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