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