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 #include <linux/aio.h>
  25
  26 #include <asm/uaccess.h>
  27 #include <asm/ioctls.h>
  28
  29 #include "internal.h"
  30
  31 /*
  32  * The max size that a non-root user is allowed to grow the pipe. Can
  33  * be set by root in /proc/sys/fs/pipe-max-size
  34  */
  35 unsigned int pipe_max_size = 1048576;
  36
  37 /*
  38  * Minimum pipe size, as required by POSIX
  39  */
  40 unsigned int pipe_min_size = PAGE_SIZE;
  41
  42 /*
  43  * We use a start+len construction, which provides full use of the
  44  * allocated memory.
  45  * -- Florian Coosmann (FGC)
  46  *
  47  * Reads with count = 0 should always return 0.
  48  * -- Julian Bradfield 1999-06-07.
  49  *
  50  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  51  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  52  *
  53  * pipe_read & write cleanup
  54  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  55  */
  56
  57 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  58 {
  59         if (pipe->files)
  60                 mutex_lock_nested(&pipe->mutex, subclass);
  61 }
  62
  63 void pipe_lock(struct pipe_inode_info *pipe)
  64 {
  65         /*
  66          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  67          */
  68         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  69 }
  70 EXPORT_SYMBOL(pipe_lock);
  71
  72 void pipe_unlock(struct pipe_inode_info *pipe)
  73 {
  74         if (pipe->files)
  75                 mutex_unlock(&pipe->mutex);
  76 }
  77 EXPORT_SYMBOL(pipe_unlock);
  78
  79 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  80 {
  81         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  82 }
  83
  84 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  85 {
  86         mutex_unlock(&pipe->mutex);
  87 }
  88
  89 void pipe_double_lock(struct pipe_inode_info *pipe1,
  90                       struct pipe_inode_info *pipe2)
  91 {
  92         BUG_ON(pipe1 == pipe2);
  93
  94         if (pipe1 < pipe2) {
  95                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  96                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  97         } else {
  98                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
  99                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 100         }
 101 }
 102
 103 /* Drop the inode semaphore and wait for a pipe event, atomically */
 104 void pipe_wait(struct pipe_inode_info *pipe)
 105 {
 106         DEFINE_WAIT(wait);
 107
 108         /*
 109          * Pipes are system-local resources, so sleeping on them
 110          * is considered a noninteractive wait:
 111          */
 112         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 113         pipe_unlock(pipe);
 114         schedule();
 115         finish_wait(&pipe->wait, &wait);
 116         pipe_lock(pipe);
 117 }
 118
 119 static int
 120 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
 121                         int atomic)
 122 {
 123         unsigned long copy;
 124
 125         while (len > 0) {
 126                 while (!iov->iov_len)
 127                         iov++;
 128                 copy = min_t(unsigned long, len, iov->iov_len);
 129
 130                 if (atomic) {
 131                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
 132                                 return -EFAULT;
 133                 } else {
 134                         if (copy_from_user(to, iov->iov_base, copy))
 135                                 return -EFAULT;
 136                 }
 137                 to += copy;
 138                 len -= copy;
 139                 iov->iov_base += copy;
 140                 iov->iov_len -= copy;
 141         }
 142         return 0;
 143 }
 144
 145 /*
 146  * Pre-fault in the user memory, so we can use atomic copies.
 147  */
 148 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 149 {
 150         while (!iov->iov_len)
 151                 iov++;
 152
 153         while (len > 0) {
 154                 unsigned long this_len;
 155
 156                 this_len = min_t(unsigned long, len, iov->iov_len);
 157                 fault_in_pages_readable(iov->iov_base, this_len);
 158                 len -= this_len;
 159                 iov++;
 160         }
 161 }
 162
 163 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 164                                   struct pipe_buffer *buf)
 165 {
 166         struct page *page = buf->page;
 167
 168         /*
 169          * If nobody else uses this page, and we don't already have a
 170          * temporary page, let's keep track of it as a one-deep
 171          * allocation cache. (Otherwise just release our reference to it)
 172          */
 173         if (page_count(page) == 1 && !pipe->tmp_page)
 174                 pipe->tmp_page = page;
 175         else
 176                 page_cache_release(page);
 177 }
 178
 179 /**
 180  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 181  * @pipe:       the pipe that the buffer belongs to
 182  * @buf:        the buffer to attempt to steal
 183  *
 184  * Description:
 185  *      This function attempts to steal the &struct page attached to
 186  *      @buf. If successful, this function returns 0 and returns with
 187  *      the page locked. The caller may then reuse the page for whatever
 188  *      he wishes; the typical use is insertion into a different file
 189  *      page cache.
 190  */
 191 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 192                            struct pipe_buffer *buf)
 193 {
 194         struct page *page = buf->page;
 195
 196         /*
 197          * A reference of one is golden, that means that the owner of this
 198          * page is the only one holding a reference to it. lock the page
 199          * and return OK.
 200          */
 201         if (page_count(page) == 1) {
 202                 lock_page(page);
 203                 return 0;
 204         }
 205
 206         return 1;
 207 }
 208 EXPORT_SYMBOL(generic_pipe_buf_steal);
 209
 210 /**
 211  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 212  * @pipe:       the pipe that the buffer belongs to
 213  * @buf:        the buffer to get a reference to
 214  *
 215  * Description:
 216  *      This function grabs an extra reference to @buf. It's used in
 217  *      in the tee() system call, when we duplicate the buffers in one
 218  *      pipe into another.
 219  */
 220 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 221 {
 222         page_cache_get(buf->page);
 223 }
 224 EXPORT_SYMBOL(generic_pipe_buf_get);
 225
 226 /**
 227  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 228  * @info:       the pipe that the buffer belongs to
 229  * @buf:        the buffer to confirm
 230  *
 231  * Description:
 232  *      This function does nothing, because the generic pipe code uses
 233  *      pages that are always good when inserted into the pipe.
 234  */
 235 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 236                              struct pipe_buffer *buf)
 237 {
 238         return 0;
 239 }
 240 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 241
 242 /**
 243  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 244  * @pipe:       the pipe that the buffer belongs to
 245  * @buf:        the buffer to put a reference to
 246  *
 247  * Description:
 248  *      This function releases a reference to @buf.
 249  */
 250 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 251                               struct pipe_buffer *buf)
 252 {
 253         page_cache_release(buf->page);
 254 }
 255 EXPORT_SYMBOL(generic_pipe_buf_release);
 256
 257 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 258         .can_merge = 1,
 259         .confirm = generic_pipe_buf_confirm,
 260         .release = anon_pipe_buf_release,
 261         .steal = generic_pipe_buf_steal,
 262         .get = generic_pipe_buf_get,
 263 };
 264
 265 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 266         .can_merge = 0,
 267         .confirm = generic_pipe_buf_confirm,
 268         .release = anon_pipe_buf_release,
 269         .steal = generic_pipe_buf_steal,
 270         .get = generic_pipe_buf_get,
 271 };
 272
 273 static ssize_t
 274 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 275            unsigned long nr_segs, loff_t pos)
 276 {
 277         struct file *filp = iocb->ki_filp;
 278         struct pipe_inode_info *pipe = filp->private_data;
 279         int do_wakeup;
 280         ssize_t ret;
 281         struct iovec *iov = (struct iovec *)_iov;
 282         size_t total_len;
 283         struct iov_iter iter;
 284
 285         total_len = iov_length(iov, nr_segs);
 286         /* Null read succeeds. */
 287         if (unlikely(total_len == 0))
 288                 return 0;
 289
 290         iov_iter_init(&iter, iov, nr_segs, total_len, 0);
 291
 292         do_wakeup = 0;
 293         ret = 0;
 294         __pipe_lock(pipe);
 295         for (;;) {
 296                 int bufs = pipe->nrbufs;
 297                 if (bufs) {
 298                         int curbuf = pipe->curbuf;
 299                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 300                         const struct pipe_buf_operations *ops = buf->ops;
 301                         size_t chars = buf->len;
 302                         size_t written;
 303                         int error;
 304
 305                         if (chars > total_len)
 306                                 chars = total_len;
 307
 308                         error = ops->confirm(pipe, buf);
 309                         if (error) {
 310                                 if (!ret)
 311                                         ret = error;
 312                                 break;
 313                         }
 314
 315                         written = copy_page_to_iter(buf->page, buf->offset, chars, &iter);
 316                         if (unlikely(written < chars)) {
 317                                 if (!ret)
 318                                         ret = -EFAULT;
 319                                 break;
 320                         }
 321                         ret += chars;
 322                         buf->offset += chars;
 323                         buf->len -= chars;
 324
 325                         /* Was it a packet buffer? Clean up and exit */
 326                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 327                                 total_len = chars;
 328                                 buf->len = 0;
 329                         }
 330
 331                         if (!buf->len) {
 332                                 buf->ops = NULL;
 333                                 ops->release(pipe, buf);
 334                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 335                                 pipe->curbuf = curbuf;
 336                                 pipe->nrbufs = --bufs;
 337                                 do_wakeup = 1;
 338                         }
 339                         total_len -= chars;
 340                         if (!total_len)
 341                                 break;  /* common path: read succeeded */
 342                 }
 343                 if (bufs)       /* More to do? */
 344                         continue;
 345                 if (!pipe->writers)
 346                         break;
 347                 if (!pipe->waiting_writers) {
 348                         /* syscall merging: Usually we must not sleep
 349                          * if O_NONBLOCK is set, or if we got some data.
 350                          * But if a writer sleeps in kernel space, then
 351                          * we can wait for that data without violating POSIX.
 352                          */
 353                         if (ret)
 354                                 break;
 355                         if (filp->f_flags & O_NONBLOCK) {
 356                                 ret = -EAGAIN;
 357                                 break;
 358                         }
 359                 }
 360                 if (signal_pending(current)) {
 361                         if (!ret)
 362                                 ret = -ERESTARTSYS;
 363                         break;
 364                 }
 365                 if (do_wakeup) {
 366                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 367                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 368                 }
 369                 pipe_wait(pipe);
 370         }
 371         __pipe_unlock(pipe);
 372
 373         /* Signal writers asynchronously that there is more room. */
 374         if (do_wakeup) {
 375                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 376                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 377         }
 378         if (ret > 0)
 379                 file_accessed(filp);
 380         return ret;
 381 }
 382
 383 static inline int is_packetized(struct file *file)
 384 {
 385         return (file->f_flags & O_DIRECT) != 0;
 386 }
 387
 388 static ssize_t
 389 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 390             unsigned long nr_segs, loff_t ppos)
 391 {
 392         struct file *filp = iocb->ki_filp;
 393         struct pipe_inode_info *pipe = filp->private_data;
 394         ssize_t ret;
 395         int do_wakeup;
 396         struct iovec *iov = (struct iovec *)_iov;
 397         size_t total_len;
 398         ssize_t chars;
 399
 400         total_len = iov_length(iov, nr_segs);
 401         /* Null write succeeds. */
 402         if (unlikely(total_len == 0))
 403                 return 0;
 404
 405         do_wakeup = 0;
 406         ret = 0;
 407         __pipe_lock(pipe);
 408
 409         if (!pipe->readers) {
 410                 send_sig(SIGPIPE, current, 0);
 411                 ret = -EPIPE;
 412                 goto out;
 413         }
 414
 415         /* We try to merge small writes */
 416         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 417         if (pipe->nrbufs && chars != 0) {
 418                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 419                                                         (pipe->buffers - 1);
 420                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 421                 const struct pipe_buf_operations *ops = buf->ops;
 422                 int offset = buf->offset + buf->len;
 423
 424                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 425                         int error, atomic = 1;
 426                         void *addr;
 427
 428                         error = ops->confirm(pipe, buf);
 429                         if (error)
 430                                 goto out;
 431
 432                         iov_fault_in_pages_read(iov, chars);
 433 redo1:
 434                         if (atomic)
 435                                 addr = kmap_atomic(buf->page);
 436                         else
 437                                 addr = kmap(buf->page);
 438                         error = pipe_iov_copy_from_user(offset + addr, iov,
 439                                                         chars, atomic);
 440                         if (atomic)
 441                                 kunmap_atomic(addr);
 442                         else
 443                                 kunmap(buf->page);
 444                         ret = error;
 445                         do_wakeup = 1;
 446                         if (error) {
 447                                 if (atomic) {
 448                                         atomic = 0;
 449                                         goto redo1;
 450                                 }
 451                                 goto out;
 452                         }
 453                         buf->len += chars;
 454                         total_len -= chars;
 455                         ret = chars;
 456                         if (!total_len)
 457                                 goto out;
 458                 }
 459         }
 460
 461         for (;;) {
 462                 int bufs;
 463
 464                 if (!pipe->readers) {
 465                         send_sig(SIGPIPE, current, 0);
 466                         if (!ret)
 467                                 ret = -EPIPE;
 468                         break;
 469                 }
 470                 bufs = pipe->nrbufs;
 471                 if (bufs < pipe->buffers) {
 472                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 473                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 474                         struct page *page = pipe->tmp_page;
 475                         char *src;
 476                         int error, atomic = 1;
 477
 478                         if (!page) {
 479                                 page = alloc_page(GFP_HIGHUSER);
 480                                 if (unlikely(!page)) {
 481                                         ret = ret ? : -ENOMEM;
 482                                         break;
 483                                 }
 484                                 pipe->tmp_page = page;
 485                         }
 486                         /* Always wake up, even if the copy fails. Otherwise
 487                          * we lock up (O_NONBLOCK-)readers that sleep due to
 488                          * syscall merging.
 489                          * FIXME! Is this really true?
 490                          */
 491                         do_wakeup = 1;
 492                         chars = PAGE_SIZE;
 493                         if (chars > total_len)
 494                                 chars = total_len;
 495
 496                         iov_fault_in_pages_read(iov, chars);
 497 redo2:
 498                         if (atomic)
 499                                 src = kmap_atomic(page);
 500                         else
 501                                 src = kmap(page);
 502
 503                         error = pipe_iov_copy_from_user(src, iov, chars,
 504                                                         atomic);
 505                         if (atomic)
 506                                 kunmap_atomic(src);
 507                         else
 508                                 kunmap(page);
 509
 510                         if (unlikely(error)) {
 511                                 if (atomic) {
 512                                         atomic = 0;
 513                                         goto redo2;
 514                                 }
 515                                 if (!ret)
 516                                         ret = error;
 517                                 break;
 518                         }
 519                         ret += chars;
 520
 521                         /* Insert it into the buffer array */
 522                         buf->page = page;
 523                         buf->ops = &anon_pipe_buf_ops;
 524                         buf->offset = 0;
 525                         buf->len = chars;
 526                         buf->flags = 0;
 527                         if (is_packetized(filp)) {
 528                                 buf->ops = &packet_pipe_buf_ops;
 529                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 530                         }
 531                         pipe->nrbufs = ++bufs;
 532                         pipe->tmp_page = NULL;
 533
 534                         total_len -= chars;
 535                         if (!total_len)
 536                                 break;
 537                 }
 538                 if (bufs < pipe->buffers)
 539                         continue;
 540                 if (filp->f_flags & O_NONBLOCK) {
 541                         if (!ret)
 542                                 ret = -EAGAIN;
 543                         break;
 544                 }
 545                 if (signal_pending(current)) {
 546                         if (!ret)
 547                                 ret = -ERESTARTSYS;
 548                         break;
 549                 }
 550                 if (do_wakeup) {
 551                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 552                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 553                         do_wakeup = 0;
 554                 }
 555                 pipe->waiting_writers++;
 556                 pipe_wait(pipe);
 557                 pipe->waiting_writers--;
 558         }
 559 out:
 560         __pipe_unlock(pipe);
 561         if (do_wakeup) {
 562                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 563                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 564         }
 565         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 566                 int err = file_update_time(filp);
 567                 if (err)
 568                         ret = err;
 569                 sb_end_write(file_inode(filp)->i_sb);
 570         }
 571         return ret;
 572 }
 573
 574 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 575 {
 576         struct pipe_inode_info *pipe = filp->private_data;
 577         int count, buf, nrbufs;
 578
 579         switch (cmd) {
 580                 case FIONREAD:
 581                         __pipe_lock(pipe);
 582                         count = 0;
 583                         buf = pipe->curbuf;
 584                         nrbufs = pipe->nrbufs;
 585                         while (--nrbufs >= 0) {
 586                                 count += pipe->bufs[buf].len;
 587                                 buf = (buf+1) & (pipe->buffers - 1);
 588                         }
 589                         __pipe_unlock(pipe);
 590
 591                         return put_user(count, (int __user *)arg);
 592                 default:
 593                         return -ENOIOCTLCMD;
 594         }
 595 }
 596
 597 /* No kernel lock held - fine */
 598 static unsigned int
 599 pipe_poll(struct file *filp, poll_table *wait)
 600 {
 601         unsigned int mask;
 602         struct pipe_inode_info *pipe = filp->private_data;
 603         int nrbufs;
 604
 605         poll_wait(filp, &pipe->wait, wait);
 606
 607         /* Reading only -- no need for acquiring the semaphore.  */
 608         nrbufs = pipe->nrbufs;
 609         mask = 0;
 610         if (filp->f_mode & FMODE_READ) {
 611                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 612                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 613                         mask |= POLLHUP;
 614         }
 615
 616         if (filp->f_mode & FMODE_WRITE) {
 617                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 618                 /*
 619                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 620                  * behave exactly like pipes for poll().
 621                  */
 622                 if (!pipe->readers)
 623                         mask |= POLLERR;
 624         }
 625
 626         return mask;
 627 }
 628
 629 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 630 {
 631         int kill = 0;
 632
 633         spin_lock(&inode->i_lock);
 634         if (!--pipe->files) {
 635                 inode->i_pipe = NULL;
 636                 kill = 1;
 637         }
 638         spin_unlock(&inode->i_lock);
 639
 640         if (kill)
 641                 free_pipe_info(pipe);
 642 }
 643
 644 static int
 645 pipe_release(struct inode *inode, struct file *file)
 646 {
 647         struct pipe_inode_info *pipe = file->private_data;
 648
 649         __pipe_lock(pipe);
 650         if (file->f_mode & FMODE_READ)
 651                 pipe->readers--;
 652         if (file->f_mode & FMODE_WRITE)
 653                 pipe->writers--;
 654
 655         if (pipe->readers || pipe->writers) {
 656                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 657                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 658                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 659         }
 660         __pipe_unlock(pipe);
 661
 662         put_pipe_info(inode, pipe);
 663         return 0;
 664 }
 665
 666 static int
 667 pipe_fasync(int fd, struct file *filp, int on)
 668 {
 669         struct pipe_inode_info *pipe = filp->private_data;
 670         int retval = 0;
 671
 672         __pipe_lock(pipe);
 673         if (filp->f_mode & FMODE_READ)
 674                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 675         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 676                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 677                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 678                         /* this can happen only if on == T */
 679                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 680         }
 681         __pipe_unlock(pipe);
 682         return retval;
 683 }
 684
 685 struct pipe_inode_info *alloc_pipe_info(void)
 686 {
 687         struct pipe_inode_info *pipe;
 688
 689         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 690         if (pipe) {
 691                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
 692                 if (pipe->bufs) {
 693                         init_waitqueue_head(&pipe->wait);
 694                         pipe->r_counter = pipe->w_counter = 1;
 695                         pipe->buffers = PIPE_DEF_BUFFERS;
 696                         mutex_init(&pipe->mutex);
 697                         return pipe;
 698                 }
 699                 kfree(pipe);
 700         }
 701
 702         return NULL;
 703 }
 704
 705 void free_pipe_info(struct pipe_inode_info *pipe)
 706 {
 707         int i;
 708
 709         for (i = 0; i < pipe->buffers; i++) {
 710                 struct pipe_buffer *buf = pipe->bufs + i;
 711                 if (buf->ops)
 712                         buf->ops->release(pipe, buf);
 713         }
 714         if (pipe->tmp_page)
 715                 __free_page(pipe->tmp_page);
 716         kfree(pipe->bufs);
 717         kfree(pipe);
 718 }
 719
 720 static struct vfsmount *pipe_mnt __read_mostly;
 721
 722 /*
 723  * pipefs_dname() is called from d_path().
 724  */
 725 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 726 {
 727         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 728                                 dentry->d_inode->i_ino);
 729 }
 730
 731 static const struct dentry_operations pipefs_dentry_operations = {
 732         .d_dname        = pipefs_dname,
 733 };
 734
 735 static struct inode * get_pipe_inode(void)
 736 {
 737         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 738         struct pipe_inode_info *pipe;
 739
 740         if (!inode)
 741                 goto fail_inode;
 742
 743         inode->i_ino = get_next_ino();
 744
 745         pipe = alloc_pipe_info();
 746         if (!pipe)
 747                 goto fail_iput;
 748
 749         inode->i_pipe = pipe;
 750         pipe->files = 2;
 751         pipe->readers = pipe->writers = 1;
 752         inode->i_fop = &pipefifo_fops;
 753
 754         /*
 755          * Mark the inode dirty from the very beginning,
 756          * that way it will never be moved to the dirty
 757          * list because "mark_inode_dirty()" will think
 758          * that it already _is_ on the dirty list.
 759          */
 760         inode->i_state = I_DIRTY;
 761         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 762         inode->i_uid = current_fsuid();
 763         inode->i_gid = current_fsgid();
 764         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 765
 766         return inode;
 767
 768 fail_iput:
 769         iput(inode);
 770
 771 fail_inode:
 772         return NULL;
 773 }
 774
 775 int create_pipe_files(struct file **res, int flags)
 776 {
 777         int err;
 778         struct inode *inode = get_pipe_inode();
 779         struct file *f;
 780         struct path path;
 781         static struct qstr name = { .name = "" };
 782
 783         if (!inode)
 784                 return -ENFILE;
 785
 786         err = -ENOMEM;
 787         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
 788         if (!path.dentry)
 789                 goto err_inode;
 790         path.mnt = mntget(pipe_mnt);
 791
 792         d_instantiate(path.dentry, inode);
 793
 794         err = -ENFILE;
 795         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 796         if (IS_ERR(f))
 797                 goto err_dentry;
 798
 799         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 800         f->private_data = inode->i_pipe;
 801
 802         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 803         if (IS_ERR(res[0]))
 804                 goto err_file;
 805
 806         path_get(&path);
 807         res[0]->private_data = inode->i_pipe;
 808         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 809         res[1] = f;
 810         return 0;
 811
 812 err_file:
 813         put_filp(f);
 814 err_dentry:
 815         free_pipe_info(inode->i_pipe);
 816         path_put(&path);
 817         return err;
 818
 819 err_inode:
 820         free_pipe_info(inode->i_pipe);
 821         iput(inode);
 822         return err;
 823 }
 824
 825 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 826 {
 827         int error;
 828         int fdw, fdr;
 829
 830         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 831                 return -EINVAL;
 832
 833         error = create_pipe_files(files, flags);
 834         if (error)
 835                 return error;
 836
 837         error = get_unused_fd_flags(flags);
 838         if (error < 0)
 839                 goto err_read_pipe;
 840         fdr = error;
 841
 842         error = get_unused_fd_flags(flags);
 843         if (error < 0)
 844                 goto err_fdr;
 845         fdw = error;
 846
 847         audit_fd_pair(fdr, fdw);
 848         fd[0] = fdr;
 849         fd[1] = fdw;
 850         return 0;
 851
 852  err_fdr:
 853         put_unused_fd(fdr);
 854  err_read_pipe:
 855         fput(files[0]);
 856         fput(files[1]);
 857         return error;
 858 }
 859
 860 int do_pipe_flags(int *fd, int flags)
 861 {
 862         struct file *files[2];
 863         int error = __do_pipe_flags(fd, files, flags);
 864         if (!error) {
 865                 fd_install(fd[0], files[0]);
 866                 fd_install(fd[1], files[1]);
 867         }
 868         return error;
 869 }
 870
 871 /*
 872  * sys_pipe() is the normal C calling standard for creating
 873  * a pipe. It's not the way Unix traditionally does this, though.
 874  */
 875 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 876 {
 877         struct file *files[2];
 878         int fd[2];
 879         int error;
 880
 881         error = __do_pipe_flags(fd, files, flags);
 882         if (!error) {
 883                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 884                         fput(files[0]);
 885                         fput(files[1]);
 886                         put_unused_fd(fd[0]);
 887                         put_unused_fd(fd[1]);
 888                         error = -EFAULT;
 889                 } else {
 890                         fd_install(fd[0], files[0]);
 891                         fd_install(fd[1], files[1]);
 892                 }
 893         }
 894         return error;
 895 }
 896
 897 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 898 {
 899         return sys_pipe2(fildes, 0);
 900 }
 901
 902 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 903 {
 904         int cur = *cnt;
 905
 906         while (cur == *cnt) {
 907                 pipe_wait(pipe);
 908                 if (signal_pending(current))
 909                         break;
 910         }
 911         return cur == *cnt ? -ERESTARTSYS : 0;
 912 }
 913
 914 static void wake_up_partner(struct pipe_inode_info *pipe)
 915 {
 916         wake_up_interruptible(&pipe->wait);
 917 }
 918
 919 static int fifo_open(struct inode *inode, struct file *filp)
 920 {
 921         struct pipe_inode_info *pipe;
 922         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 923         int ret;
 924
 925         filp->f_version = 0;
 926
 927         spin_lock(&inode->i_lock);
 928         if (inode->i_pipe) {
 929                 pipe = inode->i_pipe;
 930                 pipe->files++;
 931                 spin_unlock(&inode->i_lock);
 932         } else {
 933                 spin_unlock(&inode->i_lock);
 934                 pipe = alloc_pipe_info();
 935                 if (!pipe)
 936                         return -ENOMEM;
 937                 pipe->files = 1;
 938                 spin_lock(&inode->i_lock);
 939                 if (unlikely(inode->i_pipe)) {
 940                         inode->i_pipe->files++;
 941                         spin_unlock(&inode->i_lock);
 942                         free_pipe_info(pipe);
 943                         pipe = inode->i_pipe;
 944                 } else {
 945                         inode->i_pipe = pipe;
 946                         spin_unlock(&inode->i_lock);
 947                 }
 948         }
 949         filp->private_data = pipe;
 950         /* OK, we have a pipe and it's pinned down */
 951
 952         __pipe_lock(pipe);
 953
 954         /* We can only do regular read/write on fifos */
 955         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 956
 957         switch (filp->f_mode) {
 958         case FMODE_READ:
 959         /*
 960          *  O_RDONLY
 961          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 962          *  opened, even when there is no process writing the FIFO.
 963          */
 964                 pipe->r_counter++;
 965                 if (pipe->readers++ == 0)
 966                         wake_up_partner(pipe);
 967
 968                 if (!is_pipe && !pipe->writers) {
 969                         if ((filp->f_flags & O_NONBLOCK)) {
 970                                 /* suppress POLLHUP until we have
 971                                  * seen a writer */
 972                                 filp->f_version = pipe->w_counter;
 973                         } else {
 974                                 if (wait_for_partner(pipe, &pipe->w_counter))
 975                                         goto err_rd;
 976                         }
 977                 }
 978                 break;
 979
 980         case FMODE_WRITE:
 981         /*
 982          *  O_WRONLY
 983          *  POSIX.1 says that O_NONBLOCK means return -1 with
 984          *  errno=ENXIO when there is no process reading the FIFO.
 985          */
 986                 ret = -ENXIO;
 987                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 988                         goto err;
 989
 990                 pipe->w_counter++;
 991                 if (!pipe->writers++)
 992                         wake_up_partner(pipe);
 993
 994                 if (!is_pipe && !pipe->readers) {
 995                         if (wait_for_partner(pipe, &pipe->r_counter))
 996                                 goto err_wr;
 997                 }
 998                 break;
 999
1000         case FMODE_READ | FMODE_WRITE:
1001         /*
1002          *  O_RDWR
1003          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1004          *  This implementation will NEVER block on a O_RDWR open, since
1005          *  the process can at least talk to itself.
1006          */
1007
1008                 pipe->readers++;
1009                 pipe->writers++;
1010                 pipe->r_counter++;
1011                 pipe->w_counter++;
1012                 if (pipe->readers == 1 || pipe->writers == 1)
1013                         wake_up_partner(pipe);
1014                 break;
1015
1016         default:
1017                 ret = -EINVAL;
1018                 goto err;
1019         }
1020
1021         /* Ok! */
1022         __pipe_unlock(pipe);
1023         return 0;
1024
1025 err_rd:
1026         if (!--pipe->readers)
1027                 wake_up_interruptible(&pipe->wait);
1028         ret = -ERESTARTSYS;
1029         goto err;
1030
1031 err_wr:
1032         if (!--pipe->writers)
1033                 wake_up_interruptible(&pipe->wait);
1034         ret = -ERESTARTSYS;
1035         goto err;
1036
1037 err:
1038         __pipe_unlock(pipe);
1039
1040         put_pipe_info(inode, pipe);
1041         return ret;
1042 }
1043
1044 const struct file_operations pipefifo_fops = {
1045         .open           = fifo_open,
1046         .llseek         = no_llseek,
1047         .read           = do_sync_read,
1048         .aio_read       = pipe_read,
1049         .write          = do_sync_write,
1050         .aio_write      = pipe_write,
1051         .poll           = pipe_poll,
1052         .unlocked_ioctl = pipe_ioctl,
1053         .release        = pipe_release,
1054         .fasync         = pipe_fasync,
1055 };
1056
1057 /*
1058  * Allocate a new array of pipe buffers and copy the info over. Returns the
1059  * pipe size if successful, or return -ERROR on error.
1060  */
1061 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1062 {
1063         struct pipe_buffer *bufs;
1064
1065         /*
1066          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1067          * expect a lot of shrink+grow operations, just free and allocate
1068          * again like we would do for growing. If the pipe currently
1069          * contains more buffers than arg, then return busy.
1070          */
1071         if (nr_pages < pipe->nrbufs)
1072                 return -EBUSY;
1073
1074         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1075         if (unlikely(!bufs))
1076                 return -ENOMEM;
1077
1078         /*
1079          * The pipe array wraps around, so just start the new one at zero
1080          * and adjust the indexes.
1081          */
1082         if (pipe->nrbufs) {
1083                 unsigned int tail;
1084                 unsigned int head;
1085
1086                 tail = pipe->curbuf + pipe->nrbufs;
1087                 if (tail < pipe->buffers)
1088                         tail = 0;
1089                 else
1090                         tail &= (pipe->buffers - 1);
1091
1092                 head = pipe->nrbufs - tail;
1093                 if (head)
1094                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1095                 if (tail)
1096                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1097         }
1098
1099         pipe->curbuf = 0;
1100         kfree(pipe->bufs);
1101         pipe->bufs = bufs;
1102         pipe->buffers = nr_pages;
1103         return nr_pages * PAGE_SIZE;
1104 }
1105
1106 /*
1107  * Currently we rely on the pipe array holding a power-of-2 number
1108  * of pages.
1109  */
1110 static inline unsigned int round_pipe_size(unsigned int size)
1111 {
1112         unsigned long nr_pages;
1113
1114         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1115         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1116 }
1117
1118 /*
1119  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1120  * will return an error.
1121  */
1122 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1123                  size_t *lenp, loff_t *ppos)
1124 {
1125         int ret;
1126
1127         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1128         if (ret < 0 || !write)
1129                 return ret;
1130
1131         pipe_max_size = round_pipe_size(pipe_max_size);
1132         return ret;
1133 }
1134
1135 /*
1136  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1137  * location, so checking ->i_pipe is not enough to verify that this is a
1138  * pipe.
1139  */
1140 struct pipe_inode_info *get_pipe_info(struct file *file)
1141 {
1142         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1143 }
1144
1145 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1146 {
1147         struct pipe_inode_info *pipe;
1148         long ret;
1149
1150         pipe = get_pipe_info(file);
1151         if (!pipe)
1152                 return -EBADF;
1153
1154         __pipe_lock(pipe);
1155
1156         switch (cmd) {
1157         case F_SETPIPE_SZ: {
1158                 unsigned int size, nr_pages;
1159
1160                 size = round_pipe_size(arg);
1161                 nr_pages = size >> PAGE_SHIFT;
1162
1163                 ret = -EINVAL;
1164                 if (!nr_pages)
1165                         goto out;
1166
1167                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1168                         ret = -EPERM;
1169                         goto out;
1170                 }
1171                 ret = pipe_set_size(pipe, nr_pages);
1172                 break;
1173                 }
1174         case F_GETPIPE_SZ:
1175                 ret = pipe->buffers * PAGE_SIZE;
1176                 break;
1177         default:
1178                 ret = -EINVAL;
1179                 break;
1180         }
1181
1182 out:
1183         __pipe_unlock(pipe);
1184         return ret;
1185 }
1186
1187 static const struct super_operations pipefs_ops = {
1188         .destroy_inode = free_inode_nonrcu,
1189         .statfs = simple_statfs,
1190 };
1191
1192 /*
1193  * pipefs should _never_ be mounted by userland - too much of security hassle,
1194  * no real gain from having the whole whorehouse mounted. So we don't need
1195  * any operations on the root directory. However, we need a non-trivial
1196  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1197  */
1198 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1199                          int flags, const char *dev_name, void *data)
1200 {
1201         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1202                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1203 }
1204
1205 static struct file_system_type pipe_fs_type = {
1206         .name           = "pipefs",
1207         .mount          = pipefs_mount,
1208         .kill_sb        = kill_anon_super,
1209 };
1210
1211 static int __init init_pipe_fs(void)
1212 {
1213         int err = register_filesystem(&pipe_fs_type);
1214
1215         if (!err) {
1216                 pipe_mnt = kern_mount(&pipe_fs_type);
1217                 if (IS_ERR(pipe_mnt)) {
1218                         err = PTR_ERR(pipe_mnt);
1219                         unregister_filesystem(&pipe_fs_type);
1220                 }
1221         }
1222         return err;
1223 }
1224
1225 fs_initcall(init_pipe_fs);