3  *  linux/fs/pipe.c
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
39  * than two, then a write to a non-empty pipe may block even if the pipe is not
42  * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
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
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
89 void pipe_lock(struct pipe_inode_info *pipe)  in pipe_lock()  argument
91 	if (pipe->files)  in pipe_lock()
92 		mutex_lock(&pipe->mutex);  in pipe_lock()
96 void pipe_unlock(struct pipe_inode_info *pipe)  in pipe_unlock()  argument
98 	if (pipe->files)  in pipe_unlock()
99 		mutex_unlock(&pipe->mutex);  in pipe_unlock()
115 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,  in anon_pipe_buf_release()  argument
125 	if (page_count(page) == 1 && !pipe->tmp_page)  in anon_pipe_buf_release()
126 		pipe->tmp_page = page;  in anon_pipe_buf_release()
131 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,  in anon_pipe_buf_try_steal()  argument
145  * @pipe:	the pipe that the buffer belongs to
155 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,  in generic_pipe_buf_try_steal()  argument
175  * @pipe:	the pipe that the buffer belongs to
181  *	pipe into another.
183 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)  in generic_pipe_buf_get()  argument
191  * @pipe:	the pipe that the buffer belongs to
197 void generic_pipe_buf_release(struct pipe_inode_info *pipe,  in generic_pipe_buf_release()  argument
210 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
211 static inline bool pipe_readable(const struct pipe_inode_info *pipe)  in pipe_readable()  argument
213 	unsigned int head = READ_ONCE(pipe->head);  in pipe_readable()
214 	unsigned int tail = READ_ONCE(pipe->tail);  in pipe_readable()
215 	unsigned int writers = READ_ONCE(pipe->writers);  in pipe_readable()
220 static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,  in pipe_update_tail()  argument
224 	pipe_buf_release(pipe, buf);  in pipe_update_tail()
227 	 * If the pipe has a watch_queue, we need additional protection  in pipe_update_tail()
231 	if (pipe_has_watch_queue(pipe)) {  in pipe_update_tail()
232 		spin_lock_irq(&pipe->rd_wait.lock);  in pipe_update_tail()
235 			pipe->note_loss = true;  in pipe_update_tail()
237 		pipe->tail = ++tail;  in pipe_update_tail()
238 		spin_unlock_irq(&pipe->rd_wait.lock);  in pipe_update_tail()
246 	pipe->tail = ++tail;  in pipe_update_tail()
255 	struct pipe_inode_info *pipe = filp->private_data;  in pipe_read()  local
264 	mutex_lock(&pipe->mutex);  in pipe_read()
267 	 * We only wake up writers if the pipe was full when we started  in pipe_read()
274 	was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);  in pipe_read()
277 		unsigned int head = smp_load_acquire(&pipe->head);  in pipe_read()
278 		unsigned int tail = pipe->tail;  in pipe_read()
279 		unsigned int mask = pipe->ring_size - 1;  in pipe_read()
282 		if (pipe->note_loss) {  in pipe_read()
301 			pipe->note_loss = false;  in pipe_read()
306 			struct pipe_buffer *buf = &pipe->bufs[tail & mask];  in pipe_read()
320 			error = pipe_buf_confirm(pipe, buf);  in pipe_read()
344 				tail = pipe_update_tail(pipe, buf, tail);  in pipe_read()
352 		if (!pipe->writers)  in pipe_read()
361 		mutex_unlock(&pipe->mutex);  in pipe_read()
367 		 * pipe buffer, and might have made space in the buffers  in pipe_read()
370 		 * You can't make zero-sized pipe buffers by doing an empty  in pipe_read()
377 		 * _very_ unlikely case that the pipe was full, but we got  in pipe_read()
381 			wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);  in pipe_read()
382 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);  in pipe_read()
390 		if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)  in pipe_read()
393 		mutex_lock(&pipe->mutex);  in pipe_read()
394 		was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);  in pipe_read()
397 	if (pipe_empty(pipe->head, pipe->tail))  in pipe_read()
399 	mutex_unlock(&pipe->mutex);  in pipe_read()
402 		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);  in pipe_read()
404 		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);  in pipe_read()
405 	kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);  in pipe_read()
416 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
417 static inline bool pipe_writable(const struct pipe_inode_info *pipe)  in pipe_writable()  argument
419 	unsigned int head = READ_ONCE(pipe->head);  in pipe_writable()
420 	unsigned int tail = READ_ONCE(pipe->tail);  in pipe_writable()
421 	unsigned int max_usage = READ_ONCE(pipe->max_usage);  in pipe_writable()
424 		!READ_ONCE(pipe->readers);  in pipe_writable()
431 	struct pipe_inode_info *pipe = filp->private_data;  in pipe_write()  local
441 	 * the pipe.  in pipe_write()
443 	 * pipe locked.  in pipe_write()
444 	 * If we had to support locking a normal pipe and a notification pipe at  in pipe_write()
448 	if (pipe_has_watch_queue(pipe))  in pipe_write()
455 	mutex_lock(&pipe->mutex);  in pipe_write()
457 	if (!pipe->readers) {  in pipe_write()
471 	head = pipe->head;  in pipe_write()
472 	was_empty = pipe_empty(head, pipe->tail);  in pipe_write()
475 		unsigned int mask = pipe->ring_size - 1;  in pipe_write()
476 		struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];  in pipe_write()
481 			ret = pipe_buf_confirm(pipe, buf);  in pipe_write()
498 		if (!pipe->readers) {  in pipe_write()
505 		head = pipe->head;  in pipe_write()
506 		if (!pipe_full(head, pipe->tail, pipe->max_usage)) {  in pipe_write()
507 			unsigned int mask = pipe->ring_size - 1;  in pipe_write()
509 			struct page *page = pipe->tmp_page;  in pipe_write()
518 				pipe->tmp_page = page;  in pipe_write()
526 			pipe->head = head + 1;  in pipe_write()
529 			buf = &pipe->bufs[head & mask];  in pipe_write()
538 			pipe->tmp_page = NULL;  in pipe_write()
553 		if (!pipe_full(head, pipe->tail, pipe->max_usage))  in pipe_write()
570 		 * We're going to release the pipe lock and wait for more  in pipe_write()
572 		 * after waiting we need to re-check whether the pipe  in pipe_write()
575 		mutex_unlock(&pipe->mutex);  in pipe_write()
577 			wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);  in pipe_write()
578 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);  in pipe_write()
579 		wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));  in pipe_write()
580 		mutex_lock(&pipe->mutex);  in pipe_write()
581 		was_empty = pipe_empty(pipe->head, pipe->tail);  in pipe_write()
585 	if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))  in pipe_write()
587 	mutex_unlock(&pipe->mutex);  in pipe_write()
598 	 * Epoll nonsensically wants a wakeup whether the pipe  in pipe_write()
601 	if (was_empty || pipe->poll_usage)  in pipe_write()
602 		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);  in pipe_write()
603 	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);  in pipe_write()
605 		wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);  in pipe_write()
617 	struct pipe_inode_info *pipe = filp->private_data;  in pipe_ioctl()  local
622 		mutex_lock(&pipe->mutex);  in pipe_ioctl()
624 		head = pipe->head;  in pipe_ioctl()
625 		tail = pipe->tail;  in pipe_ioctl()
626 		mask = pipe->ring_size - 1;  in pipe_ioctl()
629 			count += pipe->bufs[tail & mask].len;  in pipe_ioctl()
632 		mutex_unlock(&pipe->mutex);  in pipe_ioctl()
639 		mutex_lock(&pipe->mutex);  in pipe_ioctl()
640 		ret = watch_queue_set_size(pipe, arg);  in pipe_ioctl()
641 		mutex_unlock(&pipe->mutex);  in pipe_ioctl()
647 			pipe, (struct watch_notification_filter __user *)arg);  in pipe_ioctl()
660 	struct pipe_inode_info *pipe = filp->private_data;  in pipe_poll()  local
664 	WRITE_ONCE(pipe->poll_usage, true);  in pipe_poll()
667 	 * Reading pipe state only -- no need for acquiring the semaphore.  in pipe_poll()
673 		poll_wait(filp, &pipe->rd_wait, wait);  in pipe_poll()
675 		poll_wait(filp, &pipe->wr_wait, wait);  in pipe_poll()
682 	head = READ_ONCE(pipe->head);  in pipe_poll()
683 	tail = READ_ONCE(pipe->tail);  in pipe_poll()
689 		if (!pipe->writers && filp->f_pipe != pipe->w_counter)  in pipe_poll()
694 		if (!pipe_full(head, tail, pipe->max_usage))  in pipe_poll()
700 		if (!pipe->readers)  in pipe_poll()
707 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)  in put_pipe_info()  argument
712 	if (!--pipe->files) {  in put_pipe_info()
719 		free_pipe_info(pipe);  in put_pipe_info()
725 	struct pipe_inode_info *pipe = file->private_data;  in pipe_release()  local
727 	mutex_lock(&pipe->mutex);  in pipe_release()
729 		pipe->readers--;  in pipe_release()
731 		pipe->writers--;  in pipe_release()
734 	if (!pipe->readers != !pipe->writers) {  in pipe_release()
735 		wake_up_interruptible_all(&pipe->rd_wait);  in pipe_release()
736 		wake_up_interruptible_all(&pipe->wr_wait);  in pipe_release()
737 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);  in pipe_release()
738 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);  in pipe_release()
740 	mutex_unlock(&pipe->mutex);  in pipe_release()
742 	put_pipe_info(inode, pipe);  in pipe_release()
749 	struct pipe_inode_info *pipe = filp->private_data;  in pipe_fasync()  local
752 	mutex_lock(&pipe->mutex);  in pipe_fasync()
754 		retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);  in pipe_fasync()
756 		retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);  in pipe_fasync()
759 			fasync_helper(-1, filp, 0, &pipe->fasync_readers);  in pipe_fasync()
761 	mutex_unlock(&pipe->mutex);  in pipe_fasync()
792 	struct pipe_inode_info *pipe;  in alloc_pipe_info()  local
798 	pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);  in alloc_pipe_info()
799 	if (pipe == NULL)  in alloc_pipe_info()
815 	pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),  in alloc_pipe_info()
818 	if (pipe->bufs) {  in alloc_pipe_info()
819 		init_waitqueue_head(&pipe->rd_wait);  in alloc_pipe_info()
820 		init_waitqueue_head(&pipe->wr_wait);  in alloc_pipe_info()
821 		pipe->r_counter = pipe->w_counter = 1;  in alloc_pipe_info()
822 		pipe->max_usage = pipe_bufs;  in alloc_pipe_info()
823 		pipe->ring_size = pipe_bufs;  in alloc_pipe_info()
824 		pipe->nr_accounted = pipe_bufs;  in alloc_pipe_info()
825 		pipe->user = user;  in alloc_pipe_info()
826 		mutex_init(&pipe->mutex);  in alloc_pipe_info()
827 		lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);  in alloc_pipe_info()
828 		return pipe;  in alloc_pipe_info()
833 	kfree(pipe);  in alloc_pipe_info()
839 void free_pipe_info(struct pipe_inode_info *pipe)  in free_pipe_info()  argument
844 	if (pipe->watch_queue)  in free_pipe_info()
845 		watch_queue_clear(pipe->watch_queue);  in free_pipe_info()
848 	(void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);  in free_pipe_info()
849 	free_uid(pipe->user);  in free_pipe_info()
850 	for (i = 0; i < pipe->ring_size; i++) {  in free_pipe_info()
851 		struct pipe_buffer *buf = pipe->bufs + i;  in free_pipe_info()
853 			pipe_buf_release(pipe, buf);  in free_pipe_info()
856 	if (pipe->watch_queue)  in free_pipe_info()
857 		put_watch_queue(pipe->watch_queue);  in free_pipe_info()
859 	if (pipe->tmp_page)  in free_pipe_info()
860 		__free_page(pipe->tmp_page);  in free_pipe_info()
861 	kfree(pipe->bufs);  in free_pipe_info()
862 	kfree(pipe);  in free_pipe_info()
872 	return dynamic_dname(buffer, buflen, "pipe:[%lu]",  in pipefs_dname()
883 	struct pipe_inode_info *pipe;  in get_pipe_inode()  local
890 	pipe = alloc_pipe_info();  in get_pipe_inode()
891 	if (!pipe)  in get_pipe_inode()
894 	inode->i_pipe = pipe;  in get_pipe_inode()
895 	pipe->files = 2;  in get_pipe_inode()
896 	pipe->readers = pipe->writers = 1;  in get_pipe_inode()
990 	/* pipe groks IOCB_NOWAIT */  in __do_pipe_flags()
1016  * a pipe. It's not the way Unix traditionally does this, though.
1045 SYSCALL_DEFINE1(pipe, int __user *, fildes)  in SYSCALL_DEFINE1()  argument
1051  * This is the stupid "wait for pipe to be readable or writable"
1059 void pipe_wait_readable(struct pipe_inode_info *pipe)  in pipe_wait_readable()  argument
1061 	pipe_unlock(pipe);  in pipe_wait_readable()
1062 	wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));  in pipe_wait_readable()
1063 	pipe_lock(pipe);  in pipe_wait_readable()
1066 void pipe_wait_writable(struct pipe_inode_info *pipe)  in pipe_wait_writable()  argument
1068 	pipe_unlock(pipe);  in pipe_wait_writable()
1069 	wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));  in pipe_wait_writable()
1070 	pipe_lock(pipe);  in pipe_wait_writable()
1075  * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1080  * because of the pipe lock, we can check the condition before being on
1083  * We use the 'rd_wait' waitqueue for pipe partner waiting.
1085 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)  in wait_for_partner()  argument
1091 		prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);  in wait_for_partner()
1092 		pipe_unlock(pipe);  in wait_for_partner()
1094 		finish_wait(&pipe->rd_wait, &rdwait);  in wait_for_partner()
1095 		pipe_lock(pipe);  in wait_for_partner()
1102 static void wake_up_partner(struct pipe_inode_info *pipe)  in wake_up_partner()  argument
1104 	wake_up_interruptible_all(&pipe->rd_wait);  in wake_up_partner()
1109 	struct pipe_inode_info *pipe;  in fifo_open()  local
1117 		pipe = inode->i_pipe;  in fifo_open()
1118 		pipe->files++;  in fifo_open()
1122 		pipe = alloc_pipe_info();  in fifo_open()
1123 		if (!pipe)  in fifo_open()
1125 		pipe->files = 1;  in fifo_open()
1130 			free_pipe_info(pipe);  in fifo_open()
1131 			pipe = inode->i_pipe;  in fifo_open()
1133 			inode->i_pipe = pipe;  in fifo_open()
1137 	filp->private_data = pipe;  in fifo_open()
1138 	/* OK, we have a pipe and it's pinned down */  in fifo_open()
1140 	mutex_lock(&pipe->mutex);  in fifo_open()
1152 		pipe->r_counter++;  in fifo_open()
1153 		if (pipe->readers++ == 0)  in fifo_open()
1154 			wake_up_partner(pipe);  in fifo_open()
1156 		if (!is_pipe && !pipe->writers) {  in fifo_open()
1160 				filp->f_pipe = pipe->w_counter;  in fifo_open()
1162 				if (wait_for_partner(pipe, &pipe->w_counter))  in fifo_open()
1175 		if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)  in fifo_open()
1178 		pipe->w_counter++;  in fifo_open()
1179 		if (!pipe->writers++)  in fifo_open()
1180 			wake_up_partner(pipe);  in fifo_open()
1182 		if (!is_pipe && !pipe->readers) {  in fifo_open()
1183 			if (wait_for_partner(pipe, &pipe->r_counter))  in fifo_open()
1196 		pipe->readers++;  in fifo_open()
1197 		pipe->writers++;  in fifo_open()
1198 		pipe->r_counter++;  in fifo_open()
1199 		pipe->w_counter++;  in fifo_open()
1200 		if (pipe->readers == 1 || pipe->writers == 1)  in fifo_open()
1201 			wake_up_partner(pipe);  in fifo_open()
1210 	mutex_unlock(&pipe->mutex);  in fifo_open()
1214 	if (!--pipe->readers)  in fifo_open()
1215 		wake_up_interruptible(&pipe->wr_wait);  in fifo_open()
1220 	if (!--pipe->writers)  in fifo_open()
1221 		wake_up_interruptible_all(&pipe->rd_wait);  in fifo_open()
1226 	mutex_unlock(&pipe->mutex);  in fifo_open()
1228 	put_pipe_info(inode, pipe);  in fifo_open()
1244  * Currently we rely on the pipe array holding a power-of-2 number
1252 	/* Minimum pipe size, as required by POSIX */  in round_pipe_size()
1260  * Resize the pipe ring to a number of slots.
1262  * Note the pipe can be reduced in capacity, but only if the current
1266 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)  in pipe_resize_ring()  argument
1276 	spin_lock_irq(&pipe->rd_wait.lock);  in pipe_resize_ring()
1277 	mask = pipe->ring_size - 1;  in pipe_resize_ring()
1278 	head = pipe->head;  in pipe_resize_ring()
1279 	tail = pipe->tail;  in pipe_resize_ring()
1283 		spin_unlock_irq(&pipe->rd_wait.lock);  in pipe_resize_ring()
1289 	 * The pipe array wraps around, so just start the new one at zero  in pipe_resize_ring()
1296 			memcpy(bufs, pipe->bufs + t,  in pipe_resize_ring()
1299 			unsigned int tsize = pipe->ring_size - t;  in pipe_resize_ring()
1301 				memcpy(bufs + tsize, pipe->bufs,  in pipe_resize_ring()
1303 			memcpy(bufs, pipe->bufs + t,  in pipe_resize_ring()
1311 	kfree(pipe->bufs);  in pipe_resize_ring()
1312 	pipe->bufs = bufs;  in pipe_resize_ring()
1313 	pipe->ring_size = nr_slots;  in pipe_resize_ring()
1314 	if (pipe->max_usage > nr_slots)  in pipe_resize_ring()
1315 		pipe->max_usage = nr_slots;  in pipe_resize_ring()
1316 	pipe->tail = tail;  in pipe_resize_ring()
1317 	pipe->head = head;  in pipe_resize_ring()
1319 	if (!pipe_has_watch_queue(pipe)) {  in pipe_resize_ring()
1320 		pipe->max_usage = nr_slots;  in pipe_resize_ring()
1321 		pipe->nr_accounted = nr_slots;  in pipe_resize_ring()
1324 	spin_unlock_irq(&pipe->rd_wait.lock);  in pipe_resize_ring()
1327 	wake_up_interruptible(&pipe->wr_wait);  in pipe_resize_ring()
1332  * Allocate a new array of pipe buffers and copy the info over. Returns the
1333  * pipe size if successful, or return -ERROR on error.
1335 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)  in pipe_set_size()  argument
1341 	if (pipe_has_watch_queue(pipe))  in pipe_set_size()
1351 	 * If trying to increase the pipe capacity, check that an  in pipe_set_size()
1354 	 * Decreasing the pipe capacity is always permitted, even  in pipe_set_size()
1357 	if (nr_slots > pipe->max_usage &&  in pipe_set_size()
1361 	user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);  in pipe_set_size()
1363 	if (nr_slots > pipe->max_usage &&  in pipe_set_size()
1371 	ret = pipe_resize_ring(pipe, nr_slots);  in pipe_set_size()
1375 	return pipe->max_usage * PAGE_SIZE;  in pipe_set_size()
1378 	(void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);  in pipe_set_size()
1384  * not enough to verify that this is a pipe.
1388 	struct pipe_inode_info *pipe = file->private_data;  in get_pipe_info()  local
1390 	if (file->f_op != &pipefifo_fops || !pipe)  in get_pipe_info()
1392 	if (for_splice && pipe_has_watch_queue(pipe))  in get_pipe_info()
1394 	return pipe;  in get_pipe_info()
1399 	struct pipe_inode_info *pipe;  in pipe_fcntl()  local
1402 	pipe = get_pipe_info(file, false);  in pipe_fcntl()
1403 	if (!pipe)  in pipe_fcntl()
1406 	mutex_lock(&pipe->mutex);  in pipe_fcntl()
1410 		ret = pipe_set_size(pipe, arg);  in pipe_fcntl()
1413 		ret = pipe->max_usage * PAGE_SIZE;  in pipe_fcntl()
1420 	mutex_unlock(&pipe->mutex);  in pipe_fcntl()
1433  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1482 		.procname	= "pipe-max-size",
1489 		.procname	= "pipe-user-pages-hard",
1496 		.procname	= "pipe-user-pages-soft",