xref: /freebsd/sys/kern/sys_pipe.c (revision 8ef24a0d4b28fe230e20637f56869cc4148cd2ca)
1 /*-
2  * Copyright (c) 1996 John S. Dyson
3  * Copyright (c) 2012 Giovanni Trematerra
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice immediately at the beginning of the file, without modification,
11  *    this list of conditions, and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Absolutely no warranty of function or purpose is made by the author
16  *    John S. Dyson.
17  * 4. Modifications may be freely made to this file if the above conditions
18  *    are met.
19  */
20 
21 /*
22  * This file contains a high-performance replacement for the socket-based
23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
24  * all features of sockets, but does do everything that pipes normally
25  * do.
26  */
27 
28 /*
29  * This code has two modes of operation, a small write mode and a large
30  * write mode.  The small write mode acts like conventional pipes with
31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
34  * memory, and the receiving process copies directly from these pinned pages
35  * in the sending process.
36  *
37  * If the sending process receives a signal, it is possible that it will
38  * go away, and certainly its address space can change, because control
39  * is returned back to the user-mode side.  In that case, the pipe code
40  * arranges to copy the buffer supplied by the user process, to a pageable
41  * kernel buffer, and the receiving process will grab the data from the
42  * pageable kernel buffer.  Since signals don't happen all that often,
43  * the copy operation is normally eliminated.
44  *
45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
46  * happen for small transfers so that the system will not spend all of
47  * its time context switching.
48  *
49  * In order to limit the resource use of pipes, two sysctls exist:
50  *
51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
52  * address space available to us in pipe_map. This value is normally
53  * autotuned, but may also be loader tuned.
54  *
55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56  * memory in use by pipes.
57  *
58  * Based on how large pipekva is relative to maxpipekva, the following
59  * will happen:
60  *
61  * 0% - 50%:
62  *     New pipes are given 16K of memory backing, pipes may dynamically
63  *     grow to as large as 64K where needed.
64  * 50% - 75%:
65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
66  *     existing pipes may NOT grow.
67  * 75% - 100%:
68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
69  *     existing pipes will be shrunk down to 4K whenever possible.
70  *
71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
73  * resize which MUST occur for reverse-direction pipes when they are
74  * first used.
75  *
76  * Additional information about the current state of pipes may be obtained
77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
78  * and kern.ipc.piperesizefail.
79  *
80  * Locking rules:  There are two locks present here:  A mutex, used via
81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
82  * the flag, as mutexes can not persist over uiomove.  The mutex
83  * exists only to guard access to the flag, and is not in itself a
84  * locking mechanism.  Also note that there is only a single mutex for
85  * both directions of a pipe.
86  *
87  * As pipelock() may have to sleep before it can acquire the flag, it
88  * is important to reread all data after a call to pipelock(); everything
89  * in the structure may have changed.
90  */
91 
92 #include <sys/cdefs.h>
93 __FBSDID("$FreeBSD$");
94 
95 #include <sys/param.h>
96 #include <sys/systm.h>
97 #include <sys/conf.h>
98 #include <sys/fcntl.h>
99 #include <sys/file.h>
100 #include <sys/filedesc.h>
101 #include <sys/filio.h>
102 #include <sys/kernel.h>
103 #include <sys/lock.h>
104 #include <sys/mutex.h>
105 #include <sys/ttycom.h>
106 #include <sys/stat.h>
107 #include <sys/malloc.h>
108 #include <sys/poll.h>
109 #include <sys/selinfo.h>
110 #include <sys/signalvar.h>
111 #include <sys/syscallsubr.h>
112 #include <sys/sysctl.h>
113 #include <sys/sysproto.h>
114 #include <sys/pipe.h>
115 #include <sys/proc.h>
116 #include <sys/vnode.h>
117 #include <sys/uio.h>
118 #include <sys/user.h>
119 #include <sys/event.h>
120 
121 #include <security/mac/mac_framework.h>
122 
123 #include <vm/vm.h>
124 #include <vm/vm_param.h>
125 #include <vm/vm_object.h>
126 #include <vm/vm_kern.h>
127 #include <vm/vm_extern.h>
128 #include <vm/pmap.h>
129 #include <vm/vm_map.h>
130 #include <vm/vm_page.h>
131 #include <vm/uma.h>
132 
133 /*
134  * Use this define if you want to disable *fancy* VM things.  Expect an
135  * approx 30% decrease in transfer rate.  This could be useful for
136  * NetBSD or OpenBSD.
137  */
138 /* #define PIPE_NODIRECT */
139 
140 #define PIPE_PEER(pipe)	\
141 	(((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
142 
143 /*
144  * interfaces to the outside world
145  */
146 static fo_rdwr_t	pipe_read;
147 static fo_rdwr_t	pipe_write;
148 static fo_truncate_t	pipe_truncate;
149 static fo_ioctl_t	pipe_ioctl;
150 static fo_poll_t	pipe_poll;
151 static fo_kqfilter_t	pipe_kqfilter;
152 static fo_stat_t	pipe_stat;
153 static fo_close_t	pipe_close;
154 static fo_chmod_t	pipe_chmod;
155 static fo_chown_t	pipe_chown;
156 static fo_fill_kinfo_t	pipe_fill_kinfo;
157 
158 struct fileops pipeops = {
159 	.fo_read = pipe_read,
160 	.fo_write = pipe_write,
161 	.fo_truncate = pipe_truncate,
162 	.fo_ioctl = pipe_ioctl,
163 	.fo_poll = pipe_poll,
164 	.fo_kqfilter = pipe_kqfilter,
165 	.fo_stat = pipe_stat,
166 	.fo_close = pipe_close,
167 	.fo_chmod = pipe_chmod,
168 	.fo_chown = pipe_chown,
169 	.fo_sendfile = invfo_sendfile,
170 	.fo_fill_kinfo = pipe_fill_kinfo,
171 	.fo_flags = DFLAG_PASSABLE
172 };
173 
174 static void	filt_pipedetach(struct knote *kn);
175 static void	filt_pipedetach_notsup(struct knote *kn);
176 static int	filt_pipenotsup(struct knote *kn, long hint);
177 static int	filt_piperead(struct knote *kn, long hint);
178 static int	filt_pipewrite(struct knote *kn, long hint);
179 
180 static struct filterops pipe_nfiltops = {
181 	.f_isfd = 1,
182 	.f_detach = filt_pipedetach_notsup,
183 	.f_event = filt_pipenotsup
184 };
185 static struct filterops pipe_rfiltops = {
186 	.f_isfd = 1,
187 	.f_detach = filt_pipedetach,
188 	.f_event = filt_piperead
189 };
190 static struct filterops pipe_wfiltops = {
191 	.f_isfd = 1,
192 	.f_detach = filt_pipedetach,
193 	.f_event = filt_pipewrite
194 };
195 
196 /*
197  * Default pipe buffer size(s), this can be kind-of large now because pipe
198  * space is pageable.  The pipe code will try to maintain locality of
199  * reference for performance reasons, so small amounts of outstanding I/O
200  * will not wipe the cache.
201  */
202 #define MINPIPESIZE (PIPE_SIZE/3)
203 #define MAXPIPESIZE (2*PIPE_SIZE/3)
204 
205 static long amountpipekva;
206 static int pipefragretry;
207 static int pipeallocfail;
208 static int piperesizefail;
209 static int piperesizeallowed = 1;
210 
211 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
212 	   &maxpipekva, 0, "Pipe KVA limit");
213 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
214 	   &amountpipekva, 0, "Pipe KVA usage");
215 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
216 	  &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
218 	  &pipeallocfail, 0, "Pipe allocation failures");
219 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
220 	  &piperesizefail, 0, "Pipe resize failures");
221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
222 	  &piperesizeallowed, 0, "Pipe resizing allowed");
223 
224 static void pipeinit(void *dummy __unused);
225 static void pipeclose(struct pipe *cpipe);
226 static void pipe_free_kmem(struct pipe *cpipe);
227 static void pipe_create(struct pipe *pipe, int backing);
228 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
229 static __inline int pipelock(struct pipe *cpipe, int catch);
230 static __inline void pipeunlock(struct pipe *cpipe);
231 #ifndef PIPE_NODIRECT
232 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
233 static void pipe_destroy_write_buffer(struct pipe *wpipe);
234 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
235 static void pipe_clone_write_buffer(struct pipe *wpipe);
236 #endif
237 static int pipespace(struct pipe *cpipe, int size);
238 static int pipespace_new(struct pipe *cpipe, int size);
239 
240 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
241 static int	pipe_zone_init(void *mem, int size, int flags);
242 static void	pipe_zone_fini(void *mem, int size);
243 
244 static uma_zone_t pipe_zone;
245 static struct unrhdr *pipeino_unr;
246 static dev_t pipedev_ino;
247 
248 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
249 
250 static void
251 pipeinit(void *dummy __unused)
252 {
253 
254 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
255 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
256 	    UMA_ALIGN_PTR, 0);
257 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
258 	pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
259 	KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
260 	pipedev_ino = devfs_alloc_cdp_inode();
261 	KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
262 }
263 
264 static int
265 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
266 {
267 	struct pipepair *pp;
268 	struct pipe *rpipe, *wpipe;
269 
270 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
271 
272 	pp = (struct pipepair *)mem;
273 
274 	/*
275 	 * We zero both pipe endpoints to make sure all the kmem pointers
276 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
277 	 * endpoints with the same time.
278 	 */
279 	rpipe = &pp->pp_rpipe;
280 	bzero(rpipe, sizeof(*rpipe));
281 	vfs_timestamp(&rpipe->pipe_ctime);
282 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
283 
284 	wpipe = &pp->pp_wpipe;
285 	bzero(wpipe, sizeof(*wpipe));
286 	wpipe->pipe_ctime = rpipe->pipe_ctime;
287 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
288 
289 	rpipe->pipe_peer = wpipe;
290 	rpipe->pipe_pair = pp;
291 	wpipe->pipe_peer = rpipe;
292 	wpipe->pipe_pair = pp;
293 
294 	/*
295 	 * Mark both endpoints as present; they will later get free'd
296 	 * one at a time.  When both are free'd, then the whole pair
297 	 * is released.
298 	 */
299 	rpipe->pipe_present = PIPE_ACTIVE;
300 	wpipe->pipe_present = PIPE_ACTIVE;
301 
302 	/*
303 	 * Eventually, the MAC Framework may initialize the label
304 	 * in ctor or init, but for now we do it elswhere to avoid
305 	 * blocking in ctor or init.
306 	 */
307 	pp->pp_label = NULL;
308 
309 	return (0);
310 }
311 
312 static int
313 pipe_zone_init(void *mem, int size, int flags)
314 {
315 	struct pipepair *pp;
316 
317 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
318 
319 	pp = (struct pipepair *)mem;
320 
321 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
322 	return (0);
323 }
324 
325 static void
326 pipe_zone_fini(void *mem, int size)
327 {
328 	struct pipepair *pp;
329 
330 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
331 
332 	pp = (struct pipepair *)mem;
333 
334 	mtx_destroy(&pp->pp_mtx);
335 }
336 
337 static void
338 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
339 {
340 	struct pipepair *pp;
341 	struct pipe *rpipe, *wpipe;
342 
343 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
344 #ifdef MAC
345 	/*
346 	 * The MAC label is shared between the connected endpoints.  As a
347 	 * result mac_pipe_init() and mac_pipe_create() are called once
348 	 * for the pair, and not on the endpoints.
349 	 */
350 	mac_pipe_init(pp);
351 	mac_pipe_create(td->td_ucred, pp);
352 #endif
353 	rpipe = &pp->pp_rpipe;
354 	wpipe = &pp->pp_wpipe;
355 
356 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
357 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
358 
359 	/* Only the forward direction pipe is backed by default */
360 	pipe_create(rpipe, 1);
361 	pipe_create(wpipe, 0);
362 
363 	rpipe->pipe_state |= PIPE_DIRECTOK;
364 	wpipe->pipe_state |= PIPE_DIRECTOK;
365 }
366 
367 void
368 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
369 {
370 	struct pipepair *pp;
371 
372 	pipe_paircreate(td, &pp);
373 	pp->pp_rpipe.pipe_state |= PIPE_NAMED;
374 	*ppipe = &pp->pp_rpipe;
375 }
376 
377 void
378 pipe_dtor(struct pipe *dpipe)
379 {
380 	struct pipe *peer;
381 	ino_t ino;
382 
383 	ino = dpipe->pipe_ino;
384 	peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
385 	funsetown(&dpipe->pipe_sigio);
386 	pipeclose(dpipe);
387 	if (peer != NULL) {
388 		funsetown(&peer->pipe_sigio);
389 		pipeclose(peer);
390 	}
391 	if (ino != 0 && ino != (ino_t)-1)
392 		free_unr(pipeino_unr, ino);
393 }
394 
395 /*
396  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
397  * the zone pick up the pieces via pipeclose().
398  */
399 int
400 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
401     struct filecaps *fcaps2)
402 {
403 	struct file *rf, *wf;
404 	struct pipe *rpipe, *wpipe;
405 	struct pipepair *pp;
406 	int fd, fflags, error;
407 
408 	pipe_paircreate(td, &pp);
409 	rpipe = &pp->pp_rpipe;
410 	wpipe = &pp->pp_wpipe;
411 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
412 	if (error) {
413 		pipeclose(rpipe);
414 		pipeclose(wpipe);
415 		return (error);
416 	}
417 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
418 	fildes[0] = fd;
419 
420 	fflags = FREAD | FWRITE;
421 	if ((flags & O_NONBLOCK) != 0)
422 		fflags |= FNONBLOCK;
423 
424 	/*
425 	 * Warning: once we've gotten past allocation of the fd for the
426 	 * read-side, we can only drop the read side via fdrop() in order
427 	 * to avoid races against processes which manage to dup() the read
428 	 * side while we are blocked trying to allocate the write side.
429 	 */
430 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
431 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
432 	if (error) {
433 		fdclose(td, rf, fildes[0]);
434 		fdrop(rf, td);
435 		/* rpipe has been closed by fdrop(). */
436 		pipeclose(wpipe);
437 		return (error);
438 	}
439 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
440 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
441 	fdrop(wf, td);
442 	fildes[1] = fd;
443 	fdrop(rf, td);
444 
445 	return (0);
446 }
447 
448 /* ARGSUSED */
449 int
450 sys_pipe(struct thread *td, struct pipe_args *uap)
451 {
452 	int error;
453 	int fildes[2];
454 
455 	error = kern_pipe(td, fildes, 0, NULL, NULL);
456 	if (error)
457 		return (error);
458 
459 	td->td_retval[0] = fildes[0];
460 	td->td_retval[1] = fildes[1];
461 
462 	return (0);
463 }
464 
465 int
466 sys_pipe2(struct thread *td, struct pipe2_args *uap)
467 {
468 	int error, fildes[2];
469 
470 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
471 		return (EINVAL);
472 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
473 	if (error)
474 		return (error);
475 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
476 	if (error) {
477 		(void)kern_close(td, fildes[0]);
478 		(void)kern_close(td, fildes[1]);
479 	}
480 	return (error);
481 }
482 
483 /*
484  * Allocate kva for pipe circular buffer, the space is pageable
485  * This routine will 'realloc' the size of a pipe safely, if it fails
486  * it will retain the old buffer.
487  * If it fails it will return ENOMEM.
488  */
489 static int
490 pipespace_new(cpipe, size)
491 	struct pipe *cpipe;
492 	int size;
493 {
494 	caddr_t buffer;
495 	int error, cnt, firstseg;
496 	static int curfail = 0;
497 	static struct timeval lastfail;
498 
499 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
500 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
501 		("pipespace: resize of direct writes not allowed"));
502 retry:
503 	cnt = cpipe->pipe_buffer.cnt;
504 	if (cnt > size)
505 		size = cnt;
506 
507 	size = round_page(size);
508 	buffer = (caddr_t) vm_map_min(pipe_map);
509 
510 	error = vm_map_find(pipe_map, NULL, 0,
511 		(vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
512 		VM_PROT_ALL, VM_PROT_ALL, 0);
513 	if (error != KERN_SUCCESS) {
514 		if ((cpipe->pipe_buffer.buffer == NULL) &&
515 			(size > SMALL_PIPE_SIZE)) {
516 			size = SMALL_PIPE_SIZE;
517 			pipefragretry++;
518 			goto retry;
519 		}
520 		if (cpipe->pipe_buffer.buffer == NULL) {
521 			pipeallocfail++;
522 			if (ppsratecheck(&lastfail, &curfail, 1))
523 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
524 		} else {
525 			piperesizefail++;
526 		}
527 		return (ENOMEM);
528 	}
529 
530 	/* copy data, then free old resources if we're resizing */
531 	if (cnt > 0) {
532 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
533 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
534 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
535 				buffer, firstseg);
536 			if ((cnt - firstseg) > 0)
537 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
538 					cpipe->pipe_buffer.in);
539 		} else {
540 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
541 				buffer, cnt);
542 		}
543 	}
544 	pipe_free_kmem(cpipe);
545 	cpipe->pipe_buffer.buffer = buffer;
546 	cpipe->pipe_buffer.size = size;
547 	cpipe->pipe_buffer.in = cnt;
548 	cpipe->pipe_buffer.out = 0;
549 	cpipe->pipe_buffer.cnt = cnt;
550 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
551 	return (0);
552 }
553 
554 /*
555  * Wrapper for pipespace_new() that performs locking assertions.
556  */
557 static int
558 pipespace(cpipe, size)
559 	struct pipe *cpipe;
560 	int size;
561 {
562 
563 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
564 		("Unlocked pipe passed to pipespace"));
565 	return (pipespace_new(cpipe, size));
566 }
567 
568 /*
569  * lock a pipe for I/O, blocking other access
570  */
571 static __inline int
572 pipelock(cpipe, catch)
573 	struct pipe *cpipe;
574 	int catch;
575 {
576 	int error;
577 
578 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
579 	while (cpipe->pipe_state & PIPE_LOCKFL) {
580 		cpipe->pipe_state |= PIPE_LWANT;
581 		error = msleep(cpipe, PIPE_MTX(cpipe),
582 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
583 		    "pipelk", 0);
584 		if (error != 0)
585 			return (error);
586 	}
587 	cpipe->pipe_state |= PIPE_LOCKFL;
588 	return (0);
589 }
590 
591 /*
592  * unlock a pipe I/O lock
593  */
594 static __inline void
595 pipeunlock(cpipe)
596 	struct pipe *cpipe;
597 {
598 
599 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
600 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
601 		("Unlocked pipe passed to pipeunlock"));
602 	cpipe->pipe_state &= ~PIPE_LOCKFL;
603 	if (cpipe->pipe_state & PIPE_LWANT) {
604 		cpipe->pipe_state &= ~PIPE_LWANT;
605 		wakeup(cpipe);
606 	}
607 }
608 
609 void
610 pipeselwakeup(cpipe)
611 	struct pipe *cpipe;
612 {
613 
614 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
615 	if (cpipe->pipe_state & PIPE_SEL) {
616 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
617 		if (!SEL_WAITING(&cpipe->pipe_sel))
618 			cpipe->pipe_state &= ~PIPE_SEL;
619 	}
620 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
621 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
622 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
623 }
624 
625 /*
626  * Initialize and allocate VM and memory for pipe.  The structure
627  * will start out zero'd from the ctor, so we just manage the kmem.
628  */
629 static void
630 pipe_create(pipe, backing)
631 	struct pipe *pipe;
632 	int backing;
633 {
634 
635 	if (backing) {
636 		/*
637 		 * Note that these functions can fail if pipe map is exhausted
638 		 * (as a result of too many pipes created), but we ignore the
639 		 * error as it is not fatal and could be provoked by
640 		 * unprivileged users. The only consequence is worse performance
641 		 * with given pipe.
642 		 */
643 		if (amountpipekva > maxpipekva / 2)
644 			(void)pipespace_new(pipe, SMALL_PIPE_SIZE);
645 		else
646 			(void)pipespace_new(pipe, PIPE_SIZE);
647 	}
648 
649 	pipe->pipe_ino = -1;
650 }
651 
652 /* ARGSUSED */
653 static int
654 pipe_read(fp, uio, active_cred, flags, td)
655 	struct file *fp;
656 	struct uio *uio;
657 	struct ucred *active_cred;
658 	struct thread *td;
659 	int flags;
660 {
661 	struct pipe *rpipe;
662 	int error;
663 	int nread = 0;
664 	int size;
665 
666 	rpipe = fp->f_data;
667 	PIPE_LOCK(rpipe);
668 	++rpipe->pipe_busy;
669 	error = pipelock(rpipe, 1);
670 	if (error)
671 		goto unlocked_error;
672 
673 #ifdef MAC
674 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
675 	if (error)
676 		goto locked_error;
677 #endif
678 	if (amountpipekva > (3 * maxpipekva) / 4) {
679 		if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
680 			(rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
681 			(rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
682 			(piperesizeallowed == 1)) {
683 			PIPE_UNLOCK(rpipe);
684 			pipespace(rpipe, SMALL_PIPE_SIZE);
685 			PIPE_LOCK(rpipe);
686 		}
687 	}
688 
689 	while (uio->uio_resid) {
690 		/*
691 		 * normal pipe buffer receive
692 		 */
693 		if (rpipe->pipe_buffer.cnt > 0) {
694 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
695 			if (size > rpipe->pipe_buffer.cnt)
696 				size = rpipe->pipe_buffer.cnt;
697 			if (size > uio->uio_resid)
698 				size = uio->uio_resid;
699 
700 			PIPE_UNLOCK(rpipe);
701 			error = uiomove(
702 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
703 			    size, uio);
704 			PIPE_LOCK(rpipe);
705 			if (error)
706 				break;
707 
708 			rpipe->pipe_buffer.out += size;
709 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
710 				rpipe->pipe_buffer.out = 0;
711 
712 			rpipe->pipe_buffer.cnt -= size;
713 
714 			/*
715 			 * If there is no more to read in the pipe, reset
716 			 * its pointers to the beginning.  This improves
717 			 * cache hit stats.
718 			 */
719 			if (rpipe->pipe_buffer.cnt == 0) {
720 				rpipe->pipe_buffer.in = 0;
721 				rpipe->pipe_buffer.out = 0;
722 			}
723 			nread += size;
724 #ifndef PIPE_NODIRECT
725 		/*
726 		 * Direct copy, bypassing a kernel buffer.
727 		 */
728 		} else if ((size = rpipe->pipe_map.cnt) &&
729 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
730 			if (size > uio->uio_resid)
731 				size = (u_int) uio->uio_resid;
732 
733 			PIPE_UNLOCK(rpipe);
734 			error = uiomove_fromphys(rpipe->pipe_map.ms,
735 			    rpipe->pipe_map.pos, size, uio);
736 			PIPE_LOCK(rpipe);
737 			if (error)
738 				break;
739 			nread += size;
740 			rpipe->pipe_map.pos += size;
741 			rpipe->pipe_map.cnt -= size;
742 			if (rpipe->pipe_map.cnt == 0) {
743 				rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
744 				wakeup(rpipe);
745 			}
746 #endif
747 		} else {
748 			/*
749 			 * detect EOF condition
750 			 * read returns 0 on EOF, no need to set error
751 			 */
752 			if (rpipe->pipe_state & PIPE_EOF)
753 				break;
754 
755 			/*
756 			 * If the "write-side" has been blocked, wake it up now.
757 			 */
758 			if (rpipe->pipe_state & PIPE_WANTW) {
759 				rpipe->pipe_state &= ~PIPE_WANTW;
760 				wakeup(rpipe);
761 			}
762 
763 			/*
764 			 * Break if some data was read.
765 			 */
766 			if (nread > 0)
767 				break;
768 
769 			/*
770 			 * Unlock the pipe buffer for our remaining processing.
771 			 * We will either break out with an error or we will
772 			 * sleep and relock to loop.
773 			 */
774 			pipeunlock(rpipe);
775 
776 			/*
777 			 * Handle non-blocking mode operation or
778 			 * wait for more data.
779 			 */
780 			if (fp->f_flag & FNONBLOCK) {
781 				error = EAGAIN;
782 			} else {
783 				rpipe->pipe_state |= PIPE_WANTR;
784 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
785 				    PRIBIO | PCATCH,
786 				    "piperd", 0)) == 0)
787 					error = pipelock(rpipe, 1);
788 			}
789 			if (error)
790 				goto unlocked_error;
791 		}
792 	}
793 #ifdef MAC
794 locked_error:
795 #endif
796 	pipeunlock(rpipe);
797 
798 	/* XXX: should probably do this before getting any locks. */
799 	if (error == 0)
800 		vfs_timestamp(&rpipe->pipe_atime);
801 unlocked_error:
802 	--rpipe->pipe_busy;
803 
804 	/*
805 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
806 	 */
807 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
808 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
809 		wakeup(rpipe);
810 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
811 		/*
812 		 * Handle write blocking hysteresis.
813 		 */
814 		if (rpipe->pipe_state & PIPE_WANTW) {
815 			rpipe->pipe_state &= ~PIPE_WANTW;
816 			wakeup(rpipe);
817 		}
818 	}
819 
820 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
821 		pipeselwakeup(rpipe);
822 
823 	PIPE_UNLOCK(rpipe);
824 	return (error);
825 }
826 
827 #ifndef PIPE_NODIRECT
828 /*
829  * Map the sending processes' buffer into kernel space and wire it.
830  * This is similar to a physical write operation.
831  */
832 static int
833 pipe_build_write_buffer(wpipe, uio)
834 	struct pipe *wpipe;
835 	struct uio *uio;
836 {
837 	u_int size;
838 	int i;
839 
840 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
841 	KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
842 		("Clone attempt on non-direct write pipe!"));
843 
844 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
845                 size = wpipe->pipe_buffer.size;
846 	else
847                 size = uio->uio_iov->iov_len;
848 
849 	if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
850 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
851 	    wpipe->pipe_map.ms, PIPENPAGES)) < 0)
852 		return (EFAULT);
853 
854 /*
855  * set up the control block
856  */
857 	wpipe->pipe_map.npages = i;
858 	wpipe->pipe_map.pos =
859 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
860 	wpipe->pipe_map.cnt = size;
861 
862 /*
863  * and update the uio data
864  */
865 
866 	uio->uio_iov->iov_len -= size;
867 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
868 	if (uio->uio_iov->iov_len == 0)
869 		uio->uio_iov++;
870 	uio->uio_resid -= size;
871 	uio->uio_offset += size;
872 	return (0);
873 }
874 
875 /*
876  * unmap and unwire the process buffer
877  */
878 static void
879 pipe_destroy_write_buffer(wpipe)
880 	struct pipe *wpipe;
881 {
882 
883 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
884 	vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
885 	wpipe->pipe_map.npages = 0;
886 }
887 
888 /*
889  * In the case of a signal, the writing process might go away.  This
890  * code copies the data into the circular buffer so that the source
891  * pages can be freed without loss of data.
892  */
893 static void
894 pipe_clone_write_buffer(wpipe)
895 	struct pipe *wpipe;
896 {
897 	struct uio uio;
898 	struct iovec iov;
899 	int size;
900 	int pos;
901 
902 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
903 	size = wpipe->pipe_map.cnt;
904 	pos = wpipe->pipe_map.pos;
905 
906 	wpipe->pipe_buffer.in = size;
907 	wpipe->pipe_buffer.out = 0;
908 	wpipe->pipe_buffer.cnt = size;
909 	wpipe->pipe_state &= ~PIPE_DIRECTW;
910 
911 	PIPE_UNLOCK(wpipe);
912 	iov.iov_base = wpipe->pipe_buffer.buffer;
913 	iov.iov_len = size;
914 	uio.uio_iov = &iov;
915 	uio.uio_iovcnt = 1;
916 	uio.uio_offset = 0;
917 	uio.uio_resid = size;
918 	uio.uio_segflg = UIO_SYSSPACE;
919 	uio.uio_rw = UIO_READ;
920 	uio.uio_td = curthread;
921 	uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
922 	PIPE_LOCK(wpipe);
923 	pipe_destroy_write_buffer(wpipe);
924 }
925 
926 /*
927  * This implements the pipe buffer write mechanism.  Note that only
928  * a direct write OR a normal pipe write can be pending at any given time.
929  * If there are any characters in the pipe buffer, the direct write will
930  * be deferred until the receiving process grabs all of the bytes from
931  * the pipe buffer.  Then the direct mapping write is set-up.
932  */
933 static int
934 pipe_direct_write(wpipe, uio)
935 	struct pipe *wpipe;
936 	struct uio *uio;
937 {
938 	int error;
939 
940 retry:
941 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
942 	error = pipelock(wpipe, 1);
943 	if (error != 0)
944 		goto error1;
945 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
946 		error = EPIPE;
947 		pipeunlock(wpipe);
948 		goto error1;
949 	}
950 	while (wpipe->pipe_state & PIPE_DIRECTW) {
951 		if (wpipe->pipe_state & PIPE_WANTR) {
952 			wpipe->pipe_state &= ~PIPE_WANTR;
953 			wakeup(wpipe);
954 		}
955 		pipeselwakeup(wpipe);
956 		wpipe->pipe_state |= PIPE_WANTW;
957 		pipeunlock(wpipe);
958 		error = msleep(wpipe, PIPE_MTX(wpipe),
959 		    PRIBIO | PCATCH, "pipdww", 0);
960 		if (error)
961 			goto error1;
962 		else
963 			goto retry;
964 	}
965 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
966 	if (wpipe->pipe_buffer.cnt > 0) {
967 		if (wpipe->pipe_state & PIPE_WANTR) {
968 			wpipe->pipe_state &= ~PIPE_WANTR;
969 			wakeup(wpipe);
970 		}
971 		pipeselwakeup(wpipe);
972 		wpipe->pipe_state |= PIPE_WANTW;
973 		pipeunlock(wpipe);
974 		error = msleep(wpipe, PIPE_MTX(wpipe),
975 		    PRIBIO | PCATCH, "pipdwc", 0);
976 		if (error)
977 			goto error1;
978 		else
979 			goto retry;
980 	}
981 
982 	wpipe->pipe_state |= PIPE_DIRECTW;
983 
984 	PIPE_UNLOCK(wpipe);
985 	error = pipe_build_write_buffer(wpipe, uio);
986 	PIPE_LOCK(wpipe);
987 	if (error) {
988 		wpipe->pipe_state &= ~PIPE_DIRECTW;
989 		pipeunlock(wpipe);
990 		goto error1;
991 	}
992 
993 	error = 0;
994 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
995 		if (wpipe->pipe_state & PIPE_EOF) {
996 			pipe_destroy_write_buffer(wpipe);
997 			pipeselwakeup(wpipe);
998 			pipeunlock(wpipe);
999 			error = EPIPE;
1000 			goto error1;
1001 		}
1002 		if (wpipe->pipe_state & PIPE_WANTR) {
1003 			wpipe->pipe_state &= ~PIPE_WANTR;
1004 			wakeup(wpipe);
1005 		}
1006 		pipeselwakeup(wpipe);
1007 		wpipe->pipe_state |= PIPE_WANTW;
1008 		pipeunlock(wpipe);
1009 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1010 		    "pipdwt", 0);
1011 		pipelock(wpipe, 0);
1012 	}
1013 
1014 	if (wpipe->pipe_state & PIPE_EOF)
1015 		error = EPIPE;
1016 	if (wpipe->pipe_state & PIPE_DIRECTW) {
1017 		/*
1018 		 * this bit of trickery substitutes a kernel buffer for
1019 		 * the process that might be going away.
1020 		 */
1021 		pipe_clone_write_buffer(wpipe);
1022 	} else {
1023 		pipe_destroy_write_buffer(wpipe);
1024 	}
1025 	pipeunlock(wpipe);
1026 	return (error);
1027 
1028 error1:
1029 	wakeup(wpipe);
1030 	return (error);
1031 }
1032 #endif
1033 
1034 static int
1035 pipe_write(fp, uio, active_cred, flags, td)
1036 	struct file *fp;
1037 	struct uio *uio;
1038 	struct ucred *active_cred;
1039 	struct thread *td;
1040 	int flags;
1041 {
1042 	int error = 0;
1043 	int desiredsize;
1044 	ssize_t orig_resid;
1045 	struct pipe *wpipe, *rpipe;
1046 
1047 	rpipe = fp->f_data;
1048 	wpipe = PIPE_PEER(rpipe);
1049 	PIPE_LOCK(rpipe);
1050 	error = pipelock(wpipe, 1);
1051 	if (error) {
1052 		PIPE_UNLOCK(rpipe);
1053 		return (error);
1054 	}
1055 	/*
1056 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1057 	 */
1058 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1059 	    (wpipe->pipe_state & PIPE_EOF)) {
1060 		pipeunlock(wpipe);
1061 		PIPE_UNLOCK(rpipe);
1062 		return (EPIPE);
1063 	}
1064 #ifdef MAC
1065 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1066 	if (error) {
1067 		pipeunlock(wpipe);
1068 		PIPE_UNLOCK(rpipe);
1069 		return (error);
1070 	}
1071 #endif
1072 	++wpipe->pipe_busy;
1073 
1074 	/* Choose a larger size if it's advantageous */
1075 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1076 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1077 		if (piperesizeallowed != 1)
1078 			break;
1079 		if (amountpipekva > maxpipekva / 2)
1080 			break;
1081 		if (desiredsize == BIG_PIPE_SIZE)
1082 			break;
1083 		desiredsize = desiredsize * 2;
1084 	}
1085 
1086 	/* Choose a smaller size if we're in a OOM situation */
1087 	if ((amountpipekva > (3 * maxpipekva) / 4) &&
1088 		(wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1089 		(wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1090 		(piperesizeallowed == 1))
1091 		desiredsize = SMALL_PIPE_SIZE;
1092 
1093 	/* Resize if the above determined that a new size was necessary */
1094 	if ((desiredsize != wpipe->pipe_buffer.size) &&
1095 		((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1096 		PIPE_UNLOCK(wpipe);
1097 		pipespace(wpipe, desiredsize);
1098 		PIPE_LOCK(wpipe);
1099 	}
1100 	if (wpipe->pipe_buffer.size == 0) {
1101 		/*
1102 		 * This can only happen for reverse direction use of pipes
1103 		 * in a complete OOM situation.
1104 		 */
1105 		error = ENOMEM;
1106 		--wpipe->pipe_busy;
1107 		pipeunlock(wpipe);
1108 		PIPE_UNLOCK(wpipe);
1109 		return (error);
1110 	}
1111 
1112 	pipeunlock(wpipe);
1113 
1114 	orig_resid = uio->uio_resid;
1115 
1116 	while (uio->uio_resid) {
1117 		int space;
1118 
1119 		pipelock(wpipe, 0);
1120 		if (wpipe->pipe_state & PIPE_EOF) {
1121 			pipeunlock(wpipe);
1122 			error = EPIPE;
1123 			break;
1124 		}
1125 #ifndef PIPE_NODIRECT
1126 		/*
1127 		 * If the transfer is large, we can gain performance if
1128 		 * we do process-to-process copies directly.
1129 		 * If the write is non-blocking, we don't use the
1130 		 * direct write mechanism.
1131 		 *
1132 		 * The direct write mechanism will detect the reader going
1133 		 * away on us.
1134 		 */
1135 		if (uio->uio_segflg == UIO_USERSPACE &&
1136 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1137 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1138 		    (fp->f_flag & FNONBLOCK) == 0) {
1139 			pipeunlock(wpipe);
1140 			error = pipe_direct_write(wpipe, uio);
1141 			if (error)
1142 				break;
1143 			continue;
1144 		}
1145 #endif
1146 
1147 		/*
1148 		 * Pipe buffered writes cannot be coincidental with
1149 		 * direct writes.  We wait until the currently executing
1150 		 * direct write is completed before we start filling the
1151 		 * pipe buffer.  We break out if a signal occurs or the
1152 		 * reader goes away.
1153 		 */
1154 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1155 			if (wpipe->pipe_state & PIPE_WANTR) {
1156 				wpipe->pipe_state &= ~PIPE_WANTR;
1157 				wakeup(wpipe);
1158 			}
1159 			pipeselwakeup(wpipe);
1160 			wpipe->pipe_state |= PIPE_WANTW;
1161 			pipeunlock(wpipe);
1162 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1163 			    "pipbww", 0);
1164 			if (error)
1165 				break;
1166 			else
1167 				continue;
1168 		}
1169 
1170 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1171 
1172 		/* Writes of size <= PIPE_BUF must be atomic. */
1173 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1174 			space = 0;
1175 
1176 		if (space > 0) {
1177 			int size;	/* Transfer size */
1178 			int segsize;	/* first segment to transfer */
1179 
1180 			/*
1181 			 * Transfer size is minimum of uio transfer
1182 			 * and free space in pipe buffer.
1183 			 */
1184 			if (space > uio->uio_resid)
1185 				size = uio->uio_resid;
1186 			else
1187 				size = space;
1188 			/*
1189 			 * First segment to transfer is minimum of
1190 			 * transfer size and contiguous space in
1191 			 * pipe buffer.  If first segment to transfer
1192 			 * is less than the transfer size, we've got
1193 			 * a wraparound in the buffer.
1194 			 */
1195 			segsize = wpipe->pipe_buffer.size -
1196 				wpipe->pipe_buffer.in;
1197 			if (segsize > size)
1198 				segsize = size;
1199 
1200 			/* Transfer first segment */
1201 
1202 			PIPE_UNLOCK(rpipe);
1203 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1204 					segsize, uio);
1205 			PIPE_LOCK(rpipe);
1206 
1207 			if (error == 0 && segsize < size) {
1208 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1209 					wpipe->pipe_buffer.size,
1210 					("Pipe buffer wraparound disappeared"));
1211 				/*
1212 				 * Transfer remaining part now, to
1213 				 * support atomic writes.  Wraparound
1214 				 * happened.
1215 				 */
1216 
1217 				PIPE_UNLOCK(rpipe);
1218 				error = uiomove(
1219 				    &wpipe->pipe_buffer.buffer[0],
1220 				    size - segsize, uio);
1221 				PIPE_LOCK(rpipe);
1222 			}
1223 			if (error == 0) {
1224 				wpipe->pipe_buffer.in += size;
1225 				if (wpipe->pipe_buffer.in >=
1226 				    wpipe->pipe_buffer.size) {
1227 					KASSERT(wpipe->pipe_buffer.in ==
1228 						size - segsize +
1229 						wpipe->pipe_buffer.size,
1230 						("Expected wraparound bad"));
1231 					wpipe->pipe_buffer.in = size - segsize;
1232 				}
1233 
1234 				wpipe->pipe_buffer.cnt += size;
1235 				KASSERT(wpipe->pipe_buffer.cnt <=
1236 					wpipe->pipe_buffer.size,
1237 					("Pipe buffer overflow"));
1238 			}
1239 			pipeunlock(wpipe);
1240 			if (error != 0)
1241 				break;
1242 		} else {
1243 			/*
1244 			 * If the "read-side" has been blocked, wake it up now.
1245 			 */
1246 			if (wpipe->pipe_state & PIPE_WANTR) {
1247 				wpipe->pipe_state &= ~PIPE_WANTR;
1248 				wakeup(wpipe);
1249 			}
1250 
1251 			/*
1252 			 * don't block on non-blocking I/O
1253 			 */
1254 			if (fp->f_flag & FNONBLOCK) {
1255 				error = EAGAIN;
1256 				pipeunlock(wpipe);
1257 				break;
1258 			}
1259 
1260 			/*
1261 			 * We have no more space and have something to offer,
1262 			 * wake up select/poll.
1263 			 */
1264 			pipeselwakeup(wpipe);
1265 
1266 			wpipe->pipe_state |= PIPE_WANTW;
1267 			pipeunlock(wpipe);
1268 			error = msleep(wpipe, PIPE_MTX(rpipe),
1269 			    PRIBIO | PCATCH, "pipewr", 0);
1270 			if (error != 0)
1271 				break;
1272 		}
1273 	}
1274 
1275 	pipelock(wpipe, 0);
1276 	--wpipe->pipe_busy;
1277 
1278 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1279 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1280 		wakeup(wpipe);
1281 	} else if (wpipe->pipe_buffer.cnt > 0) {
1282 		/*
1283 		 * If we have put any characters in the buffer, we wake up
1284 		 * the reader.
1285 		 */
1286 		if (wpipe->pipe_state & PIPE_WANTR) {
1287 			wpipe->pipe_state &= ~PIPE_WANTR;
1288 			wakeup(wpipe);
1289 		}
1290 	}
1291 
1292 	/*
1293 	 * Don't return EPIPE if any byte was written.
1294 	 * EINTR and other interrupts are handled by generic I/O layer.
1295 	 * Do not pretend that I/O succeeded for obvious user error
1296 	 * like EFAULT.
1297 	 */
1298 	if (uio->uio_resid != orig_resid && error == EPIPE)
1299 		error = 0;
1300 
1301 	if (error == 0)
1302 		vfs_timestamp(&wpipe->pipe_mtime);
1303 
1304 	/*
1305 	 * We have something to offer,
1306 	 * wake up select/poll.
1307 	 */
1308 	if (wpipe->pipe_buffer.cnt)
1309 		pipeselwakeup(wpipe);
1310 
1311 	pipeunlock(wpipe);
1312 	PIPE_UNLOCK(rpipe);
1313 	return (error);
1314 }
1315 
1316 /* ARGSUSED */
1317 static int
1318 pipe_truncate(fp, length, active_cred, td)
1319 	struct file *fp;
1320 	off_t length;
1321 	struct ucred *active_cred;
1322 	struct thread *td;
1323 {
1324 	struct pipe *cpipe;
1325 	int error;
1326 
1327 	cpipe = fp->f_data;
1328 	if (cpipe->pipe_state & PIPE_NAMED)
1329 		error = vnops.fo_truncate(fp, length, active_cred, td);
1330 	else
1331 		error = invfo_truncate(fp, length, active_cred, td);
1332 	return (error);
1333 }
1334 
1335 /*
1336  * we implement a very minimal set of ioctls for compatibility with sockets.
1337  */
1338 static int
1339 pipe_ioctl(fp, cmd, data, active_cred, td)
1340 	struct file *fp;
1341 	u_long cmd;
1342 	void *data;
1343 	struct ucred *active_cred;
1344 	struct thread *td;
1345 {
1346 	struct pipe *mpipe = fp->f_data;
1347 	int error;
1348 
1349 	PIPE_LOCK(mpipe);
1350 
1351 #ifdef MAC
1352 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1353 	if (error) {
1354 		PIPE_UNLOCK(mpipe);
1355 		return (error);
1356 	}
1357 #endif
1358 
1359 	error = 0;
1360 	switch (cmd) {
1361 
1362 	case FIONBIO:
1363 		break;
1364 
1365 	case FIOASYNC:
1366 		if (*(int *)data) {
1367 			mpipe->pipe_state |= PIPE_ASYNC;
1368 		} else {
1369 			mpipe->pipe_state &= ~PIPE_ASYNC;
1370 		}
1371 		break;
1372 
1373 	case FIONREAD:
1374 		if (!(fp->f_flag & FREAD)) {
1375 			*(int *)data = 0;
1376 			PIPE_UNLOCK(mpipe);
1377 			return (0);
1378 		}
1379 		if (mpipe->pipe_state & PIPE_DIRECTW)
1380 			*(int *)data = mpipe->pipe_map.cnt;
1381 		else
1382 			*(int *)data = mpipe->pipe_buffer.cnt;
1383 		break;
1384 
1385 	case FIOSETOWN:
1386 		PIPE_UNLOCK(mpipe);
1387 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1388 		goto out_unlocked;
1389 
1390 	case FIOGETOWN:
1391 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1392 		break;
1393 
1394 	/* This is deprecated, FIOSETOWN should be used instead. */
1395 	case TIOCSPGRP:
1396 		PIPE_UNLOCK(mpipe);
1397 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1398 		goto out_unlocked;
1399 
1400 	/* This is deprecated, FIOGETOWN should be used instead. */
1401 	case TIOCGPGRP:
1402 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1403 		break;
1404 
1405 	default:
1406 		error = ENOTTY;
1407 		break;
1408 	}
1409 	PIPE_UNLOCK(mpipe);
1410 out_unlocked:
1411 	return (error);
1412 }
1413 
1414 static int
1415 pipe_poll(fp, events, active_cred, td)
1416 	struct file *fp;
1417 	int events;
1418 	struct ucred *active_cred;
1419 	struct thread *td;
1420 {
1421 	struct pipe *rpipe;
1422 	struct pipe *wpipe;
1423 	int levents, revents;
1424 #ifdef MAC
1425 	int error;
1426 #endif
1427 
1428 	revents = 0;
1429 	rpipe = fp->f_data;
1430 	wpipe = PIPE_PEER(rpipe);
1431 	PIPE_LOCK(rpipe);
1432 #ifdef MAC
1433 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1434 	if (error)
1435 		goto locked_error;
1436 #endif
1437 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1438 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1439 		    (rpipe->pipe_buffer.cnt > 0))
1440 			revents |= events & (POLLIN | POLLRDNORM);
1441 
1442 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1443 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1444 		    (wpipe->pipe_state & PIPE_EOF) ||
1445 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1446 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1447 			 wpipe->pipe_buffer.size == 0)))
1448 			revents |= events & (POLLOUT | POLLWRNORM);
1449 
1450 	levents = events &
1451 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1452 	if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1453 	    fp->f_seqcount == rpipe->pipe_wgen)
1454 		events |= POLLINIGNEOF;
1455 
1456 	if ((events & POLLINIGNEOF) == 0) {
1457 		if (rpipe->pipe_state & PIPE_EOF) {
1458 			revents |= (events & (POLLIN | POLLRDNORM));
1459 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1460 			    (wpipe->pipe_state & PIPE_EOF))
1461 				revents |= POLLHUP;
1462 		}
1463 	}
1464 
1465 	if (revents == 0) {
1466 		if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1467 			selrecord(td, &rpipe->pipe_sel);
1468 			if (SEL_WAITING(&rpipe->pipe_sel))
1469 				rpipe->pipe_state |= PIPE_SEL;
1470 		}
1471 
1472 		if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1473 			selrecord(td, &wpipe->pipe_sel);
1474 			if (SEL_WAITING(&wpipe->pipe_sel))
1475 				wpipe->pipe_state |= PIPE_SEL;
1476 		}
1477 	}
1478 #ifdef MAC
1479 locked_error:
1480 #endif
1481 	PIPE_UNLOCK(rpipe);
1482 
1483 	return (revents);
1484 }
1485 
1486 /*
1487  * We shouldn't need locks here as we're doing a read and this should
1488  * be a natural race.
1489  */
1490 static int
1491 pipe_stat(fp, ub, active_cred, td)
1492 	struct file *fp;
1493 	struct stat *ub;
1494 	struct ucred *active_cred;
1495 	struct thread *td;
1496 {
1497 	struct pipe *pipe;
1498 	int new_unr;
1499 #ifdef MAC
1500 	int error;
1501 #endif
1502 
1503 	pipe = fp->f_data;
1504 	PIPE_LOCK(pipe);
1505 #ifdef MAC
1506 	error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1507 	if (error) {
1508 		PIPE_UNLOCK(pipe);
1509 		return (error);
1510 	}
1511 #endif
1512 
1513 	/* For named pipes ask the underlying filesystem. */
1514 	if (pipe->pipe_state & PIPE_NAMED) {
1515 		PIPE_UNLOCK(pipe);
1516 		return (vnops.fo_stat(fp, ub, active_cred, td));
1517 	}
1518 
1519 	/*
1520 	 * Lazily allocate an inode number for the pipe.  Most pipe
1521 	 * users do not call fstat(2) on the pipe, which means that
1522 	 * postponing the inode allocation until it is must be
1523 	 * returned to userland is useful.  If alloc_unr failed,
1524 	 * assign st_ino zero instead of returning an error.
1525 	 * Special pipe_ino values:
1526 	 *  -1 - not yet initialized;
1527 	 *  0  - alloc_unr failed, return 0 as st_ino forever.
1528 	 */
1529 	if (pipe->pipe_ino == (ino_t)-1) {
1530 		new_unr = alloc_unr(pipeino_unr);
1531 		if (new_unr != -1)
1532 			pipe->pipe_ino = new_unr;
1533 		else
1534 			pipe->pipe_ino = 0;
1535 	}
1536 	PIPE_UNLOCK(pipe);
1537 
1538 	bzero(ub, sizeof(*ub));
1539 	ub->st_mode = S_IFIFO;
1540 	ub->st_blksize = PAGE_SIZE;
1541 	if (pipe->pipe_state & PIPE_DIRECTW)
1542 		ub->st_size = pipe->pipe_map.cnt;
1543 	else
1544 		ub->st_size = pipe->pipe_buffer.cnt;
1545 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1546 	ub->st_atim = pipe->pipe_atime;
1547 	ub->st_mtim = pipe->pipe_mtime;
1548 	ub->st_ctim = pipe->pipe_ctime;
1549 	ub->st_uid = fp->f_cred->cr_uid;
1550 	ub->st_gid = fp->f_cred->cr_gid;
1551 	ub->st_dev = pipedev_ino;
1552 	ub->st_ino = pipe->pipe_ino;
1553 	/*
1554 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1555 	 */
1556 	return (0);
1557 }
1558 
1559 /* ARGSUSED */
1560 static int
1561 pipe_close(fp, td)
1562 	struct file *fp;
1563 	struct thread *td;
1564 {
1565 
1566 	if (fp->f_vnode != NULL)
1567 		return vnops.fo_close(fp, td);
1568 	fp->f_ops = &badfileops;
1569 	pipe_dtor(fp->f_data);
1570 	fp->f_data = NULL;
1571 	return (0);
1572 }
1573 
1574 static int
1575 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1576 {
1577 	struct pipe *cpipe;
1578 	int error;
1579 
1580 	cpipe = fp->f_data;
1581 	if (cpipe->pipe_state & PIPE_NAMED)
1582 		error = vn_chmod(fp, mode, active_cred, td);
1583 	else
1584 		error = invfo_chmod(fp, mode, active_cred, td);
1585 	return (error);
1586 }
1587 
1588 static int
1589 pipe_chown(fp, uid, gid, active_cred, td)
1590 	struct file *fp;
1591 	uid_t uid;
1592 	gid_t gid;
1593 	struct ucred *active_cred;
1594 	struct thread *td;
1595 {
1596 	struct pipe *cpipe;
1597 	int error;
1598 
1599 	cpipe = fp->f_data;
1600 	if (cpipe->pipe_state & PIPE_NAMED)
1601 		error = vn_chown(fp, uid, gid, active_cred, td);
1602 	else
1603 		error = invfo_chown(fp, uid, gid, active_cred, td);
1604 	return (error);
1605 }
1606 
1607 static int
1608 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1609 {
1610 	struct pipe *pi;
1611 
1612 	if (fp->f_type == DTYPE_FIFO)
1613 		return (vn_fill_kinfo(fp, kif, fdp));
1614 	kif->kf_type = KF_TYPE_PIPE;
1615 	pi = fp->f_data;
1616 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1617 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1618 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1619 	return (0);
1620 }
1621 
1622 static void
1623 pipe_free_kmem(cpipe)
1624 	struct pipe *cpipe;
1625 {
1626 
1627 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1628 	    ("pipe_free_kmem: pipe mutex locked"));
1629 
1630 	if (cpipe->pipe_buffer.buffer != NULL) {
1631 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1632 		vm_map_remove(pipe_map,
1633 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1634 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1635 		cpipe->pipe_buffer.buffer = NULL;
1636 	}
1637 #ifndef PIPE_NODIRECT
1638 	{
1639 		cpipe->pipe_map.cnt = 0;
1640 		cpipe->pipe_map.pos = 0;
1641 		cpipe->pipe_map.npages = 0;
1642 	}
1643 #endif
1644 }
1645 
1646 /*
1647  * shutdown the pipe
1648  */
1649 static void
1650 pipeclose(cpipe)
1651 	struct pipe *cpipe;
1652 {
1653 	struct pipepair *pp;
1654 	struct pipe *ppipe;
1655 
1656 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1657 
1658 	PIPE_LOCK(cpipe);
1659 	pipelock(cpipe, 0);
1660 	pp = cpipe->pipe_pair;
1661 
1662 	pipeselwakeup(cpipe);
1663 
1664 	/*
1665 	 * If the other side is blocked, wake it up saying that
1666 	 * we want to close it down.
1667 	 */
1668 	cpipe->pipe_state |= PIPE_EOF;
1669 	while (cpipe->pipe_busy) {
1670 		wakeup(cpipe);
1671 		cpipe->pipe_state |= PIPE_WANT;
1672 		pipeunlock(cpipe);
1673 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1674 		pipelock(cpipe, 0);
1675 	}
1676 
1677 
1678 	/*
1679 	 * Disconnect from peer, if any.
1680 	 */
1681 	ppipe = cpipe->pipe_peer;
1682 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1683 		pipeselwakeup(ppipe);
1684 
1685 		ppipe->pipe_state |= PIPE_EOF;
1686 		wakeup(ppipe);
1687 		KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1688 	}
1689 
1690 	/*
1691 	 * Mark this endpoint as free.  Release kmem resources.  We
1692 	 * don't mark this endpoint as unused until we've finished
1693 	 * doing that, or the pipe might disappear out from under
1694 	 * us.
1695 	 */
1696 	PIPE_UNLOCK(cpipe);
1697 	pipe_free_kmem(cpipe);
1698 	PIPE_LOCK(cpipe);
1699 	cpipe->pipe_present = PIPE_CLOSING;
1700 	pipeunlock(cpipe);
1701 
1702 	/*
1703 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1704 	 * PIPE_FINALIZED, that allows other end to free the
1705 	 * pipe_pair, only after the knotes are completely dismantled.
1706 	 */
1707 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1708 	cpipe->pipe_present = PIPE_FINALIZED;
1709 	seldrain(&cpipe->pipe_sel);
1710 	knlist_destroy(&cpipe->pipe_sel.si_note);
1711 
1712 	/*
1713 	 * If both endpoints are now closed, release the memory for the
1714 	 * pipe pair.  If not, unlock.
1715 	 */
1716 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1717 		PIPE_UNLOCK(cpipe);
1718 #ifdef MAC
1719 		mac_pipe_destroy(pp);
1720 #endif
1721 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1722 	} else
1723 		PIPE_UNLOCK(cpipe);
1724 }
1725 
1726 /*ARGSUSED*/
1727 static int
1728 pipe_kqfilter(struct file *fp, struct knote *kn)
1729 {
1730 	struct pipe *cpipe;
1731 
1732 	/*
1733 	 * If a filter is requested that is not supported by this file
1734 	 * descriptor, don't return an error, but also don't ever generate an
1735 	 * event.
1736 	 */
1737 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1738 		kn->kn_fop = &pipe_nfiltops;
1739 		return (0);
1740 	}
1741 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1742 		kn->kn_fop = &pipe_nfiltops;
1743 		return (0);
1744 	}
1745 	cpipe = fp->f_data;
1746 	PIPE_LOCK(cpipe);
1747 	switch (kn->kn_filter) {
1748 	case EVFILT_READ:
1749 		kn->kn_fop = &pipe_rfiltops;
1750 		break;
1751 	case EVFILT_WRITE:
1752 		kn->kn_fop = &pipe_wfiltops;
1753 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1754 			/* other end of pipe has been closed */
1755 			PIPE_UNLOCK(cpipe);
1756 			return (EPIPE);
1757 		}
1758 		cpipe = PIPE_PEER(cpipe);
1759 		break;
1760 	default:
1761 		PIPE_UNLOCK(cpipe);
1762 		return (EINVAL);
1763 	}
1764 
1765 	kn->kn_hook = cpipe;
1766 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1767 	PIPE_UNLOCK(cpipe);
1768 	return (0);
1769 }
1770 
1771 static void
1772 filt_pipedetach(struct knote *kn)
1773 {
1774 	struct pipe *cpipe = kn->kn_hook;
1775 
1776 	PIPE_LOCK(cpipe);
1777 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1778 	PIPE_UNLOCK(cpipe);
1779 }
1780 
1781 /*ARGSUSED*/
1782 static int
1783 filt_piperead(struct knote *kn, long hint)
1784 {
1785 	struct pipe *rpipe = kn->kn_hook;
1786 	struct pipe *wpipe = rpipe->pipe_peer;
1787 	int ret;
1788 
1789 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1790 	kn->kn_data = rpipe->pipe_buffer.cnt;
1791 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1792 		kn->kn_data = rpipe->pipe_map.cnt;
1793 
1794 	if ((rpipe->pipe_state & PIPE_EOF) ||
1795 	    wpipe->pipe_present != PIPE_ACTIVE ||
1796 	    (wpipe->pipe_state & PIPE_EOF)) {
1797 		kn->kn_flags |= EV_EOF;
1798 		return (1);
1799 	}
1800 	ret = kn->kn_data > 0;
1801 	return ret;
1802 }
1803 
1804 /*ARGSUSED*/
1805 static int
1806 filt_pipewrite(struct knote *kn, long hint)
1807 {
1808 	struct pipe *wpipe;
1809 
1810 	wpipe = kn->kn_hook;
1811 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1812 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1813 	    (wpipe->pipe_state & PIPE_EOF)) {
1814 		kn->kn_data = 0;
1815 		kn->kn_flags |= EV_EOF;
1816 		return (1);
1817 	}
1818 	kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1819 	    (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1820 	if (wpipe->pipe_state & PIPE_DIRECTW)
1821 		kn->kn_data = 0;
1822 
1823 	return (kn->kn_data >= PIPE_BUF);
1824 }
1825 
1826 static void
1827 filt_pipedetach_notsup(struct knote *kn)
1828 {
1829 
1830 }
1831 
1832 static int
1833 filt_pipenotsup(struct knote *kn, long hint)
1834 {
1835 
1836 	return (0);
1837 }
1838