xref: /netbsd-src/sys/kern/sys_pipe.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30)
1 /*	$NetBSD: sys_pipe.c,v 1.103 2008/09/17 14:00:41 pooka Exp $	*/
2 
3 /*-
4  * Copyright (c) 2003, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Paul Kranenburg, and by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1996 John S. Dyson
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice immediately at the beginning of the file, without modification,
41  *    this list of conditions, and the following disclaimer.
42  * 2. Redistributions in binary form must reproduce the above copyright
43  *    notice, this list of conditions and the following disclaimer in the
44  *    documentation and/or other materials provided with the distribution.
45  * 3. Absolutely no warranty of function or purpose is made by the author
46  *    John S. Dyson.
47  * 4. Modifications may be freely made to this file if the above conditions
48  *    are met.
49  *
50  * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.95 2002/03/09 22:06:31 alfred Exp $
51  */
52 
53 /*
54  * This file contains a high-performance replacement for the socket-based
55  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
56  * all features of sockets, but does do everything that pipes normally
57  * do.
58  *
59  * Adaption for NetBSD UVM, including uvm_loan() based direct write, was
60  * written by Jaromir Dolecek.
61  */
62 
63 /*
64  * This code has two modes of operation, a small write mode and a large
65  * write mode.  The small write mode acts like conventional pipes with
66  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
67  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
68  * and PIPE_SIZE in size it is mapped read-only into the kernel address space
69  * using the UVM page loan facility from where the receiving process can copy
70  * the data directly from the pages in the sending process.
71  *
72  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
73  * happen for small transfers so that the system will not spend all of
74  * its time context switching.  PIPE_SIZE is constrained by the
75  * amount of kernel virtual memory.
76  */
77 
78 #include <sys/cdefs.h>
79 __KERNEL_RCSID(0, "$NetBSD: sys_pipe.c,v 1.103 2008/09/17 14:00:41 pooka Exp $");
80 
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/proc.h>
84 #include <sys/fcntl.h>
85 #include <sys/file.h>
86 #include <sys/filedesc.h>
87 #include <sys/filio.h>
88 #include <sys/kernel.h>
89 #include <sys/ttycom.h>
90 #include <sys/stat.h>
91 #include <sys/poll.h>
92 #include <sys/signalvar.h>
93 #include <sys/vnode.h>
94 #include <sys/uio.h>
95 #include <sys/select.h>
96 #include <sys/mount.h>
97 #include <sys/syscallargs.h>
98 #include <sys/sysctl.h>
99 #include <sys/kauth.h>
100 #include <sys/atomic.h>
101 #include <sys/pipe.h>
102 
103 #include <uvm/uvm.h>
104 
105 /*
106  * Use this define if you want to disable *fancy* VM things.  Expect an
107  * approx 30% decrease in transfer rate.
108  */
109 /* #define PIPE_NODIRECT */
110 
111 /*
112  * interfaces to the outside world
113  */
114 static int pipe_read(struct file *fp, off_t *offset, struct uio *uio,
115 		kauth_cred_t cred, int flags);
116 static int pipe_write(struct file *fp, off_t *offset, struct uio *uio,
117 		kauth_cred_t cred, int flags);
118 static int pipe_close(struct file *fp);
119 static int pipe_poll(struct file *fp, int events);
120 static int pipe_kqfilter(struct file *fp, struct knote *kn);
121 static int pipe_stat(struct file *fp, struct stat *sb);
122 static int pipe_ioctl(struct file *fp, u_long cmd, void *data);
123 
124 static const struct fileops pipeops = {
125 	pipe_read, pipe_write, pipe_ioctl, fnullop_fcntl, pipe_poll,
126 	pipe_stat, pipe_close, pipe_kqfilter
127 };
128 
129 /*
130  * Single mutex shared between both ends of the pipe.
131  */
132 
133 struct pipe_mutex {
134 	kmutex_t	pm_mutex;
135 	u_int		pm_refcnt;
136 };
137 
138 /*
139  * Default pipe buffer size(s), this can be kind-of large now because pipe
140  * space is pageable.  The pipe code will try to maintain locality of
141  * reference for performance reasons, so small amounts of outstanding I/O
142  * will not wipe the cache.
143  */
144 #define MINPIPESIZE (PIPE_SIZE/3)
145 #define MAXPIPESIZE (2*PIPE_SIZE/3)
146 
147 /*
148  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
149  * is there so that on large systems, we don't exhaust it.
150  */
151 #define MAXPIPEKVA (8*1024*1024)
152 static u_int maxpipekva = MAXPIPEKVA;
153 
154 /*
155  * Limit for direct transfers, we cannot, of course limit
156  * the amount of kva for pipes in general though.
157  */
158 #define LIMITPIPEKVA (16*1024*1024)
159 static u_int limitpipekva = LIMITPIPEKVA;
160 
161 /*
162  * Limit the number of "big" pipes
163  */
164 #define LIMITBIGPIPES  32
165 static u_int maxbigpipes = LIMITBIGPIPES;
166 static u_int nbigpipe = 0;
167 
168 /*
169  * Amount of KVA consumed by pipe buffers.
170  */
171 static u_int amountpipekva = 0;
172 
173 static void pipeclose(struct file *fp, struct pipe *pipe);
174 static void pipe_free_kmem(struct pipe *pipe);
175 static int pipe_create(struct pipe **pipep, int allockva, struct pipe_mutex *);
176 static int pipelock(struct pipe *pipe, int catch);
177 static inline void pipeunlock(struct pipe *pipe);
178 static void pipeselwakeup(struct pipe *pipe, struct pipe *sigp, int code);
179 #ifndef PIPE_NODIRECT
180 static int pipe_direct_write(struct file *fp, struct pipe *wpipe,
181     struct uio *uio);
182 #endif
183 static int pipespace(struct pipe *pipe, int size);
184 
185 #ifndef PIPE_NODIRECT
186 static int pipe_loan_alloc(struct pipe *, int);
187 static void pipe_loan_free(struct pipe *);
188 #endif /* PIPE_NODIRECT */
189 
190 static int pipe_mutex_ctor(void *, void *, int);
191 static void pipe_mutex_dtor(void *, void *);
192 
193 static pool_cache_t pipe_cache;
194 static pool_cache_t pipe_mutex_cache;
195 
196 void
197 pipe_init(void)
198 {
199 
200 	pipe_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "pipepl",
201 	    NULL, IPL_NONE, NULL, NULL, NULL);
202 	KASSERT(pipe_cache != NULL);
203 
204 	pipe_mutex_cache = pool_cache_init(sizeof(struct pipe_mutex),
205 	    coherency_unit, 0, 0, "pipemtxpl", NULL, IPL_NONE, pipe_mutex_ctor,
206 	    pipe_mutex_dtor, NULL);
207 	KASSERT(pipe_cache != NULL);
208 }
209 
210 static int
211 pipe_mutex_ctor(void *arg, void *obj, int flag)
212 {
213 	struct pipe_mutex *pm = obj;
214 
215 	mutex_init(&pm->pm_mutex, MUTEX_DEFAULT, IPL_NONE);
216 	pm->pm_refcnt = 0;
217 
218 	return 0;
219 }
220 
221 static void
222 pipe_mutex_dtor(void *arg, void *obj)
223 {
224 	struct pipe_mutex *pm = obj;
225 
226 	KASSERT(pm->pm_refcnt == 0);
227 
228 	mutex_destroy(&pm->pm_mutex);
229 }
230 
231 /*
232  * The pipe system call for the DTYPE_PIPE type of pipes
233  */
234 
235 /* ARGSUSED */
236 int
237 sys_pipe(struct lwp *l, const void *v, register_t *retval)
238 {
239 	struct file *rf, *wf;
240 	struct pipe *rpipe, *wpipe;
241 	struct pipe_mutex *mutex;
242 	int fd, error;
243 	proc_t *p;
244 
245 	p = curproc;
246 	rpipe = wpipe = NULL;
247 	mutex = pool_cache_get(pipe_mutex_cache, PR_WAITOK);
248 	if (mutex == NULL)
249 		return (ENOMEM);
250 	if (pipe_create(&rpipe, 1, mutex) || pipe_create(&wpipe, 0, mutex)) {
251 		pipeclose(NULL, rpipe);
252 		pipeclose(NULL, wpipe);
253 		return (ENFILE);
254 	}
255 
256 	error = fd_allocfile(&rf, &fd);
257 	if (error)
258 		goto free2;
259 	retval[0] = fd;
260 	rf->f_flag = FREAD;
261 	rf->f_type = DTYPE_PIPE;
262 	rf->f_data = (void *)rpipe;
263 	rf->f_ops = &pipeops;
264 
265 	error = fd_allocfile(&wf, &fd);
266 	if (error)
267 		goto free3;
268 	retval[1] = fd;
269 	wf->f_flag = FWRITE;
270 	wf->f_type = DTYPE_PIPE;
271 	wf->f_data = (void *)wpipe;
272 	wf->f_ops = &pipeops;
273 
274 	rpipe->pipe_peer = wpipe;
275 	wpipe->pipe_peer = rpipe;
276 
277 	fd_affix(p, rf, (int)retval[0]);
278 	fd_affix(p, wf, (int)retval[1]);
279 	return (0);
280 free3:
281 	fd_abort(p, rf, (int)retval[0]);
282 free2:
283 	pipeclose(NULL, wpipe);
284 	pipeclose(NULL, rpipe);
285 
286 	return (error);
287 }
288 
289 /*
290  * Allocate kva for pipe circular buffer, the space is pageable
291  * This routine will 'realloc' the size of a pipe safely, if it fails
292  * it will retain the old buffer.
293  * If it fails it will return ENOMEM.
294  */
295 static int
296 pipespace(struct pipe *pipe, int size)
297 {
298 	void *buffer;
299 	/*
300 	 * Allocate pageable virtual address space. Physical memory is
301 	 * allocated on demand.
302 	 */
303 	buffer = (void *) uvm_km_alloc(kernel_map, round_page(size), 0,
304 	    UVM_KMF_PAGEABLE);
305 	if (buffer == NULL)
306 		return (ENOMEM);
307 
308 	/* free old resources if we're resizing */
309 	pipe_free_kmem(pipe);
310 	pipe->pipe_buffer.buffer = buffer;
311 	pipe->pipe_buffer.size = size;
312 	pipe->pipe_buffer.in = 0;
313 	pipe->pipe_buffer.out = 0;
314 	pipe->pipe_buffer.cnt = 0;
315 	atomic_add_int(&amountpipekva, pipe->pipe_buffer.size);
316 	return (0);
317 }
318 
319 /*
320  * Initialize and allocate VM and memory for pipe.
321  */
322 static int
323 pipe_create(struct pipe **pipep, int allockva, struct pipe_mutex *mutex)
324 {
325 	struct pipe *pipe;
326 	int error;
327 
328 	pipe = *pipep = pool_cache_get(pipe_cache, PR_WAITOK);
329 	mutex->pm_refcnt++;
330 
331 	/* Initialize */
332 	memset(pipe, 0, sizeof(struct pipe));
333 	pipe->pipe_state = PIPE_SIGNALR;
334 
335 	getmicrotime(&pipe->pipe_ctime);
336 	pipe->pipe_atime = pipe->pipe_ctime;
337 	pipe->pipe_mtime = pipe->pipe_ctime;
338 	pipe->pipe_lock = &mutex->pm_mutex;
339 	cv_init(&pipe->pipe_rcv, "piperd");
340 	cv_init(&pipe->pipe_wcv, "pipewr");
341 	cv_init(&pipe->pipe_draincv, "pipedrain");
342 	cv_init(&pipe->pipe_lkcv, "pipelk");
343 	selinit(&pipe->pipe_sel);
344 
345 	if (allockva && (error = pipespace(pipe, PIPE_SIZE)))
346 		return (error);
347 
348 	return (0);
349 }
350 
351 
352 /*
353  * Lock a pipe for I/O, blocking other access
354  * Called with pipe spin lock held.
355  * Return with pipe spin lock released on success.
356  */
357 static int
358 pipelock(struct pipe *pipe, int catch)
359 {
360 	int error;
361 
362 	KASSERT(mutex_owned(pipe->pipe_lock));
363 
364 	while (pipe->pipe_state & PIPE_LOCKFL) {
365 		pipe->pipe_state |= PIPE_LWANT;
366 		if (catch) {
367 			error = cv_wait_sig(&pipe->pipe_lkcv, pipe->pipe_lock);
368 			if (error != 0)
369 				return error;
370 		} else
371 			cv_wait(&pipe->pipe_lkcv, pipe->pipe_lock);
372 	}
373 
374 	pipe->pipe_state |= PIPE_LOCKFL;
375 
376 	return 0;
377 }
378 
379 /*
380  * unlock a pipe I/O lock
381  */
382 static inline void
383 pipeunlock(struct pipe *pipe)
384 {
385 
386 	KASSERT(pipe->pipe_state & PIPE_LOCKFL);
387 
388 	pipe->pipe_state &= ~PIPE_LOCKFL;
389 	if (pipe->pipe_state & PIPE_LWANT) {
390 		pipe->pipe_state &= ~PIPE_LWANT;
391 		cv_broadcast(&pipe->pipe_lkcv);
392 	}
393 }
394 
395 /*
396  * Select/poll wakup. This also sends SIGIO to peer connected to
397  * 'sigpipe' side of pipe.
398  */
399 static void
400 pipeselwakeup(struct pipe *selp, struct pipe *sigp, int code)
401 {
402 	int band;
403 
404 	switch (code) {
405 	case POLL_IN:
406 		band = POLLIN|POLLRDNORM;
407 		break;
408 	case POLL_OUT:
409 		band = POLLOUT|POLLWRNORM;
410 		break;
411 	case POLL_HUP:
412 		band = POLLHUP;
413 		break;
414 #if POLL_HUP != POLL_ERR
415 	case POLL_ERR:
416 		band = POLLERR;
417 		break;
418 #endif
419 	default:
420 		band = 0;
421 #ifdef DIAGNOSTIC
422 		printf("bad siginfo code %d in pipe notification.\n", code);
423 #endif
424 		break;
425 	}
426 
427 	selnotify(&selp->pipe_sel, band, NOTE_SUBMIT);
428 
429 	if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0)
430 		return;
431 
432 	fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp);
433 }
434 
435 /* ARGSUSED */
436 static int
437 pipe_read(struct file *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
438     int flags)
439 {
440 	struct pipe *rpipe = (struct pipe *) fp->f_data;
441 	struct pipebuf *bp = &rpipe->pipe_buffer;
442 	kmutex_t *lock = rpipe->pipe_lock;
443 	int error;
444 	size_t nread = 0;
445 	size_t size;
446 	size_t ocnt;
447 
448 	mutex_enter(lock);
449 	++rpipe->pipe_busy;
450 	ocnt = bp->cnt;
451 
452 again:
453 	error = pipelock(rpipe, 1);
454 	if (error)
455 		goto unlocked_error;
456 
457 	while (uio->uio_resid) {
458 		/*
459 		 * normal pipe buffer receive
460 		 */
461 		if (bp->cnt > 0) {
462 			size = bp->size - bp->out;
463 			if (size > bp->cnt)
464 				size = bp->cnt;
465 			if (size > uio->uio_resid)
466 				size = uio->uio_resid;
467 
468 			mutex_exit(lock);
469 			error = uiomove((char *)bp->buffer + bp->out, size, uio);
470 			mutex_enter(lock);
471 			if (error)
472 				break;
473 
474 			bp->out += size;
475 			if (bp->out >= bp->size)
476 				bp->out = 0;
477 
478 			bp->cnt -= size;
479 
480 			/*
481 			 * If there is no more to read in the pipe, reset
482 			 * its pointers to the beginning.  This improves
483 			 * cache hit stats.
484 			 */
485 			if (bp->cnt == 0) {
486 				bp->in = 0;
487 				bp->out = 0;
488 			}
489 			nread += size;
490 			continue;
491 		}
492 
493 #ifndef PIPE_NODIRECT
494 		if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) {
495 			/*
496 			 * Direct copy, bypassing a kernel buffer.
497 			 */
498 			void *	va;
499 
500 			KASSERT(rpipe->pipe_state & PIPE_DIRECTW);
501 
502 			size = rpipe->pipe_map.cnt;
503 			if (size > uio->uio_resid)
504 				size = uio->uio_resid;
505 
506 			va = (char *)rpipe->pipe_map.kva + rpipe->pipe_map.pos;
507 			mutex_exit(lock);
508 			error = uiomove(va, size, uio);
509 			mutex_enter(lock);
510 			if (error)
511 				break;
512 			nread += size;
513 			rpipe->pipe_map.pos += size;
514 			rpipe->pipe_map.cnt -= size;
515 			if (rpipe->pipe_map.cnt == 0) {
516 				rpipe->pipe_state &= ~PIPE_DIRECTR;
517 				cv_broadcast(&rpipe->pipe_wcv);
518 			}
519 			continue;
520 		}
521 #endif
522 		/*
523 		 * Break if some data was read.
524 		 */
525 		if (nread > 0)
526 			break;
527 
528 		/*
529 		 * detect EOF condition
530 		 * read returns 0 on EOF, no need to set error
531 		 */
532 		if (rpipe->pipe_state & PIPE_EOF)
533 			break;
534 
535 		/*
536 		 * don't block on non-blocking I/O
537 		 */
538 		if (fp->f_flag & FNONBLOCK) {
539 			error = EAGAIN;
540 			break;
541 		}
542 
543 		/*
544 		 * Unlock the pipe buffer for our remaining processing.
545 		 * We will either break out with an error or we will
546 		 * sleep and relock to loop.
547 		 */
548 		pipeunlock(rpipe);
549 
550 		/*
551 		 * Re-check to see if more direct writes are pending.
552 		 */
553 		if ((rpipe->pipe_state & PIPE_DIRECTR) != 0)
554 			goto again;
555 
556 		/*
557 		 * We want to read more, wake up select/poll.
558 		 */
559 		pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_IN);
560 
561 		/*
562 		 * If the "write-side" is blocked, wake it up now.
563 		 */
564 		cv_broadcast(&rpipe->pipe_wcv);
565 
566 		/* Now wait until the pipe is filled */
567 		error = cv_wait_sig(&rpipe->pipe_rcv, lock);
568 		if (error != 0)
569 			goto unlocked_error;
570 		goto again;
571 	}
572 
573 	if (error == 0)
574 		getmicrotime(&rpipe->pipe_atime);
575 	pipeunlock(rpipe);
576 
577 unlocked_error:
578 	--rpipe->pipe_busy;
579 	if (rpipe->pipe_busy == 0) {
580 		cv_broadcast(&rpipe->pipe_draincv);
581 	}
582 	if (bp->cnt < MINPIPESIZE) {
583 		cv_broadcast(&rpipe->pipe_wcv);
584 	}
585 
586 	/*
587 	 * If anything was read off the buffer, signal to the writer it's
588 	 * possible to write more data. Also send signal if we are here for the
589 	 * first time after last write.
590 	 */
591 	if ((bp->size - bp->cnt) >= PIPE_BUF
592 	    && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) {
593 		pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT);
594 		rpipe->pipe_state &= ~PIPE_SIGNALR;
595 	}
596 
597 	mutex_exit(lock);
598 	return (error);
599 }
600 
601 #ifndef PIPE_NODIRECT
602 /*
603  * Allocate structure for loan transfer.
604  */
605 static int
606 pipe_loan_alloc(struct pipe *wpipe, int npages)
607 {
608 	vsize_t len;
609 
610 	len = (vsize_t)npages << PAGE_SHIFT;
611 	atomic_add_int(&amountpipekva, len);
612 	wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0,
613 	    UVM_KMF_VAONLY | UVM_KMF_WAITVA);
614 	if (wpipe->pipe_map.kva == 0) {
615 		atomic_add_int(&amountpipekva, -len);
616 		return (ENOMEM);
617 	}
618 
619 	wpipe->pipe_map.npages = npages;
620 	wpipe->pipe_map.pgs = kmem_alloc(npages * sizeof(struct vm_page *),
621 	    KM_SLEEP);
622 	return (0);
623 }
624 
625 /*
626  * Free resources allocated for loan transfer.
627  */
628 static void
629 pipe_loan_free(struct pipe *wpipe)
630 {
631 	vsize_t len;
632 
633 	len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT;
634 	uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY);
635 	wpipe->pipe_map.kva = 0;
636 	atomic_add_int(&amountpipekva, -len);
637 	kmem_free(wpipe->pipe_map.pgs,
638 	    wpipe->pipe_map.npages * sizeof(struct vm_page *));
639 	wpipe->pipe_map.pgs = NULL;
640 }
641 
642 /*
643  * NetBSD direct write, using uvm_loan() mechanism.
644  * This implements the pipe buffer write mechanism.  Note that only
645  * a direct write OR a normal pipe write can be pending at any given time.
646  * If there are any characters in the pipe buffer, the direct write will
647  * be deferred until the receiving process grabs all of the bytes from
648  * the pipe buffer.  Then the direct mapping write is set-up.
649  *
650  * Called with the long-term pipe lock held.
651  */
652 static int
653 pipe_direct_write(struct file *fp, struct pipe *wpipe, struct uio *uio)
654 {
655 	int error, npages, j;
656 	struct vm_page **pgs;
657 	vaddr_t bbase, kva, base, bend;
658 	vsize_t blen, bcnt;
659 	voff_t bpos;
660 	kmutex_t *lock = wpipe->pipe_lock;
661 
662 	KASSERT(mutex_owned(wpipe->pipe_lock));
663 	KASSERT(wpipe->pipe_map.cnt == 0);
664 
665 	mutex_exit(lock);
666 
667 	/*
668 	 * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers
669 	 * not aligned to PAGE_SIZE.
670 	 */
671 	bbase = (vaddr_t)uio->uio_iov->iov_base;
672 	base = trunc_page(bbase);
673 	bend = round_page(bbase + uio->uio_iov->iov_len);
674 	blen = bend - base;
675 	bpos = bbase - base;
676 
677 	if (blen > PIPE_DIRECT_CHUNK) {
678 		blen = PIPE_DIRECT_CHUNK;
679 		bend = base + blen;
680 		bcnt = PIPE_DIRECT_CHUNK - bpos;
681 	} else {
682 		bcnt = uio->uio_iov->iov_len;
683 	}
684 	npages = blen >> PAGE_SHIFT;
685 
686 	/*
687 	 * Free the old kva if we need more pages than we have
688 	 * allocated.
689 	 */
690 	if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages)
691 		pipe_loan_free(wpipe);
692 
693 	/* Allocate new kva. */
694 	if (wpipe->pipe_map.kva == 0) {
695 		error = pipe_loan_alloc(wpipe, npages);
696 		if (error) {
697 			mutex_enter(lock);
698 			return (error);
699 		}
700 	}
701 
702 	/* Loan the write buffer memory from writer process */
703 	pgs = wpipe->pipe_map.pgs;
704 	error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen,
705 			 pgs, UVM_LOAN_TOPAGE);
706 	if (error) {
707 		pipe_loan_free(wpipe);
708 		mutex_enter(lock);
709 		return (ENOMEM); /* so that caller fallback to ordinary write */
710 	}
711 
712 	/* Enter the loaned pages to kva */
713 	kva = wpipe->pipe_map.kva;
714 	for (j = 0; j < npages; j++, kva += PAGE_SIZE) {
715 		pmap_kenter_pa(kva, VM_PAGE_TO_PHYS(pgs[j]), VM_PROT_READ);
716 	}
717 	pmap_update(pmap_kernel());
718 
719 	/* Now we can put the pipe in direct write mode */
720 	wpipe->pipe_map.pos = bpos;
721 	wpipe->pipe_map.cnt = bcnt;
722 
723 	/*
724 	 * But before we can let someone do a direct read, we
725 	 * have to wait until the pipe is drained.  Release the
726 	 * pipe lock while we wait.
727 	 */
728 	mutex_enter(lock);
729 	wpipe->pipe_state |= PIPE_DIRECTW;
730 	pipeunlock(wpipe);
731 
732 	while (error == 0 && wpipe->pipe_buffer.cnt > 0) {
733 		cv_broadcast(&wpipe->pipe_rcv);
734 		error = cv_wait_sig(&wpipe->pipe_wcv, lock);
735 		if (error == 0 && wpipe->pipe_state & PIPE_EOF)
736 			error = EPIPE;
737 	}
738 
739 	/* Pipe is drained; next read will off the direct buffer */
740 	wpipe->pipe_state |= PIPE_DIRECTR;
741 
742 	/* Wait until the reader is done */
743 	while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) {
744 		cv_broadcast(&wpipe->pipe_rcv);
745 		pipeselwakeup(wpipe, wpipe, POLL_IN);
746 		error = cv_wait_sig(&wpipe->pipe_wcv, lock);
747 		if (error == 0 && wpipe->pipe_state & PIPE_EOF)
748 			error = EPIPE;
749 	}
750 
751 	/* Take pipe out of direct write mode */
752 	wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR);
753 
754 	/* Acquire the pipe lock and cleanup */
755 	(void)pipelock(wpipe, 0);
756 	mutex_exit(lock);
757 
758 	if (pgs != NULL) {
759 		pmap_kremove(wpipe->pipe_map.kva, blen);
760 		pmap_update(pmap_kernel());
761 		uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE);
762 	}
763 	if (error || amountpipekva > maxpipekva)
764 		pipe_loan_free(wpipe);
765 
766 	mutex_enter(lock);
767 	if (error) {
768 		pipeselwakeup(wpipe, wpipe, POLL_ERR);
769 
770 		/*
771 		 * If nothing was read from what we offered, return error
772 		 * straight on. Otherwise update uio resid first. Caller
773 		 * will deal with the error condition, returning short
774 		 * write, error, or restarting the write(2) as appropriate.
775 		 */
776 		if (wpipe->pipe_map.cnt == bcnt) {
777 			wpipe->pipe_map.cnt = 0;
778 			cv_broadcast(&wpipe->pipe_wcv);
779 			return (error);
780 		}
781 
782 		bcnt -= wpipe->pipe_map.cnt;
783 	}
784 
785 	uio->uio_resid -= bcnt;
786 	/* uio_offset not updated, not set/used for write(2) */
787 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt;
788 	uio->uio_iov->iov_len -= bcnt;
789 	if (uio->uio_iov->iov_len == 0) {
790 		uio->uio_iov++;
791 		uio->uio_iovcnt--;
792 	}
793 
794 	wpipe->pipe_map.cnt = 0;
795 	return (error);
796 }
797 #endif /* !PIPE_NODIRECT */
798 
799 static int
800 pipe_write(struct file *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
801     int flags)
802 {
803 	struct pipe *wpipe, *rpipe;
804 	struct pipebuf *bp;
805 	kmutex_t *lock;
806 	int error;
807 
808 	/* We want to write to our peer */
809 	rpipe = (struct pipe *) fp->f_data;
810 	lock = rpipe->pipe_lock;
811 	error = 0;
812 
813 	mutex_enter(lock);
814 	wpipe = rpipe->pipe_peer;
815 
816 	/*
817 	 * Detect loss of pipe read side, issue SIGPIPE if lost.
818 	 */
819 	if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) {
820 		mutex_exit(lock);
821 		return EPIPE;
822 	}
823 	++wpipe->pipe_busy;
824 
825 	/* Aquire the long-term pipe lock */
826 	if ((error = pipelock(wpipe, 1)) != 0) {
827 		--wpipe->pipe_busy;
828 		if (wpipe->pipe_busy == 0) {
829 			cv_broadcast(&wpipe->pipe_draincv);
830 		}
831 		mutex_exit(lock);
832 		return (error);
833 	}
834 
835 	bp = &wpipe->pipe_buffer;
836 
837 	/*
838 	 * If it is advantageous to resize the pipe buffer, do so.
839 	 */
840 	if ((uio->uio_resid > PIPE_SIZE) &&
841 	    (nbigpipe < maxbigpipes) &&
842 #ifndef PIPE_NODIRECT
843 	    (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
844 #endif
845 	    (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) {
846 
847 		if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
848 			atomic_inc_uint(&nbigpipe);
849 	}
850 
851 	while (uio->uio_resid) {
852 		size_t space;
853 
854 #ifndef PIPE_NODIRECT
855 		/*
856 		 * Pipe buffered writes cannot be coincidental with
857 		 * direct writes.  Also, only one direct write can be
858 		 * in progress at any one time.  We wait until the currently
859 		 * executing direct write is completed before continuing.
860 		 *
861 		 * We break out if a signal occurs or the reader goes away.
862 		 */
863 		while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) {
864 			cv_broadcast(&wpipe->pipe_rcv);
865 			pipeunlock(wpipe);
866 			error = cv_wait_sig(&wpipe->pipe_wcv, lock);
867 			(void)pipelock(wpipe, 0);
868 			if (wpipe->pipe_state & PIPE_EOF)
869 				error = EPIPE;
870 		}
871 		if (error)
872 			break;
873 
874 		/*
875 		 * If the transfer is large, we can gain performance if
876 		 * we do process-to-process copies directly.
877 		 * If the write is non-blocking, we don't use the
878 		 * direct write mechanism.
879 		 *
880 		 * The direct write mechanism will detect the reader going
881 		 * away on us.
882 		 */
883 		if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
884 		    (fp->f_flag & FNONBLOCK) == 0 &&
885 		    (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) {
886 			error = pipe_direct_write(fp, wpipe, uio);
887 
888 			/*
889 			 * Break out if error occurred, unless it's ENOMEM.
890 			 * ENOMEM means we failed to allocate some resources
891 			 * for direct write, so we just fallback to ordinary
892 			 * write. If the direct write was successful,
893 			 * process rest of data via ordinary write.
894 			 */
895 			if (error == 0)
896 				continue;
897 
898 			if (error != ENOMEM)
899 				break;
900 		}
901 #endif /* PIPE_NODIRECT */
902 
903 		space = bp->size - bp->cnt;
904 
905 		/* Writes of size <= PIPE_BUF must be atomic. */
906 		if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF))
907 			space = 0;
908 
909 		if (space > 0) {
910 			int size;	/* Transfer size */
911 			int segsize;	/* first segment to transfer */
912 
913 			/*
914 			 * Transfer size is minimum of uio transfer
915 			 * and free space in pipe buffer.
916 			 */
917 			if (space > uio->uio_resid)
918 				size = uio->uio_resid;
919 			else
920 				size = space;
921 			/*
922 			 * First segment to transfer is minimum of
923 			 * transfer size and contiguous space in
924 			 * pipe buffer.  If first segment to transfer
925 			 * is less than the transfer size, we've got
926 			 * a wraparound in the buffer.
927 			 */
928 			segsize = bp->size - bp->in;
929 			if (segsize > size)
930 				segsize = size;
931 
932 			/* Transfer first segment */
933 			mutex_exit(lock);
934 			error = uiomove((char *)bp->buffer + bp->in, segsize,
935 			    uio);
936 
937 			if (error == 0 && segsize < size) {
938 				/*
939 				 * Transfer remaining part now, to
940 				 * support atomic writes.  Wraparound
941 				 * happened.
942 				 */
943 #ifdef DEBUG
944 				if (bp->in + segsize != bp->size)
945 					panic("Expected pipe buffer wraparound disappeared");
946 #endif
947 
948 				error = uiomove(bp->buffer,
949 				    size - segsize, uio);
950 			}
951 			mutex_enter(lock);
952 			if (error)
953 				break;
954 
955 			bp->in += size;
956 			if (bp->in >= bp->size) {
957 #ifdef DEBUG
958 				if (bp->in != size - segsize + bp->size)
959 					panic("Expected wraparound bad");
960 #endif
961 				bp->in = size - segsize;
962 			}
963 
964 			bp->cnt += size;
965 #ifdef DEBUG
966 			if (bp->cnt > bp->size)
967 				panic("Pipe buffer overflow");
968 #endif
969 		} else {
970 			/*
971 			 * If the "read-side" has been blocked, wake it up now.
972 			 */
973 			cv_broadcast(&wpipe->pipe_rcv);
974 
975 			/*
976 			 * don't block on non-blocking I/O
977 			 */
978 			if (fp->f_flag & FNONBLOCK) {
979 				error = EAGAIN;
980 				break;
981 			}
982 
983 			/*
984 			 * We have no more space and have something to offer,
985 			 * wake up select/poll.
986 			 */
987 			if (bp->cnt)
988 				pipeselwakeup(wpipe, wpipe, POLL_OUT);
989 
990 			pipeunlock(wpipe);
991 			error = cv_wait_sig(&wpipe->pipe_wcv, lock);
992 			(void)pipelock(wpipe, 0);
993 			if (error != 0)
994 				break;
995 			/*
996 			 * If read side wants to go away, we just issue a signal
997 			 * to ourselves.
998 			 */
999 			if (wpipe->pipe_state & PIPE_EOF) {
1000 				error = EPIPE;
1001 				break;
1002 			}
1003 		}
1004 	}
1005 
1006 	--wpipe->pipe_busy;
1007 	if (wpipe->pipe_busy == 0) {
1008 		cv_broadcast(&wpipe->pipe_draincv);
1009 	}
1010 	if (bp->cnt > 0) {
1011 		cv_broadcast(&wpipe->pipe_rcv);
1012 	}
1013 
1014 	/*
1015 	 * Don't return EPIPE if I/O was successful
1016 	 */
1017 	if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0)
1018 		error = 0;
1019 
1020 	if (error == 0)
1021 		getmicrotime(&wpipe->pipe_mtime);
1022 
1023 	/*
1024 	 * We have something to offer, wake up select/poll.
1025 	 * wpipe->pipe_map.cnt is always 0 in this point (direct write
1026 	 * is only done synchronously), so check only wpipe->pipe_buffer.cnt
1027 	 */
1028 	if (bp->cnt)
1029 		pipeselwakeup(wpipe, wpipe, POLL_OUT);
1030 
1031 	/*
1032 	 * Arrange for next read(2) to do a signal.
1033 	 */
1034 	wpipe->pipe_state |= PIPE_SIGNALR;
1035 
1036 	pipeunlock(wpipe);
1037 	mutex_exit(lock);
1038 	return (error);
1039 }
1040 
1041 /*
1042  * we implement a very minimal set of ioctls for compatibility with sockets.
1043  */
1044 int
1045 pipe_ioctl(struct file *fp, u_long cmd, void *data)
1046 {
1047 	struct pipe *pipe = fp->f_data;
1048 	kmutex_t *lock = pipe->pipe_lock;
1049 
1050 	switch (cmd) {
1051 
1052 	case FIONBIO:
1053 		return (0);
1054 
1055 	case FIOASYNC:
1056 		mutex_enter(lock);
1057 		if (*(int *)data) {
1058 			pipe->pipe_state |= PIPE_ASYNC;
1059 		} else {
1060 			pipe->pipe_state &= ~PIPE_ASYNC;
1061 		}
1062 		mutex_exit(lock);
1063 		return (0);
1064 
1065 	case FIONREAD:
1066 		mutex_enter(lock);
1067 #ifndef PIPE_NODIRECT
1068 		if (pipe->pipe_state & PIPE_DIRECTW)
1069 			*(int *)data = pipe->pipe_map.cnt;
1070 		else
1071 #endif
1072 			*(int *)data = pipe->pipe_buffer.cnt;
1073 		mutex_exit(lock);
1074 		return (0);
1075 
1076 	case FIONWRITE:
1077 		/* Look at other side */
1078 		pipe = pipe->pipe_peer;
1079 		mutex_enter(lock);
1080 #ifndef PIPE_NODIRECT
1081 		if (pipe->pipe_state & PIPE_DIRECTW)
1082 			*(int *)data = pipe->pipe_map.cnt;
1083 		else
1084 #endif
1085 			*(int *)data = pipe->pipe_buffer.cnt;
1086 		mutex_exit(lock);
1087 		return (0);
1088 
1089 	case FIONSPACE:
1090 		/* Look at other side */
1091 		pipe = pipe->pipe_peer;
1092 		mutex_enter(lock);
1093 #ifndef PIPE_NODIRECT
1094 		/*
1095 		 * If we're in direct-mode, we don't really have a
1096 		 * send queue, and any other write will block. Thus
1097 		 * zero seems like the best answer.
1098 		 */
1099 		if (pipe->pipe_state & PIPE_DIRECTW)
1100 			*(int *)data = 0;
1101 		else
1102 #endif
1103 			*(int *)data = pipe->pipe_buffer.size -
1104 			    pipe->pipe_buffer.cnt;
1105 		mutex_exit(lock);
1106 		return (0);
1107 
1108 	case TIOCSPGRP:
1109 	case FIOSETOWN:
1110 		return fsetown(&pipe->pipe_pgid, cmd, data);
1111 
1112 	case TIOCGPGRP:
1113 	case FIOGETOWN:
1114 		return fgetown(pipe->pipe_pgid, cmd, data);
1115 
1116 	}
1117 	return (EPASSTHROUGH);
1118 }
1119 
1120 int
1121 pipe_poll(struct file *fp, int events)
1122 {
1123 	struct pipe *rpipe = fp->f_data;
1124 	struct pipe *wpipe;
1125 	int eof = 0;
1126 	int revents = 0;
1127 
1128 	mutex_enter(rpipe->pipe_lock);
1129 	wpipe = rpipe->pipe_peer;
1130 
1131 	if (events & (POLLIN | POLLRDNORM))
1132 		if ((rpipe->pipe_buffer.cnt > 0) ||
1133 #ifndef PIPE_NODIRECT
1134 		    (rpipe->pipe_state & PIPE_DIRECTR) ||
1135 #endif
1136 		    (rpipe->pipe_state & PIPE_EOF))
1137 			revents |= events & (POLLIN | POLLRDNORM);
1138 
1139 	eof |= (rpipe->pipe_state & PIPE_EOF);
1140 
1141 	if (wpipe == NULL)
1142 		revents |= events & (POLLOUT | POLLWRNORM);
1143 	else {
1144 		if (events & (POLLOUT | POLLWRNORM))
1145 			if ((wpipe->pipe_state & PIPE_EOF) || (
1146 #ifndef PIPE_NODIRECT
1147 			     (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1148 #endif
1149 			     (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1150 				revents |= events & (POLLOUT | POLLWRNORM);
1151 
1152 		eof |= (wpipe->pipe_state & PIPE_EOF);
1153 	}
1154 
1155 	if (wpipe == NULL || eof)
1156 		revents |= POLLHUP;
1157 
1158 	if (revents == 0) {
1159 		if (events & (POLLIN | POLLRDNORM))
1160 			selrecord(curlwp, &rpipe->pipe_sel);
1161 
1162 		if (events & (POLLOUT | POLLWRNORM))
1163 			selrecord(curlwp, &wpipe->pipe_sel);
1164 	}
1165 	mutex_exit(rpipe->pipe_lock);
1166 
1167 	return (revents);
1168 }
1169 
1170 static int
1171 pipe_stat(struct file *fp, struct stat *ub)
1172 {
1173 	struct pipe *pipe = fp->f_data;
1174 
1175 	memset((void *)ub, 0, sizeof(*ub));
1176 	ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1177 	ub->st_blksize = pipe->pipe_buffer.size;
1178 	if (ub->st_blksize == 0 && pipe->pipe_peer)
1179 		ub->st_blksize = pipe->pipe_peer->pipe_buffer.size;
1180 	ub->st_size = pipe->pipe_buffer.cnt;
1181 	ub->st_blocks = (ub->st_size) ? 1 : 0;
1182 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_atime, &ub->st_atimespec);
1183 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_mtime, &ub->st_mtimespec);
1184 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_ctime, &ub->st_ctimespec);
1185 	ub->st_uid = kauth_cred_geteuid(fp->f_cred);
1186 	ub->st_gid = kauth_cred_getegid(fp->f_cred);
1187 
1188 	/*
1189 	 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1190 	 * XXX (st_dev, st_ino) should be unique.
1191 	 */
1192 	return (0);
1193 }
1194 
1195 /* ARGSUSED */
1196 static int
1197 pipe_close(struct file *fp)
1198 {
1199 	struct pipe *pipe = fp->f_data;
1200 
1201 	fp->f_data = NULL;
1202 	pipeclose(fp, pipe);
1203 	return (0);
1204 }
1205 
1206 static void
1207 pipe_free_kmem(struct pipe *pipe)
1208 {
1209 
1210 	if (pipe->pipe_buffer.buffer != NULL) {
1211 		if (pipe->pipe_buffer.size > PIPE_SIZE)
1212 			atomic_dec_uint(&nbigpipe);
1213 		uvm_km_free(kernel_map,
1214 			(vaddr_t)pipe->pipe_buffer.buffer,
1215 			pipe->pipe_buffer.size, UVM_KMF_PAGEABLE);
1216 		atomic_add_int(&amountpipekva, -pipe->pipe_buffer.size);
1217 		pipe->pipe_buffer.buffer = NULL;
1218 	}
1219 #ifndef PIPE_NODIRECT
1220 	if (pipe->pipe_map.kva != 0) {
1221 		pipe_loan_free(pipe);
1222 		pipe->pipe_map.cnt = 0;
1223 		pipe->pipe_map.kva = 0;
1224 		pipe->pipe_map.pos = 0;
1225 		pipe->pipe_map.npages = 0;
1226 	}
1227 #endif /* !PIPE_NODIRECT */
1228 }
1229 
1230 /*
1231  * shutdown the pipe
1232  */
1233 static void
1234 pipeclose(struct file *fp, struct pipe *pipe)
1235 {
1236 	struct pipe_mutex *mutex;
1237 	kmutex_t *lock;
1238 	struct pipe *ppipe;
1239 	u_int refcnt;
1240 
1241 	if (pipe == NULL)
1242 		return;
1243 
1244 	KASSERT(cv_is_valid(&pipe->pipe_rcv));
1245 	KASSERT(cv_is_valid(&pipe->pipe_wcv));
1246 	KASSERT(cv_is_valid(&pipe->pipe_draincv));
1247 	KASSERT(cv_is_valid(&pipe->pipe_lkcv));
1248 
1249 	lock = pipe->pipe_lock;
1250 	mutex_enter(lock);
1251 	pipeselwakeup(pipe, pipe, POLL_HUP);
1252 
1253 	/*
1254 	 * If the other side is blocked, wake it up saying that
1255 	 * we want to close it down.
1256 	 */
1257 	pipe->pipe_state |= PIPE_EOF;
1258 	if (pipe->pipe_busy) {
1259 		while (pipe->pipe_busy) {
1260 			cv_broadcast(&pipe->pipe_wcv);
1261 			cv_wait_sig(&pipe->pipe_draincv, lock);
1262 		}
1263 	}
1264 
1265 	/*
1266 	 * Disconnect from peer
1267 	 */
1268 	if ((ppipe = pipe->pipe_peer) != NULL) {
1269 		pipeselwakeup(ppipe, ppipe, POLL_HUP);
1270 		ppipe->pipe_state |= PIPE_EOF;
1271 		cv_broadcast(&ppipe->pipe_rcv);
1272 		ppipe->pipe_peer = NULL;
1273 	}
1274 
1275 	KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0);
1276 
1277 	mutex = (struct pipe_mutex *)lock;
1278 	refcnt = --(mutex->pm_refcnt);
1279 	KASSERT(refcnt == 0 || refcnt == 1);
1280 	mutex_exit(lock);
1281 
1282 	/*
1283 	 * free resources
1284 	 */
1285 	pipe_free_kmem(pipe);
1286 	cv_destroy(&pipe->pipe_rcv);
1287 	cv_destroy(&pipe->pipe_wcv);
1288 	cv_destroy(&pipe->pipe_draincv);
1289 	cv_destroy(&pipe->pipe_lkcv);
1290 	seldestroy(&pipe->pipe_sel);
1291 	pool_cache_put(pipe_cache, pipe);
1292 	if (refcnt == 0)
1293 		pool_cache_put(pipe_mutex_cache, mutex);
1294 }
1295 
1296 static void
1297 filt_pipedetach(struct knote *kn)
1298 {
1299 	struct pipe *pipe;
1300 	kmutex_t *lock;
1301 
1302 	pipe = ((file_t *)kn->kn_obj)->f_data;
1303 	lock = pipe->pipe_lock;
1304 
1305 	mutex_enter(lock);
1306 
1307 	switch(kn->kn_filter) {
1308 	case EVFILT_WRITE:
1309 		/* need the peer structure, not our own */
1310 		pipe = pipe->pipe_peer;
1311 
1312 		/* if reader end already closed, just return */
1313 		if (pipe == NULL) {
1314 			mutex_exit(lock);
1315 			return;
1316 		}
1317 
1318 		break;
1319 	default:
1320 		/* nothing to do */
1321 		break;
1322 	}
1323 
1324 #ifdef DIAGNOSTIC
1325 	if (kn->kn_hook != pipe)
1326 		panic("filt_pipedetach: inconsistent knote");
1327 #endif
1328 
1329 	SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext);
1330 	mutex_exit(lock);
1331 }
1332 
1333 /*ARGSUSED*/
1334 static int
1335 filt_piperead(struct knote *kn, long hint)
1336 {
1337 	struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
1338 	struct pipe *wpipe;
1339 
1340 	if ((hint & NOTE_SUBMIT) == 0) {
1341 		mutex_enter(rpipe->pipe_lock);
1342 	}
1343 	wpipe = rpipe->pipe_peer;
1344 	kn->kn_data = rpipe->pipe_buffer.cnt;
1345 
1346 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1347 		kn->kn_data = rpipe->pipe_map.cnt;
1348 
1349 	if ((rpipe->pipe_state & PIPE_EOF) ||
1350 	    (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1351 		kn->kn_flags |= EV_EOF;
1352 		if ((hint & NOTE_SUBMIT) == 0) {
1353 			mutex_exit(rpipe->pipe_lock);
1354 		}
1355 		return (1);
1356 	}
1357 
1358 	if ((hint & NOTE_SUBMIT) == 0) {
1359 		mutex_exit(rpipe->pipe_lock);
1360 	}
1361 	return (kn->kn_data > 0);
1362 }
1363 
1364 /*ARGSUSED*/
1365 static int
1366 filt_pipewrite(struct knote *kn, long hint)
1367 {
1368 	struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
1369 	struct pipe *wpipe;
1370 
1371 	if ((hint & NOTE_SUBMIT) == 0) {
1372 		mutex_enter(rpipe->pipe_lock);
1373 	}
1374 	wpipe = rpipe->pipe_peer;
1375 
1376 	if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1377 		kn->kn_data = 0;
1378 		kn->kn_flags |= EV_EOF;
1379 		if ((hint & NOTE_SUBMIT) == 0) {
1380 			mutex_exit(rpipe->pipe_lock);
1381 		}
1382 		return (1);
1383 	}
1384 	kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1385 	if (wpipe->pipe_state & PIPE_DIRECTW)
1386 		kn->kn_data = 0;
1387 
1388 	if ((hint & NOTE_SUBMIT) == 0) {
1389 		mutex_exit(rpipe->pipe_lock);
1390 	}
1391 	return (kn->kn_data >= PIPE_BUF);
1392 }
1393 
1394 static const struct filterops pipe_rfiltops =
1395 	{ 1, NULL, filt_pipedetach, filt_piperead };
1396 static const struct filterops pipe_wfiltops =
1397 	{ 1, NULL, filt_pipedetach, filt_pipewrite };
1398 
1399 /*ARGSUSED*/
1400 static int
1401 pipe_kqfilter(struct file *fp, struct knote *kn)
1402 {
1403 	struct pipe *pipe;
1404 	kmutex_t *lock;
1405 
1406 	pipe = ((file_t *)kn->kn_obj)->f_data;
1407 	lock = pipe->pipe_lock;
1408 
1409 	mutex_enter(lock);
1410 
1411 	switch (kn->kn_filter) {
1412 	case EVFILT_READ:
1413 		kn->kn_fop = &pipe_rfiltops;
1414 		break;
1415 	case EVFILT_WRITE:
1416 		kn->kn_fop = &pipe_wfiltops;
1417 		pipe = pipe->pipe_peer;
1418 		if (pipe == NULL) {
1419 			/* other end of pipe has been closed */
1420 			mutex_exit(lock);
1421 			return (EBADF);
1422 		}
1423 		break;
1424 	default:
1425 		mutex_exit(lock);
1426 		return (EINVAL);
1427 	}
1428 
1429 	kn->kn_hook = pipe;
1430 	SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext);
1431 	mutex_exit(lock);
1432 
1433 	return (0);
1434 }
1435 
1436 /*
1437  * Handle pipe sysctls.
1438  */
1439 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup")
1440 {
1441 
1442 	sysctl_createv(clog, 0, NULL, NULL,
1443 		       CTLFLAG_PERMANENT,
1444 		       CTLTYPE_NODE, "kern", NULL,
1445 		       NULL, 0, NULL, 0,
1446 		       CTL_KERN, CTL_EOL);
1447 	sysctl_createv(clog, 0, NULL, NULL,
1448 		       CTLFLAG_PERMANENT,
1449 		       CTLTYPE_NODE, "pipe",
1450 		       SYSCTL_DESCR("Pipe settings"),
1451 		       NULL, 0, NULL, 0,
1452 		       CTL_KERN, KERN_PIPE, CTL_EOL);
1453 
1454 	sysctl_createv(clog, 0, NULL, NULL,
1455 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1456 		       CTLTYPE_INT, "maxkvasz",
1457 		       SYSCTL_DESCR("Maximum amount of kernel memory to be "
1458 				    "used for pipes"),
1459 		       NULL, 0, &maxpipekva, 0,
1460 		       CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL);
1461 	sysctl_createv(clog, 0, NULL, NULL,
1462 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1463 		       CTLTYPE_INT, "maxloankvasz",
1464 		       SYSCTL_DESCR("Limit for direct transfers via page loan"),
1465 		       NULL, 0, &limitpipekva, 0,
1466 		       CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL);
1467 	sysctl_createv(clog, 0, NULL, NULL,
1468 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1469 		       CTLTYPE_INT, "maxbigpipes",
1470 		       SYSCTL_DESCR("Maximum number of \"big\" pipes"),
1471 		       NULL, 0, &maxbigpipes, 0,
1472 		       CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL);
1473 	sysctl_createv(clog, 0, NULL, NULL,
1474 		       CTLFLAG_PERMANENT,
1475 		       CTLTYPE_INT, "nbigpipes",
1476 		       SYSCTL_DESCR("Number of \"big\" pipes"),
1477 		       NULL, 0, &nbigpipe, 0,
1478 		       CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL);
1479 	sysctl_createv(clog, 0, NULL, NULL,
1480 		       CTLFLAG_PERMANENT,
1481 		       CTLTYPE_INT, "kvasize",
1482 		       SYSCTL_DESCR("Amount of kernel memory consumed by pipe "
1483 				    "buffers"),
1484 		       NULL, 0, &amountpipekva, 0,
1485 		       CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL);
1486 }
1487