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