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