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/security.h>
  34 #include <linux/gfp.h>
  35 #include <linux/socket.h>
  36 #include <linux/sched/signal.h>
  37
  38 #include "internal.h"
  39
  40 /*
  41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  42  * a vm helper function, it's already simplified quite a bit by the
  43  * addition of remove_mapping(). If success is returned, the caller may
  44  * attempt to reuse this page for another destination.
  45  */
  46 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  47                 struct pipe_buffer *buf)
  48 {
  49         struct folio *folio = page_folio(buf->page);
  50         struct address_space *mapping;
  51
  52         folio_lock(folio);
  53
  54         mapping = folio_mapping(folio);
  55         if (mapping) {
  56                 WARN_ON(!folio_test_uptodate(folio));
  57
  58                 /*
  59                  * At least for ext2 with nobh option, we need to wait on
  60                  * writeback completing on this folio, since we'll remove it
  61                  * from the pagecache.  Otherwise truncate wont wait on the
  62                  * folio, allowing the disk blocks to be reused by someone else
  63                  * before we actually wrote our data to them. fs corruption
  64                  * ensues.
  65                  */
  66                 folio_wait_writeback(folio);
  67
  68                 if (folio_has_private(folio) &&
  69                     !filemap_release_folio(folio, GFP_KERNEL))
  70                         goto out_unlock;
  71
  72                 /*
  73                  * If we succeeded in removing the mapping, set LRU flag
  74                  * and return good.
  75                  */
  76                 if (remove_mapping(mapping, folio)) {
  77                         buf->flags |= PIPE_BUF_FLAG_LRU;
  78                         return true;
  79                 }
  80         }
  81
  82         /*
  83          * Raced with truncate or failed to remove folio from current
  84          * address space, unlock and return failure.
  85          */
  86 out_unlock:
  87         folio_unlock(folio);
  88         return false;
  89 }
  90
  91 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  92                                         struct pipe_buffer *buf)
  93 {
  94         put_page(buf->page);
  95         buf->flags &= ~PIPE_BUF_FLAG_LRU;
  96 }
  97
  98 /*
  99  * Check whether the contents of buf is OK to access. Since the content
 100  * is a page cache page, IO may be in flight.
 101  */
 102 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 103                                        struct pipe_buffer *buf)
 104 {
 105         struct page *page = buf->page;
 106         int err;
 107
 108         if (!PageUptodate(page)) {
 109                 lock_page(page);
 110
 111                 /*
 112                  * Page got truncated/unhashed. This will cause a 0-byte
 113                  * splice, if this is the first page.
 114                  */
 115                 if (!page->mapping) {
 116                         err = -ENODATA;
 117                         goto error;
 118                 }
 119
 120                 /*
 121                  * Uh oh, read-error from disk.
 122                  */
 123                 if (!PageUptodate(page)) {
 124                         err = -EIO;
 125                         goto error;
 126                 }
 127
 128                 /*
 129                  * Page is ok afterall, we are done.
 130                  */
 131                 unlock_page(page);
 132         }
 133
 134         return 0;
 135 error:
 136         unlock_page(page);
 137         return err;
 138 }
 139
 140 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 141         .confirm        = page_cache_pipe_buf_confirm,
 142         .release        = page_cache_pipe_buf_release,
 143         .try_steal      = page_cache_pipe_buf_try_steal,
 144         .get            = generic_pipe_buf_get,
 145 };
 146
 147 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 148                 struct pipe_buffer *buf)
 149 {
 150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 151                 return false;
 152
 153         buf->flags |= PIPE_BUF_FLAG_LRU;
 154         return generic_pipe_buf_try_steal(pipe, buf);
 155 }
 156
 157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 158         .release        = page_cache_pipe_buf_release,
 159         .try_steal      = user_page_pipe_buf_try_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->rd_wait))
 167                 wake_up_interruptible(&pipe->rd_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         unsigned int tail = pipe->tail;
 187         unsigned int head = pipe->head;
 188         unsigned int mask = pipe->ring_size - 1;
 189         int ret = 0, page_nr = 0;
 190
 191         if (!spd_pages)
 192                 return 0;
 193
 194         if (unlikely(!pipe->readers)) {
 195                 send_sig(SIGPIPE, current, 0);
 196                 ret = -EPIPE;
 197                 goto out;
 198         }
 199
 200         while (!pipe_full(head, tail, pipe->max_usage)) {
 201                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 202
 203                 buf->page = spd->pages[page_nr];
 204                 buf->offset = spd->partial[page_nr].offset;
 205                 buf->len = spd->partial[page_nr].len;
 206                 buf->private = spd->partial[page_nr].private;
 207                 buf->ops = spd->ops;
 208                 buf->flags = 0;
 209
 210                 head++;
 211                 pipe->head = head;
 212                 page_nr++;
 213                 ret += buf->len;
 214
 215                 if (!--spd->nr_pages)
 216                         break;
 217         }
 218
 219         if (!ret)
 220                 ret = -EAGAIN;
 221
 222 out:
 223         while (page_nr < spd_pages)
 224                 spd->spd_release(spd, page_nr++);
 225
 226         return ret;
 227 }
 228 EXPORT_SYMBOL_GPL(splice_to_pipe);
 229
 230 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 231 {
 232         unsigned int head = pipe->head;
 233         unsigned int tail = pipe->tail;
 234         unsigned int mask = pipe->ring_size - 1;
 235         int ret;
 236
 237         if (unlikely(!pipe->readers)) {
 238                 send_sig(SIGPIPE, current, 0);
 239                 ret = -EPIPE;
 240         } else if (pipe_full(head, tail, pipe->max_usage)) {
 241                 ret = -EAGAIN;
 242         } else {
 243                 pipe->bufs[head & mask] = *buf;
 244                 pipe->head = head + 1;
 245                 return buf->len;
 246         }
 247         pipe_buf_release(pipe, buf);
 248         return ret;
 249 }
 250 EXPORT_SYMBOL(add_to_pipe);
 251
 252 /*
 253  * Check if we need to grow the arrays holding pages and partial page
 254  * descriptions.
 255  */
 256 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 257 {
 258         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 259
 260         spd->nr_pages_max = max_usage;
 261         if (max_usage <= PIPE_DEF_BUFFERS)
 262                 return 0;
 263
 264         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 265         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 266                                      GFP_KERNEL);
 267
 268         if (spd->pages && spd->partial)
 269                 return 0;
 270
 271         kfree(spd->pages);
 272         kfree(spd->partial);
 273         return -ENOMEM;
 274 }
 275
 276 void splice_shrink_spd(struct splice_pipe_desc *spd)
 277 {
 278         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 279                 return;
 280
 281         kfree(spd->pages);
 282         kfree(spd->partial);
 283 }
 284
 285 /*
 286  * Splice data from an O_DIRECT file into pages and then add them to the output
 287  * pipe.
 288  */
 289 ssize_t direct_splice_read(struct file *in, loff_t *ppos,
 290                            struct pipe_inode_info *pipe,
 291                            size_t len, unsigned int flags)
 292 {
 293         struct iov_iter to;
 294         struct bio_vec *bv;
 295         struct kiocb kiocb;
 296         struct page **pages;
 297         ssize_t ret;
 298         size_t used, npages, chunk, remain, reclaim;
 299         int i;
 300
 301         /* Work out how much data we can actually add into the pipe */
 302         used = pipe_occupancy(pipe->head, pipe->tail);
 303         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 304         len = min_t(size_t, len, npages * PAGE_SIZE);
 305         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 306
 307         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 308                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 309         if (!bv)
 310                 return -ENOMEM;
 311
 312         pages = (void *)(bv + npages);
 313         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 314         if (!npages) {
 315                 kfree(bv);
 316                 return -ENOMEM;
 317         }
 318
 319         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 320
 321         for (i = 0; i < npages; i++) {
 322                 chunk = min_t(size_t, PAGE_SIZE, remain);
 323                 bv[i].bv_page = pages[i];
 324                 bv[i].bv_offset = 0;
 325                 bv[i].bv_len = chunk;
 326                 remain -= chunk;
 327         }
 328
 329         /* Do the I/O */
 330         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 331         init_sync_kiocb(&kiocb, in);
 332         kiocb.ki_pos = *ppos;
 333         ret = call_read_iter(in, &kiocb, &to);
 334
 335         reclaim = npages * PAGE_SIZE;
 336         remain = 0;
 337         if (ret > 0) {
 338                 reclaim -= ret;
 339                 remain = ret;
 340                 *ppos = kiocb.ki_pos;
 341                 file_accessed(in);
 342         } else if (ret < 0) {
 343                 /*
 344                  * callers of ->splice_read() expect -EAGAIN on
 345                  * "can't put anything in there", rather than -EFAULT.
 346                  */
 347                 if (ret == -EFAULT)
 348                         ret = -EAGAIN;
 349         }
 350
 351         /* Free any pages that didn't get touched at all. */
 352         reclaim /= PAGE_SIZE;
 353         if (reclaim) {
 354                 npages -= reclaim;
 355                 release_pages(pages + npages, reclaim);
 356         }
 357
 358         /* Push the remaining pages into the pipe. */
 359         for (i = 0; i < npages; i++) {
 360                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 361
 362                 chunk = min_t(size_t, remain, PAGE_SIZE);
 363                 *buf = (struct pipe_buffer) {
 364                         .ops    = &default_pipe_buf_ops,
 365                         .page   = bv[i].bv_page,
 366                         .offset = 0,
 367                         .len    = chunk,
 368                 };
 369                 pipe->head++;
 370                 remain -= chunk;
 371         }
 372
 373         kfree(bv);
 374         return ret;
 375 }
 376 EXPORT_SYMBOL(direct_splice_read);
 377
 378 /**
 379  * generic_file_splice_read - splice data from file to a pipe
 380  * @in:         file to splice from
 381  * @ppos:       position in @in
 382  * @pipe:       pipe to splice to
 383  * @len:        number of bytes to splice
 384  * @flags:      splice modifier flags
 385  *
 386  * Description:
 387  *    Will read pages from given file and fill them into a pipe. Can be
 388  *    used as long as it has more or less sane ->read_iter().
 389  *
 390  */
 391 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 392                                  struct pipe_inode_info *pipe, size_t len,
 393                                  unsigned int flags)
 394 {
 395         struct iov_iter to;
 396         struct kiocb kiocb;
 397         int ret;
 398
 399         iov_iter_pipe(&to, ITER_DEST, pipe, len);
 400         init_sync_kiocb(&kiocb, in);
 401         kiocb.ki_pos = *ppos;
 402         ret = call_read_iter(in, &kiocb, &to);
 403         if (ret > 0) {
 404                 *ppos = kiocb.ki_pos;
 405                 file_accessed(in);
 406         } else if (ret < 0) {
 407                 /* free what was emitted */
 408                 pipe_discard_from(pipe, to.start_head);
 409                 /*
 410                  * callers of ->splice_read() expect -EAGAIN on
 411                  * "can't put anything in there", rather than -EFAULT.
 412                  */
 413                 if (ret == -EFAULT)
 414                         ret = -EAGAIN;
 415         }
 416
 417         return ret;
 418 }
 419 EXPORT_SYMBOL(generic_file_splice_read);
 420
 421 const struct pipe_buf_operations default_pipe_buf_ops = {
 422         .release        = generic_pipe_buf_release,
 423         .try_steal      = generic_pipe_buf_try_steal,
 424         .get            = generic_pipe_buf_get,
 425 };
 426
 427 /* Pipe buffer operations for a socket and similar. */
 428 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 429         .release        = generic_pipe_buf_release,
 430         .get            = generic_pipe_buf_get,
 431 };
 432 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 433
 434 /*
 435  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 436  * using sendpage(). Return the number of bytes sent.
 437  */
 438 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 439                             struct pipe_buffer *buf, struct splice_desc *sd)
 440 {
 441         struct file *file = sd->u.file;
 442         loff_t pos = sd->pos;
 443         int more;
 444
 445         if (!likely(file->f_op->sendpage))
 446                 return -EINVAL;
 447
 448         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 449
 450         if (sd->len < sd->total_len &&
 451             pipe_occupancy(pipe->head, pipe->tail) > 1)
 452                 more |= MSG_SENDPAGE_NOTLAST;
 453
 454         return file->f_op->sendpage(file, buf->page, buf->offset,
 455                                     sd->len, &pos, more);
 456 }
 457
 458 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 459 {
 460         smp_mb();
 461         if (waitqueue_active(&pipe->wr_wait))
 462                 wake_up_interruptible(&pipe->wr_wait);
 463         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 464 }
 465
 466 /**
 467  * splice_from_pipe_feed - feed available data from a pipe to a file
 468  * @pipe:       pipe to splice from
 469  * @sd:         information to @actor
 470  * @actor:      handler that splices the data
 471  *
 472  * Description:
 473  *    This function loops over the pipe and calls @actor to do the
 474  *    actual moving of a single struct pipe_buffer to the desired
 475  *    destination.  It returns when there's no more buffers left in
 476  *    the pipe or if the requested number of bytes (@sd->total_len)
 477  *    have been copied.  It returns a positive number (one) if the
 478  *    pipe needs to be filled with more data, zero if the required
 479  *    number of bytes have been copied and -errno on error.
 480  *
 481  *    This, together with splice_from_pipe_{begin,end,next}, may be
 482  *    used to implement the functionality of __splice_from_pipe() when
 483  *    locking is required around copying the pipe buffers to the
 484  *    destination.
 485  */
 486 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 487                           splice_actor *actor)
 488 {
 489         unsigned int head = pipe->head;
 490         unsigned int tail = pipe->tail;
 491         unsigned int mask = pipe->ring_size - 1;
 492         int ret;
 493
 494         while (!pipe_empty(head, tail)) {
 495                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 496
 497                 sd->len = buf->len;
 498                 if (sd->len > sd->total_len)
 499                         sd->len = sd->total_len;
 500
 501                 ret = pipe_buf_confirm(pipe, buf);
 502                 if (unlikely(ret)) {
 503                         if (ret == -ENODATA)
 504                                 ret = 0;
 505                         return ret;
 506                 }
 507
 508                 ret = actor(pipe, buf, sd);
 509                 if (ret <= 0)
 510                         return ret;
 511
 512                 buf->offset += ret;
 513                 buf->len -= ret;
 514
 515                 sd->num_spliced += ret;
 516                 sd->len -= ret;
 517                 sd->pos += ret;
 518                 sd->total_len -= ret;
 519
 520                 if (!buf->len) {
 521                         pipe_buf_release(pipe, buf);
 522                         tail++;
 523                         pipe->tail = tail;
 524                         if (pipe->files)
 525                                 sd->need_wakeup = true;
 526                 }
 527
 528                 if (!sd->total_len)
 529                         return 0;
 530         }
 531
 532         return 1;
 533 }
 534
 535 /* We know we have a pipe buffer, but maybe it's empty? */
 536 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 537 {
 538         unsigned int tail = pipe->tail;
 539         unsigned int mask = pipe->ring_size - 1;
 540         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 541
 542         if (unlikely(!buf->len)) {
 543                 pipe_buf_release(pipe, buf);
 544                 pipe->tail = tail+1;
 545                 return true;
 546         }
 547
 548         return false;
 549 }
 550
 551 /**
 552  * splice_from_pipe_next - wait for some data to splice from
 553  * @pipe:       pipe to splice from
 554  * @sd:         information about the splice operation
 555  *
 556  * Description:
 557  *    This function will wait for some data and return a positive
 558  *    value (one) if pipe buffers are available.  It will return zero
 559  *    or -errno if no more data needs to be spliced.
 560  */
 561 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 562 {
 563         /*
 564          * Check for signal early to make process killable when there are
 565          * always buffers available
 566          */
 567         if (signal_pending(current))
 568                 return -ERESTARTSYS;
 569
 570 repeat:
 571         while (pipe_empty(pipe->head, pipe->tail)) {
 572                 if (!pipe->writers)
 573                         return 0;
 574
 575                 if (sd->num_spliced)
 576                         return 0;
 577
 578                 if (sd->flags & SPLICE_F_NONBLOCK)
 579                         return -EAGAIN;
 580
 581                 if (signal_pending(current))
 582                         return -ERESTARTSYS;
 583
 584                 if (sd->need_wakeup) {
 585                         wakeup_pipe_writers(pipe);
 586                         sd->need_wakeup = false;
 587                 }
 588
 589                 pipe_wait_readable(pipe);
 590         }
 591
 592         if (eat_empty_buffer(pipe))
 593                 goto repeat;
 594
 595         return 1;
 596 }
 597
 598 /**
 599  * splice_from_pipe_begin - start splicing from pipe
 600  * @sd:         information about the splice operation
 601  *
 602  * Description:
 603  *    This function should be called before a loop containing
 604  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 605  *    initialize the necessary fields of @sd.
 606  */
 607 static void splice_from_pipe_begin(struct splice_desc *sd)
 608 {
 609         sd->num_spliced = 0;
 610         sd->need_wakeup = false;
 611 }
 612
 613 /**
 614  * splice_from_pipe_end - finish splicing from pipe
 615  * @pipe:       pipe to splice from
 616  * @sd:         information about the splice operation
 617  *
 618  * Description:
 619  *    This function will wake up pipe writers if necessary.  It should
 620  *    be called after a loop containing splice_from_pipe_next() and
 621  *    splice_from_pipe_feed().
 622  */
 623 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 624 {
 625         if (sd->need_wakeup)
 626                 wakeup_pipe_writers(pipe);
 627 }
 628
 629 /**
 630  * __splice_from_pipe - splice data from a pipe to given actor
 631  * @pipe:       pipe to splice from
 632  * @sd:         information to @actor
 633  * @actor:      handler that splices the data
 634  *
 635  * Description:
 636  *    This function does little more than loop over the pipe and call
 637  *    @actor to do the actual moving of a single struct pipe_buffer to
 638  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 639  *    pipe_to_user.
 640  *
 641  */
 642 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 643                            splice_actor *actor)
 644 {
 645         int ret;
 646
 647         splice_from_pipe_begin(sd);
 648         do {
 649                 cond_resched();
 650                 ret = splice_from_pipe_next(pipe, sd);
 651                 if (ret > 0)
 652                         ret = splice_from_pipe_feed(pipe, sd, actor);
 653         } while (ret > 0);
 654         splice_from_pipe_end(pipe, sd);
 655
 656         return sd->num_spliced ? sd->num_spliced : ret;
 657 }
 658 EXPORT_SYMBOL(__splice_from_pipe);
 659
 660 /**
 661  * splice_from_pipe - splice data from a pipe to a file
 662  * @pipe:       pipe to splice from
 663  * @out:        file to splice to
 664  * @ppos:       position in @out
 665  * @len:        how many bytes to splice
 666  * @flags:      splice modifier flags
 667  * @actor:      handler that splices the data
 668  *
 669  * Description:
 670  *    See __splice_from_pipe. This function locks the pipe inode,
 671  *    otherwise it's identical to __splice_from_pipe().
 672  *
 673  */
 674 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 675                          loff_t *ppos, size_t len, unsigned int flags,
 676                          splice_actor *actor)
 677 {
 678         ssize_t ret;
 679         struct splice_desc sd = {
 680                 .total_len = len,
 681                 .flags = flags,
 682                 .pos = *ppos,
 683                 .u.file = out,
 684         };
 685
 686         pipe_lock(pipe);
 687         ret = __splice_from_pipe(pipe, &sd, actor);
 688         pipe_unlock(pipe);
 689
 690         return ret;
 691 }
 692
 693 /**
 694  * iter_file_splice_write - splice data from a pipe to a file
 695  * @pipe:       pipe info
 696  * @out:        file to write to
 697  * @ppos:       position in @out
 698  * @len:        number of bytes to splice
 699  * @flags:      splice modifier flags
 700  *
 701  * Description:
 702  *    Will either move or copy pages (determined by @flags options) from
 703  *    the given pipe inode to the given file.
 704  *    This one is ->write_iter-based.
 705  *
 706  */
 707 ssize_t
 708 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 709                           loff_t *ppos, size_t len, unsigned int flags)
 710 {
 711         struct splice_desc sd = {
 712                 .total_len = len,
 713                 .flags = flags,
 714                 .pos = *ppos,
 715                 .u.file = out,
 716         };
 717         int nbufs = pipe->max_usage;
 718         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 719                                         GFP_KERNEL);
 720         ssize_t ret;
 721
 722         if (unlikely(!array))
 723                 return -ENOMEM;
 724
 725         pipe_lock(pipe);
 726
 727         splice_from_pipe_begin(&sd);
 728         while (sd.total_len) {
 729                 struct iov_iter from;
 730                 unsigned int head, tail, mask;
 731                 size_t left;
 732                 int n;
 733
 734                 ret = splice_from_pipe_next(pipe, &sd);
 735                 if (ret <= 0)
 736                         break;
 737
 738                 if (unlikely(nbufs < pipe->max_usage)) {
 739                         kfree(array);
 740                         nbufs = pipe->max_usage;
 741                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 742                                         GFP_KERNEL);
 743                         if (!array) {
 744                                 ret = -ENOMEM;
 745                                 break;
 746                         }
 747                 }
 748
 749                 head = pipe->head;
 750                 tail = pipe->tail;
 751                 mask = pipe->ring_size - 1;
 752
 753                 /* build the vector */
 754                 left = sd.total_len;
 755                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 756                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 757                         size_t this_len = buf->len;
 758
 759                         /* zero-length bvecs are not supported, skip them */
 760                         if (!this_len)
 761                                 continue;
 762                         this_len = min(this_len, left);
 763
 764                         ret = pipe_buf_confirm(pipe, buf);
 765                         if (unlikely(ret)) {
 766                                 if (ret == -ENODATA)
 767                                         ret = 0;
 768                                 goto done;
 769                         }
 770
 771                         array[n].bv_page = buf->page;
 772                         array[n].bv_len = this_len;
 773                         array[n].bv_offset = buf->offset;
 774                         left -= this_len;
 775                         n++;
 776                 }
 777
 778                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 779                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 780                 if (ret <= 0)
 781                         break;
 782
 783                 sd.num_spliced += ret;
 784                 sd.total_len -= ret;
 785                 *ppos = sd.pos;
 786
 787                 /* dismiss the fully eaten buffers, adjust the partial one */
 788                 tail = pipe->tail;
 789                 while (ret) {
 790                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 791                         if (ret >= buf->len) {
 792                                 ret -= buf->len;
 793                                 buf->len = 0;
 794                                 pipe_buf_release(pipe, buf);
 795                                 tail++;
 796                                 pipe->tail = tail;
 797                                 if (pipe->files)
 798                                         sd.need_wakeup = true;
 799                         } else {
 800                                 buf->offset += ret;
 801                                 buf->len -= ret;
 802                                 ret = 0;
 803                         }
 804                 }
 805         }
 806 done:
 807         kfree(array);
 808         splice_from_pipe_end(pipe, &sd);
 809
 810         pipe_unlock(pipe);
 811
 812         if (sd.num_spliced)
 813                 ret = sd.num_spliced;
 814
 815         return ret;
 816 }
 817
 818 EXPORT_SYMBOL(iter_file_splice_write);
 819
 820 /**
 821  * generic_splice_sendpage - splice data from a pipe to a socket
 822  * @pipe:       pipe to splice from
 823  * @out:        socket to write to
 824  * @ppos:       position in @out
 825  * @len:        number of bytes to splice
 826  * @flags:      splice modifier flags
 827  *
 828  * Description:
 829  *    Will send @len bytes from the pipe to a network socket. No data copying
 830  *    is involved.
 831  *
 832  */
 833 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 834                                 loff_t *ppos, size_t len, unsigned int flags)
 835 {
 836         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 837 }
 838
 839 EXPORT_SYMBOL(generic_splice_sendpage);
 840
 841 static int warn_unsupported(struct file *file, const char *op)
 842 {
 843         pr_debug_ratelimited(
 844                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 845                 op, file, current->pid, current->comm);
 846         return -EINVAL;
 847 }
 848
 849 /*
 850  * Attempt to initiate a splice from pipe to file.
 851  */
 852 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 853                            loff_t *ppos, size_t len, unsigned int flags)
 854 {
 855         if (unlikely(!out->f_op->splice_write))
 856                 return warn_unsupported(out, "write");
 857         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 858 }
 859
 860 /*
 861  * Attempt to initiate a splice from a file to a pipe.
 862  */
 863 static long do_splice_to(struct file *in, loff_t *ppos,
 864                          struct pipe_inode_info *pipe, size_t len,
 865                          unsigned int flags)
 866 {
 867         unsigned int p_space;
 868         int ret;
 869
 870         if (unlikely(!(in->f_mode & FMODE_READ)))
 871                 return -EBADF;
 872
 873         /* Don't try to read more the pipe has space for. */
 874         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 875         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 876
 877         ret = rw_verify_area(READ, in, ppos, len);
 878         if (unlikely(ret < 0))
 879                 return ret;
 880
 881         if (unlikely(len > MAX_RW_COUNT))
 882                 len = MAX_RW_COUNT;
 883
 884         if (unlikely(!in->f_op->splice_read))
 885                 return warn_unsupported(in, "read");
 886         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 887 }
 888
 889 /**
 890  * splice_direct_to_actor - splices data directly between two non-pipes
 891  * @in:         file to splice from
 892  * @sd:         actor information on where to splice to
 893  * @actor:      handles the data splicing
 894  *
 895  * Description:
 896  *    This is a special case helper to splice directly between two
 897  *    points, without requiring an explicit pipe. Internally an allocated
 898  *    pipe is cached in the process, and reused during the lifetime of
 899  *    that process.
 900  *
 901  */
 902 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 903                                splice_direct_actor *actor)
 904 {
 905         struct pipe_inode_info *pipe;
 906         long ret, bytes;
 907         size_t len;
 908         int i, flags, more;
 909
 910         /*
 911          * We require the input to be seekable, as we don't want to randomly
 912          * drop data for eg socket -> socket splicing. Use the piped splicing
 913          * for that!
 914          */
 915         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
 916                 return -EINVAL;
 917
 918         /*
 919          * neither in nor out is a pipe, setup an internal pipe attached to
 920          * 'out' and transfer the wanted data from 'in' to 'out' through that
 921          */
 922         pipe = current->splice_pipe;
 923         if (unlikely(!pipe)) {
 924                 pipe = alloc_pipe_info();
 925                 if (!pipe)
 926                         return -ENOMEM;
 927
 928                 /*
 929                  * We don't have an immediate reader, but we'll read the stuff
 930                  * out of the pipe right after the splice_to_pipe(). So set
 931                  * PIPE_READERS appropriately.
 932                  */
 933                 pipe->readers = 1;
 934
 935                 current->splice_pipe = pipe;
 936         }
 937
 938         /*
 939          * Do the splice.
 940          */
 941         ret = 0;
 942         bytes = 0;
 943         len = sd->total_len;
 944         flags = sd->flags;
 945
 946         /*
 947          * Don't block on output, we have to drain the direct pipe.
 948          */
 949         sd->flags &= ~SPLICE_F_NONBLOCK;
 950         more = sd->flags & SPLICE_F_MORE;
 951
 952         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 953
 954         while (len) {
 955                 size_t read_len;
 956                 loff_t pos = sd->pos, prev_pos = pos;
 957
 958                 ret = do_splice_to(in, &pos, pipe, len, flags);
 959                 if (unlikely(ret <= 0))
 960                         goto out_release;
 961
 962                 read_len = ret;
 963                 sd->total_len = read_len;
 964
 965                 /*
 966                  * If more data is pending, set SPLICE_F_MORE
 967                  * If this is the last data and SPLICE_F_MORE was not set
 968                  * initially, clears it.
 969                  */
 970                 if (read_len < len)
 971                         sd->flags |= SPLICE_F_MORE;
 972                 else if (!more)
 973                         sd->flags &= ~SPLICE_F_MORE;
 974                 /*
 975                  * NOTE: nonblocking mode only applies to the input. We
 976                  * must not do the output in nonblocking mode as then we
 977                  * could get stuck data in the internal pipe:
 978                  */
 979                 ret = actor(pipe, sd);
 980                 if (unlikely(ret <= 0)) {
 981                         sd->pos = prev_pos;
 982                         goto out_release;
 983                 }
 984
 985                 bytes += ret;
 986                 len -= ret;
 987                 sd->pos = pos;
 988
 989                 if (ret < read_len) {
 990                         sd->pos = prev_pos + ret;
 991                         goto out_release;
 992                 }
 993         }
 994
 995 done:
 996         pipe->tail = pipe->head = 0;
 997         file_accessed(in);
 998         return bytes;
 999
1000 out_release:
1001         /*
1002          * If we did an incomplete transfer we must release
1003          * the pipe buffers in question:
1004          */
1005         for (i = 0; i < pipe->ring_size; i++) {
1006                 struct pipe_buffer *buf = &pipe->bufs[i];
1007
1008                 if (buf->ops)
1009                         pipe_buf_release(pipe, buf);
1010         }
1011
1012         if (!bytes)
1013                 bytes = ret;
1014
1015         goto done;
1016 }
1017 EXPORT_SYMBOL(splice_direct_to_actor);
1018
1019 static int direct_splice_actor(struct pipe_inode_info *pipe,
1020                                struct splice_desc *sd)
1021 {
1022         struct file *file = sd->u.file;
1023
1024         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1025                               sd->flags);
1026 }
1027
1028 /**
1029  * do_splice_direct - splices data directly between two files
1030  * @in:         file to splice from
1031  * @ppos:       input file offset
1032  * @out:        file to splice to
1033  * @opos:       output file offset
1034  * @len:        number of bytes to splice
1035  * @flags:      splice modifier flags
1036  *
1037  * Description:
1038  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1039  *    doing it in the application would incur an extra system call
1040  *    (splice in + splice out, as compared to just sendfile()). So this helper
1041  *    can splice directly through a process-private pipe.
1042  *
1043  */
1044 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1045                       loff_t *opos, size_t len, unsigned int flags)
1046 {
1047         struct splice_desc sd = {
1048                 .len            = len,
1049                 .total_len      = len,
1050                 .flags          = flags,
1051                 .pos            = *ppos,
1052                 .u.file         = out,
1053                 .opos           = opos,
1054         };
1055         long ret;
1056
1057         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1058                 return -EBADF;
1059
1060         if (unlikely(out->f_flags & O_APPEND))
1061                 return -EINVAL;
1062
1063         ret = rw_verify_area(WRITE, out, opos, len);
1064         if (unlikely(ret < 0))
1065                 return ret;
1066
1067         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1068         if (ret > 0)
1069                 *ppos = sd.pos;
1070
1071         return ret;
1072 }
1073 EXPORT_SYMBOL(do_splice_direct);
1074
1075 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1076 {
1077         for (;;) {
1078                 if (unlikely(!pipe->readers)) {
1079                         send_sig(SIGPIPE, current, 0);
1080                         return -EPIPE;
1081                 }
1082                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1083                         return 0;
1084                 if (flags & SPLICE_F_NONBLOCK)
1085                         return -EAGAIN;
1086                 if (signal_pending(current))
1087                         return -ERESTARTSYS;
1088                 pipe_wait_writable(pipe);
1089         }
1090 }
1091
1092 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1093                                struct pipe_inode_info *opipe,
1094                                size_t len, unsigned int flags);
1095
1096 long splice_file_to_pipe(struct file *in,
1097                          struct pipe_inode_info *opipe,
1098                          loff_t *offset,
1099                          size_t len, unsigned int flags)
1100 {
1101         long ret;
1102
1103         pipe_lock(opipe);
1104         ret = wait_for_space(opipe, flags);
1105         if (!ret)
1106                 ret = do_splice_to(in, offset, opipe, len, flags);
1107         pipe_unlock(opipe);
1108         if (ret > 0)
1109                 wakeup_pipe_readers(opipe);
1110         return ret;
1111 }
1112
1113 /*
1114  * Determine where to splice to/from.
1115  */
1116 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1117                loff_t *off_out, size_t len, unsigned int flags)
1118 {
1119         struct pipe_inode_info *ipipe;
1120         struct pipe_inode_info *opipe;
1121         loff_t offset;
1122         long ret;
1123
1124         if (unlikely(!(in->f_mode & FMODE_READ) ||
1125                      !(out->f_mode & FMODE_WRITE)))
1126                 return -EBADF;
1127
1128         ipipe = get_pipe_info(in, true);
1129         opipe = get_pipe_info(out, true);
1130
1131         if (ipipe && opipe) {
1132                 if (off_in || off_out)
1133                         return -ESPIPE;
1134
1135                 /* Splicing to self would be fun, but... */
1136                 if (ipipe == opipe)
1137                         return -EINVAL;
1138
1139                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1140                         flags |= SPLICE_F_NONBLOCK;
1141
1142                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1143         }
1144
1145         if (ipipe) {
1146                 if (off_in)
1147                         return -ESPIPE;
1148                 if (off_out) {
1149                         if (!(out->f_mode & FMODE_PWRITE))
1150                                 return -EINVAL;
1151                         offset = *off_out;
1152                 } else {
1153                         offset = out->f_pos;
1154                 }
1155
1156                 if (unlikely(out->f_flags & O_APPEND))
1157                         return -EINVAL;
1158
1159                 ret = rw_verify_area(WRITE, out, &offset, len);
1160                 if (unlikely(ret < 0))
1161                         return ret;
1162
1163                 if (in->f_flags & O_NONBLOCK)
1164                         flags |= SPLICE_F_NONBLOCK;
1165
1166                 file_start_write(out);
1167                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1168                 file_end_write(out);
1169
1170                 if (!off_out)
1171                         out->f_pos = offset;
1172                 else
1173                         *off_out = offset;
1174
1175                 return ret;
1176         }
1177
1178         if (opipe) {
1179                 if (off_out)
1180                         return -ESPIPE;
1181                 if (off_in) {
1182                         if (!(in->f_mode & FMODE_PREAD))
1183                                 return -EINVAL;
1184                         offset = *off_in;
1185                 } else {
1186                         offset = in->f_pos;
1187                 }
1188
1189                 if (out->f_flags & O_NONBLOCK)
1190                         flags |= SPLICE_F_NONBLOCK;
1191
1192                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1193                 if (!off_in)
1194                         in->f_pos = offset;
1195                 else
1196                         *off_in = offset;
1197
1198                 return ret;
1199         }
1200
1201         return -EINVAL;
1202 }
1203
1204 static long __do_splice(struct file *in, loff_t __user *off_in,
1205                         struct file *out, loff_t __user *off_out,
1206                         size_t len, unsigned int flags)
1207 {
1208         struct pipe_inode_info *ipipe;
1209         struct pipe_inode_info *opipe;
1210         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1211         long ret;
1212
1213         ipipe = get_pipe_info(in, true);
1214         opipe = get_pipe_info(out, true);
1215
1216         if (ipipe && off_in)
1217                 return -ESPIPE;
1218         if (opipe && off_out)
1219                 return -ESPIPE;
1220
1221         if (off_out) {
1222                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1223                         return -EFAULT;
1224                 __off_out = &offset;
1225         }
1226         if (off_in) {
1227                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1228                         return -EFAULT;
1229                 __off_in = &offset;
1230         }
1231
1232         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1233         if (ret < 0)
1234                 return ret;
1235
1236         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1237                 return -EFAULT;
1238         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1239                 return -EFAULT;
1240
1241         return ret;
1242 }
1243
1244 static int iter_to_pipe(struct iov_iter *from,
1245                         struct pipe_inode_info *pipe,
1246                         unsigned flags)
1247 {
1248         struct pipe_buffer buf = {
1249                 .ops = &user_page_pipe_buf_ops,
1250                 .flags = flags
1251         };
1252         size_t total = 0;
1253         int ret = 0;
1254
1255         while (iov_iter_count(from)) {
1256                 struct page *pages[16];
1257                 ssize_t left;
1258                 size_t start;
1259                 int i, n;
1260
1261                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1262                 if (left <= 0) {
1263                         ret = left;
1264                         break;
1265                 }
1266
1267                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1268                 for (i = 0; i < n; i++) {
1269                         int size = min_t(int, left, PAGE_SIZE - start);
1270
1271                         buf.page = pages[i];
1272                         buf.offset = start;
1273                         buf.len = size;
1274                         ret = add_to_pipe(pipe, &buf);
1275                         if (unlikely(ret < 0)) {
1276                                 iov_iter_revert(from, left);
1277                                 // this one got dropped by add_to_pipe()
1278                                 while (++i < n)
1279                                         put_page(pages[i]);
1280                                 goto out;
1281                         }
1282                         total += ret;
1283                         left -= size;
1284                         start = 0;
1285                 }
1286         }
1287 out:
1288         return total ? total : ret;
1289 }
1290
1291 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1292                         struct splice_desc *sd)
1293 {
1294         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1295         return n == sd->len ? n : -EFAULT;
1296 }
1297
1298 /*
1299  * For lack of a better implementation, implement vmsplice() to userspace
1300  * as a simple copy of the pipes pages to the user iov.
1301  */
1302 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1303                              unsigned int flags)
1304 {
1305         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1306         struct splice_desc sd = {
1307                 .total_len = iov_iter_count(iter),
1308                 .flags = flags,
1309                 .u.data = iter
1310         };
1311         long ret = 0;
1312
1313         if (!pipe)
1314                 return -EBADF;
1315
1316         if (sd.total_len) {
1317                 pipe_lock(pipe);
1318                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1319                 pipe_unlock(pipe);
1320         }
1321
1322         return ret;
1323 }
1324
1325 /*
1326  * vmsplice splices a user address range into a pipe. It can be thought of
1327  * as splice-from-memory, where the regular splice is splice-from-file (or
1328  * to file). In both cases the output is a pipe, naturally.
1329  */
1330 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1331                              unsigned int flags)
1332 {
1333         struct pipe_inode_info *pipe;
1334         long ret = 0;
1335         unsigned buf_flag = 0;
1336
1337         if (flags & SPLICE_F_GIFT)
1338                 buf_flag = PIPE_BUF_FLAG_GIFT;
1339
1340         pipe = get_pipe_info(file, true);
1341         if (!pipe)
1342                 return -EBADF;
1343
1344         pipe_lock(pipe);
1345         ret = wait_for_space(pipe, flags);
1346         if (!ret)
1347                 ret = iter_to_pipe(iter, pipe, buf_flag);
1348         pipe_unlock(pipe);
1349         if (ret > 0)
1350                 wakeup_pipe_readers(pipe);
1351         return ret;
1352 }
1353
1354 static int vmsplice_type(struct fd f, int *type)
1355 {
1356         if (!f.file)
1357                 return -EBADF;
1358         if (f.file->f_mode & FMODE_WRITE) {
1359                 *type = ITER_SOURCE;
1360         } else if (f.file->f_mode & FMODE_READ) {
1361                 *type = ITER_DEST;
1362         } else {
1363                 fdput(f);
1364                 return -EBADF;
1365         }
1366         return 0;
1367 }
1368
1369 /*
1370  * Note that vmsplice only really supports true splicing _from_ user memory
1371  * to a pipe, not the other way around. Splicing from user memory is a simple
1372  * operation that can be supported without any funky alignment restrictions
1373  * or nasty vm tricks. We simply map in the user memory and fill them into
1374  * a pipe. The reverse isn't quite as easy, though. There are two possible
1375  * solutions for that:
1376  *
1377  *      - memcpy() the data internally, at which point we might as well just
1378  *        do a regular read() on the buffer anyway.
1379  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1380  *        has restriction limitations on both ends of the pipe).
1381  *
1382  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1383  *
1384  */
1385 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1386                 unsigned long, nr_segs, unsigned int, flags)
1387 {
1388         struct iovec iovstack[UIO_FASTIOV];
1389         struct iovec *iov = iovstack;
1390         struct iov_iter iter;
1391         ssize_t error;
1392         struct fd f;
1393         int type;
1394
1395         if (unlikely(flags & ~SPLICE_F_ALL))
1396                 return -EINVAL;
1397
1398         f = fdget(fd);
1399         error = vmsplice_type(f, &type);
1400         if (error)
1401                 return error;
1402
1403         error = import_iovec(type, uiov, nr_segs,
1404                              ARRAY_SIZE(iovstack), &iov, &iter);
1405         if (error < 0)
1406                 goto out_fdput;
1407
1408         if (!iov_iter_count(&iter))
1409                 error = 0;
1410         else if (type == ITER_SOURCE)
1411                 error = vmsplice_to_pipe(f.file, &iter, flags);
1412         else
1413                 error = vmsplice_to_user(f.file, &iter, flags);
1414
1415         kfree(iov);
1416 out_fdput:
1417         fdput(f);
1418         return error;
1419 }
1420
1421 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1422                 int, fd_out, loff_t __user *, off_out,
1423                 size_t, len, unsigned int, flags)
1424 {
1425         struct fd in, out;
1426         long error;
1427
1428         if (unlikely(!len))
1429                 return 0;
1430
1431         if (unlikely(flags & ~SPLICE_F_ALL))
1432                 return -EINVAL;
1433
1434         error = -EBADF;
1435         in = fdget(fd_in);
1436         if (in.file) {
1437                 out = fdget(fd_out);
1438                 if (out.file) {
1439                         error = __do_splice(in.file, off_in, out.file, off_out,
1440                                                 len, flags);
1441                         fdput(out);
1442                 }
1443                 fdput(in);
1444         }
1445         return error;
1446 }
1447
1448 /*
1449  * Make sure there's data to read. Wait for input if we can, otherwise
1450  * return an appropriate error.
1451  */
1452 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1453 {
1454         int ret;
1455
1456         /*
1457          * Check the pipe occupancy without the inode lock first. This function
1458          * is speculative anyways, so missing one is ok.
1459          */
1460         if (!pipe_empty(pipe->head, pipe->tail))
1461                 return 0;
1462
1463         ret = 0;
1464         pipe_lock(pipe);
1465
1466         while (pipe_empty(pipe->head, pipe->tail)) {
1467                 if (signal_pending(current)) {
1468                         ret = -ERESTARTSYS;
1469                         break;
1470                 }
1471                 if (!pipe->writers)
1472                         break;
1473                 if (flags & SPLICE_F_NONBLOCK) {
1474                         ret = -EAGAIN;
1475                         break;
1476                 }
1477                 pipe_wait_readable(pipe);
1478         }
1479
1480         pipe_unlock(pipe);
1481         return ret;
1482 }
1483
1484 /*
1485  * Make sure there's writeable room. Wait for room if we can, otherwise
1486  * return an appropriate error.
1487  */
1488 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1489 {
1490         int ret;
1491
1492         /*
1493          * Check pipe occupancy without the inode lock first. This function
1494          * is speculative anyways, so missing one is ok.
1495          */
1496         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1497                 return 0;
1498
1499         ret = 0;
1500         pipe_lock(pipe);
1501
1502         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1503                 if (!pipe->readers) {
1504                         send_sig(SIGPIPE, current, 0);
1505                         ret = -EPIPE;
1506                         break;
1507                 }
1508                 if (flags & SPLICE_F_NONBLOCK) {
1509                         ret = -EAGAIN;
1510                         break;
1511                 }
1512                 if (signal_pending(current)) {
1513                         ret = -ERESTARTSYS;
1514                         break;
1515                 }
1516                 pipe_wait_writable(pipe);
1517         }
1518
1519         pipe_unlock(pipe);
1520         return ret;
1521 }
1522
1523 /*
1524  * Splice contents of ipipe to opipe.
1525  */
1526 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1527                                struct pipe_inode_info *opipe,
1528                                size_t len, unsigned int flags)
1529 {
1530         struct pipe_buffer *ibuf, *obuf;
1531         unsigned int i_head, o_head;
1532         unsigned int i_tail, o_tail;
1533         unsigned int i_mask, o_mask;
1534         int ret = 0;
1535         bool input_wakeup = false;
1536
1537
1538 retry:
1539         ret = ipipe_prep(ipipe, flags);
1540         if (ret)
1541                 return ret;
1542
1543         ret = opipe_prep(opipe, flags);
1544         if (ret)
1545                 return ret;
1546
1547         /*
1548          * Potential ABBA deadlock, work around it by ordering lock
1549          * grabbing by pipe info address. Otherwise two different processes
1550          * could deadlock (one doing tee from A -> B, the other from B -> A).
1551          */
1552         pipe_double_lock(ipipe, opipe);
1553
1554         i_tail = ipipe->tail;
1555         i_mask = ipipe->ring_size - 1;
1556         o_head = opipe->head;
1557         o_mask = opipe->ring_size - 1;
1558
1559         do {
1560                 size_t o_len;
1561
1562                 if (!opipe->readers) {
1563                         send_sig(SIGPIPE, current, 0);
1564                         if (!ret)
1565                                 ret = -EPIPE;
1566                         break;
1567                 }
1568
1569                 i_head = ipipe->head;
1570                 o_tail = opipe->tail;
1571
1572                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1573                         break;
1574
1575                 /*
1576                  * Cannot make any progress, because either the input
1577                  * pipe is empty or the output pipe is full.
1578                  */
1579                 if (pipe_empty(i_head, i_tail) ||
1580                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1581                         /* Already processed some buffers, break */
1582                         if (ret)
1583                                 break;
1584
1585                         if (flags & SPLICE_F_NONBLOCK) {
1586                                 ret = -EAGAIN;
1587                                 break;
1588                         }
1589
1590                         /*
1591                          * We raced with another reader/writer and haven't
1592                          * managed to process any buffers.  A zero return
1593                          * value means EOF, so retry instead.
1594                          */
1595                         pipe_unlock(ipipe);
1596                         pipe_unlock(opipe);
1597                         goto retry;
1598                 }
1599
1600                 ibuf = &ipipe->bufs[i_tail & i_mask];
1601                 obuf = &opipe->bufs[o_head & o_mask];
1602
1603                 if (len >= ibuf->len) {
1604                         /*
1605                          * Simply move the whole buffer from ipipe to opipe
1606                          */
1607                         *obuf = *ibuf;
1608                         ibuf->ops = NULL;
1609                         i_tail++;
1610                         ipipe->tail = i_tail;
1611                         input_wakeup = true;
1612                         o_len = obuf->len;
1613                         o_head++;
1614                         opipe->head = o_head;
1615                 } else {
1616                         /*
1617                          * Get a reference to this pipe buffer,
1618                          * so we can copy the contents over.
1619                          */
1620                         if (!pipe_buf_get(ipipe, ibuf)) {
1621                                 if (ret == 0)
1622                                         ret = -EFAULT;
1623                                 break;
1624                         }
1625                         *obuf = *ibuf;
1626
1627                         /*
1628                          * Don't inherit the gift and merge flags, we need to
1629                          * prevent multiple steals of this page.
1630                          */
1631                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1632                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1633
1634                         obuf->len = len;
1635                         ibuf->offset += len;
1636                         ibuf->len -= len;
1637                         o_len = len;
1638                         o_head++;
1639                         opipe->head = o_head;
1640                 }
1641                 ret += o_len;
1642                 len -= o_len;
1643         } while (len);
1644
1645         pipe_unlock(ipipe);
1646         pipe_unlock(opipe);
1647
1648         /*
1649          * If we put data in the output pipe, wakeup any potential readers.
1650          */
1651         if (ret > 0)
1652                 wakeup_pipe_readers(opipe);
1653
1654         if (input_wakeup)
1655                 wakeup_pipe_writers(ipipe);
1656
1657         return ret;
1658 }
1659
1660 /*
1661  * Link contents of ipipe to opipe.
1662  */
1663 static int link_pipe(struct pipe_inode_info *ipipe,
1664                      struct pipe_inode_info *opipe,
1665                      size_t len, unsigned int flags)
1666 {
1667         struct pipe_buffer *ibuf, *obuf;
1668         unsigned int i_head, o_head;
1669         unsigned int i_tail, o_tail;
1670         unsigned int i_mask, o_mask;
1671         int ret = 0;
1672
1673         /*
1674          * Potential ABBA deadlock, work around it by ordering lock
1675          * grabbing by pipe info address. Otherwise two different processes
1676          * could deadlock (one doing tee from A -> B, the other from B -> A).
1677          */
1678         pipe_double_lock(ipipe, opipe);
1679
1680         i_tail = ipipe->tail;
1681         i_mask = ipipe->ring_size - 1;
1682         o_head = opipe->head;
1683         o_mask = opipe->ring_size - 1;
1684
1685         do {
1686                 if (!opipe->readers) {
1687                         send_sig(SIGPIPE, current, 0);
1688                         if (!ret)
1689                                 ret = -EPIPE;
1690                         break;
1691                 }
1692
1693                 i_head = ipipe->head;
1694                 o_tail = opipe->tail;
1695
1696                 /*
1697                  * If we have iterated all input buffers or run out of
1698                  * output room, break.
1699                  */
1700                 if (pipe_empty(i_head, i_tail) ||
1701                     pipe_full(o_head, o_tail, opipe->max_usage))
1702                         break;
1703
1704                 ibuf = &ipipe->bufs[i_tail & i_mask];
1705                 obuf = &opipe->bufs[o_head & o_mask];
1706
1707                 /*
1708                  * Get a reference to this pipe buffer,
1709                  * so we can copy the contents over.
1710                  */
1711                 if (!pipe_buf_get(ipipe, ibuf)) {
1712                         if (ret == 0)
1713                                 ret = -EFAULT;
1714                         break;
1715                 }
1716
1717                 *obuf = *ibuf;
1718
1719                 /*
1720                  * Don't inherit the gift and merge flag, we need to prevent
1721                  * multiple steals of this page.
1722                  */
1723                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1724                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1725
1726                 if (obuf->len > len)
1727                         obuf->len = len;
1728                 ret += obuf->len;
1729                 len -= obuf->len;
1730
1731                 o_head++;
1732                 opipe->head = o_head;
1733                 i_tail++;
1734         } while (len);
1735
1736         pipe_unlock(ipipe);
1737         pipe_unlock(opipe);
1738
1739         /*
1740          * If we put data in the output pipe, wakeup any potential readers.
1741          */
1742         if (ret > 0)
1743                 wakeup_pipe_readers(opipe);
1744
1745         return ret;
1746 }
1747
1748 /*
1749  * This is a tee(1) implementation that works on pipes. It doesn't copy
1750  * any data, it simply references the 'in' pages on the 'out' pipe.
1751  * The 'flags' used are the SPLICE_F_* variants, currently the only
1752  * applicable one is SPLICE_F_NONBLOCK.
1753  */
1754 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1755 {
1756         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1757         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1758         int ret = -EINVAL;
1759
1760         if (unlikely(!(in->f_mode & FMODE_READ) ||
1761                      !(out->f_mode & FMODE_WRITE)))
1762                 return -EBADF;
1763
1764         /*
1765          * Duplicate the contents of ipipe to opipe without actually
1766          * copying the data.
1767          */
1768         if (ipipe && opipe && ipipe != opipe) {
1769                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1770                         flags |= SPLICE_F_NONBLOCK;
1771
1772                 /*
1773                  * Keep going, unless we encounter an error. The ipipe/opipe
1774                  * ordering doesn't really matter.
1775                  */
1776                 ret = ipipe_prep(ipipe, flags);
1777                 if (!ret) {
1778                         ret = opipe_prep(opipe, flags);
1779                         if (!ret)
1780                                 ret = link_pipe(ipipe, opipe, len, flags);
1781                 }
1782         }
1783
1784         return ret;
1785 }
1786
1787 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1788 {
1789         struct fd in, out;
1790         int error;
1791
1792         if (unlikely(flags & ~SPLICE_F_ALL))
1793                 return -EINVAL;
1794
1795         if (unlikely(!len))
1796                 return 0;
1797
1798         error = -EBADF;
1799         in = fdget(fdin);
1800         if (in.file) {
1801                 out = fdget(fdout);
1802                 if (out.file) {
1803                         error = do_tee(in.file, out.file, len, flags);
1804                         fdput(out);
1805                 }
1806                 fdput(in);
1807         }
1808
1809         return error;
1810 }