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