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                                 spin_lock_irq(&pipe->wait.lock);
 329                                 tail++;
 330                                 pipe->tail = tail;
 331                                 do_wakeup = 0;
 332                                 wake_up_interruptible_sync_poll_locked(
 333                                         &pipe->wait, EPOLLOUT | EPOLLWRNORM);
 334                                 spin_unlock_irq(&pipe->wait.lock);
 335                                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 336                         }
 337                         total_len -= chars;
 338                         if (!total_len)
 339                                 break;  /* common path: read succeeded */
 340                         if (!pipe_empty(head, tail))    /* More to do? */
 341                                 continue;
 342                 }
 343
 344                 if (!pipe->writers)
 345                         break;
 346                 if (!pipe->waiting_writers) {
 347                         /* syscall merging: Usually we must not sleep
 348                          * if O_NONBLOCK is set, or if we got some data.
 349                          * But if a writer sleeps in kernel space, then
 350                          * we can wait for that data without violating POSIX.
 351                          */
 352                         if (ret)
 353                                 break;
 354                         if (filp->f_flags & O_NONBLOCK) {
 355                                 ret = -EAGAIN;
 356                                 break;
 357                         }
 358                 }
 359                 if (signal_pending(current)) {
 360                         if (!ret)
 361                                 ret = -ERESTARTSYS;
 362                         break;
 363                 }
 364                 if (do_wakeup) {
 365                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 366                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 367                         do_wakeup = 0;
 368                 }
 369                 pipe_wait(pipe);
 370         }
 371         __pipe_unlock(pipe);
 372
 373         /* Signal writers asynchronously that there is more room. */
 374         if (do_wakeup) {
 375                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 376                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 377         }
 378         if (ret > 0)
 379                 file_accessed(filp);
 380         return ret;
 381 }
 382
 383 static inline int is_packetized(struct file *file)
 384 {
 385         return (file->f_flags & O_DIRECT) != 0;
 386 }
 387
 388 static ssize_t
 389 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 390 {
 391         struct file *filp = iocb->ki_filp;
 392         struct pipe_inode_info *pipe = filp->private_data;
 393         unsigned int head, tail, max_usage, mask;
 394         ssize_t ret = 0;
 395         int do_wakeup = 0;
 396         size_t total_len = iov_iter_count(from);
 397         ssize_t chars;
 398
 399         /* Null write succeeds. */
 400         if (unlikely(total_len == 0))
 401                 return 0;
 402
 403         __pipe_lock(pipe);
 404
 405         if (!pipe->readers) {
 406                 send_sig(SIGPIPE, current, 0);
 407                 ret = -EPIPE;
 408                 goto out;
 409         }
 410
 411         tail = pipe->tail;
 412         head = pipe->head;
 413         max_usage = pipe->max_usage;
 414         mask = pipe->ring_size - 1;
 415
 416         /* We try to merge small writes */
 417         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 418         if (!pipe_empty(head, tail) && chars != 0) {
 419                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 420                 int offset = buf->offset + buf->len;
 421
 422                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 423                         ret = pipe_buf_confirm(pipe, buf);
 424                         if (ret)
 425                                 goto out;
 426
 427                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 428                         if (unlikely(ret < chars)) {
 429                                 ret = -EFAULT;
 430                                 goto out;
 431                         }
 432                         do_wakeup = 1;
 433                         buf->len += ret;
 434                         if (!iov_iter_count(from))
 435                                 goto out;
 436                 }
 437         }
 438
 439         for (;;) {
 440                 if (!pipe->readers) {
 441                         send_sig(SIGPIPE, current, 0);
 442                         if (!ret)
 443                                 ret = -EPIPE;
 444                         break;
 445                 }
 446
 447                 tail = pipe->tail;
 448                 if (!pipe_full(head, tail, max_usage)) {
 449                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 450                         struct page *page = pipe->tmp_page;
 451                         int copied;
 452
 453                         if (!page) {
 454                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 455                                 if (unlikely(!page)) {
 456                                         ret = ret ? : -ENOMEM;
 457                                         break;
 458                                 }
 459                                 pipe->tmp_page = page;
 460                         }
 461                         /* Always wake up, even if the copy fails. Otherwise
 462                          * we lock up (O_NONBLOCK-)readers that sleep due to
 463                          * syscall merging.
 464                          * FIXME! Is this really true?
 465                          */
 466                         do_wakeup = 1;
 467                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 468                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 469                                 if (!ret)
 470                                         ret = -EFAULT;
 471                                 break;
 472                         }
 473                         ret += copied;
 474
 475                         /* Insert it into the buffer array */
 476                         buf->page = page;
 477                         buf->ops = &anon_pipe_buf_ops;
 478                         buf->offset = 0;
 479                         buf->len = copied;
 480                         buf->flags = 0;
 481                         if (is_packetized(filp)) {
 482                                 buf->ops = &packet_pipe_buf_ops;
 483                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 484                         }
 485
 486                         head++;
 487                         pipe->head = head;
 488                         pipe->tmp_page = NULL;
 489
 490                         if (!iov_iter_count(from))
 491                                 break;
 492                 }
 493
 494                 if (!pipe_full(head, tail, max_usage))
 495                         continue;
 496
 497                 /* Wait for buffer space to become available. */
 498                 if (filp->f_flags & O_NONBLOCK) {
 499                         if (!ret)
 500                                 ret = -EAGAIN;
 501                         break;
 502                 }
 503                 if (signal_pending(current)) {
 504                         if (!ret)
 505                                 ret = -ERESTARTSYS;
 506                         break;
 507                 }
 508                 if (do_wakeup) {
 509                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 510                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 511                         do_wakeup = 0;
 512                 }
 513                 pipe->waiting_writers++;
 514                 pipe_wait(pipe);
 515                 pipe->waiting_writers--;
 516         }
 517 out:
 518         __pipe_unlock(pipe);
 519         if (do_wakeup) {
 520                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 521                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 522         }
 523         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 524                 int err = file_update_time(filp);
 525                 if (err)
 526                         ret = err;
 527                 sb_end_write(file_inode(filp)->i_sb);
 528         }
 529         return ret;
 530 }
 531
 532 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 533 {
 534         struct pipe_inode_info *pipe = filp->private_data;
 535         int count, head, tail, mask;
 536
 537         switch (cmd) {
 538                 case FIONREAD:
 539                         __pipe_lock(pipe);
 540                         count = 0;
 541                         head = pipe->head;
 542                         tail = pipe->tail;
 543                         mask = pipe->ring_size - 1;
 544
 545                         while (tail != head) {
 546                                 count += pipe->bufs[tail & mask].len;
 547                                 tail++;
 548                         }
 549                         __pipe_unlock(pipe);
 550
 551                         return put_user(count, (int __user *)arg);
 552                 default:
 553                         return -ENOIOCTLCMD;
 554         }
 555 }
 556
 557 /* No kernel lock held - fine */
 558 static __poll_t
 559 pipe_poll(struct file *filp, poll_table *wait)
 560 {
 561         __poll_t mask;
 562         struct pipe_inode_info *pipe = filp->private_data;
 563         unsigned int head = READ_ONCE(pipe->head);
 564         unsigned int tail = READ_ONCE(pipe->tail);
 565
 566         poll_wait(filp, &pipe->wait, wait);
 567
 568         BUG_ON(pipe_occupancy(head, tail) > pipe->ring_size);
 569
 570         /* Reading only -- no need for acquiring the semaphore.  */
 571         mask = 0;
 572         if (filp->f_mode & FMODE_READ) {
 573                 if (!pipe_empty(head, tail))
 574                         mask |= EPOLLIN | EPOLLRDNORM;
 575                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 576                         mask |= EPOLLHUP;
 577         }
 578
 579         if (filp->f_mode & FMODE_WRITE) {
 580                 if (!pipe_full(head, tail, pipe->max_usage))
 581                         mask |= EPOLLOUT | EPOLLWRNORM;
 582                 /*
 583                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 584                  * behave exactly like pipes for poll().
 585                  */
 586                 if (!pipe->readers)
 587                         mask |= EPOLLERR;
 588         }
 589
 590         return mask;
 591 }
 592
 593 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 594 {
 595         int kill = 0;
 596
 597         spin_lock(&inode->i_lock);
 598         if (!--pipe->files) {
 599                 inode->i_pipe = NULL;
 600                 kill = 1;
 601         }
 602         spin_unlock(&inode->i_lock);
 603
 604         if (kill)
 605                 free_pipe_info(pipe);
 606 }
 607
 608 static int
 609 pipe_release(struct inode *inode, struct file *file)
 610 {
 611         struct pipe_inode_info *pipe = file->private_data;
 612
 613         __pipe_lock(pipe);
 614         if (file->f_mode & FMODE_READ)
 615                 pipe->readers--;
 616         if (file->f_mode & FMODE_WRITE)
 617                 pipe->writers--;
 618
 619         if (pipe->readers || pipe->writers) {
 620                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 621                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 622                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 623         }
 624         __pipe_unlock(pipe);
 625
 626         put_pipe_info(inode, pipe);
 627         return 0;
 628 }
 629
 630 static int
 631 pipe_fasync(int fd, struct file *filp, int on)
 632 {
 633         struct pipe_inode_info *pipe = filp->private_data;
 634         int retval = 0;
 635
 636         __pipe_lock(pipe);
 637         if (filp->f_mode & FMODE_READ)
 638                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 639         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 640                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 641                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 642                         /* this can happen only if on == T */
 643                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 644         }
 645         __pipe_unlock(pipe);
 646         return retval;
 647 }
 648
 649 static unsigned long account_pipe_buffers(struct user_struct *user,
 650                                  unsigned long old, unsigned long new)
 651 {
 652         return atomic_long_add_return(new - old, &user->pipe_bufs);
 653 }
 654
 655 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 656 {
 657         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 658
 659         return soft_limit && user_bufs > soft_limit;
 660 }
 661
 662 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 663 {
 664         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 665
 666         return hard_limit && user_bufs > hard_limit;
 667 }
 668
 669 static bool is_unprivileged_user(void)
 670 {
 671         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 672 }
 673
 674 struct pipe_inode_info *alloc_pipe_info(void)
 675 {
 676         struct pipe_inode_info *pipe;
 677         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 678         struct user_struct *user = get_current_user();
 679         unsigned long user_bufs;
 680         unsigned int max_size = READ_ONCE(pipe_max_size);
 681
 682         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 683         if (pipe == NULL)
 684                 goto out_free_uid;
 685
 686         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 687                 pipe_bufs = max_size >> PAGE_SHIFT;
 688
 689         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 690
 691         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 692                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 693                 pipe_bufs = 1;
 694         }
 695
 696         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 697                 goto out_revert_acct;
 698
 699         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 700                              GFP_KERNEL_ACCOUNT);
 701
 702         if (pipe->bufs) {
 703                 init_waitqueue_head(&pipe->wait);
 704                 pipe->r_counter = pipe->w_counter = 1;
 705                 pipe->max_usage = pipe_bufs;
 706                 pipe->ring_size = pipe_bufs;
 707                 pipe->user = user;
 708                 mutex_init(&pipe->mutex);
 709                 return pipe;
 710         }
 711
 712 out_revert_acct:
 713         (void) account_pipe_buffers(user, pipe_bufs, 0);
 714         kfree(pipe);
 715 out_free_uid:
 716         free_uid(user);
 717         return NULL;
 718 }
 719
 720 void free_pipe_info(struct pipe_inode_info *pipe)
 721 {
 722         int i;
 723
 724         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 725         free_uid(pipe->user);
 726         for (i = 0; i < pipe->ring_size; i++) {
 727                 struct pipe_buffer *buf = pipe->bufs + i;
 728                 if (buf->ops)
 729                         pipe_buf_release(pipe, buf);
 730         }
 731         if (pipe->tmp_page)
 732                 __free_page(pipe->tmp_page);
 733         kfree(pipe->bufs);
 734         kfree(pipe);
 735 }
 736
 737 static struct vfsmount *pipe_mnt __read_mostly;
 738
 739 /*
 740  * pipefs_dname() is called from d_path().
 741  */
 742 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 743 {
 744         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 745                                 d_inode(dentry)->i_ino);
 746 }
 747
 748 static const struct dentry_operations pipefs_dentry_operations = {
 749         .d_dname        = pipefs_dname,
 750 };
 751
 752 static struct inode * get_pipe_inode(void)
 753 {
 754         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 755         struct pipe_inode_info *pipe;
 756
 757         if (!inode)
 758                 goto fail_inode;
 759
 760         inode->i_ino = get_next_ino();
 761
 762         pipe = alloc_pipe_info();
 763         if (!pipe)
 764                 goto fail_iput;
 765
 766         inode->i_pipe = pipe;
 767         pipe->files = 2;
 768         pipe->readers = pipe->writers = 1;
 769         inode->i_fop = &pipefifo_fops;
 770
 771         /*
 772          * Mark the inode dirty from the very beginning,
 773          * that way it will never be moved to the dirty
 774          * list because "mark_inode_dirty()" will think
 775          * that it already _is_ on the dirty list.
 776          */
 777         inode->i_state = I_DIRTY;
 778         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 779         inode->i_uid = current_fsuid();
 780         inode->i_gid = current_fsgid();
 781         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 782
 783         return inode;
 784
 785 fail_iput:
 786         iput(inode);
 787
 788 fail_inode:
 789         return NULL;
 790 }
 791
 792 int create_pipe_files(struct file **res, int flags)
 793 {
 794         struct inode *inode = get_pipe_inode();
 795         struct file *f;
 796
 797         if (!inode)
 798                 return -ENFILE;
 799
 800         f = alloc_file_pseudo(inode, pipe_mnt, "",
 801                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 802                                 &pipefifo_fops);
 803         if (IS_ERR(f)) {
 804                 free_pipe_info(inode->i_pipe);
 805                 iput(inode);
 806                 return PTR_ERR(f);
 807         }
 808
 809         f->private_data = inode->i_pipe;
 810
 811         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 812                                   &pipefifo_fops);
 813         if (IS_ERR(res[0])) {
 814                 put_pipe_info(inode, inode->i_pipe);
 815                 fput(f);
 816                 return PTR_ERR(res[0]);
 817         }
 818         res[0]->private_data = inode->i_pipe;
 819         res[1] = f;
 820         return 0;
 821 }
 822
 823 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 824 {
 825         int error;
 826         int fdw, fdr;
 827
 828         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 829                 return -EINVAL;
 830
 831         error = create_pipe_files(files, flags);
 832         if (error)
 833                 return error;
 834
 835         error = get_unused_fd_flags(flags);
 836         if (error < 0)
 837                 goto err_read_pipe;
 838         fdr = error;
 839
 840         error = get_unused_fd_flags(flags);
 841         if (error < 0)
 842                 goto err_fdr;
 843         fdw = error;
 844
 845         audit_fd_pair(fdr, fdw);
 846         fd[0] = fdr;
 847         fd[1] = fdw;
 848         return 0;
 849
 850  err_fdr:
 851         put_unused_fd(fdr);
 852  err_read_pipe:
 853         fput(files[0]);
 854         fput(files[1]);
 855         return error;
 856 }
 857
 858 int do_pipe_flags(int *fd, int flags)
 859 {
 860         struct file *files[2];
 861         int error = __do_pipe_flags(fd, files, flags);
 862         if (!error) {
 863                 fd_install(fd[0], files[0]);
 864                 fd_install(fd[1], files[1]);
 865         }
 866         return error;
 867 }
 868
 869 /*
 870  * sys_pipe() is the normal C calling standard for creating
 871  * a pipe. It's not the way Unix traditionally does this, though.
 872  */
 873 static int do_pipe2(int __user *fildes, int flags)
 874 {
 875         struct file *files[2];
 876         int fd[2];
 877         int error;
 878
 879         error = __do_pipe_flags(fd, files, flags);
 880         if (!error) {
 881                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 882                         fput(files[0]);
 883                         fput(files[1]);
 884                         put_unused_fd(fd[0]);
 885                         put_unused_fd(fd[1]);
 886                         error = -EFAULT;
 887                 } else {
 888                         fd_install(fd[0], files[0]);
 889                         fd_install(fd[1], files[1]);
 890                 }
 891         }
 892         return error;
 893 }
 894
 895 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 896 {
 897         return do_pipe2(fildes, flags);
 898 }
 899
 900 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 901 {
 902         return do_pipe2(fildes, 0);
 903 }
 904
 905 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 906 {
 907         int cur = *cnt;
 908
 909         while (cur == *cnt) {
 910                 pipe_wait(pipe);
 911                 if (signal_pending(current))
 912                         break;
 913         }
 914         return cur == *cnt ? -ERESTARTSYS : 0;
 915 }
 916
 917 static void wake_up_partner(struct pipe_inode_info *pipe)
 918 {
 919         wake_up_interruptible(&pipe->wait);
 920 }
 921
 922 static int fifo_open(struct inode *inode, struct file *filp)
 923 {
 924         struct pipe_inode_info *pipe;
 925         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 926         int ret;
 927
 928         filp->f_version = 0;
 929
 930         spin_lock(&inode->i_lock);
 931         if (inode->i_pipe) {
 932                 pipe = inode->i_pipe;
 933                 pipe->files++;
 934                 spin_unlock(&inode->i_lock);
 935         } else {
 936                 spin_unlock(&inode->i_lock);
 937                 pipe = alloc_pipe_info();
 938                 if (!pipe)
 939                         return -ENOMEM;
 940                 pipe->files = 1;
 941                 spin_lock(&inode->i_lock);
 942                 if (unlikely(inode->i_pipe)) {
 943                         inode->i_pipe->files++;
 944                         spin_unlock(&inode->i_lock);
 945                         free_pipe_info(pipe);
 946                         pipe = inode->i_pipe;
 947                 } else {
 948                         inode->i_pipe = pipe;
 949                         spin_unlock(&inode->i_lock);
 950                 }
 951         }
 952         filp->private_data = pipe;
 953         /* OK, we have a pipe and it's pinned down */
 954
 955         __pipe_lock(pipe);
 956
 957         /* We can only do regular read/write on fifos */
 958         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 959
 960         switch (filp->f_mode) {
 961         case FMODE_READ:
 962         /*
 963          *  O_RDONLY
 964          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 965          *  opened, even when there is no process writing the FIFO.
 966          */
 967                 pipe->r_counter++;
 968                 if (pipe->readers++ == 0)
 969                         wake_up_partner(pipe);
 970
 971                 if (!is_pipe && !pipe->writers) {
 972                         if ((filp->f_flags & O_NONBLOCK)) {
 973                                 /* suppress EPOLLHUP until we have
 974                                  * seen a writer */
 975                                 filp->f_version = pipe->w_counter;
 976                         } else {
 977                                 if (wait_for_partner(pipe, &pipe->w_counter))
 978                                         goto err_rd;
 979                         }
 980                 }
 981                 break;
 982
 983         case FMODE_WRITE:
 984         /*
 985          *  O_WRONLY
 986          *  POSIX.1 says that O_NONBLOCK means return -1 with
 987          *  errno=ENXIO when there is no process reading the FIFO.
 988          */
 989                 ret = -ENXIO;
 990                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 991                         goto err;
 992
 993                 pipe->w_counter++;
 994                 if (!pipe->writers++)
 995                         wake_up_partner(pipe);
 996
 997                 if (!is_pipe && !pipe->readers) {
 998                         if (wait_for_partner(pipe, &pipe->r_counter))
 999                                 goto err_wr;
1000                 }
1001                 break;
1002
1003         case FMODE_READ | FMODE_WRITE:
1004         /*
1005          *  O_RDWR
1006          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1007          *  This implementation will NEVER block on a O_RDWR open, since
1008          *  the process can at least talk to itself.
1009          */
1010
1011                 pipe->readers++;
1012                 pipe->writers++;
1013                 pipe->r_counter++;
1014                 pipe->w_counter++;
1015                 if (pipe->readers == 1 || pipe->writers == 1)
1016                         wake_up_partner(pipe);
1017                 break;
1018
1019         default:
1020                 ret = -EINVAL;
1021                 goto err;
1022         }
1023
1024         /* Ok! */
1025         __pipe_unlock(pipe);
1026         return 0;
1027
1028 err_rd:
1029         if (!--pipe->readers)
1030                 wake_up_interruptible(&pipe->wait);
1031         ret = -ERESTARTSYS;
1032         goto err;
1033
1034 err_wr:
1035         if (!--pipe->writers)
1036                 wake_up_interruptible(&pipe->wait);
1037         ret = -ERESTARTSYS;
1038         goto err;
1039
1040 err:
1041         __pipe_unlock(pipe);
1042
1043         put_pipe_info(inode, pipe);
1044         return ret;
1045 }
1046
1047 const struct file_operations pipefifo_fops = {
1048         .open           = fifo_open,
1049         .llseek         = no_llseek,
1050         .read_iter      = pipe_read,
1051         .write_iter     = pipe_write,
1052         .poll           = pipe_poll,
1053         .unlocked_ioctl = pipe_ioctl,
1054         .release        = pipe_release,
1055         .fasync         = pipe_fasync,
1056 };
1057
1058 /*
1059  * Currently we rely on the pipe array holding a power-of-2 number
1060  * of pages. Returns 0 on error.
1061  */
1062 unsigned int round_pipe_size(unsigned long size)
1063 {
1064         if (size > (1U << 31))
1065                 return 0;
1066
1067         /* Minimum pipe size, as required by POSIX */
1068         if (size < PAGE_SIZE)
1069                 return PAGE_SIZE;
1070
1071         return roundup_pow_of_two(size);
1072 }
1073
1074 /*
1075  * Allocate a new array of pipe buffers and copy the info over. Returns the
1076  * pipe size if successful, or return -ERROR on error.
1077  */
1078 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1079 {
1080         struct pipe_buffer *bufs;
1081         unsigned int size, nr_slots, head, tail, mask, n;
1082         unsigned long user_bufs;
1083         long ret = 0;
1084
1085         size = round_pipe_size(arg);
1086         nr_slots = size >> PAGE_SHIFT;
1087
1088         if (!nr_slots)
1089                 return -EINVAL;
1090
1091         /*
1092          * If trying to increase the pipe capacity, check that an
1093          * unprivileged user is not trying to exceed various limits
1094          * (soft limit check here, hard limit check just below).
1095          * Decreasing the pipe capacity is always permitted, even
1096          * if the user is currently over a limit.
1097          */
1098         if (nr_slots > pipe->ring_size &&
1099                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1100                 return -EPERM;
1101
1102         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1103
1104         if (nr_slots > pipe->ring_size &&
1105                         (too_many_pipe_buffers_hard(user_bufs) ||
1106                          too_many_pipe_buffers_soft(user_bufs)) &&
1107                         is_unprivileged_user()) {
1108                 ret = -EPERM;
1109                 goto out_revert_acct;
1110         }
1111
1112         /*
1113          * We can shrink the pipe, if arg is greater than the ring occupancy.
1114          * Since we don't expect a lot of shrink+grow operations, just free and
1115          * allocate again like we would do for growing.  If the pipe currently
1116          * contains more buffers than arg, then return busy.
1117          */
1118         mask = pipe->ring_size - 1;
1119         head = pipe->head;
1120         tail = pipe->tail;
1121         n = pipe_occupancy(pipe->head, pipe->tail);
1122         if (nr_slots < n) {
1123                 ret = -EBUSY;
1124                 goto out_revert_acct;
1125         }
1126
1127         bufs = kcalloc(nr_slots, sizeof(*bufs),
1128                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1129         if (unlikely(!bufs)) {
1130                 ret = -ENOMEM;
1131                 goto out_revert_acct;
1132         }
1133
1134         /*
1135          * The pipe array wraps around, so just start the new one at zero
1136          * and adjust the indices.
1137          */
1138         if (n > 0) {
1139                 unsigned int h = head & mask;
1140                 unsigned int t = tail & mask;
1141                 if (h > t) {
1142                         memcpy(bufs, pipe->bufs + t,
1143                                n * sizeof(struct pipe_buffer));
1144                 } else {
1145                         unsigned int tsize = pipe->ring_size - t;
1146                         if (h > 0)
1147                                 memcpy(bufs + tsize, pipe->bufs,
1148                                        h * sizeof(struct pipe_buffer));
1149                         memcpy(bufs, pipe->bufs + t,
1150                                tsize * sizeof(struct pipe_buffer));
1151                 }
1152         }
1153
1154         head = n;
1155         tail = 0;
1156
1157         kfree(pipe->bufs);
1158         pipe->bufs = bufs;
1159         pipe->ring_size = nr_slots;
1160         pipe->max_usage = nr_slots;
1161         pipe->tail = tail;
1162         pipe->head = head;
1163         return pipe->max_usage * PAGE_SIZE;
1164
1165 out_revert_acct:
1166         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1167         return ret;
1168 }
1169
1170 /*
1171  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1172  * location, so checking ->i_pipe is not enough to verify that this is a
1173  * pipe.
1174  */
1175 struct pipe_inode_info *get_pipe_info(struct file *file)
1176 {
1177         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1178 }
1179
1180 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1181 {
1182         struct pipe_inode_info *pipe;
1183         long ret;
1184
1185         pipe = get_pipe_info(file);
1186         if (!pipe)
1187                 return -EBADF;
1188
1189         __pipe_lock(pipe);
1190
1191         switch (cmd) {
1192         case F_SETPIPE_SZ:
1193                 ret = pipe_set_size(pipe, arg);
1194                 break;
1195         case F_GETPIPE_SZ:
1196                 ret = pipe->max_usage * PAGE_SIZE;
1197                 break;
1198         default:
1199                 ret = -EINVAL;
1200                 break;
1201         }
1202
1203         __pipe_unlock(pipe);
1204         return ret;
1205 }
1206
1207 static const struct super_operations pipefs_ops = {
1208         .destroy_inode = free_inode_nonrcu,
1209         .statfs = simple_statfs,
1210 };
1211
1212 /*
1213  * pipefs should _never_ be mounted by userland - too much of security hassle,
1214  * no real gain from having the whole whorehouse mounted. So we don't need
1215  * any operations on the root directory. However, we need a non-trivial
1216  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1217  */
1218
1219 static int pipefs_init_fs_context(struct fs_context *fc)
1220 {
1221         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1222         if (!ctx)
1223                 return -ENOMEM;
1224         ctx->ops = &pipefs_ops;
1225         ctx->dops = &pipefs_dentry_operations;
1226         return 0;
1227 }
1228
1229 static struct file_system_type pipe_fs_type = {
1230         .name           = "pipefs",
1231         .init_fs_context = pipefs_init_fs_context,
1232         .kill_sb        = kill_anon_super,
1233 };
1234
1235 static int __init init_pipe_fs(void)
1236 {
1237         int err = register_filesystem(&pipe_fs_type);
1238
1239         if (!err) {
1240                 pipe_mnt = kern_mount(&pipe_fs_type);
1241                 if (IS_ERR(pipe_mnt)) {
1242                         err = PTR_ERR(pipe_mnt);
1243                         unregister_filesystem(&pipe_fs_type);
1244                 }
1245         }
1246         return err;
1247 }
1248
1249 fs_initcall(init_pipe_fs);