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                 /* Read ->head with a barrier vs post_one_notification() */
 257                 unsigned int head = smp_load_acquire(&pipe->head);
 258                 unsigned int tail = pipe->tail;
 259                 unsigned int mask = pipe->ring_size - 1;
 260
 261 #ifdef CONFIG_WATCH_QUEUE
 262                 if (pipe->note_loss) {
 263                         struct watch_notification n;
 264
 265                         if (total_len < 8) {
 266                                 if (ret == 0)
 267                                         ret = -ENOBUFS;
 268                                 break;
 269                         }
 270
 271                         n.type = WATCH_TYPE_META;
 272                         n.subtype = WATCH_META_LOSS_NOTIFICATION;
 273                         n.info = watch_sizeof(n);
 274                         if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
 275                                 if (ret == 0)
 276                                         ret = -EFAULT;
 277                                 break;
 278                         }
 279                         ret += sizeof(n);
 280                         total_len -= sizeof(n);
 281                         pipe->note_loss = false;
 282                 }
 283 #endif
 284
 285                 if (!pipe_empty(head, tail)) {
 286                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 287                         size_t chars = buf->len;
 288                         size_t written;
 289                         int error;
 290
 291                         if (chars > total_len) {
 292                                 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
 293                                         if (ret == 0)
 294                                                 ret = -ENOBUFS;
 295                                         break;
 296                                 }
 297                                 chars = total_len;
 298                         }
 299
 300                         error = pipe_buf_confirm(pipe, buf);
 301                         if (error) {
 302                                 if (!ret)
 303                                         ret = error;
 304                                 break;
 305                         }
 306
 307                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 308                         if (unlikely(written < chars)) {
 309                                 if (!ret)
 310                                         ret = -EFAULT;
 311                                 break;
 312                         }
 313                         ret += chars;
 314                         buf->offset += chars;
 315                         buf->len -= chars;
 316
 317                         /* Was it a packet buffer? Clean up and exit */
 318                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 319                                 total_len = chars;
 320                                 buf->len = 0;
 321                         }
 322
 323                         if (!buf->len) {
 324                                 pipe_buf_release(pipe, buf);
 325                                 spin_lock_irq(&pipe->rd_wait.lock);
 326 #ifdef CONFIG_WATCH_QUEUE
 327                                 if (buf->flags & PIPE_BUF_FLAG_LOSS)
 328                                         pipe->note_loss = true;
 329 #endif
 330                                 tail++;
 331                                 pipe->tail = tail;
 332                                 spin_unlock_irq(&pipe->rd_wait.lock);
 333                         }
 334                         total_len -= chars;
 335                         if (!total_len)
 336                                 break;  /* common path: read succeeded */
 337                         if (!pipe_empty(head, tail))    /* More to do? */
 338                                 continue;
 339                 }
 340
 341                 if (!pipe->writers)
 342                         break;
 343                 if (ret)
 344                         break;
 345                 if (filp->f_flags & O_NONBLOCK) {
 346                         ret = -EAGAIN;
 347                         break;
 348                 }
 349                 __pipe_unlock(pipe);
 350
 351                 /*
 352                  * We only get here if we didn't actually read anything.
 353                  *
 354                  * However, we could have seen (and removed) a zero-sized
 355                  * pipe buffer, and might have made space in the buffers
 356                  * that way.
 357                  *
 358                  * You can't make zero-sized pipe buffers by doing an empty
 359                  * write (not even in packet mode), but they can happen if
 360                  * the writer gets an EFAULT when trying to fill a buffer
 361                  * that already got allocated and inserted in the buffer
 362                  * array.
 363                  *
 364                  * So we still need to wake up any pending writers in the
 365                  * _very_ unlikely case that the pipe was full, but we got
 366                  * no data.
 367                  */
 368                 if (unlikely(was_full))
 369                         wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 370                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 371
 372                 /*
 373                  * But because we didn't read anything, at this point we can
 374                  * just return directly with -ERESTARTSYS if we're interrupted,
 375                  * since we've done any required wakeups and there's no need
 376                  * to mark anything accessed. And we've dropped the lock.
 377                  */
 378                 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 379                         return -ERESTARTSYS;
 380
 381                 __pipe_lock(pipe);
 382                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 383                 wake_next_reader = true;
 384         }
 385         if (pipe_empty(pipe->head, pipe->tail))
 386                 wake_next_reader = false;
 387         __pipe_unlock(pipe);
 388
 389         if (was_full)
 390                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 391         if (wake_next_reader)
 392                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 393         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 394         if (ret > 0)
 395                 file_accessed(filp);
 396         return ret;
 397 }
 398
 399 static inline int is_packetized(struct file *file)
 400 {
 401         return (file->f_flags & O_DIRECT) != 0;
 402 }
 403
 404 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 405 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 406 {
 407         unsigned int head = READ_ONCE(pipe->head);
 408         unsigned int tail = READ_ONCE(pipe->tail);
 409         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 410
 411         return !pipe_full(head, tail, max_usage) ||
 412                 !READ_ONCE(pipe->readers);
 413 }
 414
 415 static ssize_t
 416 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 417 {
 418         struct file *filp = iocb->ki_filp;
 419         struct pipe_inode_info *pipe = filp->private_data;
 420         unsigned int head;
 421         ssize_t ret = 0;
 422         size_t total_len = iov_iter_count(from);
 423         ssize_t chars;
 424         bool was_empty = false;
 425         bool wake_next_writer = false;
 426
 427         /*
 428          * Reject writing to watch queue pipes before the point where we lock
 429          * the pipe.
 430          * Otherwise, lockdep would be unhappy if the caller already has another
 431          * pipe locked.
 432          * If we had to support locking a normal pipe and a notification pipe at
 433          * the same time, we could set up lockdep annotations for that, but
 434          * since we don't actually need that, it's simpler to just bail here.
 435          */
 436         if (pipe_has_watch_queue(pipe))
 437                 return -EXDEV;
 438
 439         /* Null write succeeds. */
 440         if (unlikely(total_len == 0))
 441                 return 0;
 442
 443         __pipe_lock(pipe);
 444
 445         if (!pipe->readers) {
 446                 send_sig(SIGPIPE, current, 0);
 447                 ret = -EPIPE;
 448                 goto out;
 449         }
 450
 451         /*
 452          * If it wasn't empty we try to merge new data into
 453          * the last buffer.
 454          *
 455          * That naturally merges small writes, but it also
 456          * page-aligns the rest of the writes for large writes
 457          * spanning multiple pages.
 458          */
 459         head = pipe->head;
 460         was_empty = pipe_empty(head, pipe->tail);
 461         chars = total_len & (PAGE_SIZE-1);
 462         if (chars && !was_empty) {
 463                 unsigned int mask = pipe->ring_size - 1;
 464                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 465                 int offset = buf->offset + buf->len;
 466
 467                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 468                     offset + chars <= PAGE_SIZE) {
 469                         ret = pipe_buf_confirm(pipe, buf);
 470                         if (ret)
 471                                 goto out;
 472
 473                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 474                         if (unlikely(ret < chars)) {
 475                                 ret = -EFAULT;
 476                                 goto out;
 477                         }
 478
 479                         buf->len += ret;
 480                         if (!iov_iter_count(from))
 481                                 goto out;
 482                 }
 483         }
 484
 485         for (;;) {
 486                 if (!pipe->readers) {
 487                         send_sig(SIGPIPE, current, 0);
 488                         if (!ret)
 489                                 ret = -EPIPE;
 490                         break;
 491                 }
 492
 493                 head = pipe->head;
 494                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 495                         unsigned int mask = pipe->ring_size - 1;
 496                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 497                         struct page *page = pipe->tmp_page;
 498                         int copied;
 499
 500                         if (!page) {
 501                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 502                                 if (unlikely(!page)) {
 503                                         ret = ret ? : -ENOMEM;
 504                                         break;
 505                                 }
 506                                 pipe->tmp_page = page;
 507                         }
 508
 509                         /* Allocate a slot in the ring in advance and attach an
 510                          * empty buffer.  If we fault or otherwise fail to use
 511                          * it, either the reader will consume it or it'll still
 512                          * be there for the next write.
 513                          */
 514                         spin_lock_irq(&pipe->rd_wait.lock);
 515
 516                         head = pipe->head;
 517                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 518                                 spin_unlock_irq(&pipe->rd_wait.lock);
 519                                 continue;
 520                         }
 521
 522                         pipe->head = head + 1;
 523                         spin_unlock_irq(&pipe->rd_wait.lock);
 524
 525                         /* Insert it into the buffer array */
 526                         buf = &pipe->bufs[head & mask];
 527                         buf->page = page;
 528                         buf->ops = &anon_pipe_buf_ops;
 529                         buf->offset = 0;
 530                         buf->len = 0;
 531                         if (is_packetized(filp))
 532                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 533                         else
 534                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 535                         pipe->tmp_page = NULL;
 536
 537                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 538                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 539                                 if (!ret)
 540                                         ret = -EFAULT;
 541                                 break;
 542                         }
 543                         ret += copied;
 544                         buf->offset = 0;
 545                         buf->len = copied;
 546
 547                         if (!iov_iter_count(from))
 548                                 break;
 549                 }
 550
 551                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 552                         continue;
 553
 554                 /* Wait for buffer space to become available. */
 555                 if (filp->f_flags & O_NONBLOCK) {
 556                         if (!ret)
 557                                 ret = -EAGAIN;
 558                         break;
 559                 }
 560                 if (signal_pending(current)) {
 561                         if (!ret)
 562                                 ret = -ERESTARTSYS;
 563                         break;
 564                 }
 565
 566                 /*
 567                  * We're going to release the pipe lock and wait for more
 568                  * space. We wake up any readers if necessary, and then
 569                  * after waiting we need to re-check whether the pipe
 570                  * become empty while we dropped the lock.
 571                  */
 572                 __pipe_unlock(pipe);
 573                 if (was_empty)
 574                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 575                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 576                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 577                 __pipe_lock(pipe);
 578                 was_empty = pipe_empty(pipe->head, pipe->tail);
 579                 wake_next_writer = true;
 580         }
 581 out:
 582         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 583                 wake_next_writer = false;
 584         __pipe_unlock(pipe);
 585
 586         /*
 587          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 588          * want the reader to start processing things asap, rather than
 589          * leave the data pending.
 590          *
 591          * This is particularly important for small writes, because of
 592          * how (for example) the GNU make jobserver uses small writes to
 593          * wake up pending jobs
 594          *
 595          * Epoll nonsensically wants a wakeup whether the pipe
 596          * was already empty or not.
 597          */
 598         if (was_empty || pipe->poll_usage)
 599                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 600         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 601         if (wake_next_writer)
 602                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 603         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 604                 int err = file_update_time(filp);
 605                 if (err)
 606                         ret = err;
 607                 sb_end_write(file_inode(filp)->i_sb);
 608         }
 609         return ret;
 610 }
 611
 612 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 613 {
 614         struct pipe_inode_info *pipe = filp->private_data;
 615         unsigned int count, head, tail, mask;
 616
 617         switch (cmd) {
 618         case FIONREAD:
 619                 __pipe_lock(pipe);
 620                 count = 0;
 621                 head = pipe->head;
 622                 tail = pipe->tail;
 623                 mask = pipe->ring_size - 1;
 624
 625                 while (tail != head) {
 626                         count += pipe->bufs[tail & mask].len;
 627                         tail++;
 628                 }
 629                 __pipe_unlock(pipe);
 630
 631                 return put_user(count, (int __user *)arg);
 632
 633 #ifdef CONFIG_WATCH_QUEUE
 634         case IOC_WATCH_QUEUE_SET_SIZE: {
 635                 int ret;
 636                 __pipe_lock(pipe);
 637                 ret = watch_queue_set_size(pipe, arg);
 638                 __pipe_unlock(pipe);
 639                 return ret;
 640         }
 641
 642         case IOC_WATCH_QUEUE_SET_FILTER:
 643                 return watch_queue_set_filter(
 644                         pipe, (struct watch_notification_filter __user *)arg);
 645 #endif
 646
 647         default:
 648                 return -ENOIOCTLCMD;
 649         }
 650 }
 651
 652 /* No kernel lock held - fine */
 653 static __poll_t
 654 pipe_poll(struct file *filp, poll_table *wait)
 655 {
 656         __poll_t mask;
 657         struct pipe_inode_info *pipe = filp->private_data;
 658         unsigned int head, tail;
 659
 660         /* Epoll has some historical nasty semantics, this enables them */
 661         WRITE_ONCE(pipe->poll_usage, true);
 662
 663         /*
 664          * Reading pipe state only -- no need for acquiring the semaphore.
 665          *
 666          * But because this is racy, the code has to add the
 667          * entry to the poll table _first_ ..
 668          */
 669         if (filp->f_mode & FMODE_READ)
 670                 poll_wait(filp, &pipe->rd_wait, wait);
 671         if (filp->f_mode & FMODE_WRITE)
 672                 poll_wait(filp, &pipe->wr_wait, wait);
 673
 674         /*
 675          * .. and only then can you do the racy tests. That way,
 676          * if something changes and you got it wrong, the poll
 677          * table entry will wake you up and fix it.
 678          */
 679         head = READ_ONCE(pipe->head);
 680         tail = READ_ONCE(pipe->tail);
 681
 682         mask = 0;
 683         if (filp->f_mode & FMODE_READ) {
 684                 if (!pipe_empty(head, tail))
 685                         mask |= EPOLLIN | EPOLLRDNORM;
 686                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 687                         mask |= EPOLLHUP;
 688         }
 689
 690         if (filp->f_mode & FMODE_WRITE) {
 691                 if (!pipe_full(head, tail, pipe->max_usage))
 692                         mask |= EPOLLOUT | EPOLLWRNORM;
 693                 /*
 694                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 695                  * behave exactly like pipes for poll().
 696                  */
 697                 if (!pipe->readers)
 698                         mask |= EPOLLERR;
 699         }
 700
 701         return mask;
 702 }
 703
 704 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 705 {
 706         int kill = 0;
 707
 708         spin_lock(&inode->i_lock);
 709         if (!--pipe->files) {
 710                 inode->i_pipe = NULL;
 711                 kill = 1;
 712         }
 713         spin_unlock(&inode->i_lock);
 714
 715         if (kill)
 716                 free_pipe_info(pipe);
 717 }
 718
 719 static int
 720 pipe_release(struct inode *inode, struct file *file)
 721 {
 722         struct pipe_inode_info *pipe = file->private_data;
 723
 724         __pipe_lock(pipe);
 725         if (file->f_mode & FMODE_READ)
 726                 pipe->readers--;
 727         if (file->f_mode & FMODE_WRITE)
 728                 pipe->writers--;
 729
 730         /* Was that the last reader or writer, but not the other side? */
 731         if (!pipe->readers != !pipe->writers) {
 732                 wake_up_interruptible_all(&pipe->rd_wait);
 733                 wake_up_interruptible_all(&pipe->wr_wait);
 734                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 735                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 736         }
 737         __pipe_unlock(pipe);
 738
 739         put_pipe_info(inode, pipe);
 740         return 0;
 741 }
 742
 743 static int
 744 pipe_fasync(int fd, struct file *filp, int on)
 745 {
 746         struct pipe_inode_info *pipe = filp->private_data;
 747         int retval = 0;
 748
 749         __pipe_lock(pipe);
 750         if (filp->f_mode & FMODE_READ)
 751                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 752         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 753                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 754                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 755                         /* this can happen only if on == T */
 756                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 757         }
 758         __pipe_unlock(pipe);
 759         return retval;
 760 }
 761
 762 unsigned long account_pipe_buffers(struct user_struct *user,
 763                                    unsigned long old, unsigned long new)
 764 {
 765         return atomic_long_add_return(new - old, &user->pipe_bufs);
 766 }
 767
 768 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 769 {
 770         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 771
 772         return soft_limit && user_bufs > soft_limit;
 773 }
 774
 775 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 776 {
 777         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 778
 779         return hard_limit && user_bufs > hard_limit;
 780 }
 781
 782 bool pipe_is_unprivileged_user(void)
 783 {
 784         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 785 }
 786
 787 struct pipe_inode_info *alloc_pipe_info(void)
 788 {
 789         struct pipe_inode_info *pipe;
 790         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 791         struct user_struct *user = get_current_user();
 792         unsigned long user_bufs;
 793         unsigned int max_size = READ_ONCE(pipe_max_size);
 794
 795         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 796         if (pipe == NULL)
 797                 goto out_free_uid;
 798
 799         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 800                 pipe_bufs = max_size >> PAGE_SHIFT;
 801
 802         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 803
 804         if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
 805                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 806                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 807         }
 808
 809         if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
 810                 goto out_revert_acct;
 811
 812         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 813                              GFP_KERNEL_ACCOUNT);
 814
 815         if (pipe->bufs) {
 816                 init_waitqueue_head(&pipe->rd_wait);
 817                 init_waitqueue_head(&pipe->wr_wait);
 818                 pipe->r_counter = pipe->w_counter = 1;
 819                 pipe->max_usage = pipe_bufs;
 820                 pipe->ring_size = pipe_bufs;
 821                 pipe->nr_accounted = pipe_bufs;
 822                 pipe->user = user;
 823                 mutex_init(&pipe->mutex);
 824                 return pipe;
 825         }
 826
 827 out_revert_acct:
 828         (void) account_pipe_buffers(user, pipe_bufs, 0);
 829         kfree(pipe);
 830 out_free_uid:
 831         free_uid(user);
 832         return NULL;
 833 }
 834
 835 void free_pipe_info(struct pipe_inode_info *pipe)
 836 {
 837         unsigned int i;
 838
 839 #ifdef CONFIG_WATCH_QUEUE
 840         if (pipe->watch_queue)
 841                 watch_queue_clear(pipe->watch_queue);
 842 #endif
 843
 844         (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 845         free_uid(pipe->user);
 846         for (i = 0; i < pipe->ring_size; i++) {
 847                 struct pipe_buffer *buf = pipe->bufs + i;
 848                 if (buf->ops)
 849                         pipe_buf_release(pipe, buf);
 850         }
 851 #ifdef CONFIG_WATCH_QUEUE
 852         if (pipe->watch_queue)
 853                 put_watch_queue(pipe->watch_queue);
 854 #endif
 855         if (pipe->tmp_page)
 856                 __free_page(pipe->tmp_page);
 857         kfree(pipe->bufs);
 858         kfree(pipe);
 859 }
 860
 861 static struct vfsmount *pipe_mnt __read_mostly;
 862
 863 /*
 864  * pipefs_dname() is called from d_path().
 865  */
 866 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 867 {
 868         return dynamic_dname(buffer, buflen, "pipe:[%lu]",
 869                                 d_inode(dentry)->i_ino);
 870 }
 871
 872 static const struct dentry_operations pipefs_dentry_operations = {
 873         .d_dname        = pipefs_dname,
 874 };
 875
 876 static struct inode * get_pipe_inode(void)
 877 {
 878         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 879         struct pipe_inode_info *pipe;
 880
 881         if (!inode)
 882                 goto fail_inode;
 883
 884         inode->i_ino = get_next_ino();
 885
 886         pipe = alloc_pipe_info();
 887         if (!pipe)
 888                 goto fail_iput;
 889
 890         inode->i_pipe = pipe;
 891         pipe->files = 2;
 892         pipe->readers = pipe->writers = 1;
 893         inode->i_fop = &pipefifo_fops;
 894
 895         /*
 896          * Mark the inode dirty from the very beginning,
 897          * that way it will never be moved to the dirty
 898          * list because "mark_inode_dirty()" will think
 899          * that it already _is_ on the dirty list.
 900          */
 901         inode->i_state = I_DIRTY;
 902         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 903         inode->i_uid = current_fsuid();
 904         inode->i_gid = current_fsgid();
 905         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 906
 907         return inode;
 908
 909 fail_iput:
 910         iput(inode);
 911
 912 fail_inode:
 913         return NULL;
 914 }
 915
 916 int create_pipe_files(struct file **res, int flags)
 917 {
 918         struct inode *inode = get_pipe_inode();
 919         struct file *f;
 920         int error;
 921
 922         if (!inode)
 923                 return -ENFILE;
 924
 925         if (flags & O_NOTIFICATION_PIPE) {
 926                 error = watch_queue_init(inode->i_pipe);
 927                 if (error) {
 928                         free_pipe_info(inode->i_pipe);
 929                         iput(inode);
 930                         return error;
 931                 }
 932         }
 933
 934         f = alloc_file_pseudo(inode, pipe_mnt, "",
 935                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 936                                 &pipefifo_fops);
 937         if (IS_ERR(f)) {
 938                 free_pipe_info(inode->i_pipe);
 939                 iput(inode);
 940                 return PTR_ERR(f);
 941         }
 942
 943         f->private_data = inode->i_pipe;
 944
 945         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 946                                   &pipefifo_fops);
 947         if (IS_ERR(res[0])) {
 948                 put_pipe_info(inode, inode->i_pipe);
 949                 fput(f);
 950                 return PTR_ERR(res[0]);
 951         }
 952         res[0]->private_data = inode->i_pipe;
 953         res[1] = f;
 954         stream_open(inode, res[0]);
 955         stream_open(inode, res[1]);
 956         return 0;
 957 }
 958
 959 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 960 {
 961         int error;
 962         int fdw, fdr;
 963
 964         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
 965                 return -EINVAL;
 966
 967         error = create_pipe_files(files, flags);
 968         if (error)
 969                 return error;
 970
 971         error = get_unused_fd_flags(flags);
 972         if (error < 0)
 973                 goto err_read_pipe;
 974         fdr = error;
 975
 976         error = get_unused_fd_flags(flags);
 977         if (error < 0)
 978                 goto err_fdr;
 979         fdw = error;
 980
 981         audit_fd_pair(fdr, fdw);
 982         fd[0] = fdr;
 983         fd[1] = fdw;
 984         return 0;
 985
 986  err_fdr:
 987         put_unused_fd(fdr);
 988  err_read_pipe:
 989         fput(files[0]);
 990         fput(files[1]);
 991         return error;
 992 }
 993
 994 int do_pipe_flags(int *fd, int flags)
 995 {
 996         struct file *files[2];
 997         int error = __do_pipe_flags(fd, files, flags);
 998         if (!error) {
 999                 fd_install(fd[0], files[0]);
1000                 fd_install(fd[1], files[1]);
1001         }
1002         return error;
1003 }
1004
1005 /*
1006  * sys_pipe() is the normal C calling standard for creating
1007  * a pipe. It's not the way Unix traditionally does this, though.
1008  */
1009 static int do_pipe2(int __user *fildes, int flags)
1010 {
1011         struct file *files[2];
1012         int fd[2];
1013         int error;
1014
1015         error = __do_pipe_flags(fd, files, flags);
1016         if (!error) {
1017                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1018                         fput(files[0]);
1019                         fput(files[1]);
1020                         put_unused_fd(fd[0]);
1021                         put_unused_fd(fd[1]);
1022                         error = -EFAULT;
1023                 } else {
1024                         fd_install(fd[0], files[0]);
1025                         fd_install(fd[1], files[1]);
1026                 }
1027         }
1028         return error;
1029 }
1030
1031 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1032 {
1033         return do_pipe2(fildes, flags);
1034 }
1035
1036 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1037 {
1038         return do_pipe2(fildes, 0);
1039 }
1040
1041 /*
1042  * This is the stupid "wait for pipe to be readable or writable"
1043  * model.
1044  *
1045  * See pipe_read/write() for the proper kind of exclusive wait,
1046  * but that requires that we wake up any other readers/writers
1047  * if we then do not end up reading everything (ie the whole
1048  * "wake_next_reader/writer" logic in pipe_read/write()).
1049  */
1050 void pipe_wait_readable(struct pipe_inode_info *pipe)
1051 {
1052         pipe_unlock(pipe);
1053         wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1054         pipe_lock(pipe);
1055 }
1056
1057 void pipe_wait_writable(struct pipe_inode_info *pipe)
1058 {
1059         pipe_unlock(pipe);
1060         wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1061         pipe_lock(pipe);
1062 }
1063
1064 /*
1065  * This depends on both the wait (here) and the wakeup (wake_up_partner)
1066  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1067  * race with the count check and waitqueue prep.
1068  *
1069  * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1070  * then check the condition you're waiting for, and only then sleep. But
1071  * because of the pipe lock, we can check the condition before being on
1072  * the wait queue.
1073  *
1074  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1075  */
1076 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1077 {
1078         DEFINE_WAIT(rdwait);
1079         int cur = *cnt;
1080
1081         while (cur == *cnt) {
1082                 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1083                 pipe_unlock(pipe);
1084                 schedule();
1085                 finish_wait(&pipe->rd_wait, &rdwait);
1086                 pipe_lock(pipe);
1087                 if (signal_pending(current))
1088                         break;
1089         }
1090         return cur == *cnt ? -ERESTARTSYS : 0;
1091 }
1092
1093 static void wake_up_partner(struct pipe_inode_info *pipe)
1094 {
1095         wake_up_interruptible_all(&pipe->rd_wait);
1096 }
1097
1098 static int fifo_open(struct inode *inode, struct file *filp)
1099 {
1100         struct pipe_inode_info *pipe;
1101         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1102         int ret;
1103
1104         filp->f_version = 0;
1105
1106         spin_lock(&inode->i_lock);
1107         if (inode->i_pipe) {
1108                 pipe = inode->i_pipe;
1109                 pipe->files++;
1110                 spin_unlock(&inode->i_lock);
1111         } else {
1112                 spin_unlock(&inode->i_lock);
1113                 pipe = alloc_pipe_info();
1114                 if (!pipe)
1115                         return -ENOMEM;
1116                 pipe->files = 1;
1117                 spin_lock(&inode->i_lock);
1118                 if (unlikely(inode->i_pipe)) {
1119                         inode->i_pipe->files++;
1120                         spin_unlock(&inode->i_lock);
1121                         free_pipe_info(pipe);
1122                         pipe = inode->i_pipe;
1123                 } else {
1124                         inode->i_pipe = pipe;
1125                         spin_unlock(&inode->i_lock);
1126                 }
1127         }
1128         filp->private_data = pipe;
1129         /* OK, we have a pipe and it's pinned down */
1130
1131         __pipe_lock(pipe);
1132
1133         /* We can only do regular read/write on fifos */
1134         stream_open(inode, filp);
1135
1136         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1137         case FMODE_READ:
1138         /*
1139          *  O_RDONLY
1140          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1141          *  opened, even when there is no process writing the FIFO.
1142          */
1143                 pipe->r_counter++;
1144                 if (pipe->readers++ == 0)
1145                         wake_up_partner(pipe);
1146
1147                 if (!is_pipe && !pipe->writers) {
1148                         if ((filp->f_flags & O_NONBLOCK)) {
1149                                 /* suppress EPOLLHUP until we have
1150                                  * seen a writer */
1151                                 filp->f_version = pipe->w_counter;
1152                         } else {
1153                                 if (wait_for_partner(pipe, &pipe->w_counter))
1154                                         goto err_rd;
1155                         }
1156                 }
1157                 break;
1158
1159         case FMODE_WRITE:
1160         /*
1161          *  O_WRONLY
1162          *  POSIX.1 says that O_NONBLOCK means return -1 with
1163          *  errno=ENXIO when there is no process reading the FIFO.
1164          */
1165                 ret = -ENXIO;
1166                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1167                         goto err;
1168
1169                 pipe->w_counter++;
1170                 if (!pipe->writers++)
1171                         wake_up_partner(pipe);
1172
1173                 if (!is_pipe && !pipe->readers) {
1174                         if (wait_for_partner(pipe, &pipe->r_counter))
1175                                 goto err_wr;
1176                 }
1177                 break;
1178
1179         case FMODE_READ | FMODE_WRITE:
1180         /*
1181          *  O_RDWR
1182          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1183          *  This implementation will NEVER block on a O_RDWR open, since
1184          *  the process can at least talk to itself.
1185          */
1186
1187                 pipe->readers++;
1188                 pipe->writers++;
1189                 pipe->r_counter++;
1190                 pipe->w_counter++;
1191                 if (pipe->readers == 1 || pipe->writers == 1)
1192                         wake_up_partner(pipe);
1193                 break;
1194
1195         default:
1196                 ret = -EINVAL;
1197                 goto err;
1198         }
1199
1200         /* Ok! */
1201         __pipe_unlock(pipe);
1202         return 0;
1203
1204 err_rd:
1205         if (!--pipe->readers)
1206                 wake_up_interruptible(&pipe->wr_wait);
1207         ret = -ERESTARTSYS;
1208         goto err;
1209
1210 err_wr:
1211         if (!--pipe->writers)
1212                 wake_up_interruptible_all(&pipe->rd_wait);
1213         ret = -ERESTARTSYS;
1214         goto err;
1215
1216 err:
1217         __pipe_unlock(pipe);
1218
1219         put_pipe_info(inode, pipe);
1220         return ret;
1221 }
1222
1223 const struct file_operations pipefifo_fops = {
1224         .open           = fifo_open,
1225         .llseek         = no_llseek,
1226         .read_iter      = pipe_read,
1227         .write_iter     = pipe_write,
1228         .poll           = pipe_poll,
1229         .unlocked_ioctl = pipe_ioctl,
1230         .release        = pipe_release,
1231         .fasync         = pipe_fasync,
1232         .splice_write   = iter_file_splice_write,
1233 };
1234
1235 /*
1236  * Currently we rely on the pipe array holding a power-of-2 number
1237  * of pages. Returns 0 on error.
1238  */
1239 unsigned int round_pipe_size(unsigned long size)
1240 {
1241         if (size > (1U << 31))
1242                 return 0;
1243
1244         /* Minimum pipe size, as required by POSIX */
1245         if (size < PAGE_SIZE)
1246                 return PAGE_SIZE;
1247
1248         return roundup_pow_of_two(size);
1249 }
1250
1251 /*
1252  * Resize the pipe ring to a number of slots.
1253  *
1254  * Note the pipe can be reduced in capacity, but only if the current
1255  * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1256  * returned instead.
1257  */
1258 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1259 {
1260         struct pipe_buffer *bufs;
1261         unsigned int head, tail, mask, n;
1262
1263         bufs = kcalloc(nr_slots, sizeof(*bufs),
1264                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1265         if (unlikely(!bufs))
1266                 return -ENOMEM;
1267
1268         spin_lock_irq(&pipe->rd_wait.lock);
1269         mask = pipe->ring_size - 1;
1270         head = pipe->head;
1271         tail = pipe->tail;
1272
1273         n = pipe_occupancy(head, tail);
1274         if (nr_slots < n) {
1275                 spin_unlock_irq(&pipe->rd_wait.lock);
1276                 kfree(bufs);
1277                 return -EBUSY;
1278         }
1279
1280         /*
1281          * The pipe array wraps around, so just start the new one at zero
1282          * and adjust the indices.
1283          */
1284         if (n > 0) {
1285                 unsigned int h = head & mask;
1286                 unsigned int t = tail & mask;
1287                 if (h > t) {
1288                         memcpy(bufs, pipe->bufs + t,
1289                                n * sizeof(struct pipe_buffer));
1290                 } else {
1291                         unsigned int tsize = pipe->ring_size - t;
1292                         if (h > 0)
1293                                 memcpy(bufs + tsize, pipe->bufs,
1294                                        h * sizeof(struct pipe_buffer));
1295                         memcpy(bufs, pipe->bufs + t,
1296                                tsize * sizeof(struct pipe_buffer));
1297                 }
1298         }
1299
1300         head = n;
1301         tail = 0;
1302
1303         kfree(pipe->bufs);
1304         pipe->bufs = bufs;
1305         pipe->ring_size = nr_slots;
1306         if (pipe->max_usage > nr_slots)
1307                 pipe->max_usage = nr_slots;
1308         pipe->tail = tail;
1309         pipe->head = head;
1310
1311         if (!pipe_has_watch_queue(pipe)) {
1312                 pipe->max_usage = nr_slots;
1313                 pipe->nr_accounted = nr_slots;
1314         }
1315
1316         spin_unlock_irq(&pipe->rd_wait.lock);
1317
1318         /* This might have made more room for writers */
1319         wake_up_interruptible(&pipe->wr_wait);
1320         return 0;
1321 }
1322
1323 /*
1324  * Allocate a new array of pipe buffers and copy the info over. Returns the
1325  * pipe size if successful, or return -ERROR on error.
1326  */
1327 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1328 {
1329         unsigned long user_bufs;
1330         unsigned int nr_slots, size;
1331         long ret = 0;
1332
1333         if (pipe_has_watch_queue(pipe))
1334                 return -EBUSY;
1335
1336         size = round_pipe_size(arg);
1337         nr_slots = size >> PAGE_SHIFT;
1338
1339         if (!nr_slots)
1340                 return -EINVAL;
1341
1342         /*
1343          * If trying to increase the pipe capacity, check that an
1344          * unprivileged user is not trying to exceed various limits
1345          * (soft limit check here, hard limit check just below).
1346          * Decreasing the pipe capacity is always permitted, even
1347          * if the user is currently over a limit.
1348          */
1349         if (nr_slots > pipe->max_usage &&
1350                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1351                 return -EPERM;
1352
1353         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1354
1355         if (nr_slots > pipe->max_usage &&
1356                         (too_many_pipe_buffers_hard(user_bufs) ||
1357                          too_many_pipe_buffers_soft(user_bufs)) &&
1358                         pipe_is_unprivileged_user()) {
1359                 ret = -EPERM;
1360                 goto out_revert_acct;
1361         }
1362
1363         ret = pipe_resize_ring(pipe, nr_slots);
1364         if (ret < 0)
1365                 goto out_revert_acct;
1366
1367         return pipe->max_usage * PAGE_SIZE;
1368
1369 out_revert_acct:
1370         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1371         return ret;
1372 }
1373
1374 /*
1375  * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1376  * not enough to verify that this is a pipe.
1377  */
1378 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1379 {
1380         struct pipe_inode_info *pipe = file->private_data;
1381
1382         if (file->f_op != &pipefifo_fops || !pipe)
1383                 return NULL;
1384         if (for_splice && pipe_has_watch_queue(pipe))
1385                 return NULL;
1386         return pipe;
1387 }
1388
1389 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1390 {
1391         struct pipe_inode_info *pipe;
1392         long ret;
1393
1394         pipe = get_pipe_info(file, false);
1395         if (!pipe)
1396                 return -EBADF;
1397
1398         __pipe_lock(pipe);
1399
1400         switch (cmd) {
1401         case F_SETPIPE_SZ:
1402                 ret = pipe_set_size(pipe, arg);
1403                 break;
1404         case F_GETPIPE_SZ:
1405                 ret = pipe->max_usage * PAGE_SIZE;
1406                 break;
1407         default:
1408                 ret = -EINVAL;
1409                 break;
1410         }
1411
1412         __pipe_unlock(pipe);
1413         return ret;
1414 }
1415
1416 static const struct super_operations pipefs_ops = {
1417         .destroy_inode = free_inode_nonrcu,
1418         .statfs = simple_statfs,
1419 };
1420
1421 /*
1422  * pipefs should _never_ be mounted by userland - too much of security hassle,
1423  * no real gain from having the whole whorehouse mounted. So we don't need
1424  * any operations on the root directory. However, we need a non-trivial
1425  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1426  */
1427
1428 static int pipefs_init_fs_context(struct fs_context *fc)
1429 {
1430         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1431         if (!ctx)
1432                 return -ENOMEM;
1433         ctx->ops = &pipefs_ops;
1434         ctx->dops = &pipefs_dentry_operations;
1435         return 0;
1436 }
1437
1438 static struct file_system_type pipe_fs_type = {
1439         .name           = "pipefs",
1440         .init_fs_context = pipefs_init_fs_context,
1441         .kill_sb        = kill_anon_super,
1442 };
1443
1444 #ifdef CONFIG_SYSCTL
1445 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1446                                         unsigned int *valp,
1447                                         int write, void *data)
1448 {
1449         if (write) {
1450                 unsigned int val;
1451
1452                 val = round_pipe_size(*lvalp);
1453                 if (val == 0)
1454                         return -EINVAL;
1455
1456                 *valp = val;
1457         } else {
1458                 unsigned int val = *valp;
1459                 *lvalp = (unsigned long) val;
1460         }
1461
1462         return 0;
1463 }
1464
1465 static int proc_dopipe_max_size(struct ctl_table *table, int write,
1466                                 void *buffer, size_t *lenp, loff_t *ppos)
1467 {
1468         return do_proc_douintvec(table, write, buffer, lenp, ppos,
1469                                  do_proc_dopipe_max_size_conv, NULL);
1470 }
1471
1472 static struct ctl_table fs_pipe_sysctls[] = {
1473         {
1474                 .procname       = "pipe-max-size",
1475                 .data           = &pipe_max_size,
1476                 .maxlen         = sizeof(pipe_max_size),
1477                 .mode           = 0644,
1478                 .proc_handler   = proc_dopipe_max_size,
1479         },
1480         {
1481                 .procname       = "pipe-user-pages-hard",
1482                 .data           = &pipe_user_pages_hard,
1483                 .maxlen         = sizeof(pipe_user_pages_hard),
1484                 .mode           = 0644,
1485                 .proc_handler   = proc_doulongvec_minmax,
1486         },
1487         {
1488                 .procname       = "pipe-user-pages-soft",
1489                 .data           = &pipe_user_pages_soft,
1490                 .maxlen         = sizeof(pipe_user_pages_soft),
1491                 .mode           = 0644,
1492                 .proc_handler   = proc_doulongvec_minmax,
1493         },
1494         { }
1495 };
1496 #endif
1497
1498 static int __init init_pipe_fs(void)
1499 {
1500         int err = register_filesystem(&pipe_fs_type);
1501
1502         if (!err) {
1503                 pipe_mnt = kern_mount(&pipe_fs_type);
1504                 if (IS_ERR(pipe_mnt)) {
1505                         err = PTR_ERR(pipe_mnt);
1506                         unregister_filesystem(&pipe_fs_type);
1507                 }
1508         }
1509 #ifdef CONFIG_SYSCTL
1510         register_sysctl_init("fs", fs_pipe_sysctls);
1511 #endif
1512         return err;
1513 }
1514
1515 fs_initcall(init_pipe_fs);