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 = 1;
 332                                 if (head - (tail - 1) == pipe->max_usage)
 333                                         wake_up_interruptible_sync_poll_locked(
 334                                                 &pipe->wait, EPOLLOUT | EPOLLWRNORM);
 335                                 spin_unlock_irq(&pipe->wait.lock);
 336                                 if (head - (tail - 1) == pipe->max_usage)
 337                                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 338                         }
 339                         total_len -= chars;
 340                         if (!total_len)
 341                                 break;  /* common path: read succeeded */
 342                         if (!pipe_empty(head, tail))    /* More to do? */
 343                                 continue;
 344                 }
 345
 346                 if (!pipe->writers)
 347                         break;
 348                 if (!pipe->waiting_writers) {
 349                         /* syscall merging: Usually we must not sleep
 350                          * if O_NONBLOCK is set, or if we got some data.
 351                          * But if a writer sleeps in kernel space, then
 352                          * we can wait for that data without violating POSIX.
 353                          */
 354                         if (ret)
 355                                 break;
 356                         if (filp->f_flags & O_NONBLOCK) {
 357                                 ret = -EAGAIN;
 358                                 break;
 359                         }
 360                 }
 361                 if (signal_pending(current)) {
 362                         if (!ret)
 363                                 ret = -ERESTARTSYS;
 364                         break;
 365                 }
 366                 pipe_wait(pipe);
 367         }
 368         __pipe_unlock(pipe);
 369
 370         /* Signal writers asynchronously that there is more room. */
 371         if (do_wakeup) {
 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, max_usage, mask;
 391         ssize_t ret = 0;
 392         int do_wakeup = 0;
 393         size_t total_len = iov_iter_count(from);
 394         ssize_t chars;
 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         head = pipe->head;
 409         max_usage = pipe->max_usage;
 410         mask = pipe->ring_size - 1;
 411
 412         /* We try to merge small writes */
 413         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 414         if (!pipe_empty(head, pipe->tail) && chars != 0) {
 415                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 416                 int offset = buf->offset + buf->len;
 417
 418                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 419                         ret = pipe_buf_confirm(pipe, buf);
 420                         if (ret)
 421                                 goto out;
 422
 423                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 424                         if (unlikely(ret < chars)) {
 425                                 ret = -EFAULT;
 426                                 goto out;
 427                         }
 428                         do_wakeup = 1;
 429                         buf->len += ret;
 430                         if (!iov_iter_count(from))
 431                                 goto out;
 432                 }
 433         }
 434
 435         for (;;) {
 436                 if (!pipe->readers) {
 437                         send_sig(SIGPIPE, current, 0);
 438                         if (!ret)
 439                                 ret = -EPIPE;
 440                         break;
 441                 }
 442
 443                 head = pipe->head;
 444                 if (!pipe_full(head, pipe->tail, max_usage)) {
 445                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 446                         struct page *page = pipe->tmp_page;
 447                         int copied;
 448
 449                         if (!page) {
 450                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 451                                 if (unlikely(!page)) {
 452                                         ret = ret ? : -ENOMEM;
 453                                         break;
 454                                 }
 455                                 pipe->tmp_page = page;
 456                         }
 457
 458                         /* Allocate a slot in the ring in advance and attach an
 459                          * empty buffer.  If we fault or otherwise fail to use
 460                          * it, either the reader will consume it or it'll still
 461                          * be there for the next write.
 462                          */
 463                         spin_lock_irq(&pipe->wait.lock);
 464
 465                         head = pipe->head;
 466                         pipe->head = head + 1;
 467
 468                         /* Always wake up, even if the copy fails. Otherwise
 469                          * we lock up (O_NONBLOCK-)readers that sleep due to
 470                          * syscall merging.
 471                          * FIXME! Is this really true?
 472                          */
 473                         wake_up_interruptible_sync_poll_locked(
 474                                 &pipe->wait, EPOLLIN | EPOLLRDNORM);
 475
 476                         spin_unlock_irq(&pipe->wait.lock);
 477                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 478
 479                         /* Insert it into the buffer array */
 480                         buf = &pipe->bufs[head & mask];
 481                         buf->page = page;
 482                         buf->ops = &anon_pipe_buf_ops;
 483                         buf->offset = 0;
 484                         buf->len = 0;
 485                         buf->flags = 0;
 486                         if (is_packetized(filp)) {
 487                                 buf->ops = &packet_pipe_buf_ops;
 488                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 489                         }
 490                         pipe->tmp_page = NULL;
 491
 492                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 493                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 494                                 if (!ret)
 495                                         ret = -EFAULT;
 496                                 break;
 497                         }
 498                         ret += copied;
 499                         buf->offset = 0;
 500                         buf->len = copied;
 501
 502                         if (!iov_iter_count(from))
 503                                 break;
 504                 }
 505
 506                 if (!pipe_full(head, pipe->tail, max_usage))
 507                         continue;
 508
 509                 /* Wait for buffer space to become available. */
 510                 if (filp->f_flags & O_NONBLOCK) {
 511                         if (!ret)
 512                                 ret = -EAGAIN;
 513                         break;
 514                 }
 515                 if (signal_pending(current)) {
 516                         if (!ret)
 517                                 ret = -ERESTARTSYS;
 518                         break;
 519                 }
 520                 pipe->waiting_writers++;
 521                 pipe_wait(pipe);
 522                 pipe->waiting_writers--;
 523         }
 524 out:
 525         __pipe_unlock(pipe);
 526         if (do_wakeup) {
 527                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 528                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 529         }
 530         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 531                 int err = file_update_time(filp);
 532                 if (err)
 533                         ret = err;
 534                 sb_end_write(file_inode(filp)->i_sb);
 535         }
 536         return ret;
 537 }
 538
 539 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 540 {
 541         struct pipe_inode_info *pipe = filp->private_data;
 542         int count, head, tail, mask;
 543
 544         switch (cmd) {
 545                 case FIONREAD:
 546                         __pipe_lock(pipe);
 547                         count = 0;
 548                         head = pipe->head;
 549                         tail = pipe->tail;
 550                         mask = pipe->ring_size - 1;
 551
 552                         while (tail != head) {
 553                                 count += pipe->bufs[tail & mask].len;
 554                                 tail++;
 555                         }
 556                         __pipe_unlock(pipe);
 557
 558                         return put_user(count, (int __user *)arg);
 559                 default:
 560                         return -ENOIOCTLCMD;
 561         }
 562 }
 563
 564 /* No kernel lock held - fine */
 565 static __poll_t
 566 pipe_poll(struct file *filp, poll_table *wait)
 567 {
 568         __poll_t mask;
 569         struct pipe_inode_info *pipe = filp->private_data;
 570         unsigned int head = READ_ONCE(pipe->head);
 571         unsigned int tail = READ_ONCE(pipe->tail);
 572
 573         poll_wait(filp, &pipe->wait, wait);
 574
 575         BUG_ON(pipe_occupancy(head, tail) > pipe->ring_size);
 576
 577         /* Reading only -- no need for acquiring the semaphore.  */
 578         mask = 0;
 579         if (filp->f_mode & FMODE_READ) {
 580                 if (!pipe_empty(head, tail))
 581                         mask |= EPOLLIN | EPOLLRDNORM;
 582                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 583                         mask |= EPOLLHUP;
 584         }
 585
 586         if (filp->f_mode & FMODE_WRITE) {
 587                 if (!pipe_full(head, tail, pipe->max_usage))
 588                         mask |= EPOLLOUT | EPOLLWRNORM;
 589                 /*
 590                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 591                  * behave exactly like pipes for poll().
 592                  */
 593                 if (!pipe->readers)
 594                         mask |= EPOLLERR;
 595         }
 596
 597         return mask;
 598 }
 599
 600 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 601 {
 602         int kill = 0;
 603
 604         spin_lock(&inode->i_lock);
 605         if (!--pipe->files) {
 606                 inode->i_pipe = NULL;
 607                 kill = 1;
 608         }
 609         spin_unlock(&inode->i_lock);
 610
 611         if (kill)
 612                 free_pipe_info(pipe);
 613 }
 614
 615 static int
 616 pipe_release(struct inode *inode, struct file *file)
 617 {
 618         struct pipe_inode_info *pipe = file->private_data;
 619
 620         __pipe_lock(pipe);
 621         if (file->f_mode & FMODE_READ)
 622                 pipe->readers--;
 623         if (file->f_mode & FMODE_WRITE)
 624                 pipe->writers--;
 625
 626         if (pipe->readers || pipe->writers) {
 627                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 628                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 629                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 630         }
 631         __pipe_unlock(pipe);
 632
 633         put_pipe_info(inode, pipe);
 634         return 0;
 635 }
 636
 637 static int
 638 pipe_fasync(int fd, struct file *filp, int on)
 639 {
 640         struct pipe_inode_info *pipe = filp->private_data;
 641         int retval = 0;
 642
 643         __pipe_lock(pipe);
 644         if (filp->f_mode & FMODE_READ)
 645                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 646         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 647                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 648                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 649                         /* this can happen only if on == T */
 650                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 651         }
 652         __pipe_unlock(pipe);
 653         return retval;
 654 }
 655
 656 static unsigned long account_pipe_buffers(struct user_struct *user,
 657                                  unsigned long old, unsigned long new)
 658 {
 659         return atomic_long_add_return(new - old, &user->pipe_bufs);
 660 }
 661
 662 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 663 {
 664         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 665
 666         return soft_limit && user_bufs > soft_limit;
 667 }
 668
 669 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 670 {
 671         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 672
 673         return hard_limit && user_bufs > hard_limit;
 674 }
 675
 676 static bool is_unprivileged_user(void)
 677 {
 678         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 679 }
 680
 681 struct pipe_inode_info *alloc_pipe_info(void)
 682 {
 683         struct pipe_inode_info *pipe;
 684         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 685         struct user_struct *user = get_current_user();
 686         unsigned long user_bufs;
 687         unsigned int max_size = READ_ONCE(pipe_max_size);
 688
 689         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 690         if (pipe == NULL)
 691                 goto out_free_uid;
 692
 693         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 694                 pipe_bufs = max_size >> PAGE_SHIFT;
 695
 696         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 697
 698         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 699                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 700                 pipe_bufs = 1;
 701         }
 702
 703         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 704                 goto out_revert_acct;
 705
 706         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 707                              GFP_KERNEL_ACCOUNT);
 708
 709         if (pipe->bufs) {
 710                 init_waitqueue_head(&pipe->wait);
 711                 pipe->r_counter = pipe->w_counter = 1;
 712                 pipe->max_usage = pipe_bufs;
 713                 pipe->ring_size = pipe_bufs;
 714                 pipe->user = user;
 715                 mutex_init(&pipe->mutex);
 716                 return pipe;
 717         }
 718
 719 out_revert_acct:
 720         (void) account_pipe_buffers(user, pipe_bufs, 0);
 721         kfree(pipe);
 722 out_free_uid:
 723         free_uid(user);
 724         return NULL;
 725 }
 726
 727 void free_pipe_info(struct pipe_inode_info *pipe)
 728 {
 729         int i;
 730
 731         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 732         free_uid(pipe->user);
 733         for (i = 0; i < pipe->ring_size; i++) {
 734                 struct pipe_buffer *buf = pipe->bufs + i;
 735                 if (buf->ops)
 736                         pipe_buf_release(pipe, buf);
 737         }
 738         if (pipe->tmp_page)
 739                 __free_page(pipe->tmp_page);
 740         kfree(pipe->bufs);
 741         kfree(pipe);
 742 }
 743
 744 static struct vfsmount *pipe_mnt __read_mostly;
 745
 746 /*
 747  * pipefs_dname() is called from d_path().
 748  */
 749 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 750 {
 751         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 752                                 d_inode(dentry)->i_ino);
 753 }
 754
 755 static const struct dentry_operations pipefs_dentry_operations = {
 756         .d_dname        = pipefs_dname,
 757 };
 758
 759 static struct inode * get_pipe_inode(void)
 760 {
 761         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 762         struct pipe_inode_info *pipe;
 763
 764         if (!inode)
 765                 goto fail_inode;
 766
 767         inode->i_ino = get_next_ino();
 768
 769         pipe = alloc_pipe_info();
 770         if (!pipe)
 771                 goto fail_iput;
 772
 773         inode->i_pipe = pipe;
 774         pipe->files = 2;
 775         pipe->readers = pipe->writers = 1;
 776         inode->i_fop = &pipefifo_fops;
 777
 778         /*
 779          * Mark the inode dirty from the very beginning,
 780          * that way it will never be moved to the dirty
 781          * list because "mark_inode_dirty()" will think
 782          * that it already _is_ on the dirty list.
 783          */
 784         inode->i_state = I_DIRTY;
 785         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 786         inode->i_uid = current_fsuid();
 787         inode->i_gid = current_fsgid();
 788         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 789
 790         return inode;
 791
 792 fail_iput:
 793         iput(inode);
 794
 795 fail_inode:
 796         return NULL;
 797 }
 798
 799 int create_pipe_files(struct file **res, int flags)
 800 {
 801         struct inode *inode = get_pipe_inode();
 802         struct file *f;
 803
 804         if (!inode)
 805                 return -ENFILE;
 806
 807         f = alloc_file_pseudo(inode, pipe_mnt, "",
 808                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 809                                 &pipefifo_fops);
 810         if (IS_ERR(f)) {
 811                 free_pipe_info(inode->i_pipe);
 812                 iput(inode);
 813                 return PTR_ERR(f);
 814         }
 815
 816         f->private_data = inode->i_pipe;
 817
 818         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 819                                   &pipefifo_fops);
 820         if (IS_ERR(res[0])) {
 821                 put_pipe_info(inode, inode->i_pipe);
 822                 fput(f);
 823                 return PTR_ERR(res[0]);
 824         }
 825         res[0]->private_data = inode->i_pipe;
 826         res[1] = f;
 827         return 0;
 828 }
 829
 830 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 831 {
 832         int error;
 833         int fdw, fdr;
 834
 835         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 836                 return -EINVAL;
 837
 838         error = create_pipe_files(files, flags);
 839         if (error)
 840                 return error;
 841
 842         error = get_unused_fd_flags(flags);
 843         if (error < 0)
 844                 goto err_read_pipe;
 845         fdr = error;
 846
 847         error = get_unused_fd_flags(flags);
 848         if (error < 0)
 849                 goto err_fdr;
 850         fdw = error;
 851
 852         audit_fd_pair(fdr, fdw);
 853         fd[0] = fdr;
 854         fd[1] = fdw;
 855         return 0;
 856
 857  err_fdr:
 858         put_unused_fd(fdr);
 859  err_read_pipe:
 860         fput(files[0]);
 861         fput(files[1]);
 862         return error;
 863 }
 864
 865 int do_pipe_flags(int *fd, int flags)
 866 {
 867         struct file *files[2];
 868         int error = __do_pipe_flags(fd, files, flags);
 869         if (!error) {
 870                 fd_install(fd[0], files[0]);
 871                 fd_install(fd[1], files[1]);
 872         }
 873         return error;
 874 }
 875
 876 /*
 877  * sys_pipe() is the normal C calling standard for creating
 878  * a pipe. It's not the way Unix traditionally does this, though.
 879  */
 880 static int do_pipe2(int __user *fildes, int flags)
 881 {
 882         struct file *files[2];
 883         int fd[2];
 884         int error;
 885
 886         error = __do_pipe_flags(fd, files, flags);
 887         if (!error) {
 888                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 889                         fput(files[0]);
 890                         fput(files[1]);
 891                         put_unused_fd(fd[0]);
 892                         put_unused_fd(fd[1]);
 893                         error = -EFAULT;
 894                 } else {
 895                         fd_install(fd[0], files[0]);
 896                         fd_install(fd[1], files[1]);
 897                 }
 898         }
 899         return error;
 900 }
 901
 902 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 903 {
 904         return do_pipe2(fildes, flags);
 905 }
 906
 907 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 908 {
 909         return do_pipe2(fildes, 0);
 910 }
 911
 912 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 913 {
 914         int cur = *cnt;
 915
 916         while (cur == *cnt) {
 917                 pipe_wait(pipe);
 918                 if (signal_pending(current))
 919                         break;
 920         }
 921         return cur == *cnt ? -ERESTARTSYS : 0;
 922 }
 923
 924 static void wake_up_partner(struct pipe_inode_info *pipe)
 925 {
 926         wake_up_interruptible(&pipe->wait);
 927 }
 928
 929 static int fifo_open(struct inode *inode, struct file *filp)
 930 {
 931         struct pipe_inode_info *pipe;
 932         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 933         int ret;
 934
 935         filp->f_version = 0;
 936
 937         spin_lock(&inode->i_lock);
 938         if (inode->i_pipe) {
 939                 pipe = inode->i_pipe;
 940                 pipe->files++;
 941                 spin_unlock(&inode->i_lock);
 942         } else {
 943                 spin_unlock(&inode->i_lock);
 944                 pipe = alloc_pipe_info();
 945                 if (!pipe)
 946                         return -ENOMEM;
 947                 pipe->files = 1;
 948                 spin_lock(&inode->i_lock);
 949                 if (unlikely(inode->i_pipe)) {
 950                         inode->i_pipe->files++;
 951                         spin_unlock(&inode->i_lock);
 952                         free_pipe_info(pipe);
 953                         pipe = inode->i_pipe;
 954                 } else {
 955                         inode->i_pipe = pipe;
 956                         spin_unlock(&inode->i_lock);
 957                 }
 958         }
 959         filp->private_data = pipe;
 960         /* OK, we have a pipe and it's pinned down */
 961
 962         __pipe_lock(pipe);
 963
 964         /* We can only do regular read/write on fifos */
 965         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 966
 967         switch (filp->f_mode) {
 968         case FMODE_READ:
 969         /*
 970          *  O_RDONLY
 971          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 972          *  opened, even when there is no process writing the FIFO.
 973          */
 974                 pipe->r_counter++;
 975                 if (pipe->readers++ == 0)
 976                         wake_up_partner(pipe);
 977
 978                 if (!is_pipe && !pipe->writers) {
 979                         if ((filp->f_flags & O_NONBLOCK)) {
 980                                 /* suppress EPOLLHUP until we have
 981                                  * seen a writer */
 982                                 filp->f_version = pipe->w_counter;
 983                         } else {
 984                                 if (wait_for_partner(pipe, &pipe->w_counter))
 985                                         goto err_rd;
 986                         }
 987                 }
 988                 break;
 989
 990         case FMODE_WRITE:
 991         /*
 992          *  O_WRONLY
 993          *  POSIX.1 says that O_NONBLOCK means return -1 with
 994          *  errno=ENXIO when there is no process reading the FIFO.
 995          */
 996                 ret = -ENXIO;
 997                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 998                         goto err;
 999
1000                 pipe->w_counter++;
1001                 if (!pipe->writers++)
1002                         wake_up_partner(pipe);
1003
1004                 if (!is_pipe && !pipe->readers) {
1005                         if (wait_for_partner(pipe, &pipe->r_counter))
1006                                 goto err_wr;
1007                 }
1008                 break;
1009
1010         case FMODE_READ | FMODE_WRITE:
1011         /*
1012          *  O_RDWR
1013          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1014          *  This implementation will NEVER block on a O_RDWR open, since
1015          *  the process can at least talk to itself.
1016          */
1017
1018                 pipe->readers++;
1019                 pipe->writers++;
1020                 pipe->r_counter++;
1021                 pipe->w_counter++;
1022                 if (pipe->readers == 1 || pipe->writers == 1)
1023                         wake_up_partner(pipe);
1024                 break;
1025
1026         default:
1027                 ret = -EINVAL;
1028                 goto err;
1029         }
1030
1031         /* Ok! */
1032         __pipe_unlock(pipe);
1033         return 0;
1034
1035 err_rd:
1036         if (!--pipe->readers)
1037                 wake_up_interruptible(&pipe->wait);
1038         ret = -ERESTARTSYS;
1039         goto err;
1040
1041 err_wr:
1042         if (!--pipe->writers)
1043                 wake_up_interruptible(&pipe->wait);
1044         ret = -ERESTARTSYS;
1045         goto err;
1046
1047 err:
1048         __pipe_unlock(pipe);
1049
1050         put_pipe_info(inode, pipe);
1051         return ret;
1052 }
1053
1054 const struct file_operations pipefifo_fops = {
1055         .open           = fifo_open,
1056         .llseek         = no_llseek,
1057         .read_iter      = pipe_read,
1058         .write_iter     = pipe_write,
1059         .poll           = pipe_poll,
1060         .unlocked_ioctl = pipe_ioctl,
1061         .release        = pipe_release,
1062         .fasync         = pipe_fasync,
1063 };
1064
1065 /*
1066  * Currently we rely on the pipe array holding a power-of-2 number
1067  * of pages. Returns 0 on error.
1068  */
1069 unsigned int round_pipe_size(unsigned long size)
1070 {
1071         if (size > (1U << 31))
1072                 return 0;
1073
1074         /* Minimum pipe size, as required by POSIX */
1075         if (size < PAGE_SIZE)
1076                 return PAGE_SIZE;
1077
1078         return roundup_pow_of_two(size);
1079 }
1080
1081 /*
1082  * Allocate a new array of pipe buffers and copy the info over. Returns the
1083  * pipe size if successful, or return -ERROR on error.
1084  */
1085 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1086 {
1087         struct pipe_buffer *bufs;
1088         unsigned int size, nr_slots, head, tail, mask, n;
1089         unsigned long user_bufs;
1090         long ret = 0;
1091
1092         size = round_pipe_size(arg);
1093         nr_slots = size >> PAGE_SHIFT;
1094
1095         if (!nr_slots)
1096                 return -EINVAL;
1097
1098         /*
1099          * If trying to increase the pipe capacity, check that an
1100          * unprivileged user is not trying to exceed various limits
1101          * (soft limit check here, hard limit check just below).
1102          * Decreasing the pipe capacity is always permitted, even
1103          * if the user is currently over a limit.
1104          */
1105         if (nr_slots > pipe->ring_size &&
1106                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1107                 return -EPERM;
1108
1109         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1110
1111         if (nr_slots > pipe->ring_size &&
1112                         (too_many_pipe_buffers_hard(user_bufs) ||
1113                          too_many_pipe_buffers_soft(user_bufs)) &&
1114                         is_unprivileged_user()) {
1115                 ret = -EPERM;
1116                 goto out_revert_acct;
1117         }
1118
1119         /*
1120          * We can shrink the pipe, if arg is greater than the ring occupancy.
1121          * Since we don't expect a lot of shrink+grow operations, just free and
1122          * allocate again like we would do for growing.  If the pipe currently
1123          * contains more buffers than arg, then return busy.
1124          */
1125         mask = pipe->ring_size - 1;
1126         head = pipe->head;
1127         tail = pipe->tail;
1128         n = pipe_occupancy(pipe->head, pipe->tail);
1129         if (nr_slots < n) {
1130                 ret = -EBUSY;
1131                 goto out_revert_acct;
1132         }
1133
1134         bufs = kcalloc(nr_slots, sizeof(*bufs),
1135                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1136         if (unlikely(!bufs)) {
1137                 ret = -ENOMEM;
1138                 goto out_revert_acct;
1139         }
1140
1141         /*
1142          * The pipe array wraps around, so just start the new one at zero
1143          * and adjust the indices.
1144          */
1145         if (n > 0) {
1146                 unsigned int h = head & mask;
1147                 unsigned int t = tail & mask;
1148                 if (h > t) {
1149                         memcpy(bufs, pipe->bufs + t,
1150                                n * sizeof(struct pipe_buffer));
1151                 } else {
1152                         unsigned int tsize = pipe->ring_size - t;
1153                         if (h > 0)
1154                                 memcpy(bufs + tsize, pipe->bufs,
1155                                        h * sizeof(struct pipe_buffer));
1156                         memcpy(bufs, pipe->bufs + t,
1157                                tsize * sizeof(struct pipe_buffer));
1158                 }
1159         }
1160
1161         head = n;
1162         tail = 0;
1163
1164         kfree(pipe->bufs);
1165         pipe->bufs = bufs;
1166         pipe->ring_size = nr_slots;
1167         pipe->max_usage = nr_slots;
1168         pipe->tail = tail;
1169         pipe->head = head;
1170         return pipe->max_usage * PAGE_SIZE;
1171
1172 out_revert_acct:
1173         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1174         return ret;
1175 }
1176
1177 /*
1178  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1179  * location, so checking ->i_pipe is not enough to verify that this is a
1180  * pipe.
1181  */
1182 struct pipe_inode_info *get_pipe_info(struct file *file)
1183 {
1184         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1185 }
1186
1187 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1188 {
1189         struct pipe_inode_info *pipe;
1190         long ret;
1191
1192         pipe = get_pipe_info(file);
1193         if (!pipe)
1194                 return -EBADF;
1195
1196         __pipe_lock(pipe);
1197
1198         switch (cmd) {
1199         case F_SETPIPE_SZ:
1200                 ret = pipe_set_size(pipe, arg);
1201                 break;
1202         case F_GETPIPE_SZ:
1203                 ret = pipe->max_usage * PAGE_SIZE;
1204                 break;
1205         default:
1206                 ret = -EINVAL;
1207                 break;
1208         }
1209
1210         __pipe_unlock(pipe);
1211         return ret;
1212 }
1213
1214 static const struct super_operations pipefs_ops = {
1215         .destroy_inode = free_inode_nonrcu,
1216         .statfs = simple_statfs,
1217 };
1218
1219 /*
1220  * pipefs should _never_ be mounted by userland - too much of security hassle,
1221  * no real gain from having the whole whorehouse mounted. So we don't need
1222  * any operations on the root directory. However, we need a non-trivial
1223  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1224  */
1225
1226 static int pipefs_init_fs_context(struct fs_context *fc)
1227 {
1228         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1229         if (!ctx)
1230                 return -ENOMEM;
1231         ctx->ops = &pipefs_ops;
1232         ctx->dops = &pipefs_dentry_operations;
1233         return 0;
1234 }
1235
1236 static struct file_system_type pipe_fs_type = {
1237         .name           = "pipefs",
1238         .init_fs_context = pipefs_init_fs_context,
1239         .kill_sb        = kill_anon_super,
1240 };
1241
1242 static int __init init_pipe_fs(void)
1243 {
1244         int err = register_filesystem(&pipe_fs_type);
1245
1246         if (!err) {
1247                 pipe_mnt = kern_mount(&pipe_fs_type);
1248                 if (IS_ERR(pipe_mnt)) {
1249                         err = PTR_ERR(pipe_mnt);
1250                         unregister_filesystem(&pipe_fs_type);
1251                 }
1252         }
1253         return err;
1254 }
1255
1256 fs_initcall(init_pipe_fs);