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