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 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         int do_wakeup;
 280         ssize_t ret;
 281
 282         /* Null read succeeds. */
 283         if (unlikely(total_len == 0))
 284                 return 0;
 285
 286         do_wakeup = 0;
 287         ret = 0;
 288         __pipe_lock(pipe);
 289         for (;;) {
 290                 unsigned int head = pipe->head;
 291                 unsigned int tail = pipe->tail;
 292                 unsigned int mask = pipe->ring_size - 1;
 293
 294                 if (!pipe_empty(head, tail)) {
 295                         struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 296                         size_t chars = buf->len;
 297                         size_t written;
 298                         int error;
 299
 300                         if (chars > total_len)
 301                                 chars = total_len;
 302
 303                         error = pipe_buf_confirm(pipe, buf);
 304                         if (error) {
 305                                 if (!ret)
 306                                         ret = error;
 307                                 break;
 308                         }
 309
 310                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 311                         if (unlikely(written < chars)) {
 312                                 if (!ret)
 313                                         ret = -EFAULT;
 314                                 break;
 315                         }
 316                         ret += chars;
 317                         buf->offset += chars;
 318                         buf->len -= chars;
 319
 320                         /* Was it a packet buffer? Clean up and exit */
 321                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 322                                 total_len = chars;
 323                                 buf->len = 0;
 324                         }
 325
 326                         if (!buf->len) {
 327                                 bool wake;
 328                                 pipe_buf_release(pipe, buf);
 329                                 spin_lock_irq(&pipe->wait.lock);
 330                                 tail++;
 331                                 pipe->tail = tail;
 332                                 do_wakeup = 1;
 333                                 wake = head - (tail - 1) == pipe->max_usage / 2;
 334                                 if (wake)
 335                                         wake_up_interruptible_sync_poll_locked(
 336                                                 &pipe->wait, EPOLLOUT | EPOLLWRNORM);
 337                                 spin_unlock_irq(&pipe->wait.lock);
 338                                 if (wake)
 339                                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 340                         }
 341                         total_len -= chars;
 342                         if (!total_len)
 343                                 break;  /* common path: read succeeded */
 344                         if (!pipe_empty(head, tail))    /* More to do? */
 345                                 continue;
 346                 }
 347
 348                 if (!pipe->writers)
 349                         break;
 350                 if (!pipe->waiting_writers) {
 351                         /* syscall merging: Usually we must not sleep
 352                          * if O_NONBLOCK is set, or if we got some data.
 353                          * But if a writer sleeps in kernel space, then
 354                          * we can wait for that data without violating POSIX.
 355                          */
 356                         if (ret)
 357                                 break;
 358                         if (filp->f_flags & O_NONBLOCK) {
 359                                 ret = -EAGAIN;
 360                                 break;
 361                         }
 362                 }
 363                 if (signal_pending(current)) {
 364                         if (!ret)
 365                                 ret = -ERESTARTSYS;
 366                         break;
 367                 }
 368                 pipe_wait(pipe);
 369         }
 370         __pipe_unlock(pipe);
 371
 372         /* Signal writers asynchronously that there is more room. */
 373         if (do_wakeup) {
 374                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 375                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 376         }
 377         if (ret > 0)
 378                 file_accessed(filp);
 379         return ret;
 380 }
 381
 382 static inline int is_packetized(struct file *file)
 383 {
 384         return (file->f_flags & O_DIRECT) != 0;
 385 }
 386
 387 static ssize_t
 388 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 389 {
 390         struct file *filp = iocb->ki_filp;
 391         struct pipe_inode_info *pipe = filp->private_data;
 392         unsigned int head, max_usage, mask;
 393         ssize_t ret = 0;
 394         int do_wakeup = 0;
 395         size_t total_len = iov_iter_count(from);
 396         ssize_t chars;
 397
 398         /* Null write succeeds. */
 399         if (unlikely(total_len == 0))
 400                 return 0;
 401
 402         __pipe_lock(pipe);
 403
 404         if (!pipe->readers) {
 405                 send_sig(SIGPIPE, current, 0);
 406                 ret = -EPIPE;
 407                 goto out;
 408         }
 409
 410         head = pipe->head;
 411         max_usage = pipe->max_usage;
 412         mask = pipe->ring_size - 1;
 413
 414         /* We try to merge small writes */
 415         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 416         if (!pipe_empty(head, pipe->tail) && chars != 0) {
 417                 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 418                 int offset = buf->offset + buf->len;
 419
 420                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 421                         ret = pipe_buf_confirm(pipe, buf);
 422                         if (ret)
 423                                 goto out;
 424
 425                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 426                         if (unlikely(ret < chars)) {
 427                                 ret = -EFAULT;
 428                                 goto out;
 429                         }
 430                         do_wakeup = 1;
 431                         buf->len += ret;
 432                         if (!iov_iter_count(from))
 433                                 goto out;
 434                 }
 435         }
 436
 437         for (;;) {
 438                 if (!pipe->readers) {
 439                         send_sig(SIGPIPE, current, 0);
 440                         if (!ret)
 441                                 ret = -EPIPE;
 442                         break;
 443                 }
 444
 445                 head = pipe->head;
 446                 if (!pipe_full(head, pipe->tail, max_usage)) {
 447                         struct pipe_buffer *buf = &pipe->bufs[head & mask];
 448                         struct page *page = pipe->tmp_page;
 449                         int copied;
 450
 451                         if (!page) {
 452                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 453                                 if (unlikely(!page)) {
 454                                         ret = ret ? : -ENOMEM;
 455                                         break;
 456                                 }
 457                                 pipe->tmp_page = page;
 458                         }
 459
 460                         /* Allocate a slot in the ring in advance and attach an
 461                          * empty buffer.  If we fault or otherwise fail to use
 462                          * it, either the reader will consume it or it'll still
 463                          * be there for the next write.
 464                          */
 465                         spin_lock_irq(&pipe->wait.lock);
 466
 467                         head = pipe->head;
 468                         if (pipe_full(head, pipe->tail, max_usage)) {
 469                                 spin_unlock_irq(&pipe->wait.lock);
 470                                 continue;
 471                         }
 472
 473                         pipe->head = head + 1;
 474
 475                         /* Always wake up, even if the copy fails. Otherwise
 476                          * we lock up (O_NONBLOCK-)readers that sleep due to
 477                          * syscall merging.
 478                          * FIXME! Is this really true?
 479                          */
 480                         wake_up_interruptible_sync_poll_locked(
 481                                 &pipe->wait, EPOLLIN | EPOLLRDNORM);
 482
 483                         spin_unlock_irq(&pipe->wait.lock);
 484                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 485
 486                         /* Insert it into the buffer array */
 487                         buf = &pipe->bufs[head & mask];
 488                         buf->page = page;
 489                         buf->ops = &anon_pipe_buf_ops;
 490                         buf->offset = 0;
 491                         buf->len = 0;
 492                         buf->flags = 0;
 493                         if (is_packetized(filp)) {
 494                                 buf->ops = &packet_pipe_buf_ops;
 495                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 496                         }
 497                         pipe->tmp_page = NULL;
 498
 499                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 500                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 501                                 if (!ret)
 502                                         ret = -EFAULT;
 503                                 break;
 504                         }
 505                         ret += copied;
 506                         buf->offset = 0;
 507                         buf->len = copied;
 508
 509                         if (!iov_iter_count(from))
 510                                 break;
 511                 }
 512
 513                 if (!pipe_full(head, pipe->tail, max_usage))
 514                         continue;
 515
 516                 /* Wait for buffer space to become available. */
 517                 if (filp->f_flags & O_NONBLOCK) {
 518                         if (!ret)
 519                                 ret = -EAGAIN;
 520                         break;
 521                 }
 522                 if (signal_pending(current)) {
 523                         if (!ret)
 524                                 ret = -ERESTARTSYS;
 525                         break;
 526                 }
 527                 pipe->waiting_writers++;
 528                 pipe_wait(pipe);
 529                 pipe->waiting_writers--;
 530         }
 531 out:
 532         __pipe_unlock(pipe);
 533         if (do_wakeup) {
 534                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 535                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 536         }
 537         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 538                 int err = file_update_time(filp);
 539                 if (err)
 540                         ret = err;
 541                 sb_end_write(file_inode(filp)->i_sb);
 542         }
 543         return ret;
 544 }
 545
 546 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 547 {
 548         struct pipe_inode_info *pipe = filp->private_data;
 549         int count, head, tail, mask;
 550
 551         switch (cmd) {
 552                 case FIONREAD:
 553                         __pipe_lock(pipe);
 554                         count = 0;
 555                         head = pipe->head;
 556                         tail = pipe->tail;
 557                         mask = pipe->ring_size - 1;
 558
 559                         while (tail != head) {
 560                                 count += pipe->bufs[tail & mask].len;
 561                                 tail++;
 562                         }
 563                         __pipe_unlock(pipe);
 564
 565                         return put_user(count, (int __user *)arg);
 566                 default:
 567                         return -ENOIOCTLCMD;
 568         }
 569 }
 570
 571 /* No kernel lock held - fine */
 572 static __poll_t
 573 pipe_poll(struct file *filp, poll_table *wait)
 574 {
 575         __poll_t mask;
 576         struct pipe_inode_info *pipe = filp->private_data;
 577         unsigned int head = READ_ONCE(pipe->head);
 578         unsigned int tail = READ_ONCE(pipe->tail);
 579
 580         poll_wait(filp, &pipe->wait, wait);
 581
 582         BUG_ON(pipe_occupancy(head, tail) > pipe->ring_size);
 583
 584         /* Reading only -- no need for acquiring the semaphore.  */
 585         mask = 0;
 586         if (filp->f_mode & FMODE_READ) {
 587                 if (!pipe_empty(head, tail))
 588                         mask |= EPOLLIN | EPOLLRDNORM;
 589                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 590                         mask |= EPOLLHUP;
 591         }
 592
 593         if (filp->f_mode & FMODE_WRITE) {
 594                 if (!pipe_full(head, tail, pipe->max_usage))
 595                         mask |= EPOLLOUT | EPOLLWRNORM;
 596                 /*
 597                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 598                  * behave exactly like pipes for poll().
 599                  */
 600                 if (!pipe->readers)
 601                         mask |= EPOLLERR;
 602         }
 603
 604         return mask;
 605 }
 606
 607 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 608 {
 609         int kill = 0;
 610
 611         spin_lock(&inode->i_lock);
 612         if (!--pipe->files) {
 613                 inode->i_pipe = NULL;
 614                 kill = 1;
 615         }
 616         spin_unlock(&inode->i_lock);
 617
 618         if (kill)
 619                 free_pipe_info(pipe);
 620 }
 621
 622 static int
 623 pipe_release(struct inode *inode, struct file *file)
 624 {
 625         struct pipe_inode_info *pipe = file->private_data;
 626
 627         __pipe_lock(pipe);
 628         if (file->f_mode & FMODE_READ)
 629                 pipe->readers--;
 630         if (file->f_mode & FMODE_WRITE)
 631                 pipe->writers--;
 632
 633         if (pipe->readers || pipe->writers) {
 634                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 635                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 636                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 637         }
 638         __pipe_unlock(pipe);
 639
 640         put_pipe_info(inode, pipe);
 641         return 0;
 642 }
 643
 644 static int
 645 pipe_fasync(int fd, struct file *filp, int on)
 646 {
 647         struct pipe_inode_info *pipe = filp->private_data;
 648         int retval = 0;
 649
 650         __pipe_lock(pipe);
 651         if (filp->f_mode & FMODE_READ)
 652                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 653         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 654                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 655                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 656                         /* this can happen only if on == T */
 657                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 658         }
 659         __pipe_unlock(pipe);
 660         return retval;
 661 }
 662
 663 static unsigned long account_pipe_buffers(struct user_struct *user,
 664                                  unsigned long old, unsigned long new)
 665 {
 666         return atomic_long_add_return(new - old, &user->pipe_bufs);
 667 }
 668
 669 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 670 {
 671         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 672
 673         return soft_limit && user_bufs > soft_limit;
 674 }
 675
 676 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 677 {
 678         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 679
 680         return hard_limit && user_bufs > hard_limit;
 681 }
 682
 683 static bool is_unprivileged_user(void)
 684 {
 685         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 686 }
 687
 688 struct pipe_inode_info *alloc_pipe_info(void)
 689 {
 690         struct pipe_inode_info *pipe;
 691         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 692         struct user_struct *user = get_current_user();
 693         unsigned long user_bufs;
 694         unsigned int max_size = READ_ONCE(pipe_max_size);
 695
 696         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 697         if (pipe == NULL)
 698                 goto out_free_uid;
 699
 700         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 701                 pipe_bufs = max_size >> PAGE_SHIFT;
 702
 703         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 704
 705         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 706                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 707                 pipe_bufs = 1;
 708         }
 709
 710         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 711                 goto out_revert_acct;
 712
 713         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 714                              GFP_KERNEL_ACCOUNT);
 715
 716         if (pipe->bufs) {
 717                 init_waitqueue_head(&pipe->wait);
 718                 pipe->r_counter = pipe->w_counter = 1;
 719                 pipe->max_usage = pipe_bufs;
 720                 pipe->ring_size = pipe_bufs;
 721                 pipe->user = user;
 722                 mutex_init(&pipe->mutex);
 723                 return pipe;
 724         }
 725
 726 out_revert_acct:
 727         (void) account_pipe_buffers(user, pipe_bufs, 0);
 728         kfree(pipe);
 729 out_free_uid:
 730         free_uid(user);
 731         return NULL;
 732 }
 733
 734 void free_pipe_info(struct pipe_inode_info *pipe)
 735 {
 736         int i;
 737
 738         (void) account_pipe_buffers(pipe->user, pipe->ring_size, 0);
 739         free_uid(pipe->user);
 740         for (i = 0; i < pipe->ring_size; i++) {
 741                 struct pipe_buffer *buf = pipe->bufs + i;
 742                 if (buf->ops)
 743                         pipe_buf_release(pipe, buf);
 744         }
 745         if (pipe->tmp_page)
 746                 __free_page(pipe->tmp_page);
 747         kfree(pipe->bufs);
 748         kfree(pipe);
 749 }
 750
 751 static struct vfsmount *pipe_mnt __read_mostly;
 752
 753 /*
 754  * pipefs_dname() is called from d_path().
 755  */
 756 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 757 {
 758         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 759                                 d_inode(dentry)->i_ino);
 760 }
 761
 762 static const struct dentry_operations pipefs_dentry_operations = {
 763         .d_dname        = pipefs_dname,
 764 };
 765
 766 static struct inode * get_pipe_inode(void)
 767 {
 768         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 769         struct pipe_inode_info *pipe;
 770
 771         if (!inode)
 772                 goto fail_inode;
 773
 774         inode->i_ino = get_next_ino();
 775
 776         pipe = alloc_pipe_info();
 777         if (!pipe)
 778                 goto fail_iput;
 779
 780         inode->i_pipe = pipe;
 781         pipe->files = 2;
 782         pipe->readers = pipe->writers = 1;
 783         inode->i_fop = &pipefifo_fops;
 784
 785         /*
 786          * Mark the inode dirty from the very beginning,
 787          * that way it will never be moved to the dirty
 788          * list because "mark_inode_dirty()" will think
 789          * that it already _is_ on the dirty list.
 790          */
 791         inode->i_state = I_DIRTY;
 792         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 793         inode->i_uid = current_fsuid();
 794         inode->i_gid = current_fsgid();
 795         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 796
 797         return inode;
 798
 799 fail_iput:
 800         iput(inode);
 801
 802 fail_inode:
 803         return NULL;
 804 }
 805
 806 int create_pipe_files(struct file **res, int flags)
 807 {
 808         struct inode *inode = get_pipe_inode();
 809         struct file *f;
 810
 811         if (!inode)
 812                 return -ENFILE;
 813
 814         f = alloc_file_pseudo(inode, pipe_mnt, "",
 815                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 816                                 &pipefifo_fops);
 817         if (IS_ERR(f)) {
 818                 free_pipe_info(inode->i_pipe);
 819                 iput(inode);
 820                 return PTR_ERR(f);
 821         }
 822
 823         f->private_data = inode->i_pipe;
 824
 825         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 826                                   &pipefifo_fops);
 827         if (IS_ERR(res[0])) {
 828                 put_pipe_info(inode, inode->i_pipe);
 829                 fput(f);
 830                 return PTR_ERR(res[0]);
 831         }
 832         res[0]->private_data = inode->i_pipe;
 833         res[1] = f;
 834         return 0;
 835 }
 836
 837 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 838 {
 839         int error;
 840         int fdw, fdr;
 841
 842         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 843                 return -EINVAL;
 844
 845         error = create_pipe_files(files, flags);
 846         if (error)
 847                 return error;
 848
 849         error = get_unused_fd_flags(flags);
 850         if (error < 0)
 851                 goto err_read_pipe;
 852         fdr = error;
 853
 854         error = get_unused_fd_flags(flags);
 855         if (error < 0)
 856                 goto err_fdr;
 857         fdw = error;
 858
 859         audit_fd_pair(fdr, fdw);
 860         fd[0] = fdr;
 861         fd[1] = fdw;
 862         return 0;
 863
 864  err_fdr:
 865         put_unused_fd(fdr);
 866  err_read_pipe:
 867         fput(files[0]);
 868         fput(files[1]);
 869         return error;
 870 }
 871
 872 int do_pipe_flags(int *fd, int flags)
 873 {
 874         struct file *files[2];
 875         int error = __do_pipe_flags(fd, files, flags);
 876         if (!error) {
 877                 fd_install(fd[0], files[0]);
 878                 fd_install(fd[1], files[1]);
 879         }
 880         return error;
 881 }
 882
 883 /*
 884  * sys_pipe() is the normal C calling standard for creating
 885  * a pipe. It's not the way Unix traditionally does this, though.
 886  */
 887 static int do_pipe2(int __user *fildes, int flags)
 888 {
 889         struct file *files[2];
 890         int fd[2];
 891         int error;
 892
 893         error = __do_pipe_flags(fd, files, flags);
 894         if (!error) {
 895                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 896                         fput(files[0]);
 897                         fput(files[1]);
 898                         put_unused_fd(fd[0]);
 899                         put_unused_fd(fd[1]);
 900                         error = -EFAULT;
 901                 } else {
 902                         fd_install(fd[0], files[0]);
 903                         fd_install(fd[1], files[1]);
 904                 }
 905         }
 906         return error;
 907 }
 908
 909 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 910 {
 911         return do_pipe2(fildes, flags);
 912 }
 913
 914 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 915 {
 916         return do_pipe2(fildes, 0);
 917 }
 918
 919 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 920 {
 921         int cur = *cnt;
 922
 923         while (cur == *cnt) {
 924                 pipe_wait(pipe);
 925                 if (signal_pending(current))
 926                         break;
 927         }
 928         return cur == *cnt ? -ERESTARTSYS : 0;
 929 }
 930
 931 static void wake_up_partner(struct pipe_inode_info *pipe)
 932 {
 933         wake_up_interruptible(&pipe->wait);
 934 }
 935
 936 static int fifo_open(struct inode *inode, struct file *filp)
 937 {
 938         struct pipe_inode_info *pipe;
 939         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 940         int ret;
 941
 942         filp->f_version = 0;
 943
 944         spin_lock(&inode->i_lock);
 945         if (inode->i_pipe) {
 946                 pipe = inode->i_pipe;
 947                 pipe->files++;
 948                 spin_unlock(&inode->i_lock);
 949         } else {
 950                 spin_unlock(&inode->i_lock);
 951                 pipe = alloc_pipe_info();
 952                 if (!pipe)
 953                         return -ENOMEM;
 954                 pipe->files = 1;
 955                 spin_lock(&inode->i_lock);
 956                 if (unlikely(inode->i_pipe)) {
 957                         inode->i_pipe->files++;
 958                         spin_unlock(&inode->i_lock);
 959                         free_pipe_info(pipe);
 960                         pipe = inode->i_pipe;
 961                 } else {
 962                         inode->i_pipe = pipe;
 963                         spin_unlock(&inode->i_lock);
 964                 }
 965         }
 966         filp->private_data = pipe;
 967         /* OK, we have a pipe and it's pinned down */
 968
 969         __pipe_lock(pipe);
 970
 971         /* We can only do regular read/write on fifos */
 972         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 973
 974         switch (filp->f_mode) {
 975         case FMODE_READ:
 976         /*
 977          *  O_RDONLY
 978          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 979          *  opened, even when there is no process writing the FIFO.
 980          */
 981                 pipe->r_counter++;
 982                 if (pipe->readers++ == 0)
 983                         wake_up_partner(pipe);
 984
 985                 if (!is_pipe && !pipe->writers) {
 986                         if ((filp->f_flags & O_NONBLOCK)) {
 987                                 /* suppress EPOLLHUP until we have
 988                                  * seen a writer */
 989                                 filp->f_version = pipe->w_counter;
 990                         } else {
 991                                 if (wait_for_partner(pipe, &pipe->w_counter))
 992                                         goto err_rd;
 993                         }
 994                 }
 995                 break;
 996
 997         case FMODE_WRITE:
 998         /*
 999          *  O_WRONLY
1000          *  POSIX.1 says that O_NONBLOCK means return -1 with
1001          *  errno=ENXIO when there is no process reading the FIFO.
1002          */
1003                 ret = -ENXIO;
1004                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1005                         goto err;
1006
1007                 pipe->w_counter++;
1008                 if (!pipe->writers++)
1009                         wake_up_partner(pipe);
1010
1011                 if (!is_pipe && !pipe->readers) {
1012                         if (wait_for_partner(pipe, &pipe->r_counter))
1013                                 goto err_wr;
1014                 }
1015                 break;
1016
1017         case FMODE_READ | FMODE_WRITE:
1018         /*
1019          *  O_RDWR
1020          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1021          *  This implementation will NEVER block on a O_RDWR open, since
1022          *  the process can at least talk to itself.
1023          */
1024
1025                 pipe->readers++;
1026                 pipe->writers++;
1027                 pipe->r_counter++;
1028                 pipe->w_counter++;
1029                 if (pipe->readers == 1 || pipe->writers == 1)
1030                         wake_up_partner(pipe);
1031                 break;
1032
1033         default:
1034                 ret = -EINVAL;
1035                 goto err;
1036         }
1037
1038         /* Ok! */
1039         __pipe_unlock(pipe);
1040         return 0;
1041
1042 err_rd:
1043         if (!--pipe->readers)
1044                 wake_up_interruptible(&pipe->wait);
1045         ret = -ERESTARTSYS;
1046         goto err;
1047
1048 err_wr:
1049         if (!--pipe->writers)
1050                 wake_up_interruptible(&pipe->wait);
1051         ret = -ERESTARTSYS;
1052         goto err;
1053
1054 err:
1055         __pipe_unlock(pipe);
1056
1057         put_pipe_info(inode, pipe);
1058         return ret;
1059 }
1060
1061 const struct file_operations pipefifo_fops = {
1062         .open           = fifo_open,
1063         .llseek         = no_llseek,
1064         .read_iter      = pipe_read,
1065         .write_iter     = pipe_write,
1066         .poll           = pipe_poll,
1067         .unlocked_ioctl = pipe_ioctl,
1068         .release        = pipe_release,
1069         .fasync         = pipe_fasync,
1070 };
1071
1072 /*
1073  * Currently we rely on the pipe array holding a power-of-2 number
1074  * of pages. Returns 0 on error.
1075  */
1076 unsigned int round_pipe_size(unsigned long size)
1077 {
1078         if (size > (1U << 31))
1079                 return 0;
1080
1081         /* Minimum pipe size, as required by POSIX */
1082         if (size < PAGE_SIZE)
1083                 return PAGE_SIZE;
1084
1085         return roundup_pow_of_two(size);
1086 }
1087
1088 /*
1089  * Allocate a new array of pipe buffers and copy the info over. Returns the
1090  * pipe size if successful, or return -ERROR on error.
1091  */
1092 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1093 {
1094         struct pipe_buffer *bufs;
1095         unsigned int size, nr_slots, head, tail, mask, n;
1096         unsigned long user_bufs;
1097         long ret = 0;
1098
1099         size = round_pipe_size(arg);
1100         nr_slots = size >> PAGE_SHIFT;
1101
1102         if (!nr_slots)
1103                 return -EINVAL;
1104
1105         /*
1106          * If trying to increase the pipe capacity, check that an
1107          * unprivileged user is not trying to exceed various limits
1108          * (soft limit check here, hard limit check just below).
1109          * Decreasing the pipe capacity is always permitted, even
1110          * if the user is currently over a limit.
1111          */
1112         if (nr_slots > pipe->ring_size &&
1113                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1114                 return -EPERM;
1115
1116         user_bufs = account_pipe_buffers(pipe->user, pipe->ring_size, nr_slots);
1117
1118         if (nr_slots > pipe->ring_size &&
1119                         (too_many_pipe_buffers_hard(user_bufs) ||
1120                          too_many_pipe_buffers_soft(user_bufs)) &&
1121                         is_unprivileged_user()) {
1122                 ret = -EPERM;
1123                 goto out_revert_acct;
1124         }
1125
1126         /*
1127          * We can shrink the pipe, if arg is greater than the ring occupancy.
1128          * Since we don't expect a lot of shrink+grow operations, just free and
1129          * allocate again like we would do for growing.  If the pipe currently
1130          * contains more buffers than arg, then return busy.
1131          */
1132         mask = pipe->ring_size - 1;
1133         head = pipe->head;
1134         tail = pipe->tail;
1135         n = pipe_occupancy(pipe->head, pipe->tail);
1136         if (nr_slots < n) {
1137                 ret = -EBUSY;
1138                 goto out_revert_acct;
1139         }
1140
1141         bufs = kcalloc(nr_slots, sizeof(*bufs),
1142                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1143         if (unlikely(!bufs)) {
1144                 ret = -ENOMEM;
1145                 goto out_revert_acct;
1146         }
1147
1148         /*
1149          * The pipe array wraps around, so just start the new one at zero
1150          * and adjust the indices.
1151          */
1152         if (n > 0) {
1153                 unsigned int h = head & mask;
1154                 unsigned int t = tail & mask;
1155                 if (h > t) {
1156                         memcpy(bufs, pipe->bufs + t,
1157                                n * sizeof(struct pipe_buffer));
1158                 } else {
1159                         unsigned int tsize = pipe->ring_size - t;
1160                         if (h > 0)
1161                                 memcpy(bufs + tsize, pipe->bufs,
1162                                        h * sizeof(struct pipe_buffer));
1163                         memcpy(bufs, pipe->bufs + t,
1164                                tsize * sizeof(struct pipe_buffer));
1165                 }
1166         }
1167
1168         head = n;
1169         tail = 0;
1170
1171         kfree(pipe->bufs);
1172         pipe->bufs = bufs;
1173         pipe->ring_size = nr_slots;
1174         pipe->max_usage = nr_slots;
1175         pipe->tail = tail;
1176         pipe->head = head;
1177         return pipe->max_usage * PAGE_SIZE;
1178
1179 out_revert_acct:
1180         (void) account_pipe_buffers(pipe->user, nr_slots, pipe->ring_size);
1181         return ret;
1182 }
1183
1184 /*
1185  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1186  * location, so checking ->i_pipe is not enough to verify that this is a
1187  * pipe.
1188  */
1189 struct pipe_inode_info *get_pipe_info(struct file *file)
1190 {
1191         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1192 }
1193
1194 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1195 {
1196         struct pipe_inode_info *pipe;
1197         long ret;
1198
1199         pipe = get_pipe_info(file);
1200         if (!pipe)
1201                 return -EBADF;
1202
1203         __pipe_lock(pipe);
1204
1205         switch (cmd) {
1206         case F_SETPIPE_SZ:
1207                 ret = pipe_set_size(pipe, arg);
1208                 break;
1209         case F_GETPIPE_SZ:
1210                 ret = pipe->max_usage * PAGE_SIZE;
1211                 break;
1212         default:
1213                 ret = -EINVAL;
1214                 break;
1215         }
1216
1217         __pipe_unlock(pipe);
1218         return ret;
1219 }
1220
1221 static const struct super_operations pipefs_ops = {
1222         .destroy_inode = free_inode_nonrcu,
1223         .statfs = simple_statfs,
1224 };
1225
1226 /*
1227  * pipefs should _never_ be mounted by userland - too much of security hassle,
1228  * no real gain from having the whole whorehouse mounted. So we don't need
1229  * any operations on the root directory. However, we need a non-trivial
1230  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1231  */
1232
1233 static int pipefs_init_fs_context(struct fs_context *fc)
1234 {
1235         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1236         if (!ctx)
1237                 return -ENOMEM;
1238         ctx->ops = &pipefs_ops;
1239         ctx->dops = &pipefs_dentry_operations;
1240         return 0;
1241 }
1242
1243 static struct file_system_type pipe_fs_type = {
1244         .name           = "pipefs",
1245         .init_fs_context = pipefs_init_fs_context,
1246         .kill_sb        = kill_anon_super,
1247 };
1248
1249 static int __init init_pipe_fs(void)
1250 {
1251         int err = register_filesystem(&pipe_fs_type);
1252
1253         if (!err) {
1254                 pipe_mnt = kern_mount(&pipe_fs_type);
1255                 if (IS_ERR(pipe_mnt)) {
1256                         err = PTR_ERR(pipe_mnt);
1257                         unregister_filesystem(&pipe_fs_type);
1258                 }
1259         }
1260         return err;
1261 }
1262
1263 fs_initcall(init_pipe_fs);