xref: /freebsd/sys/kern/sys_pipe.c (revision 1603881667360c015f6685131f2f25474fa67a72)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1996 John S. Dyson
5  * Copyright (c) 2012 Giovanni Trematerra
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice immediately at the beginning of the file, without modification,
13  *    this list of conditions, and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Absolutely no warranty of function or purpose is made by the author
18  *    John S. Dyson.
19  * 4. Modifications may be freely made to this file if the above conditions
20  *    are met.
21  */
22 
23 /*
24  * This file contains a high-performance replacement for the socket-based
25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
26  * all features of sockets, but does do everything that pipes normally
27  * do.
28  */
29 
30 /*
31  * This code has two modes of operation, a small write mode and a large
32  * write mode.  The small write mode acts like conventional pipes with
33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
36  * memory, and the receiving process copies directly from these pinned pages
37  * in the sending process.
38  *
39  * If the sending process receives a signal, it is possible that it will
40  * go away, and certainly its address space can change, because control
41  * is returned back to the user-mode side.  In that case, the pipe code
42  * arranges to copy the buffer supplied by the user process, to a pageable
43  * kernel buffer, and the receiving process will grab the data from the
44  * pageable kernel buffer.  Since signals don't happen all that often,
45  * the copy operation is normally eliminated.
46  *
47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
48  * happen for small transfers so that the system will not spend all of
49  * its time context switching.
50  *
51  * In order to limit the resource use of pipes, two sysctls exist:
52  *
53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
54  * address space available to us in pipe_map. This value is normally
55  * autotuned, but may also be loader tuned.
56  *
57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
58  * memory in use by pipes.
59  *
60  * Based on how large pipekva is relative to maxpipekva, the following
61  * will happen:
62  *
63  * 0% - 50%:
64  *     New pipes are given 16K of memory backing, pipes may dynamically
65  *     grow to as large as 64K where needed.
66  * 50% - 75%:
67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
68  *     existing pipes may NOT grow.
69  * 75% - 100%:
70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
71  *     existing pipes will be shrunk down to 4K whenever possible.
72  *
73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
75  * resize which MUST occur for reverse-direction pipes when they are
76  * first used.
77  *
78  * Additional information about the current state of pipes may be obtained
79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
80  * and kern.ipc.piperesizefail.
81  *
82  * Locking rules:  There are two locks present here:  A mutex, used via
83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
84  * the flag, as mutexes can not persist over uiomove.  The mutex
85  * exists only to guard access to the flag, and is not in itself a
86  * locking mechanism.  Also note that there is only a single mutex for
87  * both directions of a pipe.
88  *
89  * As pipelock() may have to sleep before it can acquire the flag, it
90  * is important to reread all data after a call to pipelock(); everything
91  * in the structure may have changed.
92  */
93 
94 #include <sys/cdefs.h>
95 __FBSDID("$FreeBSD$");
96 
97 #include <sys/param.h>
98 #include <sys/systm.h>
99 #include <sys/conf.h>
100 #include <sys/fcntl.h>
101 #include <sys/file.h>
102 #include <sys/filedesc.h>
103 #include <sys/filio.h>
104 #include <sys/kernel.h>
105 #include <sys/lock.h>
106 #include <sys/mutex.h>
107 #include <sys/ttycom.h>
108 #include <sys/stat.h>
109 #include <sys/malloc.h>
110 #include <sys/poll.h>
111 #include <sys/selinfo.h>
112 #include <sys/signalvar.h>
113 #include <sys/syscallsubr.h>
114 #include <sys/sysctl.h>
115 #include <sys/sysproto.h>
116 #include <sys/pipe.h>
117 #include <sys/proc.h>
118 #include <sys/vnode.h>
119 #include <sys/uio.h>
120 #include <sys/user.h>
121 #include <sys/event.h>
122 
123 #include <security/mac/mac_framework.h>
124 
125 #include <vm/vm.h>
126 #include <vm/vm_param.h>
127 #include <vm/vm_object.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_extern.h>
130 #include <vm/pmap.h>
131 #include <vm/vm_map.h>
132 #include <vm/vm_page.h>
133 #include <vm/uma.h>
134 
135 /*
136  * Use this define if you want to disable *fancy* VM things.  Expect an
137  * approx 30% decrease in transfer rate.  This could be useful for
138  * NetBSD or OpenBSD.
139  */
140 /* #define PIPE_NODIRECT */
141 
142 #define PIPE_PEER(pipe)	\
143 	(((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
144 
145 /*
146  * interfaces to the outside world
147  */
148 static fo_rdwr_t	pipe_read;
149 static fo_rdwr_t	pipe_write;
150 static fo_truncate_t	pipe_truncate;
151 static fo_ioctl_t	pipe_ioctl;
152 static fo_poll_t	pipe_poll;
153 static fo_kqfilter_t	pipe_kqfilter;
154 static fo_stat_t	pipe_stat;
155 static fo_close_t	pipe_close;
156 static fo_chmod_t	pipe_chmod;
157 static fo_chown_t	pipe_chown;
158 static fo_fill_kinfo_t	pipe_fill_kinfo;
159 
160 struct fileops pipeops = {
161 	.fo_read = pipe_read,
162 	.fo_write = pipe_write,
163 	.fo_truncate = pipe_truncate,
164 	.fo_ioctl = pipe_ioctl,
165 	.fo_poll = pipe_poll,
166 	.fo_kqfilter = pipe_kqfilter,
167 	.fo_stat = pipe_stat,
168 	.fo_close = pipe_close,
169 	.fo_chmod = pipe_chmod,
170 	.fo_chown = pipe_chown,
171 	.fo_sendfile = invfo_sendfile,
172 	.fo_fill_kinfo = pipe_fill_kinfo,
173 	.fo_flags = DFLAG_PASSABLE
174 };
175 
176 static void	filt_pipedetach(struct knote *kn);
177 static void	filt_pipedetach_notsup(struct knote *kn);
178 static int	filt_pipenotsup(struct knote *kn, long hint);
179 static int	filt_piperead(struct knote *kn, long hint);
180 static int	filt_pipewrite(struct knote *kn, long hint);
181 
182 static struct filterops pipe_nfiltops = {
183 	.f_isfd = 1,
184 	.f_detach = filt_pipedetach_notsup,
185 	.f_event = filt_pipenotsup
186 };
187 static struct filterops pipe_rfiltops = {
188 	.f_isfd = 1,
189 	.f_detach = filt_pipedetach,
190 	.f_event = filt_piperead
191 };
192 static struct filterops pipe_wfiltops = {
193 	.f_isfd = 1,
194 	.f_detach = filt_pipedetach,
195 	.f_event = filt_pipewrite
196 };
197 
198 /*
199  * Default pipe buffer size(s), this can be kind-of large now because pipe
200  * space is pageable.  The pipe code will try to maintain locality of
201  * reference for performance reasons, so small amounts of outstanding I/O
202  * will not wipe the cache.
203  */
204 #define MINPIPESIZE (PIPE_SIZE/3)
205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
206 
207 static long amountpipekva;
208 static int pipefragretry;
209 static int pipeallocfail;
210 static int piperesizefail;
211 static int piperesizeallowed = 1;
212 static long pipe_mindirect = PIPE_MINDIRECT;
213 
214 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
215 	   &maxpipekva, 0, "Pipe KVA limit");
216 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
217 	   &amountpipekva, 0, "Pipe KVA usage");
218 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
219 	  &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
220 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
221 	  &pipeallocfail, 0, "Pipe allocation failures");
222 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
223 	  &piperesizefail, 0, "Pipe resize failures");
224 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
225 	  &piperesizeallowed, 0, "Pipe resizing allowed");
226 
227 static void pipeinit(void *dummy __unused);
228 static void pipeclose(struct pipe *cpipe);
229 static void pipe_free_kmem(struct pipe *cpipe);
230 static int pipe_create(struct pipe *pipe, bool backing);
231 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
232 static __inline int pipelock(struct pipe *cpipe, int catch);
233 static __inline void pipeunlock(struct pipe *cpipe);
234 static void pipe_timestamp(struct timespec *tsp);
235 #ifndef PIPE_NODIRECT
236 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
237 static void pipe_destroy_write_buffer(struct pipe *wpipe);
238 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
239 static void pipe_clone_write_buffer(struct pipe *wpipe);
240 #endif
241 static int pipespace(struct pipe *cpipe, int size);
242 static int pipespace_new(struct pipe *cpipe, int size);
243 
244 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
245 static int	pipe_zone_init(void *mem, int size, int flags);
246 static void	pipe_zone_fini(void *mem, int size);
247 
248 static uma_zone_t pipe_zone;
249 static struct unrhdr64 pipeino_unr;
250 static dev_t pipedev_ino;
251 
252 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
253 
254 static void
255 pipeinit(void *dummy __unused)
256 {
257 
258 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
259 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
260 	    UMA_ALIGN_PTR, 0);
261 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
262 	new_unrhdr64(&pipeino_unr, 1);
263 	pipedev_ino = devfs_alloc_cdp_inode();
264 	KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
265 }
266 
267 static int
268 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS)
269 {
270 	int error = 0;
271 	long tmp_pipe_mindirect = pipe_mindirect;
272 
273 	error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req);
274 	if (error != 0 || req->newptr == NULL)
275 		return (error);
276 
277 	/*
278 	 * Don't allow pipe_mindirect to be set so low that we violate
279 	 * atomicity requirements.
280 	 */
281 	if (tmp_pipe_mindirect <= PIPE_BUF)
282 		return (EINVAL);
283 	pipe_mindirect = tmp_pipe_mindirect;
284 	return (0);
285 }
286 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW,
287     &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L",
288     "Minimum write size triggering VM optimization");
289 
290 static int
291 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
292 {
293 	struct pipepair *pp;
294 	struct pipe *rpipe, *wpipe;
295 
296 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
297 
298 	pp = (struct pipepair *)mem;
299 
300 	/*
301 	 * We zero both pipe endpoints to make sure all the kmem pointers
302 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
303 	 * endpoints with the same time.
304 	 */
305 	rpipe = &pp->pp_rpipe;
306 	bzero(rpipe, sizeof(*rpipe));
307 	pipe_timestamp(&rpipe->pipe_ctime);
308 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
309 
310 	wpipe = &pp->pp_wpipe;
311 	bzero(wpipe, sizeof(*wpipe));
312 	wpipe->pipe_ctime = rpipe->pipe_ctime;
313 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
314 
315 	rpipe->pipe_peer = wpipe;
316 	rpipe->pipe_pair = pp;
317 	wpipe->pipe_peer = rpipe;
318 	wpipe->pipe_pair = pp;
319 
320 	/*
321 	 * Mark both endpoints as present; they will later get free'd
322 	 * one at a time.  When both are free'd, then the whole pair
323 	 * is released.
324 	 */
325 	rpipe->pipe_present = PIPE_ACTIVE;
326 	wpipe->pipe_present = PIPE_ACTIVE;
327 
328 	/*
329 	 * Eventually, the MAC Framework may initialize the label
330 	 * in ctor or init, but for now we do it elswhere to avoid
331 	 * blocking in ctor or init.
332 	 */
333 	pp->pp_label = NULL;
334 
335 	return (0);
336 }
337 
338 static int
339 pipe_zone_init(void *mem, int size, int flags)
340 {
341 	struct pipepair *pp;
342 
343 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
344 
345 	pp = (struct pipepair *)mem;
346 
347 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
348 	return (0);
349 }
350 
351 static void
352 pipe_zone_fini(void *mem, int size)
353 {
354 	struct pipepair *pp;
355 
356 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
357 
358 	pp = (struct pipepair *)mem;
359 
360 	mtx_destroy(&pp->pp_mtx);
361 }
362 
363 static int
364 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
365 {
366 	struct pipepair *pp;
367 	struct pipe *rpipe, *wpipe;
368 	int error;
369 
370 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
371 #ifdef MAC
372 	/*
373 	 * The MAC label is shared between the connected endpoints.  As a
374 	 * result mac_pipe_init() and mac_pipe_create() are called once
375 	 * for the pair, and not on the endpoints.
376 	 */
377 	mac_pipe_init(pp);
378 	mac_pipe_create(td->td_ucred, pp);
379 #endif
380 	rpipe = &pp->pp_rpipe;
381 	wpipe = &pp->pp_wpipe;
382 
383 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
384 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
385 
386 	/*
387 	 * Only the forward direction pipe is backed by big buffer by
388 	 * default.
389 	 */
390 	error = pipe_create(rpipe, true);
391 	if (error != 0)
392 		goto fail;
393 	error = pipe_create(wpipe, false);
394 	if (error != 0) {
395 		/*
396 		 * This cleanup leaves the pipe inode number for rpipe
397 		 * still allocated, but never used.  We do not free
398 		 * inode numbers for opened pipes, which is required
399 		 * for correctness because numbers must be unique.
400 		 * But also it avoids any memory use by the unr
401 		 * allocator, so stashing away the transient inode
402 		 * number is reasonable.
403 		 */
404 		pipe_free_kmem(rpipe);
405 		goto fail;
406 	}
407 
408 	rpipe->pipe_state |= PIPE_DIRECTOK;
409 	wpipe->pipe_state |= PIPE_DIRECTOK;
410 	return (0);
411 
412 fail:
413 	knlist_destroy(&rpipe->pipe_sel.si_note);
414 	knlist_destroy(&wpipe->pipe_sel.si_note);
415 #ifdef MAC
416 	mac_pipe_destroy(pp);
417 #endif
418 	return (error);
419 }
420 
421 int
422 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
423 {
424 	struct pipepair *pp;
425 	int error;
426 
427 	error = pipe_paircreate(td, &pp);
428 	if (error != 0)
429 		return (error);
430 	pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
431 	*ppipe = &pp->pp_rpipe;
432 	return (0);
433 }
434 
435 void
436 pipe_dtor(struct pipe *dpipe)
437 {
438 	struct pipe *peer;
439 
440 	peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
441 	funsetown(&dpipe->pipe_sigio);
442 	pipeclose(dpipe);
443 	if (peer != NULL) {
444 		funsetown(&peer->pipe_sigio);
445 		pipeclose(peer);
446 	}
447 }
448 
449 /*
450  * Get a timestamp.
451  *
452  * This used to be vfs_timestamp but the higher precision is unnecessary and
453  * can very negatively affect performance in virtualized environments (e.g., on
454  * vms running on amd64 when using the rdtscp instruction).
455  */
456 static void
457 pipe_timestamp(struct timespec *tsp)
458 {
459 
460 	getnanotime(tsp);
461 }
462 
463 /*
464  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
465  * the zone pick up the pieces via pipeclose().
466  */
467 int
468 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
469     struct filecaps *fcaps2)
470 {
471 	struct file *rf, *wf;
472 	struct pipe *rpipe, *wpipe;
473 	struct pipepair *pp;
474 	int fd, fflags, error;
475 
476 	error = pipe_paircreate(td, &pp);
477 	if (error != 0)
478 		return (error);
479 	rpipe = &pp->pp_rpipe;
480 	wpipe = &pp->pp_wpipe;
481 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
482 	if (error) {
483 		pipeclose(rpipe);
484 		pipeclose(wpipe);
485 		return (error);
486 	}
487 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
488 	fildes[0] = fd;
489 
490 	fflags = FREAD | FWRITE;
491 	if ((flags & O_NONBLOCK) != 0)
492 		fflags |= FNONBLOCK;
493 
494 	/*
495 	 * Warning: once we've gotten past allocation of the fd for the
496 	 * read-side, we can only drop the read side via fdrop() in order
497 	 * to avoid races against processes which manage to dup() the read
498 	 * side while we are blocked trying to allocate the write side.
499 	 */
500 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
501 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
502 	if (error) {
503 		fdclose(td, rf, fildes[0]);
504 		fdrop(rf, td);
505 		/* rpipe has been closed by fdrop(). */
506 		pipeclose(wpipe);
507 		return (error);
508 	}
509 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
510 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
511 	fdrop(wf, td);
512 	fildes[1] = fd;
513 	fdrop(rf, td);
514 
515 	return (0);
516 }
517 
518 #ifdef COMPAT_FREEBSD10
519 /* ARGSUSED */
520 int
521 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
522 {
523 	int error;
524 	int fildes[2];
525 
526 	error = kern_pipe(td, fildes, 0, NULL, NULL);
527 	if (error)
528 		return (error);
529 
530 	td->td_retval[0] = fildes[0];
531 	td->td_retval[1] = fildes[1];
532 
533 	return (0);
534 }
535 #endif
536 
537 int
538 sys_pipe2(struct thread *td, struct pipe2_args *uap)
539 {
540 	int error, fildes[2];
541 
542 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
543 		return (EINVAL);
544 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
545 	if (error)
546 		return (error);
547 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
548 	if (error) {
549 		(void)kern_close(td, fildes[0]);
550 		(void)kern_close(td, fildes[1]);
551 	}
552 	return (error);
553 }
554 
555 /*
556  * Allocate kva for pipe circular buffer, the space is pageable
557  * This routine will 'realloc' the size of a pipe safely, if it fails
558  * it will retain the old buffer.
559  * If it fails it will return ENOMEM.
560  */
561 static int
562 pipespace_new(struct pipe *cpipe, int size)
563 {
564 	caddr_t buffer;
565 	int error, cnt, firstseg;
566 	static int curfail = 0;
567 	static struct timeval lastfail;
568 
569 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
570 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
571 		("pipespace: resize of direct writes not allowed"));
572 retry:
573 	cnt = cpipe->pipe_buffer.cnt;
574 	if (cnt > size)
575 		size = cnt;
576 
577 	size = round_page(size);
578 	buffer = (caddr_t) vm_map_min(pipe_map);
579 
580 	error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
581 	    VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
582 	if (error != KERN_SUCCESS) {
583 		if (cpipe->pipe_buffer.buffer == NULL &&
584 		    size > SMALL_PIPE_SIZE) {
585 			size = SMALL_PIPE_SIZE;
586 			pipefragretry++;
587 			goto retry;
588 		}
589 		if (cpipe->pipe_buffer.buffer == NULL) {
590 			pipeallocfail++;
591 			if (ppsratecheck(&lastfail, &curfail, 1))
592 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
593 		} else {
594 			piperesizefail++;
595 		}
596 		return (ENOMEM);
597 	}
598 
599 	/* copy data, then free old resources if we're resizing */
600 	if (cnt > 0) {
601 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
602 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
603 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
604 				buffer, firstseg);
605 			if ((cnt - firstseg) > 0)
606 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
607 					cpipe->pipe_buffer.in);
608 		} else {
609 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
610 				buffer, cnt);
611 		}
612 	}
613 	pipe_free_kmem(cpipe);
614 	cpipe->pipe_buffer.buffer = buffer;
615 	cpipe->pipe_buffer.size = size;
616 	cpipe->pipe_buffer.in = cnt;
617 	cpipe->pipe_buffer.out = 0;
618 	cpipe->pipe_buffer.cnt = cnt;
619 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
620 	return (0);
621 }
622 
623 /*
624  * Wrapper for pipespace_new() that performs locking assertions.
625  */
626 static int
627 pipespace(struct pipe *cpipe, int size)
628 {
629 
630 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
631 	    ("Unlocked pipe passed to pipespace"));
632 	return (pipespace_new(cpipe, size));
633 }
634 
635 /*
636  * lock a pipe for I/O, blocking other access
637  */
638 static __inline int
639 pipelock(struct pipe *cpipe, int catch)
640 {
641 	int error, prio;
642 
643 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
644 
645 	prio = PRIBIO;
646 	if (catch)
647 		prio |= PCATCH;
648 	while (cpipe->pipe_state & PIPE_LOCKFL) {
649 		KASSERT(cpipe->pipe_waiters >= 0,
650 		    ("%s: bad waiter count %d", __func__,
651 		    cpipe->pipe_waiters));
652 		cpipe->pipe_waiters++;
653 		error = msleep(cpipe, PIPE_MTX(cpipe),
654 		    prio, "pipelk", 0);
655 		cpipe->pipe_waiters--;
656 		if (error != 0)
657 			return (error);
658 	}
659 	cpipe->pipe_state |= PIPE_LOCKFL;
660 	return (0);
661 }
662 
663 /*
664  * unlock a pipe I/O lock
665  */
666 static __inline void
667 pipeunlock(struct pipe *cpipe)
668 {
669 
670 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
671 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
672 		("Unlocked pipe passed to pipeunlock"));
673 	KASSERT(cpipe->pipe_waiters >= 0,
674 	    ("%s: bad waiter count %d", __func__,
675 	    cpipe->pipe_waiters));
676 	cpipe->pipe_state &= ~PIPE_LOCKFL;
677 	if (cpipe->pipe_waiters > 0) {
678 		wakeup_one(cpipe);
679 	}
680 }
681 
682 void
683 pipeselwakeup(struct pipe *cpipe)
684 {
685 
686 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
687 	if (cpipe->pipe_state & PIPE_SEL) {
688 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
689 		if (!SEL_WAITING(&cpipe->pipe_sel))
690 			cpipe->pipe_state &= ~PIPE_SEL;
691 	}
692 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
693 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
694 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
695 }
696 
697 /*
698  * Initialize and allocate VM and memory for pipe.  The structure
699  * will start out zero'd from the ctor, so we just manage the kmem.
700  */
701 static int
702 pipe_create(struct pipe *pipe, bool large_backing)
703 {
704 	int error;
705 
706 	error = pipespace_new(pipe, !large_backing || amountpipekva >
707 	    maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
708 	if (error == 0)
709 		pipe->pipe_ino = alloc_unr64(&pipeino_unr);
710 	return (error);
711 }
712 
713 /* ARGSUSED */
714 static int
715 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
716     int flags, struct thread *td)
717 {
718 	struct pipe *rpipe;
719 	int error;
720 	int nread = 0;
721 	int size;
722 
723 	rpipe = fp->f_data;
724 	PIPE_LOCK(rpipe);
725 	++rpipe->pipe_busy;
726 	error = pipelock(rpipe, 1);
727 	if (error)
728 		goto unlocked_error;
729 
730 #ifdef MAC
731 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
732 	if (error)
733 		goto locked_error;
734 #endif
735 	if (amountpipekva > (3 * maxpipekva) / 4) {
736 		if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
737 		    rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
738 		    rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
739 		    piperesizeallowed == 1) {
740 			PIPE_UNLOCK(rpipe);
741 			pipespace(rpipe, SMALL_PIPE_SIZE);
742 			PIPE_LOCK(rpipe);
743 		}
744 	}
745 
746 	while (uio->uio_resid) {
747 		/*
748 		 * normal pipe buffer receive
749 		 */
750 		if (rpipe->pipe_buffer.cnt > 0) {
751 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
752 			if (size > rpipe->pipe_buffer.cnt)
753 				size = rpipe->pipe_buffer.cnt;
754 			if (size > uio->uio_resid)
755 				size = uio->uio_resid;
756 
757 			PIPE_UNLOCK(rpipe);
758 			error = uiomove(
759 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
760 			    size, uio);
761 			PIPE_LOCK(rpipe);
762 			if (error)
763 				break;
764 
765 			rpipe->pipe_buffer.out += size;
766 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
767 				rpipe->pipe_buffer.out = 0;
768 
769 			rpipe->pipe_buffer.cnt -= size;
770 
771 			/*
772 			 * If there is no more to read in the pipe, reset
773 			 * its pointers to the beginning.  This improves
774 			 * cache hit stats.
775 			 */
776 			if (rpipe->pipe_buffer.cnt == 0) {
777 				rpipe->pipe_buffer.in = 0;
778 				rpipe->pipe_buffer.out = 0;
779 			}
780 			nread += size;
781 #ifndef PIPE_NODIRECT
782 		/*
783 		 * Direct copy, bypassing a kernel buffer.
784 		 */
785 		} else if ((size = rpipe->pipe_pages.cnt) != 0) {
786 			if (size > uio->uio_resid)
787 				size = (u_int) uio->uio_resid;
788 			PIPE_UNLOCK(rpipe);
789 			error = uiomove_fromphys(rpipe->pipe_pages.ms,
790 			    rpipe->pipe_pages.pos, size, uio);
791 			PIPE_LOCK(rpipe);
792 			if (error)
793 				break;
794 			nread += size;
795 			rpipe->pipe_pages.pos += size;
796 			rpipe->pipe_pages.cnt -= size;
797 			if (rpipe->pipe_pages.cnt == 0) {
798 				rpipe->pipe_state &= ~PIPE_WANTW;
799 				wakeup(rpipe);
800 			}
801 #endif
802 		} else {
803 			/*
804 			 * detect EOF condition
805 			 * read returns 0 on EOF, no need to set error
806 			 */
807 			if (rpipe->pipe_state & PIPE_EOF)
808 				break;
809 
810 			/*
811 			 * If the "write-side" has been blocked, wake it up now.
812 			 */
813 			if (rpipe->pipe_state & PIPE_WANTW) {
814 				rpipe->pipe_state &= ~PIPE_WANTW;
815 				wakeup(rpipe);
816 			}
817 
818 			/*
819 			 * Break if some data was read.
820 			 */
821 			if (nread > 0)
822 				break;
823 
824 			/*
825 			 * Unlock the pipe buffer for our remaining processing.
826 			 * We will either break out with an error or we will
827 			 * sleep and relock to loop.
828 			 */
829 			pipeunlock(rpipe);
830 
831 			/*
832 			 * Handle non-blocking mode operation or
833 			 * wait for more data.
834 			 */
835 			if (fp->f_flag & FNONBLOCK) {
836 				error = EAGAIN;
837 			} else {
838 				rpipe->pipe_state |= PIPE_WANTR;
839 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
840 				    PRIBIO | PCATCH,
841 				    "piperd", 0)) == 0)
842 					error = pipelock(rpipe, 1);
843 			}
844 			if (error)
845 				goto unlocked_error;
846 		}
847 	}
848 #ifdef MAC
849 locked_error:
850 #endif
851 	pipeunlock(rpipe);
852 
853 	/* XXX: should probably do this before getting any locks. */
854 	if (error == 0)
855 		pipe_timestamp(&rpipe->pipe_atime);
856 unlocked_error:
857 	--rpipe->pipe_busy;
858 
859 	/*
860 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
861 	 */
862 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
863 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
864 		wakeup(rpipe);
865 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
866 		/*
867 		 * Handle write blocking hysteresis.
868 		 */
869 		if (rpipe->pipe_state & PIPE_WANTW) {
870 			rpipe->pipe_state &= ~PIPE_WANTW;
871 			wakeup(rpipe);
872 		}
873 	}
874 
875 	/*
876 	 * Only wake up writers if there was actually something read.
877 	 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
878 	 */
879 	if (nread > 0 &&
880 	    rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
881 		pipeselwakeup(rpipe);
882 
883 	PIPE_UNLOCK(rpipe);
884 	if (nread > 0)
885 		td->td_ru.ru_msgrcv++;
886 	return (error);
887 }
888 
889 #ifndef PIPE_NODIRECT
890 /*
891  * Map the sending processes' buffer into kernel space and wire it.
892  * This is similar to a physical write operation.
893  */
894 static int
895 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
896 {
897 	u_int size;
898 	int i;
899 
900 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
901 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
902 	    ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
903 	KASSERT(wpipe->pipe_pages.cnt == 0,
904 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
905 
906 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
907                 size = wpipe->pipe_buffer.size;
908 	else
909                 size = uio->uio_iov->iov_len;
910 
911 	wpipe->pipe_state |= PIPE_DIRECTW;
912 	PIPE_UNLOCK(wpipe);
913 	i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
914 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
915 	    wpipe->pipe_pages.ms, PIPENPAGES);
916 	PIPE_LOCK(wpipe);
917 	if (i < 0) {
918 		wpipe->pipe_state &= ~PIPE_DIRECTW;
919 		return (EFAULT);
920 	}
921 
922 	wpipe->pipe_pages.npages = i;
923 	wpipe->pipe_pages.pos =
924 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
925 	wpipe->pipe_pages.cnt = size;
926 
927 	uio->uio_iov->iov_len -= size;
928 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
929 	if (uio->uio_iov->iov_len == 0)
930 		uio->uio_iov++;
931 	uio->uio_resid -= size;
932 	uio->uio_offset += size;
933 	return (0);
934 }
935 
936 /*
937  * Unwire the process buffer.
938  */
939 static void
940 pipe_destroy_write_buffer(struct pipe *wpipe)
941 {
942 
943 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
944 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
945 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
946 	KASSERT(wpipe->pipe_pages.cnt == 0,
947 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
948 
949 	wpipe->pipe_state &= ~PIPE_DIRECTW;
950 	vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
951 	wpipe->pipe_pages.npages = 0;
952 }
953 
954 /*
955  * In the case of a signal, the writing process might go away.  This
956  * code copies the data into the circular buffer so that the source
957  * pages can be freed without loss of data.
958  */
959 static void
960 pipe_clone_write_buffer(struct pipe *wpipe)
961 {
962 	struct uio uio;
963 	struct iovec iov;
964 	int size;
965 	int pos;
966 
967 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
968 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
969 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
970 
971 	size = wpipe->pipe_pages.cnt;
972 	pos = wpipe->pipe_pages.pos;
973 	wpipe->pipe_pages.cnt = 0;
974 
975 	wpipe->pipe_buffer.in = size;
976 	wpipe->pipe_buffer.out = 0;
977 	wpipe->pipe_buffer.cnt = size;
978 
979 	PIPE_UNLOCK(wpipe);
980 	iov.iov_base = wpipe->pipe_buffer.buffer;
981 	iov.iov_len = size;
982 	uio.uio_iov = &iov;
983 	uio.uio_iovcnt = 1;
984 	uio.uio_offset = 0;
985 	uio.uio_resid = size;
986 	uio.uio_segflg = UIO_SYSSPACE;
987 	uio.uio_rw = UIO_READ;
988 	uio.uio_td = curthread;
989 	uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
990 	PIPE_LOCK(wpipe);
991 	pipe_destroy_write_buffer(wpipe);
992 }
993 
994 /*
995  * This implements the pipe buffer write mechanism.  Note that only
996  * a direct write OR a normal pipe write can be pending at any given time.
997  * If there are any characters in the pipe buffer, the direct write will
998  * be deferred until the receiving process grabs all of the bytes from
999  * the pipe buffer.  Then the direct mapping write is set-up.
1000  */
1001 static int
1002 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
1003 {
1004 	int error;
1005 
1006 retry:
1007 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1008 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1009 		error = EPIPE;
1010 		goto error1;
1011 	}
1012 	if (wpipe->pipe_state & PIPE_DIRECTW) {
1013 		if (wpipe->pipe_state & PIPE_WANTR) {
1014 			wpipe->pipe_state &= ~PIPE_WANTR;
1015 			wakeup(wpipe);
1016 		}
1017 		pipeselwakeup(wpipe);
1018 		wpipe->pipe_state |= PIPE_WANTW;
1019 		pipeunlock(wpipe);
1020 		error = msleep(wpipe, PIPE_MTX(wpipe),
1021 		    PRIBIO | PCATCH, "pipdww", 0);
1022 		pipelock(wpipe, 0);
1023 		if (error != 0)
1024 			goto error1;
1025 		goto retry;
1026 	}
1027 	if (wpipe->pipe_buffer.cnt > 0) {
1028 		if (wpipe->pipe_state & PIPE_WANTR) {
1029 			wpipe->pipe_state &= ~PIPE_WANTR;
1030 			wakeup(wpipe);
1031 		}
1032 		pipeselwakeup(wpipe);
1033 		wpipe->pipe_state |= PIPE_WANTW;
1034 		pipeunlock(wpipe);
1035 		error = msleep(wpipe, PIPE_MTX(wpipe),
1036 		    PRIBIO | PCATCH, "pipdwc", 0);
1037 		pipelock(wpipe, 0);
1038 		if (error != 0)
1039 			goto error1;
1040 		goto retry;
1041 	}
1042 
1043 	error = pipe_build_write_buffer(wpipe, uio);
1044 	if (error) {
1045 		goto error1;
1046 	}
1047 
1048 	while (wpipe->pipe_pages.cnt != 0 &&
1049 	    (wpipe->pipe_state & PIPE_EOF) == 0) {
1050 		if (wpipe->pipe_state & PIPE_WANTR) {
1051 			wpipe->pipe_state &= ~PIPE_WANTR;
1052 			wakeup(wpipe);
1053 		}
1054 		pipeselwakeup(wpipe);
1055 		wpipe->pipe_state |= PIPE_WANTW;
1056 		pipeunlock(wpipe);
1057 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1058 		    "pipdwt", 0);
1059 		pipelock(wpipe, 0);
1060 		if (error != 0)
1061 			break;
1062 	}
1063 
1064 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1065 		wpipe->pipe_pages.cnt = 0;
1066 		pipe_destroy_write_buffer(wpipe);
1067 		pipeselwakeup(wpipe);
1068 		error = EPIPE;
1069 	} else if (error == EINTR || error == ERESTART) {
1070 		pipe_clone_write_buffer(wpipe);
1071 	} else {
1072 		pipe_destroy_write_buffer(wpipe);
1073 	}
1074 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1075 	    ("pipe %p leaked PIPE_DIRECTW", wpipe));
1076 	return (error);
1077 
1078 error1:
1079 	wakeup(wpipe);
1080 	return (error);
1081 }
1082 #endif
1083 
1084 static int
1085 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1086     int flags, struct thread *td)
1087 {
1088 	struct pipe *wpipe, *rpipe;
1089 	ssize_t orig_resid;
1090 	int desiredsize, error;
1091 
1092 	rpipe = fp->f_data;
1093 	wpipe = PIPE_PEER(rpipe);
1094 	PIPE_LOCK(rpipe);
1095 	error = pipelock(wpipe, 1);
1096 	if (error) {
1097 		PIPE_UNLOCK(rpipe);
1098 		return (error);
1099 	}
1100 	/*
1101 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1102 	 */
1103 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1104 	    (wpipe->pipe_state & PIPE_EOF)) {
1105 		pipeunlock(wpipe);
1106 		PIPE_UNLOCK(rpipe);
1107 		return (EPIPE);
1108 	}
1109 #ifdef MAC
1110 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1111 	if (error) {
1112 		pipeunlock(wpipe);
1113 		PIPE_UNLOCK(rpipe);
1114 		return (error);
1115 	}
1116 #endif
1117 	++wpipe->pipe_busy;
1118 
1119 	/* Choose a larger size if it's advantageous */
1120 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1121 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1122 		if (piperesizeallowed != 1)
1123 			break;
1124 		if (amountpipekva > maxpipekva / 2)
1125 			break;
1126 		if (desiredsize == BIG_PIPE_SIZE)
1127 			break;
1128 		desiredsize = desiredsize * 2;
1129 	}
1130 
1131 	/* Choose a smaller size if we're in a OOM situation */
1132 	if (amountpipekva > (3 * maxpipekva) / 4 &&
1133 	    wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1134 	    wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1135 	    piperesizeallowed == 1)
1136 		desiredsize = SMALL_PIPE_SIZE;
1137 
1138 	/* Resize if the above determined that a new size was necessary */
1139 	if (desiredsize != wpipe->pipe_buffer.size &&
1140 	    (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1141 		PIPE_UNLOCK(wpipe);
1142 		pipespace(wpipe, desiredsize);
1143 		PIPE_LOCK(wpipe);
1144 	}
1145 	MPASS(wpipe->pipe_buffer.size != 0);
1146 
1147 	orig_resid = uio->uio_resid;
1148 
1149 	while (uio->uio_resid) {
1150 		int space;
1151 
1152 		if (wpipe->pipe_state & PIPE_EOF) {
1153 			error = EPIPE;
1154 			break;
1155 		}
1156 #ifndef PIPE_NODIRECT
1157 		/*
1158 		 * If the transfer is large, we can gain performance if
1159 		 * we do process-to-process copies directly.
1160 		 * If the write is non-blocking, we don't use the
1161 		 * direct write mechanism.
1162 		 *
1163 		 * The direct write mechanism will detect the reader going
1164 		 * away on us.
1165 		 */
1166 		if (uio->uio_segflg == UIO_USERSPACE &&
1167 		    uio->uio_iov->iov_len >= pipe_mindirect &&
1168 		    wpipe->pipe_buffer.size >= pipe_mindirect &&
1169 		    (fp->f_flag & FNONBLOCK) == 0) {
1170 			error = pipe_direct_write(wpipe, uio);
1171 			if (error != 0)
1172 				break;
1173 			continue;
1174 		}
1175 #endif
1176 
1177 		/*
1178 		 * Pipe buffered writes cannot be coincidental with
1179 		 * direct writes.  We wait until the currently executing
1180 		 * direct write is completed before we start filling the
1181 		 * pipe buffer.  We break out if a signal occurs or the
1182 		 * reader goes away.
1183 		 */
1184 		if (wpipe->pipe_pages.cnt != 0) {
1185 			if (wpipe->pipe_state & PIPE_WANTR) {
1186 				wpipe->pipe_state &= ~PIPE_WANTR;
1187 				wakeup(wpipe);
1188 			}
1189 			pipeselwakeup(wpipe);
1190 			wpipe->pipe_state |= PIPE_WANTW;
1191 			pipeunlock(wpipe);
1192 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1193 			    "pipbww", 0);
1194 			pipelock(wpipe, 0);
1195 			if (error != 0)
1196 				break;
1197 			continue;
1198 		}
1199 
1200 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1201 
1202 		/* Writes of size <= PIPE_BUF must be atomic. */
1203 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1204 			space = 0;
1205 
1206 		if (space > 0) {
1207 			int size;	/* Transfer size */
1208 			int segsize;	/* first segment to transfer */
1209 
1210 			/*
1211 			 * Transfer size is minimum of uio transfer
1212 			 * and free space in pipe buffer.
1213 			 */
1214 			if (space > uio->uio_resid)
1215 				size = uio->uio_resid;
1216 			else
1217 				size = space;
1218 			/*
1219 			 * First segment to transfer is minimum of
1220 			 * transfer size and contiguous space in
1221 			 * pipe buffer.  If first segment to transfer
1222 			 * is less than the transfer size, we've got
1223 			 * a wraparound in the buffer.
1224 			 */
1225 			segsize = wpipe->pipe_buffer.size -
1226 				wpipe->pipe_buffer.in;
1227 			if (segsize > size)
1228 				segsize = size;
1229 
1230 			/* Transfer first segment */
1231 
1232 			PIPE_UNLOCK(rpipe);
1233 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1234 					segsize, uio);
1235 			PIPE_LOCK(rpipe);
1236 
1237 			if (error == 0 && segsize < size) {
1238 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1239 					wpipe->pipe_buffer.size,
1240 					("Pipe buffer wraparound disappeared"));
1241 				/*
1242 				 * Transfer remaining part now, to
1243 				 * support atomic writes.  Wraparound
1244 				 * happened.
1245 				 */
1246 
1247 				PIPE_UNLOCK(rpipe);
1248 				error = uiomove(
1249 				    &wpipe->pipe_buffer.buffer[0],
1250 				    size - segsize, uio);
1251 				PIPE_LOCK(rpipe);
1252 			}
1253 			if (error == 0) {
1254 				wpipe->pipe_buffer.in += size;
1255 				if (wpipe->pipe_buffer.in >=
1256 				    wpipe->pipe_buffer.size) {
1257 					KASSERT(wpipe->pipe_buffer.in ==
1258 						size - segsize +
1259 						wpipe->pipe_buffer.size,
1260 						("Expected wraparound bad"));
1261 					wpipe->pipe_buffer.in = size - segsize;
1262 				}
1263 
1264 				wpipe->pipe_buffer.cnt += size;
1265 				KASSERT(wpipe->pipe_buffer.cnt <=
1266 					wpipe->pipe_buffer.size,
1267 					("Pipe buffer overflow"));
1268 			}
1269 			if (error != 0)
1270 				break;
1271 			continue;
1272 		} else {
1273 			/*
1274 			 * If the "read-side" has been blocked, wake it up now.
1275 			 */
1276 			if (wpipe->pipe_state & PIPE_WANTR) {
1277 				wpipe->pipe_state &= ~PIPE_WANTR;
1278 				wakeup(wpipe);
1279 			}
1280 
1281 			/*
1282 			 * don't block on non-blocking I/O
1283 			 */
1284 			if (fp->f_flag & FNONBLOCK) {
1285 				error = EAGAIN;
1286 				break;
1287 			}
1288 
1289 			/*
1290 			 * We have no more space and have something to offer,
1291 			 * wake up select/poll.
1292 			 */
1293 			pipeselwakeup(wpipe);
1294 
1295 			wpipe->pipe_state |= PIPE_WANTW;
1296 			pipeunlock(wpipe);
1297 			error = msleep(wpipe, PIPE_MTX(rpipe),
1298 			    PRIBIO | PCATCH, "pipewr", 0);
1299 			pipelock(wpipe, 0);
1300 			if (error != 0)
1301 				break;
1302 			continue;
1303 		}
1304 	}
1305 
1306 	--wpipe->pipe_busy;
1307 
1308 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1309 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1310 		wakeup(wpipe);
1311 	} else if (wpipe->pipe_buffer.cnt > 0) {
1312 		/*
1313 		 * If we have put any characters in the buffer, we wake up
1314 		 * the reader.
1315 		 */
1316 		if (wpipe->pipe_state & PIPE_WANTR) {
1317 			wpipe->pipe_state &= ~PIPE_WANTR;
1318 			wakeup(wpipe);
1319 		}
1320 	}
1321 
1322 	/*
1323 	 * Don't return EPIPE if any byte was written.
1324 	 * EINTR and other interrupts are handled by generic I/O layer.
1325 	 * Do not pretend that I/O succeeded for obvious user error
1326 	 * like EFAULT.
1327 	 */
1328 	if (uio->uio_resid != orig_resid && error == EPIPE)
1329 		error = 0;
1330 
1331 	if (error == 0)
1332 		pipe_timestamp(&wpipe->pipe_mtime);
1333 
1334 	/*
1335 	 * We have something to offer,
1336 	 * wake up select/poll.
1337 	 */
1338 	if (wpipe->pipe_buffer.cnt)
1339 		pipeselwakeup(wpipe);
1340 
1341 	pipeunlock(wpipe);
1342 	PIPE_UNLOCK(rpipe);
1343 	if (uio->uio_resid != orig_resid)
1344 		td->td_ru.ru_msgsnd++;
1345 	return (error);
1346 }
1347 
1348 /* ARGSUSED */
1349 static int
1350 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1351     struct thread *td)
1352 {
1353 	struct pipe *cpipe;
1354 	int error;
1355 
1356 	cpipe = fp->f_data;
1357 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1358 		error = vnops.fo_truncate(fp, length, active_cred, td);
1359 	else
1360 		error = invfo_truncate(fp, length, active_cred, td);
1361 	return (error);
1362 }
1363 
1364 /*
1365  * we implement a very minimal set of ioctls for compatibility with sockets.
1366  */
1367 static int
1368 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1369     struct thread *td)
1370 {
1371 	struct pipe *mpipe = fp->f_data;
1372 	int error;
1373 
1374 	PIPE_LOCK(mpipe);
1375 
1376 #ifdef MAC
1377 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1378 	if (error) {
1379 		PIPE_UNLOCK(mpipe);
1380 		return (error);
1381 	}
1382 #endif
1383 
1384 	error = 0;
1385 	switch (cmd) {
1386 	case FIONBIO:
1387 		break;
1388 
1389 	case FIOASYNC:
1390 		if (*(int *)data) {
1391 			mpipe->pipe_state |= PIPE_ASYNC;
1392 		} else {
1393 			mpipe->pipe_state &= ~PIPE_ASYNC;
1394 		}
1395 		break;
1396 
1397 	case FIONREAD:
1398 		if (!(fp->f_flag & FREAD)) {
1399 			*(int *)data = 0;
1400 			PIPE_UNLOCK(mpipe);
1401 			return (0);
1402 		}
1403 		if (mpipe->pipe_pages.cnt != 0)
1404 			*(int *)data = mpipe->pipe_pages.cnt;
1405 		else
1406 			*(int *)data = mpipe->pipe_buffer.cnt;
1407 		break;
1408 
1409 	case FIOSETOWN:
1410 		PIPE_UNLOCK(mpipe);
1411 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1412 		goto out_unlocked;
1413 
1414 	case FIOGETOWN:
1415 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1416 		break;
1417 
1418 	/* This is deprecated, FIOSETOWN should be used instead. */
1419 	case TIOCSPGRP:
1420 		PIPE_UNLOCK(mpipe);
1421 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1422 		goto out_unlocked;
1423 
1424 	/* This is deprecated, FIOGETOWN should be used instead. */
1425 	case TIOCGPGRP:
1426 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1427 		break;
1428 
1429 	default:
1430 		error = ENOTTY;
1431 		break;
1432 	}
1433 	PIPE_UNLOCK(mpipe);
1434 out_unlocked:
1435 	return (error);
1436 }
1437 
1438 static int
1439 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1440     struct thread *td)
1441 {
1442 	struct pipe *rpipe;
1443 	struct pipe *wpipe;
1444 	int levents, revents;
1445 #ifdef MAC
1446 	int error;
1447 #endif
1448 
1449 	revents = 0;
1450 	rpipe = fp->f_data;
1451 	wpipe = PIPE_PEER(rpipe);
1452 	PIPE_LOCK(rpipe);
1453 #ifdef MAC
1454 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1455 	if (error)
1456 		goto locked_error;
1457 #endif
1458 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1459 		if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1460 			revents |= events & (POLLIN | POLLRDNORM);
1461 
1462 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1463 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1464 		    (wpipe->pipe_state & PIPE_EOF) ||
1465 		    ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1466 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1467 			 wpipe->pipe_buffer.size == 0)))
1468 			revents |= events & (POLLOUT | POLLWRNORM);
1469 
1470 	levents = events &
1471 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1472 	if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1473 	    fp->f_pipegen == rpipe->pipe_wgen)
1474 		events |= POLLINIGNEOF;
1475 
1476 	if ((events & POLLINIGNEOF) == 0) {
1477 		if (rpipe->pipe_state & PIPE_EOF) {
1478 			if (fp->f_flag & FREAD)
1479 				revents |= (events & (POLLIN | POLLRDNORM));
1480 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1481 			    (wpipe->pipe_state & PIPE_EOF))
1482 				revents |= POLLHUP;
1483 		}
1484 	}
1485 
1486 	if (revents == 0) {
1487 		/*
1488 		 * Add ourselves regardless of eventmask as we have to return
1489 		 * POLLHUP even if it was not asked for.
1490 		 */
1491 		if ((fp->f_flag & FREAD) != 0) {
1492 			selrecord(td, &rpipe->pipe_sel);
1493 			if (SEL_WAITING(&rpipe->pipe_sel))
1494 				rpipe->pipe_state |= PIPE_SEL;
1495 		}
1496 
1497 		if ((fp->f_flag & FWRITE) != 0 &&
1498 		    wpipe->pipe_present == PIPE_ACTIVE) {
1499 			selrecord(td, &wpipe->pipe_sel);
1500 			if (SEL_WAITING(&wpipe->pipe_sel))
1501 				wpipe->pipe_state |= PIPE_SEL;
1502 		}
1503 	}
1504 #ifdef MAC
1505 locked_error:
1506 #endif
1507 	PIPE_UNLOCK(rpipe);
1508 
1509 	return (revents);
1510 }
1511 
1512 /*
1513  * We shouldn't need locks here as we're doing a read and this should
1514  * be a natural race.
1515  */
1516 static int
1517 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1518     struct thread *td)
1519 {
1520 	struct pipe *pipe;
1521 #ifdef MAC
1522 	int error;
1523 #endif
1524 
1525 	pipe = fp->f_data;
1526 #ifdef MAC
1527 	if (mac_pipe_check_stat_enabled()) {
1528 		PIPE_LOCK(pipe);
1529 		error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1530 		PIPE_UNLOCK(pipe);
1531 		if (error) {
1532 			return (error);
1533 		}
1534 	}
1535 #endif
1536 
1537 	/* For named pipes ask the underlying filesystem. */
1538 	if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1539 		return (vnops.fo_stat(fp, ub, active_cred, td));
1540 	}
1541 
1542 	bzero(ub, sizeof(*ub));
1543 	ub->st_mode = S_IFIFO;
1544 	ub->st_blksize = PAGE_SIZE;
1545 	if (pipe->pipe_pages.cnt != 0)
1546 		ub->st_size = pipe->pipe_pages.cnt;
1547 	else
1548 		ub->st_size = pipe->pipe_buffer.cnt;
1549 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1550 	ub->st_atim = pipe->pipe_atime;
1551 	ub->st_mtim = pipe->pipe_mtime;
1552 	ub->st_ctim = pipe->pipe_ctime;
1553 	ub->st_uid = fp->f_cred->cr_uid;
1554 	ub->st_gid = fp->f_cred->cr_gid;
1555 	ub->st_dev = pipedev_ino;
1556 	ub->st_ino = pipe->pipe_ino;
1557 	/*
1558 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1559 	 */
1560 	return (0);
1561 }
1562 
1563 /* ARGSUSED */
1564 static int
1565 pipe_close(struct file *fp, struct thread *td)
1566 {
1567 
1568 	if (fp->f_vnode != NULL)
1569 		return vnops.fo_close(fp, td);
1570 	fp->f_ops = &badfileops;
1571 	pipe_dtor(fp->f_data);
1572 	fp->f_data = NULL;
1573 	return (0);
1574 }
1575 
1576 static int
1577 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1578 {
1579 	struct pipe *cpipe;
1580 	int error;
1581 
1582 	cpipe = fp->f_data;
1583 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1584 		error = vn_chmod(fp, mode, active_cred, td);
1585 	else
1586 		error = invfo_chmod(fp, mode, active_cred, td);
1587 	return (error);
1588 }
1589 
1590 static int
1591 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1592     struct thread *td)
1593 {
1594 	struct pipe *cpipe;
1595 	int error;
1596 
1597 	cpipe = fp->f_data;
1598 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1599 		error = vn_chown(fp, uid, gid, active_cred, td);
1600 	else
1601 		error = invfo_chown(fp, uid, gid, active_cred, td);
1602 	return (error);
1603 }
1604 
1605 static int
1606 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1607 {
1608 	struct pipe *pi;
1609 
1610 	if (fp->f_type == DTYPE_FIFO)
1611 		return (vn_fill_kinfo(fp, kif, fdp));
1612 	kif->kf_type = KF_TYPE_PIPE;
1613 	pi = fp->f_data;
1614 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1615 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1616 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1617 	return (0);
1618 }
1619 
1620 static void
1621 pipe_free_kmem(struct pipe *cpipe)
1622 {
1623 
1624 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1625 	    ("pipe_free_kmem: pipe mutex locked"));
1626 
1627 	if (cpipe->pipe_buffer.buffer != NULL) {
1628 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1629 		vm_map_remove(pipe_map,
1630 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1631 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1632 		cpipe->pipe_buffer.buffer = NULL;
1633 	}
1634 #ifndef PIPE_NODIRECT
1635 	{
1636 		cpipe->pipe_pages.cnt = 0;
1637 		cpipe->pipe_pages.pos = 0;
1638 		cpipe->pipe_pages.npages = 0;
1639 	}
1640 #endif
1641 }
1642 
1643 /*
1644  * shutdown the pipe
1645  */
1646 static void
1647 pipeclose(struct pipe *cpipe)
1648 {
1649 	struct pipepair *pp;
1650 	struct pipe *ppipe;
1651 
1652 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1653 
1654 	PIPE_LOCK(cpipe);
1655 	pipelock(cpipe, 0);
1656 	pp = cpipe->pipe_pair;
1657 
1658 	/*
1659 	 * If the other side is blocked, wake it up saying that
1660 	 * we want to close it down.
1661 	 */
1662 	cpipe->pipe_state |= PIPE_EOF;
1663 	while (cpipe->pipe_busy) {
1664 		wakeup(cpipe);
1665 		cpipe->pipe_state |= PIPE_WANT;
1666 		pipeunlock(cpipe);
1667 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1668 		pipelock(cpipe, 0);
1669 	}
1670 
1671 	pipeselwakeup(cpipe);
1672 
1673 	/*
1674 	 * Disconnect from peer, if any.
1675 	 */
1676 	ppipe = cpipe->pipe_peer;
1677 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1678 		ppipe->pipe_state |= PIPE_EOF;
1679 		wakeup(ppipe);
1680 		pipeselwakeup(ppipe);
1681 	}
1682 
1683 	/*
1684 	 * Mark this endpoint as free.  Release kmem resources.  We
1685 	 * don't mark this endpoint as unused until we've finished
1686 	 * doing that, or the pipe might disappear out from under
1687 	 * us.
1688 	 */
1689 	PIPE_UNLOCK(cpipe);
1690 	pipe_free_kmem(cpipe);
1691 	PIPE_LOCK(cpipe);
1692 	cpipe->pipe_present = PIPE_CLOSING;
1693 	pipeunlock(cpipe);
1694 
1695 	/*
1696 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1697 	 * PIPE_FINALIZED, that allows other end to free the
1698 	 * pipe_pair, only after the knotes are completely dismantled.
1699 	 */
1700 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1701 	cpipe->pipe_present = PIPE_FINALIZED;
1702 	seldrain(&cpipe->pipe_sel);
1703 	knlist_destroy(&cpipe->pipe_sel.si_note);
1704 
1705 	/*
1706 	 * If both endpoints are now closed, release the memory for the
1707 	 * pipe pair.  If not, unlock.
1708 	 */
1709 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1710 		PIPE_UNLOCK(cpipe);
1711 #ifdef MAC
1712 		mac_pipe_destroy(pp);
1713 #endif
1714 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1715 	} else
1716 		PIPE_UNLOCK(cpipe);
1717 }
1718 
1719 /*ARGSUSED*/
1720 static int
1721 pipe_kqfilter(struct file *fp, struct knote *kn)
1722 {
1723 	struct pipe *cpipe;
1724 
1725 	/*
1726 	 * If a filter is requested that is not supported by this file
1727 	 * descriptor, don't return an error, but also don't ever generate an
1728 	 * event.
1729 	 */
1730 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1731 		kn->kn_fop = &pipe_nfiltops;
1732 		return (0);
1733 	}
1734 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1735 		kn->kn_fop = &pipe_nfiltops;
1736 		return (0);
1737 	}
1738 	cpipe = fp->f_data;
1739 	PIPE_LOCK(cpipe);
1740 	switch (kn->kn_filter) {
1741 	case EVFILT_READ:
1742 		kn->kn_fop = &pipe_rfiltops;
1743 		break;
1744 	case EVFILT_WRITE:
1745 		kn->kn_fop = &pipe_wfiltops;
1746 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1747 			/* other end of pipe has been closed */
1748 			PIPE_UNLOCK(cpipe);
1749 			return (EPIPE);
1750 		}
1751 		cpipe = PIPE_PEER(cpipe);
1752 		break;
1753 	default:
1754 		PIPE_UNLOCK(cpipe);
1755 		return (EINVAL);
1756 	}
1757 
1758 	kn->kn_hook = cpipe;
1759 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1760 	PIPE_UNLOCK(cpipe);
1761 	return (0);
1762 }
1763 
1764 static void
1765 filt_pipedetach(struct knote *kn)
1766 {
1767 	struct pipe *cpipe = kn->kn_hook;
1768 
1769 	PIPE_LOCK(cpipe);
1770 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1771 	PIPE_UNLOCK(cpipe);
1772 }
1773 
1774 /*ARGSUSED*/
1775 static int
1776 filt_piperead(struct knote *kn, long hint)
1777 {
1778 	struct file *fp = kn->kn_fp;
1779 	struct pipe *rpipe = kn->kn_hook;
1780 
1781 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1782 	kn->kn_data = rpipe->pipe_buffer.cnt;
1783 	if (kn->kn_data == 0)
1784 		kn->kn_data = rpipe->pipe_pages.cnt;
1785 
1786 	if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1787 	    ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1788 	    fp->f_pipegen != rpipe->pipe_wgen)) {
1789 		kn->kn_flags |= EV_EOF;
1790 		return (1);
1791 	}
1792 	kn->kn_flags &= ~EV_EOF;
1793 	return (kn->kn_data > 0);
1794 }
1795 
1796 /*ARGSUSED*/
1797 static int
1798 filt_pipewrite(struct knote *kn, long hint)
1799 {
1800 	struct pipe *wpipe = kn->kn_hook;
1801 
1802 	/*
1803 	 * If this end of the pipe is closed, the knote was removed from the
1804 	 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1805 	 */
1806 	if (wpipe->pipe_present == PIPE_ACTIVE ||
1807 	    (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1808 		PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1809 
1810 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1811 			kn->kn_data = 0;
1812 		} else if (wpipe->pipe_buffer.size > 0) {
1813 			kn->kn_data = wpipe->pipe_buffer.size -
1814 			    wpipe->pipe_buffer.cnt;
1815 		} else {
1816 			kn->kn_data = PIPE_BUF;
1817 		}
1818 	}
1819 
1820 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1821 	    (wpipe->pipe_state & PIPE_EOF)) {
1822 		kn->kn_flags |= EV_EOF;
1823 		return (1);
1824 	}
1825 	kn->kn_flags &= ~EV_EOF;
1826 	return (kn->kn_data >= PIPE_BUF);
1827 }
1828 
1829 static void
1830 filt_pipedetach_notsup(struct knote *kn)
1831 {
1832 
1833 }
1834 
1835 static int
1836 filt_pipenotsup(struct knote *kn, long hint)
1837 {
1838 
1839 	return (0);
1840 }
1841