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