git.oblomov.eu Git - linux-2.6/blob - 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
  21 #include <asm/uaccess.h>
  22 #include <asm/ioctls.h>
  23
  24 /*
  25  * We use a start+len construction, which provides full use of the
  26  * allocated memory.
  27  * -- Florian Coosmann (FGC)
  28  *
  29  * Reads with count = 0 should always return 0.
  30  * -- Julian Bradfield 1999-06-07.
  31  *
  32  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  33  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  34  *
  35  * pipe_read & write cleanup
  36  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  37  */
  38
  39 /* Drop the inode semaphore and wait for a pipe event, atomically */
  40 void pipe_wait(struct pipe_inode_info *pipe)
  41 {
  42         DEFINE_WAIT(wait);
  43
  44         /*
  45          * Pipes are system-local resources, so sleeping on them
  46          * is considered a noninteractive wait:
  47          */
  48         prepare_to_wait(&pipe->wait, &wait,
  49                         TASK_INTERRUPTIBLE | TASK_NONINTERACTIVE);
  50         if (pipe->inode)
  51                 mutex_unlock(&pipe->inode->i_mutex);
  52         schedule();
  53         finish_wait(&pipe->wait, &wait);
  54         if (pipe->inode)
  55                 mutex_lock(&pipe->inode->i_mutex);
  56 }
  57
  58 static int
  59 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
  60                         int atomic)
  61 {
  62         unsigned long copy;
  63
  64         while (len > 0) {
  65                 while (!iov->iov_len)
  66                         iov++;
  67                 copy = min_t(unsigned long, len, iov->iov_len);
  68
  69                 if (atomic) {
  70                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
  71                                 return -EFAULT;
  72                 } else {
  73                         if (copy_from_user(to, iov->iov_base, copy))
  74                                 return -EFAULT;
  75                 }
  76                 to += copy;
  77                 len -= copy;
  78                 iov->iov_base += copy;
  79                 iov->iov_len -= copy;
  80         }
  81         return 0;
  82 }
  83
  84 static int
  85 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
  86                       int atomic)
  87 {
  88         unsigned long copy;
  89
  90         while (len > 0) {
  91                 while (!iov->iov_len)
  92                         iov++;
  93                 copy = min_t(unsigned long, len, iov->iov_len);
  94
  95                 if (atomic) {
  96                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
  97                                 return -EFAULT;
  98                 } else {
  99                         if (copy_to_user(iov->iov_base, from, copy))
 100                                 return -EFAULT;
 101                 }
 102                 from += copy;
 103                 len -= copy;
 104                 iov->iov_base += copy;
 105                 iov->iov_len -= copy;
 106         }
 107         return 0;
 108 }
 109
 110 /*
 111  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 112  * Returns the number of bytes not faulted in.
 113  */
 114 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 115 {
 116         while (!iov->iov_len)
 117                 iov++;
 118
 119         while (len > 0) {
 120                 unsigned long this_len;
 121
 122                 this_len = min_t(unsigned long, len, iov->iov_len);
 123                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 124                         break;
 125
 126                 len -= this_len;
 127                 iov++;
 128         }
 129
 130         return len;
 131 }
 132
 133 /*
 134  * Pre-fault in the user memory, so we can use atomic copies.
 135  */
 136 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 137 {
 138         while (!iov->iov_len)
 139                 iov++;
 140
 141         while (len > 0) {
 142                 unsigned long this_len;
 143
 144                 this_len = min_t(unsigned long, len, iov->iov_len);
 145                 fault_in_pages_readable(iov->iov_base, this_len);
 146                 len -= this_len;
 147                 iov++;
 148         }
 149 }
 150
 151 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 152                                   struct pipe_buffer *buf)
 153 {
 154         struct page *page = buf->page;
 155
 156         /*
 157          * If nobody else uses this page, and we don't already have a
 158          * temporary page, let's keep track of it as a one-deep
 159          * allocation cache. (Otherwise just release our reference to it)
 160          */
 161         if (page_count(page) == 1 && !pipe->tmp_page)
 162                 pipe->tmp_page = page;
 163         else
 164                 page_cache_release(page);
 165 }
 166
 167 /**
 168  * generic_pipe_buf_map - virtually map a pipe buffer
 169  * @pipe:       the pipe that the buffer belongs to
 170  * @buf:        the buffer that should be mapped
 171  * @atomic:     whether to use an atomic map
 172  *
 173  * Description:
 174  *      This function returns a kernel virtual address mapping for the
 175  *      passed in @pipe_buffer. If @atomic is set, an atomic map is provided
 176  *      and the caller has to be careful not to fault before calling
 177  *      the unmap function.
 178  *
 179  *      Note that this function occupies KM_USER0 if @atomic != 0.
 180  */
 181 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 182                            struct pipe_buffer *buf, int atomic)
 183 {
 184         if (atomic) {
 185                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 186                 return kmap_atomic(buf->page, KM_USER0);
 187         }
 188
 189         return kmap(buf->page);
 190 }
 191
 192 /**
 193  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 194  * @pipe:       the pipe that the buffer belongs to
 195  * @buf:        the buffer that should be unmapped
 196  * @map_data:   the data that the mapping function returned
 197  *
 198  * Description:
 199  *      This function undoes the mapping that ->map() provided.
 200  */
 201 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 202                             struct pipe_buffer *buf, void *map_data)
 203 {
 204         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 205                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 206                 kunmap_atomic(map_data, KM_USER0);
 207         } else
 208                 kunmap(buf->page);
 209 }
 210
 211 /**
 212  * generic_pipe_buf_steal - attempt to take ownership of a @pipe_buffer
 213  * @pipe:       the pipe that the buffer belongs to
 214  * @buf:        the buffer to attempt to steal
 215  *
 216  * Description:
 217  *      This function attempts to steal the @struct page attached to
 218  *      @buf. If successful, this function returns 0 and returns with
 219  *      the page locked. The caller may then reuse the page for whatever
 220  *      he wishes, the typical use is insertion into a different file
 221  *      page cache.
 222  */
 223 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 224                            struct pipe_buffer *buf)
 225 {
 226         struct page *page = buf->page;
 227
 228         /*
 229          * A reference of one is golden, that means that the owner of this
 230          * page is the only one holding a reference to it. lock the page
 231          * and return OK.
 232          */
 233         if (page_count(page) == 1) {
 234                 lock_page(page);
 235                 return 0;
 236         }
 237
 238         return 1;
 239 }
 240
 241 /**
 242  * generic_pipe_buf_get - get a reference to a @struct pipe_buffer
 243  * @pipe:       the pipe that the buffer belongs to
 244  * @buf:        the buffer to get a reference to
 245  *
 246  * Description:
 247  *      This function grabs an extra reference to @buf. It's used in
 248  *      in the tee() system call, when we duplicate the buffers in one
 249  *      pipe into another.
 250  */
 251 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 252 {
 253         page_cache_get(buf->page);
 254 }
 255
 256 /**
 257  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 258  * @info:       the pipe that the buffer belongs to
 259  * @buf:        the buffer to confirm
 260  *
 261  * Description:
 262  *      This function does nothing, because the generic pipe code uses
 263  *      pages that are always good when inserted into the pipe.
 264  */
 265 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 266                              struct pipe_buffer *buf)
 267 {
 268         return 0;
 269 }
 270
 271 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 272         .can_merge = 1,
 273         .map = generic_pipe_buf_map,
 274         .unmap = generic_pipe_buf_unmap,
 275         .confirm = generic_pipe_buf_confirm,
 276         .release = anon_pipe_buf_release,
 277         .steal = generic_pipe_buf_steal,
 278         .get = generic_pipe_buf_get,
 279 };
 280
 281 static ssize_t
 282 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 283            unsigned long nr_segs, loff_t pos)
 284 {
 285         struct file *filp = iocb->ki_filp;
 286         struct inode *inode = filp->f_path.dentry->d_inode;
 287         struct pipe_inode_info *pipe;
 288         int do_wakeup;
 289         ssize_t ret;
 290         struct iovec *iov = (struct iovec *)_iov;
 291         size_t total_len;
 292
 293         total_len = iov_length(iov, nr_segs);
 294         /* Null read succeeds. */
 295         if (unlikely(total_len == 0))
 296                 return 0;
 297
 298         do_wakeup = 0;
 299         ret = 0;
 300         mutex_lock(&inode->i_mutex);
 301         pipe = inode->i_pipe;
 302         for (;;) {
 303                 int bufs = pipe->nrbufs;
 304                 if (bufs) {
 305                         int curbuf = pipe->curbuf;
 306                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 307                         const struct pipe_buf_operations *ops = buf->ops;
 308                         void *addr;
 309                         size_t chars = buf->len;
 310                         int error, atomic;
 311
 312                         if (chars > total_len)
 313                                 chars = total_len;
 314
 315                         error = ops->confirm(pipe, buf);
 316                         if (error) {
 317                                 if (!ret)
 318                                         error = ret;
 319                                 break;
 320                         }
 321
 322                         atomic = !iov_fault_in_pages_write(iov, chars);
 323 redo:
 324                         addr = ops->map(pipe, buf, atomic);
 325                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
 326                         ops->unmap(pipe, buf, addr);
 327                         if (unlikely(error)) {
 328                                 /*
 329                                  * Just retry with the slow path if we failed.
 330                                  */
 331                                 if (atomic) {
 332                                         atomic = 0;
 333                                         goto redo;
 334                                 }
 335                                 if (!ret)
 336                                         ret = error;
 337                                 break;
 338                         }
 339                         ret += chars;
 340                         buf->offset += chars;
 341                         buf->len -= chars;
 342                         if (!buf->len) {
 343                                 buf->ops = NULL;
 344                                 ops->release(pipe, buf);
 345                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
 346                                 pipe->curbuf = curbuf;
 347                                 pipe->nrbufs = --bufs;
 348                                 do_wakeup = 1;
 349                         }
 350                         total_len -= chars;
 351                         if (!total_len)
 352                                 break;  /* common path: read succeeded */
 353                 }
 354                 if (bufs)       /* More to do? */
 355                         continue;
 356                 if (!pipe->writers)
 357                         break;
 358                 if (!pipe->waiting_writers) {
 359                         /* syscall merging: Usually we must not sleep
 360                          * if O_NONBLOCK is set, or if we got some data.
 361                          * But if a writer sleeps in kernel space, then
 362                          * we can wait for that data without violating POSIX.
 363                          */
 364                         if (ret)
 365                                 break;
 366                         if (filp->f_flags & O_NONBLOCK) {
 367                                 ret = -EAGAIN;
 368                                 break;
 369                         }
 370                 }
 371                 if (signal_pending(current)) {
 372                         if (!ret)
 373                                 ret = -ERESTARTSYS;
 374                         break;
 375                 }
 376                 if (do_wakeup) {
 377                         wake_up_interruptible_sync(&pipe->wait);
 378                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 379                 }
 380                 pipe_wait(pipe);
 381         }
 382         mutex_unlock(&inode->i_mutex);
 383
 384         /* Signal writers asynchronously that there is more room. */
 385         if (do_wakeup) {
 386                 wake_up_interruptible(&pipe->wait);
 387                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 388         }
 389         if (ret > 0)
 390                 file_accessed(filp);
 391         return ret;
 392 }
 393
 394 static ssize_t
 395 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 396             unsigned long nr_segs, loff_t ppos)
 397 {
 398         struct file *filp = iocb->ki_filp;
 399         struct inode *inode = filp->f_path.dentry->d_inode;
 400         struct pipe_inode_info *pipe;
 401         ssize_t ret;
 402         int do_wakeup;
 403         struct iovec *iov = (struct iovec *)_iov;
 404         size_t total_len;
 405         ssize_t chars;
 406
 407         total_len = iov_length(iov, nr_segs);
 408         /* Null write succeeds. */
 409         if (unlikely(total_len == 0))
 410                 return 0;
 411
 412         do_wakeup = 0;
 413         ret = 0;
 414         mutex_lock(&inode->i_mutex);
 415         pipe = inode->i_pipe;
 416
 417         if (!pipe->readers) {
 418                 send_sig(SIGPIPE, current, 0);
 419                 ret = -EPIPE;
 420                 goto out;
 421         }
 422
 423         /* We try to merge small writes */
 424         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 425         if (pipe->nrbufs && chars != 0) {
 426                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 427                                                         (PIPE_BUFFERS-1);
 428                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 429                 const struct pipe_buf_operations *ops = buf->ops;
 430                 int offset = buf->offset + buf->len;
 431
 432                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 433                         int error, atomic = 1;
 434                         void *addr;
 435
 436                         error = ops->confirm(pipe, buf);
 437                         if (error)
 438                                 goto out;
 439
 440                         iov_fault_in_pages_read(iov, chars);
 441 redo1:
 442                         addr = ops->map(pipe, buf, atomic);
 443                         error = pipe_iov_copy_from_user(offset + addr, iov,
 444                                                         chars, atomic);
 445                         ops->unmap(pipe, buf, addr);
 446                         ret = error;
 447                         do_wakeup = 1;
 448                         if (error) {
 449                                 if (atomic) {
 450                                         atomic = 0;
 451                                         goto redo1;
 452                                 }
 453                                 goto out;
 454                         }
 455                         buf->len += chars;
 456                         total_len -= chars;
 457                         ret = chars;
 458                         if (!total_len)
 459                                 goto out;
 460                 }
 461         }
 462
 463         for (;;) {
 464                 int bufs;
 465
 466                 if (!pipe->readers) {
 467                         send_sig(SIGPIPE, current, 0);
 468                         if (!ret)
 469                                 ret = -EPIPE;
 470                         break;
 471                 }
 472                 bufs = pipe->nrbufs;
 473                 if (bufs < PIPE_BUFFERS) {
 474                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS-1);
 475                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 476                         struct page *page = pipe->tmp_page;
 477                         char *src;
 478                         int error, atomic = 1;
 479
 480                         if (!page) {
 481                                 page = alloc_page(GFP_HIGHUSER);
 482                                 if (unlikely(!page)) {
 483                                         ret = ret ? : -ENOMEM;
 484                                         break;
 485                                 }
 486                                 pipe->tmp_page = page;
 487                         }
 488                         /* Always wake up, even if the copy fails. Otherwise
 489                          * we lock up (O_NONBLOCK-)readers that sleep due to
 490                          * syscall merging.
 491                          * FIXME! Is this really true?
 492                          */
 493                         do_wakeup = 1;
 494                         chars = PAGE_SIZE;
 495                         if (chars > total_len)
 496                                 chars = total_len;
 497
 498                         iov_fault_in_pages_read(iov, chars);
 499 redo2:
 500                         if (atomic)
 501                                 src = kmap_atomic(page, KM_USER0);
 502                         else
 503                                 src = kmap(page);
 504
 505                         error = pipe_iov_copy_from_user(src, iov, chars,
 506                                                         atomic);
 507                         if (atomic)
 508                                 kunmap_atomic(src, KM_USER0);
 509                         else
 510                                 kunmap(page);
 511
 512                         if (unlikely(error)) {
 513                                 if (atomic) {
 514                                         atomic = 0;
 515                                         goto redo2;
 516                                 }
 517                                 if (!ret)
 518                                         ret = error;
 519                                 break;
 520                         }
 521                         ret += chars;
 522
 523                         /* Insert it into the buffer array */
 524                         buf->page = page;
 525                         buf->ops = &anon_pipe_buf_ops;
 526                         buf->offset = 0;
 527                         buf->len = chars;
 528                         pipe->nrbufs = ++bufs;
 529                         pipe->tmp_page = NULL;
 530
 531                         total_len -= chars;
 532                         if (!total_len)
 533                                 break;
 534                 }
 535                 if (bufs < PIPE_BUFFERS)
 536                         continue;
 537                 if (filp->f_flags & O_NONBLOCK) {
 538                         if (!ret)
 539                                 ret = -EAGAIN;
 540                         break;
 541                 }
 542                 if (signal_pending(current)) {
 543                         if (!ret)
 544                                 ret = -ERESTARTSYS;
 545                         break;
 546                 }
 547                 if (do_wakeup) {
 548                         wake_up_interruptible_sync(&pipe->wait);
 549                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 550                         do_wakeup = 0;
 551                 }
 552                 pipe->waiting_writers++;
 553                 pipe_wait(pipe);
 554                 pipe->waiting_writers--;
 555         }
 556 out:
 557         mutex_unlock(&inode->i_mutex);
 558         if (do_wakeup) {
 559                 wake_up_interruptible(&pipe->wait);
 560                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 561         }
 562         if (ret > 0)
 563                 file_update_time(filp);
 564         return ret;
 565 }
 566
 567 static ssize_t
 568 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 569 {
 570         return -EBADF;
 571 }
 572
 573 static ssize_t
 574 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 575            loff_t *ppos)
 576 {
 577         return -EBADF;
 578 }
 579
 580 static int
 581 pipe_ioctl(struct inode *pino, struct file *filp,
 582            unsigned int cmd, unsigned long arg)
 583 {
 584         struct inode *inode = filp->f_path.dentry->d_inode;
 585         struct pipe_inode_info *pipe;
 586         int count, buf, nrbufs;
 587
 588         switch (cmd) {
 589                 case FIONREAD:
 590                         mutex_lock(&inode->i_mutex);
 591                         pipe = inode->i_pipe;
 592                         count = 0;
 593                         buf = pipe->curbuf;
 594                         nrbufs = pipe->nrbufs;
 595                         while (--nrbufs >= 0) {
 596                                 count += pipe->bufs[buf].len;
 597                                 buf = (buf+1) & (PIPE_BUFFERS-1);
 598                         }
 599                         mutex_unlock(&inode->i_mutex);
 600
 601                         return put_user(count, (int __user *)arg);
 602                 default:
 603                         return -EINVAL;
 604         }
 605 }
 606
 607 /* No kernel lock held - fine */
 608 static unsigned int
 609 pipe_poll(struct file *filp, poll_table *wait)
 610 {
 611         unsigned int mask;
 612         struct inode *inode = filp->f_path.dentry->d_inode;
 613         struct pipe_inode_info *pipe = inode->i_pipe;
 614         int nrbufs;
 615
 616         poll_wait(filp, &pipe->wait, wait);
 617
 618         /* Reading only -- no need for acquiring the semaphore.  */
 619         nrbufs = pipe->nrbufs;
 620         mask = 0;
 621         if (filp->f_mode & FMODE_READ) {
 622                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 623                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 624                         mask |= POLLHUP;
 625         }
 626
 627         if (filp->f_mode & FMODE_WRITE) {
 628                 mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
 629                 /*
 630                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 631                  * behave exactly like pipes for poll().
 632                  */
 633                 if (!pipe->readers)
 634                         mask |= POLLERR;
 635         }
 636
 637         return mask;
 638 }
 639
 640 static int
 641 pipe_release(struct inode *inode, int decr, int decw)
 642 {
 643         struct pipe_inode_info *pipe;
 644
 645         mutex_lock(&inode->i_mutex);
 646         pipe = inode->i_pipe;
 647         pipe->readers -= decr;
 648         pipe->writers -= decw;
 649
 650         if (!pipe->readers && !pipe->writers) {
 651                 free_pipe_info(inode);
 652         } else {
 653                 wake_up_interruptible(&pipe->wait);
 654                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 655                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 656         }
 657         mutex_unlock(&inode->i_mutex);
 658
 659         return 0;
 660 }
 661
 662 static int
 663 pipe_read_fasync(int fd, struct file *filp, int on)
 664 {
 665         struct inode *inode = filp->f_path.dentry->d_inode;
 666         int retval;
 667
 668         mutex_lock(&inode->i_mutex);
 669         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 670         mutex_unlock(&inode->i_mutex);
 671
 672         if (retval < 0)
 673                 return retval;
 674
 675         return 0;
 676 }
 677
 678
 679 static int
 680 pipe_write_fasync(int fd, struct file *filp, int on)
 681 {
 682         struct inode *inode = filp->f_path.dentry->d_inode;
 683         int retval;
 684
 685         mutex_lock(&inode->i_mutex);
 686         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 687         mutex_unlock(&inode->i_mutex);
 688
 689         if (retval < 0)
 690                 return retval;
 691
 692         return 0;
 693 }
 694
 695
 696 static int
 697 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 698 {
 699         struct inode *inode = filp->f_path.dentry->d_inode;
 700         struct pipe_inode_info *pipe = inode->i_pipe;
 701         int retval;
 702
 703         mutex_lock(&inode->i_mutex);
 704
 705         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 706
 707         if (retval >= 0)
 708                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 709
 710         mutex_unlock(&inode->i_mutex);
 711
 712         if (retval < 0)
 713                 return retval;
 714
 715         return 0;
 716 }
 717
 718
 719 static int
 720 pipe_read_release(struct inode *inode, struct file *filp)
 721 {
 722         pipe_read_fasync(-1, filp, 0);
 723         return pipe_release(inode, 1, 0);
 724 }
 725
 726 static int
 727 pipe_write_release(struct inode *inode, struct file *filp)
 728 {
 729         pipe_write_fasync(-1, filp, 0);
 730         return pipe_release(inode, 0, 1);
 731 }
 732
 733 static int
 734 pipe_rdwr_release(struct inode *inode, struct file *filp)
 735 {
 736         int decr, decw;
 737
 738         pipe_rdwr_fasync(-1, filp, 0);
 739         decr = (filp->f_mode & FMODE_READ) != 0;
 740         decw = (filp->f_mode & FMODE_WRITE) != 0;
 741         return pipe_release(inode, decr, decw);
 742 }
 743
 744 static int
 745 pipe_read_open(struct inode *inode, struct file *filp)
 746 {
 747         /* We could have perhaps used atomic_t, but this and friends
 748            below are the only places.  So it doesn't seem worthwhile.  */
 749         mutex_lock(&inode->i_mutex);
 750         inode->i_pipe->readers++;
 751         mutex_unlock(&inode->i_mutex);
 752
 753         return 0;
 754 }
 755
 756 static int
 757 pipe_write_open(struct inode *inode, struct file *filp)
 758 {
 759         mutex_lock(&inode->i_mutex);
 760         inode->i_pipe->writers++;
 761         mutex_unlock(&inode->i_mutex);
 762
 763         return 0;
 764 }
 765
 766 static int
 767 pipe_rdwr_open(struct inode *inode, struct file *filp)
 768 {
 769         mutex_lock(&inode->i_mutex);
 770         if (filp->f_mode & FMODE_READ)
 771                 inode->i_pipe->readers++;
 772         if (filp->f_mode & FMODE_WRITE)
 773                 inode->i_pipe->writers++;
 774         mutex_unlock(&inode->i_mutex);
 775
 776         return 0;
 777 }
 778
 779 /*
 780  * The file_operations structs are not static because they
 781  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 782  */
 783 const struct file_operations read_fifo_fops = {
 784         .llseek         = no_llseek,
 785         .read           = do_sync_read,
 786         .aio_read       = pipe_read,
 787         .write          = bad_pipe_w,
 788         .poll           = pipe_poll,
 789         .ioctl          = pipe_ioctl,
 790         .open           = pipe_read_open,
 791         .release        = pipe_read_release,
 792         .fasync         = pipe_read_fasync,
 793 };
 794
 795 const struct file_operations write_fifo_fops = {
 796         .llseek         = no_llseek,
 797         .read           = bad_pipe_r,
 798         .write          = do_sync_write,
 799         .aio_write      = pipe_write,
 800         .poll           = pipe_poll,
 801         .ioctl          = pipe_ioctl,
 802         .open           = pipe_write_open,
 803         .release        = pipe_write_release,
 804         .fasync         = pipe_write_fasync,
 805 };
 806
 807 const struct file_operations rdwr_fifo_fops = {
 808         .llseek         = no_llseek,
 809         .read           = do_sync_read,
 810         .aio_read       = pipe_read,
 811         .write          = do_sync_write,
 812         .aio_write      = pipe_write,
 813         .poll           = pipe_poll,
 814         .ioctl          = pipe_ioctl,
 815         .open           = pipe_rdwr_open,
 816         .release        = pipe_rdwr_release,
 817         .fasync         = pipe_rdwr_fasync,
 818 };
 819
 820 static const struct file_operations read_pipe_fops = {
 821         .llseek         = no_llseek,
 822         .read           = do_sync_read,
 823         .aio_read       = pipe_read,
 824         .write          = bad_pipe_w,
 825         .poll           = pipe_poll,
 826         .ioctl          = pipe_ioctl,
 827         .open           = pipe_read_open,
 828         .release        = pipe_read_release,
 829         .fasync         = pipe_read_fasync,
 830 };
 831
 832 static const struct file_operations write_pipe_fops = {
 833         .llseek         = no_llseek,
 834         .read           = bad_pipe_r,
 835         .write          = do_sync_write,
 836         .aio_write      = pipe_write,
 837         .poll           = pipe_poll,
 838         .ioctl          = pipe_ioctl,
 839         .open           = pipe_write_open,
 840         .release        = pipe_write_release,
 841         .fasync         = pipe_write_fasync,
 842 };
 843
 844 static const struct file_operations rdwr_pipe_fops = {
 845         .llseek         = no_llseek,
 846         .read           = do_sync_read,
 847         .aio_read       = pipe_read,
 848         .write          = do_sync_write,
 849         .aio_write      = pipe_write,
 850         .poll           = pipe_poll,
 851         .ioctl          = pipe_ioctl,
 852         .open           = pipe_rdwr_open,
 853         .release        = pipe_rdwr_release,
 854         .fasync         = pipe_rdwr_fasync,
 855 };
 856
 857 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 858 {
 859         struct pipe_inode_info *pipe;
 860
 861         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 862         if (pipe) {
 863                 init_waitqueue_head(&pipe->wait);
 864                 pipe->r_counter = pipe->w_counter = 1;
 865                 pipe->inode = inode;
 866         }
 867
 868         return pipe;
 869 }
 870
 871 void __free_pipe_info(struct pipe_inode_info *pipe)
 872 {
 873         int i;
 874
 875         for (i = 0; i < PIPE_BUFFERS; i++) {
 876                 struct pipe_buffer *buf = pipe->bufs + i;
 877                 if (buf->ops)
 878                         buf->ops->release(pipe, buf);
 879         }
 880         if (pipe->tmp_page)
 881                 __free_page(pipe->tmp_page);
 882         kfree(pipe);
 883 }
 884
 885 void free_pipe_info(struct inode *inode)
 886 {
 887         __free_pipe_info(inode->i_pipe);
 888         inode->i_pipe = NULL;
 889 }
 890
 891 static struct vfsmount *pipe_mnt __read_mostly;
 892 static int pipefs_delete_dentry(struct dentry *dentry)
 893 {
 894         /*
 895          * At creation time, we pretended this dentry was hashed
 896          * (by clearing DCACHE_UNHASHED bit in d_flags)
 897          * At delete time, we restore the truth : not hashed.
 898          * (so that dput() can proceed correctly)
 899          */
 900         dentry->d_flags |= DCACHE_UNHASHED;
 901         return 0;
 902 }
 903
 904 /*
 905  * pipefs_dname() is called from d_path().
 906  */
 907 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 908 {
 909         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 910                                 dentry->d_inode->i_ino);
 911 }
 912
 913 static struct dentry_operations pipefs_dentry_operations = {
 914         .d_delete       = pipefs_delete_dentry,
 915         .d_dname        = pipefs_dname,
 916 };
 917
 918 static struct inode * get_pipe_inode(void)
 919 {
 920         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
 921         struct pipe_inode_info *pipe;
 922
 923         if (!inode)
 924                 goto fail_inode;
 925
 926         pipe = alloc_pipe_info(inode);
 927         if (!pipe)
 928                 goto fail_iput;
 929         inode->i_pipe = pipe;
 930
 931         pipe->readers = pipe->writers = 1;
 932         inode->i_fop = &rdwr_pipe_fops;
 933
 934         /*
 935          * Mark the inode dirty from the very beginning,
 936          * that way it will never be moved to the dirty
 937          * list because "mark_inode_dirty()" will think
 938          * that it already _is_ on the dirty list.
 939          */
 940         inode->i_state = I_DIRTY;
 941         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 942         inode->i_uid = current->fsuid;
 943         inode->i_gid = current->fsgid;
 944         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 945
 946         return inode;
 947
 948 fail_iput:
 949         iput(inode);
 950
 951 fail_inode:
 952         return NULL;
 953 }
 954
 955 struct file *create_write_pipe(void)
 956 {
 957         int err;
 958         struct inode *inode;
 959         struct file *f;
 960         struct dentry *dentry;
 961         struct qstr name = { .name = "" };
 962
 963         f = get_empty_filp();
 964         if (!f)
 965                 return ERR_PTR(-ENFILE);
 966         err = -ENFILE;
 967         inode = get_pipe_inode();
 968         if (!inode)
 969                 goto err_file;
 970
 971         err = -ENOMEM;
 972         dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
 973         if (!dentry)
 974                 goto err_inode;
 975
 976         dentry->d_op = &pipefs_dentry_operations;
 977         /*
 978          * We dont want to publish this dentry into global dentry hash table.
 979          * We pretend dentry is already hashed, by unsetting DCACHE_UNHASHED
 980          * This permits a working /proc/$pid/fd/XXX on pipes
 981          */
 982         dentry->d_flags &= ~DCACHE_UNHASHED;
 983         d_instantiate(dentry, inode);
 984         f->f_path.mnt = mntget(pipe_mnt);
 985         f->f_path.dentry = dentry;
 986         f->f_mapping = inode->i_mapping;
 987
 988         f->f_flags = O_WRONLY;
 989         f->f_op = &write_pipe_fops;
 990         f->f_mode = FMODE_WRITE;
 991         f->f_version = 0;
 992
 993         return f;
 994
 995  err_inode:
 996         free_pipe_info(inode);
 997         iput(inode);
 998  err_file:
 999         put_filp(f);
1000         return ERR_PTR(err);
1001 }
1002
1003 void free_write_pipe(struct file *f)
1004 {
1005         free_pipe_info(f->f_dentry->d_inode);
1006         dput(f->f_path.dentry);
1007         mntput(f->f_path.mnt);
1008         put_filp(f);
1009 }
1010
1011 struct file *create_read_pipe(struct file *wrf)
1012 {
1013         struct file *f = get_empty_filp();
1014         if (!f)
1015                 return ERR_PTR(-ENFILE);
1016
1017         /* Grab pipe from the writer */
1018         f->f_path.mnt = mntget(wrf->f_path.mnt);
1019         f->f_path.dentry = dget(wrf->f_path.dentry);
1020         f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
1021
1022         f->f_pos = 0;
1023         f->f_flags = O_RDONLY;
1024         f->f_op = &read_pipe_fops;
1025         f->f_mode = FMODE_READ;
1026         f->f_version = 0;
1027
1028         return f;
1029 }
1030
1031 int do_pipe(int *fd)
1032 {
1033         struct file *fw, *fr;
1034         int error;
1035         int fdw, fdr;
1036
1037         fw = create_write_pipe();
1038         if (IS_ERR(fw))
1039                 return PTR_ERR(fw);
1040         fr = create_read_pipe(fw);
1041         error = PTR_ERR(fr);
1042         if (IS_ERR(fr))
1043                 goto err_write_pipe;
1044
1045         error = get_unused_fd();
1046         if (error < 0)
1047                 goto err_read_pipe;
1048         fdr = error;
1049
1050         error = get_unused_fd();
1051         if (error < 0)
1052                 goto err_fdr;
1053         fdw = error;
1054
1055         error = audit_fd_pair(fdr, fdw);
1056         if (error < 0)
1057                 goto err_fdw;
1058
1059         fd_install(fdr, fr);
1060         fd_install(fdw, fw);
1061         fd[0] = fdr;
1062         fd[1] = fdw;
1063
1064         return 0;
1065
1066  err_fdw:
1067         put_unused_fd(fdw);
1068  err_fdr:
1069         put_unused_fd(fdr);
1070  err_read_pipe:
1071         dput(fr->f_dentry);
1072         mntput(fr->f_vfsmnt);
1073         put_filp(fr);
1074  err_write_pipe:
1075         free_write_pipe(fw);
1076         return error;
1077 }
1078
1079 /*
1080  * pipefs should _never_ be mounted by userland - too much of security hassle,
1081  * no real gain from having the whole whorehouse mounted. So we don't need
1082  * any operations on the root directory. However, we need a non-trivial
1083  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1084  */
1085 static int pipefs_get_sb(struct file_system_type *fs_type,
1086                          int flags, const char *dev_name, void *data,
1087                          struct vfsmount *mnt)
1088 {
1089         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1090 }
1091
1092 static struct file_system_type pipe_fs_type = {
1093         .name           = "pipefs",
1094         .get_sb         = pipefs_get_sb,
1095         .kill_sb        = kill_anon_super,
1096 };
1097
1098 static int __init init_pipe_fs(void)
1099 {
1100         int err = register_filesystem(&pipe_fs_type);
1101
1102         if (!err) {
1103                 pipe_mnt = kern_mount(&pipe_fs_type);
1104                 if (IS_ERR(pipe_mnt)) {
1105                         err = PTR_ERR(pipe_mnt);
1106                         unregister_filesystem(&pipe_fs_type);
1107                 }
1108         }
1109         return err;
1110 }
1111
1112 static void __exit exit_pipe_fs(void)
1113 {
1114         unregister_filesystem(&pipe_fs_type);
1115         mntput(pipe_mnt);
1116 }
1117
1118 fs_initcall(init_pipe_fs);
1119 module_exit(exit_pipe_fs);