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 <linux/aio.h>
  36 #include "internal.h"
  37
  38 /*
  39  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  40  * a vm helper function, it's already simplified quite a bit by the
  41  * addition of remove_mapping(). If success is returned, the caller may
  42  * attempt to reuse this page for another destination.
  43  */
  44 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  45                                      struct pipe_buffer *buf)
  46 {
  47         struct page *page = buf->page;
  48         struct address_space *mapping;
  49
  50         lock_page(page);
  51
  52         mapping = page_mapping(page);
  53         if (mapping) {
  54                 WARN_ON(!PageUptodate(page));
  55
  56                 /*
  57                  * At least for ext2 with nobh option, we need to wait on
  58                  * writeback completing on this page, since we'll remove it
  59                  * from the pagecache.  Otherwise truncate wont wait on the
  60                  * page, allowing the disk blocks to be reused by someone else
  61                  * before we actually wrote our data to them. fs corruption
  62                  * ensues.
  63                  */
  64                 wait_on_page_writeback(page);
  65
  66                 if (page_has_private(page) &&
  67                     !try_to_release_page(page, GFP_KERNEL))
  68                         goto out_unlock;
  69
  70                 /*
  71                  * If we succeeded in removing the mapping, set LRU flag
  72                  * and return good.
  73                  */
  74                 if (remove_mapping(mapping, page)) {
  75                         buf->flags |= PIPE_BUF_FLAG_LRU;
  76                         return 0;
  77                 }
  78         }
  79
  80         /*
  81          * Raced with truncate or failed to remove page from current
  82          * address space, unlock and return failure.
  83          */
  84 out_unlock:
  85         unlock_page(page);
  86         return 1;
  87 }
  88
  89 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  90                                         struct pipe_buffer *buf)
  91 {
  92         page_cache_release(buf->page);
  93         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  94 }
  95
  96 /*
  97  * Check whether the contents of buf is OK to access. Since the content
  98  * is a page cache page, IO may be in flight.
  99  */
 100 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 101                                        struct pipe_buffer *buf)
 102 {
 103         struct page *page = buf->page;
 104         int err;
 105
 106         if (!PageUptodate(page)) {
 107                 lock_page(page);
 108
 109                 /*
 110                  * Page got truncated/unhashed. This will cause a 0-byte
 111                  * splice, if this is the first page.
 112                  */
 113                 if (!page->mapping) {
 114                         err = -ENODATA;
 115                         goto error;
 116                 }
 117
 118                 /*
 119                  * Uh oh, read-error from disk.
 120                  */
 121                 if (!PageUptodate(page)) {
 122                         err = -EIO;
 123                         goto error;
 124                 }
 125
 126                 /*
 127                  * Page is ok afterall, we are done.
 128                  */
 129                 unlock_page(page);
 130         }
 131
 132         return 0;
 133 error:
 134         unlock_page(page);
 135         return err;
 136 }
 137
 138 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 139         .can_merge = 0,
 140         .confirm = page_cache_pipe_buf_confirm,
 141         .release = page_cache_pipe_buf_release,
 142         .steal = page_cache_pipe_buf_steal,
 143         .get = generic_pipe_buf_get,
 144 };
 145
 146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 147                                     struct pipe_buffer *buf)
 148 {
 149         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 150                 return 1;
 151
 152         buf->flags |= PIPE_BUF_FLAG_LRU;
 153         return generic_pipe_buf_steal(pipe, buf);
 154 }
 155
 156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 157         .can_merge = 0,
 158         .confirm = generic_pipe_buf_confirm,
 159         .release = page_cache_pipe_buf_release,
 160         .steal = user_page_pipe_buf_steal,
 161         .get = generic_pipe_buf_get,
 162 };
 163
 164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 165 {
 166         smp_mb();
 167         if (waitqueue_active(&pipe->wait))
 168                 wake_up_interruptible(&pipe->wait);
 169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 170 }
 171
 172 /**
 173  * splice_to_pipe - fill passed data into a pipe
 174  * @pipe:       pipe to fill
 175  * @spd:        data to fill
 176  *
 177  * Description:
 178  *    @spd contains a map of pages and len/offset tuples, along with
 179  *    the struct pipe_buf_operations associated with these pages. This
 180  *    function will link that data to the pipe.
 181  *
 182  */
 183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 184                        struct splice_pipe_desc *spd)
 185 {
 186         unsigned int spd_pages = spd->nr_pages;
 187         int ret, do_wakeup, page_nr;
 188
 189         ret = 0;
 190         do_wakeup = 0;
 191         page_nr = 0;
 192
 193         pipe_lock(pipe);
 194
 195         for (;;) {
 196                 if (!pipe->readers) {
 197                         send_sig(SIGPIPE, current, 0);
 198                         if (!ret)
 199                                 ret = -EPIPE;
 200                         break;
 201                 }
 202
 203                 if (pipe->nrbufs < pipe->buffers) {
 204                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 205                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 206
 207                         buf->page = spd->pages[page_nr];
 208                         buf->offset = spd->partial[page_nr].offset;
 209                         buf->len = spd->partial[page_nr].len;
 210                         buf->private = spd->partial[page_nr].private;
 211                         buf->ops = spd->ops;
 212                         if (spd->flags & SPLICE_F_GIFT)
 213                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 214
 215                         pipe->nrbufs++;
 216                         page_nr++;
 217                         ret += buf->len;
 218
 219                         if (pipe->files)
 220                                 do_wakeup = 1;
 221
 222                         if (!--spd->nr_pages)
 223                                 break;
 224                         if (pipe->nrbufs < pipe->buffers)
 225                                 continue;
 226
 227                         break;
 228                 }
 229
 230                 if (spd->flags & SPLICE_F_NONBLOCK) {
 231                         if (!ret)
 232                                 ret = -EAGAIN;
 233                         break;
 234                 }
 235
 236                 if (signal_pending(current)) {
 237                         if (!ret)
 238                                 ret = -ERESTARTSYS;
 239                         break;
 240                 }
 241
 242                 if (do_wakeup) {
 243                         smp_mb();
 244                         if (waitqueue_active(&pipe->wait))
 245                                 wake_up_interruptible_sync(&pipe->wait);
 246                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 247                         do_wakeup = 0;
 248                 }
 249
 250                 pipe->waiting_writers++;
 251                 pipe_wait(pipe);
 252                 pipe->waiting_writers--;
 253         }
 254
 255         pipe_unlock(pipe);
 256
 257         if (do_wakeup)
 258                 wakeup_pipe_readers(pipe);
 259
 260         while (page_nr < spd_pages)
 261                 spd->spd_release(spd, page_nr++);
 262
 263         return ret;
 264 }
 265
 266 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 267 {
 268         page_cache_release(spd->pages[i]);
 269 }
 270
 271 /*
 272  * Check if we need to grow the arrays holding pages and partial page
 273  * descriptions.
 274  */
 275 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 276 {
 277         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 278
 279         spd->nr_pages_max = buffers;
 280         if (buffers <= PIPE_DEF_BUFFERS)
 281                 return 0;
 282
 283         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 284         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 285
 286         if (spd->pages && spd->partial)
 287                 return 0;
 288
 289         kfree(spd->pages);
 290         kfree(spd->partial);
 291         return -ENOMEM;
 292 }
 293
 294 void splice_shrink_spd(struct splice_pipe_desc *spd)
 295 {
 296         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 297                 return;
 298
 299         kfree(spd->pages);
 300         kfree(spd->partial);
 301 }
 302
 303 static int
 304 __generic_file_splice_read(struct file *in, loff_t *ppos,
 305                            struct pipe_inode_info *pipe, size_t len,
 306                            unsigned int flags)
 307 {
 308         struct address_space *mapping = in->f_mapping;
 309         unsigned int loff, nr_pages, req_pages;
 310         struct page *pages[PIPE_DEF_BUFFERS];
 311         struct partial_page partial[PIPE_DEF_BUFFERS];
 312         struct page *page;
 313         pgoff_t index, end_index;
 314         loff_t isize;
 315         int error, page_nr;
 316         struct splice_pipe_desc spd = {
 317                 .pages = pages,
 318                 .partial = partial,
 319                 .nr_pages_max = PIPE_DEF_BUFFERS,
 320                 .flags = flags,
 321                 .ops = &page_cache_pipe_buf_ops,
 322                 .spd_release = spd_release_page,
 323         };
 324
 325         if (splice_grow_spd(pipe, &spd))
 326                 return -ENOMEM;
 327
 328         index = *ppos >> PAGE_CACHE_SHIFT;
 329         loff = *ppos & ~PAGE_CACHE_MASK;
 330         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 331         nr_pages = min(req_pages, spd.nr_pages_max);
 332
 333         /*
 334          * Lookup the (hopefully) full range of pages we need.
 335          */
 336         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 337         index += spd.nr_pages;
 338
 339         /*
 340          * If find_get_pages_contig() returned fewer pages than we needed,
 341          * readahead/allocate the rest and fill in the holes.
 342          */
 343         if (spd.nr_pages < nr_pages)
 344                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 345                                 index, req_pages - spd.nr_pages);
 346
 347         error = 0;
 348         while (spd.nr_pages < nr_pages) {
 349                 /*
 350                  * Page could be there, find_get_pages_contig() breaks on
 351                  * the first hole.
 352                  */
 353                 page = find_get_page(mapping, index);
 354                 if (!page) {
 355                         /*
 356                          * page didn't exist, allocate one.
 357                          */
 358                         page = page_cache_alloc_cold(mapping);
 359                         if (!page)
 360                                 break;
 361
 362                         error = add_to_page_cache_lru(page, mapping, index,
 363                                                 GFP_KERNEL);
 364                         if (unlikely(error)) {
 365                                 page_cache_release(page);
 366                                 if (error == -EEXIST)
 367                                         continue;
 368                                 break;
 369                         }
 370                         /*
 371                          * add_to_page_cache() locks the page, unlock it
 372                          * to avoid convoluting the logic below even more.
 373                          */
 374                         unlock_page(page);
 375                 }
 376
 377                 spd.pages[spd.nr_pages++] = page;
 378                 index++;
 379         }
 380
 381         /*
 382          * Now loop over the map and see if we need to start IO on any
 383          * pages, fill in the partial map, etc.
 384          */
 385         index = *ppos >> PAGE_CACHE_SHIFT;
 386         nr_pages = spd.nr_pages;
 387         spd.nr_pages = 0;
 388         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 389                 unsigned int this_len;
 390
 391                 if (!len)
 392                         break;
 393
 394                 /*
 395                  * this_len is the max we'll use from this page
 396                  */
 397                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 398                 page = spd.pages[page_nr];
 399
 400                 if (PageReadahead(page))
 401                         page_cache_async_readahead(mapping, &in->f_ra, in,
 402                                         page, index, req_pages - page_nr);
 403
 404                 /*
 405                  * If the page isn't uptodate, we may need to start io on it
 406                  */
 407                 if (!PageUptodate(page)) {
 408                         lock_page(page);
 409
 410                         /*
 411                          * Page was truncated, or invalidated by the
 412                          * filesystem.  Redo the find/create, but this time the
 413                          * page is kept locked, so there's no chance of another
 414                          * race with truncate/invalidate.
 415                          */
 416                         if (!page->mapping) {
 417                                 unlock_page(page);
 418                                 page = find_or_create_page(mapping, index,
 419                                                 mapping_gfp_mask(mapping));
 420
 421                                 if (!page) {
 422                                         error = -ENOMEM;
 423                                         break;
 424                                 }
 425                                 page_cache_release(spd.pages[page_nr]);
 426                                 spd.pages[page_nr] = page;
 427                         }
 428                         /*
 429                          * page was already under io and is now done, great
 430                          */
 431                         if (PageUptodate(page)) {
 432                                 unlock_page(page);
 433                                 goto fill_it;
 434                         }
 435
 436                         /*
 437                          * need to read in the page
 438                          */
 439                         error = mapping->a_ops->readpage(in, page);
 440                         if (unlikely(error)) {
 441                                 /*
 442                                  * We really should re-lookup the page here,
 443                                  * but it complicates things a lot. Instead
 444                                  * lets just do what we already stored, and
 445                                  * we'll get it the next time we are called.
 446                                  */
 447                                 if (error == AOP_TRUNCATED_PAGE)
 448                                         error = 0;
 449
 450                                 break;
 451                         }
 452                 }
 453 fill_it:
 454                 /*
 455                  * i_size must be checked after PageUptodate.
 456                  */
 457                 isize = i_size_read(mapping->host);
 458                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 459                 if (unlikely(!isize || index > end_index))
 460                         break;
 461
 462                 /*
 463                  * if this is the last page, see if we need to shrink
 464                  * the length and stop
 465                  */
 466                 if (end_index == index) {
 467                         unsigned int plen;
 468
 469                         /*
 470                          * max good bytes in this page
 471                          */
 472                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 473                         if (plen <= loff)
 474                                 break;
 475
 476                         /*
 477                          * force quit after adding this page
 478                          */
 479                         this_len = min(this_len, plen - loff);
 480                         len = this_len;
 481                 }
 482
 483                 spd.partial[page_nr].offset = loff;
 484                 spd.partial[page_nr].len = this_len;
 485                 len -= this_len;
 486                 loff = 0;
 487                 spd.nr_pages++;
 488                 index++;
 489         }
 490
 491         /*
 492          * Release any pages at the end, if we quit early. 'page_nr' is how far
 493          * we got, 'nr_pages' is how many pages are in the map.
 494          */
 495         while (page_nr < nr_pages)
 496                 page_cache_release(spd.pages[page_nr++]);
 497         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 498
 499         if (spd.nr_pages)
 500                 error = splice_to_pipe(pipe, &spd);
 501
 502         splice_shrink_spd(&spd);
 503         return error;
 504 }
 505
 506 /**
 507  * generic_file_splice_read - splice data from file to a pipe
 508  * @in:         file to splice from
 509  * @ppos:       position in @in
 510  * @pipe:       pipe to splice to
 511  * @len:        number of bytes to splice
 512  * @flags:      splice modifier flags
 513  *
 514  * Description:
 515  *    Will read pages from given file and fill them into a pipe. Can be
 516  *    used as long as the address_space operations for the source implements
 517  *    a readpage() hook.
 518  *
 519  */
 520 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 521                                  struct pipe_inode_info *pipe, size_t len,
 522                                  unsigned int flags)
 523 {
 524         loff_t isize, left;
 525         int ret;
 526
 527         isize = i_size_read(in->f_mapping->host);
 528         if (unlikely(*ppos >= isize))
 529                 return 0;
 530
 531         left = isize - *ppos;
 532         if (unlikely(left < len))
 533                 len = left;
 534
 535         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 536         if (ret > 0) {
 537                 *ppos += ret;
 538                 file_accessed(in);
 539         }
 540
 541         return ret;
 542 }
 543 EXPORT_SYMBOL(generic_file_splice_read);
 544
 545 static const struct pipe_buf_operations default_pipe_buf_ops = {
 546         .can_merge = 0,
 547         .confirm = generic_pipe_buf_confirm,
 548         .release = generic_pipe_buf_release,
 549         .steal = generic_pipe_buf_steal,
 550         .get = generic_pipe_buf_get,
 551 };
 552
 553 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 554                                     struct pipe_buffer *buf)
 555 {
 556         return 1;
 557 }
 558
 559 /* Pipe buffer operations for a socket and similar. */
 560 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 561         .can_merge = 0,
 562         .confirm = generic_pipe_buf_confirm,
 563         .release = generic_pipe_buf_release,
 564         .steal = generic_pipe_buf_nosteal,
 565         .get = generic_pipe_buf_get,
 566 };
 567 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 568
 569 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 570                             unsigned long vlen, loff_t offset)
 571 {
 572         mm_segment_t old_fs;
 573         loff_t pos = offset;
 574         ssize_t res;
 575
 576         old_fs = get_fs();
 577         set_fs(get_ds());
 578         /* The cast to a user pointer is valid due to the set_fs() */
 579         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 580         set_fs(old_fs);
 581
 582         return res;
 583 }
 584
 585 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 586                             loff_t pos)
 587 {
 588         mm_segment_t old_fs;
 589         ssize_t res;
 590
 591         old_fs = get_fs();
 592         set_fs(get_ds());
 593         /* The cast to a user pointer is valid due to the set_fs() */
 594         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 595         set_fs(old_fs);
 596
 597         return res;
 598 }
 599 EXPORT_SYMBOL(kernel_write);
 600
 601 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 602                                  struct pipe_inode_info *pipe, size_t len,
 603                                  unsigned int flags)
 604 {
 605         unsigned int nr_pages;
 606         unsigned int nr_freed;
 607         size_t offset;
 608         struct page *pages[PIPE_DEF_BUFFERS];
 609         struct partial_page partial[PIPE_DEF_BUFFERS];
 610         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 611         ssize_t res;
 612         size_t this_len;
 613         int error;
 614         int i;
 615         struct splice_pipe_desc spd = {
 616                 .pages = pages,
 617                 .partial = partial,
 618                 .nr_pages_max = PIPE_DEF_BUFFERS,
 619                 .flags = flags,
 620                 .ops = &default_pipe_buf_ops,
 621                 .spd_release = spd_release_page,
 622         };
 623
 624         if (splice_grow_spd(pipe, &spd))
 625                 return -ENOMEM;
 626
 627         res = -ENOMEM;
 628         vec = __vec;
 629         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 630                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 631                 if (!vec)
 632                         goto shrink_ret;
 633         }
 634
 635         offset = *ppos & ~PAGE_CACHE_MASK;
 636         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 637
 638         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 639                 struct page *page;
 640
 641                 page = alloc_page(GFP_USER);
 642                 error = -ENOMEM;
 643                 if (!page)
 644                         goto err;
 645
 646                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 647                 vec[i].iov_base = (void __user *) page_address(page);
 648                 vec[i].iov_len = this_len;
 649                 spd.pages[i] = page;
 650                 spd.nr_pages++;
 651                 len -= this_len;
 652                 offset = 0;
 653         }
 654
 655         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 656         if (res < 0) {
 657                 error = res;
 658                 goto err;
 659         }
 660
 661         error = 0;
 662         if (!res)
 663                 goto err;
 664
 665         nr_freed = 0;
 666         for (i = 0; i < spd.nr_pages; i++) {
 667                 this_len = min_t(size_t, vec[i].iov_len, res);
 668                 spd.partial[i].offset = 0;
 669                 spd.partial[i].len = this_len;
 670                 if (!this_len) {
 671                         __free_page(spd.pages[i]);
 672                         spd.pages[i] = NULL;
 673                         nr_freed++;
 674                 }
 675                 res -= this_len;
 676         }
 677         spd.nr_pages -= nr_freed;
 678
 679         res = splice_to_pipe(pipe, &spd);
 680         if (res > 0)
 681                 *ppos += res;
 682
 683 shrink_ret:
 684         if (vec != __vec)
 685                 kfree(vec);
 686         splice_shrink_spd(&spd);
 687         return res;
 688
 689 err:
 690         for (i = 0; i < spd.nr_pages; i++)
 691                 __free_page(spd.pages[i]);
 692
 693         res = error;
 694         goto shrink_ret;
 695 }
 696 EXPORT_SYMBOL(default_file_splice_read);
 697
 698 /*
 699  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 700  * using sendpage(). Return the number of bytes sent.
 701  */
 702 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 703                             struct pipe_buffer *buf, struct splice_desc *sd)
 704 {
 705         struct file *file = sd->u.file;
 706         loff_t pos = sd->pos;
 707         int more;
 708
 709         if (!likely(file->f_op->sendpage))
 710                 return -EINVAL;
 711
 712         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 713
 714         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 715                 more |= MSG_SENDPAGE_NOTLAST;
 716
 717         return file->f_op->sendpage(file, buf->page, buf->offset,
 718                                     sd->len, &pos, more);
 719 }
 720
 721 /*
 722  * This is a little more tricky than the file -> pipe splicing. There are
 723  * basically three cases:
 724  *
 725  *      - Destination page already exists in the address space and there
 726  *        are users of it. For that case we have no other option that
 727  *        copying the data. Tough luck.
 728  *      - Destination page already exists in the address space, but there
 729  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 730  *        through to last case.
 731  *      - Destination page does not exist, we can add the pipe page to
 732  *        the page cache and avoid the copy.
 733  *
 734  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 735  * sd->flags), we attempt to migrate pages from the pipe to the output
 736  * file address space page cache. This is possible if no one else has
 737  * the pipe page referenced outside of the pipe and page cache. If
 738  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 739  * a new page in the output file page cache and fill/dirty that.
 740  */
 741 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 742                  struct splice_desc *sd)
 743 {
 744         struct file *file = sd->u.file;
 745         struct address_space *mapping = file->f_mapping;
 746         unsigned int offset, this_len;
 747         struct page *page;
 748         void *fsdata;
 749         int ret;
 750
 751         offset = sd->pos & ~PAGE_CACHE_MASK;
 752
 753         this_len = sd->len;
 754         if (this_len + offset > PAGE_CACHE_SIZE)
 755                 this_len = PAGE_CACHE_SIZE - offset;
 756
 757         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 758                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 759         if (unlikely(ret))
 760                 goto out;
 761
 762         if (buf->page != page) {
 763                 char *src = kmap_atomic(buf->page);
 764                 char *dst = kmap_atomic(page);
 765
 766                 memcpy(dst + offset, src + buf->offset, this_len);
 767                 flush_dcache_page(page);
 768                 kunmap_atomic(dst);
 769                 kunmap_atomic(src);
 770         }
 771         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 772                                 page, fsdata);
 773 out:
 774         return ret;
 775 }
 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 static 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
 853 /**
 854  * splice_from_pipe_next - wait for some data to splice from
 855  * @pipe:       pipe to splice from
 856  * @sd:         information about the splice operation
 857  *
 858  * Description:
 859  *    This function will wait for some data and return a positive
 860  *    value (one) if pipe buffers are available.  It will return zero
 861  *    or -errno if no more data needs to be spliced.
 862  */
 863 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 864 {
 865         while (!pipe->nrbufs) {
 866                 if (!pipe->writers)
 867                         return 0;
 868
 869                 if (!pipe->waiting_writers && sd->num_spliced)
 870                         return 0;
 871
 872                 if (sd->flags & SPLICE_F_NONBLOCK)
 873                         return -EAGAIN;
 874
 875                 if (signal_pending(current))
 876                         return -ERESTARTSYS;
 877
 878                 if (sd->need_wakeup) {
 879                         wakeup_pipe_writers(pipe);
 880                         sd->need_wakeup = false;
 881                 }
 882
 883                 pipe_wait(pipe);
 884         }
 885
 886         return 1;
 887 }
 888
 889 /**
 890  * splice_from_pipe_begin - start splicing from pipe
 891  * @sd:         information about the splice operation
 892  *
 893  * Description:
 894  *    This function should be called before a loop containing
 895  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 896  *    initialize the necessary fields of @sd.
 897  */
 898 static void splice_from_pipe_begin(struct splice_desc *sd)
 899 {
 900         sd->num_spliced = 0;
 901         sd->need_wakeup = false;
 902 }
 903
 904 /**
 905  * splice_from_pipe_end - finish splicing from pipe
 906  * @pipe:       pipe to splice from
 907  * @sd:         information about the splice operation
 908  *
 909  * Description:
 910  *    This function will wake up pipe writers if necessary.  It should
 911  *    be called after a loop containing splice_from_pipe_next() and
 912  *    splice_from_pipe_feed().
 913  */
 914 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 915 {
 916         if (sd->need_wakeup)
 917                 wakeup_pipe_writers(pipe);
 918 }
 919
 920 /**
 921  * __splice_from_pipe - splice data from a pipe to given actor
 922  * @pipe:       pipe to splice from
 923  * @sd:         information to @actor
 924  * @actor:      handler that splices the data
 925  *
 926  * Description:
 927  *    This function does little more than loop over the pipe and call
 928  *    @actor to do the actual moving of a single struct pipe_buffer to
 929  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 930  *    pipe_to_user.
 931  *
 932  */
 933 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 934                            splice_actor *actor)
 935 {
 936         int ret;
 937
 938         splice_from_pipe_begin(sd);
 939         do {
 940                 ret = splice_from_pipe_next(pipe, sd);
 941                 if (ret > 0)
 942                         ret = splice_from_pipe_feed(pipe, sd, actor);
 943         } while (ret > 0);
 944         splice_from_pipe_end(pipe, sd);
 945
 946         return sd->num_spliced ? sd->num_spliced : ret;
 947 }
 948 EXPORT_SYMBOL(__splice_from_pipe);
 949
 950 /**
 951  * splice_from_pipe - splice data from a pipe to a file
 952  * @pipe:       pipe to splice from
 953  * @out:        file to splice to
 954  * @ppos:       position in @out
 955  * @len:        how many bytes to splice
 956  * @flags:      splice modifier flags
 957  * @actor:      handler that splices the data
 958  *
 959  * Description:
 960  *    See __splice_from_pipe. This function locks the pipe inode,
 961  *    otherwise it's identical to __splice_from_pipe().
 962  *
 963  */
 964 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 965                          loff_t *ppos, size_t len, unsigned int flags,
 966                          splice_actor *actor)
 967 {
 968         ssize_t ret;
 969         struct splice_desc sd = {
 970                 .total_len = len,
 971                 .flags = flags,
 972                 .pos = *ppos,
 973                 .u.file = out,
 974         };
 975
 976         pipe_lock(pipe);
 977         ret = __splice_from_pipe(pipe, &sd, actor);
 978         pipe_unlock(pipe);
 979
 980         return ret;
 981 }
 982
 983 /**
 984  * generic_file_splice_write - splice data from a pipe to a file
 985  * @pipe:       pipe info
 986  * @out:        file to write to
 987  * @ppos:       position in @out
 988  * @len:        number of bytes to splice
 989  * @flags:      splice modifier flags
 990  *
 991  * Description:
 992  *    Will either move or copy pages (determined by @flags options) from
 993  *    the given pipe inode to the given file.
 994  *
 995  */
 996 ssize_t
 997 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 998                           loff_t *ppos, size_t len, unsigned int flags)
 999 {
1000         struct address_space *mapping = out->f_mapping;
1001         struct inode *inode = mapping->host;
1002         struct splice_desc sd = {
1003                 .total_len = len,
1004                 .flags = flags,
1005                 .pos = *ppos,
1006                 .u.file = out,
1007         };
1008         ssize_t ret;
1009
1010         pipe_lock(pipe);
1011
1012         splice_from_pipe_begin(&sd);
1013         do {
1014                 ret = splice_from_pipe_next(pipe, &sd);
1015                 if (ret <= 0)
1016                         break;
1017
1018                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1019                 ret = file_remove_suid(out);
1020                 if (!ret) {
1021                         ret = file_update_time(out);
1022                         if (!ret)
1023                                 ret = splice_from_pipe_feed(pipe, &sd,
1024                                                             pipe_to_file);
1025                 }
1026                 mutex_unlock(&inode->i_mutex);
1027         } while (ret > 0);
1028         splice_from_pipe_end(pipe, &sd);
1029
1030         pipe_unlock(pipe);
1031
1032         if (sd.num_spliced)
1033                 ret = sd.num_spliced;
1034
1035         if (ret > 0) {
1036                 int err;
1037
1038                 err = generic_write_sync(out, *ppos, ret);
1039                 if (err)
1040                         ret = err;
1041                 else
1042                         *ppos += ret;
1043                 balance_dirty_pages_ratelimited(mapping);
1044         }
1045
1046         return ret;
1047 }
1048
1049 EXPORT_SYMBOL(generic_file_splice_write);
1050
1051 /**
1052  * iter_file_splice_write - splice data from a pipe to a file
1053  * @pipe:       pipe info
1054  * @out:        file to write to
1055  * @ppos:       position in @out
1056  * @len:        number of bytes to splice
1057  * @flags:      splice modifier flags
1058  *
1059  * Description:
1060  *    Will either move or copy pages (determined by @flags options) from
1061  *    the given pipe inode to the given file.
1062  *    This one is ->write_iter-based.
1063  *
1064  */
1065 ssize_t
1066 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1067                           loff_t *ppos, size_t len, unsigned int flags)
1068 {
1069         struct splice_desc sd = {
1070                 .total_len = len,
1071                 .flags = flags,
1072                 .pos = *ppos,
1073                 .u.file = out,
1074         };
1075         int nbufs = pipe->buffers;
1076         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
1077                                         GFP_KERNEL);
1078         ssize_t ret;
1079
1080         if (unlikely(!array))
1081                 return -ENOMEM;
1082
1083         pipe_lock(pipe);
1084
1085         splice_from_pipe_begin(&sd);
1086         while (sd.total_len) {
1087                 struct iov_iter from;
1088                 struct kiocb kiocb;
1089                 size_t left;
1090                 int n, idx;
1091
1092                 ret = splice_from_pipe_next(pipe, &sd);
1093                 if (ret <= 0)
1094                         break;
1095
1096                 if (unlikely(nbufs < pipe->buffers)) {
1097                         kfree(array);
1098                         nbufs = pipe->buffers;
1099                         array = kcalloc(nbufs, sizeof(struct bio_vec),
1100                                         GFP_KERNEL);
1101                         if (!array) {
1102                                 ret = -ENOMEM;
1103                                 break;
1104                         }
1105                 }
1106
1107                 /* build the vector */
1108                 left = sd.total_len;
1109                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
1110                         struct pipe_buffer *buf = pipe->bufs + idx;
1111                         size_t this_len = buf->len;
1112
1113                         if (this_len > left)
1114                                 this_len = left;
1115
1116                         if (idx == pipe->buffers - 1)
1117                                 idx = -1;
1118
1119                         ret = buf->ops->confirm(pipe, buf);
1120                         if (unlikely(ret)) {
1121                                 if (ret == -ENODATA)
1122                                         ret = 0;
1123                                 goto done;
1124                         }
1125
1126                         array[n].bv_page = buf->page;
1127                         array[n].bv_len = this_len;
1128                         array[n].bv_offset = buf->offset;
1129                         left -= this_len;
1130                 }
1131
1132                 /* ... iov_iter */
1133                 from.type = ITER_BVEC | WRITE;
1134                 from.bvec = array;
1135                 from.nr_segs = n;
1136                 from.count = sd.total_len - left;
1137                 from.iov_offset = 0;
1138
1139                 /* ... and iocb */
1140                 init_sync_kiocb(&kiocb, out);
1141                 kiocb.ki_pos = sd.pos;
1142                 kiocb.ki_nbytes = sd.total_len - left;
1143
1144                 /* now, send it */
1145                 ret = out->f_op->write_iter(&kiocb, &from);
1146                 if (-EIOCBQUEUED == ret)
1147                         ret = wait_on_sync_kiocb(&kiocb);
1148
1149                 if (ret <= 0)
1150                         break;
1151
1152                 sd.num_spliced += ret;
1153                 sd.total_len -= ret;
1154                 *ppos = sd.pos = kiocb.ki_pos;
1155
1156                 /* dismiss the fully eaten buffers, adjust the partial one */
1157                 while (ret) {
1158                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1159                         if (ret >= buf->len) {
1160                                 const struct pipe_buf_operations *ops = buf->ops;
1161                                 ret -= buf->len;
1162                                 buf->len = 0;
1163                                 buf->ops = NULL;
1164                                 ops->release(pipe, buf);
1165                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1166                                 pipe->nrbufs--;
1167                                 if (pipe->files)
1168                                         sd.need_wakeup = true;
1169                         } else {
1170                                 buf->offset += ret;
1171                                 buf->len -= ret;
1172                                 ret = 0;
1173                         }
1174                 }
1175         }
1176 done:
1177         kfree(array);
1178         splice_from_pipe_end(pipe, &sd);
1179
1180         pipe_unlock(pipe);
1181
1182         if (sd.num_spliced)
1183                 ret = sd.num_spliced;
1184
1185         return ret;
1186 }
1187
1188 EXPORT_SYMBOL(iter_file_splice_write);
1189
1190 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1191                           struct splice_desc *sd)
1192 {
1193         int ret;
1194         void *data;
1195         loff_t tmp = sd->pos;
1196
1197         data = kmap(buf->page);
1198         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1199         kunmap(buf->page);
1200
1201         return ret;
1202 }
1203
1204 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1205                                          struct file *out, loff_t *ppos,
1206                                          size_t len, unsigned int flags)
1207 {
1208         ssize_t ret;
1209
1210         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1211         if (ret > 0)
1212                 *ppos += ret;
1213
1214         return ret;
1215 }
1216
1217 /**
1218  * generic_splice_sendpage - splice data from a pipe to a socket
1219  * @pipe:       pipe to splice from
1220  * @out:        socket to write to
1221  * @ppos:       position in @out
1222  * @len:        number of bytes to splice
1223  * @flags:      splice modifier flags
1224  *
1225  * Description:
1226  *    Will send @len bytes from the pipe to a network socket. No data copying
1227  *    is involved.
1228  *
1229  */
1230 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1231                                 loff_t *ppos, size_t len, unsigned int flags)
1232 {
1233         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1234 }
1235
1236 EXPORT_SYMBOL(generic_splice_sendpage);
1237
1238 /*
1239  * Attempt to initiate a splice from pipe to file.
1240  */
1241 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1242                            loff_t *ppos, size_t len, unsigned int flags)
1243 {
1244         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1245                                 loff_t *, size_t, unsigned int);
1246
1247         if (out->f_op->splice_write)
1248                 splice_write = out->f_op->splice_write;
1249         else
1250                 splice_write = default_file_splice_write;
1251
1252         return splice_write(pipe, out, ppos, len, flags);
1253 }
1254
1255 /*
1256  * Attempt to initiate a splice from a file to a pipe.
1257  */
1258 static long do_splice_to(struct file *in, loff_t *ppos,
1259                          struct pipe_inode_info *pipe, size_t len,
1260                          unsigned int flags)
1261 {
1262         ssize_t (*splice_read)(struct file *, loff_t *,
1263                                struct pipe_inode_info *, size_t, unsigned int);
1264         int ret;
1265
1266         if (unlikely(!(in->f_mode & FMODE_READ)))
1267                 return -EBADF;
1268
1269         ret = rw_verify_area(READ, in, ppos, len);
1270         if (unlikely(ret < 0))
1271                 return ret;
1272
1273         if (in->f_op->splice_read)
1274                 splice_read = in->f_op->splice_read;
1275         else
1276                 splice_read = default_file_splice_read;
1277
1278         return splice_read(in, ppos, pipe, len, flags);
1279 }
1280
1281 /**
1282  * splice_direct_to_actor - splices data directly between two non-pipes
1283  * @in:         file to splice from
1284  * @sd:         actor information on where to splice to
1285  * @actor:      handles the data splicing
1286  *
1287  * Description:
1288  *    This is a special case helper to splice directly between two
1289  *    points, without requiring an explicit pipe. Internally an allocated
1290  *    pipe is cached in the process, and reused during the lifetime of
1291  *    that process.
1292  *
1293  */
1294 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1295                                splice_direct_actor *actor)
1296 {
1297         struct pipe_inode_info *pipe;
1298         long ret, bytes;
1299         umode_t i_mode;
1300         size_t len;
1301         int i, flags;
1302
1303         /*
1304          * We require the input being a regular file, as we don't want to
1305          * randomly drop data for eg socket -> socket splicing. Use the
1306          * piped splicing for that!
1307          */
1308         i_mode = file_inode(in)->i_mode;
1309         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1310                 return -EINVAL;
1311
1312         /*
1313          * neither in nor out is a pipe, setup an internal pipe attached to
1314          * 'out' and transfer the wanted data from 'in' to 'out' through that
1315          */
1316         pipe = current->splice_pipe;
1317         if (unlikely(!pipe)) {
1318                 pipe = alloc_pipe_info();
1319                 if (!pipe)
1320                         return -ENOMEM;
1321
1322                 /*
1323                  * We don't have an immediate reader, but we'll read the stuff
1324                  * out of the pipe right after the splice_to_pipe(). So set
1325                  * PIPE_READERS appropriately.
1326                  */
1327                 pipe->readers = 1;
1328
1329                 current->splice_pipe = pipe;
1330         }
1331
1332         /*
1333          * Do the splice.
1334          */
1335         ret = 0;
1336         bytes = 0;
1337         len = sd->total_len;
1338         flags = sd->flags;
1339
1340         /*
1341          * Don't block on output, we have to drain the direct pipe.
1342          */
1343         sd->flags &= ~SPLICE_F_NONBLOCK;
1344
1345         while (len) {
1346                 size_t read_len;
1347                 loff_t pos = sd->pos, prev_pos = pos;
1348
1349                 ret = do_splice_to(in, &pos, pipe, len, flags);
1350                 if (unlikely(ret <= 0))
1351                         goto out_release;
1352
1353                 read_len = ret;
1354                 sd->total_len = read_len;
1355
1356                 /*
1357                  * NOTE: nonblocking mode only applies to the input. We
1358                  * must not do the output in nonblocking mode as then we
1359                  * could get stuck data in the internal pipe:
1360                  */
1361                 ret = actor(pipe, sd);
1362                 if (unlikely(ret <= 0)) {
1363                         sd->pos = prev_pos;
1364                         goto out_release;
1365                 }
1366
1367                 bytes += ret;
1368                 len -= ret;
1369                 sd->pos = pos;
1370
1371                 if (ret < read_len) {
1372                         sd->pos = prev_pos + ret;
1373                         goto out_release;
1374                 }
1375         }
1376
1377 done:
1378         pipe->nrbufs = pipe->curbuf = 0;
1379         file_accessed(in);
1380         return bytes;
1381
1382 out_release:
1383         /*
1384          * If we did an incomplete transfer we must release
1385          * the pipe buffers in question:
1386          */
1387         for (i = 0; i < pipe->buffers; i++) {
1388                 struct pipe_buffer *buf = pipe->bufs + i;
1389
1390                 if (buf->ops) {
1391                         buf->ops->release(pipe, buf);
1392                         buf->ops = NULL;
1393                 }
1394         }
1395
1396         if (!bytes)
1397                 bytes = ret;
1398
1399         goto done;
1400 }
1401 EXPORT_SYMBOL(splice_direct_to_actor);
1402
1403 static int direct_splice_actor(struct pipe_inode_info *pipe,
1404                                struct splice_desc *sd)
1405 {
1406         struct file *file = sd->u.file;
1407
1408         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1409                               sd->flags);
1410 }
1411
1412 /**
1413  * do_splice_direct - splices data directly between two files
1414  * @in:         file to splice from
1415  * @ppos:       input file offset
1416  * @out:        file to splice to
1417  * @opos:       output file offset
1418  * @len:        number of bytes to splice
1419  * @flags:      splice modifier flags
1420  *
1421  * Description:
1422  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1423  *    doing it in the application would incur an extra system call
1424  *    (splice in + splice out, as compared to just sendfile()). So this helper
1425  *    can splice directly through a process-private pipe.
1426  *
1427  */
1428 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1429                       loff_t *opos, size_t len, unsigned int flags)
1430 {
1431         struct splice_desc sd = {
1432                 .len            = len,
1433                 .total_len      = len,
1434                 .flags          = flags,
1435                 .pos            = *ppos,
1436                 .u.file         = out,
1437                 .opos           = opos,
1438         };
1439         long ret;
1440
1441         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1442                 return -EBADF;
1443
1444         if (unlikely(out->f_flags & O_APPEND))
1445                 return -EINVAL;
1446
1447         ret = rw_verify_area(WRITE, out, opos, len);
1448         if (unlikely(ret < 0))
1449                 return ret;
1450
1451         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1452         if (ret > 0)
1453                 *ppos = sd.pos;
1454
1455         return ret;
1456 }
1457
1458 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1459                                struct pipe_inode_info *opipe,
1460                                size_t len, unsigned int flags);
1461
1462 /*
1463  * Determine where to splice to/from.
1464  */
1465 static long do_splice(struct file *in, loff_t __user *off_in,
1466                       struct file *out, loff_t __user *off_out,
1467                       size_t len, unsigned int flags)
1468 {
1469         struct pipe_inode_info *ipipe;
1470         struct pipe_inode_info *opipe;
1471         loff_t offset;
1472         long ret;
1473
1474         ipipe = get_pipe_info(in);
1475         opipe = get_pipe_info(out);
1476
1477         if (ipipe && opipe) {
1478                 if (off_in || off_out)
1479                         return -ESPIPE;
1480
1481                 if (!(in->f_mode & FMODE_READ))
1482                         return -EBADF;
1483
1484                 if (!(out->f_mode & FMODE_WRITE))
1485                         return -EBADF;
1486
1487                 /* Splicing to self would be fun, but... */
1488                 if (ipipe == opipe)
1489                         return -EINVAL;
1490
1491                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1492         }
1493
1494         if (ipipe) {
1495                 if (off_in)
1496                         return -ESPIPE;
1497                 if (off_out) {
1498                         if (!(out->f_mode & FMODE_PWRITE))
1499                                 return -EINVAL;
1500                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1501                                 return -EFAULT;
1502                 } else {
1503                         offset = out->f_pos;
1504                 }
1505
1506                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1507                         return -EBADF;
1508
1509                 if (unlikely(out->f_flags & O_APPEND))
1510                         return -EINVAL;
1511
1512                 ret = rw_verify_area(WRITE, out, &offset, len);
1513                 if (unlikely(ret < 0))
1514                         return ret;
1515
1516                 file_start_write(out);
1517                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1518                 file_end_write(out);
1519
1520                 if (!off_out)
1521                         out->f_pos = offset;
1522                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1523                         ret = -EFAULT;
1524
1525                 return ret;
1526         }
1527
1528         if (opipe) {
1529                 if (off_out)
1530                         return -ESPIPE;
1531                 if (off_in) {
1532                         if (!(in->f_mode & FMODE_PREAD))
1533                                 return -EINVAL;
1534                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1535                                 return -EFAULT;
1536                 } else {
1537                         offset = in->f_pos;
1538                 }
1539
1540                 ret = do_splice_to(in, &offset, opipe, len, flags);
1541
1542                 if (!off_in)
1543                         in->f_pos = offset;
1544                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1545                         ret = -EFAULT;
1546
1547                 return ret;
1548         }
1549
1550         return -EINVAL;
1551 }
1552
1553 /*
1554  * Map an iov into an array of pages and offset/length tupples. With the
1555  * partial_page structure, we can map several non-contiguous ranges into
1556  * our ones pages[] map instead of splitting that operation into pieces.
1557  * Could easily be exported as a generic helper for other users, in which
1558  * case one would probably want to add a 'max_nr_pages' parameter as well.
1559  */
1560 static int get_iovec_page_array(const struct iovec __user *iov,
1561                                 unsigned int nr_vecs, struct page **pages,
1562                                 struct partial_page *partial, bool aligned,
1563                                 unsigned int pipe_buffers)
1564 {
1565         int buffers = 0, error = 0;
1566
1567         while (nr_vecs) {
1568                 unsigned long off, npages;
1569                 struct iovec entry;
1570                 void __user *base;
1571                 size_t len;
1572                 int i;
1573
1574                 error = -EFAULT;
1575                 if (copy_from_user(&entry, iov, sizeof(entry)))
1576                         break;
1577
1578                 base = entry.iov_base;
1579                 len = entry.iov_len;
1580
1581                 /*
1582                  * Sanity check this iovec. 0 read succeeds.
1583                  */
1584                 error = 0;
1585                 if (unlikely(!len))
1586                         break;
1587                 error = -EFAULT;
1588                 if (!access_ok(VERIFY_READ, base, len))
1589                         break;
1590
1591                 /*
1592                  * Get this base offset and number of pages, then map
1593                  * in the user pages.
1594                  */
1595                 off = (unsigned long) base & ~PAGE_MASK;
1596
1597                 /*
1598                  * If asked for alignment, the offset must be zero and the
1599                  * length a multiple of the PAGE_SIZE.
1600                  */
1601                 error = -EINVAL;
1602                 if (aligned && (off || len & ~PAGE_MASK))
1603                         break;
1604
1605                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1606                 if (npages > pipe_buffers - buffers)
1607                         npages = pipe_buffers - buffers;
1608
1609                 error = get_user_pages_fast((unsigned long)base, npages,
1610                                         0, &pages[buffers]);
1611
1612                 if (unlikely(error <= 0))
1613                         break;
1614
1615                 /*
1616                  * Fill this contiguous range into the partial page map.
1617                  */
1618                 for (i = 0; i < error; i++) {
1619                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1620
1621                         partial[buffers].offset = off;
1622                         partial[buffers].len = plen;
1623
1624                         off = 0;
1625                         len -= plen;
1626                         buffers++;
1627                 }
1628
1629                 /*
1630                  * We didn't complete this iov, stop here since it probably
1631                  * means we have to move some of this into a pipe to
1632                  * be able to continue.
1633                  */
1634                 if (len)
1635                         break;
1636
1637                 /*
1638                  * Don't continue if we mapped fewer pages than we asked for,
1639                  * or if we mapped the max number of pages that we have
1640                  * room for.
1641                  */
1642                 if (error < npages || buffers == pipe_buffers)
1643                         break;
1644
1645                 nr_vecs--;
1646                 iov++;
1647         }
1648
1649         if (buffers)
1650                 return buffers;
1651
1652         return error;
1653 }
1654
1655 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1656                         struct splice_desc *sd)
1657 {
1658         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1659         return n == sd->len ? n : -EFAULT;
1660 }
1661
1662 /*
1663  * For lack of a better implementation, implement vmsplice() to userspace
1664  * as a simple copy of the pipes pages to the user iov.
1665  */
1666 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1667                              unsigned long nr_segs, unsigned int flags)
1668 {
1669         struct pipe_inode_info *pipe;
1670         struct splice_desc sd;
1671         long ret;
1672         struct iovec iovstack[UIO_FASTIOV];
1673         struct iovec *iov = iovstack;
1674         struct iov_iter iter;
1675         ssize_t count = 0;
1676
1677         pipe = get_pipe_info(file);
1678         if (!pipe)
1679                 return -EBADF;
1680
1681         ret = rw_copy_check_uvector(READ, uiov, nr_segs,
1682                                     ARRAY_SIZE(iovstack), iovstack, &iov);
1683         if (ret <= 0)
1684                 return ret;
1685
1686         iov_iter_init(&iter, READ, iov, nr_segs, count);
1687
1688         sd.len = 0;
1689         sd.total_len = count;
1690         sd.flags = flags;
1691         sd.u.data = &iter;
1692         sd.pos = 0;
1693
1694         pipe_lock(pipe);
1695         ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1696         pipe_unlock(pipe);
1697
1698         if (iov != iovstack)
1699                 kfree(iov);
1700
1701         return ret;
1702 }
1703
1704 /*
1705  * vmsplice splices a user address range into a pipe. It can be thought of
1706  * as splice-from-memory, where the regular splice is splice-from-file (or
1707  * to file). In both cases the output is a pipe, naturally.
1708  */
1709 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1710                              unsigned long nr_segs, unsigned int flags)
1711 {
1712         struct pipe_inode_info *pipe;
1713         struct page *pages[PIPE_DEF_BUFFERS];
1714         struct partial_page partial[PIPE_DEF_BUFFERS];
1715         struct splice_pipe_desc spd = {
1716                 .pages = pages,
1717                 .partial = partial,
1718                 .nr_pages_max = PIPE_DEF_BUFFERS,
1719                 .flags = flags,
1720                 .ops = &user_page_pipe_buf_ops,
1721                 .spd_release = spd_release_page,
1722         };
1723         long ret;
1724
1725         pipe = get_pipe_info(file);
1726         if (!pipe)
1727                 return -EBADF;
1728
1729         if (splice_grow_spd(pipe, &spd))
1730                 return -ENOMEM;
1731
1732         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1733                                             spd.partial, false,
1734                                             spd.nr_pages_max);
1735         if (spd.nr_pages <= 0)
1736                 ret = spd.nr_pages;
1737         else
1738                 ret = splice_to_pipe(pipe, &spd);
1739
1740         splice_shrink_spd(&spd);
1741         return ret;
1742 }
1743
1744 /*
1745  * Note that vmsplice only really supports true splicing _from_ user memory
1746  * to a pipe, not the other way around. Splicing from user memory is a simple
1747  * operation that can be supported without any funky alignment restrictions
1748  * or nasty vm tricks. We simply map in the user memory and fill them into
1749  * a pipe. The reverse isn't quite as easy, though. There are two possible
1750  * solutions for that:
1751  *
1752  *      - memcpy() the data internally, at which point we might as well just
1753  *        do a regular read() on the buffer anyway.
1754  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1755  *        has restriction limitations on both ends of the pipe).
1756  *
1757  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1758  *
1759  */
1760 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1761                 unsigned long, nr_segs, unsigned int, flags)
1762 {
1763         struct fd f;
1764         long error;
1765
1766         if (unlikely(nr_segs > UIO_MAXIOV))
1767                 return -EINVAL;
1768         else if (unlikely(!nr_segs))
1769                 return 0;
1770
1771         error = -EBADF;
1772         f = fdget(fd);
1773         if (f.file) {
1774                 if (f.file->f_mode & FMODE_WRITE)
1775                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1776                 else if (f.file->f_mode & FMODE_READ)
1777                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1778
1779                 fdput(f);
1780         }
1781
1782         return error;
1783 }
1784
1785 #ifdef CONFIG_COMPAT
1786 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1787                     unsigned int, nr_segs, unsigned int, flags)
1788 {
1789         unsigned i;
1790         struct iovec __user *iov;
1791         if (nr_segs > UIO_MAXIOV)
1792                 return -EINVAL;
1793         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1794         for (i = 0; i < nr_segs; i++) {
1795                 struct compat_iovec v;
1796                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1797                     get_user(v.iov_len, &iov32[i].iov_len) ||
1798                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1799                     put_user(v.iov_len, &iov[i].iov_len))
1800                         return -EFAULT;
1801         }
1802         return sys_vmsplice(fd, iov, nr_segs, flags);
1803 }
1804 #endif
1805
1806 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1807                 int, fd_out, loff_t __user *, off_out,
1808                 size_t, len, unsigned int, flags)
1809 {
1810         struct fd in, out;
1811         long error;
1812
1813         if (unlikely(!len))
1814                 return 0;
1815
1816         error = -EBADF;
1817         in = fdget(fd_in);
1818         if (in.file) {
1819                 if (in.file->f_mode & FMODE_READ) {
1820                         out = fdget(fd_out);
1821                         if (out.file) {
1822                                 if (out.file->f_mode & FMODE_WRITE)
1823                                         error = do_splice(in.file, off_in,
1824                                                           out.file, off_out,
1825                                                           len, flags);
1826                                 fdput(out);
1827                         }
1828                 }
1829                 fdput(in);
1830         }
1831         return error;
1832 }
1833
1834 /*
1835  * Make sure there's data to read. Wait for input if we can, otherwise
1836  * return an appropriate error.
1837  */
1838 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1839 {
1840         int ret;
1841
1842         /*
1843          * Check ->nrbufs without the inode lock first. This function
1844          * is speculative anyways, so missing one is ok.
1845          */
1846         if (pipe->nrbufs)
1847                 return 0;
1848
1849         ret = 0;
1850         pipe_lock(pipe);
1851
1852         while (!pipe->nrbufs) {
1853                 if (signal_pending(current)) {
1854                         ret = -ERESTARTSYS;
1855                         break;
1856                 }
1857                 if (!pipe->writers)
1858                         break;
1859                 if (!pipe->waiting_writers) {
1860                         if (flags & SPLICE_F_NONBLOCK) {
1861                                 ret = -EAGAIN;
1862                                 break;
1863                         }
1864                 }
1865                 pipe_wait(pipe);
1866         }
1867
1868         pipe_unlock(pipe);
1869         return ret;
1870 }
1871
1872 /*
1873  * Make sure there's writeable room. Wait for room if we can, otherwise
1874  * return an appropriate error.
1875  */
1876 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1877 {
1878         int ret;
1879
1880         /*
1881          * Check ->nrbufs without the inode lock first. This function
1882          * is speculative anyways, so missing one is ok.
1883          */
1884         if (pipe->nrbufs < pipe->buffers)
1885                 return 0;
1886
1887         ret = 0;
1888         pipe_lock(pipe);
1889
1890         while (pipe->nrbufs >= pipe->buffers) {
1891                 if (!pipe->readers) {
1892                         send_sig(SIGPIPE, current, 0);
1893                         ret = -EPIPE;
1894                         break;
1895                 }
1896                 if (flags & SPLICE_F_NONBLOCK) {
1897                         ret = -EAGAIN;
1898                         break;
1899                 }
1900                 if (signal_pending(current)) {
1901                         ret = -ERESTARTSYS;
1902                         break;
1903                 }
1904                 pipe->waiting_writers++;
1905                 pipe_wait(pipe);
1906                 pipe->waiting_writers--;
1907         }
1908
1909         pipe_unlock(pipe);
1910         return ret;
1911 }
1912
1913 /*
1914  * Splice contents of ipipe to opipe.
1915  */
1916 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1917                                struct pipe_inode_info *opipe,
1918                                size_t len, unsigned int flags)
1919 {
1920         struct pipe_buffer *ibuf, *obuf;
1921         int ret = 0, nbuf;
1922         bool input_wakeup = false;
1923
1924
1925 retry:
1926         ret = ipipe_prep(ipipe, flags);
1927         if (ret)
1928                 return ret;
1929
1930         ret = opipe_prep(opipe, flags);
1931         if (ret)
1932                 return ret;
1933
1934         /*
1935          * Potential ABBA deadlock, work around it by ordering lock
1936          * grabbing by pipe info address. Otherwise two different processes
1937          * could deadlock (one doing tee from A -> B, the other from B -> A).
1938          */
1939         pipe_double_lock(ipipe, opipe);
1940
1941         do {
1942                 if (!opipe->readers) {
1943                         send_sig(SIGPIPE, current, 0);
1944                         if (!ret)
1945                                 ret = -EPIPE;
1946                         break;
1947                 }
1948
1949                 if (!ipipe->nrbufs && !ipipe->writers)
1950                         break;
1951
1952                 /*
1953                  * Cannot make any progress, because either the input
1954                  * pipe is empty or the output pipe is full.
1955                  */
1956                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1957                         /* Already processed some buffers, break */
1958                         if (ret)
1959                                 break;
1960
1961                         if (flags & SPLICE_F_NONBLOCK) {
1962                                 ret = -EAGAIN;
1963                                 break;
1964                         }
1965
1966                         /*
1967                          * We raced with another reader/writer and haven't
1968                          * managed to process any buffers.  A zero return
1969                          * value means EOF, so retry instead.
1970                          */
1971                         pipe_unlock(ipipe);
1972                         pipe_unlock(opipe);
1973                         goto retry;
1974                 }
1975
1976                 ibuf = ipipe->bufs + ipipe->curbuf;
1977                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1978                 obuf = opipe->bufs + nbuf;
1979
1980                 if (len >= ibuf->len) {
1981                         /*
1982                          * Simply move the whole buffer from ipipe to opipe
1983                          */
1984                         *obuf = *ibuf;
1985                         ibuf->ops = NULL;
1986                         opipe->nrbufs++;
1987                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1988                         ipipe->nrbufs--;
1989                         input_wakeup = true;
1990                 } else {
1991                         /*
1992                          * Get a reference to this pipe buffer,
1993                          * so we can copy the contents over.
1994                          */
1995                         ibuf->ops->get(ipipe, ibuf);
1996                         *obuf = *ibuf;
1997
1998                         /*
1999                          * Don't inherit the gift flag, we need to
2000                          * prevent multiple steals of this page.
2001                          */
2002                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2003
2004                         obuf->len = len;
2005                         opipe->nrbufs++;
2006                         ibuf->offset += obuf->len;
2007                         ibuf->len -= obuf->len;
2008                 }
2009                 ret += obuf->len;
2010                 len -= obuf->len;
2011         } while (len);
2012
2013         pipe_unlock(ipipe);
2014         pipe_unlock(opipe);
2015
2016         /*
2017          * If we put data in the output pipe, wakeup any potential readers.
2018          */
2019         if (ret > 0)
2020                 wakeup_pipe_readers(opipe);
2021
2022         if (input_wakeup)
2023                 wakeup_pipe_writers(ipipe);
2024
2025         return ret;
2026 }
2027
2028 /*
2029  * Link contents of ipipe to opipe.
2030  */
2031 static int link_pipe(struct pipe_inode_info *ipipe,
2032                      struct pipe_inode_info *opipe,
2033                      size_t len, unsigned int flags)
2034 {
2035         struct pipe_buffer *ibuf, *obuf;
2036         int ret = 0, i = 0, nbuf;
2037
2038         /*
2039          * Potential ABBA deadlock, work around it by ordering lock
2040          * grabbing by pipe info address. Otherwise two different processes
2041          * could deadlock (one doing tee from A -> B, the other from B -> A).
2042          */
2043         pipe_double_lock(ipipe, opipe);
2044
2045         do {
2046                 if (!opipe->readers) {
2047                         send_sig(SIGPIPE, current, 0);
2048                         if (!ret)
2049                                 ret = -EPIPE;
2050                         break;
2051                 }
2052
2053                 /*
2054                  * If we have iterated all input buffers or ran out of
2055                  * output room, break.
2056                  */
2057                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
2058                         break;
2059
2060                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
2061                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2062
2063                 /*
2064                  * Get a reference to this pipe buffer,
2065                  * so we can copy the contents over.
2066                  */
2067                 ibuf->ops->get(ipipe, ibuf);
2068
2069                 obuf = opipe->bufs + nbuf;
2070                 *obuf = *ibuf;
2071
2072                 /*
2073                  * Don't inherit the gift flag, we need to
2074                  * prevent multiple steals of this page.
2075                  */
2076                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2077
2078                 if (obuf->len > len)
2079                         obuf->len = len;
2080
2081                 opipe->nrbufs++;
2082                 ret += obuf->len;
2083                 len -= obuf->len;
2084                 i++;
2085         } while (len);
2086
2087         /*
2088          * return EAGAIN if we have the potential of some data in the
2089          * future, otherwise just return 0
2090          */
2091         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2092                 ret = -EAGAIN;
2093
2094         pipe_unlock(ipipe);
2095         pipe_unlock(opipe);
2096
2097         /*
2098          * If we put data in the output pipe, wakeup any potential readers.
2099          */
2100         if (ret > 0)
2101                 wakeup_pipe_readers(opipe);
2102
2103         return ret;
2104 }
2105
2106 /*
2107  * This is a tee(1) implementation that works on pipes. It doesn't copy
2108  * any data, it simply references the 'in' pages on the 'out' pipe.
2109  * The 'flags' used are the SPLICE_F_* variants, currently the only
2110  * applicable one is SPLICE_F_NONBLOCK.
2111  */
2112 static long do_tee(struct file *in, struct file *out, size_t len,
2113                    unsigned int flags)
2114 {
2115         struct pipe_inode_info *ipipe = get_pipe_info(in);
2116         struct pipe_inode_info *opipe = get_pipe_info(out);
2117         int ret = -EINVAL;
2118
2119         /*
2120          * Duplicate the contents of ipipe to opipe without actually
2121          * copying the data.
2122          */
2123         if (ipipe && opipe && ipipe != opipe) {
2124                 /*
2125                  * Keep going, unless we encounter an error. The ipipe/opipe
2126                  * ordering doesn't really matter.
2127                  */
2128                 ret = ipipe_prep(ipipe, flags);
2129                 if (!ret) {
2130                         ret = opipe_prep(opipe, flags);
2131                         if (!ret)
2132                                 ret = link_pipe(ipipe, opipe, len, flags);
2133                 }
2134         }
2135
2136         return ret;
2137 }
2138
2139 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2140 {
2141         struct fd in;
2142         int error;
2143
2144         if (unlikely(!len))
2145                 return 0;
2146
2147         error = -EBADF;
2148         in = fdget(fdin);
2149         if (in.file) {
2150                 if (in.file->f_mode & FMODE_READ) {
2151                         struct fd out = fdget(fdout);
2152                         if (out.file) {
2153                                 if (out.file->f_mode & FMODE_WRITE)
2154                                         error = do_tee(in.file, out.file,
2155                                                         len, flags);
2156                                 fdput(out);
2157                         }
2158                 }
2159                 fdput(in);
2160         }
2161
2162         return error;
2163 }