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/pseudo_fs.h>
  18 #include <linux/magic.h>
  19 #include <linux/pipe_fs_i.h>
  20 #include <linux/uio.h>
  21 #include <linux/highmem.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/audit.h>
  24 #include <linux/syscalls.h>
  25 #include <linux/fcntl.h>
  26 #include <linux/memcontrol.h>
  27
  28 #include <linux/uaccess.h>
  29 #include <asm/ioctls.h>
  30
  31 #include "internal.h"
  32
  33 /*
  34  * The max size that a non-root user is allowed to grow the pipe. Can
  35  * be set by root in /proc/sys/fs/pipe-max-size
  36  */
  37 unsigned int pipe_max_size = 1048576;
  38
  39 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  40  * matches default values.
  41  */
  42 unsigned long pipe_user_pages_hard;
  43 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  44
  45 /*
  46  * We use head and tail indices that aren't masked off, except at the point of
  47  * dereference, but rather they're allowed to wrap naturally.  This means there
  48  * isn't a dead spot in the buffer, but the ring has to be a power of two and
  49  * <= 2^31.
  50  * -- David Howells 2019-09-23.
  51  *
  52  * Reads with count = 0 should always return 0.
  53  * -- Julian Bradfield 1999-06-07.
  54  *
  55  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  56  * -- Jeremy Elson <[email protected]> 2001-08-16
  57  *
  58  * pipe_read & write cleanup
  59  * -- Manfred Spraul <[email protected]> 2002-05-09
  60  */
  61
  62 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  63 {
  64         if (pipe->files)
  65                 mutex_lock_nested(&pipe->mutex, subclass);
  66 }
  67
  68 void pipe_lock(struct pipe_inode_info *pipe)
  69 {
  70         /*
  71          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  72          */
  73         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  74 }
  75 EXPORT_SYMBOL(pipe_lock);
  76
  77 void pipe_unlock(struct pipe_inode_info *pipe)
  78 {
  79         if (pipe->files)
  80                 mutex_unlock(&pipe->mutex);
  81 }
  82 EXPORT_SYMBOL(pipe_unlock);
  83
  84 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  85 {
  86         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  87 }
  88
  89 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  90 {
  91         mutex_unlock(&pipe->mutex);
  92 }
  93
  94 void pipe_double_lock(struct pipe_inode_info *pipe1,
  95                       struct pipe_inode_info *pipe2)
  96 {
  97         BUG_ON(pipe1 == pipe2);
  98
  99         if (pipe1 < pipe2) {
 100                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 101                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 102         } else {
 103                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 104                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 105         }
 106 }
 107
 108 /* Drop the inode semaphore and wait for a pipe event, atomically */
 109 void pipe_wait(struct pipe_inode_info *pipe)
 110 {
 111         DEFINE_WAIT(rdwait);
 112         DEFINE_WAIT(wrwait);
 113
 114         /*
 115          * Pipes are system-local resources, so sleeping on them
 116          * is considered a noninteractive wait:
 117          */
 118         prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
 119         prepare_to_wait(&pipe->wr_wait, &wrwait, TASK_INTERRUPTIBLE);
 120         pipe_unlock(pipe);
 121         schedule();
 122         finish_wait(&pipe->rd_wait, &rdwait);
 123         finish_wait(&pipe->wr_wait, &wrwait);
 124         pipe_lock(pipe);
 125 }
 126
 127 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 128                                   struct pipe_buffer *buf)
 129 {
 130         struct page *page = buf->page;
 131
 132         /*
 133          * If nobody else uses this page, and we don't already have a
 134          * temporary page, let's keep track of it as a one-deep
 135          * allocation cache. (Otherwise just release our reference to it)
 136          */
 137         if (page_count(page) == 1 && !pipe->tmp_page)
 138                 pipe->tmp_page = page;
 139         else
 140                 put_page(page);
 141 }
 142
 143 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 144                 struct pipe_buffer *buf)
 145 {
 146         struct page *page = buf->page;
 147
 148         if (page_count(page) != 1)
 149                 return false;
 150         memcg_kmem_uncharge_page(page, 0);
 151         __SetPageLocked(page);
 152         return true;
 153 }
 154
 155 /**
 156  * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
 157  * @pipe:       the pipe that the buffer belongs to
 158  * @buf:        the buffer to attempt to steal
 159  *
 160  * Description:
 161  *      This function attempts to steal the &struct page attached to
 162  *      @buf. If successful, this function returns 0 and returns with
 163  *      the page locked. The caller may then reuse the page for whatever
 164  *      he wishes; the typical use is insertion into a different file
 165  *      page cache.
 166  */
 167 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 168                 struct pipe_buffer *buf)
 169 {
 170         struct page *page = buf->page;
 171
 172         /*
 173          * A reference of one is golden, that means that the owner of this
 174          * page is the only one holding a reference to it. lock the page
 175          * and return OK.
 176          */
 177         if (page_count(page) == 1) {
 178                 lock_page(page);
 179                 return true;
 180         }
 181         return false;
 182 }
 183 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
 184
 185 /**
 186  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 187  * @pipe:       the pipe that the buffer belongs to
 188  * @buf:        the buffer to get a reference to
 189  *
 190  * Description:
 191  *      This function grabs an extra reference to @buf. It's used in
 192  *      in the tee() system call, when we duplicate the buffers in one
 193  *      pipe into another.
 194  */
 195 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 196 {
 197         return try_get_page(buf->page);
 198 }
 199 EXPORT_SYMBOL(generic_pipe_buf_get);
 200
 201 /**
 202  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 203  * @pipe:       the pipe that the buffer belongs to
 204  * @buf:        the buffer to put a reference to
 205  *
 206  * Description:
 207  *      This function releases a reference to @buf.
 208  */
 209 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 210                               struct pipe_buffer *buf)
 211 {
 212         put_page(buf->page);
 213 }
 214 EXPORT_SYMBOL(generic_pipe_buf_release);
 215
 216 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 217         .release        = anon_pipe_buf_release,
 218         .try_steal      = anon_pipe_buf_try_steal,
 219         .get            = generic_pipe_buf_get,
 220 };
 221
 222 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 223 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 224 {
 225         unsigned int head = READ_ONCE(pipe->head);
 226         unsigned int tail = READ_ONCE(pipe->tail);
 227         unsigned int writers = READ_ONCE(pipe->writers);
 228
 229         return !pipe_empty(head, tail) || !writers;
 230 }
 231
 232 static ssize_t
 233 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 234 {
 235         size_t total_len = iov_iter_count(to);
 236         struct file *filp = iocb->ki_filp;
 237         struct pipe_inode_info *pipe = filp->private_data;
 238         bool was_full, wake_next_reader = false;
 239         ssize_t ret;
 240
 241         /* Null read succeeds. */
 242         if (unlikely(total_len == 0))
 243                 return 0;
 244
 245         ret = 0;
 246         __pipe_lock(pipe);
 247
 248         /*
 249          * We only wake up writers if the pipe was full when we started
 250          * reading in order to avoid unnecessary wakeups.
 251          *
 252          * But when we do wake up writers, we do so using a sync wakeup
 253          * (WF_SYNC), because we want them to get going and generate more
 254          * data for us.
 255          */
 256         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 257         for (;;) {
 258                 unsigned int head = pipe->head;
 259                 unsigned int tail = pipe->tail;
 260                 unsigned int mask = pipe->ring_size - 1;
 261
 262                 if (!pipe_empty(head, tail)) {
 263                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 264                         size_t chars = buf->len;
 265                         size_t written;
 266                         int error;
 267
 268                         if (chars > total_len)
 269                                 chars = total_len;
 270
 271                         error = pipe_buf_confirm(pipe, buf);
 272                         if (error) {
 273                                 if (!ret)
 274                                         ret = error;
 275                                 break;
 276                         }
 277
 278                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 279                         if (unlikely(written < chars)) {
 280                                 if (!ret)
 281                                         ret = -EFAULT;
 282                                 break;
 283                         }
 284                         ret += chars;
 285                         buf->offset += chars;
 286                         buf->len -= chars;
 287
 288                         /* Was it a packet buffer? Clean up and exit */
 289                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 290                                 total_len = chars;
 291                                 buf->len = 0;
 292                         }
 293
 294                         if (!buf->len) {
 295                                 pipe_buf_release(pipe, buf);
 296                                 spin_lock_irq(&pipe->rd_wait.lock);
 297                                 tail++;
 298                                 pipe->tail = tail;
 299                                 spin_unlock_irq(&pipe->rd_wait.lock);
 300                         }
 301                         total_len -= chars;
 302                         if (!total_len)
 303                                 break;  /* common path: read succeeded */
 304                         if (!pipe_empty(head, tail))    /* More to do? */
 305                                 continue;
 306                 }
 307
 308                 if (!pipe->writers)
 309                         break;
 310                 if (ret)
 311                         break;
 312                 if (filp->f_flags & O_NONBLOCK) {
 313                         ret = -EAGAIN;
 314                         break;
 315                 }
 316                 __pipe_unlock(pipe);
 317
 318                 /*
 319                  * We only get here if we didn't actually read anything.
 320                  *
 321                  * However, we could have seen (and removed) a zero-sized
 322                  * pipe buffer, and might have made space in the buffers
 323                  * that way.
 324                  *
 325                  * You can't make zero-sized pipe buffers by doing an empty
 326                  * write (not even in packet mode), but they can happen if
 327                  * the writer gets an EFAULT when trying to fill a buffer
 328                  * that already got allocated and inserted in the buffer
 329                  * array.
 330                  *
 331                  * So we still need to wake up any pending writers in the
 332                  * _very_ unlikely case that the pipe was full, but we got
 333                  * no data.
 334                  */
 335                 if (unlikely(was_full)) {
 336                         wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 337                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 338                 }
 339
 340                 /*
 341                  * But because we didn't read anything, at this point we can
 342                  * just return directly with -ERESTARTSYS if we're interrupted,
 343                  * since we've done any required wakeups and there's no need
 344                  * to mark anything accessed. And we've dropped the lock.
 345                  */
 346                 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 347                         return -ERESTARTSYS;
 348
 349                 __pipe_lock(pipe);
 350                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 351                 wake_next_reader = true;
 352         }
 353         if (pipe_empty(pipe->head, pipe->tail))
 354                 wake_next_reader = false;
 355         __pipe_unlock(pipe);
 356
 357         if (was_full) {
 358                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 359                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 360         }
 361         if (wake_next_reader)
 362                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 363         if (ret > 0)
 364                 file_accessed(filp);
 365         return ret;
 366 }
 367
 368 static inline int is_packetized(struct file *file)
 369 {
 370         return (file->f_flags & O_DIRECT) != 0;
 371 }
 372
 373 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 374 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 375 {
 376         unsigned int head = READ_ONCE(pipe->head);
 377         unsigned int tail = READ_ONCE(pipe->tail);
 378         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 379
 380         return !pipe_full(head, tail, max_usage) ||
 381                 !READ_ONCE(pipe->readers);
 382 }
 383
 384 static ssize_t
 385 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 386 {
 387         struct file *filp = iocb->ki_filp;
 388         struct pipe_inode_info *pipe = filp->private_data;
 389         unsigned int head;
 390         ssize_t ret = 0;
 391         size_t total_len = iov_iter_count(from);
 392         ssize_t chars;
 393         bool was_empty = false;
 394         bool wake_next_writer = false;
 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         /*
 409          * Only wake up if the pipe started out empty, since
 410          * otherwise there should be no readers waiting.
 411          *
 412          * If it wasn't empty we try to merge new data into
 413          * the last buffer.
 414          *
 415          * That naturally merges small writes, but it also
 416          * page-aligs the rest of the writes for large writes
 417          * spanning multiple pages.
 418          */
 419         head = pipe->head;
 420         was_empty = pipe_empty(head, pipe->tail);
 421         chars = total_len & (PAGE_SIZE-1);
 422         if (chars && !was_empty) {
 423                 unsigned int mask = pipe->ring_size - 1;
 424                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 425                 int offset = buf->offset + buf->len;
 426
 427                 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 428                     offset + chars <= PAGE_SIZE) {
 429                         ret = pipe_buf_confirm(pipe, buf);
 430                         if (ret)
 431                                 goto out;
 432
 433                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 434                         if (unlikely(ret < chars)) {
 435                                 ret = -EFAULT;
 436                                 goto out;
 437                         }
 438
 439                         buf->len += ret;
 440                         if (!iov_iter_count(from))
 441                                 goto out;
 442                 }
 443         }
 444
 445         for (;;) {
 446                 if (!pipe->readers) {
 447                         send_sig(SIGPIPE, current, 0);
 448                         if (!ret)
 449                                 ret = -EPIPE;
 450                         break;
 451                 }
 452
 453                 head = pipe->head;
 454                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 455                         unsigned int mask = pipe->ring_size - 1;
 456                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 457                         struct page *page = pipe->tmp_page;
 458                         int copied;
 459
 460                         if (!page) {
 461                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 462                                 if (unlikely(!page)) {
 463                                         ret = ret ? : -ENOMEM;
 464                                         break;
 465                                 }
 466                                 pipe->tmp_page = page;
 467                         }
 468
 469                         /* Allocate a slot in the ring in advance and attach an
 470                          * empty buffer.  If we fault or otherwise fail to use
 471                          * it, either the reader will consume it or it'll still
 472                          * be there for the next write.
 473                          */
 474                         spin_lock_irq(&pipe->rd_wait.lock);
 475
 476                         head = pipe->head;
 477                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 478                                 spin_unlock_irq(&pipe->rd_wait.lock);
 479                                 continue;
 480                         }
 481
 482                         pipe->head = head + 1;
 483                         spin_unlock_irq(&pipe->rd_wait.lock);
 484
 485                         /* Insert it into the buffer array */
 486                         buf = &pipe->bufs[head & mask];
 487                         buf->page = page;
 488                         buf->ops = &anon_pipe_buf_ops;
 489                         buf->offset = 0;
 490                         buf->len = 0;
 491                         if (is_packetized(filp))
 492                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 493                         else
 494                                 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 495                         pipe->tmp_page = NULL;
 496
 497                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 498                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 499                                 if (!ret)
 500                                         ret = -EFAULT;
 501                                 break;
 502                         }
 503                         ret += copied;
 504                         buf->offset = 0;
 505                         buf->len = copied;
 506
 507                         if (!iov_iter_count(from))
 508                                 break;
 509                 }
 510
 511                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 512                         continue;
 513
 514                 /* Wait for buffer space to become available. */
 515                 if (filp->f_flags & O_NONBLOCK) {
 516                         if (!ret)
 517                                 ret = -EAGAIN;
 518                         break;
 519                 }
 520                 if (signal_pending(current)) {
 521                         if (!ret)
 522                                 ret = -ERESTARTSYS;
 523                         break;
 524                 }
 525
 526                 /*
 527                  * We're going to release the pipe lock and wait for more
 528                  * space. We wake up any readers if necessary, and then
 529                  * after waiting we need to re-check whether the pipe
 530                  * become empty while we dropped the lock.
 531                  */
 532                 __pipe_unlock(pipe);
 533                 if (was_empty) {
 534                         wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 535                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 536                 }
 537                 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 538                 __pipe_lock(pipe);
 539                 was_empty = pipe_empty(pipe->head, pipe->tail);
 540                 wake_next_writer = true;
 541         }
 542 out:
 543         if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 544                 wake_next_writer = false;
 545         __pipe_unlock(pipe);
 546
 547         /*
 548          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 549          * want the reader to start processing things asap, rather than
 550          * leave the data pending.
 551          *
 552          * This is particularly important for small writes, because of
 553          * how (for example) the GNU make jobserver uses small writes to
 554          * wake up pending jobs
 555          */
 556         if (was_empty) {
 557                 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 558                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 559         }
 560         if (wake_next_writer)
 561                 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 562         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 563                 int err = file_update_time(filp);
 564                 if (err)
 565                         ret = err;
 566                 sb_end_write(file_inode(filp)->i_sb);
 567         }
 568         return ret;
 569 }
 570
 571 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 572 {
 573         struct pipe_inode_info *pipe = filp->private_data;
 574         int count, head, tail, mask;
 575
 576         switch (cmd) {
 577                 case FIONREAD:
 578                         __pipe_lock(pipe);
 579                         count = 0;
 580                         head = pipe->head;
 581                         tail = pipe->tail;
 582                         mask = pipe->ring_size - 1;
 583
 584                         while (tail != head) {
 585                                 count += pipe->bufs[tail & mask].len;
 586                                 tail++;
 587                         }
 588                         __pipe_unlock(pipe);
 589
 590                         return put_user(count, (int __user *)arg);
 591                 default:
 592                         return -ENOIOCTLCMD;
 593         }
 594 }
 595
 596 /* No kernel lock held - fine */
 597 static __poll_t
 598 pipe_poll(struct file *filp, poll_table *wait)
 599 {
 600         __poll_t mask;
 601         struct pipe_inode_info *pipe = filp->private_data;
 602         unsigned int head, tail;
 603
 604         /*
 605          * Reading pipe state only -- no need for acquiring the semaphore.
 606          *
 607          * But because this is racy, the code has to add the
 608          * entry to the poll table _first_ ..
 609          */
 610         if (filp->f_mode & FMODE_READ)
 611                 poll_wait(filp, &pipe->rd_wait, wait);
 612         if (filp->f_mode & FMODE_WRITE)
 613                 poll_wait(filp, &pipe->wr_wait, wait);
 614
 615         /*
 616          * .. and only then can you do the racy tests. That way,
 617          * if something changes and you got it wrong, the poll
 618          * table entry will wake you up and fix it.
 619          */
 620         head = READ_ONCE(pipe->head);
 621         tail = READ_ONCE(pipe->tail);
 622
 623         mask = 0;
 624         if (filp->f_mode & FMODE_READ) {
 625                 if (!pipe_empty(head, tail))
 626                         mask |= EPOLLIN | EPOLLRDNORM;
 627                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 628                         mask |= EPOLLHUP;
 629         }
 630
 631         if (filp->f_mode & FMODE_WRITE) {
 632                 if (!pipe_full(head, tail, pipe->max_usage))
 633                         mask |= EPOLLOUT | EPOLLWRNORM;
 634                 /*
 635                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 636                  * behave exactly like pipes for poll().
 637                  */
 638                 if (!pipe->readers)
 639                         mask |= EPOLLERR;
 640         }
 641
 642         return mask;
 643 }
 644
 645 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 646 {
 647         int kill = 0;
 648
 649         spin_lock(&inode->i_lock);
 650         if (!--pipe->files) {
 651                 inode->i_pipe = NULL;
 652                 kill = 1;
 653         }
 654         spin_unlock(&inode->i_lock);
 655
 656         if (kill)
 657                 free_pipe_info(pipe);
 658 }
 659
 660 static int
 661 pipe_release(struct inode *inode, struct file *file)
 662 {
 663         struct pipe_inode_info *pipe = file->private_data;
 664
 665         __pipe_lock(pipe);
 666         if (file->f_mode & FMODE_READ)
 667                 pipe->readers--;
 668         if (file->f_mode & FMODE_WRITE)
 669                 pipe->writers--;
 670
 671         /* Was that the last reader or writer, but not the other side? */
 672         if (!pipe->readers != !pipe->writers) {
 673                 wake_up_interruptible_all(&pipe->rd_wait);
 674                 wake_up_interruptible_all(&pipe->wr_wait);
 675                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 676                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 677         }
 678         __pipe_unlock(pipe);
 679
 680         put_pipe_info(inode, pipe);
 681         return 0;
 682 }
 683
 684 static int
 685 pipe_fasync(int fd, struct file *filp, int on)
 686 {
 687         struct pipe_inode_info *pipe = filp->private_data;
 688         int retval = 0;
 689
 690         __pipe_lock(pipe);
 691         if (filp->f_mode & FMODE_READ)
 692                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 693         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 694                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 695                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 696                         /* this can happen only if on == T */
 697                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 698         }
 699         __pipe_unlock(pipe);
 700         return retval;
 701 }
 702
 703 static unsigned long account_pipe_buffers(struct user_struct *user,
 704                                  unsigned long old, unsigned long new)
 705 {
 706         return atomic_long_add_return(new - old, &user->pipe_bufs);
 707 }
 708
 709 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 710 {
 711         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 712
 713         return soft_limit && user_bufs > soft_limit;
 714 }
 715
 716 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 717 {
 718         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 719
 720         return hard_limit && user_bufs > hard_limit;
 721 }
 722
 723 static bool is_unprivileged_user(void)
 724 {
 725         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 726 }
 727
 728 struct pipe_inode_info *alloc_pipe_info(void)
 729 {
 730         struct pipe_inode_info *pipe;
 731         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 732         struct user_struct *user = get_current_user();
 733         unsigned long user_bufs;
 734         unsigned int max_size = READ_ONCE(pipe_max_size);
 735
 736         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 737         if (pipe == NULL)
 738                 goto out_free_uid;
 739
 740         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 741                 pipe_bufs = max_size >> PAGE_SHIFT;
 742
 743         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 744
 745         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 746                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 747                 pipe_bufs = 1;
 748         }
 749
 750         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 751                 goto out_revert_acct;
 752
 753         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 754                              GFP_KERNEL_ACCOUNT);
 755
 756         if (pipe->bufs) {
 757                 init_waitqueue_head(&pipe->rd_wait);
 758                 init_waitqueue_head(&pipe->wr_wait);
 759                 pipe->r_counter = pipe->w_counter = 1;
 760                 pipe->max_usage = pipe_bufs;
 761                 pipe->ring_size = pipe_bufs;
 762                 pipe->user = user;
 763                 mutex_init(&pipe->mutex);
 764                 return pipe;
 765         }
 766
 767 out_revert_acct:
 768         (void) account_pipe_buffers(user, pipe_bufs, 0);
 769         kfree(pipe);
 770 out_free_uid:
 771         free_uid(user);
 772         return NULL;
 773 }
 774
 775 void free_pipe_info(struct pipe_inode_info *pipe)
 776 {
 777         int i;
 778
 779         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 780         free_uid(pipe->user);
 781         for (i = 0; i < pipe->ring_size; i++) {
 782                 struct pipe_buffer *buf = pipe->bufs + i;
 783                 if (buf->ops)
 784                         pipe_buf_release(pipe, buf);
 785         }
 786         if (pipe->tmp_page)
 787                 __free_page(pipe->tmp_page);
 788         kfree(pipe->bufs);
 789         kfree(pipe);
 790 }
 791
 792 static struct vfsmount *pipe_mnt __read_mostly;
 793
 794 /*
 795  * pipefs_dname() is called from d_path().
 796  */
 797 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 798 {
 799         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 800                                 d_inode(dentry)->i_ino);
 801 }
 802
 803 static const struct dentry_operations pipefs_dentry_operations = {
 804         .d_dname        = pipefs_dname,
 805 };
 806
 807 static struct inode * get_pipe_inode(void)
 808 {
 809         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 810         struct pipe_inode_info *pipe;
 811
 812         if (!inode)
 813                 goto fail_inode;
 814
 815         inode->i_ino = get_next_ino();
 816
 817         pipe = alloc_pipe_info();
 818         if (!pipe)
 819                 goto fail_iput;
 820
 821         inode->i_pipe = pipe;
 822         pipe->files = 2;
 823         pipe->readers = pipe->writers = 1;
 824         inode->i_fop = &pipefifo_fops;
 825
 826         /*
 827          * Mark the inode dirty from the very beginning,
 828          * that way it will never be moved to the dirty
 829          * list because "mark_inode_dirty()" will think
 830          * that it already _is_ on the dirty list.
 831          */
 832         inode->i_state = I_DIRTY;
 833         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 834         inode->i_uid = current_fsuid();
 835         inode->i_gid = current_fsgid();
 836         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 837
 838         return inode;
 839
 840 fail_iput:
 841         iput(inode);
 842
 843 fail_inode:
 844         return NULL;
 845 }
 846
 847 int create_pipe_files(struct file **res, int flags)
 848 {
 849         struct inode *inode = get_pipe_inode();
 850         struct file *f;
 851
 852         if (!inode)
 853                 return -ENFILE;
 854
 855         f = alloc_file_pseudo(inode, pipe_mnt, "",
 856                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 857                                 &pipefifo_fops);
 858         if (IS_ERR(f)) {
 859                 free_pipe_info(inode->i_pipe);
 860                 iput(inode);
 861                 return PTR_ERR(f);
 862         }
 863
 864         f->private_data = inode->i_pipe;
 865
 866         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 867                                   &pipefifo_fops);
 868         if (IS_ERR(res[0])) {
 869                 put_pipe_info(inode, inode->i_pipe);
 870                 fput(f);
 871                 return PTR_ERR(res[0]);
 872         }
 873         res[0]->private_data = inode->i_pipe;
 874         res[1] = f;
 875         stream_open(inode, res[0]);
 876         stream_open(inode, res[1]);
 877         return 0;
 878 }
 879
 880 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 881 {
 882         int error;
 883         int fdw, fdr;
 884
 885         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 886                 return -EINVAL;
 887
 888         error = create_pipe_files(files, flags);
 889         if (error)
 890                 return error;
 891
 892         error = get_unused_fd_flags(flags);
 893         if (error < 0)
 894                 goto err_read_pipe;
 895         fdr = error;
 896
 897         error = get_unused_fd_flags(flags);
 898         if (error < 0)
 899                 goto err_fdr;
 900         fdw = error;
 901
 902         audit_fd_pair(fdr, fdw);
 903         fd[0] = fdr;
 904         fd[1] = fdw;
 905         return 0;
 906
 907  err_fdr:
 908         put_unused_fd(fdr);
 909  err_read_pipe:
 910         fput(files[0]);
 911         fput(files[1]);
 912         return error;
 913 }
 914
 915 int do_pipe_flags(int *fd, int flags)
 916 {
 917         struct file *files[2];
 918         int error = __do_pipe_flags(fd, files, flags);
 919         if (!error) {
 920                 fd_install(fd[0], files[0]);
 921                 fd_install(fd[1], files[1]);
 922         }
 923         return error;
 924 }
 925
 926 /*
 927  * sys_pipe() is the normal C calling standard for creating
 928  * a pipe. It's not the way Unix traditionally does this, though.
 929  */
 930 static int do_pipe2(int __user *fildes, int flags)
 931 {
 932         struct file *files[2];
 933         int fd[2];
 934         int error;
 935
 936         error = __do_pipe_flags(fd, files, flags);
 937         if (!error) {
 938                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 939                         fput(files[0]);
 940                         fput(files[1]);
 941                         put_unused_fd(fd[0]);
 942                         put_unused_fd(fd[1]);
 943                         error = -EFAULT;
 944                 } else {
 945                         fd_install(fd[0], files[0]);
 946                         fd_install(fd[1], files[1]);
 947                 }
 948         }
 949         return error;
 950 }
 951
 952 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 953 {
 954         return do_pipe2(fildes, flags);
 955 }
 956
 957 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 958 {
 959         return do_pipe2(fildes, 0);
 960 }
 961
 962 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 963 {
 964         int cur = *cnt;
 965
 966         while (cur == *cnt) {
 967                 pipe_wait(pipe);
 968                 if (signal_pending(current))
 969                         break;
 970         }
 971         return cur == *cnt ? -ERESTARTSYS : 0;
 972 }
 973
 974 static void wake_up_partner(struct pipe_inode_info *pipe)
 975 {
 976         wake_up_interruptible_all(&pipe->rd_wait);
 977         wake_up_interruptible_all(&pipe->wr_wait);
 978 }
 979
 980 static int fifo_open(struct inode *inode, struct file *filp)
 981 {
 982         struct pipe_inode_info *pipe;
 983         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 984         int ret;
 985
 986         filp->f_version = 0;
 987
 988         spin_lock(&inode->i_lock);
 989         if (inode->i_pipe) {
 990                 pipe = inode->i_pipe;
 991                 pipe->files++;
 992                 spin_unlock(&inode->i_lock);
 993         } else {
 994                 spin_unlock(&inode->i_lock);
 995                 pipe = alloc_pipe_info();
 996                 if (!pipe)
 997                         return -ENOMEM;
 998                 pipe->files = 1;
 999                 spin_lock(&inode->i_lock);
1000                 if (unlikely(inode->i_pipe)) {
1001                         inode->i_pipe->files++;
1002                         spin_unlock(&inode->i_lock);
1003                         free_pipe_info(pipe);
1004                         pipe = inode->i_pipe;
1005                 } else {
1006                         inode->i_pipe = pipe;
1007                         spin_unlock(&inode->i_lock);
1008                 }
1009         }
1010         filp->private_data = pipe;
1011         /* OK, we have a pipe and it's pinned down */
1012
1013         __pipe_lock(pipe);
1014
1015         /* We can only do regular read/write on fifos */
1016         stream_open(inode, filp);
1017
1018         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1019         case FMODE_READ:
1020         /*
1021          *  O_RDONLY
1022          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1023          *  opened, even when there is no process writing the FIFO.
1024          */
1025                 pipe->r_counter++;
1026                 if (pipe->readers++ == 0)
1027                         wake_up_partner(pipe);
1028
1029                 if (!is_pipe && !pipe->writers) {
1030                         if ((filp->f_flags & O_NONBLOCK)) {
1031                                 /* suppress EPOLLHUP until we have
1032                                  * seen a writer */
1033                                 filp->f_version = pipe->w_counter;
1034                         } else {
1035                                 if (wait_for_partner(pipe, &pipe->w_counter))
1036                                         goto err_rd;
1037                         }
1038                 }
1039                 break;
1040
1041         case FMODE_WRITE:
1042         /*
1043          *  O_WRONLY
1044          *  POSIX.1 says that O_NONBLOCK means return -1 with
1045          *  errno=ENXIO when there is no process reading the FIFO.
1046          */
1047                 ret = -ENXIO;
1048                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1049                         goto err;
1050
1051                 pipe->w_counter++;
1052                 if (!pipe->writers++)
1053                         wake_up_partner(pipe);
1054
1055                 if (!is_pipe && !pipe->readers) {
1056                         if (wait_for_partner(pipe, &pipe->r_counter))
1057                                 goto err_wr;
1058                 }
1059                 break;
1060
1061         case FMODE_READ | FMODE_WRITE:
1062         /*
1063          *  O_RDWR
1064          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1065          *  This implementation will NEVER block on a O_RDWR open, since
1066          *  the process can at least talk to itself.
1067          */
1068
1069                 pipe->readers++;
1070                 pipe->writers++;
1071                 pipe->r_counter++;
1072                 pipe->w_counter++;
1073                 if (pipe->readers == 1 || pipe->writers == 1)
1074                         wake_up_partner(pipe);
1075                 break;
1076
1077         default:
1078                 ret = -EINVAL;
1079                 goto err;
1080         }
1081
1082         /* Ok! */
1083         __pipe_unlock(pipe);
1084         return 0;
1085
1086 err_rd:
1087         if (!--pipe->readers)
1088                 wake_up_interruptible(&pipe->wr_wait);
1089         ret = -ERESTARTSYS;
1090         goto err;
1091
1092 err_wr:
1093         if (!--pipe->writers)
1094                 wake_up_interruptible_all(&pipe->rd_wait);
1095         ret = -ERESTARTSYS;
1096         goto err;
1097
1098 err:
1099         __pipe_unlock(pipe);
1100
1101         put_pipe_info(inode, pipe);
1102         return ret;
1103 }
1104
1105 const struct file_operations pipefifo_fops = {
1106         .open           = fifo_open,
1107         .llseek         = no_llseek,
1108         .read_iter      = pipe_read,
1109         .write_iter     = pipe_write,
1110         .poll           = pipe_poll,
1111         .unlocked_ioctl = pipe_ioctl,
1112         .release        = pipe_release,
1113         .fasync         = pipe_fasync,
1114 };
1115
1116 /*
1117  * Currently we rely on the pipe array holding a power-of-2 number
1118  * of pages. Returns 0 on error.
1119  */
1120 unsigned int round_pipe_size(unsigned long size)
1121 {
1122         if (size > (1U << 31))
1123                 return 0;
1124
1125         /* Minimum pipe size, as required by POSIX */
1126         if (size < PAGE_SIZE)
1127                 return PAGE_SIZE;
1128
1129         return roundup_pow_of_two(size);
1130 }
1131
1132 /*
1133  * Allocate a new array of pipe buffers and copy the info over. Returns the
1134  * pipe size if successful, or return -ERROR on error.
1135  */
1136 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1137 {
1138         struct pipe_buffer *bufs;
1139         unsigned int size, nr_slots, head, tail, mask, n;
1140         unsigned long user_bufs;
1141         long ret = 0;
1142
1143         size = round_pipe_size(arg);
1144         nr_slots = size >> PAGE_SHIFT;
1145
1146         if (!nr_slots)
1147                 return -EINVAL;
1148
1149         /*
1150          * If trying to increase the pipe capacity, check that an
1151          * unprivileged user is not trying to exceed various limits
1152          * (soft limit check here, hard limit check just below).
1153          * Decreasing the pipe capacity is always permitted, even
1154          * if the user is currently over a limit.
1155          */
1156         if (nr_slots > pipe->ring_size &&
1157                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1158                 return -EPERM;
1159
1160         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1161
1162         if (nr_slots > pipe->ring_size &&
1163                         (too_many_pipe_buffers_hard(user_bufs) ||
1164                          too_many_pipe_buffers_soft(user_bufs)) &&
1165                         is_unprivileged_user()) {
1166                 ret = -EPERM;
1167                 goto out_revert_acct;
1168         }
1169
1170         /*
1171          * We can shrink the pipe, if arg is greater than the ring occupancy.
1172          * Since we don't expect a lot of shrink+grow operations, just free and
1173          * allocate again like we would do for growing.  If the pipe currently
1174          * contains more buffers than arg, then return busy.
1175          */
1176         mask = pipe->ring_size - 1;
1177         head = pipe->head;
1178         tail = pipe->tail;
1179         n = pipe_occupancy(pipe->head, pipe->tail);
1180         if (nr_slots < n) {
1181                 ret = -EBUSY;
1182                 goto out_revert_acct;
1183         }
1184
1185         bufs = kcalloc(nr_slots, sizeof(*bufs),
1186                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1187         if (unlikely(!bufs)) {
1188                 ret = -ENOMEM;
1189                 goto out_revert_acct;
1190         }
1191
1192         /*
1193          * The pipe array wraps around, so just start the new one at zero
1194          * and adjust the indices.
1195          */
1196         if (n > 0) {
1197                 unsigned int h = head & mask;
1198                 unsigned int t = tail & mask;
1199                 if (h > t) {
1200                         memcpy(bufs, pipe->bufs + t,
1201                                n * sizeof(struct pipe_buffer));
1202                 } else {
1203                         unsigned int tsize = pipe->ring_size - t;
1204                         if (h > 0)
1205                                 memcpy(bufs + tsize, pipe->bufs,
1206                                        h * sizeof(struct pipe_buffer));
1207                         memcpy(bufs, pipe->bufs + t,
1208                                tsize * sizeof(struct pipe_buffer));
1209                 }
1210         }
1211
1212         head = n;
1213         tail = 0;
1214
1215         kfree(pipe->bufs);
1216         pipe->bufs = bufs;
1217         pipe->ring_size = nr_slots;
1218         pipe->max_usage = nr_slots;
1219         pipe->tail = tail;
1220         pipe->head = head;
1221
1222         /* This might have made more room for writers */
1223         wake_up_interruptible(&pipe->wr_wait);
1224         return pipe->max_usage * PAGE_SIZE;
1225
1226 out_revert_acct:
1227         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1228         return ret;
1229 }
1230
1231 /*
1232  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1233  * location, so checking ->i_pipe is not enough to verify that this is a
1234  * pipe.
1235  */
1236 struct pipe_inode_info *get_pipe_info(struct file *file)
1237 {
1238         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1239 }
1240
1241 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1242 {
1243         struct pipe_inode_info *pipe;
1244         long ret;
1245
1246         pipe = get_pipe_info(file);
1247         if (!pipe)
1248                 return -EBADF;
1249
1250         __pipe_lock(pipe);
1251
1252         switch (cmd) {
1253         case F_SETPIPE_SZ:
1254                 ret = pipe_set_size(pipe, arg);
1255                 break;
1256         case F_GETPIPE_SZ:
1257                 ret = pipe->max_usage * PAGE_SIZE;
1258                 break;
1259         default:
1260                 ret = -EINVAL;
1261                 break;
1262         }
1263
1264         __pipe_unlock(pipe);
1265         return ret;
1266 }
1267
1268 static const struct super_operations pipefs_ops = {
1269         .destroy_inode = free_inode_nonrcu,
1270         .statfs = simple_statfs,
1271 };
1272
1273 /*
1274  * pipefs should _never_ be mounted by userland - too much of security hassle,
1275  * no real gain from having the whole whorehouse mounted. So we don't need
1276  * any operations on the root directory. However, we need a non-trivial
1277  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1278  */
1279
1280 static int pipefs_init_fs_context(struct fs_context *fc)
1281 {
1282         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1283         if (!ctx)
1284                 return -ENOMEM;
1285         ctx->ops = &pipefs_ops;
1286         ctx->dops = &pipefs_dentry_operations;
1287         return 0;
1288 }
1289
1290 static struct file_system_type pipe_fs_type = {
1291         .name           = "pipefs",
1292         .init_fs_context = pipefs_init_fs_context,
1293         .kill_sb        = kill_anon_super,
1294 };
1295
1296 static int __init init_pipe_fs(void)
1297 {
1298         int err = register_filesystem(&pipe_fs_type);
1299
1300         if (!err) {
1301                 pipe_mnt = kern_mount(&pipe_fs_type);
1302                 if (IS_ERR(pipe_mnt)) {
1303                         err = PTR_ERR(pipe_mnt);
1304                         unregister_filesystem(&pipe_fs_type);
1305                 }
1306         }
1307         return err;
1308 }
1309
1310 fs_initcall(init_pipe_fs);