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         .map = generic_pipe_buf_map,
 140         .unmap = generic_pipe_buf_unmap,
 141         .confirm = page_cache_pipe_buf_confirm,
 142         .release = page_cache_pipe_buf_release,
 143         .steal = page_cache_pipe_buf_steal,
 144         .get = generic_pipe_buf_get,
 145 };
 146
 147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 148                                     struct pipe_buffer *buf)
 149 {
 150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 151                 return 1;
 152
 153         buf->flags |= PIPE_BUF_FLAG_LRU;
 154         return generic_pipe_buf_steal(pipe, buf);
 155 }
 156
 157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 158         .can_merge = 0,
 159         .map = generic_pipe_buf_map,
 160         .unmap = generic_pipe_buf_unmap,
 161         .confirm = generic_pipe_buf_confirm,
 162         .release = page_cache_pipe_buf_release,
 163         .steal = user_page_pipe_buf_steal,
 164         .get = generic_pipe_buf_get,
 165 };
 166
 167 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 168 {
 169         smp_mb();
 170         if (waitqueue_active(&pipe->wait))
 171                 wake_up_interruptible(&pipe->wait);
 172         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 173 }
 174
 175 /**
 176  * splice_to_pipe - fill passed data into a pipe
 177  * @pipe:       pipe to fill
 178  * @spd:        data to fill
 179  *
 180  * Description:
 181  *    @spd contains a map of pages and len/offset tuples, along with
 182  *    the struct pipe_buf_operations associated with these pages. This
 183  *    function will link that data to the pipe.
 184  *
 185  */
 186 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 187                        struct splice_pipe_desc *spd)
 188 {
 189         unsigned int spd_pages = spd->nr_pages;
 190         int ret, do_wakeup, page_nr;
 191
 192         ret = 0;
 193         do_wakeup = 0;
 194         page_nr = 0;
 195
 196         pipe_lock(pipe);
 197
 198         for (;;) {
 199                 if (!pipe->readers) {
 200                         send_sig(SIGPIPE, current, 0);
 201                         if (!ret)
 202                                 ret = -EPIPE;
 203                         break;
 204                 }
 205
 206                 if (pipe->nrbufs < pipe->buffers) {
 207                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 208                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 209
 210                         buf->page = spd->pages[page_nr];
 211                         buf->offset = spd->partial[page_nr].offset;
 212                         buf->len = spd->partial[page_nr].len;
 213                         buf->private = spd->partial[page_nr].private;
 214                         buf->ops = spd->ops;
 215                         if (spd->flags & SPLICE_F_GIFT)
 216                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 217
 218                         pipe->nrbufs++;
 219                         page_nr++;
 220                         ret += buf->len;
 221
 222                         if (pipe->files)
 223                                 do_wakeup = 1;
 224
 225                         if (!--spd->nr_pages)
 226                                 break;
 227                         if (pipe->nrbufs < pipe->buffers)
 228                                 continue;
 229
 230                         break;
 231                 }
 232
 233                 if (spd->flags & SPLICE_F_NONBLOCK) {
 234                         if (!ret)
 235                                 ret = -EAGAIN;
 236                         break;
 237                 }
 238
 239                 if (signal_pending(current)) {
 240                         if (!ret)
 241                                 ret = -ERESTARTSYS;
 242                         break;
 243                 }
 244
 245                 if (do_wakeup) {
 246                         smp_mb();
 247                         if (waitqueue_active(&pipe->wait))
 248                                 wake_up_interruptible_sync(&pipe->wait);
 249                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 250                         do_wakeup = 0;
 251                 }
 252
 253                 pipe->waiting_writers++;
 254                 pipe_wait(pipe);
 255                 pipe->waiting_writers--;
 256         }
 257
 258         pipe_unlock(pipe);
 259
 260         if (do_wakeup)
 261                 wakeup_pipe_readers(pipe);
 262
 263         while (page_nr < spd_pages)
 264                 spd->spd_release(spd, page_nr++);
 265
 266         return ret;
 267 }
 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                                                 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         isize = i_size_read(in->f_mapping->host);
 531         if (unlikely(*ppos >= isize))
 532                 return 0;
 533
 534         left = isize - *ppos;
 535         if (unlikely(left < len))
 536                 len = left;
 537
 538         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 539         if (ret > 0) {
 540                 *ppos += ret;
 541                 file_accessed(in);
 542         }
 543
 544         return ret;
 545 }
 546 EXPORT_SYMBOL(generic_file_splice_read);
 547
 548 static const struct pipe_buf_operations default_pipe_buf_ops = {
 549         .can_merge = 0,
 550         .map = generic_pipe_buf_map,
 551         .unmap = generic_pipe_buf_unmap,
 552         .confirm = generic_pipe_buf_confirm,
 553         .release = generic_pipe_buf_release,
 554         .steal = generic_pipe_buf_steal,
 555         .get = generic_pipe_buf_get,
 556 };
 557
 558 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 559                                     struct pipe_buffer *buf)
 560 {
 561         return 1;
 562 }
 563
 564 /* Pipe buffer operations for a socket and similar. */
 565 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 566         .can_merge = 0,
 567         .map = generic_pipe_buf_map,
 568         .unmap = generic_pipe_buf_unmap,
 569         .confirm = generic_pipe_buf_confirm,
 570         .release = generic_pipe_buf_release,
 571         .steal = generic_pipe_buf_nosteal,
 572         .get = generic_pipe_buf_get,
 573 };
 574 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 575
 576 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 577                             unsigned long vlen, loff_t offset)
 578 {
 579         mm_segment_t old_fs;
 580         loff_t pos = offset;
 581         ssize_t res;
 582
 583         old_fs = get_fs();
 584         set_fs(get_ds());
 585         /* The cast to a user pointer is valid due to the set_fs() */
 586         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 587         set_fs(old_fs);
 588
 589         return res;
 590 }
 591
 592 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 593                             loff_t pos)
 594 {
 595         mm_segment_t old_fs;
 596         ssize_t res;
 597
 598         old_fs = get_fs();
 599         set_fs(get_ds());
 600         /* The cast to a user pointer is valid due to the set_fs() */
 601         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 602         set_fs(old_fs);
 603
 604         return res;
 605 }
 606 EXPORT_SYMBOL(kernel_write);
 607
 608 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 609                                  struct pipe_inode_info *pipe, size_t len,
 610                                  unsigned int flags)
 611 {
 612         unsigned int nr_pages;
 613         unsigned int nr_freed;
 614         size_t offset;
 615         struct page *pages[PIPE_DEF_BUFFERS];
 616         struct partial_page partial[PIPE_DEF_BUFFERS];
 617         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 618         ssize_t res;
 619         size_t this_len;
 620         int error;
 621         int i;
 622         struct splice_pipe_desc spd = {
 623                 .pages = pages,
 624                 .partial = partial,
 625                 .nr_pages_max = PIPE_DEF_BUFFERS,
 626                 .flags = flags,
 627                 .ops = &default_pipe_buf_ops,
 628                 .spd_release = spd_release_page,
 629         };
 630
 631         if (splice_grow_spd(pipe, &spd))
 632                 return -ENOMEM;
 633
 634         res = -ENOMEM;
 635         vec = __vec;
 636         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 637                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 638                 if (!vec)
 639                         goto shrink_ret;
 640         }
 641
 642         offset = *ppos & ~PAGE_CACHE_MASK;
 643         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 644
 645         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 646                 struct page *page;
 647
 648                 page = alloc_page(GFP_USER);
 649                 error = -ENOMEM;
 650                 if (!page)
 651                         goto err;
 652
 653                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 654                 vec[i].iov_base = (void __user *) page_address(page);
 655                 vec[i].iov_len = this_len;
 656                 spd.pages[i] = page;
 657                 spd.nr_pages++;
 658                 len -= this_len;
 659                 offset = 0;
 660         }
 661
 662         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 663         if (res < 0) {
 664                 error = res;
 665                 goto err;
 666         }
 667
 668         error = 0;
 669         if (!res)
 670                 goto err;
 671
 672         nr_freed = 0;
 673         for (i = 0; i < spd.nr_pages; i++) {
 674                 this_len = min_t(size_t, vec[i].iov_len, res);
 675                 spd.partial[i].offset = 0;
 676                 spd.partial[i].len = this_len;
 677                 if (!this_len) {
 678                         __free_page(spd.pages[i]);
 679                         spd.pages[i] = NULL;
 680                         nr_freed++;
 681                 }
 682                 res -= this_len;
 683         }
 684         spd.nr_pages -= nr_freed;
 685
 686         res = splice_to_pipe(pipe, &spd);
 687         if (res > 0)
 688                 *ppos += res;
 689
 690 shrink_ret:
 691         if (vec != __vec)
 692                 kfree(vec);
 693         splice_shrink_spd(&spd);
 694         return res;
 695
 696 err:
 697         for (i = 0; i < spd.nr_pages; i++)
 698                 __free_page(spd.pages[i]);
 699
 700         res = error;
 701         goto shrink_ret;
 702 }
 703 EXPORT_SYMBOL(default_file_splice_read);
 704
 705 /*
 706  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 707  * using sendpage(). Return the number of bytes sent.
 708  */
 709 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 710                             struct pipe_buffer *buf, struct splice_desc *sd)
 711 {
 712         struct file *file = sd->u.file;
 713         loff_t pos = sd->pos;
 714         int more;
 715
 716         if (!likely(file->f_op && file->f_op->sendpage))
 717                 return -EINVAL;
 718
 719         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 720
 721         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 722                 more |= MSG_SENDPAGE_NOTLAST;
 723
 724         return file->f_op->sendpage(file, buf->page, buf->offset,
 725                                     sd->len, &pos, more);
 726 }
 727
 728 /*
 729  * This is a little more tricky than the file -> pipe splicing. There are
 730  * basically three cases:
 731  *
 732  *      - Destination page already exists in the address space and there
 733  *        are users of it. For that case we have no other option that
 734  *        copying the data. Tough luck.
 735  *      - Destination page already exists in the address space, but there
 736  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 737  *        through to last case.
 738  *      - Destination page does not exist, we can add the pipe page to
 739  *        the page cache and avoid the copy.
 740  *
 741  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 742  * sd->flags), we attempt to migrate pages from the pipe to the output
 743  * file address space page cache. This is possible if no one else has
 744  * the pipe page referenced outside of the pipe and page cache. If
 745  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 746  * a new page in the output file page cache and fill/dirty that.
 747  */
 748 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 749                  struct splice_desc *sd)
 750 {
 751         struct file *file = sd->u.file;
 752         struct address_space *mapping = file->f_mapping;
 753         unsigned int offset, this_len;
 754         struct page *page;
 755         void *fsdata;
 756         int ret;
 757
 758         offset = sd->pos & ~PAGE_CACHE_MASK;
 759
 760         this_len = sd->len;
 761         if (this_len + offset > PAGE_CACHE_SIZE)
 762                 this_len = PAGE_CACHE_SIZE - offset;
 763
 764         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 765                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 766         if (unlikely(ret))
 767                 goto out;
 768
 769         if (buf->page != page) {
 770                 char *src = buf->ops->map(pipe, buf, 1);
 771                 char *dst = kmap_atomic(page);
 772
 773                 memcpy(dst + offset, src + buf->offset, this_len);
 774                 flush_dcache_page(page);
 775                 kunmap_atomic(dst);
 776                 buf->ops->unmap(pipe, buf, src);
 777         }
 778         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 779                                 page, fsdata);
 780 out:
 781         return ret;
 782 }
 783 EXPORT_SYMBOL(pipe_to_file);
 784
 785 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 786 {
 787         smp_mb();
 788         if (waitqueue_active(&pipe->wait))
 789                 wake_up_interruptible(&pipe->wait);
 790         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 791 }
 792
 793 /**
 794  * splice_from_pipe_feed - feed available data from a pipe to a file
 795  * @pipe:       pipe to splice from
 796  * @sd:         information to @actor
 797  * @actor:      handler that splices the data
 798  *
 799  * Description:
 800  *    This function loops over the pipe and calls @actor to do the
 801  *    actual moving of a single struct pipe_buffer to the desired
 802  *    destination.  It returns when there's no more buffers left in
 803  *    the pipe or if the requested number of bytes (@sd->total_len)
 804  *    have been copied.  It returns a positive number (one) if the
 805  *    pipe needs to be filled with more data, zero if the required
 806  *    number of bytes have been copied and -errno on error.
 807  *
 808  *    This, together with splice_from_pipe_{begin,end,next}, may be
 809  *    used to implement the functionality of __splice_from_pipe() when
 810  *    locking is required around copying the pipe buffers to the
 811  *    destination.
 812  */
 813 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 814                           splice_actor *actor)
 815 {
 816         int ret;
 817
 818         while (pipe->nrbufs) {
 819                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 820                 const struct pipe_buf_operations *ops = buf->ops;
 821
 822                 sd->len = buf->len;
 823                 if (sd->len > sd->total_len)
 824                         sd->len = sd->total_len;
 825
 826                 ret = buf->ops->confirm(pipe, buf);
 827                 if (unlikely(ret)) {
 828                         if (ret == -ENODATA)
 829                                 ret = 0;
 830                         return ret;
 831                 }
 832
 833                 ret = actor(pipe, buf, sd);
 834                 if (ret <= 0)
 835                         return ret;
 836
 837                 buf->offset += ret;
 838                 buf->len -= ret;
 839
 840                 sd->num_spliced += ret;
 841                 sd->len -= ret;
 842                 sd->pos += ret;
 843                 sd->total_len -= ret;
 844
 845                 if (!buf->len) {
 846                         buf->ops = NULL;
 847                         ops->release(pipe, buf);
 848                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 849                         pipe->nrbufs--;
 850                         if (pipe->files)
 851                                 sd->need_wakeup = true;
 852                 }
 853
 854                 if (!sd->total_len)
 855                         return 0;
 856         }
 857
 858         return 1;
 859 }
 860 EXPORT_SYMBOL(splice_from_pipe_feed);
 861
 862 /**
 863  * splice_from_pipe_next - wait for some data to splice from
 864  * @pipe:       pipe to splice from
 865  * @sd:         information about the splice operation
 866  *
 867  * Description:
 868  *    This function will wait for some data and return a positive
 869  *    value (one) if pipe buffers are available.  It will return zero
 870  *    or -errno if no more data needs to be spliced.
 871  */
 872 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 873 {
 874         while (!pipe->nrbufs) {
 875                 if (!pipe->writers)
 876                         return 0;
 877
 878                 if (!pipe->waiting_writers && sd->num_spliced)
 879                         return 0;
 880
 881                 if (sd->flags & SPLICE_F_NONBLOCK)
 882                         return -EAGAIN;
 883
 884                 if (signal_pending(current))
 885                         return -ERESTARTSYS;
 886
 887                 if (sd->need_wakeup) {
 888                         wakeup_pipe_writers(pipe);
 889                         sd->need_wakeup = false;
 890                 }
 891
 892                 pipe_wait(pipe);
 893         }
 894
 895         return 1;
 896 }
 897 EXPORT_SYMBOL(splice_from_pipe_next);
 898
 899 /**
 900  * splice_from_pipe_begin - start splicing from pipe
 901  * @sd:         information about the splice operation
 902  *
 903  * Description:
 904  *    This function should be called before a loop containing
 905  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 906  *    initialize the necessary fields of @sd.
 907  */
 908 void splice_from_pipe_begin(struct splice_desc *sd)
 909 {
 910         sd->num_spliced = 0;
 911         sd->need_wakeup = false;
 912 }
 913 EXPORT_SYMBOL(splice_from_pipe_begin);
 914
 915 /**
 916  * splice_from_pipe_end - finish splicing from pipe
 917  * @pipe:       pipe to splice from
 918  * @sd:         information about the splice operation
 919  *
 920  * Description:
 921  *    This function will wake up pipe writers if necessary.  It should
 922  *    be called after a loop containing splice_from_pipe_next() and
 923  *    splice_from_pipe_feed().
 924  */
 925 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 926 {
 927         if (sd->need_wakeup)
 928                 wakeup_pipe_writers(pipe);
 929 }
 930 EXPORT_SYMBOL(splice_from_pipe_end);
 931
 932 /**
 933  * __splice_from_pipe - splice data from a pipe to given actor
 934  * @pipe:       pipe to splice from
 935  * @sd:         information to @actor
 936  * @actor:      handler that splices the data
 937  *
 938  * Description:
 939  *    This function does little more than loop over the pipe and call
 940  *    @actor to do the actual moving of a single struct pipe_buffer to
 941  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 942  *    pipe_to_user.
 943  *
 944  */
 945 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 946                            splice_actor *actor)
 947 {
 948         int ret;
 949
 950         splice_from_pipe_begin(sd);
 951         do {
 952                 ret = splice_from_pipe_next(pipe, sd);
 953                 if (ret > 0)
 954                         ret = splice_from_pipe_feed(pipe, sd, actor);
 955         } while (ret > 0);
 956         splice_from_pipe_end(pipe, sd);
 957
 958         return sd->num_spliced ? sd->num_spliced : ret;
 959 }
 960 EXPORT_SYMBOL(__splice_from_pipe);
 961
 962 /**
 963  * splice_from_pipe - splice data from a pipe to a file
 964  * @pipe:       pipe to splice from
 965  * @out:        file to splice to
 966  * @ppos:       position in @out
 967  * @len:        how many bytes to splice
 968  * @flags:      splice modifier flags
 969  * @actor:      handler that splices the data
 970  *
 971  * Description:
 972  *    See __splice_from_pipe. This function locks the pipe inode,
 973  *    otherwise it's identical to __splice_from_pipe().
 974  *
 975  */
 976 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 977                          loff_t *ppos, size_t len, unsigned int flags,
 978                          splice_actor *actor)
 979 {
 980         ssize_t ret;
 981         struct splice_desc sd = {
 982                 .total_len = len,
 983                 .flags = flags,
 984                 .pos = *ppos,
 985                 .u.file = out,
 986         };
 987
 988         pipe_lock(pipe);
 989         ret = __splice_from_pipe(pipe, &sd, actor);
 990         pipe_unlock(pipe);
 991
 992         return ret;
 993 }
 994
 995 /**
 996  * generic_file_splice_write - splice data from a pipe to a file
 997  * @pipe:       pipe info
 998  * @out:        file to write to
 999  * @ppos:       position in @out
1000  * @len:        number of bytes to splice
1001  * @flags:      splice modifier flags
1002  *
1003  * Description:
1004  *    Will either move or copy pages (determined by @flags options) from
1005  *    the given pipe inode to the given file.
1006  *
1007  */
1008 ssize_t
1009 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1010                           loff_t *ppos, size_t len, unsigned int flags)
1011 {
1012         struct address_space *mapping = out->f_mapping;
1013         struct inode *inode = mapping->host;
1014         struct splice_desc sd = {
1015                 .total_len = len,
1016                 .flags = flags,
1017                 .pos = *ppos,
1018                 .u.file = out,
1019         };
1020         ssize_t ret;
1021
1022         pipe_lock(pipe);
1023
1024         splice_from_pipe_begin(&sd);
1025         do {
1026                 ret = splice_from_pipe_next(pipe, &sd);
1027                 if (ret <= 0)
1028                         break;
1029
1030                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1031                 ret = file_remove_suid(out);
1032                 if (!ret) {
1033                         ret = file_update_time(out);
1034                         if (!ret)
1035                                 ret = splice_from_pipe_feed(pipe, &sd,
1036                                                             pipe_to_file);
1037                 }
1038                 mutex_unlock(&inode->i_mutex);
1039         } while (ret > 0);
1040         splice_from_pipe_end(pipe, &sd);
1041
1042         pipe_unlock(pipe);
1043
1044         if (sd.num_spliced)
1045                 ret = sd.num_spliced;
1046
1047         if (ret > 0) {
1048                 int err;
1049
1050                 err = generic_write_sync(out, *ppos, ret);
1051                 if (err)
1052                         ret = err;
1053                 else
1054                         *ppos += ret;
1055                 balance_dirty_pages_ratelimited(mapping);
1056         }
1057
1058         return ret;
1059 }
1060
1061 EXPORT_SYMBOL(generic_file_splice_write);
1062
1063 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1064                           struct splice_desc *sd)
1065 {
1066         int ret;
1067         void *data;
1068         loff_t tmp = sd->pos;
1069
1070         data = buf->ops->map(pipe, buf, 0);
1071         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1072         buf->ops->unmap(pipe, buf, data);
1073
1074         return ret;
1075 }
1076
1077 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1078                                          struct file *out, loff_t *ppos,
1079                                          size_t len, unsigned int flags)
1080 {
1081         ssize_t ret;
1082
1083         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1084         if (ret > 0)
1085                 *ppos += ret;
1086
1087         return ret;
1088 }
1089
1090 /**
1091  * generic_splice_sendpage - splice data from a pipe to a socket
1092  * @pipe:       pipe to splice from
1093  * @out:        socket to write to
1094  * @ppos:       position in @out
1095  * @len:        number of bytes to splice
1096  * @flags:      splice modifier flags
1097  *
1098  * Description:
1099  *    Will send @len bytes from the pipe to a network socket. No data copying
1100  *    is involved.
1101  *
1102  */
1103 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1104                                 loff_t *ppos, size_t len, unsigned int flags)
1105 {
1106         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1107 }
1108
1109 EXPORT_SYMBOL(generic_splice_sendpage);
1110
1111 /*
1112  * Attempt to initiate a splice from pipe to file.
1113  */
1114 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1115                            loff_t *ppos, size_t len, unsigned int flags)
1116 {
1117         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1118                                 loff_t *, size_t, unsigned int);
1119         int ret;
1120
1121         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1122                 return -EBADF;
1123
1124         if (unlikely(out->f_flags & O_APPEND))
1125                 return -EINVAL;
1126
1127         ret = rw_verify_area(WRITE, out, ppos, len);
1128         if (unlikely(ret < 0))
1129                 return ret;
1130
1131         if (out->f_op && out->f_op->splice_write)
1132                 splice_write = out->f_op->splice_write;
1133         else
1134                 splice_write = default_file_splice_write;
1135
1136         file_start_write(out);
1137         ret = splice_write(pipe, out, ppos, len, flags);
1138         file_end_write(out);
1139         return ret;
1140 }
1141
1142 /*
1143  * Attempt to initiate a splice from a file to a pipe.
1144  */
1145 static long do_splice_to(struct file *in, loff_t *ppos,
1146                          struct pipe_inode_info *pipe, size_t len,
1147                          unsigned int flags)
1148 {
1149         ssize_t (*splice_read)(struct file *, loff_t *,
1150                                struct pipe_inode_info *, size_t, unsigned int);
1151         int ret;
1152
1153         if (unlikely(!(in->f_mode & FMODE_READ)))
1154                 return -EBADF;
1155
1156         ret = rw_verify_area(READ, in, ppos, len);
1157         if (unlikely(ret < 0))
1158                 return ret;
1159
1160         if (in->f_op && in->f_op->splice_read)
1161                 splice_read = in->f_op->splice_read;
1162         else
1163                 splice_read = default_file_splice_read;
1164
1165         return splice_read(in, ppos, pipe, len, flags);
1166 }
1167
1168 /**
1169  * splice_direct_to_actor - splices data directly between two non-pipes
1170  * @in:         file to splice from
1171  * @sd:         actor information on where to splice to
1172  * @actor:      handles the data splicing
1173  *
1174  * Description:
1175  *    This is a special case helper to splice directly between two
1176  *    points, without requiring an explicit pipe. Internally an allocated
1177  *    pipe is cached in the process, and reused during the lifetime of
1178  *    that process.
1179  *
1180  */
1181 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1182                                splice_direct_actor *actor)
1183 {
1184         struct pipe_inode_info *pipe;
1185         long ret, bytes;
1186         umode_t i_mode;
1187         size_t len;
1188         int i, flags;
1189
1190         /*
1191          * We require the input being a regular file, as we don't want to
1192          * randomly drop data for eg socket -> socket splicing. Use the
1193          * piped splicing for that!
1194          */
1195         i_mode = file_inode(in)->i_mode;
1196         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1197                 return -EINVAL;
1198
1199         /*
1200          * neither in nor out is a pipe, setup an internal pipe attached to
1201          * 'out' and transfer the wanted data from 'in' to 'out' through that
1202          */
1203         pipe = current->splice_pipe;
1204         if (unlikely(!pipe)) {
1205                 pipe = alloc_pipe_info();
1206                 if (!pipe)
1207                         return -ENOMEM;
1208
1209                 /*
1210                  * We don't have an immediate reader, but we'll read the stuff
1211                  * out of the pipe right after the splice_to_pipe(). So set
1212                  * PIPE_READERS appropriately.
1213                  */
1214                 pipe->readers = 1;
1215
1216                 current->splice_pipe = pipe;
1217         }
1218
1219         /*
1220          * Do the splice.
1221          */
1222         ret = 0;
1223         bytes = 0;
1224         len = sd->total_len;
1225         flags = sd->flags;
1226
1227         /*
1228          * Don't block on output, we have to drain the direct pipe.
1229          */
1230         sd->flags &= ~SPLICE_F_NONBLOCK;
1231
1232         while (len) {
1233                 size_t read_len;
1234                 loff_t pos = sd->pos, prev_pos = pos;
1235
1236                 ret = do_splice_to(in, &pos, pipe, len, flags);
1237                 if (unlikely(ret <= 0))
1238                         goto out_release;
1239
1240                 read_len = ret;
1241                 sd->total_len = read_len;
1242
1243                 /*
1244                  * NOTE: nonblocking mode only applies to the input. We
1245                  * must not do the output in nonblocking mode as then we
1246                  * could get stuck data in the internal pipe:
1247                  */
1248                 ret = actor(pipe, sd);
1249                 if (unlikely(ret <= 0)) {
1250                         sd->pos = prev_pos;
1251                         goto out_release;
1252                 }
1253
1254                 bytes += ret;
1255                 len -= ret;
1256                 sd->pos = pos;
1257
1258                 if (ret < read_len) {
1259                         sd->pos = prev_pos + ret;
1260                         goto out_release;
1261                 }
1262         }
1263
1264 done:
1265         pipe->nrbufs = pipe->curbuf = 0;
1266         file_accessed(in);
1267         return bytes;
1268
1269 out_release:
1270         /*
1271          * If we did an incomplete transfer we must release
1272          * the pipe buffers in question:
1273          */
1274         for (i = 0; i < pipe->buffers; i++) {
1275                 struct pipe_buffer *buf = pipe->bufs + i;
1276
1277                 if (buf->ops) {
1278                         buf->ops->release(pipe, buf);
1279                         buf->ops = NULL;
1280                 }
1281         }
1282
1283         if (!bytes)
1284                 bytes = ret;
1285
1286         goto done;
1287 }
1288 EXPORT_SYMBOL(splice_direct_to_actor);
1289
1290 static int direct_splice_actor(struct pipe_inode_info *pipe,
1291                                struct splice_desc *sd)
1292 {
1293         struct file *file = sd->u.file;
1294
1295         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1296                               sd->flags);
1297 }
1298
1299 /**
1300  * do_splice_direct - splices data directly between two files
1301  * @in:         file to splice from
1302  * @ppos:       input file offset
1303  * @out:        file to splice to
1304  * @opos:       output file offset
1305  * @len:        number of bytes to splice
1306  * @flags:      splice modifier flags
1307  *
1308  * Description:
1309  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1310  *    doing it in the application would incur an extra system call
1311  *    (splice in + splice out, as compared to just sendfile()). So this helper
1312  *    can splice directly through a process-private pipe.
1313  *
1314  */
1315 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1316                       loff_t *opos, size_t len, unsigned int flags)
1317 {
1318         struct splice_desc sd = {
1319                 .len            = len,
1320                 .total_len      = len,
1321                 .flags          = flags,
1322                 .pos            = *ppos,
1323                 .u.file         = out,
1324                 .opos           = opos,
1325         };
1326         long ret;
1327
1328         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1329         if (ret > 0)
1330                 *ppos = sd.pos;
1331
1332         return ret;
1333 }
1334
1335 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1336                                struct pipe_inode_info *opipe,
1337                                size_t len, unsigned int flags);
1338
1339 /*
1340  * Determine where to splice to/from.
1341  */
1342 static long do_splice(struct file *in, loff_t __user *off_in,
1343                       struct file *out, loff_t __user *off_out,
1344                       size_t len, unsigned int flags)
1345 {
1346         struct pipe_inode_info *ipipe;
1347         struct pipe_inode_info *opipe;
1348         loff_t offset;
1349         long ret;
1350
1351         ipipe = get_pipe_info(in);
1352         opipe = get_pipe_info(out);
1353
1354         if (ipipe && opipe) {
1355                 if (off_in || off_out)
1356                         return -ESPIPE;
1357
1358                 if (!(in->f_mode & FMODE_READ))
1359                         return -EBADF;
1360
1361                 if (!(out->f_mode & FMODE_WRITE))
1362                         return -EBADF;
1363
1364                 /* Splicing to self would be fun, but... */
1365                 if (ipipe == opipe)
1366                         return -EINVAL;
1367
1368                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1369         }
1370
1371         if (ipipe) {
1372                 if (off_in)
1373                         return -ESPIPE;
1374                 if (off_out) {
1375                         if (!(out->f_mode & FMODE_PWRITE))
1376                                 return -EINVAL;
1377                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1378                                 return -EFAULT;
1379                 } else {
1380                         offset = out->f_pos;
1381                 }
1382
1383                 ret = do_splice_from(ipipe, out, &offset, len, flags);
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         char *src;
1524         int ret;
1525
1526         /*
1527          * See if we can use the atomic maps, by prefaulting in the
1528          * pages and doing an atomic copy
1529          */
1530         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1531                 src = buf->ops->map(pipe, buf, 1);
1532                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1533                                                         sd->len);
1534                 buf->ops->unmap(pipe, buf, src);
1535                 if (!ret) {
1536                         ret = sd->len;
1537                         goto out;
1538                 }
1539         }
1540
1541         /*
1542          * No dice, use slow non-atomic map and copy
1543          */
1544         src = buf->ops->map(pipe, buf, 0);
1545
1546         ret = sd->len;
1547         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1548                 ret = -EFAULT;
1549
1550         buf->ops->unmap(pipe, buf, src);
1551 out:
1552         if (ret > 0)
1553                 sd->u.userptr += ret;
1554         return ret;
1555 }
1556
1557 /*
1558  * For lack of a better implementation, implement vmsplice() to userspace
1559  * as a simple copy of the pipes pages to the user iov.
1560  */
1561 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1562                              unsigned long nr_segs, unsigned int flags)
1563 {
1564         struct pipe_inode_info *pipe;
1565         struct splice_desc sd;
1566         ssize_t size;
1567         int error;
1568         long ret;
1569
1570         pipe = get_pipe_info(file);
1571         if (!pipe)
1572                 return -EBADF;
1573
1574         pipe_lock(pipe);
1575
1576         error = ret = 0;
1577         while (nr_segs) {
1578                 void __user *base;
1579                 size_t len;
1580
1581                 /*
1582                  * Get user address base and length for this iovec.
1583                  */
1584                 error = get_user(base, &iov->iov_base);
1585                 if (unlikely(error))
1586                         break;
1587                 error = get_user(len, &iov->iov_len);
1588                 if (unlikely(error))
1589                         break;
1590
1591                 /*
1592                  * Sanity check this iovec. 0 read succeeds.
1593                  */
1594                 if (unlikely(!len))
1595                         break;
1596                 if (unlikely(!base)) {
1597                         error = -EFAULT;
1598                         break;
1599                 }
1600
1601                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1602                         error = -EFAULT;
1603                         break;
1604                 }
1605
1606                 sd.len = 0;
1607                 sd.total_len = len;
1608                 sd.flags = flags;
1609                 sd.u.userptr = base;
1610                 sd.pos = 0;
1611
1612                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1613                 if (size < 0) {
1614                         if (!ret)
1615                                 ret = size;
1616
1617                         break;
1618                 }
1619
1620                 ret += size;
1621
1622                 if (size < len)
1623                         break;
1624
1625                 nr_segs--;
1626                 iov++;
1627         }
1628
1629         pipe_unlock(pipe);
1630
1631         if (!ret)
1632                 ret = error;
1633
1634         return ret;
1635 }
1636
1637 /*
1638  * vmsplice splices a user address range into a pipe. It can be thought of
1639  * as splice-from-memory, where the regular splice is splice-from-file (or
1640  * to file). In both cases the output is a pipe, naturally.
1641  */
1642 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1643                              unsigned long nr_segs, unsigned int flags)
1644 {
1645         struct pipe_inode_info *pipe;
1646         struct page *pages[PIPE_DEF_BUFFERS];
1647         struct partial_page partial[PIPE_DEF_BUFFERS];
1648         struct splice_pipe_desc spd = {
1649                 .pages = pages,
1650                 .partial = partial,
1651                 .nr_pages_max = PIPE_DEF_BUFFERS,
1652                 .flags = flags,
1653                 .ops = &user_page_pipe_buf_ops,
1654                 .spd_release = spd_release_page,
1655         };
1656         long ret;
1657
1658         pipe = get_pipe_info(file);
1659         if (!pipe)
1660                 return -EBADF;
1661
1662         if (splice_grow_spd(pipe, &spd))
1663                 return -ENOMEM;
1664
1665         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1666                                             spd.partial, false,
1667                                             spd.nr_pages_max);
1668         if (spd.nr_pages <= 0)
1669                 ret = spd.nr_pages;
1670         else
1671                 ret = splice_to_pipe(pipe, &spd);
1672
1673         splice_shrink_spd(&spd);
1674         return ret;
1675 }
1676
1677 /*
1678  * Note that vmsplice only really supports true splicing _from_ user memory
1679  * to a pipe, not the other way around. Splicing from user memory is a simple
1680  * operation that can be supported without any funky alignment restrictions
1681  * or nasty vm tricks. We simply map in the user memory and fill them into
1682  * a pipe. The reverse isn't quite as easy, though. There are two possible
1683  * solutions for that:
1684  *
1685  *      - memcpy() the data internally, at which point we might as well just
1686  *        do a regular read() on the buffer anyway.
1687  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1688  *        has restriction limitations on both ends of the pipe).
1689  *
1690  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1691  *
1692  */
1693 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1694                 unsigned long, nr_segs, unsigned int, flags)
1695 {
1696         struct fd f;
1697         long error;
1698
1699         if (unlikely(nr_segs > UIO_MAXIOV))
1700                 return -EINVAL;
1701         else if (unlikely(!nr_segs))
1702                 return 0;
1703
1704         error = -EBADF;
1705         f = fdget(fd);
1706         if (f.file) {
1707                 if (f.file->f_mode & FMODE_WRITE)
1708                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1709                 else if (f.file->f_mode & FMODE_READ)
1710                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1711
1712                 fdput(f);
1713         }
1714
1715         return error;
1716 }
1717
1718 #ifdef CONFIG_COMPAT
1719 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1720                     unsigned int, nr_segs, unsigned int, flags)
1721 {
1722         unsigned i;
1723         struct iovec __user *iov;
1724         if (nr_segs > UIO_MAXIOV)
1725                 return -EINVAL;
1726         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1727         for (i = 0; i < nr_segs; i++) {
1728                 struct compat_iovec v;
1729                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1730                     get_user(v.iov_len, &iov32[i].iov_len) ||
1731                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1732                     put_user(v.iov_len, &iov[i].iov_len))
1733                         return -EFAULT;
1734         }
1735         return sys_vmsplice(fd, iov, nr_segs, flags);
1736 }
1737 #endif
1738
1739 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1740                 int, fd_out, loff_t __user *, off_out,
1741                 size_t, len, unsigned int, flags)
1742 {
1743         struct fd in, out;
1744         long error;
1745
1746         if (unlikely(!len))
1747                 return 0;
1748
1749         error = -EBADF;
1750         in = fdget(fd_in);
1751         if (in.file) {
1752                 if (in.file->f_mode & FMODE_READ) {
1753                         out = fdget(fd_out);
1754                         if (out.file) {
1755                                 if (out.file->f_mode & FMODE_WRITE)
1756                                         error = do_splice(in.file, off_in,
1757                                                           out.file, off_out,
1758                                                           len, flags);
1759                                 fdput(out);
1760                         }
1761                 }
1762                 fdput(in);
1763         }
1764         return error;
1765 }
1766
1767 /*
1768  * Make sure there's data to read. Wait for input if we can, otherwise
1769  * return an appropriate error.
1770  */
1771 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1772 {
1773         int ret;
1774
1775         /*
1776          * Check ->nrbufs without the inode lock first. This function
1777          * is speculative anyways, so missing one is ok.
1778          */
1779         if (pipe->nrbufs)
1780                 return 0;
1781
1782         ret = 0;
1783         pipe_lock(pipe);
1784
1785         while (!pipe->nrbufs) {
1786                 if (signal_pending(current)) {
1787                         ret = -ERESTARTSYS;
1788                         break;
1789                 }
1790                 if (!pipe->writers)
1791                         break;
1792                 if (!pipe->waiting_writers) {
1793                         if (flags & SPLICE_F_NONBLOCK) {
1794                                 ret = -EAGAIN;
1795                                 break;
1796                         }
1797                 }
1798                 pipe_wait(pipe);
1799         }
1800
1801         pipe_unlock(pipe);
1802         return ret;
1803 }
1804
1805 /*
1806  * Make sure there's writeable room. Wait for room if we can, otherwise
1807  * return an appropriate error.
1808  */
1809 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1810 {
1811         int ret;
1812
1813         /*
1814          * Check ->nrbufs without the inode lock first. This function
1815          * is speculative anyways, so missing one is ok.
1816          */
1817         if (pipe->nrbufs < pipe->buffers)
1818                 return 0;
1819
1820         ret = 0;
1821         pipe_lock(pipe);
1822
1823         while (pipe->nrbufs >= pipe->buffers) {
1824                 if (!pipe->readers) {
1825                         send_sig(SIGPIPE, current, 0);
1826                         ret = -EPIPE;
1827                         break;
1828                 }
1829                 if (flags & SPLICE_F_NONBLOCK) {
1830                         ret = -EAGAIN;
1831                         break;
1832                 }
1833                 if (signal_pending(current)) {
1834                         ret = -ERESTARTSYS;
1835                         break;
1836                 }
1837                 pipe->waiting_writers++;
1838                 pipe_wait(pipe);
1839                 pipe->waiting_writers--;
1840         }
1841
1842         pipe_unlock(pipe);
1843         return ret;
1844 }
1845
1846 /*
1847  * Splice contents of ipipe to opipe.
1848  */
1849 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1850                                struct pipe_inode_info *opipe,
1851                                size_t len, unsigned int flags)
1852 {
1853         struct pipe_buffer *ibuf, *obuf;
1854         int ret = 0, nbuf;
1855         bool input_wakeup = false;
1856
1857
1858 retry:
1859         ret = ipipe_prep(ipipe, flags);
1860         if (ret)
1861                 return ret;
1862
1863         ret = opipe_prep(opipe, flags);
1864         if (ret)
1865                 return ret;
1866
1867         /*
1868          * Potential ABBA deadlock, work around it by ordering lock
1869          * grabbing by pipe info address. Otherwise two different processes
1870          * could deadlock (one doing tee from A -> B, the other from B -> A).
1871          */
1872         pipe_double_lock(ipipe, opipe);
1873
1874         do {
1875                 if (!opipe->readers) {
1876                         send_sig(SIGPIPE, current, 0);
1877                         if (!ret)
1878                                 ret = -EPIPE;
1879                         break;
1880                 }
1881
1882                 if (!ipipe->nrbufs && !ipipe->writers)
1883                         break;
1884
1885                 /*
1886                  * Cannot make any progress, because either the input
1887                  * pipe is empty or the output pipe is full.
1888                  */
1889                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1890                         /* Already processed some buffers, break */
1891                         if (ret)
1892                                 break;
1893
1894                         if (flags & SPLICE_F_NONBLOCK) {
1895                                 ret = -EAGAIN;
1896                                 break;
1897                         }
1898
1899                         /*
1900                          * We raced with another reader/writer and haven't
1901                          * managed to process any buffers.  A zero return
1902                          * value means EOF, so retry instead.
1903                          */
1904                         pipe_unlock(ipipe);
1905                         pipe_unlock(opipe);
1906                         goto retry;
1907                 }
1908
1909                 ibuf = ipipe->bufs + ipipe->curbuf;
1910                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1911                 obuf = opipe->bufs + nbuf;
1912
1913                 if (len >= ibuf->len) {
1914                         /*
1915                          * Simply move the whole buffer from ipipe to opipe
1916                          */
1917                         *obuf = *ibuf;
1918                         ibuf->ops = NULL;
1919                         opipe->nrbufs++;
1920                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1921                         ipipe->nrbufs--;
1922                         input_wakeup = true;
1923                 } else {
1924                         /*
1925                          * Get a reference to this pipe buffer,
1926                          * so we can copy the contents over.
1927                          */
1928                         ibuf->ops->get(ipipe, ibuf);
1929                         *obuf = *ibuf;
1930
1931                         /*
1932                          * Don't inherit the gift flag, we need to
1933                          * prevent multiple steals of this page.
1934                          */
1935                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1936
1937                         obuf->len = len;
1938                         opipe->nrbufs++;
1939                         ibuf->offset += obuf->len;
1940                         ibuf->len -= obuf->len;
1941                 }
1942                 ret += obuf->len;
1943                 len -= obuf->len;
1944         } while (len);
1945
1946         pipe_unlock(ipipe);
1947         pipe_unlock(opipe);
1948
1949         /*
1950          * If we put data in the output pipe, wakeup any potential readers.
1951          */
1952         if (ret > 0)
1953                 wakeup_pipe_readers(opipe);
1954
1955         if (input_wakeup)
1956                 wakeup_pipe_writers(ipipe);
1957
1958         return ret;
1959 }
1960
1961 /*
1962  * Link contents of ipipe to opipe.
1963  */
1964 static int link_pipe(struct pipe_inode_info *ipipe,
1965                      struct pipe_inode_info *opipe,
1966                      size_t len, unsigned int flags)
1967 {
1968         struct pipe_buffer *ibuf, *obuf;
1969         int ret = 0, i = 0, nbuf;
1970
1971         /*
1972          * Potential ABBA deadlock, work around it by ordering lock
1973          * grabbing by pipe info address. Otherwise two different processes
1974          * could deadlock (one doing tee from A -> B, the other from B -> A).
1975          */
1976         pipe_double_lock(ipipe, opipe);
1977
1978         do {
1979                 if (!opipe->readers) {
1980                         send_sig(SIGPIPE, current, 0);
1981                         if (!ret)
1982                                 ret = -EPIPE;
1983                         break;
1984                 }
1985
1986                 /*
1987                  * If we have iterated all input buffers or ran out of
1988                  * output room, break.
1989                  */
1990                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1991                         break;
1992
1993                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1994                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1995
1996                 /*
1997                  * Get a reference to this pipe buffer,
1998                  * so we can copy the contents over.
1999                  */
2000                 ibuf->ops->get(ipipe, ibuf);
2001
2002                 obuf = opipe->bufs + nbuf;
2003                 *obuf = *ibuf;
2004
2005                 /*
2006                  * Don't inherit the gift flag, we need to
2007                  * prevent multiple steals of this page.
2008                  */
2009                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2010
2011                 if (obuf->len > len)
2012                         obuf->len = len;
2013
2014                 opipe->nrbufs++;
2015                 ret += obuf->len;
2016                 len -= obuf->len;
2017                 i++;
2018         } while (len);
2019
2020         /*
2021          * return EAGAIN if we have the potential of some data in the
2022          * future, otherwise just return 0
2023          */
2024         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2025                 ret = -EAGAIN;
2026
2027         pipe_unlock(ipipe);
2028         pipe_unlock(opipe);
2029
2030         /*
2031          * If we put data in the output pipe, wakeup any potential readers.
2032          */
2033         if (ret > 0)
2034                 wakeup_pipe_readers(opipe);
2035
2036         return ret;
2037 }
2038
2039 /*
2040  * This is a tee(1) implementation that works on pipes. It doesn't copy
2041  * any data, it simply references the 'in' pages on the 'out' pipe.
2042  * The 'flags' used are the SPLICE_F_* variants, currently the only
2043  * applicable one is SPLICE_F_NONBLOCK.
2044  */
2045 static long do_tee(struct file *in, struct file *out, size_t len,
2046                    unsigned int flags)
2047 {
2048         struct pipe_inode_info *ipipe = get_pipe_info(in);
2049         struct pipe_inode_info *opipe = get_pipe_info(out);
2050         int ret = -EINVAL;
2051
2052         /*
2053          * Duplicate the contents of ipipe to opipe without actually
2054          * copying the data.
2055          */
2056         if (ipipe && opipe && ipipe != opipe) {
2057                 /*
2058                  * Keep going, unless we encounter an error. The ipipe/opipe
2059                  * ordering doesn't really matter.
2060                  */
2061                 ret = ipipe_prep(ipipe, flags);
2062                 if (!ret) {
2063                         ret = opipe_prep(opipe, flags);
2064                         if (!ret)
2065                                 ret = link_pipe(ipipe, opipe, len, flags);
2066                 }
2067         }
2068
2069         return ret;
2070 }
2071
2072 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2073 {
2074         struct fd in;
2075         int error;
2076
2077         if (unlikely(!len))
2078                 return 0;
2079
2080         error = -EBADF;
2081         in = fdget(fdin);
2082         if (in.file) {
2083                 if (in.file->f_mode & FMODE_READ) {
2084                         struct fd out = fdget(fdout);
2085                         if (out.file) {
2086                                 if (out.file->f_mode & FMODE_WRITE)
2087                                         error = do_tee(in.file, out.file,
2088                                                         len, flags);
2089                                 fdput(out);
2090                         }
2091                 }
2092                 fdput(in);
2093         }
2094
2095         return error;
2096 }