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/magic.h>
  17 #include <linux/pipe_fs_i.h>
  18 #include <linux/uio.h>
  19 #include <linux/highmem.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/audit.h>
  22 #include <linux/syscalls.h>
  23 #include <linux/fcntl.h>
  24
  25 #include <asm/uaccess.h>
  26 #include <asm/ioctls.h>
  27
  28 /*
  29  * The max size that a non-root user is allowed to grow the pipe. Can
  30  * be set by root in /proc/sys/fs/pipe-max-size
  31  */
  32 unsigned int pipe_max_size = 1048576;
  33
  34 /*
  35  * Minimum pipe size, as required by POSIX
  36  */
  37 unsigned int pipe_min_size = PAGE_SIZE;
  38
  39 /*
  40  * We use a start+len construction, which provides full use of the
  41  * allocated memory.
  42  * -- Florian Coosmann (FGC)
  43  *
  44  * Reads with count = 0 should always return 0.
  45  * -- Julian Bradfield 1999-06-07.
  46  *
  47  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  48  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  49  *
  50  * pipe_read & write cleanup
  51  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  52  */
  53
  54 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  55 {
  56         if (pipe->inode)
  57                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
  58 }
  59
  60 void pipe_lock(struct pipe_inode_info *pipe)
  61 {
  62         /*
  63          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  64          */
  65         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  66 }
  67 EXPORT_SYMBOL(pipe_lock);
  68
  69 void pipe_unlock(struct pipe_inode_info *pipe)
  70 {
  71         if (pipe->inode)
  72                 mutex_unlock(&pipe->inode->i_mutex);
  73 }
  74 EXPORT_SYMBOL(pipe_unlock);
  75
  76 void pipe_double_lock(struct pipe_inode_info *pipe1,
  77                       struct pipe_inode_info *pipe2)
  78 {
  79         BUG_ON(pipe1 == pipe2);
  80
  81         if (pipe1 < pipe2) {
  82                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  83                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  84         } else {
  85                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
  86                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
  87         }
  88 }
  89
  90 /* Drop the inode semaphore and wait for a pipe event, atomically */
  91 void pipe_wait(struct pipe_inode_info *pipe)
  92 {
  93         DEFINE_WAIT(wait);
  94
  95         /*
  96          * Pipes are system-local resources, so sleeping on them
  97          * is considered a noninteractive wait:
  98          */
  99         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 100         pipe_unlock(pipe);
 101         schedule();
 102         finish_wait(&pipe->wait, &wait);
 103         pipe_lock(pipe);
 104 }
 105
 106 static int
 107 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
 108                         int atomic)
 109 {
 110         unsigned long copy;
 111
 112         while (len > 0) {
 113                 while (!iov->iov_len)
 114                         iov++;
 115                 copy = min_t(unsigned long, len, iov->iov_len);
 116
 117                 if (atomic) {
 118                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
 119                                 return -EFAULT;
 120                 } else {
 121                         if (copy_from_user(to, iov->iov_base, copy))
 122                                 return -EFAULT;
 123                 }
 124                 to += copy;
 125                 len -= copy;
 126                 iov->iov_base += copy;
 127                 iov->iov_len -= copy;
 128         }
 129         return 0;
 130 }
 131
 132 static int
 133 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
 134                       int atomic)
 135 {
 136         unsigned long copy;
 137
 138         while (len > 0) {
 139                 while (!iov->iov_len)
 140                         iov++;
 141                 copy = min_t(unsigned long, len, iov->iov_len);
 142
 143                 if (atomic) {
 144                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
 145                                 return -EFAULT;
 146                 } else {
 147                         if (copy_to_user(iov->iov_base, from, copy))
 148                                 return -EFAULT;
 149                 }
 150                 from += copy;
 151                 len -= copy;
 152                 iov->iov_base += copy;
 153                 iov->iov_len -= copy;
 154         }
 155         return 0;
 156 }
 157
 158 /*
 159  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 160  * Returns the number of bytes not faulted in.
 161  */
 162 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 163 {
 164         while (!iov->iov_len)
 165                 iov++;
 166
 167         while (len > 0) {
 168                 unsigned long this_len;
 169
 170                 this_len = min_t(unsigned long, len, iov->iov_len);
 171                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 172                         break;
 173
 174                 len -= this_len;
 175                 iov++;
 176         }
 177
 178         return len;
 179 }
 180
 181 /*
 182  * Pre-fault in the user memory, so we can use atomic copies.
 183  */
 184 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 185 {
 186         while (!iov->iov_len)
 187                 iov++;
 188
 189         while (len > 0) {
 190                 unsigned long this_len;
 191
 192                 this_len = min_t(unsigned long, len, iov->iov_len);
 193                 fault_in_pages_readable(iov->iov_base, this_len);
 194                 len -= this_len;
 195                 iov++;
 196         }
 197 }
 198
 199 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 200                                   struct pipe_buffer *buf)
 201 {
 202         struct page *page = buf->page;
 203
 204         /*
 205          * If nobody else uses this page, and we don't already have a
 206          * temporary page, let's keep track of it as a one-deep
 207          * allocation cache. (Otherwise just release our reference to it)
 208          */
 209         if (page_count(page) == 1 && !pipe->tmp_page)
 210                 pipe->tmp_page = page;
 211         else
 212                 page_cache_release(page);
 213 }
 214
 215 /**
 216  * generic_pipe_buf_map - virtually map a pipe buffer
 217  * @pipe:       the pipe that the buffer belongs to
 218  * @buf:        the buffer that should be mapped
 219  * @atomic:     whether to use an atomic map
 220  *
 221  * Description:
 222  *      This function returns a kernel virtual address mapping for the
 223  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 224  *      and the caller has to be careful not to fault before calling
 225  *      the unmap function.
 226  *
 227  *      Note that this function occupies KM_USER0 if @atomic != 0.
 228  */
 229 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 230                            struct pipe_buffer *buf, int atomic)
 231 {
 232         if (atomic) {
 233                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 234                 return kmap_atomic(buf->page);
 235         }
 236
 237         return kmap(buf->page);
 238 }
 239 EXPORT_SYMBOL(generic_pipe_buf_map);
 240
 241 /**
 242  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 243  * @pipe:       the pipe that the buffer belongs to
 244  * @buf:        the buffer that should be unmapped
 245  * @map_data:   the data that the mapping function returned
 246  *
 247  * Description:
 248  *      This function undoes the mapping that ->map() provided.
 249  */
 250 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 251                             struct pipe_buffer *buf, void *map_data)
 252 {
 253         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 254                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 255                 kunmap_atomic(map_data);
 256         } else
 257                 kunmap(buf->page);
 258 }
 259 EXPORT_SYMBOL(generic_pipe_buf_unmap);
 260
 261 /**
 262  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 263  * @pipe:       the pipe that the buffer belongs to
 264  * @buf:        the buffer to attempt to steal
 265  *
 266  * Description:
 267  *      This function attempts to steal the &struct page attached to
 268  *      @buf. If successful, this function returns 0 and returns with
 269  *      the page locked. The caller may then reuse the page for whatever
 270  *      he wishes; the typical use is insertion into a different file
 271  *      page cache.
 272  */
 273 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 274                            struct pipe_buffer *buf)
 275 {
 276         struct page *page = buf->page;
 277
 278         /*
 279          * A reference of one is golden, that means that the owner of this
 280          * page is the only one holding a reference to it. lock the page
 281          * and return OK.
 282          */
 283         if (page_count(page) == 1) {
 284                 lock_page(page);
 285                 return 0;
 286         }
 287
 288         return 1;
 289 }
 290 EXPORT_SYMBOL(generic_pipe_buf_steal);
 291
 292 /**
 293  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 294  * @pipe:       the pipe that the buffer belongs to
 295  * @buf:        the buffer to get a reference to
 296  *
 297  * Description:
 298  *      This function grabs an extra reference to @buf. It's used in
 299  *      in the tee() system call, when we duplicate the buffers in one
 300  *      pipe into another.
 301  */
 302 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 303 {
 304         page_cache_get(buf->page);
 305 }
 306 EXPORT_SYMBOL(generic_pipe_buf_get);
 307
 308 /**
 309  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 310  * @info:       the pipe that the buffer belongs to
 311  * @buf:        the buffer to confirm
 312  *
 313  * Description:
 314  *      This function does nothing, because the generic pipe code uses
 315  *      pages that are always good when inserted into the pipe.
 316  */
 317 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 318                              struct pipe_buffer *buf)
 319 {
 320         return 0;
 321 }
 322 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 323
 324 /**
 325  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 326  * @pipe:       the pipe that the buffer belongs to
 327  * @buf:        the buffer to put a reference to
 328  *
 329  * Description:
 330  *      This function releases a reference to @buf.
 331  */
 332 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 333                               struct pipe_buffer *buf)
 334 {
 335         page_cache_release(buf->page);
 336 }
 337 EXPORT_SYMBOL(generic_pipe_buf_release);
 338
 339 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 340         .can_merge = 1,
 341         .map = generic_pipe_buf_map,
 342         .unmap = generic_pipe_buf_unmap,
 343         .confirm = generic_pipe_buf_confirm,
 344         .release = anon_pipe_buf_release,
 345         .steal = generic_pipe_buf_steal,
 346         .get = generic_pipe_buf_get,
 347 };
 348
 349 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 350         .can_merge = 0,
 351         .map = generic_pipe_buf_map,
 352         .unmap = generic_pipe_buf_unmap,
 353         .confirm = generic_pipe_buf_confirm,
 354         .release = anon_pipe_buf_release,
 355         .steal = generic_pipe_buf_steal,
 356         .get = generic_pipe_buf_get,
 357 };
 358
 359 static ssize_t
 360 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 361            unsigned long nr_segs, loff_t pos)
 362 {
 363         struct file *filp = iocb->ki_filp;
 364         struct inode *inode = filp->f_path.dentry->d_inode;
 365         struct pipe_inode_info *pipe;
 366         int do_wakeup;
 367         ssize_t ret;
 368         struct iovec *iov = (struct iovec *)_iov;
 369         size_t total_len;
 370
 371         total_len = iov_length(iov, nr_segs);
 372         /* Null read succeeds. */
 373         if (unlikely(total_len == 0))
 374                 return 0;
 375
 376         do_wakeup = 0;
 377         ret = 0;
 378         mutex_lock(&inode->i_mutex);
 379         pipe = inode->i_pipe;
 380         for (;;) {
 381                 int bufs = pipe->nrbufs;
 382                 if (bufs) {
 383                         int curbuf = pipe->curbuf;
 384                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 385                         const struct pipe_buf_operations *ops = buf->ops;
 386                         void *addr;
 387                         size_t chars = buf->len;
 388                         int error, atomic;
 389
 390                         if (chars > total_len)
 391                                 chars = total_len;
 392
 393                         error = ops->confirm(pipe, buf);
 394                         if (error) {
 395                                 if (!ret)
 396                                         ret = error;
 397                                 break;
 398                         }
 399
 400                         atomic = !iov_fault_in_pages_write(iov, chars);
 401 redo:
 402                         addr = ops->map(pipe, buf, atomic);
 403                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
 404                         ops->unmap(pipe, buf, addr);
 405                         if (unlikely(error)) {
 406                                 /*
 407                                  * Just retry with the slow path if we failed.
 408                                  */
 409                                 if (atomic) {
 410                                         atomic = 0;
 411                                         goto redo;
 412                                 }
 413                                 if (!ret)
 414                                         ret = error;
 415                                 break;
 416                         }
 417                         ret += chars;
 418                         buf->offset += chars;
 419                         buf->len -= chars;
 420
 421                         /* Was it a packet buffer? Clean up and exit */
 422                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 423                                 total_len = chars;
 424                                 buf->len = 0;
 425                         }
 426
 427                         if (!buf->len) {
 428                                 buf->ops = NULL;
 429                                 ops->release(pipe, buf);
 430                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 431                                 pipe->curbuf = curbuf;
 432                                 pipe->nrbufs = --bufs;
 433                                 do_wakeup = 1;
 434                         }
 435                         total_len -= chars;
 436                         if (!total_len)
 437                                 break;  /* common path: read succeeded */
 438                 }
 439                 if (bufs)       /* More to do? */
 440                         continue;
 441                 if (!pipe->writers)
 442                         break;
 443                 if (!pipe->waiting_writers) {
 444                         /* syscall merging: Usually we must not sleep
 445                          * if O_NONBLOCK is set, or if we got some data.
 446                          * But if a writer sleeps in kernel space, then
 447                          * we can wait for that data without violating POSIX.
 448                          */
 449                         if (ret)
 450                                 break;
 451                         if (filp->f_flags & O_NONBLOCK) {
 452                                 ret = -EAGAIN;
 453                                 break;
 454                         }
 455                 }
 456                 if (signal_pending(current)) {
 457                         if (!ret)
 458                                 ret = -ERESTARTSYS;
 459                         break;
 460                 }
 461                 if (do_wakeup) {
 462                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 463                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 464                 }
 465                 pipe_wait(pipe);
 466         }
 467         mutex_unlock(&inode->i_mutex);
 468
 469         /* Signal writers asynchronously that there is more room. */
 470         if (do_wakeup) {
 471                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 472                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 473         }
 474         if (ret > 0)
 475                 file_accessed(filp);
 476         return ret;
 477 }
 478
 479 static inline int is_packetized(struct file *file)
 480 {
 481         return (file->f_flags & O_DIRECT) != 0;
 482 }
 483
 484 static ssize_t
 485 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 486             unsigned long nr_segs, loff_t ppos)
 487 {
 488         struct file *filp = iocb->ki_filp;
 489         struct inode *inode = filp->f_path.dentry->d_inode;
 490         struct pipe_inode_info *pipe;
 491         ssize_t ret;
 492         int do_wakeup;
 493         struct iovec *iov = (struct iovec *)_iov;
 494         size_t total_len;
 495         ssize_t chars;
 496
 497         total_len = iov_length(iov, nr_segs);
 498         /* Null write succeeds. */
 499         if (unlikely(total_len == 0))
 500                 return 0;
 501
 502         do_wakeup = 0;
 503         ret = 0;
 504         mutex_lock(&inode->i_mutex);
 505         pipe = inode->i_pipe;
 506
 507         if (!pipe->readers) {
 508                 send_sig(SIGPIPE, current, 0);
 509                 ret = -EPIPE;
 510                 goto out;
 511         }
 512
 513         /* We try to merge small writes */
 514         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 515         if (pipe->nrbufs && chars != 0) {
 516                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 517                                                         (pipe->buffers - 1);
 518                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 519                 const struct pipe_buf_operations *ops = buf->ops;
 520                 int offset = buf->offset + buf->len;
 521
 522                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 523                         int error, atomic = 1;
 524                         void *addr;
 525
 526                         error = ops->confirm(pipe, buf);
 527                         if (error)
 528                                 goto out;
 529
 530                         iov_fault_in_pages_read(iov, chars);
 531 redo1:
 532                         addr = ops->map(pipe, buf, atomic);
 533                         error = pipe_iov_copy_from_user(offset + addr, iov,
 534                                                         chars, atomic);
 535                         ops->unmap(pipe, buf, addr);
 536                         ret = error;
 537                         do_wakeup = 1;
 538                         if (error) {
 539                                 if (atomic) {
 540                                         atomic = 0;
 541                                         goto redo1;
 542                                 }
 543                                 goto out;
 544                         }
 545                         buf->len += chars;
 546                         total_len -= chars;
 547                         ret = chars;
 548                         if (!total_len)
 549                                 goto out;
 550                 }
 551         }
 552
 553         for (;;) {
 554                 int bufs;
 555
 556                 if (!pipe->readers) {
 557                         send_sig(SIGPIPE, current, 0);
 558                         if (!ret)
 559                                 ret = -EPIPE;
 560                         break;
 561                 }
 562                 bufs = pipe->nrbufs;
 563                 if (bufs < pipe->buffers) {
 564                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 565                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 566                         struct page *page = pipe->tmp_page;
 567                         char *src;
 568                         int error, atomic = 1;
 569
 570                         if (!page) {
 571                                 page = alloc_page(GFP_HIGHUSER);
 572                                 if (unlikely(!page)) {
 573                                         ret = ret ? : -ENOMEM;
 574                                         break;
 575                                 }
 576                                 pipe->tmp_page = page;
 577                         }
 578                         /* Always wake up, even if the copy fails. Otherwise
 579                          * we lock up (O_NONBLOCK-)readers that sleep due to
 580                          * syscall merging.
 581                          * FIXME! Is this really true?
 582                          */
 583                         do_wakeup = 1;
 584                         chars = PAGE_SIZE;
 585                         if (chars > total_len)
 586                                 chars = total_len;
 587
 588                         iov_fault_in_pages_read(iov, chars);
 589 redo2:
 590                         if (atomic)
 591                                 src = kmap_atomic(page);
 592                         else
 593                                 src = kmap(page);
 594
 595                         error = pipe_iov_copy_from_user(src, iov, chars,
 596                                                         atomic);
 597                         if (atomic)
 598                                 kunmap_atomic(src);
 599                         else
 600                                 kunmap(page);
 601
 602                         if (unlikely(error)) {
 603                                 if (atomic) {
 604                                         atomic = 0;
 605                                         goto redo2;
 606                                 }
 607                                 if (!ret)
 608                                         ret = error;
 609                                 break;
 610                         }
 611                         ret += chars;
 612
 613                         /* Insert it into the buffer array */
 614                         buf->page = page;
 615                         buf->ops = &anon_pipe_buf_ops;
 616                         buf->offset = 0;
 617                         buf->len = chars;
 618                         buf->flags = 0;
 619                         if (is_packetized(filp)) {
 620                                 buf->ops = &packet_pipe_buf_ops;
 621                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 622                         }
 623                         pipe->nrbufs = ++bufs;
 624                         pipe->tmp_page = NULL;
 625
 626                         total_len -= chars;
 627                         if (!total_len)
 628                                 break;
 629                 }
 630                 if (bufs < pipe->buffers)
 631                         continue;
 632                 if (filp->f_flags & O_NONBLOCK) {
 633                         if (!ret)
 634                                 ret = -EAGAIN;
 635                         break;
 636                 }
 637                 if (signal_pending(current)) {
 638                         if (!ret)
 639                                 ret = -ERESTARTSYS;
 640                         break;
 641                 }
 642                 if (do_wakeup) {
 643                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 644                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 645                         do_wakeup = 0;
 646                 }
 647                 pipe->waiting_writers++;
 648                 pipe_wait(pipe);
 649                 pipe->waiting_writers--;
 650         }
 651 out:
 652         mutex_unlock(&inode->i_mutex);
 653         if (do_wakeup) {
 654                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 655                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 656         }
 657         if (ret > 0) {
 658                 int err = file_update_time(filp);
 659                 if (err)
 660                         ret = err;
 661         }
 662         return ret;
 663 }
 664
 665 static ssize_t
 666 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 667 {
 668         return -EBADF;
 669 }
 670
 671 static ssize_t
 672 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 673            loff_t *ppos)
 674 {
 675         return -EBADF;
 676 }
 677
 678 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 679 {
 680         struct inode *inode = filp->f_path.dentry->d_inode;
 681         struct pipe_inode_info *pipe;
 682         int count, buf, nrbufs;
 683
 684         switch (cmd) {
 685                 case FIONREAD:
 686                         mutex_lock(&inode->i_mutex);
 687                         pipe = inode->i_pipe;
 688                         count = 0;
 689                         buf = pipe->curbuf;
 690                         nrbufs = pipe->nrbufs;
 691                         while (--nrbufs >= 0) {
 692                                 count += pipe->bufs[buf].len;
 693                                 buf = (buf+1) & (pipe->buffers - 1);
 694                         }
 695                         mutex_unlock(&inode->i_mutex);
 696
 697                         return put_user(count, (int __user *)arg);
 698                 default:
 699                         return -ENOIOCTLCMD;
 700         }
 701 }
 702
 703 /* No kernel lock held - fine */
 704 static unsigned int
 705 pipe_poll(struct file *filp, poll_table *wait)
 706 {
 707         unsigned int mask;
 708         struct inode *inode = filp->f_path.dentry->d_inode;
 709         struct pipe_inode_info *pipe = inode->i_pipe;
 710         int nrbufs;
 711
 712         poll_wait(filp, &pipe->wait, wait);
 713
 714         /* Reading only -- no need for acquiring the semaphore.  */
 715         nrbufs = pipe->nrbufs;
 716         mask = 0;
 717         if (filp->f_mode & FMODE_READ) {
 718                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 719                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 720                         mask |= POLLHUP;
 721         }
 722
 723         if (filp->f_mode & FMODE_WRITE) {
 724                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 725                 /*
 726                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 727                  * behave exactly like pipes for poll().
 728                  */
 729                 if (!pipe->readers)
 730                         mask |= POLLERR;
 731         }
 732
 733         return mask;
 734 }
 735
 736 static int
 737 pipe_release(struct inode *inode, int decr, int decw)
 738 {
 739         struct pipe_inode_info *pipe;
 740
 741         mutex_lock(&inode->i_mutex);
 742         pipe = inode->i_pipe;
 743         pipe->readers -= decr;
 744         pipe->writers -= decw;
 745
 746         if (!pipe->readers && !pipe->writers) {
 747                 free_pipe_info(inode);
 748         } else {
 749                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 750                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 751                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 752         }
 753         mutex_unlock(&inode->i_mutex);
 754
 755         return 0;
 756 }
 757
 758 static int
 759 pipe_read_fasync(int fd, struct file *filp, int on)
 760 {
 761         struct inode *inode = filp->f_path.dentry->d_inode;
 762         int retval;
 763
 764         mutex_lock(&inode->i_mutex);
 765         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 766         mutex_unlock(&inode->i_mutex);
 767
 768         return retval;
 769 }
 770
 771
 772 static int
 773 pipe_write_fasync(int fd, struct file *filp, int on)
 774 {
 775         struct inode *inode = filp->f_path.dentry->d_inode;
 776         int retval;
 777
 778         mutex_lock(&inode->i_mutex);
 779         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 780         mutex_unlock(&inode->i_mutex);
 781
 782         return retval;
 783 }
 784
 785
 786 static int
 787 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 788 {
 789         struct inode *inode = filp->f_path.dentry->d_inode;
 790         struct pipe_inode_info *pipe = inode->i_pipe;
 791         int retval;
 792
 793         mutex_lock(&inode->i_mutex);
 794         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 795         if (retval >= 0) {
 796                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 797                 if (retval < 0) /* this can happen only if on == T */
 798                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 799         }
 800         mutex_unlock(&inode->i_mutex);
 801         return retval;
 802 }
 803
 804
 805 static int
 806 pipe_read_release(struct inode *inode, struct file *filp)
 807 {
 808         return pipe_release(inode, 1, 0);
 809 }
 810
 811 static int
 812 pipe_write_release(struct inode *inode, struct file *filp)
 813 {
 814         return pipe_release(inode, 0, 1);
 815 }
 816
 817 static int
 818 pipe_rdwr_release(struct inode *inode, struct file *filp)
 819 {
 820         int decr, decw;
 821
 822         decr = (filp->f_mode & FMODE_READ) != 0;
 823         decw = (filp->f_mode & FMODE_WRITE) != 0;
 824         return pipe_release(inode, decr, decw);
 825 }
 826
 827 static int
 828 pipe_read_open(struct inode *inode, struct file *filp)
 829 {
 830         int ret = -ENOENT;
 831
 832         mutex_lock(&inode->i_mutex);
 833
 834         if (inode->i_pipe) {
 835                 ret = 0;
 836                 inode->i_pipe->readers++;
 837         }
 838
 839         mutex_unlock(&inode->i_mutex);
 840
 841         return ret;
 842 }
 843
 844 static int
 845 pipe_write_open(struct inode *inode, struct file *filp)
 846 {
 847         int ret = -ENOENT;
 848
 849         mutex_lock(&inode->i_mutex);
 850
 851         if (inode->i_pipe) {
 852                 ret = 0;
 853                 inode->i_pipe->writers++;
 854         }
 855
 856         mutex_unlock(&inode->i_mutex);
 857
 858         return ret;
 859 }
 860
 861 static int
 862 pipe_rdwr_open(struct inode *inode, struct file *filp)
 863 {
 864         int ret = -ENOENT;
 865
 866         mutex_lock(&inode->i_mutex);
 867
 868         if (inode->i_pipe) {
 869                 ret = 0;
 870                 if (filp->f_mode & FMODE_READ)
 871                         inode->i_pipe->readers++;
 872                 if (filp->f_mode & FMODE_WRITE)
 873                         inode->i_pipe->writers++;
 874         }
 875
 876         mutex_unlock(&inode->i_mutex);
 877
 878         return ret;
 879 }
 880
 881 /*
 882  * The file_operations structs are not static because they
 883  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 884  *
 885  * Pipes reuse fifos' file_operations structs.
 886  */
 887 const struct file_operations read_pipefifo_fops = {
 888         .llseek         = no_llseek,
 889         .read           = do_sync_read,
 890         .aio_read       = pipe_read,
 891         .write          = bad_pipe_w,
 892         .poll           = pipe_poll,
 893         .unlocked_ioctl = pipe_ioctl,
 894         .open           = pipe_read_open,
 895         .release        = pipe_read_release,
 896         .fasync         = pipe_read_fasync,
 897 };
 898
 899 const struct file_operations write_pipefifo_fops = {
 900         .llseek         = no_llseek,
 901         .read           = bad_pipe_r,
 902         .write          = do_sync_write,
 903         .aio_write      = pipe_write,
 904         .poll           = pipe_poll,
 905         .unlocked_ioctl = pipe_ioctl,
 906         .open           = pipe_write_open,
 907         .release        = pipe_write_release,
 908         .fasync         = pipe_write_fasync,
 909 };
 910
 911 const struct file_operations rdwr_pipefifo_fops = {
 912         .llseek         = no_llseek,
 913         .read           = do_sync_read,
 914         .aio_read       = pipe_read,
 915         .write          = do_sync_write,
 916         .aio_write      = pipe_write,
 917         .poll           = pipe_poll,
 918         .unlocked_ioctl = pipe_ioctl,
 919         .open           = pipe_rdwr_open,
 920         .release        = pipe_rdwr_release,
 921         .fasync         = pipe_rdwr_fasync,
 922 };
 923
 924 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 925 {
 926         struct pipe_inode_info *pipe;
 927
 928         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 929         if (pipe) {
 930                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
 931                 if (pipe->bufs) {
 932                         init_waitqueue_head(&pipe->wait);
 933                         pipe->r_counter = pipe->w_counter = 1;
 934                         pipe->inode = inode;
 935                         pipe->buffers = PIPE_DEF_BUFFERS;
 936                         return pipe;
 937                 }
 938                 kfree(pipe);
 939         }
 940
 941         return NULL;
 942 }
 943
 944 void __free_pipe_info(struct pipe_inode_info *pipe)
 945 {
 946         int i;
 947
 948         for (i = 0; i < pipe->buffers; i++) {
 949                 struct pipe_buffer *buf = pipe->bufs + i;
 950                 if (buf->ops)
 951                         buf->ops->release(pipe, buf);
 952         }
 953         if (pipe->tmp_page)
 954                 __free_page(pipe->tmp_page);
 955         kfree(pipe->bufs);
 956         kfree(pipe);
 957 }
 958
 959 void free_pipe_info(struct inode *inode)
 960 {
 961         __free_pipe_info(inode->i_pipe);
 962         inode->i_pipe = NULL;
 963 }
 964
 965 static struct vfsmount *pipe_mnt __read_mostly;
 966
 967 /*
 968  * pipefs_dname() is called from d_path().
 969  */
 970 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 971 {
 972         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 973                                 dentry->d_inode->i_ino);
 974 }
 975
 976 static const struct dentry_operations pipefs_dentry_operations = {
 977         .d_dname        = pipefs_dname,
 978 };
 979
 980 static struct inode * get_pipe_inode(void)
 981 {
 982         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 983         struct pipe_inode_info *pipe;
 984
 985         if (!inode)
 986                 goto fail_inode;
 987
 988         inode->i_ino = get_next_ino();
 989
 990         pipe = alloc_pipe_info(inode);
 991         if (!pipe)
 992                 goto fail_iput;
 993         inode->i_pipe = pipe;
 994
 995         pipe->readers = pipe->writers = 1;
 996         inode->i_fop = &rdwr_pipefifo_fops;
 997
 998         /*
 999          * Mark the inode dirty from the very beginning,
1000          * that way it will never be moved to the dirty
1001          * list because "mark_inode_dirty()" will think
1002          * that it already _is_ on the dirty list.
1003          */
1004         inode->i_state = I_DIRTY;
1005         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1006         inode->i_uid = current_fsuid();
1007         inode->i_gid = current_fsgid();
1008         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1009
1010         return inode;
1011
1012 fail_iput:
1013         iput(inode);
1014
1015 fail_inode:
1016         return NULL;
1017 }
1018
1019 struct file *create_write_pipe(int flags)
1020 {
1021         int err;
1022         struct inode *inode;
1023         struct file *f;
1024         struct path path;
1025         struct qstr name = { .name = "" };
1026
1027         err = -ENFILE;
1028         inode = get_pipe_inode();
1029         if (!inode)
1030                 goto err;
1031
1032         err = -ENOMEM;
1033         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1034         if (!path.dentry)
1035                 goto err_inode;
1036         path.mnt = mntget(pipe_mnt);
1037
1038         d_instantiate(path.dentry, inode);
1039
1040         err = -ENFILE;
1041         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1042         if (!f)
1043                 goto err_dentry;
1044         f->f_mapping = inode->i_mapping;
1045
1046         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
1047         f->f_version = 0;
1048
1049         return f;
1050
1051  err_dentry:
1052         free_pipe_info(inode);
1053         path_put(&path);
1054         return ERR_PTR(err);
1055
1056  err_inode:
1057         free_pipe_info(inode);
1058         iput(inode);
1059  err:
1060         return ERR_PTR(err);
1061 }
1062
1063 void free_write_pipe(struct file *f)
1064 {
1065         free_pipe_info(f->f_dentry->d_inode);
1066         path_put(&f->f_path);
1067         put_filp(f);
1068 }
1069
1070 struct file *create_read_pipe(struct file *wrf, int flags)
1071 {
1072         /* Grab pipe from the writer */
1073         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1074                                     &read_pipefifo_fops);
1075         if (!f)
1076                 return ERR_PTR(-ENFILE);
1077
1078         path_get(&wrf->f_path);
1079         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1080
1081         return f;
1082 }
1083
1084 int do_pipe_flags(int *fd, int flags)
1085 {
1086         struct file *fw, *fr;
1087         int error;
1088         int fdw, fdr;
1089
1090         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
1091                 return -EINVAL;
1092
1093         fw = create_write_pipe(flags);
1094         if (IS_ERR(fw))
1095                 return PTR_ERR(fw);
1096         fr = create_read_pipe(fw, flags);
1097         error = PTR_ERR(fr);
1098         if (IS_ERR(fr))
1099                 goto err_write_pipe;
1100
1101         error = get_unused_fd_flags(flags);
1102         if (error < 0)
1103                 goto err_read_pipe;
1104         fdr = error;
1105
1106         error = get_unused_fd_flags(flags);
1107         if (error < 0)
1108                 goto err_fdr;
1109         fdw = error;
1110
1111         audit_fd_pair(fdr, fdw);
1112         fd_install(fdr, fr);
1113         fd_install(fdw, fw);
1114         fd[0] = fdr;
1115         fd[1] = fdw;
1116
1117         return 0;
1118
1119  err_fdr:
1120         put_unused_fd(fdr);
1121  err_read_pipe:
1122         path_put(&fr->f_path);
1123         put_filp(fr);
1124  err_write_pipe:
1125         free_write_pipe(fw);
1126         return error;
1127 }
1128
1129 /*
1130  * sys_pipe() is the normal C calling standard for creating
1131  * a pipe. It's not the way Unix traditionally does this, though.
1132  */
1133 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1134 {
1135         int fd[2];
1136         int error;
1137
1138         error = do_pipe_flags(fd, flags);
1139         if (!error) {
1140                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1141                         sys_close(fd[0]);
1142                         sys_close(fd[1]);
1143                         error = -EFAULT;
1144                 }
1145         }
1146         return error;
1147 }
1148
1149 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1150 {
1151         return sys_pipe2(fildes, 0);
1152 }
1153
1154 /*
1155  * Allocate a new array of pipe buffers and copy the info over. Returns the
1156  * pipe size if successful, or return -ERROR on error.
1157  */
1158 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1159 {
1160         struct pipe_buffer *bufs;
1161
1162         /*
1163          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1164          * expect a lot of shrink+grow operations, just free and allocate
1165          * again like we would do for growing. If the pipe currently
1166          * contains more buffers than arg, then return busy.
1167          */
1168         if (nr_pages < pipe->nrbufs)
1169                 return -EBUSY;
1170
1171         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1172         if (unlikely(!bufs))
1173                 return -ENOMEM;
1174
1175         /*
1176          * The pipe array wraps around, so just start the new one at zero
1177          * and adjust the indexes.
1178          */
1179         if (pipe->nrbufs) {
1180                 unsigned int tail;
1181                 unsigned int head;
1182
1183                 tail = pipe->curbuf + pipe->nrbufs;
1184                 if (tail < pipe->buffers)
1185                         tail = 0;
1186                 else
1187                         tail &= (pipe->buffers - 1);
1188
1189                 head = pipe->nrbufs - tail;
1190                 if (head)
1191                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1192                 if (tail)
1193                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1194         }
1195
1196         pipe->curbuf = 0;
1197         kfree(pipe->bufs);
1198         pipe->bufs = bufs;
1199         pipe->buffers = nr_pages;
1200         return nr_pages * PAGE_SIZE;
1201 }
1202
1203 /*
1204  * Currently we rely on the pipe array holding a power-of-2 number
1205  * of pages.
1206  */
1207 static inline unsigned int round_pipe_size(unsigned int size)
1208 {
1209         unsigned long nr_pages;
1210
1211         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1212         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1213 }
1214
1215 /*
1216  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1217  * will return an error.
1218  */
1219 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1220                  size_t *lenp, loff_t *ppos)
1221 {
1222         int ret;
1223
1224         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1225         if (ret < 0 || !write)
1226                 return ret;
1227
1228         pipe_max_size = round_pipe_size(pipe_max_size);
1229         return ret;
1230 }
1231
1232 /*
1233  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1234  * location, so checking ->i_pipe is not enough to verify that this is a
1235  * pipe.
1236  */
1237 struct pipe_inode_info *get_pipe_info(struct file *file)
1238 {
1239         struct inode *i = file->f_path.dentry->d_inode;
1240
1241         return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1242 }
1243
1244 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1245 {
1246         struct pipe_inode_info *pipe;
1247         long ret;
1248
1249         pipe = get_pipe_info(file);
1250         if (!pipe)
1251                 return -EBADF;
1252
1253         mutex_lock(&pipe->inode->i_mutex);
1254
1255         switch (cmd) {
1256         case F_SETPIPE_SZ: {
1257                 unsigned int size, nr_pages;
1258
1259                 size = round_pipe_size(arg);
1260                 nr_pages = size >> PAGE_SHIFT;
1261
1262                 ret = -EINVAL;
1263                 if (!nr_pages)
1264                         goto out;
1265
1266                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1267                         ret = -EPERM;
1268                         goto out;
1269                 }
1270                 ret = pipe_set_size(pipe, nr_pages);
1271                 break;
1272                 }
1273         case F_GETPIPE_SZ:
1274                 ret = pipe->buffers * PAGE_SIZE;
1275                 break;
1276         default:
1277                 ret = -EINVAL;
1278                 break;
1279         }
1280
1281 out:
1282         mutex_unlock(&pipe->inode->i_mutex);
1283         return ret;
1284 }
1285
1286 static const struct super_operations pipefs_ops = {
1287         .destroy_inode = free_inode_nonrcu,
1288         .statfs = simple_statfs,
1289 };
1290
1291 /*
1292  * pipefs should _never_ be mounted by userland - too much of security hassle,
1293  * no real gain from having the whole whorehouse mounted. So we don't need
1294  * any operations on the root directory. However, we need a non-trivial
1295  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1296  */
1297 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1298                          int flags, const char *dev_name, void *data)
1299 {
1300         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1301                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1302 }
1303
1304 static struct file_system_type pipe_fs_type = {
1305         .name           = "pipefs",
1306         .mount          = pipefs_mount,
1307         .kill_sb        = kill_anon_super,
1308 };
1309
1310 static int __init init_pipe_fs(void)
1311 {
1312         int err = register_filesystem(&pipe_fs_type);
1313
1314         if (!err) {
1315                 pipe_mnt = kern_mount(&pipe_fs_type);
1316                 if (IS_ERR(pipe_mnt)) {
1317                         err = PTR_ERR(pipe_mnt);
1318                         unregister_filesystem(&pipe_fs_type);
1319                 }
1320         }
1321         return err;
1322 }
1323
1324 fs_initcall(init_pipe_fs);