fs/pipe.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/pipe.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/file.h>
  10 #include <linux/poll.h>
  11 #include <linux/slab.h>
  12 #include <linux/module.h>
  13 #include <linux/init.h>
  14 #include <linux/fs.h>
  15 #include <linux/log2.h>
  16 #include <linux/mount.h>
  17 #include <linux/pseudo_fs.h>
  18 #include <linux/magic.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/uio.h>
  21 #include <linux/highmem.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/audit.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/fcntl.h>
  26 #include <linux/memcontrol.h>
  27
  28 #include <linux/uaccess.h>
  29 #include <asm/ioctls.h>
  30
  31 #include "internal.h"
  32
  33 /*
  34  * The max size that a non-root user is allowed to grow the pipe. Can
  35  * be set by root in /proc/sys/fs/pipe-max-size
  36  */
  37 unsigned int pipe_max_size = 1048576;
  38
  39 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  40  * matches default values.
  41  */
  42 unsigned long pipe_user_pages_hard;
  43 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  44
  45 /*
  46  * We use head and tail indices that aren't masked off, except at the point of
  47  * dereference, but rather they're allowed to wrap naturally.  This means there
  48  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  49  * <= 2^31.
  50  * -- David Howells 2019-09-23.
  51  *
  52  * Reads with count = 0 should always return 0.
  53  * -- Julian Bradfield 1999-06-07.
  54  *
  55  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  56  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  57  *
  58  * pipe_read & write cleanup
  59  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  60  */
  61
  62 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  63 {
  64         if (pipe->files)
  65                 mutex_lock_nested(&pipe->mutex, subclass);
  66 }
  67
  68 void pipe_lock(struct pipe_inode_info *pipe)
  69 {
  70         /*
  71          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  72          */
  73         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  74 }
  75 EXPORT_SYMBOL(pipe_lock);
  76
  77 void pipe_unlock(struct pipe_inode_info *pipe)
  78 {
  79         if (pipe->files)
  80                 mutex_unlock(&pipe->mutex);
  81 }
  82 EXPORT_SYMBOL(pipe_unlock);
  83
  84 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  85 {
  86         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  87 }
  88
  89 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  90 {
  91         mutex_unlock(&pipe->mutex);
  92 }
  93
  94 void pipe_double_lock(struct pipe_inode_info *pipe1,
  95                       struct pipe_inode_info *pipe2)
  96 {
  97         BUG_ON(pipe1 == pipe2);
  98
  99         if (pipe1 < pipe2) {
 100                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 101                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 102         } else {
 103                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 104                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 105         }
 106 }
 107
 108 /* Drop the inode semaphore and wait for a pipe event, atomically */
 109 void pipe_wait(struct pipe_inode_info *pipe)
 110 {
 111         DEFINE_WAIT(rdwait);
 112         DEFINE_WAIT(wrwait);
 113
 114         /*
 115          * Pipes are system-local resources, so sleeping on them
 116          * is considered a noninteractive wait:
 117          */
 118         prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
 119         prepare_to_wait(&pipe->wr_wait, &wrwait, TASK_INTERRUPTIBLE);
 120         pipe_unlock(pipe);
 121         schedule();
 122         finish_wait(&pipe->rd_wait, &rdwait);
 123         finish_wait(&pipe->wr_wait, &wrwait);
 124         pipe_lock(pipe);
 125 }
 126
 127 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 128                                   struct pipe_buffer *buf)
 129 {
 130         struct page *page = buf->page;
 131
 132         /*
 133          * If nobody else uses this page, and we don't already have a
 134          * temporary page, let's keep track of it as a one-deep
 135          * allocation cache. (Otherwise just release our reference to it)
 136          */
 137         if (page_count(page) == 1 && !pipe->tmp_page)
 138                 pipe->tmp_page = page;
 139         else
 140                 put_page(page);
 141 }
 142
 143 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 144                                struct pipe_buffer *buf)
 145 {
 146         struct page *page = buf->page;
 147
 148         if (page_count(page) == 1) {
 149                 memcg_kmem_uncharge(page, 0);
 150                 __SetPageLocked(page);
 151                 return 0;
 152         }
 153         return 1;
 154 }
 155
 156 /**
 157  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 158  * @pipe:       the pipe that the buffer belongs to
 159  * @buf:        the buffer to attempt to steal
 160  *
 161  * Description:
 162  *      This function attempts to steal the &struct page attached to
 163  *      @buf. If successful, this function returns 0 and returns with
 164  *      the page locked. The caller may then reuse the page for whatever
 165  *      he wishes; the typical use is insertion into a different file
 166  *      page cache.
 167  */
 168 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 169                            struct pipe_buffer *buf)
 170 {
 171         struct page *page = buf->page;
 172
 173         /*
 174          * A reference of one is golden, that means that the owner of this
 175          * page is the only one holding a reference to it. lock the page
 176          * and return OK.
 177          */
 178         if (page_count(page) == 1) {
 179                 lock_page(page);
 180                 return 0;
 181         }
 182
 183         return 1;
 184 }
 185 EXPORT_SYMBOL(generic_pipe_buf_steal);
 186
 187 /**
 188  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 189  * @pipe:       the pipe that the buffer belongs to
 190  * @buf:        the buffer to get a reference to
 191  *
 192  * Description:
 193  *      This function grabs an extra reference to @buf. It's used in
 194  *      in the tee() system call, when we duplicate the buffers in one
 195  *      pipe into another.
 196  */
 197 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 198 {
 199         return try_get_page(buf->page);
 200 }
 201 EXPORT_SYMBOL(generic_pipe_buf_get);
 202
 203 /**
 204  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 205  * @info:       the pipe that the buffer belongs to
 206  * @buf:        the buffer to confirm
 207  *
 208  * Description:
 209  *      This function does nothing, because the generic pipe code uses
 210  *      pages that are always good when inserted into the pipe.
 211  */
 212 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 213                              struct pipe_buffer *buf)
 214 {
 215         return 0;
 216 }
 217 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 218
 219 /**
 220  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 221  * @pipe:       the pipe that the buffer belongs to
 222  * @buf:        the buffer to put a reference to
 223  *
 224  * Description:
 225  *      This function releases a reference to @buf.
 226  */
 227 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 228                               struct pipe_buffer *buf)
 229 {
 230         put_page(buf->page);
 231 }
 232 EXPORT_SYMBOL(generic_pipe_buf_release);
 233
 234 /* New data written to a pipe may be appended to a buffer with this type. */
 235 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 236         .confirm = generic_pipe_buf_confirm,
 237         .release = anon_pipe_buf_release,
 238         .steal = anon_pipe_buf_steal,
 239         .get = generic_pipe_buf_get,
 240 };
 241
 242 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 243         .confirm = generic_pipe_buf_confirm,
 244         .release = anon_pipe_buf_release,
 245         .steal = anon_pipe_buf_steal,
 246         .get = generic_pipe_buf_get,
 247 };
 248
 249 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 250         .confirm = generic_pipe_buf_confirm,
 251         .release = anon_pipe_buf_release,
 252         .steal = anon_pipe_buf_steal,
 253         .get = generic_pipe_buf_get,
 254 };
 255
 256 /**
 257  * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
 258  * @buf:        the buffer to mark
 259  *
 260  * Description:
 261  *      This function ensures that no future writes will be merged into the
 262  *      given &struct pipe_buffer. This is necessary when multiple pipe buffers
 263  *      share the same backing page.
 264  */
 265 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 266 {
 267         if (buf->ops == &anon_pipe_buf_ops)
 268                 buf->ops = &anon_pipe_buf_nomerge_ops;
 269 }
 270
 271 static bool pipe_buf_can_merge(struct pipe_buffer *buf)
 272 {
 273         return buf->ops == &anon_pipe_buf_ops;
 274 }
 275
 276 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 277 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 278 {
 279         unsigned int head = READ_ONCE(pipe->head);
 280         unsigned int tail = READ_ONCE(pipe->tail);
 281         unsigned int writers = READ_ONCE(pipe->writers);
 282
 283         return !pipe_empty(head, tail) || !writers;
 284 }
 285
 286 static ssize_t
 287 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 288 {
 289         size_t total_len = iov_iter_count(to);
 290         struct file *filp = iocb->ki_filp;
 291         struct pipe_inode_info *pipe = filp->private_data;
 292         bool was_full, wake_next_reader = false;
 293         ssize_t ret;
 294
 295         /* Null read succeeds. */
 296         if (unlikely(total_len == 0))
 297                 return 0;
 298
 299         ret = 0;
 300         __pipe_lock(pipe);
 301
 302         /*
 303          * We only wake up writers if the pipe was full when we started
 304          * reading in order to avoid unnecessary wakeups.
 305          *
 306          * But when we do wake up writers, we do so using a sync wakeup
 307          * (WF_SYNC), because we want them to get going and generate more
 308          * data for us.
 309          */
 310         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 311         for (;;) {
 312                 unsigned int head = pipe->head;
 313                 unsigned int tail = pipe->tail;
 314                 unsigned int mask = pipe->ring_size - 1;
 315
 316                 if (!pipe_empty(head, tail)) {
 317                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 318                         size_t chars = buf->len;
 319                         size_t written;
 320                         int error;
 321
 322                         if (chars > total_len)
 323                                 chars = total_len;
 324
 325                         error = pipe_buf_confirm(pipe, buf);
 326                         if (error) {
 327                                 if (!ret)
 328                                         ret = error;
 329                                 break;
 330                         }
 331
 332                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 333                         if (unlikely(written < chars)) {
 334                                 if (!ret)
 335                                         ret = -EFAULT;
 336                                 break;
 337                         }
 338                         ret += chars;
 339                         buf->offset += chars;
 340                         buf->len -= chars;
 341
 342                         /* Was it a packet buffer? Clean up and exit */
 343                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 344                                 total_len = chars;
 345                                 buf->len = 0;
 346                         }
 347
 348                         if (!buf->len) {
 349                                 pipe_buf_release(pipe, buf);
 350                                 spin_lock_irq(&pipe->rd_wait.lock);
 351                                 tail++;
 352                                 pipe->tail = tail;
 353                                 spin_unlock_irq(&pipe->rd_wait.lock);
 354                         }
 355                         total_len -= chars;
 356                         if (!total_len)
 357                                 break;  /* common path: read succeeded */
 358                         if (!pipe_empty(head, tail))    /* More to do? */
 359                                 continue;
 360                 }
 361
 362                 if (!pipe->writers)
 363                         break;
 364                 if (ret)
 365                         break;
 366                 if (filp->f_flags & O_NONBLOCK) {
 367                         ret = -EAGAIN;
 368                         break;
 369                 }
 370                 __pipe_unlock(pipe);
 371
 372                 /*
 373                  * We only get here if we didn't actually read anything.
 374                  *
 375                  * However, we could have seen (and removed) a zero-sized
 376                  * pipe buffer, and might have made space in the buffers
 377                  * that way.
 378                  *
 379                  * You can't make zero-sized pipe buffers by doing an empty
 380                  * write (not even in packet mode), but they can happen if
 381                  * the writer gets an EFAULT when trying to fill a buffer
 382                  * that already got allocated and inserted in the buffer
 383                  * array.
 384                  *
 385                  * So we still need to wake up any pending writers in the
 386                  * _very_ unlikely case that the pipe was full, but we got
 387                  * no data.
 388                  */
 389                 if (unlikely(was_full)) {
 390                         wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 391                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 392                 }
 393
 394                 /*
 395                  * But because we didn't read anything, at this point we can
 396                  * just return directly with -ERESTARTSYS if we're interrupted,
 397                  * since we've done any required wakeups and there's no need
 398                  * to mark anything accessed. And we've dropped the lock.
 399                  */
 400                 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 401                         return -ERESTARTSYS;
 402
 403                 __pipe_lock(pipe);
 404                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 405                 wake_next_reader = true;
 406         }
 407         if (pipe_empty(pipe->head, pipe->tail))
 408                 wake_next_reader = false;
 409         __pipe_unlock(pipe);
 410
 411         if (was_full) {
 412                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 413                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 414         }
 415         if (wake_next_reader)
 416                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 417         if (ret > 0)
 418                 file_accessed(filp);
 419         return ret;
 420 }
 421
 422 static inline int is_packetized(struct file *file)
 423 {
 424         return (file->f_flags & O_DIRECT) != 0;
 425 }
 426
 427 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 428 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 429 {
 430         unsigned int head = READ_ONCE(pipe->head);
 431         unsigned int tail = READ_ONCE(pipe->tail);
 432         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 433
 434         return !pipe_full(head, tail, max_usage) ||
 435                 !READ_ONCE(pipe->readers);
 436 }
 437
 438 static ssize_t
 439 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 440 {
 441         struct file *filp = iocb->ki_filp;
 442         struct pipe_inode_info *pipe = filp->private_data;
 443         unsigned int head;
 444         ssize_t ret = 0;
 445         size_t total_len = iov_iter_count(from);
 446         ssize_t chars;
 447         bool was_empty = false;
 448         bool wake_next_writer = false;
 449
 450         /* Null write succeeds. */
 451         if (unlikely(total_len == 0))
 452                 return 0;
 453
 454         __pipe_lock(pipe);
 455
 456         if (!pipe->readers) {
 457                 send_sig(SIGPIPE, current, 0);
 458                 ret = -EPIPE;
 459                 goto out;
 460         }
 461
 462         /*
 463          * Only wake up if the pipe started out empty, since
 464          * otherwise there should be no readers waiting.
 465          *
 466          * If it wasn't empty we try to merge new data into
 467          * the last buffer.
 468          *
 469          * That naturally merges small writes, but it also
 470          * page-aligs the rest of the writes for large writes
 471          * spanning multiple pages.
 472          */
 473         head = pipe->head;
 474         was_empty = pipe_empty(head, pipe->tail);
 475         chars = total_len & (PAGE_SIZE-1);
 476         if (chars && !was_empty) {
 477                 unsigned int mask = pipe->ring_size - 1;
 478                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 479                 int offset = buf->offset + buf->len;
 480
 481                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 482                         ret = pipe_buf_confirm(pipe, buf);
 483                         if (ret)
 484                                 goto out;
 485
 486                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 487                         if (unlikely(ret < chars)) {
 488                                 ret = -EFAULT;
 489                                 goto out;
 490                         }
 491
 492                         buf->len += ret;
 493                         if (!iov_iter_count(from))
 494                                 goto out;
 495                 }
 496         }
 497
 498         for (;;) {
 499                 if (!pipe->readers) {
 500                         send_sig(SIGPIPE, current, 0);
 501                         if (!ret)
 502                                 ret = -EPIPE;
 503                         break;
 504                 }
 505
 506                 head = pipe->head;
 507                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 508                         unsigned int mask = pipe->ring_size - 1;
 509                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 510                         struct page *page = pipe->tmp_page;
 511                         int copied;
 512
 513                         if (!page) {
 514                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 515                                 if (unlikely(!page)) {
 516                                         ret = ret ? : -ENOMEM;
 517                                         break;
 518                                 }
 519                                 pipe->tmp_page = page;
 520                         }
 521
 522                         /* Allocate a slot in the ring in advance and attach an
 523                          * empty buffer.  If we fault or otherwise fail to use
 524                          * it, either the reader will consume it or it'll still
 525                          * be there for the next write.
 526                          */
 527                         spin_lock_irq(&pipe->rd_wait.lock);
 528
 529                         head = pipe->head;
 530                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 531                                 spin_unlock_irq(&pipe->rd_wait.lock);
 532                                 continue;
 533                         }
 534
 535                         pipe->head = head + 1;
 536                         spin_unlock_irq(&pipe->rd_wait.lock);
 537
 538                         /* Insert it into the buffer array */
 539                         buf = &pipe->bufs[head & mask];
 540                         buf->page = page;
 541                         buf->ops = &anon_pipe_buf_ops;
 542                         buf->offset = 0;
 543                         buf->len = 0;
 544                         buf->flags = 0;
 545                         if (is_packetized(filp)) {
 546                                 buf->ops = &packet_pipe_buf_ops;
 547                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 548                         }
 549                         pipe->tmp_page = NULL;
 550
 551                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 552                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 553                                 if (!ret)
 554                                         ret = -EFAULT;
 555                                 break;
 556                         }
 557                         ret += copied;
 558                         buf->offset = 0;
 559                         buf->len = copied;
 560
 561                         if (!iov_iter_count(from))
 562                                 break;
 563                 }
 564
 565                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 566                         continue;
 567
 568                 /* Wait for buffer space to become available. */
 569                 if (filp->f_flags & O_NONBLOCK) {
 570                         if (!ret)
 571                                 ret = -EAGAIN;
 572                         break;
 573                 }
 574                 if (signal_pending(current)) {
 575                         if (!ret)
 576                                 ret = -ERESTARTSYS;
 577                         break;
 578                 }
 579
 580                 /*
 581                  * We're going to release the pipe lock and wait for more
 582                  * space. We wake up any readers if necessary, and then
 583                  * after waiting we need to re-check whether the pipe
 584                  * become empty while we dropped the lock.
 585                  */
 586                 __pipe_unlock(pipe);
 587                 if (was_empty) {
 588                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 589                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 590                 }
 591                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 592                 __pipe_lock(pipe);
 593                 was_empty = pipe_empty(pipe->head, pipe->tail);
 594                 wake_next_writer = true;
 595         }
 596 out:
 597         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 598                 wake_next_writer = false;
 599         __pipe_unlock(pipe);
 600
 601         /*
 602          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 603          * want the reader to start processing things asap, rather than
 604          * leave the data pending.
 605          *
 606          * This is particularly important for small writes, because of
 607          * how (for example) the GNU make jobserver uses small writes to
 608          * wake up pending jobs
 609          */
 610         if (was_empty) {
 611                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 612                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 613         }
 614         if (wake_next_writer)
 615                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 616         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 617                 int err = file_update_time(filp);
 618                 if (err)
 619                         ret = err;
 620                 sb_end_write(file_inode(filp)->i_sb);
 621         }
 622         return ret;
 623 }
 624
 625 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 626 {
 627         struct pipe_inode_info *pipe = filp->private_data;
 628         int count, head, tail, mask;
 629
 630         switch (cmd) {
 631                 case FIONREAD:
 632                         __pipe_lock(pipe);
 633                         count = 0;
 634                         head = pipe->head;
 635                         tail = pipe->tail;
 636                         mask = pipe->ring_size - 1;
 637
 638                         while (tail != head) {
 639                                 count += pipe->bufs[tail & mask].len;
 640                                 tail++;
 641                         }
 642                         __pipe_unlock(pipe);
 643
 644                         return put_user(count, (int __user *)arg);
 645                 default:
 646                         return -ENOIOCTLCMD;
 647         }
 648 }
 649
 650 /* No kernel lock held - fine */
 651 static __poll_t
 652 pipe_poll(struct file *filp, poll_table *wait)
 653 {
 654         __poll_t mask;
 655         struct pipe_inode_info *pipe = filp->private_data;
 656         unsigned int head, tail;
 657
 658         /*
 659          * Reading pipe state only -- no need for acquiring the semaphore.
 660          *
 661          * But because this is racy, the code has to add the
 662          * entry to the poll table _first_ ..
 663          */
 664         if (filp->f_mode & FMODE_READ)
 665                 poll_wait(filp, &pipe->rd_wait, wait);
 666         if (filp->f_mode & FMODE_WRITE)
 667                 poll_wait(filp, &pipe->wr_wait, wait);
 668
 669         /*
 670          * .. and only then can you do the racy tests. That way,
 671          * if something changes and you got it wrong, the poll
 672          * table entry will wake you up and fix it.
 673          */
 674         head = READ_ONCE(pipe->head);
 675         tail = READ_ONCE(pipe->tail);
 676
 677         mask = 0;
 678         if (filp->f_mode & FMODE_READ) {
 679                 if (!pipe_empty(head, tail))
 680                         mask |= EPOLLIN | EPOLLRDNORM;
 681                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 682                         mask |= EPOLLHUP;
 683         }
 684
 685         if (filp->f_mode & FMODE_WRITE) {
 686                 if (!pipe_full(head, tail, pipe->max_usage))
 687                         mask |= EPOLLOUT | EPOLLWRNORM;
 688                 /*
 689                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 690                  * behave exactly like pipes for poll().
 691                  */
 692                 if (!pipe->readers)
 693                         mask |= EPOLLERR;
 694         }
 695
 696         return mask;
 697 }
 698
 699 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 700 {
 701         int kill = 0;
 702
 703         spin_lock(&inode->i_lock);
 704         if (!--pipe->files) {
 705                 inode->i_pipe = NULL;
 706                 kill = 1;
 707         }
 708         spin_unlock(&inode->i_lock);
 709
 710         if (kill)
 711                 free_pipe_info(pipe);
 712 }
 713
 714 static int
 715 pipe_release(struct inode *inode, struct file *file)
 716 {
 717         struct pipe_inode_info *pipe = file->private_data;
 718
 719         __pipe_lock(pipe);
 720         if (file->f_mode & FMODE_READ)
 721                 pipe->readers--;
 722         if (file->f_mode & FMODE_WRITE)
 723                 pipe->writers--;
 724
 725         if (pipe->readers || pipe->writers) {
 726                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLHUP);
 727                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 728                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 729                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 730         }
 731         __pipe_unlock(pipe);
 732
 733         put_pipe_info(inode, pipe);
 734         return 0;
 735 }
 736
 737 static int
 738 pipe_fasync(int fd, struct file *filp, int on)
 739 {
 740         struct pipe_inode_info *pipe = filp->private_data;
 741         int retval = 0;
 742
 743         __pipe_lock(pipe);
 744         if (filp->f_mode & FMODE_READ)
 745                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 746         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 747                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 748                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 749                         /* this can happen only if on == T */
 750                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 751         }
 752         __pipe_unlock(pipe);
 753         return retval;
 754 }
 755
 756 static unsigned long account_pipe_buffers(struct user_struct *user,
 757                                  unsigned long old, unsigned long new)
 758 {
 759         return atomic_long_add_return(new - old, &user->pipe_bufs);
 760 }
 761
 762 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 763 {
 764         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 765
 766         return soft_limit && user_bufs > soft_limit;
 767 }
 768
 769 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 770 {
 771         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 772
 773         return hard_limit && user_bufs > hard_limit;
 774 }
 775
 776 static bool is_unprivileged_user(void)
 777 {
 778         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 779 }
 780
 781 struct pipe_inode_info *alloc_pipe_info(void)
 782 {
 783         struct pipe_inode_info *pipe;
 784         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 785         struct user_struct *user = get_current_user();
 786         unsigned long user_bufs;
 787         unsigned int max_size = READ_ONCE(pipe_max_size);
 788
 789         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 790         if (pipe == NULL)
 791                 goto out_free_uid;
 792
 793         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 794                 pipe_bufs = max_size >> PAGE_SHIFT;
 795
 796         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 797
 798         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 799                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 800                 pipe_bufs = 1;
 801         }
 802
 803         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 804                 goto out_revert_acct;
 805
 806         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 807                              GFP_KERNEL_ACCOUNT);
 808
 809         if (pipe->bufs) {
 810                 init_waitqueue_head(&pipe->rd_wait);
 811                 init_waitqueue_head(&pipe->wr_wait);
 812                 pipe->r_counter = pipe->w_counter = 1;
 813                 pipe->max_usage = pipe_bufs;
 814                 pipe->ring_size = pipe_bufs;
 815                 pipe->user = user;
 816                 mutex_init(&pipe->mutex);
 817                 return pipe;
 818         }
 819
 820 out_revert_acct:
 821         (void) account_pipe_buffers(user, pipe_bufs, 0);
 822         kfree(pipe);
 823 out_free_uid:
 824         free_uid(user);
 825         return NULL;
 826 }
 827
 828 void free_pipe_info(struct pipe_inode_info *pipe)
 829 {
 830         int i;
 831
 832         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 833         free_uid(pipe->user);
 834         for (i = 0; i < pipe->ring_size; i++) {
 835                 struct pipe_buffer *buf = pipe->bufs + i;
 836                 if (buf->ops)
 837                         pipe_buf_release(pipe, buf);
 838         }
 839         if (pipe->tmp_page)
 840                 __free_page(pipe->tmp_page);
 841         kfree(pipe->bufs);
 842         kfree(pipe);
 843 }
 844
 845 static struct vfsmount *pipe_mnt __read_mostly;
 846
 847 /*
 848  * pipefs_dname() is called from d_path().
 849  */
 850 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 851 {
 852         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 853                                 d_inode(dentry)->i_ino);
 854 }
 855
 856 static const struct dentry_operations pipefs_dentry_operations = {
 857         .d_dname        = pipefs_dname,
 858 };
 859
 860 static struct inode * get_pipe_inode(void)
 861 {
 862         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 863         struct pipe_inode_info *pipe;
 864
 865         if (!inode)
 866                 goto fail_inode;
 867
 868         inode->i_ino = get_next_ino();
 869
 870         pipe = alloc_pipe_info();
 871         if (!pipe)
 872                 goto fail_iput;
 873
 874         inode->i_pipe = pipe;
 875         pipe->files = 2;
 876         pipe->readers = pipe->writers = 1;
 877         inode->i_fop = &pipefifo_fops;
 878
 879         /*
 880          * Mark the inode dirty from the very beginning,
 881          * that way it will never be moved to the dirty
 882          * list because "mark_inode_dirty()" will think
 883          * that it already _is_ on the dirty list.
 884          */
 885         inode->i_state = I_DIRTY;
 886         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 887         inode->i_uid = current_fsuid();
 888         inode->i_gid = current_fsgid();
 889         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 890
 891         return inode;
 892
 893 fail_iput:
 894         iput(inode);
 895
 896 fail_inode:
 897         return NULL;
 898 }
 899
 900 int create_pipe_files(struct file **res, int flags)
 901 {
 902         struct inode *inode = get_pipe_inode();
 903         struct file *f;
 904
 905         if (!inode)
 906                 return -ENFILE;
 907
 908         f = alloc_file_pseudo(inode, pipe_mnt, "",
 909                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 910                                 &pipefifo_fops);
 911         if (IS_ERR(f)) {
 912                 free_pipe_info(inode->i_pipe);
 913                 iput(inode);
 914                 return PTR_ERR(f);
 915         }
 916
 917         f->private_data = inode->i_pipe;
 918
 919         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 920                                   &pipefifo_fops);
 921         if (IS_ERR(res[0])) {
 922                 put_pipe_info(inode, inode->i_pipe);
 923                 fput(f);
 924                 return PTR_ERR(res[0]);
 925         }
 926         res[0]->private_data = inode->i_pipe;
 927         res[1] = f;
 928         stream_open(inode, res[0]);
 929         stream_open(inode, res[1]);
 930         return 0;
 931 }
 932
 933 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 934 {
 935         int error;
 936         int fdw, fdr;
 937
 938         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 939                 return -EINVAL;
 940
 941         error = create_pipe_files(files, flags);
 942         if (error)
 943                 return error;
 944
 945         error = get_unused_fd_flags(flags);
 946         if (error < 0)
 947                 goto err_read_pipe;
 948         fdr = error;
 949
 950         error = get_unused_fd_flags(flags);
 951         if (error < 0)
 952                 goto err_fdr;
 953         fdw = error;
 954
 955         audit_fd_pair(fdr, fdw);
 956         fd[0] = fdr;
 957         fd[1] = fdw;
 958         return 0;
 959
 960  err_fdr:
 961         put_unused_fd(fdr);
 962  err_read_pipe:
 963         fput(files[0]);
 964         fput(files[1]);
 965         return error;
 966 }
 967
 968 int do_pipe_flags(int *fd, int flags)
 969 {
 970         struct file *files[2];
 971         int error = __do_pipe_flags(fd, files, flags);
 972         if (!error) {
 973                 fd_install(fd[0], files[0]);
 974                 fd_install(fd[1], files[1]);
 975         }
 976         return error;
 977 }
 978
 979 /*
 980  * sys_pipe() is the normal C calling standard for creating
 981  * a pipe. It's not the way Unix traditionally does this, though.
 982  */
 983 static int do_pipe2(int __user *fildes, int flags)
 984 {
 985         struct file *files[2];
 986         int fd[2];
 987         int error;
 988
 989         error = __do_pipe_flags(fd, files, flags);
 990         if (!error) {
 991                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 992                         fput(files[0]);
 993                         fput(files[1]);
 994                         put_unused_fd(fd[0]);
 995                         put_unused_fd(fd[1]);
 996                         error = -EFAULT;
 997                 } else {
 998                         fd_install(fd[0], files[0]);
 999                         fd_install(fd[1], files[1]);
1000                 }
1001         }
1002         return error;
1003 }
1004
1005 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1006 {
1007         return do_pipe2(fildes, flags);
1008 }
1009
1010 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1011 {
1012         return do_pipe2(fildes, 0);
1013 }
1014
1015 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1016 {
1017         int cur = *cnt;
1018
1019         while (cur == *cnt) {
1020                 pipe_wait(pipe);
1021                 if (signal_pending(current))
1022                         break;
1023         }
1024         return cur == *cnt ? -ERESTARTSYS : 0;
1025 }
1026
1027 static void wake_up_partner(struct pipe_inode_info *pipe)
1028 {
1029         wake_up_interruptible(&pipe->rd_wait);
1030         wake_up_interruptible(&pipe->wr_wait);
1031 }
1032
1033 static int fifo_open(struct inode *inode, struct file *filp)
1034 {
1035         struct pipe_inode_info *pipe;
1036         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1037         int ret;
1038
1039         filp->f_version = 0;
1040
1041         spin_lock(&inode->i_lock);
1042         if (inode->i_pipe) {
1043                 pipe = inode->i_pipe;
1044                 pipe->files++;
1045                 spin_unlock(&inode->i_lock);
1046         } else {
1047                 spin_unlock(&inode->i_lock);
1048                 pipe = alloc_pipe_info();
1049                 if (!pipe)
1050                         return -ENOMEM;
1051                 pipe->files = 1;
1052                 spin_lock(&inode->i_lock);
1053                 if (unlikely(inode->i_pipe)) {
1054                         inode->i_pipe->files++;
1055                         spin_unlock(&inode->i_lock);
1056                         free_pipe_info(pipe);
1057                         pipe = inode->i_pipe;
1058                 } else {
1059                         inode->i_pipe = pipe;
1060                         spin_unlock(&inode->i_lock);
1061                 }
1062         }
1063         filp->private_data = pipe;
1064         /* OK, we have a pipe and it's pinned down */
1065
1066         __pipe_lock(pipe);
1067
1068         /* We can only do regular read/write on fifos */
1069         stream_open(inode, filp);
1070
1071         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1072         case FMODE_READ:
1073         /*
1074          *  O_RDONLY
1075          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1076          *  opened, even when there is no process writing the FIFO.
1077          */
1078                 pipe->r_counter++;
1079                 if (pipe->readers++ == 0)
1080                         wake_up_partner(pipe);
1081
1082                 if (!is_pipe && !pipe->writers) {
1083                         if ((filp->f_flags & O_NONBLOCK)) {
1084                                 /* suppress EPOLLHUP until we have
1085                                  * seen a writer */
1086                                 filp->f_version = pipe->w_counter;
1087                         } else {
1088                                 if (wait_for_partner(pipe, &pipe->w_counter))
1089                                         goto err_rd;
1090                         }
1091                 }
1092                 break;
1093
1094         case FMODE_WRITE:
1095         /*
1096          *  O_WRONLY
1097          *  POSIX.1 says that O_NONBLOCK means return -1 with
1098          *  errno=ENXIO when there is no process reading the FIFO.
1099          */
1100                 ret = -ENXIO;
1101                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1102                         goto err;
1103
1104                 pipe->w_counter++;
1105                 if (!pipe->writers++)
1106                         wake_up_partner(pipe);
1107
1108                 if (!is_pipe && !pipe->readers) {
1109                         if (wait_for_partner(pipe, &pipe->r_counter))
1110                                 goto err_wr;
1111                 }
1112                 break;
1113
1114         case FMODE_READ | FMODE_WRITE:
1115         /*
1116          *  O_RDWR
1117          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1118          *  This implementation will NEVER block on a O_RDWR open, since
1119          *  the process can at least talk to itself.
1120          */
1121
1122                 pipe->readers++;
1123                 pipe->writers++;
1124                 pipe->r_counter++;
1125                 pipe->w_counter++;
1126                 if (pipe->readers == 1 || pipe->writers == 1)
1127                         wake_up_partner(pipe);
1128                 break;
1129
1130         default:
1131                 ret = -EINVAL;
1132                 goto err;
1133         }
1134
1135         /* Ok! */
1136         __pipe_unlock(pipe);
1137         return 0;
1138
1139 err_rd:
1140         if (!--pipe->readers)
1141                 wake_up_interruptible(&pipe->wr_wait);
1142         ret = -ERESTARTSYS;
1143         goto err;
1144
1145 err_wr:
1146         if (!--pipe->writers)
1147                 wake_up_interruptible(&pipe->rd_wait);
1148         ret = -ERESTARTSYS;
1149         goto err;
1150
1151 err:
1152         __pipe_unlock(pipe);
1153
1154         put_pipe_info(inode, pipe);
1155         return ret;
1156 }
1157
1158 const struct file_operations pipefifo_fops = {
1159         .open           = fifo_open,
1160         .llseek         = no_llseek,
1161         .read_iter      = pipe_read,
1162         .write_iter     = pipe_write,
1163         .poll           = pipe_poll,
1164         .unlocked_ioctl = pipe_ioctl,
1165         .release        = pipe_release,
1166         .fasync         = pipe_fasync,
1167 };
1168
1169 /*
1170  * Currently we rely on the pipe array holding a power-of-2 number
1171  * of pages. Returns 0 on error.
1172  */
1173 unsigned int round_pipe_size(unsigned long size)
1174 {
1175         if (size > (1U << 31))
1176                 return 0;
1177
1178         /* Minimum pipe size, as required by POSIX */
1179         if (size < PAGE_SIZE)
1180                 return PAGE_SIZE;
1181
1182         return roundup_pow_of_two(size);
1183 }
1184
1185 /*
1186  * Allocate a new array of pipe buffers and copy the info over. Returns the
1187  * pipe size if successful, or return -ERROR on error.
1188  */
1189 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1190 {
1191         struct pipe_buffer *bufs;
1192         unsigned int size, nr_slots, head, tail, mask, n;
1193         unsigned long user_bufs;
1194         long ret = 0;
1195
1196         size = round_pipe_size(arg);
1197         nr_slots = size >> PAGE_SHIFT;
1198
1199         if (!nr_slots)
1200                 return -EINVAL;
1201
1202         /*
1203          * If trying to increase the pipe capacity, check that an
1204          * unprivileged user is not trying to exceed various limits
1205          * (soft limit check here, hard limit check just below).
1206          * Decreasing the pipe capacity is always permitted, even
1207          * if the user is currently over a limit.
1208          */
1209         if (nr_slots > pipe->ring_size &&
1210                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1211                 return -EPERM;
1212
1213         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1214
1215         if (nr_slots > pipe->ring_size &&
1216                         (too_many_pipe_buffers_hard(user_bufs) ||
1217                          too_many_pipe_buffers_soft(user_bufs)) &&
1218                         is_unprivileged_user()) {
1219                 ret = -EPERM;
1220                 goto out_revert_acct;
1221         }
1222
1223         /*
1224          * We can shrink the pipe, if arg is greater than the ring occupancy.
1225          * Since we don't expect a lot of shrink+grow operations, just free and
1226          * allocate again like we would do for growing.  If the pipe currently
1227          * contains more buffers than arg, then return busy.
1228          */
1229         mask = pipe->ring_size - 1;
1230         head = pipe->head;
1231         tail = pipe->tail;
1232         n = pipe_occupancy(pipe->head, pipe->tail);
1233         if (nr_slots < n) {
1234                 ret = -EBUSY;
1235                 goto out_revert_acct;
1236         }
1237
1238         bufs = kcalloc(nr_slots, sizeof(*bufs),
1239                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1240         if (unlikely(!bufs)) {
1241                 ret = -ENOMEM;
1242                 goto out_revert_acct;
1243         }
1244
1245         /*
1246          * The pipe array wraps around, so just start the new one at zero
1247          * and adjust the indices.
1248          */
1249         if (n > 0) {
1250                 unsigned int h = head & mask;
1251                 unsigned int t = tail & mask;
1252                 if (h > t) {
1253                         memcpy(bufs, pipe->bufs + t,
1254                                n * sizeof(struct pipe_buffer));
1255                 } else {
1256                         unsigned int tsize = pipe->ring_size - t;
1257                         if (h > 0)
1258                                 memcpy(bufs + tsize, pipe->bufs,
1259                                        h * sizeof(struct pipe_buffer));
1260                         memcpy(bufs, pipe->bufs + t,
1261                                tsize * sizeof(struct pipe_buffer));
1262                 }
1263         }
1264
1265         head = n;
1266         tail = 0;
1267
1268         kfree(pipe->bufs);
1269         pipe->bufs = bufs;
1270         pipe->ring_size = nr_slots;
1271         pipe->max_usage = nr_slots;
1272         pipe->tail = tail;
1273         pipe->head = head;
1274         wake_up_interruptible_all(&pipe->rd_wait);
1275         wake_up_interruptible_all(&pipe->wr_wait);
1276         return pipe->max_usage * PAGE_SIZE;
1277
1278 out_revert_acct:
1279         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1280         return ret;
1281 }
1282
1283 /*
1284  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1285  * location, so checking ->i_pipe is not enough to verify that this is a
1286  * pipe.
1287  */
1288 struct pipe_inode_info *get_pipe_info(struct file *file)
1289 {
1290         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1291 }
1292
1293 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1294 {
1295         struct pipe_inode_info *pipe;
1296         long ret;
1297
1298         pipe = get_pipe_info(file);
1299         if (!pipe)
1300                 return -EBADF;
1301
1302         __pipe_lock(pipe);
1303
1304         switch (cmd) {
1305         case F_SETPIPE_SZ:
1306                 ret = pipe_set_size(pipe, arg);
1307                 break;
1308         case F_GETPIPE_SZ:
1309                 ret = pipe->max_usage * PAGE_SIZE;
1310                 break;
1311         default:
1312                 ret = -EINVAL;
1313                 break;
1314         }
1315
1316         __pipe_unlock(pipe);
1317         return ret;
1318 }
1319
1320 static const struct super_operations pipefs_ops = {
1321         .destroy_inode = free_inode_nonrcu,
1322         .statfs = simple_statfs,
1323 };
1324
1325 /*
1326  * pipefs should _never_ be mounted by userland - too much of security hassle,
1327  * no real gain from having the whole whorehouse mounted. So we don't need
1328  * any operations on the root directory. However, we need a non-trivial
1329  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1330  */
1331
1332 static int pipefs_init_fs_context(struct fs_context *fc)
1333 {
1334         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1335         if (!ctx)
1336                 return -ENOMEM;
1337         ctx->ops = &pipefs_ops;
1338         ctx->dops = &pipefs_dentry_operations;
1339         return 0;
1340 }
1341
1342 static struct file_system_type pipe_fs_type = {
1343         .name           = "pipefs",
1344         .init_fs_context = pipefs_init_fs_context,
1345         .kill_sb        = kill_anon_super,
1346 };
1347
1348 static int __init init_pipe_fs(void)
1349 {
1350         int err = register_filesystem(&pipe_fs_type);
1351
1352         if (!err) {
1353                 pipe_mnt = kern_mount(&pipe_fs_type);
1354                 if (IS_ERR(pipe_mnt)) {
1355                         err = PTR_ERR(pipe_mnt);
1356                         unregister_filesystem(&pipe_fs_type);
1357                 }
1358         }
1359         return err;
1360 }
1361
1362 fs_initcall(init_pipe_fs);