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/mount.h>
  15 #include <linux/pipe_fs_i.h>
  16 #include <linux/uio.h>
  17 #include <linux/highmem.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/audit.h>
  20 #include <linux/syscalls.h>
  21
  22 #include <asm/uaccess.h>
  23 #include <asm/ioctls.h>
  24
  25 /*
  26  * We use a start+len construction, which provides full use of the
  27  * allocated memory.
  28  * -- Florian Coosmann (FGC)
  29  *
  30  * Reads with count = 0 should always return 0.
  31  * -- Julian Bradfield 1999-06-07.
  32  *
  33  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  34  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  35  *
  36  * pipe_read & write cleanup
  37  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  38  */
  39
  40 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  41 {
  42         if (pipe->inode)
  43                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
  44 }
  45
  46 void pipe_lock(struct pipe_inode_info *pipe)
  47 {
  48         /*
  49          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  50          */
  51         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  52 }
  53 EXPORT_SYMBOL(pipe_lock);
  54
  55 void pipe_unlock(struct pipe_inode_info *pipe)
  56 {
  57         if (pipe->inode)
  58                 mutex_unlock(&pipe->inode->i_mutex);
  59 }
  60 EXPORT_SYMBOL(pipe_unlock);
  61
  62 void pipe_double_lock(struct pipe_inode_info *pipe1,
  63                       struct pipe_inode_info *pipe2)
  64 {
  65         BUG_ON(pipe1 == pipe2);
  66
  67         if (pipe1 < pipe2) {
  68                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  69                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  70         } else {
  71                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  72                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  73         }
  74 }
  75
  76 /* Drop the inode semaphore and wait for a pipe event, atomically */
  77 void pipe_wait(struct pipe_inode_info *pipe)
  78 {
  79         DEFINE_WAIT(wait);
  80
  81         /*
  82          * Pipes are system-local resources, so sleeping on them
  83          * is considered a noninteractive wait:
  84          */
  85         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
  86         pipe_unlock(pipe);
  87         schedule();
  88         finish_wait(&pipe->wait, &wait);
  89         pipe_lock(pipe);
  90 }
  91
  92 static int
  93 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
  94                         int atomic)
  95 {
  96         unsigned long copy;
  97
  98         while (len > 0) {
  99                 while (!iov->iov_len)
 100                         iov++;
 101                 copy = min_t(unsigned long, len, iov->iov_len);
 102
 103                 if (atomic) {
 104                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
 105                                 return -EFAULT;
 106                 } else {
 107                         if (copy_from_user(to, iov->iov_base, copy))
 108                                 return -EFAULT;
 109                 }
 110                 to += copy;
 111                 len -= copy;
 112                 iov->iov_base += copy;
 113                 iov->iov_len -= copy;
 114         }
 115         return 0;
 116 }
 117
 118 static int
 119 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
 120                       int atomic)
 121 {
 122         unsigned long copy;
 123
 124         while (len > 0) {
 125                 while (!iov->iov_len)
 126                         iov++;
 127                 copy = min_t(unsigned long, len, iov->iov_len);
 128
 129                 if (atomic) {
 130                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
 131                                 return -EFAULT;
 132                 } else {
 133                         if (copy_to_user(iov->iov_base, from, copy))
 134                                 return -EFAULT;
 135                 }
 136                 from += copy;
 137                 len -= copy;
 138                 iov->iov_base += copy;
 139                 iov->iov_len -= copy;
 140         }
 141         return 0;
 142 }
 143
 144 /*
 145  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 146  * Returns the number of bytes not faulted in.
 147  */
 148 static int iov_fault_in_pages_write(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                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 158                         break;
 159
 160                 len -= this_len;
 161                 iov++;
 162         }
 163
 164         return len;
 165 }
 166
 167 /*
 168  * Pre-fault in the user memory, so we can use atomic copies.
 169  */
 170 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 171 {
 172         while (!iov->iov_len)
 173                 iov++;
 174
 175         while (len > 0) {
 176                 unsigned long this_len;
 177
 178                 this_len = min_t(unsigned long, len, iov->iov_len);
 179                 fault_in_pages_readable(iov->iov_base, this_len);
 180                 len -= this_len;
 181                 iov++;
 182         }
 183 }
 184
 185 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 186                                   struct pipe_buffer *buf)
 187 {
 188         struct page *page = buf->page;
 189
 190         /*
 191          * If nobody else uses this page, and we don't already have a
 192          * temporary page, let's keep track of it as a one-deep
 193          * allocation cache. (Otherwise just release our reference to it)
 194          */
 195         if (page_count(page) == 1 && !pipe->tmp_page)
 196                 pipe->tmp_page = page;
 197         else
 198                 page_cache_release(page);
 199 }
 200
 201 /**
 202  * generic_pipe_buf_map - virtually map a pipe buffer
 203  * @pipe:       the pipe that the buffer belongs to
 204  * @buf:        the buffer that should be mapped
 205  * @atomic:     whether to use an atomic map
 206  *
 207  * Description:
 208  *      This function returns a kernel virtual address mapping for the
 209  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 210  *      and the caller has to be careful not to fault before calling
 211  *      the unmap function.
 212  *
 213  *      Note that this function occupies KM_USER0 if @atomic != 0.
 214  */
 215 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 216                            struct pipe_buffer *buf, int atomic)
 217 {
 218         if (atomic) {
 219                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 220                 return kmap_atomic(buf->page, KM_USER0);
 221         }
 222
 223         return kmap(buf->page);
 224 }
 225
 226 /**
 227  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 228  * @pipe:       the pipe that the buffer belongs to
 229  * @buf:        the buffer that should be unmapped
 230  * @map_data:   the data that the mapping function returned
 231  *
 232  * Description:
 233  *      This function undoes the mapping that ->map() provided.
 234  */
 235 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 236                             struct pipe_buffer *buf, void *map_data)
 237 {
 238         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 239                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 240                 kunmap_atomic(map_data, KM_USER0);
 241         } else
 242                 kunmap(buf->page);
 243 }
 244
 245 /**
 246  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 247  * @pipe:       the pipe that the buffer belongs to
 248  * @buf:        the buffer to attempt to steal
 249  *
 250  * Description:
 251  *      This function attempts to steal the &struct page attached to
 252  *      @buf. If successful, this function returns 0 and returns with
 253  *      the page locked. The caller may then reuse the page for whatever
 254  *      he wishes; the typical use is insertion into a different file
 255  *      page cache.
 256  */
 257 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 258                            struct pipe_buffer *buf)
 259 {
 260         struct page *page = buf->page;
 261
 262         /*
 263          * A reference of one is golden, that means that the owner of this
 264          * page is the only one holding a reference to it. lock the page
 265          * and return OK.
 266          */
 267         if (page_count(page) == 1) {
 268                 lock_page(page);
 269                 return 0;
 270         }
 271
 272         return 1;
 273 }
 274
 275 /**
 276  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 277  * @pipe:       the pipe that the buffer belongs to
 278  * @buf:        the buffer to get a reference to
 279  *
 280  * Description:
 281  *      This function grabs an extra reference to @buf. It's used in
 282  *      in the tee() system call, when we duplicate the buffers in one
 283  *      pipe into another.
 284  */
 285 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 286 {
 287         page_cache_get(buf->page);
 288 }
 289
 290 /**
 291  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 292  * @info:       the pipe that the buffer belongs to
 293  * @buf:        the buffer to confirm
 294  *
 295  * Description:
 296  *      This function does nothing, because the generic pipe code uses
 297  *      pages that are always good when inserted into the pipe.
 298  */
 299 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 300                              struct pipe_buffer *buf)
 301 {
 302         return 0;
 303 }
 304
 305 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 306         .can_merge = 1,
 307         .map = generic_pipe_buf_map,
 308         .unmap = generic_pipe_buf_unmap,
 309         .confirm = generic_pipe_buf_confirm,
 310         .release = anon_pipe_buf_release,
 311         .steal = generic_pipe_buf_steal,
 312         .get = generic_pipe_buf_get,
 313 };
 314
 315 static ssize_t
 316 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 317            unsigned long nr_segs, loff_t pos)
 318 {
 319         struct file *filp = iocb->ki_filp;
 320         struct inode *inode = filp->f_path.dentry->d_inode;
 321         struct pipe_inode_info *pipe;
 322         int do_wakeup;
 323         ssize_t ret;
 324         struct iovec *iov = (struct iovec *)_iov;
 325         size_t total_len;
 326
 327         total_len = iov_length(iov, nr_segs);
 328         /* Null read succeeds. */
 329         if (unlikely(total_len == 0))
 330                 return 0;
 331
 332         do_wakeup = 0;
 333         ret = 0;
 334         mutex_lock(&inode->i_mutex);
 335         pipe = inode->i_pipe;
 336         for (;;) {
 337                 int bufs = pipe->nrbufs;
 338                 if (bufs) {
 339                         int curbuf = pipe->curbuf;
 340                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 341                         const struct pipe_buf_operations *ops = buf->ops;
 342                         void *addr;
 343                         size_t chars = buf->len;
 344                         int error, atomic;
 345
 346                         if (chars > total_len)
 347                                 chars = total_len;
 348
 349                         error = ops->confirm(pipe, buf);
 350                         if (error) {
 351                                 if (!ret)
 352                                         error = ret;
 353                                 break;
 354                         }
 355
 356                         atomic = !iov_fault_in_pages_write(iov, chars);
 357 redo:
 358                         addr = ops->map(pipe, buf, atomic);
 359                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
 360                         ops->unmap(pipe, buf, addr);
 361                         if (unlikely(error)) {
 362                                 /*
 363                                  * Just retry with the slow path if we failed.
 364                                  */
 365                                 if (atomic) {
 366                                         atomic = 0;
 367                                         goto redo;
 368                                 }
 369                                 if (!ret)
 370                                         ret = error;
 371                                 break;
 372                         }
 373                         ret += chars;
 374                         buf->offset += chars;
 375                         buf->len -= chars;
 376                         if (!buf->len) {
 377                                 buf->ops = NULL;
 378                                 ops->release(pipe, buf);
 379                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
 380                                 pipe->curbuf = curbuf;
 381                                 pipe->nrbufs = --bufs;
 382                                 do_wakeup = 1;
 383                         }
 384                         total_len -= chars;
 385                         if (!total_len)
 386                                 break;  /* common path: read succeeded */
 387                 }
 388                 if (bufs)       /* More to do? */
 389                         continue;
 390                 if (!pipe->writers)
 391                         break;
 392                 if (!pipe->waiting_writers) {
 393                         /* syscall merging: Usually we must not sleep
 394                          * if O_NONBLOCK is set, or if we got some data.
 395                          * But if a writer sleeps in kernel space, then
 396                          * we can wait for that data without violating POSIX.
 397                          */
 398                         if (ret)
 399                                 break;
 400                         if (filp->f_flags & O_NONBLOCK) {
 401                                 ret = -EAGAIN;
 402                                 break;
 403                         }
 404                 }
 405                 if (signal_pending(current)) {
 406                         if (!ret)
 407                                 ret = -ERESTARTSYS;
 408                         break;
 409                 }
 410                 if (do_wakeup) {
 411                         wake_up_interruptible_sync(&pipe->wait);
 412                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 413                 }
 414                 pipe_wait(pipe);
 415         }
 416         mutex_unlock(&inode->i_mutex);
 417
 418         /* Signal writers asynchronously that there is more room. */
 419         if (do_wakeup) {
 420                 wake_up_interruptible_sync(&pipe->wait);
 421                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 422         }
 423         if (ret > 0)
 424                 file_accessed(filp);
 425         return ret;
 426 }
 427
 428 static ssize_t
 429 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 430             unsigned long nr_segs, loff_t ppos)
 431 {
 432         struct file *filp = iocb->ki_filp;
 433         struct inode *inode = filp->f_path.dentry->d_inode;
 434         struct pipe_inode_info *pipe;
 435         ssize_t ret;
 436         int do_wakeup;
 437         struct iovec *iov = (struct iovec *)_iov;
 438         size_t total_len;
 439         ssize_t chars;
 440
 441         total_len = iov_length(iov, nr_segs);
 442         /* Null write succeeds. */
 443         if (unlikely(total_len == 0))
 444                 return 0;
 445
 446         do_wakeup = 0;
 447         ret = 0;
 448         mutex_lock(&inode->i_mutex);
 449         pipe = inode->i_pipe;
 450
 451         if (!pipe->readers) {
 452                 send_sig(SIGPIPE, current, 0);
 453                 ret = -EPIPE;
 454                 goto out;
 455         }
 456
 457         /* We try to merge small writes */
 458         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 459         if (pipe->nrbufs && chars != 0) {
 460                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 461                                                         (PIPE_BUFFERS-1);
 462                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 463                 const struct pipe_buf_operations *ops = buf->ops;
 464                 int offset = buf->offset + buf->len;
 465
 466                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 467                         int error, atomic = 1;
 468                         void *addr;
 469
 470                         error = ops->confirm(pipe, buf);
 471                         if (error)
 472                                 goto out;
 473
 474                         iov_fault_in_pages_read(iov, chars);
 475 redo1:
 476                         addr = ops->map(pipe, buf, atomic);
 477                         error = pipe_iov_copy_from_user(offset + addr, iov,
 478                                                         chars, atomic);
 479                         ops->unmap(pipe, buf, addr);
 480                         ret = error;
 481                         do_wakeup = 1;
 482                         if (error) {
 483                                 if (atomic) {
 484                                         atomic = 0;
 485                                         goto redo1;
 486                                 }
 487                                 goto out;
 488                         }
 489                         buf->len += chars;
 490                         total_len -= chars;
 491                         ret = chars;
 492                         if (!total_len)
 493                                 goto out;
 494                 }
 495         }
 496
 497         for (;;) {
 498                 int bufs;
 499
 500                 if (!pipe->readers) {
 501                         send_sig(SIGPIPE, current, 0);
 502                         if (!ret)
 503                                 ret = -EPIPE;
 504                         break;
 505                 }
 506                 bufs = pipe->nrbufs;
 507                 if (bufs < PIPE_BUFFERS) {
 508                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS-1);
 509                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 510                         struct page *page = pipe->tmp_page;
 511                         char *src;
 512                         int error, atomic = 1;
 513
 514                         if (!page) {
 515                                 page = alloc_page(GFP_HIGHUSER);
 516                                 if (unlikely(!page)) {
 517                                         ret = ret ? : -ENOMEM;
 518                                         break;
 519                                 }
 520                                 pipe->tmp_page = page;
 521                         }
 522                         /* Always wake up, even if the copy fails. Otherwise
 523                          * we lock up (O_NONBLOCK-)readers that sleep due to
 524                          * syscall merging.
 525                          * FIXME! Is this really true?
 526                          */
 527                         do_wakeup = 1;
 528                         chars = PAGE_SIZE;
 529                         if (chars > total_len)
 530                                 chars = total_len;
 531
 532                         iov_fault_in_pages_read(iov, chars);
 533 redo2:
 534                         if (atomic)
 535                                 src = kmap_atomic(page, KM_USER0);
 536                         else
 537                                 src = kmap(page);
 538
 539                         error = pipe_iov_copy_from_user(src, iov, chars,
 540                                                         atomic);
 541                         if (atomic)
 542                                 kunmap_atomic(src, KM_USER0);
 543                         else
 544                                 kunmap(page);
 545
 546                         if (unlikely(error)) {
 547                                 if (atomic) {
 548                                         atomic = 0;
 549                                         goto redo2;
 550                                 }
 551                                 if (!ret)
 552                                         ret = error;
 553                                 break;
 554                         }
 555                         ret += chars;
 556
 557                         /* Insert it into the buffer array */
 558                         buf->page = page;
 559                         buf->ops = &anon_pipe_buf_ops;
 560                         buf->offset = 0;
 561                         buf->len = chars;
 562                         pipe->nrbufs = ++bufs;
 563                         pipe->tmp_page = NULL;
 564
 565                         total_len -= chars;
 566                         if (!total_len)
 567                                 break;
 568                 }
 569                 if (bufs < PIPE_BUFFERS)
 570                         continue;
 571                 if (filp->f_flags & O_NONBLOCK) {
 572                         if (!ret)
 573                                 ret = -EAGAIN;
 574                         break;
 575                 }
 576                 if (signal_pending(current)) {
 577                         if (!ret)
 578                                 ret = -ERESTARTSYS;
 579                         break;
 580                 }
 581                 if (do_wakeup) {
 582                         wake_up_interruptible_sync(&pipe->wait);
 583                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 584                         do_wakeup = 0;
 585                 }
 586                 pipe->waiting_writers++;
 587                 pipe_wait(pipe);
 588                 pipe->waiting_writers--;
 589         }
 590 out:
 591         mutex_unlock(&inode->i_mutex);
 592         if (do_wakeup) {
 593                 wake_up_interruptible_sync(&pipe->wait);
 594                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 595         }
 596         if (ret > 0)
 597                 file_update_time(filp);
 598         return ret;
 599 }
 600
 601 static ssize_t
 602 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 603 {
 604         return -EBADF;
 605 }
 606
 607 static ssize_t
 608 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 609            loff_t *ppos)
 610 {
 611         return -EBADF;
 612 }
 613
 614 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 615 {
 616         struct inode *inode = filp->f_path.dentry->d_inode;
 617         struct pipe_inode_info *pipe;
 618         int count, buf, nrbufs;
 619
 620         switch (cmd) {
 621                 case FIONREAD:
 622                         mutex_lock(&inode->i_mutex);
 623                         pipe = inode->i_pipe;
 624                         count = 0;
 625                         buf = pipe->curbuf;
 626                         nrbufs = pipe->nrbufs;
 627                         while (--nrbufs >= 0) {
 628                                 count += pipe->bufs[buf].len;
 629                                 buf = (buf+1) & (PIPE_BUFFERS-1);
 630                         }
 631                         mutex_unlock(&inode->i_mutex);
 632
 633                         return put_user(count, (int __user *)arg);
 634                 default:
 635                         return -EINVAL;
 636         }
 637 }
 638
 639 /* No kernel lock held - fine */
 640 static unsigned int
 641 pipe_poll(struct file *filp, poll_table *wait)
 642 {
 643         unsigned int mask;
 644         struct inode *inode = filp->f_path.dentry->d_inode;
 645         struct pipe_inode_info *pipe = inode->i_pipe;
 646         int nrbufs;
 647
 648         poll_wait(filp, &pipe->wait, wait);
 649
 650         /* Reading only -- no need for acquiring the semaphore.  */
 651         nrbufs = pipe->nrbufs;
 652         mask = 0;
 653         if (filp->f_mode & FMODE_READ) {
 654                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 655                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 656                         mask |= POLLHUP;
 657         }
 658
 659         if (filp->f_mode & FMODE_WRITE) {
 660                 mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
 661                 /*
 662                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 663                  * behave exactly like pipes for poll().
 664                  */
 665                 if (!pipe->readers)
 666                         mask |= POLLERR;
 667         }
 668
 669         return mask;
 670 }
 671
 672 static int
 673 pipe_release(struct inode *inode, int decr, int decw)
 674 {
 675         struct pipe_inode_info *pipe;
 676
 677         mutex_lock(&inode->i_mutex);
 678         pipe = inode->i_pipe;
 679         pipe->readers -= decr;
 680         pipe->writers -= decw;
 681
 682         if (!pipe->readers && !pipe->writers) {
 683                 free_pipe_info(inode);
 684         } else {
 685                 wake_up_interruptible_sync(&pipe->wait);
 686                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 687                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 688         }
 689         mutex_unlock(&inode->i_mutex);
 690
 691         return 0;
 692 }
 693
 694 static int
 695 pipe_read_fasync(int fd, struct file *filp, int on)
 696 {
 697         struct inode *inode = filp->f_path.dentry->d_inode;
 698         int retval;
 699
 700         mutex_lock(&inode->i_mutex);
 701         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 702         mutex_unlock(&inode->i_mutex);
 703
 704         return retval;
 705 }
 706
 707
 708 static int
 709 pipe_write_fasync(int fd, struct file *filp, int on)
 710 {
 711         struct inode *inode = filp->f_path.dentry->d_inode;
 712         int retval;
 713
 714         mutex_lock(&inode->i_mutex);
 715         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 716         mutex_unlock(&inode->i_mutex);
 717
 718         return retval;
 719 }
 720
 721
 722 static int
 723 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 724 {
 725         struct inode *inode = filp->f_path.dentry->d_inode;
 726         struct pipe_inode_info *pipe = inode->i_pipe;
 727         int retval;
 728
 729         mutex_lock(&inode->i_mutex);
 730         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 731         if (retval >= 0) {
 732                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 733                 if (retval < 0) /* this can happen only if on == T */
 734                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 735         }
 736         mutex_unlock(&inode->i_mutex);
 737         return retval;
 738 }
 739
 740
 741 static int
 742 pipe_read_release(struct inode *inode, struct file *filp)
 743 {
 744         return pipe_release(inode, 1, 0);
 745 }
 746
 747 static int
 748 pipe_write_release(struct inode *inode, struct file *filp)
 749 {
 750         return pipe_release(inode, 0, 1);
 751 }
 752
 753 static int
 754 pipe_rdwr_release(struct inode *inode, struct file *filp)
 755 {
 756         int decr, decw;
 757
 758         decr = (filp->f_mode & FMODE_READ) != 0;
 759         decw = (filp->f_mode & FMODE_WRITE) != 0;
 760         return pipe_release(inode, decr, decw);
 761 }
 762
 763 static int
 764 pipe_read_open(struct inode *inode, struct file *filp)
 765 {
 766         /* We could have perhaps used atomic_t, but this and friends
 767            below are the only places.  So it doesn't seem worthwhile.  */
 768         mutex_lock(&inode->i_mutex);
 769         inode->i_pipe->readers++;
 770         mutex_unlock(&inode->i_mutex);
 771
 772         return 0;
 773 }
 774
 775 static int
 776 pipe_write_open(struct inode *inode, struct file *filp)
 777 {
 778         mutex_lock(&inode->i_mutex);
 779         inode->i_pipe->writers++;
 780         mutex_unlock(&inode->i_mutex);
 781
 782         return 0;
 783 }
 784
 785 static int
 786 pipe_rdwr_open(struct inode *inode, struct file *filp)
 787 {
 788         mutex_lock(&inode->i_mutex);
 789         if (filp->f_mode & FMODE_READ)
 790                 inode->i_pipe->readers++;
 791         if (filp->f_mode & FMODE_WRITE)
 792                 inode->i_pipe->writers++;
 793         mutex_unlock(&inode->i_mutex);
 794
 795         return 0;
 796 }
 797
 798 /*
 799  * The file_operations structs are not static because they
 800  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 801  *
 802  * Pipes reuse fifos' file_operations structs.
 803  */
 804 const struct file_operations read_pipefifo_fops = {
 805         .llseek         = no_llseek,
 806         .read           = do_sync_read,
 807         .aio_read       = pipe_read,
 808         .write          = bad_pipe_w,
 809         .poll           = pipe_poll,
 810         .unlocked_ioctl = pipe_ioctl,
 811         .open           = pipe_read_open,
 812         .release        = pipe_read_release,
 813         .fasync         = pipe_read_fasync,
 814 };
 815
 816 const struct file_operations write_pipefifo_fops = {
 817         .llseek         = no_llseek,
 818         .read           = bad_pipe_r,
 819         .write          = do_sync_write,
 820         .aio_write      = pipe_write,
 821         .poll           = pipe_poll,
 822         .unlocked_ioctl = pipe_ioctl,
 823         .open           = pipe_write_open,
 824         .release        = pipe_write_release,
 825         .fasync         = pipe_write_fasync,
 826 };
 827
 828 const struct file_operations rdwr_pipefifo_fops = {
 829         .llseek         = no_llseek,
 830         .read           = do_sync_read,
 831         .aio_read       = pipe_read,
 832         .write          = do_sync_write,
 833         .aio_write      = pipe_write,
 834         .poll           = pipe_poll,
 835         .unlocked_ioctl = pipe_ioctl,
 836         .open           = pipe_rdwr_open,
 837         .release        = pipe_rdwr_release,
 838         .fasync         = pipe_rdwr_fasync,
 839 };
 840
 841 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 842 {
 843         struct pipe_inode_info *pipe;
 844
 845         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 846         if (pipe) {
 847                 init_waitqueue_head(&pipe->wait);
 848                 pipe->r_counter = pipe->w_counter = 1;
 849                 pipe->inode = inode;
 850         }
 851
 852         return pipe;
 853 }
 854
 855 void __free_pipe_info(struct pipe_inode_info *pipe)
 856 {
 857         int i;
 858
 859         for (i = 0; i < PIPE_BUFFERS; i++) {
 860                 struct pipe_buffer *buf = pipe->bufs + i;
 861                 if (buf->ops)
 862                         buf->ops->release(pipe, buf);
 863         }
 864         if (pipe->tmp_page)
 865                 __free_page(pipe->tmp_page);
 866         kfree(pipe);
 867 }
 868
 869 void free_pipe_info(struct inode *inode)
 870 {
 871         __free_pipe_info(inode->i_pipe);
 872         inode->i_pipe = NULL;
 873 }
 874
 875 static struct vfsmount *pipe_mnt __read_mostly;
 876 static int pipefs_delete_dentry(struct dentry *dentry)
 877 {
 878         /*
 879          * At creation time, we pretended this dentry was hashed
 880          * (by clearing DCACHE_UNHASHED bit in d_flags)
 881          * At delete time, we restore the truth : not hashed.
 882          * (so that dput() can proceed correctly)
 883          */
 884         dentry->d_flags |= DCACHE_UNHASHED;
 885         return 0;
 886 }
 887
 888 /*
 889  * pipefs_dname() is called from d_path().
 890  */
 891 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 892 {
 893         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 894                                 dentry->d_inode->i_ino);
 895 }
 896
 897 static const struct dentry_operations pipefs_dentry_operations = {
 898         .d_delete       = pipefs_delete_dentry,
 899         .d_dname        = pipefs_dname,
 900 };
 901
 902 static struct inode * get_pipe_inode(void)
 903 {
 904         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
 905         struct pipe_inode_info *pipe;
 906
 907         if (!inode)
 908                 goto fail_inode;
 909
 910         pipe = alloc_pipe_info(inode);
 911         if (!pipe)
 912                 goto fail_iput;
 913         inode->i_pipe = pipe;
 914
 915         pipe->readers = pipe->writers = 1;
 916         inode->i_fop = &rdwr_pipefifo_fops;
 917
 918         /*
 919          * Mark the inode dirty from the very beginning,
 920          * that way it will never be moved to the dirty
 921          * list because "mark_inode_dirty()" will think
 922          * that it already _is_ on the dirty list.
 923          */
 924         inode->i_state = I_DIRTY;
 925         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 926         inode->i_uid = current_fsuid();
 927         inode->i_gid = current_fsgid();
 928         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 929
 930         return inode;
 931
 932 fail_iput:
 933         iput(inode);
 934
 935 fail_inode:
 936         return NULL;
 937 }
 938
 939 struct file *create_write_pipe(int flags)
 940 {
 941         int err;
 942         struct inode *inode;
 943         struct file *f;
 944         struct dentry *dentry;
 945         struct qstr name = { .name = "" };
 946
 947         err = -ENFILE;
 948         inode = get_pipe_inode();
 949         if (!inode)
 950                 goto err;
 951
 952         err = -ENOMEM;
 953         dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
 954         if (!dentry)
 955                 goto err_inode;
 956
 957         dentry->d_op = &pipefs_dentry_operations;
 958         /*
 959          * We dont want to publish this dentry into global dentry hash table.
 960          * We pretend dentry is already hashed, by unsetting DCACHE_UNHASHED
 961          * This permits a working /proc/$pid/fd/XXX on pipes
 962          */
 963         dentry->d_flags &= ~DCACHE_UNHASHED;
 964         d_instantiate(dentry, inode);
 965
 966         err = -ENFILE;
 967         f = alloc_file(pipe_mnt, dentry, FMODE_WRITE, &write_pipefifo_fops);
 968         if (!f)
 969                 goto err_dentry;
 970         f->f_mapping = inode->i_mapping;
 971
 972         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
 973         f->f_version = 0;
 974
 975         return f;
 976
 977  err_dentry:
 978         free_pipe_info(inode);
 979         dput(dentry);
 980         return ERR_PTR(err);
 981
 982  err_inode:
 983         free_pipe_info(inode);
 984         iput(inode);
 985  err:
 986         return ERR_PTR(err);
 987 }
 988
 989 void free_write_pipe(struct file *f)
 990 {
 991         free_pipe_info(f->f_dentry->d_inode);
 992         path_put(&f->f_path);
 993         put_filp(f);
 994 }
 995
 996 struct file *create_read_pipe(struct file *wrf, int flags)
 997 {
 998         struct file *f = get_empty_filp();
 999         if (!f)
1000                 return ERR_PTR(-ENFILE);
1001
1002         /* Grab pipe from the writer */
1003         f->f_path = wrf->f_path;
1004         path_get(&wrf->f_path);
1005         f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
1006
1007         f->f_pos = 0;
1008         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1009         f->f_op = &read_pipefifo_fops;
1010         f->f_mode = FMODE_READ;
1011         f->f_version = 0;
1012
1013         return f;
1014 }
1015
1016 int do_pipe_flags(int *fd, int flags)
1017 {
1018         struct file *fw, *fr;
1019         int error;
1020         int fdw, fdr;
1021
1022         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1023                 return -EINVAL;
1024
1025         fw = create_write_pipe(flags);
1026         if (IS_ERR(fw))
1027                 return PTR_ERR(fw);
1028         fr = create_read_pipe(fw, flags);
1029         error = PTR_ERR(fr);
1030         if (IS_ERR(fr))
1031                 goto err_write_pipe;
1032
1033         error = get_unused_fd_flags(flags);
1034         if (error < 0)
1035                 goto err_read_pipe;
1036         fdr = error;
1037
1038         error = get_unused_fd_flags(flags);
1039         if (error < 0)
1040                 goto err_fdr;
1041         fdw = error;
1042
1043         audit_fd_pair(fdr, fdw);
1044         fd_install(fdr, fr);
1045         fd_install(fdw, fw);
1046         fd[0] = fdr;
1047         fd[1] = fdw;
1048
1049         return 0;
1050
1051  err_fdr:
1052         put_unused_fd(fdr);
1053  err_read_pipe:
1054         path_put(&fr->f_path);
1055         put_filp(fr);
1056  err_write_pipe:
1057         free_write_pipe(fw);
1058         return error;
1059 }
1060
1061 /*
1062  * sys_pipe() is the normal C calling standard for creating
1063  * a pipe. It's not the way Unix traditionally does this, though.
1064  */
1065 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1066 {
1067         int fd[2];
1068         int error;
1069
1070         error = do_pipe_flags(fd, flags);
1071         if (!error) {
1072                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1073                         sys_close(fd[0]);
1074                         sys_close(fd[1]);
1075                         error = -EFAULT;
1076                 }
1077         }
1078         return error;
1079 }
1080
1081 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1082 {
1083         return sys_pipe2(fildes, 0);
1084 }
1085
1086 /*
1087  * pipefs should _never_ be mounted by userland - too much of security hassle,
1088  * no real gain from having the whole whorehouse mounted. So we don't need
1089  * any operations on the root directory. However, we need a non-trivial
1090  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1091  */
1092 static int pipefs_get_sb(struct file_system_type *fs_type,
1093                          int flags, const char *dev_name, void *data,
1094                          struct vfsmount *mnt)
1095 {
1096         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1097 }
1098
1099 static struct file_system_type pipe_fs_type = {
1100         .name           = "pipefs",
1101         .get_sb         = pipefs_get_sb,
1102         .kill_sb        = kill_anon_super,
1103 };
1104
1105 static int __init init_pipe_fs(void)
1106 {
1107         int err = register_filesystem(&pipe_fs_type);
1108
1109         if (!err) {
1110                 pipe_mnt = kern_mount(&pipe_fs_type);
1111                 if (IS_ERR(pipe_mnt)) {
1112                         err = PTR_ERR(pipe_mnt);
1113                         unregister_filesystem(&pipe_fs_type);
1114                 }
1115         }
1116         return err;
1117 }
1118
1119 static void __exit exit_pipe_fs(void)
1120 {
1121         unregister_filesystem(&pipe_fs_type);
1122         mntput(pipe_mnt);
1123 }
1124
1125 fs_initcall(init_pipe_fs);
1126 module_exit(exit_pipe_fs);