xref: /netbsd-src/sys/kern/sys_pipe.c (revision cac8e449158efc7261bebc8657cbb0125a2cfdde)
1 /*	$NetBSD: sys_pipe.c,v 1.101 2008/04/28 20:24:04 martin 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 1.101 2008/04/28 20:24:04 martin 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/malloc.h>
92 #include <sys/poll.h>
93 #include <sys/signalvar.h>
94 #include <sys/vnode.h>
95 #include <sys/uio.h>
96 #include <sys/select.h>
97 #include <sys/mount.h>
98 #include <sys/syscallargs.h>
99 #include <sys/sysctl.h>
100 #include <sys/kauth.h>
101 #include <sys/atomic.h>
102 #include <sys/pipe.h>
103 
104 #include <uvm/uvm.h>
105 
106 /*
107  * Use this define if you want to disable *fancy* VM things.  Expect an
108  * approx 30% decrease in transfer rate.
109  */
110 /* #define PIPE_NODIRECT */
111 
112 /*
113  * interfaces to the outside world
114  */
115 static int pipe_read(struct file *fp, off_t *offset, struct uio *uio,
116 		kauth_cred_t cred, int flags);
117 static int pipe_write(struct file *fp, off_t *offset, struct uio *uio,
118 		kauth_cred_t cred, int flags);
119 static int pipe_close(struct file *fp);
120 static int pipe_poll(struct file *fp, int events);
121 static int pipe_kqfilter(struct file *fp, struct knote *kn);
122 static int pipe_stat(struct file *fp, struct stat *sb);
123 static int pipe_ioctl(struct file *fp, u_long cmd, void *data);
124 
125 static const struct fileops pipeops = {
126 	pipe_read, pipe_write, pipe_ioctl, fnullop_fcntl, pipe_poll,
127 	pipe_stat, pipe_close, pipe_kqfilter
128 };
129 
130 /*
131  * Single mutex shared between both ends of the pipe.
132  */
133 
134 struct pipe_mutex {
135 	kmutex_t	pm_mutex;
136 	u_int		pm_refcnt;
137 };
138 
139 /*
140  * Default pipe buffer size(s), this can be kind-of large now because pipe
141  * space is pageable.  The pipe code will try to maintain locality of
142  * reference for performance reasons, so small amounts of outstanding I/O
143  * will not wipe the cache.
144  */
145 #define MINPIPESIZE (PIPE_SIZE/3)
146 #define MAXPIPESIZE (2*PIPE_SIZE/3)
147 
148 /*
149  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
150  * is there so that on large systems, we don't exhaust it.
151  */
152 #define MAXPIPEKVA (8*1024*1024)
153 static u_int maxpipekva = MAXPIPEKVA;
154 
155 /*
156  * Limit for direct transfers, we cannot, of course limit
157  * the amount of kva for pipes in general though.
158  */
159 #define LIMITPIPEKVA (16*1024*1024)
160 static u_int limitpipekva = LIMITPIPEKVA;
161 
162 /*
163  * Limit the number of "big" pipes
164  */
165 #define LIMITBIGPIPES  32
166 static u_int maxbigpipes = LIMITBIGPIPES;
167 static u_int nbigpipe = 0;
168 
169 /*
170  * Amount of KVA consumed by pipe buffers.
171  */
172 static u_int amountpipekva = 0;
173 
174 MALLOC_DEFINE(M_PIPE, "pipe", "Pipe structures");
175 
176 static void pipeclose(struct file *fp, struct pipe *pipe);
177 static void pipe_free_kmem(struct pipe *pipe);
178 static int pipe_create(struct pipe **pipep, int allockva, struct pipe_mutex *);
179 static int pipelock(struct pipe *pipe, int catch);
180 static inline void pipeunlock(struct pipe *pipe);
181 static void pipeselwakeup(struct pipe *pipe, struct pipe *sigp, int code);
182 #ifndef PIPE_NODIRECT
183 static int pipe_direct_write(struct file *fp, struct pipe *wpipe,
184     struct uio *uio);
185 #endif
186 static int pipespace(struct pipe *pipe, int size);
187 
188 #ifndef PIPE_NODIRECT
189 static int pipe_loan_alloc(struct pipe *, int);
190 static void pipe_loan_free(struct pipe *);
191 #endif /* PIPE_NODIRECT */
192 
193 static int pipe_mutex_ctor(void *, void *, int);
194 static void pipe_mutex_dtor(void *, void *);
195 
196 static pool_cache_t pipe_cache;
197 static pool_cache_t pipe_mutex_cache;
198 
199 void
200 pipe_init(void)
201 {
202 
203 	pipe_cache = pool_cache_init(sizeof(struct pipe), 0, 0, 0, "pipepl",
204 	    NULL, IPL_NONE, NULL, NULL, NULL);
205 	KASSERT(pipe_cache != NULL);
206 
207 	pipe_mutex_cache = pool_cache_init(sizeof(struct pipe_mutex),
208 	    coherency_unit, 0, 0, "pipemtxpl", NULL, IPL_NONE, pipe_mutex_ctor,
209 	    pipe_mutex_dtor, NULL);
210 	KASSERT(pipe_cache != NULL);
211 }
212 
213 static int
214 pipe_mutex_ctor(void *arg, void *obj, int flag)
215 {
216 	struct pipe_mutex *pm = obj;
217 
218 	mutex_init(&pm->pm_mutex, MUTEX_DEFAULT, IPL_NONE);
219 	pm->pm_refcnt = 0;
220 
221 	return 0;
222 }
223 
224 static void
225 pipe_mutex_dtor(void *arg, void *obj)
226 {
227 	struct pipe_mutex *pm = obj;
228 
229 	KASSERT(pm->pm_refcnt == 0);
230 
231 	mutex_destroy(&pm->pm_mutex);
232 }
233 
234 /*
235  * The pipe system call for the DTYPE_PIPE type of pipes
236  */
237 
238 /* ARGSUSED */
239 int
240 sys_pipe(struct lwp *l, const void *v, register_t *retval)
241 {
242 	struct file *rf, *wf;
243 	struct pipe *rpipe, *wpipe;
244 	struct pipe_mutex *mutex;
245 	int fd, error;
246 	proc_t *p;
247 
248 	p = curproc;
249 	rpipe = wpipe = NULL;
250 	mutex = pool_cache_get(pipe_mutex_cache, PR_WAITOK);
251 	if (mutex == NULL)
252 		return (ENOMEM);
253 	if (pipe_create(&rpipe, 1, mutex) || pipe_create(&wpipe, 0, mutex)) {
254 		pipeclose(NULL, rpipe);
255 		pipeclose(NULL, wpipe);
256 		return (ENFILE);
257 	}
258 
259 	error = fd_allocfile(&rf, &fd);
260 	if (error)
261 		goto free2;
262 	retval[0] = fd;
263 	rf->f_flag = FREAD;
264 	rf->f_type = DTYPE_PIPE;
265 	rf->f_data = (void *)rpipe;
266 	rf->f_ops = &pipeops;
267 
268 	error = fd_allocfile(&wf, &fd);
269 	if (error)
270 		goto free3;
271 	retval[1] = fd;
272 	wf->f_flag = FWRITE;
273 	wf->f_type = DTYPE_PIPE;
274 	wf->f_data = (void *)wpipe;
275 	wf->f_ops = &pipeops;
276 
277 	rpipe->pipe_peer = wpipe;
278 	wpipe->pipe_peer = rpipe;
279 
280 	fd_affix(p, rf, (int)retval[0]);
281 	fd_affix(p, wf, (int)retval[1]);
282 	return (0);
283 free3:
284 	fd_abort(p, rf, (int)retval[0]);
285 free2:
286 	pipeclose(NULL, wpipe);
287 	pipeclose(NULL, rpipe);
288 
289 	return (error);
290 }
291 
292 /*
293  * Allocate kva for pipe circular buffer, the space is pageable
294  * This routine will 'realloc' the size of a pipe safely, if it fails
295  * it will retain the old buffer.
296  * If it fails it will return ENOMEM.
297  */
298 static int
299 pipespace(struct pipe *pipe, int size)
300 {
301 	void *buffer;
302 	/*
303 	 * Allocate pageable virtual address space. Physical memory is
304 	 * allocated on demand.
305 	 */
306 	buffer = (void *) uvm_km_alloc(kernel_map, round_page(size), 0,
307 	    UVM_KMF_PAGEABLE);
308 	if (buffer == NULL)
309 		return (ENOMEM);
310 
311 	/* free old resources if we're resizing */
312 	pipe_free_kmem(pipe);
313 	pipe->pipe_buffer.buffer = buffer;
314 	pipe->pipe_buffer.size = size;
315 	pipe->pipe_buffer.in = 0;
316 	pipe->pipe_buffer.out = 0;
317 	pipe->pipe_buffer.cnt = 0;
318 	atomic_add_int(&amountpipekva, pipe->pipe_buffer.size);
319 	return (0);
320 }
321 
322 /*
323  * Initialize and allocate VM and memory for pipe.
324  */
325 static int
326 pipe_create(struct pipe **pipep, int allockva, struct pipe_mutex *mutex)
327 {
328 	struct pipe *pipe;
329 	int error;
330 
331 	pipe = *pipep = pool_cache_get(pipe_cache, PR_WAITOK);
332 	mutex->pm_refcnt++;
333 
334 	/* Initialize */
335 	memset(pipe, 0, sizeof(struct pipe));
336 	pipe->pipe_state = PIPE_SIGNALR;
337 
338 	getmicrotime(&pipe->pipe_ctime);
339 	pipe->pipe_atime = pipe->pipe_ctime;
340 	pipe->pipe_mtime = pipe->pipe_ctime;
341 	pipe->pipe_lock = &mutex->pm_mutex;
342 	cv_init(&pipe->pipe_rcv, "piperd");
343 	cv_init(&pipe->pipe_wcv, "pipewr");
344 	cv_init(&pipe->pipe_draincv, "pipedrain");
345 	cv_init(&pipe->pipe_lkcv, "pipelk");
346 	selinit(&pipe->pipe_sel);
347 
348 	if (allockva && (error = pipespace(pipe, PIPE_SIZE)))
349 		return (error);
350 
351 	return (0);
352 }
353 
354 
355 /*
356  * Lock a pipe for I/O, blocking other access
357  * Called with pipe spin lock held.
358  * Return with pipe spin lock released on success.
359  */
360 static int
361 pipelock(struct pipe *pipe, int catch)
362 {
363 	int error;
364 
365 	KASSERT(mutex_owned(pipe->pipe_lock));
366 
367 	while (pipe->pipe_state & PIPE_LOCKFL) {
368 		pipe->pipe_state |= PIPE_LWANT;
369 		if (catch) {
370 			error = cv_wait_sig(&pipe->pipe_lkcv, pipe->pipe_lock);
371 			if (error != 0)
372 				return error;
373 		} else
374 			cv_wait(&pipe->pipe_lkcv, pipe->pipe_lock);
375 	}
376 
377 	pipe->pipe_state |= PIPE_LOCKFL;
378 
379 	return 0;
380 }
381 
382 /*
383  * unlock a pipe I/O lock
384  */
385 static inline void
386 pipeunlock(struct pipe *pipe)
387 {
388 
389 	KASSERT(pipe->pipe_state & PIPE_LOCKFL);
390 
391 	pipe->pipe_state &= ~PIPE_LOCKFL;
392 	if (pipe->pipe_state & PIPE_LWANT) {
393 		pipe->pipe_state &= ~PIPE_LWANT;
394 		cv_broadcast(&pipe->pipe_lkcv);
395 	}
396 }
397 
398 /*
399  * Select/poll wakup. This also sends SIGIO to peer connected to
400  * 'sigpipe' side of pipe.
401  */
402 static void
403 pipeselwakeup(struct pipe *selp, struct pipe *sigp, int code)
404 {
405 	int band;
406 
407 	switch (code) {
408 	case POLL_IN:
409 		band = POLLIN|POLLRDNORM;
410 		break;
411 	case POLL_OUT:
412 		band = POLLOUT|POLLWRNORM;
413 		break;
414 	case POLL_HUP:
415 		band = POLLHUP;
416 		break;
417 #if POLL_HUP != POLL_ERR
418 	case POLL_ERR:
419 		band = POLLERR;
420 		break;
421 #endif
422 	default:
423 		band = 0;
424 #ifdef DIAGNOSTIC
425 		printf("bad siginfo code %d in pipe notification.\n", code);
426 #endif
427 		break;
428 	}
429 
430 	selnotify(&selp->pipe_sel, band, NOTE_SUBMIT);
431 
432 	if (sigp == NULL || (sigp->pipe_state & PIPE_ASYNC) == 0)
433 		return;
434 
435 	fownsignal(sigp->pipe_pgid, SIGIO, code, band, selp);
436 }
437 
438 /* ARGSUSED */
439 static int
440 pipe_read(struct file *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
441     int flags)
442 {
443 	struct pipe *rpipe = (struct pipe *) fp->f_data;
444 	struct pipebuf *bp = &rpipe->pipe_buffer;
445 	kmutex_t *lock = rpipe->pipe_lock;
446 	int error;
447 	size_t nread = 0;
448 	size_t size;
449 	size_t ocnt;
450 
451 	mutex_enter(lock);
452 	++rpipe->pipe_busy;
453 	ocnt = bp->cnt;
454 
455 again:
456 	error = pipelock(rpipe, 1);
457 	if (error)
458 		goto unlocked_error;
459 
460 	while (uio->uio_resid) {
461 		/*
462 		 * normal pipe buffer receive
463 		 */
464 		if (bp->cnt > 0) {
465 			size = bp->size - bp->out;
466 			if (size > bp->cnt)
467 				size = bp->cnt;
468 			if (size > uio->uio_resid)
469 				size = uio->uio_resid;
470 
471 			mutex_exit(lock);
472 			error = uiomove((char *)bp->buffer + bp->out, size, uio);
473 			mutex_enter(lock);
474 			if (error)
475 				break;
476 
477 			bp->out += size;
478 			if (bp->out >= bp->size)
479 				bp->out = 0;
480 
481 			bp->cnt -= size;
482 
483 			/*
484 			 * If there is no more to read in the pipe, reset
485 			 * its pointers to the beginning.  This improves
486 			 * cache hit stats.
487 			 */
488 			if (bp->cnt == 0) {
489 				bp->in = 0;
490 				bp->out = 0;
491 			}
492 			nread += size;
493 			continue;
494 		}
495 
496 #ifndef PIPE_NODIRECT
497 		if ((rpipe->pipe_state & PIPE_DIRECTR) != 0) {
498 			/*
499 			 * Direct copy, bypassing a kernel buffer.
500 			 */
501 			void *	va;
502 
503 			KASSERT(rpipe->pipe_state & PIPE_DIRECTW);
504 
505 			size = rpipe->pipe_map.cnt;
506 			if (size > uio->uio_resid)
507 				size = uio->uio_resid;
508 
509 			va = (char *)rpipe->pipe_map.kva + rpipe->pipe_map.pos;
510 			mutex_exit(lock);
511 			error = uiomove(va, size, uio);
512 			mutex_enter(lock);
513 			if (error)
514 				break;
515 			nread += size;
516 			rpipe->pipe_map.pos += size;
517 			rpipe->pipe_map.cnt -= size;
518 			if (rpipe->pipe_map.cnt == 0) {
519 				rpipe->pipe_state &= ~PIPE_DIRECTR;
520 				cv_broadcast(&rpipe->pipe_wcv);
521 			}
522 			continue;
523 		}
524 #endif
525 		/*
526 		 * Break if some data was read.
527 		 */
528 		if (nread > 0)
529 			break;
530 
531 		/*
532 		 * detect EOF condition
533 		 * read returns 0 on EOF, no need to set error
534 		 */
535 		if (rpipe->pipe_state & PIPE_EOF)
536 			break;
537 
538 		/*
539 		 * don't block on non-blocking I/O
540 		 */
541 		if (fp->f_flag & FNONBLOCK) {
542 			error = EAGAIN;
543 			break;
544 		}
545 
546 		/*
547 		 * Unlock the pipe buffer for our remaining processing.
548 		 * We will either break out with an error or we will
549 		 * sleep and relock to loop.
550 		 */
551 		pipeunlock(rpipe);
552 
553 		/*
554 		 * Re-check to see if more direct writes are pending.
555 		 */
556 		if ((rpipe->pipe_state & PIPE_DIRECTR) != 0)
557 			goto again;
558 
559 		/*
560 		 * We want to read more, wake up select/poll.
561 		 */
562 		pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_IN);
563 
564 		/*
565 		 * If the "write-side" is blocked, wake it up now.
566 		 */
567 		cv_broadcast(&rpipe->pipe_wcv);
568 
569 		/* Now wait until the pipe is filled */
570 		error = cv_wait_sig(&rpipe->pipe_rcv, lock);
571 		if (error != 0)
572 			goto unlocked_error;
573 		goto again;
574 	}
575 
576 	if (error == 0)
577 		getmicrotime(&rpipe->pipe_atime);
578 	pipeunlock(rpipe);
579 
580 unlocked_error:
581 	--rpipe->pipe_busy;
582 	if (rpipe->pipe_busy == 0) {
583 		cv_broadcast(&rpipe->pipe_draincv);
584 	}
585 	if (bp->cnt < MINPIPESIZE) {
586 		cv_broadcast(&rpipe->pipe_wcv);
587 	}
588 
589 	/*
590 	 * If anything was read off the buffer, signal to the writer it's
591 	 * possible to write more data. Also send signal if we are here for the
592 	 * first time after last write.
593 	 */
594 	if ((bp->size - bp->cnt) >= PIPE_BUF
595 	    && (ocnt != bp->cnt || (rpipe->pipe_state & PIPE_SIGNALR))) {
596 		pipeselwakeup(rpipe, rpipe->pipe_peer, POLL_OUT);
597 		rpipe->pipe_state &= ~PIPE_SIGNALR;
598 	}
599 
600 	mutex_exit(lock);
601 	return (error);
602 }
603 
604 #ifndef PIPE_NODIRECT
605 /*
606  * Allocate structure for loan transfer.
607  */
608 static int
609 pipe_loan_alloc(struct pipe *wpipe, int npages)
610 {
611 	vsize_t len;
612 
613 	len = (vsize_t)npages << PAGE_SHIFT;
614 	atomic_add_int(&amountpipekva, len);
615 	wpipe->pipe_map.kva = uvm_km_alloc(kernel_map, len, 0,
616 	    UVM_KMF_VAONLY | UVM_KMF_WAITVA);
617 	if (wpipe->pipe_map.kva == 0) {
618 		atomic_add_int(&amountpipekva, -len);
619 		return (ENOMEM);
620 	}
621 
622 	wpipe->pipe_map.npages = npages;
623 	wpipe->pipe_map.pgs = malloc(npages * sizeof(struct vm_page *), M_PIPE,
624 	    M_WAITOK);
625 	return (0);
626 }
627 
628 /*
629  * Free resources allocated for loan transfer.
630  */
631 static void
632 pipe_loan_free(struct pipe *wpipe)
633 {
634 	vsize_t len;
635 
636 	len = (vsize_t)wpipe->pipe_map.npages << PAGE_SHIFT;
637 	uvm_km_free(kernel_map, wpipe->pipe_map.kva, len, UVM_KMF_VAONLY);
638 	wpipe->pipe_map.kva = 0;
639 	atomic_add_int(&amountpipekva, -len);
640 	free(wpipe->pipe_map.pgs, M_PIPE);
641 	wpipe->pipe_map.pgs = NULL;
642 }
643 
644 /*
645  * NetBSD direct write, using uvm_loan() mechanism.
646  * This implements the pipe buffer write mechanism.  Note that only
647  * a direct write OR a normal pipe write can be pending at any given time.
648  * If there are any characters in the pipe buffer, the direct write will
649  * be deferred until the receiving process grabs all of the bytes from
650  * the pipe buffer.  Then the direct mapping write is set-up.
651  *
652  * Called with the long-term pipe lock held.
653  */
654 static int
655 pipe_direct_write(struct file *fp, struct pipe *wpipe, struct uio *uio)
656 {
657 	int error, npages, j;
658 	struct vm_page **pgs;
659 	vaddr_t bbase, kva, base, bend;
660 	vsize_t blen, bcnt;
661 	voff_t bpos;
662 	kmutex_t *lock = wpipe->pipe_lock;
663 
664 	KASSERT(mutex_owned(wpipe->pipe_lock));
665 	KASSERT(wpipe->pipe_map.cnt == 0);
666 
667 	mutex_exit(lock);
668 
669 	/*
670 	 * Handle first PIPE_CHUNK_SIZE bytes of buffer. Deal with buffers
671 	 * not aligned to PAGE_SIZE.
672 	 */
673 	bbase = (vaddr_t)uio->uio_iov->iov_base;
674 	base = trunc_page(bbase);
675 	bend = round_page(bbase + uio->uio_iov->iov_len);
676 	blen = bend - base;
677 	bpos = bbase - base;
678 
679 	if (blen > PIPE_DIRECT_CHUNK) {
680 		blen = PIPE_DIRECT_CHUNK;
681 		bend = base + blen;
682 		bcnt = PIPE_DIRECT_CHUNK - bpos;
683 	} else {
684 		bcnt = uio->uio_iov->iov_len;
685 	}
686 	npages = blen >> PAGE_SHIFT;
687 
688 	/*
689 	 * Free the old kva if we need more pages than we have
690 	 * allocated.
691 	 */
692 	if (wpipe->pipe_map.kva != 0 && npages > wpipe->pipe_map.npages)
693 		pipe_loan_free(wpipe);
694 
695 	/* Allocate new kva. */
696 	if (wpipe->pipe_map.kva == 0) {
697 		error = pipe_loan_alloc(wpipe, npages);
698 		if (error) {
699 			mutex_enter(lock);
700 			return (error);
701 		}
702 	}
703 
704 	/* Loan the write buffer memory from writer process */
705 	pgs = wpipe->pipe_map.pgs;
706 	error = uvm_loan(&uio->uio_vmspace->vm_map, base, blen,
707 			 pgs, UVM_LOAN_TOPAGE);
708 	if (error) {
709 		pipe_loan_free(wpipe);
710 		mutex_enter(lock);
711 		return (ENOMEM); /* so that caller fallback to ordinary write */
712 	}
713 
714 	/* Enter the loaned pages to kva */
715 	kva = wpipe->pipe_map.kva;
716 	for (j = 0; j < npages; j++, kva += PAGE_SIZE) {
717 		pmap_kenter_pa(kva, VM_PAGE_TO_PHYS(pgs[j]), VM_PROT_READ);
718 	}
719 	pmap_update(pmap_kernel());
720 
721 	/* Now we can put the pipe in direct write mode */
722 	wpipe->pipe_map.pos = bpos;
723 	wpipe->pipe_map.cnt = bcnt;
724 
725 	/*
726 	 * But before we can let someone do a direct read, we
727 	 * have to wait until the pipe is drained.  Release the
728 	 * pipe lock while we wait.
729 	 */
730 	mutex_enter(lock);
731 	wpipe->pipe_state |= PIPE_DIRECTW;
732 	pipeunlock(wpipe);
733 
734 	while (error == 0 && wpipe->pipe_buffer.cnt > 0) {
735 		cv_broadcast(&wpipe->pipe_rcv);
736 		error = cv_wait_sig(&wpipe->pipe_wcv, lock);
737 		if (error == 0 && wpipe->pipe_state & PIPE_EOF)
738 			error = EPIPE;
739 	}
740 
741 	/* Pipe is drained; next read will off the direct buffer */
742 	wpipe->pipe_state |= PIPE_DIRECTR;
743 
744 	/* Wait until the reader is done */
745 	while (error == 0 && (wpipe->pipe_state & PIPE_DIRECTR)) {
746 		cv_broadcast(&wpipe->pipe_rcv);
747 		pipeselwakeup(wpipe, wpipe, POLL_IN);
748 		error = cv_wait_sig(&wpipe->pipe_wcv, lock);
749 		if (error == 0 && wpipe->pipe_state & PIPE_EOF)
750 			error = EPIPE;
751 	}
752 
753 	/* Take pipe out of direct write mode */
754 	wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTR);
755 
756 	/* Acquire the pipe lock and cleanup */
757 	(void)pipelock(wpipe, 0);
758 	mutex_exit(lock);
759 
760 	if (pgs != NULL) {
761 		pmap_kremove(wpipe->pipe_map.kva, blen);
762 		pmap_update(pmap_kernel());
763 		uvm_unloan(pgs, npages, UVM_LOAN_TOPAGE);
764 	}
765 	if (error || amountpipekva > maxpipekva)
766 		pipe_loan_free(wpipe);
767 
768 	mutex_enter(lock);
769 	if (error) {
770 		pipeselwakeup(wpipe, wpipe, POLL_ERR);
771 
772 		/*
773 		 * If nothing was read from what we offered, return error
774 		 * straight on. Otherwise update uio resid first. Caller
775 		 * will deal with the error condition, returning short
776 		 * write, error, or restarting the write(2) as appropriate.
777 		 */
778 		if (wpipe->pipe_map.cnt == bcnt) {
779 			wpipe->pipe_map.cnt = 0;
780 			cv_broadcast(&wpipe->pipe_wcv);
781 			return (error);
782 		}
783 
784 		bcnt -= wpipe->pipe_map.cnt;
785 	}
786 
787 	uio->uio_resid -= bcnt;
788 	/* uio_offset not updated, not set/used for write(2) */
789 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + bcnt;
790 	uio->uio_iov->iov_len -= bcnt;
791 	if (uio->uio_iov->iov_len == 0) {
792 		uio->uio_iov++;
793 		uio->uio_iovcnt--;
794 	}
795 
796 	wpipe->pipe_map.cnt = 0;
797 	return (error);
798 }
799 #endif /* !PIPE_NODIRECT */
800 
801 static int
802 pipe_write(struct file *fp, off_t *offset, struct uio *uio, kauth_cred_t cred,
803     int flags)
804 {
805 	struct pipe *wpipe, *rpipe;
806 	struct pipebuf *bp;
807 	kmutex_t *lock;
808 	int error;
809 
810 	/* We want to write to our peer */
811 	rpipe = (struct pipe *) fp->f_data;
812 	lock = rpipe->pipe_lock;
813 	error = 0;
814 
815 	mutex_enter(lock);
816 	wpipe = rpipe->pipe_peer;
817 
818 	/*
819 	 * Detect loss of pipe read side, issue SIGPIPE if lost.
820 	 */
821 	if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) != 0) {
822 		mutex_exit(lock);
823 		return EPIPE;
824 	}
825 	++wpipe->pipe_busy;
826 
827 	/* Aquire the long-term pipe lock */
828 	if ((error = pipelock(wpipe, 1)) != 0) {
829 		--wpipe->pipe_busy;
830 		if (wpipe->pipe_busy == 0) {
831 			cv_broadcast(&wpipe->pipe_draincv);
832 		}
833 		mutex_exit(lock);
834 		return (error);
835 	}
836 
837 	bp = &wpipe->pipe_buffer;
838 
839 	/*
840 	 * If it is advantageous to resize the pipe buffer, do so.
841 	 */
842 	if ((uio->uio_resid > PIPE_SIZE) &&
843 	    (nbigpipe < maxbigpipes) &&
844 #ifndef PIPE_NODIRECT
845 	    (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
846 #endif
847 	    (bp->size <= PIPE_SIZE) && (bp->cnt == 0)) {
848 
849 		if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
850 			atomic_inc_uint(&nbigpipe);
851 	}
852 
853 	while (uio->uio_resid) {
854 		size_t space;
855 
856 #ifndef PIPE_NODIRECT
857 		/*
858 		 * Pipe buffered writes cannot be coincidental with
859 		 * direct writes.  Also, only one direct write can be
860 		 * in progress at any one time.  We wait until the currently
861 		 * executing direct write is completed before continuing.
862 		 *
863 		 * We break out if a signal occurs or the reader goes away.
864 		 */
865 		while (error == 0 && wpipe->pipe_state & PIPE_DIRECTW) {
866 			cv_broadcast(&wpipe->pipe_rcv);
867 			pipeunlock(wpipe);
868 			error = cv_wait_sig(&wpipe->pipe_wcv, lock);
869 			(void)pipelock(wpipe, 0);
870 			if (wpipe->pipe_state & PIPE_EOF)
871 				error = EPIPE;
872 		}
873 		if (error)
874 			break;
875 
876 		/*
877 		 * If the transfer is large, we can gain performance if
878 		 * we do process-to-process copies directly.
879 		 * If the write is non-blocking, we don't use the
880 		 * direct write mechanism.
881 		 *
882 		 * The direct write mechanism will detect the reader going
883 		 * away on us.
884 		 */
885 		if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
886 		    (fp->f_flag & FNONBLOCK) == 0 &&
887 		    (wpipe->pipe_map.kva || (amountpipekva < limitpipekva))) {
888 			error = pipe_direct_write(fp, wpipe, uio);
889 
890 			/*
891 			 * Break out if error occurred, unless it's ENOMEM.
892 			 * ENOMEM means we failed to allocate some resources
893 			 * for direct write, so we just fallback to ordinary
894 			 * write. If the direct write was successful,
895 			 * process rest of data via ordinary write.
896 			 */
897 			if (error == 0)
898 				continue;
899 
900 			if (error != ENOMEM)
901 				break;
902 		}
903 #endif /* PIPE_NODIRECT */
904 
905 		space = bp->size - bp->cnt;
906 
907 		/* Writes of size <= PIPE_BUF must be atomic. */
908 		if ((space < uio->uio_resid) && (uio->uio_resid <= PIPE_BUF))
909 			space = 0;
910 
911 		if (space > 0) {
912 			int size;	/* Transfer size */
913 			int segsize;	/* first segment to transfer */
914 
915 			/*
916 			 * Transfer size is minimum of uio transfer
917 			 * and free space in pipe buffer.
918 			 */
919 			if (space > uio->uio_resid)
920 				size = uio->uio_resid;
921 			else
922 				size = space;
923 			/*
924 			 * First segment to transfer is minimum of
925 			 * transfer size and contiguous space in
926 			 * pipe buffer.  If first segment to transfer
927 			 * is less than the transfer size, we've got
928 			 * a wraparound in the buffer.
929 			 */
930 			segsize = bp->size - bp->in;
931 			if (segsize > size)
932 				segsize = size;
933 
934 			/* Transfer first segment */
935 			mutex_exit(lock);
936 			error = uiomove((char *)bp->buffer + bp->in, segsize,
937 			    uio);
938 
939 			if (error == 0 && segsize < size) {
940 				/*
941 				 * Transfer remaining part now, to
942 				 * support atomic writes.  Wraparound
943 				 * happened.
944 				 */
945 #ifdef DEBUG
946 				if (bp->in + segsize != bp->size)
947 					panic("Expected pipe buffer wraparound disappeared");
948 #endif
949 
950 				error = uiomove(bp->buffer,
951 				    size - segsize, uio);
952 			}
953 			mutex_enter(lock);
954 			if (error)
955 				break;
956 
957 			bp->in += size;
958 			if (bp->in >= bp->size) {
959 #ifdef DEBUG
960 				if (bp->in != size - segsize + bp->size)
961 					panic("Expected wraparound bad");
962 #endif
963 				bp->in = size - segsize;
964 			}
965 
966 			bp->cnt += size;
967 #ifdef DEBUG
968 			if (bp->cnt > bp->size)
969 				panic("Pipe buffer overflow");
970 #endif
971 		} else {
972 			/*
973 			 * If the "read-side" has been blocked, wake it up now.
974 			 */
975 			cv_broadcast(&wpipe->pipe_rcv);
976 
977 			/*
978 			 * don't block on non-blocking I/O
979 			 */
980 			if (fp->f_flag & FNONBLOCK) {
981 				error = EAGAIN;
982 				break;
983 			}
984 
985 			/*
986 			 * We have no more space and have something to offer,
987 			 * wake up select/poll.
988 			 */
989 			if (bp->cnt)
990 				pipeselwakeup(wpipe, wpipe, POLL_OUT);
991 
992 			pipeunlock(wpipe);
993 			error = cv_wait_sig(&wpipe->pipe_wcv, lock);
994 			(void)pipelock(wpipe, 0);
995 			if (error != 0)
996 				break;
997 			/*
998 			 * If read side wants to go away, we just issue a signal
999 			 * to ourselves.
1000 			 */
1001 			if (wpipe->pipe_state & PIPE_EOF) {
1002 				error = EPIPE;
1003 				break;
1004 			}
1005 		}
1006 	}
1007 
1008 	--wpipe->pipe_busy;
1009 	if (wpipe->pipe_busy == 0) {
1010 		cv_broadcast(&wpipe->pipe_draincv);
1011 	}
1012 	if (bp->cnt > 0) {
1013 		cv_broadcast(&wpipe->pipe_rcv);
1014 	}
1015 
1016 	/*
1017 	 * Don't return EPIPE if I/O was successful
1018 	 */
1019 	if (error == EPIPE && bp->cnt == 0 && uio->uio_resid == 0)
1020 		error = 0;
1021 
1022 	if (error == 0)
1023 		getmicrotime(&wpipe->pipe_mtime);
1024 
1025 	/*
1026 	 * We have something to offer, wake up select/poll.
1027 	 * wpipe->pipe_map.cnt is always 0 in this point (direct write
1028 	 * is only done synchronously), so check only wpipe->pipe_buffer.cnt
1029 	 */
1030 	if (bp->cnt)
1031 		pipeselwakeup(wpipe, wpipe, POLL_OUT);
1032 
1033 	/*
1034 	 * Arrange for next read(2) to do a signal.
1035 	 */
1036 	wpipe->pipe_state |= PIPE_SIGNALR;
1037 
1038 	pipeunlock(wpipe);
1039 	mutex_exit(lock);
1040 	return (error);
1041 }
1042 
1043 /*
1044  * we implement a very minimal set of ioctls for compatibility with sockets.
1045  */
1046 int
1047 pipe_ioctl(struct file *fp, u_long cmd, void *data)
1048 {
1049 	struct pipe *pipe = fp->f_data;
1050 	kmutex_t *lock = pipe->pipe_lock;
1051 
1052 	switch (cmd) {
1053 
1054 	case FIONBIO:
1055 		return (0);
1056 
1057 	case FIOASYNC:
1058 		mutex_enter(lock);
1059 		if (*(int *)data) {
1060 			pipe->pipe_state |= PIPE_ASYNC;
1061 		} else {
1062 			pipe->pipe_state &= ~PIPE_ASYNC;
1063 		}
1064 		mutex_exit(lock);
1065 		return (0);
1066 
1067 	case FIONREAD:
1068 		mutex_enter(lock);
1069 #ifndef PIPE_NODIRECT
1070 		if (pipe->pipe_state & PIPE_DIRECTW)
1071 			*(int *)data = pipe->pipe_map.cnt;
1072 		else
1073 #endif
1074 			*(int *)data = pipe->pipe_buffer.cnt;
1075 		mutex_exit(lock);
1076 		return (0);
1077 
1078 	case FIONWRITE:
1079 		/* Look at other side */
1080 		pipe = pipe->pipe_peer;
1081 		mutex_enter(lock);
1082 #ifndef PIPE_NODIRECT
1083 		if (pipe->pipe_state & PIPE_DIRECTW)
1084 			*(int *)data = pipe->pipe_map.cnt;
1085 		else
1086 #endif
1087 			*(int *)data = pipe->pipe_buffer.cnt;
1088 		mutex_exit(lock);
1089 		return (0);
1090 
1091 	case FIONSPACE:
1092 		/* Look at other side */
1093 		pipe = pipe->pipe_peer;
1094 		mutex_enter(lock);
1095 #ifndef PIPE_NODIRECT
1096 		/*
1097 		 * If we're in direct-mode, we don't really have a
1098 		 * send queue, and any other write will block. Thus
1099 		 * zero seems like the best answer.
1100 		 */
1101 		if (pipe->pipe_state & PIPE_DIRECTW)
1102 			*(int *)data = 0;
1103 		else
1104 #endif
1105 			*(int *)data = pipe->pipe_buffer.size -
1106 			    pipe->pipe_buffer.cnt;
1107 		mutex_exit(lock);
1108 		return (0);
1109 
1110 	case TIOCSPGRP:
1111 	case FIOSETOWN:
1112 		return fsetown(&pipe->pipe_pgid, cmd, data);
1113 
1114 	case TIOCGPGRP:
1115 	case FIOGETOWN:
1116 		return fgetown(pipe->pipe_pgid, cmd, data);
1117 
1118 	}
1119 	return (EPASSTHROUGH);
1120 }
1121 
1122 int
1123 pipe_poll(struct file *fp, int events)
1124 {
1125 	struct pipe *rpipe = fp->f_data;
1126 	struct pipe *wpipe;
1127 	int eof = 0;
1128 	int revents = 0;
1129 
1130 	mutex_enter(rpipe->pipe_lock);
1131 	wpipe = rpipe->pipe_peer;
1132 
1133 	if (events & (POLLIN | POLLRDNORM))
1134 		if ((rpipe->pipe_buffer.cnt > 0) ||
1135 #ifndef PIPE_NODIRECT
1136 		    (rpipe->pipe_state & PIPE_DIRECTR) ||
1137 #endif
1138 		    (rpipe->pipe_state & PIPE_EOF))
1139 			revents |= events & (POLLIN | POLLRDNORM);
1140 
1141 	eof |= (rpipe->pipe_state & PIPE_EOF);
1142 
1143 	if (wpipe == NULL)
1144 		revents |= events & (POLLOUT | POLLWRNORM);
1145 	else {
1146 		if (events & (POLLOUT | POLLWRNORM))
1147 			if ((wpipe->pipe_state & PIPE_EOF) || (
1148 #ifndef PIPE_NODIRECT
1149 			     (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1150 #endif
1151 			     (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1152 				revents |= events & (POLLOUT | POLLWRNORM);
1153 
1154 		eof |= (wpipe->pipe_state & PIPE_EOF);
1155 	}
1156 
1157 	if (wpipe == NULL || eof)
1158 		revents |= POLLHUP;
1159 
1160 	if (revents == 0) {
1161 		if (events & (POLLIN | POLLRDNORM))
1162 			selrecord(curlwp, &rpipe->pipe_sel);
1163 
1164 		if (events & (POLLOUT | POLLWRNORM))
1165 			selrecord(curlwp, &wpipe->pipe_sel);
1166 	}
1167 	mutex_exit(rpipe->pipe_lock);
1168 
1169 	return (revents);
1170 }
1171 
1172 static int
1173 pipe_stat(struct file *fp, struct stat *ub)
1174 {
1175 	struct pipe *pipe = fp->f_data;
1176 
1177 	memset((void *)ub, 0, sizeof(*ub));
1178 	ub->st_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1179 	ub->st_blksize = pipe->pipe_buffer.size;
1180 	if (ub->st_blksize == 0 && pipe->pipe_peer)
1181 		ub->st_blksize = pipe->pipe_peer->pipe_buffer.size;
1182 	ub->st_size = pipe->pipe_buffer.cnt;
1183 	ub->st_blocks = (ub->st_size) ? 1 : 0;
1184 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_atime, &ub->st_atimespec);
1185 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_mtime, &ub->st_mtimespec);
1186 	TIMEVAL_TO_TIMESPEC(&pipe->pipe_ctime, &ub->st_ctimespec);
1187 	ub->st_uid = kauth_cred_geteuid(fp->f_cred);
1188 	ub->st_gid = kauth_cred_getegid(fp->f_cred);
1189 
1190 	/*
1191 	 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1192 	 * XXX (st_dev, st_ino) should be unique.
1193 	 */
1194 	return (0);
1195 }
1196 
1197 /* ARGSUSED */
1198 static int
1199 pipe_close(struct file *fp)
1200 {
1201 	struct pipe *pipe = fp->f_data;
1202 
1203 	fp->f_data = NULL;
1204 	pipeclose(fp, pipe);
1205 	return (0);
1206 }
1207 
1208 static void
1209 pipe_free_kmem(struct pipe *pipe)
1210 {
1211 
1212 	if (pipe->pipe_buffer.buffer != NULL) {
1213 		if (pipe->pipe_buffer.size > PIPE_SIZE)
1214 			atomic_dec_uint(&nbigpipe);
1215 		uvm_km_free(kernel_map,
1216 			(vaddr_t)pipe->pipe_buffer.buffer,
1217 			pipe->pipe_buffer.size, UVM_KMF_PAGEABLE);
1218 		atomic_add_int(&amountpipekva, -pipe->pipe_buffer.size);
1219 		pipe->pipe_buffer.buffer = NULL;
1220 	}
1221 #ifndef PIPE_NODIRECT
1222 	if (pipe->pipe_map.kva != 0) {
1223 		pipe_loan_free(pipe);
1224 		pipe->pipe_map.cnt = 0;
1225 		pipe->pipe_map.kva = 0;
1226 		pipe->pipe_map.pos = 0;
1227 		pipe->pipe_map.npages = 0;
1228 	}
1229 #endif /* !PIPE_NODIRECT */
1230 }
1231 
1232 /*
1233  * shutdown the pipe
1234  */
1235 static void
1236 pipeclose(struct file *fp, struct pipe *pipe)
1237 {
1238 	struct pipe_mutex *mutex;
1239 	kmutex_t *lock;
1240 	struct pipe *ppipe;
1241 	u_int refcnt;
1242 
1243 	if (pipe == NULL)
1244 		return;
1245 
1246 	KASSERT(cv_is_valid(&pipe->pipe_rcv));
1247 	KASSERT(cv_is_valid(&pipe->pipe_wcv));
1248 	KASSERT(cv_is_valid(&pipe->pipe_draincv));
1249 	KASSERT(cv_is_valid(&pipe->pipe_lkcv));
1250 
1251 	lock = pipe->pipe_lock;
1252 	mutex_enter(lock);
1253 	pipeselwakeup(pipe, pipe, POLL_HUP);
1254 
1255 	/*
1256 	 * If the other side is blocked, wake it up saying that
1257 	 * we want to close it down.
1258 	 */
1259 	pipe->pipe_state |= PIPE_EOF;
1260 	if (pipe->pipe_busy) {
1261 		while (pipe->pipe_busy) {
1262 			cv_broadcast(&pipe->pipe_wcv);
1263 			cv_wait_sig(&pipe->pipe_draincv, lock);
1264 		}
1265 	}
1266 
1267 	/*
1268 	 * Disconnect from peer
1269 	 */
1270 	if ((ppipe = pipe->pipe_peer) != NULL) {
1271 		pipeselwakeup(ppipe, ppipe, POLL_HUP);
1272 		ppipe->pipe_state |= PIPE_EOF;
1273 		cv_broadcast(&ppipe->pipe_rcv);
1274 		ppipe->pipe_peer = NULL;
1275 	}
1276 
1277 	KASSERT((pipe->pipe_state & PIPE_LOCKFL) == 0);
1278 
1279 	mutex = (struct pipe_mutex *)lock;
1280 	refcnt = --(mutex->pm_refcnt);
1281 	KASSERT(refcnt == 0 || refcnt == 1);
1282 	mutex_exit(lock);
1283 
1284 	/*
1285 	 * free resources
1286 	 */
1287 	pipe_free_kmem(pipe);
1288 	cv_destroy(&pipe->pipe_rcv);
1289 	cv_destroy(&pipe->pipe_wcv);
1290 	cv_destroy(&pipe->pipe_draincv);
1291 	cv_destroy(&pipe->pipe_lkcv);
1292 	seldestroy(&pipe->pipe_sel);
1293 	pool_cache_put(pipe_cache, pipe);
1294 	if (refcnt == 0)
1295 		pool_cache_put(pipe_mutex_cache, mutex);
1296 }
1297 
1298 static void
1299 filt_pipedetach(struct knote *kn)
1300 {
1301 	struct pipe *pipe;
1302 	kmutex_t *lock;
1303 
1304 	pipe = ((file_t *)kn->kn_obj)->f_data;
1305 	lock = pipe->pipe_lock;
1306 
1307 	mutex_enter(lock);
1308 
1309 	switch(kn->kn_filter) {
1310 	case EVFILT_WRITE:
1311 		/* need the peer structure, not our own */
1312 		pipe = pipe->pipe_peer;
1313 
1314 		/* if reader end already closed, just return */
1315 		if (pipe == NULL) {
1316 			mutex_exit(lock);
1317 			return;
1318 		}
1319 
1320 		break;
1321 	default:
1322 		/* nothing to do */
1323 		break;
1324 	}
1325 
1326 #ifdef DIAGNOSTIC
1327 	if (kn->kn_hook != pipe)
1328 		panic("filt_pipedetach: inconsistent knote");
1329 #endif
1330 
1331 	SLIST_REMOVE(&pipe->pipe_sel.sel_klist, kn, knote, kn_selnext);
1332 	mutex_exit(lock);
1333 }
1334 
1335 /*ARGSUSED*/
1336 static int
1337 filt_piperead(struct knote *kn, long hint)
1338 {
1339 	struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
1340 	struct pipe *wpipe;
1341 
1342 	if ((hint & NOTE_SUBMIT) == 0) {
1343 		mutex_enter(rpipe->pipe_lock);
1344 	}
1345 	wpipe = rpipe->pipe_peer;
1346 	kn->kn_data = rpipe->pipe_buffer.cnt;
1347 
1348 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1349 		kn->kn_data = rpipe->pipe_map.cnt;
1350 
1351 	if ((rpipe->pipe_state & PIPE_EOF) ||
1352 	    (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1353 		kn->kn_flags |= EV_EOF;
1354 		if ((hint & NOTE_SUBMIT) == 0) {
1355 			mutex_exit(rpipe->pipe_lock);
1356 		}
1357 		return (1);
1358 	}
1359 
1360 	if ((hint & NOTE_SUBMIT) == 0) {
1361 		mutex_exit(rpipe->pipe_lock);
1362 	}
1363 	return (kn->kn_data > 0);
1364 }
1365 
1366 /*ARGSUSED*/
1367 static int
1368 filt_pipewrite(struct knote *kn, long hint)
1369 {
1370 	struct pipe *rpipe = ((file_t *)kn->kn_obj)->f_data;
1371 	struct pipe *wpipe;
1372 
1373 	if ((hint & NOTE_SUBMIT) == 0) {
1374 		mutex_enter(rpipe->pipe_lock);
1375 	}
1376 	wpipe = rpipe->pipe_peer;
1377 
1378 	if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1379 		kn->kn_data = 0;
1380 		kn->kn_flags |= EV_EOF;
1381 		if ((hint & NOTE_SUBMIT) == 0) {
1382 			mutex_exit(rpipe->pipe_lock);
1383 		}
1384 		return (1);
1385 	}
1386 	kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1387 	if (wpipe->pipe_state & PIPE_DIRECTW)
1388 		kn->kn_data = 0;
1389 
1390 	if ((hint & NOTE_SUBMIT) == 0) {
1391 		mutex_exit(rpipe->pipe_lock);
1392 	}
1393 	return (kn->kn_data >= PIPE_BUF);
1394 }
1395 
1396 static const struct filterops pipe_rfiltops =
1397 	{ 1, NULL, filt_pipedetach, filt_piperead };
1398 static const struct filterops pipe_wfiltops =
1399 	{ 1, NULL, filt_pipedetach, filt_pipewrite };
1400 
1401 /*ARGSUSED*/
1402 static int
1403 pipe_kqfilter(struct file *fp, struct knote *kn)
1404 {
1405 	struct pipe *pipe;
1406 	kmutex_t *lock;
1407 
1408 	pipe = ((file_t *)kn->kn_obj)->f_data;
1409 	lock = pipe->pipe_lock;
1410 
1411 	mutex_enter(lock);
1412 
1413 	switch (kn->kn_filter) {
1414 	case EVFILT_READ:
1415 		kn->kn_fop = &pipe_rfiltops;
1416 		break;
1417 	case EVFILT_WRITE:
1418 		kn->kn_fop = &pipe_wfiltops;
1419 		pipe = pipe->pipe_peer;
1420 		if (pipe == NULL) {
1421 			/* other end of pipe has been closed */
1422 			mutex_exit(lock);
1423 			return (EBADF);
1424 		}
1425 		break;
1426 	default:
1427 		mutex_exit(lock);
1428 		return (EINVAL);
1429 	}
1430 
1431 	kn->kn_hook = pipe;
1432 	SLIST_INSERT_HEAD(&pipe->pipe_sel.sel_klist, kn, kn_selnext);
1433 	mutex_exit(lock);
1434 
1435 	return (0);
1436 }
1437 
1438 /*
1439  * Handle pipe sysctls.
1440  */
1441 SYSCTL_SETUP(sysctl_kern_pipe_setup, "sysctl kern.pipe subtree setup")
1442 {
1443 
1444 	sysctl_createv(clog, 0, NULL, NULL,
1445 		       CTLFLAG_PERMANENT,
1446 		       CTLTYPE_NODE, "kern", NULL,
1447 		       NULL, 0, NULL, 0,
1448 		       CTL_KERN, CTL_EOL);
1449 	sysctl_createv(clog, 0, NULL, NULL,
1450 		       CTLFLAG_PERMANENT,
1451 		       CTLTYPE_NODE, "pipe",
1452 		       SYSCTL_DESCR("Pipe settings"),
1453 		       NULL, 0, NULL, 0,
1454 		       CTL_KERN, KERN_PIPE, CTL_EOL);
1455 
1456 	sysctl_createv(clog, 0, NULL, NULL,
1457 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1458 		       CTLTYPE_INT, "maxkvasz",
1459 		       SYSCTL_DESCR("Maximum amount of kernel memory to be "
1460 				    "used for pipes"),
1461 		       NULL, 0, &maxpipekva, 0,
1462 		       CTL_KERN, KERN_PIPE, KERN_PIPE_MAXKVASZ, CTL_EOL);
1463 	sysctl_createv(clog, 0, NULL, NULL,
1464 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1465 		       CTLTYPE_INT, "maxloankvasz",
1466 		       SYSCTL_DESCR("Limit for direct transfers via page loan"),
1467 		       NULL, 0, &limitpipekva, 0,
1468 		       CTL_KERN, KERN_PIPE, KERN_PIPE_LIMITKVA, CTL_EOL);
1469 	sysctl_createv(clog, 0, NULL, NULL,
1470 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1471 		       CTLTYPE_INT, "maxbigpipes",
1472 		       SYSCTL_DESCR("Maximum number of \"big\" pipes"),
1473 		       NULL, 0, &maxbigpipes, 0,
1474 		       CTL_KERN, KERN_PIPE, KERN_PIPE_MAXBIGPIPES, CTL_EOL);
1475 	sysctl_createv(clog, 0, NULL, NULL,
1476 		       CTLFLAG_PERMANENT,
1477 		       CTLTYPE_INT, "nbigpipes",
1478 		       SYSCTL_DESCR("Number of \"big\" pipes"),
1479 		       NULL, 0, &nbigpipe, 0,
1480 		       CTL_KERN, KERN_PIPE, KERN_PIPE_NBIGPIPES, CTL_EOL);
1481 	sysctl_createv(clog, 0, NULL, NULL,
1482 		       CTLFLAG_PERMANENT,
1483 		       CTLTYPE_INT, "kvasize",
1484 		       SYSCTL_DESCR("Amount of kernel memory consumed by pipe "
1485 				    "buffers"),
1486 		       NULL, 0, &amountpipekva, 0,
1487 		       CTL_KERN, KERN_PIPE, KERN_PIPE_KVASIZE, CTL_EOL);
1488 }
1489