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 (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 620                         pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 621
 622                 if (!too_many_pipe_buffers_hard(user)) {
 623                         if (too_many_pipe_buffers_soft(user))
 624                                 pipe_bufs = 1;
 625                         pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * pipe_bufs, GFP_KERNEL);
 626                 }
 627
 628                 if (pipe->bufs) {
 629                         init_waitqueue_head(&pipe->wait);
 630                         pipe->r_counter = pipe->w_counter = 1;
 631                         pipe->buffers = pipe_bufs;
 632                         pipe->user = user;
 633                         account_pipe_buffers(pipe, 0, pipe_bufs);
 634                         mutex_init(&pipe->mutex);
 635                         return pipe;
 636                 }
 637                 free_uid(user);
 638                 kfree(pipe);
 639         }
 640
 641         return NULL;
 642 }
 643
 644 void free_pipe_info(struct pipe_inode_info *pipe)
 645 {
 646         int i;
 647
 648         account_pipe_buffers(pipe, pipe->buffers, 0);
 649         free_uid(pipe->user);
 650         for (i = 0; i < pipe->buffers; i++) {
 651                 struct pipe_buffer *buf = pipe->bufs + i;
 652                 if (buf->ops)
 653                         buf->ops->release(pipe, buf);
 654         }
 655         if (pipe->tmp_page)
 656                 __free_page(pipe->tmp_page);
 657         kfree(pipe->bufs);
 658         kfree(pipe);
 659 }
 660
 661 static struct vfsmount *pipe_mnt __read_mostly;
 662
 663 /*
 664  * pipefs_dname() is called from d_path().
 665  */
 666 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 667 {
 668         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 669                                 d_inode(dentry)->i_ino);
 670 }
 671
 672 static const struct dentry_operations pipefs_dentry_operations = {
 673         .d_dname        = pipefs_dname,
 674 };
 675
 676 static struct inode * get_pipe_inode(void)
 677 {
 678         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 679         struct pipe_inode_info *pipe;
 680
 681         if (!inode)
 682                 goto fail_inode;
 683
 684         inode->i_ino = get_next_ino();
 685
 686         pipe = alloc_pipe_info();
 687         if (!pipe)
 688                 goto fail_iput;
 689
 690         inode->i_pipe = pipe;
 691         pipe->files = 2;
 692         pipe->readers = pipe->writers = 1;
 693         inode->i_fop = &pipefifo_fops;
 694
 695         /*
 696          * Mark the inode dirty from the very beginning,
 697          * that way it will never be moved to the dirty
 698          * list because "mark_inode_dirty()" will think
 699          * that it already _is_ on the dirty list.
 700          */
 701         inode->i_state = I_DIRTY;
 702         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 703         inode->i_uid = current_fsuid();
 704         inode->i_gid = current_fsgid();
 705         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 706
 707         return inode;
 708
 709 fail_iput:
 710         iput(inode);
 711
 712 fail_inode:
 713         return NULL;
 714 }
 715
 716 int create_pipe_files(struct file **res, int flags)
 717 {
 718         int err;
 719         struct inode *inode = get_pipe_inode();
 720         struct file *f;
 721         struct path path;
 722         static struct qstr name = { .name = "" };
 723
 724         if (!inode)
 725                 return -ENFILE;
 726
 727         err = -ENOMEM;
 728         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
 729         if (!path.dentry)
 730                 goto err_inode;
 731         path.mnt = mntget(pipe_mnt);
 732
 733         d_instantiate(path.dentry, inode);
 734
 735         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 736         if (IS_ERR(f)) {
 737                 err = PTR_ERR(f);
 738                 goto err_dentry;
 739         }
 740
 741         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 742         f->private_data = inode->i_pipe;
 743
 744         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 745         if (IS_ERR(res[0])) {
 746                 err = PTR_ERR(res[0]);
 747                 goto err_file;
 748         }
 749
 750         path_get(&path);
 751         res[0]->private_data = inode->i_pipe;
 752         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 753         res[1] = f;
 754         return 0;
 755
 756 err_file:
 757         put_filp(f);
 758 err_dentry:
 759         free_pipe_info(inode->i_pipe);
 760         path_put(&path);
 761         return err;
 762
 763 err_inode:
 764         free_pipe_info(inode->i_pipe);
 765         iput(inode);
 766         return err;
 767 }
 768
 769 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 770 {
 771         int error;
 772         int fdw, fdr;
 773
 774         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 775                 return -EINVAL;
 776
 777         error = create_pipe_files(files, flags);
 778         if (error)
 779                 return error;
 780
 781         error = get_unused_fd_flags(flags);
 782         if (error < 0)
 783                 goto err_read_pipe;
 784         fdr = error;
 785
 786         error = get_unused_fd_flags(flags);
 787         if (error < 0)
 788                 goto err_fdr;
 789         fdw = error;
 790
 791         audit_fd_pair(fdr, fdw);
 792         fd[0] = fdr;
 793         fd[1] = fdw;
 794         return 0;
 795
 796  err_fdr:
 797         put_unused_fd(fdr);
 798  err_read_pipe:
 799         fput(files[0]);
 800         fput(files[1]);
 801         return error;
 802 }
 803
 804 int do_pipe_flags(int *fd, int flags)
 805 {
 806         struct file *files[2];
 807         int error = __do_pipe_flags(fd, files, flags);
 808         if (!error) {
 809                 fd_install(fd[0], files[0]);
 810                 fd_install(fd[1], files[1]);
 811         }
 812         return error;
 813 }
 814
 815 /*
 816  * sys_pipe() is the normal C calling standard for creating
 817  * a pipe. It's not the way Unix traditionally does this, though.
 818  */
 819 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 820 {
 821         struct file *files[2];
 822         int fd[2];
 823         int error;
 824
 825         error = __do_pipe_flags(fd, files, flags);
 826         if (!error) {
 827                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 828                         fput(files[0]);
 829                         fput(files[1]);
 830                         put_unused_fd(fd[0]);
 831                         put_unused_fd(fd[1]);
 832                         error = -EFAULT;
 833                 } else {
 834                         fd_install(fd[0], files[0]);
 835                         fd_install(fd[1], files[1]);
 836                 }
 837         }
 838         return error;
 839 }
 840
 841 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 842 {
 843         return sys_pipe2(fildes, 0);
 844 }
 845
 846 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 847 {
 848         int cur = *cnt;
 849
 850         while (cur == *cnt) {
 851                 pipe_wait(pipe);
 852                 if (signal_pending(current))
 853                         break;
 854         }
 855         return cur == *cnt ? -ERESTARTSYS : 0;
 856 }
 857
 858 static void wake_up_partner(struct pipe_inode_info *pipe)
 859 {
 860         wake_up_interruptible(&pipe->wait);
 861 }
 862
 863 static int fifo_open(struct inode *inode, struct file *filp)
 864 {
 865         struct pipe_inode_info *pipe;
 866         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 867         int ret;
 868
 869         filp->f_version = 0;
 870
 871         spin_lock(&inode->i_lock);
 872         if (inode->i_pipe) {
 873                 pipe = inode->i_pipe;
 874                 pipe->files++;
 875                 spin_unlock(&inode->i_lock);
 876         } else {
 877                 spin_unlock(&inode->i_lock);
 878                 pipe = alloc_pipe_info();
 879                 if (!pipe)
 880                         return -ENOMEM;
 881                 pipe->files = 1;
 882                 spin_lock(&inode->i_lock);
 883                 if (unlikely(inode->i_pipe)) {
 884                         inode->i_pipe->files++;
 885                         spin_unlock(&inode->i_lock);
 886                         free_pipe_info(pipe);
 887                         pipe = inode->i_pipe;
 888                 } else {
 889                         inode->i_pipe = pipe;
 890                         spin_unlock(&inode->i_lock);
 891                 }
 892         }
 893         filp->private_data = pipe;
 894         /* OK, we have a pipe and it's pinned down */
 895
 896         __pipe_lock(pipe);
 897
 898         /* We can only do regular read/write on fifos */
 899         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 900
 901         switch (filp->f_mode) {
 902         case FMODE_READ:
 903         /*
 904          *  O_RDONLY
 905          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 906          *  opened, even when there is no process writing the FIFO.
 907          */
 908                 pipe->r_counter++;
 909                 if (pipe->readers++ == 0)
 910                         wake_up_partner(pipe);
 911
 912                 if (!is_pipe && !pipe->writers) {
 913                         if ((filp->f_flags & O_NONBLOCK)) {
 914                                 /* suppress POLLHUP until we have
 915                                  * seen a writer */
 916                                 filp->f_version = pipe->w_counter;
 917                         } else {
 918                                 if (wait_for_partner(pipe, &pipe->w_counter))
 919                                         goto err_rd;
 920                         }
 921                 }
 922                 break;
 923
 924         case FMODE_WRITE:
 925         /*
 926          *  O_WRONLY
 927          *  POSIX.1 says that O_NONBLOCK means return -1 with
 928          *  errno=ENXIO when there is no process reading the FIFO.
 929          */
 930                 ret = -ENXIO;
 931                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 932                         goto err;
 933
 934                 pipe->w_counter++;
 935                 if (!pipe->writers++)
 936                         wake_up_partner(pipe);
 937
 938                 if (!is_pipe && !pipe->readers) {
 939                         if (wait_for_partner(pipe, &pipe->r_counter))
 940                                 goto err_wr;
 941                 }
 942                 break;
 943
 944         case FMODE_READ | FMODE_WRITE:
 945         /*
 946          *  O_RDWR
 947          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 948          *  This implementation will NEVER block on a O_RDWR open, since
 949          *  the process can at least talk to itself.
 950          */
 951
 952                 pipe->readers++;
 953                 pipe->writers++;
 954                 pipe->r_counter++;
 955                 pipe->w_counter++;
 956                 if (pipe->readers == 1 || pipe->writers == 1)
 957                         wake_up_partner(pipe);
 958                 break;
 959
 960         default:
 961                 ret = -EINVAL;
 962                 goto err;
 963         }
 964
 965         /* Ok! */
 966         __pipe_unlock(pipe);
 967         return 0;
 968
 969 err_rd:
 970         if (!--pipe->readers)
 971                 wake_up_interruptible(&pipe->wait);
 972         ret = -ERESTARTSYS;
 973         goto err;
 974
 975 err_wr:
 976         if (!--pipe->writers)
 977                 wake_up_interruptible(&pipe->wait);
 978         ret = -ERESTARTSYS;
 979         goto err;
 980
 981 err:
 982         __pipe_unlock(pipe);
 983
 984         put_pipe_info(inode, pipe);
 985         return ret;
 986 }
 987
 988 const struct file_operations pipefifo_fops = {
 989         .open           = fifo_open,
 990         .llseek         = no_llseek,
 991         .read_iter      = pipe_read,
 992         .write_iter     = pipe_write,
 993         .poll           = pipe_poll,
 994         .unlocked_ioctl = pipe_ioctl,
 995         .release        = pipe_release,
 996         .fasync         = pipe_fasync,
 997 };
 998
 999 /*
1000  * Allocate a new array of pipe buffers and copy the info over. Returns the
1001  * pipe size if successful, or return -ERROR on error.
1002  */
1003 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1004 {
1005         struct pipe_buffer *bufs;
1006
1007         if (!nr_pages)
1008                 return -EINVAL;
1009
1010         /*
1011          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1012          * expect a lot of shrink+grow operations, just free and allocate
1013          * again like we would do for growing. If the pipe currently
1014          * contains more buffers than arg, then return busy.
1015          */
1016         if (nr_pages < pipe->nrbufs)
1017                 return -EBUSY;
1018
1019         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1020         if (unlikely(!bufs))
1021                 return -ENOMEM;
1022
1023         /*
1024          * The pipe array wraps around, so just start the new one at zero
1025          * and adjust the indexes.
1026          */
1027         if (pipe->nrbufs) {
1028                 unsigned int tail;
1029                 unsigned int head;
1030
1031                 tail = pipe->curbuf + pipe->nrbufs;
1032                 if (tail < pipe->buffers)
1033                         tail = 0;
1034                 else
1035                         tail &= (pipe->buffers - 1);
1036
1037                 head = pipe->nrbufs - tail;
1038                 if (head)
1039                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1040                 if (tail)
1041                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1042         }
1043
1044         account_pipe_buffers(pipe, pipe->buffers, nr_pages);
1045         pipe->curbuf = 0;
1046         kfree(pipe->bufs);
1047         pipe->bufs = bufs;
1048         pipe->buffers = nr_pages;
1049         return nr_pages * PAGE_SIZE;
1050 }
1051
1052 /*
1053  * Currently we rely on the pipe array holding a power-of-2 number
1054  * of pages. Returns 0 on error.
1055  */
1056 static inline unsigned int round_pipe_size(unsigned int size)
1057 {
1058         unsigned long nr_pages;
1059
1060         if (size < pipe_min_size)
1061                 size = pipe_min_size;
1062
1063         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1064         if (nr_pages == 0)
1065                 return 0;
1066
1067         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1068 }
1069
1070 /*
1071  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1072  * will return an error.
1073  */
1074 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1075                  size_t *lenp, loff_t *ppos)
1076 {
1077         unsigned int rounded_pipe_max_size;
1078         int ret;
1079
1080         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1081         if (ret < 0 || !write)
1082                 return ret;
1083
1084         rounded_pipe_max_size = round_pipe_size(pipe_max_size);
1085         if (rounded_pipe_max_size == 0)
1086                 return -EINVAL;
1087
1088         pipe_max_size = rounded_pipe_max_size;
1089         return ret;
1090 }
1091
1092 /*
1093  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1094  * location, so checking ->i_pipe is not enough to verify that this is a
1095  * pipe.
1096  */
1097 struct pipe_inode_info *get_pipe_info(struct file *file)
1098 {
1099         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1100 }
1101
1102 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1103 {
1104         struct pipe_inode_info *pipe;
1105         long ret;
1106
1107         pipe = get_pipe_info(file);
1108         if (!pipe)
1109                 return -EBADF;
1110
1111         __pipe_lock(pipe);
1112
1113         switch (cmd) {
1114         case F_SETPIPE_SZ: {
1115                 unsigned int size, nr_pages;
1116
1117                 size = round_pipe_size(arg);
1118                 nr_pages = size >> PAGE_SHIFT;
1119
1120                 ret = -EINVAL;
1121                 if (!nr_pages)
1122                         goto out;
1123
1124                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1125                         ret = -EPERM;
1126                         goto out;
1127                 } else if ((too_many_pipe_buffers_hard(pipe->user) ||
1128                             too_many_pipe_buffers_soft(pipe->user)) &&
1129                            !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1130                         ret = -EPERM;
1131                         goto out;
1132                 }
1133                 ret = pipe_set_size(pipe, nr_pages);
1134                 break;
1135                 }
1136         case F_GETPIPE_SZ:
1137                 ret = pipe->buffers * PAGE_SIZE;
1138                 break;
1139         default:
1140                 ret = -EINVAL;
1141                 break;
1142         }
1143
1144 out:
1145         __pipe_unlock(pipe);
1146         return ret;
1147 }
1148
1149 static const struct super_operations pipefs_ops = {
1150         .destroy_inode = free_inode_nonrcu,
1151         .statfs = simple_statfs,
1152 };
1153
1154 /*
1155  * pipefs should _never_ be mounted by userland - too much of security hassle,
1156  * no real gain from having the whole whorehouse mounted. So we don't need
1157  * any operations on the root directory. However, we need a non-trivial
1158  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1159  */
1160 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1161                          int flags, const char *dev_name, void *data)
1162 {
1163         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1164                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1165 }
1166
1167 static struct file_system_type pipe_fs_type = {
1168         .name           = "pipefs",
1169         .mount          = pipefs_mount,
1170         .kill_sb        = kill_anon_super,
1171 };
1172
1173 static int __init init_pipe_fs(void)
1174 {
1175         int err = register_filesystem(&pipe_fs_type);
1176
1177         if (!err) {
1178                 pipe_mnt = kern_mount(&pipe_fs_type);
1179                 if (IS_ERR(pipe_mnt)) {
1180                         err = PTR_ERR(pipe_mnt);
1181                         unregister_filesystem(&pipe_fs_type);
1182                 }
1183         }
1184         return err;
1185 }
1186
1187 fs_initcall(init_pipe_fs);