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 <jelson@circlemud.org> 2001-08-16
  57  *
  58  * pipe_read & write cleanup
  59  * -- Manfred Spraul <manfred@colorfullife.com> 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(wait);
 112
 113         /*
 114          * Pipes are system-local resources, so sleeping on them
 115          * is considered a noninteractive wait:
 116          */
 117         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 118         pipe_unlock(pipe);
 119         schedule();
 120         finish_wait(&pipe->wait, &wait);
 121         pipe_lock(pipe);
 122 }
 123
 124 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 125                                   struct pipe_buffer *buf)
 126 {
 127         struct page *page = buf->page;
 128
 129         /*
 130          * If nobody else uses this page, and we don't already have a
 131          * temporary page, let's keep track of it as a one-deep
 132          * allocation cache. (Otherwise just release our reference to it)
 133          */
 134         if (page_count(page) == 1 && !pipe->tmp_page)
 135                 pipe->tmp_page = page;
 136         else
 137                 put_page(page);
 138 }
 139
 140 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 141                                struct pipe_buffer *buf)
 142 {
 143         struct page *page = buf->page;
 144
 145         if (page_count(page) == 1) {
 146                 memcg_kmem_uncharge(page, 0);
 147                 __SetPageLocked(page);
 148                 return 0;
 149         }
 150         return 1;
 151 }
 152
 153 /**
 154  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 155  * @pipe:       the pipe that the buffer belongs to
 156  * @buf:        the buffer to attempt to steal
 157  *
 158  * Description:
 159  *      This function attempts to steal the &struct page attached to
 160  *      @buf. If successful, this function returns 0 and returns with
 161  *      the page locked. The caller may then reuse the page for whatever
 162  *      he wishes; the typical use is insertion into a different file
 163  *      page cache.
 164  */
 165 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 166                            struct pipe_buffer *buf)
 167 {
 168         struct page *page = buf->page;
 169
 170         /*
 171          * A reference of one is golden, that means that the owner of this
 172          * page is the only one holding a reference to it. lock the page
 173          * and return OK.
 174          */
 175         if (page_count(page) == 1) {
 176                 lock_page(page);
 177                 return 0;
 178         }
 179
 180         return 1;
 181 }
 182 EXPORT_SYMBOL(generic_pipe_buf_steal);
 183
 184 /**
 185  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 186  * @pipe:       the pipe that the buffer belongs to
 187  * @buf:        the buffer to get a reference to
 188  *
 189  * Description:
 190  *      This function grabs an extra reference to @buf. It's used in
 191  *      in the tee() system call, when we duplicate the buffers in one
 192  *      pipe into another.
 193  */
 194 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 195 {
 196         return try_get_page(buf->page);
 197 }
 198 EXPORT_SYMBOL(generic_pipe_buf_get);
 199
 200 /**
 201  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 202  * @info:       the pipe that the buffer belongs to
 203  * @buf:        the buffer to confirm
 204  *
 205  * Description:
 206  *      This function does nothing, because the generic pipe code uses
 207  *      pages that are always good when inserted into the pipe.
 208  */
 209 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 210                              struct pipe_buffer *buf)
 211 {
 212         return 0;
 213 }
 214 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 215
 216 /**
 217  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 218  * @pipe:       the pipe that the buffer belongs to
 219  * @buf:        the buffer to put a reference to
 220  *
 221  * Description:
 222  *      This function releases a reference to @buf.
 223  */
 224 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 225                               struct pipe_buffer *buf)
 226 {
 227         put_page(buf->page);
 228 }
 229 EXPORT_SYMBOL(generic_pipe_buf_release);
 230
 231 /* New data written to a pipe may be appended to a buffer with this type. */
 232 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 233         .confirm = generic_pipe_buf_confirm,
 234         .release = anon_pipe_buf_release,
 235         .steal = anon_pipe_buf_steal,
 236         .get = generic_pipe_buf_get,
 237 };
 238
 239 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 240         .confirm = generic_pipe_buf_confirm,
 241         .release = anon_pipe_buf_release,
 242         .steal = anon_pipe_buf_steal,
 243         .get = generic_pipe_buf_get,
 244 };
 245
 246 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 247         .confirm = generic_pipe_buf_confirm,
 248         .release = anon_pipe_buf_release,
 249         .steal = anon_pipe_buf_steal,
 250         .get = generic_pipe_buf_get,
 251 };
 252
 253 /**
 254  * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
 255  * @buf:        the buffer to mark
 256  *
 257  * Description:
 258  *      This function ensures that no future writes will be merged into the
 259  *      given &struct pipe_buffer. This is necessary when multiple pipe buffers
 260  *      share the same backing page.
 261  */
 262 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 263 {
 264         if (buf->ops == &anon_pipe_buf_ops)
 265                 buf->ops = &anon_pipe_buf_nomerge_ops;
 266 }
 267
 268 static bool pipe_buf_can_merge(struct pipe_buffer *buf)
 269 {
 270         return buf->ops == &anon_pipe_buf_ops;
 271 }
 272
 273 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 274 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 275 {
 276         unsigned int head = READ_ONCE(pipe->head);
 277         unsigned int tail = READ_ONCE(pipe->tail);
 278         unsigned int writers = READ_ONCE(pipe->writers);
 279
 280         return !pipe_empty(head, tail) || !writers;
 281 }
 282
 283 static ssize_t
 284 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 285 {
 286         size_t total_len = iov_iter_count(to);
 287         struct file *filp = iocb->ki_filp;
 288         struct pipe_inode_info *pipe = filp->private_data;
 289         bool was_full;
 290         ssize_t ret;
 291
 292         /* Null read succeeds. */
 293         if (unlikely(total_len == 0))
 294                 return 0;
 295
 296         ret = 0;
 297         __pipe_lock(pipe);
 298
 299         /*
 300          * We only wake up writers if the pipe was full when we started
 301          * reading in order to avoid unnecessary wakeups.
 302          *
 303          * But when we do wake up writers, we do so using a sync wakeup
 304          * (WF_SYNC), because we want them to get going and generate more
 305          * data for us.
 306          */
 307         was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 308         for (;;) {
 309                 unsigned int head = pipe->head;
 310                 unsigned int tail = pipe->tail;
 311                 unsigned int mask = pipe->ring_size - 1;
 312
 313                 if (!pipe_empty(head, tail)) {
 314                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 315                         size_t chars = buf->len;
 316                         size_t written;
 317                         int error;
 318
 319                         if (chars > total_len)
 320                                 chars = total_len;
 321
 322                         error = pipe_buf_confirm(pipe, buf);
 323                         if (error) {
 324                                 if (!ret)
 325                                         ret = error;
 326                                 break;
 327                         }
 328
 329                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 330                         if (unlikely(written < chars)) {
 331                                 if (!ret)
 332                                         ret = -EFAULT;
 333                                 break;
 334                         }
 335                         ret += chars;
 336                         buf->offset += chars;
 337                         buf->len -= chars;
 338
 339                         /* Was it a packet buffer? Clean up and exit */
 340                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 341                                 total_len = chars;
 342                                 buf->len = 0;
 343                         }
 344
 345                         if (!buf->len) {
 346                                 pipe_buf_release(pipe, buf);
 347                                 spin_lock_irq(&pipe->wait.lock);
 348                                 tail++;
 349                                 pipe->tail = tail;
 350                                 spin_unlock_irq(&pipe->wait.lock);
 351                         }
 352                         total_len -= chars;
 353                         if (!total_len)
 354                                 break;  /* common path: read succeeded */
 355                         if (!pipe_empty(head, tail))    /* More to do? */
 356                                 continue;
 357                 }
 358
 359                 if (!pipe->writers)
 360                         break;
 361                 if (ret)
 362                         break;
 363                 if (filp->f_flags & O_NONBLOCK) {
 364                         ret = -EAGAIN;
 365                         break;
 366                 }
 367                 if (signal_pending(current)) {
 368                         if (!ret)
 369                                 ret = -ERESTARTSYS;
 370                         break;
 371                 }
 372                 __pipe_unlock(pipe);
 373                 if (was_full) {
 374                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 375                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 376                 }
 377                 wait_event_interruptible(pipe->wait, pipe_readable(pipe));
 378                 __pipe_lock(pipe);
 379                 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 380         }
 381         __pipe_unlock(pipe);
 382
 383         if (was_full) {
 384                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 385                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 386         }
 387         if (ret > 0)
 388                 file_accessed(filp);
 389         return ret;
 390 }
 391
 392 static inline int is_packetized(struct file *file)
 393 {
 394         return (file->f_flags & O_DIRECT) != 0;
 395 }
 396
 397 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 398 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 399 {
 400         unsigned int head = READ_ONCE(pipe->head);
 401         unsigned int tail = READ_ONCE(pipe->tail);
 402         unsigned int max_usage = READ_ONCE(pipe->max_usage);
 403
 404         return !pipe_full(head, tail, max_usage) ||
 405                 !READ_ONCE(pipe->readers);
 406 }
 407
 408 static ssize_t
 409 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 410 {
 411         struct file *filp = iocb->ki_filp;
 412         struct pipe_inode_info *pipe = filp->private_data;
 413         unsigned int head;
 414         ssize_t ret = 0;
 415         size_t total_len = iov_iter_count(from);
 416         ssize_t chars;
 417         bool was_empty = false;
 418
 419         /* Null write succeeds. */
 420         if (unlikely(total_len == 0))
 421                 return 0;
 422
 423         __pipe_lock(pipe);
 424
 425         if (!pipe->readers) {
 426                 send_sig(SIGPIPE, current, 0);
 427                 ret = -EPIPE;
 428                 goto out;
 429         }
 430
 431         /*
 432          * Only wake up if the pipe started out empty, since
 433          * otherwise there should be no readers waiting.
 434          *
 435          * If it wasn't empty we try to merge new data into
 436          * the last buffer.
 437          *
 438          * That naturally merges small writes, but it also
 439          * page-aligs the rest of the writes for large writes
 440          * spanning multiple pages.
 441          */
 442         head = pipe->head;
 443         was_empty = pipe_empty(head, pipe->tail);
 444         chars = total_len & (PAGE_SIZE-1);
 445         if (chars && !was_empty) {
 446                 unsigned int mask = pipe->ring_size - 1;
 447                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 448                 int offset = buf->offset + buf->len;
 449
 450                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 451                         ret = pipe_buf_confirm(pipe, buf);
 452                         if (ret)
 453                                 goto out;
 454
 455                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 456                         if (unlikely(ret < chars)) {
 457                                 ret = -EFAULT;
 458                                 goto out;
 459                         }
 460
 461                         buf->len += ret;
 462                         if (!iov_iter_count(from))
 463                                 goto out;
 464                 }
 465         }
 466
 467         for (;;) {
 468                 if (!pipe->readers) {
 469                         send_sig(SIGPIPE, current, 0);
 470                         if (!ret)
 471                                 ret = -EPIPE;
 472                         break;
 473                 }
 474
 475                 head = pipe->head;
 476                 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 477                         unsigned int mask = pipe->ring_size - 1;
 478                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 479                         struct page *page = pipe->tmp_page;
 480                         int copied;
 481
 482                         if (!page) {
 483                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 484                                 if (unlikely(!page)) {
 485                                         ret = ret ? : -ENOMEM;
 486                                         break;
 487                                 }
 488                                 pipe->tmp_page = page;
 489                         }
 490
 491                         /* Allocate a slot in the ring in advance and attach an
 492                          * empty buffer.  If we fault or otherwise fail to use
 493                          * it, either the reader will consume it or it'll still
 494                          * be there for the next write.
 495                          */
 496                         spin_lock_irq(&pipe->wait.lock);
 497
 498                         head = pipe->head;
 499                         if (pipe_full(head, pipe->tail, pipe->max_usage)) {
 500                                 spin_unlock_irq(&pipe->wait.lock);
 501                                 continue;
 502                         }
 503
 504                         pipe->head = head + 1;
 505                         spin_unlock_irq(&pipe->wait.lock);
 506
 507                         /* Insert it into the buffer array */
 508                         buf = &pipe->bufs[head & mask];
 509                         buf->page = page;
 510                         buf->ops = &anon_pipe_buf_ops;
 511                         buf->offset = 0;
 512                         buf->len = 0;
 513                         buf->flags = 0;
 514                         if (is_packetized(filp)) {
 515                                 buf->ops = &packet_pipe_buf_ops;
 516                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 517                         }
 518                         pipe->tmp_page = NULL;
 519
 520                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 521                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 522                                 if (!ret)
 523                                         ret = -EFAULT;
 524                                 break;
 525                         }
 526                         ret += copied;
 527                         buf->offset = 0;
 528                         buf->len = copied;
 529
 530                         if (!iov_iter_count(from))
 531                                 break;
 532                 }
 533
 534                 if (!pipe_full(head, pipe->tail, pipe->max_usage))
 535                         continue;
 536
 537                 /* Wait for buffer space to become available. */
 538                 if (filp->f_flags & O_NONBLOCK) {
 539                         if (!ret)
 540                                 ret = -EAGAIN;
 541                         break;
 542                 }
 543                 if (signal_pending(current)) {
 544                         if (!ret)
 545                                 ret = -ERESTARTSYS;
 546                         break;
 547                 }
 548
 549                 /*
 550                  * We're going to release the pipe lock and wait for more
 551                  * space. We wake up any readers if necessary, and then
 552                  * after waiting we need to re-check whether the pipe
 553                  * become empty while we dropped the lock.
 554                  */
 555                 __pipe_unlock(pipe);
 556                 if (was_empty) {
 557                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 558                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 559                 }
 560                 wait_event_interruptible(pipe->wait, pipe_writable(pipe));
 561                 __pipe_lock(pipe);
 562                 was_empty = pipe_empty(head, pipe->tail);
 563         }
 564 out:
 565         __pipe_unlock(pipe);
 566
 567         /*
 568          * If we do do a wakeup event, we do a 'sync' wakeup, because we
 569          * want the reader to start processing things asap, rather than
 570          * leave the data pending.
 571          *
 572          * This is particularly important for small writes, because of
 573          * how (for example) the GNU make jobserver uses small writes to
 574          * wake up pending jobs
 575          */
 576         if (was_empty) {
 577                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 578                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 579         }
 580         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 581                 int err = file_update_time(filp);
 582                 if (err)
 583                         ret = err;
 584                 sb_end_write(file_inode(filp)->i_sb);
 585         }
 586         return ret;
 587 }
 588
 589 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 590 {
 591         struct pipe_inode_info *pipe = filp->private_data;
 592         int count, head, tail, mask;
 593
 594         switch (cmd) {
 595                 case FIONREAD:
 596                         __pipe_lock(pipe);
 597                         count = 0;
 598                         head = pipe->head;
 599                         tail = pipe->tail;
 600                         mask = pipe->ring_size - 1;
 601
 602                         while (tail != head) {
 603                                 count += pipe->bufs[tail & mask].len;
 604                                 tail++;
 605                         }
 606                         __pipe_unlock(pipe);
 607
 608                         return put_user(count, (int __user *)arg);
 609                 default:
 610                         return -ENOIOCTLCMD;
 611         }
 612 }
 613
 614 /* No kernel lock held - fine */
 615 static __poll_t
 616 pipe_poll(struct file *filp, poll_table *wait)
 617 {
 618         __poll_t mask;
 619         struct pipe_inode_info *pipe = filp->private_data;
 620         unsigned int head, tail;
 621
 622         /*
 623          * Reading only -- no need for acquiring the semaphore.
 624          *
 625          * But because this is racy, the code has to add the
 626          * entry to the poll table _first_ ..
 627          */
 628         poll_wait(filp, &pipe->wait, wait);
 629
 630         /*
 631          * .. and only then can you do the racy tests. That way,
 632          * if something changes and you got it wrong, the poll
 633          * table entry will wake you up and fix it.
 634          */
 635         head = READ_ONCE(pipe->head);
 636         tail = READ_ONCE(pipe->tail);
 637
 638         mask = 0;
 639         if (filp->f_mode & FMODE_READ) {
 640                 if (!pipe_empty(head, tail))
 641                         mask |= EPOLLIN | EPOLLRDNORM;
 642                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 643                         mask |= EPOLLHUP;
 644         }
 645
 646         if (filp->f_mode & FMODE_WRITE) {
 647                 if (!pipe_full(head, tail, pipe->max_usage))
 648                         mask |= EPOLLOUT | EPOLLWRNORM;
 649                 /*
 650                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 651                  * behave exactly like pipes for poll().
 652                  */
 653                 if (!pipe->readers)
 654                         mask |= EPOLLERR;
 655         }
 656
 657         return mask;
 658 }
 659
 660 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 661 {
 662         int kill = 0;
 663
 664         spin_lock(&inode->i_lock);
 665         if (!--pipe->files) {
 666                 inode->i_pipe = NULL;
 667                 kill = 1;
 668         }
 669         spin_unlock(&inode->i_lock);
 670
 671         if (kill)
 672                 free_pipe_info(pipe);
 673 }
 674
 675 static int
 676 pipe_release(struct inode *inode, struct file *file)
 677 {
 678         struct pipe_inode_info *pipe = file->private_data;
 679
 680         __pipe_lock(pipe);
 681         if (file->f_mode & FMODE_READ)
 682                 pipe->readers--;
 683         if (file->f_mode & FMODE_WRITE)
 684                 pipe->writers--;
 685
 686         if (pipe->readers || pipe->writers) {
 687                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 688                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 689                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 690         }
 691         __pipe_unlock(pipe);
 692
 693         put_pipe_info(inode, pipe);
 694         return 0;
 695 }
 696
 697 static int
 698 pipe_fasync(int fd, struct file *filp, int on)
 699 {
 700         struct pipe_inode_info *pipe = filp->private_data;
 701         int retval = 0;
 702
 703         __pipe_lock(pipe);
 704         if (filp->f_mode & FMODE_READ)
 705                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 706         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 707                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 708                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 709                         /* this can happen only if on == T */
 710                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 711         }
 712         __pipe_unlock(pipe);
 713         return retval;
 714 }
 715
 716 static unsigned long account_pipe_buffers(struct user_struct *user,
 717                                  unsigned long old, unsigned long new)
 718 {
 719         return atomic_long_add_return(new - old, &user->pipe_bufs);
 720 }
 721
 722 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 723 {
 724         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 725
 726         return soft_limit && user_bufs > soft_limit;
 727 }
 728
 729 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 730 {
 731         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 732
 733         return hard_limit && user_bufs > hard_limit;
 734 }
 735
 736 static bool is_unprivileged_user(void)
 737 {
 738         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 739 }
 740
 741 struct pipe_inode_info *alloc_pipe_info(void)
 742 {
 743         struct pipe_inode_info *pipe;
 744         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 745         struct user_struct *user = get_current_user();
 746         unsigned long user_bufs;
 747         unsigned int max_size = READ_ONCE(pipe_max_size);
 748
 749         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 750         if (pipe == NULL)
 751                 goto out_free_uid;
 752
 753         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 754                 pipe_bufs = max_size >> PAGE_SHIFT;
 755
 756         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 757
 758         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 759                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 760                 pipe_bufs = 1;
 761         }
 762
 763         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 764                 goto out_revert_acct;
 765
 766         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 767                              GFP_KERNEL_ACCOUNT);
 768
 769         if (pipe->bufs) {
 770                 init_waitqueue_head(&pipe->wait);
 771                 pipe->r_counter = pipe->w_counter = 1;
 772                 pipe->max_usage = pipe_bufs;
 773                 pipe->ring_size = pipe_bufs;
 774                 pipe->user = user;
 775                 mutex_init(&pipe->mutex);
 776                 return pipe;
 777         }
 778
 779 out_revert_acct:
 780         (void) account_pipe_buffers(user, pipe_bufs, 0);
 781         kfree(pipe);
 782 out_free_uid:
 783         free_uid(user);
 784         return NULL;
 785 }
 786
 787 void free_pipe_info(struct pipe_inode_info *pipe)
 788 {
 789         int i;
 790
 791         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 792         free_uid(pipe->user);
 793         for (i = 0; i < pipe->ring_size; i++) {
 794                 struct pipe_buffer *buf = pipe->bufs + i;
 795                 if (buf->ops)
 796                         pipe_buf_release(pipe, buf);
 797         }
 798         if (pipe->tmp_page)
 799                 __free_page(pipe->tmp_page);
 800         kfree(pipe->bufs);
 801         kfree(pipe);
 802 }
 803
 804 static struct vfsmount *pipe_mnt __read_mostly;
 805
 806 /*
 807  * pipefs_dname() is called from d_path().
 808  */
 809 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 810 {
 811         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 812                                 d_inode(dentry)->i_ino);
 813 }
 814
 815 static const struct dentry_operations pipefs_dentry_operations = {
 816         .d_dname        = pipefs_dname,
 817 };
 818
 819 static struct inode * get_pipe_inode(void)
 820 {
 821         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 822         struct pipe_inode_info *pipe;
 823
 824         if (!inode)
 825                 goto fail_inode;
 826
 827         inode->i_ino = get_next_ino();
 828
 829         pipe = alloc_pipe_info();
 830         if (!pipe)
 831                 goto fail_iput;
 832
 833         inode->i_pipe = pipe;
 834         pipe->files = 2;
 835         pipe->readers = pipe->writers = 1;
 836         inode->i_fop = &pipefifo_fops;
 837
 838         /*
 839          * Mark the inode dirty from the very beginning,
 840          * that way it will never be moved to the dirty
 841          * list because "mark_inode_dirty()" will think
 842          * that it already _is_ on the dirty list.
 843          */
 844         inode->i_state = I_DIRTY;
 845         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 846         inode->i_uid = current_fsuid();
 847         inode->i_gid = current_fsgid();
 848         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 849
 850         return inode;
 851
 852 fail_iput:
 853         iput(inode);
 854
 855 fail_inode:
 856         return NULL;
 857 }
 858
 859 int create_pipe_files(struct file **res, int flags)
 860 {
 861         struct inode *inode = get_pipe_inode();
 862         struct file *f;
 863
 864         if (!inode)
 865                 return -ENFILE;
 866
 867         f = alloc_file_pseudo(inode, pipe_mnt, "",
 868                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 869                                 &pipefifo_fops);
 870         if (IS_ERR(f)) {
 871                 free_pipe_info(inode->i_pipe);
 872                 iput(inode);
 873                 return PTR_ERR(f);
 874         }
 875
 876         f->private_data = inode->i_pipe;
 877
 878         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 879                                   &pipefifo_fops);
 880         if (IS_ERR(res[0])) {
 881                 put_pipe_info(inode, inode->i_pipe);
 882                 fput(f);
 883                 return PTR_ERR(res[0]);
 884         }
 885         res[0]->private_data = inode->i_pipe;
 886         res[1] = f;
 887         stream_open(inode, res[0]);
 888         stream_open(inode, res[1]);
 889         return 0;
 890 }
 891
 892 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 893 {
 894         int error;
 895         int fdw, fdr;
 896
 897         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 898                 return -EINVAL;
 899
 900         error = create_pipe_files(files, flags);
 901         if (error)
 902                 return error;
 903
 904         error = get_unused_fd_flags(flags);
 905         if (error < 0)
 906                 goto err_read_pipe;
 907         fdr = error;
 908
 909         error = get_unused_fd_flags(flags);
 910         if (error < 0)
 911                 goto err_fdr;
 912         fdw = error;
 913
 914         audit_fd_pair(fdr, fdw);
 915         fd[0] = fdr;
 916         fd[1] = fdw;
 917         return 0;
 918
 919  err_fdr:
 920         put_unused_fd(fdr);
 921  err_read_pipe:
 922         fput(files[0]);
 923         fput(files[1]);
 924         return error;
 925 }
 926
 927 int do_pipe_flags(int *fd, int flags)
 928 {
 929         struct file *files[2];
 930         int error = __do_pipe_flags(fd, files, flags);
 931         if (!error) {
 932                 fd_install(fd[0], files[0]);
 933                 fd_install(fd[1], files[1]);
 934         }
 935         return error;
 936 }
 937
 938 /*
 939  * sys_pipe() is the normal C calling standard for creating
 940  * a pipe. It's not the way Unix traditionally does this, though.
 941  */
 942 static int do_pipe2(int __user *fildes, int flags)
 943 {
 944         struct file *files[2];
 945         int fd[2];
 946         int error;
 947
 948         error = __do_pipe_flags(fd, files, flags);
 949         if (!error) {
 950                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 951                         fput(files[0]);
 952                         fput(files[1]);
 953                         put_unused_fd(fd[0]);
 954                         put_unused_fd(fd[1]);
 955                         error = -EFAULT;
 956                 } else {
 957                         fd_install(fd[0], files[0]);
 958                         fd_install(fd[1], files[1]);
 959                 }
 960         }
 961         return error;
 962 }
 963
 964 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 965 {
 966         return do_pipe2(fildes, flags);
 967 }
 968
 969 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 970 {
 971         return do_pipe2(fildes, 0);
 972 }
 973
 974 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 975 {
 976         int cur = *cnt;
 977
 978         while (cur == *cnt) {
 979                 pipe_wait(pipe);
 980                 if (signal_pending(current))
 981                         break;
 982         }
 983         return cur == *cnt ? -ERESTARTSYS : 0;
 984 }
 985
 986 static void wake_up_partner(struct pipe_inode_info *pipe)
 987 {
 988         wake_up_interruptible(&pipe->wait);
 989 }
 990
 991 static int fifo_open(struct inode *inode, struct file *filp)
 992 {
 993         struct pipe_inode_info *pipe;
 994         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 995         int ret;
 996
 997         filp->f_version = 0;
 998
 999         spin_lock(&inode->i_lock);
1000         if (inode->i_pipe) {
1001                 pipe = inode->i_pipe;
1002                 pipe->files++;
1003                 spin_unlock(&inode->i_lock);
1004         } else {
1005                 spin_unlock(&inode->i_lock);
1006                 pipe = alloc_pipe_info();
1007                 if (!pipe)
1008                         return -ENOMEM;
1009                 pipe->files = 1;
1010                 spin_lock(&inode->i_lock);
1011                 if (unlikely(inode->i_pipe)) {
1012                         inode->i_pipe->files++;
1013                         spin_unlock(&inode->i_lock);
1014                         free_pipe_info(pipe);
1015                         pipe = inode->i_pipe;
1016                 } else {
1017                         inode->i_pipe = pipe;
1018                         spin_unlock(&inode->i_lock);
1019                 }
1020         }
1021         filp->private_data = pipe;
1022         /* OK, we have a pipe and it's pinned down */
1023
1024         __pipe_lock(pipe);
1025
1026         /* We can only do regular read/write on fifos */
1027         stream_open(inode, filp);
1028
1029         switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1030         case FMODE_READ:
1031         /*
1032          *  O_RDONLY
1033          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1034          *  opened, even when there is no process writing the FIFO.
1035          */
1036                 pipe->r_counter++;
1037                 if (pipe->readers++ == 0)
1038                         wake_up_partner(pipe);
1039
1040                 if (!is_pipe && !pipe->writers) {
1041                         if ((filp->f_flags & O_NONBLOCK)) {
1042                                 /* suppress EPOLLHUP until we have
1043                                  * seen a writer */
1044                                 filp->f_version = pipe->w_counter;
1045                         } else {
1046                                 if (wait_for_partner(pipe, &pipe->w_counter))
1047                                         goto err_rd;
1048                         }
1049                 }
1050                 break;
1051
1052         case FMODE_WRITE:
1053         /*
1054          *  O_WRONLY
1055          *  POSIX.1 says that O_NONBLOCK means return -1 with
1056          *  errno=ENXIO when there is no process reading the FIFO.
1057          */
1058                 ret = -ENXIO;
1059                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1060                         goto err;
1061
1062                 pipe->w_counter++;
1063                 if (!pipe->writers++)
1064                         wake_up_partner(pipe);
1065
1066                 if (!is_pipe && !pipe->readers) {
1067                         if (wait_for_partner(pipe, &pipe->r_counter))
1068                                 goto err_wr;
1069                 }
1070                 break;
1071
1072         case FMODE_READ | FMODE_WRITE:
1073         /*
1074          *  O_RDWR
1075          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1076          *  This implementation will NEVER block on a O_RDWR open, since
1077          *  the process can at least talk to itself.
1078          */
1079
1080                 pipe->readers++;
1081                 pipe->writers++;
1082                 pipe->r_counter++;
1083                 pipe->w_counter++;
1084                 if (pipe->readers == 1 || pipe->writers == 1)
1085                         wake_up_partner(pipe);
1086                 break;
1087
1088         default:
1089                 ret = -EINVAL;
1090                 goto err;
1091         }
1092
1093         /* Ok! */
1094         __pipe_unlock(pipe);
1095         return 0;
1096
1097 err_rd:
1098         if (!--pipe->readers)
1099                 wake_up_interruptible(&pipe->wait);
1100         ret = -ERESTARTSYS;
1101         goto err;
1102
1103 err_wr:
1104         if (!--pipe->writers)
1105                 wake_up_interruptible(&pipe->wait);
1106         ret = -ERESTARTSYS;
1107         goto err;
1108
1109 err:
1110         __pipe_unlock(pipe);
1111
1112         put_pipe_info(inode, pipe);
1113         return ret;
1114 }
1115
1116 const struct file_operations pipefifo_fops = {
1117         .open           = fifo_open,
1118         .llseek         = no_llseek,
1119         .read_iter      = pipe_read,
1120         .write_iter     = pipe_write,
1121         .poll           = pipe_poll,
1122         .unlocked_ioctl = pipe_ioctl,
1123         .release        = pipe_release,
1124         .fasync         = pipe_fasync,
1125 };
1126
1127 /*
1128  * Currently we rely on the pipe array holding a power-of-2 number
1129  * of pages. Returns 0 on error.
1130  */
1131 unsigned int round_pipe_size(unsigned long size)
1132 {
1133         if (size > (1U << 31))
1134                 return 0;
1135
1136         /* Minimum pipe size, as required by POSIX */
1137         if (size < PAGE_SIZE)
1138                 return PAGE_SIZE;
1139
1140         return roundup_pow_of_two(size);
1141 }
1142
1143 /*
1144  * Allocate a new array of pipe buffers and copy the info over. Returns the
1145  * pipe size if successful, or return -ERROR on error.
1146  */
1147 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1148 {
1149         struct pipe_buffer *bufs;
1150         unsigned int size, nr_slots, head, tail, mask, n;
1151         unsigned long user_bufs;
1152         long ret = 0;
1153
1154         size = round_pipe_size(arg);
1155         nr_slots = size >> PAGE_SHIFT;
1156
1157         if (!nr_slots)
1158                 return -EINVAL;
1159
1160         /*
1161          * If trying to increase the pipe capacity, check that an
1162          * unprivileged user is not trying to exceed various limits
1163          * (soft limit check here, hard limit check just below).
1164          * Decreasing the pipe capacity is always permitted, even
1165          * if the user is currently over a limit.
1166          */
1167         if (nr_slots > pipe->ring_size &&
1168                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1169                 return -EPERM;
1170
1171         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1172
1173         if (nr_slots > pipe->ring_size &&
1174                         (too_many_pipe_buffers_hard(user_bufs) ||
1175                          too_many_pipe_buffers_soft(user_bufs)) &&
1176                         is_unprivileged_user()) {
1177                 ret = -EPERM;
1178                 goto out_revert_acct;
1179         }
1180
1181         /*
1182          * We can shrink the pipe, if arg is greater than the ring occupancy.
1183          * Since we don't expect a lot of shrink+grow operations, just free and
1184          * allocate again like we would do for growing.  If the pipe currently
1185          * contains more buffers than arg, then return busy.
1186          */
1187         mask = pipe->ring_size - 1;
1188         head = pipe->head;
1189         tail = pipe->tail;
1190         n = pipe_occupancy(pipe->head, pipe->tail);
1191         if (nr_slots < n) {
1192                 ret = -EBUSY;
1193                 goto out_revert_acct;
1194         }
1195
1196         bufs = kcalloc(nr_slots, sizeof(*bufs),
1197                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1198         if (unlikely(!bufs)) {
1199                 ret = -ENOMEM;
1200                 goto out_revert_acct;
1201         }
1202
1203         /*
1204          * The pipe array wraps around, so just start the new one at zero
1205          * and adjust the indices.
1206          */
1207         if (n > 0) {
1208                 unsigned int h = head & mask;
1209                 unsigned int t = tail & mask;
1210                 if (h > t) {
1211                         memcpy(bufs, pipe->bufs + t,
1212                                n * sizeof(struct pipe_buffer));
1213                 } else {
1214                         unsigned int tsize = pipe->ring_size - t;
1215                         if (h > 0)
1216                                 memcpy(bufs + tsize, pipe->bufs,
1217                                        h * sizeof(struct pipe_buffer));
1218                         memcpy(bufs, pipe->bufs + t,
1219                                tsize * sizeof(struct pipe_buffer));
1220                 }
1221         }
1222
1223         head = n;
1224         tail = 0;
1225
1226         kfree(pipe->bufs);
1227         pipe->bufs = bufs;
1228         pipe->ring_size = nr_slots;
1229         pipe->max_usage = nr_slots;
1230         pipe->tail = tail;
1231         pipe->head = head;
1232         wake_up_interruptible_all(&pipe->wait);
1233         return pipe->max_usage * PAGE_SIZE;
1234
1235 out_revert_acct:
1236         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1237         return ret;
1238 }
1239
1240 /*
1241  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1242  * location, so checking ->i_pipe is not enough to verify that this is a
1243  * pipe.
1244  */
1245 struct pipe_inode_info *get_pipe_info(struct file *file)
1246 {
1247         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1248 }
1249
1250 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1251 {
1252         struct pipe_inode_info *pipe;
1253         long ret;
1254
1255         pipe = get_pipe_info(file);
1256         if (!pipe)
1257                 return -EBADF;
1258
1259         __pipe_lock(pipe);
1260
1261         switch (cmd) {
1262         case F_SETPIPE_SZ:
1263                 ret = pipe_set_size(pipe, arg);
1264                 break;
1265         case F_GETPIPE_SZ:
1266                 ret = pipe->max_usage * PAGE_SIZE;
1267                 break;
1268         default:
1269                 ret = -EINVAL;
1270                 break;
1271         }
1272
1273         __pipe_unlock(pipe);
1274         return ret;
1275 }
1276
1277 static const struct super_operations pipefs_ops = {
1278         .destroy_inode = free_inode_nonrcu,
1279         .statfs = simple_statfs,
1280 };
1281
1282 /*
1283  * pipefs should _never_ be mounted by userland - too much of security hassle,
1284  * no real gain from having the whole whorehouse mounted. So we don't need
1285  * any operations on the root directory. However, we need a non-trivial
1286  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1287  */
1288
1289 static int pipefs_init_fs_context(struct fs_context *fc)
1290 {
1291         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1292         if (!ctx)
1293                 return -ENOMEM;
1294         ctx->ops = &pipefs_ops;
1295         ctx->dops = &pipefs_dentry_operations;
1296         return 0;
1297 }
1298
1299 static struct file_system_type pipe_fs_type = {
1300         .name           = "pipefs",
1301         .init_fs_context = pipefs_init_fs_context,
1302         .kill_sb        = kill_anon_super,
1303 };
1304
1305 static int __init init_pipe_fs(void)
1306 {
1307         int err = register_filesystem(&pipe_fs_type);
1308
1309         if (!err) {
1310                 pipe_mnt = kern_mount(&pipe_fs_type);
1311                 if (IS_ERR(pipe_mnt)) {
1312                         err = PTR_ERR(pipe_mnt);
1313                         unregister_filesystem(&pipe_fs_type);
1314                 }
1315         }
1316         return err;
1317 }
1318
1319 fs_initcall(init_pipe_fs);