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