fs/pipe.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/pipe.c
   4  *
   5  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/file.h>
  10 #include <linux/poll.h>
  11 #include <linux/slab.h>
  12 #include <linux/module.h>
  13 #include <linux/init.h>
  14 #include <linux/fs.h>
  15 #include <linux/log2.h>
  16 #include <linux/mount.h>
  17 #include <linux/magic.h>
  18 #include <linux/pipe_fs_i.h>
  19 #include <linux/uio.h>
  20 #include <linux/highmem.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/audit.h>
  23 #include <linux/syscalls.h>
  24 #include <linux/fcntl.h>
  25 #include <linux/memcontrol.h>
  26
  27 #include <linux/uaccess.h>
  28 #include <asm/ioctls.h>
  29
  30 #include "internal.h"
  31
  32 /*
  33  * New pipe buffers will be restricted to this size while the user is exceeding
  34  * their pipe buffer quota. The general pipe use case needs at least two
  35  * buffers: one for data yet to be read, and one for new data. If this is less
  36  * than two, then a write to a non-empty pipe may block even if the pipe is not
  37  * full. This can occur with GNU make jobserver or similar uses of pipes as
  38  * semaphores: multiple processes may be waiting to write tokens back to the
  39  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  40  *
  41  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  42  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  43  * emptied.
  44  */
  45 #define PIPE_MIN_DEF_BUFFERS 2
  46
  47 /*
  48  * The max size that a non-root user is allowed to grow the pipe. Can
  49  * be set by root in /proc/sys/fs/pipe-max-size
  50  */
  51 unsigned int pipe_max_size = 1048576;
  52
  53 /*
  54  * Minimum pipe size, as required by POSIX
  55  */
  56 unsigned int pipe_min_size = PAGE_SIZE;
  57
  58 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  59  * matches default values.
  60  */
  61 unsigned long pipe_user_pages_hard;
  62 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  63
  64 /*
  65  * We use a start+len construction, which provides full use of the
  66  * allocated memory.
  67  * -- Florian Coosmann (FGC)
  68  *
  69  * Reads with count = 0 should always return 0.
  70  * -- Julian Bradfield 1999-06-07.
  71  *
  72  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  73  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  74  *
  75  * pipe_read & write cleanup
  76  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  77  */
  78
  79 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  80 {
  81         if (pipe->files)
  82                 mutex_lock_nested(&pipe->mutex, subclass);
  83 }
  84
  85 void pipe_lock(struct pipe_inode_info *pipe)
  86 {
  87         /*
  88          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  89          */
  90         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  91 }
  92 EXPORT_SYMBOL(pipe_lock);
  93
  94 void pipe_unlock(struct pipe_inode_info *pipe)
  95 {
  96         if (pipe->files)
  97                 mutex_unlock(&pipe->mutex);
  98 }
  99 EXPORT_SYMBOL(pipe_unlock);
 100
 101 static inline void __pipe_lock(struct pipe_inode_info *pipe)
 102 {
 103         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 104 }
 105
 106 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 107 {
 108         mutex_unlock(&pipe->mutex);
 109 }
 110
 111 void pipe_double_lock(struct pipe_inode_info *pipe1,
 112                       struct pipe_inode_info *pipe2)
 113 {
 114         BUG_ON(pipe1 == pipe2);
 115
 116         if (pipe1 < pipe2) {
 117                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 118                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 119         } else {
 120                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 121                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 122         }
 123 }
 124
 125 /* Drop the inode semaphore and wait for a pipe event, atomically */
 126 void pipe_wait(struct pipe_inode_info *pipe)
 127 {
 128         DEFINE_WAIT(wait);
 129
 130         /*
 131          * Pipes are system-local resources, so sleeping on them
 132          * is considered a noninteractive wait:
 133          */
 134         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 135         pipe_unlock(pipe);
 136         schedule();
 137         finish_wait(&pipe->wait, &wait);
 138         pipe_lock(pipe);
 139 }
 140
 141 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 142                                   struct pipe_buffer *buf)
 143 {
 144         struct page *page = buf->page;
 145
 146         /*
 147          * If nobody else uses this page, and we don't already have a
 148          * temporary page, let's keep track of it as a one-deep
 149          * allocation cache. (Otherwise just release our reference to it)
 150          */
 151         if (page_count(page) == 1 && !pipe->tmp_page)
 152                 pipe->tmp_page = page;
 153         else
 154                 put_page(page);
 155 }
 156
 157 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 158                                struct pipe_buffer *buf)
 159 {
 160         struct page *page = buf->page;
 161
 162         if (page_count(page) == 1) {
 163                 if (memcg_kmem_enabled())
 164                         memcg_kmem_uncharge(page, 0);
 165                 __SetPageLocked(page);
 166                 return 0;
 167         }
 168         return 1;
 169 }
 170
 171 /**
 172  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 173  * @pipe:       the pipe that the buffer belongs to
 174  * @buf:        the buffer to attempt to steal
 175  *
 176  * Description:
 177  *      This function attempts to steal the &struct page attached to
 178  *      @buf. If successful, this function returns 0 and returns with
 179  *      the page locked. The caller may then reuse the page for whatever
 180  *      he wishes; the typical use is insertion into a different file
 181  *      page cache.
 182  */
 183 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 184                            struct pipe_buffer *buf)
 185 {
 186         struct page *page = buf->page;
 187
 188         /*
 189          * A reference of one is golden, that means that the owner of this
 190          * page is the only one holding a reference to it. lock the page
 191          * and return OK.
 192          */
 193         if (page_count(page) == 1) {
 194                 lock_page(page);
 195                 return 0;
 196         }
 197
 198         return 1;
 199 }
 200 EXPORT_SYMBOL(generic_pipe_buf_steal);
 201
 202 /**
 203  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 204  * @pipe:       the pipe that the buffer belongs to
 205  * @buf:        the buffer to get a reference to
 206  *
 207  * Description:
 208  *      This function grabs an extra reference to @buf. It's used in
 209  *      in the tee() system call, when we duplicate the buffers in one
 210  *      pipe into another.
 211  */
 212 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 213 {
 214         return try_get_page(buf->page);
 215 }
 216 EXPORT_SYMBOL(generic_pipe_buf_get);
 217
 218 /**
 219  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 220  * @info:       the pipe that the buffer belongs to
 221  * @buf:        the buffer to confirm
 222  *
 223  * Description:
 224  *      This function does nothing, because the generic pipe code uses
 225  *      pages that are always good when inserted into the pipe.
 226  */
 227 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 228                              struct pipe_buffer *buf)
 229 {
 230         return 0;
 231 }
 232 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 233
 234 /**
 235  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 236  * @pipe:       the pipe that the buffer belongs to
 237  * @buf:        the buffer to put a reference to
 238  *
 239  * Description:
 240  *      This function releases a reference to @buf.
 241  */
 242 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 243                               struct pipe_buffer *buf)
 244 {
 245         put_page(buf->page);
 246 }
 247 EXPORT_SYMBOL(generic_pipe_buf_release);
 248
 249 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 250         .can_merge = 1,
 251         .confirm = generic_pipe_buf_confirm,
 252         .release = anon_pipe_buf_release,
 253         .steal = anon_pipe_buf_steal,
 254         .get = generic_pipe_buf_get,
 255 };
 256
 257 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 258         .can_merge = 0,
 259         .confirm = generic_pipe_buf_confirm,
 260         .release = anon_pipe_buf_release,
 261         .steal = anon_pipe_buf_steal,
 262         .get = generic_pipe_buf_get,
 263 };
 264
 265 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 266         .can_merge = 0,
 267         .confirm = generic_pipe_buf_confirm,
 268         .release = anon_pipe_buf_release,
 269         .steal = anon_pipe_buf_steal,
 270         .get = generic_pipe_buf_get,
 271 };
 272
 273 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 274 {
 275         if (buf->ops == &anon_pipe_buf_ops)
 276                 buf->ops = &anon_pipe_buf_nomerge_ops;
 277 }
 278
 279 static ssize_t
 280 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 281 {
 282         size_t total_len = iov_iter_count(to);
 283         struct file *filp = iocb->ki_filp;
 284         struct pipe_inode_info *pipe = filp->private_data;
 285         int do_wakeup;
 286         ssize_t ret;
 287
 288         /* Null read succeeds. */
 289         if (unlikely(total_len == 0))
 290                 return 0;
 291
 292         do_wakeup = 0;
 293         ret = 0;
 294         __pipe_lock(pipe);
 295         for (;;) {
 296                 int bufs = pipe->nrbufs;
 297                 if (bufs) {
 298                         int curbuf = pipe->curbuf;
 299                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 300                         size_t chars = buf->len;
 301                         size_t written;
 302                         int error;
 303
 304                         if (chars > total_len)
 305                                 chars = total_len;
 306
 307                         error = pipe_buf_confirm(pipe, buf);
 308                         if (error) {
 309                                 if (!ret)
 310                                         ret = error;
 311                                 break;
 312                         }
 313
 314                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 315                         if (unlikely(written < chars)) {
 316                                 if (!ret)
 317                                         ret = -EFAULT;
 318                                 break;
 319                         }
 320                         ret += chars;
 321                         buf->offset += chars;
 322                         buf->len -= chars;
 323
 324                         /* Was it a packet buffer? Clean up and exit */
 325                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 326                                 total_len = chars;
 327                                 buf->len = 0;
 328                         }
 329
 330                         if (!buf->len) {
 331                                 pipe_buf_release(pipe, buf);
 332                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 333                                 pipe->curbuf = curbuf;
 334                                 pipe->nrbufs = --bufs;
 335                                 do_wakeup = 1;
 336                         }
 337                         total_len -= chars;
 338                         if (!total_len)
 339                                 break;  /* common path: read succeeded */
 340                 }
 341                 if (bufs)       /* More to do? */
 342                         continue;
 343                 if (!pipe->writers)
 344                         break;
 345                 if (!pipe->waiting_writers) {
 346                         /* syscall merging: Usually we must not sleep
 347                          * if O_NONBLOCK is set, or if we got some data.
 348                          * But if a writer sleeps in kernel space, then
 349                          * we can wait for that data without violating POSIX.
 350                          */
 351                         if (ret)
 352                                 break;
 353                         if (filp->f_flags & O_NONBLOCK) {
 354                                 ret = -EAGAIN;
 355                                 break;
 356                         }
 357                 }
 358                 if (signal_pending(current)) {
 359                         if (!ret)
 360                                 ret = -ERESTARTSYS;
 361                         break;
 362                 }
 363                 if (do_wakeup) {
 364                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 365                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 366                 }
 367                 pipe_wait(pipe);
 368         }
 369         __pipe_unlock(pipe);
 370
 371         /* Signal writers asynchronously that there is more room. */
 372         if (do_wakeup) {
 373                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 374                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 375         }
 376         if (ret > 0)
 377                 file_accessed(filp);
 378         return ret;
 379 }
 380
 381 static inline int is_packetized(struct file *file)
 382 {
 383         return (file->f_flags & O_DIRECT) != 0;
 384 }
 385
 386 static ssize_t
 387 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 388 {
 389         struct file *filp = iocb->ki_filp;
 390         struct pipe_inode_info *pipe = filp->private_data;
 391         ssize_t ret = 0;
 392         int do_wakeup = 0;
 393         size_t total_len = iov_iter_count(from);
 394         ssize_t chars;
 395
 396         /* Null write succeeds. */
 397         if (unlikely(total_len == 0))
 398                 return 0;
 399
 400         __pipe_lock(pipe);
 401
 402         if (!pipe->readers) {
 403                 send_sig(SIGPIPE, current, 0);
 404                 ret = -EPIPE;
 405                 goto out;
 406         }
 407
 408         /* We try to merge small writes */
 409         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 410         if (pipe->nrbufs && chars != 0) {
 411                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 412                                                         (pipe->buffers - 1);
 413                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 414                 int offset = buf->offset + buf->len;
 415
 416                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 417                         ret = pipe_buf_confirm(pipe, buf);
 418                         if (ret)
 419                                 goto out;
 420
 421                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 422                         if (unlikely(ret < chars)) {
 423                                 ret = -EFAULT;
 424                                 goto out;
 425                         }
 426                         do_wakeup = 1;
 427                         buf->len += ret;
 428                         if (!iov_iter_count(from))
 429                                 goto out;
 430                 }
 431         }
 432
 433         for (;;) {
 434                 int bufs;
 435
 436                 if (!pipe->readers) {
 437                         send_sig(SIGPIPE, current, 0);
 438                         if (!ret)
 439                                 ret = -EPIPE;
 440                         break;
 441                 }
 442                 bufs = pipe->nrbufs;
 443                 if (bufs < pipe->buffers) {
 444                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 445                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 446                         struct page *page = pipe->tmp_page;
 447                         int copied;
 448
 449                         if (!page) {
 450                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 451                                 if (unlikely(!page)) {
 452                                         ret = ret ? : -ENOMEM;
 453                                         break;
 454                                 }
 455                                 pipe->tmp_page = page;
 456                         }
 457                         /* Always wake up, even if the copy fails. Otherwise
 458                          * we lock up (O_NONBLOCK-)readers that sleep due to
 459                          * syscall merging.
 460                          * FIXME! Is this really true?
 461                          */
 462                         do_wakeup = 1;
 463                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 464                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 465                                 if (!ret)
 466                                         ret = -EFAULT;
 467                                 break;
 468                         }
 469                         ret += copied;
 470
 471                         /* Insert it into the buffer array */
 472                         buf->page = page;
 473                         buf->ops = &anon_pipe_buf_ops;
 474                         buf->offset = 0;
 475                         buf->len = copied;
 476                         buf->flags = 0;
 477                         if (is_packetized(filp)) {
 478                                 buf->ops = &packet_pipe_buf_ops;
 479                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 480                         }
 481                         pipe->nrbufs = ++bufs;
 482                         pipe->tmp_page = NULL;
 483
 484                         if (!iov_iter_count(from))
 485                                 break;
 486                 }
 487                 if (bufs < pipe->buffers)
 488                         continue;
 489                 if (filp->f_flags & O_NONBLOCK) {
 490                         if (!ret)
 491                                 ret = -EAGAIN;
 492                         break;
 493                 }
 494                 if (signal_pending(current)) {
 495                         if (!ret)
 496                                 ret = -ERESTARTSYS;
 497                         break;
 498                 }
 499                 if (do_wakeup) {
 500                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 501                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 502                         do_wakeup = 0;
 503                 }
 504                 pipe->waiting_writers++;
 505                 pipe_wait(pipe);
 506                 pipe->waiting_writers--;
 507         }
 508 out:
 509         __pipe_unlock(pipe);
 510         if (do_wakeup) {
 511                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 512                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 513         }
 514         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 515                 int err = file_update_time(filp);
 516                 if (err)
 517                         ret = err;
 518                 sb_end_write(file_inode(filp)->i_sb);
 519         }
 520         return ret;
 521 }
 522
 523 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 524 {
 525         struct pipe_inode_info *pipe = filp->private_data;
 526         int count, buf, nrbufs;
 527
 528         switch (cmd) {
 529                 case FIONREAD:
 530                         __pipe_lock(pipe);
 531                         count = 0;
 532                         buf = pipe->curbuf;
 533                         nrbufs = pipe->nrbufs;
 534                         while (--nrbufs >= 0) {
 535                                 count += pipe->bufs[buf].len;
 536                                 buf = (buf+1) & (pipe->buffers - 1);
 537                         }
 538                         __pipe_unlock(pipe);
 539
 540                         return put_user(count, (int __user *)arg);
 541                 default:
 542                         return -ENOIOCTLCMD;
 543         }
 544 }
 545
 546 /* No kernel lock held - fine */
 547 static unsigned int
 548 pipe_poll(struct file *filp, poll_table *wait)
 549 {
 550         unsigned int mask;
 551         struct pipe_inode_info *pipe = filp->private_data;
 552         int nrbufs;
 553
 554         poll_wait(filp, &pipe->wait, wait);
 555
 556         /* Reading only -- no need for acquiring the semaphore.  */
 557         nrbufs = pipe->nrbufs;
 558         mask = 0;
 559         if (filp->f_mode & FMODE_READ) {
 560                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 561                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 562                         mask |= POLLHUP;
 563         }
 564
 565         if (filp->f_mode & FMODE_WRITE) {
 566                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 567                 /*
 568                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 569                  * behave exactly like pipes for poll().
 570                  */
 571                 if (!pipe->readers)
 572                         mask |= POLLERR;
 573         }
 574
 575         return mask;
 576 }
 577
 578 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 579 {
 580         int kill = 0;
 581
 582         spin_lock(&inode->i_lock);
 583         if (!--pipe->files) {
 584                 inode->i_pipe = NULL;
 585                 kill = 1;
 586         }
 587         spin_unlock(&inode->i_lock);
 588
 589         if (kill)
 590                 free_pipe_info(pipe);
 591 }
 592
 593 static int
 594 pipe_release(struct inode *inode, struct file *file)
 595 {
 596         struct pipe_inode_info *pipe = file->private_data;
 597
 598         __pipe_lock(pipe);
 599         if (file->f_mode & FMODE_READ)
 600                 pipe->readers--;
 601         if (file->f_mode & FMODE_WRITE)
 602                 pipe->writers--;
 603
 604         if (pipe->readers || pipe->writers) {
 605                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 606                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 607                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 608         }
 609         __pipe_unlock(pipe);
 610
 611         put_pipe_info(inode, pipe);
 612         return 0;
 613 }
 614
 615 static int
 616 pipe_fasync(int fd, struct file *filp, int on)
 617 {
 618         struct pipe_inode_info *pipe = filp->private_data;
 619         int retval = 0;
 620
 621         __pipe_lock(pipe);
 622         if (filp->f_mode & FMODE_READ)
 623                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 624         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 625                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 626                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 627                         /* this can happen only if on == T */
 628                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 629         }
 630         __pipe_unlock(pipe);
 631         return retval;
 632 }
 633
 634 static unsigned long account_pipe_buffers(struct user_struct *user,
 635                                  unsigned long old, unsigned long new)
 636 {
 637         return atomic_long_add_return(new - old, &user->pipe_bufs);
 638 }
 639
 640 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 641 {
 642         return pipe_user_pages_soft && user_bufs > pipe_user_pages_soft;
 643 }
 644
 645 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 646 {
 647         return pipe_user_pages_hard && user_bufs > pipe_user_pages_hard;
 648 }
 649
 650 static bool is_unprivileged_user(void)
 651 {
 652         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 653 }
 654
 655 struct pipe_inode_info *alloc_pipe_info(void)
 656 {
 657         struct pipe_inode_info *pipe;
 658         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 659         struct user_struct *user = get_current_user();
 660         unsigned long user_bufs;
 661
 662         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 663         if (pipe == NULL)
 664                 goto out_free_uid;
 665
 666         if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 667                 pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 668
 669         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 670
 671         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 672                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 673                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 674         }
 675
 676         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 677                 goto out_revert_acct;
 678
 679         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 680                              GFP_KERNEL_ACCOUNT);
 681
 682         if (pipe->bufs) {
 683                 init_waitqueue_head(&pipe->wait);
 684                 pipe->r_counter = pipe->w_counter = 1;
 685                 pipe->buffers = pipe_bufs;
 686                 pipe->user = user;
 687                 mutex_init(&pipe->mutex);
 688                 return pipe;
 689         }
 690
 691 out_revert_acct:
 692         (void) account_pipe_buffers(user, pipe_bufs, 0);
 693         kfree(pipe);
 694 out_free_uid:
 695         free_uid(user);
 696         return NULL;
 697 }
 698
 699 void free_pipe_info(struct pipe_inode_info *pipe)
 700 {
 701         int i;
 702
 703         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 704         free_uid(pipe->user);
 705         for (i = 0; i < pipe->buffers; i++) {
 706                 struct pipe_buffer *buf = pipe->bufs + i;
 707                 if (buf->ops)
 708                         pipe_buf_release(pipe, buf);
 709         }
 710         if (pipe->tmp_page)
 711                 __free_page(pipe->tmp_page);
 712         kfree(pipe->bufs);
 713         kfree(pipe);
 714 }
 715
 716 static struct vfsmount *pipe_mnt __read_mostly;
 717
 718 /*
 719  * pipefs_dname() is called from d_path().
 720  */
 721 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 722 {
 723         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 724                                 d_inode(dentry)->i_ino);
 725 }
 726
 727 static const struct dentry_operations pipefs_dentry_operations = {
 728         .d_dname        = pipefs_dname,
 729 };
 730
 731 static struct inode * get_pipe_inode(void)
 732 {
 733         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 734         struct pipe_inode_info *pipe;
 735
 736         if (!inode)
 737                 goto fail_inode;
 738
 739         inode->i_ino = get_next_ino();
 740
 741         pipe = alloc_pipe_info();
 742         if (!pipe)
 743                 goto fail_iput;
 744
 745         inode->i_pipe = pipe;
 746         pipe->files = 2;
 747         pipe->readers = pipe->writers = 1;
 748         inode->i_fop = &pipefifo_fops;
 749
 750         /*
 751          * Mark the inode dirty from the very beginning,
 752          * that way it will never be moved to the dirty
 753          * list because "mark_inode_dirty()" will think
 754          * that it already _is_ on the dirty list.
 755          */
 756         inode->i_state = I_DIRTY;
 757         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 758         inode->i_uid = current_fsuid();
 759         inode->i_gid = current_fsgid();
 760         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 761
 762         return inode;
 763
 764 fail_iput:
 765         iput(inode);
 766
 767 fail_inode:
 768         return NULL;
 769 }
 770
 771 int create_pipe_files(struct file **res, int flags)
 772 {
 773         int err;
 774         struct inode *inode = get_pipe_inode();
 775         struct file *f;
 776         struct path path;
 777
 778         if (!inode)
 779                 return -ENFILE;
 780
 781         err = -ENOMEM;
 782         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &empty_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_douintvec_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);