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
  28 #include <linux/uaccess.h>
  29 #include <asm/ioctls.h>
  30
  31 #include "internal.h"
  32
  33 /*
  34  * The max size that a non-root user is allowed to grow the pipe. Can
  35  * be set by root in /proc/sys/fs/pipe-max-size
  36  */
  37 unsigned int pipe_max_size = 1048576;
  38
  39 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  40  * matches default values.
  41  */
  42 unsigned long pipe_user_pages_hard;
  43 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  44
  45 /*
  46  * We use head and tail indices that aren't masked off, except at the point of
  47  * dereference, but rather they're allowed to wrap naturally.  This means there
  48  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  49  * <= 2^31.
  50  * -- David Howells 2019-09-23.
  51  *
  52  * Reads with count = 0 should always return 0.
  53  * -- Julian Bradfield 1999-06-07.
  54  *
  55  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  56  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  57  *
  58  * pipe_read & write cleanup
  59  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  60  */
  61
  62 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  63 {
  64         if (pipe->files)
  65                 mutex_lock_nested(&pipe->mutex, subclass);
  66 }
  67
  68 void pipe_lock(struct pipe_inode_info *pipe)
  69 {
  70         /*
  71          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  72          */
  73         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  74 }
  75 EXPORT_SYMBOL(pipe_lock);
  76
  77 void pipe_unlock(struct pipe_inode_info *pipe)
  78 {
  79         if (pipe->files)
  80                 mutex_unlock(&pipe->mutex);
  81 }
  82 EXPORT_SYMBOL(pipe_unlock);
  83
  84 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  85 {
  86         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  87 }
  88
  89 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  90 {
  91         mutex_unlock(&pipe->mutex);
  92 }
  93
  94 void pipe_double_lock(struct pipe_inode_info *pipe1,
  95                       struct pipe_inode_info *pipe2)
  96 {
  97         BUG_ON(pipe1 == pipe2);
  98
  99         if (pipe1 < pipe2) {
 100                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 101                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 102         } else {
 103                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 104                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 105         }
 106 }
 107
 108 /* Drop the inode semaphore and wait for a pipe event, atomically */
 109 void pipe_wait(struct pipe_inode_info *pipe)
 110 {
 111         DEFINE_WAIT(wait);
 112
 113         /*
 114          * Pipes are system-local resources, so sleeping on them
 115          * is considered a noninteractive wait:
 116          */
 117         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 118         pipe_unlock(pipe);
 119         schedule();
 120         finish_wait(&pipe->wait, &wait);
 121         pipe_lock(pipe);
 122 }
 123
 124 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 125                                   struct pipe_buffer *buf)
 126 {
 127         struct page *page = buf->page;
 128
 129         /*
 130          * If nobody else uses this page, and we don't already have a
 131          * temporary page, let's keep track of it as a one-deep
 132          * allocation cache. (Otherwise just release our reference to it)
 133          */
 134         if (page_count(page) == 1 && !pipe->tmp_page)
 135                 pipe->tmp_page = page;
 136         else
 137                 put_page(page);
 138 }
 139
 140 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 141                                struct pipe_buffer *buf)
 142 {
 143         struct page *page = buf->page;
 144
 145         if (page_count(page) == 1) {
 146                 memcg_kmem_uncharge(page, 0);
 147                 __SetPageLocked(page);
 148                 return 0;
 149         }
 150         return 1;
 151 }
 152
 153 /**
 154  * generic_pipe_buf_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 int generic_pipe_buf_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 0;
 178         }
 179
 180         return 1;
 181 }
 182 EXPORT_SYMBOL(generic_pipe_buf_steal);
 183
 184 /**
 185  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 186  * @pipe:       the pipe that the buffer belongs to
 187  * @buf:        the buffer to get a reference to
 188  *
 189  * Description:
 190  *      This function grabs an extra reference to @buf. It's used in
 191  *      in the tee() system call, when we duplicate the buffers in one
 192  *      pipe into another.
 193  */
 194 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 195 {
 196         return try_get_page(buf->page);
 197 }
 198 EXPORT_SYMBOL(generic_pipe_buf_get);
 199
 200 /**
 201  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 202  * @info:       the pipe that the buffer belongs to
 203  * @buf:        the buffer to confirm
 204  *
 205  * Description:
 206  *      This function does nothing, because the generic pipe code uses
 207  *      pages that are always good when inserted into the pipe.
 208  */
 209 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 210                              struct pipe_buffer *buf)
 211 {
 212         return 0;
 213 }
 214 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 215
 216 /**
 217  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 218  * @pipe:       the pipe that the buffer belongs to
 219  * @buf:        the buffer to put a reference to
 220  *
 221  * Description:
 222  *      This function releases a reference to @buf.
 223  */
 224 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 225                               struct pipe_buffer *buf)
 226 {
 227         put_page(buf->page);
 228 }
 229 EXPORT_SYMBOL(generic_pipe_buf_release);
 230
 231 /* New data written to a pipe may be appended to a buffer with this type. */
 232 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 233         .confirm = generic_pipe_buf_confirm,
 234         .release = anon_pipe_buf_release,
 235         .steal = anon_pipe_buf_steal,
 236         .get = generic_pipe_buf_get,
 237 };
 238
 239 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 240         .confirm = generic_pipe_buf_confirm,
 241         .release = anon_pipe_buf_release,
 242         .steal = anon_pipe_buf_steal,
 243         .get = generic_pipe_buf_get,
 244 };
 245
 246 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 247         .confirm = generic_pipe_buf_confirm,
 248         .release = anon_pipe_buf_release,
 249         .steal = anon_pipe_buf_steal,
 250         .get = generic_pipe_buf_get,
 251 };
 252
 253 /**
 254  * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
 255  * @buf:        the buffer to mark
 256  *
 257  * Description:
 258  *      This function ensures that no future writes will be merged into the
 259  *      given &struct pipe_buffer. This is necessary when multiple pipe buffers
 260  *      share the same backing page.
 261  */
 262 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 263 {
 264         if (buf->ops == &anon_pipe_buf_ops)
 265                 buf->ops = &anon_pipe_buf_nomerge_ops;
 266 }
 267
 268 static bool pipe_buf_can_merge(struct pipe_buffer *buf)
 269 {
 270         return buf->ops == &anon_pipe_buf_ops;
 271 }
 272
 273 static ssize_t
 274 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 275 {
 276         size_t total_len = iov_iter_count(to);
 277         struct file *filp = iocb->ki_filp;
 278         struct pipe_inode_info *pipe = filp->private_data;
 279         bool was_full;
 280         ssize_t ret;
 281
 282         /* Null read succeeds. */
 283         if (unlikely(total_len == 0))
 284                 return 0;
 285
 286         ret = 0;
 287         __pipe_lock(pipe);
 288
 289         /*
 290          * We only wake up writers if the pipe was full when we started
 291          * reading in order to avoid unnecessary wakeups.
 292          *
 293          * But when we do wake up writers, we do so using a sync wakeup
 294          * (WF_SYNC), because we want them to get going and generate more
 295          * data for us.
 296          */
 297         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 298         for (;;) {
 299                 unsigned int head = pipe->head;
 300                 unsigned int tail = pipe->tail;
 301                 unsigned int mask = pipe->ring_size - 1;
 302
 303                 if (!pipe_empty(head, tail)) {
 304                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 305                         size_t chars = buf->len;
 306                         size_t written;
 307                         int error;
 308
 309                         if (chars > total_len)
 310                                 chars = total_len;
 311
 312                         error = pipe_buf_confirm(pipe, buf);
 313                         if (error) {
 314                                 if (!ret)
 315                                         ret = error;
 316                                 break;
 317                         }
 318
 319                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 320                         if (unlikely(written < chars)) {
 321                                 if (!ret)
 322                                         ret = -EFAULT;
 323                                 break;
 324                         }
 325                         ret += chars;
 326                         buf->offset += chars;
 327                         buf->len -= chars;
 328
 329                         /* Was it a packet buffer? Clean up and exit */
 330                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 331                                 total_len = chars;
 332                                 buf->len = 0;
 333                         }
 334
 335                         if (!buf->len) {
 336                                 pipe_buf_release(pipe, buf);
 337                                 spin_lock_irq(&pipe->wait.lock);
 338                                 tail++;
 339                                 pipe->tail = tail;
 340                                 spin_unlock_irq(&pipe->wait.lock);
 341                         }
 342                         total_len -= chars;
 343                         if (!total_len)
 344                                 break;  /* common path: read succeeded */
 345                         if (!pipe_empty(head, tail))    /* More to do? */
 346                                 continue;
 347                 }
 348
 349                 if (!pipe->writers)
 350                         break;
 351                 if (ret)
 352                         break;
 353                 if (filp->f_flags & O_NONBLOCK) {
 354                         ret = -EAGAIN;
 355                         break;
 356                 }
 357                 if (signal_pending(current)) {
 358                         if (!ret)
 359                                 ret = -ERESTARTSYS;
 360                         break;
 361                 }
 362                 if (was_full) {
 363                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 364                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 365                 }
 366                 pipe_wait(pipe);
 367                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 368         }
 369         __pipe_unlock(pipe);
 370
 371         if (was_full) {
 372                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 373                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 374         }
 375         if (ret > 0)
 376                 file_accessed(filp);
 377         return ret;
 378 }
 379
 380 static inline int is_packetized(struct file *file)
 381 {
 382         return (file->f_flags & O_DIRECT) != 0;
 383 }
 384
 385 static ssize_t
 386 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 387 {
 388         struct file *filp = iocb->ki_filp;
 389         struct pipe_inode_info *pipe = filp->private_data;
 390         unsigned int head;
 391         ssize_t ret = 0;
 392         size_t total_len = iov_iter_count(from);
 393         ssize_t chars;
 394         bool was_empty = false;
 395
 396         /* Null write succeeds. */
 397         if (unlikely(total_len == 0))
 398                 return 0;
 399
 400         __pipe_lock(pipe);
 401
 402         if (!pipe->readers) {
 403                 send_sig(SIGPIPE, current, 0);
 404                 ret = -EPIPE;
 405                 goto out;
 406         }
 407
 408         /*
 409          * Only wake up if the pipe started out empty, since
 410          * otherwise there should be no readers waiting.
 411          *
 412          * If it wasn't empty we try to merge new data into
 413          * the last buffer.
 414          *
 415          * That naturally merges small writes, but it also
 416          * page-aligs the rest of the writes for large writes
 417          * spanning multiple pages.
 418          */
 419         head = pipe->head;
 420         was_empty = pipe_empty(head, pipe->tail);
 421         chars = total_len & (PAGE_SIZE-1);
 422         if (chars && !was_empty) {
 423                 unsigned int mask = pipe->ring_size - 1;
 424                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 425                 int offset = buf->offset + buf->len;
 426
 427                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 428                         ret = pipe_buf_confirm(pipe, buf);
 429                         if (ret)
 430                                 goto out;
 431
 432                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 433                         if (unlikely(ret < chars)) {
 434                                 ret = -EFAULT;
 435                                 goto out;
 436                         }
 437
 438                         buf->len += ret;
 439                         if (!iov_iter_count(from))
 440                                 goto out;
 441                 }
 442         }
 443
 444         for (;;) {
 445                 if (!pipe->readers) {
 446                         send_sig(SIGPIPE, current, 0);
 447                         if (!ret)
 448                                 ret = -EPIPE;
 449                         break;
 450                 }
 451
 452                 head = pipe->head;
 453                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 454                         unsigned int mask = pipe->ring_size - 1;
 455                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 456                         struct page *page = pipe->tmp_page;
 457                         int copied;
 458
 459                         if (!page) {
 460                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 461                                 if (unlikely(!page)) {
 462                                         ret = ret ? : -ENOMEM;
 463                                         break;
 464                                 }
 465                                 pipe->tmp_page = page;
 466                         }
 467
 468                         /* Allocate a slot in the ring in advance and attach an
 469                          * empty buffer.  If we fault or otherwise fail to use
 470                          * it, either the reader will consume it or it'll still
 471                          * be there for the next write.
 472                          */
 473                         spin_lock_irq(&pipe->wait.lock);
 474
 475                         head = pipe->head;
 476                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 477                                 spin_unlock_irq(&pipe->wait.lock);
 478                                 continue;
 479                         }
 480
 481                         pipe->head = head + 1;
 482                         spin_unlock_irq(&pipe->wait.lock);
 483
 484                         /* Insert it into the buffer array */
 485                         buf = &pipe->bufs[head & mask];
 486                         buf->page = page;
 487                         buf->ops = &anon_pipe_buf_ops;
 488                         buf->offset = 0;
 489                         buf->len = 0;
 490                         buf->flags = 0;
 491                         if (is_packetized(filp)) {
 492                                 buf->ops = &packet_pipe_buf_ops;
 493                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 494                         }
 495                         pipe->tmp_page = NULL;
 496
 497                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 498                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 499                                 if (!ret)
 500                                         ret = -EFAULT;
 501                                 break;
 502                         }
 503                         ret += copied;
 504                         buf->offset = 0;
 505                         buf->len = copied;
 506
 507                         if (!iov_iter_count(from))
 508                                 break;
 509                 }
 510
 511                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 512                         continue;
 513
 514                 /* Wait for buffer space to become available. */
 515                 if (filp->f_flags & O_NONBLOCK) {
 516                         if (!ret)
 517                                 ret = -EAGAIN;
 518                         break;
 519                 }
 520                 if (signal_pending(current)) {
 521                         if (!ret)
 522                                 ret = -ERESTARTSYS;
 523                         break;
 524                 }
 525
 526                 /*
 527                  * We're going to release the pipe lock and wait for more
 528                  * space. We wake up any readers if necessary, and then
 529                  * after waiting we need to re-check whether the pipe
 530                  * become empty while we dropped the lock.
 531                  */
 532                 if (was_empty) {
 533                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 534                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 535                 }
 536                 pipe_wait(pipe);
 537
 538                 was_empty = pipe_empty(head, pipe->tail);
 539         }
 540 out:
 541         __pipe_unlock(pipe);
 542
 543         /*
 544          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 545          * want the reader to start processing things asap, rather than
 546          * leave the data pending.
 547          *
 548          * This is particularly important for small writes, because of
 549          * how (for example) the GNU make jobserver uses small writes to
 550          * wake up pending jobs
 551          */
 552         if (was_empty) {
 553                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 554                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 555         }
 556         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 557                 int err = file_update_time(filp);
 558                 if (err)
 559                         ret = err;
 560                 sb_end_write(file_inode(filp)->i_sb);
 561         }
 562         return ret;
 563 }
 564
 565 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 566 {
 567         struct pipe_inode_info *pipe = filp->private_data;
 568         int count, head, tail, mask;
 569
 570         switch (cmd) {
 571                 case FIONREAD:
 572                         __pipe_lock(pipe);
 573                         count = 0;
 574                         head = pipe->head;
 575                         tail = pipe->tail;
 576                         mask = pipe->ring_size - 1;
 577
 578                         while (tail != head) {
 579                                 count += pipe->bufs[tail & mask].len;
 580                                 tail++;
 581                         }
 582                         __pipe_unlock(pipe);
 583
 584                         return put_user(count, (int __user *)arg);
 585                 default:
 586                         return -ENOIOCTLCMD;
 587         }
 588 }
 589
 590 /* No kernel lock held - fine */
 591 static __poll_t
 592 pipe_poll(struct file *filp, poll_table *wait)
 593 {
 594         __poll_t mask;
 595         struct pipe_inode_info *pipe = filp->private_data;
 596         unsigned int head, tail;
 597
 598         /*
 599          * Reading only -- no need for acquiring the semaphore.
 600          *
 601          * But because this is racy, the code has to add the
 602          * entry to the poll table _first_ ..
 603          */
 604         poll_wait(filp, &pipe->wait, wait);
 605
 606         /*
 607          * .. and only then can you do the racy tests. That way,
 608          * if something changes and you got it wrong, the poll
 609          * table entry will wake you up and fix it.
 610          */
 611         head = READ_ONCE(pipe->head);
 612         tail = READ_ONCE(pipe->tail);
 613
 614         mask = 0;
 615         if (filp->f_mode & FMODE_READ) {
 616                 if (!pipe_empty(head, tail))
 617                         mask |= EPOLLIN | EPOLLRDNORM;
 618                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 619                         mask |= EPOLLHUP;
 620         }
 621
 622         if (filp->f_mode & FMODE_WRITE) {
 623                 if (!pipe_full(head, tail, pipe->max_usage))
 624                         mask |= EPOLLOUT | EPOLLWRNORM;
 625                 /*
 626                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 627                  * behave exactly like pipes for poll().
 628                  */
 629                 if (!pipe->readers)
 630                         mask |= EPOLLERR;
 631         }
 632
 633         return mask;
 634 }
 635
 636 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 637 {
 638         int kill = 0;
 639
 640         spin_lock(&inode->i_lock);
 641         if (!--pipe->files) {
 642                 inode->i_pipe = NULL;
 643                 kill = 1;
 644         }
 645         spin_unlock(&inode->i_lock);
 646
 647         if (kill)
 648                 free_pipe_info(pipe);
 649 }
 650
 651 static int
 652 pipe_release(struct inode *inode, struct file *file)
 653 {
 654         struct pipe_inode_info *pipe = file->private_data;
 655
 656         __pipe_lock(pipe);
 657         if (file->f_mode & FMODE_READ)
 658                 pipe->readers--;
 659         if (file->f_mode & FMODE_WRITE)
 660                 pipe->writers--;
 661
 662         if (pipe->readers || pipe->writers) {
 663                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 664                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 665                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 666         }
 667         __pipe_unlock(pipe);
 668
 669         put_pipe_info(inode, pipe);
 670         return 0;
 671 }
 672
 673 static int
 674 pipe_fasync(int fd, struct file *filp, int on)
 675 {
 676         struct pipe_inode_info *pipe = filp->private_data;
 677         int retval = 0;
 678
 679         __pipe_lock(pipe);
 680         if (filp->f_mode & FMODE_READ)
 681                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 682         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 683                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 684                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 685                         /* this can happen only if on == T */
 686                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 687         }
 688         __pipe_unlock(pipe);
 689         return retval;
 690 }
 691
 692 static unsigned long account_pipe_buffers(struct user_struct *user,
 693                                  unsigned long old, unsigned long new)
 694 {
 695         return atomic_long_add_return(new - old, &user->pipe_bufs);
 696 }
 697
 698 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 699 {
 700         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 701
 702         return soft_limit && user_bufs > soft_limit;
 703 }
 704
 705 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 706 {
 707         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 708
 709         return hard_limit && user_bufs > hard_limit;
 710 }
 711
 712 static bool is_unprivileged_user(void)
 713 {
 714         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 715 }
 716
 717 struct pipe_inode_info *alloc_pipe_info(void)
 718 {
 719         struct pipe_inode_info *pipe;
 720         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 721         struct user_struct *user = get_current_user();
 722         unsigned long user_bufs;
 723         unsigned int max_size = READ_ONCE(pipe_max_size);
 724
 725         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 726         if (pipe == NULL)
 727                 goto out_free_uid;
 728
 729         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 730                 pipe_bufs = max_size >> PAGE_SHIFT;
 731
 732         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 733
 734         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 735                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 736                 pipe_bufs = 1;
 737         }
 738
 739         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 740                 goto out_revert_acct;
 741
 742         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 743                              GFP_KERNEL_ACCOUNT);
 744
 745         if (pipe->bufs) {
 746                 init_waitqueue_head(&pipe->wait);
 747                 pipe->r_counter = pipe->w_counter = 1;
 748                 pipe->max_usage = pipe_bufs;
 749                 pipe->ring_size = pipe_bufs;
 750                 pipe->user = user;
 751                 mutex_init(&pipe->mutex);
 752                 return pipe;
 753         }
 754
 755 out_revert_acct:
 756         (void) account_pipe_buffers(user, pipe_bufs, 0);
 757         kfree(pipe);
 758 out_free_uid:
 759         free_uid(user);
 760         return NULL;
 761 }
 762
 763 void free_pipe_info(struct pipe_inode_info *pipe)
 764 {
 765         int i;
 766
 767         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 768         free_uid(pipe->user);
 769         for (i = 0; i < pipe->ring_size; i++) {
 770                 struct pipe_buffer *buf = pipe->bufs + i;
 771                 if (buf->ops)
 772                         pipe_buf_release(pipe, buf);
 773         }
 774         if (pipe->tmp_page)
 775                 __free_page(pipe->tmp_page);
 776         kfree(pipe->bufs);
 777         kfree(pipe);
 778 }
 779
 780 static struct vfsmount *pipe_mnt __read_mostly;
 781
 782 /*
 783  * pipefs_dname() is called from d_path().
 784  */
 785 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 786 {
 787         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 788                                 d_inode(dentry)->i_ino);
 789 }
 790
 791 static const struct dentry_operations pipefs_dentry_operations = {
 792         .d_dname        = pipefs_dname,
 793 };
 794
 795 static struct inode * get_pipe_inode(void)
 796 {
 797         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 798         struct pipe_inode_info *pipe;
 799
 800         if (!inode)
 801                 goto fail_inode;
 802
 803         inode->i_ino = get_next_ino();
 804
 805         pipe = alloc_pipe_info();
 806         if (!pipe)
 807                 goto fail_iput;
 808
 809         inode->i_pipe = pipe;
 810         pipe->files = 2;
 811         pipe->readers = pipe->writers = 1;
 812         inode->i_fop = &pipefifo_fops;
 813
 814         /*
 815          * Mark the inode dirty from the very beginning,
 816          * that way it will never be moved to the dirty
 817          * list because "mark_inode_dirty()" will think
 818          * that it already _is_ on the dirty list.
 819          */
 820         inode->i_state = I_DIRTY;
 821         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 822         inode->i_uid = current_fsuid();
 823         inode->i_gid = current_fsgid();
 824         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 825
 826         return inode;
 827
 828 fail_iput:
 829         iput(inode);
 830
 831 fail_inode:
 832         return NULL;
 833 }
 834
 835 int create_pipe_files(struct file **res, int flags)
 836 {
 837         struct inode *inode = get_pipe_inode();
 838         struct file *f;
 839
 840         if (!inode)
 841                 return -ENFILE;
 842
 843         f = alloc_file_pseudo(inode, pipe_mnt, "",
 844                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 845                                 &pipefifo_fops);
 846         if (IS_ERR(f)) {
 847                 free_pipe_info(inode->i_pipe);
 848                 iput(inode);
 849                 return PTR_ERR(f);
 850         }
 851
 852         f->private_data = inode->i_pipe;
 853
 854         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 855                                   &pipefifo_fops);
 856         if (IS_ERR(res[0])) {
 857                 put_pipe_info(inode, inode->i_pipe);
 858                 fput(f);
 859                 return PTR_ERR(res[0]);
 860         }
 861         res[0]->private_data = inode->i_pipe;
 862         res[1] = f;
 863         stream_open(inode, res[0]);
 864         stream_open(inode, res[1]);
 865         return 0;
 866 }
 867
 868 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 869 {
 870         int error;
 871         int fdw, fdr;
 872
 873         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 874                 return -EINVAL;
 875
 876         error = create_pipe_files(files, flags);
 877         if (error)
 878                 return error;
 879
 880         error = get_unused_fd_flags(flags);
 881         if (error < 0)
 882                 goto err_read_pipe;
 883         fdr = error;
 884
 885         error = get_unused_fd_flags(flags);
 886         if (error < 0)
 887                 goto err_fdr;
 888         fdw = error;
 889
 890         audit_fd_pair(fdr, fdw);
 891         fd[0] = fdr;
 892         fd[1] = fdw;
 893         return 0;
 894
 895  err_fdr:
 896         put_unused_fd(fdr);
 897  err_read_pipe:
 898         fput(files[0]);
 899         fput(files[1]);
 900         return error;
 901 }
 902
 903 int do_pipe_flags(int *fd, int flags)
 904 {
 905         struct file *files[2];
 906         int error = __do_pipe_flags(fd, files, flags);
 907         if (!error) {
 908                 fd_install(fd[0], files[0]);
 909                 fd_install(fd[1], files[1]);
 910         }
 911         return error;
 912 }
 913
 914 /*
 915  * sys_pipe() is the normal C calling standard for creating
 916  * a pipe. It's not the way Unix traditionally does this, though.
 917  */
 918 static int do_pipe2(int __user *fildes, int flags)
 919 {
 920         struct file *files[2];
 921         int fd[2];
 922         int error;
 923
 924         error = __do_pipe_flags(fd, files, flags);
 925         if (!error) {
 926                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 927                         fput(files[0]);
 928                         fput(files[1]);
 929                         put_unused_fd(fd[0]);
 930                         put_unused_fd(fd[1]);
 931                         error = -EFAULT;
 932                 } else {
 933                         fd_install(fd[0], files[0]);
 934                         fd_install(fd[1], files[1]);
 935                 }
 936         }
 937         return error;
 938 }
 939
 940 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 941 {
 942         return do_pipe2(fildes, flags);
 943 }
 944
 945 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 946 {
 947         return do_pipe2(fildes, 0);
 948 }
 949
 950 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 951 {
 952         int cur = *cnt;
 953
 954         while (cur == *cnt) {
 955                 pipe_wait(pipe);
 956                 if (signal_pending(current))
 957                         break;
 958         }
 959         return cur == *cnt ? -ERESTARTSYS : 0;
 960 }
 961
 962 static void wake_up_partner(struct pipe_inode_info *pipe)
 963 {
 964         wake_up_interruptible(&pipe->wait);
 965 }
 966
 967 static int fifo_open(struct inode *inode, struct file *filp)
 968 {
 969         struct pipe_inode_info *pipe;
 970         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 971         int ret;
 972
 973         filp->f_version = 0;
 974
 975         spin_lock(&inode->i_lock);
 976         if (inode->i_pipe) {
 977                 pipe = inode->i_pipe;
 978                 pipe->files++;
 979                 spin_unlock(&inode->i_lock);
 980         } else {
 981                 spin_unlock(&inode->i_lock);
 982                 pipe = alloc_pipe_info();
 983                 if (!pipe)
 984                         return -ENOMEM;
 985                 pipe->files = 1;
 986                 spin_lock(&inode->i_lock);
 987                 if (unlikely(inode->i_pipe)) {
 988                         inode->i_pipe->files++;
 989                         spin_unlock(&inode->i_lock);
 990                         free_pipe_info(pipe);
 991                         pipe = inode->i_pipe;
 992                 } else {
 993                         inode->i_pipe = pipe;
 994                         spin_unlock(&inode->i_lock);
 995                 }
 996         }
 997         filp->private_data = pipe;
 998         /* OK, we have a pipe and it's pinned down */
 999
1000         __pipe_lock(pipe);
1001
1002         /* We can only do regular read/write on fifos */
1003         stream_open(inode, filp);
1004
1005         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1006         case FMODE_READ:
1007         /*
1008          *  O_RDONLY
1009          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1010          *  opened, even when there is no process writing the FIFO.
1011          */
1012                 pipe->r_counter++;
1013                 if (pipe->readers++ == 0)
1014                         wake_up_partner(pipe);
1015
1016                 if (!is_pipe && !pipe->writers) {
1017                         if ((filp->f_flags & O_NONBLOCK)) {
1018                                 /* suppress EPOLLHUP until we have
1019                                  * seen a writer */
1020                                 filp->f_version = pipe->w_counter;
1021                         } else {
1022                                 if (wait_for_partner(pipe, &pipe->w_counter))
1023                                         goto err_rd;
1024                         }
1025                 }
1026                 break;
1027
1028         case FMODE_WRITE:
1029         /*
1030          *  O_WRONLY
1031          *  POSIX.1 says that O_NONBLOCK means return -1 with
1032          *  errno=ENXIO when there is no process reading the FIFO.
1033          */
1034                 ret = -ENXIO;
1035                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1036                         goto err;
1037
1038                 pipe->w_counter++;
1039                 if (!pipe->writers++)
1040                         wake_up_partner(pipe);
1041
1042                 if (!is_pipe && !pipe->readers) {
1043                         if (wait_for_partner(pipe, &pipe->r_counter))
1044                                 goto err_wr;
1045                 }
1046                 break;
1047
1048         case FMODE_READ | FMODE_WRITE:
1049         /*
1050          *  O_RDWR
1051          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1052          *  This implementation will NEVER block on a O_RDWR open, since
1053          *  the process can at least talk to itself.
1054          */
1055
1056                 pipe->readers++;
1057                 pipe->writers++;
1058                 pipe->r_counter++;
1059                 pipe->w_counter++;
1060                 if (pipe->readers == 1 || pipe->writers == 1)
1061                         wake_up_partner(pipe);
1062                 break;
1063
1064         default:
1065                 ret = -EINVAL;
1066                 goto err;
1067         }
1068
1069         /* Ok! */
1070         __pipe_unlock(pipe);
1071         return 0;
1072
1073 err_rd:
1074         if (!--pipe->readers)
1075                 wake_up_interruptible(&pipe->wait);
1076         ret = -ERESTARTSYS;
1077         goto err;
1078
1079 err_wr:
1080         if (!--pipe->writers)
1081                 wake_up_interruptible(&pipe->wait);
1082         ret = -ERESTARTSYS;
1083         goto err;
1084
1085 err:
1086         __pipe_unlock(pipe);
1087
1088         put_pipe_info(inode, pipe);
1089         return ret;
1090 }
1091
1092 const struct file_operations pipefifo_fops = {
1093         .open           = fifo_open,
1094         .llseek         = no_llseek,
1095         .read_iter      = pipe_read,
1096         .write_iter     = pipe_write,
1097         .poll           = pipe_poll,
1098         .unlocked_ioctl = pipe_ioctl,
1099         .release        = pipe_release,
1100         .fasync         = pipe_fasync,
1101 };
1102
1103 /*
1104  * Currently we rely on the pipe array holding a power-of-2 number
1105  * of pages. Returns 0 on error.
1106  */
1107 unsigned int round_pipe_size(unsigned long size)
1108 {
1109         if (size > (1U << 31))
1110                 return 0;
1111
1112         /* Minimum pipe size, as required by POSIX */
1113         if (size < PAGE_SIZE)
1114                 return PAGE_SIZE;
1115
1116         return roundup_pow_of_two(size);
1117 }
1118
1119 /*
1120  * Allocate a new array of pipe buffers and copy the info over. Returns the
1121  * pipe size if successful, or return -ERROR on error.
1122  */
1123 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1124 {
1125         struct pipe_buffer *bufs;
1126         unsigned int size, nr_slots, head, tail, mask, n;
1127         unsigned long user_bufs;
1128         long ret = 0;
1129
1130         size = round_pipe_size(arg);
1131         nr_slots = size >> PAGE_SHIFT;
1132
1133         if (!nr_slots)
1134                 return -EINVAL;
1135
1136         /*
1137          * If trying to increase the pipe capacity, check that an
1138          * unprivileged user is not trying to exceed various limits
1139          * (soft limit check here, hard limit check just below).
1140          * Decreasing the pipe capacity is always permitted, even
1141          * if the user is currently over a limit.
1142          */
1143         if (nr_slots > pipe->ring_size &&
1144                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1145                 return -EPERM;
1146
1147         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1148
1149         if (nr_slots > pipe->ring_size &&
1150                         (too_many_pipe_buffers_hard(user_bufs) ||
1151                          too_many_pipe_buffers_soft(user_bufs)) &&
1152                         is_unprivileged_user()) {
1153                 ret = -EPERM;
1154                 goto out_revert_acct;
1155         }
1156
1157         /*
1158          * We can shrink the pipe, if arg is greater than the ring occupancy.
1159          * Since we don't expect a lot of shrink+grow operations, just free and
1160          * allocate again like we would do for growing.  If the pipe currently
1161          * contains more buffers than arg, then return busy.
1162          */
1163         mask = pipe->ring_size - 1;
1164         head = pipe->head;
1165         tail = pipe->tail;
1166         n = pipe_occupancy(pipe->head, pipe->tail);
1167         if (nr_slots < n) {
1168                 ret = -EBUSY;
1169                 goto out_revert_acct;
1170         }
1171
1172         bufs = kcalloc(nr_slots, sizeof(*bufs),
1173                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1174         if (unlikely(!bufs)) {
1175                 ret = -ENOMEM;
1176                 goto out_revert_acct;
1177         }
1178
1179         /*
1180          * The pipe array wraps around, so just start the new one at zero
1181          * and adjust the indices.
1182          */
1183         if (n > 0) {
1184                 unsigned int h = head & mask;
1185                 unsigned int t = tail & mask;
1186                 if (h > t) {
1187                         memcpy(bufs, pipe->bufs + t,
1188                                n * sizeof(struct pipe_buffer));
1189                 } else {
1190                         unsigned int tsize = pipe->ring_size - t;
1191                         if (h > 0)
1192                                 memcpy(bufs + tsize, pipe->bufs,
1193                                        h * sizeof(struct pipe_buffer));
1194                         memcpy(bufs, pipe->bufs + t,
1195                                tsize * sizeof(struct pipe_buffer));
1196                 }
1197         }
1198
1199         head = n;
1200         tail = 0;
1201
1202         kfree(pipe->bufs);
1203         pipe->bufs = bufs;
1204         pipe->ring_size = nr_slots;
1205         pipe->max_usage = nr_slots;
1206         pipe->tail = tail;
1207         pipe->head = head;
1208         wake_up_interruptible_all(&pipe->wait);
1209         return pipe->max_usage * PAGE_SIZE;
1210
1211 out_revert_acct:
1212         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1213         return ret;
1214 }
1215
1216 /*
1217  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1218  * location, so checking ->i_pipe is not enough to verify that this is a
1219  * pipe.
1220  */
1221 struct pipe_inode_info *get_pipe_info(struct file *file)
1222 {
1223         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1224 }
1225
1226 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1227 {
1228         struct pipe_inode_info *pipe;
1229         long ret;
1230
1231         pipe = get_pipe_info(file);
1232         if (!pipe)
1233                 return -EBADF;
1234
1235         __pipe_lock(pipe);
1236
1237         switch (cmd) {
1238         case F_SETPIPE_SZ:
1239                 ret = pipe_set_size(pipe, arg);
1240                 break;
1241         case F_GETPIPE_SZ:
1242                 ret = pipe->max_usage * PAGE_SIZE;
1243                 break;
1244         default:
1245                 ret = -EINVAL;
1246                 break;
1247         }
1248
1249         __pipe_unlock(pipe);
1250         return ret;
1251 }
1252
1253 static const struct super_operations pipefs_ops = {
1254         .destroy_inode = free_inode_nonrcu,
1255         .statfs = simple_statfs,
1256 };
1257
1258 /*
1259  * pipefs should _never_ be mounted by userland - too much of security hassle,
1260  * no real gain from having the whole whorehouse mounted. So we don't need
1261  * any operations on the root directory. However, we need a non-trivial
1262  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1263  */
1264
1265 static int pipefs_init_fs_context(struct fs_context *fc)
1266 {
1267         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1268         if (!ctx)
1269                 return -ENOMEM;
1270         ctx->ops = &pipefs_ops;
1271         ctx->dops = &pipefs_dentry_operations;
1272         return 0;
1273 }
1274
1275 static struct file_system_type pipe_fs_type = {
1276         .name           = "pipefs",
1277         .init_fs_context = pipefs_init_fs_context,
1278         .kill_sb        = kill_anon_super,
1279 };
1280
1281 static int __init init_pipe_fs(void)
1282 {
1283         int err = register_filesystem(&pipe_fs_type);
1284
1285         if (!err) {
1286                 pipe_mnt = kern_mount(&pipe_fs_type);
1287                 if (IS_ERR(pipe_mnt)) {
1288                         err = PTR_ERR(pipe_mnt);
1289                         unregister_filesystem(&pipe_fs_type);
1290                 }
1291         }
1292         return err;
1293 }
1294
1295 fs_initcall(init_pipe_fs);