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 #include <linux/sysctl.h>
  29
  30 #include <linux/uaccess.h>
  31 #include <asm/ioctls.h>
  32
  33 #include "internal.h"
  34
  35 /*
  36  * New pipe buffers will be restricted to this size while the user is exceeding
  37  * their pipe buffer quota. The general pipe use case needs at least two
  38  * buffers: one for data yet to be read, and one for new data. If this is less
  39  * than two, then a write to a non-empty pipe may block even if the pipe is not
  40  * full. This can occur with GNU make jobserver or similar uses of pipes as
  41  * semaphores: multiple processes may be waiting to write tokens back to the
  42  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  43  *
  44  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  45  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  46  * emptied.
  47  */
  48 #define PIPE_MIN_DEF_BUFFERS 2
  49
  50 /*
  51  * The max size that a non-root user is allowed to grow the pipe. Can
  52  * be set by root in /proc/sys/fs/pipe-max-size
  53  */
  54 static unsigned int pipe_max_size = 1048576;
  55
  56 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  57  * matches default values.
  58  */
  59 static unsigned long pipe_user_pages_hard;
  60 static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  61
  62 /*
  63  * We use head and tail indices that aren't masked off, except at the point of
  64  * dereference, but rather they're allowed to wrap naturally.  This means there
  65  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  66  * <= 2^31.
  67  * -- David Howells 2019-09-23.
  68  *
  69  * Reads with count = 0 should always return 0.
  70  * -- Julian Bradfield 1999-06-07.
  71  *
  72  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  73  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  74  *
  75  * pipe_read & write cleanup
  76  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  77  */
  78
  79 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  80 {
  81         if (pipe->files)
  82                 mutex_lock_nested(&pipe->mutex, subclass);
  83 }
  84
  85 void pipe_lock(struct pipe_inode_info *pipe)
  86 {
  87         /*
  88          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  89          */
  90         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  91 }
  92 EXPORT_SYMBOL(pipe_lock);
  93
  94 void pipe_unlock(struct pipe_inode_info *pipe)
  95 {
  96         if (pipe->files)
  97                 mutex_unlock(&pipe->mutex);
  98 }
  99 EXPORT_SYMBOL(pipe_unlock);
 100
 101 static inline void __pipe_lock(struct pipe_inode_info *pipe)
 102 {
 103         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 104 }
 105
 106 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 107 {
 108         mutex_unlock(&pipe->mutex);
 109 }
 110
 111 void pipe_double_lock(struct pipe_inode_info *pipe1,
 112                       struct pipe_inode_info *pipe2)
 113 {
 114         BUG_ON(pipe1 == pipe2);
 115
 116         if (pipe1 < pipe2) {
 117                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 118                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 119         } else {
 120                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 121                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 122         }
 123 }
 124
 125 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 126                                   struct pipe_buffer *buf)
 127 {
 128         struct page *page = buf->page;
 129
 130         /*
 131          * If nobody else uses this page, and we don't already have a
 132          * temporary page, let's keep track of it as a one-deep
 133          * allocation cache. (Otherwise just release our reference to it)
 134          */
 135         if (page_count(page) == 1 && !pipe->tmp_page)
 136                 pipe->tmp_page = page;
 137         else
 138                 put_page(page);
 139 }
 140
 141 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 142                 struct pipe_buffer *buf)
 143 {
 144         struct page *page = buf->page;
 145
 146         if (page_count(page) != 1)
 147                 return false;
 148         memcg_kmem_uncharge_page(page, 0);
 149         __SetPageLocked(page);
 150         return true;
 151 }
 152
 153 /**
 154  * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
 155  * @pipe:       the pipe that the buffer belongs to
 156  * @buf:        the buffer to attempt to steal
 157  *
 158  * Description:
 159  *      This function attempts to steal the &struct page attached to
 160  *      @buf. If successful, this function returns 0 and returns with
 161  *      the page locked. The caller may then reuse the page for whatever
 162  *      he wishes; the typical use is insertion into a different file
 163  *      page cache.
 164  */
 165 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 166                 struct pipe_buffer *buf)
 167 {
 168         struct page *page = buf->page;
 169
 170         /*
 171          * A reference of one is golden, that means that the owner of this
 172          * page is the only one holding a reference to it. lock the page
 173          * and return OK.
 174          */
 175         if (page_count(page) == 1) {
 176                 lock_page(page);
 177                 return true;
 178         }
 179         return false;
 180 }
 181 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
 182
 183 /**
 184  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 185  * @pipe:       the pipe that the buffer belongs to
 186  * @buf:        the buffer to get a reference to
 187  *
 188  * Description:
 189  *      This function grabs an extra reference to @buf. It's used in
 190  *      the tee() system call, when we duplicate the buffers in one
 191  *      pipe into another.
 192  */
 193 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 194 {
 195         return try_get_page(buf->page);
 196 }
 197 EXPORT_SYMBOL(generic_pipe_buf_get);
 198
 199 /**
 200  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 201  * @pipe:       the pipe that the buffer belongs to
 202  * @buf:        the buffer to put a reference to
 203  *
 204  * Description:
 205  *      This function releases a reference to @buf.
 206  */
 207 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 208                               struct pipe_buffer *buf)
 209 {
 210         put_page(buf->page);
 211 }
 212 EXPORT_SYMBOL(generic_pipe_buf_release);
 213
 214 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 215         .release        = anon_pipe_buf_release,
 216         .try_steal      = anon_pipe_buf_try_steal,
 217         .get            = generic_pipe_buf_get,
 218 };
 219
 220 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 221 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 222 {
 223         unsigned int head = READ_ONCE(pipe->head);
 224         unsigned int tail = READ_ONCE(pipe->tail);
 225         unsigned int writers = READ_ONCE(pipe->writers);
 226
 227         return !pipe_empty(head, tail) || !writers;
 228 }
 229
 230 static ssize_t
 231 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 232 {
 233         size_t total_len = iov_iter_count(to);
 234         struct file *filp = iocb->ki_filp;
 235         struct pipe_inode_info *pipe = filp->private_data;
 236         bool was_full, wake_next_reader = false;
 237         ssize_t ret;
 238
 239         /* Null read succeeds. */
 240         if (unlikely(total_len == 0))
 241                 return 0;
 242
 243         ret = 0;
 244         __pipe_lock(pipe);
 245
 246         /*
 247          * We only wake up writers if the pipe was full when we started
 248          * reading in order to avoid unnecessary wakeups.
 249          *
 250          * But when we do wake up writers, we do so using a sync wakeup
 251          * (WF_SYNC), because we want them to get going and generate more
 252          * data for us.
 253          */
 254         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 255         for (;;) {
 256                 unsigned int head = pipe->head;
 257                 unsigned int tail = pipe->tail;
 258                 unsigned int mask = pipe->ring_size - 1;
 259
 260 #ifdef CONFIG_WATCH_QUEUE
 261                 if (pipe->note_loss) {
 262                         struct watch_notification n;
 263
 264                         if (total_len < 8) {
 265                                 if (ret == 0)
 266                                         ret = -ENOBUFS;
 267                                 break;
 268                         }
 269
 270                         n.type = WATCH_TYPE_META;
 271                         n.subtype = WATCH_META_LOSS_NOTIFICATION;
 272                         n.info = watch_sizeof(n);
 273                         if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
 274                                 if (ret == 0)
 275                                         ret = -EFAULT;
 276                                 break;
 277                         }
 278                         ret += sizeof(n);
 279                         total_len -= sizeof(n);
 280                         pipe->note_loss = false;
 281                 }
 282 #endif
 283
 284                 if (!pipe_empty(head, tail)) {
 285                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 286                         size_t chars = buf->len;
 287                         size_t written;
 288                         int error;
 289
 290                         if (chars > total_len) {
 291                                 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
 292                                         if (ret == 0)
 293                                                 ret = -ENOBUFS;
 294                                         break;
 295                                 }
 296                                 chars = total_len;
 297                         }
 298
 299                         error = pipe_buf_confirm(pipe, buf);
 300                         if (error) {
 301                                 if (!ret)
 302                                         ret = error;
 303                                 break;
 304                         }
 305
 306                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 307                         if (unlikely(written < chars)) {
 308                                 if (!ret)
 309                                         ret = -EFAULT;
 310                                 break;
 311                         }
 312                         ret += chars;
 313                         buf->offset += chars;
 314                         buf->len -= chars;
 315
 316                         /* Was it a packet buffer? Clean up and exit */
 317                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 318                                 total_len = chars;
 319                                 buf->len = 0;
 320                         }
 321
 322                         if (!buf->len) {
 323                                 pipe_buf_release(pipe, buf);
 324                                 spin_lock_irq(&pipe->rd_wait.lock);
 325 #ifdef CONFIG_WATCH_QUEUE
 326                                 if (buf->flags & PIPE_BUF_FLAG_LOSS)
 327                                         pipe->note_loss = true;
 328 #endif
 329                                 tail++;
 330                                 pipe->tail = tail;
 331                                 spin_unlock_irq(&pipe->rd_wait.lock);
 332                         }
 333                         total_len -= chars;
 334                         if (!total_len)
 335                                 break;  /* common path: read succeeded */
 336                         if (!pipe_empty(head, tail))    /* More to do? */
 337                                 continue;
 338                 }
 339
 340                 if (!pipe->writers)
 341                         break;
 342                 if (ret)
 343                         break;
 344                 if (filp->f_flags & O_NONBLOCK) {
 345                         ret = -EAGAIN;
 346                         break;
 347                 }
 348                 __pipe_unlock(pipe);
 349
 350                 /*
 351                  * We only get here if we didn't actually read anything.
 352                  *
 353                  * However, we could have seen (and removed) a zero-sized
 354                  * pipe buffer, and might have made space in the buffers
 355                  * that way.
 356                  *
 357                  * You can't make zero-sized pipe buffers by doing an empty
 358                  * write (not even in packet mode), but they can happen if
 359                  * the writer gets an EFAULT when trying to fill a buffer
 360                  * that already got allocated and inserted in the buffer
 361                  * array.
 362                  *
 363                  * So we still need to wake up any pending writers in the
 364                  * _very_ unlikely case that the pipe was full, but we got
 365                  * no data.
 366                  */
 367                 if (unlikely(was_full))
 368                         wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 369                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 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         if (wake_next_reader)
 391                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 392         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 393         if (ret > 0)
 394                 file_accessed(filp);
 395         return ret;
 396 }
 397
 398 static inline int is_packetized(struct file *file)
 399 {
 400         return (file->f_flags & O_DIRECT) != 0;
 401 }
 402
 403 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 404 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 405 {
 406         unsigned int head = READ_ONCE(pipe->head);
 407         unsigned int tail = READ_ONCE(pipe->tail);
 408         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 409
 410         return !pipe_full(head, tail, max_usage) ||
 411                 !READ_ONCE(pipe->readers);
 412 }
 413
 414 static ssize_t
 415 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 416 {
 417         struct file *filp = iocb->ki_filp;
 418         struct pipe_inode_info *pipe = filp->private_data;
 419         unsigned int head;
 420         ssize_t ret = 0;
 421         size_t total_len = iov_iter_count(from);
 422         ssize_t chars;
 423         bool was_empty = false;
 424         bool wake_next_writer = false;
 425
 426         /* Null write succeeds. */
 427         if (unlikely(total_len == 0))
 428                 return 0;
 429
 430         __pipe_lock(pipe);
 431
 432         if (!pipe->readers) {
 433                 send_sig(SIGPIPE, current, 0);
 434                 ret = -EPIPE;
 435                 goto out;
 436         }
 437
 438 #ifdef CONFIG_WATCH_QUEUE
 439         if (pipe->watch_queue) {
 440                 ret = -EXDEV;
 441                 goto out;
 442         }
 443 #endif
 444
 445         /*
 446          * If it wasn't empty we try to merge new data into
 447          * the last buffer.
 448          *
 449          * That naturally merges small writes, but it also
 450          * page-aligns the rest of the writes for large writes
 451          * spanning multiple pages.
 452          */
 453         head = pipe->head;
 454         was_empty = pipe_empty(head, pipe->tail);
 455         chars = total_len & (PAGE_SIZE-1);
 456         if (chars && !was_empty) {
 457                 unsigned int mask = pipe->ring_size - 1;
 458                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 459                 int offset = buf->offset + buf->len;
 460
 461                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 462                     offset + chars <= PAGE_SIZE) {
 463                         ret = pipe_buf_confirm(pipe, buf);
 464                         if (ret)
 465                                 goto out;
 466
 467                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 468                         if (unlikely(ret < chars)) {
 469                                 ret = -EFAULT;
 470                                 goto out;
 471                         }
 472
 473                         buf->len += ret;
 474                         if (!iov_iter_count(from))
 475                                 goto out;
 476                 }
 477         }
 478
 479         for (;;) {
 480                 if (!pipe->readers) {
 481                         send_sig(SIGPIPE, current, 0);
 482                         if (!ret)
 483                                 ret = -EPIPE;
 484                         break;
 485                 }
 486
 487                 head = pipe->head;
 488                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 489                         unsigned int mask = pipe->ring_size - 1;
 490                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 491                         struct page *page = pipe->tmp_page;
 492                         int copied;
 493
 494                         if (!page) {
 495                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 496                                 if (unlikely(!page)) {
 497                                         ret = ret ? : -ENOMEM;
 498                                         break;
 499                                 }
 500                                 pipe->tmp_page = page;
 501                         }
 502
 503                         /* Allocate a slot in the ring in advance and attach an
 504                          * empty buffer.  If we fault or otherwise fail to use
 505                          * it, either the reader will consume it or it'll still
 506                          * be there for the next write.
 507                          */
 508                         spin_lock_irq(&pipe->rd_wait.lock);
 509
 510                         head = pipe->head;
 511                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 512                                 spin_unlock_irq(&pipe->rd_wait.lock);
 513                                 continue;
 514                         }
 515
 516                         pipe->head = head + 1;
 517                         spin_unlock_irq(&pipe->rd_wait.lock);
 518
 519                         /* Insert it into the buffer array */
 520                         buf = &pipe->bufs[head & mask];
 521                         buf->page = page;
 522                         buf->ops = &anon_pipe_buf_ops;
 523                         buf->offset = 0;
 524                         buf->len = 0;
 525                         if (is_packetized(filp))
 526                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 527                         else
 528                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 529                         pipe->tmp_page = NULL;
 530
 531                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 532                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 533                                 if (!ret)
 534                                         ret = -EFAULT;
 535                                 break;
 536                         }
 537                         ret += copied;
 538                         buf->offset = 0;
 539                         buf->len = copied;
 540
 541                         if (!iov_iter_count(from))
 542                                 break;
 543                 }
 544
 545                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 546                         continue;
 547
 548                 /* Wait for buffer space to become available. */
 549                 if (filp->f_flags & O_NONBLOCK) {
 550                         if (!ret)
 551                                 ret = -EAGAIN;
 552                         break;
 553                 }
 554                 if (signal_pending(current)) {
 555                         if (!ret)
 556                                 ret = -ERESTARTSYS;
 557                         break;
 558                 }
 559
 560                 /*
 561                  * We're going to release the pipe lock and wait for more
 562                  * space. We wake up any readers if necessary, and then
 563                  * after waiting we need to re-check whether the pipe
 564                  * become empty while we dropped the lock.
 565                  */
 566                 __pipe_unlock(pipe);
 567                 if (was_empty)
 568                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 569                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 570                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 571                 __pipe_lock(pipe);
 572                 was_empty = pipe_empty(pipe->head, pipe->tail);
 573                 wake_next_writer = true;
 574         }
 575 out:
 576         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 577                 wake_next_writer = false;
 578         __pipe_unlock(pipe);
 579
 580         /*
 581          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 582          * want the reader to start processing things asap, rather than
 583          * leave the data pending.
 584          *
 585          * This is particularly important for small writes, because of
 586          * how (for example) the GNU make jobserver uses small writes to
 587          * wake up pending jobs
 588          *
 589          * Epoll nonsensically wants a wakeup whether the pipe
 590          * was already empty or not.
 591          */
 592         if (was_empty || pipe->poll_usage)
 593                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 594         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 595         if (wake_next_writer)
 596                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 597         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 598                 int err = file_update_time(filp);
 599                 if (err)
 600                         ret = err;
 601                 sb_end_write(file_inode(filp)->i_sb);
 602         }
 603         return ret;
 604 }
 605
 606 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 607 {
 608         struct pipe_inode_info *pipe = filp->private_data;
 609         int count, head, tail, mask;
 610
 611         switch (cmd) {
 612         case FIONREAD:
 613                 __pipe_lock(pipe);
 614                 count = 0;
 615                 head = pipe->head;
 616                 tail = pipe->tail;
 617                 mask = pipe->ring_size - 1;
 618
 619                 while (tail != head) {
 620                         count += pipe->bufs[tail & mask].len;
 621                         tail++;
 622                 }
 623                 __pipe_unlock(pipe);
 624
 625                 return put_user(count, (int __user *)arg);
 626
 627 #ifdef CONFIG_WATCH_QUEUE
 628         case IOC_WATCH_QUEUE_SET_SIZE: {
 629                 int ret;
 630                 __pipe_lock(pipe);
 631                 ret = watch_queue_set_size(pipe, arg);
 632                 __pipe_unlock(pipe);
 633                 return ret;
 634         }
 635
 636         case IOC_WATCH_QUEUE_SET_FILTER:
 637                 return watch_queue_set_filter(
 638                         pipe, (struct watch_notification_filter __user *)arg);
 639 #endif
 640
 641         default:
 642                 return -ENOIOCTLCMD;
 643         }
 644 }
 645
 646 /* No kernel lock held - fine */
 647 static __poll_t
 648 pipe_poll(struct file *filp, poll_table *wait)
 649 {
 650         __poll_t mask;
 651         struct pipe_inode_info *pipe = filp->private_data;
 652         unsigned int head, tail;
 653
 654         /* Epoll has some historical nasty semantics, this enables them */
 655         pipe->poll_usage = 1;
 656
 657         /*
 658          * Reading pipe state only -- no need for acquiring the semaphore.
 659          *
 660          * But because this is racy, the code has to add the
 661          * entry to the poll table _first_ ..
 662          */
 663         if (filp->f_mode & FMODE_READ)
 664                 poll_wait(filp, &pipe->rd_wait, wait);
 665         if (filp->f_mode & FMODE_WRITE)
 666                 poll_wait(filp, &pipe->wr_wait, wait);
 667
 668         /*
 669          * .. and only then can you do the racy tests. That way,
 670          * if something changes and you got it wrong, the poll
 671          * table entry will wake you up and fix it.
 672          */
 673         head = READ_ONCE(pipe->head);
 674         tail = READ_ONCE(pipe->tail);
 675
 676         mask = 0;
 677         if (filp->f_mode & FMODE_READ) {
 678                 if (!pipe_empty(head, tail))
 679                         mask |= EPOLLIN | EPOLLRDNORM;
 680                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 681                         mask |= EPOLLHUP;
 682         }
 683
 684         if (filp->f_mode & FMODE_WRITE) {
 685                 if (!pipe_full(head, tail, pipe->max_usage))
 686                         mask |= EPOLLOUT | EPOLLWRNORM;
 687                 /*
 688                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 689                  * behave exactly like pipes for poll().
 690                  */
 691                 if (!pipe->readers)
 692                         mask |= EPOLLERR;
 693         }
 694
 695         return mask;
 696 }
 697
 698 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 699 {
 700         int kill = 0;
 701
 702         spin_lock(&inode->i_lock);
 703         if (!--pipe->files) {
 704                 inode->i_pipe = NULL;
 705                 kill = 1;
 706         }
 707         spin_unlock(&inode->i_lock);
 708
 709         if (kill)
 710                 free_pipe_info(pipe);
 711 }
 712
 713 static int
 714 pipe_release(struct inode *inode, struct file *file)
 715 {
 716         struct pipe_inode_info *pipe = file->private_data;
 717
 718         __pipe_lock(pipe);
 719         if (file->f_mode & FMODE_READ)
 720                 pipe->readers--;
 721         if (file->f_mode & FMODE_WRITE)
 722                 pipe->writers--;
 723
 724         /* Was that the last reader or writer, but not the other side? */
 725         if (!pipe->readers != !pipe->writers) {
 726                 wake_up_interruptible_all(&pipe->rd_wait);
 727                 wake_up_interruptible_all(&pipe->wr_wait);
 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 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 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 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 bool pipe_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) && pipe_is_unprivileged_user()) {
 799                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 800                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 801         }
 802
 803         if (too_many_pipe_buffers_hard(user_bufs) && pipe_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->nr_accounted = pipe_bufs;
 816                 pipe->user = user;
 817                 mutex_init(&pipe->mutex);
 818                 return pipe;
 819         }
 820
 821 out_revert_acct:
 822         (void) account_pipe_buffers(user, pipe_bufs, 0);
 823         kfree(pipe);
 824 out_free_uid:
 825         free_uid(user);
 826         return NULL;
 827 }
 828
 829 void free_pipe_info(struct pipe_inode_info *pipe)
 830 {
 831         int i;
 832
 833 #ifdef CONFIG_WATCH_QUEUE
 834         if (pipe->watch_queue) {
 835                 watch_queue_clear(pipe->watch_queue);
 836                 put_watch_queue(pipe->watch_queue);
 837         }
 838 #endif
 839
 840         (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 841         free_uid(pipe->user);
 842         for (i = 0; i < pipe->ring_size; i++) {
 843                 struct pipe_buffer *buf = pipe->bufs + i;
 844                 if (buf->ops)
 845                         pipe_buf_release(pipe, buf);
 846         }
 847         if (pipe->tmp_page)
 848                 __free_page(pipe->tmp_page);
 849         kfree(pipe->bufs);
 850         kfree(pipe);
 851 }
 852
 853 static struct vfsmount *pipe_mnt __read_mostly;
 854
 855 /*
 856  * pipefs_dname() is called from d_path().
 857  */
 858 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 859 {
 860         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 861                                 d_inode(dentry)->i_ino);
 862 }
 863
 864 static const struct dentry_operations pipefs_dentry_operations = {
 865         .d_dname        = pipefs_dname,
 866 };
 867
 868 static struct inode * get_pipe_inode(void)
 869 {
 870         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 871         struct pipe_inode_info *pipe;
 872
 873         if (!inode)
 874                 goto fail_inode;
 875
 876         inode->i_ino = get_next_ino();
 877
 878         pipe = alloc_pipe_info();
 879         if (!pipe)
 880                 goto fail_iput;
 881
 882         inode->i_pipe = pipe;
 883         pipe->files = 2;
 884         pipe->readers = pipe->writers = 1;
 885         inode->i_fop = &pipefifo_fops;
 886
 887         /*
 888          * Mark the inode dirty from the very beginning,
 889          * that way it will never be moved to the dirty
 890          * list because "mark_inode_dirty()" will think
 891          * that it already _is_ on the dirty list.
 892          */
 893         inode->i_state = I_DIRTY;
 894         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 895         inode->i_uid = current_fsuid();
 896         inode->i_gid = current_fsgid();
 897         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 898
 899         return inode;
 900
 901 fail_iput:
 902         iput(inode);
 903
 904 fail_inode:
 905         return NULL;
 906 }
 907
 908 int create_pipe_files(struct file **res, int flags)
 909 {
 910         struct inode *inode = get_pipe_inode();
 911         struct file *f;
 912         int error;
 913
 914         if (!inode)
 915                 return -ENFILE;
 916
 917         if (flags & O_NOTIFICATION_PIPE) {
 918                 error = watch_queue_init(inode->i_pipe);
 919                 if (error) {
 920                         free_pipe_info(inode->i_pipe);
 921                         iput(inode);
 922                         return error;
 923                 }
 924         }
 925
 926         f = alloc_file_pseudo(inode, pipe_mnt, "",
 927                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 928                                 &pipefifo_fops);
 929         if (IS_ERR(f)) {
 930                 free_pipe_info(inode->i_pipe);
 931                 iput(inode);
 932                 return PTR_ERR(f);
 933         }
 934
 935         f->private_data = inode->i_pipe;
 936
 937         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 938                                   &pipefifo_fops);
 939         if (IS_ERR(res[0])) {
 940                 put_pipe_info(inode, inode->i_pipe);
 941                 fput(f);
 942                 return PTR_ERR(res[0]);
 943         }
 944         res[0]->private_data = inode->i_pipe;
 945         res[1] = f;
 946         stream_open(inode, res[0]);
 947         stream_open(inode, res[1]);
 948         return 0;
 949 }
 950
 951 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 952 {
 953         int error;
 954         int fdw, fdr;
 955
 956         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
 957                 return -EINVAL;
 958
 959         error = create_pipe_files(files, flags);
 960         if (error)
 961                 return error;
 962
 963         error = get_unused_fd_flags(flags);
 964         if (error < 0)
 965                 goto err_read_pipe;
 966         fdr = error;
 967
 968         error = get_unused_fd_flags(flags);
 969         if (error < 0)
 970                 goto err_fdr;
 971         fdw = error;
 972
 973         audit_fd_pair(fdr, fdw);
 974         fd[0] = fdr;
 975         fd[1] = fdw;
 976         return 0;
 977
 978  err_fdr:
 979         put_unused_fd(fdr);
 980  err_read_pipe:
 981         fput(files[0]);
 982         fput(files[1]);
 983         return error;
 984 }
 985
 986 int do_pipe_flags(int *fd, int flags)
 987 {
 988         struct file *files[2];
 989         int error = __do_pipe_flags(fd, files, flags);
 990         if (!error) {
 991                 fd_install(fd[0], files[0]);
 992                 fd_install(fd[1], files[1]);
 993         }
 994         return error;
 995 }
 996
 997 /*
 998  * sys_pipe() is the normal C calling standard for creating
 999  * a pipe. It's not the way Unix traditionally does this, though.
1000  */
1001 static int do_pipe2(int __user *fildes, int flags)
1002 {
1003         struct file *files[2];
1004         int fd[2];
1005         int error;
1006
1007         error = __do_pipe_flags(fd, files, flags);
1008         if (!error) {
1009                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1010                         fput(files[0]);
1011                         fput(files[1]);
1012                         put_unused_fd(fd[0]);
1013                         put_unused_fd(fd[1]);
1014                         error = -EFAULT;
1015                 } else {
1016                         fd_install(fd[0], files[0]);
1017                         fd_install(fd[1], files[1]);
1018                 }
1019         }
1020         return error;
1021 }
1022
1023 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1024 {
1025         return do_pipe2(fildes, flags);
1026 }
1027
1028 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1029 {
1030         return do_pipe2(fildes, 0);
1031 }
1032
1033 /*
1034  * This is the stupid "wait for pipe to be readable or writable"
1035  * model.
1036  *
1037  * See pipe_read/write() for the proper kind of exclusive wait,
1038  * but that requires that we wake up any other readers/writers
1039  * if we then do not end up reading everything (ie the whole
1040  * "wake_next_reader/writer" logic in pipe_read/write()).
1041  */
1042 void pipe_wait_readable(struct pipe_inode_info *pipe)
1043 {
1044         pipe_unlock(pipe);
1045         wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1046         pipe_lock(pipe);
1047 }
1048
1049 void pipe_wait_writable(struct pipe_inode_info *pipe)
1050 {
1051         pipe_unlock(pipe);
1052         wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1053         pipe_lock(pipe);
1054 }
1055
1056 /*
1057  * This depends on both the wait (here) and the wakeup (wake_up_partner)
1058  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1059  * race with the count check and waitqueue prep.
1060  *
1061  * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1062  * then check the condition you're waiting for, and only then sleep. But
1063  * because of the pipe lock, we can check the condition before being on
1064  * the wait queue.
1065  *
1066  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1067  */
1068 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1069 {
1070         DEFINE_WAIT(rdwait);
1071         int cur = *cnt;
1072
1073         while (cur == *cnt) {
1074                 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1075                 pipe_unlock(pipe);
1076                 schedule();
1077                 finish_wait(&pipe->rd_wait, &rdwait);
1078                 pipe_lock(pipe);
1079                 if (signal_pending(current))
1080                         break;
1081         }
1082         return cur == *cnt ? -ERESTARTSYS : 0;
1083 }
1084
1085 static void wake_up_partner(struct pipe_inode_info *pipe)
1086 {
1087         wake_up_interruptible_all(&pipe->rd_wait);
1088 }
1089
1090 static int fifo_open(struct inode *inode, struct file *filp)
1091 {
1092         struct pipe_inode_info *pipe;
1093         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1094         int ret;
1095
1096         filp->f_version = 0;
1097
1098         spin_lock(&inode->i_lock);
1099         if (inode->i_pipe) {
1100                 pipe = inode->i_pipe;
1101                 pipe->files++;
1102                 spin_unlock(&inode->i_lock);
1103         } else {
1104                 spin_unlock(&inode->i_lock);
1105                 pipe = alloc_pipe_info();
1106                 if (!pipe)
1107                         return -ENOMEM;
1108                 pipe->files = 1;
1109                 spin_lock(&inode->i_lock);
1110                 if (unlikely(inode->i_pipe)) {
1111                         inode->i_pipe->files++;
1112                         spin_unlock(&inode->i_lock);
1113                         free_pipe_info(pipe);
1114                         pipe = inode->i_pipe;
1115                 } else {
1116                         inode->i_pipe = pipe;
1117                         spin_unlock(&inode->i_lock);
1118                 }
1119         }
1120         filp->private_data = pipe;
1121         /* OK, we have a pipe and it's pinned down */
1122
1123         __pipe_lock(pipe);
1124
1125         /* We can only do regular read/write on fifos */
1126         stream_open(inode, filp);
1127
1128         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1129         case FMODE_READ:
1130         /*
1131          *  O_RDONLY
1132          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1133          *  opened, even when there is no process writing the FIFO.
1134          */
1135                 pipe->r_counter++;
1136                 if (pipe->readers++ == 0)
1137                         wake_up_partner(pipe);
1138
1139                 if (!is_pipe && !pipe->writers) {
1140                         if ((filp->f_flags & O_NONBLOCK)) {
1141                                 /* suppress EPOLLHUP until we have
1142                                  * seen a writer */
1143                                 filp->f_version = pipe->w_counter;
1144                         } else {
1145                                 if (wait_for_partner(pipe, &pipe->w_counter))
1146                                         goto err_rd;
1147                         }
1148                 }
1149                 break;
1150
1151         case FMODE_WRITE:
1152         /*
1153          *  O_WRONLY
1154          *  POSIX.1 says that O_NONBLOCK means return -1 with
1155          *  errno=ENXIO when there is no process reading the FIFO.
1156          */
1157                 ret = -ENXIO;
1158                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1159                         goto err;
1160
1161                 pipe->w_counter++;
1162                 if (!pipe->writers++)
1163                         wake_up_partner(pipe);
1164
1165                 if (!is_pipe && !pipe->readers) {
1166                         if (wait_for_partner(pipe, &pipe->r_counter))
1167                                 goto err_wr;
1168                 }
1169                 break;
1170
1171         case FMODE_READ | FMODE_WRITE:
1172         /*
1173          *  O_RDWR
1174          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1175          *  This implementation will NEVER block on a O_RDWR open, since
1176          *  the process can at least talk to itself.
1177          */
1178
1179                 pipe->readers++;
1180                 pipe->writers++;
1181                 pipe->r_counter++;
1182                 pipe->w_counter++;
1183                 if (pipe->readers == 1 || pipe->writers == 1)
1184                         wake_up_partner(pipe);
1185                 break;
1186
1187         default:
1188                 ret = -EINVAL;
1189                 goto err;
1190         }
1191
1192         /* Ok! */
1193         __pipe_unlock(pipe);
1194         return 0;
1195
1196 err_rd:
1197         if (!--pipe->readers)
1198                 wake_up_interruptible(&pipe->wr_wait);
1199         ret = -ERESTARTSYS;
1200         goto err;
1201
1202 err_wr:
1203         if (!--pipe->writers)
1204                 wake_up_interruptible_all(&pipe->rd_wait);
1205         ret = -ERESTARTSYS;
1206         goto err;
1207
1208 err:
1209         __pipe_unlock(pipe);
1210
1211         put_pipe_info(inode, pipe);
1212         return ret;
1213 }
1214
1215 const struct file_operations pipefifo_fops = {
1216         .open           = fifo_open,
1217         .llseek         = no_llseek,
1218         .read_iter      = pipe_read,
1219         .write_iter     = pipe_write,
1220         .poll           = pipe_poll,
1221         .unlocked_ioctl = pipe_ioctl,
1222         .release        = pipe_release,
1223         .fasync         = pipe_fasync,
1224         .splice_write   = iter_file_splice_write,
1225 };
1226
1227 /*
1228  * Currently we rely on the pipe array holding a power-of-2 number
1229  * of pages. Returns 0 on error.
1230  */
1231 unsigned int round_pipe_size(unsigned long size)
1232 {
1233         if (size > (1U << 31))
1234                 return 0;
1235
1236         /* Minimum pipe size, as required by POSIX */
1237         if (size < PAGE_SIZE)
1238                 return PAGE_SIZE;
1239
1240         return roundup_pow_of_two(size);
1241 }
1242
1243 /*
1244  * Resize the pipe ring to a number of slots.
1245  */
1246 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1247 {
1248         struct pipe_buffer *bufs;
1249         unsigned int head, tail, mask, n;
1250
1251         /*
1252          * We can shrink the pipe, if arg is greater than the ring occupancy.
1253          * Since we don't expect a lot of shrink+grow operations, just free and
1254          * allocate again like we would do for growing.  If the pipe currently
1255          * contains more buffers than arg, then return busy.
1256          */
1257         mask = pipe->ring_size - 1;
1258         head = pipe->head;
1259         tail = pipe->tail;
1260         n = pipe_occupancy(pipe->head, pipe->tail);
1261         if (nr_slots < n)
1262                 return -EBUSY;
1263
1264         bufs = kcalloc(nr_slots, sizeof(*bufs),
1265                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1266         if (unlikely(!bufs))
1267                 return -ENOMEM;
1268
1269         /*
1270          * The pipe array wraps around, so just start the new one at zero
1271          * and adjust the indices.
1272          */
1273         if (n > 0) {
1274                 unsigned int h = head & mask;
1275                 unsigned int t = tail & mask;
1276                 if (h > t) {
1277                         memcpy(bufs, pipe->bufs + t,
1278                                n * sizeof(struct pipe_buffer));
1279                 } else {
1280                         unsigned int tsize = pipe->ring_size - t;
1281                         if (h > 0)
1282                                 memcpy(bufs + tsize, pipe->bufs,
1283                                        h * sizeof(struct pipe_buffer));
1284                         memcpy(bufs, pipe->bufs + t,
1285                                tsize * sizeof(struct pipe_buffer));
1286                 }
1287         }
1288
1289         head = n;
1290         tail = 0;
1291
1292         kfree(pipe->bufs);
1293         pipe->bufs = bufs;
1294         pipe->ring_size = nr_slots;
1295         if (pipe->max_usage > nr_slots)
1296                 pipe->max_usage = nr_slots;
1297         pipe->tail = tail;
1298         pipe->head = head;
1299
1300         /* This might have made more room for writers */
1301         wake_up_interruptible(&pipe->wr_wait);
1302         return 0;
1303 }
1304
1305 /*
1306  * Allocate a new array of pipe buffers and copy the info over. Returns the
1307  * pipe size if successful, or return -ERROR on error.
1308  */
1309 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1310 {
1311         unsigned long user_bufs;
1312         unsigned int nr_slots, size;
1313         long ret = 0;
1314
1315 #ifdef CONFIG_WATCH_QUEUE
1316         if (pipe->watch_queue)
1317                 return -EBUSY;
1318 #endif
1319
1320         size = round_pipe_size(arg);
1321         nr_slots = size >> PAGE_SHIFT;
1322
1323         if (!nr_slots)
1324                 return -EINVAL;
1325
1326         /*
1327          * If trying to increase the pipe capacity, check that an
1328          * unprivileged user is not trying to exceed various limits
1329          * (soft limit check here, hard limit check just below).
1330          * Decreasing the pipe capacity is always permitted, even
1331          * if the user is currently over a limit.
1332          */
1333         if (nr_slots > pipe->max_usage &&
1334                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1335                 return -EPERM;
1336
1337         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1338
1339         if (nr_slots > pipe->max_usage &&
1340                         (too_many_pipe_buffers_hard(user_bufs) ||
1341                          too_many_pipe_buffers_soft(user_bufs)) &&
1342                         pipe_is_unprivileged_user()) {
1343                 ret = -EPERM;
1344                 goto out_revert_acct;
1345         }
1346
1347         ret = pipe_resize_ring(pipe, nr_slots);
1348         if (ret < 0)
1349                 goto out_revert_acct;
1350
1351         pipe->max_usage = nr_slots;
1352         pipe->nr_accounted = nr_slots;
1353         return pipe->max_usage * PAGE_SIZE;
1354
1355 out_revert_acct:
1356         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1357         return ret;
1358 }
1359
1360 /*
1361  * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1362  * not enough to verify that this is a pipe.
1363  */
1364 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1365 {
1366         struct pipe_inode_info *pipe = file->private_data;
1367
1368         if (file->f_op != &pipefifo_fops || !pipe)
1369                 return NULL;
1370 #ifdef CONFIG_WATCH_QUEUE
1371         if (for_splice && pipe->watch_queue)
1372                 return NULL;
1373 #endif
1374         return pipe;
1375 }
1376
1377 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1378 {
1379         struct pipe_inode_info *pipe;
1380         long ret;
1381
1382         pipe = get_pipe_info(file, false);
1383         if (!pipe)
1384                 return -EBADF;
1385
1386         __pipe_lock(pipe);
1387
1388         switch (cmd) {
1389         case F_SETPIPE_SZ:
1390                 ret = pipe_set_size(pipe, arg);
1391                 break;
1392         case F_GETPIPE_SZ:
1393                 ret = pipe->max_usage * PAGE_SIZE;
1394                 break;
1395         default:
1396                 ret = -EINVAL;
1397                 break;
1398         }
1399
1400         __pipe_unlock(pipe);
1401         return ret;
1402 }
1403
1404 static const struct super_operations pipefs_ops = {
1405         .destroy_inode = free_inode_nonrcu,
1406         .statfs = simple_statfs,
1407 };
1408
1409 /*
1410  * pipefs should _never_ be mounted by userland - too much of security hassle,
1411  * no real gain from having the whole whorehouse mounted. So we don't need
1412  * any operations on the root directory. However, we need a non-trivial
1413  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1414  */
1415
1416 static int pipefs_init_fs_context(struct fs_context *fc)
1417 {
1418         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1419         if (!ctx)
1420                 return -ENOMEM;
1421         ctx->ops = &pipefs_ops;
1422         ctx->dops = &pipefs_dentry_operations;
1423         return 0;
1424 }
1425
1426 static struct file_system_type pipe_fs_type = {
1427         .name           = "pipefs",
1428         .init_fs_context = pipefs_init_fs_context,
1429         .kill_sb        = kill_anon_super,
1430 };
1431
1432 #ifdef CONFIG_SYSCTL
1433 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1434                                         unsigned int *valp,
1435                                         int write, void *data)
1436 {
1437         if (write) {
1438                 unsigned int val;
1439
1440                 val = round_pipe_size(*lvalp);
1441                 if (val == 0)
1442                         return -EINVAL;
1443
1444                 *valp = val;
1445         } else {
1446                 unsigned int val = *valp;
1447                 *lvalp = (unsigned long) val;
1448         }
1449
1450         return 0;
1451 }
1452
1453 static int proc_dopipe_max_size(struct ctl_table *table, int write,
1454                                 void *buffer, size_t *lenp, loff_t *ppos)
1455 {
1456         return do_proc_douintvec(table, write, buffer, lenp, ppos,
1457                                  do_proc_dopipe_max_size_conv, NULL);
1458 }
1459
1460 static struct ctl_table fs_pipe_sysctls[] = {
1461         {
1462                 .procname       = "pipe-max-size",
1463                 .data           = &pipe_max_size,
1464                 .maxlen         = sizeof(pipe_max_size),
1465                 .mode           = 0644,
1466                 .proc_handler   = proc_dopipe_max_size,
1467         },
1468         {
1469                 .procname       = "pipe-user-pages-hard",
1470                 .data           = &pipe_user_pages_hard,
1471                 .maxlen         = sizeof(pipe_user_pages_hard),
1472                 .mode           = 0644,
1473                 .proc_handler   = proc_doulongvec_minmax,
1474         },
1475         {
1476                 .procname       = "pipe-user-pages-soft",
1477                 .data           = &pipe_user_pages_soft,
1478                 .maxlen         = sizeof(pipe_user_pages_soft),
1479                 .mode           = 0644,
1480                 .proc_handler   = proc_doulongvec_minmax,
1481         },
1482         { }
1483 };
1484 #endif
1485
1486 static int __init init_pipe_fs(void)
1487 {
1488         int err = register_filesystem(&pipe_fs_type);
1489
1490         if (!err) {
1491                 pipe_mnt = kern_mount(&pipe_fs_type);
1492                 if (IS_ERR(pipe_mnt)) {
1493                         err = PTR_ERR(pipe_mnt);
1494                         unregister_filesystem(&pipe_fs_type);
1495                 }
1496         }
1497 #ifdef CONFIG_SYSCTL
1498         register_sysctl_init("fs", fs_pipe_sysctls);
1499 #endif
1500         return err;
1501 }
1502
1503 fs_initcall(init_pipe_fs);