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 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  39  * matches default values.
  40  */
  41 unsigned long pipe_user_pages_hard;
  42 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  43
  44 /*
  45  * We use a start+len construction, which provides full use of the
  46  * allocated memory.
  47  * -- Florian Coosmann (FGC)
  48  *
  49  * Reads with count = 0 should always return 0.
  50  * -- Julian Bradfield 1999-06-07.
  51  *
  52  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  53  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  54  *
  55  * pipe_read & write cleanup
  56  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  57  */
  58
  59 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  60 {
  61         if (pipe->files)
  62                 mutex_lock_nested(&pipe->mutex, subclass);
  63 }
  64
  65 void pipe_lock(struct pipe_inode_info *pipe)
  66 {
  67         /*
  68          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  69          */
  70         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  71 }
  72 EXPORT_SYMBOL(pipe_lock);
  73
  74 void pipe_unlock(struct pipe_inode_info *pipe)
  75 {
  76         if (pipe->files)
  77                 mutex_unlock(&pipe->mutex);
  78 }
  79 EXPORT_SYMBOL(pipe_unlock);
  80
  81 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  82 {
  83         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
  84 }
  85
  86 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
  87 {
  88         mutex_unlock(&pipe->mutex);
  89 }
  90
  91 void pipe_double_lock(struct pipe_inode_info *pipe1,
  92                       struct pipe_inode_info *pipe2)
  93 {
  94         BUG_ON(pipe1 == pipe2);
  95
  96         if (pipe1 < pipe2) {
  97                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  98                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  99         } else {
 100                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 101                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 102         }
 103 }
 104
 105 /* Drop the inode semaphore and wait for a pipe event, atomically */
 106 void pipe_wait(struct pipe_inode_info *pipe)
 107 {
 108         DEFINE_WAIT(wait);
 109
 110         /*
 111          * Pipes are system-local resources, so sleeping on them
 112          * is considered a noninteractive wait:
 113          */
 114         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 115         pipe_unlock(pipe);
 116         schedule();
 117         finish_wait(&pipe->wait, &wait);
 118         pipe_lock(pipe);
 119 }
 120
 121 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 122                                   struct pipe_buffer *buf)
 123 {
 124         struct page *page = buf->page;
 125
 126         /*
 127          * If nobody else uses this page, and we don't already have a
 128          * temporary page, let's keep track of it as a one-deep
 129          * allocation cache. (Otherwise just release our reference to it)
 130          */
 131         if (page_count(page) == 1 && !pipe->tmp_page)
 132                 pipe->tmp_page = page;
 133         else
 134                 put_page(page);
 135 }
 136
 137 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 138                                struct pipe_buffer *buf)
 139 {
 140         struct page *page = buf->page;
 141
 142         if (page_count(page) == 1) {
 143                 memcg_kmem_uncharge(page, 0);
 144                 __SetPageLocked(page);
 145                 return 0;
 146         }
 147         return 1;
 148 }
 149
 150 /**
 151  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 152  * @pipe:       the pipe that the buffer belongs to
 153  * @buf:        the buffer to attempt to steal
 154  *
 155  * Description:
 156  *      This function attempts to steal the &struct page attached to
 157  *      @buf. If successful, this function returns 0 and returns with
 158  *      the page locked. The caller may then reuse the page for whatever
 159  *      he wishes; the typical use is insertion into a different file
 160  *      page cache.
 161  */
 162 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 163                            struct pipe_buffer *buf)
 164 {
 165         struct page *page = buf->page;
 166
 167         /*
 168          * A reference of one is golden, that means that the owner of this
 169          * page is the only one holding a reference to it. lock the page
 170          * and return OK.
 171          */
 172         if (page_count(page) == 1) {
 173                 lock_page(page);
 174                 return 0;
 175         }
 176
 177         return 1;
 178 }
 179 EXPORT_SYMBOL(generic_pipe_buf_steal);
 180
 181 /**
 182  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 183  * @pipe:       the pipe that the buffer belongs to
 184  * @buf:        the buffer to get a reference to
 185  *
 186  * Description:
 187  *      This function grabs an extra reference to @buf. It's used in
 188  *      in the tee() system call, when we duplicate the buffers in one
 189  *      pipe into another.
 190  */
 191 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 192 {
 193         get_page(buf->page);
 194 }
 195 EXPORT_SYMBOL(generic_pipe_buf_get);
 196
 197 /**
 198  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 199  * @info:       the pipe that the buffer belongs to
 200  * @buf:        the buffer to confirm
 201  *
 202  * Description:
 203  *      This function does nothing, because the generic pipe code uses
 204  *      pages that are always good when inserted into the pipe.
 205  */
 206 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 207                              struct pipe_buffer *buf)
 208 {
 209         return 0;
 210 }
 211 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 212
 213 /**
 214  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 215  * @pipe:       the pipe that the buffer belongs to
 216  * @buf:        the buffer to put a reference to
 217  *
 218  * Description:
 219  *      This function releases a reference to @buf.
 220  */
 221 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 222                               struct pipe_buffer *buf)
 223 {
 224         put_page(buf->page);
 225 }
 226 EXPORT_SYMBOL(generic_pipe_buf_release);
 227
 228 /* New data written to a pipe may be appended to a buffer with this type. */
 229 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 230         .confirm = generic_pipe_buf_confirm,
 231         .release = anon_pipe_buf_release,
 232         .steal = anon_pipe_buf_steal,
 233         .get = generic_pipe_buf_get,
 234 };
 235
 236 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 237         .confirm = generic_pipe_buf_confirm,
 238         .release = anon_pipe_buf_release,
 239         .steal = anon_pipe_buf_steal,
 240         .get = generic_pipe_buf_get,
 241 };
 242
 243 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 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 /**
 251  * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
 252  * @buf:        the buffer to mark
 253  *
 254  * Description:
 255  *      This function ensures that no future writes will be merged into the
 256  *      given &struct pipe_buffer. This is necessary when multiple pipe buffers
 257  *      share the same backing page.
 258  */
 259 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 260 {
 261         if (buf->ops == &anon_pipe_buf_ops)
 262                 buf->ops = &anon_pipe_buf_nomerge_ops;
 263 }
 264
 265 static bool pipe_buf_can_merge(struct pipe_buffer *buf)
 266 {
 267         return buf->ops == &anon_pipe_buf_ops;
 268 }
 269
 270 static ssize_t
 271 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 272 {
 273         size_t total_len = iov_iter_count(to);
 274         struct file *filp = iocb->ki_filp;
 275         struct pipe_inode_info *pipe = filp->private_data;
 276         int do_wakeup;
 277         ssize_t ret;
 278
 279         /* Null read succeeds. */
 280         if (unlikely(total_len == 0))
 281                 return 0;
 282
 283         do_wakeup = 0;
 284         ret = 0;
 285         __pipe_lock(pipe);
 286         for (;;) {
 287                 int bufs = pipe->nrbufs;
 288                 if (bufs) {
 289                         int curbuf = pipe->curbuf;
 290                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 291                         size_t chars = buf->len;
 292                         size_t written;
 293                         int error;
 294
 295                         if (chars > total_len)
 296                                 chars = total_len;
 297
 298                         error = pipe_buf_confirm(pipe, buf);
 299                         if (error) {
 300                                 if (!ret)
 301                                         ret = error;
 302                                 break;
 303                         }
 304
 305                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 306                         if (unlikely(written < chars)) {
 307                                 if (!ret)
 308                                         ret = -EFAULT;
 309                                 break;
 310                         }
 311                         ret += chars;
 312                         buf->offset += chars;
 313                         buf->len -= chars;
 314
 315                         /* Was it a packet buffer? Clean up and exit */
 316                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 317                                 total_len = chars;
 318                                 buf->len = 0;
 319                         }
 320
 321                         if (!buf->len) {
 322                                 pipe_buf_release(pipe, buf);
 323                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 324                                 pipe->curbuf = curbuf;
 325                                 pipe->nrbufs = --bufs;
 326                                 do_wakeup = 1;
 327                         }
 328                         total_len -= chars;
 329                         if (!total_len)
 330                                 break;  /* common path: read succeeded */
 331                 }
 332                 if (bufs)       /* More to do? */
 333                         continue;
 334                 if (!pipe->writers)
 335                         break;
 336                 if (!pipe->waiting_writers) {
 337                         /* syscall merging: Usually we must not sleep
 338                          * if O_NONBLOCK is set, or if we got some data.
 339                          * But if a writer sleeps in kernel space, then
 340                          * we can wait for that data without violating POSIX.
 341                          */
 342                         if (ret)
 343                                 break;
 344                         if (filp->f_flags & O_NONBLOCK) {
 345                                 ret = -EAGAIN;
 346                                 break;
 347                         }
 348                 }
 349                 if (signal_pending(current)) {
 350                         if (!ret)
 351                                 ret = -ERESTARTSYS;
 352                         break;
 353                 }
 354                 if (do_wakeup) {
 355                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 356                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 357                 }
 358                 pipe_wait(pipe);
 359         }
 360         __pipe_unlock(pipe);
 361
 362         /* Signal writers asynchronously that there is more room. */
 363         if (do_wakeup) {
 364                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 365                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 366         }
 367         if (ret > 0)
 368                 file_accessed(filp);
 369         return ret;
 370 }
 371
 372 static inline int is_packetized(struct file *file)
 373 {
 374         return (file->f_flags & O_DIRECT) != 0;
 375 }
 376
 377 static ssize_t
 378 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 379 {
 380         struct file *filp = iocb->ki_filp;
 381         struct pipe_inode_info *pipe = filp->private_data;
 382         ssize_t ret = 0;
 383         int do_wakeup = 0;
 384         size_t total_len = iov_iter_count(from);
 385         ssize_t chars;
 386
 387         /* Null write succeeds. */
 388         if (unlikely(total_len == 0))
 389                 return 0;
 390
 391         __pipe_lock(pipe);
 392
 393         if (!pipe->readers) {
 394                 send_sig(SIGPIPE, current, 0);
 395                 ret = -EPIPE;
 396                 goto out;
 397         }
 398
 399         /* We try to merge small writes */
 400         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 401         if (pipe->nrbufs && chars != 0) {
 402                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 403                                                         (pipe->buffers - 1);
 404                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 405                 int offset = buf->offset + buf->len;
 406
 407                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 408                         ret = pipe_buf_confirm(pipe, buf);
 409                         if (ret)
 410                                 goto out;
 411
 412                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 413                         if (unlikely(ret < chars)) {
 414                                 ret = -EFAULT;
 415                                 goto out;
 416                         }
 417                         do_wakeup = 1;
 418                         buf->len += ret;
 419                         if (!iov_iter_count(from))
 420                                 goto out;
 421                 }
 422         }
 423
 424         for (;;) {
 425                 int bufs;
 426
 427                 if (!pipe->readers) {
 428                         send_sig(SIGPIPE, current, 0);
 429                         if (!ret)
 430                                 ret = -EPIPE;
 431                         break;
 432                 }
 433                 bufs = pipe->nrbufs;
 434                 if (bufs < pipe->buffers) {
 435                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 436                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 437                         struct page *page = pipe->tmp_page;
 438                         int copied;
 439
 440                         if (!page) {
 441                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 442                                 if (unlikely(!page)) {
 443                                         ret = ret ? : -ENOMEM;
 444                                         break;
 445                                 }
 446                                 pipe->tmp_page = page;
 447                         }
 448                         /* Always wake up, even if the copy fails. Otherwise
 449                          * we lock up (O_NONBLOCK-)readers that sleep due to
 450                          * syscall merging.
 451                          * FIXME! Is this really true?
 452                          */
 453                         do_wakeup = 1;
 454                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 455                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 456                                 if (!ret)
 457                                         ret = -EFAULT;
 458                                 break;
 459                         }
 460                         ret += copied;
 461
 462                         /* Insert it into the buffer array */
 463                         buf->page = page;
 464                         buf->ops = &anon_pipe_buf_ops;
 465                         buf->offset = 0;
 466                         buf->len = copied;
 467                         buf->flags = 0;
 468                         if (is_packetized(filp)) {
 469                                 buf->ops = &packet_pipe_buf_ops;
 470                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 471                         }
 472                         pipe->nrbufs = ++bufs;
 473                         pipe->tmp_page = NULL;
 474
 475                         if (!iov_iter_count(from))
 476                                 break;
 477                 }
 478                 if (bufs < pipe->buffers)
 479                         continue;
 480                 if (filp->f_flags & O_NONBLOCK) {
 481                         if (!ret)
 482                                 ret = -EAGAIN;
 483                         break;
 484                 }
 485                 if (signal_pending(current)) {
 486                         if (!ret)
 487                                 ret = -ERESTARTSYS;
 488                         break;
 489                 }
 490                 if (do_wakeup) {
 491                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 492                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 493                         do_wakeup = 0;
 494                 }
 495                 pipe->waiting_writers++;
 496                 pipe_wait(pipe);
 497                 pipe->waiting_writers--;
 498         }
 499 out:
 500         __pipe_unlock(pipe);
 501         if (do_wakeup) {
 502                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 503                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 504         }
 505         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 506                 int err = file_update_time(filp);
 507                 if (err)
 508                         ret = err;
 509                 sb_end_write(file_inode(filp)->i_sb);
 510         }
 511         return ret;
 512 }
 513
 514 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 515 {
 516         struct pipe_inode_info *pipe = filp->private_data;
 517         int count, buf, nrbufs;
 518
 519         switch (cmd) {
 520                 case FIONREAD:
 521                         __pipe_lock(pipe);
 522                         count = 0;
 523                         buf = pipe->curbuf;
 524                         nrbufs = pipe->nrbufs;
 525                         while (--nrbufs >= 0) {
 526                                 count += pipe->bufs[buf].len;
 527                                 buf = (buf+1) & (pipe->buffers - 1);
 528                         }
 529                         __pipe_unlock(pipe);
 530
 531                         return put_user(count, (int __user *)arg);
 532                 default:
 533                         return -ENOIOCTLCMD;
 534         }
 535 }
 536
 537 /* No kernel lock held - fine */
 538 static __poll_t
 539 pipe_poll(struct file *filp, poll_table *wait)
 540 {
 541         __poll_t mask;
 542         struct pipe_inode_info *pipe = filp->private_data;
 543         int nrbufs;
 544
 545         poll_wait(filp, &pipe->wait, wait);
 546
 547         /* Reading only -- no need for acquiring the semaphore.  */
 548         nrbufs = pipe->nrbufs;
 549         mask = 0;
 550         if (filp->f_mode & FMODE_READ) {
 551                 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
 552                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 553                         mask |= EPOLLHUP;
 554         }
 555
 556         if (filp->f_mode & FMODE_WRITE) {
 557                 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
 558                 /*
 559                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 560                  * behave exactly like pipes for poll().
 561                  */
 562                 if (!pipe->readers)
 563                         mask |= EPOLLERR;
 564         }
 565
 566         return mask;
 567 }
 568
 569 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 570 {
 571         int kill = 0;
 572
 573         spin_lock(&inode->i_lock);
 574         if (!--pipe->files) {
 575                 inode->i_pipe = NULL;
 576                 kill = 1;
 577         }
 578         spin_unlock(&inode->i_lock);
 579
 580         if (kill)
 581                 free_pipe_info(pipe);
 582 }
 583
 584 static int
 585 pipe_release(struct inode *inode, struct file *file)
 586 {
 587         struct pipe_inode_info *pipe = file->private_data;
 588
 589         __pipe_lock(pipe);
 590         if (file->f_mode & FMODE_READ)
 591                 pipe->readers--;
 592         if (file->f_mode & FMODE_WRITE)
 593                 pipe->writers--;
 594
 595         if (pipe->readers || pipe->writers) {
 596                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 597                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 598                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 599         }
 600         __pipe_unlock(pipe);
 601
 602         put_pipe_info(inode, pipe);
 603         return 0;
 604 }
 605
 606 static int
 607 pipe_fasync(int fd, struct file *filp, int on)
 608 {
 609         struct pipe_inode_info *pipe = filp->private_data;
 610         int retval = 0;
 611
 612         __pipe_lock(pipe);
 613         if (filp->f_mode & FMODE_READ)
 614                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 615         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 616                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 617                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 618                         /* this can happen only if on == T */
 619                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 620         }
 621         __pipe_unlock(pipe);
 622         return retval;
 623 }
 624
 625 static unsigned long account_pipe_buffers(struct user_struct *user,
 626                                  unsigned long old, unsigned long new)
 627 {
 628         return atomic_long_add_return(new - old, &user->pipe_bufs);
 629 }
 630
 631 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 632 {
 633         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 634
 635         return soft_limit && user_bufs > soft_limit;
 636 }
 637
 638 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 639 {
 640         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 641
 642         return hard_limit && user_bufs > hard_limit;
 643 }
 644
 645 static bool is_unprivileged_user(void)
 646 {
 647         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 648 }
 649
 650 struct pipe_inode_info *alloc_pipe_info(void)
 651 {
 652         struct pipe_inode_info *pipe;
 653         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 654         struct user_struct *user = get_current_user();
 655         unsigned long user_bufs;
 656         unsigned int max_size = READ_ONCE(pipe_max_size);
 657
 658         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 659         if (pipe == NULL)
 660                 goto out_free_uid;
 661
 662         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 663                 pipe_bufs = max_size >> PAGE_SHIFT;
 664
 665         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 666
 667         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 668                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
 669                 pipe_bufs = 1;
 670         }
 671
 672         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 673                 goto out_revert_acct;
 674
 675         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 676                              GFP_KERNEL_ACCOUNT);
 677
 678         if (pipe->bufs) {
 679                 init_waitqueue_head(&pipe->wait);
 680                 pipe->r_counter = pipe->w_counter = 1;
 681                 pipe->buffers = pipe_bufs;
 682                 pipe->user = user;
 683                 mutex_init(&pipe->mutex);
 684                 return pipe;
 685         }
 686
 687 out_revert_acct:
 688         (void) account_pipe_buffers(user, pipe_bufs, 0);
 689         kfree(pipe);
 690 out_free_uid:
 691         free_uid(user);
 692         return NULL;
 693 }
 694
 695 void free_pipe_info(struct pipe_inode_info *pipe)
 696 {
 697         int i;
 698
 699         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 700         free_uid(pipe->user);
 701         for (i = 0; i < pipe->buffers; i++) {
 702                 struct pipe_buffer *buf = pipe->bufs + i;
 703                 if (buf->ops)
 704                         pipe_buf_release(pipe, buf);
 705         }
 706         if (pipe->tmp_page)
 707                 __free_page(pipe->tmp_page);
 708         kfree(pipe->bufs);
 709         kfree(pipe);
 710 }
 711
 712 static struct vfsmount *pipe_mnt __read_mostly;
 713
 714 /*
 715  * pipefs_dname() is called from d_path().
 716  */
 717 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 718 {
 719         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 720                                 d_inode(dentry)->i_ino);
 721 }
 722
 723 static const struct dentry_operations pipefs_dentry_operations = {
 724         .d_dname        = pipefs_dname,
 725 };
 726
 727 static struct inode * get_pipe_inode(void)
 728 {
 729         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 730         struct pipe_inode_info *pipe;
 731
 732         if (!inode)
 733                 goto fail_inode;
 734
 735         inode->i_ino = get_next_ino();
 736
 737         pipe = alloc_pipe_info();
 738         if (!pipe)
 739                 goto fail_iput;
 740
 741         inode->i_pipe = pipe;
 742         pipe->files = 2;
 743         pipe->readers = pipe->writers = 1;
 744         inode->i_fop = &pipefifo_fops;
 745
 746         /*
 747          * Mark the inode dirty from the very beginning,
 748          * that way it will never be moved to the dirty
 749          * list because "mark_inode_dirty()" will think
 750          * that it already _is_ on the dirty list.
 751          */
 752         inode->i_state = I_DIRTY;
 753         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 754         inode->i_uid = current_fsuid();
 755         inode->i_gid = current_fsgid();
 756         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 757
 758         return inode;
 759
 760 fail_iput:
 761         iput(inode);
 762
 763 fail_inode:
 764         return NULL;
 765 }
 766
 767 int create_pipe_files(struct file **res, int flags)
 768 {
 769         struct inode *inode = get_pipe_inode();
 770         struct file *f;
 771
 772         if (!inode)
 773                 return -ENFILE;
 774
 775         f = alloc_file_pseudo(inode, pipe_mnt, "",
 776                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 777                                 &pipefifo_fops);
 778         if (IS_ERR(f)) {
 779                 free_pipe_info(inode->i_pipe);
 780                 iput(inode);
 781                 return PTR_ERR(f);
 782         }
 783
 784         f->private_data = inode->i_pipe;
 785
 786         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 787                                   &pipefifo_fops);
 788         if (IS_ERR(res[0])) {
 789                 put_pipe_info(inode, inode->i_pipe);
 790                 fput(f);
 791                 return PTR_ERR(res[0]);
 792         }
 793         res[0]->private_data = inode->i_pipe;
 794         res[1] = f;
 795         return 0;
 796 }
 797
 798 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 799 {
 800         int error;
 801         int fdw, fdr;
 802
 803         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 804                 return -EINVAL;
 805
 806         error = create_pipe_files(files, flags);
 807         if (error)
 808                 return error;
 809
 810         error = get_unused_fd_flags(flags);
 811         if (error < 0)
 812                 goto err_read_pipe;
 813         fdr = error;
 814
 815         error = get_unused_fd_flags(flags);
 816         if (error < 0)
 817                 goto err_fdr;
 818         fdw = error;
 819
 820         audit_fd_pair(fdr, fdw);
 821         fd[0] = fdr;
 822         fd[1] = fdw;
 823         return 0;
 824
 825  err_fdr:
 826         put_unused_fd(fdr);
 827  err_read_pipe:
 828         fput(files[0]);
 829         fput(files[1]);
 830         return error;
 831 }
 832
 833 int do_pipe_flags(int *fd, int flags)
 834 {
 835         struct file *files[2];
 836         int error = __do_pipe_flags(fd, files, flags);
 837         if (!error) {
 838                 fd_install(fd[0], files[0]);
 839                 fd_install(fd[1], files[1]);
 840         }
 841         return error;
 842 }
 843
 844 /*
 845  * sys_pipe() is the normal C calling standard for creating
 846  * a pipe. It's not the way Unix traditionally does this, though.
 847  */
 848 static int do_pipe2(int __user *fildes, int flags)
 849 {
 850         struct file *files[2];
 851         int fd[2];
 852         int error;
 853
 854         error = __do_pipe_flags(fd, files, flags);
 855         if (!error) {
 856                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 857                         fput(files[0]);
 858                         fput(files[1]);
 859                         put_unused_fd(fd[0]);
 860                         put_unused_fd(fd[1]);
 861                         error = -EFAULT;
 862                 } else {
 863                         fd_install(fd[0], files[0]);
 864                         fd_install(fd[1], files[1]);
 865                 }
 866         }
 867         return error;
 868 }
 869
 870 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 871 {
 872         return do_pipe2(fildes, flags);
 873 }
 874
 875 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 876 {
 877         return do_pipe2(fildes, 0);
 878 }
 879
 880 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 881 {
 882         int cur = *cnt;
 883
 884         while (cur == *cnt) {
 885                 pipe_wait(pipe);
 886                 if (signal_pending(current))
 887                         break;
 888         }
 889         return cur == *cnt ? -ERESTARTSYS : 0;
 890 }
 891
 892 static void wake_up_partner(struct pipe_inode_info *pipe)
 893 {
 894         wake_up_interruptible(&pipe->wait);
 895 }
 896
 897 static int fifo_open(struct inode *inode, struct file *filp)
 898 {
 899         struct pipe_inode_info *pipe;
 900         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 901         int ret;
 902
 903         filp->f_version = 0;
 904
 905         spin_lock(&inode->i_lock);
 906         if (inode->i_pipe) {
 907                 pipe = inode->i_pipe;
 908                 pipe->files++;
 909                 spin_unlock(&inode->i_lock);
 910         } else {
 911                 spin_unlock(&inode->i_lock);
 912                 pipe = alloc_pipe_info();
 913                 if (!pipe)
 914                         return -ENOMEM;
 915                 pipe->files = 1;
 916                 spin_lock(&inode->i_lock);
 917                 if (unlikely(inode->i_pipe)) {
 918                         inode->i_pipe->files++;
 919                         spin_unlock(&inode->i_lock);
 920                         free_pipe_info(pipe);
 921                         pipe = inode->i_pipe;
 922                 } else {
 923                         inode->i_pipe = pipe;
 924                         spin_unlock(&inode->i_lock);
 925                 }
 926         }
 927         filp->private_data = pipe;
 928         /* OK, we have a pipe and it's pinned down */
 929
 930         __pipe_lock(pipe);
 931
 932         /* We can only do regular read/write on fifos */
 933         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 934
 935         switch (filp->f_mode) {
 936         case FMODE_READ:
 937         /*
 938          *  O_RDONLY
 939          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 940          *  opened, even when there is no process writing the FIFO.
 941          */
 942                 pipe->r_counter++;
 943                 if (pipe->readers++ == 0)
 944                         wake_up_partner(pipe);
 945
 946                 if (!is_pipe && !pipe->writers) {
 947                         if ((filp->f_flags & O_NONBLOCK)) {
 948                                 /* suppress EPOLLHUP until we have
 949                                  * seen a writer */
 950                                 filp->f_version = pipe->w_counter;
 951                         } else {
 952                                 if (wait_for_partner(pipe, &pipe->w_counter))
 953                                         goto err_rd;
 954                         }
 955                 }
 956                 break;
 957
 958         case FMODE_WRITE:
 959         /*
 960          *  O_WRONLY
 961          *  POSIX.1 says that O_NONBLOCK means return -1 with
 962          *  errno=ENXIO when there is no process reading the FIFO.
 963          */
 964                 ret = -ENXIO;
 965                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 966                         goto err;
 967
 968                 pipe->w_counter++;
 969                 if (!pipe->writers++)
 970                         wake_up_partner(pipe);
 971
 972                 if (!is_pipe && !pipe->readers) {
 973                         if (wait_for_partner(pipe, &pipe->r_counter))
 974                                 goto err_wr;
 975                 }
 976                 break;
 977
 978         case FMODE_READ | FMODE_WRITE:
 979         /*
 980          *  O_RDWR
 981          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 982          *  This implementation will NEVER block on a O_RDWR open, since
 983          *  the process can at least talk to itself.
 984          */
 985
 986                 pipe->readers++;
 987                 pipe->writers++;
 988                 pipe->r_counter++;
 989                 pipe->w_counter++;
 990                 if (pipe->readers == 1 || pipe->writers == 1)
 991                         wake_up_partner(pipe);
 992                 break;
 993
 994         default:
 995                 ret = -EINVAL;
 996                 goto err;
 997         }
 998
 999         /* Ok! */
1000         __pipe_unlock(pipe);
1001         return 0;
1002
1003 err_rd:
1004         if (!--pipe->readers)
1005                 wake_up_interruptible(&pipe->wait);
1006         ret = -ERESTARTSYS;
1007         goto err;
1008
1009 err_wr:
1010         if (!--pipe->writers)
1011                 wake_up_interruptible(&pipe->wait);
1012         ret = -ERESTARTSYS;
1013         goto err;
1014
1015 err:
1016         __pipe_unlock(pipe);
1017
1018         put_pipe_info(inode, pipe);
1019         return ret;
1020 }
1021
1022 const struct file_operations pipefifo_fops = {
1023         .open           = fifo_open,
1024         .llseek         = no_llseek,
1025         .read_iter      = pipe_read,
1026         .write_iter     = pipe_write,
1027         .poll           = pipe_poll,
1028         .unlocked_ioctl = pipe_ioctl,
1029         .release        = pipe_release,
1030         .fasync         = pipe_fasync,
1031 };
1032
1033 /*
1034  * Currently we rely on the pipe array holding a power-of-2 number
1035  * of pages. Returns 0 on error.
1036  */
1037 unsigned int round_pipe_size(unsigned long size)
1038 {
1039         if (size > (1U << 31))
1040                 return 0;
1041
1042         /* Minimum pipe size, as required by POSIX */
1043         if (size < PAGE_SIZE)
1044                 return PAGE_SIZE;
1045
1046         return roundup_pow_of_two(size);
1047 }
1048
1049 /*
1050  * Allocate a new array of pipe buffers and copy the info over. Returns the
1051  * pipe size if successful, or return -ERROR on error.
1052  */
1053 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1054 {
1055         struct pipe_buffer *bufs;
1056         unsigned int size, nr_pages;
1057         unsigned long user_bufs;
1058         long ret = 0;
1059
1060         size = round_pipe_size(arg);
1061         nr_pages = size >> PAGE_SHIFT;
1062
1063         if (!nr_pages)
1064                 return -EINVAL;
1065
1066         /*
1067          * If trying to increase the pipe capacity, check that an
1068          * unprivileged user is not trying to exceed various limits
1069          * (soft limit check here, hard limit check just below).
1070          * Decreasing the pipe capacity is always permitted, even
1071          * if the user is currently over a limit.
1072          */
1073         if (nr_pages > pipe->buffers &&
1074                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1075                 return -EPERM;
1076
1077         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1078
1079         if (nr_pages > pipe->buffers &&
1080                         (too_many_pipe_buffers_hard(user_bufs) ||
1081                          too_many_pipe_buffers_soft(user_bufs)) &&
1082                         is_unprivileged_user()) {
1083                 ret = -EPERM;
1084                 goto out_revert_acct;
1085         }
1086
1087         /*
1088          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1089          * expect a lot of shrink+grow operations, just free and allocate
1090          * again like we would do for growing. If the pipe currently
1091          * contains more buffers than arg, then return busy.
1092          */
1093         if (nr_pages < pipe->nrbufs) {
1094                 ret = -EBUSY;
1095                 goto out_revert_acct;
1096         }
1097
1098         bufs = kcalloc(nr_pages, sizeof(*bufs),
1099                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1100         if (unlikely(!bufs)) {
1101                 ret = -ENOMEM;
1102                 goto out_revert_acct;
1103         }
1104
1105         /*
1106          * The pipe array wraps around, so just start the new one at zero
1107          * and adjust the indexes.
1108          */
1109         if (pipe->nrbufs) {
1110                 unsigned int tail;
1111                 unsigned int head;
1112
1113                 tail = pipe->curbuf + pipe->nrbufs;
1114                 if (tail < pipe->buffers)
1115                         tail = 0;
1116                 else
1117                         tail &= (pipe->buffers - 1);
1118
1119                 head = pipe->nrbufs - tail;
1120                 if (head)
1121                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1122                 if (tail)
1123                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1124         }
1125
1126         pipe->curbuf = 0;
1127         kfree(pipe->bufs);
1128         pipe->bufs = bufs;
1129         pipe->buffers = nr_pages;
1130         return nr_pages * PAGE_SIZE;
1131
1132 out_revert_acct:
1133         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1134         return ret;
1135 }
1136
1137 /*
1138  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1139  * location, so checking ->i_pipe is not enough to verify that this is a
1140  * pipe.
1141  */
1142 struct pipe_inode_info *get_pipe_info(struct file *file)
1143 {
1144         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1145 }
1146
1147 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1148 {
1149         struct pipe_inode_info *pipe;
1150         long ret;
1151
1152         pipe = get_pipe_info(file);
1153         if (!pipe)
1154                 return -EBADF;
1155
1156         __pipe_lock(pipe);
1157
1158         switch (cmd) {
1159         case F_SETPIPE_SZ:
1160                 ret = pipe_set_size(pipe, arg);
1161                 break;
1162         case F_GETPIPE_SZ:
1163                 ret = pipe->buffers * PAGE_SIZE;
1164                 break;
1165         default:
1166                 ret = -EINVAL;
1167                 break;
1168         }
1169
1170         __pipe_unlock(pipe);
1171         return ret;
1172 }
1173
1174 static const struct super_operations pipefs_ops = {
1175         .destroy_inode = free_inode_nonrcu,
1176         .statfs = simple_statfs,
1177 };
1178
1179 /*
1180  * pipefs should _never_ be mounted by userland - too much of security hassle,
1181  * no real gain from having the whole whorehouse mounted. So we don't need
1182  * any operations on the root directory. However, we need a non-trivial
1183  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1184  */
1185 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1186                          int flags, const char *dev_name, void *data)
1187 {
1188         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1189                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1190 }
1191
1192 static struct file_system_type pipe_fs_type = {
1193         .name           = "pipefs",
1194         .mount          = pipefs_mount,
1195         .kill_sb        = kill_anon_super,
1196 };
1197
1198 static int __init init_pipe_fs(void)
1199 {
1200         int err = register_filesystem(&pipe_fs_type);
1201
1202         if (!err) {
1203                 pipe_mnt = kern_mount(&pipe_fs_type);
1204                 if (IS_ERR(pipe_mnt)) {
1205                         err = PTR_ERR(pipe_mnt);
1206                         unregister_filesystem(&pipe_fs_type);
1207                 }
1208         }
1209         return err;
1210 }
1211
1212 fs_initcall(init_pipe_fs);