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 #include <linux/sort.h>
  30
  31 #include <linux/uaccess.h>
  32 #include <asm/ioctls.h>
  33
  34 #include "internal.h"
  35
  36 /*
  37  * New pipe buffers will be restricted to this size while the user is exceeding
  38  * their pipe buffer quota. The general pipe use case needs at least two
  39  * buffers: one for data yet to be read, and one for new data. If this is less
  40  * than two, then a write to a non-empty pipe may block even if the pipe is not
  41  * full. This can occur with GNU make jobserver or similar uses of pipes as
  42  * semaphores: multiple processes may be waiting to write tokens back to the
  43  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  44  *
  45  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  46  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  47  * emptied.
  48  */
  49 #define PIPE_MIN_DEF_BUFFERS 2
  50
  51 /*
  52  * The max size that a non-root user is allowed to grow the pipe. Can
  53  * be set by root in /proc/sys/fs/pipe-max-size
  54  */
  55 static unsigned int pipe_max_size = 1048576;
  56
  57 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  58  * matches default values.
  59  */
  60 static unsigned long pipe_user_pages_hard;
  61 static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  62
  63 /*
  64  * We use head and tail indices that aren't masked off, except at the point of
  65  * dereference, but rather they're allowed to wrap naturally.  This means there
  66  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  67  * <= 2^31.
  68  * -- David Howells 2019-09-23.
  69  *
  70  * Reads with count = 0 should always return 0.
  71  * -- Julian Bradfield 1999-06-07.
  72  *
  73  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  74  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  75  *
  76  * pipe_read & write cleanup
  77  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  78  */
  79
  80 #ifdef CONFIG_PROVE_LOCKING
  81 static int pipe_lock_cmp_fn(const struct lockdep_map *a,
  82                             const struct lockdep_map *b)
  83 {
  84         return cmp_int((unsigned long) a, (unsigned long) b);
  85 }
  86 #endif
  87
  88 void pipe_lock(struct pipe_inode_info *pipe)
  89 {
  90         if (pipe->files)
  91                 mutex_lock(&pipe->mutex);
  92 }
  93 EXPORT_SYMBOL(pipe_lock);
  94
  95 void pipe_unlock(struct pipe_inode_info *pipe)
  96 {
  97         if (pipe->files)
  98                 mutex_unlock(&pipe->mutex);
  99 }
 100 EXPORT_SYMBOL(pipe_unlock);
 101
 102 void pipe_double_lock(struct pipe_inode_info *pipe1,
 103                       struct pipe_inode_info *pipe2)
 104 {
 105         BUG_ON(pipe1 == pipe2);
 106
 107         if (pipe1 > pipe2)
 108                 swap(pipe1, pipe2);
 109
 110         pipe_lock(pipe1);
 111         pipe_lock(pipe2);
 112 }
 113
 114 static struct page *anon_pipe_get_page(struct pipe_inode_info *pipe)
 115 {
 116         for (int i = 0; i < ARRAY_SIZE(pipe->tmp_page); i++) {
 117                 if (pipe->tmp_page[i]) {
 118                         struct page *page = pipe->tmp_page[i];
 119                         pipe->tmp_page[i] = NULL;
 120                         return page;
 121                 }
 122         }
 123
 124         return alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 125 }
 126
 127 static void anon_pipe_put_page(struct pipe_inode_info *pipe,
 128                                struct page *page)
 129 {
 130         if (page_count(page) == 1) {
 131                 for (int i = 0; i < ARRAY_SIZE(pipe->tmp_page); i++) {
 132                         if (!pipe->tmp_page[i]) {
 133                                 pipe->tmp_page[i] = page;
 134                                 return;
 135                         }
 136                 }
 137         }
 138
 139         put_page(page);
 140 }
 141
 142 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 143                                   struct pipe_buffer *buf)
 144 {
 145         struct page *page = buf->page;
 146
 147         anon_pipe_put_page(pipe, page);
 148 }
 149
 150 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 151                 struct pipe_buffer *buf)
 152 {
 153         struct page *page = buf->page;
 154
 155         if (page_count(page) != 1)
 156                 return false;
 157         memcg_kmem_uncharge_page(page, 0);
 158         __SetPageLocked(page);
 159         return true;
 160 }
 161
 162 /**
 163  * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
 164  * @pipe:       the pipe that the buffer belongs to
 165  * @buf:        the buffer to attempt to steal
 166  *
 167  * Description:
 168  *      This function attempts to steal the &struct page attached to
 169  *      @buf. If successful, this function returns 0 and returns with
 170  *      the page locked. The caller may then reuse the page for whatever
 171  *      he wishes; the typical use is insertion into a different file
 172  *      page cache.
 173  */
 174 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 175                 struct pipe_buffer *buf)
 176 {
 177         struct page *page = buf->page;
 178
 179         /*
 180          * A reference of one is golden, that means that the owner of this
 181          * page is the only one holding a reference to it. lock the page
 182          * and return OK.
 183          */
 184         if (page_count(page) == 1) {
 185                 lock_page(page);
 186                 return true;
 187         }
 188         return false;
 189 }
 190 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
 191
 192 /**
 193  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 194  * @pipe:       the pipe that the buffer belongs to
 195  * @buf:        the buffer to get a reference to
 196  *
 197  * Description:
 198  *      This function grabs an extra reference to @buf. It's used in
 199  *      the tee() system call, when we duplicate the buffers in one
 200  *      pipe into another.
 201  */
 202 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 203 {
 204         return try_get_page(buf->page);
 205 }
 206 EXPORT_SYMBOL(generic_pipe_buf_get);
 207
 208 /**
 209  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 210  * @pipe:       the pipe that the buffer belongs to
 211  * @buf:        the buffer to put a reference to
 212  *
 213  * Description:
 214  *      This function releases a reference to @buf.
 215  */
 216 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 217                               struct pipe_buffer *buf)
 218 {
 219         put_page(buf->page);
 220 }
 221 EXPORT_SYMBOL(generic_pipe_buf_release);
 222
 223 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 224         .release        = anon_pipe_buf_release,
 225         .try_steal      = anon_pipe_buf_try_steal,
 226         .get            = generic_pipe_buf_get,
 227 };
 228
 229 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 230 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 231 {
 232         union pipe_index idx = { .head_tail = READ_ONCE(pipe->head_tail) };
 233         unsigned int writers = READ_ONCE(pipe->writers);
 234
 235         return !pipe_empty(idx.head, idx.tail) || !writers;
 236 }
 237
 238 static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
 239                                             struct pipe_buffer *buf,
 240                                             unsigned int tail)
 241 {
 242         pipe_buf_release(pipe, buf);
 243
 244         /*
 245          * If the pipe has a watch_queue, we need additional protection
 246          * by the spinlock because notifications get posted with only
 247          * this spinlock, no mutex
 248          */
 249         if (pipe_has_watch_queue(pipe)) {
 250                 spin_lock_irq(&pipe->rd_wait.lock);
 251 #ifdef CONFIG_WATCH_QUEUE
 252                 if (buf->flags & PIPE_BUF_FLAG_LOSS)
 253                         pipe->note_loss = true;
 254 #endif
 255                 pipe->tail = ++tail;
 256                 spin_unlock_irq(&pipe->rd_wait.lock);
 257                 return tail;
 258         }
 259
 260         /*
 261          * Without a watch_queue, we can simply increment the tail
 262          * without the spinlock - the mutex is enough.
 263          */
 264         pipe->tail = ++tail;
 265         return tail;
 266 }
 267
 268 static ssize_t
 269 anon_pipe_read(struct kiocb *iocb, struct iov_iter *to)
 270 {
 271         size_t total_len = iov_iter_count(to);
 272         struct file *filp = iocb->ki_filp;
 273         struct pipe_inode_info *pipe = filp->private_data;
 274         bool wake_writer = false, wake_next_reader = false;
 275         ssize_t ret;
 276
 277         /* Null read succeeds. */
 278         if (unlikely(total_len == 0))
 279                 return 0;
 280
 281         ret = 0;
 282         mutex_lock(&pipe->mutex);
 283
 284         /*
 285          * We only wake up writers if the pipe was full when we started reading
 286          * and it is no longer full after reading to avoid unnecessary wakeups.
 287          *
 288          * But when we do wake up writers, we do so using a sync wakeup
 289          * (WF_SYNC), because we want them to get going and generate more
 290          * data for us.
 291          */
 292         for (;;) {
 293                 /* Read ->head with a barrier vs post_one_notification() */
 294                 unsigned int head = smp_load_acquire(&pipe->head);
 295                 unsigned int tail = pipe->tail;
 296
 297 #ifdef CONFIG_WATCH_QUEUE
 298                 if (pipe->note_loss) {
 299                         struct watch_notification n;
 300
 301                         if (total_len < 8) {
 302                                 if (ret == 0)
 303                                         ret = -ENOBUFS;
 304                                 break;
 305                         }
 306
 307                         n.type = WATCH_TYPE_META;
 308                         n.subtype = WATCH_META_LOSS_NOTIFICATION;
 309                         n.info = watch_sizeof(n);
 310                         if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
 311                                 if (ret == 0)
 312                                         ret = -EFAULT;
 313                                 break;
 314                         }
 315                         ret += sizeof(n);
 316                         total_len -= sizeof(n);
 317                         pipe->note_loss = false;
 318                 }
 319 #endif
 320
 321                 if (!pipe_empty(head, tail)) {
 322                         struct pipe_buffer *buf = pipe_buf(pipe, tail);
 323                         size_t chars = buf->len;
 324                         size_t written;
 325                         int error;
 326
 327                         if (chars > total_len) {
 328                                 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
 329                                         if (ret == 0)
 330                                                 ret = -ENOBUFS;
 331                                         break;
 332                                 }
 333                                 chars = total_len;
 334                         }
 335
 336                         error = pipe_buf_confirm(pipe, buf);
 337                         if (error) {
 338                                 if (!ret)
 339                                         ret = error;
 340                                 break;
 341                         }
 342
 343                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 344                         if (unlikely(written < chars)) {
 345                                 if (!ret)
 346                                         ret = -EFAULT;
 347                                 break;
 348                         }
 349                         ret += chars;
 350                         buf->offset += chars;
 351                         buf->len -= chars;
 352
 353                         /* Was it a packet buffer? Clean up and exit */
 354                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 355                                 total_len = chars;
 356                                 buf->len = 0;
 357                         }
 358
 359                         if (!buf->len) {
 360                                 wake_writer |= pipe_full(head, tail, pipe->max_usage);
 361                                 tail = pipe_update_tail(pipe, buf, tail);
 362                         }
 363                         total_len -= chars;
 364                         if (!total_len)
 365                                 break;  /* common path: read succeeded */
 366                         if (!pipe_empty(head, tail))    /* More to do? */
 367                                 continue;
 368                 }
 369
 370                 if (!pipe->writers)
 371                         break;
 372                 if (ret)
 373                         break;
 374                 if ((filp->f_flags & O_NONBLOCK) ||
 375                     (iocb->ki_flags & IOCB_NOWAIT)) {
 376                         ret = -EAGAIN;
 377                         break;
 378                 }
 379                 mutex_unlock(&pipe->mutex);
 380                 /*
 381                  * We only get here if we didn't actually read anything.
 382                  *
 383                  * But because we didn't read anything, at this point we can
 384                  * just return directly with -ERESTARTSYS if we're interrupted,
 385                  * since we've done any required wakeups and there's no need
 386                  * to mark anything accessed. And we've dropped the lock.
 387                  */
 388                 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 389                         return -ERESTARTSYS;
 390
 391                 wake_next_reader = true;
 392                 mutex_lock(&pipe->mutex);
 393         }
 394         if (pipe_is_empty(pipe))
 395                 wake_next_reader = false;
 396         mutex_unlock(&pipe->mutex);
 397
 398         if (wake_writer)
 399                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 400         if (wake_next_reader)
 401                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 402         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 403         return ret;
 404 }
 405
 406 static ssize_t
 407 fifo_pipe_read(struct kiocb *iocb, struct iov_iter *to)
 408 {
 409         int ret = anon_pipe_read(iocb, to);
 410         if (ret > 0)
 411                 file_accessed(iocb->ki_filp);
 412         return ret;
 413 }
 414
 415 static inline int is_packetized(struct file *file)
 416 {
 417         return (file->f_flags & O_DIRECT) != 0;
 418 }
 419
 420 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 421 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 422 {
 423         union pipe_index idx = { .head_tail = READ_ONCE(pipe->head_tail) };
 424         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 425
 426         return !pipe_full(idx.head, idx.tail, max_usage) ||
 427                 !READ_ONCE(pipe->readers);
 428 }
 429
 430 static ssize_t
 431 anon_pipe_write(struct kiocb *iocb, struct iov_iter *from)
 432 {
 433         struct file *filp = iocb->ki_filp;
 434         struct pipe_inode_info *pipe = filp->private_data;
 435         unsigned int head;
 436         ssize_t ret = 0;
 437         size_t total_len = iov_iter_count(from);
 438         ssize_t chars;
 439         bool was_empty = false;
 440         bool wake_next_writer = false;
 441
 442         /*
 443          * Reject writing to watch queue pipes before the point where we lock
 444          * the pipe.
 445          * Otherwise, lockdep would be unhappy if the caller already has another
 446          * pipe locked.
 447          * If we had to support locking a normal pipe and a notification pipe at
 448          * the same time, we could set up lockdep annotations for that, but
 449          * since we don't actually need that, it's simpler to just bail here.
 450          */
 451         if (pipe_has_watch_queue(pipe))
 452                 return -EXDEV;
 453
 454         /* Null write succeeds. */
 455         if (unlikely(total_len == 0))
 456                 return 0;
 457
 458         mutex_lock(&pipe->mutex);
 459
 460         if (!pipe->readers) {
 461                 send_sig(SIGPIPE, current, 0);
 462                 ret = -EPIPE;
 463                 goto out;
 464         }
 465
 466         /*
 467          * If it wasn't empty we try to merge new data into
 468          * the last buffer.
 469          *
 470          * That naturally merges small writes, but it also
 471          * page-aligns the rest of the writes for large writes
 472          * spanning multiple pages.
 473          */
 474         head = pipe->head;
 475         was_empty = pipe_empty(head, pipe->tail);
 476         chars = total_len & (PAGE_SIZE-1);
 477         if (chars && !was_empty) {
 478                 struct pipe_buffer *buf = pipe_buf(pipe, head - 1);
 479                 int offset = buf->offset + buf->len;
 480
 481                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 482                     offset + chars <= PAGE_SIZE) {
 483                         ret = pipe_buf_confirm(pipe, buf);
 484                         if (ret)
 485                                 goto out;
 486
 487                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 488                         if (unlikely(ret < chars)) {
 489                                 ret = -EFAULT;
 490                                 goto out;
 491                         }
 492
 493                         buf->len += ret;
 494                         if (!iov_iter_count(from))
 495                                 goto out;
 496                 }
 497         }
 498
 499         for (;;) {
 500                 if (!pipe->readers) {
 501                         send_sig(SIGPIPE, current, 0);
 502                         if (!ret)
 503                                 ret = -EPIPE;
 504                         break;
 505                 }
 506
 507                 head = pipe->head;
 508                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 509                         struct pipe_buffer *buf;
 510                         struct page *page;
 511                         int copied;
 512
 513                         page = anon_pipe_get_page(pipe);
 514                         if (unlikely(!page)) {
 515                                 if (!ret)
 516                                         ret = -ENOMEM;
 517                                 break;
 518                         }
 519
 520                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 521                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 522                                 anon_pipe_put_page(pipe, page);
 523                                 if (!ret)
 524                                         ret = -EFAULT;
 525                                 break;
 526                         }
 527
 528                         pipe->head = head + 1;
 529                         /* Insert it into the buffer array */
 530                         buf = pipe_buf(pipe, head);
 531                         buf->page = page;
 532                         buf->ops = &anon_pipe_buf_ops;
 533                         buf->offset = 0;
 534                         if (is_packetized(filp))
 535                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 536                         else
 537                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 538
 539                         buf->len = copied;
 540                         ret += copied;
 541
 542                         if (!iov_iter_count(from))
 543                                 break;
 544
 545                         continue;
 546                 }
 547
 548                 /* Wait for buffer space to become available. */
 549                 if ((filp->f_flags & O_NONBLOCK) ||
 550                     (iocb->ki_flags & IOCB_NOWAIT)) {
 551                         if (!ret)
 552                                 ret = -EAGAIN;
 553                         break;
 554                 }
 555                 if (signal_pending(current)) {
 556                         if (!ret)
 557                                 ret = -ERESTARTSYS;
 558                         break;
 559                 }
 560
 561                 /*
 562                  * We're going to release the pipe lock and wait for more
 563                  * space. We wake up any readers if necessary, and then
 564                  * after waiting we need to re-check whether the pipe
 565                  * become empty while we dropped the lock.
 566                  */
 567                 mutex_unlock(&pipe->mutex);
 568                 if (was_empty)
 569                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 570                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 571                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 572                 mutex_lock(&pipe->mutex);
 573                 was_empty = pipe_is_empty(pipe);
 574                 wake_next_writer = true;
 575         }
 576 out:
 577         if (pipe_is_full(pipe))
 578                 wake_next_writer = false;
 579         mutex_unlock(&pipe->mutex);
 580
 581         /*
 582          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 583          * want the reader to start processing things asap, rather than
 584          * leave the data pending.
 585          *
 586          * This is particularly important for small writes, because of
 587          * how (for example) the GNU make jobserver uses small writes to
 588          * wake up pending jobs
 589          *
 590          * Epoll nonsensically wants a wakeup whether the pipe
 591          * was already empty or not.
 592          */
 593         if (was_empty || pipe->poll_usage)
 594                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 595         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 596         if (wake_next_writer)
 597                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 598         return ret;
 599 }
 600
 601 static ssize_t
 602 fifo_pipe_write(struct kiocb *iocb, struct iov_iter *from)
 603 {
 604         int ret = anon_pipe_write(iocb, from);
 605         if (ret > 0) {
 606                 struct file *filp = iocb->ki_filp;
 607                 if (sb_start_write_trylock(file_inode(filp)->i_sb)) {
 608                         int err = file_update_time(filp);
 609                         if (err)
 610                                 ret = err;
 611                         sb_end_write(file_inode(filp)->i_sb);
 612                 }
 613         }
 614         return ret;
 615 }
 616
 617 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 618 {
 619         struct pipe_inode_info *pipe = filp->private_data;
 620         unsigned int count, head, tail;
 621
 622         switch (cmd) {
 623         case FIONREAD:
 624                 mutex_lock(&pipe->mutex);
 625                 count = 0;
 626                 head = pipe->head;
 627                 tail = pipe->tail;
 628
 629                 while (!pipe_empty(head, tail)) {
 630                         count += pipe_buf(pipe, tail)->len;
 631                         tail++;
 632                 }
 633                 mutex_unlock(&pipe->mutex);
 634
 635                 return put_user(count, (int __user *)arg);
 636
 637 #ifdef CONFIG_WATCH_QUEUE
 638         case IOC_WATCH_QUEUE_SET_SIZE: {
 639                 int ret;
 640                 mutex_lock(&pipe->mutex);
 641                 ret = watch_queue_set_size(pipe, arg);
 642                 mutex_unlock(&pipe->mutex);
 643                 return ret;
 644         }
 645
 646         case IOC_WATCH_QUEUE_SET_FILTER:
 647                 return watch_queue_set_filter(
 648                         pipe, (struct watch_notification_filter __user *)arg);
 649 #endif
 650
 651         default:
 652                 return -ENOIOCTLCMD;
 653         }
 654 }
 655
 656 /* No kernel lock held - fine */
 657 static __poll_t
 658 pipe_poll(struct file *filp, poll_table *wait)
 659 {
 660         __poll_t mask;
 661         struct pipe_inode_info *pipe = filp->private_data;
 662         union pipe_index idx;
 663
 664         /* Epoll has some historical nasty semantics, this enables them */
 665         WRITE_ONCE(pipe->poll_usage, true);
 666
 667         /*
 668          * Reading pipe state only -- no need for acquiring the semaphore.
 669          *
 670          * But because this is racy, the code has to add the
 671          * entry to the poll table _first_ ..
 672          */
 673         if (filp->f_mode & FMODE_READ)
 674                 poll_wait(filp, &pipe->rd_wait, wait);
 675         if (filp->f_mode & FMODE_WRITE)
 676                 poll_wait(filp, &pipe->wr_wait, wait);
 677
 678         /*
 679          * .. and only then can you do the racy tests. That way,
 680          * if something changes and you got it wrong, the poll
 681          * table entry will wake you up and fix it.
 682          */
 683         idx.head_tail = READ_ONCE(pipe->head_tail);
 684
 685         mask = 0;
 686         if (filp->f_mode & FMODE_READ) {
 687                 if (!pipe_empty(idx.head, idx.tail))
 688                         mask |= EPOLLIN | EPOLLRDNORM;
 689                 if (!pipe->writers && filp->f_pipe != pipe->w_counter)
 690                         mask |= EPOLLHUP;
 691         }
 692
 693         if (filp->f_mode & FMODE_WRITE) {
 694                 if (!pipe_full(idx.head, idx.tail, pipe->max_usage))
 695                         mask |= EPOLLOUT | EPOLLWRNORM;
 696                 /*
 697                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 698                  * behave exactly like pipes for poll().
 699                  */
 700                 if (!pipe->readers)
 701                         mask |= EPOLLERR;
 702         }
 703
 704         return mask;
 705 }
 706
 707 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 708 {
 709         int kill = 0;
 710
 711         spin_lock(&inode->i_lock);
 712         if (!--pipe->files) {
 713                 inode->i_pipe = NULL;
 714                 kill = 1;
 715         }
 716         spin_unlock(&inode->i_lock);
 717
 718         if (kill)
 719                 free_pipe_info(pipe);
 720 }
 721
 722 static int
 723 pipe_release(struct inode *inode, struct file *file)
 724 {
 725         struct pipe_inode_info *pipe = file->private_data;
 726
 727         mutex_lock(&pipe->mutex);
 728         if (file->f_mode & FMODE_READ)
 729                 pipe->readers--;
 730         if (file->f_mode & FMODE_WRITE)
 731                 pipe->writers--;
 732
 733         /* Was that the last reader or writer, but not the other side? */
 734         if (!pipe->readers != !pipe->writers) {
 735                 wake_up_interruptible_all(&pipe->rd_wait);
 736                 wake_up_interruptible_all(&pipe->wr_wait);
 737                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 738                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 739         }
 740         mutex_unlock(&pipe->mutex);
 741
 742         put_pipe_info(inode, pipe);
 743         return 0;
 744 }
 745
 746 static int
 747 pipe_fasync(int fd, struct file *filp, int on)
 748 {
 749         struct pipe_inode_info *pipe = filp->private_data;
 750         int retval = 0;
 751
 752         mutex_lock(&pipe->mutex);
 753         if (filp->f_mode & FMODE_READ)
 754                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 755         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 756                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 757                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 758                         /* this can happen only if on == T */
 759                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 760         }
 761         mutex_unlock(&pipe->mutex);
 762         return retval;
 763 }
 764
 765 unsigned long account_pipe_buffers(struct user_struct *user,
 766                                    unsigned long old, unsigned long new)
 767 {
 768         return atomic_long_add_return(new - old, &user->pipe_bufs);
 769 }
 770
 771 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 772 {
 773         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 774
 775         return soft_limit && user_bufs > soft_limit;
 776 }
 777
 778 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 779 {
 780         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 781
 782         return hard_limit && user_bufs > hard_limit;
 783 }
 784
 785 bool pipe_is_unprivileged_user(void)
 786 {
 787         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 788 }
 789
 790 struct pipe_inode_info *alloc_pipe_info(void)
 791 {
 792         struct pipe_inode_info *pipe;
 793         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 794         struct user_struct *user = get_current_user();
 795         unsigned long user_bufs;
 796         unsigned int max_size = READ_ONCE(pipe_max_size);
 797
 798         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 799         if (pipe == NULL)
 800                 goto out_free_uid;
 801
 802         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 803                 pipe_bufs = max_size >> PAGE_SHIFT;
 804
 805         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 806
 807         if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
 808                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 809                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 810         }
 811
 812         if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
 813                 goto out_revert_acct;
 814
 815         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 816                              GFP_KERNEL_ACCOUNT);
 817
 818         if (pipe->bufs) {
 819                 init_waitqueue_head(&pipe->rd_wait);
 820                 init_waitqueue_head(&pipe->wr_wait);
 821                 pipe->r_counter = pipe->w_counter = 1;
 822                 pipe->max_usage = pipe_bufs;
 823                 pipe->ring_size = pipe_bufs;
 824                 pipe->nr_accounted = pipe_bufs;
 825                 pipe->user = user;
 826                 mutex_init(&pipe->mutex);
 827                 lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);
 828                 return pipe;
 829         }
 830
 831 out_revert_acct:
 832         (void) account_pipe_buffers(user, pipe_bufs, 0);
 833         kfree(pipe);
 834 out_free_uid:
 835         free_uid(user);
 836         return NULL;
 837 }
 838
 839 void free_pipe_info(struct pipe_inode_info *pipe)
 840 {
 841         unsigned int i;
 842
 843 #ifdef CONFIG_WATCH_QUEUE
 844         if (pipe->watch_queue)
 845                 watch_queue_clear(pipe->watch_queue);
 846 #endif
 847
 848         (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 849         free_uid(pipe->user);
 850         for (i = 0; i < pipe->ring_size; i++) {
 851                 struct pipe_buffer *buf = pipe->bufs + i;
 852                 if (buf->ops)
 853                         pipe_buf_release(pipe, buf);
 854         }
 855 #ifdef CONFIG_WATCH_QUEUE
 856         if (pipe->watch_queue)
 857                 put_watch_queue(pipe->watch_queue);
 858 #endif
 859         for (i = 0; i < ARRAY_SIZE(pipe->tmp_page); i++) {
 860                 if (pipe->tmp_page[i])
 861                         __free_page(pipe->tmp_page[i]);
 862         }
 863         kfree(pipe->bufs);
 864         kfree(pipe);
 865 }
 866
 867 static struct vfsmount *pipe_mnt __ro_after_init;
 868
 869 /*
 870  * pipefs_dname() is called from d_path().
 871  */
 872 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 873 {
 874         return dynamic_dname(buffer, buflen, "pipe:[%lu]",
 875                                 d_inode(dentry)->i_ino);
 876 }
 877
 878 static const struct dentry_operations pipefs_dentry_operations = {
 879         .d_dname        = pipefs_dname,
 880 };
 881
 882 static const struct file_operations pipeanon_fops;
 883
 884 static struct inode * get_pipe_inode(void)
 885 {
 886         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 887         struct pipe_inode_info *pipe;
 888
 889         if (!inode)
 890                 goto fail_inode;
 891
 892         inode->i_ino = get_next_ino();
 893
 894         pipe = alloc_pipe_info();
 895         if (!pipe)
 896                 goto fail_iput;
 897
 898         inode->i_pipe = pipe;
 899         pipe->files = 2;
 900         pipe->readers = pipe->writers = 1;
 901         inode->i_fop = &pipeanon_fops;
 902
 903         /*
 904          * Mark the inode dirty from the very beginning,
 905          * that way it will never be moved to the dirty
 906          * list because "mark_inode_dirty()" will think
 907          * that it already _is_ on the dirty list.
 908          */
 909         inode->i_state = I_DIRTY;
 910         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 911         inode->i_uid = current_fsuid();
 912         inode->i_gid = current_fsgid();
 913         simple_inode_init_ts(inode);
 914
 915         return inode;
 916
 917 fail_iput:
 918         iput(inode);
 919
 920 fail_inode:
 921         return NULL;
 922 }
 923
 924 int create_pipe_files(struct file **res, int flags)
 925 {
 926         struct inode *inode = get_pipe_inode();
 927         struct file *f;
 928         int error;
 929
 930         if (!inode)
 931                 return -ENFILE;
 932
 933         if (flags & O_NOTIFICATION_PIPE) {
 934                 error = watch_queue_init(inode->i_pipe);
 935                 if (error) {
 936                         free_pipe_info(inode->i_pipe);
 937                         iput(inode);
 938                         return error;
 939                 }
 940         }
 941
 942         f = alloc_file_pseudo(inode, pipe_mnt, "",
 943                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 944                                 &pipeanon_fops);
 945         if (IS_ERR(f)) {
 946                 free_pipe_info(inode->i_pipe);
 947                 iput(inode);
 948                 return PTR_ERR(f);
 949         }
 950
 951         f->private_data = inode->i_pipe;
 952         f->f_pipe = 0;
 953
 954         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 955                                   &pipeanon_fops);
 956         if (IS_ERR(res[0])) {
 957                 put_pipe_info(inode, inode->i_pipe);
 958                 fput(f);
 959                 return PTR_ERR(res[0]);
 960         }
 961         res[0]->private_data = inode->i_pipe;
 962         res[0]->f_pipe = 0;
 963         res[1] = f;
 964         stream_open(inode, res[0]);
 965         stream_open(inode, res[1]);
 966
 967         /* pipe groks IOCB_NOWAIT */
 968         res[0]->f_mode |= FMODE_NOWAIT;
 969         res[1]->f_mode |= FMODE_NOWAIT;
 970
 971         /*
 972          * Disable permission and pre-content events, but enable legacy
 973          * inotify events for legacy users.
 974          */
 975         file_set_fsnotify_mode(res[0], FMODE_NONOTIFY_PERM);
 976         file_set_fsnotify_mode(res[1], FMODE_NONOTIFY_PERM);
 977         return 0;
 978 }
 979
 980 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 981 {
 982         int error;
 983         int fdw, fdr;
 984
 985         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
 986                 return -EINVAL;
 987
 988         error = create_pipe_files(files, flags);
 989         if (error)
 990                 return error;
 991
 992         error = get_unused_fd_flags(flags);
 993         if (error < 0)
 994                 goto err_read_pipe;
 995         fdr = error;
 996
 997         error = get_unused_fd_flags(flags);
 998         if (error < 0)
 999                 goto err_fdr;
1000         fdw = error;
1001
1002         audit_fd_pair(fdr, fdw);
1003         fd[0] = fdr;
1004         fd[1] = fdw;
1005         return 0;
1006
1007  err_fdr:
1008         put_unused_fd(fdr);
1009  err_read_pipe:
1010         fput(files[0]);
1011         fput(files[1]);
1012         return error;
1013 }
1014
1015 int do_pipe_flags(int *fd, int flags)
1016 {
1017         struct file *files[2];
1018         int error = __do_pipe_flags(fd, files, flags);
1019         if (!error) {
1020                 fd_install(fd[0], files[0]);
1021                 fd_install(fd[1], files[1]);
1022         }
1023         return error;
1024 }
1025
1026 /*
1027  * sys_pipe() is the normal C calling standard for creating
1028  * a pipe. It's not the way Unix traditionally does this, though.
1029  */
1030 static int do_pipe2(int __user *fildes, int flags)
1031 {
1032         struct file *files[2];
1033         int fd[2];
1034         int error;
1035
1036         error = __do_pipe_flags(fd, files, flags);
1037         if (!error) {
1038                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1039                         fput(files[0]);
1040                         fput(files[1]);
1041                         put_unused_fd(fd[0]);
1042                         put_unused_fd(fd[1]);
1043                         error = -EFAULT;
1044                 } else {
1045                         fd_install(fd[0], files[0]);
1046                         fd_install(fd[1], files[1]);
1047                 }
1048         }
1049         return error;
1050 }
1051
1052 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1053 {
1054         return do_pipe2(fildes, flags);
1055 }
1056
1057 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1058 {
1059         return do_pipe2(fildes, 0);
1060 }
1061
1062 /*
1063  * This is the stupid "wait for pipe to be readable or writable"
1064  * model.
1065  *
1066  * See pipe_read/write() for the proper kind of exclusive wait,
1067  * but that requires that we wake up any other readers/writers
1068  * if we then do not end up reading everything (ie the whole
1069  * "wake_next_reader/writer" logic in pipe_read/write()).
1070  */
1071 void pipe_wait_readable(struct pipe_inode_info *pipe)
1072 {
1073         pipe_unlock(pipe);
1074         wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1075         pipe_lock(pipe);
1076 }
1077
1078 void pipe_wait_writable(struct pipe_inode_info *pipe)
1079 {
1080         pipe_unlock(pipe);
1081         wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1082         pipe_lock(pipe);
1083 }
1084
1085 /*
1086  * This depends on both the wait (here) and the wakeup (wake_up_partner)
1087  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1088  * race with the count check and waitqueue prep.
1089  *
1090  * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1091  * then check the condition you're waiting for, and only then sleep. But
1092  * because of the pipe lock, we can check the condition before being on
1093  * the wait queue.
1094  *
1095  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1096  */
1097 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1098 {
1099         DEFINE_WAIT(rdwait);
1100         int cur = *cnt;
1101
1102         while (cur == *cnt) {
1103                 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1104                 pipe_unlock(pipe);
1105                 schedule();
1106                 finish_wait(&pipe->rd_wait, &rdwait);
1107                 pipe_lock(pipe);
1108                 if (signal_pending(current))
1109                         break;
1110         }
1111         return cur == *cnt ? -ERESTARTSYS : 0;
1112 }
1113
1114 static void wake_up_partner(struct pipe_inode_info *pipe)
1115 {
1116         wake_up_interruptible_all(&pipe->rd_wait);
1117 }
1118
1119 static int fifo_open(struct inode *inode, struct file *filp)
1120 {
1121         bool is_pipe = inode->i_fop == &pipeanon_fops;
1122         struct pipe_inode_info *pipe;
1123         int ret;
1124
1125         filp->f_pipe = 0;
1126
1127         spin_lock(&inode->i_lock);
1128         if (inode->i_pipe) {
1129                 pipe = inode->i_pipe;
1130                 pipe->files++;
1131                 spin_unlock(&inode->i_lock);
1132         } else {
1133                 spin_unlock(&inode->i_lock);
1134                 pipe = alloc_pipe_info();
1135                 if (!pipe)
1136                         return -ENOMEM;
1137                 pipe->files = 1;
1138                 spin_lock(&inode->i_lock);
1139                 if (unlikely(inode->i_pipe)) {
1140                         inode->i_pipe->files++;
1141                         spin_unlock(&inode->i_lock);
1142                         free_pipe_info(pipe);
1143                         pipe = inode->i_pipe;
1144                 } else {
1145                         inode->i_pipe = pipe;
1146                         spin_unlock(&inode->i_lock);
1147                 }
1148         }
1149         filp->private_data = pipe;
1150         /* OK, we have a pipe and it's pinned down */
1151
1152         mutex_lock(&pipe->mutex);
1153
1154         /* We can only do regular read/write on fifos */
1155         stream_open(inode, filp);
1156
1157         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1158         case FMODE_READ:
1159         /*
1160          *  O_RDONLY
1161          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1162          *  opened, even when there is no process writing the FIFO.
1163          */
1164                 pipe->r_counter++;
1165                 if (pipe->readers++ == 0)
1166                         wake_up_partner(pipe);
1167
1168                 if (!is_pipe && !pipe->writers) {
1169                         if ((filp->f_flags & O_NONBLOCK)) {
1170                                 /* suppress EPOLLHUP until we have
1171                                  * seen a writer */
1172                                 filp->f_pipe = pipe->w_counter;
1173                         } else {
1174                                 if (wait_for_partner(pipe, &pipe->w_counter))
1175                                         goto err_rd;
1176                         }
1177                 }
1178                 break;
1179
1180         case FMODE_WRITE:
1181         /*
1182          *  O_WRONLY
1183          *  POSIX.1 says that O_NONBLOCK means return -1 with
1184          *  errno=ENXIO when there is no process reading the FIFO.
1185          */
1186                 ret = -ENXIO;
1187                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1188                         goto err;
1189
1190                 pipe->w_counter++;
1191                 if (!pipe->writers++)
1192                         wake_up_partner(pipe);
1193
1194                 if (!is_pipe && !pipe->readers) {
1195                         if (wait_for_partner(pipe, &pipe->r_counter))
1196                                 goto err_wr;
1197                 }
1198                 break;
1199
1200         case FMODE_READ | FMODE_WRITE:
1201         /*
1202          *  O_RDWR
1203          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1204          *  This implementation will NEVER block on a O_RDWR open, since
1205          *  the process can at least talk to itself.
1206          */
1207
1208                 pipe->readers++;
1209                 pipe->writers++;
1210                 pipe->r_counter++;
1211                 pipe->w_counter++;
1212                 if (pipe->readers == 1 || pipe->writers == 1)
1213                         wake_up_partner(pipe);
1214                 break;
1215
1216         default:
1217                 ret = -EINVAL;
1218                 goto err;
1219         }
1220
1221         /* Ok! */
1222         mutex_unlock(&pipe->mutex);
1223         return 0;
1224
1225 err_rd:
1226         if (!--pipe->readers)
1227                 wake_up_interruptible(&pipe->wr_wait);
1228         ret = -ERESTARTSYS;
1229         goto err;
1230
1231 err_wr:
1232         if (!--pipe->writers)
1233                 wake_up_interruptible_all(&pipe->rd_wait);
1234         ret = -ERESTARTSYS;
1235         goto err;
1236
1237 err:
1238         mutex_unlock(&pipe->mutex);
1239
1240         put_pipe_info(inode, pipe);
1241         return ret;
1242 }
1243
1244 const struct file_operations pipefifo_fops = {
1245         .open           = fifo_open,
1246         .read_iter      = fifo_pipe_read,
1247         .write_iter     = fifo_pipe_write,
1248         .poll           = pipe_poll,
1249         .unlocked_ioctl = pipe_ioctl,
1250         .release        = pipe_release,
1251         .fasync         = pipe_fasync,
1252         .splice_write   = iter_file_splice_write,
1253 };
1254
1255 static const struct file_operations pipeanon_fops = {
1256         .open           = fifo_open,
1257         .read_iter      = anon_pipe_read,
1258         .write_iter     = anon_pipe_write,
1259         .poll           = pipe_poll,
1260         .unlocked_ioctl = pipe_ioctl,
1261         .release        = pipe_release,
1262         .fasync         = pipe_fasync,
1263         .splice_write   = iter_file_splice_write,
1264 };
1265
1266 /*
1267  * Currently we rely on the pipe array holding a power-of-2 number
1268  * of pages. Returns 0 on error.
1269  */
1270 unsigned int round_pipe_size(unsigned int size)
1271 {
1272         if (size > (1U << 31))
1273                 return 0;
1274
1275         /* Minimum pipe size, as required by POSIX */
1276         if (size < PAGE_SIZE)
1277                 return PAGE_SIZE;
1278
1279         return roundup_pow_of_two(size);
1280 }
1281
1282 /*
1283  * Resize the pipe ring to a number of slots.
1284  *
1285  * Note the pipe can be reduced in capacity, but only if the current
1286  * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1287  * returned instead.
1288  */
1289 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1290 {
1291         struct pipe_buffer *bufs;
1292         unsigned int head, tail, mask, n;
1293
1294         /* nr_slots larger than limits of pipe->{head,tail} */
1295         if (unlikely(nr_slots > (pipe_index_t)-1u))
1296                 return -EINVAL;
1297
1298         bufs = kcalloc(nr_slots, sizeof(*bufs),
1299                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1300         if (unlikely(!bufs))
1301                 return -ENOMEM;
1302
1303         spin_lock_irq(&pipe->rd_wait.lock);
1304         mask = pipe->ring_size - 1;
1305         head = pipe->head;
1306         tail = pipe->tail;
1307
1308         n = pipe_occupancy(head, tail);
1309         if (nr_slots < n) {
1310                 spin_unlock_irq(&pipe->rd_wait.lock);
1311                 kfree(bufs);
1312                 return -EBUSY;
1313         }
1314
1315         /*
1316          * The pipe array wraps around, so just start the new one at zero
1317          * and adjust the indices.
1318          */
1319         if (n > 0) {
1320                 unsigned int h = head & mask;
1321                 unsigned int t = tail & mask;
1322                 if (h > t) {
1323                         memcpy(bufs, pipe->bufs + t,
1324                                n * sizeof(struct pipe_buffer));
1325                 } else {
1326                         unsigned int tsize = pipe->ring_size - t;
1327                         if (h > 0)
1328                                 memcpy(bufs + tsize, pipe->bufs,
1329                                        h * sizeof(struct pipe_buffer));
1330                         memcpy(bufs, pipe->bufs + t,
1331                                tsize * sizeof(struct pipe_buffer));
1332                 }
1333         }
1334
1335         head = n;
1336         tail = 0;
1337
1338         kfree(pipe->bufs);
1339         pipe->bufs = bufs;
1340         pipe->ring_size = nr_slots;
1341         if (pipe->max_usage > nr_slots)
1342                 pipe->max_usage = nr_slots;
1343         pipe->tail = tail;
1344         pipe->head = head;
1345
1346         if (!pipe_has_watch_queue(pipe)) {
1347                 pipe->max_usage = nr_slots;
1348                 pipe->nr_accounted = nr_slots;
1349         }
1350
1351         spin_unlock_irq(&pipe->rd_wait.lock);
1352
1353         /* This might have made more room for writers */
1354         wake_up_interruptible(&pipe->wr_wait);
1355         return 0;
1356 }
1357
1358 /*
1359  * Allocate a new array of pipe buffers and copy the info over. Returns the
1360  * pipe size if successful, or return -ERROR on error.
1361  */
1362 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1363 {
1364         unsigned long user_bufs;
1365         unsigned int nr_slots, size;
1366         long ret = 0;
1367
1368         if (pipe_has_watch_queue(pipe))
1369                 return -EBUSY;
1370
1371         size = round_pipe_size(arg);
1372         nr_slots = size >> PAGE_SHIFT;
1373
1374         if (!nr_slots)
1375                 return -EINVAL;
1376
1377         /*
1378          * If trying to increase the pipe capacity, check that an
1379          * unprivileged user is not trying to exceed various limits
1380          * (soft limit check here, hard limit check just below).
1381          * Decreasing the pipe capacity is always permitted, even
1382          * if the user is currently over a limit.
1383          */
1384         if (nr_slots > pipe->max_usage &&
1385                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1386                 return -EPERM;
1387
1388         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1389
1390         if (nr_slots > pipe->max_usage &&
1391                         (too_many_pipe_buffers_hard(user_bufs) ||
1392                          too_many_pipe_buffers_soft(user_bufs)) &&
1393                         pipe_is_unprivileged_user()) {
1394                 ret = -EPERM;
1395                 goto out_revert_acct;
1396         }
1397
1398         ret = pipe_resize_ring(pipe, nr_slots);
1399         if (ret < 0)
1400                 goto out_revert_acct;
1401
1402         return pipe->max_usage * PAGE_SIZE;
1403
1404 out_revert_acct:
1405         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1406         return ret;
1407 }
1408
1409 /*
1410  * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1411  * not enough to verify that this is a pipe.
1412  */
1413 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1414 {
1415         struct pipe_inode_info *pipe = file->private_data;
1416
1417         if (!pipe)
1418                 return NULL;
1419         if (file->f_op != &pipefifo_fops && file->f_op != &pipeanon_fops)
1420                 return NULL;
1421         if (for_splice && pipe_has_watch_queue(pipe))
1422                 return NULL;
1423         return pipe;
1424 }
1425
1426 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1427 {
1428         struct pipe_inode_info *pipe;
1429         long ret;
1430
1431         pipe = get_pipe_info(file, false);
1432         if (!pipe)
1433                 return -EBADF;
1434
1435         mutex_lock(&pipe->mutex);
1436
1437         switch (cmd) {
1438         case F_SETPIPE_SZ:
1439                 ret = pipe_set_size(pipe, arg);
1440                 break;
1441         case F_GETPIPE_SZ:
1442                 ret = pipe->max_usage * PAGE_SIZE;
1443                 break;
1444         default:
1445                 ret = -EINVAL;
1446                 break;
1447         }
1448
1449         mutex_unlock(&pipe->mutex);
1450         return ret;
1451 }
1452
1453 static const struct super_operations pipefs_ops = {
1454         .destroy_inode = free_inode_nonrcu,
1455         .statfs = simple_statfs,
1456 };
1457
1458 /*
1459  * pipefs should _never_ be mounted by userland - too much of security hassle,
1460  * no real gain from having the whole file system mounted. So we don't need
1461  * any operations on the root directory. However, we need a non-trivial
1462  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1463  */
1464
1465 static int pipefs_init_fs_context(struct fs_context *fc)
1466 {
1467         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1468         if (!ctx)
1469                 return -ENOMEM;
1470         ctx->ops = &pipefs_ops;
1471         ctx->dops = &pipefs_dentry_operations;
1472         return 0;
1473 }
1474
1475 static struct file_system_type pipe_fs_type = {
1476         .name           = "pipefs",
1477         .init_fs_context = pipefs_init_fs_context,
1478         .kill_sb        = kill_anon_super,
1479 };
1480
1481 #ifdef CONFIG_SYSCTL
1482 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1483                                         unsigned int *valp,
1484                                         int write, void *data)
1485 {
1486         if (write) {
1487                 unsigned int val;
1488
1489                 val = round_pipe_size(*lvalp);
1490                 if (val == 0)
1491                         return -EINVAL;
1492
1493                 *valp = val;
1494         } else {
1495                 unsigned int val = *valp;
1496                 *lvalp = (unsigned long) val;
1497         }
1498
1499         return 0;
1500 }
1501
1502 static int proc_dopipe_max_size(const struct ctl_table *table, int write,
1503                                 void *buffer, size_t *lenp, loff_t *ppos)
1504 {
1505         return do_proc_douintvec(table, write, buffer, lenp, ppos,
1506                                  do_proc_dopipe_max_size_conv, NULL);
1507 }
1508
1509 static const struct ctl_table fs_pipe_sysctls[] = {
1510         {
1511                 .procname       = "pipe-max-size",
1512                 .data           = &pipe_max_size,
1513                 .maxlen         = sizeof(pipe_max_size),
1514                 .mode           = 0644,
1515                 .proc_handler   = proc_dopipe_max_size,
1516         },
1517         {
1518                 .procname       = "pipe-user-pages-hard",
1519                 .data           = &pipe_user_pages_hard,
1520                 .maxlen         = sizeof(pipe_user_pages_hard),
1521                 .mode           = 0644,
1522                 .proc_handler   = proc_doulongvec_minmax,
1523         },
1524         {
1525                 .procname       = "pipe-user-pages-soft",
1526                 .data           = &pipe_user_pages_soft,
1527                 .maxlen         = sizeof(pipe_user_pages_soft),
1528                 .mode           = 0644,
1529                 .proc_handler   = proc_doulongvec_minmax,
1530         },
1531 };
1532 #endif
1533
1534 static int __init init_pipe_fs(void)
1535 {
1536         int err = register_filesystem(&pipe_fs_type);
1537
1538         if (!err) {
1539                 pipe_mnt = kern_mount(&pipe_fs_type);
1540                 if (IS_ERR(pipe_mnt)) {
1541                         err = PTR_ERR(pipe_mnt);
1542                         unregister_filesystem(&pipe_fs_type);
1543                 }
1544         }
1545 #ifdef CONFIG_SYSCTL
1546         register_sysctl_init("fs", fs_pipe_sysctls);
1547 #endif
1548         return err;
1549 }
1550
1551 fs_initcall(init_pipe_fs);