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