fs/splice.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * "splice": joining two ropes together by interweaving their strands.
   4  *
   5  * This is the "extended pipe" functionality, where a pipe is used as
   6  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   7  * buffer that you can use to transfer data from one end to the other.
   8  *
   9  * The traditional unix read/write is extended with a "splice()" operation
  10  * that transfers data buffers to or from a pipe buffer.
  11  *
  12  * Named by Larry McVoy, original implementation from Linus, extended by
  13  * Jens to support splicing to files, network, direct splicing, etc and
  14  * fixing lots of bugs.
  15  *
  16  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  17  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  18  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  19  *
  20  */
  21 #include <linux/bvec.h>
  22 #include <linux/fs.h>
  23 #include <linux/file.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/splice.h>
  26 #include <linux/memcontrol.h>
  27 #include <linux/mm_inline.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/export.h>
  31 #include <linux/syscalls.h>
  32 #include <linux/uio.h>
  33 #include <linux/fsnotify.h>
  34 #include <linux/security.h>
  35 #include <linux/gfp.h>
  36 #include <linux/socket.h>
  37 #include <linux/sched/signal.h>
  38
  39 #include "internal.h"
  40
  41 /*
  42  * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
  43  * indicate they support non-blocking reads or writes, we must clear it
  44  * here if set to avoid blocking other users of this pipe if splice is
  45  * being done on it.
  46  */
  47 static noinline void noinline pipe_clear_nowait(struct file *file)
  48 {
  49         fmode_t fmode = READ_ONCE(file->f_mode);
  50
  51         do {
  52                 if (!(fmode & FMODE_NOWAIT))
  53                         break;
  54         } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
  55 }
  56
  57 /*
  58  * Attempt to steal a page from a pipe buffer. This should perhaps go into
  59  * a vm helper function, it's already simplified quite a bit by the
  60  * addition of remove_mapping(). If success is returned, the caller may
  61  * attempt to reuse this page for another destination.
  62  */
  63 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
  64                 struct pipe_buffer *buf)
  65 {
  66         struct folio *folio = page_folio(buf->page);
  67         struct address_space *mapping;
  68
  69         folio_lock(folio);
  70
  71         mapping = folio_mapping(folio);
  72         if (mapping) {
  73                 WARN_ON(!folio_test_uptodate(folio));
  74
  75                 /*
  76                  * At least for ext2 with nobh option, we need to wait on
  77                  * writeback completing on this folio, since we'll remove it
  78                  * from the pagecache.  Otherwise truncate wont wait on the
  79                  * folio, allowing the disk blocks to be reused by someone else
  80                  * before we actually wrote our data to them. fs corruption
  81                  * ensues.
  82                  */
  83                 folio_wait_writeback(folio);
  84
  85                 if (folio_has_private(folio) &&
  86                     !filemap_release_folio(folio, GFP_KERNEL))
  87                         goto out_unlock;
  88
  89                 /*
  90                  * If we succeeded in removing the mapping, set LRU flag
  91                  * and return good.
  92                  */
  93                 if (remove_mapping(mapping, folio)) {
  94                         buf->flags |= PIPE_BUF_FLAG_LRU;
  95                         return true;
  96                 }
  97         }
  98
  99         /*
 100          * Raced with truncate or failed to remove folio from current
 101          * address space, unlock and return failure.
 102          */
 103 out_unlock:
 104         folio_unlock(folio);
 105         return false;
 106 }
 107
 108 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
 109                                         struct pipe_buffer *buf)
 110 {
 111         put_page(buf->page);
 112         buf->flags &= ~PIPE_BUF_FLAG_LRU;
 113 }
 114
 115 /*
 116  * Check whether the contents of buf is OK to access. Since the content
 117  * is a page cache page, IO may be in flight.
 118  */
 119 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
 120                                        struct pipe_buffer *buf)
 121 {
 122         struct page *page = buf->page;
 123         int err;
 124
 125         if (!PageUptodate(page)) {
 126                 lock_page(page);
 127
 128                 /*
 129                  * Page got truncated/unhashed. This will cause a 0-byte
 130                  * splice, if this is the first page.
 131                  */
 132                 if (!page->mapping) {
 133                         err = -ENODATA;
 134                         goto error;
 135                 }
 136
 137                 /*
 138                  * Uh oh, read-error from disk.
 139                  */
 140                 if (!PageUptodate(page)) {
 141                         err = -EIO;
 142                         goto error;
 143                 }
 144
 145                 /*
 146                  * Page is ok afterall, we are done.
 147                  */
 148                 unlock_page(page);
 149         }
 150
 151         return 0;
 152 error:
 153         unlock_page(page);
 154         return err;
 155 }
 156
 157 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 158         .confirm        = page_cache_pipe_buf_confirm,
 159         .release        = page_cache_pipe_buf_release,
 160         .try_steal      = page_cache_pipe_buf_try_steal,
 161         .get            = generic_pipe_buf_get,
 162 };
 163
 164 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 165                 struct pipe_buffer *buf)
 166 {
 167         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 168                 return false;
 169
 170         buf->flags |= PIPE_BUF_FLAG_LRU;
 171         return generic_pipe_buf_try_steal(pipe, buf);
 172 }
 173
 174 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 175         .release        = page_cache_pipe_buf_release,
 176         .try_steal      = user_page_pipe_buf_try_steal,
 177         .get            = generic_pipe_buf_get,
 178 };
 179
 180 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 181 {
 182         smp_mb();
 183         if (waitqueue_active(&pipe->rd_wait))
 184                 wake_up_interruptible(&pipe->rd_wait);
 185         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 186 }
 187
 188 /**
 189  * splice_to_pipe - fill passed data into a pipe
 190  * @pipe:       pipe to fill
 191  * @spd:        data to fill
 192  *
 193  * Description:
 194  *    @spd contains a map of pages and len/offset tuples, along with
 195  *    the struct pipe_buf_operations associated with these pages. This
 196  *    function will link that data to the pipe.
 197  *
 198  */
 199 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 200                        struct splice_pipe_desc *spd)
 201 {
 202         unsigned int spd_pages = spd->nr_pages;
 203         unsigned int tail = pipe->tail;
 204         unsigned int head = pipe->head;
 205         unsigned int mask = pipe->ring_size - 1;
 206         int ret = 0, page_nr = 0;
 207
 208         if (!spd_pages)
 209                 return 0;
 210
 211         if (unlikely(!pipe->readers)) {
 212                 send_sig(SIGPIPE, current, 0);
 213                 ret = -EPIPE;
 214                 goto out;
 215         }
 216
 217         while (!pipe_full(head, tail, pipe->max_usage)) {
 218                 struct pipe_buffer *buf = &pipe->bufs[head & mask];
 219
 220                 buf->page = spd->pages[page_nr];
 221                 buf->offset = spd->partial[page_nr].offset;
 222                 buf->len = spd->partial[page_nr].len;
 223                 buf->private = spd->partial[page_nr].private;
 224                 buf->ops = spd->ops;
 225                 buf->flags = 0;
 226
 227                 head++;
 228                 pipe->head = head;
 229                 page_nr++;
 230                 ret += buf->len;
 231
 232                 if (!--spd->nr_pages)
 233                         break;
 234         }
 235
 236         if (!ret)
 237                 ret = -EAGAIN;
 238
 239 out:
 240         while (page_nr < spd_pages)
 241                 spd->spd_release(spd, page_nr++);
 242
 243         return ret;
 244 }
 245 EXPORT_SYMBOL_GPL(splice_to_pipe);
 246
 247 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 248 {
 249         unsigned int head = pipe->head;
 250         unsigned int tail = pipe->tail;
 251         unsigned int mask = pipe->ring_size - 1;
 252         int ret;
 253
 254         if (unlikely(!pipe->readers)) {
 255                 send_sig(SIGPIPE, current, 0);
 256                 ret = -EPIPE;
 257         } else if (pipe_full(head, tail, pipe->max_usage)) {
 258                 ret = -EAGAIN;
 259         } else {
 260                 pipe->bufs[head & mask] = *buf;
 261                 pipe->head = head + 1;
 262                 return buf->len;
 263         }
 264         pipe_buf_release(pipe, buf);
 265         return ret;
 266 }
 267 EXPORT_SYMBOL(add_to_pipe);
 268
 269 /*
 270  * Check if we need to grow the arrays holding pages and partial page
 271  * descriptions.
 272  */
 273 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 274 {
 275         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 276
 277         spd->nr_pages_max = max_usage;
 278         if (max_usage <= PIPE_DEF_BUFFERS)
 279                 return 0;
 280
 281         spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
 282         spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
 283                                      GFP_KERNEL);
 284
 285         if (spd->pages && spd->partial)
 286                 return 0;
 287
 288         kfree(spd->pages);
 289         kfree(spd->partial);
 290         return -ENOMEM;
 291 }
 292
 293 void splice_shrink_spd(struct splice_pipe_desc *spd)
 294 {
 295         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 296                 return;
 297
 298         kfree(spd->pages);
 299         kfree(spd->partial);
 300 }
 301
 302 /*
 303  * Copy data from a file into pages and then splice those into the output pipe.
 304  */
 305 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
 306                          struct pipe_inode_info *pipe,
 307                          size_t len, unsigned int flags)
 308 {
 309         struct iov_iter to;
 310         struct bio_vec *bv;
 311         struct kiocb kiocb;
 312         struct page **pages;
 313         ssize_t ret;
 314         size_t used, npages, chunk, remain, keep = 0;
 315         int i;
 316
 317         /* Work out how much data we can actually add into the pipe */
 318         used = pipe_occupancy(pipe->head, pipe->tail);
 319         npages = max_t(ssize_t, pipe->max_usage - used, 0);
 320         len = min_t(size_t, len, npages * PAGE_SIZE);
 321         npages = DIV_ROUND_UP(len, PAGE_SIZE);
 322
 323         bv = kzalloc(array_size(npages, sizeof(bv[0])) +
 324                      array_size(npages, sizeof(struct page *)), GFP_KERNEL);
 325         if (!bv)
 326                 return -ENOMEM;
 327
 328         pages = (struct page **)(bv + npages);
 329         npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
 330         if (!npages) {
 331                 kfree(bv);
 332                 return -ENOMEM;
 333         }
 334
 335         remain = len = min_t(size_t, len, npages * PAGE_SIZE);
 336
 337         for (i = 0; i < npages; i++) {
 338                 chunk = min_t(size_t, PAGE_SIZE, remain);
 339                 bv[i].bv_page = pages[i];
 340                 bv[i].bv_offset = 0;
 341                 bv[i].bv_len = chunk;
 342                 remain -= chunk;
 343         }
 344
 345         /* Do the I/O */
 346         iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
 347         init_sync_kiocb(&kiocb, in);
 348         kiocb.ki_pos = *ppos;
 349         ret = call_read_iter(in, &kiocb, &to);
 350
 351         if (ret > 0) {
 352                 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
 353                 *ppos = kiocb.ki_pos;
 354                 file_accessed(in);
 355         } else if (ret < 0) {
 356                 /*
 357                  * callers of ->splice_read() expect -EAGAIN on
 358                  * "can't put anything in there", rather than -EFAULT.
 359                  */
 360                 if (ret == -EFAULT)
 361                         ret = -EAGAIN;
 362         }
 363
 364         /* Free any pages that didn't get touched at all. */
 365         if (keep < npages)
 366                 release_pages(pages + keep, npages - keep);
 367
 368         /* Push the remaining pages into the pipe. */
 369         remain = ret;
 370         for (i = 0; i < keep; i++) {
 371                 struct pipe_buffer *buf = pipe_head_buf(pipe);
 372
 373                 chunk = min_t(size_t, remain, PAGE_SIZE);
 374                 *buf = (struct pipe_buffer) {
 375                         .ops    = &default_pipe_buf_ops,
 376                         .page   = bv[i].bv_page,
 377                         .offset = 0,
 378                         .len    = chunk,
 379                 };
 380                 pipe->head++;
 381                 remain -= chunk;
 382         }
 383
 384         kfree(bv);
 385         return ret;
 386 }
 387 EXPORT_SYMBOL(copy_splice_read);
 388
 389 /**
 390  * generic_file_splice_read - splice data from file to a pipe
 391  * @in:         file to splice from
 392  * @ppos:       position in @in
 393  * @pipe:       pipe to splice to
 394  * @len:        number of bytes to splice
 395  * @flags:      splice modifier flags
 396  *
 397  * Description:
 398  *    Will read pages from given file and fill them into a pipe. Can be
 399  *    used as long as it has more or less sane ->read_iter().
 400  *
 401  */
 402 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 403                                  struct pipe_inode_info *pipe, size_t len,
 404                                  unsigned int flags)
 405 {
 406         struct iov_iter to;
 407         struct kiocb kiocb;
 408         int ret;
 409
 410         iov_iter_pipe(&to, ITER_DEST, pipe, len);
 411         init_sync_kiocb(&kiocb, in);
 412         kiocb.ki_pos = *ppos;
 413         ret = call_read_iter(in, &kiocb, &to);
 414         if (ret > 0) {
 415                 *ppos = kiocb.ki_pos;
 416                 file_accessed(in);
 417         } else if (ret < 0) {
 418                 /* free what was emitted */
 419                 pipe_discard_from(pipe, to.start_head);
 420                 /*
 421                  * callers of ->splice_read() expect -EAGAIN on
 422                  * "can't put anything in there", rather than -EFAULT.
 423                  */
 424                 if (ret == -EFAULT)
 425                         ret = -EAGAIN;
 426         }
 427
 428         return ret;
 429 }
 430 EXPORT_SYMBOL(generic_file_splice_read);
 431
 432 const struct pipe_buf_operations default_pipe_buf_ops = {
 433         .release        = generic_pipe_buf_release,
 434         .try_steal      = generic_pipe_buf_try_steal,
 435         .get            = generic_pipe_buf_get,
 436 };
 437
 438 /* Pipe buffer operations for a socket and similar. */
 439 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
 440         .release        = generic_pipe_buf_release,
 441         .get            = generic_pipe_buf_get,
 442 };
 443 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
 444
 445 /*
 446  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 447  * using sendpage(). Return the number of bytes sent.
 448  */
 449 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 450                             struct pipe_buffer *buf, struct splice_desc *sd)
 451 {
 452         struct file *file = sd->u.file;
 453         loff_t pos = sd->pos;
 454         int more;
 455
 456         if (!likely(file->f_op->sendpage))
 457                 return -EINVAL;
 458
 459         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 460
 461         if (sd->len < sd->total_len &&
 462             pipe_occupancy(pipe->head, pipe->tail) > 1)
 463                 more |= MSG_SENDPAGE_NOTLAST;
 464
 465         return file->f_op->sendpage(file, buf->page, buf->offset,
 466                                     sd->len, &pos, more);
 467 }
 468
 469 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 470 {
 471         smp_mb();
 472         if (waitqueue_active(&pipe->wr_wait))
 473                 wake_up_interruptible(&pipe->wr_wait);
 474         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 475 }
 476
 477 /**
 478  * splice_from_pipe_feed - feed available data from a pipe to a file
 479  * @pipe:       pipe to splice from
 480  * @sd:         information to @actor
 481  * @actor:      handler that splices the data
 482  *
 483  * Description:
 484  *    This function loops over the pipe and calls @actor to do the
 485  *    actual moving of a single struct pipe_buffer to the desired
 486  *    destination.  It returns when there's no more buffers left in
 487  *    the pipe or if the requested number of bytes (@sd->total_len)
 488  *    have been copied.  It returns a positive number (one) if the
 489  *    pipe needs to be filled with more data, zero if the required
 490  *    number of bytes have been copied and -errno on error.
 491  *
 492  *    This, together with splice_from_pipe_{begin,end,next}, may be
 493  *    used to implement the functionality of __splice_from_pipe() when
 494  *    locking is required around copying the pipe buffers to the
 495  *    destination.
 496  */
 497 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 498                           splice_actor *actor)
 499 {
 500         unsigned int head = pipe->head;
 501         unsigned int tail = pipe->tail;
 502         unsigned int mask = pipe->ring_size - 1;
 503         int ret;
 504
 505         while (!pipe_empty(head, tail)) {
 506                 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 507
 508                 sd->len = buf->len;
 509                 if (sd->len > sd->total_len)
 510                         sd->len = sd->total_len;
 511
 512                 ret = pipe_buf_confirm(pipe, buf);
 513                 if (unlikely(ret)) {
 514                         if (ret == -ENODATA)
 515                                 ret = 0;
 516                         return ret;
 517                 }
 518
 519                 ret = actor(pipe, buf, sd);
 520                 if (ret <= 0)
 521                         return ret;
 522
 523                 buf->offset += ret;
 524                 buf->len -= ret;
 525
 526                 sd->num_spliced += ret;
 527                 sd->len -= ret;
 528                 sd->pos += ret;
 529                 sd->total_len -= ret;
 530
 531                 if (!buf->len) {
 532                         pipe_buf_release(pipe, buf);
 533                         tail++;
 534                         pipe->tail = tail;
 535                         if (pipe->files)
 536                                 sd->need_wakeup = true;
 537                 }
 538
 539                 if (!sd->total_len)
 540                         return 0;
 541         }
 542
 543         return 1;
 544 }
 545
 546 /* We know we have a pipe buffer, but maybe it's empty? */
 547 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
 548 {
 549         unsigned int tail = pipe->tail;
 550         unsigned int mask = pipe->ring_size - 1;
 551         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 552
 553         if (unlikely(!buf->len)) {
 554                 pipe_buf_release(pipe, buf);
 555                 pipe->tail = tail+1;
 556                 return true;
 557         }
 558
 559         return false;
 560 }
 561
 562 /**
 563  * splice_from_pipe_next - wait for some data to splice from
 564  * @pipe:       pipe to splice from
 565  * @sd:         information about the splice operation
 566  *
 567  * Description:
 568  *    This function will wait for some data and return a positive
 569  *    value (one) if pipe buffers are available.  It will return zero
 570  *    or -errno if no more data needs to be spliced.
 571  */
 572 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 573 {
 574         /*
 575          * Check for signal early to make process killable when there are
 576          * always buffers available
 577          */
 578         if (signal_pending(current))
 579                 return -ERESTARTSYS;
 580
 581 repeat:
 582         while (pipe_empty(pipe->head, pipe->tail)) {
 583                 if (!pipe->writers)
 584                         return 0;
 585
 586                 if (sd->num_spliced)
 587                         return 0;
 588
 589                 if (sd->flags & SPLICE_F_NONBLOCK)
 590                         return -EAGAIN;
 591
 592                 if (signal_pending(current))
 593                         return -ERESTARTSYS;
 594
 595                 if (sd->need_wakeup) {
 596                         wakeup_pipe_writers(pipe);
 597                         sd->need_wakeup = false;
 598                 }
 599
 600                 pipe_wait_readable(pipe);
 601         }
 602
 603         if (eat_empty_buffer(pipe))
 604                 goto repeat;
 605
 606         return 1;
 607 }
 608
 609 /**
 610  * splice_from_pipe_begin - start splicing from pipe
 611  * @sd:         information about the splice operation
 612  *
 613  * Description:
 614  *    This function should be called before a loop containing
 615  *    splice_from_pipe_next() and splice_from_pipe_feed() to
 616  *    initialize the necessary fields of @sd.
 617  */
 618 static void splice_from_pipe_begin(struct splice_desc *sd)
 619 {
 620         sd->num_spliced = 0;
 621         sd->need_wakeup = false;
 622 }
 623
 624 /**
 625  * splice_from_pipe_end - finish splicing from pipe
 626  * @pipe:       pipe to splice from
 627  * @sd:         information about the splice operation
 628  *
 629  * Description:
 630  *    This function will wake up pipe writers if necessary.  It should
 631  *    be called after a loop containing splice_from_pipe_next() and
 632  *    splice_from_pipe_feed().
 633  */
 634 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 635 {
 636         if (sd->need_wakeup)
 637                 wakeup_pipe_writers(pipe);
 638 }
 639
 640 /**
 641  * __splice_from_pipe - splice data from a pipe to given actor
 642  * @pipe:       pipe to splice from
 643  * @sd:         information to @actor
 644  * @actor:      handler that splices the data
 645  *
 646  * Description:
 647  *    This function does little more than loop over the pipe and call
 648  *    @actor to do the actual moving of a single struct pipe_buffer to
 649  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 650  *    pipe_to_user.
 651  *
 652  */
 653 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 654                            splice_actor *actor)
 655 {
 656         int ret;
 657
 658         splice_from_pipe_begin(sd);
 659         do {
 660                 cond_resched();
 661                 ret = splice_from_pipe_next(pipe, sd);
 662                 if (ret > 0)
 663                         ret = splice_from_pipe_feed(pipe, sd, actor);
 664         } while (ret > 0);
 665         splice_from_pipe_end(pipe, sd);
 666
 667         return sd->num_spliced ? sd->num_spliced : ret;
 668 }
 669 EXPORT_SYMBOL(__splice_from_pipe);
 670
 671 /**
 672  * splice_from_pipe - splice data from a pipe to a file
 673  * @pipe:       pipe to splice from
 674  * @out:        file to splice to
 675  * @ppos:       position in @out
 676  * @len:        how many bytes to splice
 677  * @flags:      splice modifier flags
 678  * @actor:      handler that splices the data
 679  *
 680  * Description:
 681  *    See __splice_from_pipe. This function locks the pipe inode,
 682  *    otherwise it's identical to __splice_from_pipe().
 683  *
 684  */
 685 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 686                          loff_t *ppos, size_t len, unsigned int flags,
 687                          splice_actor *actor)
 688 {
 689         ssize_t ret;
 690         struct splice_desc sd = {
 691                 .total_len = len,
 692                 .flags = flags,
 693                 .pos = *ppos,
 694                 .u.file = out,
 695         };
 696
 697         pipe_lock(pipe);
 698         ret = __splice_from_pipe(pipe, &sd, actor);
 699         pipe_unlock(pipe);
 700
 701         return ret;
 702 }
 703
 704 /**
 705  * iter_file_splice_write - splice data from a pipe to a file
 706  * @pipe:       pipe info
 707  * @out:        file to write to
 708  * @ppos:       position in @out
 709  * @len:        number of bytes to splice
 710  * @flags:      splice modifier flags
 711  *
 712  * Description:
 713  *    Will either move or copy pages (determined by @flags options) from
 714  *    the given pipe inode to the given file.
 715  *    This one is ->write_iter-based.
 716  *
 717  */
 718 ssize_t
 719 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 720                           loff_t *ppos, size_t len, unsigned int flags)
 721 {
 722         struct splice_desc sd = {
 723                 .total_len = len,
 724                 .flags = flags,
 725                 .pos = *ppos,
 726                 .u.file = out,
 727         };
 728         int nbufs = pipe->max_usage;
 729         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
 730                                         GFP_KERNEL);
 731         ssize_t ret;
 732
 733         if (unlikely(!array))
 734                 return -ENOMEM;
 735
 736         pipe_lock(pipe);
 737
 738         splice_from_pipe_begin(&sd);
 739         while (sd.total_len) {
 740                 struct iov_iter from;
 741                 unsigned int head, tail, mask;
 742                 size_t left;
 743                 int n;
 744
 745                 ret = splice_from_pipe_next(pipe, &sd);
 746                 if (ret <= 0)
 747                         break;
 748
 749                 if (unlikely(nbufs < pipe->max_usage)) {
 750                         kfree(array);
 751                         nbufs = pipe->max_usage;
 752                         array = kcalloc(nbufs, sizeof(struct bio_vec),
 753                                         GFP_KERNEL);
 754                         if (!array) {
 755                                 ret = -ENOMEM;
 756                                 break;
 757                         }
 758                 }
 759
 760                 head = pipe->head;
 761                 tail = pipe->tail;
 762                 mask = pipe->ring_size - 1;
 763
 764                 /* build the vector */
 765                 left = sd.total_len;
 766                 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
 767                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 768                         size_t this_len = buf->len;
 769
 770                         /* zero-length bvecs are not supported, skip them */
 771                         if (!this_len)
 772                                 continue;
 773                         this_len = min(this_len, left);
 774
 775                         ret = pipe_buf_confirm(pipe, buf);
 776                         if (unlikely(ret)) {
 777                                 if (ret == -ENODATA)
 778                                         ret = 0;
 779                                 goto done;
 780                         }
 781
 782                         bvec_set_page(&array[n], buf->page, this_len,
 783                                       buf->offset);
 784                         left -= this_len;
 785                         n++;
 786                 }
 787
 788                 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
 789                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
 790                 if (ret <= 0)
 791                         break;
 792
 793                 sd.num_spliced += ret;
 794                 sd.total_len -= ret;
 795                 *ppos = sd.pos;
 796
 797                 /* dismiss the fully eaten buffers, adjust the partial one */
 798                 tail = pipe->tail;
 799                 while (ret) {
 800                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 801                         if (ret >= buf->len) {
 802                                 ret -= buf->len;
 803                                 buf->len = 0;
 804                                 pipe_buf_release(pipe, buf);
 805                                 tail++;
 806                                 pipe->tail = tail;
 807                                 if (pipe->files)
 808                                         sd.need_wakeup = true;
 809                         } else {
 810                                 buf->offset += ret;
 811                                 buf->len -= ret;
 812                                 ret = 0;
 813                         }
 814                 }
 815         }
 816 done:
 817         kfree(array);
 818         splice_from_pipe_end(pipe, &sd);
 819
 820         pipe_unlock(pipe);
 821
 822         if (sd.num_spliced)
 823                 ret = sd.num_spliced;
 824
 825         return ret;
 826 }
 827
 828 EXPORT_SYMBOL(iter_file_splice_write);
 829
 830 /**
 831  * generic_splice_sendpage - splice data from a pipe to a socket
 832  * @pipe:       pipe to splice from
 833  * @out:        socket to write to
 834  * @ppos:       position in @out
 835  * @len:        number of bytes to splice
 836  * @flags:      splice modifier flags
 837  *
 838  * Description:
 839  *    Will send @len bytes from the pipe to a network socket. No data copying
 840  *    is involved.
 841  *
 842  */
 843 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
 844                                 loff_t *ppos, size_t len, unsigned int flags)
 845 {
 846         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
 847 }
 848
 849 EXPORT_SYMBOL(generic_splice_sendpage);
 850
 851 static int warn_unsupported(struct file *file, const char *op)
 852 {
 853         pr_debug_ratelimited(
 854                 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
 855                 op, file, current->pid, current->comm);
 856         return -EINVAL;
 857 }
 858
 859 /*
 860  * Attempt to initiate a splice from pipe to file.
 861  */
 862 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
 863                            loff_t *ppos, size_t len, unsigned int flags)
 864 {
 865         if (unlikely(!out->f_op->splice_write))
 866                 return warn_unsupported(out, "write");
 867         return out->f_op->splice_write(pipe, out, ppos, len, flags);
 868 }
 869
 870 /**
 871  * vfs_splice_read - Read data from a file and splice it into a pipe
 872  * @in:         File to splice from
 873  * @ppos:       Input file offset
 874  * @pipe:       Pipe to splice to
 875  * @len:        Number of bytes to splice
 876  * @flags:      Splice modifier flags (SPLICE_F_*)
 877  *
 878  * Splice the requested amount of data from the input file to the pipe.  This
 879  * is synchronous as the caller must hold the pipe lock across the entire
 880  * operation.
 881  *
 882  * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
 883  * a hole and a negative error code otherwise.
 884  */
 885 long vfs_splice_read(struct file *in, loff_t *ppos,
 886                      struct pipe_inode_info *pipe, size_t len,
 887                      unsigned int flags)
 888 {
 889         unsigned int p_space;
 890         int ret;
 891
 892         if (unlikely(!(in->f_mode & FMODE_READ)))
 893                 return -EBADF;
 894         if (!len)
 895                 return 0;
 896
 897         /* Don't try to read more the pipe has space for. */
 898         p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
 899         len = min_t(size_t, len, p_space << PAGE_SHIFT);
 900
 901         ret = rw_verify_area(READ, in, ppos, len);
 902         if (unlikely(ret < 0))
 903                 return ret;
 904
 905         if (unlikely(len > MAX_RW_COUNT))
 906                 len = MAX_RW_COUNT;
 907
 908         if (unlikely(!in->f_op->splice_read))
 909                 return warn_unsupported(in, "read");
 910         /*
 911          * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
 912          * buffer, copy into it and splice that into the pipe.
 913          */
 914         if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
 915                 return copy_splice_read(in, ppos, pipe, len, flags);
 916         return in->f_op->splice_read(in, ppos, pipe, len, flags);
 917 }
 918 EXPORT_SYMBOL_GPL(vfs_splice_read);
 919
 920 /**
 921  * splice_direct_to_actor - splices data directly between two non-pipes
 922  * @in:         file to splice from
 923  * @sd:         actor information on where to splice to
 924  * @actor:      handles the data splicing
 925  *
 926  * Description:
 927  *    This is a special case helper to splice directly between two
 928  *    points, without requiring an explicit pipe. Internally an allocated
 929  *    pipe is cached in the process, and reused during the lifetime of
 930  *    that process.
 931  *
 932  */
 933 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 934                                splice_direct_actor *actor)
 935 {
 936         struct pipe_inode_info *pipe;
 937         long ret, bytes;
 938         size_t len;
 939         int i, flags, more;
 940
 941         /*
 942          * We require the input to be seekable, as we don't want to randomly
 943          * drop data for eg socket -> socket splicing. Use the piped splicing
 944          * for that!
 945          */
 946         if (unlikely(!(in->f_mode & FMODE_LSEEK)))
 947                 return -EINVAL;
 948
 949         /*
 950          * neither in nor out is a pipe, setup an internal pipe attached to
 951          * 'out' and transfer the wanted data from 'in' to 'out' through that
 952          */
 953         pipe = current->splice_pipe;
 954         if (unlikely(!pipe)) {
 955                 pipe = alloc_pipe_info();
 956                 if (!pipe)
 957                         return -ENOMEM;
 958
 959                 /*
 960                  * We don't have an immediate reader, but we'll read the stuff
 961                  * out of the pipe right after the splice_to_pipe(). So set
 962                  * PIPE_READERS appropriately.
 963                  */
 964                 pipe->readers = 1;
 965
 966                 current->splice_pipe = pipe;
 967         }
 968
 969         /*
 970          * Do the splice.
 971          */
 972         bytes = 0;
 973         len = sd->total_len;
 974         flags = sd->flags;
 975
 976         /*
 977          * Don't block on output, we have to drain the direct pipe.
 978          */
 979         sd->flags &= ~SPLICE_F_NONBLOCK;
 980         more = sd->flags & SPLICE_F_MORE;
 981
 982         WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
 983
 984         while (len) {
 985                 size_t read_len;
 986                 loff_t pos = sd->pos, prev_pos = pos;
 987
 988                 ret = vfs_splice_read(in, &pos, pipe, len, flags);
 989                 if (unlikely(ret <= 0))
 990                         goto out_release;
 991
 992                 read_len = ret;
 993                 sd->total_len = read_len;
 994
 995                 /*
 996                  * If more data is pending, set SPLICE_F_MORE
 997                  * If this is the last data and SPLICE_F_MORE was not set
 998                  * initially, clears it.
 999                  */
1000                 if (read_len < len)
1001                         sd->flags |= SPLICE_F_MORE;
1002                 else if (!more)
1003                         sd->flags &= ~SPLICE_F_MORE;
1004                 /*
1005                  * NOTE: nonblocking mode only applies to the input. We
1006                  * must not do the output in nonblocking mode as then we
1007                  * could get stuck data in the internal pipe:
1008                  */
1009                 ret = actor(pipe, sd);
1010                 if (unlikely(ret <= 0)) {
1011                         sd->pos = prev_pos;
1012                         goto out_release;
1013                 }
1014
1015                 bytes += ret;
1016                 len -= ret;
1017                 sd->pos = pos;
1018
1019                 if (ret < read_len) {
1020                         sd->pos = prev_pos + ret;
1021                         goto out_release;
1022                 }
1023         }
1024
1025 done:
1026         pipe->tail = pipe->head = 0;
1027         file_accessed(in);
1028         return bytes;
1029
1030 out_release:
1031         /*
1032          * If we did an incomplete transfer we must release
1033          * the pipe buffers in question:
1034          */
1035         for (i = 0; i < pipe->ring_size; i++) {
1036                 struct pipe_buffer *buf = &pipe->bufs[i];
1037
1038                 if (buf->ops)
1039                         pipe_buf_release(pipe, buf);
1040         }
1041
1042         if (!bytes)
1043                 bytes = ret;
1044
1045         goto done;
1046 }
1047 EXPORT_SYMBOL(splice_direct_to_actor);
1048
1049 static int direct_splice_actor(struct pipe_inode_info *pipe,
1050                                struct splice_desc *sd)
1051 {
1052         struct file *file = sd->u.file;
1053
1054         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1055                               sd->flags);
1056 }
1057
1058 /**
1059  * do_splice_direct - splices data directly between two files
1060  * @in:         file to splice from
1061  * @ppos:       input file offset
1062  * @out:        file to splice to
1063  * @opos:       output file offset
1064  * @len:        number of bytes to splice
1065  * @flags:      splice modifier flags
1066  *
1067  * Description:
1068  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1069  *    doing it in the application would incur an extra system call
1070  *    (splice in + splice out, as compared to just sendfile()). So this helper
1071  *    can splice directly through a process-private pipe.
1072  *
1073  */
1074 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1075                       loff_t *opos, size_t len, unsigned int flags)
1076 {
1077         struct splice_desc sd = {
1078                 .len            = len,
1079                 .total_len      = len,
1080                 .flags          = flags,
1081                 .pos            = *ppos,
1082                 .u.file         = out,
1083                 .opos           = opos,
1084         };
1085         long ret;
1086
1087         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1088                 return -EBADF;
1089
1090         if (unlikely(out->f_flags & O_APPEND))
1091                 return -EINVAL;
1092
1093         ret = rw_verify_area(WRITE, out, opos, len);
1094         if (unlikely(ret < 0))
1095                 return ret;
1096
1097         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1098         if (ret > 0)
1099                 *ppos = sd.pos;
1100
1101         return ret;
1102 }
1103 EXPORT_SYMBOL(do_splice_direct);
1104
1105 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1106 {
1107         for (;;) {
1108                 if (unlikely(!pipe->readers)) {
1109                         send_sig(SIGPIPE, current, 0);
1110                         return -EPIPE;
1111                 }
1112                 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1113                         return 0;
1114                 if (flags & SPLICE_F_NONBLOCK)
1115                         return -EAGAIN;
1116                 if (signal_pending(current))
1117                         return -ERESTARTSYS;
1118                 pipe_wait_writable(pipe);
1119         }
1120 }
1121
1122 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1123                                struct pipe_inode_info *opipe,
1124                                size_t len, unsigned int flags);
1125
1126 long splice_file_to_pipe(struct file *in,
1127                          struct pipe_inode_info *opipe,
1128                          loff_t *offset,
1129                          size_t len, unsigned int flags)
1130 {
1131         long ret;
1132
1133         pipe_lock(opipe);
1134         ret = wait_for_space(opipe, flags);
1135         if (!ret)
1136                 ret = vfs_splice_read(in, offset, opipe, len, flags);
1137         pipe_unlock(opipe);
1138         if (ret > 0)
1139                 wakeup_pipe_readers(opipe);
1140         return ret;
1141 }
1142
1143 /*
1144  * Determine where to splice to/from.
1145  */
1146 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1147                loff_t *off_out, size_t len, unsigned int flags)
1148 {
1149         struct pipe_inode_info *ipipe;
1150         struct pipe_inode_info *opipe;
1151         loff_t offset;
1152         long ret;
1153
1154         if (unlikely(!(in->f_mode & FMODE_READ) ||
1155                      !(out->f_mode & FMODE_WRITE)))
1156                 return -EBADF;
1157
1158         ipipe = get_pipe_info(in, true);
1159         opipe = get_pipe_info(out, true);
1160
1161         if (ipipe && opipe) {
1162                 if (off_in || off_out)
1163                         return -ESPIPE;
1164
1165                 /* Splicing to self would be fun, but... */
1166                 if (ipipe == opipe)
1167                         return -EINVAL;
1168
1169                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1170                         flags |= SPLICE_F_NONBLOCK;
1171
1172                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1173         }
1174
1175         if (ipipe) {
1176                 if (off_in)
1177                         return -ESPIPE;
1178                 if (off_out) {
1179                         if (!(out->f_mode & FMODE_PWRITE))
1180                                 return -EINVAL;
1181                         offset = *off_out;
1182                 } else {
1183                         offset = out->f_pos;
1184                 }
1185
1186                 if (unlikely(out->f_flags & O_APPEND))
1187                         return -EINVAL;
1188
1189                 ret = rw_verify_area(WRITE, out, &offset, len);
1190                 if (unlikely(ret < 0))
1191                         return ret;
1192
1193                 if (in->f_flags & O_NONBLOCK)
1194                         flags |= SPLICE_F_NONBLOCK;
1195
1196                 file_start_write(out);
1197                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1198                 file_end_write(out);
1199
1200                 if (ret > 0)
1201                         fsnotify_modify(out);
1202
1203                 if (!off_out)
1204                         out->f_pos = offset;
1205                 else
1206                         *off_out = offset;
1207
1208                 return ret;
1209         }
1210
1211         if (opipe) {
1212                 if (off_out)
1213                         return -ESPIPE;
1214                 if (off_in) {
1215                         if (!(in->f_mode & FMODE_PREAD))
1216                                 return -EINVAL;
1217                         offset = *off_in;
1218                 } else {
1219                         offset = in->f_pos;
1220                 }
1221
1222                 if (out->f_flags & O_NONBLOCK)
1223                         flags |= SPLICE_F_NONBLOCK;
1224
1225                 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1226
1227                 if (ret > 0)
1228                         fsnotify_access(in);
1229
1230                 if (!off_in)
1231                         in->f_pos = offset;
1232                 else
1233                         *off_in = offset;
1234
1235                 return ret;
1236         }
1237
1238         return -EINVAL;
1239 }
1240
1241 static long __do_splice(struct file *in, loff_t __user *off_in,
1242                         struct file *out, loff_t __user *off_out,
1243                         size_t len, unsigned int flags)
1244 {
1245         struct pipe_inode_info *ipipe;
1246         struct pipe_inode_info *opipe;
1247         loff_t offset, *__off_in = NULL, *__off_out = NULL;
1248         long ret;
1249
1250         ipipe = get_pipe_info(in, true);
1251         opipe = get_pipe_info(out, true);
1252
1253         if (ipipe) {
1254                 if (off_in)
1255                         return -ESPIPE;
1256                 pipe_clear_nowait(in);
1257         }
1258         if (opipe) {
1259                 if (off_out)
1260                         return -ESPIPE;
1261                 pipe_clear_nowait(out);
1262         }
1263
1264         if (off_out) {
1265                 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1266                         return -EFAULT;
1267                 __off_out = &offset;
1268         }
1269         if (off_in) {
1270                 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1271                         return -EFAULT;
1272                 __off_in = &offset;
1273         }
1274
1275         ret = do_splice(in, __off_in, out, __off_out, len, flags);
1276         if (ret < 0)
1277                 return ret;
1278
1279         if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1280                 return -EFAULT;
1281         if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1282                 return -EFAULT;
1283
1284         return ret;
1285 }
1286
1287 static int iter_to_pipe(struct iov_iter *from,
1288                         struct pipe_inode_info *pipe,
1289                         unsigned flags)
1290 {
1291         struct pipe_buffer buf = {
1292                 .ops = &user_page_pipe_buf_ops,
1293                 .flags = flags
1294         };
1295         size_t total = 0;
1296         int ret = 0;
1297
1298         while (iov_iter_count(from)) {
1299                 struct page *pages[16];
1300                 ssize_t left;
1301                 size_t start;
1302                 int i, n;
1303
1304                 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1305                 if (left <= 0) {
1306                         ret = left;
1307                         break;
1308                 }
1309
1310                 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1311                 for (i = 0; i < n; i++) {
1312                         int size = min_t(int, left, PAGE_SIZE - start);
1313
1314                         buf.page = pages[i];
1315                         buf.offset = start;
1316                         buf.len = size;
1317                         ret = add_to_pipe(pipe, &buf);
1318                         if (unlikely(ret < 0)) {
1319                                 iov_iter_revert(from, left);
1320                                 // this one got dropped by add_to_pipe()
1321                                 while (++i < n)
1322                                         put_page(pages[i]);
1323                                 goto out;
1324                         }
1325                         total += ret;
1326                         left -= size;
1327                         start = 0;
1328                 }
1329         }
1330 out:
1331         return total ? total : ret;
1332 }
1333
1334 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1335                         struct splice_desc *sd)
1336 {
1337         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1338         return n == sd->len ? n : -EFAULT;
1339 }
1340
1341 /*
1342  * For lack of a better implementation, implement vmsplice() to userspace
1343  * as a simple copy of the pipes pages to the user iov.
1344  */
1345 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1346                              unsigned int flags)
1347 {
1348         struct pipe_inode_info *pipe = get_pipe_info(file, true);
1349         struct splice_desc sd = {
1350                 .total_len = iov_iter_count(iter),
1351                 .flags = flags,
1352                 .u.data = iter
1353         };
1354         long ret = 0;
1355
1356         if (!pipe)
1357                 return -EBADF;
1358
1359         pipe_clear_nowait(file);
1360
1361         if (sd.total_len) {
1362                 pipe_lock(pipe);
1363                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1364                 pipe_unlock(pipe);
1365         }
1366
1367         return ret;
1368 }
1369
1370 /*
1371  * vmsplice splices a user address range into a pipe. It can be thought of
1372  * as splice-from-memory, where the regular splice is splice-from-file (or
1373  * to file). In both cases the output is a pipe, naturally.
1374  */
1375 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1376                              unsigned int flags)
1377 {
1378         struct pipe_inode_info *pipe;
1379         long ret = 0;
1380         unsigned buf_flag = 0;
1381
1382         if (flags & SPLICE_F_GIFT)
1383                 buf_flag = PIPE_BUF_FLAG_GIFT;
1384
1385         pipe = get_pipe_info(file, true);
1386         if (!pipe)
1387                 return -EBADF;
1388
1389         pipe_clear_nowait(file);
1390
1391         pipe_lock(pipe);
1392         ret = wait_for_space(pipe, flags);
1393         if (!ret)
1394                 ret = iter_to_pipe(iter, pipe, buf_flag);
1395         pipe_unlock(pipe);
1396         if (ret > 0)
1397                 wakeup_pipe_readers(pipe);
1398         return ret;
1399 }
1400
1401 static int vmsplice_type(struct fd f, int *type)
1402 {
1403         if (!f.file)
1404                 return -EBADF;
1405         if (f.file->f_mode & FMODE_WRITE) {
1406                 *type = ITER_SOURCE;
1407         } else if (f.file->f_mode & FMODE_READ) {
1408                 *type = ITER_DEST;
1409         } else {
1410                 fdput(f);
1411                 return -EBADF;
1412         }
1413         return 0;
1414 }
1415
1416 /*
1417  * Note that vmsplice only really supports true splicing _from_ user memory
1418  * to a pipe, not the other way around. Splicing from user memory is a simple
1419  * operation that can be supported without any funky alignment restrictions
1420  * or nasty vm tricks. We simply map in the user memory and fill them into
1421  * a pipe. The reverse isn't quite as easy, though. There are two possible
1422  * solutions for that:
1423  *
1424  *      - memcpy() the data internally, at which point we might as well just
1425  *        do a regular read() on the buffer anyway.
1426  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1427  *        has restriction limitations on both ends of the pipe).
1428  *
1429  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1430  *
1431  */
1432 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1433                 unsigned long, nr_segs, unsigned int, flags)
1434 {
1435         struct iovec iovstack[UIO_FASTIOV];
1436         struct iovec *iov = iovstack;
1437         struct iov_iter iter;
1438         ssize_t error;
1439         struct fd f;
1440         int type;
1441
1442         if (unlikely(flags & ~SPLICE_F_ALL))
1443                 return -EINVAL;
1444
1445         f = fdget(fd);
1446         error = vmsplice_type(f, &type);
1447         if (error)
1448                 return error;
1449
1450         error = import_iovec(type, uiov, nr_segs,
1451                              ARRAY_SIZE(iovstack), &iov, &iter);
1452         if (error < 0)
1453                 goto out_fdput;
1454
1455         if (!iov_iter_count(&iter))
1456                 error = 0;
1457         else if (type == ITER_SOURCE)
1458                 error = vmsplice_to_pipe(f.file, &iter, flags);
1459         else
1460                 error = vmsplice_to_user(f.file, &iter, flags);
1461
1462         kfree(iov);
1463 out_fdput:
1464         fdput(f);
1465         return error;
1466 }
1467
1468 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1469                 int, fd_out, loff_t __user *, off_out,
1470                 size_t, len, unsigned int, flags)
1471 {
1472         struct fd in, out;
1473         long error;
1474
1475         if (unlikely(!len))
1476                 return 0;
1477
1478         if (unlikely(flags & ~SPLICE_F_ALL))
1479                 return -EINVAL;
1480
1481         error = -EBADF;
1482         in = fdget(fd_in);
1483         if (in.file) {
1484                 out = fdget(fd_out);
1485                 if (out.file) {
1486                         error = __do_splice(in.file, off_in, out.file, off_out,
1487                                                 len, flags);
1488                         fdput(out);
1489                 }
1490                 fdput(in);
1491         }
1492         return error;
1493 }
1494
1495 /*
1496  * Make sure there's data to read. Wait for input if we can, otherwise
1497  * return an appropriate error.
1498  */
1499 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1500 {
1501         int ret;
1502
1503         /*
1504          * Check the pipe occupancy without the inode lock first. This function
1505          * is speculative anyways, so missing one is ok.
1506          */
1507         if (!pipe_empty(pipe->head, pipe->tail))
1508                 return 0;
1509
1510         ret = 0;
1511         pipe_lock(pipe);
1512
1513         while (pipe_empty(pipe->head, pipe->tail)) {
1514                 if (signal_pending(current)) {
1515                         ret = -ERESTARTSYS;
1516                         break;
1517                 }
1518                 if (!pipe->writers)
1519                         break;
1520                 if (flags & SPLICE_F_NONBLOCK) {
1521                         ret = -EAGAIN;
1522                         break;
1523                 }
1524                 pipe_wait_readable(pipe);
1525         }
1526
1527         pipe_unlock(pipe);
1528         return ret;
1529 }
1530
1531 /*
1532  * Make sure there's writeable room. Wait for room if we can, otherwise
1533  * return an appropriate error.
1534  */
1535 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1536 {
1537         int ret;
1538
1539         /*
1540          * Check pipe occupancy without the inode lock first. This function
1541          * is speculative anyways, so missing one is ok.
1542          */
1543         if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1544                 return 0;
1545
1546         ret = 0;
1547         pipe_lock(pipe);
1548
1549         while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1550                 if (!pipe->readers) {
1551                         send_sig(SIGPIPE, current, 0);
1552                         ret = -EPIPE;
1553                         break;
1554                 }
1555                 if (flags & SPLICE_F_NONBLOCK) {
1556                         ret = -EAGAIN;
1557                         break;
1558                 }
1559                 if (signal_pending(current)) {
1560                         ret = -ERESTARTSYS;
1561                         break;
1562                 }
1563                 pipe_wait_writable(pipe);
1564         }
1565
1566         pipe_unlock(pipe);
1567         return ret;
1568 }
1569
1570 /*
1571  * Splice contents of ipipe to opipe.
1572  */
1573 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1574                                struct pipe_inode_info *opipe,
1575                                size_t len, unsigned int flags)
1576 {
1577         struct pipe_buffer *ibuf, *obuf;
1578         unsigned int i_head, o_head;
1579         unsigned int i_tail, o_tail;
1580         unsigned int i_mask, o_mask;
1581         int ret = 0;
1582         bool input_wakeup = false;
1583
1584
1585 retry:
1586         ret = ipipe_prep(ipipe, flags);
1587         if (ret)
1588                 return ret;
1589
1590         ret = opipe_prep(opipe, flags);
1591         if (ret)
1592                 return ret;
1593
1594         /*
1595          * Potential ABBA deadlock, work around it by ordering lock
1596          * grabbing by pipe info address. Otherwise two different processes
1597          * could deadlock (one doing tee from A -> B, the other from B -> A).
1598          */
1599         pipe_double_lock(ipipe, opipe);
1600
1601         i_tail = ipipe->tail;
1602         i_mask = ipipe->ring_size - 1;
1603         o_head = opipe->head;
1604         o_mask = opipe->ring_size - 1;
1605
1606         do {
1607                 size_t o_len;
1608
1609                 if (!opipe->readers) {
1610                         send_sig(SIGPIPE, current, 0);
1611                         if (!ret)
1612                                 ret = -EPIPE;
1613                         break;
1614                 }
1615
1616                 i_head = ipipe->head;
1617                 o_tail = opipe->tail;
1618
1619                 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1620                         break;
1621
1622                 /*
1623                  * Cannot make any progress, because either the input
1624                  * pipe is empty or the output pipe is full.
1625                  */
1626                 if (pipe_empty(i_head, i_tail) ||
1627                     pipe_full(o_head, o_tail, opipe->max_usage)) {
1628                         /* Already processed some buffers, break */
1629                         if (ret)
1630                                 break;
1631
1632                         if (flags & SPLICE_F_NONBLOCK) {
1633                                 ret = -EAGAIN;
1634                                 break;
1635                         }
1636
1637                         /*
1638                          * We raced with another reader/writer and haven't
1639                          * managed to process any buffers.  A zero return
1640                          * value means EOF, so retry instead.
1641                          */
1642                         pipe_unlock(ipipe);
1643                         pipe_unlock(opipe);
1644                         goto retry;
1645                 }
1646
1647                 ibuf = &ipipe->bufs[i_tail & i_mask];
1648                 obuf = &opipe->bufs[o_head & o_mask];
1649
1650                 if (len >= ibuf->len) {
1651                         /*
1652                          * Simply move the whole buffer from ipipe to opipe
1653                          */
1654                         *obuf = *ibuf;
1655                         ibuf->ops = NULL;
1656                         i_tail++;
1657                         ipipe->tail = i_tail;
1658                         input_wakeup = true;
1659                         o_len = obuf->len;
1660                         o_head++;
1661                         opipe->head = o_head;
1662                 } else {
1663                         /*
1664                          * Get a reference to this pipe buffer,
1665                          * so we can copy the contents over.
1666                          */
1667                         if (!pipe_buf_get(ipipe, ibuf)) {
1668                                 if (ret == 0)
1669                                         ret = -EFAULT;
1670                                 break;
1671                         }
1672                         *obuf = *ibuf;
1673
1674                         /*
1675                          * Don't inherit the gift and merge flags, we need to
1676                          * prevent multiple steals of this page.
1677                          */
1678                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1679                         obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1680
1681                         obuf->len = len;
1682                         ibuf->offset += len;
1683                         ibuf->len -= len;
1684                         o_len = len;
1685                         o_head++;
1686                         opipe->head = o_head;
1687                 }
1688                 ret += o_len;
1689                 len -= o_len;
1690         } while (len);
1691
1692         pipe_unlock(ipipe);
1693         pipe_unlock(opipe);
1694
1695         /*
1696          * If we put data in the output pipe, wakeup any potential readers.
1697          */
1698         if (ret > 0)
1699                 wakeup_pipe_readers(opipe);
1700
1701         if (input_wakeup)
1702                 wakeup_pipe_writers(ipipe);
1703
1704         return ret;
1705 }
1706
1707 /*
1708  * Link contents of ipipe to opipe.
1709  */
1710 static int link_pipe(struct pipe_inode_info *ipipe,
1711                      struct pipe_inode_info *opipe,
1712                      size_t len, unsigned int flags)
1713 {
1714         struct pipe_buffer *ibuf, *obuf;
1715         unsigned int i_head, o_head;
1716         unsigned int i_tail, o_tail;
1717         unsigned int i_mask, o_mask;
1718         int ret = 0;
1719
1720         /*
1721          * Potential ABBA deadlock, work around it by ordering lock
1722          * grabbing by pipe info address. Otherwise two different processes
1723          * could deadlock (one doing tee from A -> B, the other from B -> A).
1724          */
1725         pipe_double_lock(ipipe, opipe);
1726
1727         i_tail = ipipe->tail;
1728         i_mask = ipipe->ring_size - 1;
1729         o_head = opipe->head;
1730         o_mask = opipe->ring_size - 1;
1731
1732         do {
1733                 if (!opipe->readers) {
1734                         send_sig(SIGPIPE, current, 0);
1735                         if (!ret)
1736                                 ret = -EPIPE;
1737                         break;
1738                 }
1739
1740                 i_head = ipipe->head;
1741                 o_tail = opipe->tail;
1742
1743                 /*
1744                  * If we have iterated all input buffers or run out of
1745                  * output room, break.
1746                  */
1747                 if (pipe_empty(i_head, i_tail) ||
1748                     pipe_full(o_head, o_tail, opipe->max_usage))
1749                         break;
1750
1751                 ibuf = &ipipe->bufs[i_tail & i_mask];
1752                 obuf = &opipe->bufs[o_head & o_mask];
1753
1754                 /*
1755                  * Get a reference to this pipe buffer,
1756                  * so we can copy the contents over.
1757                  */
1758                 if (!pipe_buf_get(ipipe, ibuf)) {
1759                         if (ret == 0)
1760                                 ret = -EFAULT;
1761                         break;
1762                 }
1763
1764                 *obuf = *ibuf;
1765
1766                 /*
1767                  * Don't inherit the gift and merge flag, we need to prevent
1768                  * multiple steals of this page.
1769                  */
1770                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1771                 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1772
1773                 if (obuf->len > len)
1774                         obuf->len = len;
1775                 ret += obuf->len;
1776                 len -= obuf->len;
1777
1778                 o_head++;
1779                 opipe->head = o_head;
1780                 i_tail++;
1781         } while (len);
1782
1783         pipe_unlock(ipipe);
1784         pipe_unlock(opipe);
1785
1786         /*
1787          * If we put data in the output pipe, wakeup any potential readers.
1788          */
1789         if (ret > 0)
1790                 wakeup_pipe_readers(opipe);
1791
1792         return ret;
1793 }
1794
1795 /*
1796  * This is a tee(1) implementation that works on pipes. It doesn't copy
1797  * any data, it simply references the 'in' pages on the 'out' pipe.
1798  * The 'flags' used are the SPLICE_F_* variants, currently the only
1799  * applicable one is SPLICE_F_NONBLOCK.
1800  */
1801 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1802 {
1803         struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1804         struct pipe_inode_info *opipe = get_pipe_info(out, true);
1805         int ret = -EINVAL;
1806
1807         if (unlikely(!(in->f_mode & FMODE_READ) ||
1808                      !(out->f_mode & FMODE_WRITE)))
1809                 return -EBADF;
1810
1811         /*
1812          * Duplicate the contents of ipipe to opipe without actually
1813          * copying the data.
1814          */
1815         if (ipipe && opipe && ipipe != opipe) {
1816                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1817                         flags |= SPLICE_F_NONBLOCK;
1818
1819                 /*
1820                  * Keep going, unless we encounter an error. The ipipe/opipe
1821                  * ordering doesn't really matter.
1822                  */
1823                 ret = ipipe_prep(ipipe, flags);
1824                 if (!ret) {
1825                         ret = opipe_prep(opipe, flags);
1826                         if (!ret)
1827                                 ret = link_pipe(ipipe, opipe, len, flags);
1828                 }
1829         }
1830
1831         return ret;
1832 }
1833
1834 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1835 {
1836         struct fd in, out;
1837         int error;
1838
1839         if (unlikely(flags & ~SPLICE_F_ALL))
1840                 return -EINVAL;
1841
1842         if (unlikely(!len))
1843                 return 0;
1844
1845         error = -EBADF;
1846         in = fdget(fdin);
1847         if (in.file) {
1848                 out = fdget(fdout);
1849                 if (out.file) {
1850                         error = do_tee(in.file, out.file, len, flags);
1851                         fdput(out);
1852                 }
1853                 fdput(in);
1854         }
1855
1856         return error;
1857 }