blk-mq: don't count completed flush data request as inflight in case of quiesce
[linux-block.git] / fs / pipe.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
1da177e4
LT
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>
35f3d14d 15#include <linux/log2.h>
1da177e4 16#include <linux/mount.h>
4fa7ec5d 17#include <linux/pseudo_fs.h>
b502bd11 18#include <linux/magic.h>
1da177e4
LT
19#include <linux/pipe_fs_i.h>
20#include <linux/uio.h>
21#include <linux/highmem.h>
5274f052 22#include <linux/pagemap.h>
db349509 23#include <linux/audit.h>
ba719bae 24#include <linux/syscalls.h>
b492e95b 25#include <linux/fcntl.h>
d86133bd 26#include <linux/memcontrol.h>
c73be61c 27#include <linux/watch_queue.h>
1998f193 28#include <linux/sysctl.h>
1da177e4 29
7c0f6ba6 30#include <linux/uaccess.h>
1da177e4
LT
31#include <asm/ioctls.h>
32
599a0ac1
AV
33#include "internal.h"
34
46c4c9d1
AXH
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
b492e95b
JA
50/*
51 * The max size that a non-root user is allowed to grow the pipe. Can
ff9da691 52 * be set by root in /proc/sys/fs/pipe-max-size
b492e95b 53 */
1998f193 54static unsigned int pipe_max_size = 1048576;
ff9da691 55
759c0114
WT
56/* Maximum allocatable pages per user. Hard limit is unset by default, soft
57 * matches default values.
58 */
1998f193
LC
59static unsigned long pipe_user_pages_hard;
60static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
759c0114 61
1da177e4 62/*
8cefc107
DH
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 *
1da177e4
LT
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
61e0d47c
MS
79static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
80{
6447a3cf 81 if (pipe->files)
72b0d9aa 82 mutex_lock_nested(&pipe->mutex, subclass);
61e0d47c
MS
83}
84
85void pipe_lock(struct pipe_inode_info *pipe)
86{
87 /*
88 * pipe_lock() nests non-pipe inode locks (for writing to a file)
89 */
90 pipe_lock_nested(pipe, I_MUTEX_PARENT);
91}
92EXPORT_SYMBOL(pipe_lock);
93
94void pipe_unlock(struct pipe_inode_info *pipe)
95{
6447a3cf 96 if (pipe->files)
72b0d9aa 97 mutex_unlock(&pipe->mutex);
61e0d47c
MS
98}
99EXPORT_SYMBOL(pipe_unlock);
100
ebec73f4
AV
101static inline void __pipe_lock(struct pipe_inode_info *pipe)
102{
103 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
104}
105
106static inline void __pipe_unlock(struct pipe_inode_info *pipe)
107{
108 mutex_unlock(&pipe->mutex);
109}
110
61e0d47c
MS
111void pipe_double_lock(struct pipe_inode_info *pipe1,
112 struct pipe_inode_info *pipe2)
113{
114 BUG_ON(pipe1 == pipe2);
115
116 if (pipe1 < pipe2) {
117 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
118 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
119 } else {
023d43c7
PZ
120 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
121 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
61e0d47c
MS
122 }
123}
124
341b446b
IM
125static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
126 struct pipe_buffer *buf)
1da177e4
LT
127{
128 struct page *page = buf->page;
129
5274f052
JA
130 /*
131 * If nobody else uses this page, and we don't already have a
132 * temporary page, let's keep track of it as a one-deep
341b446b 133 * allocation cache. (Otherwise just release our reference to it)
5274f052 134 */
341b446b 135 if (page_count(page) == 1 && !pipe->tmp_page)
923f4f23 136 pipe->tmp_page = page;
341b446b 137 else
09cbfeaf 138 put_page(page);
1da177e4
LT
139}
140
c928f642
CH
141static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
142 struct pipe_buffer *buf)
d86133bd
VD
143{
144 struct page *page = buf->page;
145
c928f642
CH
146 if (page_count(page) != 1)
147 return false;
148 memcg_kmem_uncharge_page(page, 0);
149 __SetPageLocked(page);
150 return true;
d86133bd
VD
151}
152
0845718d 153/**
c928f642 154 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
0845718d
JA
155 * @pipe: the pipe that the buffer belongs to
156 * @buf: the buffer to attempt to steal
157 *
158 * Description:
b51d63c6 159 * This function attempts to steal the &struct page attached to
0845718d
JA
160 * @buf. If successful, this function returns 0 and returns with
161 * the page locked. The caller may then reuse the page for whatever
b51d63c6 162 * he wishes; the typical use is insertion into a different file
0845718d
JA
163 * page cache.
164 */
c928f642
CH
165bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
166 struct pipe_buffer *buf)
5abc97aa 167{
46e678c9
JA
168 struct page *page = buf->page;
169
0845718d
JA
170 /*
171 * A reference of one is golden, that means that the owner of this
172 * page is the only one holding a reference to it. lock the page
173 * and return OK.
174 */
46e678c9 175 if (page_count(page) == 1) {
46e678c9 176 lock_page(page);
c928f642 177 return true;
46e678c9 178 }
c928f642 179 return false;
5abc97aa 180}
c928f642 181EXPORT_SYMBOL(generic_pipe_buf_try_steal);
5abc97aa 182
0845718d 183/**
b51d63c6 184 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
0845718d
JA
185 * @pipe: the pipe that the buffer belongs to
186 * @buf: the buffer to get a reference to
187 *
188 * Description:
189 * This function grabs an extra reference to @buf. It's used in
3d742d4b 190 * the tee() system call, when we duplicate the buffers in one
0845718d
JA
191 * pipe into another.
192 */
15fab63e 193bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
70524490 194{
cd1adf1b 195 return try_get_page(buf->page);
70524490 196}
51921cb7 197EXPORT_SYMBOL(generic_pipe_buf_get);
70524490 198
6818173b
MS
199/**
200 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
201 * @pipe: the pipe that the buffer belongs to
202 * @buf: the buffer to put a reference to
203 *
204 * Description:
205 * This function releases a reference to @buf.
206 */
207void generic_pipe_buf_release(struct pipe_inode_info *pipe,
208 struct pipe_buffer *buf)
209{
09cbfeaf 210 put_page(buf->page);
6818173b 211}
51921cb7 212EXPORT_SYMBOL(generic_pipe_buf_release);
6818173b 213
d4c3cca9 214static const struct pipe_buf_operations anon_pipe_buf_ops = {
c928f642
CH
215 .release = anon_pipe_buf_release,
216 .try_steal = anon_pipe_buf_try_steal,
217 .get = generic_pipe_buf_get,
1da177e4
LT
218};
219
85190d15
LT
220/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
221static inline bool pipe_readable(const struct pipe_inode_info *pipe)
222{
223 unsigned int head = READ_ONCE(pipe->head);
224 unsigned int tail = READ_ONCE(pipe->tail);
225 unsigned int writers = READ_ONCE(pipe->writers);
226
227 return !pipe_empty(head, tail) || !writers;
228}
229
478dbf12
MK
230static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
231 struct pipe_buffer *buf,
232 unsigned int tail)
233{
234 pipe_buf_release(pipe, buf);
235
236 /*
237 * If the pipe has a watch_queue, we need additional protection
238 * by the spinlock because notifications get posted with only
239 * this spinlock, no mutex
240 */
241 if (pipe_has_watch_queue(pipe)) {
242 spin_lock_irq(&pipe->rd_wait.lock);
243#ifdef CONFIG_WATCH_QUEUE
244 if (buf->flags & PIPE_BUF_FLAG_LOSS)
245 pipe->note_loss = true;
246#endif
247 pipe->tail = ++tail;
248 spin_unlock_irq(&pipe->rd_wait.lock);
249 return tail;
250 }
251
252 /*
253 * Without a watch_queue, we can simply increment the tail
254 * without the spinlock - the mutex is enough.
255 */
256 pipe->tail = ++tail;
257 return tail;
258}
259
1da177e4 260static ssize_t
fb9096a3 261pipe_read(struct kiocb *iocb, struct iov_iter *to)
1da177e4 262{
fb9096a3 263 size_t total_len = iov_iter_count(to);
ee0b3e67 264 struct file *filp = iocb->ki_filp;
de32ec4c 265 struct pipe_inode_info *pipe = filp->private_data;
0ddad21d 266 bool was_full, wake_next_reader = false;
1da177e4 267 ssize_t ret;
1da177e4 268
1da177e4
LT
269 /* Null read succeeds. */
270 if (unlikely(total_len == 0))
271 return 0;
272
1da177e4 273 ret = 0;
ebec73f4 274 __pipe_lock(pipe);
f467a6a6
LT
275
276 /*
277 * We only wake up writers if the pipe was full when we started
278 * reading in order to avoid unnecessary wakeups.
279 *
280 * But when we do wake up writers, we do so using a sync wakeup
281 * (WF_SYNC), because we want them to get going and generate more
282 * data for us.
283 */
284 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
1da177e4 285 for (;;) {
2ed147f0
DH
286 /* Read ->head with a barrier vs post_one_notification() */
287 unsigned int head = smp_load_acquire(&pipe->head);
8cefc107
DH
288 unsigned int tail = pipe->tail;
289 unsigned int mask = pipe->ring_size - 1;
290
e7d553d6
DH
291#ifdef CONFIG_WATCH_QUEUE
292 if (pipe->note_loss) {
293 struct watch_notification n;
294
295 if (total_len < 8) {
296 if (ret == 0)
297 ret = -ENOBUFS;
298 break;
299 }
300
301 n.type = WATCH_TYPE_META;
302 n.subtype = WATCH_META_LOSS_NOTIFICATION;
303 n.info = watch_sizeof(n);
304 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
305 if (ret == 0)
306 ret = -EFAULT;
307 break;
308 }
309 ret += sizeof(n);
310 total_len -= sizeof(n);
311 pipe->note_loss = false;
312 }
313#endif
314
8cefc107
DH
315 if (!pipe_empty(head, tail)) {
316 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
1da177e4 317 size_t chars = buf->len;
637b58c2
AV
318 size_t written;
319 int error;
1da177e4 320
8cfba763
DH
321 if (chars > total_len) {
322 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
323 if (ret == 0)
324 ret = -ENOBUFS;
325 break;
326 }
1da177e4 327 chars = total_len;
8cfba763 328 }
1da177e4 329
fba597db 330 error = pipe_buf_confirm(pipe, buf);
f84d7519 331 if (error) {
5274f052 332 if (!ret)
e5953cbd 333 ret = error;
5274f052
JA
334 break;
335 }
f84d7519 336
fb9096a3 337 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
637b58c2 338 if (unlikely(written < chars)) {
341b446b 339 if (!ret)
637b58c2 340 ret = -EFAULT;
1da177e4
LT
341 break;
342 }
343 ret += chars;
344 buf->offset += chars;
345 buf->len -= chars;
9883035a
LT
346
347 /* Was it a packet buffer? Clean up and exit */
348 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
349 total_len = chars;
350 buf->len = 0;
351 }
352
478dbf12
MK
353 if (!buf->len)
354 tail = pipe_update_tail(pipe, buf, tail);
1da177e4
LT
355 total_len -= chars;
356 if (!total_len)
357 break; /* common path: read succeeded */
8cefc107
DH
358 if (!pipe_empty(head, tail)) /* More to do? */
359 continue;
1da177e4 360 }
8cefc107 361
923f4f23 362 if (!pipe->writers)
1da177e4 363 break;
a28c8b9d
LT
364 if (ret)
365 break;
c04fe8e3
JA
366 if ((filp->f_flags & O_NONBLOCK) ||
367 (iocb->ki_flags & IOCB_NOWAIT)) {
a28c8b9d
LT
368 ret = -EAGAIN;
369 break;
1da177e4 370 }
85190d15 371 __pipe_unlock(pipe);
d1c6a2aa
LT
372
373 /*
374 * We only get here if we didn't actually read anything.
375 *
376 * However, we could have seen (and removed) a zero-sized
377 * pipe buffer, and might have made space in the buffers
378 * that way.
379 *
380 * You can't make zero-sized pipe buffers by doing an empty
381 * write (not even in packet mode), but they can happen if
382 * the writer gets an EFAULT when trying to fill a buffer
383 * that already got allocated and inserted in the buffer
384 * array.
385 *
386 * So we still need to wake up any pending writers in the
387 * _very_ unlikely case that the pipe was full, but we got
388 * no data.
389 */
fe67f4dd 390 if (unlikely(was_full))
0ddad21d 391 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
fe67f4dd 392 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
d1c6a2aa
LT
393
394 /*
395 * But because we didn't read anything, at this point we can
396 * just return directly with -ERESTARTSYS if we're interrupted,
397 * since we've done any required wakeups and there's no need
398 * to mark anything accessed. And we've dropped the lock.
399 */
0ddad21d 400 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
d1c6a2aa
LT
401 return -ERESTARTSYS;
402
85190d15 403 __pipe_lock(pipe);
f467a6a6 404 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
0ddad21d 405 wake_next_reader = true;
1da177e4 406 }
0ddad21d
LT
407 if (pipe_empty(pipe->head, pipe->tail))
408 wake_next_reader = false;
ebec73f4 409 __pipe_unlock(pipe);
341b446b 410
fe67f4dd 411 if (was_full)
0ddad21d 412 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
0ddad21d
LT
413 if (wake_next_reader)
414 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
fe67f4dd 415 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
1da177e4
LT
416 if (ret > 0)
417 file_accessed(filp);
418 return ret;
419}
420
9883035a
LT
421static inline int is_packetized(struct file *file)
422{
423 return (file->f_flags & O_DIRECT) != 0;
424}
425
85190d15
LT
426/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
427static inline bool pipe_writable(const struct pipe_inode_info *pipe)
428{
429 unsigned int head = READ_ONCE(pipe->head);
430 unsigned int tail = READ_ONCE(pipe->tail);
431 unsigned int max_usage = READ_ONCE(pipe->max_usage);
432
433 return !pipe_full(head, tail, max_usage) ||
434 !READ_ONCE(pipe->readers);
435}
436
1da177e4 437static ssize_t
f0d1bec9 438pipe_write(struct kiocb *iocb, struct iov_iter *from)
1da177e4 439{
ee0b3e67 440 struct file *filp = iocb->ki_filp;
de32ec4c 441 struct pipe_inode_info *pipe = filp->private_data;
8f868d68 442 unsigned int head;
f0d1bec9 443 ssize_t ret = 0;
f0d1bec9 444 size_t total_len = iov_iter_count(from);
1da177e4 445 ssize_t chars;
1b6b26ae 446 bool was_empty = false;
0ddad21d 447 bool wake_next_writer = false;
1da177e4 448
1da177e4
LT
449 /* Null write succeeds. */
450 if (unlikely(total_len == 0))
451 return 0;
452
ebec73f4 453 __pipe_lock(pipe);
1da177e4 454
923f4f23 455 if (!pipe->readers) {
1da177e4
LT
456 send_sig(SIGPIPE, current, 0);
457 ret = -EPIPE;
458 goto out;
459 }
460
b4bd6b4b 461 if (pipe_has_watch_queue(pipe)) {
c73be61c
DH
462 ret = -EXDEV;
463 goto out;
464 }
c73be61c 465
1b6b26ae 466 /*
1b6b26ae
LT
467 * If it wasn't empty we try to merge new data into
468 * the last buffer.
469 *
470 * That naturally merges small writes, but it also
3a34b13a 471 * page-aligns the rest of the writes for large writes
1b6b26ae
LT
472 * spanning multiple pages.
473 */
8cefc107 474 head = pipe->head;
3b844826 475 was_empty = pipe_empty(head, pipe->tail);
1b6b26ae 476 chars = total_len & (PAGE_SIZE-1);
3b844826 477 if (chars && !was_empty) {
8f868d68 478 unsigned int mask = pipe->ring_size - 1;
8cefc107 479 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
1da177e4 480 int offset = buf->offset + buf->len;
341b446b 481
f6dd9755
CH
482 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
483 offset + chars <= PAGE_SIZE) {
fba597db 484 ret = pipe_buf_confirm(pipe, buf);
6ae08069 485 if (ret)
5274f052 486 goto out;
f84d7519 487
f0d1bec9
AV
488 ret = copy_page_from_iter(buf->page, offset, chars, from);
489 if (unlikely(ret < chars)) {
6ae08069 490 ret = -EFAULT;
1da177e4 491 goto out;
f6762b7a 492 }
1b6b26ae 493
6ae08069 494 buf->len += ret;
f0d1bec9 495 if (!iov_iter_count(from))
1da177e4
LT
496 goto out;
497 }
498 }
499
500 for (;;) {
923f4f23 501 if (!pipe->readers) {
1da177e4 502 send_sig(SIGPIPE, current, 0);
341b446b
IM
503 if (!ret)
504 ret = -EPIPE;
1da177e4
LT
505 break;
506 }
8cefc107 507
a194dfe6 508 head = pipe->head;
8f868d68
DH
509 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
510 unsigned int mask = pipe->ring_size - 1;
fbaa530e 511 struct pipe_buffer *buf;
923f4f23 512 struct page *page = pipe->tmp_page;
f0d1bec9 513 int copied;
1da177e4
LT
514
515 if (!page) {
d86133bd 516 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
1da177e4
LT
517 if (unlikely(!page)) {
518 ret = ret ? : -ENOMEM;
519 break;
520 }
923f4f23 521 pipe->tmp_page = page;
1da177e4 522 }
a194dfe6
DH
523
524 /* Allocate a slot in the ring in advance and attach an
525 * empty buffer. If we fault or otherwise fail to use
526 * it, either the reader will consume it or it'll still
527 * be there for the next write.
528 */
a194dfe6 529 pipe->head = head + 1;
1da177e4
LT
530
531 /* Insert it into the buffer array */
a194dfe6 532 buf = &pipe->bufs[head & mask];
1da177e4
LT
533 buf->page = page;
534 buf->ops = &anon_pipe_buf_ops;
535 buf->offset = 0;
a194dfe6 536 buf->len = 0;
f6dd9755 537 if (is_packetized(filp))
9883035a 538 buf->flags = PIPE_BUF_FLAG_PACKET;
f6dd9755
CH
539 else
540 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
923f4f23 541 pipe->tmp_page = NULL;
1da177e4 542
a194dfe6
DH
543 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
544 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
545 if (!ret)
546 ret = -EFAULT;
547 break;
548 }
549 ret += copied;
a194dfe6
DH
550 buf->len = copied;
551
f0d1bec9 552 if (!iov_iter_count(from))
1da177e4
LT
553 break;
554 }
8cefc107 555
8f868d68 556 if (!pipe_full(head, pipe->tail, pipe->max_usage))
1da177e4 557 continue;
8cefc107
DH
558
559 /* Wait for buffer space to become available. */
c04fe8e3
JA
560 if ((filp->f_flags & O_NONBLOCK) ||
561 (iocb->ki_flags & IOCB_NOWAIT)) {
341b446b
IM
562 if (!ret)
563 ret = -EAGAIN;
1da177e4
LT
564 break;
565 }
566 if (signal_pending(current)) {
341b446b
IM
567 if (!ret)
568 ret = -ERESTARTSYS;
1da177e4
LT
569 break;
570 }
1b6b26ae
LT
571
572 /*
573 * We're going to release the pipe lock and wait for more
574 * space. We wake up any readers if necessary, and then
575 * after waiting we need to re-check whether the pipe
576 * become empty while we dropped the lock.
577 */
85190d15 578 __pipe_unlock(pipe);
fe67f4dd 579 if (was_empty)
0ddad21d 580 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
fe67f4dd 581 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
0ddad21d 582 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
85190d15 583 __pipe_lock(pipe);
0dd1e377 584 was_empty = pipe_empty(pipe->head, pipe->tail);
0ddad21d 585 wake_next_writer = true;
1da177e4
LT
586 }
587out:
0ddad21d
LT
588 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
589 wake_next_writer = false;
ebec73f4 590 __pipe_unlock(pipe);
1b6b26ae
LT
591
592 /*
593 * If we do do a wakeup event, we do a 'sync' wakeup, because we
594 * want the reader to start processing things asap, rather than
595 * leave the data pending.
596 *
597 * This is particularly important for small writes, because of
598 * how (for example) the GNU make jobserver uses small writes to
599 * wake up pending jobs
3b844826
LT
600 *
601 * Epoll nonsensically wants a wakeup whether the pipe
602 * was already empty or not.
1b6b26ae 603 */
fe67f4dd 604 if (was_empty || pipe->poll_usage)
0ddad21d 605 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
fe67f4dd 606 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
0ddad21d
LT
607 if (wake_next_writer)
608 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
7e775f46 609 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
c3b2da31
JB
610 int err = file_update_time(filp);
611 if (err)
612 ret = err;
7e775f46 613 sb_end_write(file_inode(filp)->i_sb);
c3b2da31 614 }
1da177e4
LT
615 return ret;
616}
617
d59d0b1b 618static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
1da177e4 619{
de32ec4c 620 struct pipe_inode_info *pipe = filp->private_data;
aeb213cd 621 unsigned int count, head, tail, mask;
1da177e4
LT
622
623 switch (cmd) {
c73be61c
DH
624 case FIONREAD:
625 __pipe_lock(pipe);
626 count = 0;
627 head = pipe->head;
628 tail = pipe->tail;
629 mask = pipe->ring_size - 1;
8cefc107 630
c73be61c
DH
631 while (tail != head) {
632 count += pipe->bufs[tail & mask].len;
633 tail++;
634 }
635 __pipe_unlock(pipe);
923f4f23 636
c73be61c 637 return put_user(count, (int __user *)arg);
923f4f23 638
c73be61c
DH
639#ifdef CONFIG_WATCH_QUEUE
640 case IOC_WATCH_QUEUE_SET_SIZE: {
641 int ret;
642 __pipe_lock(pipe);
643 ret = watch_queue_set_size(pipe, arg);
644 __pipe_unlock(pipe);
645 return ret;
646 }
647
648 case IOC_WATCH_QUEUE_SET_FILTER:
649 return watch_queue_set_filter(
650 pipe, (struct watch_notification_filter __user *)arg);
651#endif
652
653 default:
654 return -ENOIOCTLCMD;
1da177e4
LT
655 }
656}
657
dd67081b 658/* No kernel lock held - fine */
a11e1d43
LT
659static __poll_t
660pipe_poll(struct file *filp, poll_table *wait)
dd67081b 661{
a11e1d43 662 __poll_t mask;
dd67081b 663 struct pipe_inode_info *pipe = filp->private_data;
ad910e36 664 unsigned int head, tail;
a11e1d43 665
3b844826 666 /* Epoll has some historical nasty semantics, this enables them */
f485922d 667 WRITE_ONCE(pipe->poll_usage, true);
3b844826 668
ad910e36 669 /*
0ddad21d 670 * Reading pipe state only -- no need for acquiring the semaphore.
ad910e36
LT
671 *
672 * But because this is racy, the code has to add the
673 * entry to the poll table _first_ ..
674 */
0ddad21d
LT
675 if (filp->f_mode & FMODE_READ)
676 poll_wait(filp, &pipe->rd_wait, wait);
677 if (filp->f_mode & FMODE_WRITE)
678 poll_wait(filp, &pipe->wr_wait, wait);
1da177e4 679
ad910e36
LT
680 /*
681 * .. and only then can you do the racy tests. That way,
682 * if something changes and you got it wrong, the poll
683 * table entry will wake you up and fix it.
684 */
685 head = READ_ONCE(pipe->head);
686 tail = READ_ONCE(pipe->tail);
687
a11e1d43 688 mask = 0;
1da177e4 689 if (filp->f_mode & FMODE_READ) {
8cefc107
DH
690 if (!pipe_empty(head, tail))
691 mask |= EPOLLIN | EPOLLRDNORM;
923f4f23 692 if (!pipe->writers && filp->f_version != pipe->w_counter)
a9a08845 693 mask |= EPOLLHUP;
1da177e4
LT
694 }
695
696 if (filp->f_mode & FMODE_WRITE) {
6718b6f8 697 if (!pipe_full(head, tail, pipe->max_usage))
8cefc107 698 mask |= EPOLLOUT | EPOLLWRNORM;
5e5d7a22 699 /*
a9a08845 700 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
5e5d7a22
PE
701 * behave exactly like pipes for poll().
702 */
923f4f23 703 if (!pipe->readers)
a9a08845 704 mask |= EPOLLERR;
1da177e4
LT
705 }
706
707 return mask;
708}
709
b0d8d229
LT
710static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
711{
712 int kill = 0;
713
714 spin_lock(&inode->i_lock);
715 if (!--pipe->files) {
716 inode->i_pipe = NULL;
717 kill = 1;
718 }
719 spin_unlock(&inode->i_lock);
720
721 if (kill)
722 free_pipe_info(pipe);
723}
724
1da177e4 725static int
599a0ac1 726pipe_release(struct inode *inode, struct file *file)
1da177e4 727{
b0d8d229 728 struct pipe_inode_info *pipe = file->private_data;
923f4f23 729
ebec73f4 730 __pipe_lock(pipe);
599a0ac1
AV
731 if (file->f_mode & FMODE_READ)
732 pipe->readers--;
733 if (file->f_mode & FMODE_WRITE)
734 pipe->writers--;
341b446b 735
6551d5c5
LT
736 /* Was that the last reader or writer, but not the other side? */
737 if (!pipe->readers != !pipe->writers) {
738 wake_up_interruptible_all(&pipe->rd_wait);
739 wake_up_interruptible_all(&pipe->wr_wait);
923f4f23
IM
740 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
741 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
1da177e4 742 }
ebec73f4 743 __pipe_unlock(pipe);
ba5bb147 744
b0d8d229 745 put_pipe_info(inode, pipe);
1da177e4
LT
746 return 0;
747}
748
749static int
599a0ac1 750pipe_fasync(int fd, struct file *filp, int on)
1da177e4 751{
de32ec4c 752 struct pipe_inode_info *pipe = filp->private_data;
599a0ac1 753 int retval = 0;
1da177e4 754
ebec73f4 755 __pipe_lock(pipe);
599a0ac1
AV
756 if (filp->f_mode & FMODE_READ)
757 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
758 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
341b446b 759 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
599a0ac1
AV
760 if (retval < 0 && (filp->f_mode & FMODE_READ))
761 /* this can happen only if on == T */
e5bc49ba
ON
762 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
763 }
ebec73f4 764 __pipe_unlock(pipe);
60aa4924 765 return retval;
1da177e4
LT
766}
767
c73be61c
DH
768unsigned long account_pipe_buffers(struct user_struct *user,
769 unsigned long old, unsigned long new)
759c0114 770{
9c87bcf0 771 return atomic_long_add_return(new - old, &user->pipe_bufs);
759c0114
WT
772}
773
c73be61c 774bool too_many_pipe_buffers_soft(unsigned long user_bufs)
759c0114 775{
f7340761
EB
776 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
777
778 return soft_limit && user_bufs > soft_limit;
759c0114
WT
779}
780
c73be61c 781bool too_many_pipe_buffers_hard(unsigned long user_bufs)
759c0114 782{
f7340761
EB
783 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
784
785 return hard_limit && user_bufs > hard_limit;
759c0114
WT
786}
787
c73be61c 788bool pipe_is_unprivileged_user(void)
85c2dd54
EB
789{
790 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
791}
792
7bee130e 793struct pipe_inode_info *alloc_pipe_info(void)
3a326a2c 794{
923f4f23 795 struct pipe_inode_info *pipe;
09b4d199
MK
796 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
797 struct user_struct *user = get_current_user();
9c87bcf0 798 unsigned long user_bufs;
f7340761 799 unsigned int max_size = READ_ONCE(pipe_max_size);
3a326a2c 800
d86133bd 801 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
09b4d199
MK
802 if (pipe == NULL)
803 goto out_free_uid;
804
f7340761
EB
805 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
806 pipe_bufs = max_size >> PAGE_SHIFT;
086e774a 807
9c87bcf0 808 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
a005ca0e 809
c73be61c 810 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
46c4c9d1
AXH
811 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
812 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
09b4d199 813 }
759c0114 814
c73be61c 815 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
a005ca0e
MK
816 goto out_revert_acct;
817
906f9040 818 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
a005ca0e
MK
819 GFP_KERNEL_ACCOUNT);
820
09b4d199 821 if (pipe->bufs) {
0ddad21d
LT
822 init_waitqueue_head(&pipe->rd_wait);
823 init_waitqueue_head(&pipe->wr_wait);
09b4d199 824 pipe->r_counter = pipe->w_counter = 1;
6718b6f8 825 pipe->max_usage = pipe_bufs;
8cefc107 826 pipe->ring_size = pipe_bufs;
c73be61c 827 pipe->nr_accounted = pipe_bufs;
09b4d199 828 pipe->user = user;
09b4d199
MK
829 mutex_init(&pipe->mutex);
830 return pipe;
3a326a2c
IM
831 }
832
a005ca0e 833out_revert_acct:
9c87bcf0 834 (void) account_pipe_buffers(user, pipe_bufs, 0);
09b4d199
MK
835 kfree(pipe);
836out_free_uid:
837 free_uid(user);
35f3d14d 838 return NULL;
3a326a2c
IM
839}
840
4b8a8f1e 841void free_pipe_info(struct pipe_inode_info *pipe)
1da177e4 842{
aeb213cd 843 unsigned int i;
1da177e4 844
c73be61c 845#ifdef CONFIG_WATCH_QUEUE
db8facfc 846 if (pipe->watch_queue)
c73be61c 847 watch_queue_clear(pipe->watch_queue);
c73be61c
DH
848#endif
849
850 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
759c0114 851 free_uid(pipe->user);
8cefc107 852 for (i = 0; i < pipe->ring_size; i++) {
923f4f23 853 struct pipe_buffer *buf = pipe->bufs + i;
1da177e4 854 if (buf->ops)
a779638c 855 pipe_buf_release(pipe, buf);
1da177e4 856 }
db8facfc
DH
857#ifdef CONFIG_WATCH_QUEUE
858 if (pipe->watch_queue)
859 put_watch_queue(pipe->watch_queue);
860#endif
923f4f23
IM
861 if (pipe->tmp_page)
862 __free_page(pipe->tmp_page);
906f9040 863 kfree(pipe->bufs);
923f4f23 864 kfree(pipe);
1da177e4
LT
865}
866
68279f9c 867static struct vfsmount *pipe_mnt __ro_after_init;
341b446b 868
c23fbb6b
ED
869/*
870 * pipefs_dname() is called from d_path().
871 */
872static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
873{
0f60d288 874 return dynamic_dname(buffer, buflen, "pipe:[%lu]",
75c3cfa8 875 d_inode(dentry)->i_ino);
c23fbb6b
ED
876}
877
3ba13d17 878static const struct dentry_operations pipefs_dentry_operations = {
c23fbb6b 879 .d_dname = pipefs_dname,
1da177e4
LT
880};
881
882static struct inode * get_pipe_inode(void)
883{
a209dfc7 884 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
923f4f23 885 struct pipe_inode_info *pipe;
1da177e4
LT
886
887 if (!inode)
888 goto fail_inode;
889
85fe4025
CH
890 inode->i_ino = get_next_ino();
891
7bee130e 892 pipe = alloc_pipe_info();
923f4f23 893 if (!pipe)
1da177e4 894 goto fail_iput;
3a326a2c 895
ba5bb147
AV
896 inode->i_pipe = pipe;
897 pipe->files = 2;
923f4f23 898 pipe->readers = pipe->writers = 1;
599a0ac1 899 inode->i_fop = &pipefifo_fops;
1da177e4
LT
900
901 /*
902 * Mark the inode dirty from the very beginning,
903 * that way it will never be moved to the dirty
904 * list because "mark_inode_dirty()" will think
905 * that it already _is_ on the dirty list.
906 */
907 inode->i_state = I_DIRTY;
908 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
da9592ed
DH
909 inode->i_uid = current_fsuid();
910 inode->i_gid = current_fsgid();
16a94965 911 simple_inode_init_ts(inode);
923f4f23 912
1da177e4
LT
913 return inode;
914
915fail_iput:
916 iput(inode);
341b446b 917
1da177e4
LT
918fail_inode:
919 return NULL;
920}
921
e4fad8e5 922int create_pipe_files(struct file **res, int flags)
1da177e4 923{
e4fad8e5 924 struct inode *inode = get_pipe_inode();
d6cbd281 925 struct file *f;
8a018eb5 926 int error;
1da177e4 927
1da177e4 928 if (!inode)
e4fad8e5 929 return -ENFILE;
1da177e4 930
c73be61c 931 if (flags & O_NOTIFICATION_PIPE) {
8a018eb5
QC
932 error = watch_queue_init(inode->i_pipe);
933 if (error) {
934 free_pipe_info(inode->i_pipe);
c73be61c 935 iput(inode);
8a018eb5 936 return error;
c73be61c 937 }
c73be61c
DH
938 }
939
152b6372
AV
940 f = alloc_file_pseudo(inode, pipe_mnt, "",
941 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
942 &pipefifo_fops);
e9bb1f9b 943 if (IS_ERR(f)) {
152b6372
AV
944 free_pipe_info(inode->i_pipe);
945 iput(inode);
946 return PTR_ERR(f);
e9bb1f9b 947 }
341b446b 948
de32ec4c 949 f->private_data = inode->i_pipe;
d6cbd281 950
183266f2
AV
951 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
952 &pipefifo_fops);
e9bb1f9b 953 if (IS_ERR(res[0])) {
b10a4a9f
AV
954 put_pipe_info(inode, inode->i_pipe);
955 fput(f);
956 return PTR_ERR(res[0]);
e9bb1f9b 957 }
de32ec4c 958 res[0]->private_data = inode->i_pipe;
e4fad8e5 959 res[1] = f;
d8e464ec
LT
960 stream_open(inode, res[0]);
961 stream_open(inode, res[1]);
e4fad8e5 962 return 0;
d6cbd281
AK
963}
964
5b249b1b 965static int __do_pipe_flags(int *fd, struct file **files, int flags)
d6cbd281 966{
d6cbd281
AK
967 int error;
968 int fdw, fdr;
969
c73be61c 970 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
ed8cae8b
UD
971 return -EINVAL;
972
e4fad8e5
AV
973 error = create_pipe_files(files, flags);
974 if (error)
975 return error;
d6cbd281 976
ed8cae8b 977 error = get_unused_fd_flags(flags);
d6cbd281
AK
978 if (error < 0)
979 goto err_read_pipe;
980 fdr = error;
981
ed8cae8b 982 error = get_unused_fd_flags(flags);
d6cbd281
AK
983 if (error < 0)
984 goto err_fdr;
985 fdw = error;
986
157cf649 987 audit_fd_pair(fdr, fdw);
d6cbd281
AK
988 fd[0] = fdr;
989 fd[1] = fdw;
afed6271
JA
990 /* pipe groks IOCB_NOWAIT */
991 files[0]->f_mode |= FMODE_NOWAIT;
992 files[1]->f_mode |= FMODE_NOWAIT;
d6cbd281
AK
993 return 0;
994
995 err_fdr:
996 put_unused_fd(fdr);
997 err_read_pipe:
e4fad8e5
AV
998 fput(files[0]);
999 fput(files[1]);
d6cbd281 1000 return error;
1da177e4
LT
1001}
1002
5b249b1b
AV
1003int do_pipe_flags(int *fd, int flags)
1004{
1005 struct file *files[2];
1006 int error = __do_pipe_flags(fd, files, flags);
1007 if (!error) {
1008 fd_install(fd[0], files[0]);
1009 fd_install(fd[1], files[1]);
1010 }
1011 return error;
1012}
1013
d35c7b0e
UD
1014/*
1015 * sys_pipe() is the normal C calling standard for creating
1016 * a pipe. It's not the way Unix traditionally does this, though.
1017 */
0a216dd1 1018static int do_pipe2(int __user *fildes, int flags)
d35c7b0e 1019{
5b249b1b 1020 struct file *files[2];
d35c7b0e
UD
1021 int fd[2];
1022 int error;
1023
5b249b1b 1024 error = __do_pipe_flags(fd, files, flags);
d35c7b0e 1025 if (!error) {
5b249b1b
AV
1026 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1027 fput(files[0]);
1028 fput(files[1]);
1029 put_unused_fd(fd[0]);
1030 put_unused_fd(fd[1]);
d35c7b0e 1031 error = -EFAULT;
5b249b1b
AV
1032 } else {
1033 fd_install(fd[0], files[0]);
1034 fd_install(fd[1], files[1]);
ba719bae 1035 }
d35c7b0e
UD
1036 }
1037 return error;
1038}
1039
0a216dd1
DB
1040SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1041{
1042 return do_pipe2(fildes, flags);
1043}
1044
2b664219 1045SYSCALL_DEFINE1(pipe, int __user *, fildes)
ed8cae8b 1046{
0a216dd1 1047 return do_pipe2(fildes, 0);
ed8cae8b
UD
1048}
1049
472e5b05
LT
1050/*
1051 * This is the stupid "wait for pipe to be readable or writable"
1052 * model.
1053 *
1054 * See pipe_read/write() for the proper kind of exclusive wait,
1055 * but that requires that we wake up any other readers/writers
1056 * if we then do not end up reading everything (ie the whole
1057 * "wake_next_reader/writer" logic in pipe_read/write()).
1058 */
1059void pipe_wait_readable(struct pipe_inode_info *pipe)
1060{
1061 pipe_unlock(pipe);
1062 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1063 pipe_lock(pipe);
1064}
1065
1066void pipe_wait_writable(struct pipe_inode_info *pipe)
1067{
1068 pipe_unlock(pipe);
1069 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1070 pipe_lock(pipe);
1071}
1072
1073/*
1074 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1075 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1076 * race with the count check and waitqueue prep.
1077 *
1078 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1079 * then check the condition you're waiting for, and only then sleep. But
1080 * because of the pipe lock, we can check the condition before being on
1081 * the wait queue.
1082 *
1083 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1084 */
fc7478a2 1085static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
f776c738 1086{
472e5b05 1087 DEFINE_WAIT(rdwait);
8cefc107 1088 int cur = *cnt;
f776c738
AV
1089
1090 while (cur == *cnt) {
472e5b05
LT
1091 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1092 pipe_unlock(pipe);
1093 schedule();
1094 finish_wait(&pipe->rd_wait, &rdwait);
1095 pipe_lock(pipe);
f776c738
AV
1096 if (signal_pending(current))
1097 break;
1098 }
1099 return cur == *cnt ? -ERESTARTSYS : 0;
1100}
1101
fc7478a2 1102static void wake_up_partner(struct pipe_inode_info *pipe)
f776c738 1103{
6551d5c5 1104 wake_up_interruptible_all(&pipe->rd_wait);
f776c738
AV
1105}
1106
1107static int fifo_open(struct inode *inode, struct file *filp)
1108{
1109 struct pipe_inode_info *pipe;
599a0ac1 1110 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
f776c738
AV
1111 int ret;
1112
ba5bb147
AV
1113 filp->f_version = 0;
1114
1115 spin_lock(&inode->i_lock);
1116 if (inode->i_pipe) {
1117 pipe = inode->i_pipe;
1118 pipe->files++;
1119 spin_unlock(&inode->i_lock);
1120 } else {
1121 spin_unlock(&inode->i_lock);
7bee130e 1122 pipe = alloc_pipe_info();
f776c738 1123 if (!pipe)
ba5bb147
AV
1124 return -ENOMEM;
1125 pipe->files = 1;
1126 spin_lock(&inode->i_lock);
1127 if (unlikely(inode->i_pipe)) {
1128 inode->i_pipe->files++;
1129 spin_unlock(&inode->i_lock);
4b8a8f1e 1130 free_pipe_info(pipe);
ba5bb147
AV
1131 pipe = inode->i_pipe;
1132 } else {
1133 inode->i_pipe = pipe;
1134 spin_unlock(&inode->i_lock);
1135 }
f776c738 1136 }
de32ec4c 1137 filp->private_data = pipe;
ba5bb147
AV
1138 /* OK, we have a pipe and it's pinned down */
1139
ebec73f4 1140 __pipe_lock(pipe);
f776c738
AV
1141
1142 /* We can only do regular read/write on fifos */
d8e464ec 1143 stream_open(inode, filp);
f776c738 1144
d8e464ec 1145 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
f776c738
AV
1146 case FMODE_READ:
1147 /*
1148 * O_RDONLY
1149 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1150 * opened, even when there is no process writing the FIFO.
1151 */
f776c738
AV
1152 pipe->r_counter++;
1153 if (pipe->readers++ == 0)
fc7478a2 1154 wake_up_partner(pipe);
f776c738 1155
599a0ac1 1156 if (!is_pipe && !pipe->writers) {
f776c738 1157 if ((filp->f_flags & O_NONBLOCK)) {
a9a08845 1158 /* suppress EPOLLHUP until we have
f776c738
AV
1159 * seen a writer */
1160 filp->f_version = pipe->w_counter;
1161 } else {
fc7478a2 1162 if (wait_for_partner(pipe, &pipe->w_counter))
f776c738
AV
1163 goto err_rd;
1164 }
1165 }
1166 break;
8cefc107 1167
f776c738
AV
1168 case FMODE_WRITE:
1169 /*
1170 * O_WRONLY
1171 * POSIX.1 says that O_NONBLOCK means return -1 with
1172 * errno=ENXIO when there is no process reading the FIFO.
1173 */
1174 ret = -ENXIO;
599a0ac1 1175 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
f776c738
AV
1176 goto err;
1177
f776c738
AV
1178 pipe->w_counter++;
1179 if (!pipe->writers++)
fc7478a2 1180 wake_up_partner(pipe);
f776c738 1181
599a0ac1 1182 if (!is_pipe && !pipe->readers) {
fc7478a2 1183 if (wait_for_partner(pipe, &pipe->r_counter))
f776c738
AV
1184 goto err_wr;
1185 }
1186 break;
8cefc107 1187
f776c738
AV
1188 case FMODE_READ | FMODE_WRITE:
1189 /*
1190 * O_RDWR
1191 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1192 * This implementation will NEVER block on a O_RDWR open, since
1193 * the process can at least talk to itself.
1194 */
f776c738
AV
1195
1196 pipe->readers++;
1197 pipe->writers++;
1198 pipe->r_counter++;
1199 pipe->w_counter++;
1200 if (pipe->readers == 1 || pipe->writers == 1)
fc7478a2 1201 wake_up_partner(pipe);
f776c738
AV
1202 break;
1203
1204 default:
1205 ret = -EINVAL;
1206 goto err;
1207 }
1208
1209 /* Ok! */
ebec73f4 1210 __pipe_unlock(pipe);
f776c738
AV
1211 return 0;
1212
1213err_rd:
1214 if (!--pipe->readers)
0ddad21d 1215 wake_up_interruptible(&pipe->wr_wait);
f776c738
AV
1216 ret = -ERESTARTSYS;
1217 goto err;
1218
1219err_wr:
1220 if (!--pipe->writers)
6551d5c5 1221 wake_up_interruptible_all(&pipe->rd_wait);
f776c738
AV
1222 ret = -ERESTARTSYS;
1223 goto err;
1224
1225err:
ebec73f4 1226 __pipe_unlock(pipe);
b0d8d229
LT
1227
1228 put_pipe_info(inode, pipe);
f776c738
AV
1229 return ret;
1230}
1231
599a0ac1
AV
1232const struct file_operations pipefifo_fops = {
1233 .open = fifo_open,
1234 .llseek = no_llseek,
fb9096a3 1235 .read_iter = pipe_read,
f0d1bec9 1236 .write_iter = pipe_write,
a11e1d43 1237 .poll = pipe_poll,
599a0ac1
AV
1238 .unlocked_ioctl = pipe_ioctl,
1239 .release = pipe_release,
1240 .fasync = pipe_fasync,
f8ad8187 1241 .splice_write = iter_file_splice_write,
f776c738
AV
1242};
1243
f491bd71
MK
1244/*
1245 * Currently we rely on the pipe array holding a power-of-2 number
d3f14c48 1246 * of pages. Returns 0 on error.
f491bd71 1247 */
515c5046 1248unsigned int round_pipe_size(unsigned int size)
f491bd71 1249{
c4fed5a9 1250 if (size > (1U << 31))
96e99be4
EB
1251 return 0;
1252
4c2e4bef
EB
1253 /* Minimum pipe size, as required by POSIX */
1254 if (size < PAGE_SIZE)
c4fed5a9 1255 return PAGE_SIZE;
d3f14c48 1256
c4fed5a9 1257 return roundup_pow_of_two(size);
f491bd71
MK
1258}
1259
35f3d14d 1260/*
c73be61c 1261 * Resize the pipe ring to a number of slots.
189b0ddc
DH
1262 *
1263 * Note the pipe can be reduced in capacity, but only if the current
1264 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1265 * returned instead.
35f3d14d 1266 */
c73be61c 1267int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
35f3d14d
JA
1268{
1269 struct pipe_buffer *bufs;
c73be61c 1270 unsigned int head, tail, mask, n;
35f3d14d 1271
906f9040
LT
1272 bufs = kcalloc(nr_slots, sizeof(*bufs),
1273 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
c73be61c
DH
1274 if (unlikely(!bufs))
1275 return -ENOMEM;
35f3d14d 1276
189b0ddc
DH
1277 spin_lock_irq(&pipe->rd_wait.lock);
1278 mask = pipe->ring_size - 1;
1279 head = pipe->head;
1280 tail = pipe->tail;
1281
1282 n = pipe_occupancy(head, tail);
1283 if (nr_slots < n) {
1284 spin_unlock_irq(&pipe->rd_wait.lock);
1285 kfree(bufs);
1286 return -EBUSY;
1287 }
1288
35f3d14d
JA
1289 /*
1290 * The pipe array wraps around, so just start the new one at zero
8cefc107 1291 * and adjust the indices.
35f3d14d 1292 */
8cefc107
DH
1293 if (n > 0) {
1294 unsigned int h = head & mask;
1295 unsigned int t = tail & mask;
1296 if (h > t) {
1297 memcpy(bufs, pipe->bufs + t,
1298 n * sizeof(struct pipe_buffer));
1299 } else {
1300 unsigned int tsize = pipe->ring_size - t;
1301 if (h > 0)
1302 memcpy(bufs + tsize, pipe->bufs,
1303 h * sizeof(struct pipe_buffer));
1304 memcpy(bufs, pipe->bufs + t,
1305 tsize * sizeof(struct pipe_buffer));
1306 }
35f3d14d
JA
1307 }
1308
8cefc107
DH
1309 head = n;
1310 tail = 0;
1311
906f9040 1312 kfree(pipe->bufs);
35f3d14d 1313 pipe->bufs = bufs;
8cefc107 1314 pipe->ring_size = nr_slots;
c73be61c
DH
1315 if (pipe->max_usage > nr_slots)
1316 pipe->max_usage = nr_slots;
8cefc107
DH
1317 pipe->tail = tail;
1318 pipe->head = head;
6551d5c5 1319
189b0ddc
DH
1320 spin_unlock_irq(&pipe->rd_wait.lock);
1321
6551d5c5
LT
1322 /* This might have made more room for writers */
1323 wake_up_interruptible(&pipe->wr_wait);
c73be61c
DH
1324 return 0;
1325}
1326
1327/*
1328 * Allocate a new array of pipe buffers and copy the info over. Returns the
1329 * pipe size if successful, or return -ERROR on error.
1330 */
515c5046 1331static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
c73be61c
DH
1332{
1333 unsigned long user_bufs;
1334 unsigned int nr_slots, size;
1335 long ret = 0;
1336
b4bd6b4b 1337 if (pipe_has_watch_queue(pipe))
c73be61c 1338 return -EBUSY;
c73be61c
DH
1339
1340 size = round_pipe_size(arg);
1341 nr_slots = size >> PAGE_SHIFT;
1342
1343 if (!nr_slots)
1344 return -EINVAL;
1345
1346 /*
1347 * If trying to increase the pipe capacity, check that an
1348 * unprivileged user is not trying to exceed various limits
1349 * (soft limit check here, hard limit check just below).
1350 * Decreasing the pipe capacity is always permitted, even
1351 * if the user is currently over a limit.
1352 */
1353 if (nr_slots > pipe->max_usage &&
1354 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1355 return -EPERM;
1356
1357 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1358
1359 if (nr_slots > pipe->max_usage &&
1360 (too_many_pipe_buffers_hard(user_bufs) ||
1361 too_many_pipe_buffers_soft(user_bufs)) &&
1362 pipe_is_unprivileged_user()) {
1363 ret = -EPERM;
1364 goto out_revert_acct;
1365 }
1366
1367 ret = pipe_resize_ring(pipe, nr_slots);
1368 if (ret < 0)
1369 goto out_revert_acct;
1370
1371 pipe->max_usage = nr_slots;
1372 pipe->nr_accounted = nr_slots;
6718b6f8 1373 return pipe->max_usage * PAGE_SIZE;
b0b91d18
MK
1374
1375out_revert_acct:
c73be61c 1376 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
b0b91d18 1377 return ret;
35f3d14d
JA
1378}
1379
72083646 1380/*
4e7b5671
CH
1381 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1382 * not enough to verify that this is a pipe.
72083646 1383 */
c73be61c 1384struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
72083646 1385{
c73be61c
DH
1386 struct pipe_inode_info *pipe = file->private_data;
1387
1388 if (file->f_op != &pipefifo_fops || !pipe)
1389 return NULL;
b4bd6b4b 1390 if (for_splice && pipe_has_watch_queue(pipe))
c73be61c 1391 return NULL;
c73be61c 1392 return pipe;
72083646
LT
1393}
1394
515c5046 1395long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
35f3d14d
JA
1396{
1397 struct pipe_inode_info *pipe;
1398 long ret;
1399
c73be61c 1400 pipe = get_pipe_info(file, false);
35f3d14d
JA
1401 if (!pipe)
1402 return -EBADF;
1403
ebec73f4 1404 __pipe_lock(pipe);
35f3d14d
JA
1405
1406 switch (cmd) {
d37d4166
MK
1407 case F_SETPIPE_SZ:
1408 ret = pipe_set_size(pipe, arg);
35f3d14d
JA
1409 break;
1410 case F_GETPIPE_SZ:
6718b6f8 1411 ret = pipe->max_usage * PAGE_SIZE;
35f3d14d
JA
1412 break;
1413 default:
1414 ret = -EINVAL;
1415 break;
1416 }
1417
ebec73f4 1418 __pipe_unlock(pipe);
35f3d14d
JA
1419 return ret;
1420}
1421
ff0c7d15
NP
1422static const struct super_operations pipefs_ops = {
1423 .destroy_inode = free_inode_nonrcu,
d70ef97b 1424 .statfs = simple_statfs,
ff0c7d15
NP
1425};
1426
1da177e4
LT
1427/*
1428 * pipefs should _never_ be mounted by userland - too much of security hassle,
1429 * no real gain from having the whole whorehouse mounted. So we don't need
1430 * any operations on the root directory. However, we need a non-trivial
1431 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1432 */
4fa7ec5d
DH
1433
1434static int pipefs_init_fs_context(struct fs_context *fc)
1da177e4 1435{
4fa7ec5d
DH
1436 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1437 if (!ctx)
1438 return -ENOMEM;
1439 ctx->ops = &pipefs_ops;
1440 ctx->dops = &pipefs_dentry_operations;
1441 return 0;
1da177e4
LT
1442}
1443
1444static struct file_system_type pipe_fs_type = {
1445 .name = "pipefs",
4fa7ec5d 1446 .init_fs_context = pipefs_init_fs_context,
1da177e4
LT
1447 .kill_sb = kill_anon_super,
1448};
1449
1998f193
LC
1450#ifdef CONFIG_SYSCTL
1451static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1452 unsigned int *valp,
1453 int write, void *data)
1454{
1455 if (write) {
1456 unsigned int val;
1457
1458 val = round_pipe_size(*lvalp);
1459 if (val == 0)
1460 return -EINVAL;
1461
1462 *valp = val;
1463 } else {
1464 unsigned int val = *valp;
1465 *lvalp = (unsigned long) val;
1466 }
1467
1468 return 0;
1469}
1470
1471static int proc_dopipe_max_size(struct ctl_table *table, int write,
1472 void *buffer, size_t *lenp, loff_t *ppos)
1473{
1474 return do_proc_douintvec(table, write, buffer, lenp, ppos,
1475 do_proc_dopipe_max_size_conv, NULL);
1476}
1477
1478static struct ctl_table fs_pipe_sysctls[] = {
1479 {
1480 .procname = "pipe-max-size",
1481 .data = &pipe_max_size,
1482 .maxlen = sizeof(pipe_max_size),
1483 .mode = 0644,
1484 .proc_handler = proc_dopipe_max_size,
1485 },
1486 {
1487 .procname = "pipe-user-pages-hard",
1488 .data = &pipe_user_pages_hard,
1489 .maxlen = sizeof(pipe_user_pages_hard),
1490 .mode = 0644,
1491 .proc_handler = proc_doulongvec_minmax,
1492 },
1493 {
1494 .procname = "pipe-user-pages-soft",
1495 .data = &pipe_user_pages_soft,
1496 .maxlen = sizeof(pipe_user_pages_soft),
1497 .mode = 0644,
1498 .proc_handler = proc_doulongvec_minmax,
1499 },
1500 { }
1501};
1502#endif
1503
1da177e4
LT
1504static int __init init_pipe_fs(void)
1505{
1506 int err = register_filesystem(&pipe_fs_type);
341b446b 1507
1da177e4
LT
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 }
1998f193
LC
1515#ifdef CONFIG_SYSCTL
1516 register_sysctl_init("fs", fs_pipe_sysctls);
1517#endif
1da177e4
LT
1518 return err;
1519}
1520
1da177e4 1521fs_initcall(init_pipe_fs);