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