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 #include <linux/watch_queue.h>
  28
  29 #include <linux/uaccess.h>
  30 #include <asm/ioctls.h>
  31
  32 #include "internal.h"
  33
  34 /*
  35  * New pipe buffers will be restricted to this size while the user is exceeding
  36  * their pipe buffer quota. The general pipe use case needs at least two
  37  * buffers: one for data yet to be read, and one for new data. If this is less
  38  * than two, then a write to a non-empty pipe may block even if the pipe is not
  39  * full. This can occur with GNU make jobserver or similar uses of pipes as
  40  * semaphores: multiple processes may be waiting to write tokens back to the
  41  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  42  *
  43  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  44  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  45  * emptied.
  46  */
  47 #define PIPE_MIN_DEF_BUFFERS 2
  48
  49 /*
  50  * The max size that a non-root user is allowed to grow the pipe. Can
  51  * be set by root in /proc/sys/fs/pipe-max-size
  52  */
  53 unsigned int pipe_max_size = 1048576;
  54
  55 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  56  * matches default values.
  57  */
  58 unsigned long pipe_user_pages_hard;
  59 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  60
  61 /*
  62  * We use head and tail indices that aren't masked off, except at the point of
  63  * dereference, but rather they're allowed to wrap naturally.  This means there
  64  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  65  * <= 2^31.
  66  * -- David Howells 2019-09-23.
  67  *
  68  * Reads with count = 0 should always return 0.
  69  * -- Julian Bradfield 1999-06-07.
  70  *
  71  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  72  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  73  *
  74  * pipe_read & write cleanup
  75  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  76  */
  77
  78 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  79 {
  80         if (pipe->files)
  81                 mutex_lock_nested(&pipe->mutex, subclass);
  82 }
  83
  84 void pipe_lock(struct pipe_inode_info *pipe)
  85 {
  86         /*
  87          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  88          */
  89         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  90 }
  91 EXPORT_SYMBOL(pipe_lock);
  92
  93 void pipe_unlock(struct pipe_inode_info *pipe)
  94 {
  95         if (pipe->files)
  96                 mutex_unlock(&pipe->mutex);
  97 }
  98 EXPORT_SYMBOL(pipe_unlock);
  99
 100 static inline void __pipe_lock(struct pipe_inode_info *pipe)
 101 {
 102         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 103 }
 104
 105 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 106 {
 107         mutex_unlock(&pipe->mutex);
 108 }
 109
 110 void pipe_double_lock(struct pipe_inode_info *pipe1,
 111                       struct pipe_inode_info *pipe2)
 112 {
 113         BUG_ON(pipe1 == pipe2);
 114
 115         if (pipe1 < pipe2) {
 116                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 117                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 118         } else {
 119                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 120                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 121         }
 122 }
 123
 124 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 125                                   struct pipe_buffer *buf)
 126 {
 127         struct page *page = buf->page;
 128
 129         /*
 130          * If nobody else uses this page, and we don't already have a
 131          * temporary page, let's keep track of it as a one-deep
 132          * allocation cache. (Otherwise just release our reference to it)
 133          */
 134         if (page_count(page) == 1 && !pipe->tmp_page)
 135                 pipe->tmp_page = page;
 136         else
 137                 put_page(page);
 138 }
 139
 140 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 141                 struct pipe_buffer *buf)
 142 {
 143         struct page *page = buf->page;
 144
 145         if (page_count(page) != 1)
 146                 return false;
 147         memcg_kmem_uncharge_page(page, 0);
 148         __SetPageLocked(page);
 149         return true;
 150 }
 151
 152 /**
 153  * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
 154  * @pipe:       the pipe that the buffer belongs to
 155  * @buf:        the buffer to attempt to steal
 156  *
 157  * Description:
 158  *      This function attempts to steal the &struct page attached to
 159  *      @buf. If successful, this function returns 0 and returns with
 160  *      the page locked. The caller may then reuse the page for whatever
 161  *      he wishes; the typical use is insertion into a different file
 162  *      page cache.
 163  */
 164 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 165                 struct pipe_buffer *buf)
 166 {
 167         struct page *page = buf->page;
 168
 169         /*
 170          * A reference of one is golden, that means that the owner of this
 171          * page is the only one holding a reference to it. lock the page
 172          * and return OK.
 173          */
 174         if (page_count(page) == 1) {
 175                 lock_page(page);
 176                 return true;
 177         }
 178         return false;
 179 }
 180 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
 181
 182 /**
 183  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 184  * @pipe:       the pipe that the buffer belongs to
 185  * @buf:        the buffer to get a reference to
 186  *
 187  * Description:
 188  *      This function grabs an extra reference to @buf. It's used in
 189  *      the tee() system call, when we duplicate the buffers in one
 190  *      pipe into another.
 191  */
 192 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 193 {
 194         return try_get_page(buf->page);
 195 }
 196 EXPORT_SYMBOL(generic_pipe_buf_get);
 197
 198 /**
 199  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 200  * @pipe:       the pipe that the buffer belongs to
 201  * @buf:        the buffer to put a reference to
 202  *
 203  * Description:
 204  *      This function releases a reference to @buf.
 205  */
 206 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 207                               struct pipe_buffer *buf)
 208 {
 209         put_page(buf->page);
 210 }
 211 EXPORT_SYMBOL(generic_pipe_buf_release);
 212
 213 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 214         .release        = anon_pipe_buf_release,
 215         .try_steal      = anon_pipe_buf_try_steal,
 216         .get            = generic_pipe_buf_get,
 217 };
 218
 219 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 220 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 221 {
 222         unsigned int head = READ_ONCE(pipe->head);
 223         unsigned int tail = READ_ONCE(pipe->tail);
 224         unsigned int writers = READ_ONCE(pipe->writers);
 225
 226         return !pipe_empty(head, tail) || !writers;
 227 }
 228
 229 static ssize_t
 230 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 231 {
 232         size_t total_len = iov_iter_count(to);
 233         struct file *filp = iocb->ki_filp;
 234         struct pipe_inode_info *pipe = filp->private_data;
 235         bool was_full, wake_next_reader = false;
 236         ssize_t ret;
 237
 238         /* Null read succeeds. */
 239         if (unlikely(total_len == 0))
 240                 return 0;
 241
 242         ret = 0;
 243         __pipe_lock(pipe);
 244
 245         /*
 246          * We only wake up writers if the pipe was full when we started
 247          * reading in order to avoid unnecessary wakeups.
 248          *
 249          * But when we do wake up writers, we do so using a sync wakeup
 250          * (WF_SYNC), because we want them to get going and generate more
 251          * data for us.
 252          */
 253         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 254         for (;;) {
 255                 unsigned int head = pipe->head;
 256                 unsigned int tail = pipe->tail;
 257                 unsigned int mask = pipe->ring_size - 1;
 258
 259 #ifdef CONFIG_WATCH_QUEUE
 260                 if (pipe->note_loss) {
 261                         struct watch_notification n;
 262
 263                         if (total_len < 8) {
 264                                 if (ret == 0)
 265                                         ret = -ENOBUFS;
 266                                 break;
 267                         }
 268
 269                         n.type = WATCH_TYPE_META;
 270                         n.subtype = WATCH_META_LOSS_NOTIFICATION;
 271                         n.info = watch_sizeof(n);
 272                         if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
 273                                 if (ret == 0)
 274                                         ret = -EFAULT;
 275                                 break;
 276                         }
 277                         ret += sizeof(n);
 278                         total_len -= sizeof(n);
 279                         pipe->note_loss = false;
 280                 }
 281 #endif
 282
 283                 if (!pipe_empty(head, tail)) {
 284                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 285                         size_t chars = buf->len;
 286                         size_t written;
 287                         int error;
 288
 289                         if (chars > total_len) {
 290                                 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
 291                                         if (ret == 0)
 292                                                 ret = -ENOBUFS;
 293                                         break;
 294                                 }
 295                                 chars = total_len;
 296                         }
 297
 298                         error = pipe_buf_confirm(pipe, buf);
 299                         if (error) {
 300                                 if (!ret)
 301                                         ret = error;
 302                                 break;
 303                         }
 304
 305                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 306                         if (unlikely(written < chars)) {
 307                                 if (!ret)
 308                                         ret = -EFAULT;
 309                                 break;
 310                         }
 311                         ret += chars;
 312                         buf->offset += chars;
 313                         buf->len -= chars;
 314
 315                         /* Was it a packet buffer? Clean up and exit */
 316                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 317                                 total_len = chars;
 318                                 buf->len = 0;
 319                         }
 320
 321                         if (!buf->len) {
 322                                 pipe_buf_release(pipe, buf);
 323                                 spin_lock_irq(&pipe->rd_wait.lock);
 324 #ifdef CONFIG_WATCH_QUEUE
 325                                 if (buf->flags & PIPE_BUF_FLAG_LOSS)
 326                                         pipe->note_loss = true;
 327 #endif
 328                                 tail++;
 329                                 pipe->tail = tail;
 330                                 spin_unlock_irq(&pipe->rd_wait.lock);
 331                         }
 332                         total_len -= chars;
 333                         if (!total_len)
 334                                 break;  /* common path: read succeeded */
 335                         if (!pipe_empty(head, tail))    /* More to do? */
 336                                 continue;
 337                 }
 338
 339                 if (!pipe->writers)
 340                         break;
 341                 if (ret)
 342                         break;
 343                 if (filp->f_flags & O_NONBLOCK) {
 344                         ret = -EAGAIN;
 345                         break;
 346                 }
 347                 __pipe_unlock(pipe);
 348
 349                 /*
 350                  * We only get here if we didn't actually read anything.
 351                  *
 352                  * However, we could have seen (and removed) a zero-sized
 353                  * pipe buffer, and might have made space in the buffers
 354                  * that way.
 355                  *
 356                  * You can't make zero-sized pipe buffers by doing an empty
 357                  * write (not even in packet mode), but they can happen if
 358                  * the writer gets an EFAULT when trying to fill a buffer
 359                  * that already got allocated and inserted in the buffer
 360                  * array.
 361                  *
 362                  * So we still need to wake up any pending writers in the
 363                  * _very_ unlikely case that the pipe was full, but we got
 364                  * no data.
 365                  */
 366                 if (unlikely(was_full)) {
 367                         wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 368                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 369                 }
 370
 371                 /*
 372                  * But because we didn't read anything, at this point we can
 373                  * just return directly with -ERESTARTSYS if we're interrupted,
 374                  * since we've done any required wakeups and there's no need
 375                  * to mark anything accessed. And we've dropped the lock.
 376                  */
 377                 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 378                         return -ERESTARTSYS;
 379
 380                 __pipe_lock(pipe);
 381                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 382                 wake_next_reader = true;
 383         }
 384         if (pipe_empty(pipe->head, pipe->tail))
 385                 wake_next_reader = false;
 386         __pipe_unlock(pipe);
 387
 388         if (was_full) {
 389                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 390                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 391         }
 392         if (wake_next_reader)
 393                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 394         if (ret > 0)
 395                 file_accessed(filp);
 396         return ret;
 397 }
 398
 399 static inline int is_packetized(struct file *file)
 400 {
 401         return (file->f_flags & O_DIRECT) != 0;
 402 }
 403
 404 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 405 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 406 {
 407         unsigned int head = READ_ONCE(pipe->head);
 408         unsigned int tail = READ_ONCE(pipe->tail);
 409         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 410
 411         return !pipe_full(head, tail, max_usage) ||
 412                 !READ_ONCE(pipe->readers);
 413 }
 414
 415 static ssize_t
 416 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 417 {
 418         struct file *filp = iocb->ki_filp;
 419         struct pipe_inode_info *pipe = filp->private_data;
 420         unsigned int head;
 421         ssize_t ret = 0;
 422         size_t total_len = iov_iter_count(from);
 423         ssize_t chars;
 424         bool was_empty = false;
 425         bool wake_next_writer = false;
 426
 427         /* Null write succeeds. */
 428         if (unlikely(total_len == 0))
 429                 return 0;
 430
 431         __pipe_lock(pipe);
 432
 433         if (!pipe->readers) {
 434                 send_sig(SIGPIPE, current, 0);
 435                 ret = -EPIPE;
 436                 goto out;
 437         }
 438
 439 #ifdef CONFIG_WATCH_QUEUE
 440         if (pipe->watch_queue) {
 441                 ret = -EXDEV;
 442                 goto out;
 443         }
 444 #endif
 445
 446         /*
 447          * If it wasn't empty we try to merge new data into
 448          * the last buffer.
 449          *
 450          * That naturally merges small writes, but it also
 451          * page-aligns the rest of the writes for large writes
 452          * spanning multiple pages.
 453          */
 454         head = pipe->head;
 455         was_empty = pipe_empty(head, pipe->tail);
 456         chars = total_len & (PAGE_SIZE-1);
 457         if (chars && !was_empty) {
 458                 unsigned int mask = pipe->ring_size - 1;
 459                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 460                 int offset = buf->offset + buf->len;
 461
 462                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 463                     offset + chars <= PAGE_SIZE) {
 464                         ret = pipe_buf_confirm(pipe, buf);
 465                         if (ret)
 466                                 goto out;
 467
 468                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 469                         if (unlikely(ret < chars)) {
 470                                 ret = -EFAULT;
 471                                 goto out;
 472                         }
 473
 474                         buf->len += ret;
 475                         if (!iov_iter_count(from))
 476                                 goto out;
 477                 }
 478         }
 479
 480         for (;;) {
 481                 if (!pipe->readers) {
 482                         send_sig(SIGPIPE, current, 0);
 483                         if (!ret)
 484                                 ret = -EPIPE;
 485                         break;
 486                 }
 487
 488                 head = pipe->head;
 489                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 490                         unsigned int mask = pipe->ring_size - 1;
 491                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 492                         struct page *page = pipe->tmp_page;
 493                         int copied;
 494
 495                         if (!page) {
 496                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 497                                 if (unlikely(!page)) {
 498                                         ret = ret ? : -ENOMEM;
 499                                         break;
 500                                 }
 501                                 pipe->tmp_page = page;
 502                         }
 503
 504                         /* Allocate a slot in the ring in advance and attach an
 505                          * empty buffer.  If we fault or otherwise fail to use
 506                          * it, either the reader will consume it or it'll still
 507                          * be there for the next write.
 508                          */
 509                         spin_lock_irq(&pipe->rd_wait.lock);
 510
 511                         head = pipe->head;
 512                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 513                                 spin_unlock_irq(&pipe->rd_wait.lock);
 514                                 continue;
 515                         }
 516
 517                         pipe->head = head + 1;
 518                         spin_unlock_irq(&pipe->rd_wait.lock);
 519
 520                         /* Insert it into the buffer array */
 521                         buf = &pipe->bufs[head & mask];
 522                         buf->page = page;
 523                         buf->ops = &anon_pipe_buf_ops;
 524                         buf->offset = 0;
 525                         buf->len = 0;
 526                         if (is_packetized(filp))
 527                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 528                         else
 529                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 530                         pipe->tmp_page = NULL;
 531
 532                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 533                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 534                                 if (!ret)
 535                                         ret = -EFAULT;
 536                                 break;
 537                         }
 538                         ret += copied;
 539                         buf->offset = 0;
 540                         buf->len = copied;
 541
 542                         if (!iov_iter_count(from))
 543                                 break;
 544                 }
 545
 546                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 547                         continue;
 548
 549                 /* Wait for buffer space to become available. */
 550                 if (filp->f_flags & O_NONBLOCK) {
 551                         if (!ret)
 552                                 ret = -EAGAIN;
 553                         break;
 554                 }
 555                 if (signal_pending(current)) {
 556                         if (!ret)
 557                                 ret = -ERESTARTSYS;
 558                         break;
 559                 }
 560
 561                 /*
 562                  * We're going to release the pipe lock and wait for more
 563                  * space. We wake up any readers if necessary, and then
 564                  * after waiting we need to re-check whether the pipe
 565                  * become empty while we dropped the lock.
 566                  */
 567                 __pipe_unlock(pipe);
 568                 if (was_empty) {
 569                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 570                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 571                 }
 572                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 573                 __pipe_lock(pipe);
 574                 was_empty = pipe_empty(pipe->head, pipe->tail);
 575                 wake_next_writer = true;
 576         }
 577 out:
 578         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 579                 wake_next_writer = false;
 580         __pipe_unlock(pipe);
 581
 582         /*
 583          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 584          * want the reader to start processing things asap, rather than
 585          * leave the data pending.
 586          *
 587          * This is particularly important for small writes, because of
 588          * how (for example) the GNU make jobserver uses small writes to
 589          * wake up pending jobs
 590          *
 591          * Epoll nonsensically wants a wakeup whether the pipe
 592          * was already empty or not.
 593          */
 594         if (was_empty || pipe->poll_usage) {
 595                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 596                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 597         }
 598         if (wake_next_writer)
 599                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 600         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 601                 int err = file_update_time(filp);
 602                 if (err)
 603                         ret = err;
 604                 sb_end_write(file_inode(filp)->i_sb);
 605         }
 606         return ret;
 607 }
 608
 609 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 610 {
 611         struct pipe_inode_info *pipe = filp->private_data;
 612         int count, head, tail, mask;
 613
 614         switch (cmd) {
 615         case FIONREAD:
 616                 __pipe_lock(pipe);
 617                 count = 0;
 618                 head = pipe->head;
 619                 tail = pipe->tail;
 620                 mask = pipe->ring_size - 1;
 621
 622                 while (tail != head) {
 623                         count += pipe->bufs[tail & mask].len;
 624                         tail++;
 625                 }
 626                 __pipe_unlock(pipe);
 627
 628                 return put_user(count, (int __user *)arg);
 629
 630 #ifdef CONFIG_WATCH_QUEUE
 631         case IOC_WATCH_QUEUE_SET_SIZE: {
 632                 int ret;
 633                 __pipe_lock(pipe);
 634                 ret = watch_queue_set_size(pipe, arg);
 635                 __pipe_unlock(pipe);
 636                 return ret;
 637         }
 638
 639         case IOC_WATCH_QUEUE_SET_FILTER:
 640                 return watch_queue_set_filter(
 641                         pipe, (struct watch_notification_filter __user *)arg);
 642 #endif
 643
 644         default:
 645                 return -ENOIOCTLCMD;
 646         }
 647 }
 648
 649 /* No kernel lock held - fine */
 650 static __poll_t
 651 pipe_poll(struct file *filp, poll_table *wait)
 652 {
 653         __poll_t mask;
 654         struct pipe_inode_info *pipe = filp->private_data;
 655         unsigned int head, tail;
 656
 657         /* Epoll has some historical nasty semantics, this enables them */
 658         pipe->poll_usage = 1;
 659
 660         /*
 661          * Reading pipe state only -- no need for acquiring the semaphore.
 662          *
 663          * But because this is racy, the code has to add the
 664          * entry to the poll table _first_ ..
 665          */
 666         if (filp->f_mode & FMODE_READ)
 667                 poll_wait(filp, &pipe->rd_wait, wait);
 668         if (filp->f_mode & FMODE_WRITE)
 669                 poll_wait(filp, &pipe->wr_wait, wait);
 670
 671         /*
 672          * .. and only then can you do the racy tests. That way,
 673          * if something changes and you got it wrong, the poll
 674          * table entry will wake you up and fix it.
 675          */
 676         head = READ_ONCE(pipe->head);
 677         tail = READ_ONCE(pipe->tail);
 678
 679         mask = 0;
 680         if (filp->f_mode & FMODE_READ) {
 681                 if (!pipe_empty(head, tail))
 682                         mask |= EPOLLIN | EPOLLRDNORM;
 683                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 684                         mask |= EPOLLHUP;
 685         }
 686
 687         if (filp->f_mode & FMODE_WRITE) {
 688                 if (!pipe_full(head, tail, pipe->max_usage))
 689                         mask |= EPOLLOUT | EPOLLWRNORM;
 690                 /*
 691                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 692                  * behave exactly like pipes for poll().
 693                  */
 694                 if (!pipe->readers)
 695                         mask |= EPOLLERR;
 696         }
 697
 698         return mask;
 699 }
 700
 701 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 702 {
 703         int kill = 0;
 704
 705         spin_lock(&inode->i_lock);
 706         if (!--pipe->files) {
 707                 inode->i_pipe = NULL;
 708                 kill = 1;
 709         }
 710         spin_unlock(&inode->i_lock);
 711
 712         if (kill)
 713                 free_pipe_info(pipe);
 714 }
 715
 716 static int
 717 pipe_release(struct inode *inode, struct file *file)
 718 {
 719         struct pipe_inode_info *pipe = file->private_data;
 720
 721         __pipe_lock(pipe);
 722         if (file->f_mode & FMODE_READ)
 723                 pipe->readers--;
 724         if (file->f_mode & FMODE_WRITE)
 725                 pipe->writers--;
 726
 727         /* Was that the last reader or writer, but not the other side? */
 728         if (!pipe->readers != !pipe->writers) {
 729                 wake_up_interruptible_all(&pipe->rd_wait);
 730                 wake_up_interruptible_all(&pipe->wr_wait);
 731                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 732                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 733         }
 734         __pipe_unlock(pipe);
 735
 736         put_pipe_info(inode, pipe);
 737         return 0;
 738 }
 739
 740 static int
 741 pipe_fasync(int fd, struct file *filp, int on)
 742 {
 743         struct pipe_inode_info *pipe = filp->private_data;
 744         int retval = 0;
 745
 746         __pipe_lock(pipe);
 747         if (filp->f_mode & FMODE_READ)
 748                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 749         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 750                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 751                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 752                         /* this can happen only if on == T */
 753                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 754         }
 755         __pipe_unlock(pipe);
 756         return retval;
 757 }
 758
 759 unsigned long account_pipe_buffers(struct user_struct *user,
 760                                    unsigned long old, unsigned long new)
 761 {
 762         return atomic_long_add_return(new - old, &user->pipe_bufs);
 763 }
 764
 765 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 766 {
 767         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 768
 769         return soft_limit && user_bufs > soft_limit;
 770 }
 771
 772 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 773 {
 774         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 775
 776         return hard_limit && user_bufs > hard_limit;
 777 }
 778
 779 bool pipe_is_unprivileged_user(void)
 780 {
 781         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 782 }
 783
 784 struct pipe_inode_info *alloc_pipe_info(void)
 785 {
 786         struct pipe_inode_info *pipe;
 787         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 788         struct user_struct *user = get_current_user();
 789         unsigned long user_bufs;
 790         unsigned int max_size = READ_ONCE(pipe_max_size);
 791
 792         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 793         if (pipe == NULL)
 794                 goto out_free_uid;
 795
 796         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 797                 pipe_bufs = max_size >> PAGE_SHIFT;
 798
 799         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 800
 801         if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
 802                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 803                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 804         }
 805
 806         if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
 807                 goto out_revert_acct;
 808
 809         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 810                              GFP_KERNEL_ACCOUNT);
 811
 812         if (pipe->bufs) {
 813                 init_waitqueue_head(&pipe->rd_wait);
 814                 init_waitqueue_head(&pipe->wr_wait);
 815                 pipe->r_counter = pipe->w_counter = 1;
 816                 pipe->max_usage = pipe_bufs;
 817                 pipe->ring_size = pipe_bufs;
 818                 pipe->nr_accounted = pipe_bufs;
 819                 pipe->user = user;
 820                 mutex_init(&pipe->mutex);
 821                 return pipe;
 822         }
 823
 824 out_revert_acct:
 825         (void) account_pipe_buffers(user, pipe_bufs, 0);
 826         kfree(pipe);
 827 out_free_uid:
 828         free_uid(user);
 829         return NULL;
 830 }
 831
 832 void free_pipe_info(struct pipe_inode_info *pipe)
 833 {
 834         int i;
 835
 836 #ifdef CONFIG_WATCH_QUEUE
 837         if (pipe->watch_queue) {
 838                 watch_queue_clear(pipe->watch_queue);
 839                 put_watch_queue(pipe->watch_queue);
 840         }
 841 #endif
 842
 843         (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 844         free_uid(pipe->user);
 845         for (i = 0; i < pipe->ring_size; i++) {
 846                 struct pipe_buffer *buf = pipe->bufs + i;
 847                 if (buf->ops)
 848                         pipe_buf_release(pipe, buf);
 849         }
 850         if (pipe->tmp_page)
 851                 __free_page(pipe->tmp_page);
 852         kfree(pipe->bufs);
 853         kfree(pipe);
 854 }
 855
 856 static struct vfsmount *pipe_mnt __read_mostly;
 857
 858 /*
 859  * pipefs_dname() is called from d_path().
 860  */
 861 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 862 {
 863         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 864                                 d_inode(dentry)->i_ino);
 865 }
 866
 867 static const struct dentry_operations pipefs_dentry_operations = {
 868         .d_dname        = pipefs_dname,
 869 };
 870
 871 static struct inode * get_pipe_inode(void)
 872 {
 873         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 874         struct pipe_inode_info *pipe;
 875
 876         if (!inode)
 877                 goto fail_inode;
 878
 879         inode->i_ino = get_next_ino();
 880
 881         pipe = alloc_pipe_info();
 882         if (!pipe)
 883                 goto fail_iput;
 884
 885         inode->i_pipe = pipe;
 886         pipe->files = 2;
 887         pipe->readers = pipe->writers = 1;
 888         inode->i_fop = &pipefifo_fops;
 889
 890         /*
 891          * Mark the inode dirty from the very beginning,
 892          * that way it will never be moved to the dirty
 893          * list because "mark_inode_dirty()" will think
 894          * that it already _is_ on the dirty list.
 895          */
 896         inode->i_state = I_DIRTY;
 897         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 898         inode->i_uid = current_fsuid();
 899         inode->i_gid = current_fsgid();
 900         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 901
 902         return inode;
 903
 904 fail_iput:
 905         iput(inode);
 906
 907 fail_inode:
 908         return NULL;
 909 }
 910
 911 int create_pipe_files(struct file **res, int flags)
 912 {
 913         struct inode *inode = get_pipe_inode();
 914         struct file *f;
 915         int error;
 916
 917         if (!inode)
 918                 return -ENFILE;
 919
 920         if (flags & O_NOTIFICATION_PIPE) {
 921                 error = watch_queue_init(inode->i_pipe);
 922                 if (error) {
 923                         free_pipe_info(inode->i_pipe);
 924                         iput(inode);
 925                         return error;
 926                 }
 927         }
 928
 929         f = alloc_file_pseudo(inode, pipe_mnt, "",
 930                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 931                                 &pipefifo_fops);
 932         if (IS_ERR(f)) {
 933                 free_pipe_info(inode->i_pipe);
 934                 iput(inode);
 935                 return PTR_ERR(f);
 936         }
 937
 938         f->private_data = inode->i_pipe;
 939
 940         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 941                                   &pipefifo_fops);
 942         if (IS_ERR(res[0])) {
 943                 put_pipe_info(inode, inode->i_pipe);
 944                 fput(f);
 945                 return PTR_ERR(res[0]);
 946         }
 947         res[0]->private_data = inode->i_pipe;
 948         res[1] = f;
 949         stream_open(inode, res[0]);
 950         stream_open(inode, res[1]);
 951         return 0;
 952 }
 953
 954 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 955 {
 956         int error;
 957         int fdw, fdr;
 958
 959         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
 960                 return -EINVAL;
 961
 962         error = create_pipe_files(files, flags);
 963         if (error)
 964                 return error;
 965
 966         error = get_unused_fd_flags(flags);
 967         if (error < 0)
 968                 goto err_read_pipe;
 969         fdr = error;
 970
 971         error = get_unused_fd_flags(flags);
 972         if (error < 0)
 973                 goto err_fdr;
 974         fdw = error;
 975
 976         audit_fd_pair(fdr, fdw);
 977         fd[0] = fdr;
 978         fd[1] = fdw;
 979         return 0;
 980
 981  err_fdr:
 982         put_unused_fd(fdr);
 983  err_read_pipe:
 984         fput(files[0]);
 985         fput(files[1]);
 986         return error;
 987 }
 988
 989 int do_pipe_flags(int *fd, int flags)
 990 {
 991         struct file *files[2];
 992         int error = __do_pipe_flags(fd, files, flags);
 993         if (!error) {
 994                 fd_install(fd[0], files[0]);
 995                 fd_install(fd[1], files[1]);
 996         }
 997         return error;
 998 }
 999
1000 /*
1001  * sys_pipe() is the normal C calling standard for creating
1002  * a pipe. It's not the way Unix traditionally does this, though.
1003  */
1004 static int do_pipe2(int __user *fildes, int flags)
1005 {
1006         struct file *files[2];
1007         int fd[2];
1008         int error;
1009
1010         error = __do_pipe_flags(fd, files, flags);
1011         if (!error) {
1012                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1013                         fput(files[0]);
1014                         fput(files[1]);
1015                         put_unused_fd(fd[0]);
1016                         put_unused_fd(fd[1]);
1017                         error = -EFAULT;
1018                 } else {
1019                         fd_install(fd[0], files[0]);
1020                         fd_install(fd[1], files[1]);
1021                 }
1022         }
1023         return error;
1024 }
1025
1026 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1027 {
1028         return do_pipe2(fildes, flags);
1029 }
1030
1031 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1032 {
1033         return do_pipe2(fildes, 0);
1034 }
1035
1036 /*
1037  * This is the stupid "wait for pipe to be readable or writable"
1038  * model.
1039  *
1040  * See pipe_read/write() for the proper kind of exclusive wait,
1041  * but that requires that we wake up any other readers/writers
1042  * if we then do not end up reading everything (ie the whole
1043  * "wake_next_reader/writer" logic in pipe_read/write()).
1044  */
1045 void pipe_wait_readable(struct pipe_inode_info *pipe)
1046 {
1047         pipe_unlock(pipe);
1048         wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1049         pipe_lock(pipe);
1050 }
1051
1052 void pipe_wait_writable(struct pipe_inode_info *pipe)
1053 {
1054         pipe_unlock(pipe);
1055         wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1056         pipe_lock(pipe);
1057 }
1058
1059 /*
1060  * This depends on both the wait (here) and the wakeup (wake_up_partner)
1061  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1062  * race with the count check and waitqueue prep.
1063  *
1064  * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1065  * then check the condition you're waiting for, and only then sleep. But
1066  * because of the pipe lock, we can check the condition before being on
1067  * the wait queue.
1068  *
1069  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1070  */
1071 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1072 {
1073         DEFINE_WAIT(rdwait);
1074         int cur = *cnt;
1075
1076         while (cur == *cnt) {
1077                 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1078                 pipe_unlock(pipe);
1079                 schedule();
1080                 finish_wait(&pipe->rd_wait, &rdwait);
1081                 pipe_lock(pipe);
1082                 if (signal_pending(current))
1083                         break;
1084         }
1085         return cur == *cnt ? -ERESTARTSYS : 0;
1086 }
1087
1088 static void wake_up_partner(struct pipe_inode_info *pipe)
1089 {
1090         wake_up_interruptible_all(&pipe->rd_wait);
1091 }
1092
1093 static int fifo_open(struct inode *inode, struct file *filp)
1094 {
1095         struct pipe_inode_info *pipe;
1096         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1097         int ret;
1098
1099         filp->f_version = 0;
1100
1101         spin_lock(&inode->i_lock);
1102         if (inode->i_pipe) {
1103                 pipe = inode->i_pipe;
1104                 pipe->files++;
1105                 spin_unlock(&inode->i_lock);
1106         } else {
1107                 spin_unlock(&inode->i_lock);
1108                 pipe = alloc_pipe_info();
1109                 if (!pipe)
1110                         return -ENOMEM;
1111                 pipe->files = 1;
1112                 spin_lock(&inode->i_lock);
1113                 if (unlikely(inode->i_pipe)) {
1114                         inode->i_pipe->files++;
1115                         spin_unlock(&inode->i_lock);
1116                         free_pipe_info(pipe);
1117                         pipe = inode->i_pipe;
1118                 } else {
1119                         inode->i_pipe = pipe;
1120                         spin_unlock(&inode->i_lock);
1121                 }
1122         }
1123         filp->private_data = pipe;
1124         /* OK, we have a pipe and it's pinned down */
1125
1126         __pipe_lock(pipe);
1127
1128         /* We can only do regular read/write on fifos */
1129         stream_open(inode, filp);
1130
1131         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1132         case FMODE_READ:
1133         /*
1134          *  O_RDONLY
1135          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1136          *  opened, even when there is no process writing the FIFO.
1137          */
1138                 pipe->r_counter++;
1139                 if (pipe->readers++ == 0)
1140                         wake_up_partner(pipe);
1141
1142                 if (!is_pipe && !pipe->writers) {
1143                         if ((filp->f_flags & O_NONBLOCK)) {
1144                                 /* suppress EPOLLHUP until we have
1145                                  * seen a writer */
1146                                 filp->f_version = pipe->w_counter;
1147                         } else {
1148                                 if (wait_for_partner(pipe, &pipe->w_counter))
1149                                         goto err_rd;
1150                         }
1151                 }
1152                 break;
1153
1154         case FMODE_WRITE:
1155         /*
1156          *  O_WRONLY
1157          *  POSIX.1 says that O_NONBLOCK means return -1 with
1158          *  errno=ENXIO when there is no process reading the FIFO.
1159          */
1160                 ret = -ENXIO;
1161                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1162                         goto err;
1163
1164                 pipe->w_counter++;
1165                 if (!pipe->writers++)
1166                         wake_up_partner(pipe);
1167
1168                 if (!is_pipe && !pipe->readers) {
1169                         if (wait_for_partner(pipe, &pipe->r_counter))
1170                                 goto err_wr;
1171                 }
1172                 break;
1173
1174         case FMODE_READ | FMODE_WRITE:
1175         /*
1176          *  O_RDWR
1177          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1178          *  This implementation will NEVER block on a O_RDWR open, since
1179          *  the process can at least talk to itself.
1180          */
1181
1182                 pipe->readers++;
1183                 pipe->writers++;
1184                 pipe->r_counter++;
1185                 pipe->w_counter++;
1186                 if (pipe->readers == 1 || pipe->writers == 1)
1187                         wake_up_partner(pipe);
1188                 break;
1189
1190         default:
1191                 ret = -EINVAL;
1192                 goto err;
1193         }
1194
1195         /* Ok! */
1196         __pipe_unlock(pipe);
1197         return 0;
1198
1199 err_rd:
1200         if (!--pipe->readers)
1201                 wake_up_interruptible(&pipe->wr_wait);
1202         ret = -ERESTARTSYS;
1203         goto err;
1204
1205 err_wr:
1206         if (!--pipe->writers)
1207                 wake_up_interruptible_all(&pipe->rd_wait);
1208         ret = -ERESTARTSYS;
1209         goto err;
1210
1211 err:
1212         __pipe_unlock(pipe);
1213
1214         put_pipe_info(inode, pipe);
1215         return ret;
1216 }
1217
1218 const struct file_operations pipefifo_fops = {
1219         .open           = fifo_open,
1220         .llseek         = no_llseek,
1221         .read_iter      = pipe_read,
1222         .write_iter     = pipe_write,
1223         .poll           = pipe_poll,
1224         .unlocked_ioctl = pipe_ioctl,
1225         .release        = pipe_release,
1226         .fasync         = pipe_fasync,
1227         .splice_write   = iter_file_splice_write,
1228 };
1229
1230 /*
1231  * Currently we rely on the pipe array holding a power-of-2 number
1232  * of pages. Returns 0 on error.
1233  */
1234 unsigned int round_pipe_size(unsigned long size)
1235 {
1236         if (size > (1U << 31))
1237                 return 0;
1238
1239         /* Minimum pipe size, as required by POSIX */
1240         if (size < PAGE_SIZE)
1241                 return PAGE_SIZE;
1242
1243         return roundup_pow_of_two(size);
1244 }
1245
1246 /*
1247  * Resize the pipe ring to a number of slots.
1248  */
1249 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1250 {
1251         struct pipe_buffer *bufs;
1252         unsigned int head, tail, mask, n;
1253
1254         /*
1255          * We can shrink the pipe, if arg is greater than the ring occupancy.
1256          * Since we don't expect a lot of shrink+grow operations, just free and
1257          * allocate again like we would do for growing.  If the pipe currently
1258          * contains more buffers than arg, then return busy.
1259          */
1260         mask = pipe->ring_size - 1;
1261         head = pipe->head;
1262         tail = pipe->tail;
1263         n = pipe_occupancy(pipe->head, pipe->tail);
1264         if (nr_slots < n)
1265                 return -EBUSY;
1266
1267         bufs = kcalloc(nr_slots, sizeof(*bufs),
1268                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1269         if (unlikely(!bufs))
1270                 return -ENOMEM;
1271
1272         /*
1273          * The pipe array wraps around, so just start the new one at zero
1274          * and adjust the indices.
1275          */
1276         if (n > 0) {
1277                 unsigned int h = head & mask;
1278                 unsigned int t = tail & mask;
1279                 if (h > t) {
1280                         memcpy(bufs, pipe->bufs + t,
1281                                n * sizeof(struct pipe_buffer));
1282                 } else {
1283                         unsigned int tsize = pipe->ring_size - t;
1284                         if (h > 0)
1285                                 memcpy(bufs + tsize, pipe->bufs,
1286                                        h * sizeof(struct pipe_buffer));
1287                         memcpy(bufs, pipe->bufs + t,
1288                                tsize * sizeof(struct pipe_buffer));
1289                 }
1290         }
1291
1292         head = n;
1293         tail = 0;
1294
1295         kfree(pipe->bufs);
1296         pipe->bufs = bufs;
1297         pipe->ring_size = nr_slots;
1298         if (pipe->max_usage > nr_slots)
1299                 pipe->max_usage = nr_slots;
1300         pipe->tail = tail;
1301         pipe->head = head;
1302
1303         /* This might have made more room for writers */
1304         wake_up_interruptible(&pipe->wr_wait);
1305         return 0;
1306 }
1307
1308 /*
1309  * Allocate a new array of pipe buffers and copy the info over. Returns the
1310  * pipe size if successful, or return -ERROR on error.
1311  */
1312 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1313 {
1314         unsigned long user_bufs;
1315         unsigned int nr_slots, size;
1316         long ret = 0;
1317
1318 #ifdef CONFIG_WATCH_QUEUE
1319         if (pipe->watch_queue)
1320                 return -EBUSY;
1321 #endif
1322
1323         size = round_pipe_size(arg);
1324         nr_slots = size >> PAGE_SHIFT;
1325
1326         if (!nr_slots)
1327                 return -EINVAL;
1328
1329         /*
1330          * If trying to increase the pipe capacity, check that an
1331          * unprivileged user is not trying to exceed various limits
1332          * (soft limit check here, hard limit check just below).
1333          * Decreasing the pipe capacity is always permitted, even
1334          * if the user is currently over a limit.
1335          */
1336         if (nr_slots > pipe->max_usage &&
1337                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1338                 return -EPERM;
1339
1340         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1341
1342         if (nr_slots > pipe->max_usage &&
1343                         (too_many_pipe_buffers_hard(user_bufs) ||
1344                          too_many_pipe_buffers_soft(user_bufs)) &&
1345                         pipe_is_unprivileged_user()) {
1346                 ret = -EPERM;
1347                 goto out_revert_acct;
1348         }
1349
1350         ret = pipe_resize_ring(pipe, nr_slots);
1351         if (ret < 0)
1352                 goto out_revert_acct;
1353
1354         pipe->max_usage = nr_slots;
1355         pipe->nr_accounted = nr_slots;
1356         return pipe->max_usage * PAGE_SIZE;
1357
1358 out_revert_acct:
1359         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1360         return ret;
1361 }
1362
1363 /*
1364  * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1365  * not enough to verify that this is a pipe.
1366  */
1367 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1368 {
1369         struct pipe_inode_info *pipe = file->private_data;
1370
1371         if (file->f_op != &pipefifo_fops || !pipe)
1372                 return NULL;
1373 #ifdef CONFIG_WATCH_QUEUE
1374         if (for_splice && pipe->watch_queue)
1375                 return NULL;
1376 #endif
1377         return pipe;
1378 }
1379
1380 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1381 {
1382         struct pipe_inode_info *pipe;
1383         long ret;
1384
1385         pipe = get_pipe_info(file, false);
1386         if (!pipe)
1387                 return -EBADF;
1388
1389         __pipe_lock(pipe);
1390
1391         switch (cmd) {
1392         case F_SETPIPE_SZ:
1393                 ret = pipe_set_size(pipe, arg);
1394                 break;
1395         case F_GETPIPE_SZ:
1396                 ret = pipe->max_usage * PAGE_SIZE;
1397                 break;
1398         default:
1399                 ret = -EINVAL;
1400                 break;
1401         }
1402
1403         __pipe_unlock(pipe);
1404         return ret;
1405 }
1406
1407 static const struct super_operations pipefs_ops = {
1408         .destroy_inode = free_inode_nonrcu,
1409         .statfs = simple_statfs,
1410 };
1411
1412 /*
1413  * pipefs should _never_ be mounted by userland - too much of security hassle,
1414  * no real gain from having the whole whorehouse mounted. So we don't need
1415  * any operations on the root directory. However, we need a non-trivial
1416  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1417  */
1418
1419 static int pipefs_init_fs_context(struct fs_context *fc)
1420 {
1421         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1422         if (!ctx)
1423                 return -ENOMEM;
1424         ctx->ops = &pipefs_ops;
1425         ctx->dops = &pipefs_dentry_operations;
1426         return 0;
1427 }
1428
1429 static struct file_system_type pipe_fs_type = {
1430         .name           = "pipefs",
1431         .init_fs_context = pipefs_init_fs_context,
1432         .kill_sb        = kill_anon_super,
1433 };
1434
1435 static int __init init_pipe_fs(void)
1436 {
1437         int err = register_filesystem(&pipe_fs_type);
1438
1439         if (!err) {
1440                 pipe_mnt = kern_mount(&pipe_fs_type);
1441                 if (IS_ERR(pipe_mnt)) {
1442                         err = PTR_ERR(pipe_mnt);
1443                         unregister_filesystem(&pipe_fs_type);
1444                 }
1445         }
1446         return err;
1447 }
1448
1449 fs_initcall(init_pipe_fs);