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