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