fs/splice.c

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