fs/pipe.c

   1 /*
   2  *  linux/fs/pipe.c
   3  *
   4  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
   5  */
   6
   7 #include <linux/mm.h>
   8 #include <linux/file.h>
   9 #include <linux/poll.h>
  10 #include <linux/slab.h>
  11 #include <linux/module.h>
  12 #include <linux/init.h>
  13 #include <linux/fs.h>
  14 #include <linux/log2.h>
  15 #include <linux/mount.h>
  16 #include <linux/magic.h>
  17 #include <linux/pipe_fs_i.h>
  18 #include <linux/uio.h>
  19 #include <linux/highmem.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/audit.h>
  22 #include <linux/syscalls.h>
  23 #include <linux/fcntl.h>
  24 #include <linux/memcontrol.h>
  25
  26 #include <asm/uaccess.h>
  27 #include <asm/ioctls.h>
  28
  29 #include "internal.h"
  30
  31 /*
  32  * New pipe buffers will be restricted to this size while the user is exceeding
  33  * their pipe buffer quota. The general pipe use case needs at least two
  34  * buffers: one for data yet to be read, and one for new data. If this is less
  35  * than two, then a write to a non-empty pipe may block even if the pipe is not
  36  * full. This can occur with GNU make jobserver or similar uses of pipes as
  37  * semaphores: multiple processes may be waiting to write tokens back to the
  38  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  39  *
  40  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  41  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  42  * emptied.
  43  */
  44 #define PIPE_MIN_DEF_BUFFERS 2
  45
  46 /*
  47  * The max size that a non-root user is allowed to grow the pipe. Can
  48  * be set by root in /proc/sys/fs/pipe-max-size
  49  */
  50 unsigned int pipe_max_size = 1048576;
  51
  52 /*
  53  * Minimum pipe size, as required by POSIX
  54  */
  55 unsigned int pipe_min_size = PAGE_SIZE;
  56
  57 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  58  * matches default values.
  59  */
  60 unsigned long pipe_user_pages_hard;
  61 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  62
  63 /*
  64  * We use a start+len construction, which provides full use of the
  65  * allocated memory.
  66  * -- Florian Coosmann (FGC)
  67  *
  68  * Reads with count = 0 should always return 0.
  69  * -- Julian Bradfield 1999-06-07.
  70  *
  71  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  72  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  73  *
  74  * pipe_read & write cleanup
  75  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  76  */
  77
  78 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  79 {
  80         if (pipe->files)
  81                 mutex_lock_nested(&pipe->mutex, subclass);
  82 }
  83
  84 void pipe_lock(struct pipe_inode_info *pipe)
  85 {
  86         /*
  87          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  88          */
  89         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  90 }
  91 EXPORT_SYMBOL(pipe_lock);
  92
  93 void pipe_unlock(struct pipe_inode_info *pipe)
  94 {
  95         if (pipe->files)
  96                 mutex_unlock(&pipe->mutex);
  97 }
  98 EXPORT_SYMBOL(pipe_unlock);
  99
 100 static inline void __pipe_lock(struct pipe_inode_info *pipe)
 101 {
 102         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 103 }
 104
 105 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 106 {
 107         mutex_unlock(&pipe->mutex);
 108 }
 109
 110 void pipe_double_lock(struct pipe_inode_info *pipe1,
 111                       struct pipe_inode_info *pipe2)
 112 {
 113         BUG_ON(pipe1 == pipe2);
 114
 115         if (pipe1 < pipe2) {
 116                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 117                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 118         } else {
 119                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 120                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 121         }
 122 }
 123
 124 /* Drop the inode semaphore and wait for a pipe event, atomically */
 125 void pipe_wait(struct pipe_inode_info *pipe)
 126 {
 127         DEFINE_WAIT(wait);
 128
 129         /*
 130          * Pipes are system-local resources, so sleeping on them
 131          * is considered a noninteractive wait:
 132          */
 133         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 134         pipe_unlock(pipe);
 135         schedule();
 136         finish_wait(&pipe->wait, &wait);
 137         pipe_lock(pipe);
 138 }
 139
 140 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 141                                   struct pipe_buffer *buf)
 142 {
 143         struct page *page = buf->page;
 144
 145         /*
 146          * If nobody else uses this page, and we don't already have a
 147          * temporary page, let's keep track of it as a one-deep
 148          * allocation cache. (Otherwise just release our reference to it)
 149          */
 150         if (page_count(page) == 1 && !pipe->tmp_page)
 151                 pipe->tmp_page = page;
 152         else
 153                 put_page(page);
 154 }
 155
 156 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 157                                struct pipe_buffer *buf)
 158 {
 159         struct page *page = buf->page;
 160
 161         if (page_count(page) == 1) {
 162                 if (memcg_kmem_enabled())
 163                         memcg_kmem_uncharge(page, 0);
 164                 __SetPageLocked(page);
 165                 return 0;
 166         }
 167         return 1;
 168 }
 169
 170 /**
 171  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 172  * @pipe:       the pipe that the buffer belongs to
 173  * @buf:        the buffer to attempt to steal
 174  *
 175  * Description:
 176  *      This function attempts to steal the &struct page attached to
 177  *      @buf. If successful, this function returns 0 and returns with
 178  *      the page locked. The caller may then reuse the page for whatever
 179  *      he wishes; the typical use is insertion into a different file
 180  *      page cache.
 181  */
 182 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 183                            struct pipe_buffer *buf)
 184 {
 185         struct page *page = buf->page;
 186
 187         /*
 188          * A reference of one is golden, that means that the owner of this
 189          * page is the only one holding a reference to it. lock the page
 190          * and return OK.
 191          */
 192         if (page_count(page) == 1) {
 193                 lock_page(page);
 194                 return 0;
 195         }
 196
 197         return 1;
 198 }
 199 EXPORT_SYMBOL(generic_pipe_buf_steal);
 200
 201 /**
 202  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 203  * @pipe:       the pipe that the buffer belongs to
 204  * @buf:        the buffer to get a reference to
 205  *
 206  * Description:
 207  *      This function grabs an extra reference to @buf. It's used in
 208  *      in the tee() system call, when we duplicate the buffers in one
 209  *      pipe into another.
 210  */
 211 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 212 {
 213         return try_get_page(buf->page);
 214 }
 215 EXPORT_SYMBOL(generic_pipe_buf_get);
 216
 217 /**
 218  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 219  * @info:       the pipe that the buffer belongs to
 220  * @buf:        the buffer to confirm
 221  *
 222  * Description:
 223  *      This function does nothing, because the generic pipe code uses
 224  *      pages that are always good when inserted into the pipe.
 225  */
 226 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 227                              struct pipe_buffer *buf)
 228 {
 229         return 0;
 230 }
 231 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 232
 233 /**
 234  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 235  * @pipe:       the pipe that the buffer belongs to
 236  * @buf:        the buffer to put a reference to
 237  *
 238  * Description:
 239  *      This function releases a reference to @buf.
 240  */
 241 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 242                               struct pipe_buffer *buf)
 243 {
 244         put_page(buf->page);
 245 }
 246 EXPORT_SYMBOL(generic_pipe_buf_release);
 247
 248 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 249         .can_merge = 1,
 250         .confirm = generic_pipe_buf_confirm,
 251         .release = anon_pipe_buf_release,
 252         .steal = anon_pipe_buf_steal,
 253         .get = generic_pipe_buf_get,
 254 };
 255
 256 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 257         .can_merge = 0,
 258         .confirm = generic_pipe_buf_confirm,
 259         .release = anon_pipe_buf_release,
 260         .steal = anon_pipe_buf_steal,
 261         .get = generic_pipe_buf_get,
 262 };
 263
 264 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 265         .can_merge = 0,
 266         .confirm = generic_pipe_buf_confirm,
 267         .release = anon_pipe_buf_release,
 268         .steal = anon_pipe_buf_steal,
 269         .get = generic_pipe_buf_get,
 270 };
 271
 272 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 273 {
 274         if (buf->ops == &anon_pipe_buf_ops)
 275                 buf->ops = &anon_pipe_buf_nomerge_ops;
 276 }
 277
 278 static ssize_t
 279 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 280 {
 281         size_t total_len = iov_iter_count(to);
 282         struct file *filp = iocb->ki_filp;
 283         struct pipe_inode_info *pipe = filp->private_data;
 284         int do_wakeup;
 285         ssize_t ret;
 286
 287         /* Null read succeeds. */
 288         if (unlikely(total_len == 0))
 289                 return 0;
 290
 291         do_wakeup = 0;
 292         ret = 0;
 293         __pipe_lock(pipe);
 294         for (;;) {
 295                 int bufs = pipe->nrbufs;
 296                 if (bufs) {
 297                         int curbuf = pipe->curbuf;
 298                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 299                         size_t chars = buf->len;
 300                         size_t written;
 301                         int error;
 302
 303                         if (chars > total_len)
 304                                 chars = total_len;
 305
 306                         error = pipe_buf_confirm(pipe, buf);
 307                         if (error) {
 308                                 if (!ret)
 309                                         ret = error;
 310                                 break;
 311                         }
 312
 313                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 314                         if (unlikely(written < chars)) {
 315                                 if (!ret)
 316                                         ret = -EFAULT;
 317                                 break;
 318                         }
 319                         ret += chars;
 320                         buf->offset += chars;
 321                         buf->len -= chars;
 322
 323                         /* Was it a packet buffer? Clean up and exit */
 324                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 325                                 total_len = chars;
 326                                 buf->len = 0;
 327                         }
 328
 329                         if (!buf->len) {
 330                                 pipe_buf_release(pipe, buf);
 331                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 332                                 pipe->curbuf = curbuf;
 333                                 pipe->nrbufs = --bufs;
 334                                 do_wakeup = 1;
 335                         }
 336                         total_len -= chars;
 337                         if (!total_len)
 338                                 break;  /* common path: read succeeded */
 339                 }
 340                 if (bufs)       /* More to do? */
 341                         continue;
 342                 if (!pipe->writers)
 343                         break;
 344                 if (!pipe->waiting_writers) {
 345                         /* syscall merging: Usually we must not sleep
 346                          * if O_NONBLOCK is set, or if we got some data.
 347                          * But if a writer sleeps in kernel space, then
 348                          * we can wait for that data without violating POSIX.
 349                          */
 350                         if (ret)
 351                                 break;
 352                         if (filp->f_flags & O_NONBLOCK) {
 353                                 ret = -EAGAIN;
 354                                 break;
 355                         }
 356                 }
 357                 if (signal_pending(current)) {
 358                         if (!ret)
 359                                 ret = -ERESTARTSYS;
 360                         break;
 361                 }
 362                 if (do_wakeup) {
 363                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 364                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 365                 }
 366                 pipe_wait(pipe);
 367         }
 368         __pipe_unlock(pipe);
 369
 370         /* Signal writers asynchronously that there is more room. */
 371         if (do_wakeup) {
 372                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 373                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 374         }
 375         if (ret > 0)
 376                 file_accessed(filp);
 377         return ret;
 378 }
 379
 380 static inline int is_packetized(struct file *file)
 381 {
 382         return (file->f_flags & O_DIRECT) != 0;
 383 }
 384
 385 static ssize_t
 386 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 387 {
 388         struct file *filp = iocb->ki_filp;
 389         struct pipe_inode_info *pipe = filp->private_data;
 390         ssize_t ret = 0;
 391         int do_wakeup = 0;
 392         size_t total_len = iov_iter_count(from);
 393         ssize_t chars;
 394
 395         /* Null write succeeds. */
 396         if (unlikely(total_len == 0))
 397                 return 0;
 398
 399         __pipe_lock(pipe);
 400
 401         if (!pipe->readers) {
 402                 send_sig(SIGPIPE, current, 0);
 403                 ret = -EPIPE;
 404                 goto out;
 405         }
 406
 407         /* We try to merge small writes */
 408         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 409         if (pipe->nrbufs && chars != 0) {
 410                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 411                                                         (pipe->buffers - 1);
 412                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 413                 int offset = buf->offset + buf->len;
 414
 415                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 416                         ret = pipe_buf_confirm(pipe, buf);
 417                         if (ret)
 418                                 goto out;
 419
 420                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 421                         if (unlikely(ret < chars)) {
 422                                 ret = -EFAULT;
 423                                 goto out;
 424                         }
 425                         do_wakeup = 1;
 426                         buf->len += ret;
 427                         if (!iov_iter_count(from))
 428                                 goto out;
 429                 }
 430         }
 431
 432         for (;;) {
 433                 int bufs;
 434
 435                 if (!pipe->readers) {
 436                         send_sig(SIGPIPE, current, 0);
 437                         if (!ret)
 438                                 ret = -EPIPE;
 439                         break;
 440                 }
 441                 bufs = pipe->nrbufs;
 442                 if (bufs < pipe->buffers) {
 443                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 444                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 445                         struct page *page = pipe->tmp_page;
 446                         int copied;
 447
 448                         if (!page) {
 449                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 450                                 if (unlikely(!page)) {
 451                                         ret = ret ? : -ENOMEM;
 452                                         break;
 453                                 }
 454                                 pipe->tmp_page = page;
 455                         }
 456                         /* Always wake up, even if the copy fails. Otherwise
 457                          * we lock up (O_NONBLOCK-)readers that sleep due to
 458                          * syscall merging.
 459                          * FIXME! Is this really true?
 460                          */
 461                         do_wakeup = 1;
 462                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 463                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 464                                 if (!ret)
 465                                         ret = -EFAULT;
 466                                 break;
 467                         }
 468                         ret += copied;
 469
 470                         /* Insert it into the buffer array */
 471                         buf->page = page;
 472                         buf->ops = &anon_pipe_buf_ops;
 473                         buf->offset = 0;
 474                         buf->len = copied;
 475                         buf->flags = 0;
 476                         if (is_packetized(filp)) {
 477                                 buf->ops = &packet_pipe_buf_ops;
 478                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 479                         }
 480                         pipe->nrbufs = ++bufs;
 481                         pipe->tmp_page = NULL;
 482
 483                         if (!iov_iter_count(from))
 484                                 break;
 485                 }
 486                 if (bufs < pipe->buffers)
 487                         continue;
 488                 if (filp->f_flags & O_NONBLOCK) {
 489                         if (!ret)
 490                                 ret = -EAGAIN;
 491                         break;
 492                 }
 493                 if (signal_pending(current)) {
 494                         if (!ret)
 495                                 ret = -ERESTARTSYS;
 496                         break;
 497                 }
 498                 if (do_wakeup) {
 499                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 500                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 501                         do_wakeup = 0;
 502                 }
 503                 pipe->waiting_writers++;
 504                 pipe_wait(pipe);
 505                 pipe->waiting_writers--;
 506         }
 507 out:
 508         __pipe_unlock(pipe);
 509         if (do_wakeup) {
 510                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 511                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 512         }
 513         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 514                 int err = file_update_time(filp);
 515                 if (err)
 516                         ret = err;
 517                 sb_end_write(file_inode(filp)->i_sb);
 518         }
 519         return ret;
 520 }
 521
 522 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 523 {
 524         struct pipe_inode_info *pipe = filp->private_data;
 525         int count, buf, nrbufs;
 526
 527         switch (cmd) {
 528                 case FIONREAD:
 529                         __pipe_lock(pipe);
 530                         count = 0;
 531                         buf = pipe->curbuf;
 532                         nrbufs = pipe->nrbufs;
 533                         while (--nrbufs >= 0) {
 534                                 count += pipe->bufs[buf].len;
 535                                 buf = (buf+1) & (pipe->buffers - 1);
 536                         }
 537                         __pipe_unlock(pipe);
 538
 539                         return put_user(count, (int __user *)arg);
 540                 default:
 541                         return -ENOIOCTLCMD;
 542         }
 543 }
 544
 545 /* No kernel lock held - fine */
 546 static unsigned int
 547 pipe_poll(struct file *filp, poll_table *wait)
 548 {
 549         unsigned int mask;
 550         struct pipe_inode_info *pipe = filp->private_data;
 551         int nrbufs;
 552
 553         poll_wait(filp, &pipe->wait, wait);
 554
 555         /* Reading only -- no need for acquiring the semaphore.  */
 556         nrbufs = pipe->nrbufs;
 557         mask = 0;
 558         if (filp->f_mode & FMODE_READ) {
 559                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 560                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 561                         mask |= POLLHUP;
 562         }
 563
 564         if (filp->f_mode & FMODE_WRITE) {
 565                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 566                 /*
 567                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 568                  * behave exactly like pipes for poll().
 569                  */
 570                 if (!pipe->readers)
 571                         mask |= POLLERR;
 572         }
 573
 574         return mask;
 575 }
 576
 577 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 578 {
 579         int kill = 0;
 580
 581         spin_lock(&inode->i_lock);
 582         if (!--pipe->files) {
 583                 inode->i_pipe = NULL;
 584                 kill = 1;
 585         }
 586         spin_unlock(&inode->i_lock);
 587
 588         if (kill)
 589                 free_pipe_info(pipe);
 590 }
 591
 592 static int
 593 pipe_release(struct inode *inode, struct file *file)
 594 {
 595         struct pipe_inode_info *pipe = file->private_data;
 596
 597         __pipe_lock(pipe);
 598         if (file->f_mode & FMODE_READ)
 599                 pipe->readers--;
 600         if (file->f_mode & FMODE_WRITE)
 601                 pipe->writers--;
 602
 603         if (pipe->readers || pipe->writers) {
 604                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 605                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 606                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 607         }
 608         __pipe_unlock(pipe);
 609
 610         put_pipe_info(inode, pipe);
 611         return 0;
 612 }
 613
 614 static int
 615 pipe_fasync(int fd, struct file *filp, int on)
 616 {
 617         struct pipe_inode_info *pipe = filp->private_data;
 618         int retval = 0;
 619
 620         __pipe_lock(pipe);
 621         if (filp->f_mode & FMODE_READ)
 622                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 623         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 624                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 625                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 626                         /* this can happen only if on == T */
 627                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 628         }
 629         __pipe_unlock(pipe);
 630         return retval;
 631 }
 632
 633 static unsigned long account_pipe_buffers(struct user_struct *user,
 634                                  unsigned long old, unsigned long new)
 635 {
 636         return atomic_long_add_return(new - old, &user->pipe_bufs);
 637 }
 638
 639 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 640 {
 641         return pipe_user_pages_soft && user_bufs > pipe_user_pages_soft;
 642 }
 643
 644 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 645 {
 646         return pipe_user_pages_hard && user_bufs > pipe_user_pages_hard;
 647 }
 648
 649 static bool is_unprivileged_user(void)
 650 {
 651         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 652 }
 653
 654 struct pipe_inode_info *alloc_pipe_info(void)
 655 {
 656         struct pipe_inode_info *pipe;
 657         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 658         struct user_struct *user = get_current_user();
 659         unsigned long user_bufs;
 660
 661         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 662         if (pipe == NULL)
 663                 goto out_free_uid;
 664
 665         if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 666                 pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 667
 668         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 669
 670         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 671                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 672                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 673         }
 674
 675         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 676                 goto out_revert_acct;
 677
 678         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 679                              GFP_KERNEL_ACCOUNT);
 680
 681         if (pipe->bufs) {
 682                 init_waitqueue_head(&pipe->wait);
 683                 pipe->r_counter = pipe->w_counter = 1;
 684                 pipe->buffers = pipe_bufs;
 685                 pipe->user = user;
 686                 mutex_init(&pipe->mutex);
 687                 return pipe;
 688         }
 689
 690 out_revert_acct:
 691         (void) account_pipe_buffers(user, pipe_bufs, 0);
 692         kfree(pipe);
 693 out_free_uid:
 694         free_uid(user);
 695         return NULL;
 696 }
 697
 698 void free_pipe_info(struct pipe_inode_info *pipe)
 699 {
 700         int i;
 701
 702         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 703         free_uid(pipe->user);
 704         for (i = 0; i < pipe->buffers; i++) {
 705                 struct pipe_buffer *buf = pipe->bufs + i;
 706                 if (buf->ops)
 707                         pipe_buf_release(pipe, buf);
 708         }
 709         if (pipe->tmp_page)
 710                 __free_page(pipe->tmp_page);
 711         kfree(pipe->bufs);
 712         kfree(pipe);
 713 }
 714
 715 static struct vfsmount *pipe_mnt __read_mostly;
 716
 717 /*
 718  * pipefs_dname() is called from d_path().
 719  */
 720 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 721 {
 722         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 723                                 d_inode(dentry)->i_ino);
 724 }
 725
 726 static const struct dentry_operations pipefs_dentry_operations = {
 727         .d_dname        = pipefs_dname,
 728 };
 729
 730 static struct inode * get_pipe_inode(void)
 731 {
 732         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 733         struct pipe_inode_info *pipe;
 734
 735         if (!inode)
 736                 goto fail_inode;
 737
 738         inode->i_ino = get_next_ino();
 739
 740         pipe = alloc_pipe_info();
 741         if (!pipe)
 742                 goto fail_iput;
 743
 744         inode->i_pipe = pipe;
 745         pipe->files = 2;
 746         pipe->readers = pipe->writers = 1;
 747         inode->i_fop = &pipefifo_fops;
 748
 749         /*
 750          * Mark the inode dirty from the very beginning,
 751          * that way it will never be moved to the dirty
 752          * list because "mark_inode_dirty()" will think
 753          * that it already _is_ on the dirty list.
 754          */
 755         inode->i_state = I_DIRTY;
 756         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 757         inode->i_uid = current_fsuid();
 758         inode->i_gid = current_fsgid();
 759         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 760
 761         return inode;
 762
 763 fail_iput:
 764         iput(inode);
 765
 766 fail_inode:
 767         return NULL;
 768 }
 769
 770 int create_pipe_files(struct file **res, int flags)
 771 {
 772         int err;
 773         struct inode *inode = get_pipe_inode();
 774         struct file *f;
 775         struct path path;
 776         static struct qstr name = { .name = "" };
 777
 778         if (!inode)
 779                 return -ENFILE;
 780
 781         err = -ENOMEM;
 782         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
 783         if (!path.dentry)
 784                 goto err_inode;
 785         path.mnt = mntget(pipe_mnt);
 786
 787         d_instantiate(path.dentry, inode);
 788
 789         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 790         if (IS_ERR(f)) {
 791                 err = PTR_ERR(f);
 792                 goto err_dentry;
 793         }
 794
 795         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 796         f->private_data = inode->i_pipe;
 797
 798         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 799         if (IS_ERR(res[0])) {
 800                 err = PTR_ERR(res[0]);
 801                 goto err_file;
 802         }
 803
 804         path_get(&path);
 805         res[0]->private_data = inode->i_pipe;
 806         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 807         res[1] = f;
 808         return 0;
 809
 810 err_file:
 811         put_filp(f);
 812 err_dentry:
 813         free_pipe_info(inode->i_pipe);
 814         path_put(&path);
 815         return err;
 816
 817 err_inode:
 818         free_pipe_info(inode->i_pipe);
 819         iput(inode);
 820         return err;
 821 }
 822
 823 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 824 {
 825         int error;
 826         int fdw, fdr;
 827
 828         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 829                 return -EINVAL;
 830
 831         error = create_pipe_files(files, flags);
 832         if (error)
 833                 return error;
 834
 835         error = get_unused_fd_flags(flags);
 836         if (error < 0)
 837                 goto err_read_pipe;
 838         fdr = error;
 839
 840         error = get_unused_fd_flags(flags);
 841         if (error < 0)
 842                 goto err_fdr;
 843         fdw = error;
 844
 845         audit_fd_pair(fdr, fdw);
 846         fd[0] = fdr;
 847         fd[1] = fdw;
 848         return 0;
 849
 850  err_fdr:
 851         put_unused_fd(fdr);
 852  err_read_pipe:
 853         fput(files[0]);
 854         fput(files[1]);
 855         return error;
 856 }
 857
 858 int do_pipe_flags(int *fd, int flags)
 859 {
 860         struct file *files[2];
 861         int error = __do_pipe_flags(fd, files, flags);
 862         if (!error) {
 863                 fd_install(fd[0], files[0]);
 864                 fd_install(fd[1], files[1]);
 865         }
 866         return error;
 867 }
 868
 869 /*
 870  * sys_pipe() is the normal C calling standard for creating
 871  * a pipe. It's not the way Unix traditionally does this, though.
 872  */
 873 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 874 {
 875         struct file *files[2];
 876         int fd[2];
 877         int error;
 878
 879         error = __do_pipe_flags(fd, files, flags);
 880         if (!error) {
 881                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 882                         fput(files[0]);
 883                         fput(files[1]);
 884                         put_unused_fd(fd[0]);
 885                         put_unused_fd(fd[1]);
 886                         error = -EFAULT;
 887                 } else {
 888                         fd_install(fd[0], files[0]);
 889                         fd_install(fd[1], files[1]);
 890                 }
 891         }
 892         return error;
 893 }
 894
 895 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 896 {
 897         return sys_pipe2(fildes, 0);
 898 }
 899
 900 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 901 {
 902         int cur = *cnt;
 903
 904         while (cur == *cnt) {
 905                 pipe_wait(pipe);
 906                 if (signal_pending(current))
 907                         break;
 908         }
 909         return cur == *cnt ? -ERESTARTSYS : 0;
 910 }
 911
 912 static void wake_up_partner(struct pipe_inode_info *pipe)
 913 {
 914         wake_up_interruptible(&pipe->wait);
 915 }
 916
 917 static int fifo_open(struct inode *inode, struct file *filp)
 918 {
 919         struct pipe_inode_info *pipe;
 920         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 921         int ret;
 922
 923         filp->f_version = 0;
 924
 925         spin_lock(&inode->i_lock);
 926         if (inode->i_pipe) {
 927                 pipe = inode->i_pipe;
 928                 pipe->files++;
 929                 spin_unlock(&inode->i_lock);
 930         } else {
 931                 spin_unlock(&inode->i_lock);
 932                 pipe = alloc_pipe_info();
 933                 if (!pipe)
 934                         return -ENOMEM;
 935                 pipe->files = 1;
 936                 spin_lock(&inode->i_lock);
 937                 if (unlikely(inode->i_pipe)) {
 938                         inode->i_pipe->files++;
 939                         spin_unlock(&inode->i_lock);
 940                         free_pipe_info(pipe);
 941                         pipe = inode->i_pipe;
 942                 } else {
 943                         inode->i_pipe = pipe;
 944                         spin_unlock(&inode->i_lock);
 945                 }
 946         }
 947         filp->private_data = pipe;
 948         /* OK, we have a pipe and it's pinned down */
 949
 950         __pipe_lock(pipe);
 951
 952         /* We can only do regular read/write on fifos */
 953         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 954
 955         switch (filp->f_mode) {
 956         case FMODE_READ:
 957         /*
 958          *  O_RDONLY
 959          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 960          *  opened, even when there is no process writing the FIFO.
 961          */
 962                 pipe->r_counter++;
 963                 if (pipe->readers++ == 0)
 964                         wake_up_partner(pipe);
 965
 966                 if (!is_pipe && !pipe->writers) {
 967                         if ((filp->f_flags & O_NONBLOCK)) {
 968                                 /* suppress POLLHUP until we have
 969                                  * seen a writer */
 970                                 filp->f_version = pipe->w_counter;
 971                         } else {
 972                                 if (wait_for_partner(pipe, &pipe->w_counter))
 973                                         goto err_rd;
 974                         }
 975                 }
 976                 break;
 977
 978         case FMODE_WRITE:
 979         /*
 980          *  O_WRONLY
 981          *  POSIX.1 says that O_NONBLOCK means return -1 with
 982          *  errno=ENXIO when there is no process reading the FIFO.
 983          */
 984                 ret = -ENXIO;
 985                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 986                         goto err;
 987
 988                 pipe->w_counter++;
 989                 if (!pipe->writers++)
 990                         wake_up_partner(pipe);
 991
 992                 if (!is_pipe && !pipe->readers) {
 993                         if (wait_for_partner(pipe, &pipe->r_counter))
 994                                 goto err_wr;
 995                 }
 996                 break;
 997
 998         case FMODE_READ | FMODE_WRITE:
 999         /*
1000          *  O_RDWR
1001          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1002          *  This implementation will NEVER block on a O_RDWR open, since
1003          *  the process can at least talk to itself.
1004          */
1005
1006                 pipe->readers++;
1007                 pipe->writers++;
1008                 pipe->r_counter++;
1009                 pipe->w_counter++;
1010                 if (pipe->readers == 1 || pipe->writers == 1)
1011                         wake_up_partner(pipe);
1012                 break;
1013
1014         default:
1015                 ret = -EINVAL;
1016                 goto err;
1017         }
1018
1019         /* Ok! */
1020         __pipe_unlock(pipe);
1021         return 0;
1022
1023 err_rd:
1024         if (!--pipe->readers)
1025                 wake_up_interruptible(&pipe->wait);
1026         ret = -ERESTARTSYS;
1027         goto err;
1028
1029 err_wr:
1030         if (!--pipe->writers)
1031                 wake_up_interruptible(&pipe->wait);
1032         ret = -ERESTARTSYS;
1033         goto err;
1034
1035 err:
1036         __pipe_unlock(pipe);
1037
1038         put_pipe_info(inode, pipe);
1039         return ret;
1040 }
1041
1042 const struct file_operations pipefifo_fops = {
1043         .open           = fifo_open,
1044         .llseek         = no_llseek,
1045         .read_iter      = pipe_read,
1046         .write_iter     = pipe_write,
1047         .poll           = pipe_poll,
1048         .unlocked_ioctl = pipe_ioctl,
1049         .release        = pipe_release,
1050         .fasync         = pipe_fasync,
1051 };
1052
1053 /*
1054  * Currently we rely on the pipe array holding a power-of-2 number
1055  * of pages. Returns 0 on error.
1056  */
1057 static inline unsigned int round_pipe_size(unsigned int size)
1058 {
1059         unsigned long nr_pages;
1060
1061         if (size < pipe_min_size)
1062                 size = pipe_min_size;
1063
1064         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1065         if (nr_pages == 0)
1066                 return 0;
1067
1068         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1069 }
1070
1071 /*
1072  * Allocate a new array of pipe buffers and copy the info over. Returns the
1073  * pipe size if successful, or return -ERROR on error.
1074  */
1075 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1076 {
1077         struct pipe_buffer *bufs;
1078         unsigned int size, nr_pages;
1079         unsigned long user_bufs;
1080         long ret = 0;
1081
1082         size = round_pipe_size(arg);
1083         if (size == 0)
1084                 return -EINVAL;
1085         nr_pages = size >> PAGE_SHIFT;
1086
1087         if (!nr_pages)
1088                 return -EINVAL;
1089
1090         /*
1091          * If trying to increase the pipe capacity, check that an
1092          * unprivileged user is not trying to exceed various limits
1093          * (soft limit check here, hard limit check just below).
1094          * Decreasing the pipe capacity is always permitted, even
1095          * if the user is currently over a limit.
1096          */
1097         if (nr_pages > pipe->buffers &&
1098                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1099                 return -EPERM;
1100
1101         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1102
1103         if (nr_pages > pipe->buffers &&
1104                         (too_many_pipe_buffers_hard(user_bufs) ||
1105                          too_many_pipe_buffers_soft(user_bufs)) &&
1106                         is_unprivileged_user()) {
1107                 ret = -EPERM;
1108                 goto out_revert_acct;
1109         }
1110
1111         /*
1112          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1113          * expect a lot of shrink+grow operations, just free and allocate
1114          * again like we would do for growing. If the pipe currently
1115          * contains more buffers than arg, then return busy.
1116          */
1117         if (nr_pages < pipe->nrbufs) {
1118                 ret = -EBUSY;
1119                 goto out_revert_acct;
1120         }
1121
1122         bufs = kcalloc(nr_pages, sizeof(*bufs),
1123                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1124         if (unlikely(!bufs)) {
1125                 ret = -ENOMEM;
1126                 goto out_revert_acct;
1127         }
1128
1129         /*
1130          * The pipe array wraps around, so just start the new one at zero
1131          * and adjust the indexes.
1132          */
1133         if (pipe->nrbufs) {
1134                 unsigned int tail;
1135                 unsigned int head;
1136
1137                 tail = pipe->curbuf + pipe->nrbufs;
1138                 if (tail < pipe->buffers)
1139                         tail = 0;
1140                 else
1141                         tail &= (pipe->buffers - 1);
1142
1143                 head = pipe->nrbufs - tail;
1144                 if (head)
1145                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1146                 if (tail)
1147                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1148         }
1149
1150         pipe->curbuf = 0;
1151         kfree(pipe->bufs);
1152         pipe->bufs = bufs;
1153         pipe->buffers = nr_pages;
1154         return nr_pages * PAGE_SIZE;
1155
1156 out_revert_acct:
1157         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1158         return ret;
1159 }
1160
1161 /*
1162  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1163  * will return an error.
1164  */
1165 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1166                  size_t *lenp, loff_t *ppos)
1167 {
1168         unsigned int rounded_pipe_max_size;
1169         int ret;
1170
1171         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1172         if (ret < 0 || !write)
1173                 return ret;
1174
1175         rounded_pipe_max_size = round_pipe_size(pipe_max_size);
1176         if (rounded_pipe_max_size == 0)
1177                 return -EINVAL;
1178
1179         pipe_max_size = rounded_pipe_max_size;
1180         return ret;
1181 }
1182
1183 /*
1184  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1185  * location, so checking ->i_pipe is not enough to verify that this is a
1186  * pipe.
1187  */
1188 struct pipe_inode_info *get_pipe_info(struct file *file)
1189 {
1190         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1191 }
1192
1193 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1194 {
1195         struct pipe_inode_info *pipe;
1196         long ret;
1197
1198         pipe = get_pipe_info(file);
1199         if (!pipe)
1200                 return -EBADF;
1201
1202         __pipe_lock(pipe);
1203
1204         switch (cmd) {
1205         case F_SETPIPE_SZ:
1206                 ret = pipe_set_size(pipe, arg);
1207                 break;
1208         case F_GETPIPE_SZ:
1209                 ret = pipe->buffers * PAGE_SIZE;
1210                 break;
1211         default:
1212                 ret = -EINVAL;
1213                 break;
1214         }
1215
1216         __pipe_unlock(pipe);
1217         return ret;
1218 }
1219
1220 static const struct super_operations pipefs_ops = {
1221         .destroy_inode = free_inode_nonrcu,
1222         .statfs = simple_statfs,
1223 };
1224
1225 /*
1226  * pipefs should _never_ be mounted by userland - too much of security hassle,
1227  * no real gain from having the whole whorehouse mounted. So we don't need
1228  * any operations on the root directory. However, we need a non-trivial
1229  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1230  */
1231 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1232                          int flags, const char *dev_name, void *data)
1233 {
1234         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1235                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1236 }
1237
1238 static struct file_system_type pipe_fs_type = {
1239         .name           = "pipefs",
1240         .mount          = pipefs_mount,
1241         .kill_sb        = kill_anon_super,
1242 };
1243
1244 static int __init init_pipe_fs(void)
1245 {
1246         int err = register_filesystem(&pipe_fs_type);
1247
1248         if (!err) {
1249                 pipe_mnt = kern_mount(&pipe_fs_type);
1250                 if (IS_ERR(pipe_mnt)) {
1251                         err = PTR_ERR(pipe_mnt);
1252                         unregister_filesystem(&pipe_fs_type);
1253                 }
1254         }
1255         return err;
1256 }
1257
1258 fs_initcall(init_pipe_fs);