fs/pipe.c

   1 /*
   2  *  linux/fs/pipe.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
   5  */
   6
   7 #include <linux/mm.h>
   8 #include <linux/file.h>
   9 #include <linux/poll.h>
  10 #include <linux/slab.h>
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/fs.h>
  14 #include <linux/log2.h>
  15 #include <linux/mount.h>
  16 #include <linux/magic.h>
  17 #include <linux/pipe_fs_i.h>
  18 #include <linux/uio.h>
  19 #include <linux/highmem.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/audit.h>
  22 #include <linux/syscalls.h>
  23 #include <linux/fcntl.h>
  24
  25 #include <asm/uaccess.h>
  26 #include <asm/ioctls.h>
  27
  28 #include "internal.h"
  29
  30 /*
  31  * New pipe buffers will be restricted to this size while the user is exceeding
  32  * their pipe buffer quota. The general pipe use case needs at least two
  33  * buffers: one for data yet to be read, and one for new data. If this is less
  34  * than two, then a write to a non-empty pipe may block even if the pipe is not
  35  * full. This can occur with GNU make jobserver or similar uses of pipes as
  36  * semaphores: multiple processes may be waiting to write tokens back to the
  37  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  38  *
  39  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  40  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  41  * emptied.
  42  */
  43 #define PIPE_MIN_DEF_BUFFERS 2
  44
  45 /*
  46  * The max size that a non-root user is allowed to grow the pipe. Can
  47  * be set by root in /proc/sys/fs/pipe-max-size
  48  */
  49 unsigned int pipe_max_size = 1048576;
  50
  51 /*
  52  * Minimum pipe size, as required by POSIX
  53  */
  54 unsigned int pipe_min_size = PAGE_SIZE;
  55
  56 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  57  * matches default values.
  58  */
  59 unsigned long pipe_user_pages_hard;
  60 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  61
  62 /*
  63  * We use a start+len construction, which provides full use of the
  64  * allocated memory.
  65  * -- Florian Coosmann (FGC)
  66  *
  67  * Reads with count = 0 should always return 0.
  68  * -- Julian Bradfield 1999-06-07.
  69  *
  70  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  71  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  72  *
  73  * pipe_read & write cleanup
  74  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  75  */
  76
  77 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  78 {
  79         if (pipe->files)
  80                 mutex_lock_nested(&pipe->mutex, subclass);
  81 }
  82
  83 void pipe_lock(struct pipe_inode_info *pipe)
  84 {
  85         /*
  86          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  87          */
  88         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  89 }
  90 EXPORT_SYMBOL(pipe_lock);
  91
  92 void pipe_unlock(struct pipe_inode_info *pipe)
  93 {
  94         if (pipe->files)
  95                 mutex_unlock(&pipe->mutex);
  96 }
  97 EXPORT_SYMBOL(pipe_unlock);
  98
  99 static inline void __pipe_lock(struct pipe_inode_info *pipe)
 100 {
 101         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 102 }
 103
 104 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 105 {
 106         mutex_unlock(&pipe->mutex);
 107 }
 108
 109 void pipe_double_lock(struct pipe_inode_info *pipe1,
 110                       struct pipe_inode_info *pipe2)
 111 {
 112         BUG_ON(pipe1 == pipe2);
 113
 114         if (pipe1 < pipe2) {
 115                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 116                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 117         } else {
 118                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 119                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 120         }
 121 }
 122
 123 /* Drop the inode semaphore and wait for a pipe event, atomically */
 124 void pipe_wait(struct pipe_inode_info *pipe)
 125 {
 126         DEFINE_WAIT(wait);
 127
 128         /*
 129          * Pipes are system-local resources, so sleeping on them
 130          * is considered a noninteractive wait:
 131          */
 132         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 133         pipe_unlock(pipe);
 134         schedule();
 135         finish_wait(&pipe->wait, &wait);
 136         pipe_lock(pipe);
 137 }
 138
 139 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 140                                   struct pipe_buffer *buf)
 141 {
 142         struct page *page = buf->page;
 143
 144         /*
 145          * If nobody else uses this page, and we don't already have a
 146          * temporary page, let's keep track of it as a one-deep
 147          * allocation cache. (Otherwise just release our reference to it)
 148          */
 149         if (page_count(page) == 1 && !pipe->tmp_page)
 150                 pipe->tmp_page = page;
 151         else
 152                 page_cache_release(page);
 153 }
 154
 155 /**
 156  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 157  * @pipe:       the pipe that the buffer belongs to
 158  * @buf:        the buffer to attempt to steal
 159  *
 160  * Description:
 161  *      This function attempts to steal the &struct page attached to
 162  *      @buf. If successful, this function returns 0 and returns with
 163  *      the page locked. The caller may then reuse the page for whatever
 164  *      he wishes; the typical use is insertion into a different file
 165  *      page cache.
 166  */
 167 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 168                            struct pipe_buffer *buf)
 169 {
 170         struct page *page = buf->page;
 171
 172         /*
 173          * A reference of one is golden, that means that the owner of this
 174          * page is the only one holding a reference to it. lock the page
 175          * and return OK.
 176          */
 177         if (page_count(page) == 1) {
 178                 lock_page(page);
 179                 return 0;
 180         }
 181
 182         return 1;
 183 }
 184 EXPORT_SYMBOL(generic_pipe_buf_steal);
 185
 186 /**
 187  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 188  * @pipe:       the pipe that the buffer belongs to
 189  * @buf:        the buffer to get a reference to
 190  *
 191  * Description:
 192  *      This function grabs an extra reference to @buf. It's used in
 193  *      in the tee() system call, when we duplicate the buffers in one
 194  *      pipe into another.
 195  */
 196 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 197 {
 198         return try_get_page(buf->page);
 199 }
 200 EXPORT_SYMBOL(generic_pipe_buf_get);
 201
 202 /**
 203  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 204  * @info:       the pipe that the buffer belongs to
 205  * @buf:        the buffer to confirm
 206  *
 207  * Description:
 208  *      This function does nothing, because the generic pipe code uses
 209  *      pages that are always good when inserted into the pipe.
 210  */
 211 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 212                              struct pipe_buffer *buf)
 213 {
 214         return 0;
 215 }
 216 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 217
 218 /**
 219  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 220  * @pipe:       the pipe that the buffer belongs to
 221  * @buf:        the buffer to put a reference to
 222  *
 223  * Description:
 224  *      This function releases a reference to @buf.
 225  */
 226 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 227                               struct pipe_buffer *buf)
 228 {
 229         page_cache_release(buf->page);
 230 }
 231 EXPORT_SYMBOL(generic_pipe_buf_release);
 232
 233 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 234         .can_merge = 1,
 235         .confirm = generic_pipe_buf_confirm,
 236         .release = anon_pipe_buf_release,
 237         .steal = generic_pipe_buf_steal,
 238         .get = generic_pipe_buf_get,
 239 };
 240
 241 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 242         .can_merge = 0,
 243         .confirm = generic_pipe_buf_confirm,
 244         .release = anon_pipe_buf_release,
 245         .steal = generic_pipe_buf_steal,
 246         .get = generic_pipe_buf_get,
 247 };
 248
 249 static ssize_t
 250 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 251 {
 252         size_t total_len = iov_iter_count(to);
 253         struct file *filp = iocb->ki_filp;
 254         struct pipe_inode_info *pipe = filp->private_data;
 255         int do_wakeup;
 256         ssize_t ret;
 257
 258         /* Null read succeeds. */
 259         if (unlikely(total_len == 0))
 260                 return 0;
 261
 262         do_wakeup = 0;
 263         ret = 0;
 264         __pipe_lock(pipe);
 265         for (;;) {
 266                 int bufs = pipe->nrbufs;
 267                 if (bufs) {
 268                         int curbuf = pipe->curbuf;
 269                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 270                         const struct pipe_buf_operations *ops = buf->ops;
 271                         size_t chars = buf->len;
 272                         size_t written;
 273                         int error;
 274
 275                         if (chars > total_len)
 276                                 chars = total_len;
 277
 278                         error = ops->confirm(pipe, buf);
 279                         if (error) {
 280                                 if (!ret)
 281                                         ret = error;
 282                                 break;
 283                         }
 284
 285                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 286                         if (unlikely(written < chars)) {
 287                                 if (!ret)
 288                                         ret = -EFAULT;
 289                                 break;
 290                         }
 291                         ret += chars;
 292                         buf->offset += chars;
 293                         buf->len -= chars;
 294
 295                         /* Was it a packet buffer? Clean up and exit */
 296                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 297                                 total_len = chars;
 298                                 buf->len = 0;
 299                         }
 300
 301                         if (!buf->len) {
 302                                 buf->ops = NULL;
 303                                 ops->release(pipe, buf);
 304                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 305                                 pipe->curbuf = curbuf;
 306                                 pipe->nrbufs = --bufs;
 307                                 do_wakeup = 1;
 308                         }
 309                         total_len -= chars;
 310                         if (!total_len)
 311                                 break;  /* common path: read succeeded */
 312                 }
 313                 if (bufs)       /* More to do? */
 314                         continue;
 315                 if (!pipe->writers)
 316                         break;
 317                 if (!pipe->waiting_writers) {
 318                         /* syscall merging: Usually we must not sleep
 319                          * if O_NONBLOCK is set, or if we got some data.
 320                          * But if a writer sleeps in kernel space, then
 321                          * we can wait for that data without violating POSIX.
 322                          */
 323                         if (ret)
 324                                 break;
 325                         if (filp->f_flags & O_NONBLOCK) {
 326                                 ret = -EAGAIN;
 327                                 break;
 328                         }
 329                 }
 330                 if (signal_pending(current)) {
 331                         if (!ret)
 332                                 ret = -ERESTARTSYS;
 333                         break;
 334                 }
 335                 if (do_wakeup) {
 336                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 337                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 338                 }
 339                 pipe_wait(pipe);
 340         }
 341         __pipe_unlock(pipe);
 342
 343         /* Signal writers asynchronously that there is more room. */
 344         if (do_wakeup) {
 345                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 346                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 347         }
 348         if (ret > 0)
 349                 file_accessed(filp);
 350         return ret;
 351 }
 352
 353 static inline int is_packetized(struct file *file)
 354 {
 355         return (file->f_flags & O_DIRECT) != 0;
 356 }
 357
 358 static ssize_t
 359 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 360 {
 361         struct file *filp = iocb->ki_filp;
 362         struct pipe_inode_info *pipe = filp->private_data;
 363         ssize_t ret = 0;
 364         int do_wakeup = 0;
 365         size_t total_len = iov_iter_count(from);
 366         ssize_t chars;
 367
 368         /* Null write succeeds. */
 369         if (unlikely(total_len == 0))
 370                 return 0;
 371
 372         __pipe_lock(pipe);
 373
 374         if (!pipe->readers) {
 375                 send_sig(SIGPIPE, current, 0);
 376                 ret = -EPIPE;
 377                 goto out;
 378         }
 379
 380         /* We try to merge small writes */
 381         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 382         if (pipe->nrbufs && chars != 0) {
 383                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 384                                                         (pipe->buffers - 1);
 385                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 386                 const struct pipe_buf_operations *ops = buf->ops;
 387                 int offset = buf->offset + buf->len;
 388
 389                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 390                         ret = ops->confirm(pipe, buf);
 391                         if (ret)
 392                                 goto out;
 393
 394                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 395                         if (unlikely(ret < chars)) {
 396                                 ret = -EFAULT;
 397                                 goto out;
 398                         }
 399                         do_wakeup = 1;
 400                         buf->len += ret;
 401                         if (!iov_iter_count(from))
 402                                 goto out;
 403                 }
 404         }
 405
 406         for (;;) {
 407                 int bufs;
 408
 409                 if (!pipe->readers) {
 410                         send_sig(SIGPIPE, current, 0);
 411                         if (!ret)
 412                                 ret = -EPIPE;
 413                         break;
 414                 }
 415                 bufs = pipe->nrbufs;
 416                 if (bufs < pipe->buffers) {
 417                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 418                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 419                         struct page *page = pipe->tmp_page;
 420                         int copied;
 421
 422                         if (!page) {
 423                                 page = alloc_page(GFP_HIGHUSER);
 424                                 if (unlikely(!page)) {
 425                                         ret = ret ? : -ENOMEM;
 426                                         break;
 427                                 }
 428                                 pipe->tmp_page = page;
 429                         }
 430                         /* Always wake up, even if the copy fails. Otherwise
 431                          * we lock up (O_NONBLOCK-)readers that sleep due to
 432                          * syscall merging.
 433                          * FIXME! Is this really true?
 434                          */
 435                         do_wakeup = 1;
 436                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 437                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 438                                 if (!ret)
 439                                         ret = -EFAULT;
 440                                 break;
 441                         }
 442                         ret += copied;
 443
 444                         /* Insert it into the buffer array */
 445                         buf->page = page;
 446                         buf->ops = &anon_pipe_buf_ops;
 447                         buf->offset = 0;
 448                         buf->len = copied;
 449                         buf->flags = 0;
 450                         if (is_packetized(filp)) {
 451                                 buf->ops = &packet_pipe_buf_ops;
 452                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 453                         }
 454                         pipe->nrbufs = ++bufs;
 455                         pipe->tmp_page = NULL;
 456
 457                         if (!iov_iter_count(from))
 458                                 break;
 459                 }
 460                 if (bufs < pipe->buffers)
 461                         continue;
 462                 if (filp->f_flags & O_NONBLOCK) {
 463                         if (!ret)
 464                                 ret = -EAGAIN;
 465                         break;
 466                 }
 467                 if (signal_pending(current)) {
 468                         if (!ret)
 469                                 ret = -ERESTARTSYS;
 470                         break;
 471                 }
 472                 if (do_wakeup) {
 473                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 474                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 475                         do_wakeup = 0;
 476                 }
 477                 pipe->waiting_writers++;
 478                 pipe_wait(pipe);
 479                 pipe->waiting_writers--;
 480         }
 481 out:
 482         __pipe_unlock(pipe);
 483         if (do_wakeup) {
 484                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 485                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 486         }
 487         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 488                 int err = file_update_time(filp);
 489                 if (err)
 490                         ret = err;
 491                 sb_end_write(file_inode(filp)->i_sb);
 492         }
 493         return ret;
 494 }
 495
 496 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 497 {
 498         struct pipe_inode_info *pipe = filp->private_data;
 499         int count, buf, nrbufs;
 500
 501         switch (cmd) {
 502                 case FIONREAD:
 503                         __pipe_lock(pipe);
 504                         count = 0;
 505                         buf = pipe->curbuf;
 506                         nrbufs = pipe->nrbufs;
 507                         while (--nrbufs >= 0) {
 508                                 count += pipe->bufs[buf].len;
 509                                 buf = (buf+1) & (pipe->buffers - 1);
 510                         }
 511                         __pipe_unlock(pipe);
 512
 513                         return put_user(count, (int __user *)arg);
 514                 default:
 515                         return -ENOIOCTLCMD;
 516         }
 517 }
 518
 519 /* No kernel lock held - fine */
 520 static unsigned int
 521 pipe_poll(struct file *filp, poll_table *wait)
 522 {
 523         unsigned int mask;
 524         struct pipe_inode_info *pipe = filp->private_data;
 525         int nrbufs;
 526
 527         poll_wait(filp, &pipe->wait, wait);
 528
 529         /* Reading only -- no need for acquiring the semaphore.  */
 530         nrbufs = pipe->nrbufs;
 531         mask = 0;
 532         if (filp->f_mode & FMODE_READ) {
 533                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 534                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 535                         mask |= POLLHUP;
 536         }
 537
 538         if (filp->f_mode & FMODE_WRITE) {
 539                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 540                 /*
 541                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 542                  * behave exactly like pipes for poll().
 543                  */
 544                 if (!pipe->readers)
 545                         mask |= POLLERR;
 546         }
 547
 548         return mask;
 549 }
 550
 551 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 552 {
 553         int kill = 0;
 554
 555         spin_lock(&inode->i_lock);
 556         if (!--pipe->files) {
 557                 inode->i_pipe = NULL;
 558                 kill = 1;
 559         }
 560         spin_unlock(&inode->i_lock);
 561
 562         if (kill)
 563                 free_pipe_info(pipe);
 564 }
 565
 566 static int
 567 pipe_release(struct inode *inode, struct file *file)
 568 {
 569         struct pipe_inode_info *pipe = file->private_data;
 570
 571         __pipe_lock(pipe);
 572         if (file->f_mode & FMODE_READ)
 573                 pipe->readers--;
 574         if (file->f_mode & FMODE_WRITE)
 575                 pipe->writers--;
 576
 577         if (pipe->readers || pipe->writers) {
 578                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 579                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 580                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 581         }
 582         __pipe_unlock(pipe);
 583
 584         put_pipe_info(inode, pipe);
 585         return 0;
 586 }
 587
 588 static int
 589 pipe_fasync(int fd, struct file *filp, int on)
 590 {
 591         struct pipe_inode_info *pipe = filp->private_data;
 592         int retval = 0;
 593
 594         __pipe_lock(pipe);
 595         if (filp->f_mode & FMODE_READ)
 596                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 597         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 598                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 599                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 600                         /* this can happen only if on == T */
 601                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 602         }
 603         __pipe_unlock(pipe);
 604         return retval;
 605 }
 606
 607 static void account_pipe_buffers(struct pipe_inode_info *pipe,
 608                                  unsigned long old, unsigned long new)
 609 {
 610         atomic_long_add(new - old, &pipe->user->pipe_bufs);
 611 }
 612
 613 static bool too_many_pipe_buffers_soft(struct user_struct *user)
 614 {
 615         return pipe_user_pages_soft &&
 616                atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_soft;
 617 }
 618
 619 static bool too_many_pipe_buffers_hard(struct user_struct *user)
 620 {
 621         return pipe_user_pages_hard &&
 622                atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_hard;
 623 }
 624
 625 struct pipe_inode_info *alloc_pipe_info(void)
 626 {
 627         struct pipe_inode_info *pipe;
 628
 629         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 630         if (pipe) {
 631                 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 632                 struct user_struct *user = get_current_user();
 633
 634                 if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 635                         pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 636
 637                 if (!too_many_pipe_buffers_hard(user)) {
 638                         if (too_many_pipe_buffers_soft(user))
 639                                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 640                         pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * pipe_bufs, GFP_KERNEL);
 641                 }
 642
 643                 if (pipe->bufs) {
 644                         init_waitqueue_head(&pipe->wait);
 645                         pipe->r_counter = pipe->w_counter = 1;
 646                         pipe->buffers = pipe_bufs;
 647                         pipe->user = user;
 648                         account_pipe_buffers(pipe, 0, pipe_bufs);
 649                         mutex_init(&pipe->mutex);
 650                         return pipe;
 651                 }
 652                 free_uid(user);
 653                 kfree(pipe);
 654         }
 655
 656         return NULL;
 657 }
 658
 659 void free_pipe_info(struct pipe_inode_info *pipe)
 660 {
 661         int i;
 662
 663         account_pipe_buffers(pipe, pipe->buffers, 0);
 664         free_uid(pipe->user);
 665         for (i = 0; i < pipe->buffers; i++) {
 666                 struct pipe_buffer *buf = pipe->bufs + i;
 667                 if (buf->ops)
 668                         buf->ops->release(pipe, buf);
 669         }
 670         if (pipe->tmp_page)
 671                 __free_page(pipe->tmp_page);
 672         kfree(pipe->bufs);
 673         kfree(pipe);
 674 }
 675
 676 static struct vfsmount *pipe_mnt __read_mostly;
 677
 678 /*
 679  * pipefs_dname() is called from d_path().
 680  */
 681 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 682 {
 683         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 684                                 d_inode(dentry)->i_ino);
 685 }
 686
 687 static const struct dentry_operations pipefs_dentry_operations = {
 688         .d_dname        = pipefs_dname,
 689 };
 690
 691 static struct inode * get_pipe_inode(void)
 692 {
 693         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 694         struct pipe_inode_info *pipe;
 695
 696         if (!inode)
 697                 goto fail_inode;
 698
 699         inode->i_ino = get_next_ino();
 700
 701         pipe = alloc_pipe_info();
 702         if (!pipe)
 703                 goto fail_iput;
 704
 705         inode->i_pipe = pipe;
 706         pipe->files = 2;
 707         pipe->readers = pipe->writers = 1;
 708         inode->i_fop = &pipefifo_fops;
 709
 710         /*
 711          * Mark the inode dirty from the very beginning,
 712          * that way it will never be moved to the dirty
 713          * list because "mark_inode_dirty()" will think
 714          * that it already _is_ on the dirty list.
 715          */
 716         inode->i_state = I_DIRTY;
 717         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 718         inode->i_uid = current_fsuid();
 719         inode->i_gid = current_fsgid();
 720         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 721
 722         return inode;
 723
 724 fail_iput:
 725         iput(inode);
 726
 727 fail_inode:
 728         return NULL;
 729 }
 730
 731 int create_pipe_files(struct file **res, int flags)
 732 {
 733         int err;
 734         struct inode *inode = get_pipe_inode();
 735         struct file *f;
 736         struct path path;
 737         static struct qstr name = { .name = "" };
 738
 739         if (!inode)
 740                 return -ENFILE;
 741
 742         err = -ENOMEM;
 743         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
 744         if (!path.dentry)
 745                 goto err_inode;
 746         path.mnt = mntget(pipe_mnt);
 747
 748         d_instantiate(path.dentry, inode);
 749
 750         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 751         if (IS_ERR(f)) {
 752                 err = PTR_ERR(f);
 753                 goto err_dentry;
 754         }
 755
 756         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 757         f->private_data = inode->i_pipe;
 758
 759         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 760         if (IS_ERR(res[0])) {
 761                 err = PTR_ERR(res[0]);
 762                 goto err_file;
 763         }
 764
 765         path_get(&path);
 766         res[0]->private_data = inode->i_pipe;
 767         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 768         res[1] = f;
 769         return 0;
 770
 771 err_file:
 772         put_filp(f);
 773 err_dentry:
 774         free_pipe_info(inode->i_pipe);
 775         path_put(&path);
 776         return err;
 777
 778 err_inode:
 779         free_pipe_info(inode->i_pipe);
 780         iput(inode);
 781         return err;
 782 }
 783
 784 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 785 {
 786         int error;
 787         int fdw, fdr;
 788
 789         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 790                 return -EINVAL;
 791
 792         error = create_pipe_files(files, flags);
 793         if (error)
 794                 return error;
 795
 796         error = get_unused_fd_flags(flags);
 797         if (error < 0)
 798                 goto err_read_pipe;
 799         fdr = error;
 800
 801         error = get_unused_fd_flags(flags);
 802         if (error < 0)
 803                 goto err_fdr;
 804         fdw = error;
 805
 806         audit_fd_pair(fdr, fdw);
 807         fd[0] = fdr;
 808         fd[1] = fdw;
 809         return 0;
 810
 811  err_fdr:
 812         put_unused_fd(fdr);
 813  err_read_pipe:
 814         fput(files[0]);
 815         fput(files[1]);
 816         return error;
 817 }
 818
 819 int do_pipe_flags(int *fd, int flags)
 820 {
 821         struct file *files[2];
 822         int error = __do_pipe_flags(fd, files, flags);
 823         if (!error) {
 824                 fd_install(fd[0], files[0]);
 825                 fd_install(fd[1], files[1]);
 826         }
 827         return error;
 828 }
 829
 830 /*
 831  * sys_pipe() is the normal C calling standard for creating
 832  * a pipe. It's not the way Unix traditionally does this, though.
 833  */
 834 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 835 {
 836         struct file *files[2];
 837         int fd[2];
 838         int error;
 839
 840         error = __do_pipe_flags(fd, files, flags);
 841         if (!error) {
 842                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 843                         fput(files[0]);
 844                         fput(files[1]);
 845                         put_unused_fd(fd[0]);
 846                         put_unused_fd(fd[1]);
 847                         error = -EFAULT;
 848                 } else {
 849                         fd_install(fd[0], files[0]);
 850                         fd_install(fd[1], files[1]);
 851                 }
 852         }
 853         return error;
 854 }
 855
 856 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 857 {
 858         return sys_pipe2(fildes, 0);
 859 }
 860
 861 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 862 {
 863         int cur = *cnt;
 864
 865         while (cur == *cnt) {
 866                 pipe_wait(pipe);
 867                 if (signal_pending(current))
 868                         break;
 869         }
 870         return cur == *cnt ? -ERESTARTSYS : 0;
 871 }
 872
 873 static void wake_up_partner(struct pipe_inode_info *pipe)
 874 {
 875         wake_up_interruptible(&pipe->wait);
 876 }
 877
 878 static int fifo_open(struct inode *inode, struct file *filp)
 879 {
 880         struct pipe_inode_info *pipe;
 881         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 882         int ret;
 883
 884         filp->f_version = 0;
 885
 886         spin_lock(&inode->i_lock);
 887         if (inode->i_pipe) {
 888                 pipe = inode->i_pipe;
 889                 pipe->files++;
 890                 spin_unlock(&inode->i_lock);
 891         } else {
 892                 spin_unlock(&inode->i_lock);
 893                 pipe = alloc_pipe_info();
 894                 if (!pipe)
 895                         return -ENOMEM;
 896                 pipe->files = 1;
 897                 spin_lock(&inode->i_lock);
 898                 if (unlikely(inode->i_pipe)) {
 899                         inode->i_pipe->files++;
 900                         spin_unlock(&inode->i_lock);
 901                         free_pipe_info(pipe);
 902                         pipe = inode->i_pipe;
 903                 } else {
 904                         inode->i_pipe = pipe;
 905                         spin_unlock(&inode->i_lock);
 906                 }
 907         }
 908         filp->private_data = pipe;
 909         /* OK, we have a pipe and it's pinned down */
 910
 911         __pipe_lock(pipe);
 912
 913         /* We can only do regular read/write on fifos */
 914         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 915
 916         switch (filp->f_mode) {
 917         case FMODE_READ:
 918         /*
 919          *  O_RDONLY
 920          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 921          *  opened, even when there is no process writing the FIFO.
 922          */
 923                 pipe->r_counter++;
 924                 if (pipe->readers++ == 0)
 925                         wake_up_partner(pipe);
 926
 927                 if (!is_pipe && !pipe->writers) {
 928                         if ((filp->f_flags & O_NONBLOCK)) {
 929                                 /* suppress POLLHUP until we have
 930                                  * seen a writer */
 931                                 filp->f_version = pipe->w_counter;
 932                         } else {
 933                                 if (wait_for_partner(pipe, &pipe->w_counter))
 934                                         goto err_rd;
 935                         }
 936                 }
 937                 break;
 938
 939         case FMODE_WRITE:
 940         /*
 941          *  O_WRONLY
 942          *  POSIX.1 says that O_NONBLOCK means return -1 with
 943          *  errno=ENXIO when there is no process reading the FIFO.
 944          */
 945                 ret = -ENXIO;
 946                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 947                         goto err;
 948
 949                 pipe->w_counter++;
 950                 if (!pipe->writers++)
 951                         wake_up_partner(pipe);
 952
 953                 if (!is_pipe && !pipe->readers) {
 954                         if (wait_for_partner(pipe, &pipe->r_counter))
 955                                 goto err_wr;
 956                 }
 957                 break;
 958
 959         case FMODE_READ | FMODE_WRITE:
 960         /*
 961          *  O_RDWR
 962          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 963          *  This implementation will NEVER block on a O_RDWR open, since
 964          *  the process can at least talk to itself.
 965          */
 966
 967                 pipe->readers++;
 968                 pipe->writers++;
 969                 pipe->r_counter++;
 970                 pipe->w_counter++;
 971                 if (pipe->readers == 1 || pipe->writers == 1)
 972                         wake_up_partner(pipe);
 973                 break;
 974
 975         default:
 976                 ret = -EINVAL;
 977                 goto err;
 978         }
 979
 980         /* Ok! */
 981         __pipe_unlock(pipe);
 982         return 0;
 983
 984 err_rd:
 985         if (!--pipe->readers)
 986                 wake_up_interruptible(&pipe->wait);
 987         ret = -ERESTARTSYS;
 988         goto err;
 989
 990 err_wr:
 991         if (!--pipe->writers)
 992                 wake_up_interruptible(&pipe->wait);
 993         ret = -ERESTARTSYS;
 994         goto err;
 995
 996 err:
 997         __pipe_unlock(pipe);
 998
 999         put_pipe_info(inode, pipe);
1000         return ret;
1001 }
1002
1003 const struct file_operations pipefifo_fops = {
1004         .open           = fifo_open,
1005         .llseek         = no_llseek,
1006         .read_iter      = pipe_read,
1007         .write_iter     = pipe_write,
1008         .poll           = pipe_poll,
1009         .unlocked_ioctl = pipe_ioctl,
1010         .release        = pipe_release,
1011         .fasync         = pipe_fasync,
1012 };
1013
1014 /*
1015  * Allocate a new array of pipe buffers and copy the info over. Returns the
1016  * pipe size if successful, or return -ERROR on error.
1017  */
1018 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1019 {
1020         struct pipe_buffer *bufs;
1021
1022         if (!nr_pages)
1023                 return -EINVAL;
1024
1025         /*
1026          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1027          * expect a lot of shrink+grow operations, just free and allocate
1028          * again like we would do for growing. If the pipe currently
1029          * contains more buffers than arg, then return busy.
1030          */
1031         if (nr_pages < pipe->nrbufs)
1032                 return -EBUSY;
1033
1034         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1035         if (unlikely(!bufs))
1036                 return -ENOMEM;
1037
1038         /*
1039          * The pipe array wraps around, so just start the new one at zero
1040          * and adjust the indexes.
1041          */
1042         if (pipe->nrbufs) {
1043                 unsigned int tail;
1044                 unsigned int head;
1045
1046                 tail = pipe->curbuf + pipe->nrbufs;
1047                 if (tail < pipe->buffers)
1048                         tail = 0;
1049                 else
1050                         tail &= (pipe->buffers - 1);
1051
1052                 head = pipe->nrbufs - tail;
1053                 if (head)
1054                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1055                 if (tail)
1056                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1057         }
1058
1059         account_pipe_buffers(pipe, pipe->buffers, nr_pages);
1060         pipe->curbuf = 0;
1061         kfree(pipe->bufs);
1062         pipe->bufs = bufs;
1063         pipe->buffers = nr_pages;
1064         return nr_pages * PAGE_SIZE;
1065 }
1066
1067 /*
1068  * Currently we rely on the pipe array holding a power-of-2 number
1069  * of pages. Returns 0 on error.
1070  */
1071 static inline unsigned int round_pipe_size(unsigned int size)
1072 {
1073         unsigned long nr_pages;
1074
1075         if (size < pipe_min_size)
1076                 size = pipe_min_size;
1077
1078         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1079         if (nr_pages == 0)
1080                 return 0;
1081
1082         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1083 }
1084
1085 /*
1086  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1087  * will return an error.
1088  */
1089 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1090                  size_t *lenp, loff_t *ppos)
1091 {
1092         unsigned int rounded_pipe_max_size;
1093         int ret;
1094
1095         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1096         if (ret < 0 || !write)
1097                 return ret;
1098
1099         rounded_pipe_max_size = round_pipe_size(pipe_max_size);
1100         if (rounded_pipe_max_size == 0)
1101                 return -EINVAL;
1102
1103         pipe_max_size = rounded_pipe_max_size;
1104         return ret;
1105 }
1106
1107 /*
1108  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1109  * location, so checking ->i_pipe is not enough to verify that this is a
1110  * pipe.
1111  */
1112 struct pipe_inode_info *get_pipe_info(struct file *file)
1113 {
1114         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1115 }
1116
1117 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1118 {
1119         struct pipe_inode_info *pipe;
1120         long ret;
1121
1122         pipe = get_pipe_info(file);
1123         if (!pipe)
1124                 return -EBADF;
1125
1126         __pipe_lock(pipe);
1127
1128         switch (cmd) {
1129         case F_SETPIPE_SZ: {
1130                 unsigned int size, nr_pages;
1131
1132                 size = round_pipe_size(arg);
1133                 nr_pages = size >> PAGE_SHIFT;
1134
1135                 ret = -EINVAL;
1136                 if (!nr_pages)
1137                         goto out;
1138
1139                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1140                         ret = -EPERM;
1141                         goto out;
1142                 } else if ((too_many_pipe_buffers_hard(pipe->user) ||
1143                             too_many_pipe_buffers_soft(pipe->user)) &&
1144                            !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1145                         ret = -EPERM;
1146                         goto out;
1147                 }
1148                 ret = pipe_set_size(pipe, nr_pages);
1149                 break;
1150                 }
1151         case F_GETPIPE_SZ:
1152                 ret = pipe->buffers * PAGE_SIZE;
1153                 break;
1154         default:
1155                 ret = -EINVAL;
1156                 break;
1157         }
1158
1159 out:
1160         __pipe_unlock(pipe);
1161         return ret;
1162 }
1163
1164 static const struct super_operations pipefs_ops = {
1165         .destroy_inode = free_inode_nonrcu,
1166         .statfs = simple_statfs,
1167 };
1168
1169 /*
1170  * pipefs should _never_ be mounted by userland - too much of security hassle,
1171  * no real gain from having the whole whorehouse mounted. So we don't need
1172  * any operations on the root directory. However, we need a non-trivial
1173  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1174  */
1175 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1176                          int flags, const char *dev_name, void *data)
1177 {
1178         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1179                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1180 }
1181
1182 static struct file_system_type pipe_fs_type = {
1183         .name           = "pipefs",
1184         .mount          = pipefs_mount,
1185         .kill_sb        = kill_anon_super,
1186 };
1187
1188 static int __init init_pipe_fs(void)
1189 {
1190         int err = register_filesystem(&pipe_fs_type);
1191
1192         if (!err) {
1193                 pipe_mnt = kern_mount(&pipe_fs_type);
1194                 if (IS_ERR(pipe_mnt)) {
1195                         err = PTR_ERR(pipe_mnt);
1196                         unregister_filesystem(&pipe_fs_type);
1197                 }
1198         }
1199         return err;
1200 }
1201
1202 fs_initcall(init_pipe_fs);