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/magic.h>
  18 #include <linux/pipe_fs_i.h>
  19 #include <linux/uio.h>
  20 #include <linux/highmem.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/audit.h>
  23 #include <linux/syscalls.h>
  24 #include <linux/fcntl.h>
  25 #include <linux/memcontrol.h>
  26
  27 #include <linux/uaccess.h>
  28 #include <asm/ioctls.h>
  29
  30 #include "internal.h"
  31
  32 /*
  33  * The max size that a non-root user is allowed to grow the pipe. Can
  34  * be set by root in /proc/sys/fs/pipe-max-size
  35  */
  36 unsigned int pipe_max_size = 1048576;
  37
  38 /*
  39  * Minimum pipe size, as required by POSIX
  40  */
  41 unsigned int pipe_min_size = PAGE_SIZE;
  42
  43 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  44  * matches default values.
  45  */
  46 unsigned long pipe_user_pages_hard;
  47 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  48
  49 /*
  50  * We use a start+len construction, which provides full use of the
  51  * allocated memory.
  52  * -- Florian Coosmann (FGC)
  53  *
  54  * Reads with count = 0 should always return 0.
  55  * -- Julian Bradfield 1999-06-07.
  56  *
  57  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  58  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  59  *
  60  * pipe_read & write cleanup
  61  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  62  */
  63
  64 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  65 {
  66         if (pipe->files)
  67                 mutex_lock_nested(&pipe->mutex, subclass);
  68 }
  69
  70 void pipe_lock(struct pipe_inode_info *pipe)
  71 {
  72         /*
  73          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  74          */
  75         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  76 }
  77 EXPORT_SYMBOL(pipe_lock);
  78
  79 void pipe_unlock(struct pipe_inode_info *pipe)
  80 {
  81         if (pipe->files)
  82                 mutex_unlock(&pipe->mutex);
  83 }
  84 EXPORT_SYMBOL(pipe_unlock);
  85
  86 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  87 {
  88         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  89 }
  90
  91 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  92 {
  93         mutex_unlock(&pipe->mutex);
  94 }
  95
  96 void pipe_double_lock(struct pipe_inode_info *pipe1,
  97                       struct pipe_inode_info *pipe2)
  98 {
  99         BUG_ON(pipe1 == pipe2);
 100
 101         if (pipe1 < pipe2) {
 102                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 103                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 104         } else {
 105                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 106                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 107         }
 108 }
 109
 110 /* Drop the inode semaphore and wait for a pipe event, atomically */
 111 void pipe_wait(struct pipe_inode_info *pipe)
 112 {
 113         DEFINE_WAIT(wait);
 114
 115         /*
 116          * Pipes are system-local resources, so sleeping on them
 117          * is considered a noninteractive wait:
 118          */
 119         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 120         pipe_unlock(pipe);
 121         schedule();
 122         finish_wait(&pipe->wait, &wait);
 123         pipe_lock(pipe);
 124 }
 125
 126 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 127                                   struct pipe_buffer *buf)
 128 {
 129         struct page *page = buf->page;
 130
 131         /*
 132          * If nobody else uses this page, and we don't already have a
 133          * temporary page, let's keep track of it as a one-deep
 134          * allocation cache. (Otherwise just release our reference to it)
 135          */
 136         if (page_count(page) == 1 && !pipe->tmp_page)
 137                 pipe->tmp_page = page;
 138         else
 139                 put_page(page);
 140 }
 141
 142 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 143                                struct pipe_buffer *buf)
 144 {
 145         struct page *page = buf->page;
 146
 147         if (page_count(page) == 1) {
 148                 if (memcg_kmem_enabled())
 149                         memcg_kmem_uncharge(page, 0);
 150                 __SetPageLocked(page);
 151                 return 0;
 152         }
 153         return 1;
 154 }
 155
 156 /**
 157  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 158  * @pipe:       the pipe that the buffer belongs to
 159  * @buf:        the buffer to attempt to steal
 160  *
 161  * Description:
 162  *      This function attempts to steal the &struct page attached to
 163  *      @buf. If successful, this function returns 0 and returns with
 164  *      the page locked. The caller may then reuse the page for whatever
 165  *      he wishes; the typical use is insertion into a different file
 166  *      page cache.
 167  */
 168 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 169                            struct pipe_buffer *buf)
 170 {
 171         struct page *page = buf->page;
 172
 173         /*
 174          * A reference of one is golden, that means that the owner of this
 175          * page is the only one holding a reference to it. lock the page
 176          * and return OK.
 177          */
 178         if (page_count(page) == 1) {
 179                 lock_page(page);
 180                 return 0;
 181         }
 182
 183         return 1;
 184 }
 185 EXPORT_SYMBOL(generic_pipe_buf_steal);
 186
 187 /**
 188  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 189  * @pipe:       the pipe that the buffer belongs to
 190  * @buf:        the buffer to get a reference to
 191  *
 192  * Description:
 193  *      This function grabs an extra reference to @buf. It's used in
 194  *      in the tee() system call, when we duplicate the buffers in one
 195  *      pipe into another.
 196  */
 197 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 198 {
 199         get_page(buf->page);
 200 }
 201 EXPORT_SYMBOL(generic_pipe_buf_get);
 202
 203 /**
 204  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 205  * @info:       the pipe that the buffer belongs to
 206  * @buf:        the buffer to confirm
 207  *
 208  * Description:
 209  *      This function does nothing, because the generic pipe code uses
 210  *      pages that are always good when inserted into the pipe.
 211  */
 212 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 213                              struct pipe_buffer *buf)
 214 {
 215         return 0;
 216 }
 217 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 218
 219 /**
 220  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 221  * @pipe:       the pipe that the buffer belongs to
 222  * @buf:        the buffer to put a reference to
 223  *
 224  * Description:
 225  *      This function releases a reference to @buf.
 226  */
 227 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 228                               struct pipe_buffer *buf)
 229 {
 230         put_page(buf->page);
 231 }
 232 EXPORT_SYMBOL(generic_pipe_buf_release);
 233
 234 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 235         .can_merge = 1,
 236         .confirm = generic_pipe_buf_confirm,
 237         .release = anon_pipe_buf_release,
 238         .steal = anon_pipe_buf_steal,
 239         .get = generic_pipe_buf_get,
 240 };
 241
 242 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 243         .can_merge = 0,
 244         .confirm = generic_pipe_buf_confirm,
 245         .release = anon_pipe_buf_release,
 246         .steal = anon_pipe_buf_steal,
 247         .get = generic_pipe_buf_get,
 248 };
 249
 250 static ssize_t
 251 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 252 {
 253         size_t total_len = iov_iter_count(to);
 254         struct file *filp = iocb->ki_filp;
 255         struct pipe_inode_info *pipe = filp->private_data;
 256         int do_wakeup;
 257         ssize_t ret;
 258
 259         /* Null read succeeds. */
 260         if (unlikely(total_len == 0))
 261                 return 0;
 262
 263         do_wakeup = 0;
 264         ret = 0;
 265         __pipe_lock(pipe);
 266         for (;;) {
 267                 int bufs = pipe->nrbufs;
 268                 if (bufs) {
 269                         int curbuf = pipe->curbuf;
 270                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 271                         size_t chars = buf->len;
 272                         size_t written;
 273                         int error;
 274
 275                         if (chars > total_len)
 276                                 chars = total_len;
 277
 278                         error = pipe_buf_confirm(pipe, buf);
 279                         if (error) {
 280                                 if (!ret)
 281                                         ret = error;
 282                                 break;
 283                         }
 284
 285                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 286                         if (unlikely(written < chars)) {
 287                                 if (!ret)
 288                                         ret = -EFAULT;
 289                                 break;
 290                         }
 291                         ret += chars;
 292                         buf->offset += chars;
 293                         buf->len -= chars;
 294
 295                         /* Was it a packet buffer? Clean up and exit */
 296                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 297                                 total_len = chars;
 298                                 buf->len = 0;
 299                         }
 300
 301                         if (!buf->len) {
 302                                 pipe_buf_release(pipe, buf);
 303                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 304                                 pipe->curbuf = curbuf;
 305                                 pipe->nrbufs = --bufs;
 306                                 do_wakeup = 1;
 307                         }
 308                         total_len -= chars;
 309                         if (!total_len)
 310                                 break;  /* common path: read succeeded */
 311                 }
 312                 if (bufs)       /* More to do? */
 313                         continue;
 314                 if (!pipe->writers)
 315                         break;
 316                 if (!pipe->waiting_writers) {
 317                         /* syscall merging: Usually we must not sleep
 318                          * if O_NONBLOCK is set, or if we got some data.
 319                          * But if a writer sleeps in kernel space, then
 320                          * we can wait for that data without violating POSIX.
 321                          */
 322                         if (ret)
 323                                 break;
 324                         if (filp->f_flags & O_NONBLOCK) {
 325                                 ret = -EAGAIN;
 326                                 break;
 327                         }
 328                 }
 329                 if (signal_pending(current)) {
 330                         if (!ret)
 331                                 ret = -ERESTARTSYS;
 332                         break;
 333                 }
 334                 if (do_wakeup) {
 335                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 336                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 337                 }
 338                 pipe_wait(pipe);
 339         }
 340         __pipe_unlock(pipe);
 341
 342         /* Signal writers asynchronously that there is more room. */
 343         if (do_wakeup) {
 344                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 345                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 346         }
 347         if (ret > 0)
 348                 file_accessed(filp);
 349         return ret;
 350 }
 351
 352 static inline int is_packetized(struct file *file)
 353 {
 354         return (file->f_flags & O_DIRECT) != 0;
 355 }
 356
 357 static ssize_t
 358 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 359 {
 360         struct file *filp = iocb->ki_filp;
 361         struct pipe_inode_info *pipe = filp->private_data;
 362         ssize_t ret = 0;
 363         int do_wakeup = 0;
 364         size_t total_len = iov_iter_count(from);
 365         ssize_t chars;
 366
 367         /* Null write succeeds. */
 368         if (unlikely(total_len == 0))
 369                 return 0;
 370
 371         __pipe_lock(pipe);
 372
 373         if (!pipe->readers) {
 374                 send_sig(SIGPIPE, current, 0);
 375                 ret = -EPIPE;
 376                 goto out;
 377         }
 378
 379         /* We try to merge small writes */
 380         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 381         if (pipe->nrbufs && chars != 0) {
 382                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 383                                                         (pipe->buffers - 1);
 384                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 385                 int offset = buf->offset + buf->len;
 386
 387                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
 388                         ret = pipe_buf_confirm(pipe, buf);
 389                         if (ret)
 390                                 goto out;
 391
 392                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 393                         if (unlikely(ret < chars)) {
 394                                 ret = -EFAULT;
 395                                 goto out;
 396                         }
 397                         do_wakeup = 1;
 398                         buf->len += ret;
 399                         if (!iov_iter_count(from))
 400                                 goto out;
 401                 }
 402         }
 403
 404         for (;;) {
 405                 int bufs;
 406
 407                 if (!pipe->readers) {
 408                         send_sig(SIGPIPE, current, 0);
 409                         if (!ret)
 410                                 ret = -EPIPE;
 411                         break;
 412                 }
 413                 bufs = pipe->nrbufs;
 414                 if (bufs < pipe->buffers) {
 415                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 416                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 417                         struct page *page = pipe->tmp_page;
 418                         int copied;
 419
 420                         if (!page) {
 421                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 422                                 if (unlikely(!page)) {
 423                                         ret = ret ? : -ENOMEM;
 424                                         break;
 425                                 }
 426                                 pipe->tmp_page = page;
 427                         }
 428                         /* Always wake up, even if the copy fails. Otherwise
 429                          * we lock up (O_NONBLOCK-)readers that sleep due to
 430                          * syscall merging.
 431                          * FIXME! Is this really true?
 432                          */
 433                         do_wakeup = 1;
 434                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 435                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 436                                 if (!ret)
 437                                         ret = -EFAULT;
 438                                 break;
 439                         }
 440                         ret += copied;
 441
 442                         /* Insert it into the buffer array */
 443                         buf->page = page;
 444                         buf->ops = &anon_pipe_buf_ops;
 445                         buf->offset = 0;
 446                         buf->len = copied;
 447                         buf->flags = 0;
 448                         if (is_packetized(filp)) {
 449                                 buf->ops = &packet_pipe_buf_ops;
 450                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 451                         }
 452                         pipe->nrbufs = ++bufs;
 453                         pipe->tmp_page = NULL;
 454
 455                         if (!iov_iter_count(from))
 456                                 break;
 457                 }
 458                 if (bufs < pipe->buffers)
 459                         continue;
 460                 if (filp->f_flags & O_NONBLOCK) {
 461                         if (!ret)
 462                                 ret = -EAGAIN;
 463                         break;
 464                 }
 465                 if (signal_pending(current)) {
 466                         if (!ret)
 467                                 ret = -ERESTARTSYS;
 468                         break;
 469                 }
 470                 if (do_wakeup) {
 471                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 472                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 473                         do_wakeup = 0;
 474                 }
 475                 pipe->waiting_writers++;
 476                 pipe_wait(pipe);
 477                 pipe->waiting_writers--;
 478         }
 479 out:
 480         __pipe_unlock(pipe);
 481         if (do_wakeup) {
 482                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 483                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 484         }
 485         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 486                 int err = file_update_time(filp);
 487                 if (err)
 488                         ret = err;
 489                 sb_end_write(file_inode(filp)->i_sb);
 490         }
 491         return ret;
 492 }
 493
 494 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 495 {
 496         struct pipe_inode_info *pipe = filp->private_data;
 497         int count, buf, nrbufs;
 498
 499         switch (cmd) {
 500                 case FIONREAD:
 501                         __pipe_lock(pipe);
 502                         count = 0;
 503                         buf = pipe->curbuf;
 504                         nrbufs = pipe->nrbufs;
 505                         while (--nrbufs >= 0) {
 506                                 count += pipe->bufs[buf].len;
 507                                 buf = (buf+1) & (pipe->buffers - 1);
 508                         }
 509                         __pipe_unlock(pipe);
 510
 511                         return put_user(count, (int __user *)arg);
 512                 default:
 513                         return -ENOIOCTLCMD;
 514         }
 515 }
 516
 517 /* No kernel lock held - fine */
 518 static unsigned int
 519 pipe_poll(struct file *filp, poll_table *wait)
 520 {
 521         unsigned int mask;
 522         struct pipe_inode_info *pipe = filp->private_data;
 523         int nrbufs;
 524
 525         poll_wait(filp, &pipe->wait, wait);
 526
 527         /* Reading only -- no need for acquiring the semaphore.  */
 528         nrbufs = pipe->nrbufs;
 529         mask = 0;
 530         if (filp->f_mode & FMODE_READ) {
 531                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 532                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 533                         mask |= POLLHUP;
 534         }
 535
 536         if (filp->f_mode & FMODE_WRITE) {
 537                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 538                 /*
 539                  * Most Unices do not set POLLERR for FIFOs but on Linux they
 540                  * behave exactly like pipes for poll().
 541                  */
 542                 if (!pipe->readers)
 543                         mask |= POLLERR;
 544         }
 545
 546         return mask;
 547 }
 548
 549 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 550 {
 551         int kill = 0;
 552
 553         spin_lock(&inode->i_lock);
 554         if (!--pipe->files) {
 555                 inode->i_pipe = NULL;
 556                 kill = 1;
 557         }
 558         spin_unlock(&inode->i_lock);
 559
 560         if (kill)
 561                 free_pipe_info(pipe);
 562 }
 563
 564 static int
 565 pipe_release(struct inode *inode, struct file *file)
 566 {
 567         struct pipe_inode_info *pipe = file->private_data;
 568
 569         __pipe_lock(pipe);
 570         if (file->f_mode & FMODE_READ)
 571                 pipe->readers--;
 572         if (file->f_mode & FMODE_WRITE)
 573                 pipe->writers--;
 574
 575         if (pipe->readers || pipe->writers) {
 576                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 577                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 578                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 579         }
 580         __pipe_unlock(pipe);
 581
 582         put_pipe_info(inode, pipe);
 583         return 0;
 584 }
 585
 586 static int
 587 pipe_fasync(int fd, struct file *filp, int on)
 588 {
 589         struct pipe_inode_info *pipe = filp->private_data;
 590         int retval = 0;
 591
 592         __pipe_lock(pipe);
 593         if (filp->f_mode & FMODE_READ)
 594                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 595         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 596                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 597                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 598                         /* this can happen only if on == T */
 599                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 600         }
 601         __pipe_unlock(pipe);
 602         return retval;
 603 }
 604
 605 static unsigned long account_pipe_buffers(struct user_struct *user,
 606                                  unsigned long old, unsigned long new)
 607 {
 608         return atomic_long_add_return(new - old, &user->pipe_bufs);
 609 }
 610
 611 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 612 {
 613         return pipe_user_pages_soft && user_bufs >= pipe_user_pages_soft;
 614 }
 615
 616 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 617 {
 618         return pipe_user_pages_hard && user_bufs >= pipe_user_pages_hard;
 619 }
 620
 621 struct pipe_inode_info *alloc_pipe_info(void)
 622 {
 623         struct pipe_inode_info *pipe;
 624         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 625         struct user_struct *user = get_current_user();
 626         unsigned long user_bufs;
 627
 628         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 629         if (pipe == NULL)
 630                 goto out_free_uid;
 631
 632         if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
 633                 pipe_bufs = pipe_max_size >> PAGE_SHIFT;
 634
 635         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 636
 637         if (too_many_pipe_buffers_soft(user_bufs)) {
 638                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 639                 pipe_bufs = 1;
 640         }
 641
 642         if (too_many_pipe_buffers_hard(user_bufs))
 643                 goto out_revert_acct;
 644
 645         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 646                              GFP_KERNEL_ACCOUNT);
 647
 648         if (pipe->bufs) {
 649                 init_waitqueue_head(&pipe->wait);
 650                 pipe->r_counter = pipe->w_counter = 1;
 651                 pipe->buffers = pipe_bufs;
 652                 pipe->user = user;
 653                 mutex_init(&pipe->mutex);
 654                 return pipe;
 655         }
 656
 657 out_revert_acct:
 658         (void) account_pipe_buffers(user, pipe_bufs, 0);
 659         kfree(pipe);
 660 out_free_uid:
 661         free_uid(user);
 662         return NULL;
 663 }
 664
 665 void free_pipe_info(struct pipe_inode_info *pipe)
 666 {
 667         int i;
 668
 669         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 670         free_uid(pipe->user);
 671         for (i = 0; i < pipe->buffers; i++) {
 672                 struct pipe_buffer *buf = pipe->bufs + i;
 673                 if (buf->ops)
 674                         pipe_buf_release(pipe, buf);
 675         }
 676         if (pipe->tmp_page)
 677                 __free_page(pipe->tmp_page);
 678         kfree(pipe->bufs);
 679         kfree(pipe);
 680 }
 681
 682 static struct vfsmount *pipe_mnt __read_mostly;
 683
 684 /*
 685  * pipefs_dname() is called from d_path().
 686  */
 687 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 688 {
 689         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 690                                 d_inode(dentry)->i_ino);
 691 }
 692
 693 static const struct dentry_operations pipefs_dentry_operations = {
 694         .d_dname        = pipefs_dname,
 695 };
 696
 697 static struct inode * get_pipe_inode(void)
 698 {
 699         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 700         struct pipe_inode_info *pipe;
 701
 702         if (!inode)
 703                 goto fail_inode;
 704
 705         inode->i_ino = get_next_ino();
 706
 707         pipe = alloc_pipe_info();
 708         if (!pipe)
 709                 goto fail_iput;
 710
 711         inode->i_pipe = pipe;
 712         pipe->files = 2;
 713         pipe->readers = pipe->writers = 1;
 714         inode->i_fop = &pipefifo_fops;
 715
 716         /*
 717          * Mark the inode dirty from the very beginning,
 718          * that way it will never be moved to the dirty
 719          * list because "mark_inode_dirty()" will think
 720          * that it already _is_ on the dirty list.
 721          */
 722         inode->i_state = I_DIRTY;
 723         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 724         inode->i_uid = current_fsuid();
 725         inode->i_gid = current_fsgid();
 726         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 727
 728         return inode;
 729
 730 fail_iput:
 731         iput(inode);
 732
 733 fail_inode:
 734         return NULL;
 735 }
 736
 737 int create_pipe_files(struct file **res, int flags)
 738 {
 739         int err;
 740         struct inode *inode = get_pipe_inode();
 741         struct file *f;
 742         struct path path;
 743
 744         if (!inode)
 745                 return -ENFILE;
 746
 747         err = -ENOMEM;
 748         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &empty_name);
 749         if (!path.dentry)
 750                 goto err_inode;
 751         path.mnt = mntget(pipe_mnt);
 752
 753         d_instantiate(path.dentry, inode);
 754
 755         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
 756         if (IS_ERR(f)) {
 757                 err = PTR_ERR(f);
 758                 goto err_dentry;
 759         }
 760
 761         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
 762         f->private_data = inode->i_pipe;
 763
 764         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
 765         if (IS_ERR(res[0])) {
 766                 err = PTR_ERR(res[0]);
 767                 goto err_file;
 768         }
 769
 770         path_get(&path);
 771         res[0]->private_data = inode->i_pipe;
 772         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
 773         res[1] = f;
 774         return 0;
 775
 776 err_file:
 777         put_filp(f);
 778 err_dentry:
 779         free_pipe_info(inode->i_pipe);
 780         path_put(&path);
 781         return err;
 782
 783 err_inode:
 784         free_pipe_info(inode->i_pipe);
 785         iput(inode);
 786         return err;
 787 }
 788
 789 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 790 {
 791         int error;
 792         int fdw, fdr;
 793
 794         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 795                 return -EINVAL;
 796
 797         error = create_pipe_files(files, flags);
 798         if (error)
 799                 return error;
 800
 801         error = get_unused_fd_flags(flags);
 802         if (error < 0)
 803                 goto err_read_pipe;
 804         fdr = error;
 805
 806         error = get_unused_fd_flags(flags);
 807         if (error < 0)
 808                 goto err_fdr;
 809         fdw = error;
 810
 811         audit_fd_pair(fdr, fdw);
 812         fd[0] = fdr;
 813         fd[1] = fdw;
 814         return 0;
 815
 816  err_fdr:
 817         put_unused_fd(fdr);
 818  err_read_pipe:
 819         fput(files[0]);
 820         fput(files[1]);
 821         return error;
 822 }
 823
 824 int do_pipe_flags(int *fd, int flags)
 825 {
 826         struct file *files[2];
 827         int error = __do_pipe_flags(fd, files, flags);
 828         if (!error) {
 829                 fd_install(fd[0], files[0]);
 830                 fd_install(fd[1], files[1]);
 831         }
 832         return error;
 833 }
 834
 835 /*
 836  * sys_pipe() is the normal C calling standard for creating
 837  * a pipe. It's not the way Unix traditionally does this, though.
 838  */
 839 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 840 {
 841         struct file *files[2];
 842         int fd[2];
 843         int error;
 844
 845         error = __do_pipe_flags(fd, files, flags);
 846         if (!error) {
 847                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 848                         fput(files[0]);
 849                         fput(files[1]);
 850                         put_unused_fd(fd[0]);
 851                         put_unused_fd(fd[1]);
 852                         error = -EFAULT;
 853                 } else {
 854                         fd_install(fd[0], files[0]);
 855                         fd_install(fd[1], files[1]);
 856                 }
 857         }
 858         return error;
 859 }
 860
 861 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 862 {
 863         return sys_pipe2(fildes, 0);
 864 }
 865
 866 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 867 {
 868         int cur = *cnt;
 869
 870         while (cur == *cnt) {
 871                 pipe_wait(pipe);
 872                 if (signal_pending(current))
 873                         break;
 874         }
 875         return cur == *cnt ? -ERESTARTSYS : 0;
 876 }
 877
 878 static void wake_up_partner(struct pipe_inode_info *pipe)
 879 {
 880         wake_up_interruptible(&pipe->wait);
 881 }
 882
 883 static int fifo_open(struct inode *inode, struct file *filp)
 884 {
 885         struct pipe_inode_info *pipe;
 886         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 887         int ret;
 888
 889         filp->f_version = 0;
 890
 891         spin_lock(&inode->i_lock);
 892         if (inode->i_pipe) {
 893                 pipe = inode->i_pipe;
 894                 pipe->files++;
 895                 spin_unlock(&inode->i_lock);
 896         } else {
 897                 spin_unlock(&inode->i_lock);
 898                 pipe = alloc_pipe_info();
 899                 if (!pipe)
 900                         return -ENOMEM;
 901                 pipe->files = 1;
 902                 spin_lock(&inode->i_lock);
 903                 if (unlikely(inode->i_pipe)) {
 904                         inode->i_pipe->files++;
 905                         spin_unlock(&inode->i_lock);
 906                         free_pipe_info(pipe);
 907                         pipe = inode->i_pipe;
 908                 } else {
 909                         inode->i_pipe = pipe;
 910                         spin_unlock(&inode->i_lock);
 911                 }
 912         }
 913         filp->private_data = pipe;
 914         /* OK, we have a pipe and it's pinned down */
 915
 916         __pipe_lock(pipe);
 917
 918         /* We can only do regular read/write on fifos */
 919         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 920
 921         switch (filp->f_mode) {
 922         case FMODE_READ:
 923         /*
 924          *  O_RDONLY
 925          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 926          *  opened, even when there is no process writing the FIFO.
 927          */
 928                 pipe->r_counter++;
 929                 if (pipe->readers++ == 0)
 930                         wake_up_partner(pipe);
 931
 932                 if (!is_pipe && !pipe->writers) {
 933                         if ((filp->f_flags & O_NONBLOCK)) {
 934                                 /* suppress POLLHUP until we have
 935                                  * seen a writer */
 936                                 filp->f_version = pipe->w_counter;
 937                         } else {
 938                                 if (wait_for_partner(pipe, &pipe->w_counter))
 939                                         goto err_rd;
 940                         }
 941                 }
 942                 break;
 943
 944         case FMODE_WRITE:
 945         /*
 946          *  O_WRONLY
 947          *  POSIX.1 says that O_NONBLOCK means return -1 with
 948          *  errno=ENXIO when there is no process reading the FIFO.
 949          */
 950                 ret = -ENXIO;
 951                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 952                         goto err;
 953
 954                 pipe->w_counter++;
 955                 if (!pipe->writers++)
 956                         wake_up_partner(pipe);
 957
 958                 if (!is_pipe && !pipe->readers) {
 959                         if (wait_for_partner(pipe, &pipe->r_counter))
 960                                 goto err_wr;
 961                 }
 962                 break;
 963
 964         case FMODE_READ | FMODE_WRITE:
 965         /*
 966          *  O_RDWR
 967          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 968          *  This implementation will NEVER block on a O_RDWR open, since
 969          *  the process can at least talk to itself.
 970          */
 971
 972                 pipe->readers++;
 973                 pipe->writers++;
 974                 pipe->r_counter++;
 975                 pipe->w_counter++;
 976                 if (pipe->readers == 1 || pipe->writers == 1)
 977                         wake_up_partner(pipe);
 978                 break;
 979
 980         default:
 981                 ret = -EINVAL;
 982                 goto err;
 983         }
 984
 985         /* Ok! */
 986         __pipe_unlock(pipe);
 987         return 0;
 988
 989 err_rd:
 990         if (!--pipe->readers)
 991                 wake_up_interruptible(&pipe->wait);
 992         ret = -ERESTARTSYS;
 993         goto err;
 994
 995 err_wr:
 996         if (!--pipe->writers)
 997                 wake_up_interruptible(&pipe->wait);
 998         ret = -ERESTARTSYS;
 999         goto err;
1000
1001 err:
1002         __pipe_unlock(pipe);
1003
1004         put_pipe_info(inode, pipe);
1005         return ret;
1006 }
1007
1008 const struct file_operations pipefifo_fops = {
1009         .open           = fifo_open,
1010         .llseek         = no_llseek,
1011         .read_iter      = pipe_read,
1012         .write_iter     = pipe_write,
1013         .poll           = pipe_poll,
1014         .unlocked_ioctl = pipe_ioctl,
1015         .release        = pipe_release,
1016         .fasync         = pipe_fasync,
1017 };
1018
1019 /*
1020  * Currently we rely on the pipe array holding a power-of-2 number
1021  * of pages.
1022  */
1023 static inline unsigned int round_pipe_size(unsigned int size)
1024 {
1025         unsigned long nr_pages;
1026
1027         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1028         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1029 }
1030
1031 /*
1032  * Allocate a new array of pipe buffers and copy the info over. Returns the
1033  * pipe size if successful, or return -ERROR on error.
1034  */
1035 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1036 {
1037         struct pipe_buffer *bufs;
1038         unsigned int size, nr_pages;
1039         unsigned long user_bufs;
1040         long ret = 0;
1041
1042         size = round_pipe_size(arg);
1043         nr_pages = size >> PAGE_SHIFT;
1044
1045         if (!nr_pages)
1046                 return -EINVAL;
1047
1048         /*
1049          * If trying to increase the pipe capacity, check that an
1050          * unprivileged user is not trying to exceed various limits
1051          * (soft limit check here, hard limit check just below).
1052          * Decreasing the pipe capacity is always permitted, even
1053          * if the user is currently over a limit.
1054          */
1055         if (nr_pages > pipe->buffers &&
1056                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1057                 return -EPERM;
1058
1059         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1060
1061         if (nr_pages > pipe->buffers &&
1062                         (too_many_pipe_buffers_hard(user_bufs) ||
1063                          too_many_pipe_buffers_soft(user_bufs)) &&
1064                         !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1065                 ret = -EPERM;
1066                 goto out_revert_acct;
1067         }
1068
1069         /*
1070          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1071          * expect a lot of shrink+grow operations, just free and allocate
1072          * again like we would do for growing. If the pipe currently
1073          * contains more buffers than arg, then return busy.
1074          */
1075         if (nr_pages < pipe->nrbufs) {
1076                 ret = -EBUSY;
1077                 goto out_revert_acct;
1078         }
1079
1080         bufs = kcalloc(nr_pages, sizeof(*bufs),
1081                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1082         if (unlikely(!bufs)) {
1083                 ret = -ENOMEM;
1084                 goto out_revert_acct;
1085         }
1086
1087         /*
1088          * The pipe array wraps around, so just start the new one at zero
1089          * and adjust the indexes.
1090          */
1091         if (pipe->nrbufs) {
1092                 unsigned int tail;
1093                 unsigned int head;
1094
1095                 tail = pipe->curbuf + pipe->nrbufs;
1096                 if (tail < pipe->buffers)
1097                         tail = 0;
1098                 else
1099                         tail &= (pipe->buffers - 1);
1100
1101                 head = pipe->nrbufs - tail;
1102                 if (head)
1103                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1104                 if (tail)
1105                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1106         }
1107
1108         pipe->curbuf = 0;
1109         kfree(pipe->bufs);
1110         pipe->bufs = bufs;
1111         pipe->buffers = nr_pages;
1112         return nr_pages * PAGE_SIZE;
1113
1114 out_revert_acct:
1115         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1116         return ret;
1117 }
1118
1119 /*
1120  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1121  * will return an error.
1122  */
1123 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1124                  size_t *lenp, loff_t *ppos)
1125 {
1126         int ret;
1127
1128         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1129         if (ret < 0 || !write)
1130                 return ret;
1131
1132         pipe_max_size = round_pipe_size(pipe_max_size);
1133         return ret;
1134 }
1135
1136 /*
1137  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1138  * location, so checking ->i_pipe is not enough to verify that this is a
1139  * pipe.
1140  */
1141 struct pipe_inode_info *get_pipe_info(struct file *file)
1142 {
1143         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1144 }
1145
1146 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1147 {
1148         struct pipe_inode_info *pipe;
1149         long ret;
1150
1151         pipe = get_pipe_info(file);
1152         if (!pipe)
1153                 return -EBADF;
1154
1155         __pipe_lock(pipe);
1156
1157         switch (cmd) {
1158         case F_SETPIPE_SZ:
1159                 ret = pipe_set_size(pipe, arg);
1160                 break;
1161         case F_GETPIPE_SZ:
1162                 ret = pipe->buffers * PAGE_SIZE;
1163                 break;
1164         default:
1165                 ret = -EINVAL;
1166                 break;
1167         }
1168
1169         __pipe_unlock(pipe);
1170         return ret;
1171 }
1172
1173 static const struct super_operations pipefs_ops = {
1174         .destroy_inode = free_inode_nonrcu,
1175         .statfs = simple_statfs,
1176 };
1177
1178 /*
1179  * pipefs should _never_ be mounted by userland - too much of security hassle,
1180  * no real gain from having the whole whorehouse mounted. So we don't need
1181  * any operations on the root directory. However, we need a non-trivial
1182  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1183  */
1184 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1185                          int flags, const char *dev_name, void *data)
1186 {
1187         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1188                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1189 }
1190
1191 static struct file_system_type pipe_fs_type = {
1192         .name           = "pipefs",
1193         .mount          = pipefs_mount,
1194         .kill_sb        = kill_anon_super,
1195 };
1196
1197 static int __init init_pipe_fs(void)
1198 {
1199         int err = register_filesystem(&pipe_fs_type);
1200
1201         if (!err) {
1202                 pipe_mnt = kern_mount(&pipe_fs_type);
1203                 if (IS_ERR(pipe_mnt)) {
1204                         err = PTR_ERR(pipe_mnt);
1205                         unregister_filesystem(&pipe_fs_type);
1206                 }
1207         }
1208         return err;
1209 }
1210
1211 fs_initcall(init_pipe_fs);