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