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