fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files, network, direct splicing, etc and
  13  * fixing lots of bugs.
  14  *
  15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18  *
  19  */
  20 #include <linux/fs.h>
  21 #include <linux/file.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/splice.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm_inline.h>
  26 #include <linux/swap.h>
  27 #include <linux/writeback.h>
  28 #include <linux/export.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/uio.h>
  31 #include <linux/security.h>
  32 #include <linux/gfp.h>
  33 #include <linux/socket.h>
  34 #include <linux/compat.h>
  35 #include "internal.h"
  36
  37 /*
  38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  39  * a vm helper function, it's already simplified quite a bit by the
  40  * addition of remove_mapping(). If success is returned, the caller may
  41  * attempt to reuse this page for another destination.
  42  */
  43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  44                                      struct pipe_buffer *buf)
  45 {
  46         struct page *page = buf->page;
  47         struct address_space *mapping;
  48
  49         lock_page(page);
  50
  51         mapping = page_mapping(page);
  52         if (mapping) {
  53                 WARN_ON(!PageUptodate(page));
  54
  55                 /*
  56                  * At least for ext2 with nobh option, we need to wait on
  57                  * writeback completing on this page, since we'll remove it
  58                  * from the pagecache.  Otherwise truncate wont wait on the
  59                  * page, allowing the disk blocks to be reused by someone else
  60                  * before we actually wrote our data to them. fs corruption
  61                  * ensues.
  62                  */
  63                 wait_on_page_writeback(page);
  64
  65                 if (page_has_private(page) &&
  66                     !try_to_release_page(page, GFP_KERNEL))
  67                         goto out_unlock;
  68
  69                 /*
  70                  * If we succeeded in removing the mapping, set LRU flag
  71                  * and return good.
  72                  */
  73                 if (remove_mapping(mapping, page)) {
  74                         buf->flags |= PIPE_BUF_FLAG_LRU;
  75                         return 0;
  76                 }
  77         }
  78
  79         /*
  80          * Raced with truncate or failed to remove page from current
  81          * address space, unlock and return failure.
  82          */
  83 out_unlock:
  84         unlock_page(page);
  85         return 1;
  86 }
  87
  88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  89                                         struct pipe_buffer *buf)
  90 {
  91         page_cache_release(buf->page);
  92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  93 }
  94
  95 /*
  96  * Check whether the contents of buf is OK to access. Since the content
  97  * is a page cache page, IO may be in flight.
  98  */
  99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 100                                        struct pipe_buffer *buf)
 101 {
 102         struct page *page = buf->page;
 103         int err;
 104
 105         if (!PageUptodate(page)) {
 106                 lock_page(page);
 107
 108                 /*
 109                  * Page got truncated/unhashed. This will cause a 0-byte
 110                  * splice, if this is the first page.
 111                  */
 112                 if (!page->mapping) {
 113                         err = -ENODATA;
 114                         goto error;
 115                 }
 116
 117                 /*
 118                  * Uh oh, read-error from disk.
 119                  */
 120                 if (!PageUptodate(page)) {
 121                         err = -EIO;
 122                         goto error;
 123                 }
 124
 125                 /*
 126                  * Page is ok afterall, we are done.
 127                  */
 128                 unlock_page(page);
 129         }
 130
 131         return 0;
 132 error:
 133         unlock_page(page);
 134         return err;
 135 }
 136
 137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 138         .can_merge = 0,
 139         .map = generic_pipe_buf_map,
 140         .unmap = generic_pipe_buf_unmap,
 141         .confirm = page_cache_pipe_buf_confirm,
 142         .release = page_cache_pipe_buf_release,
 143         .steal = page_cache_pipe_buf_steal,
 144         .get = generic_pipe_buf_get,
 145 };
 146
 147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 148                                     struct pipe_buffer *buf)
 149 {
 150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 151                 return 1;
 152
 153         buf->flags |= PIPE_BUF_FLAG_LRU;
 154         return generic_pipe_buf_steal(pipe, buf);
 155 }
 156
 157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 158         .can_merge = 0,
 159         .map = generic_pipe_buf_map,
 160         .unmap = generic_pipe_buf_unmap,
 161         .confirm = generic_pipe_buf_confirm,
 162         .release = page_cache_pipe_buf_release,
 163         .steal = user_page_pipe_buf_steal,
 164         .get = generic_pipe_buf_get,
 165 };
 166
 167 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 168 {
 169         smp_mb();
 170         if (waitqueue_active(&pipe->wait))
 171                 wake_up_interruptible(&pipe->wait);
 172         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 173 }
 174
 175 /**
 176  * splice_to_pipe - fill passed data into a pipe
 177  * @pipe:       pipe to fill
 178  * @spd:        data to fill
 179  *
 180  * Description:
 181  *    @spd contains a map of pages and len/offset tuples, along with
 182  *    the struct pipe_buf_operations associated with these pages. This
 183  *    function will link that data to the pipe.
 184  *
 185  */
 186 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 187                        struct splice_pipe_desc *spd)
 188 {
 189         unsigned int spd_pages = spd->nr_pages;
 190         int ret, do_wakeup, page_nr;
 191
 192         ret = 0;
 193         do_wakeup = 0;
 194         page_nr = 0;
 195
 196         pipe_lock(pipe);
 197
 198         for (;;) {
 199                 if (!pipe->readers) {
 200                         send_sig(SIGPIPE, current, 0);
 201                         if (!ret)
 202                                 ret = -EPIPE;
 203                         break;
 204                 }
 205
 206                 if (pipe->nrbufs < pipe->buffers) {
 207                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 208                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 209
 210                         buf->page = spd->pages[page_nr];
 211                         buf->offset = spd->partial[page_nr].offset;
 212                         buf->len = spd->partial[page_nr].len;
 213                         buf->private = spd->partial[page_nr].private;
 214                         buf->ops = spd->ops;
 215                         if (spd->flags & SPLICE_F_GIFT)
 216                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
 217
 218                         pipe->nrbufs++;
 219                         page_nr++;
 220                         ret += buf->len;
 221
 222                         if (pipe->files)
 223                                 do_wakeup = 1;
 224
 225                         if (!--spd->nr_pages)
 226                                 break;
 227                         if (pipe->nrbufs < pipe->buffers)
 228                                 continue;
 229
 230                         break;
 231                 }
 232
 233                 if (spd->flags & SPLICE_F_NONBLOCK) {
 234                         if (!ret)
 235                                 ret = -EAGAIN;
 236                         break;
 237                 }
 238
 239                 if (signal_pending(current)) {
 240                         if (!ret)
 241                                 ret = -ERESTARTSYS;
 242                         break;
 243                 }
 244
 245                 if (do_wakeup) {
 246                         smp_mb();
 247                         if (waitqueue_active(&pipe->wait))
 248                                 wake_up_interruptible_sync(&pipe->wait);
 249                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 250                         do_wakeup = 0;
 251                 }
 252
 253                 pipe->waiting_writers++;
 254                 pipe_wait(pipe);
 255                 pipe->waiting_writers--;
 256         }
 257
 258         pipe_unlock(pipe);
 259
 260         if (do_wakeup)
 261                 wakeup_pipe_readers(pipe);
 262
 263         while (page_nr < spd_pages)
 264                 spd->spd_release(spd, page_nr++);
 265
 266         return ret;
 267 }
 268
 269 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 270 {
 271         page_cache_release(spd->pages[i]);
 272 }
 273
 274 /*
 275  * Check if we need to grow the arrays holding pages and partial page
 276  * descriptions.
 277  */
 278 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 279 {
 280         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 281
 282         spd->nr_pages_max = buffers;
 283         if (buffers <= PIPE_DEF_BUFFERS)
 284                 return 0;
 285
 286         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 287         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 288
 289         if (spd->pages && spd->partial)
 290                 return 0;
 291
 292         kfree(spd->pages);
 293         kfree(spd->partial);
 294         return -ENOMEM;
 295 }
 296
 297 void splice_shrink_spd(struct splice_pipe_desc *spd)
 298 {
 299         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 300                 return;
 301
 302         kfree(spd->pages);
 303         kfree(spd->partial);
 304 }
 305
 306 static int
 307 __generic_file_splice_read(struct file *in, loff_t *ppos,
 308                            struct pipe_inode_info *pipe, size_t len,
 309                            unsigned int flags)
 310 {
 311         struct address_space *mapping = in->f_mapping;
 312         unsigned int loff, nr_pages, req_pages;
 313         struct page *pages[PIPE_DEF_BUFFERS];
 314         struct partial_page partial[PIPE_DEF_BUFFERS];
 315         struct page *page;
 316         pgoff_t index, end_index;
 317         loff_t isize;
 318         int error, page_nr;
 319         struct splice_pipe_desc spd = {
 320                 .pages = pages,
 321                 .partial = partial,
 322                 .nr_pages_max = PIPE_DEF_BUFFERS,
 323                 .flags = flags,
 324                 .ops = &page_cache_pipe_buf_ops,
 325                 .spd_release = spd_release_page,
 326         };
 327
 328         if (splice_grow_spd(pipe, &spd))
 329                 return -ENOMEM;
 330
 331         index = *ppos >> PAGE_CACHE_SHIFT;
 332         loff = *ppos & ~PAGE_CACHE_MASK;
 333         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 334         nr_pages = min(req_pages, spd.nr_pages_max);
 335
 336         /*
 337          * Lookup the (hopefully) full range of pages we need.
 338          */
 339         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 340         index += spd.nr_pages;
 341
 342         /*
 343          * If find_get_pages_contig() returned fewer pages than we needed,
 344          * readahead/allocate the rest and fill in the holes.
 345          */
 346         if (spd.nr_pages < nr_pages)
 347                 page_cache_sync_readahead(mapping, &in->f_ra, in,
 348                                 index, req_pages - spd.nr_pages);
 349
 350         error = 0;
 351         while (spd.nr_pages < nr_pages) {
 352                 /*
 353                  * Page could be there, find_get_pages_contig() breaks on
 354                  * the first hole.
 355                  */
 356                 page = find_get_page(mapping, index);
 357                 if (!page) {
 358                         /*
 359                          * page didn't exist, allocate one.
 360                          */
 361                         page = page_cache_alloc_cold(mapping);
 362                         if (!page)
 363                                 break;
 364
 365                         error = add_to_page_cache_lru(page, mapping, index,
 366                                                 GFP_KERNEL);
 367                         if (unlikely(error)) {
 368                                 page_cache_release(page);
 369                                 if (error == -EEXIST)
 370                                         continue;
 371                                 break;
 372                         }
 373                         /*
 374                          * add_to_page_cache() locks the page, unlock it
 375                          * to avoid convoluting the logic below even more.
 376                          */
 377                         unlock_page(page);
 378                 }
 379
 380                 spd.pages[spd.nr_pages++] = page;
 381                 index++;
 382         }
 383
 384         /*
 385          * Now loop over the map and see if we need to start IO on any
 386          * pages, fill in the partial map, etc.
 387          */
 388         index = *ppos >> PAGE_CACHE_SHIFT;
 389         nr_pages = spd.nr_pages;
 390         spd.nr_pages = 0;
 391         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 392                 unsigned int this_len;
 393
 394                 if (!len)
 395                         break;
 396
 397                 /*
 398                  * this_len is the max we'll use from this page
 399                  */
 400                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 401                 page = spd.pages[page_nr];
 402
 403                 if (PageReadahead(page))
 404                         page_cache_async_readahead(mapping, &in->f_ra, in,
 405                                         page, index, req_pages - page_nr);
 406
 407                 /*
 408                  * If the page isn't uptodate, we may need to start io on it
 409                  */
 410                 if (!PageUptodate(page)) {
 411                         lock_page(page);
 412
 413                         /*
 414                          * Page was truncated, or invalidated by the
 415                          * filesystem.  Redo the find/create, but this time the
 416                          * page is kept locked, so there's no chance of another
 417                          * race with truncate/invalidate.
 418                          */
 419                         if (!page->mapping) {
 420                                 unlock_page(page);
 421                                 page = find_or_create_page(mapping, index,
 422                                                 mapping_gfp_mask(mapping));
 423
 424                                 if (!page) {
 425                                         error = -ENOMEM;
 426                                         break;
 427                                 }
 428                                 page_cache_release(spd.pages[page_nr]);
 429                                 spd.pages[page_nr] = page;
 430                         }
 431                         /*
 432                          * page was already under io and is now done, great
 433                          */
 434                         if (PageUptodate(page)) {
 435                                 unlock_page(page);
 436                                 goto fill_it;
 437                         }
 438
 439                         /*
 440                          * need to read in the page
 441                          */
 442                         error = mapping->a_ops->readpage(in, page);
 443                         if (unlikely(error)) {
 444                                 /*
 445                                  * We really should re-lookup the page here,
 446                                  * but it complicates things a lot. Instead
 447                                  * lets just do what we already stored, and
 448                                  * we'll get it the next time we are called.
 449                                  */
 450                                 if (error == AOP_TRUNCATED_PAGE)
 451                                         error = 0;
 452
 453                                 break;
 454                         }
 455                 }
 456 fill_it:
 457                 /*
 458                  * i_size must be checked after PageUptodate.
 459                  */
 460                 isize = i_size_read(mapping->host);
 461                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 462                 if (unlikely(!isize || index > end_index))
 463                         break;
 464
 465                 /*
 466                  * if this is the last page, see if we need to shrink
 467                  * the length and stop
 468                  */
 469                 if (end_index == index) {
 470                         unsigned int plen;
 471
 472                         /*
 473                          * max good bytes in this page
 474                          */
 475                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 476                         if (plen <= loff)
 477                                 break;
 478
 479                         /*
 480                          * force quit after adding this page
 481                          */
 482                         this_len = min(this_len, plen - loff);
 483                         len = this_len;
 484                 }
 485
 486                 spd.partial[page_nr].offset = loff;
 487                 spd.partial[page_nr].len = this_len;
 488                 len -= this_len;
 489                 loff = 0;
 490                 spd.nr_pages++;
 491                 index++;
 492         }
 493
 494         /*
 495          * Release any pages at the end, if we quit early. 'page_nr' is how far
 496          * we got, 'nr_pages' is how many pages are in the map.
 497          */
 498         while (page_nr < nr_pages)
 499                 page_cache_release(spd.pages[page_nr++]);
 500         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 501
 502         if (spd.nr_pages)
 503                 error = splice_to_pipe(pipe, &spd);
 504
 505         splice_shrink_spd(&spd);
 506         return error;
 507 }
 508
 509 /**
 510  * generic_file_splice_read - splice data from file to a pipe
 511  * @in:         file to splice from
 512  * @ppos:       position in @in
 513  * @pipe:       pipe to splice to
 514  * @len:        number of bytes to splice
 515  * @flags:      splice modifier flags
 516  *
 517  * Description:
 518  *    Will read pages from given file and fill them into a pipe. Can be
 519  *    used as long as the address_space operations for the source implements
 520  *    a readpage() hook.
 521  *
 522  */
 523 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 524                                  struct pipe_inode_info *pipe, size_t len,
 525                                  unsigned int flags)
 526 {
 527         loff_t isize, left;
 528         int ret;
 529
 530         isize = i_size_read(in->f_mapping->host);
 531         if (unlikely(*ppos >= isize))
 532                 return 0;
 533
 534         left = isize - *ppos;
 535         if (unlikely(left < len))
 536                 len = left;
 537
 538         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 539         if (ret > 0) {
 540                 *ppos += ret;
 541                 file_accessed(in);
 542         }
 543
 544         return ret;
 545 }
 546 EXPORT_SYMBOL(generic_file_splice_read);
 547
 548 static const struct pipe_buf_operations default_pipe_buf_ops = {
 549         .can_merge = 0,
 550         .map = generic_pipe_buf_map,
 551         .unmap = generic_pipe_buf_unmap,
 552         .confirm = generic_pipe_buf_confirm,
 553         .release = generic_pipe_buf_release,
 554         .steal = generic_pipe_buf_steal,
 555         .get = generic_pipe_buf_get,
 556 };
 557
 558 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 559                                     struct pipe_buffer *buf)
 560 {
 561         return 1;
 562 }
 563
 564 /* Pipe buffer operations for a socket and similar. */
 565 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 566         .can_merge = 0,
 567         .map = generic_pipe_buf_map,
 568         .unmap = generic_pipe_buf_unmap,
 569         .confirm = generic_pipe_buf_confirm,
 570         .release = generic_pipe_buf_release,
 571         .steal = generic_pipe_buf_nosteal,
 572         .get = generic_pipe_buf_get,
 573 };
 574 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 575
 576 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 577                             unsigned long vlen, loff_t offset)
 578 {
 579         mm_segment_t old_fs;
 580         loff_t pos = offset;
 581         ssize_t res;
 582
 583         old_fs = get_fs();
 584         set_fs(get_ds());
 585         /* The cast to a user pointer is valid due to the set_fs() */
 586         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 587         set_fs(old_fs);
 588
 589         return res;
 590 }
 591
 592 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 593                             loff_t pos)
 594 {
 595         mm_segment_t old_fs;
 596         ssize_t res;
 597
 598         old_fs = get_fs();
 599         set_fs(get_ds());
 600         /* The cast to a user pointer is valid due to the set_fs() */
 601         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 602         set_fs(old_fs);
 603
 604         return res;
 605 }
 606 EXPORT_SYMBOL(kernel_write);
 607
 608 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 609                                  struct pipe_inode_info *pipe, size_t len,
 610                                  unsigned int flags)
 611 {
 612         unsigned int nr_pages;
 613         unsigned int nr_freed;
 614         size_t offset;
 615         struct page *pages[PIPE_DEF_BUFFERS];
 616         struct partial_page partial[PIPE_DEF_BUFFERS];
 617         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 618         ssize_t res;
 619         size_t this_len;
 620         int error;
 621         int i;
 622         struct splice_pipe_desc spd = {
 623                 .pages = pages,
 624                 .partial = partial,
 625                 .nr_pages_max = PIPE_DEF_BUFFERS,
 626                 .flags = flags,
 627                 .ops = &default_pipe_buf_ops,
 628                 .spd_release = spd_release_page,
 629         };
 630
 631         if (splice_grow_spd(pipe, &spd))
 632                 return -ENOMEM;
 633
 634         res = -ENOMEM;
 635         vec = __vec;
 636         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 637                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 638                 if (!vec)
 639                         goto shrink_ret;
 640         }
 641
 642         offset = *ppos & ~PAGE_CACHE_MASK;
 643         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 644
 645         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 646                 struct page *page;
 647
 648                 page = alloc_page(GFP_USER);
 649                 error = -ENOMEM;
 650                 if (!page)
 651                         goto err;
 652
 653                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 654                 vec[i].iov_base = (void __user *) page_address(page);
 655                 vec[i].iov_len = this_len;
 656                 spd.pages[i] = page;
 657                 spd.nr_pages++;
 658                 len -= this_len;
 659                 offset = 0;
 660         }
 661
 662         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 663         if (res < 0) {
 664                 error = res;
 665                 goto err;
 666         }
 667
 668         error = 0;
 669         if (!res)
 670                 goto err;
 671
 672         nr_freed = 0;
 673         for (i = 0; i < spd.nr_pages; i++) {
 674                 this_len = min_t(size_t, vec[i].iov_len, res);
 675                 spd.partial[i].offset = 0;
 676                 spd.partial[i].len = this_len;
 677                 if (!this_len) {
 678                         __free_page(spd.pages[i]);
 679                         spd.pages[i] = NULL;
 680                         nr_freed++;
 681                 }
 682                 res -= this_len;
 683         }
 684         spd.nr_pages -= nr_freed;
 685
 686         res = splice_to_pipe(pipe, &spd);
 687         if (res > 0)
 688                 *ppos += res;
 689
 690 shrink_ret:
 691         if (vec != __vec)
 692                 kfree(vec);
 693         splice_shrink_spd(&spd);
 694         return res;
 695
 696 err:
 697         for (i = 0; i < spd.nr_pages; i++)
 698                 __free_page(spd.pages[i]);
 699
 700         res = error;
 701         goto shrink_ret;
 702 }
 703 EXPORT_SYMBOL(default_file_splice_read);
 704
 705 /*
 706  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 707  * using sendpage(). Return the number of bytes sent.
 708  */
 709 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 710                             struct pipe_buffer *buf, struct splice_desc *sd)
 711 {
 712         struct file *file = sd->u.file;
 713         loff_t pos = sd->pos;
 714         int more;
 715
 716         if (!likely(file->f_op->sendpage))
 717                 return -EINVAL;
 718
 719         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 720
 721         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 722                 more |= MSG_SENDPAGE_NOTLAST;
 723
 724         return file->f_op->sendpage(file, buf->page, buf->offset,
 725                                     sd->len, &pos, more);
 726 }
 727
 728 /*
 729  * This is a little more tricky than the file -> pipe splicing. There are
 730  * basically three cases:
 731  *
 732  *      - Destination page already exists in the address space and there
 733  *        are users of it. For that case we have no other option that
 734  *        copying the data. Tough luck.
 735  *      - Destination page already exists in the address space, but there
 736  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 737  *        through to last case.
 738  *      - Destination page does not exist, we can add the pipe page to
 739  *        the page cache and avoid the copy.
 740  *
 741  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 742  * sd->flags), we attempt to migrate pages from the pipe to the output
 743  * file address space page cache. This is possible if no one else has
 744  * the pipe page referenced outside of the pipe and page cache. If
 745  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 746  * a new page in the output file page cache and fill/dirty that.
 747  */
 748 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 749                  struct splice_desc *sd)
 750 {
 751         struct file *file = sd->u.file;
 752         struct address_space *mapping = file->f_mapping;
 753         unsigned int offset, this_len;
 754         struct page *page;
 755         void *fsdata;
 756         int ret;
 757
 758         offset = sd->pos & ~PAGE_CACHE_MASK;
 759
 760         this_len = sd->len;
 761         if (this_len + offset > PAGE_CACHE_SIZE)
 762                 this_len = PAGE_CACHE_SIZE - offset;
 763
 764         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 765                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 766         if (unlikely(ret))
 767                 goto out;
 768
 769         if (buf->page != page) {
 770                 char *src = buf->ops->map(pipe, buf, 1);
 771                 char *dst = kmap_atomic(page);
 772
 773                 memcpy(dst + offset, src + buf->offset, this_len);
 774                 flush_dcache_page(page);
 775                 kunmap_atomic(dst);
 776                 buf->ops->unmap(pipe, buf, src);
 777         }
 778         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 779                                 page, fsdata);
 780 out:
 781         return ret;
 782 }
 783 EXPORT_SYMBOL(pipe_to_file);
 784
 785 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 786 {
 787         smp_mb();
 788         if (waitqueue_active(&pipe->wait))
 789                 wake_up_interruptible(&pipe->wait);
 790         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 791 }
 792
 793 /**
 794  * splice_from_pipe_feed - feed available data from a pipe to a file
 795  * @pipe:       pipe to splice from
 796  * @sd:         information to @actor
 797  * @actor:      handler that splices the data
 798  *
 799  * Description:
 800  *    This function loops over the pipe and calls @actor to do the
 801  *    actual moving of a single struct pipe_buffer to the desired
 802  *    destination.  It returns when there's no more buffers left in
 803  *    the pipe or if the requested number of bytes (@sd->total_len)
 804  *    have been copied.  It returns a positive number (one) if the
 805  *    pipe needs to be filled with more data, zero if the required
 806  *    number of bytes have been copied and -errno on error.
 807  *
 808  *    This, together with splice_from_pipe_{begin,end,next}, may be
 809  *    used to implement the functionality of __splice_from_pipe() when
 810  *    locking is required around copying the pipe buffers to the
 811  *    destination.
 812  */
 813 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 814                           splice_actor *actor)
 815 {
 816         int ret;
 817
 818         while (pipe->nrbufs) {
 819                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 820                 const struct pipe_buf_operations *ops = buf->ops;
 821
 822                 sd->len = buf->len;
 823                 if (sd->len > sd->total_len)
 824                         sd->len = sd->total_len;
 825
 826                 ret = buf->ops->confirm(pipe, buf);
 827                 if (unlikely(ret)) {
 828                         if (ret == -ENODATA)
 829                                 ret = 0;
 830                         return ret;
 831                 }
 832
 833                 ret = actor(pipe, buf, sd);
 834                 if (ret <= 0)
 835                         return ret;
 836
 837                 buf->offset += ret;
 838                 buf->len -= ret;
 839
 840                 sd->num_spliced += ret;
 841                 sd->len -= ret;
 842                 sd->pos += ret;
 843                 sd->total_len -= ret;
 844
 845                 if (!buf->len) {
 846                         buf->ops = NULL;
 847                         ops->release(pipe, buf);
 848                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 849                         pipe->nrbufs--;
 850                         if (pipe->files)
 851                                 sd->need_wakeup = true;
 852                 }
 853
 854                 if (!sd->total_len)
 855                         return 0;
 856         }
 857
 858         return 1;
 859 }
 860 EXPORT_SYMBOL(splice_from_pipe_feed);
 861
 862 /**
 863  * splice_from_pipe_next - wait for some data to splice from
 864  * @pipe:       pipe to splice from
 865  * @sd:         information about the splice operation
 866  *
 867  * Description:
 868  *    This function will wait for some data and return a positive
 869  *    value (one) if pipe buffers are available.  It will return zero
 870  *    or -errno if no more data needs to be spliced.
 871  */
 872 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 873 {
 874         while (!pipe->nrbufs) {
 875                 if (!pipe->writers)
 876                         return 0;
 877
 878                 if (!pipe->waiting_writers && sd->num_spliced)
 879                         return 0;
 880
 881                 if (sd->flags & SPLICE_F_NONBLOCK)
 882                         return -EAGAIN;
 883
 884                 if (signal_pending(current))
 885                         return -ERESTARTSYS;
 886
 887                 if (sd->need_wakeup) {
 888                         wakeup_pipe_writers(pipe);
 889                         sd->need_wakeup = false;
 890                 }
 891
 892                 pipe_wait(pipe);
 893         }
 894
 895         return 1;
 896 }
 897 EXPORT_SYMBOL(splice_from_pipe_next);
 898
 899 /**
 900  * splice_from_pipe_begin - start splicing from pipe
 901  * @sd:         information about the splice operation
 902  *
 903  * Description:
 904  *    This function should be called before a loop containing
 905  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 906  *    initialize the necessary fields of @sd.
 907  */
 908 void splice_from_pipe_begin(struct splice_desc *sd)
 909 {
 910         sd->num_spliced = 0;
 911         sd->need_wakeup = false;
 912 }
 913 EXPORT_SYMBOL(splice_from_pipe_begin);
 914
 915 /**
 916  * splice_from_pipe_end - finish splicing from pipe
 917  * @pipe:       pipe to splice from
 918  * @sd:         information about the splice operation
 919  *
 920  * Description:
 921  *    This function will wake up pipe writers if necessary.  It should
 922  *    be called after a loop containing splice_from_pipe_next() and
 923  *    splice_from_pipe_feed().
 924  */
 925 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 926 {
 927         if (sd->need_wakeup)
 928                 wakeup_pipe_writers(pipe);
 929 }
 930 EXPORT_SYMBOL(splice_from_pipe_end);
 931
 932 /**
 933  * __splice_from_pipe - splice data from a pipe to given actor
 934  * @pipe:       pipe to splice from
 935  * @sd:         information to @actor
 936  * @actor:      handler that splices the data
 937  *
 938  * Description:
 939  *    This function does little more than loop over the pipe and call
 940  *    @actor to do the actual moving of a single struct pipe_buffer to
 941  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 942  *    pipe_to_user.
 943  *
 944  */
 945 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 946                            splice_actor *actor)
 947 {
 948         int ret;
 949
 950         splice_from_pipe_begin(sd);
 951         do {
 952                 ret = splice_from_pipe_next(pipe, sd);
 953                 if (ret > 0)
 954                         ret = splice_from_pipe_feed(pipe, sd, actor);
 955         } while (ret > 0);
 956         splice_from_pipe_end(pipe, sd);
 957
 958         return sd->num_spliced ? sd->num_spliced : ret;
 959 }
 960 EXPORT_SYMBOL(__splice_from_pipe);
 961
 962 /**
 963  * splice_from_pipe - splice data from a pipe to a file
 964  * @pipe:       pipe to splice from
 965  * @out:        file to splice to
 966  * @ppos:       position in @out
 967  * @len:        how many bytes to splice
 968  * @flags:      splice modifier flags
 969  * @actor:      handler that splices the data
 970  *
 971  * Description:
 972  *    See __splice_from_pipe. This function locks the pipe inode,
 973  *    otherwise it's identical to __splice_from_pipe().
 974  *
 975  */
 976 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 977                          loff_t *ppos, size_t len, unsigned int flags,
 978                          splice_actor *actor)
 979 {
 980         ssize_t ret;
 981         struct splice_desc sd = {
 982                 .total_len = len,
 983                 .flags = flags,
 984                 .pos = *ppos,
 985                 .u.file = out,
 986         };
 987
 988         pipe_lock(pipe);
 989         ret = __splice_from_pipe(pipe, &sd, actor);
 990         pipe_unlock(pipe);
 991
 992         return ret;
 993 }
 994
 995 /**
 996  * generic_file_splice_write - splice data from a pipe to a file
 997  * @pipe:       pipe info
 998  * @out:        file to write to
 999  * @ppos:       position in @out
1000  * @len:        number of bytes to splice
1001  * @flags:      splice modifier flags
1002  *
1003  * Description:
1004  *    Will either move or copy pages (determined by @flags options) from
1005  *    the given pipe inode to the given file.
1006  *
1007  */
1008 ssize_t
1009 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1010                           loff_t *ppos, size_t len, unsigned int flags)
1011 {
1012         struct address_space *mapping = out->f_mapping;
1013         struct inode *inode = mapping->host;
1014         struct splice_desc sd = {
1015                 .total_len = len,
1016                 .flags = flags,
1017                 .pos = *ppos,
1018                 .u.file = out,
1019         };
1020         ssize_t ret;
1021
1022         pipe_lock(pipe);
1023
1024         splice_from_pipe_begin(&sd);
1025         do {
1026                 ret = splice_from_pipe_next(pipe, &sd);
1027                 if (ret <= 0)
1028                         break;
1029
1030                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1031                 ret = file_remove_suid(out);
1032                 if (!ret) {
1033                         ret = file_update_time(out);
1034                         if (!ret)
1035                                 ret = splice_from_pipe_feed(pipe, &sd,
1036                                                             pipe_to_file);
1037                 }
1038                 mutex_unlock(&inode->i_mutex);
1039         } while (ret > 0);
1040         splice_from_pipe_end(pipe, &sd);
1041
1042         pipe_unlock(pipe);
1043
1044         if (sd.num_spliced)
1045                 ret = sd.num_spliced;
1046
1047         if (ret > 0) {
1048                 int err;
1049
1050                 err = generic_write_sync(out, *ppos, ret);
1051                 if (err)
1052                         ret = err;
1053                 else
1054                         *ppos += ret;
1055                 balance_dirty_pages_ratelimited(mapping);
1056         }
1057
1058         return ret;
1059 }
1060
1061 EXPORT_SYMBOL(generic_file_splice_write);
1062
1063 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1064                           struct splice_desc *sd)
1065 {
1066         int ret;
1067         void *data;
1068         loff_t tmp = sd->pos;
1069
1070         data = buf->ops->map(pipe, buf, 0);
1071         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1072         buf->ops->unmap(pipe, buf, data);
1073
1074         return ret;
1075 }
1076
1077 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1078                                          struct file *out, loff_t *ppos,
1079                                          size_t len, unsigned int flags)
1080 {
1081         ssize_t ret;
1082
1083         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1084         if (ret > 0)
1085                 *ppos += ret;
1086
1087         return ret;
1088 }
1089
1090 /**
1091  * generic_splice_sendpage - splice data from a pipe to a socket
1092  * @pipe:       pipe to splice from
1093  * @out:        socket to write to
1094  * @ppos:       position in @out
1095  * @len:        number of bytes to splice
1096  * @flags:      splice modifier flags
1097  *
1098  * Description:
1099  *    Will send @len bytes from the pipe to a network socket. No data copying
1100  *    is involved.
1101  *
1102  */
1103 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1104                                 loff_t *ppos, size_t len, unsigned int flags)
1105 {
1106         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1107 }
1108
1109 EXPORT_SYMBOL(generic_splice_sendpage);
1110
1111 /*
1112  * Attempt to initiate a splice from pipe to file.
1113  */
1114 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1115                            loff_t *ppos, size_t len, unsigned int flags)
1116 {
1117         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1118                                 loff_t *, size_t, unsigned int);
1119
1120         if (out->f_op->splice_write)
1121                 splice_write = out->f_op->splice_write;
1122         else
1123                 splice_write = default_file_splice_write;
1124
1125         return splice_write(pipe, out, ppos, len, flags);
1126 }
1127
1128 /*
1129  * Attempt to initiate a splice from a file to a pipe.
1130  */
1131 static long do_splice_to(struct file *in, loff_t *ppos,
1132                          struct pipe_inode_info *pipe, size_t len,
1133                          unsigned int flags)
1134 {
1135         ssize_t (*splice_read)(struct file *, loff_t *,
1136                                struct pipe_inode_info *, size_t, unsigned int);
1137         int ret;
1138
1139         if (unlikely(!(in->f_mode & FMODE_READ)))
1140                 return -EBADF;
1141
1142         ret = rw_verify_area(READ, in, ppos, len);
1143         if (unlikely(ret < 0))
1144                 return ret;
1145
1146         if (in->f_op->splice_read)
1147                 splice_read = in->f_op->splice_read;
1148         else
1149                 splice_read = default_file_splice_read;
1150
1151         return splice_read(in, ppos, pipe, len, flags);
1152 }
1153
1154 /**
1155  * splice_direct_to_actor - splices data directly between two non-pipes
1156  * @in:         file to splice from
1157  * @sd:         actor information on where to splice to
1158  * @actor:      handles the data splicing
1159  *
1160  * Description:
1161  *    This is a special case helper to splice directly between two
1162  *    points, without requiring an explicit pipe. Internally an allocated
1163  *    pipe is cached in the process, and reused during the lifetime of
1164  *    that process.
1165  *
1166  */
1167 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1168                                splice_direct_actor *actor)
1169 {
1170         struct pipe_inode_info *pipe;
1171         long ret, bytes;
1172         umode_t i_mode;
1173         size_t len;
1174         int i, flags;
1175
1176         /*
1177          * We require the input being a regular file, as we don't want to
1178          * randomly drop data for eg socket -> socket splicing. Use the
1179          * piped splicing for that!
1180          */
1181         i_mode = file_inode(in)->i_mode;
1182         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1183                 return -EINVAL;
1184
1185         /*
1186          * neither in nor out is a pipe, setup an internal pipe attached to
1187          * 'out' and transfer the wanted data from 'in' to 'out' through that
1188          */
1189         pipe = current->splice_pipe;
1190         if (unlikely(!pipe)) {
1191                 pipe = alloc_pipe_info();
1192                 if (!pipe)
1193                         return -ENOMEM;
1194
1195                 /*
1196                  * We don't have an immediate reader, but we'll read the stuff
1197                  * out of the pipe right after the splice_to_pipe(). So set
1198                  * PIPE_READERS appropriately.
1199                  */
1200                 pipe->readers = 1;
1201
1202                 current->splice_pipe = pipe;
1203         }
1204
1205         /*
1206          * Do the splice.
1207          */
1208         ret = 0;
1209         bytes = 0;
1210         len = sd->total_len;
1211         flags = sd->flags;
1212
1213         /*
1214          * Don't block on output, we have to drain the direct pipe.
1215          */
1216         sd->flags &= ~SPLICE_F_NONBLOCK;
1217
1218         while (len) {
1219                 size_t read_len;
1220                 loff_t pos = sd->pos, prev_pos = pos;
1221
1222                 ret = do_splice_to(in, &pos, pipe, len, flags);
1223                 if (unlikely(ret <= 0))
1224                         goto out_release;
1225
1226                 read_len = ret;
1227                 sd->total_len = read_len;
1228
1229                 /*
1230                  * NOTE: nonblocking mode only applies to the input. We
1231                  * must not do the output in nonblocking mode as then we
1232                  * could get stuck data in the internal pipe:
1233                  */
1234                 ret = actor(pipe, sd);
1235                 if (unlikely(ret <= 0)) {
1236                         sd->pos = prev_pos;
1237                         goto out_release;
1238                 }
1239
1240                 bytes += ret;
1241                 len -= ret;
1242                 sd->pos = pos;
1243
1244                 if (ret < read_len) {
1245                         sd->pos = prev_pos + ret;
1246                         goto out_release;
1247                 }
1248         }
1249
1250 done:
1251         pipe->nrbufs = pipe->curbuf = 0;
1252         file_accessed(in);
1253         return bytes;
1254
1255 out_release:
1256         /*
1257          * If we did an incomplete transfer we must release
1258          * the pipe buffers in question:
1259          */
1260         for (i = 0; i < pipe->buffers; i++) {
1261                 struct pipe_buffer *buf = pipe->bufs + i;
1262
1263                 if (buf->ops) {
1264                         buf->ops->release(pipe, buf);
1265                         buf->ops = NULL;
1266                 }
1267         }
1268
1269         if (!bytes)
1270                 bytes = ret;
1271
1272         goto done;
1273 }
1274 EXPORT_SYMBOL(splice_direct_to_actor);
1275
1276 static int direct_splice_actor(struct pipe_inode_info *pipe,
1277                                struct splice_desc *sd)
1278 {
1279         struct file *file = sd->u.file;
1280
1281         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1282                               sd->flags);
1283 }
1284
1285 /**
1286  * do_splice_direct - splices data directly between two files
1287  * @in:         file to splice from
1288  * @ppos:       input file offset
1289  * @out:        file to splice to
1290  * @opos:       output file offset
1291  * @len:        number of bytes to splice
1292  * @flags:      splice modifier flags
1293  *
1294  * Description:
1295  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1296  *    doing it in the application would incur an extra system call
1297  *    (splice in + splice out, as compared to just sendfile()). So this helper
1298  *    can splice directly through a process-private pipe.
1299  *
1300  */
1301 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1302                       loff_t *opos, size_t len, unsigned int flags)
1303 {
1304         struct splice_desc sd = {
1305                 .len            = len,
1306                 .total_len      = len,
1307                 .flags          = flags,
1308                 .pos            = *ppos,
1309                 .u.file         = out,
1310                 .opos           = opos,
1311         };
1312         long ret;
1313
1314         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1315                 return -EBADF;
1316
1317         if (unlikely(out->f_flags & O_APPEND))
1318                 return -EINVAL;
1319
1320         ret = rw_verify_area(WRITE, out, opos, len);
1321         if (unlikely(ret < 0))
1322                 return ret;
1323
1324         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1325         if (ret > 0)
1326                 *ppos = sd.pos;
1327
1328         return ret;
1329 }
1330
1331 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1332                                struct pipe_inode_info *opipe,
1333                                size_t len, unsigned int flags);
1334
1335 /*
1336  * Determine where to splice to/from.
1337  */
1338 static long do_splice(struct file *in, loff_t __user *off_in,
1339                       struct file *out, loff_t __user *off_out,
1340                       size_t len, unsigned int flags)
1341 {
1342         struct pipe_inode_info *ipipe;
1343         struct pipe_inode_info *opipe;
1344         loff_t offset;
1345         long ret;
1346
1347         ipipe = get_pipe_info(in);
1348         opipe = get_pipe_info(out);
1349
1350         if (ipipe && opipe) {
1351                 if (off_in || off_out)
1352                         return -ESPIPE;
1353
1354                 if (!(in->f_mode & FMODE_READ))
1355                         return -EBADF;
1356
1357                 if (!(out->f_mode & FMODE_WRITE))
1358                         return -EBADF;
1359
1360                 /* Splicing to self would be fun, but... */
1361                 if (ipipe == opipe)
1362                         return -EINVAL;
1363
1364                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1365         }
1366
1367         if (ipipe) {
1368                 if (off_in)
1369                         return -ESPIPE;
1370                 if (off_out) {
1371                         if (!(out->f_mode & FMODE_PWRITE))
1372                                 return -EINVAL;
1373                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1374                                 return -EFAULT;
1375                 } else {
1376                         offset = out->f_pos;
1377                 }
1378
1379                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1380                         return -EBADF;
1381
1382                 if (unlikely(out->f_flags & O_APPEND))
1383                         return -EINVAL;
1384
1385                 ret = rw_verify_area(WRITE, out, &offset, len);
1386                 if (unlikely(ret < 0))
1387                         return ret;
1388
1389                 file_start_write(out);
1390                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1391                 file_end_write(out);
1392
1393                 if (!off_out)
1394                         out->f_pos = offset;
1395                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1396                         ret = -EFAULT;
1397
1398                 return ret;
1399         }
1400
1401         if (opipe) {
1402                 if (off_out)
1403                         return -ESPIPE;
1404                 if (off_in) {
1405                         if (!(in->f_mode & FMODE_PREAD))
1406                                 return -EINVAL;
1407                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1408                                 return -EFAULT;
1409                 } else {
1410                         offset = in->f_pos;
1411                 }
1412
1413                 ret = do_splice_to(in, &offset, opipe, len, flags);
1414
1415                 if (!off_in)
1416                         in->f_pos = offset;
1417                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1418                         ret = -EFAULT;
1419
1420                 return ret;
1421         }
1422
1423         return -EINVAL;
1424 }
1425
1426 /*
1427  * Map an iov into an array of pages and offset/length tupples. With the
1428  * partial_page structure, we can map several non-contiguous ranges into
1429  * our ones pages[] map instead of splitting that operation into pieces.
1430  * Could easily be exported as a generic helper for other users, in which
1431  * case one would probably want to add a 'max_nr_pages' parameter as well.
1432  */
1433 static int get_iovec_page_array(const struct iovec __user *iov,
1434                                 unsigned int nr_vecs, struct page **pages,
1435                                 struct partial_page *partial, bool aligned,
1436                                 unsigned int pipe_buffers)
1437 {
1438         int buffers = 0, error = 0;
1439
1440         while (nr_vecs) {
1441                 unsigned long off, npages;
1442                 struct iovec entry;
1443                 void __user *base;
1444                 size_t len;
1445                 int i;
1446
1447                 error = -EFAULT;
1448                 if (copy_from_user(&entry, iov, sizeof(entry)))
1449                         break;
1450
1451                 base = entry.iov_base;
1452                 len = entry.iov_len;
1453
1454                 /*
1455                  * Sanity check this iovec. 0 read succeeds.
1456                  */
1457                 error = 0;
1458                 if (unlikely(!len))
1459                         break;
1460                 error = -EFAULT;
1461                 if (!access_ok(VERIFY_READ, base, len))
1462                         break;
1463
1464                 /*
1465                  * Get this base offset and number of pages, then map
1466                  * in the user pages.
1467                  */
1468                 off = (unsigned long) base & ~PAGE_MASK;
1469
1470                 /*
1471                  * If asked for alignment, the offset must be zero and the
1472                  * length a multiple of the PAGE_SIZE.
1473                  */
1474                 error = -EINVAL;
1475                 if (aligned && (off || len & ~PAGE_MASK))
1476                         break;
1477
1478                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1479                 if (npages > pipe_buffers - buffers)
1480                         npages = pipe_buffers - buffers;
1481
1482                 error = get_user_pages_fast((unsigned long)base, npages,
1483                                         0, &pages[buffers]);
1484
1485                 if (unlikely(error <= 0))
1486                         break;
1487
1488                 /*
1489                  * Fill this contiguous range into the partial page map.
1490                  */
1491                 for (i = 0; i < error; i++) {
1492                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1493
1494                         partial[buffers].offset = off;
1495                         partial[buffers].len = plen;
1496
1497                         off = 0;
1498                         len -= plen;
1499                         buffers++;
1500                 }
1501
1502                 /*
1503                  * We didn't complete this iov, stop here since it probably
1504                  * means we have to move some of this into a pipe to
1505                  * be able to continue.
1506                  */
1507                 if (len)
1508                         break;
1509
1510                 /*
1511                  * Don't continue if we mapped fewer pages than we asked for,
1512                  * or if we mapped the max number of pages that we have
1513                  * room for.
1514                  */
1515                 if (error < npages || buffers == pipe_buffers)
1516                         break;
1517
1518                 nr_vecs--;
1519                 iov++;
1520         }
1521
1522         if (buffers)
1523                 return buffers;
1524
1525         return error;
1526 }
1527
1528 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1529                         struct splice_desc *sd)
1530 {
1531         char *src;
1532         int ret;
1533
1534         /*
1535          * See if we can use the atomic maps, by prefaulting in the
1536          * pages and doing an atomic copy
1537          */
1538         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1539                 src = buf->ops->map(pipe, buf, 1);
1540                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1541                                                         sd->len);
1542                 buf->ops->unmap(pipe, buf, src);
1543                 if (!ret) {
1544                         ret = sd->len;
1545                         goto out;
1546                 }
1547         }
1548
1549         /*
1550          * No dice, use slow non-atomic map and copy
1551          */
1552         src = buf->ops->map(pipe, buf, 0);
1553
1554         ret = sd->len;
1555         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1556                 ret = -EFAULT;
1557
1558         buf->ops->unmap(pipe, buf, src);
1559 out:
1560         if (ret > 0)
1561                 sd->u.userptr += ret;
1562         return ret;
1563 }
1564
1565 /*
1566  * For lack of a better implementation, implement vmsplice() to userspace
1567  * as a simple copy of the pipes pages to the user iov.
1568  */
1569 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1570                              unsigned long nr_segs, unsigned int flags)
1571 {
1572         struct pipe_inode_info *pipe;
1573         struct splice_desc sd;
1574         ssize_t size;
1575         int error;
1576         long ret;
1577
1578         pipe = get_pipe_info(file);
1579         if (!pipe)
1580                 return -EBADF;
1581
1582         pipe_lock(pipe);
1583
1584         error = ret = 0;
1585         while (nr_segs) {
1586                 void __user *base;
1587                 size_t len;
1588
1589                 /*
1590                  * Get user address base and length for this iovec.
1591                  */
1592                 error = get_user(base, &iov->iov_base);
1593                 if (unlikely(error))
1594                         break;
1595                 error = get_user(len, &iov->iov_len);
1596                 if (unlikely(error))
1597                         break;
1598
1599                 /*
1600                  * Sanity check this iovec. 0 read succeeds.
1601                  */
1602                 if (unlikely(!len))
1603                         break;
1604                 if (unlikely(!base)) {
1605                         error = -EFAULT;
1606                         break;
1607                 }
1608
1609                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1610                         error = -EFAULT;
1611                         break;
1612                 }
1613
1614                 sd.len = 0;
1615                 sd.total_len = len;
1616                 sd.flags = flags;
1617                 sd.u.userptr = base;
1618                 sd.pos = 0;
1619
1620                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1621                 if (size < 0) {
1622                         if (!ret)
1623                                 ret = size;
1624
1625                         break;
1626                 }
1627
1628                 ret += size;
1629
1630                 if (size < len)
1631                         break;
1632
1633                 nr_segs--;
1634                 iov++;
1635         }
1636
1637         pipe_unlock(pipe);
1638
1639         if (!ret)
1640                 ret = error;
1641
1642         return ret;
1643 }
1644
1645 /*
1646  * vmsplice splices a user address range into a pipe. It can be thought of
1647  * as splice-from-memory, where the regular splice is splice-from-file (or
1648  * to file). In both cases the output is a pipe, naturally.
1649  */
1650 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1651                              unsigned long nr_segs, unsigned int flags)
1652 {
1653         struct pipe_inode_info *pipe;
1654         struct page *pages[PIPE_DEF_BUFFERS];
1655         struct partial_page partial[PIPE_DEF_BUFFERS];
1656         struct splice_pipe_desc spd = {
1657                 .pages = pages,
1658                 .partial = partial,
1659                 .nr_pages_max = PIPE_DEF_BUFFERS,
1660                 .flags = flags,
1661                 .ops = &user_page_pipe_buf_ops,
1662                 .spd_release = spd_release_page,
1663         };
1664         long ret;
1665
1666         pipe = get_pipe_info(file);
1667         if (!pipe)
1668                 return -EBADF;
1669
1670         if (splice_grow_spd(pipe, &spd))
1671                 return -ENOMEM;
1672
1673         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1674                                             spd.partial, false,
1675                                             spd.nr_pages_max);
1676         if (spd.nr_pages <= 0)
1677                 ret = spd.nr_pages;
1678         else
1679                 ret = splice_to_pipe(pipe, &spd);
1680
1681         splice_shrink_spd(&spd);
1682         return ret;
1683 }
1684
1685 /*
1686  * Note that vmsplice only really supports true splicing _from_ user memory
1687  * to a pipe, not the other way around. Splicing from user memory is a simple
1688  * operation that can be supported without any funky alignment restrictions
1689  * or nasty vm tricks. We simply map in the user memory and fill them into
1690  * a pipe. The reverse isn't quite as easy, though. There are two possible
1691  * solutions for that:
1692  *
1693  *      - memcpy() the data internally, at which point we might as well just
1694  *        do a regular read() on the buffer anyway.
1695  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1696  *        has restriction limitations on both ends of the pipe).
1697  *
1698  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1699  *
1700  */
1701 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1702                 unsigned long, nr_segs, unsigned int, flags)
1703 {
1704         struct fd f;
1705         long error;
1706
1707         if (unlikely(nr_segs > UIO_MAXIOV))
1708                 return -EINVAL;
1709         else if (unlikely(!nr_segs))
1710                 return 0;
1711
1712         error = -EBADF;
1713         f = fdget(fd);
1714         if (f.file) {
1715                 if (f.file->f_mode & FMODE_WRITE)
1716                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1717                 else if (f.file->f_mode & FMODE_READ)
1718                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1719
1720                 fdput(f);
1721         }
1722
1723         return error;
1724 }
1725
1726 #ifdef CONFIG_COMPAT
1727 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1728                     unsigned int, nr_segs, unsigned int, flags)
1729 {
1730         unsigned i;
1731         struct iovec __user *iov;
1732         if (nr_segs > UIO_MAXIOV)
1733                 return -EINVAL;
1734         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1735         for (i = 0; i < nr_segs; i++) {
1736                 struct compat_iovec v;
1737                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1738                     get_user(v.iov_len, &iov32[i].iov_len) ||
1739                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1740                     put_user(v.iov_len, &iov[i].iov_len))
1741                         return -EFAULT;
1742         }
1743         return sys_vmsplice(fd, iov, nr_segs, flags);
1744 }
1745 #endif
1746
1747 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1748                 int, fd_out, loff_t __user *, off_out,
1749                 size_t, len, unsigned int, flags)
1750 {
1751         struct fd in, out;
1752         long error;
1753
1754         if (unlikely(!len))
1755                 return 0;
1756
1757         error = -EBADF;
1758         in = fdget(fd_in);
1759         if (in.file) {
1760                 if (in.file->f_mode & FMODE_READ) {
1761                         out = fdget(fd_out);
1762                         if (out.file) {
1763                                 if (out.file->f_mode & FMODE_WRITE)
1764                                         error = do_splice(in.file, off_in,
1765                                                           out.file, off_out,
1766                                                           len, flags);
1767                                 fdput(out);
1768                         }
1769                 }
1770                 fdput(in);
1771         }
1772         return error;
1773 }
1774
1775 /*
1776  * Make sure there's data to read. Wait for input if we can, otherwise
1777  * return an appropriate error.
1778  */
1779 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1780 {
1781         int ret;
1782
1783         /*
1784          * Check ->nrbufs without the inode lock first. This function
1785          * is speculative anyways, so missing one is ok.
1786          */
1787         if (pipe->nrbufs)
1788                 return 0;
1789
1790         ret = 0;
1791         pipe_lock(pipe);
1792
1793         while (!pipe->nrbufs) {
1794                 if (signal_pending(current)) {
1795                         ret = -ERESTARTSYS;
1796                         break;
1797                 }
1798                 if (!pipe->writers)
1799                         break;
1800                 if (!pipe->waiting_writers) {
1801                         if (flags & SPLICE_F_NONBLOCK) {
1802                                 ret = -EAGAIN;
1803                                 break;
1804                         }
1805                 }
1806                 pipe_wait(pipe);
1807         }
1808
1809         pipe_unlock(pipe);
1810         return ret;
1811 }
1812
1813 /*
1814  * Make sure there's writeable room. Wait for room if we can, otherwise
1815  * return an appropriate error.
1816  */
1817 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1818 {
1819         int ret;
1820
1821         /*
1822          * Check ->nrbufs without the inode lock first. This function
1823          * is speculative anyways, so missing one is ok.
1824          */
1825         if (pipe->nrbufs < pipe->buffers)
1826                 return 0;
1827
1828         ret = 0;
1829         pipe_lock(pipe);
1830
1831         while (pipe->nrbufs >= pipe->buffers) {
1832                 if (!pipe->readers) {
1833                         send_sig(SIGPIPE, current, 0);
1834                         ret = -EPIPE;
1835                         break;
1836                 }
1837                 if (flags & SPLICE_F_NONBLOCK) {
1838                         ret = -EAGAIN;
1839                         break;
1840                 }
1841                 if (signal_pending(current)) {
1842                         ret = -ERESTARTSYS;
1843                         break;
1844                 }
1845                 pipe->waiting_writers++;
1846                 pipe_wait(pipe);
1847                 pipe->waiting_writers--;
1848         }
1849
1850         pipe_unlock(pipe);
1851         return ret;
1852 }
1853
1854 /*
1855  * Splice contents of ipipe to opipe.
1856  */
1857 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1858                                struct pipe_inode_info *opipe,
1859                                size_t len, unsigned int flags)
1860 {
1861         struct pipe_buffer *ibuf, *obuf;
1862         int ret = 0, nbuf;
1863         bool input_wakeup = false;
1864
1865
1866 retry:
1867         ret = ipipe_prep(ipipe, flags);
1868         if (ret)
1869                 return ret;
1870
1871         ret = opipe_prep(opipe, flags);
1872         if (ret)
1873                 return ret;
1874
1875         /*
1876          * Potential ABBA deadlock, work around it by ordering lock
1877          * grabbing by pipe info address. Otherwise two different processes
1878          * could deadlock (one doing tee from A -> B, the other from B -> A).
1879          */
1880         pipe_double_lock(ipipe, opipe);
1881
1882         do {
1883                 if (!opipe->readers) {
1884                         send_sig(SIGPIPE, current, 0);
1885                         if (!ret)
1886                                 ret = -EPIPE;
1887                         break;
1888                 }
1889
1890                 if (!ipipe->nrbufs && !ipipe->writers)
1891                         break;
1892
1893                 /*
1894                  * Cannot make any progress, because either the input
1895                  * pipe is empty or the output pipe is full.
1896                  */
1897                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1898                         /* Already processed some buffers, break */
1899                         if (ret)
1900                                 break;
1901
1902                         if (flags & SPLICE_F_NONBLOCK) {
1903                                 ret = -EAGAIN;
1904                                 break;
1905                         }
1906
1907                         /*
1908                          * We raced with another reader/writer and haven't
1909                          * managed to process any buffers.  A zero return
1910                          * value means EOF, so retry instead.
1911                          */
1912                         pipe_unlock(ipipe);
1913                         pipe_unlock(opipe);
1914                         goto retry;
1915                 }
1916
1917                 ibuf = ipipe->bufs + ipipe->curbuf;
1918                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1919                 obuf = opipe->bufs + nbuf;
1920
1921                 if (len >= ibuf->len) {
1922                         /*
1923                          * Simply move the whole buffer from ipipe to opipe
1924                          */
1925                         *obuf = *ibuf;
1926                         ibuf->ops = NULL;
1927                         opipe->nrbufs++;
1928                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1929                         ipipe->nrbufs--;
1930                         input_wakeup = true;
1931                 } else {
1932                         /*
1933                          * Get a reference to this pipe buffer,
1934                          * so we can copy the contents over.
1935                          */
1936                         ibuf->ops->get(ipipe, ibuf);
1937                         *obuf = *ibuf;
1938
1939                         /*
1940                          * Don't inherit the gift flag, we need to
1941                          * prevent multiple steals of this page.
1942                          */
1943                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1944
1945                         obuf->len = len;
1946                         opipe->nrbufs++;
1947                         ibuf->offset += obuf->len;
1948                         ibuf->len -= obuf->len;
1949                 }
1950                 ret += obuf->len;
1951                 len -= obuf->len;
1952         } while (len);
1953
1954         pipe_unlock(ipipe);
1955         pipe_unlock(opipe);
1956
1957         /*
1958          * If we put data in the output pipe, wakeup any potential readers.
1959          */
1960         if (ret > 0)
1961                 wakeup_pipe_readers(opipe);
1962
1963         if (input_wakeup)
1964                 wakeup_pipe_writers(ipipe);
1965
1966         return ret;
1967 }
1968
1969 /*
1970  * Link contents of ipipe to opipe.
1971  */
1972 static int link_pipe(struct pipe_inode_info *ipipe,
1973                      struct pipe_inode_info *opipe,
1974                      size_t len, unsigned int flags)
1975 {
1976         struct pipe_buffer *ibuf, *obuf;
1977         int ret = 0, i = 0, nbuf;
1978
1979         /*
1980          * Potential ABBA deadlock, work around it by ordering lock
1981          * grabbing by pipe info address. Otherwise two different processes
1982          * could deadlock (one doing tee from A -> B, the other from B -> A).
1983          */
1984         pipe_double_lock(ipipe, opipe);
1985
1986         do {
1987                 if (!opipe->readers) {
1988                         send_sig(SIGPIPE, current, 0);
1989                         if (!ret)
1990                                 ret = -EPIPE;
1991                         break;
1992                 }
1993
1994                 /*
1995                  * If we have iterated all input buffers or ran out of
1996                  * output room, break.
1997                  */
1998                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1999                         break;
2000
2001                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
2002                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2003
2004                 /*
2005                  * Get a reference to this pipe buffer,
2006                  * so we can copy the contents over.
2007                  */
2008                 ibuf->ops->get(ipipe, ibuf);
2009
2010                 obuf = opipe->bufs + nbuf;
2011                 *obuf = *ibuf;
2012
2013                 /*
2014                  * Don't inherit the gift flag, we need to
2015                  * prevent multiple steals of this page.
2016                  */
2017                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2018
2019                 if (obuf->len > len)
2020                         obuf->len = len;
2021
2022                 opipe->nrbufs++;
2023                 ret += obuf->len;
2024                 len -= obuf->len;
2025                 i++;
2026         } while (len);
2027
2028         /*
2029          * return EAGAIN if we have the potential of some data in the
2030          * future, otherwise just return 0
2031          */
2032         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2033                 ret = -EAGAIN;
2034
2035         pipe_unlock(ipipe);
2036         pipe_unlock(opipe);
2037
2038         /*
2039          * If we put data in the output pipe, wakeup any potential readers.
2040          */
2041         if (ret > 0)
2042                 wakeup_pipe_readers(opipe);
2043
2044         return ret;
2045 }
2046
2047 /*
2048  * This is a tee(1) implementation that works on pipes. It doesn't copy
2049  * any data, it simply references the 'in' pages on the 'out' pipe.
2050  * The 'flags' used are the SPLICE_F_* variants, currently the only
2051  * applicable one is SPLICE_F_NONBLOCK.
2052  */
2053 static long do_tee(struct file *in, struct file *out, size_t len,
2054                    unsigned int flags)
2055 {
2056         struct pipe_inode_info *ipipe = get_pipe_info(in);
2057         struct pipe_inode_info *opipe = get_pipe_info(out);
2058         int ret = -EINVAL;
2059
2060         /*
2061          * Duplicate the contents of ipipe to opipe without actually
2062          * copying the data.
2063          */
2064         if (ipipe && opipe && ipipe != opipe) {
2065                 /*
2066                  * Keep going, unless we encounter an error. The ipipe/opipe
2067                  * ordering doesn't really matter.
2068                  */
2069                 ret = ipipe_prep(ipipe, flags);
2070                 if (!ret) {
2071                         ret = opipe_prep(opipe, flags);
2072                         if (!ret)
2073                                 ret = link_pipe(ipipe, opipe, len, flags);
2074                 }
2075         }
2076
2077         return ret;
2078 }
2079
2080 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2081 {
2082         struct fd in;
2083         int error;
2084
2085         if (unlikely(!len))
2086                 return 0;
2087
2088         error = -EBADF;
2089         in = fdget(fdin);
2090         if (in.file) {
2091                 if (in.file->f_mode & FMODE_READ) {
2092                         struct fd out = fdget(fdout);
2093                         if (out.file) {
2094                                 if (out.file->f_mode & FMODE_WRITE)
2095                                         error = do_tee(in.file, out.file,
2096                                                         len, flags);
2097                                 fdput(out);
2098                         }
2099                 }
2100                 fdput(in);
2101         }
2102
2103         return error;
2104 }