fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  17  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/fsnotify.h>
  34 #include <linux/security.h>
  35 #include <linux/gfp.h>
  36 #include <linux/net.h>
  37 #include <linux/socket.h>
  38 #include <linux/sched/signal.h>
  39
  40 #include "internal.h"
  41
  42 /*
  43  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  44  * indicate they support non-blocking reads or writes, we must clear it
  45  * here if set to avoid blocking other users of this pipe if splice is
  46  * being done on it.
  47  */
  48 static noinline void noinline pipe_clear_nowait(struct file *file)
  49 {
  50         fmode_t fmode = READ_ONCE(file->f_mode);
  51
  52         do {
  53                 if (!(fmode & FMODE_NOWAIT))
  54                         break;
  55         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  56 }
  57
  58 /*
  59  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  60  * a vm helper function, it's already simplified quite a bit by the
  61  * addition of remove_mapping(). If success is returned, the caller may
  62  * attempt to reuse this page for another destination.
  63  */
  64 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  65                 struct pipe_buffer *buf)
  66 {
  67         struct folio *folio = page_folio(buf->page);
  68         struct address_space *mapping;
  69
  70         folio_lock(folio);
  71
  72         mapping = folio_mapping(folio);
  73         if (mapping) {
  74                 WARN_ON(!folio_test_uptodate(folio));
  75
  76                 /*
  77                  * At least for ext2 with nobh option, we need to wait on
  78                  * writeback completing on this folio, since we'll remove it
  79                  * from the pagecache.  Otherwise truncate wont wait on the
  80                  * folio, allowing the disk blocks to be reused by someone else
  81                  * before we actually wrote our data to them. fs corruption
  82                  * ensues.
  83                  */
  84                 folio_wait_writeback(folio);
  85
  86                 if (folio_has_private(folio) &&
  87                     !filemap_release_folio(folio, GFP_KERNEL))
  88                         goto out_unlock;
  89
  90                 /*
  91                  * If we succeeded in removing the mapping, set LRU flag
  92                  * and return good.
  93                  */
  94                 if (remove_mapping(mapping, folio)) {
  95                         buf->flags |= PIPE_BUF_FLAG_LRU;
  96                         return true;
  97                 }
  98         }
  99
 100         /*
 101          * Raced with truncate or failed to remove folio from current
 102          * address space, unlock and return failure.
 103          */
 104 out_unlock:
 105         folio_unlock(folio);
 106         return false;
 107 }
 108
 109 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 110                                         struct pipe_buffer *buf)
 111 {
 112         put_page(buf->page);
 113         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 114 }
 115
 116 /*
 117  * Check whether the contents of buf is OK to access. Since the content
 118  * is a page cache page, IO may be in flight.
 119  */
 120 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 121                                        struct pipe_buffer *buf)
 122 {
 123         struct page *page = buf->page;
 124         int err;
 125
 126         if (!PageUptodate(page)) {
 127                 lock_page(page);
 128
 129                 /*
 130                  * Page got truncated/unhashed. This will cause a 0-byte
 131                  * splice, if this is the first page.
 132                  */
 133                 if (!page->mapping) {
 134                         err = -ENODATA;
 135                         goto error;
 136                 }
 137
 138                 /*
 139                  * Uh oh, read-error from disk.
 140                  */
 141                 if (!PageUptodate(page)) {
 142                         err = -EIO;
 143                         goto error;
 144                 }
 145
 146                 /*
 147                  * Page is ok afterall, we are done.
 148                  */
 149                 unlock_page(page);
 150         }
 151
 152         return 0;
 153 error:
 154         unlock_page(page);
 155         return err;
 156 }
 157
 158 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 159         .confirm        = page_cache_pipe_buf_confirm,
 160         .release        = page_cache_pipe_buf_release,
 161         .try_steal      = page_cache_pipe_buf_try_steal,
 162         .get            = generic_pipe_buf_get,
 163 };
 164
 165 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 166                 struct pipe_buffer *buf)
 167 {
 168         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 169                 return false;
 170
 171         buf->flags |= PIPE_BUF_FLAG_LRU;
 172         return generic_pipe_buf_try_steal(pipe, buf);
 173 }
 174
 175 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 176         .release        = page_cache_pipe_buf_release,
 177         .try_steal      = user_page_pipe_buf_try_steal,
 178         .get            = generic_pipe_buf_get,
 179 };
 180
 181 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 182 {
 183         smp_mb();
 184         if (waitqueue_active(&pipe->rd_wait))
 185                 wake_up_interruptible(&pipe->rd_wait);
 186         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 187 }
 188
 189 /**
 190  * splice_to_pipe - fill passed data into a pipe
 191  * @pipe:       pipe to fill
 192  * @spd:        data to fill
 193  *
 194  * Description:
 195  *    @spd contains a map of pages and len/offset tuples, along with
 196  *    the struct pipe_buf_operations associated with these pages. This
 197  *    function will link that data to the pipe.
 198  *
 199  */
 200 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 201                        struct splice_pipe_desc *spd)
 202 {
 203         unsigned int spd_pages = spd->nr_pages;
 204         unsigned int tail = pipe->tail;
 205         unsigned int head = pipe->head;
 206         unsigned int mask = pipe->ring_size - 1;
 207         int ret = 0, page_nr = 0;
 208
 209         if (!spd_pages)
 210                 return 0;
 211
 212         if (unlikely(!pipe->readers)) {
 213                 send_sig(SIGPIPE, current, 0);
 214                 ret = -EPIPE;
 215                 goto out;
 216         }
 217
 218         while (!pipe_full(head, tail, pipe->max_usage)) {
 219                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 220
 221                 buf->page = spd->pages[page_nr];
 222                 buf->offset = spd->partial[page_nr].offset;
 223                 buf->len = spd->partial[page_nr].len;
 224                 buf->private = spd->partial[page_nr].private;
 225                 buf->ops = spd->ops;
 226                 buf->flags = 0;
 227
 228                 head++;
 229                 pipe->head = head;
 230                 page_nr++;
 231                 ret += buf->len;
 232
 233                 if (!--spd->nr_pages)
 234                         break;
 235         }
 236
 237         if (!ret)
 238                 ret = -EAGAIN;
 239
 240 out:
 241         while (page_nr < spd_pages)
 242                 spd->spd_release(spd, page_nr++);
 243
 244         return ret;
 245 }
 246 EXPORT_SYMBOL_GPL(splice_to_pipe);
 247
 248 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 249 {
 250         unsigned int head = pipe->head;
 251         unsigned int tail = pipe->tail;
 252         unsigned int mask = pipe->ring_size - 1;
 253         int ret;
 254
 255         if (unlikely(!pipe->readers)) {
 256                 send_sig(SIGPIPE, current, 0);
 257                 ret = -EPIPE;
 258         } else if (pipe_full(head, tail, pipe->max_usage)) {
 259                 ret = -EAGAIN;
 260         } else {
 261                 pipe->bufs[head & mask] = *buf;
 262                 pipe->head = head + 1;
 263                 return buf->len;
 264         }
 265         pipe_buf_release(pipe, buf);
 266         return ret;
 267 }
 268 EXPORT_SYMBOL(add_to_pipe);
 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 max_usage = READ_ONCE(pipe->max_usage);
 277
 278         spd->nr_pages_max = max_usage;
 279         if (max_usage <= PIPE_DEF_BUFFERS)
 280                 return 0;
 281
 282         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 283         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 284                                      GFP_KERNEL);
 285
 286         if (spd->pages && spd->partial)
 287                 return 0;
 288
 289         kfree(spd->pages);
 290         kfree(spd->partial);
 291         return -ENOMEM;
 292 }
 293
 294 void splice_shrink_spd(struct splice_pipe_desc *spd)
 295 {
 296         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 297                 return;
 298
 299         kfree(spd->pages);
 300         kfree(spd->partial);
 301 }
 302
 303 /*
 304  * Splice data from an O_DIRECT file into pages and then add them to the output
 305  * pipe.
 306  */
 307 ssize_t direct_splice_read(struct file *in, loff_t *ppos,
 308                            struct pipe_inode_info *pipe,
 309                            size_t len, unsigned int flags)
 310 {
 311         struct iov_iter to;
 312         struct bio_vec *bv;
 313         struct kiocb kiocb;
 314         struct page **pages;
 315         ssize_t ret;
 316         size_t used, npages, chunk, remain, reclaim;
 317         int i;
 318
 319         /* Work out how much data we can actually add into the pipe */
 320         used = pipe_occupancy(pipe->head, pipe->tail);
 321         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 322         len = min_t(size_t, len, npages * PAGE_SIZE);
 323         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 324
 325         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 326                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 327         if (!bv)
 328                 return -ENOMEM;
 329
 330         pages = (void *)(bv + npages);
 331         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 332         if (!npages) {
 333                 kfree(bv);
 334                 return -ENOMEM;
 335         }
 336
 337         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 338
 339         for (i = 0; i < npages; i++) {
 340                 chunk = min_t(size_t, PAGE_SIZE, remain);
 341                 bv[i].bv_page = pages[i];
 342                 bv[i].bv_offset = 0;
 343                 bv[i].bv_len = chunk;
 344                 remain -= chunk;
 345         }
 346
 347         /* Do the I/O */
 348         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 349         init_sync_kiocb(&kiocb, in);
 350         kiocb.ki_pos = *ppos;
 351         ret = call_read_iter(in, &kiocb, &to);
 352
 353         reclaim = npages * PAGE_SIZE;
 354         remain = 0;
 355         if (ret > 0) {
 356                 reclaim -= ret;
 357                 remain = ret;
 358                 *ppos = kiocb.ki_pos;
 359                 file_accessed(in);
 360         } else if (ret < 0) {
 361                 /*
 362                  * callers of ->splice_read() expect -EAGAIN on
 363                  * "can't put anything in there", rather than -EFAULT.
 364                  */
 365                 if (ret == -EFAULT)
 366                         ret = -EAGAIN;
 367         }
 368
 369         /* Free any pages that didn't get touched at all. */
 370         reclaim /= PAGE_SIZE;
 371         if (reclaim) {
 372                 npages -= reclaim;
 373                 release_pages(pages + npages, reclaim);
 374         }
 375
 376         /* Push the remaining pages into the pipe. */
 377         for (i = 0; i < npages; i++) {
 378                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 379
 380                 chunk = min_t(size_t, remain, PAGE_SIZE);
 381                 *buf = (struct pipe_buffer) {
 382                         .ops    = &default_pipe_buf_ops,
 383                         .page   = bv[i].bv_page,
 384                         .offset = 0,
 385                         .len    = chunk,
 386                 };
 387                 pipe->head++;
 388                 remain -= chunk;
 389         }
 390
 391         kfree(bv);
 392         return ret;
 393 }
 394 EXPORT_SYMBOL(direct_splice_read);
 395
 396 /**
 397  * generic_file_splice_read - splice data from file to a pipe
 398  * @in:         file to splice from
 399  * @ppos:       position in @in
 400  * @pipe:       pipe to splice to
 401  * @len:        number of bytes to splice
 402  * @flags:      splice modifier flags
 403  *
 404  * Description:
 405  *    Will read pages from given file and fill them into a pipe. Can be
 406  *    used as long as it has more or less sane ->read_iter().
 407  *
 408  */
 409 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 410                                  struct pipe_inode_info *pipe, size_t len,
 411                                  unsigned int flags)
 412 {
 413         struct iov_iter to;
 414         struct kiocb kiocb;
 415         int ret;
 416
 417         iov_iter_pipe(&to, ITER_DEST, pipe, len);
 418         init_sync_kiocb(&kiocb, in);
 419         kiocb.ki_pos = *ppos;
 420         ret = call_read_iter(in, &kiocb, &to);
 421         if (ret > 0) {
 422                 *ppos = kiocb.ki_pos;
 423                 file_accessed(in);
 424         } else if (ret < 0) {
 425                 /* free what was emitted */
 426                 pipe_discard_from(pipe, to.start_head);
 427                 /*
 428                  * callers of ->splice_read() expect -EAGAIN on
 429                  * "can't put anything in there", rather than -EFAULT.
 430                  */
 431                 if (ret == -EFAULT)
 432                         ret = -EAGAIN;
 433         }
 434
 435         return ret;
 436 }
 437 EXPORT_SYMBOL(generic_file_splice_read);
 438
 439 const struct pipe_buf_operations default_pipe_buf_ops = {
 440         .release        = generic_pipe_buf_release,
 441         .try_steal      = generic_pipe_buf_try_steal,
 442         .get            = generic_pipe_buf_get,
 443 };
 444
 445 /* Pipe buffer operations for a socket and similar. */
 446 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 447         .release        = generic_pipe_buf_release,
 448         .get            = generic_pipe_buf_get,
 449 };
 450 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 451
 452 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 453 {
 454         smp_mb();
 455         if (waitqueue_active(&pipe->wr_wait))
 456                 wake_up_interruptible(&pipe->wr_wait);
 457         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 458 }
 459
 460 /**
 461  * splice_from_pipe_feed - feed available data from a pipe to a file
 462  * @pipe:       pipe to splice from
 463  * @sd:         information to @actor
 464  * @actor:      handler that splices the data
 465  *
 466  * Description:
 467  *    This function loops over the pipe and calls @actor to do the
 468  *    actual moving of a single struct pipe_buffer to the desired
 469  *    destination.  It returns when there's no more buffers left in
 470  *    the pipe or if the requested number of bytes (@sd->total_len)
 471  *    have been copied.  It returns a positive number (one) if the
 472  *    pipe needs to be filled with more data, zero if the required
 473  *    number of bytes have been copied and -errno on error.
 474  *
 475  *    This, together with splice_from_pipe_{begin,end,next}, may be
 476  *    used to implement the functionality of __splice_from_pipe() when
 477  *    locking is required around copying the pipe buffers to the
 478  *    destination.
 479  */
 480 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 481                           splice_actor *actor)
 482 {
 483         unsigned int head = pipe->head;
 484         unsigned int tail = pipe->tail;
 485         unsigned int mask = pipe->ring_size - 1;
 486         int ret;
 487
 488         while (!pipe_empty(head, tail)) {
 489                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 490
 491                 sd->len = buf->len;
 492                 if (sd->len > sd->total_len)
 493                         sd->len = sd->total_len;
 494
 495                 ret = pipe_buf_confirm(pipe, buf);
 496                 if (unlikely(ret)) {
 497                         if (ret == -ENODATA)
 498                                 ret = 0;
 499                         return ret;
 500                 }
 501
 502                 ret = actor(pipe, buf, sd);
 503                 if (ret <= 0)
 504                         return ret;
 505
 506                 buf->offset += ret;
 507                 buf->len -= ret;
 508
 509                 sd->num_spliced += ret;
 510                 sd->len -= ret;
 511                 sd->pos += ret;
 512                 sd->total_len -= ret;
 513
 514                 if (!buf->len) {
 515                         pipe_buf_release(pipe, buf);
 516                         tail++;
 517                         pipe->tail = tail;
 518                         if (pipe->files)
 519                                 sd->need_wakeup = true;
 520                 }
 521
 522                 if (!sd->total_len)
 523                         return 0;
 524         }
 525
 526         return 1;
 527 }
 528
 529 /* We know we have a pipe buffer, but maybe it's empty? */
 530 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 531 {
 532         unsigned int tail = pipe->tail;
 533         unsigned int mask = pipe->ring_size - 1;
 534         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 535
 536         if (unlikely(!buf->len)) {
 537                 pipe_buf_release(pipe, buf);
 538                 pipe->tail = tail+1;
 539                 return true;
 540         }
 541
 542         return false;
 543 }
 544
 545 /**
 546  * splice_from_pipe_next - wait for some data to splice from
 547  * @pipe:       pipe to splice from
 548  * @sd:         information about the splice operation
 549  *
 550  * Description:
 551  *    This function will wait for some data and return a positive
 552  *    value (one) if pipe buffers are available.  It will return zero
 553  *    or -errno if no more data needs to be spliced.
 554  */
 555 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 556 {
 557         /*
 558          * Check for signal early to make process killable when there are
 559          * always buffers available
 560          */
 561         if (signal_pending(current))
 562                 return -ERESTARTSYS;
 563
 564 repeat:
 565         while (pipe_empty(pipe->head, pipe->tail)) {
 566                 if (!pipe->writers)
 567                         return 0;
 568
 569                 if (sd->num_spliced)
 570                         return 0;
 571
 572                 if (sd->flags & SPLICE_F_NONBLOCK)
 573                         return -EAGAIN;
 574
 575                 if (signal_pending(current))
 576                         return -ERESTARTSYS;
 577
 578                 if (sd->need_wakeup) {
 579                         wakeup_pipe_writers(pipe);
 580                         sd->need_wakeup = false;
 581                 }
 582
 583                 pipe_wait_readable(pipe);
 584         }
 585
 586         if (eat_empty_buffer(pipe))
 587                 goto repeat;
 588
 589         return 1;
 590 }
 591
 592 /**
 593  * splice_from_pipe_begin - start splicing from pipe
 594  * @sd:         information about the splice operation
 595  *
 596  * Description:
 597  *    This function should be called before a loop containing
 598  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 599  *    initialize the necessary fields of @sd.
 600  */
 601 static void splice_from_pipe_begin(struct splice_desc *sd)
 602 {
 603         sd->num_spliced = 0;
 604         sd->need_wakeup = false;
 605 }
 606
 607 /**
 608  * splice_from_pipe_end - finish splicing from pipe
 609  * @pipe:       pipe to splice from
 610  * @sd:         information about the splice operation
 611  *
 612  * Description:
 613  *    This function will wake up pipe writers if necessary.  It should
 614  *    be called after a loop containing splice_from_pipe_next() and
 615  *    splice_from_pipe_feed().
 616  */
 617 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 618 {
 619         if (sd->need_wakeup)
 620                 wakeup_pipe_writers(pipe);
 621 }
 622
 623 /**
 624  * __splice_from_pipe - splice data from a pipe to given actor
 625  * @pipe:       pipe to splice from
 626  * @sd:         information to @actor
 627  * @actor:      handler that splices the data
 628  *
 629  * Description:
 630  *    This function does little more than loop over the pipe and call
 631  *    @actor to do the actual moving of a single struct pipe_buffer to
 632  *    the desired destination. See pipe_to_file, pipe_to_sendmsg, or
 633  *    pipe_to_user.
 634  *
 635  */
 636 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 637                            splice_actor *actor)
 638 {
 639         int ret;
 640
 641         splice_from_pipe_begin(sd);
 642         do {
 643                 cond_resched();
 644                 ret = splice_from_pipe_next(pipe, sd);
 645                 if (ret > 0)
 646                         ret = splice_from_pipe_feed(pipe, sd, actor);
 647         } while (ret > 0);
 648         splice_from_pipe_end(pipe, sd);
 649
 650         return sd->num_spliced ? sd->num_spliced : ret;
 651 }
 652 EXPORT_SYMBOL(__splice_from_pipe);
 653
 654 /**
 655  * splice_from_pipe - splice data from a pipe to a file
 656  * @pipe:       pipe to splice from
 657  * @out:        file to splice to
 658  * @ppos:       position in @out
 659  * @len:        how many bytes to splice
 660  * @flags:      splice modifier flags
 661  * @actor:      handler that splices the data
 662  *
 663  * Description:
 664  *    See __splice_from_pipe. This function locks the pipe inode,
 665  *    otherwise it's identical to __splice_from_pipe().
 666  *
 667  */
 668 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 669                          loff_t *ppos, size_t len, unsigned int flags,
 670                          splice_actor *actor)
 671 {
 672         ssize_t ret;
 673         struct splice_desc sd = {
 674                 .total_len = len,
 675                 .flags = flags,
 676                 .pos = *ppos,
 677                 .u.file = out,
 678         };
 679
 680         pipe_lock(pipe);
 681         ret = __splice_from_pipe(pipe, &sd, actor);
 682         pipe_unlock(pipe);
 683
 684         return ret;
 685 }
 686
 687 /**
 688  * iter_file_splice_write - splice data from a pipe to a file
 689  * @pipe:       pipe info
 690  * @out:        file to write to
 691  * @ppos:       position in @out
 692  * @len:        number of bytes to splice
 693  * @flags:      splice modifier flags
 694  *
 695  * Description:
 696  *    Will either move or copy pages (determined by @flags options) from
 697  *    the given pipe inode to the given file.
 698  *    This one is ->write_iter-based.
 699  *
 700  */
 701 ssize_t
 702 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 703                           loff_t *ppos, size_t len, unsigned int flags)
 704 {
 705         struct splice_desc sd = {
 706                 .total_len = len,
 707                 .flags = flags,
 708                 .pos = *ppos,
 709                 .u.file = out,
 710         };
 711         int nbufs = pipe->max_usage;
 712         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 713                                         GFP_KERNEL);
 714         ssize_t ret;
 715
 716         if (unlikely(!array))
 717                 return -ENOMEM;
 718
 719         pipe_lock(pipe);
 720
 721         splice_from_pipe_begin(&sd);
 722         while (sd.total_len) {
 723                 struct iov_iter from;
 724                 unsigned int head, tail, mask;
 725                 size_t left;
 726                 int n;
 727
 728                 ret = splice_from_pipe_next(pipe, &sd);
 729                 if (ret <= 0)
 730                         break;
 731
 732                 if (unlikely(nbufs < pipe->max_usage)) {
 733                         kfree(array);
 734                         nbufs = pipe->max_usage;
 735                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 736                                         GFP_KERNEL);
 737                         if (!array) {
 738                                 ret = -ENOMEM;
 739                                 break;
 740                         }
 741                 }
 742
 743                 head = pipe->head;
 744                 tail = pipe->tail;
 745                 mask = pipe->ring_size - 1;
 746
 747                 /* build the vector */
 748                 left = sd.total_len;
 749                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 750                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 751                         size_t this_len = buf->len;
 752
 753                         /* zero-length bvecs are not supported, skip them */
 754                         if (!this_len)
 755                                 continue;
 756                         this_len = min(this_len, left);
 757
 758                         ret = pipe_buf_confirm(pipe, buf);
 759                         if (unlikely(ret)) {
 760                                 if (ret == -ENODATA)
 761                                         ret = 0;
 762                                 goto done;
 763                         }
 764
 765                         bvec_set_page(&array[n], buf->page, this_len,
 766                                       buf->offset);
 767                         left -= this_len;
 768                         n++;
 769                 }
 770
 771                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 772                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 773                 if (ret <= 0)
 774                         break;
 775
 776                 sd.num_spliced += ret;
 777                 sd.total_len -= ret;
 778                 *ppos = sd.pos;
 779
 780                 /* dismiss the fully eaten buffers, adjust the partial one */
 781                 tail = pipe->tail;
 782                 while (ret) {
 783                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 784                         if (ret >= buf->len) {
 785                                 ret -= buf->len;
 786                                 buf->len = 0;
 787                                 pipe_buf_release(pipe, buf);
 788                                 tail++;
 789                                 pipe->tail = tail;
 790                                 if (pipe->files)
 791                                         sd.need_wakeup = true;
 792                         } else {
 793                                 buf->offset += ret;
 794                                 buf->len -= ret;
 795                                 ret = 0;
 796                         }
 797                 }
 798         }
 799 done:
 800         kfree(array);
 801         splice_from_pipe_end(pipe, &sd);
 802
 803         pipe_unlock(pipe);
 804
 805         if (sd.num_spliced)
 806                 ret = sd.num_spliced;
 807
 808         return ret;
 809 }
 810
 811 EXPORT_SYMBOL(iter_file_splice_write);
 812
 813 #ifdef CONFIG_NET
 814 /**
 815  * splice_to_socket - splice data from a pipe to a socket
 816  * @pipe:       pipe to splice from
 817  * @out:        socket to write to
 818  * @ppos:       position in @out
 819  * @len:        number of bytes to splice
 820  * @flags:      splice modifier flags
 821  *
 822  * Description:
 823  *    Will send @len bytes from the pipe to a network socket. No data copying
 824  *    is involved.
 825  *
 826  */
 827 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 828                          loff_t *ppos, size_t len, unsigned int flags)
 829 {
 830         struct socket *sock = sock_from_file(out);
 831         struct bio_vec bvec[16];
 832         struct msghdr msg = {};
 833         ssize_t ret = 0;
 834         size_t spliced = 0;
 835         bool need_wakeup = false;
 836
 837         pipe_lock(pipe);
 838
 839         while (len > 0) {
 840                 unsigned int head, tail, mask, bc = 0;
 841                 size_t remain = len;
 842
 843                 /*
 844                  * Check for signal early to make process killable when there
 845                  * are always buffers available
 846                  */
 847                 ret = -ERESTARTSYS;
 848                 if (signal_pending(current))
 849                         break;
 850
 851                 while (pipe_empty(pipe->head, pipe->tail)) {
 852                         ret = 0;
 853                         if (!pipe->writers)
 854                                 goto out;
 855
 856                         if (spliced)
 857                                 goto out;
 858
 859                         ret = -EAGAIN;
 860                         if (flags & SPLICE_F_NONBLOCK)
 861                                 goto out;
 862
 863                         ret = -ERESTARTSYS;
 864                         if (signal_pending(current))
 865                                 goto out;
 866
 867                         if (need_wakeup) {
 868                                 wakeup_pipe_writers(pipe);
 869                                 need_wakeup = false;
 870                         }
 871
 872                         pipe_wait_readable(pipe);
 873                 }
 874
 875                 head = pipe->head;
 876                 tail = pipe->tail;
 877                 mask = pipe->ring_size - 1;
 878
 879                 while (!pipe_empty(head, tail)) {
 880                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 881                         size_t seg;
 882
 883                         if (!buf->len) {
 884                                 tail++;
 885                                 continue;
 886                         }
 887
 888                         seg = min_t(size_t, remain, buf->len);
 889                         seg = min_t(size_t, seg, PAGE_SIZE);
 890
 891                         ret = pipe_buf_confirm(pipe, buf);
 892                         if (unlikely(ret)) {
 893                                 if (ret == -ENODATA)
 894                                         ret = 0;
 895                                 break;
 896                         }
 897
 898                         bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
 899                         remain -= seg;
 900                         if (seg >= buf->len)
 901                                 tail++;
 902                         if (bc >= ARRAY_SIZE(bvec))
 903                                 break;
 904                 }
 905
 906                 if (!bc)
 907                         break;
 908
 909                 msg.msg_flags = MSG_SPLICE_PAGES;
 910                 if (flags & SPLICE_F_MORE)
 911                         msg.msg_flags |= MSG_MORE;
 912                 if (remain && pipe_occupancy(pipe->head, tail) > 0)
 913                         msg.msg_flags |= MSG_MORE;
 914
 915                 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 916                               len - remain);
 917                 ret = sock_sendmsg(sock, &msg);
 918                 if (ret <= 0)
 919                         break;
 920
 921                 spliced += ret;
 922                 len -= ret;
 923                 tail = pipe->tail;
 924                 while (ret > 0) {
 925                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 926                         size_t seg = min_t(size_t, ret, buf->len);
 927
 928                         buf->offset += seg;
 929                         buf->len -= seg;
 930                         ret -= seg;
 931
 932                         if (!buf->len) {
 933                                 pipe_buf_release(pipe, buf);
 934                                 tail++;
 935                         }
 936                 }
 937
 938                 if (tail != pipe->tail) {
 939                         pipe->tail = tail;
 940                         if (pipe->files)
 941                                 need_wakeup = true;
 942                 }
 943         }
 944
 945 out:
 946         pipe_unlock(pipe);
 947         if (need_wakeup)
 948                 wakeup_pipe_writers(pipe);
 949         return spliced ?: ret;
 950 }
 951 #endif
 952
 953 static int warn_unsupported(struct file *file, const char *op)
 954 {
 955         pr_debug_ratelimited(
 956                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 957                 op, file, current->pid, current->comm);
 958         return -EINVAL;
 959 }
 960
 961 /*
 962  * Attempt to initiate a splice from pipe to file.
 963  */
 964 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 965                            loff_t *ppos, size_t len, unsigned int flags)
 966 {
 967         if (unlikely(!out->f_op->splice_write))
 968                 return warn_unsupported(out, "write");
 969         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 970 }
 971
 972 /*
 973  * Indicate to the caller that there was a premature EOF when reading from the
 974  * source and the caller didn't indicate they would be sending more data after
 975  * this.
 976  */
 977 static void do_splice_eof(struct splice_desc *sd)
 978 {
 979         if (sd->splice_eof)
 980                 sd->splice_eof(sd);
 981 }
 982
 983 /*
 984  * Attempt to initiate a splice from a file to a pipe.
 985  */
 986 static long do_splice_to(struct file *in, loff_t *ppos,
 987                          struct pipe_inode_info *pipe, size_t len,
 988                          unsigned int flags)
 989 {
 990         unsigned int p_space;
 991         int ret;
 992
 993         if (unlikely(!(in->f_mode & FMODE_READ)))
 994                 return -EBADF;
 995
 996         /* Don't try to read more the pipe has space for. */
 997         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 998         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 999
1000         ret = rw_verify_area(READ, in, ppos, len);
1001         if (unlikely(ret < 0))
1002                 return ret;
1003
1004         if (unlikely(len > MAX_RW_COUNT))
1005                 len = MAX_RW_COUNT;
1006
1007         if (unlikely(!in->f_op->splice_read))
1008                 return warn_unsupported(in, "read");
1009         return in->f_op->splice_read(in, ppos, pipe, len, flags);
1010 }
1011
1012 /**
1013  * splice_direct_to_actor - splices data directly between two non-pipes
1014  * @in:         file to splice from
1015  * @sd:         actor information on where to splice to
1016  * @actor:      handles the data splicing
1017  *
1018  * Description:
1019  *    This is a special case helper to splice directly between two
1020  *    points, without requiring an explicit pipe. Internally an allocated
1021  *    pipe is cached in the process, and reused during the lifetime of
1022  *    that process.
1023  *
1024  */
1025 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1026                                splice_direct_actor *actor)
1027 {
1028         struct pipe_inode_info *pipe;
1029         long ret, bytes;
1030         size_t len;
1031         int i, flags, more;
1032
1033         /*
1034          * We require the input to be seekable, as we don't want to randomly
1035          * drop data for eg socket -> socket splicing. Use the piped splicing
1036          * for that!
1037          */
1038         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1039                 return -EINVAL;
1040
1041         /*
1042          * neither in nor out is a pipe, setup an internal pipe attached to
1043          * 'out' and transfer the wanted data from 'in' to 'out' through that
1044          */
1045         pipe = current->splice_pipe;
1046         if (unlikely(!pipe)) {
1047                 pipe = alloc_pipe_info();
1048                 if (!pipe)
1049                         return -ENOMEM;
1050
1051                 /*
1052                  * We don't have an immediate reader, but we'll read the stuff
1053                  * out of the pipe right after the splice_to_pipe(). So set
1054                  * PIPE_READERS appropriately.
1055                  */
1056                 pipe->readers = 1;
1057
1058                 current->splice_pipe = pipe;
1059         }
1060
1061         /*
1062          * Do the splice.
1063          */
1064         bytes = 0;
1065         len = sd->total_len;
1066
1067         /* Don't block on output, we have to drain the direct pipe. */
1068         flags = sd->flags;
1069         sd->flags &= ~SPLICE_F_NONBLOCK;
1070
1071         /*
1072          * We signal MORE until we've read sufficient data to fulfill the
1073          * request and we keep signalling it if the caller set it.
1074          */
1075         more = sd->flags & SPLICE_F_MORE;
1076         sd->flags |= SPLICE_F_MORE;
1077
1078         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1079
1080         while (len) {
1081                 size_t read_len;
1082                 loff_t pos = sd->pos, prev_pos = pos;
1083
1084                 ret = do_splice_to(in, &pos, pipe, len, flags);
1085                 if (unlikely(ret <= 0))
1086                         goto read_failure;
1087
1088                 read_len = ret;
1089                 sd->total_len = read_len;
1090
1091                 /*
1092                  * If we now have sufficient data to fulfill the request then
1093                  * we clear SPLICE_F_MORE if it was not set initially.
1094                  */
1095                 if (read_len >= len && !more)
1096                         sd->flags &= ~SPLICE_F_MORE;
1097
1098                 /*
1099                  * NOTE: nonblocking mode only applies to the input. We
1100                  * must not do the output in nonblocking mode as then we
1101                  * could get stuck data in the internal pipe:
1102                  */
1103                 ret = actor(pipe, sd);
1104                 if (unlikely(ret <= 0)) {
1105                         sd->pos = prev_pos;
1106                         goto out_release;
1107                 }
1108
1109                 bytes += ret;
1110                 len -= ret;
1111                 sd->pos = pos;
1112
1113                 if (ret < read_len) {
1114                         sd->pos = prev_pos + ret;
1115                         goto out_release;
1116                 }
1117         }
1118
1119 done:
1120         pipe->tail = pipe->head = 0;
1121         file_accessed(in);
1122         return bytes;
1123
1124 read_failure:
1125         /*
1126          * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1127          * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1128          * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1129          * least 1 byte *then* we will also do the ->splice_eof() call.
1130          */
1131         if (ret == 0 && !more && len > 0 && bytes)
1132                 do_splice_eof(sd);
1133 out_release:
1134         /*
1135          * If we did an incomplete transfer we must release
1136          * the pipe buffers in question:
1137          */
1138         for (i = 0; i < pipe->ring_size; i++) {
1139                 struct pipe_buffer *buf = &pipe->bufs[i];
1140
1141                 if (buf->ops)
1142                         pipe_buf_release(pipe, buf);
1143         }
1144
1145         if (!bytes)
1146                 bytes = ret;
1147
1148         goto done;
1149 }
1150 EXPORT_SYMBOL(splice_direct_to_actor);
1151
1152 static int direct_splice_actor(struct pipe_inode_info *pipe,
1153                                struct splice_desc *sd)
1154 {
1155         struct file *file = sd->u.file;
1156
1157         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1158                               sd->flags);
1159 }
1160
1161 static void direct_file_splice_eof(struct splice_desc *sd)
1162 {
1163         struct file *file = sd->u.file;
1164
1165         if (file->f_op->splice_eof)
1166                 file->f_op->splice_eof(file);
1167 }
1168
1169 /**
1170  * do_splice_direct - splices data directly between two files
1171  * @in:         file to splice from
1172  * @ppos:       input file offset
1173  * @out:        file to splice to
1174  * @opos:       output file offset
1175  * @len:        number of bytes to splice
1176  * @flags:      splice modifier flags
1177  *
1178  * Description:
1179  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1180  *    doing it in the application would incur an extra system call
1181  *    (splice in + splice out, as compared to just sendfile()). So this helper
1182  *    can splice directly through a process-private pipe.
1183  *
1184  */
1185 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1186                       loff_t *opos, size_t len, unsigned int flags)
1187 {
1188         struct splice_desc sd = {
1189                 .len            = len,
1190                 .total_len      = len,
1191                 .flags          = flags,
1192                 .pos            = *ppos,
1193                 .u.file         = out,
1194                 .splice_eof     = direct_file_splice_eof,
1195                 .opos           = opos,
1196         };
1197         long ret;
1198
1199         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1200                 return -EBADF;
1201
1202         if (unlikely(out->f_flags & O_APPEND))
1203                 return -EINVAL;
1204
1205         ret = rw_verify_area(WRITE, out, opos, len);
1206         if (unlikely(ret < 0))
1207                 return ret;
1208
1209         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1210         if (ret > 0)
1211                 *ppos = sd.pos;
1212
1213         return ret;
1214 }
1215 EXPORT_SYMBOL(do_splice_direct);
1216
1217 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1218 {
1219         for (;;) {
1220                 if (unlikely(!pipe->readers)) {
1221                         send_sig(SIGPIPE, current, 0);
1222                         return -EPIPE;
1223                 }
1224                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1225                         return 0;
1226                 if (flags & SPLICE_F_NONBLOCK)
1227                         return -EAGAIN;
1228                 if (signal_pending(current))
1229                         return -ERESTARTSYS;
1230                 pipe_wait_writable(pipe);
1231         }
1232 }
1233
1234 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1235                                struct pipe_inode_info *opipe,
1236                                size_t len, unsigned int flags);
1237
1238 long splice_file_to_pipe(struct file *in,
1239                          struct pipe_inode_info *opipe,
1240                          loff_t *offset,
1241                          size_t len, unsigned int flags)
1242 {
1243         long ret;
1244
1245         pipe_lock(opipe);
1246         ret = wait_for_space(opipe, flags);
1247         if (!ret)
1248                 ret = do_splice_to(in, offset, opipe, len, flags);
1249         pipe_unlock(opipe);
1250         if (ret > 0)
1251                 wakeup_pipe_readers(opipe);
1252         return ret;
1253 }
1254
1255 /*
1256  * Determine where to splice to/from.
1257  */
1258 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1259                loff_t *off_out, size_t len, unsigned int flags)
1260 {
1261         struct pipe_inode_info *ipipe;
1262         struct pipe_inode_info *opipe;
1263         loff_t offset;
1264         long ret;
1265
1266         if (unlikely(!(in->f_mode & FMODE_READ) ||
1267                      !(out->f_mode & FMODE_WRITE)))
1268                 return -EBADF;
1269
1270         ipipe = get_pipe_info(in, true);
1271         opipe = get_pipe_info(out, true);
1272
1273         if (ipipe && opipe) {
1274                 if (off_in || off_out)
1275                         return -ESPIPE;
1276
1277                 /* Splicing to self would be fun, but... */
1278                 if (ipipe == opipe)
1279                         return -EINVAL;
1280
1281                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1282                         flags |= SPLICE_F_NONBLOCK;
1283
1284                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1285         }
1286
1287         if (ipipe) {
1288                 if (off_in)
1289                         return -ESPIPE;
1290                 if (off_out) {
1291                         if (!(out->f_mode & FMODE_PWRITE))
1292                                 return -EINVAL;
1293                         offset = *off_out;
1294                 } else {
1295                         offset = out->f_pos;
1296                 }
1297
1298                 if (unlikely(out->f_flags & O_APPEND))
1299                         return -EINVAL;
1300
1301                 ret = rw_verify_area(WRITE, out, &offset, len);
1302                 if (unlikely(ret < 0))
1303                         return ret;
1304
1305                 if (in->f_flags & O_NONBLOCK)
1306                         flags |= SPLICE_F_NONBLOCK;
1307
1308                 file_start_write(out);
1309                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1310                 file_end_write(out);
1311
1312                 if (ret > 0)
1313                         fsnotify_modify(out);
1314
1315                 if (!off_out)
1316                         out->f_pos = offset;
1317                 else
1318                         *off_out = offset;
1319
1320                 return ret;
1321         }
1322
1323         if (opipe) {
1324                 if (off_out)
1325                         return -ESPIPE;
1326                 if (off_in) {
1327                         if (!(in->f_mode & FMODE_PREAD))
1328                                 return -EINVAL;
1329                         offset = *off_in;
1330                 } else {
1331                         offset = in->f_pos;
1332                 }
1333
1334                 if (out->f_flags & O_NONBLOCK)
1335                         flags |= SPLICE_F_NONBLOCK;
1336
1337                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1338
1339                 if (ret > 0)
1340                         fsnotify_access(in);
1341
1342                 if (!off_in)
1343                         in->f_pos = offset;
1344                 else
1345                         *off_in = offset;
1346
1347                 return ret;
1348         }
1349
1350         return -EINVAL;
1351 }
1352
1353 static long __do_splice(struct file *in, loff_t __user *off_in,
1354                         struct file *out, loff_t __user *off_out,
1355                         size_t len, unsigned int flags)
1356 {
1357         struct pipe_inode_info *ipipe;
1358         struct pipe_inode_info *opipe;
1359         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1360         long ret;
1361
1362         ipipe = get_pipe_info(in, true);
1363         opipe = get_pipe_info(out, true);
1364
1365         if (ipipe) {
1366                 if (off_in)
1367                         return -ESPIPE;
1368                 pipe_clear_nowait(in);
1369         }
1370         if (opipe) {
1371                 if (off_out)
1372                         return -ESPIPE;
1373                 pipe_clear_nowait(out);
1374         }
1375
1376         if (off_out) {
1377                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1378                         return -EFAULT;
1379                 __off_out = &offset;
1380         }
1381         if (off_in) {
1382                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1383                         return -EFAULT;
1384                 __off_in = &offset;
1385         }
1386
1387         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1388         if (ret < 0)
1389                 return ret;
1390
1391         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1392                 return -EFAULT;
1393         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1394                 return -EFAULT;
1395
1396         return ret;
1397 }
1398
1399 static int iter_to_pipe(struct iov_iter *from,
1400                         struct pipe_inode_info *pipe,
1401                         unsigned flags)
1402 {
1403         struct pipe_buffer buf = {
1404                 .ops = &user_page_pipe_buf_ops,
1405                 .flags = flags
1406         };
1407         size_t total = 0;
1408         int ret = 0;
1409
1410         while (iov_iter_count(from)) {
1411                 struct page *pages[16];
1412                 ssize_t left;
1413                 size_t start;
1414                 int i, n;
1415
1416                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1417                 if (left <= 0) {
1418                         ret = left;
1419                         break;
1420                 }
1421
1422                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1423                 for (i = 0; i < n; i++) {
1424                         int size = min_t(int, left, PAGE_SIZE - start);
1425
1426                         buf.page = pages[i];
1427                         buf.offset = start;
1428                         buf.len = size;
1429                         ret = add_to_pipe(pipe, &buf);
1430                         if (unlikely(ret < 0)) {
1431                                 iov_iter_revert(from, left);
1432                                 // this one got dropped by add_to_pipe()
1433                                 while (++i < n)
1434                                         put_page(pages[i]);
1435                                 goto out;
1436                         }
1437                         total += ret;
1438                         left -= size;
1439                         start = 0;
1440                 }
1441         }
1442 out:
1443         return total ? total : ret;
1444 }
1445
1446 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1447                         struct splice_desc *sd)
1448 {
1449         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1450         return n == sd->len ? n : -EFAULT;
1451 }
1452
1453 /*
1454  * For lack of a better implementation, implement vmsplice() to userspace
1455  * as a simple copy of the pipes pages to the user iov.
1456  */
1457 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1458                              unsigned int flags)
1459 {
1460         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1461         struct splice_desc sd = {
1462                 .total_len = iov_iter_count(iter),
1463                 .flags = flags,
1464                 .u.data = iter
1465         };
1466         long ret = 0;
1467
1468         if (!pipe)
1469                 return -EBADF;
1470
1471         pipe_clear_nowait(file);
1472
1473         if (sd.total_len) {
1474                 pipe_lock(pipe);
1475                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1476                 pipe_unlock(pipe);
1477         }
1478
1479         return ret;
1480 }
1481
1482 /*
1483  * vmsplice splices a user address range into a pipe. It can be thought of
1484  * as splice-from-memory, where the regular splice is splice-from-file (or
1485  * to file). In both cases the output is a pipe, naturally.
1486  */
1487 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1488                              unsigned int flags)
1489 {
1490         struct pipe_inode_info *pipe;
1491         long ret = 0;
1492         unsigned buf_flag = 0;
1493
1494         if (flags & SPLICE_F_GIFT)
1495                 buf_flag = PIPE_BUF_FLAG_GIFT;
1496
1497         pipe = get_pipe_info(file, true);
1498         if (!pipe)
1499                 return -EBADF;
1500
1501         pipe_clear_nowait(file);
1502
1503         pipe_lock(pipe);
1504         ret = wait_for_space(pipe, flags);
1505         if (!ret)
1506                 ret = iter_to_pipe(iter, pipe, buf_flag);
1507         pipe_unlock(pipe);
1508         if (ret > 0)
1509                 wakeup_pipe_readers(pipe);
1510         return ret;
1511 }
1512
1513 static int vmsplice_type(struct fd f, int *type)
1514 {
1515         if (!f.file)
1516                 return -EBADF;
1517         if (f.file->f_mode & FMODE_WRITE) {
1518                 *type = ITER_SOURCE;
1519         } else if (f.file->f_mode & FMODE_READ) {
1520                 *type = ITER_DEST;
1521         } else {
1522                 fdput(f);
1523                 return -EBADF;
1524         }
1525         return 0;
1526 }
1527
1528 /*
1529  * Note that vmsplice only really supports true splicing _from_ user memory
1530  * to a pipe, not the other way around. Splicing from user memory is a simple
1531  * operation that can be supported without any funky alignment restrictions
1532  * or nasty vm tricks. We simply map in the user memory and fill them into
1533  * a pipe. The reverse isn't quite as easy, though. There are two possible
1534  * solutions for that:
1535  *
1536  *      - memcpy() the data internally, at which point we might as well just
1537  *        do a regular read() on the buffer anyway.
1538  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1539  *        has restriction limitations on both ends of the pipe).
1540  *
1541  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1542  *
1543  */
1544 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1545                 unsigned long, nr_segs, unsigned int, flags)
1546 {
1547         struct iovec iovstack[UIO_FASTIOV];
1548         struct iovec *iov = iovstack;
1549         struct iov_iter iter;
1550         ssize_t error;
1551         struct fd f;
1552         int type;
1553
1554         if (unlikely(flags & ~SPLICE_F_ALL))
1555                 return -EINVAL;
1556
1557         f = fdget(fd);
1558         error = vmsplice_type(f, &type);
1559         if (error)
1560                 return error;
1561
1562         error = import_iovec(type, uiov, nr_segs,
1563                              ARRAY_SIZE(iovstack), &iov, &iter);
1564         if (error < 0)
1565                 goto out_fdput;
1566
1567         if (!iov_iter_count(&iter))
1568                 error = 0;
1569         else if (type == ITER_SOURCE)
1570                 error = vmsplice_to_pipe(f.file, &iter, flags);
1571         else
1572                 error = vmsplice_to_user(f.file, &iter, flags);
1573
1574         kfree(iov);
1575 out_fdput:
1576         fdput(f);
1577         return error;
1578 }
1579
1580 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1581                 int, fd_out, loff_t __user *, off_out,
1582                 size_t, len, unsigned int, flags)
1583 {
1584         struct fd in, out;
1585         long error;
1586
1587         if (unlikely(!len))
1588                 return 0;
1589
1590         if (unlikely(flags & ~SPLICE_F_ALL))
1591                 return -EINVAL;
1592
1593         error = -EBADF;
1594         in = fdget(fd_in);
1595         if (in.file) {
1596                 out = fdget(fd_out);
1597                 if (out.file) {
1598                         error = __do_splice(in.file, off_in, out.file, off_out,
1599                                                 len, flags);
1600                         fdput(out);
1601                 }
1602                 fdput(in);
1603         }
1604         return error;
1605 }
1606
1607 /*
1608  * Make sure there's data to read. Wait for input if we can, otherwise
1609  * return an appropriate error.
1610  */
1611 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1612 {
1613         int ret;
1614
1615         /*
1616          * Check the pipe occupancy without the inode lock first. This function
1617          * is speculative anyways, so missing one is ok.
1618          */
1619         if (!pipe_empty(pipe->head, pipe->tail))
1620                 return 0;
1621
1622         ret = 0;
1623         pipe_lock(pipe);
1624
1625         while (pipe_empty(pipe->head, pipe->tail)) {
1626                 if (signal_pending(current)) {
1627                         ret = -ERESTARTSYS;
1628                         break;
1629                 }
1630                 if (!pipe->writers)
1631                         break;
1632                 if (flags & SPLICE_F_NONBLOCK) {
1633                         ret = -EAGAIN;
1634                         break;
1635                 }
1636                 pipe_wait_readable(pipe);
1637         }
1638
1639         pipe_unlock(pipe);
1640         return ret;
1641 }
1642
1643 /*
1644  * Make sure there's writeable room. Wait for room if we can, otherwise
1645  * return an appropriate error.
1646  */
1647 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1648 {
1649         int ret;
1650
1651         /*
1652          * Check pipe occupancy without the inode lock first. This function
1653          * is speculative anyways, so missing one is ok.
1654          */
1655         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1656                 return 0;
1657
1658         ret = 0;
1659         pipe_lock(pipe);
1660
1661         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1662                 if (!pipe->readers) {
1663                         send_sig(SIGPIPE, current, 0);
1664                         ret = -EPIPE;
1665                         break;
1666                 }
1667                 if (flags & SPLICE_F_NONBLOCK) {
1668                         ret = -EAGAIN;
1669                         break;
1670                 }
1671                 if (signal_pending(current)) {
1672                         ret = -ERESTARTSYS;
1673                         break;
1674                 }
1675                 pipe_wait_writable(pipe);
1676         }
1677
1678         pipe_unlock(pipe);
1679         return ret;
1680 }
1681
1682 /*
1683  * Splice contents of ipipe to opipe.
1684  */
1685 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1686                                struct pipe_inode_info *opipe,
1687                                size_t len, unsigned int flags)
1688 {
1689         struct pipe_buffer *ibuf, *obuf;
1690         unsigned int i_head, o_head;
1691         unsigned int i_tail, o_tail;
1692         unsigned int i_mask, o_mask;
1693         int ret = 0;
1694         bool input_wakeup = false;
1695
1696
1697 retry:
1698         ret = ipipe_prep(ipipe, flags);
1699         if (ret)
1700                 return ret;
1701
1702         ret = opipe_prep(opipe, flags);
1703         if (ret)
1704                 return ret;
1705
1706         /*
1707          * Potential ABBA deadlock, work around it by ordering lock
1708          * grabbing by pipe info address. Otherwise two different processes
1709          * could deadlock (one doing tee from A -> B, the other from B -> A).
1710          */
1711         pipe_double_lock(ipipe, opipe);
1712
1713         i_tail = ipipe->tail;
1714         i_mask = ipipe->ring_size - 1;
1715         o_head = opipe->head;
1716         o_mask = opipe->ring_size - 1;
1717
1718         do {
1719                 size_t o_len;
1720
1721                 if (!opipe->readers) {
1722                         send_sig(SIGPIPE, current, 0);
1723                         if (!ret)
1724                                 ret = -EPIPE;
1725                         break;
1726                 }
1727
1728                 i_head = ipipe->head;
1729                 o_tail = opipe->tail;
1730
1731                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1732                         break;
1733
1734                 /*
1735                  * Cannot make any progress, because either the input
1736                  * pipe is empty or the output pipe is full.
1737                  */
1738                 if (pipe_empty(i_head, i_tail) ||
1739                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1740                         /* Already processed some buffers, break */
1741                         if (ret)
1742                                 break;
1743
1744                         if (flags & SPLICE_F_NONBLOCK) {
1745                                 ret = -EAGAIN;
1746                                 break;
1747                         }
1748
1749                         /*
1750                          * We raced with another reader/writer and haven't
1751                          * managed to process any buffers.  A zero return
1752                          * value means EOF, so retry instead.
1753                          */
1754                         pipe_unlock(ipipe);
1755                         pipe_unlock(opipe);
1756                         goto retry;
1757                 }
1758
1759                 ibuf = &ipipe->bufs[i_tail & i_mask];
1760                 obuf = &opipe->bufs[o_head & o_mask];
1761
1762                 if (len >= ibuf->len) {
1763                         /*
1764                          * Simply move the whole buffer from ipipe to opipe
1765                          */
1766                         *obuf = *ibuf;
1767                         ibuf->ops = NULL;
1768                         i_tail++;
1769                         ipipe->tail = i_tail;
1770                         input_wakeup = true;
1771                         o_len = obuf->len;
1772                         o_head++;
1773                         opipe->head = o_head;
1774                 } else {
1775                         /*
1776                          * Get a reference to this pipe buffer,
1777                          * so we can copy the contents over.
1778                          */
1779                         if (!pipe_buf_get(ipipe, ibuf)) {
1780                                 if (ret == 0)
1781                                         ret = -EFAULT;
1782                                 break;
1783                         }
1784                         *obuf = *ibuf;
1785
1786                         /*
1787                          * Don't inherit the gift and merge flags, we need to
1788                          * prevent multiple steals of this page.
1789                          */
1790                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1791                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1792
1793                         obuf->len = len;
1794                         ibuf->offset += len;
1795                         ibuf->len -= len;
1796                         o_len = len;
1797                         o_head++;
1798                         opipe->head = o_head;
1799                 }
1800                 ret += o_len;
1801                 len -= o_len;
1802         } while (len);
1803
1804         pipe_unlock(ipipe);
1805         pipe_unlock(opipe);
1806
1807         /*
1808          * If we put data in the output pipe, wakeup any potential readers.
1809          */
1810         if (ret > 0)
1811                 wakeup_pipe_readers(opipe);
1812
1813         if (input_wakeup)
1814                 wakeup_pipe_writers(ipipe);
1815
1816         return ret;
1817 }
1818
1819 /*
1820  * Link contents of ipipe to opipe.
1821  */
1822 static int link_pipe(struct pipe_inode_info *ipipe,
1823                      struct pipe_inode_info *opipe,
1824                      size_t len, unsigned int flags)
1825 {
1826         struct pipe_buffer *ibuf, *obuf;
1827         unsigned int i_head, o_head;
1828         unsigned int i_tail, o_tail;
1829         unsigned int i_mask, o_mask;
1830         int ret = 0;
1831
1832         /*
1833          * Potential ABBA deadlock, work around it by ordering lock
1834          * grabbing by pipe info address. Otherwise two different processes
1835          * could deadlock (one doing tee from A -> B, the other from B -> A).
1836          */
1837         pipe_double_lock(ipipe, opipe);
1838
1839         i_tail = ipipe->tail;
1840         i_mask = ipipe->ring_size - 1;
1841         o_head = opipe->head;
1842         o_mask = opipe->ring_size - 1;
1843
1844         do {
1845                 if (!opipe->readers) {
1846                         send_sig(SIGPIPE, current, 0);
1847                         if (!ret)
1848                                 ret = -EPIPE;
1849                         break;
1850                 }
1851
1852                 i_head = ipipe->head;
1853                 o_tail = opipe->tail;
1854
1855                 /*
1856                  * If we have iterated all input buffers or run out of
1857                  * output room, break.
1858                  */
1859                 if (pipe_empty(i_head, i_tail) ||
1860                     pipe_full(o_head, o_tail, opipe->max_usage))
1861                         break;
1862
1863                 ibuf = &ipipe->bufs[i_tail & i_mask];
1864                 obuf = &opipe->bufs[o_head & o_mask];
1865
1866                 /*
1867                  * Get a reference to this pipe buffer,
1868                  * so we can copy the contents over.
1869                  */
1870                 if (!pipe_buf_get(ipipe, ibuf)) {
1871                         if (ret == 0)
1872                                 ret = -EFAULT;
1873                         break;
1874                 }
1875
1876                 *obuf = *ibuf;
1877
1878                 /*
1879                  * Don't inherit the gift and merge flag, we need to prevent
1880                  * multiple steals of this page.
1881                  */
1882                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1883                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1884
1885                 if (obuf->len > len)
1886                         obuf->len = len;
1887                 ret += obuf->len;
1888                 len -= obuf->len;
1889
1890                 o_head++;
1891                 opipe->head = o_head;
1892                 i_tail++;
1893         } while (len);
1894
1895         pipe_unlock(ipipe);
1896         pipe_unlock(opipe);
1897
1898         /*
1899          * If we put data in the output pipe, wakeup any potential readers.
1900          */
1901         if (ret > 0)
1902                 wakeup_pipe_readers(opipe);
1903
1904         return ret;
1905 }
1906
1907 /*
1908  * This is a tee(1) implementation that works on pipes. It doesn't copy
1909  * any data, it simply references the 'in' pages on the 'out' pipe.
1910  * The 'flags' used are the SPLICE_F_* variants, currently the only
1911  * applicable one is SPLICE_F_NONBLOCK.
1912  */
1913 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1914 {
1915         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1916         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1917         int ret = -EINVAL;
1918
1919         if (unlikely(!(in->f_mode & FMODE_READ) ||
1920                      !(out->f_mode & FMODE_WRITE)))
1921                 return -EBADF;
1922
1923         /*
1924          * Duplicate the contents of ipipe to opipe without actually
1925          * copying the data.
1926          */
1927         if (ipipe && opipe && ipipe != opipe) {
1928                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1929                         flags |= SPLICE_F_NONBLOCK;
1930
1931                 /*
1932                  * Keep going, unless we encounter an error. The ipipe/opipe
1933                  * ordering doesn't really matter.
1934                  */
1935                 ret = ipipe_prep(ipipe, flags);
1936                 if (!ret) {
1937                         ret = opipe_prep(opipe, flags);
1938                         if (!ret)
1939                                 ret = link_pipe(ipipe, opipe, len, flags);
1940                 }
1941         }
1942
1943         return ret;
1944 }
1945
1946 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1947 {
1948         struct fd in, out;
1949         int error;
1950
1951         if (unlikely(flags & ~SPLICE_F_ALL))
1952                 return -EINVAL;
1953
1954         if (unlikely(!len))
1955                 return 0;
1956
1957         error = -EBADF;
1958         in = fdget(fdin);
1959         if (in.file) {
1960                 out = fdget(fdout);
1961                 if (out.file) {
1962                         error = do_tee(in.file, out.file, len, flags);
1963                         fdput(out);
1964                 }
1965                 fdput(in);
1966         }
1967
1968         return error;
1969 }