Linux Audio

Check our new training course

Loading...
   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#include <linux/watch_queue.h>
  28#include <linux/sysctl.h>
  29
  30#include <linux/uaccess.h>
  31#include <asm/ioctls.h>
  32
  33#include "internal.h"
  34
  35/*
  36 * New pipe buffers will be restricted to this size while the user is exceeding
  37 * their pipe buffer quota. The general pipe use case needs at least two
  38 * buffers: one for data yet to be read, and one for new data. If this is less
  39 * than two, then a write to a non-empty pipe may block even if the pipe is not
  40 * full. This can occur with GNU make jobserver or similar uses of pipes as
  41 * semaphores: multiple processes may be waiting to write tokens back to the
  42 * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
  43 *
  44 * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
  45 * own risk, namely: pipe writes to non-full pipes may block until the pipe is
  46 * emptied.
  47 */
  48#define PIPE_MIN_DEF_BUFFERS 2
  49
  50/*
  51 * The max size that a non-root user is allowed to grow the pipe. Can
  52 * be set by root in /proc/sys/fs/pipe-max-size
  53 */
  54static unsigned int pipe_max_size = 1048576;
  55
  56/* Maximum allocatable pages per user. Hard limit is unset by default, soft
  57 * matches default values.
  58 */
  59static unsigned long pipe_user_pages_hard;
  60static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
  61
  62/*
  63 * We use head and tail indices that aren't masked off, except at the point of
  64 * dereference, but rather they're allowed to wrap naturally.  This means there
  65 * isn't a dead spot in the buffer, but the ring has to be a power of two and
  66 * <= 2^31.
  67 * -- David Howells 2019-09-23.
  68 *
  69 * Reads with count = 0 should always return 0.
  70 * -- Julian Bradfield 1999-06-07.
  71 *
  72 * FIFOs and Pipes now generate SIGIO for both readers and writers.
  73 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  74 *
  75 * pipe_read & write cleanup
  76 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  77 */
  78
  79#define cmp_int(l, r)		((l > r) - (l < r))
  80
  81#ifdef CONFIG_PROVE_LOCKING
  82static int pipe_lock_cmp_fn(const struct lockdep_map *a,
  83			    const struct lockdep_map *b)
  84{
  85	return cmp_int((unsigned long) a, (unsigned long) b);
  86}
  87#endif
  88
  89void pipe_lock(struct pipe_inode_info *pipe)
  90{
  91	if (pipe->files)
  92		mutex_lock(&pipe->mutex);
  93}
  94EXPORT_SYMBOL(pipe_lock);
  95
  96void pipe_unlock(struct pipe_inode_info *pipe)
  97{
  98	if (pipe->files)
  99		mutex_unlock(&pipe->mutex);
 100}
 101EXPORT_SYMBOL(pipe_unlock);
 102
 103void pipe_double_lock(struct pipe_inode_info *pipe1,
 104		      struct pipe_inode_info *pipe2)
 105{
 106	BUG_ON(pipe1 == pipe2);
 107
 108	if (pipe1 > pipe2)
 109		swap(pipe1, pipe2);
 110
 111	pipe_lock(pipe1);
 112	pipe_lock(pipe2);
 113}
 114
 115static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 116				  struct pipe_buffer *buf)
 117{
 118	struct page *page = buf->page;
 119
 120	/*
 121	 * If nobody else uses this page, and we don't already have a
 122	 * temporary page, let's keep track of it as a one-deep
 123	 * allocation cache. (Otherwise just release our reference to it)
 124	 */
 125	if (page_count(page) == 1 && !pipe->tmp_page)
 126		pipe->tmp_page = page;
 127	else
 128		put_page(page);
 129}
 130
 131static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 132		struct pipe_buffer *buf)
 133{
 134	struct page *page = buf->page;
 135
 136	if (page_count(page) != 1)
 137		return false;
 138	memcg_kmem_uncharge_page(page, 0);
 139	__SetPageLocked(page);
 140	return true;
 141}
 142
 143/**
 144 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
 145 * @pipe:	the pipe that the buffer belongs to
 146 * @buf:	the buffer to attempt to steal
 147 *
 148 * Description:
 149 *	This function attempts to steal the &struct page attached to
 150 *	@buf. If successful, this function returns 0 and returns with
 151 *	the page locked. The caller may then reuse the page for whatever
 152 *	he wishes; the typical use is insertion into a different file
 153 *	page cache.
 154 */
 155bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
 156		struct pipe_buffer *buf)
 157{
 158	struct page *page = buf->page;
 159
 160	/*
 161	 * A reference of one is golden, that means that the owner of this
 162	 * page is the only one holding a reference to it. lock the page
 163	 * and return OK.
 164	 */
 165	if (page_count(page) == 1) {
 166		lock_page(page);
 167		return true;
 168	}
 169	return false;
 170}
 171EXPORT_SYMBOL(generic_pipe_buf_try_steal);
 172
 173/**
 174 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 175 * @pipe:	the pipe that the buffer belongs to
 176 * @buf:	the buffer to get a reference to
 177 *
 178 * Description:
 179 *	This function grabs an extra reference to @buf. It's used in
 180 *	the tee() system call, when we duplicate the buffers in one
 181 *	pipe into another.
 182 */
 183bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 184{
 185	return try_get_page(buf->page);
 186}
 187EXPORT_SYMBOL(generic_pipe_buf_get);
 188
 189/**
 190 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 191 * @pipe:	the pipe that the buffer belongs to
 192 * @buf:	the buffer to put a reference to
 193 *
 194 * Description:
 195 *	This function releases a reference to @buf.
 196 */
 197void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 198			      struct pipe_buffer *buf)
 199{
 200	put_page(buf->page);
 201}
 202EXPORT_SYMBOL(generic_pipe_buf_release);
 203
 204static const struct pipe_buf_operations anon_pipe_buf_ops = {
 205	.release	= anon_pipe_buf_release,
 206	.try_steal	= anon_pipe_buf_try_steal,
 207	.get		= generic_pipe_buf_get,
 208};
 209
 210/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 211static inline bool pipe_readable(const struct pipe_inode_info *pipe)
 212{
 213	unsigned int head = READ_ONCE(pipe->head);
 214	unsigned int tail = READ_ONCE(pipe->tail);
 215	unsigned int writers = READ_ONCE(pipe->writers);
 216
 217	return !pipe_empty(head, tail) || !writers;
 218}
 219
 220static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
 221					    struct pipe_buffer *buf,
 222					    unsigned int tail)
 223{
 224	pipe_buf_release(pipe, buf);
 225
 226	/*
 227	 * If the pipe has a watch_queue, we need additional protection
 228	 * by the spinlock because notifications get posted with only
 229	 * this spinlock, no mutex
 230	 */
 231	if (pipe_has_watch_queue(pipe)) {
 232		spin_lock_irq(&pipe->rd_wait.lock);
 233#ifdef CONFIG_WATCH_QUEUE
 234		if (buf->flags & PIPE_BUF_FLAG_LOSS)
 235			pipe->note_loss = true;
 236#endif
 237		pipe->tail = ++tail;
 238		spin_unlock_irq(&pipe->rd_wait.lock);
 239		return tail;
 240	}
 241
 242	/*
 243	 * Without a watch_queue, we can simply increment the tail
 244	 * without the spinlock - the mutex is enough.
 245	 */
 246	pipe->tail = ++tail;
 247	return tail;
 248}
 249
 250static ssize_t
 251pipe_read(struct kiocb *iocb, struct iov_iter *to)
 252{
 253	size_t total_len = iov_iter_count(to);
 254	struct file *filp = iocb->ki_filp;
 255	struct pipe_inode_info *pipe = filp->private_data;
 256	bool was_full, wake_next_reader = false;
 257	ssize_t ret;
 258
 259	/* Null read succeeds. */
 260	if (unlikely(total_len == 0))
 261		return 0;
 262
 263	ret = 0;
 264	mutex_lock(&pipe->mutex);
 265
 266	/*
 267	 * We only wake up writers if the pipe was full when we started
 268	 * reading in order to avoid unnecessary wakeups.
 269	 *
 270	 * But when we do wake up writers, we do so using a sync wakeup
 271	 * (WF_SYNC), because we want them to get going and generate more
 272	 * data for us.
 273	 */
 274	was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 275	for (;;) {
 276		/* Read ->head with a barrier vs post_one_notification() */
 277		unsigned int head = smp_load_acquire(&pipe->head);
 278		unsigned int tail = pipe->tail;
 279		unsigned int mask = pipe->ring_size - 1;
 280
 281#ifdef CONFIG_WATCH_QUEUE
 282		if (pipe->note_loss) {
 283			struct watch_notification n;
 284
 285			if (total_len < 8) {
 286				if (ret == 0)
 287					ret = -ENOBUFS;
 288				break;
 289			}
 290
 291			n.type = WATCH_TYPE_META;
 292			n.subtype = WATCH_META_LOSS_NOTIFICATION;
 293			n.info = watch_sizeof(n);
 294			if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
 295				if (ret == 0)
 296					ret = -EFAULT;
 297				break;
 298			}
 299			ret += sizeof(n);
 300			total_len -= sizeof(n);
 301			pipe->note_loss = false;
 302		}
 303#endif
 304
 305		if (!pipe_empty(head, tail)) {
 306			struct pipe_buffer *buf = &pipe->bufs[tail & mask];
 307			size_t chars = buf->len;
 308			size_t written;
 309			int error;
 310
 311			if (chars > total_len) {
 312				if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
 313					if (ret == 0)
 314						ret = -ENOBUFS;
 315					break;
 316				}
 317				chars = total_len;
 318			}
 319
 320			error = pipe_buf_confirm(pipe, buf);
 321			if (error) {
 322				if (!ret)
 323					ret = error;
 324				break;
 325			}
 326
 327			written = copy_page_to_iter(buf->page, buf->offset, chars, to);
 328			if (unlikely(written < chars)) {
 329				if (!ret)
 330					ret = -EFAULT;
 331				break;
 332			}
 333			ret += chars;
 334			buf->offset += chars;
 335			buf->len -= chars;
 336
 337			/* Was it a packet buffer? Clean up and exit */
 338			if (buf->flags & PIPE_BUF_FLAG_PACKET) {
 339				total_len = chars;
 340				buf->len = 0;
 341			}
 342
 343			if (!buf->len)
 344				tail = pipe_update_tail(pipe, buf, tail);
 345			total_len -= chars;
 346			if (!total_len)
 347				break;	/* common path: read succeeded */
 348			if (!pipe_empty(head, tail))	/* More to do? */
 349				continue;
 350		}
 351
 352		if (!pipe->writers)
 353			break;
 354		if (ret)
 355			break;
 356		if ((filp->f_flags & O_NONBLOCK) ||
 357		    (iocb->ki_flags & IOCB_NOWAIT)) {
 358			ret = -EAGAIN;
 359			break;
 360		}
 361		mutex_unlock(&pipe->mutex);
 362
 363		/*
 364		 * We only get here if we didn't actually read anything.
 365		 *
 366		 * However, we could have seen (and removed) a zero-sized
 367		 * pipe buffer, and might have made space in the buffers
 368		 * that way.
 369		 *
 370		 * You can't make zero-sized pipe buffers by doing an empty
 371		 * write (not even in packet mode), but they can happen if
 372		 * the writer gets an EFAULT when trying to fill a buffer
 373		 * that already got allocated and inserted in the buffer
 374		 * array.
 375		 *
 376		 * So we still need to wake up any pending writers in the
 377		 * _very_ unlikely case that the pipe was full, but we got
 378		 * no data.
 379		 */
 380		if (unlikely(was_full))
 381			wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 382		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 383
 384		/*
 385		 * But because we didn't read anything, at this point we can
 386		 * just return directly with -ERESTARTSYS if we're interrupted,
 387		 * since we've done any required wakeups and there's no need
 388		 * to mark anything accessed. And we've dropped the lock.
 389		 */
 390		if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
 391			return -ERESTARTSYS;
 392
 393		mutex_lock(&pipe->mutex);
 394		was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
 395		wake_next_reader = true;
 396	}
 397	if (pipe_empty(pipe->head, pipe->tail))
 398		wake_next_reader = false;
 399	mutex_unlock(&pipe->mutex);
 400
 401	if (was_full)
 402		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 403	if (wake_next_reader)
 404		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 405	kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 406	if (ret > 0)
 407		file_accessed(filp);
 408	return ret;
 409}
 410
 411static inline int is_packetized(struct file *file)
 412{
 413	return (file->f_flags & O_DIRECT) != 0;
 414}
 415
 416/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
 417static inline bool pipe_writable(const struct pipe_inode_info *pipe)
 418{
 419	unsigned int head = READ_ONCE(pipe->head);
 420	unsigned int tail = READ_ONCE(pipe->tail);
 421	unsigned int max_usage = READ_ONCE(pipe->max_usage);
 422
 423	return !pipe_full(head, tail, max_usage) ||
 424		!READ_ONCE(pipe->readers);
 425}
 426
 427static ssize_t
 428pipe_write(struct kiocb *iocb, struct iov_iter *from)
 429{
 430	struct file *filp = iocb->ki_filp;
 431	struct pipe_inode_info *pipe = filp->private_data;
 432	unsigned int head;
 433	ssize_t ret = 0;
 434	size_t total_len = iov_iter_count(from);
 435	ssize_t chars;
 436	bool was_empty = false;
 437	bool wake_next_writer = false;
 438
 439	/*
 440	 * Reject writing to watch queue pipes before the point where we lock
 441	 * the pipe.
 442	 * Otherwise, lockdep would be unhappy if the caller already has another
 443	 * pipe locked.
 444	 * If we had to support locking a normal pipe and a notification pipe at
 445	 * the same time, we could set up lockdep annotations for that, but
 446	 * since we don't actually need that, it's simpler to just bail here.
 447	 */
 448	if (pipe_has_watch_queue(pipe))
 449		return -EXDEV;
 450
 451	/* Null write succeeds. */
 452	if (unlikely(total_len == 0))
 453		return 0;
 454
 455	mutex_lock(&pipe->mutex);
 456
 457	if (!pipe->readers) {
 458		send_sig(SIGPIPE, current, 0);
 459		ret = -EPIPE;
 460		goto out;
 461	}
 462
 463	/*
 464	 * If it wasn't empty we try to merge new data into
 465	 * the last buffer.
 466	 *
 467	 * That naturally merges small writes, but it also
 468	 * page-aligns the rest of the writes for large writes
 469	 * spanning multiple pages.
 470	 */
 471	head = pipe->head;
 472	was_empty = pipe_empty(head, pipe->tail);
 473	chars = total_len & (PAGE_SIZE-1);
 474	if (chars && !was_empty) {
 475		unsigned int mask = pipe->ring_size - 1;
 476		struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
 477		int offset = buf->offset + buf->len;
 478
 479		if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 480		    offset + chars <= PAGE_SIZE) {
 481			ret = pipe_buf_confirm(pipe, buf);
 482			if (ret)
 483				goto out;
 484
 485			ret = copy_page_from_iter(buf->page, offset, chars, from);
 486			if (unlikely(ret < chars)) {
 487				ret = -EFAULT;
 488				goto out;
 489			}
 490
 491			buf->len += ret;
 492			if (!iov_iter_count(from))
 493				goto out;
 494		}
 495	}
 496
 497	for (;;) {
 498		if (!pipe->readers) {
 499			send_sig(SIGPIPE, current, 0);
 500			if (!ret)
 501				ret = -EPIPE;
 502			break;
 503		}
 504
 505		head = pipe->head;
 506		if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
 507			unsigned int mask = pipe->ring_size - 1;
 508			struct pipe_buffer *buf;
 509			struct page *page = pipe->tmp_page;
 510			int copied;
 511
 512			if (!page) {
 513				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
 514				if (unlikely(!page)) {
 515					ret = ret ? : -ENOMEM;
 516					break;
 517				}
 518				pipe->tmp_page = page;
 519			}
 520
 521			/* Allocate a slot in the ring in advance and attach an
 522			 * empty buffer.  If we fault or otherwise fail to use
 523			 * it, either the reader will consume it or it'll still
 524			 * be there for the next write.
 525			 */
 526			pipe->head = head + 1;
 527
 528			/* Insert it into the buffer array */
 529			buf = &pipe->bufs[head & mask];
 530			buf->page = page;
 531			buf->ops = &anon_pipe_buf_ops;
 532			buf->offset = 0;
 533			buf->len = 0;
 534			if (is_packetized(filp))
 535				buf->flags = PIPE_BUF_FLAG_PACKET;
 536			else
 537				buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
 538			pipe->tmp_page = NULL;
 539
 540			copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
 541			if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
 542				if (!ret)
 543					ret = -EFAULT;
 544				break;
 545			}
 546			ret += copied;
 547			buf->len = copied;
 548
 549			if (!iov_iter_count(from))
 550				break;
 551		}
 552
 553		if (!pipe_full(head, pipe->tail, pipe->max_usage))
 554			continue;
 555
 556		/* Wait for buffer space to become available. */
 557		if ((filp->f_flags & O_NONBLOCK) ||
 558		    (iocb->ki_flags & IOCB_NOWAIT)) {
 559			if (!ret)
 560				ret = -EAGAIN;
 561			break;
 562		}
 563		if (signal_pending(current)) {
 564			if (!ret)
 565				ret = -ERESTARTSYS;
 566			break;
 567		}
 568
 569		/*
 570		 * We're going to release the pipe lock and wait for more
 571		 * space. We wake up any readers if necessary, and then
 572		 * after waiting we need to re-check whether the pipe
 573		 * become empty while we dropped the lock.
 574		 */
 575		mutex_unlock(&pipe->mutex);
 576		if (was_empty)
 577			wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 578		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 579		wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
 580		mutex_lock(&pipe->mutex);
 581		was_empty = pipe_empty(pipe->head, pipe->tail);
 582		wake_next_writer = true;
 583	}
 584out:
 585	if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
 586		wake_next_writer = false;
 587	mutex_unlock(&pipe->mutex);
 588
 589	/*
 590	 * If we do do a wakeup event, we do a 'sync' wakeup, because we
 591	 * want the reader to start processing things asap, rather than
 592	 * leave the data pending.
 593	 *
 594	 * This is particularly important for small writes, because of
 595	 * how (for example) the GNU make jobserver uses small writes to
 596	 * wake up pending jobs
 597	 *
 598	 * Epoll nonsensically wants a wakeup whether the pipe
 599	 * was already empty or not.
 600	 */
 601	if (was_empty || pipe->poll_usage)
 602		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 603	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 604	if (wake_next_writer)
 605		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
 606	if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
 607		int err = file_update_time(filp);
 608		if (err)
 609			ret = err;
 610		sb_end_write(file_inode(filp)->i_sb);
 611	}
 612	return ret;
 613}
 614
 615static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 616{
 617	struct pipe_inode_info *pipe = filp->private_data;
 618	unsigned int count, head, tail, mask;
 619
 620	switch (cmd) {
 621	case FIONREAD:
 622		mutex_lock(&pipe->mutex);
 623		count = 0;
 624		head = pipe->head;
 625		tail = pipe->tail;
 626		mask = pipe->ring_size - 1;
 627
 628		while (tail != head) {
 629			count += pipe->bufs[tail & mask].len;
 630			tail++;
 631		}
 632		mutex_unlock(&pipe->mutex);
 633
 634		return put_user(count, (int __user *)arg);
 635
 636#ifdef CONFIG_WATCH_QUEUE
 637	case IOC_WATCH_QUEUE_SET_SIZE: {
 638		int ret;
 639		mutex_lock(&pipe->mutex);
 640		ret = watch_queue_set_size(pipe, arg);
 641		mutex_unlock(&pipe->mutex);
 642		return ret;
 643	}
 644
 645	case IOC_WATCH_QUEUE_SET_FILTER:
 646		return watch_queue_set_filter(
 647			pipe, (struct watch_notification_filter __user *)arg);
 648#endif
 649
 650	default:
 651		return -ENOIOCTLCMD;
 652	}
 653}
 654
 655/* No kernel lock held - fine */
 656static __poll_t
 657pipe_poll(struct file *filp, poll_table *wait)
 658{
 659	__poll_t mask;
 660	struct pipe_inode_info *pipe = filp->private_data;
 661	unsigned int head, tail;
 662
 663	/* Epoll has some historical nasty semantics, this enables them */
 664	WRITE_ONCE(pipe->poll_usage, true);
 665
 666	/*
 667	 * Reading pipe state only -- no need for acquiring the semaphore.
 668	 *
 669	 * But because this is racy, the code has to add the
 670	 * entry to the poll table _first_ ..
 671	 */
 672	if (filp->f_mode & FMODE_READ)
 673		poll_wait(filp, &pipe->rd_wait, wait);
 674	if (filp->f_mode & FMODE_WRITE)
 675		poll_wait(filp, &pipe->wr_wait, wait);
 676
 677	/*
 678	 * .. and only then can you do the racy tests. That way,
 679	 * if something changes and you got it wrong, the poll
 680	 * table entry will wake you up and fix it.
 681	 */
 682	head = READ_ONCE(pipe->head);
 683	tail = READ_ONCE(pipe->tail);
 684
 685	mask = 0;
 686	if (filp->f_mode & FMODE_READ) {
 687		if (!pipe_empty(head, tail))
 688			mask |= EPOLLIN | EPOLLRDNORM;
 689		if (!pipe->writers && filp->f_version != pipe->w_counter)
 690			mask |= EPOLLHUP;
 691	}
 692
 693	if (filp->f_mode & FMODE_WRITE) {
 694		if (!pipe_full(head, tail, pipe->max_usage))
 695			mask |= EPOLLOUT | EPOLLWRNORM;
 696		/*
 697		 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
 698		 * behave exactly like pipes for poll().
 699		 */
 700		if (!pipe->readers)
 701			mask |= EPOLLERR;
 702	}
 703
 704	return mask;
 705}
 706
 707static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
 708{
 709	int kill = 0;
 710
 711	spin_lock(&inode->i_lock);
 712	if (!--pipe->files) {
 713		inode->i_pipe = NULL;
 714		kill = 1;
 715	}
 716	spin_unlock(&inode->i_lock);
 717
 718	if (kill)
 719		free_pipe_info(pipe);
 720}
 721
 722static int
 723pipe_release(struct inode *inode, struct file *file)
 724{
 725	struct pipe_inode_info *pipe = file->private_data;
 726
 727	mutex_lock(&pipe->mutex);
 728	if (file->f_mode & FMODE_READ)
 729		pipe->readers--;
 730	if (file->f_mode & FMODE_WRITE)
 731		pipe->writers--;
 732
 733	/* Was that the last reader or writer, but not the other side? */
 734	if (!pipe->readers != !pipe->writers) {
 735		wake_up_interruptible_all(&pipe->rd_wait);
 736		wake_up_interruptible_all(&pipe->wr_wait);
 737		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 738		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 739	}
 740	mutex_unlock(&pipe->mutex);
 741
 742	put_pipe_info(inode, pipe);
 743	return 0;
 744}
 745
 746static int
 747pipe_fasync(int fd, struct file *filp, int on)
 748{
 749	struct pipe_inode_info *pipe = filp->private_data;
 750	int retval = 0;
 751
 752	mutex_lock(&pipe->mutex);
 753	if (filp->f_mode & FMODE_READ)
 754		retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 755	if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
 756		retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 757		if (retval < 0 && (filp->f_mode & FMODE_READ))
 758			/* this can happen only if on == T */
 759			fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 760	}
 761	mutex_unlock(&pipe->mutex);
 762	return retval;
 763}
 764
 765unsigned long account_pipe_buffers(struct user_struct *user,
 766				   unsigned long old, unsigned long new)
 767{
 768	return atomic_long_add_return(new - old, &user->pipe_bufs);
 769}
 770
 771bool too_many_pipe_buffers_soft(unsigned long user_bufs)
 772{
 773	unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
 774
 775	return soft_limit && user_bufs > soft_limit;
 776}
 777
 778bool too_many_pipe_buffers_hard(unsigned long user_bufs)
 779{
 780	unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
 781
 782	return hard_limit && user_bufs > hard_limit;
 783}
 784
 785bool pipe_is_unprivileged_user(void)
 786{
 787	return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
 788}
 789
 790struct pipe_inode_info *alloc_pipe_info(void)
 791{
 792	struct pipe_inode_info *pipe;
 793	unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
 794	struct user_struct *user = get_current_user();
 795	unsigned long user_bufs;
 796	unsigned int max_size = READ_ONCE(pipe_max_size);
 797
 798	pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
 799	if (pipe == NULL)
 800		goto out_free_uid;
 801
 802	if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
 803		pipe_bufs = max_size >> PAGE_SHIFT;
 804
 805	user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
 806
 807	if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
 808		user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
 809		pipe_bufs = PIPE_MIN_DEF_BUFFERS;
 810	}
 811
 812	if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
 813		goto out_revert_acct;
 814
 815	pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
 816			     GFP_KERNEL_ACCOUNT);
 817
 818	if (pipe->bufs) {
 819		init_waitqueue_head(&pipe->rd_wait);
 820		init_waitqueue_head(&pipe->wr_wait);
 821		pipe->r_counter = pipe->w_counter = 1;
 822		pipe->max_usage = pipe_bufs;
 823		pipe->ring_size = pipe_bufs;
 824		pipe->nr_accounted = pipe_bufs;
 825		pipe->user = user;
 826		mutex_init(&pipe->mutex);
 827		lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);
 828		return pipe;
 829	}
 830
 831out_revert_acct:
 832	(void) account_pipe_buffers(user, pipe_bufs, 0);
 833	kfree(pipe);
 834out_free_uid:
 835	free_uid(user);
 836	return NULL;
 837}
 838
 839void free_pipe_info(struct pipe_inode_info *pipe)
 840{
 841	unsigned int i;
 842
 843#ifdef CONFIG_WATCH_QUEUE
 844	if (pipe->watch_queue)
 845		watch_queue_clear(pipe->watch_queue);
 846#endif
 847
 848	(void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
 849	free_uid(pipe->user);
 850	for (i = 0; i < pipe->ring_size; i++) {
 851		struct pipe_buffer *buf = pipe->bufs + i;
 852		if (buf->ops)
 853			pipe_buf_release(pipe, buf);
 854	}
 855#ifdef CONFIG_WATCH_QUEUE
 856	if (pipe->watch_queue)
 857		put_watch_queue(pipe->watch_queue);
 858#endif
 859	if (pipe->tmp_page)
 860		__free_page(pipe->tmp_page);
 861	kfree(pipe->bufs);
 862	kfree(pipe);
 863}
 864
 865static struct vfsmount *pipe_mnt __ro_after_init;
 866
 867/*
 868 * pipefs_dname() is called from d_path().
 869 */
 870static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 871{
 872	return dynamic_dname(buffer, buflen, "pipe:[%lu]",
 873				d_inode(dentry)->i_ino);
 874}
 875
 876static const struct dentry_operations pipefs_dentry_operations = {
 877	.d_dname	= pipefs_dname,
 878};
 879
 880static struct inode * get_pipe_inode(void)
 881{
 882	struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 883	struct pipe_inode_info *pipe;
 884
 885	if (!inode)
 886		goto fail_inode;
 887
 888	inode->i_ino = get_next_ino();
 889
 890	pipe = alloc_pipe_info();
 891	if (!pipe)
 892		goto fail_iput;
 893
 894	inode->i_pipe = pipe;
 895	pipe->files = 2;
 896	pipe->readers = pipe->writers = 1;
 897	inode->i_fop = &pipefifo_fops;
 898
 899	/*
 900	 * Mark the inode dirty from the very beginning,
 901	 * that way it will never be moved to the dirty
 902	 * list because "mark_inode_dirty()" will think
 903	 * that it already _is_ on the dirty list.
 904	 */
 905	inode->i_state = I_DIRTY;
 906	inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 907	inode->i_uid = current_fsuid();
 908	inode->i_gid = current_fsgid();
 909	simple_inode_init_ts(inode);
 910
 911	return inode;
 912
 913fail_iput:
 914	iput(inode);
 915
 916fail_inode:
 917	return NULL;
 918}
 919
 920int create_pipe_files(struct file **res, int flags)
 921{
 922	struct inode *inode = get_pipe_inode();
 923	struct file *f;
 924	int error;
 925
 926	if (!inode)
 927		return -ENFILE;
 928
 929	if (flags & O_NOTIFICATION_PIPE) {
 930		error = watch_queue_init(inode->i_pipe);
 931		if (error) {
 932			free_pipe_info(inode->i_pipe);
 933			iput(inode);
 934			return error;
 935		}
 936	}
 937
 938	f = alloc_file_pseudo(inode, pipe_mnt, "",
 939				O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
 940				&pipefifo_fops);
 941	if (IS_ERR(f)) {
 942		free_pipe_info(inode->i_pipe);
 943		iput(inode);
 944		return PTR_ERR(f);
 945	}
 946
 947	f->private_data = inode->i_pipe;
 948
 949	res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
 950				  &pipefifo_fops);
 951	if (IS_ERR(res[0])) {
 952		put_pipe_info(inode, inode->i_pipe);
 953		fput(f);
 954		return PTR_ERR(res[0]);
 955	}
 956	res[0]->private_data = inode->i_pipe;
 957	res[1] = f;
 958	stream_open(inode, res[0]);
 959	stream_open(inode, res[1]);
 960	return 0;
 961}
 962
 963static int __do_pipe_flags(int *fd, struct file **files, int flags)
 964{
 965	int error;
 966	int fdw, fdr;
 967
 968	if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
 969		return -EINVAL;
 970
 971	error = create_pipe_files(files, flags);
 972	if (error)
 973		return error;
 974
 975	error = get_unused_fd_flags(flags);
 976	if (error < 0)
 977		goto err_read_pipe;
 978	fdr = error;
 979
 980	error = get_unused_fd_flags(flags);
 981	if (error < 0)
 982		goto err_fdr;
 983	fdw = error;
 984
 985	audit_fd_pair(fdr, fdw);
 986	fd[0] = fdr;
 987	fd[1] = fdw;
 988	/* pipe groks IOCB_NOWAIT */
 989	files[0]->f_mode |= FMODE_NOWAIT;
 990	files[1]->f_mode |= FMODE_NOWAIT;
 991	return 0;
 992
 993 err_fdr:
 994	put_unused_fd(fdr);
 995 err_read_pipe:
 996	fput(files[0]);
 997	fput(files[1]);
 998	return error;
 999}
1000
1001int do_pipe_flags(int *fd, int flags)
1002{
1003	struct file *files[2];
1004	int error = __do_pipe_flags(fd, files, flags);
1005	if (!error) {
1006		fd_install(fd[0], files[0]);
1007		fd_install(fd[1], files[1]);
1008	}
1009	return error;
1010}
1011
1012/*
1013 * sys_pipe() is the normal C calling standard for creating
1014 * a pipe. It's not the way Unix traditionally does this, though.
1015 */
1016static int do_pipe2(int __user *fildes, int flags)
1017{
1018	struct file *files[2];
1019	int fd[2];
1020	int error;
1021
1022	error = __do_pipe_flags(fd, files, flags);
1023	if (!error) {
1024		if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1025			fput(files[0]);
1026			fput(files[1]);
1027			put_unused_fd(fd[0]);
1028			put_unused_fd(fd[1]);
1029			error = -EFAULT;
1030		} else {
1031			fd_install(fd[0], files[0]);
1032			fd_install(fd[1], files[1]);
1033		}
1034	}
1035	return error;
1036}
1037
1038SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1039{
1040	return do_pipe2(fildes, flags);
1041}
1042
1043SYSCALL_DEFINE1(pipe, int __user *, fildes)
1044{
1045	return do_pipe2(fildes, 0);
1046}
1047
1048/*
1049 * This is the stupid "wait for pipe to be readable or writable"
1050 * model.
1051 *
1052 * See pipe_read/write() for the proper kind of exclusive wait,
1053 * but that requires that we wake up any other readers/writers
1054 * if we then do not end up reading everything (ie the whole
1055 * "wake_next_reader/writer" logic in pipe_read/write()).
1056 */
1057void pipe_wait_readable(struct pipe_inode_info *pipe)
1058{
1059	pipe_unlock(pipe);
1060	wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1061	pipe_lock(pipe);
1062}
1063
1064void pipe_wait_writable(struct pipe_inode_info *pipe)
1065{
1066	pipe_unlock(pipe);
1067	wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1068	pipe_lock(pipe);
1069}
1070
1071/*
1072 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1073 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1074 * race with the count check and waitqueue prep.
1075 *
1076 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1077 * then check the condition you're waiting for, and only then sleep. But
1078 * because of the pipe lock, we can check the condition before being on
1079 * the wait queue.
1080 *
1081 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1082 */
1083static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1084{
1085	DEFINE_WAIT(rdwait);
1086	int cur = *cnt;
1087
1088	while (cur == *cnt) {
1089		prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1090		pipe_unlock(pipe);
1091		schedule();
1092		finish_wait(&pipe->rd_wait, &rdwait);
1093		pipe_lock(pipe);
1094		if (signal_pending(current))
1095			break;
1096	}
1097	return cur == *cnt ? -ERESTARTSYS : 0;
1098}
1099
1100static void wake_up_partner(struct pipe_inode_info *pipe)
1101{
1102	wake_up_interruptible_all(&pipe->rd_wait);
1103}
1104
1105static int fifo_open(struct inode *inode, struct file *filp)
1106{
1107	struct pipe_inode_info *pipe;
1108	bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1109	int ret;
1110
1111	filp->f_version = 0;
1112
1113	spin_lock(&inode->i_lock);
1114	if (inode->i_pipe) {
1115		pipe = inode->i_pipe;
1116		pipe->files++;
1117		spin_unlock(&inode->i_lock);
1118	} else {
1119		spin_unlock(&inode->i_lock);
1120		pipe = alloc_pipe_info();
1121		if (!pipe)
1122			return -ENOMEM;
1123		pipe->files = 1;
1124		spin_lock(&inode->i_lock);
1125		if (unlikely(inode->i_pipe)) {
1126			inode->i_pipe->files++;
1127			spin_unlock(&inode->i_lock);
1128			free_pipe_info(pipe);
1129			pipe = inode->i_pipe;
1130		} else {
1131			inode->i_pipe = pipe;
1132			spin_unlock(&inode->i_lock);
1133		}
1134	}
1135	filp->private_data = pipe;
1136	/* OK, we have a pipe and it's pinned down */
1137
1138	mutex_lock(&pipe->mutex);
1139
1140	/* We can only do regular read/write on fifos */
1141	stream_open(inode, filp);
1142
1143	switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1144	case FMODE_READ:
1145	/*
1146	 *  O_RDONLY
1147	 *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1148	 *  opened, even when there is no process writing the FIFO.
1149	 */
1150		pipe->r_counter++;
1151		if (pipe->readers++ == 0)
1152			wake_up_partner(pipe);
1153
1154		if (!is_pipe && !pipe->writers) {
1155			if ((filp->f_flags & O_NONBLOCK)) {
1156				/* suppress EPOLLHUP until we have
1157				 * seen a writer */
1158				filp->f_version = pipe->w_counter;
1159			} else {
1160				if (wait_for_partner(pipe, &pipe->w_counter))
1161					goto err_rd;
1162			}
1163		}
1164		break;
1165
1166	case FMODE_WRITE:
1167	/*
1168	 *  O_WRONLY
1169	 *  POSIX.1 says that O_NONBLOCK means return -1 with
1170	 *  errno=ENXIO when there is no process reading the FIFO.
1171	 */
1172		ret = -ENXIO;
1173		if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1174			goto err;
1175
1176		pipe->w_counter++;
1177		if (!pipe->writers++)
1178			wake_up_partner(pipe);
1179
1180		if (!is_pipe && !pipe->readers) {
1181			if (wait_for_partner(pipe, &pipe->r_counter))
1182				goto err_wr;
1183		}
1184		break;
1185
1186	case FMODE_READ | FMODE_WRITE:
1187	/*
1188	 *  O_RDWR
1189	 *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1190	 *  This implementation will NEVER block on a O_RDWR open, since
1191	 *  the process can at least talk to itself.
1192	 */
1193
1194		pipe->readers++;
1195		pipe->writers++;
1196		pipe->r_counter++;
1197		pipe->w_counter++;
1198		if (pipe->readers == 1 || pipe->writers == 1)
1199			wake_up_partner(pipe);
1200		break;
1201
1202	default:
1203		ret = -EINVAL;
1204		goto err;
1205	}
1206
1207	/* Ok! */
1208	mutex_unlock(&pipe->mutex);
1209	return 0;
1210
1211err_rd:
1212	if (!--pipe->readers)
1213		wake_up_interruptible(&pipe->wr_wait);
1214	ret = -ERESTARTSYS;
1215	goto err;
1216
1217err_wr:
1218	if (!--pipe->writers)
1219		wake_up_interruptible_all(&pipe->rd_wait);
1220	ret = -ERESTARTSYS;
1221	goto err;
1222
1223err:
1224	mutex_unlock(&pipe->mutex);
1225
1226	put_pipe_info(inode, pipe);
1227	return ret;
1228}
1229
1230const struct file_operations pipefifo_fops = {
1231	.open		= fifo_open,
1232	.llseek		= no_llseek,
1233	.read_iter	= pipe_read,
1234	.write_iter	= pipe_write,
1235	.poll		= pipe_poll,
1236	.unlocked_ioctl	= pipe_ioctl,
1237	.release	= pipe_release,
1238	.fasync		= pipe_fasync,
1239	.splice_write	= iter_file_splice_write,
1240};
1241
1242/*
1243 * Currently we rely on the pipe array holding a power-of-2 number
1244 * of pages. Returns 0 on error.
1245 */
1246unsigned int round_pipe_size(unsigned int size)
1247{
1248	if (size > (1U << 31))
1249		return 0;
1250
1251	/* Minimum pipe size, as required by POSIX */
1252	if (size < PAGE_SIZE)
1253		return PAGE_SIZE;
1254
1255	return roundup_pow_of_two(size);
1256}
1257
1258/*
1259 * Resize the pipe ring to a number of slots.
1260 *
1261 * Note the pipe can be reduced in capacity, but only if the current
1262 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1263 * returned instead.
1264 */
1265int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1266{
1267	struct pipe_buffer *bufs;
1268	unsigned int head, tail, mask, n;
1269
1270	bufs = kcalloc(nr_slots, sizeof(*bufs),
1271		       GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1272	if (unlikely(!bufs))
1273		return -ENOMEM;
1274
1275	spin_lock_irq(&pipe->rd_wait.lock);
1276	mask = pipe->ring_size - 1;
1277	head = pipe->head;
1278	tail = pipe->tail;
1279
1280	n = pipe_occupancy(head, tail);
1281	if (nr_slots < n) {
1282		spin_unlock_irq(&pipe->rd_wait.lock);
1283		kfree(bufs);
1284		return -EBUSY;
1285	}
1286
1287	/*
1288	 * The pipe array wraps around, so just start the new one at zero
1289	 * and adjust the indices.
1290	 */
1291	if (n > 0) {
1292		unsigned int h = head & mask;
1293		unsigned int t = tail & mask;
1294		if (h > t) {
1295			memcpy(bufs, pipe->bufs + t,
1296			       n * sizeof(struct pipe_buffer));
1297		} else {
1298			unsigned int tsize = pipe->ring_size - t;
1299			if (h > 0)
1300				memcpy(bufs + tsize, pipe->bufs,
1301				       h * sizeof(struct pipe_buffer));
1302			memcpy(bufs, pipe->bufs + t,
1303			       tsize * sizeof(struct pipe_buffer));
1304		}
1305	}
1306
1307	head = n;
1308	tail = 0;
1309
1310	kfree(pipe->bufs);
1311	pipe->bufs = bufs;
1312	pipe->ring_size = nr_slots;
1313	if (pipe->max_usage > nr_slots)
1314		pipe->max_usage = nr_slots;
1315	pipe->tail = tail;
1316	pipe->head = head;
1317
1318	if (!pipe_has_watch_queue(pipe)) {
1319		pipe->max_usage = nr_slots;
1320		pipe->nr_accounted = nr_slots;
1321	}
1322
1323	spin_unlock_irq(&pipe->rd_wait.lock);
1324
1325	/* This might have made more room for writers */
1326	wake_up_interruptible(&pipe->wr_wait);
1327	return 0;
1328}
1329
1330/*
1331 * Allocate a new array of pipe buffers and copy the info over. Returns the
1332 * pipe size if successful, or return -ERROR on error.
1333 */
1334static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1335{
1336	unsigned long user_bufs;
1337	unsigned int nr_slots, size;
1338	long ret = 0;
1339
1340	if (pipe_has_watch_queue(pipe))
1341		return -EBUSY;
1342
1343	size = round_pipe_size(arg);
1344	nr_slots = size >> PAGE_SHIFT;
1345
1346	if (!nr_slots)
1347		return -EINVAL;
1348
1349	/*
1350	 * If trying to increase the pipe capacity, check that an
1351	 * unprivileged user is not trying to exceed various limits
1352	 * (soft limit check here, hard limit check just below).
1353	 * Decreasing the pipe capacity is always permitted, even
1354	 * if the user is currently over a limit.
1355	 */
1356	if (nr_slots > pipe->max_usage &&
1357			size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1358		return -EPERM;
1359
1360	user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1361
1362	if (nr_slots > pipe->max_usage &&
1363			(too_many_pipe_buffers_hard(user_bufs) ||
1364			 too_many_pipe_buffers_soft(user_bufs)) &&
1365			pipe_is_unprivileged_user()) {
1366		ret = -EPERM;
1367		goto out_revert_acct;
1368	}
1369
1370	ret = pipe_resize_ring(pipe, nr_slots);
1371	if (ret < 0)
1372		goto out_revert_acct;
1373
1374	return pipe->max_usage * PAGE_SIZE;
1375
1376out_revert_acct:
1377	(void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1378	return ret;
1379}
1380
1381/*
1382 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1383 * not enough to verify that this is a pipe.
1384 */
1385struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1386{
1387	struct pipe_inode_info *pipe = file->private_data;
1388
1389	if (file->f_op != &pipefifo_fops || !pipe)
1390		return NULL;
1391	if (for_splice && pipe_has_watch_queue(pipe))
1392		return NULL;
1393	return pipe;
1394}
1395
1396long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1397{
1398	struct pipe_inode_info *pipe;
1399	long ret;
1400
1401	pipe = get_pipe_info(file, false);
1402	if (!pipe)
1403		return -EBADF;
1404
1405	mutex_lock(&pipe->mutex);
1406
1407	switch (cmd) {
1408	case F_SETPIPE_SZ:
1409		ret = pipe_set_size(pipe, arg);
1410		break;
1411	case F_GETPIPE_SZ:
1412		ret = pipe->max_usage * PAGE_SIZE;
1413		break;
1414	default:
1415		ret = -EINVAL;
1416		break;
1417	}
1418
1419	mutex_unlock(&pipe->mutex);
1420	return ret;
1421}
1422
1423static const struct super_operations pipefs_ops = {
1424	.destroy_inode = free_inode_nonrcu,
1425	.statfs = simple_statfs,
1426};
1427
1428/*
1429 * pipefs should _never_ be mounted by userland - too much of security hassle,
1430 * no real gain from having the whole whorehouse mounted. So we don't need
1431 * any operations on the root directory. However, we need a non-trivial
1432 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1433 */
1434
1435static int pipefs_init_fs_context(struct fs_context *fc)
1436{
1437	struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1438	if (!ctx)
1439		return -ENOMEM;
1440	ctx->ops = &pipefs_ops;
1441	ctx->dops = &pipefs_dentry_operations;
1442	return 0;
1443}
1444
1445static struct file_system_type pipe_fs_type = {
1446	.name		= "pipefs",
1447	.init_fs_context = pipefs_init_fs_context,
1448	.kill_sb	= kill_anon_super,
1449};
1450
1451#ifdef CONFIG_SYSCTL
1452static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1453					unsigned int *valp,
1454					int write, void *data)
1455{
1456	if (write) {
1457		unsigned int val;
1458
1459		val = round_pipe_size(*lvalp);
1460		if (val == 0)
1461			return -EINVAL;
1462
1463		*valp = val;
1464	} else {
1465		unsigned int val = *valp;
1466		*lvalp = (unsigned long) val;
1467	}
1468
1469	return 0;
1470}
1471
1472static int proc_dopipe_max_size(struct ctl_table *table, int write,
1473				void *buffer, size_t *lenp, loff_t *ppos)
1474{
1475	return do_proc_douintvec(table, write, buffer, lenp, ppos,
1476				 do_proc_dopipe_max_size_conv, NULL);
1477}
1478
1479static struct ctl_table fs_pipe_sysctls[] = {
1480	{
1481		.procname	= "pipe-max-size",
1482		.data		= &pipe_max_size,
1483		.maxlen		= sizeof(pipe_max_size),
1484		.mode		= 0644,
1485		.proc_handler	= proc_dopipe_max_size,
1486	},
1487	{
1488		.procname	= "pipe-user-pages-hard",
1489		.data		= &pipe_user_pages_hard,
1490		.maxlen		= sizeof(pipe_user_pages_hard),
1491		.mode		= 0644,
1492		.proc_handler	= proc_doulongvec_minmax,
1493	},
1494	{
1495		.procname	= "pipe-user-pages-soft",
1496		.data		= &pipe_user_pages_soft,
1497		.maxlen		= sizeof(pipe_user_pages_soft),
1498		.mode		= 0644,
1499		.proc_handler	= proc_doulongvec_minmax,
1500	},
1501};
1502#endif
1503
1504static int __init init_pipe_fs(void)
1505{
1506	int err = register_filesystem(&pipe_fs_type);
1507
1508	if (!err) {
1509		pipe_mnt = kern_mount(&pipe_fs_type);
1510		if (IS_ERR(pipe_mnt)) {
1511			err = PTR_ERR(pipe_mnt);
1512			unregister_filesystem(&pipe_fs_type);
1513		}
1514	}
1515#ifdef CONFIG_SYSCTL
1516	register_sysctl_init("fs", fs_pipe_sysctls);
1517#endif
1518	return err;
1519}
1520
1521fs_initcall(init_pipe_fs);
   1/*
   2 *  linux/fs/pipe.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
   5 */
   6
   7#include <linux/mm.h>
   8#include <linux/file.h>
   9#include <linux/poll.h>
  10#include <linux/slab.h>
  11#include <linux/module.h>
  12#include <linux/init.h>
  13#include <linux/fs.h>
  14#include <linux/log2.h>
  15#include <linux/mount.h>
  16#include <linux/pipe_fs_i.h>
  17#include <linux/uio.h>
  18#include <linux/highmem.h>
  19#include <linux/pagemap.h>
  20#include <linux/audit.h>
  21#include <linux/syscalls.h>
  22#include <linux/fcntl.h>
  23
  24#include <asm/uaccess.h>
  25#include <asm/ioctls.h>
  26
  27/*
  28 * The max size that a non-root user is allowed to grow the pipe. Can
  29 * be set by root in /proc/sys/fs/pipe-max-size
  30 */
  31unsigned int pipe_max_size = 1048576;
  32
  33/*
  34 * Minimum pipe size, as required by POSIX
  35 */
  36unsigned int pipe_min_size = PAGE_SIZE;
  37
  38/*
  39 * We use a start+len construction, which provides full use of the 
  40 * allocated memory.
  41 * -- Florian Coosmann (FGC)
  42 * 
  43 * Reads with count = 0 should always return 0.
  44 * -- Julian Bradfield 1999-06-07.
  45 *
  46 * FIFOs and Pipes now generate SIGIO for both readers and writers.
  47 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
  48 *
  49 * pipe_read & write cleanup
  50 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
  51 */
  52
  53static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
  54{
  55	if (pipe->inode)
  56		mutex_lock_nested(&pipe->inode->i_mutex, subclass);
  57}
  58
  59void pipe_lock(struct pipe_inode_info *pipe)
  60{
  61	/*
  62	 * pipe_lock() nests non-pipe inode locks (for writing to a file)
  63	 */
  64	pipe_lock_nested(pipe, I_MUTEX_PARENT);
  65}
  66EXPORT_SYMBOL(pipe_lock);
  67
  68void pipe_unlock(struct pipe_inode_info *pipe)
  69{
  70	if (pipe->inode)
  71		mutex_unlock(&pipe->inode->i_mutex);
  72}
  73EXPORT_SYMBOL(pipe_unlock);
  74
  75void pipe_double_lock(struct pipe_inode_info *pipe1,
  76		      struct pipe_inode_info *pipe2)
  77{
  78	BUG_ON(pipe1 == pipe2);
  79
  80	if (pipe1 < pipe2) {
  81		pipe_lock_nested(pipe1, I_MUTEX_PARENT);
  82		pipe_lock_nested(pipe2, I_MUTEX_CHILD);
  83	} else {
  84		pipe_lock_nested(pipe2, I_MUTEX_PARENT);
  85		pipe_lock_nested(pipe1, I_MUTEX_CHILD);
  86	}
  87}
  88
  89/* Drop the inode semaphore and wait for a pipe event, atomically */
  90void pipe_wait(struct pipe_inode_info *pipe)
  91{
  92	DEFINE_WAIT(wait);
  93
  94	/*
  95	 * Pipes are system-local resources, so sleeping on them
  96	 * is considered a noninteractive wait:
  97	 */
  98	prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
  99	pipe_unlock(pipe);
 100	schedule();
 101	finish_wait(&pipe->wait, &wait);
 102	pipe_lock(pipe);
 103}
 104
 105static int
 106pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
 107			int atomic)
 108{
 109	unsigned long copy;
 110
 111	while (len > 0) {
 112		while (!iov->iov_len)
 113			iov++;
 114		copy = min_t(unsigned long, len, iov->iov_len);
 115
 116		if (atomic) {
 117			if (__copy_from_user_inatomic(to, iov->iov_base, copy))
 118				return -EFAULT;
 119		} else {
 120			if (copy_from_user(to, iov->iov_base, copy))
 121				return -EFAULT;
 122		}
 123		to += copy;
 124		len -= copy;
 125		iov->iov_base += copy;
 126		iov->iov_len -= copy;
 127	}
 128	return 0;
 129}
 130
 131static int
 132pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
 133		      int atomic)
 134{
 135	unsigned long copy;
 136
 137	while (len > 0) {
 138		while (!iov->iov_len)
 139			iov++;
 140		copy = min_t(unsigned long, len, iov->iov_len);
 141
 142		if (atomic) {
 143			if (__copy_to_user_inatomic(iov->iov_base, from, copy))
 144				return -EFAULT;
 145		} else {
 146			if (copy_to_user(iov->iov_base, from, copy))
 147				return -EFAULT;
 148		}
 149		from += copy;
 150		len -= copy;
 151		iov->iov_base += copy;
 152		iov->iov_len -= copy;
 153	}
 154	return 0;
 155}
 156
 157/*
 158 * Attempt to pre-fault in the user memory, so we can use atomic copies.
 159 * Returns the number of bytes not faulted in.
 160 */
 161static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
 162{
 163	while (!iov->iov_len)
 164		iov++;
 165
 166	while (len > 0) {
 167		unsigned long this_len;
 168
 169		this_len = min_t(unsigned long, len, iov->iov_len);
 170		if (fault_in_pages_writeable(iov->iov_base, this_len))
 171			break;
 172
 173		len -= this_len;
 174		iov++;
 175	}
 176
 177	return len;
 178}
 179
 180/*
 181 * Pre-fault in the user memory, so we can use atomic copies.
 182 */
 183static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
 184{
 185	while (!iov->iov_len)
 186		iov++;
 187
 188	while (len > 0) {
 189		unsigned long this_len;
 190
 191		this_len = min_t(unsigned long, len, iov->iov_len);
 192		fault_in_pages_readable(iov->iov_base, this_len);
 193		len -= this_len;
 194		iov++;
 195	}
 196}
 197
 198static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
 199				  struct pipe_buffer *buf)
 200{
 201	struct page *page = buf->page;
 202
 203	/*
 204	 * If nobody else uses this page, and we don't already have a
 205	 * temporary page, let's keep track of it as a one-deep
 206	 * allocation cache. (Otherwise just release our reference to it)
 207	 */
 208	if (page_count(page) == 1 && !pipe->tmp_page)
 209		pipe->tmp_page = page;
 210	else
 211		page_cache_release(page);
 212}
 213
 214/**
 215 * generic_pipe_buf_map - virtually map a pipe buffer
 216 * @pipe:	the pipe that the buffer belongs to
 217 * @buf:	the buffer that should be mapped
 218 * @atomic:	whether to use an atomic map
 219 *
 220 * Description:
 221 *	This function returns a kernel virtual address mapping for the
 222 *	pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
 223 *	and the caller has to be careful not to fault before calling
 224 *	the unmap function.
 225 *
 226 *	Note that this function occupies KM_USER0 if @atomic != 0.
 227 */
 228void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
 229			   struct pipe_buffer *buf, int atomic)
 230{
 231	if (atomic) {
 232		buf->flags |= PIPE_BUF_FLAG_ATOMIC;
 233		return kmap_atomic(buf->page, KM_USER0);
 234	}
 235
 236	return kmap(buf->page);
 237}
 238EXPORT_SYMBOL(generic_pipe_buf_map);
 239
 240/**
 241 * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
 242 * @pipe:	the pipe that the buffer belongs to
 243 * @buf:	the buffer that should be unmapped
 244 * @map_data:	the data that the mapping function returned
 245 *
 246 * Description:
 247 *	This function undoes the mapping that ->map() provided.
 248 */
 249void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
 250			    struct pipe_buffer *buf, void *map_data)
 251{
 252	if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
 253		buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
 254		kunmap_atomic(map_data, KM_USER0);
 255	} else
 256		kunmap(buf->page);
 257}
 258EXPORT_SYMBOL(generic_pipe_buf_unmap);
 259
 260/**
 261 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
 262 * @pipe:	the pipe that the buffer belongs to
 263 * @buf:	the buffer to attempt to steal
 264 *
 265 * Description:
 266 *	This function attempts to steal the &struct page attached to
 267 *	@buf. If successful, this function returns 0 and returns with
 268 *	the page locked. The caller may then reuse the page for whatever
 269 *	he wishes; the typical use is insertion into a different file
 270 *	page cache.
 271 */
 272int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
 273			   struct pipe_buffer *buf)
 274{
 275	struct page *page = buf->page;
 276
 277	/*
 278	 * A reference of one is golden, that means that the owner of this
 279	 * page is the only one holding a reference to it. lock the page
 280	 * and return OK.
 281	 */
 282	if (page_count(page) == 1) {
 283		lock_page(page);
 284		return 0;
 285	}
 286
 287	return 1;
 288}
 289EXPORT_SYMBOL(generic_pipe_buf_steal);
 290
 291/**
 292 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
 293 * @pipe:	the pipe that the buffer belongs to
 294 * @buf:	the buffer to get a reference to
 295 *
 296 * Description:
 297 *	This function grabs an extra reference to @buf. It's used in
 298 *	in the tee() system call, when we duplicate the buffers in one
 299 *	pipe into another.
 300 */
 301void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
 302{
 303	page_cache_get(buf->page);
 304}
 305EXPORT_SYMBOL(generic_pipe_buf_get);
 306
 307/**
 308 * generic_pipe_buf_confirm - verify contents of the pipe buffer
 309 * @info:	the pipe that the buffer belongs to
 310 * @buf:	the buffer to confirm
 311 *
 312 * Description:
 313 *	This function does nothing, because the generic pipe code uses
 314 *	pages that are always good when inserted into the pipe.
 315 */
 316int generic_pipe_buf_confirm(struct pipe_inode_info *info,
 317			     struct pipe_buffer *buf)
 318{
 319	return 0;
 320}
 321EXPORT_SYMBOL(generic_pipe_buf_confirm);
 322
 323/**
 324 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
 325 * @pipe:	the pipe that the buffer belongs to
 326 * @buf:	the buffer to put a reference to
 327 *
 328 * Description:
 329 *	This function releases a reference to @buf.
 330 */
 331void generic_pipe_buf_release(struct pipe_inode_info *pipe,
 332			      struct pipe_buffer *buf)
 333{
 334	page_cache_release(buf->page);
 335}
 336EXPORT_SYMBOL(generic_pipe_buf_release);
 337
 338static const struct pipe_buf_operations anon_pipe_buf_ops = {
 339	.can_merge = 1,
 340	.map = generic_pipe_buf_map,
 341	.unmap = generic_pipe_buf_unmap,
 342	.confirm = generic_pipe_buf_confirm,
 343	.release = anon_pipe_buf_release,
 344	.steal = generic_pipe_buf_steal,
 345	.get = generic_pipe_buf_get,
 346};
 347
 348static ssize_t
 349pipe_read(struct kiocb *iocb, const struct iovec *_iov,
 350	   unsigned long nr_segs, loff_t pos)
 351{
 352	struct file *filp = iocb->ki_filp;
 353	struct inode *inode = filp->f_path.dentry->d_inode;
 354	struct pipe_inode_info *pipe;
 355	int do_wakeup;
 356	ssize_t ret;
 357	struct iovec *iov = (struct iovec *)_iov;
 358	size_t total_len;
 359
 360	total_len = iov_length(iov, nr_segs);
 361	/* Null read succeeds. */
 362	if (unlikely(total_len == 0))
 363		return 0;
 364
 365	do_wakeup = 0;
 366	ret = 0;
 367	mutex_lock(&inode->i_mutex);
 368	pipe = inode->i_pipe;
 369	for (;;) {
 370		int bufs = pipe->nrbufs;
 371		if (bufs) {
 372			int curbuf = pipe->curbuf;
 373			struct pipe_buffer *buf = pipe->bufs + curbuf;
 374			const struct pipe_buf_operations *ops = buf->ops;
 375			void *addr;
 376			size_t chars = buf->len;
 377			int error, atomic;
 378
 379			if (chars > total_len)
 380				chars = total_len;
 381
 382			error = ops->confirm(pipe, buf);
 383			if (error) {
 384				if (!ret)
 385					ret = error;
 386				break;
 387			}
 388
 389			atomic = !iov_fault_in_pages_write(iov, chars);
 390redo:
 391			addr = ops->map(pipe, buf, atomic);
 392			error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
 393			ops->unmap(pipe, buf, addr);
 394			if (unlikely(error)) {
 395				/*
 396				 * Just retry with the slow path if we failed.
 397				 */
 398				if (atomic) {
 399					atomic = 0;
 400					goto redo;
 401				}
 402				if (!ret)
 403					ret = error;
 404				break;
 405			}
 406			ret += chars;
 407			buf->offset += chars;
 408			buf->len -= chars;
 409			if (!buf->len) {
 410				buf->ops = NULL;
 411				ops->release(pipe, buf);
 412				curbuf = (curbuf + 1) & (pipe->buffers - 1);
 413				pipe->curbuf = curbuf;
 414				pipe->nrbufs = --bufs;
 415				do_wakeup = 1;
 416			}
 417			total_len -= chars;
 418			if (!total_len)
 419				break;	/* common path: read succeeded */
 420		}
 421		if (bufs)	/* More to do? */
 422			continue;
 423		if (!pipe->writers)
 424			break;
 425		if (!pipe->waiting_writers) {
 426			/* syscall merging: Usually we must not sleep
 427			 * if O_NONBLOCK is set, or if we got some data.
 428			 * But if a writer sleeps in kernel space, then
 429			 * we can wait for that data without violating POSIX.
 430			 */
 431			if (ret)
 432				break;
 433			if (filp->f_flags & O_NONBLOCK) {
 434				ret = -EAGAIN;
 435				break;
 436			}
 437		}
 438		if (signal_pending(current)) {
 439			if (!ret)
 440				ret = -ERESTARTSYS;
 441			break;
 442		}
 443		if (do_wakeup) {
 444			wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 445 			kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 446		}
 447		pipe_wait(pipe);
 448	}
 449	mutex_unlock(&inode->i_mutex);
 450
 451	/* Signal writers asynchronously that there is more room. */
 452	if (do_wakeup) {
 453		wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
 454		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 455	}
 456	if (ret > 0)
 457		file_accessed(filp);
 458	return ret;
 459}
 460
 461static ssize_t
 462pipe_write(struct kiocb *iocb, const struct iovec *_iov,
 463	    unsigned long nr_segs, loff_t ppos)
 464{
 465	struct file *filp = iocb->ki_filp;
 466	struct inode *inode = filp->f_path.dentry->d_inode;
 467	struct pipe_inode_info *pipe;
 468	ssize_t ret;
 469	int do_wakeup;
 470	struct iovec *iov = (struct iovec *)_iov;
 471	size_t total_len;
 472	ssize_t chars;
 473
 474	total_len = iov_length(iov, nr_segs);
 475	/* Null write succeeds. */
 476	if (unlikely(total_len == 0))
 477		return 0;
 478
 479	do_wakeup = 0;
 480	ret = 0;
 481	mutex_lock(&inode->i_mutex);
 482	pipe = inode->i_pipe;
 483
 484	if (!pipe->readers) {
 485		send_sig(SIGPIPE, current, 0);
 486		ret = -EPIPE;
 487		goto out;
 488	}
 489
 490	/* We try to merge small writes */
 491	chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
 492	if (pipe->nrbufs && chars != 0) {
 493		int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
 494							(pipe->buffers - 1);
 495		struct pipe_buffer *buf = pipe->bufs + lastbuf;
 496		const struct pipe_buf_operations *ops = buf->ops;
 497		int offset = buf->offset + buf->len;
 498
 499		if (ops->can_merge && offset + chars <= PAGE_SIZE) {
 500			int error, atomic = 1;
 501			void *addr;
 502
 503			error = ops->confirm(pipe, buf);
 504			if (error)
 505				goto out;
 506
 507			iov_fault_in_pages_read(iov, chars);
 508redo1:
 509			addr = ops->map(pipe, buf, atomic);
 510			error = pipe_iov_copy_from_user(offset + addr, iov,
 511							chars, atomic);
 512			ops->unmap(pipe, buf, addr);
 513			ret = error;
 514			do_wakeup = 1;
 515			if (error) {
 516				if (atomic) {
 517					atomic = 0;
 518					goto redo1;
 519				}
 520				goto out;
 521			}
 522			buf->len += chars;
 523			total_len -= chars;
 524			ret = chars;
 525			if (!total_len)
 526				goto out;
 527		}
 528	}
 529
 530	for (;;) {
 531		int bufs;
 532
 533		if (!pipe->readers) {
 534			send_sig(SIGPIPE, current, 0);
 535			if (!ret)
 536				ret = -EPIPE;
 537			break;
 538		}
 539		bufs = pipe->nrbufs;
 540		if (bufs < pipe->buffers) {
 541			int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
 542			struct pipe_buffer *buf = pipe->bufs + newbuf;
 543			struct page *page = pipe->tmp_page;
 544			char *src;
 545			int error, atomic = 1;
 546
 547			if (!page) {
 548				page = alloc_page(GFP_HIGHUSER);
 549				if (unlikely(!page)) {
 550					ret = ret ? : -ENOMEM;
 551					break;
 552				}
 553				pipe->tmp_page = page;
 554			}
 555			/* Always wake up, even if the copy fails. Otherwise
 556			 * we lock up (O_NONBLOCK-)readers that sleep due to
 557			 * syscall merging.
 558			 * FIXME! Is this really true?
 559			 */
 560			do_wakeup = 1;
 561			chars = PAGE_SIZE;
 562			if (chars > total_len)
 563				chars = total_len;
 564
 565			iov_fault_in_pages_read(iov, chars);
 566redo2:
 567			if (atomic)
 568				src = kmap_atomic(page, KM_USER0);
 569			else
 570				src = kmap(page);
 571
 572			error = pipe_iov_copy_from_user(src, iov, chars,
 573							atomic);
 574			if (atomic)
 575				kunmap_atomic(src, KM_USER0);
 576			else
 577				kunmap(page);
 578
 579			if (unlikely(error)) {
 580				if (atomic) {
 581					atomic = 0;
 582					goto redo2;
 583				}
 584				if (!ret)
 585					ret = error;
 586				break;
 587			}
 588			ret += chars;
 589
 590			/* Insert it into the buffer array */
 591			buf->page = page;
 592			buf->ops = &anon_pipe_buf_ops;
 593			buf->offset = 0;
 594			buf->len = chars;
 595			pipe->nrbufs = ++bufs;
 596			pipe->tmp_page = NULL;
 597
 598			total_len -= chars;
 599			if (!total_len)
 600				break;
 601		}
 602		if (bufs < pipe->buffers)
 603			continue;
 604		if (filp->f_flags & O_NONBLOCK) {
 605			if (!ret)
 606				ret = -EAGAIN;
 607			break;
 608		}
 609		if (signal_pending(current)) {
 610			if (!ret)
 611				ret = -ERESTARTSYS;
 612			break;
 613		}
 614		if (do_wakeup) {
 615			wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 616			kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 617			do_wakeup = 0;
 618		}
 619		pipe->waiting_writers++;
 620		pipe_wait(pipe);
 621		pipe->waiting_writers--;
 622	}
 623out:
 624	mutex_unlock(&inode->i_mutex);
 625	if (do_wakeup) {
 626		wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
 627		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 628	}
 629	if (ret > 0)
 630		file_update_time(filp);
 631	return ret;
 632}
 633
 634static ssize_t
 635bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 636{
 637	return -EBADF;
 638}
 639
 640static ssize_t
 641bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
 642	   loff_t *ppos)
 643{
 644	return -EBADF;
 645}
 646
 647static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 648{
 649	struct inode *inode = filp->f_path.dentry->d_inode;
 650	struct pipe_inode_info *pipe;
 651	int count, buf, nrbufs;
 652
 653	switch (cmd) {
 654		case FIONREAD:
 655			mutex_lock(&inode->i_mutex);
 656			pipe = inode->i_pipe;
 657			count = 0;
 658			buf = pipe->curbuf;
 659			nrbufs = pipe->nrbufs;
 660			while (--nrbufs >= 0) {
 661				count += pipe->bufs[buf].len;
 662				buf = (buf+1) & (pipe->buffers - 1);
 663			}
 664			mutex_unlock(&inode->i_mutex);
 665
 666			return put_user(count, (int __user *)arg);
 667		default:
 668			return -EINVAL;
 669	}
 670}
 671
 672/* No kernel lock held - fine */
 673static unsigned int
 674pipe_poll(struct file *filp, poll_table *wait)
 675{
 676	unsigned int mask;
 677	struct inode *inode = filp->f_path.dentry->d_inode;
 678	struct pipe_inode_info *pipe = inode->i_pipe;
 679	int nrbufs;
 680
 681	poll_wait(filp, &pipe->wait, wait);
 682
 683	/* Reading only -- no need for acquiring the semaphore.  */
 684	nrbufs = pipe->nrbufs;
 685	mask = 0;
 686	if (filp->f_mode & FMODE_READ) {
 687		mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
 688		if (!pipe->writers && filp->f_version != pipe->w_counter)
 689			mask |= POLLHUP;
 690	}
 691
 692	if (filp->f_mode & FMODE_WRITE) {
 693		mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
 694		/*
 695		 * Most Unices do not set POLLERR for FIFOs but on Linux they
 696		 * behave exactly like pipes for poll().
 697		 */
 698		if (!pipe->readers)
 699			mask |= POLLERR;
 700	}
 701
 702	return mask;
 703}
 704
 705static int
 706pipe_release(struct inode *inode, int decr, int decw)
 707{
 708	struct pipe_inode_info *pipe;
 709
 710	mutex_lock(&inode->i_mutex);
 711	pipe = inode->i_pipe;
 712	pipe->readers -= decr;
 713	pipe->writers -= decw;
 714
 715	if (!pipe->readers && !pipe->writers) {
 716		free_pipe_info(inode);
 717	} else {
 718		wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
 719		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 720		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 721	}
 722	mutex_unlock(&inode->i_mutex);
 723
 724	return 0;
 725}
 726
 727static int
 728pipe_read_fasync(int fd, struct file *filp, int on)
 729{
 730	struct inode *inode = filp->f_path.dentry->d_inode;
 731	int retval;
 732
 733	mutex_lock(&inode->i_mutex);
 734	retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
 735	mutex_unlock(&inode->i_mutex);
 736
 737	return retval;
 738}
 739
 740
 741static int
 742pipe_write_fasync(int fd, struct file *filp, int on)
 743{
 744	struct inode *inode = filp->f_path.dentry->d_inode;
 745	int retval;
 746
 747	mutex_lock(&inode->i_mutex);
 748	retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
 749	mutex_unlock(&inode->i_mutex);
 750
 751	return retval;
 752}
 753
 754
 755static int
 756pipe_rdwr_fasync(int fd, struct file *filp, int on)
 757{
 758	struct inode *inode = filp->f_path.dentry->d_inode;
 759	struct pipe_inode_info *pipe = inode->i_pipe;
 760	int retval;
 761
 762	mutex_lock(&inode->i_mutex);
 763	retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
 764	if (retval >= 0) {
 765		retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
 766		if (retval < 0) /* this can happen only if on == T */
 767			fasync_helper(-1, filp, 0, &pipe->fasync_readers);
 768	}
 769	mutex_unlock(&inode->i_mutex);
 770	return retval;
 771}
 772
 773
 774static int
 775pipe_read_release(struct inode *inode, struct file *filp)
 776{
 777	return pipe_release(inode, 1, 0);
 778}
 779
 780static int
 781pipe_write_release(struct inode *inode, struct file *filp)
 782{
 783	return pipe_release(inode, 0, 1);
 784}
 785
 786static int
 787pipe_rdwr_release(struct inode *inode, struct file *filp)
 788{
 789	int decr, decw;
 790
 791	decr = (filp->f_mode & FMODE_READ) != 0;
 792	decw = (filp->f_mode & FMODE_WRITE) != 0;
 793	return pipe_release(inode, decr, decw);
 794}
 795
 796static int
 797pipe_read_open(struct inode *inode, struct file *filp)
 798{
 799	int ret = -ENOENT;
 800
 801	mutex_lock(&inode->i_mutex);
 802
 803	if (inode->i_pipe) {
 804		ret = 0;
 805		inode->i_pipe->readers++;
 806	}
 807
 808	mutex_unlock(&inode->i_mutex);
 809
 810	return ret;
 811}
 812
 813static int
 814pipe_write_open(struct inode *inode, struct file *filp)
 815{
 816	int ret = -ENOENT;
 817
 818	mutex_lock(&inode->i_mutex);
 819
 820	if (inode->i_pipe) {
 821		ret = 0;
 822		inode->i_pipe->writers++;
 823	}
 824
 825	mutex_unlock(&inode->i_mutex);
 826
 827	return ret;
 828}
 829
 830static int
 831pipe_rdwr_open(struct inode *inode, struct file *filp)
 832{
 833	int ret = -ENOENT;
 834
 835	mutex_lock(&inode->i_mutex);
 836
 837	if (inode->i_pipe) {
 838		ret = 0;
 839		if (filp->f_mode & FMODE_READ)
 840			inode->i_pipe->readers++;
 841		if (filp->f_mode & FMODE_WRITE)
 842			inode->i_pipe->writers++;
 843	}
 844
 845	mutex_unlock(&inode->i_mutex);
 846
 847	return ret;
 848}
 849
 850/*
 851 * The file_operations structs are not static because they
 852 * are also used in linux/fs/fifo.c to do operations on FIFOs.
 853 *
 854 * Pipes reuse fifos' file_operations structs.
 855 */
 856const struct file_operations read_pipefifo_fops = {
 857	.llseek		= no_llseek,
 858	.read		= do_sync_read,
 859	.aio_read	= pipe_read,
 860	.write		= bad_pipe_w,
 861	.poll		= pipe_poll,
 862	.unlocked_ioctl	= pipe_ioctl,
 863	.open		= pipe_read_open,
 864	.release	= pipe_read_release,
 865	.fasync		= pipe_read_fasync,
 866};
 867
 868const struct file_operations write_pipefifo_fops = {
 869	.llseek		= no_llseek,
 870	.read		= bad_pipe_r,
 871	.write		= do_sync_write,
 872	.aio_write	= pipe_write,
 873	.poll		= pipe_poll,
 874	.unlocked_ioctl	= pipe_ioctl,
 875	.open		= pipe_write_open,
 876	.release	= pipe_write_release,
 877	.fasync		= pipe_write_fasync,
 878};
 879
 880const struct file_operations rdwr_pipefifo_fops = {
 881	.llseek		= no_llseek,
 882	.read		= do_sync_read,
 883	.aio_read	= pipe_read,
 884	.write		= do_sync_write,
 885	.aio_write	= pipe_write,
 886	.poll		= pipe_poll,
 887	.unlocked_ioctl	= pipe_ioctl,
 888	.open		= pipe_rdwr_open,
 889	.release	= pipe_rdwr_release,
 890	.fasync		= pipe_rdwr_fasync,
 891};
 892
 893struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
 894{
 895	struct pipe_inode_info *pipe;
 896
 897	pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
 898	if (pipe) {
 899		pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
 900		if (pipe->bufs) {
 901			init_waitqueue_head(&pipe->wait);
 902			pipe->r_counter = pipe->w_counter = 1;
 903			pipe->inode = inode;
 904			pipe->buffers = PIPE_DEF_BUFFERS;
 905			return pipe;
 906		}
 907		kfree(pipe);
 908	}
 909
 910	return NULL;
 911}
 912
 913void __free_pipe_info(struct pipe_inode_info *pipe)
 914{
 915	int i;
 916
 917	for (i = 0; i < pipe->buffers; i++) {
 918		struct pipe_buffer *buf = pipe->bufs + i;
 919		if (buf->ops)
 920			buf->ops->release(pipe, buf);
 921	}
 922	if (pipe->tmp_page)
 923		__free_page(pipe->tmp_page);
 924	kfree(pipe->bufs);
 925	kfree(pipe);
 926}
 927
 928void free_pipe_info(struct inode *inode)
 929{
 930	__free_pipe_info(inode->i_pipe);
 931	inode->i_pipe = NULL;
 932}
 933
 934static struct vfsmount *pipe_mnt __read_mostly;
 935
 936/*
 937 * pipefs_dname() is called from d_path().
 938 */
 939static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
 940{
 941	return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
 942				dentry->d_inode->i_ino);
 943}
 944
 945static const struct dentry_operations pipefs_dentry_operations = {
 946	.d_dname	= pipefs_dname,
 947};
 948
 949static struct inode * get_pipe_inode(void)
 950{
 951	struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
 952	struct pipe_inode_info *pipe;
 953
 954	if (!inode)
 955		goto fail_inode;
 956
 957	inode->i_ino = get_next_ino();
 958
 959	pipe = alloc_pipe_info(inode);
 960	if (!pipe)
 961		goto fail_iput;
 962	inode->i_pipe = pipe;
 963
 964	pipe->readers = pipe->writers = 1;
 965	inode->i_fop = &rdwr_pipefifo_fops;
 966
 967	/*
 968	 * Mark the inode dirty from the very beginning,
 969	 * that way it will never be moved to the dirty
 970	 * list because "mark_inode_dirty()" will think
 971	 * that it already _is_ on the dirty list.
 972	 */
 973	inode->i_state = I_DIRTY;
 974	inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
 975	inode->i_uid = current_fsuid();
 976	inode->i_gid = current_fsgid();
 977	inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 978
 979	return inode;
 980
 981fail_iput:
 982	iput(inode);
 983
 984fail_inode:
 985	return NULL;
 986}
 987
 988struct file *create_write_pipe(int flags)
 989{
 990	int err;
 991	struct inode *inode;
 992	struct file *f;
 993	struct path path;
 994	struct qstr name = { .name = "" };
 995
 996	err = -ENFILE;
 997	inode = get_pipe_inode();
 998	if (!inode)
 999		goto err;
1000
1001	err = -ENOMEM;
1002	path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1003	if (!path.dentry)
1004		goto err_inode;
1005	path.mnt = mntget(pipe_mnt);
1006
1007	d_instantiate(path.dentry, inode);
1008
1009	err = -ENFILE;
1010	f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1011	if (!f)
1012		goto err_dentry;
1013	f->f_mapping = inode->i_mapping;
1014
1015	f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
1016	f->f_version = 0;
1017
1018	return f;
1019
1020 err_dentry:
1021	free_pipe_info(inode);
1022	path_put(&path);
1023	return ERR_PTR(err);
1024
1025 err_inode:
1026	free_pipe_info(inode);
1027	iput(inode);
1028 err:
1029	return ERR_PTR(err);
1030}
1031
1032void free_write_pipe(struct file *f)
1033{
1034	free_pipe_info(f->f_dentry->d_inode);
1035	path_put(&f->f_path);
1036	put_filp(f);
1037}
1038
1039struct file *create_read_pipe(struct file *wrf, int flags)
1040{
1041	/* Grab pipe from the writer */
1042	struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1043				    &read_pipefifo_fops);
1044	if (!f)
1045		return ERR_PTR(-ENFILE);
1046
1047	path_get(&wrf->f_path);
1048	f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1049
1050	return f;
1051}
1052
1053int do_pipe_flags(int *fd, int flags)
1054{
1055	struct file *fw, *fr;
1056	int error;
1057	int fdw, fdr;
1058
1059	if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1060		return -EINVAL;
1061
1062	fw = create_write_pipe(flags);
1063	if (IS_ERR(fw))
1064		return PTR_ERR(fw);
1065	fr = create_read_pipe(fw, flags);
1066	error = PTR_ERR(fr);
1067	if (IS_ERR(fr))
1068		goto err_write_pipe;
1069
1070	error = get_unused_fd_flags(flags);
1071	if (error < 0)
1072		goto err_read_pipe;
1073	fdr = error;
1074
1075	error = get_unused_fd_flags(flags);
1076	if (error < 0)
1077		goto err_fdr;
1078	fdw = error;
1079
1080	audit_fd_pair(fdr, fdw);
1081	fd_install(fdr, fr);
1082	fd_install(fdw, fw);
1083	fd[0] = fdr;
1084	fd[1] = fdw;
1085
1086	return 0;
1087
1088 err_fdr:
1089	put_unused_fd(fdr);
1090 err_read_pipe:
1091	path_put(&fr->f_path);
1092	put_filp(fr);
1093 err_write_pipe:
1094	free_write_pipe(fw);
1095	return error;
1096}
1097
1098/*
1099 * sys_pipe() is the normal C calling standard for creating
1100 * a pipe. It's not the way Unix traditionally does this, though.
1101 */
1102SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1103{
1104	int fd[2];
1105	int error;
1106
1107	error = do_pipe_flags(fd, flags);
1108	if (!error) {
1109		if (copy_to_user(fildes, fd, sizeof(fd))) {
1110			sys_close(fd[0]);
1111			sys_close(fd[1]);
1112			error = -EFAULT;
1113		}
1114	}
1115	return error;
1116}
1117
1118SYSCALL_DEFINE1(pipe, int __user *, fildes)
1119{
1120	return sys_pipe2(fildes, 0);
1121}
1122
1123/*
1124 * Allocate a new array of pipe buffers and copy the info over. Returns the
1125 * pipe size if successful, or return -ERROR on error.
1126 */
1127static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1128{
1129	struct pipe_buffer *bufs;
1130
1131	/*
1132	 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1133	 * expect a lot of shrink+grow operations, just free and allocate
1134	 * again like we would do for growing. If the pipe currently
1135	 * contains more buffers than arg, then return busy.
1136	 */
1137	if (nr_pages < pipe->nrbufs)
1138		return -EBUSY;
1139
1140	bufs = kcalloc(nr_pages, sizeof(struct pipe_buffer), GFP_KERNEL);
1141	if (unlikely(!bufs))
1142		return -ENOMEM;
1143
1144	/*
1145	 * The pipe array wraps around, so just start the new one at zero
1146	 * and adjust the indexes.
1147	 */
1148	if (pipe->nrbufs) {
1149		unsigned int tail;
1150		unsigned int head;
1151
1152		tail = pipe->curbuf + pipe->nrbufs;
1153		if (tail < pipe->buffers)
1154			tail = 0;
1155		else
1156			tail &= (pipe->buffers - 1);
1157
1158		head = pipe->nrbufs - tail;
1159		if (head)
1160			memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1161		if (tail)
1162			memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1163	}
1164
1165	pipe->curbuf = 0;
1166	kfree(pipe->bufs);
1167	pipe->bufs = bufs;
1168	pipe->buffers = nr_pages;
1169	return nr_pages * PAGE_SIZE;
1170}
1171
1172/*
1173 * Currently we rely on the pipe array holding a power-of-2 number
1174 * of pages.
1175 */
1176static inline unsigned int round_pipe_size(unsigned int size)
1177{
1178	unsigned long nr_pages;
1179
1180	nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1181	return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1182}
1183
1184/*
1185 * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1186 * will return an error.
1187 */
1188int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1189		 size_t *lenp, loff_t *ppos)
1190{
1191	int ret;
1192
1193	ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1194	if (ret < 0 || !write)
1195		return ret;
1196
1197	pipe_max_size = round_pipe_size(pipe_max_size);
1198	return ret;
1199}
1200
1201/*
1202 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1203 * location, so checking ->i_pipe is not enough to verify that this is a
1204 * pipe.
1205 */
1206struct pipe_inode_info *get_pipe_info(struct file *file)
1207{
1208	struct inode *i = file->f_path.dentry->d_inode;
1209
1210	return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1211}
1212
1213long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1214{
1215	struct pipe_inode_info *pipe;
1216	long ret;
1217
1218	pipe = get_pipe_info(file);
1219	if (!pipe)
1220		return -EBADF;
1221
1222	mutex_lock(&pipe->inode->i_mutex);
1223
1224	switch (cmd) {
1225	case F_SETPIPE_SZ: {
1226		unsigned int size, nr_pages;
1227
1228		size = round_pipe_size(arg);
1229		nr_pages = size >> PAGE_SHIFT;
1230
1231		ret = -EINVAL;
1232		if (!nr_pages)
1233			goto out;
1234
1235		if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1236			ret = -EPERM;
1237			goto out;
1238		}
1239		ret = pipe_set_size(pipe, nr_pages);
1240		break;
1241		}
1242	case F_GETPIPE_SZ:
1243		ret = pipe->buffers * PAGE_SIZE;
1244		break;
1245	default:
1246		ret = -EINVAL;
1247		break;
1248	}
1249
1250out:
1251	mutex_unlock(&pipe->inode->i_mutex);
1252	return ret;
1253}
1254
1255static const struct super_operations pipefs_ops = {
1256	.destroy_inode = free_inode_nonrcu,
1257};
1258
1259/*
1260 * pipefs should _never_ be mounted by userland - too much of security hassle,
1261 * no real gain from having the whole whorehouse mounted. So we don't need
1262 * any operations on the root directory. However, we need a non-trivial
1263 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1264 */
1265static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1266			 int flags, const char *dev_name, void *data)
1267{
1268	return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1269			&pipefs_dentry_operations, PIPEFS_MAGIC);
1270}
1271
1272static struct file_system_type pipe_fs_type = {
1273	.name		= "pipefs",
1274	.mount		= pipefs_mount,
1275	.kill_sb	= kill_anon_super,
1276};
1277
1278static int __init init_pipe_fs(void)
1279{
1280	int err = register_filesystem(&pipe_fs_type);
1281
1282	if (!err) {
1283		pipe_mnt = kern_mount(&pipe_fs_type);
1284		if (IS_ERR(pipe_mnt)) {
1285			err = PTR_ERR(pipe_mnt);
1286			unregister_filesystem(&pipe_fs_type);
1287		}
1288	}
1289	return err;
1290}
1291
1292static void __exit exit_pipe_fs(void)
1293{
1294	kern_unmount(pipe_mnt);
1295	unregister_filesystem(&pipe_fs_type);
1296}
1297
1298fs_initcall(init_pipe_fs);
1299module_exit(exit_pipe_fs);