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  * New pipe buffers will be restricted to this size while the user is exceeding
  35  * their pipe buffer quota. The general pipe use case needs at least two
  36  * buffers: one for data yet to be read, and one for new data. If this is less
  37  * than two, then a write to a non-empty pipe may block even if the pipe is not
  38  * full. This can occur with GNU make jobserver or similar uses of pipes as
  39  * semaphores: multiple processes may be waiting to write tokens back to the
  40  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  41  *
  42  * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  43  * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  44  * emptied.
  45  */
  46 #define PIPE_MIN_DEF_BUFFERS 2
  47
  48 /*
  49  * The max size that a non-root user is allowed to grow the pipe. Can
  50  * be set by root in /proc/sys/fs/pipe-max-size
  51  */
  52 unsigned int pipe_max_size = 1048576;
  53
  54 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
  55  * matches default values.
  56  */
  57 unsigned long pipe_user_pages_hard;
  58 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  59
  60 /*
  61  * We use a start+len construction, which provides full use of the
  62  * allocated memory.
  63  * -- Florian Coosmann (FGC)
  64  *
  65  * Reads with count = 0 should always return 0.
  66  * -- Julian Bradfield 1999-06-07.
  67  *
  68  * FIFOs and Pipes now generate SIGIO for both readers and writers.
  69  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  70  *
  71  * pipe_read & write cleanup
  72  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  73  */
  74
  75 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  76 {
  77         if (pipe->files)
  78                 mutex_lock_nested(&pipe->mutex, subclass);
  79 }
  80
  81 void pipe_lock(struct pipe_inode_info *pipe)
  82 {
  83         /*
  84          * pipe_lock() nests non-pipe inode locks (for writing to a file)
  85          */
  86         pipe_lock_nested(pipe, I_MUTEX_PARENT);
  87 }
  88 EXPORT_SYMBOL(pipe_lock);
  89
  90 void pipe_unlock(struct pipe_inode_info *pipe)
  91 {
  92         if (pipe->files)
  93                 mutex_unlock(&pipe->mutex);
  94 }
  95 EXPORT_SYMBOL(pipe_unlock);
  96
  97 static inline void __pipe_lock(struct pipe_inode_info *pipe)
  98 {
  99         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
 100 }
 101
 102 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 103 {
 104         mutex_unlock(&pipe->mutex);
 105 }
 106
 107 void pipe_double_lock(struct pipe_inode_info *pipe1,
 108                       struct pipe_inode_info *pipe2)
 109 {
 110         BUG_ON(pipe1 == pipe2);
 111
 112         if (pipe1 < pipe2) {
 113                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
 114                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
 115         } else {
 116                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
 117                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
 118         }
 119 }
 120
 121 /* Drop the inode semaphore and wait for a pipe event, atomically */
 122 void pipe_wait(struct pipe_inode_info *pipe)
 123 {
 124         DEFINE_WAIT(wait);
 125
 126         /*
 127          * Pipes are system-local resources, so sleeping on them
 128          * is considered a noninteractive wait:
 129          */
 130         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
 131         pipe_unlock(pipe);
 132         schedule();
 133         finish_wait(&pipe->wait, &wait);
 134         pipe_lock(pipe);
 135 }
 136
 137 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 138                                   struct pipe_buffer *buf)
 139 {
 140         struct page *page = buf->page;
 141
 142         /*
 143          * If nobody else uses this page, and we don't already have a
 144          * temporary page, let's keep track of it as a one-deep
 145          * allocation cache. (Otherwise just release our reference to it)
 146          */
 147         if (page_count(page) == 1 && !pipe->tmp_page)
 148                 pipe->tmp_page = page;
 149         else
 150                 put_page(page);
 151 }
 152
 153 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
 154                                struct pipe_buffer *buf)
 155 {
 156         struct page *page = buf->page;
 157
 158         if (page_count(page) == 1) {
 159                 memcg_kmem_uncharge(page, 0);
 160                 __SetPageLocked(page);
 161                 return 0;
 162         }
 163         return 1;
 164 }
 165
 166 /**
 167  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 168  * @pipe:       the pipe that the buffer belongs to
 169  * @buf:        the buffer to attempt to steal
 170  *
 171  * Description:
 172  *      This function attempts to steal the &struct page attached to
 173  *      @buf. If successful, this function returns 0 and returns with
 174  *      the page locked. The caller may then reuse the page for whatever
 175  *      he wishes; the typical use is insertion into a different file
 176  *      page cache.
 177  */
 178 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 179                            struct pipe_buffer *buf)
 180 {
 181         struct page *page = buf->page;
 182
 183         /*
 184          * A reference of one is golden, that means that the owner of this
 185          * page is the only one holding a reference to it. lock the page
 186          * and return OK.
 187          */
 188         if (page_count(page) == 1) {
 189                 lock_page(page);
 190                 return 0;
 191         }
 192
 193         return 1;
 194 }
 195 EXPORT_SYMBOL(generic_pipe_buf_steal);
 196
 197 /**
 198  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 199  * @pipe:       the pipe that the buffer belongs to
 200  * @buf:        the buffer to get a reference to
 201  *
 202  * Description:
 203  *      This function grabs an extra reference to @buf. It's used in
 204  *      in the tee() system call, when we duplicate the buffers in one
 205  *      pipe into another.
 206  */
 207 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 208 {
 209         return try_get_page(buf->page);
 210 }
 211 EXPORT_SYMBOL(generic_pipe_buf_get);
 212
 213 /**
 214  * generic_pipe_buf_confirm - verify contents of the pipe buffer
 215  * @info:       the pipe that the buffer belongs to
 216  * @buf:        the buffer to confirm
 217  *
 218  * Description:
 219  *      This function does nothing, because the generic pipe code uses
 220  *      pages that are always good when inserted into the pipe.
 221  */
 222 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 223                              struct pipe_buffer *buf)
 224 {
 225         return 0;
 226 }
 227 EXPORT_SYMBOL(generic_pipe_buf_confirm);
 228
 229 /**
 230  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 231  * @pipe:       the pipe that the buffer belongs to
 232  * @buf:        the buffer to put a reference to
 233  *
 234  * Description:
 235  *      This function releases a reference to @buf.
 236  */
 237 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 238                               struct pipe_buffer *buf)
 239 {
 240         put_page(buf->page);
 241 }
 242 EXPORT_SYMBOL(generic_pipe_buf_release);
 243
 244 /* New data written to a pipe may be appended to a buffer with this type. */
 245 static const struct pipe_buf_operations anon_pipe_buf_ops = {
 246         .confirm = generic_pipe_buf_confirm,
 247         .release = anon_pipe_buf_release,
 248         .steal = anon_pipe_buf_steal,
 249         .get = generic_pipe_buf_get,
 250 };
 251
 252 static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
 253         .confirm = generic_pipe_buf_confirm,
 254         .release = anon_pipe_buf_release,
 255         .steal = anon_pipe_buf_steal,
 256         .get = generic_pipe_buf_get,
 257 };
 258
 259 static const struct pipe_buf_operations packet_pipe_buf_ops = {
 260         .confirm = generic_pipe_buf_confirm,
 261         .release = anon_pipe_buf_release,
 262         .steal = anon_pipe_buf_steal,
 263         .get = generic_pipe_buf_get,
 264 };
 265
 266 /**
 267  * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
 268  * @buf:        the buffer to mark
 269  *
 270  * Description:
 271  *      This function ensures that no future writes will be merged into the
 272  *      given &struct pipe_buffer. This is necessary when multiple pipe buffers
 273  *      share the same backing page.
 274  */
 275 void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
 276 {
 277         if (buf->ops == &anon_pipe_buf_ops)
 278                 buf->ops = &anon_pipe_buf_nomerge_ops;
 279 }
 280
 281 static bool pipe_buf_can_merge(struct pipe_buffer *buf)
 282 {
 283         return buf->ops == &anon_pipe_buf_ops;
 284 }
 285
 286 static ssize_t
 287 pipe_read(struct kiocb *iocb, struct iov_iter *to)
 288 {
 289         size_t total_len = iov_iter_count(to);
 290         struct file *filp = iocb->ki_filp;
 291         struct pipe_inode_info *pipe = filp->private_data;
 292         int do_wakeup;
 293         ssize_t ret;
 294
 295         /* Null read succeeds. */
 296         if (unlikely(total_len == 0))
 297                 return 0;
 298
 299         do_wakeup = 0;
 300         ret = 0;
 301         __pipe_lock(pipe);
 302         for (;;) {
 303                 int bufs = pipe->nrbufs;
 304                 if (bufs) {
 305                         int curbuf = pipe->curbuf;
 306                         struct pipe_buffer *buf = pipe->bufs + curbuf;
 307                         size_t chars = buf->len;
 308                         size_t written;
 309                         int error;
 310
 311                         if (chars > total_len)
 312                                 chars = total_len;
 313
 314                         error = pipe_buf_confirm(pipe, buf);
 315                         if (error) {
 316                                 if (!ret)
 317                                         ret = error;
 318                                 break;
 319                         }
 320
 321                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 322                         if (unlikely(written < chars)) {
 323                                 if (!ret)
 324                                         ret = -EFAULT;
 325                                 break;
 326                         }
 327                         ret += chars;
 328                         buf->offset += chars;
 329                         buf->len -= chars;
 330
 331                         /* Was it a packet buffer? Clean up and exit */
 332                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 333                                 total_len = chars;
 334                                 buf->len = 0;
 335                         }
 336
 337                         if (!buf->len) {
 338                                 pipe_buf_release(pipe, buf);
 339                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
 340                                 pipe->curbuf = curbuf;
 341                                 pipe->nrbufs = --bufs;
 342                                 do_wakeup = 1;
 343                         }
 344                         total_len -= chars;
 345                         if (!total_len)
 346                                 break;  /* common path: read succeeded */
 347                 }
 348                 if (bufs)       /* More to do? */
 349                         continue;
 350                 if (!pipe->writers)
 351                         break;
 352                 if (!pipe->waiting_writers) {
 353                         /* syscall merging: Usually we must not sleep
 354                          * if O_NONBLOCK is set, or if we got some data.
 355                          * But if a writer sleeps in kernel space, then
 356                          * we can wait for that data without violating POSIX.
 357                          */
 358                         if (ret)
 359                                 break;
 360                         if (filp->f_flags & O_NONBLOCK) {
 361                                 ret = -EAGAIN;
 362                                 break;
 363                         }
 364                 }
 365                 if (signal_pending(current)) {
 366                         if (!ret)
 367                                 ret = -ERESTARTSYS;
 368                         break;
 369                 }
 370                 if (do_wakeup) {
 371                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 372                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 373                 }
 374                 pipe_wait(pipe);
 375         }
 376         __pipe_unlock(pipe);
 377
 378         /* Signal writers asynchronously that there is more room. */
 379         if (do_wakeup) {
 380                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
 381                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 382         }
 383         if (ret > 0)
 384                 file_accessed(filp);
 385         return ret;
 386 }
 387
 388 static inline int is_packetized(struct file *file)
 389 {
 390         return (file->f_flags & O_DIRECT) != 0;
 391 }
 392
 393 static ssize_t
 394 pipe_write(struct kiocb *iocb, struct iov_iter *from)
 395 {
 396         struct file *filp = iocb->ki_filp;
 397         struct pipe_inode_info *pipe = filp->private_data;
 398         ssize_t ret = 0;
 399         int do_wakeup = 0;
 400         size_t total_len = iov_iter_count(from);
 401         ssize_t chars;
 402
 403         /* Null write succeeds. */
 404         if (unlikely(total_len == 0))
 405                 return 0;
 406
 407         __pipe_lock(pipe);
 408
 409         if (!pipe->readers) {
 410                 send_sig(SIGPIPE, current, 0);
 411                 ret = -EPIPE;
 412                 goto out;
 413         }
 414
 415         /* We try to merge small writes */
 416         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 417         if (pipe->nrbufs && chars != 0) {
 418                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 419                                                         (pipe->buffers - 1);
 420                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
 421                 int offset = buf->offset + buf->len;
 422
 423                 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
 424                         ret = pipe_buf_confirm(pipe, buf);
 425                         if (ret)
 426                                 goto out;
 427
 428                         ret = copy_page_from_iter(buf->page, offset, chars, from);
 429                         if (unlikely(ret < chars)) {
 430                                 ret = -EFAULT;
 431                                 goto out;
 432                         }
 433                         do_wakeup = 1;
 434                         buf->len += ret;
 435                         if (!iov_iter_count(from))
 436                                 goto out;
 437                 }
 438         }
 439
 440         for (;;) {
 441                 int bufs;
 442
 443                 if (!pipe->readers) {
 444                         send_sig(SIGPIPE, current, 0);
 445                         if (!ret)
 446                                 ret = -EPIPE;
 447                         break;
 448                 }
 449                 bufs = pipe->nrbufs;
 450                 if (bufs < pipe->buffers) {
 451                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 452                         struct pipe_buffer *buf = pipe->bufs + newbuf;
 453                         struct page *page = pipe->tmp_page;
 454                         int copied;
 455
 456                         if (!page) {
 457                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 458                                 if (unlikely(!page)) {
 459                                         ret = ret ? : -ENOMEM;
 460                                         break;
 461                                 }
 462                                 pipe->tmp_page = page;
 463                         }
 464                         /* Always wake up, even if the copy fails. Otherwise
 465                          * we lock up (O_NONBLOCK-)readers that sleep due to
 466                          * syscall merging.
 467                          * FIXME! Is this really true?
 468                          */
 469                         do_wakeup = 1;
 470                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 471                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 472                                 if (!ret)
 473                                         ret = -EFAULT;
 474                                 break;
 475                         }
 476                         ret += copied;
 477
 478                         /* Insert it into the buffer array */
 479                         buf->page = page;
 480                         buf->ops = &anon_pipe_buf_ops;
 481                         buf->offset = 0;
 482                         buf->len = copied;
 483                         buf->flags = 0;
 484                         if (is_packetized(filp)) {
 485                                 buf->ops = &packet_pipe_buf_ops;
 486                                 buf->flags = PIPE_BUF_FLAG_PACKET;
 487                         }
 488                         pipe->nrbufs = ++bufs;
 489                         pipe->tmp_page = NULL;
 490
 491                         if (!iov_iter_count(from))
 492                                 break;
 493                 }
 494                 if (bufs < pipe->buffers)
 495                         continue;
 496                 if (filp->f_flags & O_NONBLOCK) {
 497                         if (!ret)
 498                                 ret = -EAGAIN;
 499                         break;
 500                 }
 501                 if (signal_pending(current)) {
 502                         if (!ret)
 503                                 ret = -ERESTARTSYS;
 504                         break;
 505                 }
 506                 if (do_wakeup) {
 507                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 508                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 509                         do_wakeup = 0;
 510                 }
 511                 pipe->waiting_writers++;
 512                 pipe_wait(pipe);
 513                 pipe->waiting_writers--;
 514         }
 515 out:
 516         __pipe_unlock(pipe);
 517         if (do_wakeup) {
 518                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
 519                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 520         }
 521         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 522                 int err = file_update_time(filp);
 523                 if (err)
 524                         ret = err;
 525                 sb_end_write(file_inode(filp)->i_sb);
 526         }
 527         return ret;
 528 }
 529
 530 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 531 {
 532         struct pipe_inode_info *pipe = filp->private_data;
 533         int count, buf, nrbufs;
 534
 535         switch (cmd) {
 536                 case FIONREAD:
 537                         __pipe_lock(pipe);
 538                         count = 0;
 539                         buf = pipe->curbuf;
 540                         nrbufs = pipe->nrbufs;
 541                         while (--nrbufs >= 0) {
 542                                 count += pipe->bufs[buf].len;
 543                                 buf = (buf+1) & (pipe->buffers - 1);
 544                         }
 545                         __pipe_unlock(pipe);
 546
 547                         return put_user(count, (int __user *)arg);
 548                 default:
 549                         return -ENOIOCTLCMD;
 550         }
 551 }
 552
 553 /* No kernel lock held - fine */
 554 static __poll_t
 555 pipe_poll(struct file *filp, poll_table *wait)
 556 {
 557         __poll_t mask;
 558         struct pipe_inode_info *pipe = filp->private_data;
 559         int nrbufs;
 560
 561         poll_wait(filp, &pipe->wait, wait);
 562
 563         /* Reading only -- no need for acquiring the semaphore.  */
 564         nrbufs = pipe->nrbufs;
 565         mask = 0;
 566         if (filp->f_mode & FMODE_READ) {
 567                 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
 568                 if (!pipe->writers && filp->f_version != pipe->w_counter)
 569                         mask |= EPOLLHUP;
 570         }
 571
 572         if (filp->f_mode & FMODE_WRITE) {
 573                 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
 574                 /*
 575                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 576                  * behave exactly like pipes for poll().
 577                  */
 578                 if (!pipe->readers)
 579                         mask |= EPOLLERR;
 580         }
 581
 582         return mask;
 583 }
 584
 585 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 586 {
 587         int kill = 0;
 588
 589         spin_lock(&inode->i_lock);
 590         if (!--pipe->files) {
 591                 inode->i_pipe = NULL;
 592                 kill = 1;
 593         }
 594         spin_unlock(&inode->i_lock);
 595
 596         if (kill)
 597                 free_pipe_info(pipe);
 598 }
 599
 600 static int
 601 pipe_release(struct inode *inode, struct file *file)
 602 {
 603         struct pipe_inode_info *pipe = file->private_data;
 604
 605         __pipe_lock(pipe);
 606         if (file->f_mode & FMODE_READ)
 607                 pipe->readers--;
 608         if (file->f_mode & FMODE_WRITE)
 609                 pipe->writers--;
 610
 611         if (pipe->readers || pipe->writers) {
 612                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
 613                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 614                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 615         }
 616         __pipe_unlock(pipe);
 617
 618         put_pipe_info(inode, pipe);
 619         return 0;
 620 }
 621
 622 static int
 623 pipe_fasync(int fd, struct file *filp, int on)
 624 {
 625         struct pipe_inode_info *pipe = filp->private_data;
 626         int retval = 0;
 627
 628         __pipe_lock(pipe);
 629         if (filp->f_mode & FMODE_READ)
 630                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 631         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 632                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 633                 if (retval < 0 && (filp->f_mode & FMODE_READ))
 634                         /* this can happen only if on == T */
 635                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 636         }
 637         __pipe_unlock(pipe);
 638         return retval;
 639 }
 640
 641 static unsigned long account_pipe_buffers(struct user_struct *user,
 642                                  unsigned long old, unsigned long new)
 643 {
 644         return atomic_long_add_return(new - old, &user->pipe_bufs);
 645 }
 646
 647 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 648 {
 649         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 650
 651         return soft_limit && user_bufs > soft_limit;
 652 }
 653
 654 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 655 {
 656         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 657
 658         return hard_limit && user_bufs > hard_limit;
 659 }
 660
 661 static bool is_unprivileged_user(void)
 662 {
 663         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 664 }
 665
 666 struct pipe_inode_info *alloc_pipe_info(void)
 667 {
 668         struct pipe_inode_info *pipe;
 669         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 670         struct user_struct *user = get_current_user();
 671         unsigned long user_bufs;
 672         unsigned int max_size = READ_ONCE(pipe_max_size);
 673
 674         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 675         if (pipe == NULL)
 676                 goto out_free_uid;
 677
 678         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 679                 pipe_bufs = max_size >> PAGE_SHIFT;
 680
 681         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 682
 683         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
 684                 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 685                 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 686         }
 687
 688         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
 689                 goto out_revert_acct;
 690
 691         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 692                              GFP_KERNEL_ACCOUNT);
 693
 694         if (pipe->bufs) {
 695                 init_waitqueue_head(&pipe->wait);
 696                 pipe->r_counter = pipe->w_counter = 1;
 697                 pipe->buffers = pipe_bufs;
 698                 pipe->user = user;
 699                 mutex_init(&pipe->mutex);
 700                 return pipe;
 701         }
 702
 703 out_revert_acct:
 704         (void) account_pipe_buffers(user, pipe_bufs, 0);
 705         kfree(pipe);
 706 out_free_uid:
 707         free_uid(user);
 708         return NULL;
 709 }
 710
 711 void free_pipe_info(struct pipe_inode_info *pipe)
 712 {
 713         int i;
 714
 715         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
 716         free_uid(pipe->user);
 717         for (i = 0; i < pipe->buffers; i++) {
 718                 struct pipe_buffer *buf = pipe->bufs + i;
 719                 if (buf->ops)
 720                         pipe_buf_release(pipe, buf);
 721         }
 722         if (pipe->tmp_page)
 723                 __free_page(pipe->tmp_page);
 724         kfree(pipe->bufs);
 725         kfree(pipe);
 726 }
 727
 728 static struct vfsmount *pipe_mnt __read_mostly;
 729
 730 /*
 731  * pipefs_dname() is called from d_path().
 732  */
 733 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 734 {
 735         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 736                                 d_inode(dentry)->i_ino);
 737 }
 738
 739 static const struct dentry_operations pipefs_dentry_operations = {
 740         .d_dname        = pipefs_dname,
 741 };
 742
 743 static struct inode * get_pipe_inode(void)
 744 {
 745         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 746         struct pipe_inode_info *pipe;
 747
 748         if (!inode)
 749                 goto fail_inode;
 750
 751         inode->i_ino = get_next_ino();
 752
 753         pipe = alloc_pipe_info();
 754         if (!pipe)
 755                 goto fail_iput;
 756
 757         inode->i_pipe = pipe;
 758         pipe->files = 2;
 759         pipe->readers = pipe->writers = 1;
 760         inode->i_fop = &pipefifo_fops;
 761
 762         /*
 763          * Mark the inode dirty from the very beginning,
 764          * that way it will never be moved to the dirty
 765          * list because "mark_inode_dirty()" will think
 766          * that it already _is_ on the dirty list.
 767          */
 768         inode->i_state = I_DIRTY;
 769         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 770         inode->i_uid = current_fsuid();
 771         inode->i_gid = current_fsgid();
 772         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 773
 774         return inode;
 775
 776 fail_iput:
 777         iput(inode);
 778
 779 fail_inode:
 780         return NULL;
 781 }
 782
 783 int create_pipe_files(struct file **res, int flags)
 784 {
 785         struct inode *inode = get_pipe_inode();
 786         struct file *f;
 787
 788         if (!inode)
 789                 return -ENFILE;
 790
 791         f = alloc_file_pseudo(inode, pipe_mnt, "",
 792                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 793                                 &pipefifo_fops);
 794         if (IS_ERR(f)) {
 795                 free_pipe_info(inode->i_pipe);
 796                 iput(inode);
 797                 return PTR_ERR(f);
 798         }
 799
 800         f->private_data = inode->i_pipe;
 801
 802         res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 803                                   &pipefifo_fops);
 804         if (IS_ERR(res[0])) {
 805                 put_pipe_info(inode, inode->i_pipe);
 806                 fput(f);
 807                 return PTR_ERR(res[0]);
 808         }
 809         res[0]->private_data = inode->i_pipe;
 810         res[1] = f;
 811         return 0;
 812 }
 813
 814 static int __do_pipe_flags(int *fd, struct file **files, int flags)
 815 {
 816         int error;
 817         int fdw, fdr;
 818
 819         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
 820                 return -EINVAL;
 821
 822         error = create_pipe_files(files, flags);
 823         if (error)
 824                 return error;
 825
 826         error = get_unused_fd_flags(flags);
 827         if (error < 0)
 828                 goto err_read_pipe;
 829         fdr = error;
 830
 831         error = get_unused_fd_flags(flags);
 832         if (error < 0)
 833                 goto err_fdr;
 834         fdw = error;
 835
 836         audit_fd_pair(fdr, fdw);
 837         fd[0] = fdr;
 838         fd[1] = fdw;
 839         return 0;
 840
 841  err_fdr:
 842         put_unused_fd(fdr);
 843  err_read_pipe:
 844         fput(files[0]);
 845         fput(files[1]);
 846         return error;
 847 }
 848
 849 int do_pipe_flags(int *fd, int flags)
 850 {
 851         struct file *files[2];
 852         int error = __do_pipe_flags(fd, files, flags);
 853         if (!error) {
 854                 fd_install(fd[0], files[0]);
 855                 fd_install(fd[1], files[1]);
 856         }
 857         return error;
 858 }
 859
 860 /*
 861  * sys_pipe() is the normal C calling standard for creating
 862  * a pipe. It's not the way Unix traditionally does this, though.
 863  */
 864 static int do_pipe2(int __user *fildes, int flags)
 865 {
 866         struct file *files[2];
 867         int fd[2];
 868         int error;
 869
 870         error = __do_pipe_flags(fd, files, flags);
 871         if (!error) {
 872                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
 873                         fput(files[0]);
 874                         fput(files[1]);
 875                         put_unused_fd(fd[0]);
 876                         put_unused_fd(fd[1]);
 877                         error = -EFAULT;
 878                 } else {
 879                         fd_install(fd[0], files[0]);
 880                         fd_install(fd[1], files[1]);
 881                 }
 882         }
 883         return error;
 884 }
 885
 886 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
 887 {
 888         return do_pipe2(fildes, flags);
 889 }
 890
 891 SYSCALL_DEFINE1(pipe, int __user *, fildes)
 892 {
 893         return do_pipe2(fildes, 0);
 894 }
 895
 896 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
 897 {
 898         int cur = *cnt;
 899
 900         while (cur == *cnt) {
 901                 pipe_wait(pipe);
 902                 if (signal_pending(current))
 903                         break;
 904         }
 905         return cur == *cnt ? -ERESTARTSYS : 0;
 906 }
 907
 908 static void wake_up_partner(struct pipe_inode_info *pipe)
 909 {
 910         wake_up_interruptible(&pipe->wait);
 911 }
 912
 913 static int fifo_open(struct inode *inode, struct file *filp)
 914 {
 915         struct pipe_inode_info *pipe;
 916         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
 917         int ret;
 918
 919         filp->f_version = 0;
 920
 921         spin_lock(&inode->i_lock);
 922         if (inode->i_pipe) {
 923                 pipe = inode->i_pipe;
 924                 pipe->files++;
 925                 spin_unlock(&inode->i_lock);
 926         } else {
 927                 spin_unlock(&inode->i_lock);
 928                 pipe = alloc_pipe_info();
 929                 if (!pipe)
 930                         return -ENOMEM;
 931                 pipe->files = 1;
 932                 spin_lock(&inode->i_lock);
 933                 if (unlikely(inode->i_pipe)) {
 934                         inode->i_pipe->files++;
 935                         spin_unlock(&inode->i_lock);
 936                         free_pipe_info(pipe);
 937                         pipe = inode->i_pipe;
 938                 } else {
 939                         inode->i_pipe = pipe;
 940                         spin_unlock(&inode->i_lock);
 941                 }
 942         }
 943         filp->private_data = pipe;
 944         /* OK, we have a pipe and it's pinned down */
 945
 946         __pipe_lock(pipe);
 947
 948         /* We can only do regular read/write on fifos */
 949         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
 950
 951         switch (filp->f_mode) {
 952         case FMODE_READ:
 953         /*
 954          *  O_RDONLY
 955          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
 956          *  opened, even when there is no process writing the FIFO.
 957          */
 958                 pipe->r_counter++;
 959                 if (pipe->readers++ == 0)
 960                         wake_up_partner(pipe);
 961
 962                 if (!is_pipe && !pipe->writers) {
 963                         if ((filp->f_flags & O_NONBLOCK)) {
 964                                 /* suppress EPOLLHUP until we have
 965                                  * seen a writer */
 966                                 filp->f_version = pipe->w_counter;
 967                         } else {
 968                                 if (wait_for_partner(pipe, &pipe->w_counter))
 969                                         goto err_rd;
 970                         }
 971                 }
 972                 break;
 973
 974         case FMODE_WRITE:
 975         /*
 976          *  O_WRONLY
 977          *  POSIX.1 says that O_NONBLOCK means return -1 with
 978          *  errno=ENXIO when there is no process reading the FIFO.
 979          */
 980                 ret = -ENXIO;
 981                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
 982                         goto err;
 983
 984                 pipe->w_counter++;
 985                 if (!pipe->writers++)
 986                         wake_up_partner(pipe);
 987
 988                 if (!is_pipe && !pipe->readers) {
 989                         if (wait_for_partner(pipe, &pipe->r_counter))
 990                                 goto err_wr;
 991                 }
 992                 break;
 993
 994         case FMODE_READ | FMODE_WRITE:
 995         /*
 996          *  O_RDWR
 997          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
 998          *  This implementation will NEVER block on a O_RDWR open, since
 999          *  the process can at least talk to itself.
1000          */
1001
1002                 pipe->readers++;
1003                 pipe->writers++;
1004                 pipe->r_counter++;
1005                 pipe->w_counter++;
1006                 if (pipe->readers == 1 || pipe->writers == 1)
1007                         wake_up_partner(pipe);
1008                 break;
1009
1010         default:
1011                 ret = -EINVAL;
1012                 goto err;
1013         }
1014
1015         /* Ok! */
1016         __pipe_unlock(pipe);
1017         return 0;
1018
1019 err_rd:
1020         if (!--pipe->readers)
1021                 wake_up_interruptible(&pipe->wait);
1022         ret = -ERESTARTSYS;
1023         goto err;
1024
1025 err_wr:
1026         if (!--pipe->writers)
1027                 wake_up_interruptible(&pipe->wait);
1028         ret = -ERESTARTSYS;
1029         goto err;
1030
1031 err:
1032         __pipe_unlock(pipe);
1033
1034         put_pipe_info(inode, pipe);
1035         return ret;
1036 }
1037
1038 const struct file_operations pipefifo_fops = {
1039         .open           = fifo_open,
1040         .llseek         = no_llseek,
1041         .read_iter      = pipe_read,
1042         .write_iter     = pipe_write,
1043         .poll           = pipe_poll,
1044         .unlocked_ioctl = pipe_ioctl,
1045         .release        = pipe_release,
1046         .fasync         = pipe_fasync,
1047 };
1048
1049 /*
1050  * Currently we rely on the pipe array holding a power-of-2 number
1051  * of pages. Returns 0 on error.
1052  */
1053 unsigned int round_pipe_size(unsigned long size)
1054 {
1055         if (size > (1U << 31))
1056                 return 0;
1057
1058         /* Minimum pipe size, as required by POSIX */
1059         if (size < PAGE_SIZE)
1060                 return PAGE_SIZE;
1061
1062         return roundup_pow_of_two(size);
1063 }
1064
1065 /*
1066  * Allocate a new array of pipe buffers and copy the info over. Returns the
1067  * pipe size if successful, or return -ERROR on error.
1068  */
1069 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1070 {
1071         struct pipe_buffer *bufs;
1072         unsigned int size, nr_pages;
1073         unsigned long user_bufs;
1074         long ret = 0;
1075
1076         size = round_pipe_size(arg);
1077         nr_pages = size >> PAGE_SHIFT;
1078
1079         if (!nr_pages)
1080                 return -EINVAL;
1081
1082         /*
1083          * If trying to increase the pipe capacity, check that an
1084          * unprivileged user is not trying to exceed various limits
1085          * (soft limit check here, hard limit check just below).
1086          * Decreasing the pipe capacity is always permitted, even
1087          * if the user is currently over a limit.
1088          */
1089         if (nr_pages > pipe->buffers &&
1090                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1091                 return -EPERM;
1092
1093         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1094
1095         if (nr_pages > pipe->buffers &&
1096                         (too_many_pipe_buffers_hard(user_bufs) ||
1097                          too_many_pipe_buffers_soft(user_bufs)) &&
1098                         is_unprivileged_user()) {
1099                 ret = -EPERM;
1100                 goto out_revert_acct;
1101         }
1102
1103         /*
1104          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1105          * expect a lot of shrink+grow operations, just free and allocate
1106          * again like we would do for growing. If the pipe currently
1107          * contains more buffers than arg, then return busy.
1108          */
1109         if (nr_pages < pipe->nrbufs) {
1110                 ret = -EBUSY;
1111                 goto out_revert_acct;
1112         }
1113
1114         bufs = kcalloc(nr_pages, sizeof(*bufs),
1115                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1116         if (unlikely(!bufs)) {
1117                 ret = -ENOMEM;
1118                 goto out_revert_acct;
1119         }
1120
1121         /*
1122          * The pipe array wraps around, so just start the new one at zero
1123          * and adjust the indexes.
1124          */
1125         if (pipe->nrbufs) {
1126                 unsigned int tail;
1127                 unsigned int head;
1128
1129                 tail = pipe->curbuf + pipe->nrbufs;
1130                 if (tail < pipe->buffers)
1131                         tail = 0;
1132                 else
1133                         tail &= (pipe->buffers - 1);
1134
1135                 head = pipe->nrbufs - tail;
1136                 if (head)
1137                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1138                 if (tail)
1139                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1140         }
1141
1142         pipe->curbuf = 0;
1143         kfree(pipe->bufs);
1144         pipe->bufs = bufs;
1145         pipe->buffers = nr_pages;
1146         return nr_pages * PAGE_SIZE;
1147
1148 out_revert_acct:
1149         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1150         return ret;
1151 }
1152
1153 /*
1154  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1155  * location, so checking ->i_pipe is not enough to verify that this is a
1156  * pipe.
1157  */
1158 struct pipe_inode_info *get_pipe_info(struct file *file)
1159 {
1160         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1161 }
1162
1163 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1164 {
1165         struct pipe_inode_info *pipe;
1166         long ret;
1167
1168         pipe = get_pipe_info(file);
1169         if (!pipe)
1170                 return -EBADF;
1171
1172         __pipe_lock(pipe);
1173
1174         switch (cmd) {
1175         case F_SETPIPE_SZ:
1176                 ret = pipe_set_size(pipe, arg);
1177                 break;
1178         case F_GETPIPE_SZ:
1179                 ret = pipe->buffers * PAGE_SIZE;
1180                 break;
1181         default:
1182                 ret = -EINVAL;
1183                 break;
1184         }
1185
1186         __pipe_unlock(pipe);
1187         return ret;
1188 }
1189
1190 static const struct super_operations pipefs_ops = {
1191         .destroy_inode = free_inode_nonrcu,
1192         .statfs = simple_statfs,
1193 };
1194
1195 /*
1196  * pipefs should _never_ be mounted by userland - too much of security hassle,
1197  * no real gain from having the whole whorehouse mounted. So we don't need
1198  * any operations on the root directory. However, we need a non-trivial
1199  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1200  */
1201
1202 static int pipefs_init_fs_context(struct fs_context *fc)
1203 {
1204         struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1205         if (!ctx)
1206                 return -ENOMEM;
1207         ctx->ops = &pipefs_ops;
1208         ctx->dops = &pipefs_dentry_operations;
1209         return 0;
1210 }
1211
1212 static struct file_system_type pipe_fs_type = {
1213         .name           = "pipefs",
1214         .init_fs_context = pipefs_init_fs_context,
1215         .kill_sb        = kill_anon_super,
1216 };
1217
1218 static int __init init_pipe_fs(void)
1219 {
1220         int err = register_filesystem(&pipe_fs_type);
1221
1222         if (!err) {
1223                 pipe_mnt = kern_mount(&pipe_fs_type);
1224                 if (IS_ERR(pipe_mnt)) {
1225                         err = PTR_ERR(pipe_mnt);
1226                         unregister_filesystem(&pipe_fs_type);
1227                 }
1228         }
1229         return err;
1230 }
1231
1232 fs_initcall(init_pipe_fs);