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/pipe_fs_i.h>
  17 #include <linux/uio.h>
  18 #include <linux/highmem.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/audit.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/fcntl.h>
  23
  24 #include <asm/uaccess.h>
  25 #include <asm/ioctls.h>
  26
  27 /*
  28  * The max size that a non-root user is allowed to grow the pipe. Can
  29  * be set by root in /proc/sys/fs/pipe-max-size
  30  */
  31 unsigned int pipe_max_size = 1048576;
  32
  33 /*
  34  * Minimum pipe size, as required by POSIX
  35  */
  36 unsigned int pipe_min_size = PAGE_SIZE;
  37
  38 /*
  39  * We use a start+len construction, which provides full use of the
  40  * allocated memory.
  41  * -- Florian Coosmann (FGC)
  42  *
  43  * Reads with count = 0 should always return 0.
  44  * -- Julian Bradfield 1999-06-07.
  45  *
  46  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  47  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  48  *
  49  * pipe_read & write cleanup
  50  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  51  */
  52
  53 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  54 {
  55         if (pipe->inode)
  56                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
  57 }
  58
  59 void pipe_lock(struct pipe_inode_info *pipe)
  60 {
  61         /*
  62          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  63          */
  64         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  65 }
  66 EXPORT_SYMBOL(pipe_lock);
  67
  68 void pipe_unlock(struct pipe_inode_info *pipe)
  69 {
  70         if (pipe->inode)
  71                 mutex_unlock(&pipe->inode->i_mutex);
  72 }
  73 EXPORT_SYMBOL(pipe_unlock);
  74
  75 void pipe_double_lock(struct pipe_inode_info *pipe1,
  76                       struct pipe_inode_info *pipe2)
  77 {
  78         BUG_ON(pipe1 == pipe2);
  79
  80         if (pipe1 < pipe2) {
  81                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  82                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  83         } else {
  84                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
  85                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
  86         }
  87 }
  88
  89 /* Drop the inode semaphore and wait for a pipe event, atomically */
  90 void pipe_wait(struct pipe_inode_info *pipe)
  91 {
  92         DEFINE_WAIT(wait);
  93
  94         /*
  95          * Pipes are system-local resources, so sleeping on them
  96          * is considered a noninteractive wait:
  97          */
  98         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
  99         pipe_unlock(pipe);
 100         schedule();
 101         finish_wait(&pipe->wait, &wait);
 102         pipe_lock(pipe);
 103 }
 104
 105 static int
 106 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
 107                         int atomic)
 108 {
 109         unsigned long copy;
 110
 111         while (len > 0) {
 112                 while (!iov->iov_len)
 113                         iov++;
 114                 copy = min_t(unsigned long, len, iov->iov_len);
 115
 116                 if (atomic) {
 117                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
 118                                 return -EFAULT;
 119                 } else {
 120                         if (copy_from_user(to, iov->iov_base, copy))
 121                                 return -EFAULT;
 122                 }
 123                 to += copy;
 124                 len -= copy;
 125                 iov->iov_base += copy;
 126                 iov->iov_len -= copy;
 127         }
 128         return 0;
 129 }
 130
 131 static int
 132 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
 133                       int atomic)
 134 {
 135         unsigned long copy;
 136
 137         while (len > 0) {
 138                 while (!iov->iov_len)
 139                         iov++;
 140                 copy = min_t(unsigned long, len, iov->iov_len);
 141
 142                 if (atomic) {
 143                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
 144                                 return -EFAULT;
 145                 } else {
 146                         if (copy_to_user(iov->iov_base, from, copy))
 147                                 return -EFAULT;
 148                 }
 149                 from += copy;
 150                 len -= copy;
 151                 iov->iov_base += copy;
 152                 iov->iov_len -= copy;
 153         }
 154         return 0;
 155 }
 156
 157 /*
 158  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 159  * Returns the number of bytes not faulted in.
 160  */
 161 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 162 {
 163         while (!iov->iov_len)
 164                 iov++;
 165
 166         while (len > 0) {
 167                 unsigned long this_len;
 168
 169                 this_len = min_t(unsigned long, len, iov->iov_len);
 170                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 171                         break;
 172
 173                 len -= this_len;
 174                 iov++;
 175         }
 176
 177         return len;
 178 }
 179
 180 /*
 181  * Pre-fault in the user memory, so we can use atomic copies.
 182  */
 183 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 184 {
 185         while (!iov->iov_len)
 186                 iov++;
 187
 188         while (len > 0) {
 189                 unsigned long this_len;
 190
 191                 this_len = min_t(unsigned long, len, iov->iov_len);
 192                 fault_in_pages_readable(iov->iov_base, this_len);
 193                 len -= this_len;
 194                 iov++;
 195         }
 196 }
 197
 198 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 199                                   struct pipe_buffer *buf)
 200 {
 201         struct page *page = buf->page;
 202
 203         /*
 204          * If nobody else uses this page, and we don't already have a
 205          * temporary page, let's keep track of it as a one-deep
 206          * allocation cache. (Otherwise just release our reference to it)
 207          */
 208         if (page_count(page) == 1 && !pipe->tmp_page)
 209                 pipe->tmp_page = page;
 210         else
 211                 page_cache_release(page);
 212 }
 213
 214 /**
 215  * generic_pipe_buf_map - virtually map a pipe buffer
 216  * @pipe:       the pipe that the buffer belongs to
 217  * @buf:        the buffer that should be mapped
 218  * @atomic:     whether to use an atomic map
 219  *
 220  * Description:
 221  *      This function returns a kernel virtual address mapping for the
 222  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 223  *      and the caller has to be careful not to fault before calling
 224  *      the unmap function.
 225  *
 226  *      Note that this function occupies KM_USER0 if @atomic != 0.
 227  */
 228 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 229                            struct pipe_buffer *buf, int atomic)
 230 {
 231         if (atomic) {
 232                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 233                 return kmap_atomic(buf->page, KM_USER0);
 234         }
 235
 236         return kmap(buf->page);
 237 }
 238 EXPORT_SYMBOL(generic_pipe_buf_map);
 239
 240 /**
 241  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 242  * @pipe:       the pipe that the buffer belongs to
 243  * @buf:        the buffer that should be unmapped
 244  * @map_data:   the data that the mapping function returned
 245  *
 246  * Description:
 247  *      This function undoes the mapping that ->map() provided.
 248  */
 249 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 250                             struct pipe_buffer *buf, void *map_data)
 251 {
 252         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 253                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 254                 kunmap_atomic(map_data, KM_USER0);
 255         } else
 256                 kunmap(buf->page);
 257 }
 258 EXPORT_SYMBOL(generic_pipe_buf_unmap);
 259
 260 /**
 261  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 262  * @pipe:       the pipe that the buffer belongs to
 263  * @buf:        the buffer to attempt to steal
 264  *
 265  * Description:
 266  *      This function attempts to steal the &struct page attached to
 267  *      @buf. If successful, this function returns 0 and returns with
 268  *      the page locked. The caller may then reuse the page for whatever
 269  *      he wishes; the typical use is insertion into a different file
 270  *      page cache.
 271  */
 272 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 273                            struct pipe_buffer *buf)
 274 {
 275         struct page *page = buf->page;
 276
 277         /*
 278          * A reference of one is golden, that means that the owner of this
 279          * page is the only one holding a reference to it. lock the page
 280          * and return OK.
 281          */
 282         if (page_count(page) == 1) {
 283                 lock_page(page);
 284                 return 0;
 285         }
 286
 287         return 1;
 288 }
 289 EXPORT_SYMBOL(generic_pipe_buf_steal);
 290
 291 /**
 292  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 293  * @pipe:       the pipe that the buffer belongs to
 294  * @buf:        the buffer to get a reference to
 295  *
 296  * Description:
 297  *      This function grabs an extra reference to @buf. It's used in
 298  *      in the tee() system call, when we duplicate the buffers in one
 299  *      pipe into another.
 300  */
 301 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 302 {
 303         page_cache_get(buf->page);
 304 }
 305 EXPORT_SYMBOL(generic_pipe_buf_get);
 306
 307 /**
 308  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 309  * @info:       the pipe that the buffer belongs to
 310  * @buf:        the buffer to confirm
 311  *
 312  * Description:
 313  *      This function does nothing, because the generic pipe code uses
 314  *      pages that are always good when inserted into the pipe.
 315  */
 316 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 317                              struct pipe_buffer *buf)
 318 {
 319         return 0;
 320 }
 321 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 322
 323 /**
 324  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 325  * @pipe:       the pipe that the buffer belongs to
 326  * @buf:        the buffer to put a reference to
 327  *
 328  * Description:
 329  *      This function releases a reference to @buf.
 330  */
 331 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 332                               struct pipe_buffer *buf)
 333 {
 334         page_cache_release(buf->page);
 335 }
 336 EXPORT_SYMBOL(generic_pipe_buf_release);
 337
 338 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 339         .can_merge = 1,
 340         .map = generic_pipe_buf_map,
 341         .unmap = generic_pipe_buf_unmap,
 342         .confirm = generic_pipe_buf_confirm,
 343         .release = anon_pipe_buf_release,
 344         .steal = generic_pipe_buf_steal,
 345         .get = generic_pipe_buf_get,
 346 };
 347
 348 static ssize_t
 349 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 350            unsigned long nr_segs, loff_t pos)
 351 {
 352         struct file *filp = iocb->ki_filp;
 353         struct inode *inode = filp->f_path.dentry->d_inode;
 354         struct pipe_inode_info *pipe;
 355         int do_wakeup;
 356         ssize_t ret;
 357         struct iovec *iov = (struct iovec *)_iov;
 358         size_t total_len;
 359
 360         total_len = iov_length(iov, nr_segs);
 361         /* Null read succeeds. */
 362         if (unlikely(total_len == 0))
 363                 return 0;
 364
 365         do_wakeup = 0;
 366         ret = 0;
 367         mutex_lock(&inode->i_mutex);
 368         pipe = inode->i_pipe;
 369         for (;;) {
 370                 int bufs = pipe->nrbufs;
 371                 if (bufs) {
 372                         int curbuf = pipe->curbuf;
 373                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 374                         const struct pipe_buf_operations *ops = buf->ops;
 375                         void *addr;
 376                         size_t chars = buf->len;
 377                         int error, atomic;
 378
 379                         if (chars > total_len)
 380                                 chars = total_len;
 381
 382                         error = ops->confirm(pipe, buf);
 383                         if (error) {
 384                                 if (!ret)
 385                                         error = ret;
 386                                 break;
 387                         }
 388
 389                         atomic = !iov_fault_in_pages_write(iov, chars);
 390 redo:
 391                         addr = ops->map(pipe, buf, atomic);
 392                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
 393                         ops->unmap(pipe, buf, addr);
 394                         if (unlikely(error)) {
 395                                 /*
 396                                  * Just retry with the slow path if we failed.
 397                                  */
 398                                 if (atomic) {
 399                                         atomic = 0;
 400                                         goto redo;
 401                                 }
 402                                 if (!ret)
 403                                         ret = error;
 404                                 break;
 405                         }
 406                         ret += chars;
 407                         buf->offset += chars;
 408                         buf->len -= chars;
 409                         if (!buf->len) {
 410                                 buf->ops = NULL;
 411                                 ops->release(pipe, buf);
 412                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 413                                 pipe->curbuf = curbuf;
 414                                 pipe->nrbufs = --bufs;
 415                                 do_wakeup = 1;
 416                         }
 417                         total_len -= chars;
 418                         if (!total_len)
 419                                 break;  /* common path: read succeeded */
 420                 }
 421                 if (bufs)       /* More to do? */
 422                         continue;
 423                 if (!pipe->writers)
 424                         break;
 425                 if (!pipe->waiting_writers) {
 426                         /* syscall merging: Usually we must not sleep
 427                          * if O_NONBLOCK is set, or if we got some data.
 428                          * But if a writer sleeps in kernel space, then
 429                          * we can wait for that data without violating POSIX.
 430                          */
 431                         if (ret)
 432                                 break;
 433                         if (filp->f_flags & O_NONBLOCK) {
 434                                 ret = -EAGAIN;
 435                                 break;
 436                         }
 437                 }
 438                 if (signal_pending(current)) {
 439                         if (!ret)
 440                                 ret = -ERESTARTSYS;
 441                         break;
 442                 }
 443                 if (do_wakeup) {
 444                         wake_up_interruptible_sync(&pipe->wait);
 445                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 446                 }
 447                 pipe_wait(pipe);
 448         }
 449         mutex_unlock(&inode->i_mutex);
 450
 451         /* Signal writers asynchronously that there is more room. */
 452         if (do_wakeup) {
 453                 wake_up_interruptible_sync(&pipe->wait);
 454                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 455         }
 456         if (ret > 0)
 457                 file_accessed(filp);
 458         return ret;
 459 }
 460
 461 static ssize_t
 462 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 463             unsigned long nr_segs, loff_t ppos)
 464 {
 465         struct file *filp = iocb->ki_filp;
 466         struct inode *inode = filp->f_path.dentry->d_inode;
 467         struct pipe_inode_info *pipe;
 468         ssize_t ret;
 469         int do_wakeup;
 470         struct iovec *iov = (struct iovec *)_iov;
 471         size_t total_len;
 472         ssize_t chars;
 473
 474         total_len = iov_length(iov, nr_segs);
 475         /* Null write succeeds. */
 476         if (unlikely(total_len == 0))
 477                 return 0;
 478
 479         do_wakeup = 0;
 480         ret = 0;
 481         mutex_lock(&inode->i_mutex);
 482         pipe = inode->i_pipe;
 483
 484         if (!pipe->readers) {
 485                 send_sig(SIGPIPE, current, 0);
 486                 ret = -EPIPE;
 487                 goto out;
 488         }
 489
 490         /* We try to merge small writes */
 491         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 492         if (pipe->nrbufs && chars != 0) {
 493                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 494                                                         (pipe->buffers - 1);
 495                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 496                 const struct pipe_buf_operations *ops = buf->ops;
 497                 int offset = buf->offset + buf->len;
 498
 499                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 500                         int error, atomic = 1;
 501                         void *addr;
 502
 503                         error = ops->confirm(pipe, buf);
 504                         if (error)
 505                                 goto out;
 506
 507                         iov_fault_in_pages_read(iov, chars);
 508 redo1:
 509                         addr = ops->map(pipe, buf, atomic);
 510                         error = pipe_iov_copy_from_user(offset + addr, iov,
 511                                                         chars, atomic);
 512                         ops->unmap(pipe, buf, addr);
 513                         ret = error;
 514                         do_wakeup = 1;
 515                         if (error) {
 516                                 if (atomic) {
 517                                         atomic = 0;
 518                                         goto redo1;
 519                                 }
 520                                 goto out;
 521                         }
 522                         buf->len += chars;
 523                         total_len -= chars;
 524                         ret = chars;
 525                         if (!total_len)
 526                                 goto out;
 527                 }
 528         }
 529
 530         for (;;) {
 531                 int bufs;
 532
 533                 if (!pipe->readers) {
 534                         send_sig(SIGPIPE, current, 0);
 535                         if (!ret)
 536                                 ret = -EPIPE;
 537                         break;
 538                 }
 539                 bufs = pipe->nrbufs;
 540                 if (bufs < pipe->buffers) {
 541                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 542                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 543                         struct page *page = pipe->tmp_page;
 544                         char *src;
 545                         int error, atomic = 1;
 546
 547                         if (!page) {
 548                                 page = alloc_page(GFP_HIGHUSER);
 549                                 if (unlikely(!page)) {
 550                                         ret = ret ? : -ENOMEM;
 551                                         break;
 552                                 }
 553                                 pipe->tmp_page = page;
 554                         }
 555                         /* Always wake up, even if the copy fails. Otherwise
 556                          * we lock up (O_NONBLOCK-)readers that sleep due to
 557                          * syscall merging.
 558                          * FIXME! Is this really true?
 559                          */
 560                         do_wakeup = 1;
 561                         chars = PAGE_SIZE;
 562                         if (chars > total_len)
 563                                 chars = total_len;
 564
 565                         iov_fault_in_pages_read(iov, chars);
 566 redo2:
 567                         if (atomic)
 568                                 src = kmap_atomic(page, KM_USER0);
 569                         else
 570                                 src = kmap(page);
 571
 572                         error = pipe_iov_copy_from_user(src, iov, chars,
 573                                                         atomic);
 574                         if (atomic)
 575                                 kunmap_atomic(src, KM_USER0);
 576                         else
 577                                 kunmap(page);
 578
 579                         if (unlikely(error)) {
 580                                 if (atomic) {
 581                                         atomic = 0;
 582                                         goto redo2;
 583                                 }
 584                                 if (!ret)
 585                                         ret = error;
 586                                 break;
 587                         }
 588                         ret += chars;
 589
 590                         /* Insert it into the buffer array */
 591                         buf->page = page;
 592                         buf->ops = &anon_pipe_buf_ops;
 593                         buf->offset = 0;
 594                         buf->len = chars;
 595                         pipe->nrbufs = ++bufs;
 596                         pipe->tmp_page = NULL;
 597
 598                         total_len -= chars;
 599                         if (!total_len)
 600                                 break;
 601                 }
 602                 if (bufs < pipe->buffers)
 603                         continue;
 604                 if (filp->f_flags & O_NONBLOCK) {
 605                         if (!ret)
 606                                 ret = -EAGAIN;
 607                         break;
 608                 }
 609                 if (signal_pending(current)) {
 610                         if (!ret)
 611                                 ret = -ERESTARTSYS;
 612                         break;
 613                 }
 614                 if (do_wakeup) {
 615                         wake_up_interruptible_sync(&pipe->wait);
 616                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 617                         do_wakeup = 0;
 618                 }
 619                 pipe->waiting_writers++;
 620                 pipe_wait(pipe);
 621                 pipe->waiting_writers--;
 622         }
 623 out:
 624         mutex_unlock(&inode->i_mutex);
 625         if (do_wakeup) {
 626                 wake_up_interruptible_sync(&pipe->wait);
 627                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 628         }
 629         if (ret > 0)
 630                 file_update_time(filp);
 631         return ret;
 632 }
 633
 634 static ssize_t
 635 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 636 {
 637         return -EBADF;
 638 }
 639
 640 static ssize_t
 641 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 642            loff_t *ppos)
 643 {
 644         return -EBADF;
 645 }
 646
 647 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 648 {
 649         struct inode *inode = filp->f_path.dentry->d_inode;
 650         struct pipe_inode_info *pipe;
 651         int count, buf, nrbufs;
 652
 653         switch (cmd) {
 654                 case FIONREAD:
 655                         mutex_lock(&inode->i_mutex);
 656                         pipe = inode->i_pipe;
 657                         count = 0;
 658                         buf = pipe->curbuf;
 659                         nrbufs = pipe->nrbufs;
 660                         while (--nrbufs >= 0) {
 661                                 count += pipe->bufs[buf].len;
 662                                 buf = (buf+1) & (pipe->buffers - 1);
 663                         }
 664                         mutex_unlock(&inode->i_mutex);
 665
 666                         return put_user(count, (int __user *)arg);
 667                 default:
 668                         return -EINVAL;
 669         }
 670 }
 671
 672 /* No kernel lock held - fine */
 673 static unsigned int
 674 pipe_poll(struct file *filp, poll_table *wait)
 675 {
 676         unsigned int mask;
 677         struct inode *inode = filp->f_path.dentry->d_inode;
 678         struct pipe_inode_info *pipe = inode->i_pipe;
 679         int nrbufs;
 680
 681         poll_wait(filp, &pipe->wait, wait);
 682
 683         /* Reading only -- no need for acquiring the semaphore.  */
 684         nrbufs = pipe->nrbufs;
 685         mask = 0;
 686         if (filp->f_mode & FMODE_READ) {
 687                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 688                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 689                         mask |= POLLHUP;
 690         }
 691
 692         if (filp->f_mode & FMODE_WRITE) {
 693                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 694                 /*
 695                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 696                  * behave exactly like pipes for poll().
 697                  */
 698                 if (!pipe->readers)
 699                         mask |= POLLERR;
 700         }
 701
 702         return mask;
 703 }
 704
 705 static int
 706 pipe_release(struct inode *inode, int decr, int decw)
 707 {
 708         struct pipe_inode_info *pipe;
 709
 710         mutex_lock(&inode->i_mutex);
 711         pipe = inode->i_pipe;
 712         pipe->readers -= decr;
 713         pipe->writers -= decw;
 714
 715         if (!pipe->readers && !pipe->writers) {
 716                 free_pipe_info(inode);
 717         } else {
 718                 wake_up_interruptible_sync(&pipe->wait);
 719                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 720                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 721         }
 722         mutex_unlock(&inode->i_mutex);
 723
 724         return 0;
 725 }
 726
 727 static int
 728 pipe_read_fasync(int fd, struct file *filp, int on)
 729 {
 730         struct inode *inode = filp->f_path.dentry->d_inode;
 731         int retval;
 732
 733         mutex_lock(&inode->i_mutex);
 734         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 735         mutex_unlock(&inode->i_mutex);
 736
 737         return retval;
 738 }
 739
 740
 741 static int
 742 pipe_write_fasync(int fd, struct file *filp, int on)
 743 {
 744         struct inode *inode = filp->f_path.dentry->d_inode;
 745         int retval;
 746
 747         mutex_lock(&inode->i_mutex);
 748         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 749         mutex_unlock(&inode->i_mutex);
 750
 751         return retval;
 752 }
 753
 754
 755 static int
 756 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 757 {
 758         struct inode *inode = filp->f_path.dentry->d_inode;
 759         struct pipe_inode_info *pipe = inode->i_pipe;
 760         int retval;
 761
 762         mutex_lock(&inode->i_mutex);
 763         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 764         if (retval >= 0) {
 765                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 766                 if (retval < 0) /* this can happen only if on == T */
 767                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 768         }
 769         mutex_unlock(&inode->i_mutex);
 770         return retval;
 771 }
 772
 773
 774 static int
 775 pipe_read_release(struct inode *inode, struct file *filp)
 776 {
 777         return pipe_release(inode, 1, 0);
 778 }
 779
 780 static int
 781 pipe_write_release(struct inode *inode, struct file *filp)
 782 {
 783         return pipe_release(inode, 0, 1);
 784 }
 785
 786 static int
 787 pipe_rdwr_release(struct inode *inode, struct file *filp)
 788 {
 789         int decr, decw;
 790
 791         decr = (filp->f_mode & FMODE_READ) != 0;
 792         decw = (filp->f_mode & FMODE_WRITE) != 0;
 793         return pipe_release(inode, decr, decw);
 794 }
 795
 796 static int
 797 pipe_read_open(struct inode *inode, struct file *filp)
 798 {
 799         int ret = -ENOENT;
 800
 801         mutex_lock(&inode->i_mutex);
 802
 803         if (inode->i_pipe) {
 804                 ret = 0;
 805                 inode->i_pipe->readers++;
 806         }
 807
 808         mutex_unlock(&inode->i_mutex);
 809
 810         return ret;
 811 }
 812
 813 static int
 814 pipe_write_open(struct inode *inode, struct file *filp)
 815 {
 816         int ret = -ENOENT;
 817
 818         mutex_lock(&inode->i_mutex);
 819
 820         if (inode->i_pipe) {
 821                 ret = 0;
 822                 inode->i_pipe->writers++;
 823         }
 824
 825         mutex_unlock(&inode->i_mutex);
 826
 827         return ret;
 828 }
 829
 830 static int
 831 pipe_rdwr_open(struct inode *inode, struct file *filp)
 832 {
 833         int ret = -ENOENT;
 834
 835         mutex_lock(&inode->i_mutex);
 836
 837         if (inode->i_pipe) {
 838                 ret = 0;
 839                 if (filp->f_mode & FMODE_READ)
 840                         inode->i_pipe->readers++;
 841                 if (filp->f_mode & FMODE_WRITE)
 842                         inode->i_pipe->writers++;
 843         }
 844
 845         mutex_unlock(&inode->i_mutex);
 846
 847         return ret;
 848 }
 849
 850 /*
 851  * The file_operations structs are not static because they
 852  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 853  *
 854  * Pipes reuse fifos' file_operations structs.
 855  */
 856 const struct file_operations read_pipefifo_fops = {
 857         .llseek         = no_llseek,
 858         .read           = do_sync_read,
 859         .aio_read       = pipe_read,
 860         .write          = bad_pipe_w,
 861         .poll           = pipe_poll,
 862         .unlocked_ioctl = pipe_ioctl,
 863         .open           = pipe_read_open,
 864         .release        = pipe_read_release,
 865         .fasync         = pipe_read_fasync,
 866 };
 867
 868 const struct file_operations write_pipefifo_fops = {
 869         .llseek         = no_llseek,
 870         .read           = bad_pipe_r,
 871         .write          = do_sync_write,
 872         .aio_write      = pipe_write,
 873         .poll           = pipe_poll,
 874         .unlocked_ioctl = pipe_ioctl,
 875         .open           = pipe_write_open,
 876         .release        = pipe_write_release,
 877         .fasync         = pipe_write_fasync,
 878 };
 879
 880 const struct file_operations rdwr_pipefifo_fops = {
 881         .llseek         = no_llseek,
 882         .read           = do_sync_read,
 883         .aio_read       = pipe_read,
 884         .write          = do_sync_write,
 885         .aio_write      = pipe_write,
 886         .poll           = pipe_poll,
 887         .unlocked_ioctl = pipe_ioctl,
 888         .open           = pipe_rdwr_open,
 889         .release        = pipe_rdwr_release,
 890         .fasync         = pipe_rdwr_fasync,
 891 };
 892
 893 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 894 {
 895         struct pipe_inode_info *pipe;
 896
 897         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 898         if (pipe) {
 899                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
 900                 if (pipe->bufs) {
 901                         init_waitqueue_head(&pipe->wait);
 902                         pipe->r_counter = pipe->w_counter = 1;
 903                         pipe->inode = inode;
 904                         pipe->buffers = PIPE_DEF_BUFFERS;
 905                         return pipe;
 906                 }
 907                 kfree(pipe);
 908         }
 909
 910         return NULL;
 911 }
 912
 913 void __free_pipe_info(struct pipe_inode_info *pipe)
 914 {
 915         int i;
 916
 917         for (i = 0; i < pipe->buffers; i++) {
 918                 struct pipe_buffer *buf = pipe->bufs + i;
 919                 if (buf->ops)
 920                         buf->ops->release(pipe, buf);
 921         }
 922         if (pipe->tmp_page)
 923                 __free_page(pipe->tmp_page);
 924         kfree(pipe->bufs);
 925         kfree(pipe);
 926 }
 927
 928 void free_pipe_info(struct inode *inode)
 929 {
 930         __free_pipe_info(inode->i_pipe);
 931         inode->i_pipe = NULL;
 932 }
 933
 934 static struct vfsmount *pipe_mnt __read_mostly;
 935
 936 /*
 937  * pipefs_dname() is called from d_path().
 938  */
 939 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 940 {
 941         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 942                                 dentry->d_inode->i_ino);
 943 }
 944
 945 static const struct dentry_operations pipefs_dentry_operations = {
 946         .d_dname        = pipefs_dname,
 947 };
 948
 949 static struct inode * get_pipe_inode(void)
 950 {
 951         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
 952         struct pipe_inode_info *pipe;
 953
 954         if (!inode)
 955                 goto fail_inode;
 956
 957         pipe = alloc_pipe_info(inode);
 958         if (!pipe)
 959                 goto fail_iput;
 960         inode->i_pipe = pipe;
 961
 962         pipe->readers = pipe->writers = 1;
 963         inode->i_fop = &rdwr_pipefifo_fops;
 964
 965         /*
 966          * Mark the inode dirty from the very beginning,
 967          * that way it will never be moved to the dirty
 968          * list because "mark_inode_dirty()" will think
 969          * that it already _is_ on the dirty list.
 970          */
 971         inode->i_state = I_DIRTY;
 972         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 973         inode->i_uid = current_fsuid();
 974         inode->i_gid = current_fsgid();
 975         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 976
 977         return inode;
 978
 979 fail_iput:
 980         iput(inode);
 981
 982 fail_inode:
 983         return NULL;
 984 }
 985
 986 struct file *create_write_pipe(int flags)
 987 {
 988         int err;
 989         struct inode *inode;
 990         struct file *f;
 991         struct path path;
 992         struct qstr name = { .name = "" };
 993
 994         err = -ENFILE;
 995         inode = get_pipe_inode();
 996         if (!inode)
 997                 goto err;
 998
 999         err = -ENOMEM;
1000         path.dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
1001         if (!path.dentry)
1002                 goto err_inode;
1003         path.mnt = mntget(pipe_mnt);
1004
1005         path.dentry->d_op = &pipefs_dentry_operations;
1006         d_instantiate(path.dentry, inode);
1007
1008         err = -ENFILE;
1009         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1010         if (!f)
1011                 goto err_dentry;
1012         f->f_mapping = inode->i_mapping;
1013
1014         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
1015         f->f_version = 0;
1016
1017         return f;
1018
1019  err_dentry:
1020         free_pipe_info(inode);
1021         path_put(&path);
1022         return ERR_PTR(err);
1023
1024  err_inode:
1025         free_pipe_info(inode);
1026         iput(inode);
1027  err:
1028         return ERR_PTR(err);
1029 }
1030
1031 void free_write_pipe(struct file *f)
1032 {
1033         free_pipe_info(f->f_dentry->d_inode);
1034         path_put(&f->f_path);
1035         put_filp(f);
1036 }
1037
1038 struct file *create_read_pipe(struct file *wrf, int flags)
1039 {
1040         /* Grab pipe from the writer */
1041         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1042                                     &read_pipefifo_fops);
1043         if (!f)
1044                 return ERR_PTR(-ENFILE);
1045
1046         path_get(&wrf->f_path);
1047         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1048
1049         return f;
1050 }
1051
1052 int do_pipe_flags(int *fd, int flags)
1053 {
1054         struct file *fw, *fr;
1055         int error;
1056         int fdw, fdr;
1057
1058         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1059                 return -EINVAL;
1060
1061         fw = create_write_pipe(flags);
1062         if (IS_ERR(fw))
1063                 return PTR_ERR(fw);
1064         fr = create_read_pipe(fw, flags);
1065         error = PTR_ERR(fr);
1066         if (IS_ERR(fr))
1067                 goto err_write_pipe;
1068
1069         error = get_unused_fd_flags(flags);
1070         if (error < 0)
1071                 goto err_read_pipe;
1072         fdr = error;
1073
1074         error = get_unused_fd_flags(flags);
1075         if (error < 0)
1076                 goto err_fdr;
1077         fdw = error;
1078
1079         audit_fd_pair(fdr, fdw);
1080         fd_install(fdr, fr);
1081         fd_install(fdw, fw);
1082         fd[0] = fdr;
1083         fd[1] = fdw;
1084
1085         return 0;
1086
1087  err_fdr:
1088         put_unused_fd(fdr);
1089  err_read_pipe:
1090         path_put(&fr->f_path);
1091         put_filp(fr);
1092  err_write_pipe:
1093         free_write_pipe(fw);
1094         return error;
1095 }
1096
1097 /*
1098  * sys_pipe() is the normal C calling standard for creating
1099  * a pipe. It's not the way Unix traditionally does this, though.
1100  */
1101 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1102 {
1103         int fd[2];
1104         int error;
1105
1106         error = do_pipe_flags(fd, flags);
1107         if (!error) {
1108                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1109                         sys_close(fd[0]);
1110                         sys_close(fd[1]);
1111                         error = -EFAULT;
1112                 }
1113         }
1114         return error;
1115 }
1116
1117 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1118 {
1119         return sys_pipe2(fildes, 0);
1120 }
1121
1122 /*
1123  * Allocate a new array of pipe buffers and copy the info over. Returns the
1124  * pipe size if successful, or return -ERROR on error.
1125  */
1126 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1127 {
1128         struct pipe_buffer *bufs;
1129
1130         /*
1131          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1132          * expect a lot of shrink+grow operations, just free and allocate
1133          * again like we would do for growing. If the pipe currently
1134          * contains more buffers than arg, then return busy.
1135          */
1136         if (nr_pages < pipe->nrbufs)
1137                 return -EBUSY;
1138
1139         bufs = kcalloc(nr_pages, sizeof(struct pipe_buffer), GFP_KERNEL);
1140         if (unlikely(!bufs))
1141                 return -ENOMEM;
1142
1143         /*
1144          * The pipe array wraps around, so just start the new one at zero
1145          * and adjust the indexes.
1146          */
1147         if (pipe->nrbufs) {
1148                 unsigned int tail;
1149                 unsigned int head;
1150
1151                 tail = pipe->curbuf + pipe->nrbufs;
1152                 if (tail < pipe->buffers)
1153                         tail = 0;
1154                 else
1155                         tail &= (pipe->buffers - 1);
1156
1157                 head = pipe->nrbufs - tail;
1158                 if (head)
1159                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1160                 if (tail)
1161                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1162         }
1163
1164         pipe->curbuf = 0;
1165         kfree(pipe->bufs);
1166         pipe->bufs = bufs;
1167         pipe->buffers = nr_pages;
1168         return nr_pages * PAGE_SIZE;
1169 }
1170
1171 /*
1172  * Currently we rely on the pipe array holding a power-of-2 number
1173  * of pages.
1174  */
1175 static inline unsigned int round_pipe_size(unsigned int size)
1176 {
1177         unsigned long nr_pages;
1178
1179         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1180         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1181 }
1182
1183 /*
1184  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1185  * will return an error.
1186  */
1187 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1188                  size_t *lenp, loff_t *ppos)
1189 {
1190         int ret;
1191
1192         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1193         if (ret < 0 || !write)
1194                 return ret;
1195
1196         pipe_max_size = round_pipe_size(pipe_max_size);
1197         return ret;
1198 }
1199
1200 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1201 {
1202         struct pipe_inode_info *pipe;
1203         long ret;
1204
1205         pipe = file->f_path.dentry->d_inode->i_pipe;
1206         if (!pipe)
1207                 return -EBADF;
1208
1209         mutex_lock(&pipe->inode->i_mutex);
1210
1211         switch (cmd) {
1212         case F_SETPIPE_SZ: {
1213                 unsigned int size, nr_pages;
1214
1215                 size = round_pipe_size(arg);
1216                 nr_pages = size >> PAGE_SHIFT;
1217
1218                 ret = -EINVAL;
1219                 if (!nr_pages)
1220                         goto out;
1221
1222                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1223                         ret = -EPERM;
1224                         goto out;
1225                 }
1226                 ret = pipe_set_size(pipe, nr_pages);
1227                 break;
1228                 }
1229         case F_GETPIPE_SZ:
1230                 ret = pipe->buffers * PAGE_SIZE;
1231                 break;
1232         default:
1233                 ret = -EINVAL;
1234                 break;
1235         }
1236
1237 out:
1238         mutex_unlock(&pipe->inode->i_mutex);
1239         return ret;
1240 }
1241
1242 /*
1243  * pipefs should _never_ be mounted by userland - too much of security hassle,
1244  * no real gain from having the whole whorehouse mounted. So we don't need
1245  * any operations on the root directory. However, we need a non-trivial
1246  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1247  */
1248 static int pipefs_get_sb(struct file_system_type *fs_type,
1249                          int flags, const char *dev_name, void *data,
1250                          struct vfsmount *mnt)
1251 {
1252         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1253 }
1254
1255 static struct file_system_type pipe_fs_type = {
1256         .name           = "pipefs",
1257         .get_sb         = pipefs_get_sb,
1258         .kill_sb        = kill_anon_super,
1259 };
1260
1261 static int __init init_pipe_fs(void)
1262 {
1263         int err = register_filesystem(&pipe_fs_type);
1264
1265         if (!err) {
1266                 pipe_mnt = kern_mount(&pipe_fs_type);
1267                 if (IS_ERR(pipe_mnt)) {
1268                         err = PTR_ERR(pipe_mnt);
1269                         unregister_filesystem(&pipe_fs_type);
1270                 }
1271         }
1272         return err;
1273 }
1274
1275 static void __exit exit_pipe_fs(void)
1276 {
1277         unregister_filesystem(&pipe_fs_type);
1278         mntput(pipe_mnt);
1279 }
1280
1281 fs_initcall(init_pipe_fs);
1282 module_exit(exit_pipe_fs);