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