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 <[email protected]>
  16  * Copyright (C) 2005-2006 Linus Torvalds <[email protected]>
  17  * Copyright (C) 2006 Ingo Molnar <[email protected]>
  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                                    mapping_gfp_constraint(mapping, GFP_KERNEL));
 364                         if (unlikely(error)) {
 365                                 page_cache_release(page);
 366                                 if (error == -EEXIST)
 367                                         continue;
 368                                 break;
 369                         }
 370                         /*
 371                          * add_to_page_cache() locks the page, unlock it
 372                          * to avoid convoluting the logic below even more.
 373                          */
 374                         unlock_page(page);
 375                 }
 376
 377                 spd.pages[spd.nr_pages++] = page;
 378                 index++;
 379         }
 380
 381         /*
 382          * Now loop over the map and see if we need to start IO on any
 383          * pages, fill in the partial map, etc.
 384          */
 385         index = *ppos >> PAGE_CACHE_SHIFT;
 386         nr_pages = spd.nr_pages;
 387         spd.nr_pages = 0;
 388         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 389                 unsigned int this_len;
 390
 391                 if (!len)
 392                         break;
 393
 394                 /*
 395                  * this_len is the max we'll use from this page
 396                  */
 397                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 398                 page = spd.pages[page_nr];
 399
 400                 if (PageReadahead(page))
 401                         page_cache_async_readahead(mapping, &in->f_ra, in,
 402                                         page, index, req_pages - page_nr);
 403
 404                 /*
 405                  * If the page isn't uptodate, we may need to start io on it
 406                  */
 407                 if (!PageUptodate(page)) {
 408                         lock_page(page);
 409
 410                         /*
 411                          * Page was truncated, or invalidated by the
 412                          * filesystem.  Redo the find/create, but this time the
 413                          * page is kept locked, so there's no chance of another
 414                          * race with truncate/invalidate.
 415                          */
 416                         if (!page->mapping) {
 417                                 unlock_page(page);
 418 retry_lookup:
 419                                 page = find_or_create_page(mapping, index,
 420                                                 mapping_gfp_mask(mapping));
 421
 422                                 if (!page) {
 423                                         error = -ENOMEM;
 424                                         break;
 425                                 }
 426                                 page_cache_release(spd.pages[page_nr]);
 427                                 spd.pages[page_nr] = page;
 428                         }
 429                         /*
 430                          * page was already under io and is now done, great
 431                          */
 432                         if (PageUptodate(page)) {
 433                                 unlock_page(page);
 434                                 goto fill_it;
 435                         }
 436
 437                         /*
 438                          * need to read in the page
 439                          */
 440                         error = mapping->a_ops->readpage(in, page);
 441                         if (unlikely(error)) {
 442                                 /*
 443                                  * Re-lookup the page
 444                                  */
 445                                 if (error == AOP_TRUNCATED_PAGE)
 446                                         goto retry_lookup;
 447
 448                                 break;
 449                         }
 450                 }
 451 fill_it:
 452                 /*
 453                  * i_size must be checked after PageUptodate.
 454                  */
 455                 isize = i_size_read(mapping->host);
 456                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 457                 if (unlikely(!isize || index > end_index))
 458                         break;
 459
 460                 /*
 461                  * if this is the last page, see if we need to shrink
 462                  * the length and stop
 463                  */
 464                 if (end_index == index) {
 465                         unsigned int plen;
 466
 467                         /*
 468                          * max good bytes in this page
 469                          */
 470                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 471                         if (plen <= loff)
 472                                 break;
 473
 474                         /*
 475                          * force quit after adding this page
 476                          */
 477                         this_len = min(this_len, plen - loff);
 478                         len = this_len;
 479                 }
 480
 481                 spd.partial[page_nr].offset = loff;
 482                 spd.partial[page_nr].len = this_len;
 483                 len -= this_len;
 484                 loff = 0;
 485                 spd.nr_pages++;
 486                 index++;
 487         }
 488
 489         /*
 490          * Release any pages at the end, if we quit early. 'page_nr' is how far
 491          * we got, 'nr_pages' is how many pages are in the map.
 492          */
 493         while (page_nr < nr_pages)
 494                 page_cache_release(spd.pages[page_nr++]);
 495         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 496
 497         if (spd.nr_pages)
 498                 error = splice_to_pipe(pipe, &spd);
 499
 500         splice_shrink_spd(&spd);
 501         return error;
 502 }
 503
 504 /**
 505  * generic_file_splice_read - splice data from file to a pipe
 506  * @in:         file to splice from
 507  * @ppos:       position in @in
 508  * @pipe:       pipe to splice to
 509  * @len:        number of bytes to splice
 510  * @flags:      splice modifier flags
 511  *
 512  * Description:
 513  *    Will read pages from given file and fill them into a pipe. Can be
 514  *    used as long as the address_space operations for the source implements
 515  *    a readpage() hook.
 516  *
 517  */
 518 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 519                                  struct pipe_inode_info *pipe, size_t len,
 520                                  unsigned int flags)
 521 {
 522         loff_t isize, left;
 523         int ret;
 524
 525         if (IS_DAX(in->f_mapping->host))
 526                 return default_file_splice_read(in, ppos, pipe, len, flags);
 527
 528         isize = i_size_read(in->f_mapping->host);
 529         if (unlikely(*ppos >= isize))
 530                 return 0;
 531
 532         left = isize - *ppos;
 533         if (unlikely(left < len))
 534                 len = left;
 535
 536         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 537         if (ret > 0) {
 538                 *ppos += ret;
 539                 file_accessed(in);
 540         }
 541
 542         return ret;
 543 }
 544 EXPORT_SYMBOL(generic_file_splice_read);
 545
 546 static const struct pipe_buf_operations default_pipe_buf_ops = {
 547         .can_merge = 0,
 548         .confirm = generic_pipe_buf_confirm,
 549         .release = generic_pipe_buf_release,
 550         .steal = generic_pipe_buf_steal,
 551         .get = generic_pipe_buf_get,
 552 };
 553
 554 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 555                                     struct pipe_buffer *buf)
 556 {
 557         return 1;
 558 }
 559
 560 /* Pipe buffer operations for a socket and similar. */
 561 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 562         .can_merge = 0,
 563         .confirm = generic_pipe_buf_confirm,
 564         .release = generic_pipe_buf_release,
 565         .steal = generic_pipe_buf_nosteal,
 566         .get = generic_pipe_buf_get,
 567 };
 568 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 569
 570 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 571                             unsigned long vlen, loff_t offset)
 572 {
 573         mm_segment_t old_fs;
 574         loff_t pos = offset;
 575         ssize_t res;
 576
 577         old_fs = get_fs();
 578         set_fs(get_ds());
 579         /* The cast to a user pointer is valid due to the set_fs() */
 580         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 581         set_fs(old_fs);
 582
 583         return res;
 584 }
 585
 586 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 587                             loff_t pos)
 588 {
 589         mm_segment_t old_fs;
 590         ssize_t res;
 591
 592         old_fs = get_fs();
 593         set_fs(get_ds());
 594         /* The cast to a user pointer is valid due to the set_fs() */
 595         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
 596         set_fs(old_fs);
 597
 598         return res;
 599 }
 600 EXPORT_SYMBOL(kernel_write);
 601
 602 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 603                                  struct pipe_inode_info *pipe, size_t len,
 604                                  unsigned int flags)
 605 {
 606         unsigned int nr_pages;
 607         unsigned int nr_freed;
 608         size_t offset;
 609         struct page *pages[PIPE_DEF_BUFFERS];
 610         struct partial_page partial[PIPE_DEF_BUFFERS];
 611         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 612         ssize_t res;
 613         size_t this_len;
 614         int error;
 615         int i;
 616         struct splice_pipe_desc spd = {
 617                 .pages = pages,
 618                 .partial = partial,
 619                 .nr_pages_max = PIPE_DEF_BUFFERS,
 620                 .flags = flags,
 621                 .ops = &default_pipe_buf_ops,
 622                 .spd_release = spd_release_page,
 623         };
 624
 625         if (splice_grow_spd(pipe, &spd))
 626                 return -ENOMEM;
 627
 628         res = -ENOMEM;
 629         vec = __vec;
 630         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 631                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 632                 if (!vec)
 633                         goto shrink_ret;
 634         }
 635
 636         offset = *ppos & ~PAGE_CACHE_MASK;
 637         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 638
 639         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 640                 struct page *page;
 641
 642                 page = alloc_page(GFP_USER);
 643                 error = -ENOMEM;
 644                 if (!page)
 645                         goto err;
 646
 647                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 648                 vec[i].iov_base = (void __user *) page_address(page);
 649                 vec[i].iov_len = this_len;
 650                 spd.pages[i] = page;
 651                 spd.nr_pages++;
 652                 len -= this_len;
 653                 offset = 0;
 654         }
 655
 656         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 657         if (res < 0) {
 658                 error = res;
 659                 goto err;
 660         }
 661
 662         error = 0;
 663         if (!res)
 664                 goto err;
 665
 666         nr_freed = 0;
 667         for (i = 0; i < spd.nr_pages; i++) {
 668                 this_len = min_t(size_t, vec[i].iov_len, res);
 669                 spd.partial[i].offset = 0;
 670                 spd.partial[i].len = this_len;
 671                 if (!this_len) {
 672                         __free_page(spd.pages[i]);
 673                         spd.pages[i] = NULL;
 674                         nr_freed++;
 675                 }
 676                 res -= this_len;
 677         }
 678         spd.nr_pages -= nr_freed;
 679
 680         res = splice_to_pipe(pipe, &spd);
 681         if (res > 0)
 682                 *ppos += res;
 683
 684 shrink_ret:
 685         if (vec != __vec)
 686                 kfree(vec);
 687         splice_shrink_spd(&spd);
 688         return res;
 689
 690 err:
 691         for (i = 0; i < spd.nr_pages; i++)
 692                 __free_page(spd.pages[i]);
 693
 694         res = error;
 695         goto shrink_ret;
 696 }
 697 EXPORT_SYMBOL(default_file_splice_read);
 698
 699 /*
 700  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 701  * using sendpage(). Return the number of bytes sent.
 702  */
 703 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 704                             struct pipe_buffer *buf, struct splice_desc *sd)
 705 {
 706         struct file *file = sd->u.file;
 707         loff_t pos = sd->pos;
 708         int more;
 709
 710         if (!likely(file->f_op->sendpage))
 711                 return -EINVAL;
 712
 713         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 714
 715         if (sd->len < sd->total_len && pipe->nrbufs > 1)
 716                 more |= MSG_SENDPAGE_NOTLAST;
 717
 718         return file->f_op->sendpage(file, buf->page, buf->offset,
 719                                     sd->len, &pos, more);
 720 }
 721
 722 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 723 {
 724         smp_mb();
 725         if (waitqueue_active(&pipe->wait))
 726                 wake_up_interruptible(&pipe->wait);
 727         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 728 }
 729
 730 /**
 731  * splice_from_pipe_feed - feed available data from a pipe to a file
 732  * @pipe:       pipe to splice from
 733  * @sd:         information to @actor
 734  * @actor:      handler that splices the data
 735  *
 736  * Description:
 737  *    This function loops over the pipe and calls @actor to do the
 738  *    actual moving of a single struct pipe_buffer to the desired
 739  *    destination.  It returns when there's no more buffers left in
 740  *    the pipe or if the requested number of bytes (@sd->total_len)
 741  *    have been copied.  It returns a positive number (one) if the
 742  *    pipe needs to be filled with more data, zero if the required
 743  *    number of bytes have been copied and -errno on error.
 744  *
 745  *    This, together with splice_from_pipe_{begin,end,next}, may be
 746  *    used to implement the functionality of __splice_from_pipe() when
 747  *    locking is required around copying the pipe buffers to the
 748  *    destination.
 749  */
 750 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 751                           splice_actor *actor)
 752 {
 753         int ret;
 754
 755         while (pipe->nrbufs) {
 756                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 757                 const struct pipe_buf_operations *ops = buf->ops;
 758
 759                 sd->len = buf->len;
 760                 if (sd->len > sd->total_len)
 761                         sd->len = sd->total_len;
 762
 763                 ret = buf->ops->confirm(pipe, buf);
 764                 if (unlikely(ret)) {
 765                         if (ret == -ENODATA)
 766                                 ret = 0;
 767                         return ret;
 768                 }
 769
 770                 ret = actor(pipe, buf, sd);
 771                 if (ret <= 0)
 772                         return ret;
 773
 774                 buf->offset += ret;
 775                 buf->len -= ret;
 776
 777                 sd->num_spliced += ret;
 778                 sd->len -= ret;
 779                 sd->pos += ret;
 780                 sd->total_len -= ret;
 781
 782                 if (!buf->len) {
 783                         buf->ops = NULL;
 784                         ops->release(pipe, buf);
 785                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 786                         pipe->nrbufs--;
 787                         if (pipe->files)
 788                                 sd->need_wakeup = true;
 789                 }
 790
 791                 if (!sd->total_len)
 792                         return 0;
 793         }
 794
 795         return 1;
 796 }
 797
 798 /**
 799  * splice_from_pipe_next - wait for some data to splice from
 800  * @pipe:       pipe to splice from
 801  * @sd:         information about the splice operation
 802  *
 803  * Description:
 804  *    This function will wait for some data and return a positive
 805  *    value (one) if pipe buffers are available.  It will return zero
 806  *    or -errno if no more data needs to be spliced.
 807  */
 808 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 809 {
 810         /*
 811          * Check for signal early to make process killable when there are
 812          * always buffers available
 813          */
 814         if (signal_pending(current))
 815                 return -ERESTARTSYS;
 816
 817         while (!pipe->nrbufs) {
 818                 if (!pipe->writers)
 819                         return 0;
 820
 821                 if (!pipe->waiting_writers && sd->num_spliced)
 822                         return 0;
 823
 824                 if (sd->flags & SPLICE_F_NONBLOCK)
 825                         return -EAGAIN;
 826
 827                 if (signal_pending(current))
 828                         return -ERESTARTSYS;
 829
 830                 if (sd->need_wakeup) {
 831                         wakeup_pipe_writers(pipe);
 832                         sd->need_wakeup = false;
 833                 }
 834
 835                 pipe_wait(pipe);
 836         }
 837
 838         return 1;
 839 }
 840
 841 /**
 842  * splice_from_pipe_begin - start splicing from pipe
 843  * @sd:         information about the splice operation
 844  *
 845  * Description:
 846  *    This function should be called before a loop containing
 847  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 848  *    initialize the necessary fields of @sd.
 849  */
 850 static void splice_from_pipe_begin(struct splice_desc *sd)
 851 {
 852         sd->num_spliced = 0;
 853         sd->need_wakeup = false;
 854 }
 855
 856 /**
 857  * splice_from_pipe_end - finish splicing from pipe
 858  * @pipe:       pipe to splice from
 859  * @sd:         information about the splice operation
 860  *
 861  * Description:
 862  *    This function will wake up pipe writers if necessary.  It should
 863  *    be called after a loop containing splice_from_pipe_next() and
 864  *    splice_from_pipe_feed().
 865  */
 866 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 867 {
 868         if (sd->need_wakeup)
 869                 wakeup_pipe_writers(pipe);
 870 }
 871
 872 /**
 873  * __splice_from_pipe - splice data from a pipe to given actor
 874  * @pipe:       pipe to splice from
 875  * @sd:         information to @actor
 876  * @actor:      handler that splices the data
 877  *
 878  * Description:
 879  *    This function does little more than loop over the pipe and call
 880  *    @actor to do the actual moving of a single struct pipe_buffer to
 881  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 882  *    pipe_to_user.
 883  *
 884  */
 885 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 886                            splice_actor *actor)
 887 {
 888         int ret;
 889
 890         splice_from_pipe_begin(sd);
 891         do {
 892                 cond_resched();
 893                 ret = splice_from_pipe_next(pipe, sd);
 894                 if (ret > 0)
 895                         ret = splice_from_pipe_feed(pipe, sd, actor);
 896         } while (ret > 0);
 897         splice_from_pipe_end(pipe, sd);
 898
 899         return sd->num_spliced ? sd->num_spliced : ret;
 900 }
 901 EXPORT_SYMBOL(__splice_from_pipe);
 902
 903 /**
 904  * splice_from_pipe - splice data from a pipe to a file
 905  * @pipe:       pipe to splice from
 906  * @out:        file to splice to
 907  * @ppos:       position in @out
 908  * @len:        how many bytes to splice
 909  * @flags:      splice modifier flags
 910  * @actor:      handler that splices the data
 911  *
 912  * Description:
 913  *    See __splice_from_pipe. This function locks the pipe inode,
 914  *    otherwise it's identical to __splice_from_pipe().
 915  *
 916  */
 917 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 918                          loff_t *ppos, size_t len, unsigned int flags,
 919                          splice_actor *actor)
 920 {
 921         ssize_t ret;
 922         struct splice_desc sd = {
 923                 .total_len = len,
 924                 .flags = flags,
 925                 .pos = *ppos,
 926                 .u.file = out,
 927         };
 928
 929         pipe_lock(pipe);
 930         ret = __splice_from_pipe(pipe, &sd, actor);
 931         pipe_unlock(pipe);
 932
 933         return ret;
 934 }
 935
 936 /**
 937  * iter_file_splice_write - splice data from a pipe to a file
 938  * @pipe:       pipe info
 939  * @out:        file to write to
 940  * @ppos:       position in @out
 941  * @len:        number of bytes to splice
 942  * @flags:      splice modifier flags
 943  *
 944  * Description:
 945  *    Will either move or copy pages (determined by @flags options) from
 946  *    the given pipe inode to the given file.
 947  *    This one is ->write_iter-based.
 948  *
 949  */
 950 ssize_t
 951 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 952                           loff_t *ppos, size_t len, unsigned int flags)
 953 {
 954         struct splice_desc sd = {
 955                 .total_len = len,
 956                 .flags = flags,
 957                 .pos = *ppos,
 958                 .u.file = out,
 959         };
 960         int nbufs = pipe->buffers;
 961         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 962                                         GFP_KERNEL);
 963         ssize_t ret;
 964
 965         if (unlikely(!array))
 966                 return -ENOMEM;
 967
 968         pipe_lock(pipe);
 969
 970         splice_from_pipe_begin(&sd);
 971         while (sd.total_len) {
 972                 struct iov_iter from;
 973                 size_t left;
 974                 int n, idx;
 975
 976                 ret = splice_from_pipe_next(pipe, &sd);
 977                 if (ret <= 0)
 978                         break;
 979
 980                 if (unlikely(nbufs < pipe->buffers)) {
 981                         kfree(array);
 982                         nbufs = pipe->buffers;
 983                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 984                                         GFP_KERNEL);
 985                         if (!array) {
 986                                 ret = -ENOMEM;
 987                                 break;
 988                         }
 989                 }
 990
 991                 /* build the vector */
 992                 left = sd.total_len;
 993                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 994                         struct pipe_buffer *buf = pipe->bufs + idx;
 995                         size_t this_len = buf->len;
 996
 997                         if (this_len > left)
 998                                 this_len = left;
 999
1000                         if (idx == pipe->buffers - 1)
1001                                 idx = -1;
1002
1003                         ret = buf->ops->confirm(pipe, buf);
1004                         if (unlikely(ret)) {
1005                                 if (ret == -ENODATA)
1006                                         ret = 0;
1007                                 goto done;
1008                         }
1009
1010                         array[n].bv_page = buf->page;
1011                         array[n].bv_len = this_len;
1012                         array[n].bv_offset = buf->offset;
1013                         left -= this_len;
1014                 }
1015
1016                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
1017                               sd.total_len - left);
1018                 ret = vfs_iter_write(out, &from, &sd.pos);
1019                 if (ret <= 0)
1020                         break;
1021
1022                 sd.num_spliced += ret;
1023                 sd.total_len -= ret;
1024                 *ppos = sd.pos;
1025
1026                 /* dismiss the fully eaten buffers, adjust the partial one */
1027                 while (ret) {
1028                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1029                         if (ret >= buf->len) {
1030                                 const struct pipe_buf_operations *ops = buf->ops;
1031                                 ret -= buf->len;
1032                                 buf->len = 0;
1033                                 buf->ops = NULL;
1034                                 ops->release(pipe, buf);
1035                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1036                                 pipe->nrbufs--;
1037                                 if (pipe->files)
1038                                         sd.need_wakeup = true;
1039                         } else {
1040                                 buf->offset += ret;
1041                                 buf->len -= ret;
1042                                 ret = 0;
1043                         }
1044                 }
1045         }
1046 done:
1047         kfree(array);
1048         splice_from_pipe_end(pipe, &sd);
1049
1050         pipe_unlock(pipe);
1051
1052         if (sd.num_spliced)
1053                 ret = sd.num_spliced;
1054
1055         return ret;
1056 }
1057
1058 EXPORT_SYMBOL(iter_file_splice_write);
1059
1060 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1061                           struct splice_desc *sd)
1062 {
1063         int ret;
1064         void *data;
1065         loff_t tmp = sd->pos;
1066
1067         data = kmap(buf->page);
1068         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1069         kunmap(buf->page);
1070
1071         return ret;
1072 }
1073
1074 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1075                                          struct file *out, loff_t *ppos,
1076                                          size_t len, unsigned int flags)
1077 {
1078         ssize_t ret;
1079
1080         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1081         if (ret > 0)
1082                 *ppos += ret;
1083
1084         return ret;
1085 }
1086
1087 /**
1088  * generic_splice_sendpage - splice data from a pipe to a socket
1089  * @pipe:       pipe to splice from
1090  * @out:        socket to write to
1091  * @ppos:       position in @out
1092  * @len:        number of bytes to splice
1093  * @flags:      splice modifier flags
1094  *
1095  * Description:
1096  *    Will send @len bytes from the pipe to a network socket. No data copying
1097  *    is involved.
1098  *
1099  */
1100 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1101                                 loff_t *ppos, size_t len, unsigned int flags)
1102 {
1103         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1104 }
1105
1106 EXPORT_SYMBOL(generic_splice_sendpage);
1107
1108 /*
1109  * Attempt to initiate a splice from pipe to file.
1110  */
1111 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1112                            loff_t *ppos, size_t len, unsigned int flags)
1113 {
1114         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1115                                 loff_t *, size_t, unsigned int);
1116
1117         if (out->f_op->splice_write)
1118                 splice_write = out->f_op->splice_write;
1119         else
1120                 splice_write = default_file_splice_write;
1121
1122         return splice_write(pipe, out, ppos, len, flags);
1123 }
1124
1125 /*
1126  * Attempt to initiate a splice from a file to a pipe.
1127  */
1128 static long do_splice_to(struct file *in, loff_t *ppos,
1129                          struct pipe_inode_info *pipe, size_t len,
1130                          unsigned int flags)
1131 {
1132         ssize_t (*splice_read)(struct file *, loff_t *,
1133                                struct pipe_inode_info *, size_t, unsigned int);
1134         int ret;
1135
1136         if (unlikely(!(in->f_mode & FMODE_READ)))
1137                 return -EBADF;
1138
1139         ret = rw_verify_area(READ, in, ppos, len);
1140         if (unlikely(ret < 0))
1141                 return ret;
1142
1143         if (in->f_op->splice_read)
1144                 splice_read = in->f_op->splice_read;
1145         else
1146                 splice_read = default_file_splice_read;
1147
1148         return splice_read(in, ppos, pipe, len, flags);
1149 }
1150
1151 /**
1152  * splice_direct_to_actor - splices data directly between two non-pipes
1153  * @in:         file to splice from
1154  * @sd:         actor information on where to splice to
1155  * @actor:      handles the data splicing
1156  *
1157  * Description:
1158  *    This is a special case helper to splice directly between two
1159  *    points, without requiring an explicit pipe. Internally an allocated
1160  *    pipe is cached in the process, and reused during the lifetime of
1161  *    that process.
1162  *
1163  */
1164 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1165                                splice_direct_actor *actor)
1166 {
1167         struct pipe_inode_info *pipe;
1168         long ret, bytes;
1169         umode_t i_mode;
1170         size_t len;
1171         int i, flags, more;
1172
1173         /*
1174          * We require the input being a regular file, as we don't want to
1175          * randomly drop data for eg socket -> socket splicing. Use the
1176          * piped splicing for that!
1177          */
1178         i_mode = file_inode(in)->i_mode;
1179         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1180                 return -EINVAL;
1181
1182         /*
1183          * neither in nor out is a pipe, setup an internal pipe attached to
1184          * 'out' and transfer the wanted data from 'in' to 'out' through that
1185          */
1186         pipe = current->splice_pipe;
1187         if (unlikely(!pipe)) {
1188                 pipe = alloc_pipe_info();
1189                 if (!pipe)
1190                         return -ENOMEM;
1191
1192                 /*
1193                  * We don't have an immediate reader, but we'll read the stuff
1194                  * out of the pipe right after the splice_to_pipe(). So set
1195                  * PIPE_READERS appropriately.
1196                  */
1197                 pipe->readers = 1;
1198
1199                 current->splice_pipe = pipe;
1200         }
1201
1202         /*
1203          * Do the splice.
1204          */
1205         ret = 0;
1206         bytes = 0;
1207         len = sd->total_len;
1208         flags = sd->flags;
1209
1210         /*
1211          * Don't block on output, we have to drain the direct pipe.
1212          */
1213         sd->flags &= ~SPLICE_F_NONBLOCK;
1214         more = sd->flags & SPLICE_F_MORE;
1215
1216         while (len) {
1217                 size_t read_len;
1218                 loff_t pos = sd->pos, prev_pos = pos;
1219
1220                 ret = do_splice_to(in, &pos, pipe, len, flags);
1221                 if (unlikely(ret <= 0))
1222                         goto out_release;
1223
1224                 read_len = ret;
1225                 sd->total_len = read_len;
1226
1227                 /*
1228                  * If more data is pending, set SPLICE_F_MORE
1229                  * If this is the last data and SPLICE_F_MORE was not set
1230                  * initially, clears it.
1231                  */
1232                 if (read_len < len)
1233                         sd->flags |= SPLICE_F_MORE;
1234                 else if (!more)
1235                         sd->flags &= ~SPLICE_F_MORE;
1236                 /*
1237                  * NOTE: nonblocking mode only applies to the input. We
1238                  * must not do the output in nonblocking mode as then we
1239                  * could get stuck data in the internal pipe:
1240                  */
1241                 ret = actor(pipe, sd);
1242                 if (unlikely(ret <= 0)) {
1243                         sd->pos = prev_pos;
1244                         goto out_release;
1245                 }
1246
1247                 bytes += ret;
1248                 len -= ret;
1249                 sd->pos = pos;
1250
1251                 if (ret < read_len) {
1252                         sd->pos = prev_pos + ret;
1253                         goto out_release;
1254                 }
1255         }
1256
1257 done:
1258         pipe->nrbufs = pipe->curbuf = 0;
1259         file_accessed(in);
1260         return bytes;
1261
1262 out_release:
1263         /*
1264          * If we did an incomplete transfer we must release
1265          * the pipe buffers in question:
1266          */
1267         for (i = 0; i < pipe->buffers; i++) {
1268                 struct pipe_buffer *buf = pipe->bufs + i;
1269
1270                 if (buf->ops) {
1271                         buf->ops->release(pipe, buf);
1272                         buf->ops = NULL;
1273                 }
1274         }
1275
1276         if (!bytes)
1277                 bytes = ret;
1278
1279         goto done;
1280 }
1281 EXPORT_SYMBOL(splice_direct_to_actor);
1282
1283 static int direct_splice_actor(struct pipe_inode_info *pipe,
1284                                struct splice_desc *sd)
1285 {
1286         struct file *file = sd->u.file;
1287
1288         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1289                               sd->flags);
1290 }
1291
1292 /**
1293  * do_splice_direct - splices data directly between two files
1294  * @in:         file to splice from
1295  * @ppos:       input file offset
1296  * @out:        file to splice to
1297  * @opos:       output file offset
1298  * @len:        number of bytes to splice
1299  * @flags:      splice modifier flags
1300  *
1301  * Description:
1302  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1303  *    doing it in the application would incur an extra system call
1304  *    (splice in + splice out, as compared to just sendfile()). So this helper
1305  *    can splice directly through a process-private pipe.
1306  *
1307  */
1308 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1309                       loff_t *opos, size_t len, unsigned int flags)
1310 {
1311         struct splice_desc sd = {
1312                 .len            = len,
1313                 .total_len      = len,
1314                 .flags          = flags,
1315                 .pos            = *ppos,
1316                 .u.file         = out,
1317                 .opos           = opos,
1318         };
1319         long ret;
1320
1321         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1322                 return -EBADF;
1323
1324         if (unlikely(out->f_flags & O_APPEND))
1325                 return -EINVAL;
1326
1327         ret = rw_verify_area(WRITE, out, opos, len);
1328         if (unlikely(ret < 0))
1329                 return ret;
1330
1331         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1332         if (ret > 0)
1333                 *ppos = sd.pos;
1334
1335         return ret;
1336 }
1337 EXPORT_SYMBOL(do_splice_direct);
1338
1339 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1340                                struct pipe_inode_info *opipe,
1341                                size_t len, unsigned int flags);
1342
1343 /*
1344  * Determine where to splice to/from.
1345  */
1346 static long do_splice(struct file *in, loff_t __user *off_in,
1347                       struct file *out, loff_t __user *off_out,
1348                       size_t len, unsigned int flags)
1349 {
1350         struct pipe_inode_info *ipipe;
1351         struct pipe_inode_info *opipe;
1352         loff_t offset;
1353         long ret;
1354
1355         ipipe = get_pipe_info(in);
1356         opipe = get_pipe_info(out);
1357
1358         if (ipipe && opipe) {
1359                 if (off_in || off_out)
1360                         return -ESPIPE;
1361
1362                 if (!(in->f_mode & FMODE_READ))
1363                         return -EBADF;
1364
1365                 if (!(out->f_mode & FMODE_WRITE))
1366                         return -EBADF;
1367
1368                 /* Splicing to self would be fun, but... */
1369                 if (ipipe == opipe)
1370                         return -EINVAL;
1371
1372                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1373         }
1374
1375         if (ipipe) {
1376                 if (off_in)
1377                         return -ESPIPE;
1378                 if (off_out) {
1379                         if (!(out->f_mode & FMODE_PWRITE))
1380                                 return -EINVAL;
1381                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1382                                 return -EFAULT;
1383                 } else {
1384                         offset = out->f_pos;
1385                 }
1386
1387                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1388                         return -EBADF;
1389
1390                 if (unlikely(out->f_flags & O_APPEND))
1391                         return -EINVAL;
1392
1393                 ret = rw_verify_area(WRITE, out, &offset, len);
1394                 if (unlikely(ret < 0))
1395                         return ret;
1396
1397                 file_start_write(out);
1398                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1399                 file_end_write(out);
1400
1401                 if (!off_out)
1402                         out->f_pos = offset;
1403                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1404                         ret = -EFAULT;
1405
1406                 return ret;
1407         }
1408
1409         if (opipe) {
1410                 if (off_out)
1411                         return -ESPIPE;
1412                 if (off_in) {
1413                         if (!(in->f_mode & FMODE_PREAD))
1414                                 return -EINVAL;
1415                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1416                                 return -EFAULT;
1417                 } else {
1418                         offset = in->f_pos;
1419                 }
1420
1421                 ret = do_splice_to(in, &offset, opipe, len, flags);
1422
1423                 if (!off_in)
1424                         in->f_pos = offset;
1425                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1426                         ret = -EFAULT;
1427
1428                 return ret;
1429         }
1430
1431         return -EINVAL;
1432 }
1433
1434 /*
1435  * Map an iov into an array of pages and offset/length tupples. With the
1436  * partial_page structure, we can map several non-contiguous ranges into
1437  * our ones pages[] map instead of splitting that operation into pieces.
1438  * Could easily be exported as a generic helper for other users, in which
1439  * case one would probably want to add a 'max_nr_pages' parameter as well.
1440  */
1441 static int get_iovec_page_array(const struct iovec __user *iov,
1442                                 unsigned int nr_vecs, struct page **pages,
1443                                 struct partial_page *partial, bool aligned,
1444                                 unsigned int pipe_buffers)
1445 {
1446         int buffers = 0, error = 0;
1447
1448         while (nr_vecs) {
1449                 unsigned long off, npages;
1450                 struct iovec entry;
1451                 void __user *base;
1452                 size_t len;
1453                 int i;
1454
1455                 error = -EFAULT;
1456                 if (copy_from_user(&entry, iov, sizeof(entry)))
1457                         break;
1458
1459                 base = entry.iov_base;
1460                 len = entry.iov_len;
1461
1462                 /*
1463                  * Sanity check this iovec. 0 read succeeds.
1464                  */
1465                 error = 0;
1466                 if (unlikely(!len))
1467                         break;
1468                 error = -EFAULT;
1469                 if (!access_ok(VERIFY_READ, base, len))
1470                         break;
1471
1472                 /*
1473                  * Get this base offset and number of pages, then map
1474                  * in the user pages.
1475                  */
1476                 off = (unsigned long) base & ~PAGE_MASK;
1477
1478                 /*
1479                  * If asked for alignment, the offset must be zero and the
1480                  * length a multiple of the PAGE_SIZE.
1481                  */
1482                 error = -EINVAL;
1483                 if (aligned && (off || len & ~PAGE_MASK))
1484                         break;
1485
1486                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1487                 if (npages > pipe_buffers - buffers)
1488                         npages = pipe_buffers - buffers;
1489
1490                 error = get_user_pages_fast((unsigned long)base, npages,
1491                                         0, &pages[buffers]);
1492
1493                 if (unlikely(error <= 0))
1494                         break;
1495
1496                 /*
1497                  * Fill this contiguous range into the partial page map.
1498                  */
1499                 for (i = 0; i < error; i++) {
1500                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1501
1502                         partial[buffers].offset = off;
1503                         partial[buffers].len = plen;
1504
1505                         off = 0;
1506                         len -= plen;
1507                         buffers++;
1508                 }
1509
1510                 /*
1511                  * We didn't complete this iov, stop here since it probably
1512                  * means we have to move some of this into a pipe to
1513                  * be able to continue.
1514                  */
1515                 if (len)
1516                         break;
1517
1518                 /*
1519                  * Don't continue if we mapped fewer pages than we asked for,
1520                  * or if we mapped the max number of pages that we have
1521                  * room for.
1522                  */
1523                 if (error < npages || buffers == pipe_buffers)
1524                         break;
1525
1526                 nr_vecs--;
1527                 iov++;
1528         }
1529
1530         if (buffers)
1531                 return buffers;
1532
1533         return error;
1534 }
1535
1536 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1537                         struct splice_desc *sd)
1538 {
1539         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1540         return n == sd->len ? n : -EFAULT;
1541 }
1542
1543 /*
1544  * For lack of a better implementation, implement vmsplice() to userspace
1545  * as a simple copy of the pipes pages to the user iov.
1546  */
1547 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1548                              unsigned long nr_segs, unsigned int flags)
1549 {
1550         struct pipe_inode_info *pipe;
1551         struct splice_desc sd;
1552         long ret;
1553         struct iovec iovstack[UIO_FASTIOV];
1554         struct iovec *iov = iovstack;
1555         struct iov_iter iter;
1556
1557         pipe = get_pipe_info(file);
1558         if (!pipe)
1559                 return -EBADF;
1560
1561         ret = import_iovec(READ, uiov, nr_segs,
1562                            ARRAY_SIZE(iovstack), &iov, &iter);
1563         if (ret < 0)
1564                 return ret;
1565
1566         sd.total_len = iov_iter_count(&iter);
1567         sd.len = 0;
1568         sd.flags = flags;
1569         sd.u.data = &iter;
1570         sd.pos = 0;
1571
1572         if (sd.total_len) {
1573                 pipe_lock(pipe);
1574                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1575                 pipe_unlock(pipe);
1576         }
1577
1578         kfree(iov);
1579         return ret;
1580 }
1581
1582 /*
1583  * vmsplice splices a user address range into a pipe. It can be thought of
1584  * as splice-from-memory, where the regular splice is splice-from-file (or
1585  * to file). In both cases the output is a pipe, naturally.
1586  */
1587 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1588                              unsigned long nr_segs, unsigned int flags)
1589 {
1590         struct pipe_inode_info *pipe;
1591         struct page *pages[PIPE_DEF_BUFFERS];
1592         struct partial_page partial[PIPE_DEF_BUFFERS];
1593         struct splice_pipe_desc spd = {
1594                 .pages = pages,
1595                 .partial = partial,
1596                 .nr_pages_max = PIPE_DEF_BUFFERS,
1597                 .flags = flags,
1598                 .ops = &user_page_pipe_buf_ops,
1599                 .spd_release = spd_release_page,
1600         };
1601         long ret;
1602
1603         pipe = get_pipe_info(file);
1604         if (!pipe)
1605                 return -EBADF;
1606
1607         if (splice_grow_spd(pipe, &spd))
1608                 return -ENOMEM;
1609
1610         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1611                                             spd.partial, false,
1612                                             spd.nr_pages_max);
1613         if (spd.nr_pages <= 0)
1614                 ret = spd.nr_pages;
1615         else
1616                 ret = splice_to_pipe(pipe, &spd);
1617
1618         splice_shrink_spd(&spd);
1619         return ret;
1620 }
1621
1622 /*
1623  * Note that vmsplice only really supports true splicing _from_ user memory
1624  * to a pipe, not the other way around. Splicing from user memory is a simple
1625  * operation that can be supported without any funky alignment restrictions
1626  * or nasty vm tricks. We simply map in the user memory and fill them into
1627  * a pipe. The reverse isn't quite as easy, though. There are two possible
1628  * solutions for that:
1629  *
1630  *      - memcpy() the data internally, at which point we might as well just
1631  *        do a regular read() on the buffer anyway.
1632  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1633  *        has restriction limitations on both ends of the pipe).
1634  *
1635  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1636  *
1637  */
1638 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1639                 unsigned long, nr_segs, unsigned int, flags)
1640 {
1641         struct fd f;
1642         long error;
1643
1644         if (unlikely(nr_segs > UIO_MAXIOV))
1645                 return -EINVAL;
1646         else if (unlikely(!nr_segs))
1647                 return 0;
1648
1649         error = -EBADF;
1650         f = fdget(fd);
1651         if (f.file) {
1652                 if (f.file->f_mode & FMODE_WRITE)
1653                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1654                 else if (f.file->f_mode & FMODE_READ)
1655                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1656
1657                 fdput(f);
1658         }
1659
1660         return error;
1661 }
1662
1663 #ifdef CONFIG_COMPAT
1664 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1665                     unsigned int, nr_segs, unsigned int, flags)
1666 {
1667         unsigned i;
1668         struct iovec __user *iov;
1669         if (nr_segs > UIO_MAXIOV)
1670                 return -EINVAL;
1671         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1672         for (i = 0; i < nr_segs; i++) {
1673                 struct compat_iovec v;
1674                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1675                     get_user(v.iov_len, &iov32[i].iov_len) ||
1676                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1677                     put_user(v.iov_len, &iov[i].iov_len))
1678                         return -EFAULT;
1679         }
1680         return sys_vmsplice(fd, iov, nr_segs, flags);
1681 }
1682 #endif
1683
1684 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1685                 int, fd_out, loff_t __user *, off_out,
1686                 size_t, len, unsigned int, flags)
1687 {
1688         struct fd in, out;
1689         long error;
1690
1691         if (unlikely(!len))
1692                 return 0;
1693
1694         error = -EBADF;
1695         in = fdget(fd_in);
1696         if (in.file) {
1697                 if (in.file->f_mode & FMODE_READ) {
1698                         out = fdget(fd_out);
1699                         if (out.file) {
1700                                 if (out.file->f_mode & FMODE_WRITE)
1701                                         error = do_splice(in.file, off_in,
1702                                                           out.file, off_out,
1703                                                           len, flags);
1704                                 fdput(out);
1705                         }
1706                 }
1707                 fdput(in);
1708         }
1709         return error;
1710 }
1711
1712 /*
1713  * Make sure there's data to read. Wait for input if we can, otherwise
1714  * return an appropriate error.
1715  */
1716 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1717 {
1718         int ret;
1719
1720         /*
1721          * Check ->nrbufs without the inode lock first. This function
1722          * is speculative anyways, so missing one is ok.
1723          */
1724         if (pipe->nrbufs)
1725                 return 0;
1726
1727         ret = 0;
1728         pipe_lock(pipe);
1729
1730         while (!pipe->nrbufs) {
1731                 if (signal_pending(current)) {
1732                         ret = -ERESTARTSYS;
1733                         break;
1734                 }
1735                 if (!pipe->writers)
1736                         break;
1737                 if (!pipe->waiting_writers) {
1738                         if (flags & SPLICE_F_NONBLOCK) {
1739                                 ret = -EAGAIN;
1740                                 break;
1741                         }
1742                 }
1743                 pipe_wait(pipe);
1744         }
1745
1746         pipe_unlock(pipe);
1747         return ret;
1748 }
1749
1750 /*
1751  * Make sure there's writeable room. Wait for room if we can, otherwise
1752  * return an appropriate error.
1753  */
1754 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1755 {
1756         int ret;
1757
1758         /*
1759          * Check ->nrbufs without the inode lock first. This function
1760          * is speculative anyways, so missing one is ok.
1761          */
1762         if (pipe->nrbufs < pipe->buffers)
1763                 return 0;
1764
1765         ret = 0;
1766         pipe_lock(pipe);
1767
1768         while (pipe->nrbufs >= pipe->buffers) {
1769                 if (!pipe->readers) {
1770                         send_sig(SIGPIPE, current, 0);
1771                         ret = -EPIPE;
1772                         break;
1773                 }
1774                 if (flags & SPLICE_F_NONBLOCK) {
1775                         ret = -EAGAIN;
1776                         break;
1777                 }
1778                 if (signal_pending(current)) {
1779                         ret = -ERESTARTSYS;
1780                         break;
1781                 }
1782                 pipe->waiting_writers++;
1783                 pipe_wait(pipe);
1784                 pipe->waiting_writers--;
1785         }
1786
1787         pipe_unlock(pipe);
1788         return ret;
1789 }
1790
1791 /*
1792  * Splice contents of ipipe to opipe.
1793  */
1794 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1795                                struct pipe_inode_info *opipe,
1796                                size_t len, unsigned int flags)
1797 {
1798         struct pipe_buffer *ibuf, *obuf;
1799         int ret = 0, nbuf;
1800         bool input_wakeup = false;
1801
1802
1803 retry:
1804         ret = ipipe_prep(ipipe, flags);
1805         if (ret)
1806                 return ret;
1807
1808         ret = opipe_prep(opipe, flags);
1809         if (ret)
1810                 return ret;
1811
1812         /*
1813          * Potential ABBA deadlock, work around it by ordering lock
1814          * grabbing by pipe info address. Otherwise two different processes
1815          * could deadlock (one doing tee from A -> B, the other from B -> A).
1816          */
1817         pipe_double_lock(ipipe, opipe);
1818
1819         do {
1820                 if (!opipe->readers) {
1821                         send_sig(SIGPIPE, current, 0);
1822                         if (!ret)
1823                                 ret = -EPIPE;
1824                         break;
1825                 }
1826
1827                 if (!ipipe->nrbufs && !ipipe->writers)
1828                         break;
1829
1830                 /*
1831                  * Cannot make any progress, because either the input
1832                  * pipe is empty or the output pipe is full.
1833                  */
1834                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1835                         /* Already processed some buffers, break */
1836                         if (ret)
1837                                 break;
1838
1839                         if (flags & SPLICE_F_NONBLOCK) {
1840                                 ret = -EAGAIN;
1841                                 break;
1842                         }
1843
1844                         /*
1845                          * We raced with another reader/writer and haven't
1846                          * managed to process any buffers.  A zero return
1847                          * value means EOF, so retry instead.
1848                          */
1849                         pipe_unlock(ipipe);
1850                         pipe_unlock(opipe);
1851                         goto retry;
1852                 }
1853
1854                 ibuf = ipipe->bufs + ipipe->curbuf;
1855                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1856                 obuf = opipe->bufs + nbuf;
1857
1858                 if (len >= ibuf->len) {
1859                         /*
1860                          * Simply move the whole buffer from ipipe to opipe
1861                          */
1862                         *obuf = *ibuf;
1863                         ibuf->ops = NULL;
1864                         opipe->nrbufs++;
1865                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1866                         ipipe->nrbufs--;
1867                         input_wakeup = true;
1868                 } else {
1869                         /*
1870                          * Get a reference to this pipe buffer,
1871                          * so we can copy the contents over.
1872                          */
1873                         ibuf->ops->get(ipipe, ibuf);
1874                         *obuf = *ibuf;
1875
1876                         /*
1877                          * Don't inherit the gift flag, we need to
1878                          * prevent multiple steals of this page.
1879                          */
1880                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1881
1882                         obuf->len = len;
1883                         opipe->nrbufs++;
1884                         ibuf->offset += obuf->len;
1885                         ibuf->len -= obuf->len;
1886                 }
1887                 ret += obuf->len;
1888                 len -= obuf->len;
1889         } while (len);
1890
1891         pipe_unlock(ipipe);
1892         pipe_unlock(opipe);
1893
1894         /*
1895          * If we put data in the output pipe, wakeup any potential readers.
1896          */
1897         if (ret > 0)
1898                 wakeup_pipe_readers(opipe);
1899
1900         if (input_wakeup)
1901                 wakeup_pipe_writers(ipipe);
1902
1903         return ret;
1904 }
1905
1906 /*
1907  * Link contents of ipipe to opipe.
1908  */
1909 static int link_pipe(struct pipe_inode_info *ipipe,
1910                      struct pipe_inode_info *opipe,
1911                      size_t len, unsigned int flags)
1912 {
1913         struct pipe_buffer *ibuf, *obuf;
1914         int ret = 0, i = 0, nbuf;
1915
1916         /*
1917          * Potential ABBA deadlock, work around it by ordering lock
1918          * grabbing by pipe info address. Otherwise two different processes
1919          * could deadlock (one doing tee from A -> B, the other from B -> A).
1920          */
1921         pipe_double_lock(ipipe, opipe);
1922
1923         do {
1924                 if (!opipe->readers) {
1925                         send_sig(SIGPIPE, current, 0);
1926                         if (!ret)
1927                                 ret = -EPIPE;
1928                         break;
1929                 }
1930
1931                 /*
1932                  * If we have iterated all input buffers or ran out of
1933                  * output room, break.
1934                  */
1935                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1936                         break;
1937
1938                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1939                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1940
1941                 /*
1942                  * Get a reference to this pipe buffer,
1943                  * so we can copy the contents over.
1944                  */
1945                 ibuf->ops->get(ipipe, ibuf);
1946
1947                 obuf = opipe->bufs + nbuf;
1948                 *obuf = *ibuf;
1949
1950                 /*
1951                  * Don't inherit the gift flag, we need to
1952                  * prevent multiple steals of this page.
1953                  */
1954                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1955
1956                 if (obuf->len > len)
1957                         obuf->len = len;
1958
1959                 opipe->nrbufs++;
1960                 ret += obuf->len;
1961                 len -= obuf->len;
1962                 i++;
1963         } while (len);
1964
1965         /*
1966          * return EAGAIN if we have the potential of some data in the
1967          * future, otherwise just return 0
1968          */
1969         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1970                 ret = -EAGAIN;
1971
1972         pipe_unlock(ipipe);
1973         pipe_unlock(opipe);
1974
1975         /*
1976          * If we put data in the output pipe, wakeup any potential readers.
1977          */
1978         if (ret > 0)
1979                 wakeup_pipe_readers(opipe);
1980
1981         return ret;
1982 }
1983
1984 /*
1985  * This is a tee(1) implementation that works on pipes. It doesn't copy
1986  * any data, it simply references the 'in' pages on the 'out' pipe.
1987  * The 'flags' used are the SPLICE_F_* variants, currently the only
1988  * applicable one is SPLICE_F_NONBLOCK.
1989  */
1990 static long do_tee(struct file *in, struct file *out, size_t len,
1991                    unsigned int flags)
1992 {
1993         struct pipe_inode_info *ipipe = get_pipe_info(in);
1994         struct pipe_inode_info *opipe = get_pipe_info(out);
1995         int ret = -EINVAL;
1996
1997         /*
1998          * Duplicate the contents of ipipe to opipe without actually
1999          * copying the data.
2000          */
2001         if (ipipe && opipe && ipipe != opipe) {
2002                 /*
2003                  * Keep going, unless we encounter an error. The ipipe/opipe
2004                  * ordering doesn't really matter.
2005                  */
2006                 ret = ipipe_prep(ipipe, flags);
2007                 if (!ret) {
2008                         ret = opipe_prep(opipe, flags);
2009                         if (!ret)
2010                                 ret = link_pipe(ipipe, opipe, len, flags);
2011                 }
2012         }
2013
2014         return ret;
2015 }
2016
2017 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2018 {
2019         struct fd in;
2020         int error;
2021
2022         if (unlikely(!len))
2023                 return 0;
2024
2025         error = -EBADF;
2026         in = fdget(fdin);
2027         if (in.file) {
2028                 if (in.file->f_mode & FMODE_READ) {
2029                         struct fd out = fdget(fdout);
2030                         if (out.file) {
2031                                 if (out.file->f_mode & FMODE_WRITE)
2032                                         error = do_tee(in.file, out.file,
2033                                                         len, flags);
2034                                 fdput(out);
2035                         }
2036                 }
2037                 fdput(in);
2038         }
2039
2040         return error;
2041 }