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