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