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