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