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