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 *addr, int *offset, struct iovec *iov,
 107                         size_t *remaining, int atomic)
 108 {
 109         unsigned long copy;
 110
 111         while (*remaining > 0) {
 112                 while (!iov->iov_len)
 113                         iov++;
 114                 copy = min_t(unsigned long, *remaining, iov->iov_len);
 115
 116                 if (atomic) {
 117                         if (__copy_from_user_inatomic(addr + *offset,
 118                                                       iov->iov_base, copy))
 119                                 return -EFAULT;
 120                 } else {
 121                         if (copy_from_user(addr + *offset,
 122                                            iov->iov_base, copy))
 123                                 return -EFAULT;
 124                 }
 125                 *offset += copy;
 126                 *remaining -= copy;
 127                 iov->iov_base += copy;
 128                 iov->iov_len -= copy;
 129         }
 130         return 0;
 131 }
 132
 133 static int
 134 pipe_iov_copy_to_user(struct iovec *iov, void *addr, int *offset,
 135                       size_t *remaining, int atomic)
 136 {
 137         unsigned long copy;
 138
 139         while (*remaining > 0) {
 140                 while (!iov->iov_len)
 141                         iov++;
 142                 copy = min_t(unsigned long, *remaining, iov->iov_len);
 143
 144                 if (atomic) {
 145                         if (__copy_to_user_inatomic(iov->iov_base,
 146                                                     addr + *offset, copy))
 147                                 return -EFAULT;
 148                 } else {
 149                         if (copy_to_user(iov->iov_base,
 150                                          addr + *offset, copy))
 151                                 return -EFAULT;
 152                 }
 153                 *offset += copy;
 154                 *remaining -= copy;
 155                 iov->iov_base += copy;
 156                 iov->iov_len -= copy;
 157         }
 158         return 0;
 159 }
 160
 161 /*
 162  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 163  * Returns the number of bytes not faulted in.
 164  */
 165 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 166 {
 167         while (!iov->iov_len)
 168                 iov++;
 169
 170         while (len > 0) {
 171                 unsigned long this_len;
 172
 173                 this_len = min_t(unsigned long, len, iov->iov_len);
 174                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 175                         break;
 176
 177                 len -= this_len;
 178                 iov++;
 179         }
 180
 181         return len;
 182 }
 183
 184 /*
 185  * Pre-fault in the user memory, so we can use atomic copies.
 186  */
 187 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 188 {
 189         while (!iov->iov_len)
 190                 iov++;
 191
 192         while (len > 0) {
 193                 unsigned long this_len;
 194
 195                 this_len = min_t(unsigned long, len, iov->iov_len);
 196                 fault_in_pages_readable(iov->iov_base, this_len);
 197                 len -= this_len;
 198                 iov++;
 199         }
 200 }
 201
 202 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 203                                   struct pipe_buffer *buf)
 204 {
 205         struct page *page = buf->page;
 206
 207         /*
 208          * If nobody else uses this page, and we don't already have a
 209          * temporary page, let's keep track of it as a one-deep
 210          * allocation cache. (Otherwise just release our reference to it)
 211          */
 212         if (page_count(page) == 1 && !pipe->tmp_page)
 213                 pipe->tmp_page = page;
 214         else
 215                 page_cache_release(page);
 216 }
 217
 218 /**
 219  * generic_pipe_buf_map - virtually map a pipe buffer
 220  * @pipe:       the pipe that the buffer belongs to
 221  * @buf:        the buffer that should be mapped
 222  * @atomic:     whether to use an atomic map
 223  *
 224  * Description:
 225  *      This function returns a kernel virtual address mapping for the
 226  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 227  *      and the caller has to be careful not to fault before calling
 228  *      the unmap function.
 229  *
 230  *      Note that this function occupies KM_USER0 if @atomic != 0.
 231  */
 232 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 233                            struct pipe_buffer *buf, int atomic)
 234 {
 235         if (atomic) {
 236                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 237                 return kmap_atomic(buf->page, KM_USER0);
 238         }
 239
 240         return kmap(buf->page);
 241 }
 242 EXPORT_SYMBOL(generic_pipe_buf_map);
 243
 244 /**
 245  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 246  * @pipe:       the pipe that the buffer belongs to
 247  * @buf:        the buffer that should be unmapped
 248  * @map_data:   the data that the mapping function returned
 249  *
 250  * Description:
 251  *      This function undoes the mapping that ->map() provided.
 252  */
 253 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 254                             struct pipe_buffer *buf, void *map_data)
 255 {
 256         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 257                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 258                 kunmap_atomic(map_data, KM_USER0);
 259         } else
 260                 kunmap(buf->page);
 261 }
 262 EXPORT_SYMBOL(generic_pipe_buf_unmap);
 263
 264 /**
 265  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 266  * @pipe:       the pipe that the buffer belongs to
 267  * @buf:        the buffer to attempt to steal
 268  *
 269  * Description:
 270  *      This function attempts to steal the &struct page attached to
 271  *      @buf. If successful, this function returns 0 and returns with
 272  *      the page locked. The caller may then reuse the page for whatever
 273  *      he wishes; the typical use is insertion into a different file
 274  *      page cache.
 275  */
 276 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 277                            struct pipe_buffer *buf)
 278 {
 279         struct page *page = buf->page;
 280
 281         /*
 282          * A reference of one is golden, that means that the owner of this
 283          * page is the only one holding a reference to it. lock the page
 284          * and return OK.
 285          */
 286         if (page_count(page) == 1) {
 287                 lock_page(page);
 288                 return 0;
 289         }
 290
 291         return 1;
 292 }
 293 EXPORT_SYMBOL(generic_pipe_buf_steal);
 294
 295 /**
 296  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 297  * @pipe:       the pipe that the buffer belongs to
 298  * @buf:        the buffer to get a reference to
 299  *
 300  * Description:
 301  *      This function grabs an extra reference to @buf. It's used in
 302  *      in the tee() system call, when we duplicate the buffers in one
 303  *      pipe into another.
 304  */
 305 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 306 {
 307         page_cache_get(buf->page);
 308 }
 309 EXPORT_SYMBOL(generic_pipe_buf_get);
 310
 311 /**
 312  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 313  * @info:       the pipe that the buffer belongs to
 314  * @buf:        the buffer to confirm
 315  *
 316  * Description:
 317  *      This function does nothing, because the generic pipe code uses
 318  *      pages that are always good when inserted into the pipe.
 319  */
 320 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 321                              struct pipe_buffer *buf)
 322 {
 323         return 0;
 324 }
 325 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 326
 327 /**
 328  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 329  * @pipe:       the pipe that the buffer belongs to
 330  * @buf:        the buffer to put a reference to
 331  *
 332  * Description:
 333  *      This function releases a reference to @buf.
 334  */
 335 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 336                               struct pipe_buffer *buf)
 337 {
 338         page_cache_release(buf->page);
 339 }
 340 EXPORT_SYMBOL(generic_pipe_buf_release);
 341
 342 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 343         .can_merge = 1,
 344         .map = generic_pipe_buf_map,
 345         .unmap = generic_pipe_buf_unmap,
 346         .confirm = generic_pipe_buf_confirm,
 347         .release = anon_pipe_buf_release,
 348         .steal = generic_pipe_buf_steal,
 349         .get = generic_pipe_buf_get,
 350 };
 351
 352 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 353         .can_merge = 0,
 354         .map = generic_pipe_buf_map,
 355         .unmap = generic_pipe_buf_unmap,
 356         .confirm = generic_pipe_buf_confirm,
 357         .release = anon_pipe_buf_release,
 358         .steal = generic_pipe_buf_steal,
 359         .get = generic_pipe_buf_get,
 360 };
 361
 362 static ssize_t
 363 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 364            unsigned long nr_segs, loff_t pos)
 365 {
 366         struct file *filp = iocb->ki_filp;
 367         struct inode *inode = filp->f_path.dentry->d_inode;
 368         struct pipe_inode_info *pipe;
 369         int do_wakeup;
 370         ssize_t ret;
 371         struct iovec *iov = (struct iovec *)_iov;
 372         size_t total_len;
 373
 374         total_len = iov_length(iov, nr_segs);
 375         /* Null read succeeds. */
 376         if (unlikely(total_len == 0))
 377                 return 0;
 378
 379         do_wakeup = 0;
 380         ret = 0;
 381         mutex_lock(&inode->i_mutex);
 382         pipe = inode->i_pipe;
 383         for (;;) {
 384                 int bufs = pipe->nrbufs;
 385                 if (bufs) {
 386                         int curbuf = pipe->curbuf;
 387                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 388                         const struct pipe_buf_operations *ops = buf->ops;
 389                         void *addr;
 390                         size_t chars = buf->len, remaining;
 391                         int error, atomic;
 392
 393                         if (chars > total_len)
 394                                 chars = total_len;
 395
 396                         error = ops->confirm(pipe, buf);
 397                         if (error) {
 398                                 if (!ret)
 399                                         ret = error;
 400                                 break;
 401                         }
 402
 403                         atomic = !iov_fault_in_pages_write(iov, chars);
 404                         remaining = chars;
 405 redo:
 406                         addr = ops->map(pipe, buf, atomic);
 407                         error = pipe_iov_copy_to_user(iov, addr, &buf->offset,
 408                                                       &remaining, atomic);
 409                         ops->unmap(pipe, buf, addr);
 410                         if (unlikely(error)) {
 411                                 /*
 412                                  * Just retry with the slow path if we failed.
 413                                  */
 414                                 if (atomic) {
 415                                         atomic = 0;
 416                                         goto redo;
 417                                 }
 418                                 if (!ret)
 419                                         ret = error;
 420                                 break;
 421                         }
 422                         ret += chars;
 423                         buf->len -= chars;
 424
 425                         /* Was it a packet buffer? Clean up and exit */
 426                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 427                                 total_len = chars;
 428                                 buf->len = 0;
 429                         }
 430
 431                         if (!buf->len) {
 432                                 buf->ops = NULL;
 433                                 ops->release(pipe, buf);
 434                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 435                                 pipe->curbuf = curbuf;
 436                                 pipe->nrbufs = --bufs;
 437                                 do_wakeup = 1;
 438                         }
 439                         total_len -= chars;
 440                         if (!total_len)
 441                                 break;  /* common path: read succeeded */
 442                 }
 443                 if (bufs)       /* More to do? */
 444                         continue;
 445                 if (!pipe->writers)
 446                         break;
 447                 if (!pipe->waiting_writers) {
 448                         /* syscall merging: Usually we must not sleep
 449                          * if O_NONBLOCK is set, or if we got some data.
 450                          * But if a writer sleeps in kernel space, then
 451                          * we can wait for that data without violating POSIX.
 452                          */
 453                         if (ret)
 454                                 break;
 455                         if (filp->f_flags & O_NONBLOCK) {
 456                                 ret = -EAGAIN;
 457                                 break;
 458                         }
 459                 }
 460                 if (signal_pending(current)) {
 461                         if (!ret)
 462                                 ret = -ERESTARTSYS;
 463                         break;
 464                 }
 465                 if (do_wakeup) {
 466                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 467                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 468                 }
 469                 pipe_wait(pipe);
 470         }
 471         mutex_unlock(&inode->i_mutex);
 472
 473         /* Signal writers asynchronously that there is more room. */
 474         if (do_wakeup) {
 475                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 476                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 477         }
 478         if (ret > 0)
 479                 file_accessed(filp);
 480         return ret;
 481 }
 482
 483 static inline int is_packetized(struct file *file)
 484 {
 485         return (file->f_flags & O_DIRECT) != 0;
 486 }
 487
 488 static ssize_t
 489 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 490             unsigned long nr_segs, loff_t ppos)
 491 {
 492         struct file *filp = iocb->ki_filp;
 493         struct inode *inode = filp->f_path.dentry->d_inode;
 494         struct pipe_inode_info *pipe;
 495         ssize_t ret;
 496         int do_wakeup;
 497         struct iovec *iov = (struct iovec *)_iov;
 498         size_t total_len;
 499         ssize_t chars;
 500
 501         total_len = iov_length(iov, nr_segs);
 502         /* Null write succeeds. */
 503         if (unlikely(total_len == 0))
 504                 return 0;
 505
 506         do_wakeup = 0;
 507         ret = 0;
 508         mutex_lock(&inode->i_mutex);
 509         pipe = inode->i_pipe;
 510
 511         if (!pipe->readers) {
 512                 send_sig(SIGPIPE, current, 0);
 513                 ret = -EPIPE;
 514                 goto out;
 515         }
 516
 517         /* We try to merge small writes */
 518         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 519         if (pipe->nrbufs && chars != 0) {
 520                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 521                                                         (pipe->buffers - 1);
 522                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 523                 const struct pipe_buf_operations *ops = buf->ops;
 524                 int offset = buf->offset + buf->len;
 525
 526                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 527                         int error, atomic = 1;
 528                         void *addr;
 529                         size_t remaining = chars;
 530
 531                         error = ops->confirm(pipe, buf);
 532                         if (error)
 533                                 goto out;
 534
 535                         iov_fault_in_pages_read(iov, chars);
 536 redo1:
 537                         addr = ops->map(pipe, buf, atomic);
 538                         error = pipe_iov_copy_from_user(addr, &offset, iov,
 539                                                         &remaining, atomic);
 540                         ops->unmap(pipe, buf, addr);
 541                         ret = error;
 542                         do_wakeup = 1;
 543                         if (error) {
 544                                 if (atomic) {
 545                                         atomic = 0;
 546                                         goto redo1;
 547                                 }
 548                                 goto out;
 549                         }
 550                         buf->len += chars;
 551                         total_len -= chars;
 552                         ret = chars;
 553                         if (!total_len)
 554                                 goto out;
 555                 }
 556         }
 557
 558         for (;;) {
 559                 int bufs;
 560
 561                 if (!pipe->readers) {
 562                         send_sig(SIGPIPE, current, 0);
 563                         if (!ret)
 564                                 ret = -EPIPE;
 565                         break;
 566                 }
 567                 bufs = pipe->nrbufs;
 568                 if (bufs < pipe->buffers) {
 569                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 570                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 571                         struct page *page = pipe->tmp_page;
 572                         char *src;
 573                         int error, atomic = 1;
 574                         int offset = 0;
 575                         size_t remaining;
 576
 577                         if (!page) {
 578                                 page = alloc_page(GFP_HIGHUSER);
 579                                 if (unlikely(!page)) {
 580                                         ret = ret ? : -ENOMEM;
 581                                         break;
 582                                 }
 583                                 pipe->tmp_page = page;
 584                         }
 585                         /* Always wake up, even if the copy fails. Otherwise
 586                          * we lock up (O_NONBLOCK-)readers that sleep due to
 587                          * syscall merging.
 588                          * FIXME! Is this really true?
 589                          */
 590                         do_wakeup = 1;
 591                         chars = PAGE_SIZE;
 592                         if (chars > total_len)
 593                                 chars = total_len;
 594
 595                         iov_fault_in_pages_read(iov, chars);
 596                         remaining = chars;
 597 redo2:
 598                         if (atomic)
 599                                 src = kmap_atomic(page, KM_USER0);
 600                         else
 601                                 src = kmap(page);
 602
 603                         error = pipe_iov_copy_from_user(src, &offset, iov,
 604                                                         &remaining, atomic);
 605                         if (atomic)
 606                                 kunmap_atomic(src, KM_USER0);
 607                         else
 608                                 kunmap(page);
 609
 610                         if (unlikely(error)) {
 611                                 if (atomic) {
 612                                         atomic = 0;
 613                                         goto redo2;
 614                                 }
 615                                 if (!ret)
 616                                         ret = error;
 617                                 break;
 618                         }
 619                         ret += chars;
 620
 621                         /* Insert it into the buffer array */
 622                         buf->page = page;
 623                         buf->ops = &anon_pipe_buf_ops;
 624                         buf->offset = 0;
 625                         buf->len = chars;
 626                         buf->flags = 0;
 627                         if (is_packetized(filp)) {
 628                                 buf->ops = &packet_pipe_buf_ops;
 629                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 630                         }
 631                         pipe->nrbufs = ++bufs;
 632                         pipe->tmp_page = NULL;
 633
 634                         total_len -= chars;
 635                         if (!total_len)
 636                                 break;
 637                 }
 638                 if (bufs < pipe->buffers)
 639                         continue;
 640                 if (filp->f_flags & O_NONBLOCK) {
 641                         if (!ret)
 642                                 ret = -EAGAIN;
 643                         break;
 644                 }
 645                 if (signal_pending(current)) {
 646                         if (!ret)
 647                                 ret = -ERESTARTSYS;
 648                         break;
 649                 }
 650                 if (do_wakeup) {
 651                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 652                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 653                         do_wakeup = 0;
 654                 }
 655                 pipe->waiting_writers++;
 656                 pipe_wait(pipe);
 657                 pipe->waiting_writers--;
 658         }
 659 out:
 660         mutex_unlock(&inode->i_mutex);
 661         if (do_wakeup) {
 662                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 663                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 664         }
 665         if (ret > 0)
 666                 file_update_time(filp);
 667         return ret;
 668 }
 669
 670 static ssize_t
 671 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 672 {
 673         return -EBADF;
 674 }
 675
 676 static ssize_t
 677 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 678            loff_t *ppos)
 679 {
 680         return -EBADF;
 681 }
 682
 683 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 684 {
 685         struct inode *inode = filp->f_path.dentry->d_inode;
 686         struct pipe_inode_info *pipe;
 687         int count, buf, nrbufs;
 688
 689         switch (cmd) {
 690                 case FIONREAD:
 691                         mutex_lock(&inode->i_mutex);
 692                         pipe = inode->i_pipe;
 693                         count = 0;
 694                         buf = pipe->curbuf;
 695                         nrbufs = pipe->nrbufs;
 696                         while (--nrbufs >= 0) {
 697                                 count += pipe->bufs[buf].len;
 698                                 buf = (buf+1) & (pipe->buffers - 1);
 699                         }
 700                         mutex_unlock(&inode->i_mutex);
 701
 702                         return put_user(count, (int __user *)arg);
 703                 default:
 704                         return -EINVAL;
 705         }
 706 }
 707
 708 /* No kernel lock held - fine */
 709 static unsigned int
 710 pipe_poll(struct file *filp, poll_table *wait)
 711 {
 712         unsigned int mask;
 713         struct inode *inode = filp->f_path.dentry->d_inode;
 714         struct pipe_inode_info *pipe = inode->i_pipe;
 715         int nrbufs;
 716
 717         poll_wait(filp, &pipe->wait, wait);
 718
 719         /* Reading only -- no need for acquiring the semaphore.  */
 720         nrbufs = pipe->nrbufs;
 721         mask = 0;
 722         if (filp->f_mode & FMODE_READ) {
 723                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 724                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 725                         mask |= POLLHUP;
 726         }
 727
 728         if (filp->f_mode & FMODE_WRITE) {
 729                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 730                 /*
 731                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 732                  * behave exactly like pipes for poll().
 733                  */
 734                 if (!pipe->readers)
 735                         mask |= POLLERR;
 736         }
 737
 738         return mask;
 739 }
 740
 741 static int
 742 pipe_release(struct inode *inode, int decr, int decw)
 743 {
 744         struct pipe_inode_info *pipe;
 745
 746         mutex_lock(&inode->i_mutex);
 747         pipe = inode->i_pipe;
 748         pipe->readers -= decr;
 749         pipe->writers -= decw;
 750
 751         if (!pipe->readers && !pipe->writers) {
 752                 free_pipe_info(inode);
 753         } else {
 754                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 755                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 756                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 757         }
 758         mutex_unlock(&inode->i_mutex);
 759
 760         return 0;
 761 }
 762
 763 static int
 764 pipe_read_fasync(int fd, struct file *filp, int on)
 765 {
 766         struct inode *inode = filp->f_path.dentry->d_inode;
 767         int retval;
 768
 769         mutex_lock(&inode->i_mutex);
 770         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 771         mutex_unlock(&inode->i_mutex);
 772
 773         return retval;
 774 }
 775
 776
 777 static int
 778 pipe_write_fasync(int fd, struct file *filp, int on)
 779 {
 780         struct inode *inode = filp->f_path.dentry->d_inode;
 781         int retval;
 782
 783         mutex_lock(&inode->i_mutex);
 784         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 785         mutex_unlock(&inode->i_mutex);
 786
 787         return retval;
 788 }
 789
 790
 791 static int
 792 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 793 {
 794         struct inode *inode = filp->f_path.dentry->d_inode;
 795         struct pipe_inode_info *pipe = inode->i_pipe;
 796         int retval;
 797
 798         mutex_lock(&inode->i_mutex);
 799         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 800         if (retval >= 0) {
 801                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 802                 if (retval < 0) /* this can happen only if on == T */
 803                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 804         }
 805         mutex_unlock(&inode->i_mutex);
 806         return retval;
 807 }
 808
 809
 810 static int
 811 pipe_read_release(struct inode *inode, struct file *filp)
 812 {
 813         return pipe_release(inode, 1, 0);
 814 }
 815
 816 static int
 817 pipe_write_release(struct inode *inode, struct file *filp)
 818 {
 819         return pipe_release(inode, 0, 1);
 820 }
 821
 822 static int
 823 pipe_rdwr_release(struct inode *inode, struct file *filp)
 824 {
 825         int decr, decw;
 826
 827         decr = (filp->f_mode & FMODE_READ) != 0;
 828         decw = (filp->f_mode & FMODE_WRITE) != 0;
 829         return pipe_release(inode, decr, decw);
 830 }
 831
 832 static int
 833 pipe_read_open(struct inode *inode, struct file *filp)
 834 {
 835         int ret = -ENOENT;
 836
 837         mutex_lock(&inode->i_mutex);
 838
 839         if (inode->i_pipe) {
 840                 ret = 0;
 841                 inode->i_pipe->readers++;
 842         }
 843
 844         mutex_unlock(&inode->i_mutex);
 845
 846         return ret;
 847 }
 848
 849 static int
 850 pipe_write_open(struct inode *inode, struct file *filp)
 851 {
 852         int ret = -ENOENT;
 853
 854         mutex_lock(&inode->i_mutex);
 855
 856         if (inode->i_pipe) {
 857                 ret = 0;
 858                 inode->i_pipe->writers++;
 859         }
 860
 861         mutex_unlock(&inode->i_mutex);
 862
 863         return ret;
 864 }
 865
 866 static int
 867 pipe_rdwr_open(struct inode *inode, struct file *filp)
 868 {
 869         int ret = -ENOENT;
 870
 871         if (!(filp->f_mode & (FMODE_READ|FMODE_WRITE)))
 872                 return -EINVAL;
 873
 874         mutex_lock(&inode->i_mutex);
 875
 876         if (inode->i_pipe) {
 877                 ret = 0;
 878                 if (filp->f_mode & FMODE_READ)
 879                         inode->i_pipe->readers++;
 880                 if (filp->f_mode & FMODE_WRITE)
 881                         inode->i_pipe->writers++;
 882         }
 883
 884         mutex_unlock(&inode->i_mutex);
 885
 886         return ret;
 887 }
 888
 889 /*
 890  * The file_operations structs are not static because they
 891  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 892  *
 893  * Pipes reuse fifos' file_operations structs.
 894  */
 895 const struct file_operations read_pipefifo_fops = {
 896         .llseek         = no_llseek,
 897         .read           = do_sync_read,
 898         .aio_read       = pipe_read,
 899         .write          = bad_pipe_w,
 900         .poll           = pipe_poll,
 901         .unlocked_ioctl = pipe_ioctl,
 902         .open           = pipe_read_open,
 903         .release        = pipe_read_release,
 904         .fasync         = pipe_read_fasync,
 905 };
 906
 907 const struct file_operations write_pipefifo_fops = {
 908         .llseek         = no_llseek,
 909         .read           = bad_pipe_r,
 910         .write          = do_sync_write,
 911         .aio_write      = pipe_write,
 912         .poll           = pipe_poll,
 913         .unlocked_ioctl = pipe_ioctl,
 914         .open           = pipe_write_open,
 915         .release        = pipe_write_release,
 916         .fasync         = pipe_write_fasync,
 917 };
 918
 919 const struct file_operations rdwr_pipefifo_fops = {
 920         .llseek         = no_llseek,
 921         .read           = do_sync_read,
 922         .aio_read       = pipe_read,
 923         .write          = do_sync_write,
 924         .aio_write      = pipe_write,
 925         .poll           = pipe_poll,
 926         .unlocked_ioctl = pipe_ioctl,
 927         .open           = pipe_rdwr_open,
 928         .release        = pipe_rdwr_release,
 929         .fasync         = pipe_rdwr_fasync,
 930 };
 931
 932 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 933 {
 934         struct pipe_inode_info *pipe;
 935
 936         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 937         if (pipe) {
 938                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
 939                 if (pipe->bufs) {
 940                         init_waitqueue_head(&pipe->wait);
 941                         pipe->r_counter = pipe->w_counter = 1;
 942                         pipe->inode = inode;
 943                         pipe->buffers = PIPE_DEF_BUFFERS;
 944                         return pipe;
 945                 }
 946                 kfree(pipe);
 947         }
 948
 949         return NULL;
 950 }
 951
 952 void __free_pipe_info(struct pipe_inode_info *pipe)
 953 {
 954         int i;
 955
 956         for (i = 0; i < pipe->buffers; i++) {
 957                 struct pipe_buffer *buf = pipe->bufs + i;
 958                 if (buf->ops)
 959                         buf->ops->release(pipe, buf);
 960         }
 961         if (pipe->tmp_page)
 962                 __free_page(pipe->tmp_page);
 963         kfree(pipe->bufs);
 964         kfree(pipe);
 965 }
 966
 967 void free_pipe_info(struct inode *inode)
 968 {
 969         __free_pipe_info(inode->i_pipe);
 970         inode->i_pipe = NULL;
 971 }
 972
 973 static struct vfsmount *pipe_mnt __read_mostly;
 974
 975 /*
 976  * pipefs_dname() is called from d_path().
 977  */
 978 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 979 {
 980         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 981                                 dentry->d_inode->i_ino);
 982 }
 983
 984 static const struct dentry_operations pipefs_dentry_operations = {
 985         .d_dname        = pipefs_dname,
 986 };
 987
 988 static struct inode * get_pipe_inode(void)
 989 {
 990         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 991         struct pipe_inode_info *pipe;
 992
 993         if (!inode)
 994                 goto fail_inode;
 995
 996         inode->i_ino = get_next_ino();
 997
 998         pipe = alloc_pipe_info(inode);
 999         if (!pipe)
1000                 goto fail_iput;
1001         inode->i_pipe = pipe;
1002
1003         pipe->readers = pipe->writers = 1;
1004         inode->i_fop = &rdwr_pipefifo_fops;
1005
1006         /*
1007          * Mark the inode dirty from the very beginning,
1008          * that way it will never be moved to the dirty
1009          * list because "mark_inode_dirty()" will think
1010          * that it already _is_ on the dirty list.
1011          */
1012         inode->i_state = I_DIRTY;
1013         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1014         inode->i_uid = current_fsuid();
1015         inode->i_gid = current_fsgid();
1016         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1017
1018         return inode;
1019
1020 fail_iput:
1021         iput(inode);
1022
1023 fail_inode:
1024         return NULL;
1025 }
1026
1027 struct file *create_write_pipe(int flags)
1028 {
1029         int err;
1030         struct inode *inode;
1031         struct file *f;
1032         struct path path;
1033         struct qstr name = { .name = "" };
1034
1035         err = -ENFILE;
1036         inode = get_pipe_inode();
1037         if (!inode)
1038                 goto err;
1039
1040         err = -ENOMEM;
1041         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1042         if (!path.dentry)
1043                 goto err_inode;
1044         path.mnt = mntget(pipe_mnt);
1045
1046         d_instantiate(path.dentry, inode);
1047
1048         err = -ENFILE;
1049         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1050         if (!f)
1051                 goto err_dentry;
1052         f->f_mapping = inode->i_mapping;
1053
1054         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
1055         f->f_version = 0;
1056
1057         return f;
1058
1059  err_dentry:
1060         free_pipe_info(inode);
1061         path_put(&path);
1062         return ERR_PTR(err);
1063
1064  err_inode:
1065         free_pipe_info(inode);
1066         iput(inode);
1067  err:
1068         return ERR_PTR(err);
1069 }
1070
1071 void free_write_pipe(struct file *f)
1072 {
1073         free_pipe_info(f->f_dentry->d_inode);
1074         path_put(&f->f_path);
1075         put_filp(f);
1076 }
1077
1078 struct file *create_read_pipe(struct file *wrf, int flags)
1079 {
1080         /* Grab pipe from the writer */
1081         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1082                                     &read_pipefifo_fops);
1083         if (!f)
1084                 return ERR_PTR(-ENFILE);
1085
1086         path_get(&wrf->f_path);
1087         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1088
1089         return f;
1090 }
1091
1092 int do_pipe_flags(int *fd, int flags)
1093 {
1094         struct file *fw, *fr;
1095         int error;
1096         int fdw, fdr;
1097
1098         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
1099                 return -EINVAL;
1100
1101         fw = create_write_pipe(flags);
1102         if (IS_ERR(fw))
1103                 return PTR_ERR(fw);
1104         fr = create_read_pipe(fw, flags);
1105         error = PTR_ERR(fr);
1106         if (IS_ERR(fr))
1107                 goto err_write_pipe;
1108
1109         error = get_unused_fd_flags(flags);
1110         if (error < 0)
1111                 goto err_read_pipe;
1112         fdr = error;
1113
1114         error = get_unused_fd_flags(flags);
1115         if (error < 0)
1116                 goto err_fdr;
1117         fdw = error;
1118
1119         audit_fd_pair(fdr, fdw);
1120         fd_install(fdr, fr);
1121         fd_install(fdw, fw);
1122         fd[0] = fdr;
1123         fd[1] = fdw;
1124
1125         return 0;
1126
1127  err_fdr:
1128         put_unused_fd(fdr);
1129  err_read_pipe:
1130         path_put(&fr->f_path);
1131         put_filp(fr);
1132  err_write_pipe:
1133         free_write_pipe(fw);
1134         return error;
1135 }
1136
1137 /*
1138  * sys_pipe() is the normal C calling standard for creating
1139  * a pipe. It's not the way Unix traditionally does this, though.
1140  */
1141 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1142 {
1143         int fd[2];
1144         int error;
1145
1146         error = do_pipe_flags(fd, flags);
1147         if (!error) {
1148                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1149                         sys_close(fd[0]);
1150                         sys_close(fd[1]);
1151                         error = -EFAULT;
1152                 }
1153         }
1154         return error;
1155 }
1156
1157 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1158 {
1159         return sys_pipe2(fildes, 0);
1160 }
1161
1162 /*
1163  * Allocate a new array of pipe buffers and copy the info over. Returns the
1164  * pipe size if successful, or return -ERROR on error.
1165  */
1166 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1167 {
1168         struct pipe_buffer *bufs;
1169
1170         /*
1171          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1172          * expect a lot of shrink+grow operations, just free and allocate
1173          * again like we would do for growing. If the pipe currently
1174          * contains more buffers than arg, then return busy.
1175          */
1176         if (nr_pages < pipe->nrbufs)
1177                 return -EBUSY;
1178
1179         bufs = kcalloc(nr_pages, sizeof(struct pipe_buffer), GFP_KERNEL);
1180         if (unlikely(!bufs))
1181                 return -ENOMEM;
1182
1183         /*
1184          * The pipe array wraps around, so just start the new one at zero
1185          * and adjust the indexes.
1186          */
1187         if (pipe->nrbufs) {
1188                 unsigned int tail;
1189                 unsigned int head;
1190
1191                 tail = pipe->curbuf + pipe->nrbufs;
1192                 if (tail < pipe->buffers)
1193                         tail = 0;
1194                 else
1195                         tail &= (pipe->buffers - 1);
1196
1197                 head = pipe->nrbufs - tail;
1198                 if (head)
1199                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1200                 if (tail)
1201                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1202         }
1203
1204         pipe->curbuf = 0;
1205         kfree(pipe->bufs);
1206         pipe->bufs = bufs;
1207         pipe->buffers = nr_pages;
1208         return nr_pages * PAGE_SIZE;
1209 }
1210
1211 /*
1212  * Currently we rely on the pipe array holding a power-of-2 number
1213  * of pages.
1214  */
1215 static inline unsigned int round_pipe_size(unsigned int size)
1216 {
1217         unsigned long nr_pages;
1218
1219         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1220         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1221 }
1222
1223 /*
1224  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1225  * will return an error.
1226  */
1227 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1228                  size_t *lenp, loff_t *ppos)
1229 {
1230         int ret;
1231
1232         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1233         if (ret < 0 || !write)
1234                 return ret;
1235
1236         pipe_max_size = round_pipe_size(pipe_max_size);
1237         return ret;
1238 }
1239
1240 /*
1241  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1242  * location, so checking ->i_pipe is not enough to verify that this is a
1243  * pipe.
1244  */
1245 struct pipe_inode_info *get_pipe_info(struct file *file)
1246 {
1247         struct inode *i = file->f_path.dentry->d_inode;
1248
1249         return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1250 }
1251
1252 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1253 {
1254         struct pipe_inode_info *pipe;
1255         long ret;
1256
1257         pipe = get_pipe_info(file);
1258         if (!pipe)
1259                 return -EBADF;
1260
1261         mutex_lock(&pipe->inode->i_mutex);
1262
1263         switch (cmd) {
1264         case F_SETPIPE_SZ: {
1265                 unsigned int size, nr_pages;
1266
1267                 size = round_pipe_size(arg);
1268                 nr_pages = size >> PAGE_SHIFT;
1269
1270                 ret = -EINVAL;
1271                 if (!nr_pages)
1272                         goto out;
1273
1274                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1275                         ret = -EPERM;
1276                         goto out;
1277                 }
1278                 ret = pipe_set_size(pipe, nr_pages);
1279                 break;
1280                 }
1281         case F_GETPIPE_SZ:
1282                 ret = pipe->buffers * PAGE_SIZE;
1283                 break;
1284         default:
1285                 ret = -EINVAL;
1286                 break;
1287         }
1288
1289 out:
1290         mutex_unlock(&pipe->inode->i_mutex);
1291         return ret;
1292 }
1293
1294 static const struct super_operations pipefs_ops = {
1295         .destroy_inode = free_inode_nonrcu,
1296         .statfs = simple_statfs,
1297 };
1298
1299 /*
1300  * pipefs should _never_ be mounted by userland - too much of security hassle,
1301  * no real gain from having the whole whorehouse mounted. So we don't need
1302  * any operations on the root directory. However, we need a non-trivial
1303  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1304  */
1305 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1306                          int flags, const char *dev_name, void *data)
1307 {
1308         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1309                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1310 }
1311
1312 static struct file_system_type pipe_fs_type = {
1313         .name           = "pipefs",
1314         .mount          = pipefs_mount,
1315         .kill_sb        = kill_anon_super,
1316 };
1317
1318 static int __init init_pipe_fs(void)
1319 {
1320         int err = register_filesystem(&pipe_fs_type);
1321
1322         if (!err) {
1323                 pipe_mnt = kern_mount(&pipe_fs_type);
1324                 if (IS_ERR(pipe_mnt)) {
1325                         err = PTR_ERR(pipe_mnt);
1326                         unregister_filesystem(&pipe_fs_type);
1327                 }
1328         }
1329         return err;
1330 }
1331
1332 static void __exit exit_pipe_fs(void)
1333 {
1334         kern_unmount(pipe_mnt);
1335         unregister_filesystem(&pipe_fs_type);
1336 }
1337
1338 fs_initcall(init_pipe_fs);
1339 module_exit(exit_pipe_fs);