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         int do_wakeup;
 280         ssize_t ret;
 281
 282         /* Null read succeeds. */
 283         if (unlikely(total_len == 0))
 284                 return 0;
 285
 286         do_wakeup = 0;
 287         ret = 0;
 288         __pipe_lock(pipe);
 289         for (;;) {
 290                 unsigned int head = pipe->head;
 291                 unsigned int tail = pipe->tail;
 292                 unsigned int mask = pipe->ring_size - 1;
 293
 294                 if (!pipe_empty(head, tail)) {
 295                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 296                         size_t chars = buf->len;
 297                         size_t written;
 298                         int error;
 299
 300                         if (chars > total_len)
 301                                 chars = total_len;
 302
 303                         error = pipe_buf_confirm(pipe, buf);
 304                         if (error) {
 305                                 if (!ret)
 306                                         ret = error;
 307                                 break;
 308                         }
 309
 310                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 311                         if (unlikely(written < chars)) {
 312                                 if (!ret)
 313                                         ret = -EFAULT;
 314                                 break;
 315                         }
 316                         ret += chars;
 317                         buf->offset += chars;
 318                         buf->len -= chars;
 319
 320                         /* Was it a packet buffer? Clean up and exit */
 321                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 322                                 total_len = chars;
 323                                 buf->len = 0;
 324                         }
 325
 326                         if (!buf->len) {
 327                                 pipe_buf_release(pipe, buf);
 328                                 tail++;
 329                                 pipe->tail = tail;
 330                                 do_wakeup = 1;
 331                         }
 332                         total_len -= chars;
 333                         if (!total_len)
 334                                 break;  /* common path: read succeeded */
 335                         if (!pipe_empty(head, tail))    /* More to do? */
 336                                 continue;
 337                 }
 338
 339                 if (!pipe->writers)
 340                         break;
 341                 if (!pipe->waiting_writers) {
 342                         /* syscall merging: Usually we must not sleep
 343                          * if O_NONBLOCK is set, or if we got some data.
 344                          * But if a writer sleeps in kernel space, then
 345                          * we can wait for that data without violating POSIX.
 346                          */
 347                         if (ret)
 348                                 break;
 349                         if (filp->f_flags & O_NONBLOCK) {
 350                                 ret = -EAGAIN;
 351                                 break;
 352                         }
 353                 }
 354                 if (signal_pending(current)) {
 355                         if (!ret)
 356                                 ret = -ERESTARTSYS;
 357                         break;
 358                 }
 359                 if (do_wakeup) {
 360                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 361                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 362                 }
 363                 pipe_wait(pipe);
 364         }
 365         __pipe_unlock(pipe);
 366
 367         /* Signal writers asynchronously that there is more room. */
 368         if (do_wakeup) {
 369                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 370                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 371         }
 372         if (ret > 0)
 373                 file_accessed(filp);
 374         return ret;
 375 }
 376
 377 static inline int is_packetized(struct file *file)
 378 {
 379         return (file->f_flags & O_DIRECT) != 0;
 380 }
 381
 382 static ssize_t
 383 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 384 {
 385         struct file *filp = iocb->ki_filp;
 386         struct pipe_inode_info *pipe = filp->private_data;
 387         unsigned int head, tail, max_usage, mask;
 388         ssize_t ret = 0;
 389         int do_wakeup = 0;
 390         size_t total_len = iov_iter_count(from);
 391         ssize_t chars;
 392
 393         /* Null write succeeds. */
 394         if (unlikely(total_len == 0))
 395                 return 0;
 396
 397         __pipe_lock(pipe);
 398
 399         if (!pipe->readers) {
 400                 send_sig(SIGPIPE, current, 0);
 401                 ret = -EPIPE;
 402                 goto out;
 403         }
 404
 405         tail = pipe->tail;
 406         head = pipe->head;
 407         max_usage = pipe->ring_size;
 408         mask = pipe->ring_size - 1;
 409
 410         /* We try to merge small writes */
 411         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 412         if (!pipe_empty(head, tail) && chars != 0) {
 413                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 414                 int offset = buf->offset + buf->len;
 415
 416                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 417                         ret = pipe_buf_confirm(pipe, buf);
 418                         if (ret)
 419                                 goto out;
 420
 421                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 422                         if (unlikely(ret < chars)) {
 423                                 ret = -EFAULT;
 424                                 goto out;
 425                         }
 426                         do_wakeup = 1;
 427                         buf->len += ret;
 428                         if (!iov_iter_count(from))
 429                                 goto out;
 430                 }
 431         }
 432
 433         for (;;) {
 434                 if (!pipe->readers) {
 435                         send_sig(SIGPIPE, current, 0);
 436                         if (!ret)
 437                                 ret = -EPIPE;
 438                         break;
 439                 }
 440
 441                 tail = pipe->tail;
 442                 if (!pipe_full(head, tail, max_usage)) {
 443                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 444                         struct page *page = pipe->tmp_page;
 445                         int copied;
 446
 447                         if (!page) {
 448                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 449                                 if (unlikely(!page)) {
 450                                         ret = ret ? : -ENOMEM;
 451                                         break;
 452                                 }
 453                                 pipe->tmp_page = page;
 454                         }
 455                         /* Always wake up, even if the copy fails. Otherwise
 456                          * we lock up (O_NONBLOCK-)readers that sleep due to
 457                          * syscall merging.
 458                          * FIXME! Is this really true?
 459                          */
 460                         do_wakeup = 1;
 461                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 462                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 463                                 if (!ret)
 464                                         ret = -EFAULT;
 465                                 break;
 466                         }
 467                         ret += copied;
 468
 469                         /* Insert it into the buffer array */
 470                         buf->page = page;
 471                         buf->ops = &anon_pipe_buf_ops;
 472                         buf->offset = 0;
 473                         buf->len = copied;
 474                         buf->flags = 0;
 475                         if (is_packetized(filp)) {
 476                                 buf->ops = &packet_pipe_buf_ops;
 477                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 478                         }
 479
 480                         head++;
 481                         pipe->head = head;
 482                         pipe->tmp_page = NULL;
 483
 484                         if (!iov_iter_count(from))
 485                                 break;
 486                 }
 487
 488                 if (!pipe_full(head, tail, max_usage))
 489                         continue;
 490
 491                 /* Wait for buffer space to become available. */
 492                 if (filp->f_flags & O_NONBLOCK) {
 493                         if (!ret)
 494                                 ret = -EAGAIN;
 495                         break;
 496                 }
 497                 if (signal_pending(current)) {
 498                         if (!ret)
 499                                 ret = -ERESTARTSYS;
 500                         break;
 501                 }
 502                 if (do_wakeup) {
 503                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 504                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 505                         do_wakeup = 0;
 506                 }
 507                 pipe->waiting_writers++;
 508                 pipe_wait(pipe);
 509                 pipe->waiting_writers--;
 510         }
 511 out:
 512         __pipe_unlock(pipe);
 513         if (do_wakeup) {
 514                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 515                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 516         }
 517         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 518                 int err = file_update_time(filp);
 519                 if (err)
 520                         ret = err;
 521                 sb_end_write(file_inode(filp)->i_sb);
 522         }
 523         return ret;
 524 }
 525
 526 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 527 {
 528         struct pipe_inode_info *pipe = filp->private_data;
 529         int count, head, tail, mask;
 530
 531         switch (cmd) {
 532                 case FIONREAD:
 533                         __pipe_lock(pipe);
 534                         count = 0;
 535                         head = pipe->head;
 536                         tail = pipe->tail;
 537                         mask = pipe->ring_size - 1;
 538
 539                         while (tail != head) {
 540                                 count += pipe->bufs[tail & mask].len;
 541                                 tail++;
 542                         }
 543                         __pipe_unlock(pipe);
 544
 545                         return put_user(count, (int __user *)arg);
 546                 default:
 547                         return -ENOIOCTLCMD;
 548         }
 549 }
 550
 551 /* No kernel lock held - fine */
 552 static __poll_t
 553 pipe_poll(struct file *filp, poll_table *wait)
 554 {
 555         __poll_t mask;
 556         struct pipe_inode_info *pipe = filp->private_data;
 557         unsigned int head = READ_ONCE(pipe->head);
 558         unsigned int tail = READ_ONCE(pipe->tail);
 559
 560         poll_wait(filp, &pipe->wait, wait);
 561
 562         BUG_ON(pipe_occupancy(head, tail) > pipe->ring_size);
 563
 564         /* Reading only -- no need for acquiring the semaphore.  */
 565         mask = 0;
 566         if (filp->f_mode & FMODE_READ) {
 567                 if (!pipe_empty(head, tail))
 568                         mask |= EPOLLIN | EPOLLRDNORM;
 569                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 570                         mask |= EPOLLHUP;
 571         }
 572
 573         if (filp->f_mode & FMODE_WRITE) {
 574                 if (!pipe_full(head, tail, pipe->ring_size))
 575                         mask |= EPOLLOUT | EPOLLWRNORM;
 576                 /*
 577                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 578                  * behave exactly like pipes for poll().
 579                  */
 580                 if (!pipe->readers)
 581                         mask |= EPOLLERR;
 582         }
 583
 584         return mask;
 585 }
 586
 587 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 588 {
 589         int kill = 0;
 590
 591         spin_lock(&inode->i_lock);
 592         if (!--pipe->files) {
 593                 inode->i_pipe = NULL;
 594                 kill = 1;
 595         }
 596         spin_unlock(&inode->i_lock);
 597
 598         if (kill)
 599                 free_pipe_info(pipe);
 600 }
 601
 602 static int
 603 pipe_release(struct inode *inode, struct file *file)
 604 {
 605         struct pipe_inode_info *pipe = file->private_data;
 606
 607         __pipe_lock(pipe);
 608         if (file->f_mode & FMODE_READ)
 609                 pipe->readers--;
 610         if (file->f_mode & FMODE_WRITE)
 611                 pipe->writers--;
 612
 613         if (pipe->readers || pipe->writers) {
 614                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 615                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 616                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 617         }
 618         __pipe_unlock(pipe);
 619
 620         put_pipe_info(inode, pipe);
 621         return 0;
 622 }
 623
 624 static int
 625 pipe_fasync(int fd, struct file *filp, int on)
 626 {
 627         struct pipe_inode_info *pipe = filp->private_data;
 628         int retval = 0;
 629
 630         __pipe_lock(pipe);
 631         if (filp->f_mode & FMODE_READ)
 632                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 633         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 634                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 635                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 636                         /* this can happen only if on == T */
 637                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 638         }
 639         __pipe_unlock(pipe);
 640         return retval;
 641 }
 642
 643 static unsigned long account_pipe_buffers(struct user_struct *user,
 644                                  unsigned long old, unsigned long new)
 645 {
 646         return atomic_long_add_return(new - old, &user->pipe_bufs);
 647 }
 648
 649 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 650 {
 651         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 652
 653         return soft_limit && user_bufs > soft_limit;
 654 }
 655
 656 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 657 {
 658         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 659
 660         return hard_limit && user_bufs > hard_limit;
 661 }
 662
 663 static bool is_unprivileged_user(void)
 664 {
 665         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 666 }
 667
 668 struct pipe_inode_info *alloc_pipe_info(void)
 669 {
 670         struct pipe_inode_info *pipe;
 671         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 672         struct user_struct *user = get_current_user();
 673         unsigned long user_bufs;
 674         unsigned int max_size = READ_ONCE(pipe_max_size);
 675
 676         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 677         if (pipe == NULL)
 678                 goto out_free_uid;
 679
 680         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 681                 pipe_bufs = max_size >> PAGE_SHIFT;
 682
 683         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 684
 685         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 686                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 687                 pipe_bufs = 1;
 688         }
 689
 690         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 691                 goto out_revert_acct;
 692
 693         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 694                              GFP_KERNEL_ACCOUNT);
 695
 696         if (pipe->bufs) {
 697                 init_waitqueue_head(&pipe->wait);
 698                 pipe->r_counter = pipe->w_counter = 1;
 699                 pipe->ring_size = pipe_bufs;
 700                 pipe->user = user;
 701                 mutex_init(&pipe->mutex);
 702                 return pipe;
 703         }
 704
 705 out_revert_acct:
 706         (void) account_pipe_buffers(user, pipe_bufs, 0);
 707         kfree(pipe);
 708 out_free_uid:
 709         free_uid(user);
 710         return NULL;
 711 }
 712
 713 void free_pipe_info(struct pipe_inode_info *pipe)
 714 {
 715         int i;
 716
 717         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 718         free_uid(pipe->user);
 719         for (i = 0; i < pipe->ring_size; i++) {
 720                 struct pipe_buffer *buf = pipe->bufs + i;
 721                 if (buf->ops)
 722                         pipe_buf_release(pipe, buf);
 723         }
 724         if (pipe->tmp_page)
 725                 __free_page(pipe->tmp_page);
 726         kfree(pipe->bufs);
 727         kfree(pipe);
 728 }
 729
 730 static struct vfsmount *pipe_mnt __read_mostly;
 731
 732 /*
 733  * pipefs_dname() is called from d_path().
 734  */
 735 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 736 {
 737         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 738                                 d_inode(dentry)->i_ino);
 739 }
 740
 741 static const struct dentry_operations pipefs_dentry_operations = {
 742         .d_dname        = pipefs_dname,
 743 };
 744
 745 static struct inode * get_pipe_inode(void)
 746 {
 747         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 748         struct pipe_inode_info *pipe;
 749
 750         if (!inode)
 751                 goto fail_inode;
 752
 753         inode->i_ino = get_next_ino();
 754
 755         pipe = alloc_pipe_info();
 756         if (!pipe)
 757                 goto fail_iput;
 758
 759         inode->i_pipe = pipe;
 760         pipe->files = 2;
 761         pipe->readers = pipe->writers = 1;
 762         inode->i_fop = &pipefifo_fops;
 763
 764         /*
 765          * Mark the inode dirty from the very beginning,
 766          * that way it will never be moved to the dirty
 767          * list because "mark_inode_dirty()" will think
 768          * that it already _is_ on the dirty list.
 769          */
 770         inode->i_state = I_DIRTY;
 771         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 772         inode->i_uid = current_fsuid();
 773         inode->i_gid = current_fsgid();
 774         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 775
 776         return inode;
 777
 778 fail_iput:
 779         iput(inode);
 780
 781 fail_inode:
 782         return NULL;
 783 }
 784
 785 int create_pipe_files(struct file **res, int flags)
 786 {
 787         struct inode *inode = get_pipe_inode();
 788         struct file *f;
 789
 790         if (!inode)
 791                 return -ENFILE;
 792
 793         f = alloc_file_pseudo(inode, pipe_mnt, "",
 794                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 795                                 &pipefifo_fops);
 796         if (IS_ERR(f)) {
 797                 free_pipe_info(inode->i_pipe);
 798                 iput(inode);
 799                 return PTR_ERR(f);
 800         }
 801
 802         f->private_data = inode->i_pipe;
 803
 804         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 805                                   &pipefifo_fops);
 806         if (IS_ERR(res[0])) {
 807                 put_pipe_info(inode, inode->i_pipe);
 808                 fput(f);
 809                 return PTR_ERR(res[0]);
 810         }
 811         res[0]->private_data = inode->i_pipe;
 812         res[1] = f;
 813         return 0;
 814 }
 815
 816 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 817 {
 818         int error;
 819         int fdw, fdr;
 820
 821         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 822                 return -EINVAL;
 823
 824         error = create_pipe_files(files, flags);
 825         if (error)
 826                 return error;
 827
 828         error = get_unused_fd_flags(flags);
 829         if (error < 0)
 830                 goto err_read_pipe;
 831         fdr = error;
 832
 833         error = get_unused_fd_flags(flags);
 834         if (error < 0)
 835                 goto err_fdr;
 836         fdw = error;
 837
 838         audit_fd_pair(fdr, fdw);
 839         fd[0] = fdr;
 840         fd[1] = fdw;
 841         return 0;
 842
 843  err_fdr:
 844         put_unused_fd(fdr);
 845  err_read_pipe:
 846         fput(files[0]);
 847         fput(files[1]);
 848         return error;
 849 }
 850
 851 int do_pipe_flags(int *fd, int flags)
 852 {
 853         struct file *files[2];
 854         int error = __do_pipe_flags(fd, files, flags);
 855         if (!error) {
 856                 fd_install(fd[0], files[0]);
 857                 fd_install(fd[1], files[1]);
 858         }
 859         return error;
 860 }
 861
 862 /*
 863  * sys_pipe() is the normal C calling standard for creating
 864  * a pipe. It's not the way Unix traditionally does this, though.
 865  */
 866 static int do_pipe2(int __user *fildes, int flags)
 867 {
 868         struct file *files[2];
 869         int fd[2];
 870         int error;
 871
 872         error = __do_pipe_flags(fd, files, flags);
 873         if (!error) {
 874                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 875                         fput(files[0]);
 876                         fput(files[1]);
 877                         put_unused_fd(fd[0]);
 878                         put_unused_fd(fd[1]);
 879                         error = -EFAULT;
 880                 } else {
 881                         fd_install(fd[0], files[0]);
 882                         fd_install(fd[1], files[1]);
 883                 }
 884         }
 885         return error;
 886 }
 887
 888 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 889 {
 890         return do_pipe2(fildes, flags);
 891 }
 892
 893 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 894 {
 895         return do_pipe2(fildes, 0);
 896 }
 897
 898 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 899 {
 900         int cur = *cnt;
 901
 902         while (cur == *cnt) {
 903                 pipe_wait(pipe);
 904                 if (signal_pending(current))
 905                         break;
 906         }
 907         return cur == *cnt ? -ERESTARTSYS : 0;
 908 }
 909
 910 static void wake_up_partner(struct pipe_inode_info *pipe)
 911 {
 912         wake_up_interruptible(&pipe->wait);
 913 }
 914
 915 static int fifo_open(struct inode *inode, struct file *filp)
 916 {
 917         struct pipe_inode_info *pipe;
 918         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 919         int ret;
 920
 921         filp->f_version = 0;
 922
 923         spin_lock(&inode->i_lock);
 924         if (inode->i_pipe) {
 925                 pipe = inode->i_pipe;
 926                 pipe->files++;
 927                 spin_unlock(&inode->i_lock);
 928         } else {
 929                 spin_unlock(&inode->i_lock);
 930                 pipe = alloc_pipe_info();
 931                 if (!pipe)
 932                         return -ENOMEM;
 933                 pipe->files = 1;
 934                 spin_lock(&inode->i_lock);
 935                 if (unlikely(inode->i_pipe)) {
 936                         inode->i_pipe->files++;
 937                         spin_unlock(&inode->i_lock);
 938                         free_pipe_info(pipe);
 939                         pipe = inode->i_pipe;
 940                 } else {
 941                         inode->i_pipe = pipe;
 942                         spin_unlock(&inode->i_lock);
 943                 }
 944         }
 945         filp->private_data = pipe;
 946         /* OK, we have a pipe and it's pinned down */
 947
 948         __pipe_lock(pipe);
 949
 950         /* We can only do regular read/write on fifos */
 951         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 952
 953         switch (filp->f_mode) {
 954         case FMODE_READ:
 955         /*
 956          *  O_RDONLY
 957          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 958          *  opened, even when there is no process writing the FIFO.
 959          */
 960                 pipe->r_counter++;
 961                 if (pipe->readers++ == 0)
 962                         wake_up_partner(pipe);
 963
 964                 if (!is_pipe && !pipe->writers) {
 965                         if ((filp->f_flags & O_NONBLOCK)) {
 966                                 /* suppress EPOLLHUP until we have
 967                                  * seen a writer */
 968                                 filp->f_version = pipe->w_counter;
 969                         } else {
 970                                 if (wait_for_partner(pipe, &pipe->w_counter))
 971                                         goto err_rd;
 972                         }
 973                 }
 974                 break;
 975
 976         case FMODE_WRITE:
 977         /*
 978          *  O_WRONLY
 979          *  POSIX.1 says that O_NONBLOCK means return -1 with
 980          *  errno=ENXIO when there is no process reading the FIFO.
 981          */
 982                 ret = -ENXIO;
 983                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 984                         goto err;
 985
 986                 pipe->w_counter++;
 987                 if (!pipe->writers++)
 988                         wake_up_partner(pipe);
 989
 990                 if (!is_pipe && !pipe->readers) {
 991                         if (wait_for_partner(pipe, &pipe->r_counter))
 992                                 goto err_wr;
 993                 }
 994                 break;
 995
 996         case FMODE_READ | FMODE_WRITE:
 997         /*
 998          *  O_RDWR
 999          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1000          *  This implementation will NEVER block on a O_RDWR open, since
1001          *  the process can at least talk to itself.
1002          */
1003
1004                 pipe->readers++;
1005                 pipe->writers++;
1006                 pipe->r_counter++;
1007                 pipe->w_counter++;
1008                 if (pipe->readers == 1 || pipe->writers == 1)
1009                         wake_up_partner(pipe);
1010                 break;
1011
1012         default:
1013                 ret = -EINVAL;
1014                 goto err;
1015         }
1016
1017         /* Ok! */
1018         __pipe_unlock(pipe);
1019         return 0;
1020
1021 err_rd:
1022         if (!--pipe->readers)
1023                 wake_up_interruptible(&pipe->wait);
1024         ret = -ERESTARTSYS;
1025         goto err;
1026
1027 err_wr:
1028         if (!--pipe->writers)
1029                 wake_up_interruptible(&pipe->wait);
1030         ret = -ERESTARTSYS;
1031         goto err;
1032
1033 err:
1034         __pipe_unlock(pipe);
1035
1036         put_pipe_info(inode, pipe);
1037         return ret;
1038 }
1039
1040 const struct file_operations pipefifo_fops = {
1041         .open           = fifo_open,
1042         .llseek         = no_llseek,
1043         .read_iter      = pipe_read,
1044         .write_iter     = pipe_write,
1045         .poll           = pipe_poll,
1046         .unlocked_ioctl = pipe_ioctl,
1047         .release        = pipe_release,
1048         .fasync         = pipe_fasync,
1049 };
1050
1051 /*
1052  * Currently we rely on the pipe array holding a power-of-2 number
1053  * of pages. Returns 0 on error.
1054  */
1055 unsigned int round_pipe_size(unsigned long size)
1056 {
1057         if (size > (1U << 31))
1058                 return 0;
1059
1060         /* Minimum pipe size, as required by POSIX */
1061         if (size < PAGE_SIZE)
1062                 return PAGE_SIZE;
1063
1064         return roundup_pow_of_two(size);
1065 }
1066
1067 /*
1068  * Allocate a new array of pipe buffers and copy the info over. Returns the
1069  * pipe size if successful, or return -ERROR on error.
1070  */
1071 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1072 {
1073         struct pipe_buffer *bufs;
1074         unsigned int size, nr_slots, head, tail, mask, n;
1075         unsigned long user_bufs;
1076         long ret = 0;
1077
1078         size = round_pipe_size(arg);
1079         nr_slots = size >> PAGE_SHIFT;
1080
1081         if (!nr_slots)
1082                 return -EINVAL;
1083
1084         /*
1085          * If trying to increase the pipe capacity, check that an
1086          * unprivileged user is not trying to exceed various limits
1087          * (soft limit check here, hard limit check just below).
1088          * Decreasing the pipe capacity is always permitted, even
1089          * if the user is currently over a limit.
1090          */
1091         if (nr_slots > pipe->ring_size &&
1092                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1093                 return -EPERM;
1094
1095         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1096
1097         if (nr_slots > pipe->ring_size &&
1098                         (too_many_pipe_buffers_hard(user_bufs) ||
1099                          too_many_pipe_buffers_soft(user_bufs)) &&
1100                         is_unprivileged_user()) {
1101                 ret = -EPERM;
1102                 goto out_revert_acct;
1103         }
1104
1105         /*
1106          * We can shrink the pipe, if arg is greater than the ring occupancy.
1107          * Since we don't expect a lot of shrink+grow operations, just free and
1108          * allocate again like we would do for growing.  If the pipe currently
1109          * contains more buffers than arg, then return busy.
1110          */
1111         mask = pipe->ring_size - 1;
1112         head = pipe->head;
1113         tail = pipe->tail;
1114         n = pipe_occupancy(pipe->head, pipe->tail);
1115         if (nr_slots < n) {
1116                 ret = -EBUSY;
1117                 goto out_revert_acct;
1118         }
1119
1120         bufs = kcalloc(nr_slots, sizeof(*bufs),
1121                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1122         if (unlikely(!bufs)) {
1123                 ret = -ENOMEM;
1124                 goto out_revert_acct;
1125         }
1126
1127         /*
1128          * The pipe array wraps around, so just start the new one at zero
1129          * and adjust the indices.
1130          */
1131         if (n > 0) {
1132                 unsigned int h = head & mask;
1133                 unsigned int t = tail & mask;
1134                 if (h > t) {
1135                         memcpy(bufs, pipe->bufs + t,
1136                                n * sizeof(struct pipe_buffer));
1137                 } else {
1138                         unsigned int tsize = pipe->ring_size - t;
1139                         if (h > 0)
1140                                 memcpy(bufs + tsize, pipe->bufs,
1141                                        h * sizeof(struct pipe_buffer));
1142                         memcpy(bufs, pipe->bufs + t,
1143                                tsize * sizeof(struct pipe_buffer));
1144                 }
1145         }
1146
1147         head = n;
1148         tail = 0;
1149
1150         kfree(pipe->bufs);
1151         pipe->bufs = bufs;
1152         pipe->ring_size = nr_slots;
1153         pipe->tail = tail;
1154         pipe->head = head;
1155         return pipe->ring_size * PAGE_SIZE;
1156
1157 out_revert_acct:
1158         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1159         return ret;
1160 }
1161
1162 /*
1163  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1164  * location, so checking ->i_pipe is not enough to verify that this is a
1165  * pipe.
1166  */
1167 struct pipe_inode_info *get_pipe_info(struct file *file)
1168 {
1169         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1170 }
1171
1172 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1173 {
1174         struct pipe_inode_info *pipe;
1175         long ret;
1176
1177         pipe = get_pipe_info(file);
1178         if (!pipe)
1179                 return -EBADF;
1180
1181         __pipe_lock(pipe);
1182
1183         switch (cmd) {
1184         case F_SETPIPE_SZ:
1185                 ret = pipe_set_size(pipe, arg);
1186                 break;
1187         case F_GETPIPE_SZ:
1188                 ret = pipe->ring_size * PAGE_SIZE;
1189                 break;
1190         default:
1191                 ret = -EINVAL;
1192                 break;
1193         }
1194
1195         __pipe_unlock(pipe);
1196         return ret;
1197 }
1198
1199 static const struct super_operations pipefs_ops = {
1200         .destroy_inode = free_inode_nonrcu,
1201         .statfs = simple_statfs,
1202 };
1203
1204 /*
1205  * pipefs should _never_ be mounted by userland - too much of security hassle,
1206  * no real gain from having the whole whorehouse mounted. So we don't need
1207  * any operations on the root directory. However, we need a non-trivial
1208  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1209  */
1210
1211 static int pipefs_init_fs_context(struct fs_context *fc)
1212 {
1213         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1214         if (!ctx)
1215                 return -ENOMEM;
1216         ctx->ops = &pipefs_ops;
1217         ctx->dops = &pipefs_dentry_operations;
1218         return 0;
1219 }
1220
1221 static struct file_system_type pipe_fs_type = {
1222         .name           = "pipefs",
1223         .init_fs_context = pipefs_init_fs_context,
1224         .kill_sb        = kill_anon_super,
1225 };
1226
1227 static int __init init_pipe_fs(void)
1228 {
1229         int err = register_filesystem(&pipe_fs_type);
1230
1231         if (!err) {
1232                 pipe_mnt = kern_mount(&pipe_fs_type);
1233                 if (IS_ERR(pipe_mnt)) {
1234                         err = PTR_ERR(pipe_mnt);
1235                         unregister_filesystem(&pipe_fs_type);
1236                 }
1237         }
1238         return err;
1239 }
1240
1241 fs_initcall(init_pipe_fs);