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                                    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                                 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         /*
 813          * Check for signal early to make process killable when there are
 814          * always buffers available
 815          */
 816         if (signal_pending(current))
 817                 return -ERESTARTSYS;
 818
 819         while (!pipe->nrbufs) {
 820                 if (!pipe->writers)
 821                         return 0;
 822
 823                 if (!pipe->waiting_writers && sd->num_spliced)
 824                         return 0;
 825
 826                 if (sd->flags & SPLICE_F_NONBLOCK)
 827                         return -EAGAIN;
 828
 829                 if (signal_pending(current))
 830                         return -ERESTARTSYS;
 831
 832                 if (sd->need_wakeup) {
 833                         wakeup_pipe_writers(pipe);
 834                         sd->need_wakeup = false;
 835                 }
 836
 837                 pipe_wait(pipe);
 838         }
 839
 840         return 1;
 841 }
 842
 843 /**
 844  * splice_from_pipe_begin - start splicing from pipe
 845  * @sd:         information about the splice operation
 846  *
 847  * Description:
 848  *    This function should be called before a loop containing
 849  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 850  *    initialize the necessary fields of @sd.
 851  */
 852 static void splice_from_pipe_begin(struct splice_desc *sd)
 853 {
 854         sd->num_spliced = 0;
 855         sd->need_wakeup = false;
 856 }
 857
 858 /**
 859  * splice_from_pipe_end - finish splicing from pipe
 860  * @pipe:       pipe to splice from
 861  * @sd:         information about the splice operation
 862  *
 863  * Description:
 864  *    This function will wake up pipe writers if necessary.  It should
 865  *    be called after a loop containing splice_from_pipe_next() and
 866  *    splice_from_pipe_feed().
 867  */
 868 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 869 {
 870         if (sd->need_wakeup)
 871                 wakeup_pipe_writers(pipe);
 872 }
 873
 874 /**
 875  * __splice_from_pipe - splice data from a pipe to given actor
 876  * @pipe:       pipe to splice from
 877  * @sd:         information to @actor
 878  * @actor:      handler that splices the data
 879  *
 880  * Description:
 881  *    This function does little more than loop over the pipe and call
 882  *    @actor to do the actual moving of a single struct pipe_buffer to
 883  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 884  *    pipe_to_user.
 885  *
 886  */
 887 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 888                            splice_actor *actor)
 889 {
 890         int ret;
 891
 892         splice_from_pipe_begin(sd);
 893         do {
 894                 cond_resched();
 895                 ret = splice_from_pipe_next(pipe, sd);
 896                 if (ret > 0)
 897                         ret = splice_from_pipe_feed(pipe, sd, actor);
 898         } while (ret > 0);
 899         splice_from_pipe_end(pipe, sd);
 900
 901         return sd->num_spliced ? sd->num_spliced : ret;
 902 }
 903 EXPORT_SYMBOL(__splice_from_pipe);
 904
 905 /**
 906  * splice_from_pipe - splice data from a pipe to a file
 907  * @pipe:       pipe to splice from
 908  * @out:        file to splice to
 909  * @ppos:       position in @out
 910  * @len:        how many bytes to splice
 911  * @flags:      splice modifier flags
 912  * @actor:      handler that splices the data
 913  *
 914  * Description:
 915  *    See __splice_from_pipe. This function locks the pipe inode,
 916  *    otherwise it's identical to __splice_from_pipe().
 917  *
 918  */
 919 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 920                          loff_t *ppos, size_t len, unsigned int flags,
 921                          splice_actor *actor)
 922 {
 923         ssize_t ret;
 924         struct splice_desc sd = {
 925                 .total_len = len,
 926                 .flags = flags,
 927                 .pos = *ppos,
 928                 .u.file = out,
 929         };
 930
 931         pipe_lock(pipe);
 932         ret = __splice_from_pipe(pipe, &sd, actor);
 933         pipe_unlock(pipe);
 934
 935         return ret;
 936 }
 937
 938 /**
 939  * iter_file_splice_write - splice data from a pipe to a file
 940  * @pipe:       pipe info
 941  * @out:        file to write to
 942  * @ppos:       position in @out
 943  * @len:        number of bytes to splice
 944  * @flags:      splice modifier flags
 945  *
 946  * Description:
 947  *    Will either move or copy pages (determined by @flags options) from
 948  *    the given pipe inode to the given file.
 949  *    This one is ->write_iter-based.
 950  *
 951  */
 952 ssize_t
 953 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 954                           loff_t *ppos, size_t len, unsigned int flags)
 955 {
 956         struct splice_desc sd = {
 957                 .total_len = len,
 958                 .flags = flags,
 959                 .pos = *ppos,
 960                 .u.file = out,
 961         };
 962         int nbufs = pipe->buffers;
 963         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 964                                         GFP_KERNEL);
 965         ssize_t ret;
 966
 967         if (unlikely(!array))
 968                 return -ENOMEM;
 969
 970         pipe_lock(pipe);
 971
 972         splice_from_pipe_begin(&sd);
 973         while (sd.total_len) {
 974                 struct iov_iter from;
 975                 size_t left;
 976                 int n, idx;
 977
 978                 ret = splice_from_pipe_next(pipe, &sd);
 979                 if (ret <= 0)
 980                         break;
 981
 982                 if (unlikely(nbufs < pipe->buffers)) {
 983                         kfree(array);
 984                         nbufs = pipe->buffers;
 985                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 986                                         GFP_KERNEL);
 987                         if (!array) {
 988                                 ret = -ENOMEM;
 989                                 break;
 990                         }
 991                 }
 992
 993                 /* build the vector */
 994                 left = sd.total_len;
 995                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
 996                         struct pipe_buffer *buf = pipe->bufs + idx;
 997                         size_t this_len = buf->len;
 998
 999                         if (this_len > left)
1000                                 this_len = left;
1001
1002                         if (idx == pipe->buffers - 1)
1003                                 idx = -1;
1004
1005                         ret = buf->ops->confirm(pipe, buf);
1006                         if (unlikely(ret)) {
1007                                 if (ret == -ENODATA)
1008                                         ret = 0;
1009                                 goto done;
1010                         }
1011
1012                         array[n].bv_page = buf->page;
1013                         array[n].bv_len = this_len;
1014                         array[n].bv_offset = buf->offset;
1015                         left -= this_len;
1016                 }
1017
1018                 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
1019                               sd.total_len - left);
1020                 ret = vfs_iter_write(out, &from, &sd.pos);
1021                 if (ret <= 0)
1022                         break;
1023
1024                 sd.num_spliced += ret;
1025                 sd.total_len -= ret;
1026                 *ppos = sd.pos;
1027
1028                 /* dismiss the fully eaten buffers, adjust the partial one */
1029                 while (ret) {
1030                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1031                         if (ret >= buf->len) {
1032                                 const struct pipe_buf_operations *ops = buf->ops;
1033                                 ret -= buf->len;
1034                                 buf->len = 0;
1035                                 buf->ops = NULL;
1036                                 ops->release(pipe, buf);
1037                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1038                                 pipe->nrbufs--;
1039                                 if (pipe->files)
1040                                         sd.need_wakeup = true;
1041                         } else {
1042                                 buf->offset += ret;
1043                                 buf->len -= ret;
1044                                 ret = 0;
1045                         }
1046                 }
1047         }
1048 done:
1049         kfree(array);
1050         splice_from_pipe_end(pipe, &sd);
1051
1052         pipe_unlock(pipe);
1053
1054         if (sd.num_spliced)
1055                 ret = sd.num_spliced;
1056
1057         return ret;
1058 }
1059
1060 EXPORT_SYMBOL(iter_file_splice_write);
1061
1062 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1063                           struct splice_desc *sd)
1064 {
1065         int ret;
1066         void *data;
1067         loff_t tmp = sd->pos;
1068
1069         data = kmap(buf->page);
1070         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1071         kunmap(buf->page);
1072
1073         return ret;
1074 }
1075
1076 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1077                                          struct file *out, loff_t *ppos,
1078                                          size_t len, unsigned int flags)
1079 {
1080         ssize_t ret;
1081
1082         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1083         if (ret > 0)
1084                 *ppos += ret;
1085
1086         return ret;
1087 }
1088
1089 /**
1090  * generic_splice_sendpage - splice data from a pipe to a socket
1091  * @pipe:       pipe to splice from
1092  * @out:        socket to write to
1093  * @ppos:       position in @out
1094  * @len:        number of bytes to splice
1095  * @flags:      splice modifier flags
1096  *
1097  * Description:
1098  *    Will send @len bytes from the pipe to a network socket. No data copying
1099  *    is involved.
1100  *
1101  */
1102 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1103                                 loff_t *ppos, size_t len, unsigned int flags)
1104 {
1105         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1106 }
1107
1108 EXPORT_SYMBOL(generic_splice_sendpage);
1109
1110 /*
1111  * Attempt to initiate a splice from pipe to file.
1112  */
1113 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1114                            loff_t *ppos, size_t len, unsigned int flags)
1115 {
1116         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1117                                 loff_t *, size_t, unsigned int);
1118
1119         if (out->f_op->splice_write)
1120                 splice_write = out->f_op->splice_write;
1121         else
1122                 splice_write = default_file_splice_write;
1123
1124         return splice_write(pipe, out, ppos, len, flags);
1125 }
1126
1127 /*
1128  * Attempt to initiate a splice from a file to a pipe.
1129  */
1130 static long do_splice_to(struct file *in, loff_t *ppos,
1131                          struct pipe_inode_info *pipe, size_t len,
1132                          unsigned int flags)
1133 {
1134         ssize_t (*splice_read)(struct file *, loff_t *,
1135                                struct pipe_inode_info *, size_t, unsigned int);
1136         int ret;
1137
1138         if (unlikely(!(in->f_mode & FMODE_READ)))
1139                 return -EBADF;
1140
1141         ret = rw_verify_area(READ, in, ppos, len);
1142         if (unlikely(ret < 0))
1143                 return ret;
1144
1145         if (in->f_op->splice_read)
1146                 splice_read = in->f_op->splice_read;
1147         else
1148                 splice_read = default_file_splice_read;
1149
1150         return splice_read(in, ppos, pipe, len, flags);
1151 }
1152
1153 /**
1154  * splice_direct_to_actor - splices data directly between two non-pipes
1155  * @in:         file to splice from
1156  * @sd:         actor information on where to splice to
1157  * @actor:      handles the data splicing
1158  *
1159  * Description:
1160  *    This is a special case helper to splice directly between two
1161  *    points, without requiring an explicit pipe. Internally an allocated
1162  *    pipe is cached in the process, and reused during the lifetime of
1163  *    that process.
1164  *
1165  */
1166 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1167                                splice_direct_actor *actor)
1168 {
1169         struct pipe_inode_info *pipe;
1170         long ret, bytes;
1171         umode_t i_mode;
1172         size_t len;
1173         int i, flags, more;
1174
1175         /*
1176          * We require the input being a regular file, as we don't want to
1177          * randomly drop data for eg socket -> socket splicing. Use the
1178          * piped splicing for that!
1179          */
1180         i_mode = file_inode(in)->i_mode;
1181         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1182                 return -EINVAL;
1183
1184         /*
1185          * neither in nor out is a pipe, setup an internal pipe attached to
1186          * 'out' and transfer the wanted data from 'in' to 'out' through that
1187          */
1188         pipe = current->splice_pipe;
1189         if (unlikely(!pipe)) {
1190                 pipe = alloc_pipe_info();
1191                 if (!pipe)
1192                         return -ENOMEM;
1193
1194                 /*
1195                  * We don't have an immediate reader, but we'll read the stuff
1196                  * out of the pipe right after the splice_to_pipe(). So set
1197                  * PIPE_READERS appropriately.
1198                  */
1199                 pipe->readers = 1;
1200
1201                 current->splice_pipe = pipe;
1202         }
1203
1204         /*
1205          * Do the splice.
1206          */
1207         ret = 0;
1208         bytes = 0;
1209         len = sd->total_len;
1210         flags = sd->flags;
1211
1212         /*
1213          * Don't block on output, we have to drain the direct pipe.
1214          */
1215         sd->flags &= ~SPLICE_F_NONBLOCK;
1216         more = sd->flags & SPLICE_F_MORE;
1217
1218         while (len) {
1219                 size_t read_len;
1220                 loff_t pos = sd->pos, prev_pos = pos;
1221
1222                 ret = do_splice_to(in, &pos, pipe, len, flags);
1223                 if (unlikely(ret <= 0))
1224                         goto out_release;
1225
1226                 read_len = ret;
1227                 sd->total_len = read_len;
1228
1229                 /*
1230                  * If more data is pending, set SPLICE_F_MORE
1231                  * If this is the last data and SPLICE_F_MORE was not set
1232                  * initially, clears it.
1233                  */
1234                 if (read_len < len)
1235                         sd->flags |= SPLICE_F_MORE;
1236                 else if (!more)
1237                         sd->flags &= ~SPLICE_F_MORE;
1238                 /*
1239                  * NOTE: nonblocking mode only applies to the input. We
1240                  * must not do the output in nonblocking mode as then we
1241                  * could get stuck data in the internal pipe:
1242                  */
1243                 ret = actor(pipe, sd);
1244                 if (unlikely(ret <= 0)) {
1245                         sd->pos = prev_pos;
1246                         goto out_release;
1247                 }
1248
1249                 bytes += ret;
1250                 len -= ret;
1251                 sd->pos = pos;
1252
1253                 if (ret < read_len) {
1254                         sd->pos = prev_pos + ret;
1255                         goto out_release;
1256                 }
1257         }
1258
1259 done:
1260         pipe->nrbufs = pipe->curbuf = 0;
1261         file_accessed(in);
1262         return bytes;
1263
1264 out_release:
1265         /*
1266          * If we did an incomplete transfer we must release
1267          * the pipe buffers in question:
1268          */
1269         for (i = 0; i < pipe->buffers; i++) {
1270                 struct pipe_buffer *buf = pipe->bufs + i;
1271
1272                 if (buf->ops) {
1273                         buf->ops->release(pipe, buf);
1274                         buf->ops = NULL;
1275                 }
1276         }
1277
1278         if (!bytes)
1279                 bytes = ret;
1280
1281         goto done;
1282 }
1283 EXPORT_SYMBOL(splice_direct_to_actor);
1284
1285 static int direct_splice_actor(struct pipe_inode_info *pipe,
1286                                struct splice_desc *sd)
1287 {
1288         struct file *file = sd->u.file;
1289
1290         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1291                               sd->flags);
1292 }
1293
1294 /**
1295  * do_splice_direct - splices data directly between two files
1296  * @in:         file to splice from
1297  * @ppos:       input file offset
1298  * @out:        file to splice to
1299  * @opos:       output file offset
1300  * @len:        number of bytes to splice
1301  * @flags:      splice modifier flags
1302  *
1303  * Description:
1304  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1305  *    doing it in the application would incur an extra system call
1306  *    (splice in + splice out, as compared to just sendfile()). So this helper
1307  *    can splice directly through a process-private pipe.
1308  *
1309  */
1310 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1311                       loff_t *opos, size_t len, unsigned int flags)
1312 {
1313         struct splice_desc sd = {
1314                 .len            = len,
1315                 .total_len      = len,
1316                 .flags          = flags,
1317                 .pos            = *ppos,
1318                 .u.file         = out,
1319                 .opos           = opos,
1320         };
1321         long ret;
1322
1323         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1324                 return -EBADF;
1325
1326         if (unlikely(out->f_flags & O_APPEND))
1327                 return -EINVAL;
1328
1329         ret = rw_verify_area(WRITE, out, opos, len);
1330         if (unlikely(ret < 0))
1331                 return ret;
1332
1333         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1334         if (ret > 0)
1335                 *ppos = sd.pos;
1336
1337         return ret;
1338 }
1339 EXPORT_SYMBOL(do_splice_direct);
1340
1341 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1342                                struct pipe_inode_info *opipe,
1343                                size_t len, unsigned int flags);
1344
1345 /*
1346  * Determine where to splice to/from.
1347  */
1348 static long do_splice(struct file *in, loff_t __user *off_in,
1349                       struct file *out, loff_t __user *off_out,
1350                       size_t len, unsigned int flags)
1351 {
1352         struct pipe_inode_info *ipipe;
1353         struct pipe_inode_info *opipe;
1354         loff_t offset;
1355         long ret;
1356
1357         ipipe = get_pipe_info(in);
1358         opipe = get_pipe_info(out);
1359
1360         if (ipipe && opipe) {
1361                 if (off_in || off_out)
1362                         return -ESPIPE;
1363
1364                 if (!(in->f_mode & FMODE_READ))
1365                         return -EBADF;
1366
1367                 if (!(out->f_mode & FMODE_WRITE))
1368                         return -EBADF;
1369
1370                 /* Splicing to self would be fun, but... */
1371                 if (ipipe == opipe)
1372                         return -EINVAL;
1373
1374                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1375         }
1376
1377         if (ipipe) {
1378                 if (off_in)
1379                         return -ESPIPE;
1380                 if (off_out) {
1381                         if (!(out->f_mode & FMODE_PWRITE))
1382                                 return -EINVAL;
1383                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1384                                 return -EFAULT;
1385                 } else {
1386                         offset = out->f_pos;
1387                 }
1388
1389                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1390                         return -EBADF;
1391
1392                 if (unlikely(out->f_flags & O_APPEND))
1393                         return -EINVAL;
1394
1395                 ret = rw_verify_area(WRITE, out, &offset, len);
1396                 if (unlikely(ret < 0))
1397                         return ret;
1398
1399                 file_start_write(out);
1400                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1401                 file_end_write(out);
1402
1403                 if (!off_out)
1404                         out->f_pos = offset;
1405                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1406                         ret = -EFAULT;
1407
1408                 return ret;
1409         }
1410
1411         if (opipe) {
1412                 if (off_out)
1413                         return -ESPIPE;
1414                 if (off_in) {
1415                         if (!(in->f_mode & FMODE_PREAD))
1416                                 return -EINVAL;
1417                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1418                                 return -EFAULT;
1419                 } else {
1420                         offset = in->f_pos;
1421                 }
1422
1423                 ret = do_splice_to(in, &offset, opipe, len, flags);
1424
1425                 if (!off_in)
1426                         in->f_pos = offset;
1427                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1428                         ret = -EFAULT;
1429
1430                 return ret;
1431         }
1432
1433         return -EINVAL;
1434 }
1435
1436 /*
1437  * Map an iov into an array of pages and offset/length tupples. With the
1438  * partial_page structure, we can map several non-contiguous ranges into
1439  * our ones pages[] map instead of splitting that operation into pieces.
1440  * Could easily be exported as a generic helper for other users, in which
1441  * case one would probably want to add a 'max_nr_pages' parameter as well.
1442  */
1443 static int get_iovec_page_array(const struct iovec __user *iov,
1444                                 unsigned int nr_vecs, struct page **pages,
1445                                 struct partial_page *partial, bool aligned,
1446                                 unsigned int pipe_buffers)
1447 {
1448         int buffers = 0, error = 0;
1449
1450         while (nr_vecs) {
1451                 unsigned long off, npages;
1452                 struct iovec entry;
1453                 void __user *base;
1454                 size_t len;
1455                 int i;
1456
1457                 error = -EFAULT;
1458                 if (copy_from_user(&entry, iov, sizeof(entry)))
1459                         break;
1460
1461                 base = entry.iov_base;
1462                 len = entry.iov_len;
1463
1464                 /*
1465                  * Sanity check this iovec. 0 read succeeds.
1466                  */
1467                 error = 0;
1468                 if (unlikely(!len))
1469                         break;
1470                 error = -EFAULT;
1471                 if (!access_ok(VERIFY_READ, base, len))
1472                         break;
1473
1474                 /*
1475                  * Get this base offset and number of pages, then map
1476                  * in the user pages.
1477                  */
1478                 off = (unsigned long) base & ~PAGE_MASK;
1479
1480                 /*
1481                  * If asked for alignment, the offset must be zero and the
1482                  * length a multiple of the PAGE_SIZE.
1483                  */
1484                 error = -EINVAL;
1485                 if (aligned && (off || len & ~PAGE_MASK))
1486                         break;
1487
1488                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1489                 if (npages > pipe_buffers - buffers)
1490                         npages = pipe_buffers - buffers;
1491
1492                 error = get_user_pages_fast((unsigned long)base, npages,
1493                                         0, &pages[buffers]);
1494
1495                 if (unlikely(error <= 0))
1496                         break;
1497
1498                 /*
1499                  * Fill this contiguous range into the partial page map.
1500                  */
1501                 for (i = 0; i < error; i++) {
1502                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1503
1504                         partial[buffers].offset = off;
1505                         partial[buffers].len = plen;
1506
1507                         off = 0;
1508                         len -= plen;
1509                         buffers++;
1510                 }
1511
1512                 /*
1513                  * We didn't complete this iov, stop here since it probably
1514                  * means we have to move some of this into a pipe to
1515                  * be able to continue.
1516                  */
1517                 if (len)
1518                         break;
1519
1520                 /*
1521                  * Don't continue if we mapped fewer pages than we asked for,
1522                  * or if we mapped the max number of pages that we have
1523                  * room for.
1524                  */
1525                 if (error < npages || buffers == pipe_buffers)
1526                         break;
1527
1528                 nr_vecs--;
1529                 iov++;
1530         }
1531
1532         if (buffers)
1533                 return buffers;
1534
1535         return error;
1536 }
1537
1538 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1539                         struct splice_desc *sd)
1540 {
1541         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1542         return n == sd->len ? n : -EFAULT;
1543 }
1544
1545 /*
1546  * For lack of a better implementation, implement vmsplice() to userspace
1547  * as a simple copy of the pipes pages to the user iov.
1548  */
1549 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1550                              unsigned long nr_segs, unsigned int flags)
1551 {
1552         struct pipe_inode_info *pipe;
1553         struct splice_desc sd;
1554         long ret;
1555         struct iovec iovstack[UIO_FASTIOV];
1556         struct iovec *iov = iovstack;
1557         struct iov_iter iter;
1558
1559         pipe = get_pipe_info(file);
1560         if (!pipe)
1561                 return -EBADF;
1562
1563         ret = import_iovec(READ, uiov, nr_segs,
1564                            ARRAY_SIZE(iovstack), &iov, &iter);
1565         if (ret < 0)
1566                 return ret;
1567
1568         sd.total_len = iov_iter_count(&iter);
1569         sd.len = 0;
1570         sd.flags = flags;
1571         sd.u.data = &iter;
1572         sd.pos = 0;
1573
1574         if (sd.total_len) {
1575                 pipe_lock(pipe);
1576                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1577                 pipe_unlock(pipe);
1578         }
1579
1580         kfree(iov);
1581         return ret;
1582 }
1583
1584 /*
1585  * vmsplice splices a user address range into a pipe. It can be thought of
1586  * as splice-from-memory, where the regular splice is splice-from-file (or
1587  * to file). In both cases the output is a pipe, naturally.
1588  */
1589 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1590                              unsigned long nr_segs, unsigned int flags)
1591 {
1592         struct pipe_inode_info *pipe;
1593         struct page *pages[PIPE_DEF_BUFFERS];
1594         struct partial_page partial[PIPE_DEF_BUFFERS];
1595         struct splice_pipe_desc spd = {
1596                 .pages = pages,
1597                 .partial = partial,
1598                 .nr_pages_max = PIPE_DEF_BUFFERS,
1599                 .flags = flags,
1600                 .ops = &user_page_pipe_buf_ops,
1601                 .spd_release = spd_release_page,
1602         };
1603         long ret;
1604
1605         pipe = get_pipe_info(file);
1606         if (!pipe)
1607                 return -EBADF;
1608
1609         if (splice_grow_spd(pipe, &spd))
1610                 return -ENOMEM;
1611
1612         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1613                                             spd.partial, false,
1614                                             spd.nr_pages_max);
1615         if (spd.nr_pages <= 0)
1616                 ret = spd.nr_pages;
1617         else
1618                 ret = splice_to_pipe(pipe, &spd);
1619
1620         splice_shrink_spd(&spd);
1621         return ret;
1622 }
1623
1624 /*
1625  * Note that vmsplice only really supports true splicing _from_ user memory
1626  * to a pipe, not the other way around. Splicing from user memory is a simple
1627  * operation that can be supported without any funky alignment restrictions
1628  * or nasty vm tricks. We simply map in the user memory and fill them into
1629  * a pipe. The reverse isn't quite as easy, though. There are two possible
1630  * solutions for that:
1631  *
1632  *      - memcpy() the data internally, at which point we might as well just
1633  *        do a regular read() on the buffer anyway.
1634  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1635  *        has restriction limitations on both ends of the pipe).
1636  *
1637  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1638  *
1639  */
1640 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1641                 unsigned long, nr_segs, unsigned int, flags)
1642 {
1643         struct fd f;
1644         long error;
1645
1646         if (unlikely(nr_segs > UIO_MAXIOV))
1647                 return -EINVAL;
1648         else if (unlikely(!nr_segs))
1649                 return 0;
1650
1651         error = -EBADF;
1652         f = fdget(fd);
1653         if (f.file) {
1654                 if (f.file->f_mode & FMODE_WRITE)
1655                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1656                 else if (f.file->f_mode & FMODE_READ)
1657                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1658
1659                 fdput(f);
1660         }
1661
1662         return error;
1663 }
1664
1665 #ifdef CONFIG_COMPAT
1666 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1667                     unsigned int, nr_segs, unsigned int, flags)
1668 {
1669         unsigned i;
1670         struct iovec __user *iov;
1671         if (nr_segs > UIO_MAXIOV)
1672                 return -EINVAL;
1673         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1674         for (i = 0; i < nr_segs; i++) {
1675                 struct compat_iovec v;
1676                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1677                     get_user(v.iov_len, &iov32[i].iov_len) ||
1678                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1679                     put_user(v.iov_len, &iov[i].iov_len))
1680                         return -EFAULT;
1681         }
1682         return sys_vmsplice(fd, iov, nr_segs, flags);
1683 }
1684 #endif
1685
1686 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1687                 int, fd_out, loff_t __user *, off_out,
1688                 size_t, len, unsigned int, flags)
1689 {
1690         struct fd in, out;
1691         long error;
1692
1693         if (unlikely(!len))
1694                 return 0;
1695
1696         error = -EBADF;
1697         in = fdget(fd_in);
1698         if (in.file) {
1699                 if (in.file->f_mode & FMODE_READ) {
1700                         out = fdget(fd_out);
1701                         if (out.file) {
1702                                 if (out.file->f_mode & FMODE_WRITE)
1703                                         error = do_splice(in.file, off_in,
1704                                                           out.file, off_out,
1705                                                           len, flags);
1706                                 fdput(out);
1707                         }
1708                 }
1709                 fdput(in);
1710         }
1711         return error;
1712 }
1713
1714 /*
1715  * Make sure there's data to read. Wait for input if we can, otherwise
1716  * return an appropriate error.
1717  */
1718 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1719 {
1720         int ret;
1721
1722         /*
1723          * Check ->nrbufs without the inode lock first. This function
1724          * is speculative anyways, so missing one is ok.
1725          */
1726         if (pipe->nrbufs)
1727                 return 0;
1728
1729         ret = 0;
1730         pipe_lock(pipe);
1731
1732         while (!pipe->nrbufs) {
1733                 if (signal_pending(current)) {
1734                         ret = -ERESTARTSYS;
1735                         break;
1736                 }
1737                 if (!pipe->writers)
1738                         break;
1739                 if (!pipe->waiting_writers) {
1740                         if (flags & SPLICE_F_NONBLOCK) {
1741                                 ret = -EAGAIN;
1742                                 break;
1743                         }
1744                 }
1745                 pipe_wait(pipe);
1746         }
1747
1748         pipe_unlock(pipe);
1749         return ret;
1750 }
1751
1752 /*
1753  * Make sure there's writeable room. Wait for room if we can, otherwise
1754  * return an appropriate error.
1755  */
1756 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1757 {
1758         int ret;
1759
1760         /*
1761          * Check ->nrbufs without the inode lock first. This function
1762          * is speculative anyways, so missing one is ok.
1763          */
1764         if (pipe->nrbufs < pipe->buffers)
1765                 return 0;
1766
1767         ret = 0;
1768         pipe_lock(pipe);
1769
1770         while (pipe->nrbufs >= pipe->buffers) {
1771                 if (!pipe->readers) {
1772                         send_sig(SIGPIPE, current, 0);
1773                         ret = -EPIPE;
1774                         break;
1775                 }
1776                 if (flags & SPLICE_F_NONBLOCK) {
1777                         ret = -EAGAIN;
1778                         break;
1779                 }
1780                 if (signal_pending(current)) {
1781                         ret = -ERESTARTSYS;
1782                         break;
1783                 }
1784                 pipe->waiting_writers++;
1785                 pipe_wait(pipe);
1786                 pipe->waiting_writers--;
1787         }
1788
1789         pipe_unlock(pipe);
1790         return ret;
1791 }
1792
1793 /*
1794  * Splice contents of ipipe to opipe.
1795  */
1796 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1797                                struct pipe_inode_info *opipe,
1798                                size_t len, unsigned int flags)
1799 {
1800         struct pipe_buffer *ibuf, *obuf;
1801         int ret = 0, nbuf;
1802         bool input_wakeup = false;
1803
1804
1805 retry:
1806         ret = ipipe_prep(ipipe, flags);
1807         if (ret)
1808                 return ret;
1809
1810         ret = opipe_prep(opipe, flags);
1811         if (ret)
1812                 return ret;
1813
1814         /*
1815          * Potential ABBA deadlock, work around it by ordering lock
1816          * grabbing by pipe info address. Otherwise two different processes
1817          * could deadlock (one doing tee from A -> B, the other from B -> A).
1818          */
1819         pipe_double_lock(ipipe, opipe);
1820
1821         do {
1822                 if (!opipe->readers) {
1823                         send_sig(SIGPIPE, current, 0);
1824                         if (!ret)
1825                                 ret = -EPIPE;
1826                         break;
1827                 }
1828
1829                 if (!ipipe->nrbufs && !ipipe->writers)
1830                         break;
1831
1832                 /*
1833                  * Cannot make any progress, because either the input
1834                  * pipe is empty or the output pipe is full.
1835                  */
1836                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1837                         /* Already processed some buffers, break */
1838                         if (ret)
1839                                 break;
1840
1841                         if (flags & SPLICE_F_NONBLOCK) {
1842                                 ret = -EAGAIN;
1843                                 break;
1844                         }
1845
1846                         /*
1847                          * We raced with another reader/writer and haven't
1848                          * managed to process any buffers.  A zero return
1849                          * value means EOF, so retry instead.
1850                          */
1851                         pipe_unlock(ipipe);
1852                         pipe_unlock(opipe);
1853                         goto retry;
1854                 }
1855
1856                 ibuf = ipipe->bufs + ipipe->curbuf;
1857                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1858                 obuf = opipe->bufs + nbuf;
1859
1860                 if (len >= ibuf->len) {
1861                         /*
1862                          * Simply move the whole buffer from ipipe to opipe
1863                          */
1864                         *obuf = *ibuf;
1865                         ibuf->ops = NULL;
1866                         opipe->nrbufs++;
1867                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1868                         ipipe->nrbufs--;
1869                         input_wakeup = true;
1870                 } else {
1871                         /*
1872                          * Get a reference to this pipe buffer,
1873                          * so we can copy the contents over.
1874                          */
1875                         ibuf->ops->get(ipipe, ibuf);
1876                         *obuf = *ibuf;
1877
1878                         /*
1879                          * Don't inherit the gift flag, we need to
1880                          * prevent multiple steals of this page.
1881                          */
1882                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1883
1884                         obuf->len = len;
1885                         opipe->nrbufs++;
1886                         ibuf->offset += obuf->len;
1887                         ibuf->len -= obuf->len;
1888                 }
1889                 ret += obuf->len;
1890                 len -= obuf->len;
1891         } while (len);
1892
1893         pipe_unlock(ipipe);
1894         pipe_unlock(opipe);
1895
1896         /*
1897          * If we put data in the output pipe, wakeup any potential readers.
1898          */
1899         if (ret > 0)
1900                 wakeup_pipe_readers(opipe);
1901
1902         if (input_wakeup)
1903                 wakeup_pipe_writers(ipipe);
1904
1905         return ret;
1906 }
1907
1908 /*
1909  * Link contents of ipipe to opipe.
1910  */
1911 static int link_pipe(struct pipe_inode_info *ipipe,
1912                      struct pipe_inode_info *opipe,
1913                      size_t len, unsigned int flags)
1914 {
1915         struct pipe_buffer *ibuf, *obuf;
1916         int ret = 0, i = 0, nbuf;
1917
1918         /*
1919          * Potential ABBA deadlock, work around it by ordering lock
1920          * grabbing by pipe info address. Otherwise two different processes
1921          * could deadlock (one doing tee from A -> B, the other from B -> A).
1922          */
1923         pipe_double_lock(ipipe, opipe);
1924
1925         do {
1926                 if (!opipe->readers) {
1927                         send_sig(SIGPIPE, current, 0);
1928                         if (!ret)
1929                                 ret = -EPIPE;
1930                         break;
1931                 }
1932
1933                 /*
1934                  * If we have iterated all input buffers or ran out of
1935                  * output room, break.
1936                  */
1937                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1938                         break;
1939
1940                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1941                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1942
1943                 /*
1944                  * Get a reference to this pipe buffer,
1945                  * so we can copy the contents over.
1946                  */
1947                 ibuf->ops->get(ipipe, ibuf);
1948
1949                 obuf = opipe->bufs + nbuf;
1950                 *obuf = *ibuf;
1951
1952                 /*
1953                  * Don't inherit the gift flag, we need to
1954                  * prevent multiple steals of this page.
1955                  */
1956                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1957
1958                 if (obuf->len > len)
1959                         obuf->len = len;
1960
1961                 opipe->nrbufs++;
1962                 ret += obuf->len;
1963                 len -= obuf->len;
1964                 i++;
1965         } while (len);
1966
1967         /*
1968          * return EAGAIN if we have the potential of some data in the
1969          * future, otherwise just return 0
1970          */
1971         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1972                 ret = -EAGAIN;
1973
1974         pipe_unlock(ipipe);
1975         pipe_unlock(opipe);
1976
1977         /*
1978          * If we put data in the output pipe, wakeup any potential readers.
1979          */
1980         if (ret > 0)
1981                 wakeup_pipe_readers(opipe);
1982
1983         return ret;
1984 }
1985
1986 /*
1987  * This is a tee(1) implementation that works on pipes. It doesn't copy
1988  * any data, it simply references the 'in' pages on the 'out' pipe.
1989  * The 'flags' used are the SPLICE_F_* variants, currently the only
1990  * applicable one is SPLICE_F_NONBLOCK.
1991  */
1992 static long do_tee(struct file *in, struct file *out, size_t len,
1993                    unsigned int flags)
1994 {
1995         struct pipe_inode_info *ipipe = get_pipe_info(in);
1996         struct pipe_inode_info *opipe = get_pipe_info(out);
1997         int ret = -EINVAL;
1998
1999         /*
2000          * Duplicate the contents of ipipe to opipe without actually
2001          * copying the data.
2002          */
2003         if (ipipe && opipe && ipipe != opipe) {
2004                 /*
2005                  * Keep going, unless we encounter an error. The ipipe/opipe
2006                  * ordering doesn't really matter.
2007                  */
2008                 ret = ipipe_prep(ipipe, flags);
2009                 if (!ret) {
2010                         ret = opipe_prep(opipe, flags);
2011                         if (!ret)
2012                                 ret = link_pipe(ipipe, opipe, len, flags);
2013                 }
2014         }
2015
2016         return ret;
2017 }
2018
2019 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2020 {
2021         struct fd in;
2022         int error;
2023
2024         if (unlikely(!len))
2025                 return 0;
2026
2027         error = -EBADF;
2028         in = fdget(fdin);
2029         if (in.file) {
2030                 if (in.file->f_mode & FMODE_READ) {
2031                         struct fd out = fdget(fdout);
2032                         if (out.file) {
2033                                 if (out.file->f_mode & FMODE_WRITE)
2034                                         error = do_tee(in.file, out.file,
2035                                                         len, flags);
2036                                 fdput(out);
2037                         }
2038                 }
2039                 fdput(in);
2040         }
2041
2042         return error;
2043 }