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 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 197 {
 198         return try_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 anon_pipe_buf_nomerge_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 const struct pipe_buf_operations packet_pipe_buf_ops = {
 250         .can_merge = 0,
 251         .confirm = generic_pipe_buf_confirm,
 252         .release = anon_pipe_buf_release,
 253         .steal = anon_pipe_buf_steal,
 254         .get = generic_pipe_buf_get,
 255 };
 256
 257 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 258 {
 259         if (buf->ops == &anon_pipe_buf_ops)
 260                 buf->ops = &anon_pipe_buf_nomerge_ops;
 261 }
 262
 263 static ssize_t
 264 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 265 {
 266         size_t total_len = iov_iter_count(to);
 267         struct file *filp = iocb->ki_filp;
 268         struct pipe_inode_info *pipe = filp->private_data;
 269         int do_wakeup;
 270         ssize_t ret;
 271
 272         /* Null read succeeds. */
 273         if (unlikely(total_len == 0))
 274                 return 0;
 275
 276         do_wakeup = 0;
 277         ret = 0;
 278         __pipe_lock(pipe);
 279         for (;;) {
 280                 int bufs = pipe->nrbufs;
 281                 if (bufs) {
 282                         int curbuf = pipe->curbuf;
 283                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 284                         size_t chars = buf->len;
 285                         size_t written;
 286                         int error;
 287
 288                         if (chars > total_len)
 289                                 chars = total_len;
 290
 291                         error = pipe_buf_confirm(pipe, buf);
 292                         if (error) {
 293                                 if (!ret)
 294                                         ret = error;
 295                                 break;
 296                         }
 297
 298                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 299                         if (unlikely(written < chars)) {
 300                                 if (!ret)
 301                                         ret = -EFAULT;
 302                                 break;
 303                         }
 304                         ret += chars;
 305                         buf->offset += chars;
 306                         buf->len -= chars;
 307
 308                         /* Was it a packet buffer? Clean up and exit */
 309                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 310                                 total_len = chars;
 311                                 buf->len = 0;
 312                         }
 313
 314                         if (!buf->len) {
 315                                 pipe_buf_release(pipe, buf);
 316                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 317                                 pipe->curbuf = curbuf;
 318                                 pipe->nrbufs = --bufs;
 319                                 do_wakeup = 1;
 320                         }
 321                         total_len -= chars;
 322                         if (!total_len)
 323                                 break;  /* common path: read succeeded */
 324                 }
 325                 if (bufs)       /* More to do? */
 326                         continue;
 327                 if (!pipe->writers)
 328                         break;
 329                 if (!pipe->waiting_writers) {
 330                         /* syscall merging: Usually we must not sleep
 331                          * if O_NONBLOCK is set, or if we got some data.
 332                          * But if a writer sleeps in kernel space, then
 333                          * we can wait for that data without violating POSIX.
 334                          */
 335                         if (ret)
 336                                 break;
 337                         if (filp->f_flags & O_NONBLOCK) {
 338                                 ret = -EAGAIN;
 339                                 break;
 340                         }
 341                 }
 342                 if (signal_pending(current)) {
 343                         if (!ret)
 344                                 ret = -ERESTARTSYS;
 345                         break;
 346                 }
 347                 if (do_wakeup) {
 348                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 349                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 350                 }
 351                 pipe_wait(pipe);
 352         }
 353         __pipe_unlock(pipe);
 354
 355         /* Signal writers asynchronously that there is more room. */
 356         if (do_wakeup) {
 357                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 358                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 359         }
 360         if (ret > 0)
 361                 file_accessed(filp);
 362         return ret;
 363 }
 364
 365 static inline int is_packetized(struct file *file)
 366 {
 367         return (file->f_flags & O_DIRECT) != 0;
 368 }
 369
 370 static ssize_t
 371 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 372 {
 373         struct file *filp = iocb->ki_filp;
 374         struct pipe_inode_info *pipe = filp->private_data;
 375         ssize_t ret = 0;
 376         int do_wakeup = 0;
 377         size_t total_len = iov_iter_count(from);
 378         ssize_t chars;
 379
 380         /* Null write succeeds. */
 381         if (unlikely(total_len == 0))
 382                 return 0;
 383
 384         __pipe_lock(pipe);
 385
 386         if (!pipe->readers) {
 387                 send_sig(SIGPIPE, current, 0);
 388                 ret = -EPIPE;
 389                 goto out;
 390         }
 391
 392         /* We try to merge small writes */
 393         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 394         if (pipe->nrbufs && chars != 0) {
 395                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 396                                                         (pipe->buffers - 1);
 397                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 398                 int offset = buf->offset + buf->len;
 399
 400                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 401                         ret = pipe_buf_confirm(pipe, buf);
 402                         if (ret)
 403                                 goto out;
 404
 405                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 406                         if (unlikely(ret < chars)) {
 407                                 ret = -EFAULT;
 408                                 goto out;
 409                         }
 410                         do_wakeup = 1;
 411                         buf->len += ret;
 412                         if (!iov_iter_count(from))
 413                                 goto out;
 414                 }
 415         }
 416
 417         for (;;) {
 418                 int bufs;
 419
 420                 if (!pipe->readers) {
 421                         send_sig(SIGPIPE, current, 0);
 422                         if (!ret)
 423                                 ret = -EPIPE;
 424                         break;
 425                 }
 426                 bufs = pipe->nrbufs;
 427                 if (bufs < pipe->buffers) {
 428                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 429                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 430                         struct page *page = pipe->tmp_page;
 431                         int copied;
 432
 433                         if (!page) {
 434                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 435                                 if (unlikely(!page)) {
 436                                         ret = ret ? : -ENOMEM;
 437                                         break;
 438                                 }
 439                                 pipe->tmp_page = page;
 440                         }
 441                         /* Always wake up, even if the copy fails. Otherwise
 442                          * we lock up (O_NONBLOCK-)readers that sleep due to
 443                          * syscall merging.
 444                          * FIXME! Is this really true?
 445                          */
 446                         do_wakeup = 1;
 447                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 448                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 449                                 if (!ret)
 450                                         ret = -EFAULT;
 451                                 break;
 452                         }
 453                         ret += copied;
 454
 455                         /* Insert it into the buffer array */
 456                         buf->page = page;
 457                         buf->ops = &anon_pipe_buf_ops;
 458                         buf->offset = 0;
 459                         buf->len = copied;
 460                         buf->flags = 0;
 461                         if (is_packetized(filp)) {
 462                                 buf->ops = &packet_pipe_buf_ops;
 463                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 464                         }
 465                         pipe->nrbufs = ++bufs;
 466                         pipe->tmp_page = NULL;
 467
 468                         if (!iov_iter_count(from))
 469                                 break;
 470                 }
 471                 if (bufs < pipe->buffers)
 472                         continue;
 473                 if (filp->f_flags & O_NONBLOCK) {
 474                         if (!ret)
 475                                 ret = -EAGAIN;
 476                         break;
 477                 }
 478                 if (signal_pending(current)) {
 479                         if (!ret)
 480                                 ret = -ERESTARTSYS;
 481                         break;
 482                 }
 483                 if (do_wakeup) {
 484                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 485                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 486                         do_wakeup = 0;
 487                 }
 488                 pipe->waiting_writers++;
 489                 pipe_wait(pipe);
 490                 pipe->waiting_writers--;
 491         }
 492 out:
 493         __pipe_unlock(pipe);
 494         if (do_wakeup) {
 495                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 496                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 497         }
 498         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 499                 int err = file_update_time(filp);
 500                 if (err)
 501                         ret = err;
 502                 sb_end_write(file_inode(filp)->i_sb);
 503         }
 504         return ret;
 505 }
 506
 507 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 508 {
 509         struct pipe_inode_info *pipe = filp->private_data;
 510         int count, buf, nrbufs;
 511
 512         switch (cmd) {
 513                 case FIONREAD:
 514                         __pipe_lock(pipe);
 515                         count = 0;
 516                         buf = pipe->curbuf;
 517                         nrbufs = pipe->nrbufs;
 518                         while (--nrbufs >= 0) {
 519                                 count += pipe->bufs[buf].len;
 520                                 buf = (buf+1) & (pipe->buffers - 1);
 521                         }
 522                         __pipe_unlock(pipe);
 523
 524                         return put_user(count, (int __user *)arg);
 525                 default:
 526                         return -ENOIOCTLCMD;
 527         }
 528 }
 529
 530 /* No kernel lock held - fine */
 531 static unsigned int
 532 pipe_poll(struct file *filp, poll_table *wait)
 533 {
 534         unsigned int mask;
 535         struct pipe_inode_info *pipe = filp->private_data;
 536         int nrbufs;
 537
 538         poll_wait(filp, &pipe->wait, wait);
 539
 540         /* Reading only -- no need for acquiring the semaphore.  */
 541         nrbufs = pipe->nrbufs;
 542         mask = 0;
 543         if (filp->f_mode & FMODE_READ) {
 544                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 545                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 546                         mask |= POLLHUP;
 547         }
 548
 549         if (filp->f_mode & FMODE_WRITE) {
 550                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 551                 /*
 552                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 553                  * behave exactly like pipes for poll().
 554                  */
 555                 if (!pipe->readers)
 556                         mask |= POLLERR;
 557         }
 558
 559         return mask;
 560 }
 561
 562 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 563 {
 564         int kill = 0;
 565
 566         spin_lock(&inode->i_lock);
 567         if (!--pipe->files) {
 568                 inode->i_pipe = NULL;
 569                 kill = 1;
 570         }
 571         spin_unlock(&inode->i_lock);
 572
 573         if (kill)
 574                 free_pipe_info(pipe);
 575 }
 576
 577 static int
 578 pipe_release(struct inode *inode, struct file *file)
 579 {
 580         struct pipe_inode_info *pipe = file->private_data;
 581
 582         __pipe_lock(pipe);
 583         if (file->f_mode & FMODE_READ)
 584                 pipe->readers--;
 585         if (file->f_mode & FMODE_WRITE)
 586                 pipe->writers--;
 587
 588         if (pipe->readers || pipe->writers) {
 589                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 590                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 591                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 592         }
 593         __pipe_unlock(pipe);
 594
 595         put_pipe_info(inode, pipe);
 596         return 0;
 597 }
 598
 599 static int
 600 pipe_fasync(int fd, struct file *filp, int on)
 601 {
 602         struct pipe_inode_info *pipe = filp->private_data;
 603         int retval = 0;
 604
 605         __pipe_lock(pipe);
 606         if (filp->f_mode & FMODE_READ)
 607                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 608         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 609                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 610                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 611                         /* this can happen only if on == T */
 612                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 613         }
 614         __pipe_unlock(pipe);
 615         return retval;
 616 }
 617
 618 static unsigned long account_pipe_buffers(struct user_struct *user,
 619                                  unsigned long old, unsigned long new)
 620 {
 621         return atomic_long_add_return(new - old, &user->pipe_bufs);
 622 }
 623
 624 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 625 {
 626         return pipe_user_pages_soft && user_bufs > pipe_user_pages_soft;
 627 }
 628
 629 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 630 {
 631         return pipe_user_pages_hard && user_bufs > pipe_user_pages_hard;
 632 }
 633
 634 static bool is_unprivileged_user(void)
 635 {
 636         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 637 }
 638
 639 struct pipe_inode_info *alloc_pipe_info(void)
 640 {
 641         struct pipe_inode_info *pipe;
 642         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 643         struct user_struct *user = get_current_user();
 644         unsigned long user_bufs;
 645
 646         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 647         if (pipe == NULL)
 648                 goto out_free_uid;
 649
 650         if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 651                 pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 652
 653         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 654
 655         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 656                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 657                 pipe_bufs = 1;
 658         }
 659
 660         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 661                 goto out_revert_acct;
 662
 663         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 664                              GFP_KERNEL_ACCOUNT);
 665
 666         if (pipe->bufs) {
 667                 init_waitqueue_head(&pipe->wait);
 668                 pipe->r_counter = pipe->w_counter = 1;
 669                 pipe->buffers = pipe_bufs;
 670                 pipe->user = user;
 671                 mutex_init(&pipe->mutex);
 672                 return pipe;
 673         }
 674
 675 out_revert_acct:
 676         (void) account_pipe_buffers(user, pipe_bufs, 0);
 677         kfree(pipe);
 678 out_free_uid:
 679         free_uid(user);
 680         return NULL;
 681 }
 682
 683 void free_pipe_info(struct pipe_inode_info *pipe)
 684 {
 685         int i;
 686
 687         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 688         free_uid(pipe->user);
 689         for (i = 0; i < pipe->buffers; i++) {
 690                 struct pipe_buffer *buf = pipe->bufs + i;
 691                 if (buf->ops)
 692                         pipe_buf_release(pipe, buf);
 693         }
 694         if (pipe->tmp_page)
 695                 __free_page(pipe->tmp_page);
 696         kfree(pipe->bufs);
 697         kfree(pipe);
 698 }
 699
 700 static struct vfsmount *pipe_mnt __read_mostly;
 701
 702 /*
 703  * pipefs_dname() is called from d_path().
 704  */
 705 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 706 {
 707         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 708                                 d_inode(dentry)->i_ino);
 709 }
 710
 711 static const struct dentry_operations pipefs_dentry_operations = {
 712         .d_dname        = pipefs_dname,
 713 };
 714
 715 static struct inode * get_pipe_inode(void)
 716 {
 717         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 718         struct pipe_inode_info *pipe;
 719
 720         if (!inode)
 721                 goto fail_inode;
 722
 723         inode->i_ino = get_next_ino();
 724
 725         pipe = alloc_pipe_info();
 726         if (!pipe)
 727                 goto fail_iput;
 728
 729         inode->i_pipe = pipe;
 730         pipe->files = 2;
 731         pipe->readers = pipe->writers = 1;
 732         inode->i_fop = &pipefifo_fops;
 733
 734         /*
 735          * Mark the inode dirty from the very beginning,
 736          * that way it will never be moved to the dirty
 737          * list because "mark_inode_dirty()" will think
 738          * that it already _is_ on the dirty list.
 739          */
 740         inode->i_state = I_DIRTY;
 741         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 742         inode->i_uid = current_fsuid();
 743         inode->i_gid = current_fsgid();
 744         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 745
 746         return inode;
 747
 748 fail_iput:
 749         iput(inode);
 750
 751 fail_inode:
 752         return NULL;
 753 }
 754
 755 int create_pipe_files(struct file **res, int flags)
 756 {
 757         int err;
 758         struct inode *inode = get_pipe_inode();
 759         struct file *f;
 760         struct path path;
 761         static struct qstr name = { .name = "" };
 762
 763         if (!inode)
 764                 return -ENFILE;
 765
 766         err = -ENOMEM;
 767         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
 768         if (!path.dentry)
 769                 goto err_inode;
 770         path.mnt = mntget(pipe_mnt);
 771
 772         d_instantiate(path.dentry, inode);
 773
 774         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 775         if (IS_ERR(f)) {
 776                 err = PTR_ERR(f);
 777                 goto err_dentry;
 778         }
 779
 780         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 781         f->private_data = inode->i_pipe;
 782
 783         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 784         if (IS_ERR(res[0])) {
 785                 err = PTR_ERR(res[0]);
 786                 goto err_file;
 787         }
 788
 789         path_get(&path);
 790         res[0]->private_data = inode->i_pipe;
 791         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 792         res[1] = f;
 793         return 0;
 794
 795 err_file:
 796         put_filp(f);
 797 err_dentry:
 798         free_pipe_info(inode->i_pipe);
 799         path_put(&path);
 800         return err;
 801
 802 err_inode:
 803         free_pipe_info(inode->i_pipe);
 804         iput(inode);
 805         return err;
 806 }
 807
 808 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 809 {
 810         int error;
 811         int fdw, fdr;
 812
 813         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 814                 return -EINVAL;
 815
 816         error = create_pipe_files(files, flags);
 817         if (error)
 818                 return error;
 819
 820         error = get_unused_fd_flags(flags);
 821         if (error < 0)
 822                 goto err_read_pipe;
 823         fdr = error;
 824
 825         error = get_unused_fd_flags(flags);
 826         if (error < 0)
 827                 goto err_fdr;
 828         fdw = error;
 829
 830         audit_fd_pair(fdr, fdw);
 831         fd[0] = fdr;
 832         fd[1] = fdw;
 833         return 0;
 834
 835  err_fdr:
 836         put_unused_fd(fdr);
 837  err_read_pipe:
 838         fput(files[0]);
 839         fput(files[1]);
 840         return error;
 841 }
 842
 843 int do_pipe_flags(int *fd, int flags)
 844 {
 845         struct file *files[2];
 846         int error = __do_pipe_flags(fd, files, flags);
 847         if (!error) {
 848                 fd_install(fd[0], files[0]);
 849                 fd_install(fd[1], files[1]);
 850         }
 851         return error;
 852 }
 853
 854 /*
 855  * sys_pipe() is the normal C calling standard for creating
 856  * a pipe. It's not the way Unix traditionally does this, though.
 857  */
 858 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 859 {
 860         struct file *files[2];
 861         int fd[2];
 862         int error;
 863
 864         error = __do_pipe_flags(fd, files, flags);
 865         if (!error) {
 866                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 867                         fput(files[0]);
 868                         fput(files[1]);
 869                         put_unused_fd(fd[0]);
 870                         put_unused_fd(fd[1]);
 871                         error = -EFAULT;
 872                 } else {
 873                         fd_install(fd[0], files[0]);
 874                         fd_install(fd[1], files[1]);
 875                 }
 876         }
 877         return error;
 878 }
 879
 880 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 881 {
 882         return sys_pipe2(fildes, 0);
 883 }
 884
 885 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 886 {
 887         int cur = *cnt;
 888
 889         while (cur == *cnt) {
 890                 pipe_wait(pipe);
 891                 if (signal_pending(current))
 892                         break;
 893         }
 894         return cur == *cnt ? -ERESTARTSYS : 0;
 895 }
 896
 897 static void wake_up_partner(struct pipe_inode_info *pipe)
 898 {
 899         wake_up_interruptible(&pipe->wait);
 900 }
 901
 902 static int fifo_open(struct inode *inode, struct file *filp)
 903 {
 904         struct pipe_inode_info *pipe;
 905         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 906         int ret;
 907
 908         filp->f_version = 0;
 909
 910         spin_lock(&inode->i_lock);
 911         if (inode->i_pipe) {
 912                 pipe = inode->i_pipe;
 913                 pipe->files++;
 914                 spin_unlock(&inode->i_lock);
 915         } else {
 916                 spin_unlock(&inode->i_lock);
 917                 pipe = alloc_pipe_info();
 918                 if (!pipe)
 919                         return -ENOMEM;
 920                 pipe->files = 1;
 921                 spin_lock(&inode->i_lock);
 922                 if (unlikely(inode->i_pipe)) {
 923                         inode->i_pipe->files++;
 924                         spin_unlock(&inode->i_lock);
 925                         free_pipe_info(pipe);
 926                         pipe = inode->i_pipe;
 927                 } else {
 928                         inode->i_pipe = pipe;
 929                         spin_unlock(&inode->i_lock);
 930                 }
 931         }
 932         filp->private_data = pipe;
 933         /* OK, we have a pipe and it's pinned down */
 934
 935         __pipe_lock(pipe);
 936
 937         /* We can only do regular read/write on fifos */
 938         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 939
 940         switch (filp->f_mode) {
 941         case FMODE_READ:
 942         /*
 943          *  O_RDONLY
 944          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 945          *  opened, even when there is no process writing the FIFO.
 946          */
 947                 pipe->r_counter++;
 948                 if (pipe->readers++ == 0)
 949                         wake_up_partner(pipe);
 950
 951                 if (!is_pipe && !pipe->writers) {
 952                         if ((filp->f_flags & O_NONBLOCK)) {
 953                                 /* suppress POLLHUP until we have
 954                                  * seen a writer */
 955                                 filp->f_version = pipe->w_counter;
 956                         } else {
 957                                 if (wait_for_partner(pipe, &pipe->w_counter))
 958                                         goto err_rd;
 959                         }
 960                 }
 961                 break;
 962
 963         case FMODE_WRITE:
 964         /*
 965          *  O_WRONLY
 966          *  POSIX.1 says that O_NONBLOCK means return -1 with
 967          *  errno=ENXIO when there is no process reading the FIFO.
 968          */
 969                 ret = -ENXIO;
 970                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 971                         goto err;
 972
 973                 pipe->w_counter++;
 974                 if (!pipe->writers++)
 975                         wake_up_partner(pipe);
 976
 977                 if (!is_pipe && !pipe->readers) {
 978                         if (wait_for_partner(pipe, &pipe->r_counter))
 979                                 goto err_wr;
 980                 }
 981                 break;
 982
 983         case FMODE_READ | FMODE_WRITE:
 984         /*
 985          *  O_RDWR
 986          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 987          *  This implementation will NEVER block on a O_RDWR open, since
 988          *  the process can at least talk to itself.
 989          */
 990
 991                 pipe->readers++;
 992                 pipe->writers++;
 993                 pipe->r_counter++;
 994                 pipe->w_counter++;
 995                 if (pipe->readers == 1 || pipe->writers == 1)
 996                         wake_up_partner(pipe);
 997                 break;
 998
 999         default:
1000                 ret = -EINVAL;
1001                 goto err;
1002         }
1003
1004         /* Ok! */
1005         __pipe_unlock(pipe);
1006         return 0;
1007
1008 err_rd:
1009         if (!--pipe->readers)
1010                 wake_up_interruptible(&pipe->wait);
1011         ret = -ERESTARTSYS;
1012         goto err;
1013
1014 err_wr:
1015         if (!--pipe->writers)
1016                 wake_up_interruptible(&pipe->wait);
1017         ret = -ERESTARTSYS;
1018         goto err;
1019
1020 err:
1021         __pipe_unlock(pipe);
1022
1023         put_pipe_info(inode, pipe);
1024         return ret;
1025 }
1026
1027 const struct file_operations pipefifo_fops = {
1028         .open           = fifo_open,
1029         .llseek         = no_llseek,
1030         .read_iter      = pipe_read,
1031         .write_iter     = pipe_write,
1032         .poll           = pipe_poll,
1033         .unlocked_ioctl = pipe_ioctl,
1034         .release        = pipe_release,
1035         .fasync         = pipe_fasync,
1036 };
1037
1038 /*
1039  * Currently we rely on the pipe array holding a power-of-2 number
1040  * of pages. Returns 0 on error.
1041  */
1042 static inline unsigned int round_pipe_size(unsigned int size)
1043 {
1044         unsigned long nr_pages;
1045
1046         if (size < pipe_min_size)
1047                 size = pipe_min_size;
1048
1049         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1050         if (nr_pages == 0)
1051                 return 0;
1052
1053         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1054 }
1055
1056 /*
1057  * Allocate a new array of pipe buffers and copy the info over. Returns the
1058  * pipe size if successful, or return -ERROR on error.
1059  */
1060 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1061 {
1062         struct pipe_buffer *bufs;
1063         unsigned int size, nr_pages;
1064         unsigned long user_bufs;
1065         long ret = 0;
1066
1067         size = round_pipe_size(arg);
1068         if (size == 0)
1069                 return -EINVAL;
1070         nr_pages = size >> PAGE_SHIFT;
1071
1072         if (!nr_pages)
1073                 return -EINVAL;
1074
1075         /*
1076          * If trying to increase the pipe capacity, check that an
1077          * unprivileged user is not trying to exceed various limits
1078          * (soft limit check here, hard limit check just below).
1079          * Decreasing the pipe capacity is always permitted, even
1080          * if the user is currently over a limit.
1081          */
1082         if (nr_pages > pipe->buffers &&
1083                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1084                 return -EPERM;
1085
1086         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1087
1088         if (nr_pages > pipe->buffers &&
1089                         (too_many_pipe_buffers_hard(user_bufs) ||
1090                          too_many_pipe_buffers_soft(user_bufs)) &&
1091                         is_unprivileged_user()) {
1092                 ret = -EPERM;
1093                 goto out_revert_acct;
1094         }
1095
1096         /*
1097          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1098          * expect a lot of shrink+grow operations, just free and allocate
1099          * again like we would do for growing. If the pipe currently
1100          * contains more buffers than arg, then return busy.
1101          */
1102         if (nr_pages < pipe->nrbufs) {
1103                 ret = -EBUSY;
1104                 goto out_revert_acct;
1105         }
1106
1107         bufs = kcalloc(nr_pages, sizeof(*bufs),
1108                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1109         if (unlikely(!bufs)) {
1110                 ret = -ENOMEM;
1111                 goto out_revert_acct;
1112         }
1113
1114         /*
1115          * The pipe array wraps around, so just start the new one at zero
1116          * and adjust the indexes.
1117          */
1118         if (pipe->nrbufs) {
1119                 unsigned int tail;
1120                 unsigned int head;
1121
1122                 tail = pipe->curbuf + pipe->nrbufs;
1123                 if (tail < pipe->buffers)
1124                         tail = 0;
1125                 else
1126                         tail &= (pipe->buffers - 1);
1127
1128                 head = pipe->nrbufs - tail;
1129                 if (head)
1130                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1131                 if (tail)
1132                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1133         }
1134
1135         pipe->curbuf = 0;
1136         kfree(pipe->bufs);
1137         pipe->bufs = bufs;
1138         pipe->buffers = nr_pages;
1139         return nr_pages * PAGE_SIZE;
1140
1141 out_revert_acct:
1142         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1143         return ret;
1144 }
1145
1146 /*
1147  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1148  * will return an error.
1149  */
1150 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1151                  size_t *lenp, loff_t *ppos)
1152 {
1153         unsigned int rounded_pipe_max_size;
1154         int ret;
1155
1156         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1157         if (ret < 0 || !write)
1158                 return ret;
1159
1160         rounded_pipe_max_size = round_pipe_size(pipe_max_size);
1161         if (rounded_pipe_max_size == 0)
1162                 return -EINVAL;
1163
1164         pipe_max_size = rounded_pipe_max_size;
1165         return ret;
1166 }
1167
1168 /*
1169  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1170  * location, so checking ->i_pipe is not enough to verify that this is a
1171  * pipe.
1172  */
1173 struct pipe_inode_info *get_pipe_info(struct file *file)
1174 {
1175         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1176 }
1177
1178 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1179 {
1180         struct pipe_inode_info *pipe;
1181         long ret;
1182
1183         pipe = get_pipe_info(file);
1184         if (!pipe)
1185                 return -EBADF;
1186
1187         __pipe_lock(pipe);
1188
1189         switch (cmd) {
1190         case F_SETPIPE_SZ:
1191                 ret = pipe_set_size(pipe, arg);
1192                 break;
1193         case F_GETPIPE_SZ:
1194                 ret = pipe->buffers * PAGE_SIZE;
1195                 break;
1196         default:
1197                 ret = -EINVAL;
1198                 break;
1199         }
1200
1201         __pipe_unlock(pipe);
1202         return ret;
1203 }
1204
1205 static const struct super_operations pipefs_ops = {
1206         .destroy_inode = free_inode_nonrcu,
1207         .statfs = simple_statfs,
1208 };
1209
1210 /*
1211  * pipefs should _never_ be mounted by userland - too much of security hassle,
1212  * no real gain from having the whole whorehouse mounted. So we don't need
1213  * any operations on the root directory. However, we need a non-trivial
1214  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1215  */
1216 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1217                          int flags, const char *dev_name, void *data)
1218 {
1219         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1220                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1221 }
1222
1223 static struct file_system_type pipe_fs_type = {
1224         .name           = "pipefs",
1225         .mount          = pipefs_mount,
1226         .kill_sb        = kill_anon_super,
1227 };
1228
1229 static int __init init_pipe_fs(void)
1230 {
1231         int err = register_filesystem(&pipe_fs_type);
1232
1233         if (!err) {
1234                 pipe_mnt = kern_mount(&pipe_fs_type);
1235                 if (IS_ERR(pipe_mnt)) {
1236                         err = PTR_ERR(pipe_mnt);
1237                         unregister_filesystem(&pipe_fs_type);
1238                 }
1239         }
1240         return err;
1241 }
1242
1243 fs_initcall(init_pipe_fs);