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  *      in 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                 /* Read ->head with a barrier vs post_one_notification() */
 256                 unsigned int head = smp_load_acquire(&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         if (pipe_has_watch_queue(pipe)) {
 439                 ret = -EXDEV;
 440                 goto out;
 441         }
 442
 443         /*
 444          * If it wasn't empty we try to merge new data into
 445          * the last buffer.
 446          *
 447          * That naturally merges small writes, but it also
 448          * page-aligns the rest of the writes for large writes
 449          * spanning multiple pages.
 450          */
 451         head = pipe->head;
 452         was_empty = pipe_empty(head, pipe->tail);
 453         chars = total_len & (PAGE_SIZE-1);
 454         if (chars && !was_empty) {
 455                 unsigned int mask = pipe->ring_size - 1;
 456                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 457                 int offset = buf->offset + buf->len;
 458
 459                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 460                     offset + chars <= PAGE_SIZE) {
 461                         ret = pipe_buf_confirm(pipe, buf);
 462                         if (ret)
 463                                 goto out;
 464
 465                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 466                         if (unlikely(ret < chars)) {
 467                                 ret = -EFAULT;
 468                                 goto out;
 469                         }
 470
 471                         buf->len += ret;
 472                         if (!iov_iter_count(from))
 473                                 goto out;
 474                 }
 475         }
 476
 477         for (;;) {
 478                 if (!pipe->readers) {
 479                         send_sig(SIGPIPE, current, 0);
 480                         if (!ret)
 481                                 ret = -EPIPE;
 482                         break;
 483                 }
 484
 485                 head = pipe->head;
 486                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 487                         unsigned int mask = pipe->ring_size - 1;
 488                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 489                         struct page *page = pipe->tmp_page;
 490                         int copied;
 491
 492                         if (!page) {
 493                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 494                                 if (unlikely(!page)) {
 495                                         ret = ret ? : -ENOMEM;
 496                                         break;
 497                                 }
 498                                 pipe->tmp_page = page;
 499                         }
 500
 501                         /* Allocate a slot in the ring in advance and attach an
 502                          * empty buffer.  If we fault or otherwise fail to use
 503                          * it, either the reader will consume it or it'll still
 504                          * be there for the next write.
 505                          */
 506                         spin_lock_irq(&pipe->rd_wait.lock);
 507
 508                         head = pipe->head;
 509                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 510                                 spin_unlock_irq(&pipe->rd_wait.lock);
 511                                 continue;
 512                         }
 513
 514                         pipe->head = head + 1;
 515                         spin_unlock_irq(&pipe->rd_wait.lock);
 516
 517                         /* Insert it into the buffer array */
 518                         buf = &pipe->bufs[head & mask];
 519                         buf->page = page;
 520                         buf->ops = &anon_pipe_buf_ops;
 521                         buf->offset = 0;
 522                         buf->len = 0;
 523                         if (is_packetized(filp))
 524                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 525                         else
 526                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 527                         pipe->tmp_page = NULL;
 528
 529                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 530                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 531                                 if (!ret)
 532                                         ret = -EFAULT;
 533                                 break;
 534                         }
 535                         ret += copied;
 536                         buf->offset = 0;
 537                         buf->len = copied;
 538
 539                         if (!iov_iter_count(from))
 540                                 break;
 541                 }
 542
 543                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 544                         continue;
 545
 546                 /* Wait for buffer space to become available. */
 547                 if (filp->f_flags & O_NONBLOCK) {
 548                         if (!ret)
 549                                 ret = -EAGAIN;
 550                         break;
 551                 }
 552                 if (signal_pending(current)) {
 553                         if (!ret)
 554                                 ret = -ERESTARTSYS;
 555                         break;
 556                 }
 557
 558                 /*
 559                  * We're going to release the pipe lock and wait for more
 560                  * space. We wake up any readers if necessary, and then
 561                  * after waiting we need to re-check whether the pipe
 562                  * become empty while we dropped the lock.
 563                  */
 564                 __pipe_unlock(pipe);
 565                 if (was_empty)
 566                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 567                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 568                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 569                 __pipe_lock(pipe);
 570                 was_empty = pipe_empty(pipe->head, pipe->tail);
 571                 wake_next_writer = true;
 572         }
 573 out:
 574         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 575                 wake_next_writer = false;
 576         __pipe_unlock(pipe);
 577
 578         /*
 579          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 580          * want the reader to start processing things asap, rather than
 581          * leave the data pending.
 582          *
 583          * This is particularly important for small writes, because of
 584          * how (for example) the GNU make jobserver uses small writes to
 585          * wake up pending jobs
 586          *
 587          * Epoll nonsensically wants a wakeup whether the pipe
 588          * was already empty or not.
 589          */
 590         if (was_empty || pipe->poll_usage)
 591                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 592         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 593         if (wake_next_writer)
 594                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 595         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 596                 int err = file_update_time(filp);
 597                 if (err)
 598                         ret = err;
 599                 sb_end_write(file_inode(filp)->i_sb);
 600         }
 601         return ret;
 602 }
 603
 604 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 605 {
 606         struct pipe_inode_info *pipe = filp->private_data;
 607         int count, head, tail, mask;
 608
 609         switch (cmd) {
 610         case FIONREAD:
 611                 __pipe_lock(pipe);
 612                 count = 0;
 613                 head = pipe->head;
 614                 tail = pipe->tail;
 615                 mask = pipe->ring_size - 1;
 616
 617                 while (tail != head) {
 618                         count += pipe->bufs[tail & mask].len;
 619                         tail++;
 620                 }
 621                 __pipe_unlock(pipe);
 622
 623                 return put_user(count, (int __user *)arg);
 624
 625 #ifdef CONFIG_WATCH_QUEUE
 626         case IOC_WATCH_QUEUE_SET_SIZE: {
 627                 int ret;
 628                 __pipe_lock(pipe);
 629                 ret = watch_queue_set_size(pipe, arg);
 630                 __pipe_unlock(pipe);
 631                 return ret;
 632         }
 633
 634         case IOC_WATCH_QUEUE_SET_FILTER:
 635                 return watch_queue_set_filter(
 636                         pipe, (struct watch_notification_filter __user *)arg);
 637 #endif
 638
 639         default:
 640                 return -ENOIOCTLCMD;
 641         }
 642 }
 643
 644 /* No kernel lock held - fine */
 645 static __poll_t
 646 pipe_poll(struct file *filp, poll_table *wait)
 647 {
 648         __poll_t mask;
 649         struct pipe_inode_info *pipe = filp->private_data;
 650         unsigned int head, tail;
 651
 652         /* Epoll has some historical nasty semantics, this enables them */
 653         WRITE_ONCE(pipe->poll_usage, true);
 654
 655         /*
 656          * Reading pipe state only -- no need for acquiring the semaphore.
 657          *
 658          * But because this is racy, the code has to add the
 659          * entry to the poll table _first_ ..
 660          */
 661         if (filp->f_mode & FMODE_READ)
 662                 poll_wait(filp, &pipe->rd_wait, wait);
 663         if (filp->f_mode & FMODE_WRITE)
 664                 poll_wait(filp, &pipe->wr_wait, wait);
 665
 666         /*
 667          * .. and only then can you do the racy tests. That way,
 668          * if something changes and you got it wrong, the poll
 669          * table entry will wake you up and fix it.
 670          */
 671         head = READ_ONCE(pipe->head);
 672         tail = READ_ONCE(pipe->tail);
 673
 674         mask = 0;
 675         if (filp->f_mode & FMODE_READ) {
 676                 if (!pipe_empty(head, tail))
 677                         mask |= EPOLLIN | EPOLLRDNORM;
 678                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 679                         mask |= EPOLLHUP;
 680         }
 681
 682         if (filp->f_mode & FMODE_WRITE) {
 683                 if (!pipe_full(head, tail, pipe->max_usage))
 684                         mask |= EPOLLOUT | EPOLLWRNORM;
 685                 /*
 686                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 687                  * behave exactly like pipes for poll().
 688                  */
 689                 if (!pipe->readers)
 690                         mask |= EPOLLERR;
 691         }
 692
 693         return mask;
 694 }
 695
 696 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 697 {
 698         int kill = 0;
 699
 700         spin_lock(&inode->i_lock);
 701         if (!--pipe->files) {
 702                 inode->i_pipe = NULL;
 703                 kill = 1;
 704         }
 705         spin_unlock(&inode->i_lock);
 706
 707         if (kill)
 708                 free_pipe_info(pipe);
 709 }
 710
 711 static int
 712 pipe_release(struct inode *inode, struct file *file)
 713 {
 714         struct pipe_inode_info *pipe = file->private_data;
 715
 716         __pipe_lock(pipe);
 717         if (file->f_mode & FMODE_READ)
 718                 pipe->readers--;
 719         if (file->f_mode & FMODE_WRITE)
 720                 pipe->writers--;
 721
 722         /* Was that the last reader or writer, but not the other side? */
 723         if (!pipe->readers != !pipe->writers) {
 724                 wake_up_interruptible_all(&pipe->rd_wait);
 725                 wake_up_interruptible_all(&pipe->wr_wait);
 726                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 727                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 728         }
 729         __pipe_unlock(pipe);
 730
 731         put_pipe_info(inode, pipe);
 732         return 0;
 733 }
 734
 735 static int
 736 pipe_fasync(int fd, struct file *filp, int on)
 737 {
 738         struct pipe_inode_info *pipe = filp->private_data;
 739         int retval = 0;
 740
 741         __pipe_lock(pipe);
 742         if (filp->f_mode & FMODE_READ)
 743                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 744         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 745                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 746                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 747                         /* this can happen only if on == T */
 748                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 749         }
 750         __pipe_unlock(pipe);
 751         return retval;
 752 }
 753
 754 unsigned long account_pipe_buffers(struct user_struct *user,
 755                                    unsigned long old, unsigned long new)
 756 {
 757         return atomic_long_add_return(new - old, &user->pipe_bufs);
 758 }
 759
 760 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 761 {
 762         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 763
 764         return soft_limit && user_bufs > soft_limit;
 765 }
 766
 767 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 768 {
 769         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 770
 771         return hard_limit && user_bufs > hard_limit;
 772 }
 773
 774 bool pipe_is_unprivileged_user(void)
 775 {
 776         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 777 }
 778
 779 struct pipe_inode_info *alloc_pipe_info(void)
 780 {
 781         struct pipe_inode_info *pipe;
 782         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 783         struct user_struct *user = get_current_user();
 784         unsigned long user_bufs;
 785         unsigned int max_size = READ_ONCE(pipe_max_size);
 786
 787         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 788         if (pipe == NULL)
 789                 goto out_free_uid;
 790
 791         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 792                 pipe_bufs = max_size >> PAGE_SHIFT;
 793
 794         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 795
 796         if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
 797                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 798                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 799         }
 800
 801         if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
 802                 goto out_revert_acct;
 803
 804         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 805                              GFP_KERNEL_ACCOUNT);
 806
 807         if (pipe->bufs) {
 808                 init_waitqueue_head(&pipe->rd_wait);
 809                 init_waitqueue_head(&pipe->wr_wait);
 810                 pipe->r_counter = pipe->w_counter = 1;
 811                 pipe->max_usage = pipe_bufs;
 812                 pipe->ring_size = pipe_bufs;
 813                 pipe->nr_accounted = pipe_bufs;
 814                 pipe->user = user;
 815                 mutex_init(&pipe->mutex);
 816                 return pipe;
 817         }
 818
 819 out_revert_acct:
 820         (void) account_pipe_buffers(user, pipe_bufs, 0);
 821         kfree(pipe);
 822 out_free_uid:
 823         free_uid(user);
 824         return NULL;
 825 }
 826
 827 void free_pipe_info(struct pipe_inode_info *pipe)
 828 {
 829         int i;
 830
 831 #ifdef CONFIG_WATCH_QUEUE
 832         if (pipe->watch_queue)
 833                 watch_queue_clear(pipe->watch_queue);
 834 #endif
 835
 836         (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 837         free_uid(pipe->user);
 838         for (i = 0; i < pipe->ring_size; i++) {
 839                 struct pipe_buffer *buf = pipe->bufs + i;
 840                 if (buf->ops)
 841                         pipe_buf_release(pipe, buf);
 842         }
 843 #ifdef CONFIG_WATCH_QUEUE
 844         if (pipe->watch_queue)
 845                 put_watch_queue(pipe->watch_queue);
 846 #endif
 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  * Note the pipe can be reduced in capacity, but only if the current
1247  * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1248  * returned instead.
1249  */
1250 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1251 {
1252         struct pipe_buffer *bufs;
1253         unsigned int head, tail, mask, n;
1254
1255         bufs = kcalloc(nr_slots, sizeof(*bufs),
1256                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1257         if (unlikely(!bufs))
1258                 return -ENOMEM;
1259
1260         spin_lock_irq(&pipe->rd_wait.lock);
1261         mask = pipe->ring_size - 1;
1262         head = pipe->head;
1263         tail = pipe->tail;
1264
1265         n = pipe_occupancy(head, tail);
1266         if (nr_slots < n) {
1267                 spin_unlock_irq(&pipe->rd_wait.lock);
1268                 kfree(bufs);
1269                 return -EBUSY;
1270         }
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         if (!pipe_has_watch_queue(pipe)) {
1304                 pipe->max_usage = nr_slots;
1305                 pipe->nr_accounted = nr_slots;
1306         }
1307
1308         spin_unlock_irq(&pipe->rd_wait.lock);
1309
1310         /* This might have made more room for writers */
1311         wake_up_interruptible(&pipe->wr_wait);
1312         return 0;
1313 }
1314
1315 /*
1316  * Allocate a new array of pipe buffers and copy the info over. Returns the
1317  * pipe size if successful, or return -ERROR on error.
1318  */
1319 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1320 {
1321         unsigned long user_bufs;
1322         unsigned int nr_slots, size;
1323         long ret = 0;
1324
1325         if (pipe_has_watch_queue(pipe))
1326                 return -EBUSY;
1327
1328         size = round_pipe_size(arg);
1329         nr_slots = size >> PAGE_SHIFT;
1330
1331         if (!nr_slots)
1332                 return -EINVAL;
1333
1334         /*
1335          * If trying to increase the pipe capacity, check that an
1336          * unprivileged user is not trying to exceed various limits
1337          * (soft limit check here, hard limit check just below).
1338          * Decreasing the pipe capacity is always permitted, even
1339          * if the user is currently over a limit.
1340          */
1341         if (nr_slots > pipe->max_usage &&
1342                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1343                 return -EPERM;
1344
1345         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1346
1347         if (nr_slots > pipe->max_usage &&
1348                         (too_many_pipe_buffers_hard(user_bufs) ||
1349                          too_many_pipe_buffers_soft(user_bufs)) &&
1350                         pipe_is_unprivileged_user()) {
1351                 ret = -EPERM;
1352                 goto out_revert_acct;
1353         }
1354
1355         ret = pipe_resize_ring(pipe, nr_slots);
1356         if (ret < 0)
1357                 goto out_revert_acct;
1358
1359         return pipe->max_usage * PAGE_SIZE;
1360
1361 out_revert_acct:
1362         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1363         return ret;
1364 }
1365
1366 /*
1367  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1368  * location, so checking ->i_pipe is not enough to verify that this is a
1369  * pipe.
1370  */
1371 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1372 {
1373         struct pipe_inode_info *pipe = file->private_data;
1374
1375         if (file->f_op != &pipefifo_fops || !pipe)
1376                 return NULL;
1377         if (for_splice && pipe_has_watch_queue(pipe))
1378                 return NULL;
1379         return pipe;
1380 }
1381
1382 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1383 {
1384         struct pipe_inode_info *pipe;
1385         long ret;
1386
1387         pipe = get_pipe_info(file, false);
1388         if (!pipe)
1389                 return -EBADF;
1390
1391         __pipe_lock(pipe);
1392
1393         switch (cmd) {
1394         case F_SETPIPE_SZ:
1395                 ret = pipe_set_size(pipe, arg);
1396                 break;
1397         case F_GETPIPE_SZ:
1398                 ret = pipe->max_usage * PAGE_SIZE;
1399                 break;
1400         default:
1401                 ret = -EINVAL;
1402                 break;
1403         }
1404
1405         __pipe_unlock(pipe);
1406         return ret;
1407 }
1408
1409 static const struct super_operations pipefs_ops = {
1410         .destroy_inode = free_inode_nonrcu,
1411         .statfs = simple_statfs,
1412 };
1413
1414 /*
1415  * pipefs should _never_ be mounted by userland - too much of security hassle,
1416  * no real gain from having the whole whorehouse mounted. So we don't need
1417  * any operations on the root directory. However, we need a non-trivial
1418  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1419  */
1420
1421 static int pipefs_init_fs_context(struct fs_context *fc)
1422 {
1423         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1424         if (!ctx)
1425                 return -ENOMEM;
1426         ctx->ops = &pipefs_ops;
1427         ctx->dops = &pipefs_dentry_operations;
1428         return 0;
1429 }
1430
1431 static struct file_system_type pipe_fs_type = {
1432         .name           = "pipefs",
1433         .init_fs_context = pipefs_init_fs_context,
1434         .kill_sb        = kill_anon_super,
1435 };
1436
1437 static int __init init_pipe_fs(void)
1438 {
1439         int err = register_filesystem(&pipe_fs_type);
1440
1441         if (!err) {
1442                 pipe_mnt = kern_mount(&pipe_fs_type);
1443                 if (IS_ERR(pipe_mnt)) {
1444                         err = PTR_ERR(pipe_mnt);
1445                         unregister_filesystem(&pipe_fs_type);
1446                 }
1447         }
1448         return err;
1449 }
1450
1451 fs_initcall(init_pipe_fs);