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