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/pipe_fs_i.h>
  17 #include <linux/uio.h>
  18 #include <linux/highmem.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/audit.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/fcntl.h>
  23
  24 #include <asm/uaccess.h>
  25 #include <asm/ioctls.h>
  26
  27 /*
  28  * The max size that a non-root user is allowed to grow the pipe. Can
  29  * be set by root in /proc/sys/fs/pipe-max-size
  30  */
  31 unsigned int pipe_max_size = 1048576;
  32
  33 /*
  34  * Minimum pipe size, as required by POSIX
  35  */
  36 unsigned int pipe_min_size = PAGE_SIZE;
  37
  38 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  39  * matches default values.
  40  */
  41 unsigned long pipe_user_pages_hard;
  42 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  43
  44 /*
  45  * We use a start+len construction, which provides full use of the
  46  * allocated memory.
  47  * -- Florian Coosmann (FGC)
  48  *
  49  * Reads with count = 0 should always return 0.
  50  * -- Julian Bradfield 1999-06-07.
  51  *
  52  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  53  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  54  *
  55  * pipe_read & write cleanup
  56  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  57  */
  58
  59 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  60 {
  61         if (pipe->inode)
  62                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
  63 }
  64
  65 void pipe_lock(struct pipe_inode_info *pipe)
  66 {
  67         /*
  68          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  69          */
  70         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  71 }
  72 EXPORT_SYMBOL(pipe_lock);
  73
  74 void pipe_unlock(struct pipe_inode_info *pipe)
  75 {
  76         if (pipe->inode)
  77                 mutex_unlock(&pipe->inode->i_mutex);
  78 }
  79 EXPORT_SYMBOL(pipe_unlock);
  80
  81 void pipe_double_lock(struct pipe_inode_info *pipe1,
  82                       struct pipe_inode_info *pipe2)
  83 {
  84         BUG_ON(pipe1 == pipe2);
  85
  86         if (pipe1 < pipe2) {
  87                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  88                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  89         } else {
  90                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
  91                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
  92         }
  93 }
  94
  95 /* Drop the inode semaphore and wait for a pipe event, atomically */
  96 void pipe_wait(struct pipe_inode_info *pipe)
  97 {
  98         DEFINE_WAIT(wait);
  99
 100         /*
 101          * Pipes are system-local resources, so sleeping on them
 102          * is considered a noninteractive wait:
 103          */
 104         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 105         pipe_unlock(pipe);
 106         schedule();
 107         finish_wait(&pipe->wait, &wait);
 108         pipe_lock(pipe);
 109 }
 110
 111 static int
 112 pipe_iov_copy_from_user(void *addr, int *offset, struct iovec *iov,
 113                         size_t *remaining, int atomic)
 114 {
 115         unsigned long copy;
 116
 117         while (*remaining > 0) {
 118                 while (!iov->iov_len)
 119                         iov++;
 120                 copy = min_t(unsigned long, *remaining, iov->iov_len);
 121
 122                 if (atomic) {
 123                         if (__copy_from_user_inatomic(addr + *offset,
 124                                                       iov->iov_base, copy))
 125                                 return -EFAULT;
 126                 } else {
 127                         if (copy_from_user(addr + *offset,
 128                                            iov->iov_base, copy))
 129                                 return -EFAULT;
 130                 }
 131                 *offset += copy;
 132                 *remaining -= copy;
 133                 iov->iov_base += copy;
 134                 iov->iov_len -= copy;
 135         }
 136         return 0;
 137 }
 138
 139 static int
 140 pipe_iov_copy_to_user(struct iovec *iov, void *addr, int *offset,
 141                       size_t *remaining, int atomic)
 142 {
 143         unsigned long copy;
 144
 145         while (*remaining > 0) {
 146                 while (!iov->iov_len)
 147                         iov++;
 148                 copy = min_t(unsigned long, *remaining, iov->iov_len);
 149
 150                 if (atomic) {
 151                         if (__copy_to_user_inatomic(iov->iov_base,
 152                                                     addr + *offset, copy))
 153                                 return -EFAULT;
 154                 } else {
 155                         if (copy_to_user(iov->iov_base,
 156                                          addr + *offset, copy))
 157                                 return -EFAULT;
 158                 }
 159                 *offset += copy;
 160                 *remaining -= copy;
 161                 iov->iov_base += copy;
 162                 iov->iov_len -= copy;
 163         }
 164         return 0;
 165 }
 166
 167 /*
 168  * Attempt to pre-fault in the user memory, so we can use atomic copies.
 169  * Returns the number of bytes not faulted in.
 170  */
 171 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 172 {
 173         while (!iov->iov_len)
 174                 iov++;
 175
 176         while (len > 0) {
 177                 unsigned long this_len;
 178
 179                 this_len = min_t(unsigned long, len, iov->iov_len);
 180                 if (fault_in_pages_writeable(iov->iov_base, this_len))
 181                         break;
 182
 183                 len -= this_len;
 184                 iov++;
 185         }
 186
 187         return len;
 188 }
 189
 190 /*
 191  * Pre-fault in the user memory, so we can use atomic copies.
 192  */
 193 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 194 {
 195         while (!iov->iov_len)
 196                 iov++;
 197
 198         while (len > 0) {
 199                 unsigned long this_len;
 200
 201                 this_len = min_t(unsigned long, len, iov->iov_len);
 202                 fault_in_pages_readable(iov->iov_base, this_len);
 203                 len -= this_len;
 204                 iov++;
 205         }
 206 }
 207
 208 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 209                                   struct pipe_buffer *buf)
 210 {
 211         struct page *page = buf->page;
 212
 213         /*
 214          * If nobody else uses this page, and we don't already have a
 215          * temporary page, let's keep track of it as a one-deep
 216          * allocation cache. (Otherwise just release our reference to it)
 217          */
 218         if (page_count(page) == 1 && !pipe->tmp_page)
 219                 pipe->tmp_page = page;
 220         else
 221                 page_cache_release(page);
 222 }
 223
 224 /**
 225  * generic_pipe_buf_map - virtually map a pipe buffer
 226  * @pipe:       the pipe that the buffer belongs to
 227  * @buf:        the buffer that should be mapped
 228  * @atomic:     whether to use an atomic map
 229  *
 230  * Description:
 231  *      This function returns a kernel virtual address mapping for the
 232  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 233  *      and the caller has to be careful not to fault before calling
 234  *      the unmap function.
 235  *
 236  *      Note that this function occupies KM_USER0 if @atomic != 0.
 237  */
 238 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 239                            struct pipe_buffer *buf, int atomic)
 240 {
 241         if (atomic) {
 242                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 243                 return kmap_atomic(buf->page, KM_USER0);
 244         }
 245
 246         return kmap(buf->page);
 247 }
 248 EXPORT_SYMBOL(generic_pipe_buf_map);
 249
 250 /**
 251  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 252  * @pipe:       the pipe that the buffer belongs to
 253  * @buf:        the buffer that should be unmapped
 254  * @map_data:   the data that the mapping function returned
 255  *
 256  * Description:
 257  *      This function undoes the mapping that ->map() provided.
 258  */
 259 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 260                             struct pipe_buffer *buf, void *map_data)
 261 {
 262         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 263                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 264                 kunmap_atomic(map_data, KM_USER0);
 265         } else
 266                 kunmap(buf->page);
 267 }
 268 EXPORT_SYMBOL(generic_pipe_buf_unmap);
 269
 270 /**
 271  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 272  * @pipe:       the pipe that the buffer belongs to
 273  * @buf:        the buffer to attempt to steal
 274  *
 275  * Description:
 276  *      This function attempts to steal the &struct page attached to
 277  *      @buf. If successful, this function returns 0 and returns with
 278  *      the page locked. The caller may then reuse the page for whatever
 279  *      he wishes; the typical use is insertion into a different file
 280  *      page cache.
 281  */
 282 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 283                            struct pipe_buffer *buf)
 284 {
 285         struct page *page = buf->page;
 286
 287         /*
 288          * A reference of one is golden, that means that the owner of this
 289          * page is the only one holding a reference to it. lock the page
 290          * and return OK.
 291          */
 292         if (page_count(page) == 1) {
 293                 lock_page(page);
 294                 return 0;
 295         }
 296
 297         return 1;
 298 }
 299 EXPORT_SYMBOL(generic_pipe_buf_steal);
 300
 301 /**
 302  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 303  * @pipe:       the pipe that the buffer belongs to
 304  * @buf:        the buffer to get a reference to
 305  *
 306  * Description:
 307  *      This function grabs an extra reference to @buf. It's used in
 308  *      in the tee() system call, when we duplicate the buffers in one
 309  *      pipe into another.
 310  */
 311 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 312 {
 313         page_cache_get(buf->page);
 314 }
 315 EXPORT_SYMBOL(generic_pipe_buf_get);
 316
 317 /**
 318  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 319  * @info:       the pipe that the buffer belongs to
 320  * @buf:        the buffer to confirm
 321  *
 322  * Description:
 323  *      This function does nothing, because the generic pipe code uses
 324  *      pages that are always good when inserted into the pipe.
 325  */
 326 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 327                              struct pipe_buffer *buf)
 328 {
 329         return 0;
 330 }
 331 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 332
 333 /**
 334  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 335  * @pipe:       the pipe that the buffer belongs to
 336  * @buf:        the buffer to put a reference to
 337  *
 338  * Description:
 339  *      This function releases a reference to @buf.
 340  */
 341 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 342                               struct pipe_buffer *buf)
 343 {
 344         page_cache_release(buf->page);
 345 }
 346 EXPORT_SYMBOL(generic_pipe_buf_release);
 347
 348 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 349         .can_merge = 1,
 350         .map = generic_pipe_buf_map,
 351         .unmap = generic_pipe_buf_unmap,
 352         .confirm = generic_pipe_buf_confirm,
 353         .release = anon_pipe_buf_release,
 354         .steal = generic_pipe_buf_steal,
 355         .get = generic_pipe_buf_get,
 356 };
 357
 358 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 359         .can_merge = 0,
 360         .map = generic_pipe_buf_map,
 361         .unmap = generic_pipe_buf_unmap,
 362         .confirm = generic_pipe_buf_confirm,
 363         .release = anon_pipe_buf_release,
 364         .steal = generic_pipe_buf_steal,
 365         .get = generic_pipe_buf_get,
 366 };
 367
 368 static ssize_t
 369 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 370            unsigned long nr_segs, loff_t pos)
 371 {
 372         struct file *filp = iocb->ki_filp;
 373         struct inode *inode = filp->f_path.dentry->d_inode;
 374         struct pipe_inode_info *pipe;
 375         int do_wakeup;
 376         ssize_t ret;
 377         struct iovec *iov = (struct iovec *)_iov;
 378         size_t total_len;
 379
 380         total_len = iov_length(iov, nr_segs);
 381         /* Null read succeeds. */
 382         if (unlikely(total_len == 0))
 383                 return 0;
 384
 385         do_wakeup = 0;
 386         ret = 0;
 387         mutex_lock(&inode->i_mutex);
 388         pipe = inode->i_pipe;
 389         for (;;) {
 390                 int bufs = pipe->nrbufs;
 391                 if (bufs) {
 392                         int curbuf = pipe->curbuf;
 393                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 394                         const struct pipe_buf_operations *ops = buf->ops;
 395                         void *addr;
 396                         size_t chars = buf->len, remaining;
 397                         int error, atomic;
 398                         int offset;
 399
 400                         if (chars > total_len)
 401                                 chars = total_len;
 402
 403                         error = ops->confirm(pipe, buf);
 404                         if (error) {
 405                                 if (!ret)
 406                                         ret = error;
 407                                 break;
 408                         }
 409
 410                         atomic = !iov_fault_in_pages_write(iov, chars);
 411                         remaining = chars;
 412                         offset = buf->offset;
 413 redo:
 414                         addr = ops->map(pipe, buf, atomic);
 415                         error = pipe_iov_copy_to_user(iov, addr, &offset,
 416                                                       &remaining, atomic);
 417                         ops->unmap(pipe, buf, addr);
 418                         if (unlikely(error)) {
 419                                 /*
 420                                  * Just retry with the slow path if we failed.
 421                                  */
 422                                 if (atomic) {
 423                                         atomic = 0;
 424                                         goto redo;
 425                                 }
 426                                 if (!ret)
 427                                         ret = error;
 428                                 break;
 429                         }
 430                         ret += chars;
 431                         buf->offset += chars;
 432                         buf->len -= chars;
 433
 434                         /* Was it a packet buffer? Clean up and exit */
 435                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 436                                 total_len = chars;
 437                                 buf->len = 0;
 438                         }
 439
 440                         if (!buf->len) {
 441                                 buf->ops = NULL;
 442                                 ops->release(pipe, buf);
 443                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 444                                 pipe->curbuf = curbuf;
 445                                 pipe->nrbufs = --bufs;
 446                                 do_wakeup = 1;
 447                         }
 448                         total_len -= chars;
 449                         if (!total_len)
 450                                 break;  /* common path: read succeeded */
 451                 }
 452                 if (bufs)       /* More to do? */
 453                         continue;
 454                 if (!pipe->writers)
 455                         break;
 456                 if (!pipe->waiting_writers) {
 457                         /* syscall merging: Usually we must not sleep
 458                          * if O_NONBLOCK is set, or if we got some data.
 459                          * But if a writer sleeps in kernel space, then
 460                          * we can wait for that data without violating POSIX.
 461                          */
 462                         if (ret)
 463                                 break;
 464                         if (filp->f_flags & O_NONBLOCK) {
 465                                 ret = -EAGAIN;
 466                                 break;
 467                         }
 468                 }
 469                 if (signal_pending(current)) {
 470                         if (!ret)
 471                                 ret = -ERESTARTSYS;
 472                         break;
 473                 }
 474                 if (do_wakeup) {
 475                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 476                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 477                 }
 478                 pipe_wait(pipe);
 479         }
 480         mutex_unlock(&inode->i_mutex);
 481
 482         /* Signal writers asynchronously that there is more room. */
 483         if (do_wakeup) {
 484                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 485                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 486         }
 487         if (ret > 0)
 488                 file_accessed(filp);
 489         return ret;
 490 }
 491
 492 static inline int is_packetized(struct file *file)
 493 {
 494         return (file->f_flags & O_DIRECT) != 0;
 495 }
 496
 497 static ssize_t
 498 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 499             unsigned long nr_segs, loff_t ppos)
 500 {
 501         struct file *filp = iocb->ki_filp;
 502         struct inode *inode = filp->f_path.dentry->d_inode;
 503         struct pipe_inode_info *pipe;
 504         ssize_t ret;
 505         int do_wakeup;
 506         struct iovec *iov = (struct iovec *)_iov;
 507         size_t total_len;
 508         ssize_t chars;
 509
 510         total_len = iov_length(iov, nr_segs);
 511         /* Null write succeeds. */
 512         if (unlikely(total_len == 0))
 513                 return 0;
 514
 515         do_wakeup = 0;
 516         ret = 0;
 517         mutex_lock(&inode->i_mutex);
 518         pipe = inode->i_pipe;
 519
 520         if (!pipe->readers) {
 521                 send_sig(SIGPIPE, current, 0);
 522                 ret = -EPIPE;
 523                 goto out;
 524         }
 525
 526         /* We try to merge small writes */
 527         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 528         if (pipe->nrbufs && chars != 0) {
 529                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 530                                                         (pipe->buffers - 1);
 531                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 532                 const struct pipe_buf_operations *ops = buf->ops;
 533                 int offset = buf->offset + buf->len;
 534
 535                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 536                         int error, atomic = 1;
 537                         void *addr;
 538                         size_t remaining = chars;
 539
 540                         error = ops->confirm(pipe, buf);
 541                         if (error)
 542                                 goto out;
 543
 544                         iov_fault_in_pages_read(iov, chars);
 545 redo1:
 546                         addr = ops->map(pipe, buf, atomic);
 547                         error = pipe_iov_copy_from_user(addr, &offset, iov,
 548                                                         &remaining, atomic);
 549                         ops->unmap(pipe, buf, addr);
 550                         ret = error;
 551                         do_wakeup = 1;
 552                         if (error) {
 553                                 if (atomic) {
 554                                         atomic = 0;
 555                                         goto redo1;
 556                                 }
 557                                 goto out;
 558                         }
 559                         buf->len += chars;
 560                         total_len -= chars;
 561                         ret = chars;
 562                         if (!total_len)
 563                                 goto out;
 564                 }
 565         }
 566
 567         for (;;) {
 568                 int bufs;
 569
 570                 if (!pipe->readers) {
 571                         send_sig(SIGPIPE, current, 0);
 572                         if (!ret)
 573                                 ret = -EPIPE;
 574                         break;
 575                 }
 576                 bufs = pipe->nrbufs;
 577                 if (bufs < pipe->buffers) {
 578                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 579                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 580                         struct page *page = pipe->tmp_page;
 581                         char *src;
 582                         int error, atomic = 1;
 583                         int offset = 0;
 584                         size_t remaining;
 585
 586                         if (!page) {
 587                                 page = alloc_page(GFP_HIGHUSER);
 588                                 if (unlikely(!page)) {
 589                                         ret = ret ? : -ENOMEM;
 590                                         break;
 591                                 }
 592                                 pipe->tmp_page = page;
 593                         }
 594                         /* Always wake up, even if the copy fails. Otherwise
 595                          * we lock up (O_NONBLOCK-)readers that sleep due to
 596                          * syscall merging.
 597                          * FIXME! Is this really true?
 598                          */
 599                         do_wakeup = 1;
 600                         chars = PAGE_SIZE;
 601                         if (chars > total_len)
 602                                 chars = total_len;
 603
 604                         iov_fault_in_pages_read(iov, chars);
 605                         remaining = chars;
 606 redo2:
 607                         if (atomic)
 608                                 src = kmap_atomic(page, KM_USER0);
 609                         else
 610                                 src = kmap(page);
 611
 612                         error = pipe_iov_copy_from_user(src, &offset, iov,
 613                                                         &remaining, atomic);
 614                         if (atomic)
 615                                 kunmap_atomic(src, KM_USER0);
 616                         else
 617                                 kunmap(page);
 618
 619                         if (unlikely(error)) {
 620                                 if (atomic) {
 621                                         atomic = 0;
 622                                         goto redo2;
 623                                 }
 624                                 if (!ret)
 625                                         ret = error;
 626                                 break;
 627                         }
 628                         ret += chars;
 629
 630                         /* Insert it into the buffer array */
 631                         buf->page = page;
 632                         buf->ops = &anon_pipe_buf_ops;
 633                         buf->offset = 0;
 634                         buf->len = chars;
 635                         buf->flags = 0;
 636                         if (is_packetized(filp)) {
 637                                 buf->ops = &packet_pipe_buf_ops;
 638                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 639                         }
 640                         pipe->nrbufs = ++bufs;
 641                         pipe->tmp_page = NULL;
 642
 643                         total_len -= chars;
 644                         if (!total_len)
 645                                 break;
 646                 }
 647                 if (bufs < pipe->buffers)
 648                         continue;
 649                 if (filp->f_flags & O_NONBLOCK) {
 650                         if (!ret)
 651                                 ret = -EAGAIN;
 652                         break;
 653                 }
 654                 if (signal_pending(current)) {
 655                         if (!ret)
 656                                 ret = -ERESTARTSYS;
 657                         break;
 658                 }
 659                 if (do_wakeup) {
 660                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 661                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 662                         do_wakeup = 0;
 663                 }
 664                 pipe->waiting_writers++;
 665                 pipe_wait(pipe);
 666                 pipe->waiting_writers--;
 667         }
 668 out:
 669         mutex_unlock(&inode->i_mutex);
 670         if (do_wakeup) {
 671                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 672                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 673         }
 674         if (ret > 0)
 675                 file_update_time(filp);
 676         return ret;
 677 }
 678
 679 static ssize_t
 680 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 681 {
 682         return -EBADF;
 683 }
 684
 685 static ssize_t
 686 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 687            loff_t *ppos)
 688 {
 689         return -EBADF;
 690 }
 691
 692 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 693 {
 694         struct inode *inode = filp->f_path.dentry->d_inode;
 695         struct pipe_inode_info *pipe;
 696         int count, buf, nrbufs;
 697
 698         switch (cmd) {
 699                 case FIONREAD:
 700                         mutex_lock(&inode->i_mutex);
 701                         pipe = inode->i_pipe;
 702                         count = 0;
 703                         buf = pipe->curbuf;
 704                         nrbufs = pipe->nrbufs;
 705                         while (--nrbufs >= 0) {
 706                                 count += pipe->bufs[buf].len;
 707                                 buf = (buf+1) & (pipe->buffers - 1);
 708                         }
 709                         mutex_unlock(&inode->i_mutex);
 710
 711                         return put_user(count, (int __user *)arg);
 712                 default:
 713                         return -EINVAL;
 714         }
 715 }
 716
 717 /* No kernel lock held - fine */
 718 static unsigned int
 719 pipe_poll(struct file *filp, poll_table *wait)
 720 {
 721         unsigned int mask;
 722         struct inode *inode = filp->f_path.dentry->d_inode;
 723         struct pipe_inode_info *pipe = inode->i_pipe;
 724         int nrbufs;
 725
 726         poll_wait(filp, &pipe->wait, wait);
 727
 728         /* Reading only -- no need for acquiring the semaphore.  */
 729         nrbufs = pipe->nrbufs;
 730         mask = 0;
 731         if (filp->f_mode & FMODE_READ) {
 732                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 733                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 734                         mask |= POLLHUP;
 735         }
 736
 737         if (filp->f_mode & FMODE_WRITE) {
 738                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 739                 /*
 740                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 741                  * behave exactly like pipes for poll().
 742                  */
 743                 if (!pipe->readers)
 744                         mask |= POLLERR;
 745         }
 746
 747         return mask;
 748 }
 749
 750 static int
 751 pipe_release(struct inode *inode, int decr, int decw)
 752 {
 753         struct pipe_inode_info *pipe;
 754
 755         mutex_lock(&inode->i_mutex);
 756         pipe = inode->i_pipe;
 757         pipe->readers -= decr;
 758         pipe->writers -= decw;
 759
 760         if (!pipe->readers && !pipe->writers) {
 761                 free_pipe_info(inode);
 762         } else {
 763                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 764                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 765                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 766         }
 767         mutex_unlock(&inode->i_mutex);
 768
 769         return 0;
 770 }
 771
 772 static int
 773 pipe_read_fasync(int fd, struct file *filp, int on)
 774 {
 775         struct inode *inode = filp->f_path.dentry->d_inode;
 776         int retval;
 777
 778         mutex_lock(&inode->i_mutex);
 779         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 780         mutex_unlock(&inode->i_mutex);
 781
 782         return retval;
 783 }
 784
 785
 786 static int
 787 pipe_write_fasync(int fd, struct file *filp, int on)
 788 {
 789         struct inode *inode = filp->f_path.dentry->d_inode;
 790         int retval;
 791
 792         mutex_lock(&inode->i_mutex);
 793         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 794         mutex_unlock(&inode->i_mutex);
 795
 796         return retval;
 797 }
 798
 799
 800 static int
 801 pipe_rdwr_fasync(int fd, struct file *filp, int on)
 802 {
 803         struct inode *inode = filp->f_path.dentry->d_inode;
 804         struct pipe_inode_info *pipe = inode->i_pipe;
 805         int retval;
 806
 807         mutex_lock(&inode->i_mutex);
 808         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 809         if (retval >= 0) {
 810                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 811                 if (retval < 0) /* this can happen only if on == T */
 812                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 813         }
 814         mutex_unlock(&inode->i_mutex);
 815         return retval;
 816 }
 817
 818
 819 static int
 820 pipe_read_release(struct inode *inode, struct file *filp)
 821 {
 822         return pipe_release(inode, 1, 0);
 823 }
 824
 825 static int
 826 pipe_write_release(struct inode *inode, struct file *filp)
 827 {
 828         return pipe_release(inode, 0, 1);
 829 }
 830
 831 static int
 832 pipe_rdwr_release(struct inode *inode, struct file *filp)
 833 {
 834         int decr, decw;
 835
 836         decr = (filp->f_mode & FMODE_READ) != 0;
 837         decw = (filp->f_mode & FMODE_WRITE) != 0;
 838         return pipe_release(inode, decr, decw);
 839 }
 840
 841 static int
 842 pipe_read_open(struct inode *inode, struct file *filp)
 843 {
 844         int ret = -ENOENT;
 845
 846         mutex_lock(&inode->i_mutex);
 847
 848         if (inode->i_pipe) {
 849                 ret = 0;
 850                 inode->i_pipe->readers++;
 851         }
 852
 853         mutex_unlock(&inode->i_mutex);
 854
 855         return ret;
 856 }
 857
 858 static int
 859 pipe_write_open(struct inode *inode, struct file *filp)
 860 {
 861         int ret = -ENOENT;
 862
 863         mutex_lock(&inode->i_mutex);
 864
 865         if (inode->i_pipe) {
 866                 ret = 0;
 867                 inode->i_pipe->writers++;
 868         }
 869
 870         mutex_unlock(&inode->i_mutex);
 871
 872         return ret;
 873 }
 874
 875 static int
 876 pipe_rdwr_open(struct inode *inode, struct file *filp)
 877 {
 878         int ret = -ENOENT;
 879
 880         if (!(filp->f_mode & (FMODE_READ|FMODE_WRITE)))
 881                 return -EINVAL;
 882
 883         mutex_lock(&inode->i_mutex);
 884
 885         if (inode->i_pipe) {
 886                 ret = 0;
 887                 if (filp->f_mode & FMODE_READ)
 888                         inode->i_pipe->readers++;
 889                 if (filp->f_mode & FMODE_WRITE)
 890                         inode->i_pipe->writers++;
 891         }
 892
 893         mutex_unlock(&inode->i_mutex);
 894
 895         return ret;
 896 }
 897
 898 /*
 899  * The file_operations structs are not static because they
 900  * are also used in linux/fs/fifo.c to do operations on FIFOs.
 901  *
 902  * Pipes reuse fifos' file_operations structs.
 903  */
 904 const struct file_operations read_pipefifo_fops = {
 905         .llseek         = no_llseek,
 906         .read           = do_sync_read,
 907         .aio_read       = pipe_read,
 908         .write          = bad_pipe_w,
 909         .poll           = pipe_poll,
 910         .unlocked_ioctl = pipe_ioctl,
 911         .open           = pipe_read_open,
 912         .release        = pipe_read_release,
 913         .fasync         = pipe_read_fasync,
 914 };
 915
 916 const struct file_operations write_pipefifo_fops = {
 917         .llseek         = no_llseek,
 918         .read           = bad_pipe_r,
 919         .write          = do_sync_write,
 920         .aio_write      = pipe_write,
 921         .poll           = pipe_poll,
 922         .unlocked_ioctl = pipe_ioctl,
 923         .open           = pipe_write_open,
 924         .release        = pipe_write_release,
 925         .fasync         = pipe_write_fasync,
 926 };
 927
 928 const struct file_operations rdwr_pipefifo_fops = {
 929         .llseek         = no_llseek,
 930         .read           = do_sync_read,
 931         .aio_read       = pipe_read,
 932         .write          = do_sync_write,
 933         .aio_write      = pipe_write,
 934         .poll           = pipe_poll,
 935         .unlocked_ioctl = pipe_ioctl,
 936         .open           = pipe_rdwr_open,
 937         .release        = pipe_rdwr_release,
 938         .fasync         = pipe_rdwr_fasync,
 939 };
 940
 941 static void account_pipe_buffers(struct pipe_inode_info *pipe,
 942                                  unsigned long old, unsigned long new)
 943 {
 944         atomic_long_add(new - old, &pipe->user->pipe_bufs);
 945 }
 946
 947 static bool too_many_pipe_buffers_soft(struct user_struct *user)
 948 {
 949         return pipe_user_pages_soft &&
 950                atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_soft;
 951 }
 952
 953 static bool too_many_pipe_buffers_hard(struct user_struct *user)
 954 {
 955         return pipe_user_pages_hard &&
 956                atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_hard;
 957 }
 958
 959 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 960 {
 961         struct pipe_inode_info *pipe;
 962
 963         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 964         if (pipe) {
 965                 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 966                 struct user_struct *user = get_current_user();
 967
 968                 if (!too_many_pipe_buffers_hard(user)) {
 969                         if (too_many_pipe_buffers_soft(user))
 970                                 pipe_bufs = 1;
 971                         pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * pipe_bufs, GFP_KERNEL);
 972                 }
 973
 974                 if (pipe->bufs) {
 975                         init_waitqueue_head(&pipe->wait);
 976                         pipe->r_counter = pipe->w_counter = 1;
 977                         pipe->inode = inode;
 978                         pipe->buffers = pipe_bufs;
 979                         pipe->user = user;
 980                         account_pipe_buffers(pipe, 0, pipe_bufs);
 981                         return pipe;
 982                 }
 983                 free_uid(user);
 984                 kfree(pipe);
 985         }
 986
 987         return NULL;
 988 }
 989
 990 void __free_pipe_info(struct pipe_inode_info *pipe)
 991 {
 992         int i;
 993
 994         account_pipe_buffers(pipe, pipe->buffers, 0);
 995         free_uid(pipe->user);
 996         for (i = 0; i < pipe->buffers; i++) {
 997                 struct pipe_buffer *buf = pipe->bufs + i;
 998                 if (buf->ops)
 999                         buf->ops->release(pipe, buf);
1000         }
1001         if (pipe->tmp_page)
1002                 __free_page(pipe->tmp_page);
1003         kfree(pipe->bufs);
1004         kfree(pipe);
1005 }
1006
1007 void free_pipe_info(struct inode *inode)
1008 {
1009         __free_pipe_info(inode->i_pipe);
1010         inode->i_pipe = NULL;
1011 }
1012
1013 static struct vfsmount *pipe_mnt __read_mostly;
1014
1015 /*
1016  * pipefs_dname() is called from d_path().
1017  */
1018 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
1019 {
1020         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
1021                                 dentry->d_inode->i_ino);
1022 }
1023
1024 static const struct dentry_operations pipefs_dentry_operations = {
1025         .d_dname        = pipefs_dname,
1026 };
1027
1028 static struct inode * get_pipe_inode(void)
1029 {
1030         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
1031         struct pipe_inode_info *pipe;
1032
1033         if (!inode)
1034                 goto fail_inode;
1035
1036         inode->i_ino = get_next_ino();
1037
1038         pipe = alloc_pipe_info(inode);
1039         if (!pipe)
1040                 goto fail_iput;
1041         inode->i_pipe = pipe;
1042
1043         pipe->readers = pipe->writers = 1;
1044         inode->i_fop = &rdwr_pipefifo_fops;
1045
1046         /*
1047          * Mark the inode dirty from the very beginning,
1048          * that way it will never be moved to the dirty
1049          * list because "mark_inode_dirty()" will think
1050          * that it already _is_ on the dirty list.
1051          */
1052         inode->i_state = I_DIRTY;
1053         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1054         inode->i_uid = current_fsuid();
1055         inode->i_gid = current_fsgid();
1056         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1057
1058         return inode;
1059
1060 fail_iput:
1061         iput(inode);
1062
1063 fail_inode:
1064         return NULL;
1065 }
1066
1067 struct file *create_write_pipe(int flags)
1068 {
1069         int err;
1070         struct inode *inode;
1071         struct file *f;
1072         struct path path;
1073         struct qstr name = { .name = "" };
1074
1075         err = -ENFILE;
1076         inode = get_pipe_inode();
1077         if (!inode)
1078                 goto err;
1079
1080         err = -ENOMEM;
1081         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1082         if (!path.dentry)
1083                 goto err_inode;
1084         path.mnt = mntget(pipe_mnt);
1085
1086         d_instantiate(path.dentry, inode);
1087
1088         err = -ENFILE;
1089         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1090         if (!f)
1091                 goto err_dentry;
1092         f->f_mapping = inode->i_mapping;
1093
1094         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
1095         f->f_version = 0;
1096
1097         return f;
1098
1099  err_dentry:
1100         free_pipe_info(inode);
1101         path_put(&path);
1102         return ERR_PTR(err);
1103
1104  err_inode:
1105         free_pipe_info(inode);
1106         iput(inode);
1107  err:
1108         return ERR_PTR(err);
1109 }
1110
1111 void free_write_pipe(struct file *f)
1112 {
1113         free_pipe_info(f->f_dentry->d_inode);
1114         path_put(&f->f_path);
1115         put_filp(f);
1116 }
1117
1118 struct file *create_read_pipe(struct file *wrf, int flags)
1119 {
1120         /* Grab pipe from the writer */
1121         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1122                                     &read_pipefifo_fops);
1123         if (!f)
1124                 return ERR_PTR(-ENFILE);
1125
1126         path_get(&wrf->f_path);
1127         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1128
1129         return f;
1130 }
1131
1132 int do_pipe_flags(int *fd, int flags)
1133 {
1134         struct file *fw, *fr;
1135         int error;
1136         int fdw, fdr;
1137
1138         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
1139                 return -EINVAL;
1140
1141         fw = create_write_pipe(flags);
1142         if (IS_ERR(fw))
1143                 return PTR_ERR(fw);
1144         fr = create_read_pipe(fw, flags);
1145         error = PTR_ERR(fr);
1146         if (IS_ERR(fr))
1147                 goto err_write_pipe;
1148
1149         error = get_unused_fd_flags(flags);
1150         if (error < 0)
1151                 goto err_read_pipe;
1152         fdr = error;
1153
1154         error = get_unused_fd_flags(flags);
1155         if (error < 0)
1156                 goto err_fdr;
1157         fdw = error;
1158
1159         audit_fd_pair(fdr, fdw);
1160         fd_install(fdr, fr);
1161         fd_install(fdw, fw);
1162         fd[0] = fdr;
1163         fd[1] = fdw;
1164
1165         return 0;
1166
1167  err_fdr:
1168         put_unused_fd(fdr);
1169  err_read_pipe:
1170         path_put(&fr->f_path);
1171         put_filp(fr);
1172  err_write_pipe:
1173         free_write_pipe(fw);
1174         return error;
1175 }
1176
1177 /*
1178  * sys_pipe() is the normal C calling standard for creating
1179  * a pipe. It's not the way Unix traditionally does this, though.
1180  */
1181 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1182 {
1183         int fd[2];
1184         int error;
1185
1186         error = do_pipe_flags(fd, flags);
1187         if (!error) {
1188                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1189                         sys_close(fd[0]);
1190                         sys_close(fd[1]);
1191                         error = -EFAULT;
1192                 }
1193         }
1194         return error;
1195 }
1196
1197 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1198 {
1199         return sys_pipe2(fildes, 0);
1200 }
1201
1202 /*
1203  * Allocate a new array of pipe buffers and copy the info over. Returns the
1204  * pipe size if successful, or return -ERROR on error.
1205  */
1206 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1207 {
1208         struct pipe_buffer *bufs;
1209
1210         /*
1211          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1212          * expect a lot of shrink+grow operations, just free and allocate
1213          * again like we would do for growing. If the pipe currently
1214          * contains more buffers than arg, then return busy.
1215          */
1216         if (nr_pages < pipe->nrbufs)
1217                 return -EBUSY;
1218
1219         bufs = kcalloc(nr_pages, sizeof(struct pipe_buffer), GFP_KERNEL);
1220         if (unlikely(!bufs))
1221                 return -ENOMEM;
1222
1223         /*
1224          * The pipe array wraps around, so just start the new one at zero
1225          * and adjust the indexes.
1226          */
1227         if (pipe->nrbufs) {
1228                 unsigned int tail;
1229                 unsigned int head;
1230
1231                 tail = pipe->curbuf + pipe->nrbufs;
1232                 if (tail < pipe->buffers)
1233                         tail = 0;
1234                 else
1235                         tail &= (pipe->buffers - 1);
1236
1237                 head = pipe->nrbufs - tail;
1238                 if (head)
1239                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1240                 if (tail)
1241                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1242         }
1243
1244         account_pipe_buffers(pipe, pipe->buffers, nr_pages);
1245         pipe->curbuf = 0;
1246         kfree(pipe->bufs);
1247         pipe->bufs = bufs;
1248         pipe->buffers = nr_pages;
1249         return nr_pages * PAGE_SIZE;
1250 }
1251
1252 /*
1253  * Currently we rely on the pipe array holding a power-of-2 number
1254  * of pages.
1255  */
1256 static inline unsigned int round_pipe_size(unsigned int size)
1257 {
1258         unsigned long nr_pages;
1259
1260         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1261         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1262 }
1263
1264 /*
1265  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1266  * will return an error.
1267  */
1268 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1269                  size_t *lenp, loff_t *ppos)
1270 {
1271         int ret;
1272
1273         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1274         if (ret < 0 || !write)
1275                 return ret;
1276
1277         pipe_max_size = round_pipe_size(pipe_max_size);
1278         return ret;
1279 }
1280
1281 /*
1282  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1283  * location, so checking ->i_pipe is not enough to verify that this is a
1284  * pipe.
1285  */
1286 struct pipe_inode_info *get_pipe_info(struct file *file)
1287 {
1288         struct inode *i = file->f_path.dentry->d_inode;
1289
1290         return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1291 }
1292
1293 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1294 {
1295         struct pipe_inode_info *pipe;
1296         long ret;
1297
1298         pipe = get_pipe_info(file);
1299         if (!pipe)
1300                 return -EBADF;
1301
1302         mutex_lock(&pipe->inode->i_mutex);
1303
1304         switch (cmd) {
1305         case F_SETPIPE_SZ: {
1306                 unsigned int size, nr_pages;
1307
1308                 size = round_pipe_size(arg);
1309                 nr_pages = size >> PAGE_SHIFT;
1310
1311                 ret = -EINVAL;
1312                 if (!nr_pages)
1313                         goto out;
1314
1315                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1316                         ret = -EPERM;
1317                         goto out;
1318                 } else if ((too_many_pipe_buffers_hard(pipe->user) ||
1319                             too_many_pipe_buffers_soft(pipe->user)) &&
1320                            !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1321                         ret = -EPERM;
1322                         goto out;
1323                 }
1324                 ret = pipe_set_size(pipe, nr_pages);
1325                 break;
1326                 }
1327         case F_GETPIPE_SZ:
1328                 ret = pipe->buffers * PAGE_SIZE;
1329                 break;
1330         default:
1331                 ret = -EINVAL;
1332                 break;
1333         }
1334
1335 out:
1336         mutex_unlock(&pipe->inode->i_mutex);
1337         return ret;
1338 }
1339
1340 static const struct super_operations pipefs_ops = {
1341         .destroy_inode = free_inode_nonrcu,
1342         .statfs = simple_statfs,
1343 };
1344
1345 /*
1346  * pipefs should _never_ be mounted by userland - too much of security hassle,
1347  * no real gain from having the whole whorehouse mounted. So we don't need
1348  * any operations on the root directory. However, we need a non-trivial
1349  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1350  */
1351 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1352                          int flags, const char *dev_name, void *data)
1353 {
1354         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1355                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1356 }
1357
1358 static struct file_system_type pipe_fs_type = {
1359         .name           = "pipefs",
1360         .mount          = pipefs_mount,
1361         .kill_sb        = kill_anon_super,
1362 };
1363
1364 static int __init init_pipe_fs(void)
1365 {
1366         int err = register_filesystem(&pipe_fs_type);
1367
1368         if (!err) {
1369                 pipe_mnt = kern_mount(&pipe_fs_type);
1370                 if (IS_ERR(pipe_mnt)) {
1371                         err = PTR_ERR(pipe_mnt);
1372                         unregister_filesystem(&pipe_fs_type);
1373                 }
1374         }
1375         return err;
1376 }
1377
1378 static void __exit exit_pipe_fs(void)
1379 {
1380         kern_unmount(pipe_mnt);
1381         unregister_filesystem(&pipe_fs_type);
1382 }
1383
1384 fs_initcall(init_pipe_fs);
1385 module_exit(exit_pipe_fs);