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