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/magic.h>
  18 #include <linux/pipe_fs_i.h>
  19 #include <linux/uio.h>
  20 #include <linux/highmem.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/audit.h>
  23 #include <linux/syscalls.h>
  24 #include <linux/fcntl.h>
  25 #include <linux/memcontrol.h>
  26
  27 #include <linux/uaccess.h>
  28 #include <asm/ioctls.h>
  29
  30 #include "internal.h"
  31
  32 /*
  33  * The max size that a non-root user is allowed to grow the pipe. Can
  34  * be set by root in /proc/sys/fs/pipe-max-size
  35  */
  36 unsigned int pipe_max_size = 1048576;
  37
  38 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  39  * matches default values.
  40  */
  41 unsigned long pipe_user_pages_hard;
  42 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  43
  44 /*
  45  * We use a start+len construction, which provides full use of the
  46  * allocated memory.
  47  * -- Florian Coosmann (FGC)
  48  *
  49  * Reads with count = 0 should always return 0.
  50  * -- Julian Bradfield 1999-06-07.
  51  *
  52  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  53  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  54  *
  55  * pipe_read & write cleanup
  56  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  57  */
  58
  59 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  60 {
  61         if (pipe->files)
  62                 mutex_lock_nested(&pipe->mutex, subclass);
  63 }
  64
  65 void pipe_lock(struct pipe_inode_info *pipe)
  66 {
  67         /*
  68          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  69          */
  70         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  71 }
  72 EXPORT_SYMBOL(pipe_lock);
  73
  74 void pipe_unlock(struct pipe_inode_info *pipe)
  75 {
  76         if (pipe->files)
  77                 mutex_unlock(&pipe->mutex);
  78 }
  79 EXPORT_SYMBOL(pipe_unlock);
  80
  81 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  82 {
  83         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  84 }
  85
  86 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  87 {
  88         mutex_unlock(&pipe->mutex);
  89 }
  90
  91 void pipe_double_lock(struct pipe_inode_info *pipe1,
  92                       struct pipe_inode_info *pipe2)
  93 {
  94         BUG_ON(pipe1 == pipe2);
  95
  96         if (pipe1 < pipe2) {
  97                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  98                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  99         } else {
 100                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 101                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 102         }
 103 }
 104
 105 /* Drop the inode semaphore and wait for a pipe event, atomically */
 106 void pipe_wait(struct pipe_inode_info *pipe)
 107 {
 108         DEFINE_WAIT(wait);
 109
 110         /*
 111          * Pipes are system-local resources, so sleeping on them
 112          * is considered a noninteractive wait:
 113          */
 114         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 115         pipe_unlock(pipe);
 116         schedule();
 117         finish_wait(&pipe->wait, &wait);
 118         pipe_lock(pipe);
 119 }
 120
 121 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 122                                   struct pipe_buffer *buf)
 123 {
 124         struct page *page = buf->page;
 125
 126         /*
 127          * If nobody else uses this page, and we don't already have a
 128          * temporary page, let's keep track of it as a one-deep
 129          * allocation cache. (Otherwise just release our reference to it)
 130          */
 131         if (page_count(page) == 1 && !pipe->tmp_page)
 132                 pipe->tmp_page = page;
 133         else
 134                 put_page(page);
 135 }
 136
 137 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 138                                struct pipe_buffer *buf)
 139 {
 140         struct page *page = buf->page;
 141
 142         if (page_count(page) == 1) {
 143                 if (memcg_kmem_enabled())
 144                         memcg_kmem_uncharge(page, 0);
 145                 __SetPageLocked(page);
 146                 return 0;
 147         }
 148         return 1;
 149 }
 150
 151 /**
 152  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 153  * @pipe:       the pipe that the buffer belongs to
 154  * @buf:        the buffer to attempt to steal
 155  *
 156  * Description:
 157  *      This function attempts to steal the &struct page attached to
 158  *      @buf. If successful, this function returns 0 and returns with
 159  *      the page locked. The caller may then reuse the page for whatever
 160  *      he wishes; the typical use is insertion into a different file
 161  *      page cache.
 162  */
 163 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 164                            struct pipe_buffer *buf)
 165 {
 166         struct page *page = buf->page;
 167
 168         /*
 169          * A reference of one is golden, that means that the owner of this
 170          * page is the only one holding a reference to it. lock the page
 171          * and return OK.
 172          */
 173         if (page_count(page) == 1) {
 174                 lock_page(page);
 175                 return 0;
 176         }
 177
 178         return 1;
 179 }
 180 EXPORT_SYMBOL(generic_pipe_buf_steal);
 181
 182 /**
 183  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 184  * @pipe:       the pipe that the buffer belongs to
 185  * @buf:        the buffer to get a reference to
 186  *
 187  * Description:
 188  *      This function grabs an extra reference to @buf. It's used in
 189  *      in the tee() system call, when we duplicate the buffers in one
 190  *      pipe into another.
 191  */
 192 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 193 {
 194         get_page(buf->page);
 195 }
 196 EXPORT_SYMBOL(generic_pipe_buf_get);
 197
 198 /**
 199  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 200  * @info:       the pipe that the buffer belongs to
 201  * @buf:        the buffer to confirm
 202  *
 203  * Description:
 204  *      This function does nothing, because the generic pipe code uses
 205  *      pages that are always good when inserted into the pipe.
 206  */
 207 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 208                              struct pipe_buffer *buf)
 209 {
 210         return 0;
 211 }
 212 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 213
 214 /**
 215  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 216  * @pipe:       the pipe that the buffer belongs to
 217  * @buf:        the buffer to put a reference to
 218  *
 219  * Description:
 220  *      This function releases a reference to @buf.
 221  */
 222 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 223                               struct pipe_buffer *buf)
 224 {
 225         put_page(buf->page);
 226 }
 227 EXPORT_SYMBOL(generic_pipe_buf_release);
 228
 229 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 230         .can_merge = 1,
 231         .confirm = generic_pipe_buf_confirm,
 232         .release = anon_pipe_buf_release,
 233         .steal = anon_pipe_buf_steal,
 234         .get = generic_pipe_buf_get,
 235 };
 236
 237 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 238         .can_merge = 0,
 239         .confirm = generic_pipe_buf_confirm,
 240         .release = anon_pipe_buf_release,
 241         .steal = anon_pipe_buf_steal,
 242         .get = generic_pipe_buf_get,
 243 };
 244
 245 static ssize_t
 246 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 247 {
 248         size_t total_len = iov_iter_count(to);
 249         struct file *filp = iocb->ki_filp;
 250         struct pipe_inode_info *pipe = filp->private_data;
 251         int do_wakeup;
 252         ssize_t ret;
 253
 254         /* Null read succeeds. */
 255         if (unlikely(total_len == 0))
 256                 return 0;
 257
 258         do_wakeup = 0;
 259         ret = 0;
 260         __pipe_lock(pipe);
 261         for (;;) {
 262                 int bufs = pipe->nrbufs;
 263                 if (bufs) {
 264                         int curbuf = pipe->curbuf;
 265                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 266                         size_t chars = buf->len;
 267                         size_t written;
 268                         int error;
 269
 270                         if (chars > total_len)
 271                                 chars = total_len;
 272
 273                         error = pipe_buf_confirm(pipe, buf);
 274                         if (error) {
 275                                 if (!ret)
 276                                         ret = error;
 277                                 break;
 278                         }
 279
 280                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 281                         if (unlikely(written < chars)) {
 282                                 if (!ret)
 283                                         ret = -EFAULT;
 284                                 break;
 285                         }
 286                         ret += chars;
 287                         buf->offset += chars;
 288                         buf->len -= chars;
 289
 290                         /* Was it a packet buffer? Clean up and exit */
 291                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 292                                 total_len = chars;
 293                                 buf->len = 0;
 294                         }
 295
 296                         if (!buf->len) {
 297                                 pipe_buf_release(pipe, buf);
 298                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 299                                 pipe->curbuf = curbuf;
 300                                 pipe->nrbufs = --bufs;
 301                                 do_wakeup = 1;
 302                         }
 303                         total_len -= chars;
 304                         if (!total_len)
 305                                 break;  /* common path: read succeeded */
 306                 }
 307                 if (bufs)       /* More to do? */
 308                         continue;
 309                 if (!pipe->writers)
 310                         break;
 311                 if (!pipe->waiting_writers) {
 312                         /* syscall merging: Usually we must not sleep
 313                          * if O_NONBLOCK is set, or if we got some data.
 314                          * But if a writer sleeps in kernel space, then
 315                          * we can wait for that data without violating POSIX.
 316                          */
 317                         if (ret)
 318                                 break;
 319                         if (filp->f_flags & O_NONBLOCK) {
 320                                 ret = -EAGAIN;
 321                                 break;
 322                         }
 323                 }
 324                 if (signal_pending(current)) {
 325                         if (!ret)
 326                                 ret = -ERESTARTSYS;
 327                         break;
 328                 }
 329                 if (do_wakeup) {
 330                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 331                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 332                 }
 333                 pipe_wait(pipe);
 334         }
 335         __pipe_unlock(pipe);
 336
 337         /* Signal writers asynchronously that there is more room. */
 338         if (do_wakeup) {
 339                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 340                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 341         }
 342         if (ret > 0)
 343                 file_accessed(filp);
 344         return ret;
 345 }
 346
 347 static inline int is_packetized(struct file *file)
 348 {
 349         return (file->f_flags & O_DIRECT) != 0;
 350 }
 351
 352 static ssize_t
 353 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 354 {
 355         struct file *filp = iocb->ki_filp;
 356         struct pipe_inode_info *pipe = filp->private_data;
 357         ssize_t ret = 0;
 358         int do_wakeup = 0;
 359         size_t total_len = iov_iter_count(from);
 360         ssize_t chars;
 361
 362         /* Null write succeeds. */
 363         if (unlikely(total_len == 0))
 364                 return 0;
 365
 366         __pipe_lock(pipe);
 367
 368         if (!pipe->readers) {
 369                 send_sig(SIGPIPE, current, 0);
 370                 ret = -EPIPE;
 371                 goto out;
 372         }
 373
 374         /* We try to merge small writes */
 375         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 376         if (pipe->nrbufs && chars != 0) {
 377                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 378                                                         (pipe->buffers - 1);
 379                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 380                 int offset = buf->offset + buf->len;
 381
 382                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 383                         ret = pipe_buf_confirm(pipe, buf);
 384                         if (ret)
 385                                 goto out;
 386
 387                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 388                         if (unlikely(ret < chars)) {
 389                                 ret = -EFAULT;
 390                                 goto out;
 391                         }
 392                         do_wakeup = 1;
 393                         buf->len += ret;
 394                         if (!iov_iter_count(from))
 395                                 goto out;
 396                 }
 397         }
 398
 399         for (;;) {
 400                 int bufs;
 401
 402                 if (!pipe->readers) {
 403                         send_sig(SIGPIPE, current, 0);
 404                         if (!ret)
 405                                 ret = -EPIPE;
 406                         break;
 407                 }
 408                 bufs = pipe->nrbufs;
 409                 if (bufs < pipe->buffers) {
 410                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 411                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 412                         struct page *page = pipe->tmp_page;
 413                         int copied;
 414
 415                         if (!page) {
 416                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 417                                 if (unlikely(!page)) {
 418                                         ret = ret ? : -ENOMEM;
 419                                         break;
 420                                 }
 421                                 pipe->tmp_page = page;
 422                         }
 423                         /* Always wake up, even if the copy fails. Otherwise
 424                          * we lock up (O_NONBLOCK-)readers that sleep due to
 425                          * syscall merging.
 426                          * FIXME! Is this really true?
 427                          */
 428                         do_wakeup = 1;
 429                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 430                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 431                                 if (!ret)
 432                                         ret = -EFAULT;
 433                                 break;
 434                         }
 435                         ret += copied;
 436
 437                         /* Insert it into the buffer array */
 438                         buf->page = page;
 439                         buf->ops = &anon_pipe_buf_ops;
 440                         buf->offset = 0;
 441                         buf->len = copied;
 442                         buf->flags = 0;
 443                         if (is_packetized(filp)) {
 444                                 buf->ops = &packet_pipe_buf_ops;
 445                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 446                         }
 447                         pipe->nrbufs = ++bufs;
 448                         pipe->tmp_page = NULL;
 449
 450                         if (!iov_iter_count(from))
 451                                 break;
 452                 }
 453                 if (bufs < pipe->buffers)
 454                         continue;
 455                 if (filp->f_flags & O_NONBLOCK) {
 456                         if (!ret)
 457                                 ret = -EAGAIN;
 458                         break;
 459                 }
 460                 if (signal_pending(current)) {
 461                         if (!ret)
 462                                 ret = -ERESTARTSYS;
 463                         break;
 464                 }
 465                 if (do_wakeup) {
 466                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 467                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 468                         do_wakeup = 0;
 469                 }
 470                 pipe->waiting_writers++;
 471                 pipe_wait(pipe);
 472                 pipe->waiting_writers--;
 473         }
 474 out:
 475         __pipe_unlock(pipe);
 476         if (do_wakeup) {
 477                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 478                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 479         }
 480         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 481                 int err = file_update_time(filp);
 482                 if (err)
 483                         ret = err;
 484                 sb_end_write(file_inode(filp)->i_sb);
 485         }
 486         return ret;
 487 }
 488
 489 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 490 {
 491         struct pipe_inode_info *pipe = filp->private_data;
 492         int count, buf, nrbufs;
 493
 494         switch (cmd) {
 495                 case FIONREAD:
 496                         __pipe_lock(pipe);
 497                         count = 0;
 498                         buf = pipe->curbuf;
 499                         nrbufs = pipe->nrbufs;
 500                         while (--nrbufs >= 0) {
 501                                 count += pipe->bufs[buf].len;
 502                                 buf = (buf+1) & (pipe->buffers - 1);
 503                         }
 504                         __pipe_unlock(pipe);
 505
 506                         return put_user(count, (int __user *)arg);
 507                 default:
 508                         return -ENOIOCTLCMD;
 509         }
 510 }
 511
 512 /* No kernel lock held - fine */
 513 static __poll_t
 514 pipe_poll(struct file *filp, poll_table *wait)
 515 {
 516         __poll_t mask;
 517         struct pipe_inode_info *pipe = filp->private_data;
 518         int nrbufs;
 519
 520         poll_wait(filp, &pipe->wait, wait);
 521
 522         /* Reading only -- no need for acquiring the semaphore.  */
 523         nrbufs = pipe->nrbufs;
 524         mask = 0;
 525         if (filp->f_mode & FMODE_READ) {
 526                 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
 527                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 528                         mask |= EPOLLHUP;
 529         }
 530
 531         if (filp->f_mode & FMODE_WRITE) {
 532                 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
 533                 /*
 534                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 535                  * behave exactly like pipes for poll().
 536                  */
 537                 if (!pipe->readers)
 538                         mask |= EPOLLERR;
 539         }
 540
 541         return mask;
 542 }
 543
 544 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 545 {
 546         int kill = 0;
 547
 548         spin_lock(&inode->i_lock);
 549         if (!--pipe->files) {
 550                 inode->i_pipe = NULL;
 551                 kill = 1;
 552         }
 553         spin_unlock(&inode->i_lock);
 554
 555         if (kill)
 556                 free_pipe_info(pipe);
 557 }
 558
 559 static int
 560 pipe_release(struct inode *inode, struct file *file)
 561 {
 562         struct pipe_inode_info *pipe = file->private_data;
 563
 564         __pipe_lock(pipe);
 565         if (file->f_mode & FMODE_READ)
 566                 pipe->readers--;
 567         if (file->f_mode & FMODE_WRITE)
 568                 pipe->writers--;
 569
 570         if (pipe->readers || pipe->writers) {
 571                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 572                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 573                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 574         }
 575         __pipe_unlock(pipe);
 576
 577         put_pipe_info(inode, pipe);
 578         return 0;
 579 }
 580
 581 static int
 582 pipe_fasync(int fd, struct file *filp, int on)
 583 {
 584         struct pipe_inode_info *pipe = filp->private_data;
 585         int retval = 0;
 586
 587         __pipe_lock(pipe);
 588         if (filp->f_mode & FMODE_READ)
 589                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 590         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 591                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 592                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 593                         /* this can happen only if on == T */
 594                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 595         }
 596         __pipe_unlock(pipe);
 597         return retval;
 598 }
 599
 600 static unsigned long account_pipe_buffers(struct user_struct *user,
 601                                  unsigned long old, unsigned long new)
 602 {
 603         return atomic_long_add_return(new - old, &user->pipe_bufs);
 604 }
 605
 606 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 607 {
 608         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 609
 610         return soft_limit && user_bufs > soft_limit;
 611 }
 612
 613 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 614 {
 615         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 616
 617         return hard_limit && user_bufs > hard_limit;
 618 }
 619
 620 static bool is_unprivileged_user(void)
 621 {
 622         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 623 }
 624
 625 struct pipe_inode_info *alloc_pipe_info(void)
 626 {
 627         struct pipe_inode_info *pipe;
 628         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 629         struct user_struct *user = get_current_user();
 630         unsigned long user_bufs;
 631         unsigned int max_size = READ_ONCE(pipe_max_size);
 632
 633         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 634         if (pipe == NULL)
 635                 goto out_free_uid;
 636
 637         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 638                 pipe_bufs = max_size >> PAGE_SHIFT;
 639
 640         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 641
 642         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 643                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 644                 pipe_bufs = 1;
 645         }
 646
 647         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 648                 goto out_revert_acct;
 649
 650         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 651                              GFP_KERNEL_ACCOUNT);
 652
 653         if (pipe->bufs) {
 654                 init_waitqueue_head(&pipe->wait);
 655                 pipe->r_counter = pipe->w_counter = 1;
 656                 pipe->buffers = pipe_bufs;
 657                 pipe->user = user;
 658                 mutex_init(&pipe->mutex);
 659                 return pipe;
 660         }
 661
 662 out_revert_acct:
 663         (void) account_pipe_buffers(user, pipe_bufs, 0);
 664         kfree(pipe);
 665 out_free_uid:
 666         free_uid(user);
 667         return NULL;
 668 }
 669
 670 void free_pipe_info(struct pipe_inode_info *pipe)
 671 {
 672         int i;
 673
 674         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 675         free_uid(pipe->user);
 676         for (i = 0; i < pipe->buffers; i++) {
 677                 struct pipe_buffer *buf = pipe->bufs + i;
 678                 if (buf->ops)
 679                         pipe_buf_release(pipe, buf);
 680         }
 681         if (pipe->tmp_page)
 682                 __free_page(pipe->tmp_page);
 683         kfree(pipe->bufs);
 684         kfree(pipe);
 685 }
 686
 687 static struct vfsmount *pipe_mnt __read_mostly;
 688
 689 /*
 690  * pipefs_dname() is called from d_path().
 691  */
 692 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 693 {
 694         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 695                                 d_inode(dentry)->i_ino);
 696 }
 697
 698 static const struct dentry_operations pipefs_dentry_operations = {
 699         .d_dname        = pipefs_dname,
 700 };
 701
 702 static struct inode * get_pipe_inode(void)
 703 {
 704         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 705         struct pipe_inode_info *pipe;
 706
 707         if (!inode)
 708                 goto fail_inode;
 709
 710         inode->i_ino = get_next_ino();
 711
 712         pipe = alloc_pipe_info();
 713         if (!pipe)
 714                 goto fail_iput;
 715
 716         inode->i_pipe = pipe;
 717         pipe->files = 2;
 718         pipe->readers = pipe->writers = 1;
 719         inode->i_fop = &pipefifo_fops;
 720
 721         /*
 722          * Mark the inode dirty from the very beginning,
 723          * that way it will never be moved to the dirty
 724          * list because "mark_inode_dirty()" will think
 725          * that it already _is_ on the dirty list.
 726          */
 727         inode->i_state = I_DIRTY;
 728         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 729         inode->i_uid = current_fsuid();
 730         inode->i_gid = current_fsgid();
 731         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 732
 733         return inode;
 734
 735 fail_iput:
 736         iput(inode);
 737
 738 fail_inode:
 739         return NULL;
 740 }
 741
 742 int create_pipe_files(struct file **res, int flags)
 743 {
 744         int err;
 745         struct inode *inode = get_pipe_inode();
 746         struct file *f;
 747         struct path path;
 748
 749         if (!inode)
 750                 return -ENFILE;
 751
 752         err = -ENOMEM;
 753         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &empty_name);
 754         if (!path.dentry)
 755                 goto err_inode;
 756         path.mnt = mntget(pipe_mnt);
 757
 758         d_instantiate(path.dentry, inode);
 759
 760         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 761         if (IS_ERR(f)) {
 762                 err = PTR_ERR(f);
 763                 goto err_dentry;
 764         }
 765
 766         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 767         f->private_data = inode->i_pipe;
 768
 769         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 770         if (IS_ERR(res[0])) {
 771                 err = PTR_ERR(res[0]);
 772                 goto err_file;
 773         }
 774
 775         path_get(&path);
 776         res[0]->private_data = inode->i_pipe;
 777         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 778         res[1] = f;
 779         return 0;
 780
 781 err_file:
 782         put_filp(f);
 783 err_dentry:
 784         free_pipe_info(inode->i_pipe);
 785         path_put(&path);
 786         return err;
 787
 788 err_inode:
 789         free_pipe_info(inode->i_pipe);
 790         iput(inode);
 791         return err;
 792 }
 793
 794 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 795 {
 796         int error;
 797         int fdw, fdr;
 798
 799         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 800                 return -EINVAL;
 801
 802         error = create_pipe_files(files, flags);
 803         if (error)
 804                 return error;
 805
 806         error = get_unused_fd_flags(flags);
 807         if (error < 0)
 808                 goto err_read_pipe;
 809         fdr = error;
 810
 811         error = get_unused_fd_flags(flags);
 812         if (error < 0)
 813                 goto err_fdr;
 814         fdw = error;
 815
 816         audit_fd_pair(fdr, fdw);
 817         fd[0] = fdr;
 818         fd[1] = fdw;
 819         return 0;
 820
 821  err_fdr:
 822         put_unused_fd(fdr);
 823  err_read_pipe:
 824         fput(files[0]);
 825         fput(files[1]);
 826         return error;
 827 }
 828
 829 int do_pipe_flags(int *fd, int flags)
 830 {
 831         struct file *files[2];
 832         int error = __do_pipe_flags(fd, files, flags);
 833         if (!error) {
 834                 fd_install(fd[0], files[0]);
 835                 fd_install(fd[1], files[1]);
 836         }
 837         return error;
 838 }
 839
 840 /*
 841  * sys_pipe() is the normal C calling standard for creating
 842  * a pipe. It's not the way Unix traditionally does this, though.
 843  */
 844 static int do_pipe2(int __user *fildes, int flags)
 845 {
 846         struct file *files[2];
 847         int fd[2];
 848         int error;
 849
 850         error = __do_pipe_flags(fd, files, flags);
 851         if (!error) {
 852                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 853                         fput(files[0]);
 854                         fput(files[1]);
 855                         put_unused_fd(fd[0]);
 856                         put_unused_fd(fd[1]);
 857                         error = -EFAULT;
 858                 } else {
 859                         fd_install(fd[0], files[0]);
 860                         fd_install(fd[1], files[1]);
 861                 }
 862         }
 863         return error;
 864 }
 865
 866 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 867 {
 868         return do_pipe2(fildes, flags);
 869 }
 870
 871 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 872 {
 873         return do_pipe2(fildes, 0);
 874 }
 875
 876 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 877 {
 878         int cur = *cnt;
 879
 880         while (cur == *cnt) {
 881                 pipe_wait(pipe);
 882                 if (signal_pending(current))
 883                         break;
 884         }
 885         return cur == *cnt ? -ERESTARTSYS : 0;
 886 }
 887
 888 static void wake_up_partner(struct pipe_inode_info *pipe)
 889 {
 890         wake_up_interruptible(&pipe->wait);
 891 }
 892
 893 static int fifo_open(struct inode *inode, struct file *filp)
 894 {
 895         struct pipe_inode_info *pipe;
 896         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 897         int ret;
 898
 899         filp->f_version = 0;
 900
 901         spin_lock(&inode->i_lock);
 902         if (inode->i_pipe) {
 903                 pipe = inode->i_pipe;
 904                 pipe->files++;
 905                 spin_unlock(&inode->i_lock);
 906         } else {
 907                 spin_unlock(&inode->i_lock);
 908                 pipe = alloc_pipe_info();
 909                 if (!pipe)
 910                         return -ENOMEM;
 911                 pipe->files = 1;
 912                 spin_lock(&inode->i_lock);
 913                 if (unlikely(inode->i_pipe)) {
 914                         inode->i_pipe->files++;
 915                         spin_unlock(&inode->i_lock);
 916                         free_pipe_info(pipe);
 917                         pipe = inode->i_pipe;
 918                 } else {
 919                         inode->i_pipe = pipe;
 920                         spin_unlock(&inode->i_lock);
 921                 }
 922         }
 923         filp->private_data = pipe;
 924         /* OK, we have a pipe and it's pinned down */
 925
 926         __pipe_lock(pipe);
 927
 928         /* We can only do regular read/write on fifos */
 929         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 930
 931         switch (filp->f_mode) {
 932         case FMODE_READ:
 933         /*
 934          *  O_RDONLY
 935          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 936          *  opened, even when there is no process writing the FIFO.
 937          */
 938                 pipe->r_counter++;
 939                 if (pipe->readers++ == 0)
 940                         wake_up_partner(pipe);
 941
 942                 if (!is_pipe && !pipe->writers) {
 943                         if ((filp->f_flags & O_NONBLOCK)) {
 944                                 /* suppress EPOLLHUP until we have
 945                                  * seen a writer */
 946                                 filp->f_version = pipe->w_counter;
 947                         } else {
 948                                 if (wait_for_partner(pipe, &pipe->w_counter))
 949                                         goto err_rd;
 950                         }
 951                 }
 952                 break;
 953
 954         case FMODE_WRITE:
 955         /*
 956          *  O_WRONLY
 957          *  POSIX.1 says that O_NONBLOCK means return -1 with
 958          *  errno=ENXIO when there is no process reading the FIFO.
 959          */
 960                 ret = -ENXIO;
 961                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 962                         goto err;
 963
 964                 pipe->w_counter++;
 965                 if (!pipe->writers++)
 966                         wake_up_partner(pipe);
 967
 968                 if (!is_pipe && !pipe->readers) {
 969                         if (wait_for_partner(pipe, &pipe->r_counter))
 970                                 goto err_wr;
 971                 }
 972                 break;
 973
 974         case FMODE_READ | FMODE_WRITE:
 975         /*
 976          *  O_RDWR
 977          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 978          *  This implementation will NEVER block on a O_RDWR open, since
 979          *  the process can at least talk to itself.
 980          */
 981
 982                 pipe->readers++;
 983                 pipe->writers++;
 984                 pipe->r_counter++;
 985                 pipe->w_counter++;
 986                 if (pipe->readers == 1 || pipe->writers == 1)
 987                         wake_up_partner(pipe);
 988                 break;
 989
 990         default:
 991                 ret = -EINVAL;
 992                 goto err;
 993         }
 994
 995         /* Ok! */
 996         __pipe_unlock(pipe);
 997         return 0;
 998
 999 err_rd:
1000         if (!--pipe->readers)
1001                 wake_up_interruptible(&pipe->wait);
1002         ret = -ERESTARTSYS;
1003         goto err;
1004
1005 err_wr:
1006         if (!--pipe->writers)
1007                 wake_up_interruptible(&pipe->wait);
1008         ret = -ERESTARTSYS;
1009         goto err;
1010
1011 err:
1012         __pipe_unlock(pipe);
1013
1014         put_pipe_info(inode, pipe);
1015         return ret;
1016 }
1017
1018 const struct file_operations pipefifo_fops = {
1019         .open           = fifo_open,
1020         .llseek         = no_llseek,
1021         .read_iter      = pipe_read,
1022         .write_iter     = pipe_write,
1023         .poll           = pipe_poll,
1024         .unlocked_ioctl = pipe_ioctl,
1025         .release        = pipe_release,
1026         .fasync         = pipe_fasync,
1027 };
1028
1029 /*
1030  * Currently we rely on the pipe array holding a power-of-2 number
1031  * of pages. Returns 0 on error.
1032  */
1033 unsigned int round_pipe_size(unsigned long size)
1034 {
1035         if (size > (1U << 31))
1036                 return 0;
1037
1038         /* Minimum pipe size, as required by POSIX */
1039         if (size < PAGE_SIZE)
1040                 return PAGE_SIZE;
1041
1042         return roundup_pow_of_two(size);
1043 }
1044
1045 /*
1046  * Allocate a new array of pipe buffers and copy the info over. Returns the
1047  * pipe size if successful, or return -ERROR on error.
1048  */
1049 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1050 {
1051         struct pipe_buffer *bufs;
1052         unsigned int size, nr_pages;
1053         unsigned long user_bufs;
1054         long ret = 0;
1055
1056         size = round_pipe_size(arg);
1057         nr_pages = size >> PAGE_SHIFT;
1058
1059         if (!nr_pages)
1060                 return -EINVAL;
1061
1062         /*
1063          * If trying to increase the pipe capacity, check that an
1064          * unprivileged user is not trying to exceed various limits
1065          * (soft limit check here, hard limit check just below).
1066          * Decreasing the pipe capacity is always permitted, even
1067          * if the user is currently over a limit.
1068          */
1069         if (nr_pages > pipe->buffers &&
1070                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1071                 return -EPERM;
1072
1073         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1074
1075         if (nr_pages > pipe->buffers &&
1076                         (too_many_pipe_buffers_hard(user_bufs) ||
1077                          too_many_pipe_buffers_soft(user_bufs)) &&
1078                         is_unprivileged_user()) {
1079                 ret = -EPERM;
1080                 goto out_revert_acct;
1081         }
1082
1083         /*
1084          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1085          * expect a lot of shrink+grow operations, just free and allocate
1086          * again like we would do for growing. If the pipe currently
1087          * contains more buffers than arg, then return busy.
1088          */
1089         if (nr_pages < pipe->nrbufs) {
1090                 ret = -EBUSY;
1091                 goto out_revert_acct;
1092         }
1093
1094         bufs = kcalloc(nr_pages, sizeof(*bufs),
1095                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1096         if (unlikely(!bufs)) {
1097                 ret = -ENOMEM;
1098                 goto out_revert_acct;
1099         }
1100
1101         /*
1102          * The pipe array wraps around, so just start the new one at zero
1103          * and adjust the indexes.
1104          */
1105         if (pipe->nrbufs) {
1106                 unsigned int tail;
1107                 unsigned int head;
1108
1109                 tail = pipe->curbuf + pipe->nrbufs;
1110                 if (tail < pipe->buffers)
1111                         tail = 0;
1112                 else
1113                         tail &= (pipe->buffers - 1);
1114
1115                 head = pipe->nrbufs - tail;
1116                 if (head)
1117                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1118                 if (tail)
1119                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1120         }
1121
1122         pipe->curbuf = 0;
1123         kfree(pipe->bufs);
1124         pipe->bufs = bufs;
1125         pipe->buffers = nr_pages;
1126         return nr_pages * PAGE_SIZE;
1127
1128 out_revert_acct:
1129         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1130         return ret;
1131 }
1132
1133 /*
1134  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1135  * location, so checking ->i_pipe is not enough to verify that this is a
1136  * pipe.
1137  */
1138 struct pipe_inode_info *get_pipe_info(struct file *file)
1139 {
1140         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1141 }
1142
1143 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1144 {
1145         struct pipe_inode_info *pipe;
1146         long ret;
1147
1148         pipe = get_pipe_info(file);
1149         if (!pipe)
1150                 return -EBADF;
1151
1152         __pipe_lock(pipe);
1153
1154         switch (cmd) {
1155         case F_SETPIPE_SZ:
1156                 ret = pipe_set_size(pipe, arg);
1157                 break;
1158         case F_GETPIPE_SZ:
1159                 ret = pipe->buffers * PAGE_SIZE;
1160                 break;
1161         default:
1162                 ret = -EINVAL;
1163                 break;
1164         }
1165
1166         __pipe_unlock(pipe);
1167         return ret;
1168 }
1169
1170 static const struct super_operations pipefs_ops = {
1171         .destroy_inode = free_inode_nonrcu,
1172         .statfs = simple_statfs,
1173 };
1174
1175 /*
1176  * pipefs should _never_ be mounted by userland - too much of security hassle,
1177  * no real gain from having the whole whorehouse mounted. So we don't need
1178  * any operations on the root directory. However, we need a non-trivial
1179  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1180  */
1181 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1182                          int flags, const char *dev_name, void *data)
1183 {
1184         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1185                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1186 }
1187
1188 static struct file_system_type pipe_fs_type = {
1189         .name           = "pipefs",
1190         .mount          = pipefs_mount,
1191         .kill_sb        = kill_anon_super,
1192 };
1193
1194 static int __init init_pipe_fs(void)
1195 {
1196         int err = register_filesystem(&pipe_fs_type);
1197
1198         if (!err) {
1199                 pipe_mnt = kern_mount(&pipe_fs_type);
1200                 if (IS_ERR(pipe_mnt)) {
1201                         err = PTR_ERR(pipe_mnt);
1202                         unregister_filesystem(&pipe_fs_type);
1203                 }
1204         }
1205         return err;
1206 }
1207
1208 fs_initcall(init_pipe_fs);