT_SYMBOL(do_splice_direct);
1343
1344 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1345                                struct pipe_inode_info *opipe,
1346                                size_t len, unsigned int flags);
1347
1348 /*
1349  * Determine where to splice to/from.
1350  */
1351 static long do_splice(struct file *in, loff_t __user *off_in,
1352                       struct file *out, loff_t __user *off_out,
1353                       size_t len, unsigned int flags)
1354 {
1355         struct pipe_inode_info *ipipe;
1356         struct pipe_inode_info *opipe;
1357         loff_t offset;
1358         long ret;
1359
1360         ipipe = get_pipe_info(in);
1361         opipe = get_pipe_info(out);
1362
1363         if (ipipe && opipe) {
1364                 if (off_in || off_out)
1365                         return -ESPIPE;
1366
1367                 if (!(in->f_mode & FMODE_READ))
1368                         return -EBADF;
1369
1370                 if (!(out->f_mode & FMODE_WRITE))
1371                         return -EBADF;
1372
1373                 /* Splicing to self would be fun, but... */
1374                 if (ipipe == opipe)
1375                         return -EINVAL;
1376
1377                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1378         }
1379
1380         if (ipipe) {
1381                 if (off_in)
1382                         return -ESPIPE;
1383                 if (off_out) {
1384                         if (!(out->f_mode & FMODE_PWRITE))
1385                                 return -EINVAL;
1386                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1387                                 return -EFAULT;
1388                 } else {
1389                         offset = out->f_pos;
1390                 }
1391
1392                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1393                         return -EBADF;
1394
1395                 if (unlikely(out->f_flags & O_APPEND))
1396                         return -EINVAL;
1397
1398                 ret = rw_verify_area(WRITE, out, &offset, len);
1399                 if (unlikely(ret < 0))
1400                         return ret;
1401
1402                 file_start_write(out);
1403                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1404                 file_end_write(out);
1405
1406                 if (!off_out)
1407                         out->f_pos = offset;
1408                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1409                         ret = -EFAULT;
1410
1411                 return ret;
1412         }
1413
1414         if (opipe) {
1415                 if (off_out)
1416                         return -ESPIPE;
1417                 if (off_in) {
1418                         if (!(in->f_mode & FMODE_PREAD))
1419                                 return -EINVAL;
1420                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1421                                 return -EFAULT;
1422                 } else {
1423                         offset = in->f_pos;
1424                 }
1425
1426                 ret = do_splice_to(in, &offset, opipe, len, flags);
1427
1428                 if (!off_in)
1429                         in->f_pos = offset;
1430                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1431                         ret = -EFAULT;
1432
1433                 return ret;
1434         }
1435
1436         return -EINVAL;
1437 }
1438
1439 /*
1440  * Map an iov into an array of pages and offset/length tupples. With the
1441  * partial_page structure, we can map several non-contiguous ranges into
1442  * our ones pages[] map instead of splitting that operation into pieces.
1443  * Could easily be exported as a generic helper for other users, in which
1444  * case one would probably want to add a 'max_nr_pages' parameter as well.
1445  */
1446 static int get_iovec_page_array(const struct iovec __user *iov,
1447                                 unsigned int nr_vecs, struct page **pages,
1448                                 struct partial_page *partial, bool aligned,
1449                                 unsigned int pipe_buffers)
1450 {
1451         int buffers = 0, error = 0;
1452
1453         while (nr_vecs) {
1454                 unsigned long off, npages;
1455                 struct iovec entry;
1456                 void __user *base;
1457                 size_t len;
1458                 int i;
1459
1460                 error = -EFAULT;
1461                 if (copy_from_user(&entry, iov, sizeof(entry)))
1462                         break;
1463
1464                 base = entry.iov_base;
1465                 len = entry.iov_len;
1466
1467                 /*
1468                  * Sanity check this iovec. 0 read succeeds.
1469                  */
1470                 error = 0;
1471                 if (unlikely(!len))
1472                         break;
1473                 error = -EFAULT;
1474                 if (!access_ok(VERIFY_READ, base, len))
1475                         break;
1476
1477                 /*
1478                  * Get this base offset and number of pages, then map
1479                  * in the user pages.
1480                  */
1481                 off = (unsigned long) base & ~PAGE_MASK;
1482
1483                 /*
1484                  * If asked for alignment, the offset must be zero and the
1485                  * length a multiple of the PAGE_SIZE.
1486                  */
1487                 error = -EINVAL;
1488                 if (aligned && (off || len & ~PAGE_MASK))
1489                         break;
1490
1491                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1492                 if (npages > pipe_buffers - buffers)
1493                         npages = pipe_buffers - buffers;
1494
1495                 error = get_user_pages_fast((unsigned long)base, npages,
1496                                         0, &pages[buffers]);
1497
1498                 if (unlikely(error <= 0))
1499                         break;
1500
1501                 /*
1502                  * Fill this contiguous range into the partial page map.
1503                  */
1504                 for (i = 0; i < error; i++) {
1505                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1506
1507                         partial[buffers].offset = off;
1508                         partial[buffers].len = plen;
1509
1510                         off = 0;
1511                         len -= plen;
1512                         buffers++;
1513                 }
1514
1515                 /*
1516                  * We didn't complete this iov, stop here since it probably
1517                  * means we have to move some of this into a pipe to
1518                  * be able to continue.
1519                  */
1520                 if (len)
1521                         break;
1522
1523                 /*
1524                  * Don't continue if we mapped fewer pages than we asked for,
1525                  * or if we mapped the max number of pages that we have
1526                  * room for.
1527                  */
1528                 if (error < npages || buffers == pipe_buffers)
1529                         break;
1530
1531                 nr_vecs--;
1532                 iov++;
1533         }
1534
1535         if (buffers)
1536                 return buffers;
1537
1538         return error;
1539 }
1540
1541 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1542                         struct splice_desc *sd)
1543 {
1544         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1545         return n == sd->len ? n : -EFAULT;
1546 }
1547
1548 /*
1549  * For lack of a better implementation, implement vmsplice() to userspace
1550  * as a simple copy of the pipes pages to the user iov.
1551  */
1552 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1553                              unsigned long nr_segs, unsigned int flags)
1554 {
1555         struct pipe_inode_info *pipe;
1556         struct splice_desc sd;
1557         long ret;
1558         struct iovec iovstack[UIO_FASTIOV];
1559         struct iovec *iov = iovstack;
1560         struct iov_iter iter;
1561
1562         pipe = get_pipe_info(file);
1563         if (!pipe)
1564                 return -EBADF;
1565
1566         ret = import_iovec(READ, uiov, nr_segs,
1567                            ARRAY_SIZE(iovstack), &iov, &iter);
1568         if (ret < 0)
1569                 return ret;
1570
1571         sd.total_len = iov_iter_count(&iter);
1572         sd.len = 0;
1573         sd.flags = flags;
1574         sd.u.data = &iter;
1575         sd.pos = 0;
1576
1577         if (sd.total_len) {
1578                 pipe_lock(pipe);
1579                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1580                 pipe_unlock(pipe);
1581         }
1582
1583         kfree(iov);
1584         return ret;
1585 }
1586
1587 /*
1588  * vmsplice splices a user address range into a pipe. It can be thought of
1589  * as splice-from-memory, where the regular splice is splice-from-file (or
1590  * to file). In both cases the output is a pipe, naturally.
1591  */
1592 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1593                              unsigned long nr_segs, unsigned int flags)
1594 {
1595         struct pipe_inode_info *pipe;
1596         struct page *pages[PIPE_DEF_BUFFERS];
1597         struct partial_page partial[PIPE_DEF_BUFFERS];
1598         struct splice_pipe_desc spd = {
1599                 .pages = pages,
1600                 .partial = partial,
1601                 .nr_pages_max = PIPE_DEF_BUFFERS,
1602                 .flags = flags,
1603                 .ops = &user_page_pipe_buf_ops,
1604                 .spd_release = spd_release_page,
1605         };
1606         long ret;
1607
1608         pipe = get_pipe_info(file);
1609         if (!pipe)
1610                 return -EBADF;
1611
1612         if (splice_grow_spd(pipe, &spd))
1613                 return -ENOMEM;
1614
1615         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1616                                             spd.partial, false,
1617                                             spd.nr_pages_max);
1618         if (spd.nr_pages <= 0)
1619                 ret = spd.nr_pages;
1620         else
1621                 ret = splice_to_pipe(pipe, &spd);
1622
1623         splice_shrink_spd(&spd);
1624         return ret;
1625 }
1626
1627 /*
1628  * Note that vmsplice only really supports true splicing _from_ user memory
1629  * to a pipe, not the other way around. Splicing from user memory is a simple
1630  * operation that can be supported without any funky alignment restrictions
1631  * or nasty vm tricks. We simply map in the user memory and fill them into
1632  * a pipe. The reverse isn't quite as easy, though. There are two possible
1633  * solutions for that:
1634  *
1635  *      - memcpy() the data internally, at which point we might as well just
1636  *        do a regular read() on the buffer anyway.
1637  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1638  *        has restriction limitations on both ends of the pipe).
1639  *
1640  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1641  *
1642  */
1643 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1644                 unsigned long, nr_segs, unsigned int, flags)
1645 {
1646         struct fd f;
1647         long error;
1648
1649         if (unlikely(nr_segs > UIO_MAXIOV))
1650                 return -EINVAL;
1651         else if (unlikely(!nr_segs))
1652                 return 0;
1653
1654         error = -EBADF;
1655         f = fdget(fd);
1656         if (f.file) {
1657                 if (f.file->f_mode & FMODE_WRITE)
1658                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1659                 else if (f.file->f_mode & FMODE_READ)
1660                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1661
1662                 fdput(f);
1663         }
1664
1665         return error;
1666 }
1667
1668 #ifdef CONFIG_COMPAT
1669 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1670                     unsigned int, nr_segs, unsigned int, flags)
1671 {
1672         unsigned i;
1673         struct iovec __user *iov;
1674         if (nr_segs > UIO_MAXIOV)
1675                 return -EINVAL;
1676         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1677         for (i = 0; i < nr_segs; i++) {
1678                 struct compat_iovec v;
1679                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1680                     get_user(v.iov_len, &iov32[i].iov_len) ||
1681                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1682                     put_user(v.iov_len, &iov[i].iov_len))
1683                         return -EFAULT;
1684         }
1685         return sys_vmsplice(fd, iov, nr_segs, flags);
1686 }
1687 #endif
1688
1689 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1690                 int, fd_out, loff_t __user *, off_out,
1691                 size_t, len, unsigned int, flags)
1692 {
1693         struct fd in, out;
1694         long error;
1695
1696         if (unlikely(!len))
1697                 return 0;
1698
1699         error = -EBADF;
1700         in = fdget(fd_in);
1701         if (in.file) {
1702                 if (in.file->f_mode & FMODE_READ) {
1703                         out = fdget(fd_out);
1704                         if (out.file) {
1705                                 if (out.file->f_mode & FMODE_WRITE)
1706                                         error = do_splice(in.file, off_in,
1707                                                           out.file, off_out,
1708                                                           len, flags);
1709                                 fdput(out);
1710                         }
1711                 }
1712                 fdput(in);
1713         }
1714         return error;
1715 }
1716
1717 /*
1718  * Make sure there's data to read. Wait for input if we can, otherwise
1719  * return an appropriate error.
1720  */
1721 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1722 {
1723         int ret;
1724
1725         /*
1726          * Check ->nrbufs without the inode lock first. This function
1727          * is speculative anyways, so missing one is ok.
1728          */
1729         if (pipe->nrbufs)
1730                 return 0;
1731
1732         ret = 0;
1733         pipe_lock(pipe);
1734
1735         while (!pipe->nrbufs) {
1736                 if (signal_pending(current)) {
1737                         ret = -ERESTARTSYS;
1738                         break;
1739                 }
1740                 if (!pipe->writers)
1741                         break;
1742                 if (!pipe->waiting_writers) {
1743                         if (flags & SPLICE_F_NONBLOCK) {
1744                                 ret = -EAGAIN;
1745                                 break;
1746                         }
1747                 }
1748                 pipe_wait(pipe);
1749         }
1750
1751         pipe_unlock(pipe);
1752         return ret;
1753 }
1754
1755 /*
1756  * Make sure there's writeable room. Wait for room if we can, otherwise
1757  * return an appropriate error.
1758  */
1759 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1760 {
1761         int ret;
1762
1763         /*
1764          * Check ->nrbufs without the inode lock first. This function
1765          * is speculative anyways, so missing one is ok.
1766          */
1767         if (pipe->nrbufs < pipe->buffers)
1768                 return 0;
1769
1770         ret = 0;
1771         pipe_lock(pipe);
1772
1773         while (pipe->nrbufs >= pipe->buffers) {
1774                 if (!pipe->readers) {
1775                         send_sig(SIGPIPE, current, 0);
1776                         ret = -EPIPE;
1777                         break;
1778                 }
1779                 if (flags & SPLICE_F_NONBLOCK) {
1780                         ret = -EAGAIN;
1781                         break;
1782                 }
1783                 if (signal_pending(current)) {
1784                         ret = -ERESTARTSYS;
1785                         break;
1786                 }
1787                 pipe->waiting_writers++;
1788                 pipe_wait(pipe);
1789                 pipe->waiting_writers--;
1790         }
1791
1792         pipe_unlock(pipe);
1793         return ret;
1794 }
1795
1796 /*
1797  * Splice contents of ipipe to opipe.
1798  */
1799 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1800                                struct pipe_inode_info *opipe,
1801                                size_t len, unsigned int flags)
1802 {
1803         struct pipe_buffer *ibuf, *obuf;
1804         int ret = 0, nbuf;
1805         bool input_wakeup = false;
1806
1807
1808 retry:
1809         ret = ipipe_prep(ipipe, flags);
1810         if (ret)
1811                 return ret;
1812
1813         ret = opipe_prep(opipe, flags);
1814         if (ret)
1815                 return ret;
1816
1817         /*
1818          * Potential ABBA deadlock, work around it by ordering lock
1819          * grabbing by pipe info address. Otherwise two different processes
1820          * could deadlock (one doing tee from A -> B, the other from B -> A).
1821          */
1822         pipe_double_lock(ipipe, opipe);
1823
1824         do {
1825                 if (!opipe->readers) {
1826                         send_sig(SIGPIPE, current, 0);
1827                         if (!ret)
1828                                 ret = -EPIPE;
1829                         break;
1830                 }
1831
1832                 if (!ipipe->nrbufs && !ipipe->writers)
1833                         break;
1834
1835                 /*
1836                  * Cannot make any progress, because either the input
1837                  * pipe is empty or the output pipe is full.
1838                  */
1839                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1840                         /* Already processed some buffers, break */
1841                         if (ret)
1842                                 break;
1843
1844                         if (flags & SPLICE_F_NONBLOCK) {
1845                                 ret = -EAGAIN;
1846                                 break;
1847                         }
1848
1849                         /*
1850                          * We raced with another reader/writer and haven't
1851                          * managed to process any buffers.  A zero return
1852                          * value means EOF, so retry instead.
1853                          */
1854                         pipe_unlock(ipipe);
1855                         pipe_unlock(opipe);
1856                         goto retry;
1857                 }
1858
1859                 ibuf = ipipe->bufs + ipipe->curbuf;
1860                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1861                 obuf = opipe->bufs + nbuf;
1862
1863                 if (len >= ibuf->len) {
1864                         /*
1865                          * Simply move the whole buffer from ipipe to opipe
1866                          */
1867                         *obuf = *ibuf;
1868                         ibuf->ops = NULL;
1869                         opipe->nrbufs++;
1870                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1871                         ipipe->nrbufs--;
1872                         input_wakeup = true;
1873                 } else {
1874                         /*
1875                          * Get a reference to this pipe buffer,
1876                          * so we can copy the contents over.
1877                          */
1878                         ibuf->ops->get(ipipe, ibuf);
1879                         *obuf = *ibuf;
1880
1881                         /*
1882                          * Don't inherit the gift flag, we need to
1883                          * prevent multiple steals of this page.
1884                          */
1885                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1886
1887                         obuf->len = len;
1888                         opipe->nrbufs++;
1889                         ibuf->offset += obuf->len;
1890                         ibuf->len -= obuf->len;
1891                 }
1892                 ret += obuf->len;
1893                 len -= obuf->len;
1894         } while (len);
1895
1896         pipe_unlock(ipipe);
1897         pipe_unlock(opipe);
1898
1899         /*
1900          * If we put data in the output pipe, wakeup any potential readers.
1901          */
1902         if (ret > 0)
1903                 wakeup_pipe_readers(opipe);
1904
1905         if (input_wakeup)
1906                 wakeup_pipe_writers(ipipe);
1907
1908         return ret;
1909 }
1910
1911 /*
1912  * Link contents of ipipe to opipe.
1913  */
1914 static int link_pipe(struct pipe_inode_info *ipipe,
1915                      struct pipe_inode_info *opipe,
1916                      size_t len, unsigned int flags)
1917 {
1918         struct pipe_buffer *ibuf, *obuf;
1919         int ret = 0, i = 0, nbuf;
1920
1921         /*
1922          * Potential ABBA deadlock, work around it by ordering lock
1923          * grabbing by pipe info address. Otherwise two different processes
1924          * could deadlock (one doing tee from A -> B, the other from B -> A).
1925          */
1926         pipe_double_lock(ipipe, opipe);
1927
1928         do {
1929                 if (!opipe->readers) {
1930                         send_sig(SIGPIPE, current, 0);
1931                         if (!ret)
1932                                 ret = -EPIPE;
1933                         break;
1934                 }
1935
1936                 /*
1937                  * If we have iterated all input buffers or ran out of
1938                  * output room, break.
1939                  */
1940                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1941                         break;
1942
1943                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1944                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1945
1946                 /*
1947                  * Get a reference to this pipe buffer,
1948                  * so we can copy the contents over.
1949                  */
1950                 ibuf->ops->get(ipipe, ibuf);
1951
1952                 obuf = opipe->bufs + nbuf;
1953                 *obuf = *ibuf;
1954
1955                 /*
1956                  * Don't inherit the gift flag, we need to
1957                  * prevent multiple steals of this page.
1958                  */
1959                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1960
1961                 if (obuf->len > len)
1962                         obuf->len = len;
1963
1964                 opipe->nrbufs++;
1965                 ret += obuf->len;
1966                 len -= obuf->len;
1967                 i++;
1968         } while (len);
1969
1970         /*
1971          * return EAGAIN if we have the potential of some data in the
1972          * future, otherwise just return 0
1973          */
1974         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1975                 ret = -EAGAIN;
1976
1977         pipe_unlock(ipipe);
1978         pipe_unlock(opipe);
1979
1980         /*
1981          * If we put data in the output pipe, wakeup any potential readers.
1982          */
1983         if (ret > 0)
1984                 wakeup_pipe_readers(opipe);
1985
1986         return ret;
1987 }
1988
1989 /*
1990  * This is a tee(1) implementation that works on pipes. It doesn't copy
1991  * any data, it simply references the 'in' pages on the 'out' pipe.
1992  * The 'flags' used are the SPLICE_F_* variants, currently the only
1993  * applicable one is SPLICE_F_NONBLOCK.
1994  */
1995 static long do_tee(struct file *in, struct file *out, size_t len,
1996                    unsigned int flags)
1997 {
1998         struct pipe_inode_info *ipipe = get_pipe_info(in);
1999         struct pipe_inode_info *opipe = get_pipe_info(out);
2000         int ret = -EINVAL;
2001
2002         /*
2003          * Duplicate the contents of ipipe to opipe without actually
2004          * copying the data.
2005          */
2006         if (ipipe && opipe && ipipe != opipe) {
2007                 /*
2008                  * Keep going, unless we encounter an error. The ipipe/opipe
2009                  * ordering doesn't really matter.
2010                  */
2011                 ret = ipipe_prep(ipipe, flags);
2012                 if (!ret) {
2013                         ret = opipe_prep(opipe, flags);
2014                         if (!ret)
2015                                 ret = link_pipe(ipipe, opipe, len, flags);
2016                 }
2017         }
2018
2019         return ret;
2020 }
2021
2022 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2023 {
2024         struct fd in;
2025         int error;
2026
2027         if (unlikely(!len))
2028                 return 0;
2029
2030         error = -EBADF;
2031         in = fdget(fdin);
2032         if (in.file) {
2033                 if (in.file->f_mode & FMODE_READ) {
2034                         struct fd out = fdget(fdout);
2035                         if (out.file) {
2036                                 if (out.file->f_mode & FMODE_WRITE)
2037                                         error = do_tee(in.file, out.file,
2038                                                         len, flags);
2039                                 fdput(out);
2040                         }
2041                 }
2042                 fdput(in);
2043         }
2044
2045         return error;
2046 }