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