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 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  54  * matches default values.
  55  */
  56 unsigned long pipe_user_pages_hard;
  57 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  58
  59 /*
  60  * We use a start+len construction, which provides full use of the
  61  * allocated memory.
  62  * -- Florian Coosmann (FGC)
  63  *
  64  * Reads with count = 0 should always return 0.
  65  * -- Julian Bradfield 1999-06-07.
  66  *
  67  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  68  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  69  *
  70  * pipe_read & write cleanup
  71  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  72  */
  73
  74 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  75 {
  76         if (pipe->files)
  77                 mutex_lock_nested(&pipe->mutex, subclass);
  78 }
  79
  80 void pipe_lock(struct pipe_inode_info *pipe)
  81 {
  82         /*
  83          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  84          */
  85         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  86 }
  87 EXPORT_SYMBOL(pipe_lock);
  88
  89 void pipe_unlock(struct pipe_inode_info *pipe)
  90 {
  91         if (pipe->files)
  92                 mutex_unlock(&pipe->mutex);
  93 }
  94 EXPORT_SYMBOL(pipe_unlock);
  95
  96 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  97 {
  98         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  99 }
 100
 101 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 102 {
 103         mutex_unlock(&pipe->mutex);
 104 }
 105
 106 void pipe_double_lock(struct pipe_inode_info *pipe1,
 107                       struct pipe_inode_info *pipe2)
 108 {
 109         BUG_ON(pipe1 == pipe2);
 110
 111         if (pipe1 < pipe2) {
 112                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 113                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 114         } else {
 115                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 116                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 117         }
 118 }
 119
 120 /* Drop the inode semaphore and wait for a pipe event, atomically */
 121 void pipe_wait(struct pipe_inode_info *pipe)
 122 {
 123         DEFINE_WAIT(wait);
 124
 125         /*
 126          * Pipes are system-local resources, so sleeping on them
 127          * is considered a noninteractive wait:
 128          */
 129         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 130         pipe_unlock(pipe);
 131         schedule();
 132         finish_wait(&pipe->wait, &wait);
 133         pipe_lock(pipe);
 134 }
 135
 136 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 137                                   struct pipe_buffer *buf)
 138 {
 139         struct page *page = buf->page;
 140
 141         /*
 142          * If nobody else uses this page, and we don't already have a
 143          * temporary page, let's keep track of it as a one-deep
 144          * allocation cache. (Otherwise just release our reference to it)
 145          */
 146         if (page_count(page) == 1 && !pipe->tmp_page)
 147                 pipe->tmp_page = page;
 148         else
 149                 put_page(page);
 150 }
 151
 152 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 153                                struct pipe_buffer *buf)
 154 {
 155         struct page *page = buf->page;
 156
 157         if (page_count(page) == 1) {
 158                 if (memcg_kmem_enabled())
 159                         memcg_kmem_uncharge(page, 0);
 160                 __SetPageLocked(page);
 161                 return 0;
 162         }
 163         return 1;
 164 }
 165
 166 /**
 167  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 168  * @pipe:       the pipe that the buffer belongs to
 169  * @buf:        the buffer to attempt to steal
 170  *
 171  * Description:
 172  *      This function attempts to steal the &struct page attached to
 173  *      @buf. If successful, this function returns 0 and returns with
 174  *      the page locked. The caller may then reuse the page for whatever
 175  *      he wishes; the typical use is insertion into a different file
 176  *      page cache.
 177  */
 178 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 179                            struct pipe_buffer *buf)
 180 {
 181         struct page *page = buf->page;
 182
 183         /*
 184          * A reference of one is golden, that means that the owner of this
 185          * page is the only one holding a reference to it. lock the page
 186          * and return OK.
 187          */
 188         if (page_count(page) == 1) {
 189                 lock_page(page);
 190                 return 0;
 191         }
 192
 193         return 1;
 194 }
 195 EXPORT_SYMBOL(generic_pipe_buf_steal);
 196
 197 /**
 198  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 199  * @pipe:       the pipe that the buffer belongs to
 200  * @buf:        the buffer to get a reference to
 201  *
 202  * Description:
 203  *      This function grabs an extra reference to @buf. It's used in
 204  *      in the tee() system call, when we duplicate the buffers in one
 205  *      pipe into another.
 206  */
 207 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 208 {
 209         return try_get_page(buf->page);
 210 }
 211 EXPORT_SYMBOL(generic_pipe_buf_get);
 212
 213 /**
 214  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 215  * @info:       the pipe that the buffer belongs to
 216  * @buf:        the buffer to confirm
 217  *
 218  * Description:
 219  *      This function does nothing, because the generic pipe code uses
 220  *      pages that are always good when inserted into the pipe.
 221  */
 222 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 223                              struct pipe_buffer *buf)
 224 {
 225         return 0;
 226 }
 227 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 228
 229 /**
 230  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 231  * @pipe:       the pipe that the buffer belongs to
 232  * @buf:        the buffer to put a reference to
 233  *
 234  * Description:
 235  *      This function releases a reference to @buf.
 236  */
 237 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 238                               struct pipe_buffer *buf)
 239 {
 240         put_page(buf->page);
 241 }
 242 EXPORT_SYMBOL(generic_pipe_buf_release);
 243
 244 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 245         .can_merge = 1,
 246         .confirm = generic_pipe_buf_confirm,
 247         .release = anon_pipe_buf_release,
 248         .steal = anon_pipe_buf_steal,
 249         .get = generic_pipe_buf_get,
 250 };
 251
 252 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 253         .can_merge = 0,
 254         .confirm = generic_pipe_buf_confirm,
 255         .release = anon_pipe_buf_release,
 256         .steal = anon_pipe_buf_steal,
 257         .get = generic_pipe_buf_get,
 258 };
 259
 260 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 261         .can_merge = 0,
 262         .confirm = generic_pipe_buf_confirm,
 263         .release = anon_pipe_buf_release,
 264         .steal = anon_pipe_buf_steal,
 265         .get = generic_pipe_buf_get,
 266 };
 267
 268 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 269 {
 270         if (buf->ops == &anon_pipe_buf_ops)
 271                 buf->ops = &anon_pipe_buf_nomerge_ops;
 272 }
 273
 274 static ssize_t
 275 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 276 {
 277         size_t total_len = iov_iter_count(to);
 278         struct file *filp = iocb->ki_filp;
 279         struct pipe_inode_info *pipe = filp->private_data;
 280         int do_wakeup;
 281         ssize_t ret;
 282
 283         /* Null read succeeds. */
 284         if (unlikely(total_len == 0))
 285                 return 0;
 286
 287         do_wakeup = 0;
 288         ret = 0;
 289         __pipe_lock(pipe);
 290         for (;;) {
 291                 int bufs = pipe->nrbufs;
 292                 if (bufs) {
 293                         int curbuf = pipe->curbuf;
 294                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 295                         size_t chars = buf->len;
 296                         size_t written;
 297                         int error;
 298
 299                         if (chars > total_len)
 300                                 chars = total_len;
 301
 302                         error = pipe_buf_confirm(pipe, buf);
 303                         if (error) {
 304                                 if (!ret)
 305                                         ret = error;
 306                                 break;
 307                         }
 308
 309                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 310                         if (unlikely(written < chars)) {
 311                                 if (!ret)
 312                                         ret = -EFAULT;
 313                                 break;
 314                         }
 315                         ret += chars;
 316                         buf->offset += chars;
 317                         buf->len -= chars;
 318
 319                         /* Was it a packet buffer? Clean up and exit */
 320                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 321                                 total_len = chars;
 322                                 buf->len = 0;
 323                         }
 324
 325                         if (!buf->len) {
 326                                 pipe_buf_release(pipe, buf);
 327                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 328                                 pipe->curbuf = curbuf;
 329                                 pipe->nrbufs = --bufs;
 330                                 do_wakeup = 1;
 331                         }
 332                         total_len -= chars;
 333                         if (!total_len)
 334                                 break;  /* common path: read succeeded */
 335                 }
 336                 if (bufs)       /* More to do? */
 337                         continue;
 338                 if (!pipe->writers)
 339                         break;
 340                 if (!pipe->waiting_writers) {
 341                         /* syscall merging: Usually we must not sleep
 342                          * if O_NONBLOCK is set, or if we got some data.
 343                          * But if a writer sleeps in kernel space, then
 344                          * we can wait for that data without violating POSIX.
 345                          */
 346                         if (ret)
 347                                 break;
 348                         if (filp->f_flags & O_NONBLOCK) {
 349                                 ret = -EAGAIN;
 350                                 break;
 351                         }
 352                 }
 353                 if (signal_pending(current)) {
 354                         if (!ret)
 355                                 ret = -ERESTARTSYS;
 356                         break;
 357                 }
 358                 if (do_wakeup) {
 359                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 360                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 361                 }
 362                 pipe_wait(pipe);
 363         }
 364         __pipe_unlock(pipe);
 365
 366         /* Signal writers asynchronously that there is more room. */
 367         if (do_wakeup) {
 368                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 369                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 370         }
 371         if (ret > 0)
 372                 file_accessed(filp);
 373         return ret;
 374 }
 375
 376 static inline int is_packetized(struct file *file)
 377 {
 378         return (file->f_flags & O_DIRECT) != 0;
 379 }
 380
 381 static ssize_t
 382 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 383 {
 384         struct file *filp = iocb->ki_filp;
 385         struct pipe_inode_info *pipe = filp->private_data;
 386         ssize_t ret = 0;
 387         int do_wakeup = 0;
 388         size_t total_len = iov_iter_count(from);
 389         ssize_t chars;
 390
 391         /* Null write succeeds. */
 392         if (unlikely(total_len == 0))
 393                 return 0;
 394
 395         __pipe_lock(pipe);
 396
 397         if (!pipe->readers) {
 398                 send_sig(SIGPIPE, current, 0);
 399                 ret = -EPIPE;
 400                 goto out;
 401         }
 402
 403         /* We try to merge small writes */
 404         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 405         if (pipe->nrbufs && chars != 0) {
 406                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 407                                                         (pipe->buffers - 1);
 408                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 409                 int offset = buf->offset + buf->len;
 410
 411                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 412                         ret = pipe_buf_confirm(pipe, buf);
 413                         if (ret)
 414                                 goto out;
 415
 416                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 417                         if (unlikely(ret < chars)) {
 418                                 ret = -EFAULT;
 419                                 goto out;
 420                         }
 421                         do_wakeup = 1;
 422                         buf->len += ret;
 423                         if (!iov_iter_count(from))
 424                                 goto out;
 425                 }
 426         }
 427
 428         for (;;) {
 429                 int bufs;
 430
 431                 if (!pipe->readers) {
 432                         send_sig(SIGPIPE, current, 0);
 433                         if (!ret)
 434                                 ret = -EPIPE;
 435                         break;
 436                 }
 437                 bufs = pipe->nrbufs;
 438                 if (bufs < pipe->buffers) {
 439                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 440                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 441                         struct page *page = pipe->tmp_page;
 442                         int copied;
 443
 444                         if (!page) {
 445                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 446                                 if (unlikely(!page)) {
 447                                         ret = ret ? : -ENOMEM;
 448                                         break;
 449                                 }
 450                                 pipe->tmp_page = page;
 451                         }
 452                         /* Always wake up, even if the copy fails. Otherwise
 453                          * we lock up (O_NONBLOCK-)readers that sleep due to
 454                          * syscall merging.
 455                          * FIXME! Is this really true?
 456                          */
 457                         do_wakeup = 1;
 458                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 459                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 460                                 if (!ret)
 461                                         ret = -EFAULT;
 462                                 break;
 463                         }
 464                         ret += copied;
 465
 466                         /* Insert it into the buffer array */
 467                         buf->page = page;
 468                         buf->ops = &anon_pipe_buf_ops;
 469                         buf->offset = 0;
 470                         buf->len = copied;
 471                         buf->flags = 0;
 472                         if (is_packetized(filp)) {
 473                                 buf->ops = &packet_pipe_buf_ops;
 474                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 475                         }
 476                         pipe->nrbufs = ++bufs;
 477                         pipe->tmp_page = NULL;
 478
 479                         if (!iov_iter_count(from))
 480                                 break;
 481                 }
 482                 if (bufs < pipe->buffers)
 483                         continue;
 484                 if (filp->f_flags & O_NONBLOCK) {
 485                         if (!ret)
 486                                 ret = -EAGAIN;
 487                         break;
 488                 }
 489                 if (signal_pending(current)) {
 490                         if (!ret)
 491                                 ret = -ERESTARTSYS;
 492                         break;
 493                 }
 494                 if (do_wakeup) {
 495                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 496                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 497                         do_wakeup = 0;
 498                 }
 499                 pipe->waiting_writers++;
 500                 pipe_wait(pipe);
 501                 pipe->waiting_writers--;
 502         }
 503 out:
 504         __pipe_unlock(pipe);
 505         if (do_wakeup) {
 506                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 507                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 508         }
 509         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 510                 int err = file_update_time(filp);
 511                 if (err)
 512                         ret = err;
 513                 sb_end_write(file_inode(filp)->i_sb);
 514         }
 515         return ret;
 516 }
 517
 518 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 519 {
 520         struct pipe_inode_info *pipe = filp->private_data;
 521         int count, buf, nrbufs;
 522
 523         switch (cmd) {
 524                 case FIONREAD:
 525                         __pipe_lock(pipe);
 526                         count = 0;
 527                         buf = pipe->curbuf;
 528                         nrbufs = pipe->nrbufs;
 529                         while (--nrbufs >= 0) {
 530                                 count += pipe->bufs[buf].len;
 531                                 buf = (buf+1) & (pipe->buffers - 1);
 532                         }
 533                         __pipe_unlock(pipe);
 534
 535                         return put_user(count, (int __user *)arg);
 536                 default:
 537                         return -ENOIOCTLCMD;
 538         }
 539 }
 540
 541 /* No kernel lock held - fine */
 542 static __poll_t
 543 pipe_poll(struct file *filp, poll_table *wait)
 544 {
 545         __poll_t mask;
 546         struct pipe_inode_info *pipe = filp->private_data;
 547         int nrbufs;
 548
 549         poll_wait(filp, &pipe->wait, wait);
 550
 551         /* Reading only -- no need for acquiring the semaphore.  */
 552         nrbufs = pipe->nrbufs;
 553         mask = 0;
 554         if (filp->f_mode & FMODE_READ) {
 555                 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
 556                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 557                         mask |= EPOLLHUP;
 558         }
 559
 560         if (filp->f_mode & FMODE_WRITE) {
 561                 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
 562                 /*
 563                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 564                  * behave exactly like pipes for poll().
 565                  */
 566                 if (!pipe->readers)
 567                         mask |= EPOLLERR;
 568         }
 569
 570         return mask;
 571 }
 572
 573 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 574 {
 575         int kill = 0;
 576
 577         spin_lock(&inode->i_lock);
 578         if (!--pipe->files) {
 579                 inode->i_pipe = NULL;
 580                 kill = 1;
 581         }
 582         spin_unlock(&inode->i_lock);
 583
 584         if (kill)
 585                 free_pipe_info(pipe);
 586 }
 587
 588 static int
 589 pipe_release(struct inode *inode, struct file *file)
 590 {
 591         struct pipe_inode_info *pipe = file->private_data;
 592
 593         __pipe_lock(pipe);
 594         if (file->f_mode & FMODE_READ)
 595                 pipe->readers--;
 596         if (file->f_mode & FMODE_WRITE)
 597                 pipe->writers--;
 598
 599         if (pipe->readers || pipe->writers) {
 600                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 601                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 602                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 603         }
 604         __pipe_unlock(pipe);
 605
 606         put_pipe_info(inode, pipe);
 607         return 0;
 608 }
 609
 610 static int
 611 pipe_fasync(int fd, struct file *filp, int on)
 612 {
 613         struct pipe_inode_info *pipe = filp->private_data;
 614         int retval = 0;
 615
 616         __pipe_lock(pipe);
 617         if (filp->f_mode & FMODE_READ)
 618                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 619         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 620                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 621                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 622                         /* this can happen only if on == T */
 623                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 624         }
 625         __pipe_unlock(pipe);
 626         return retval;
 627 }
 628
 629 static unsigned long account_pipe_buffers(struct user_struct *user,
 630                                  unsigned long old, unsigned long new)
 631 {
 632         return atomic_long_add_return(new - old, &user->pipe_bufs);
 633 }
 634
 635 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 636 {
 637         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 638
 639         return soft_limit && user_bufs > soft_limit;
 640 }
 641
 642 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 643 {
 644         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 645
 646         return hard_limit && user_bufs > hard_limit;
 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         unsigned int max_size = READ_ONCE(pipe_max_size);
 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 > max_size && !capable(CAP_SYS_RESOURCE))
 667                 pipe_bufs = 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         struct inode *inode = get_pipe_inode();
 774         struct file *f;
 775
 776         if (!inode)
 777                 return -ENFILE;
 778
 779         f = alloc_file_pseudo(inode, pipe_mnt, "",
 780                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 781                                 &pipefifo_fops);
 782         if (IS_ERR(f)) {
 783                 free_pipe_info(inode->i_pipe);
 784                 iput(inode);
 785                 return PTR_ERR(f);
 786         }
 787
 788         f->private_data = inode->i_pipe;
 789
 790         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 791                                   &pipefifo_fops);
 792         if (IS_ERR(res[0])) {
 793                 put_pipe_info(inode, inode->i_pipe);
 794                 fput(f);
 795                 return PTR_ERR(res[0]);
 796         }
 797         res[0]->private_data = inode->i_pipe;
 798         res[1] = f;
 799         return 0;
 800 }
 801
 802 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 803 {
 804         int error;
 805         int fdw, fdr;
 806
 807         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 808                 return -EINVAL;
 809
 810         error = create_pipe_files(files, flags);
 811         if (error)
 812                 return error;
 813
 814         error = get_unused_fd_flags(flags);
 815         if (error < 0)
 816                 goto err_read_pipe;
 817         fdr = error;
 818
 819         error = get_unused_fd_flags(flags);
 820         if (error < 0)
 821                 goto err_fdr;
 822         fdw = error;
 823
 824         audit_fd_pair(fdr, fdw);
 825         fd[0] = fdr;
 826         fd[1] = fdw;
 827         return 0;
 828
 829  err_fdr:
 830         put_unused_fd(fdr);
 831  err_read_pipe:
 832         fput(files[0]);
 833         fput(files[1]);
 834         return error;
 835 }
 836
 837 int do_pipe_flags(int *fd, int flags)
 838 {
 839         struct file *files[2];
 840         int error = __do_pipe_flags(fd, files, flags);
 841         if (!error) {
 842                 fd_install(fd[0], files[0]);
 843                 fd_install(fd[1], files[1]);
 844         }
 845         return error;
 846 }
 847
 848 /*
 849  * sys_pipe() is the normal C calling standard for creating
 850  * a pipe. It's not the way Unix traditionally does this, though.
 851  */
 852 static int do_pipe2(int __user *fildes, int flags)
 853 {
 854         struct file *files[2];
 855         int fd[2];
 856         int error;
 857
 858         error = __do_pipe_flags(fd, files, flags);
 859         if (!error) {
 860                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 861                         fput(files[0]);
 862                         fput(files[1]);
 863                         put_unused_fd(fd[0]);
 864                         put_unused_fd(fd[1]);
 865                         error = -EFAULT;
 866                 } else {
 867                         fd_install(fd[0], files[0]);
 868                         fd_install(fd[1], files[1]);
 869                 }
 870         }
 871         return error;
 872 }
 873
 874 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 875 {
 876         return do_pipe2(fildes, flags);
 877 }
 878
 879 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 880 {
 881         return do_pipe2(fildes, 0);
 882 }
 883
 884 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 885 {
 886         int cur = *cnt;
 887
 888         while (cur == *cnt) {
 889                 pipe_wait(pipe);
 890                 if (signal_pending(current))
 891                         break;
 892         }
 893         return cur == *cnt ? -ERESTARTSYS : 0;
 894 }
 895
 896 static void wake_up_partner(struct pipe_inode_info *pipe)
 897 {
 898         wake_up_interruptible(&pipe->wait);
 899 }
 900
 901 static int fifo_open(struct inode *inode, struct file *filp)
 902 {
 903         struct pipe_inode_info *pipe;
 904         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 905         int ret;
 906
 907         filp->f_version = 0;
 908
 909         spin_lock(&inode->i_lock);
 910         if (inode->i_pipe) {
 911                 pipe = inode->i_pipe;
 912                 pipe->files++;
 913                 spin_unlock(&inode->i_lock);
 914         } else {
 915                 spin_unlock(&inode->i_lock);
 916                 pipe = alloc_pipe_info();
 917                 if (!pipe)
 918                         return -ENOMEM;
 919                 pipe->files = 1;
 920                 spin_lock(&inode->i_lock);
 921                 if (unlikely(inode->i_pipe)) {
 922                         inode->i_pipe->files++;
 923                         spin_unlock(&inode->i_lock);
 924                         free_pipe_info(pipe);
 925                         pipe = inode->i_pipe;
 926                 } else {
 927                         inode->i_pipe = pipe;
 928                         spin_unlock(&inode->i_lock);
 929                 }
 930         }
 931         filp->private_data = pipe;
 932         /* OK, we have a pipe and it's pinned down */
 933
 934         __pipe_lock(pipe);
 935
 936         /* We can only do regular read/write on fifos */
 937         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 938
 939         switch (filp->f_mode) {
 940         case FMODE_READ:
 941         /*
 942          *  O_RDONLY
 943          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 944          *  opened, even when there is no process writing the FIFO.
 945          */
 946                 pipe->r_counter++;
 947                 if (pipe->readers++ == 0)
 948                         wake_up_partner(pipe);
 949
 950                 if (!is_pipe && !pipe->writers) {
 951                         if ((filp->f_flags & O_NONBLOCK)) {
 952                                 /* suppress EPOLLHUP until we have
 953                                  * seen a writer */
 954                                 filp->f_version = pipe->w_counter;
 955                         } else {
 956                                 if (wait_for_partner(pipe, &pipe->w_counter))
 957                                         goto err_rd;
 958                         }
 959                 }
 960                 break;
 961
 962         case FMODE_WRITE:
 963         /*
 964          *  O_WRONLY
 965          *  POSIX.1 says that O_NONBLOCK means return -1 with
 966          *  errno=ENXIO when there is no process reading the FIFO.
 967          */
 968                 ret = -ENXIO;
 969                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 970                         goto err;
 971
 972                 pipe->w_counter++;
 973                 if (!pipe->writers++)
 974                         wake_up_partner(pipe);
 975
 976                 if (!is_pipe && !pipe->readers) {
 977                         if (wait_for_partner(pipe, &pipe->r_counter))
 978                                 goto err_wr;
 979                 }
 980                 break;
 981
 982         case FMODE_READ | FMODE_WRITE:
 983         /*
 984          *  O_RDWR
 985          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 986          *  This implementation will NEVER block on a O_RDWR open, since
 987          *  the process can at least talk to itself.
 988          */
 989
 990                 pipe->readers++;
 991                 pipe->writers++;
 992                 pipe->r_counter++;
 993                 pipe->w_counter++;
 994                 if (pipe->readers == 1 || pipe->writers == 1)
 995                         wake_up_partner(pipe);
 996                 break;
 997
 998         default:
 999                 ret = -EINVAL;
1000                 goto err;
1001         }
1002
1003         /* Ok! */
1004         __pipe_unlock(pipe);
1005         return 0;
1006
1007 err_rd:
1008         if (!--pipe->readers)
1009                 wake_up_interruptible(&pipe->wait);
1010         ret = -ERESTARTSYS;
1011         goto err;
1012
1013 err_wr:
1014         if (!--pipe->writers)
1015                 wake_up_interruptible(&pipe->wait);
1016         ret = -ERESTARTSYS;
1017         goto err;
1018
1019 err:
1020         __pipe_unlock(pipe);
1021
1022         put_pipe_info(inode, pipe);
1023         return ret;
1024 }
1025
1026 const struct file_operations pipefifo_fops = {
1027         .open           = fifo_open,
1028         .llseek         = no_llseek,
1029         .read_iter      = pipe_read,
1030         .write_iter     = pipe_write,
1031         .poll           = pipe_poll,
1032         .unlocked_ioctl = pipe_ioctl,
1033         .release        = pipe_release,
1034         .fasync         = pipe_fasync,
1035 };
1036
1037 /*
1038  * Currently we rely on the pipe array holding a power-of-2 number
1039  * of pages. Returns 0 on error.
1040  */
1041 unsigned int round_pipe_size(unsigned long size)
1042 {
1043         if (size > (1U << 31))
1044                 return 0;
1045
1046         /* Minimum pipe size, as required by POSIX */
1047         if (size < PAGE_SIZE)
1048                 return PAGE_SIZE;
1049
1050         return roundup_pow_of_two(size);
1051 }
1052
1053 /*
1054  * Allocate a new array of pipe buffers and copy the info over. Returns the
1055  * pipe size if successful, or return -ERROR on error.
1056  */
1057 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1058 {
1059         struct pipe_buffer *bufs;
1060         unsigned int size, nr_pages;
1061         unsigned long user_bufs;
1062         long ret = 0;
1063
1064         size = round_pipe_size(arg);
1065         nr_pages = size >> PAGE_SHIFT;
1066
1067         if (!nr_pages)
1068                 return -EINVAL;
1069
1070         /*
1071          * If trying to increase the pipe capacity, check that an
1072          * unprivileged user is not trying to exceed various limits
1073          * (soft limit check here, hard limit check just below).
1074          * Decreasing the pipe capacity is always permitted, even
1075          * if the user is currently over a limit.
1076          */
1077         if (nr_pages > pipe->buffers &&
1078                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1079                 return -EPERM;
1080
1081         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1082
1083         if (nr_pages > pipe->buffers &&
1084                         (too_many_pipe_buffers_hard(user_bufs) ||
1085                          too_many_pipe_buffers_soft(user_bufs)) &&
1086                         is_unprivileged_user()) {
1087                 ret = -EPERM;
1088                 goto out_revert_acct;
1089         }
1090
1091         /*
1092          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1093          * expect a lot of shrink+grow operations, just free and allocate
1094          * again like we would do for growing. If the pipe currently
1095          * contains more buffers than arg, then return busy.
1096          */
1097         if (nr_pages < pipe->nrbufs) {
1098                 ret = -EBUSY;
1099                 goto out_revert_acct;
1100         }
1101
1102         bufs = kcalloc(nr_pages, sizeof(*bufs),
1103                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1104         if (unlikely(!bufs)) {
1105                 ret = -ENOMEM;
1106                 goto out_revert_acct;
1107         }
1108
1109         /*
1110          * The pipe array wraps around, so just start the new one at zero
1111          * and adjust the indexes.
1112          */
1113         if (pipe->nrbufs) {
1114                 unsigned int tail;
1115                 unsigned int head;
1116
1117                 tail = pipe->curbuf + pipe->nrbufs;
1118                 if (tail < pipe->buffers)
1119                         tail = 0;
1120                 else
1121                         tail &= (pipe->buffers - 1);
1122
1123                 head = pipe->nrbufs - tail;
1124                 if (head)
1125                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1126                 if (tail)
1127                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1128         }
1129
1130         pipe->curbuf = 0;
1131         kfree(pipe->bufs);
1132         pipe->bufs = bufs;
1133         pipe->buffers = nr_pages;
1134         return nr_pages * PAGE_SIZE;
1135
1136 out_revert_acct:
1137         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1138         return ret;
1139 }
1140
1141 /*
1142  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1143  * location, so checking ->i_pipe is not enough to verify that this is a
1144  * pipe.
1145  */
1146 struct pipe_inode_info *get_pipe_info(struct file *file)
1147 {
1148         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1149 }
1150
1151 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1152 {
1153         struct pipe_inode_info *pipe;
1154         long ret;
1155
1156         pipe = get_pipe_info(file);
1157         if (!pipe)
1158                 return -EBADF;
1159
1160         __pipe_lock(pipe);
1161
1162         switch (cmd) {
1163         case F_SETPIPE_SZ:
1164                 ret = pipe_set_size(pipe, arg);
1165                 break;
1166         case F_GETPIPE_SZ:
1167                 ret = pipe->buffers * PAGE_SIZE;
1168                 break;
1169         default:
1170                 ret = -EINVAL;
1171                 break;
1172         }
1173
1174         __pipe_unlock(pipe);
1175         return ret;
1176 }
1177
1178 static const struct super_operations pipefs_ops = {
1179         .destroy_inode = free_inode_nonrcu,
1180         .statfs = simple_statfs,
1181 };
1182
1183 /*
1184  * pipefs should _never_ be mounted by userland - too much of security hassle,
1185  * no real gain from having the whole whorehouse mounted. So we don't need
1186  * any operations on the root directory. However, we need a non-trivial
1187  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1188  */
1189 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1190                          int flags, const char *dev_name, void *data)
1191 {
1192         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1193                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1194 }
1195
1196 static struct file_system_type pipe_fs_type = {
1197         .name           = "pipefs",
1198         .mount          = pipefs_mount,
1199         .kill_sb        = kill_anon_super,
1200 };
1201
1202 static int __init init_pipe_fs(void)
1203 {
1204         int err = register_filesystem(&pipe_fs_type);
1205
1206         if (!err) {
1207                 pipe_mnt = kern_mount(&pipe_fs_type);
1208                 if (IS_ERR(pipe_mnt)) {
1209                         err = PTR_ERR(pipe_mnt);
1210                         unregister_filesystem(&pipe_fs_type);
1211                 }
1212         }
1213         return err;
1214 }
1215
1216 fs_initcall(init_pipe_fs);