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