xref: /freebsd/sys/kern/sys_pipe.c (revision ff0ba87247820afbdfdc1b307c803f7923d0e4d3)
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_RECURSE);
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 	ino_t ino;
381 
382 	ino = dpipe->pipe_ino;
383 	funsetown(&dpipe->pipe_sigio);
384 	pipeclose(dpipe);
385 	if (dpipe->pipe_state & PIPE_NAMED) {
386 		dpipe = dpipe->pipe_peer;
387 		funsetown(&dpipe->pipe_sigio);
388 		pipeclose(dpipe);
389 	}
390 	if (ino != 0 && ino != (ino_t)-1)
391 		free_unr(pipeino_unr, ino);
392 }
393 
394 /*
395  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
396  * the zone pick up the pieces via pipeclose().
397  */
398 int
399 kern_pipe(struct thread *td, int fildes[2])
400 {
401 
402 	return (kern_pipe2(td, fildes, 0));
403 }
404 
405 int
406 kern_pipe2(struct thread *td, int fildes[2], int flags)
407 {
408 	struct filedesc *fdp;
409 	struct file *rf, *wf;
410 	struct pipe *rpipe, *wpipe;
411 	struct pipepair *pp;
412 	int fd, fflags, error;
413 
414 	fdp = td->td_proc->p_fd;
415 	pipe_paircreate(td, &pp);
416 	rpipe = &pp->pp_rpipe;
417 	wpipe = &pp->pp_wpipe;
418 	error = falloc(td, &rf, &fd, flags);
419 	if (error) {
420 		pipeclose(rpipe);
421 		pipeclose(wpipe);
422 		return (error);
423 	}
424 	/* An extra reference on `rf' has been held for us by falloc(). */
425 	fildes[0] = fd;
426 
427 	fflags = FREAD | FWRITE;
428 	if ((flags & O_NONBLOCK) != 0)
429 		fflags |= FNONBLOCK;
430 
431 	/*
432 	 * Warning: once we've gotten past allocation of the fd for the
433 	 * read-side, we can only drop the read side via fdrop() in order
434 	 * to avoid races against processes which manage to dup() the read
435 	 * side while we are blocked trying to allocate the write side.
436 	 */
437 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
438 	error = falloc(td, &wf, &fd, flags);
439 	if (error) {
440 		fdclose(fdp, rf, fildes[0], td);
441 		fdrop(rf, td);
442 		/* rpipe has been closed by fdrop(). */
443 		pipeclose(wpipe);
444 		return (error);
445 	}
446 	/* An extra reference on `wf' has been held for us by falloc(). */
447 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
448 	fdrop(wf, td);
449 	fildes[1] = fd;
450 	fdrop(rf, td);
451 
452 	return (0);
453 }
454 
455 /* ARGSUSED */
456 int
457 sys_pipe(struct thread *td, struct pipe_args *uap)
458 {
459 	int error;
460 	int fildes[2];
461 
462 	error = kern_pipe(td, fildes);
463 	if (error)
464 		return (error);
465 
466 	td->td_retval[0] = fildes[0];
467 	td->td_retval[1] = fildes[1];
468 
469 	return (0);
470 }
471 
472 int
473 sys_pipe2(struct thread *td, struct pipe2_args *uap)
474 {
475 	int error, fildes[2];
476 
477 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
478 		return (EINVAL);
479 	error = kern_pipe2(td, fildes, uap->flags);
480 	if (error)
481 		return (error);
482 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
483 	if (error) {
484 		(void)kern_close(td, fildes[0]);
485 		(void)kern_close(td, fildes[1]);
486 	}
487 	return (error);
488 }
489 
490 /*
491  * Allocate kva for pipe circular buffer, the space is pageable
492  * This routine will 'realloc' the size of a pipe safely, if it fails
493  * it will retain the old buffer.
494  * If it fails it will return ENOMEM.
495  */
496 static int
497 pipespace_new(cpipe, size)
498 	struct pipe *cpipe;
499 	int size;
500 {
501 	caddr_t buffer;
502 	int error, cnt, firstseg;
503 	static int curfail = 0;
504 	static struct timeval lastfail;
505 
506 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
507 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
508 		("pipespace: resize of direct writes not allowed"));
509 retry:
510 	cnt = cpipe->pipe_buffer.cnt;
511 	if (cnt > size)
512 		size = cnt;
513 
514 	size = round_page(size);
515 	buffer = (caddr_t) vm_map_min(pipe_map);
516 
517 	error = vm_map_find(pipe_map, NULL, 0,
518 		(vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
519 		VM_PROT_ALL, VM_PROT_ALL, 0);
520 	if (error != KERN_SUCCESS) {
521 		if ((cpipe->pipe_buffer.buffer == NULL) &&
522 			(size > SMALL_PIPE_SIZE)) {
523 			size = SMALL_PIPE_SIZE;
524 			pipefragretry++;
525 			goto retry;
526 		}
527 		if (cpipe->pipe_buffer.buffer == NULL) {
528 			pipeallocfail++;
529 			if (ppsratecheck(&lastfail, &curfail, 1))
530 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
531 		} else {
532 			piperesizefail++;
533 		}
534 		return (ENOMEM);
535 	}
536 
537 	/* copy data, then free old resources if we're resizing */
538 	if (cnt > 0) {
539 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
540 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
541 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
542 				buffer, firstseg);
543 			if ((cnt - firstseg) > 0)
544 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
545 					cpipe->pipe_buffer.in);
546 		} else {
547 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
548 				buffer, cnt);
549 		}
550 	}
551 	pipe_free_kmem(cpipe);
552 	cpipe->pipe_buffer.buffer = buffer;
553 	cpipe->pipe_buffer.size = size;
554 	cpipe->pipe_buffer.in = cnt;
555 	cpipe->pipe_buffer.out = 0;
556 	cpipe->pipe_buffer.cnt = cnt;
557 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
558 	return (0);
559 }
560 
561 /*
562  * Wrapper for pipespace_new() that performs locking assertions.
563  */
564 static int
565 pipespace(cpipe, size)
566 	struct pipe *cpipe;
567 	int size;
568 {
569 
570 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
571 		("Unlocked pipe passed to pipespace"));
572 	return (pipespace_new(cpipe, size));
573 }
574 
575 /*
576  * lock a pipe for I/O, blocking other access
577  */
578 static __inline int
579 pipelock(cpipe, catch)
580 	struct pipe *cpipe;
581 	int catch;
582 {
583 	int error;
584 
585 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
586 	while (cpipe->pipe_state & PIPE_LOCKFL) {
587 		cpipe->pipe_state |= PIPE_LWANT;
588 		error = msleep(cpipe, PIPE_MTX(cpipe),
589 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
590 		    "pipelk", 0);
591 		if (error != 0)
592 			return (error);
593 	}
594 	cpipe->pipe_state |= PIPE_LOCKFL;
595 	return (0);
596 }
597 
598 /*
599  * unlock a pipe I/O lock
600  */
601 static __inline void
602 pipeunlock(cpipe)
603 	struct pipe *cpipe;
604 {
605 
606 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
607 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
608 		("Unlocked pipe passed to pipeunlock"));
609 	cpipe->pipe_state &= ~PIPE_LOCKFL;
610 	if (cpipe->pipe_state & PIPE_LWANT) {
611 		cpipe->pipe_state &= ~PIPE_LWANT;
612 		wakeup(cpipe);
613 	}
614 }
615 
616 void
617 pipeselwakeup(cpipe)
618 	struct pipe *cpipe;
619 {
620 
621 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
622 	if (cpipe->pipe_state & PIPE_SEL) {
623 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
624 		if (!SEL_WAITING(&cpipe->pipe_sel))
625 			cpipe->pipe_state &= ~PIPE_SEL;
626 	}
627 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
628 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
629 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
630 }
631 
632 /*
633  * Initialize and allocate VM and memory for pipe.  The structure
634  * will start out zero'd from the ctor, so we just manage the kmem.
635  */
636 static void
637 pipe_create(pipe, backing)
638 	struct pipe *pipe;
639 	int backing;
640 {
641 
642 	if (backing) {
643 		/*
644 		 * Note that these functions can fail if pipe map is exhausted
645 		 * (as a result of too many pipes created), but we ignore the
646 		 * error as it is not fatal and could be provoked by
647 		 * unprivileged users. The only consequence is worse performance
648 		 * with given pipe.
649 		 */
650 		if (amountpipekva > maxpipekva / 2)
651 			(void)pipespace_new(pipe, SMALL_PIPE_SIZE);
652 		else
653 			(void)pipespace_new(pipe, PIPE_SIZE);
654 	}
655 
656 	pipe->pipe_ino = -1;
657 }
658 
659 /* ARGSUSED */
660 static int
661 pipe_read(fp, uio, active_cred, flags, td)
662 	struct file *fp;
663 	struct uio *uio;
664 	struct ucred *active_cred;
665 	struct thread *td;
666 	int flags;
667 {
668 	struct pipe *rpipe;
669 	int error;
670 	int nread = 0;
671 	int size;
672 
673 	rpipe = fp->f_data;
674 	PIPE_LOCK(rpipe);
675 	++rpipe->pipe_busy;
676 	error = pipelock(rpipe, 1);
677 	if (error)
678 		goto unlocked_error;
679 
680 #ifdef MAC
681 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
682 	if (error)
683 		goto locked_error;
684 #endif
685 	if (amountpipekva > (3 * maxpipekva) / 4) {
686 		if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
687 			(rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
688 			(rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
689 			(piperesizeallowed == 1)) {
690 			PIPE_UNLOCK(rpipe);
691 			pipespace(rpipe, SMALL_PIPE_SIZE);
692 			PIPE_LOCK(rpipe);
693 		}
694 	}
695 
696 	while (uio->uio_resid) {
697 		/*
698 		 * normal pipe buffer receive
699 		 */
700 		if (rpipe->pipe_buffer.cnt > 0) {
701 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
702 			if (size > rpipe->pipe_buffer.cnt)
703 				size = rpipe->pipe_buffer.cnt;
704 			if (size > uio->uio_resid)
705 				size = uio->uio_resid;
706 
707 			PIPE_UNLOCK(rpipe);
708 			error = uiomove(
709 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
710 			    size, uio);
711 			PIPE_LOCK(rpipe);
712 			if (error)
713 				break;
714 
715 			rpipe->pipe_buffer.out += size;
716 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
717 				rpipe->pipe_buffer.out = 0;
718 
719 			rpipe->pipe_buffer.cnt -= size;
720 
721 			/*
722 			 * If there is no more to read in the pipe, reset
723 			 * its pointers to the beginning.  This improves
724 			 * cache hit stats.
725 			 */
726 			if (rpipe->pipe_buffer.cnt == 0) {
727 				rpipe->pipe_buffer.in = 0;
728 				rpipe->pipe_buffer.out = 0;
729 			}
730 			nread += size;
731 #ifndef PIPE_NODIRECT
732 		/*
733 		 * Direct copy, bypassing a kernel buffer.
734 		 */
735 		} else if ((size = rpipe->pipe_map.cnt) &&
736 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
737 			if (size > uio->uio_resid)
738 				size = (u_int) uio->uio_resid;
739 
740 			PIPE_UNLOCK(rpipe);
741 			error = uiomove_fromphys(rpipe->pipe_map.ms,
742 			    rpipe->pipe_map.pos, size, uio);
743 			PIPE_LOCK(rpipe);
744 			if (error)
745 				break;
746 			nread += size;
747 			rpipe->pipe_map.pos += size;
748 			rpipe->pipe_map.cnt -= size;
749 			if (rpipe->pipe_map.cnt == 0) {
750 				rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
751 				wakeup(rpipe);
752 			}
753 #endif
754 		} else {
755 			/*
756 			 * detect EOF condition
757 			 * read returns 0 on EOF, no need to set error
758 			 */
759 			if (rpipe->pipe_state & PIPE_EOF)
760 				break;
761 
762 			/*
763 			 * If the "write-side" has been blocked, wake it up now.
764 			 */
765 			if (rpipe->pipe_state & PIPE_WANTW) {
766 				rpipe->pipe_state &= ~PIPE_WANTW;
767 				wakeup(rpipe);
768 			}
769 
770 			/*
771 			 * Break if some data was read.
772 			 */
773 			if (nread > 0)
774 				break;
775 
776 			/*
777 			 * Unlock the pipe buffer for our remaining processing.
778 			 * We will either break out with an error or we will
779 			 * sleep and relock to loop.
780 			 */
781 			pipeunlock(rpipe);
782 
783 			/*
784 			 * Handle non-blocking mode operation or
785 			 * wait for more data.
786 			 */
787 			if (fp->f_flag & FNONBLOCK) {
788 				error = EAGAIN;
789 			} else {
790 				rpipe->pipe_state |= PIPE_WANTR;
791 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
792 				    PRIBIO | PCATCH,
793 				    "piperd", 0)) == 0)
794 					error = pipelock(rpipe, 1);
795 			}
796 			if (error)
797 				goto unlocked_error;
798 		}
799 	}
800 #ifdef MAC
801 locked_error:
802 #endif
803 	pipeunlock(rpipe);
804 
805 	/* XXX: should probably do this before getting any locks. */
806 	if (error == 0)
807 		vfs_timestamp(&rpipe->pipe_atime);
808 unlocked_error:
809 	--rpipe->pipe_busy;
810 
811 	/*
812 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
813 	 */
814 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
815 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
816 		wakeup(rpipe);
817 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
818 		/*
819 		 * Handle write blocking hysteresis.
820 		 */
821 		if (rpipe->pipe_state & PIPE_WANTW) {
822 			rpipe->pipe_state &= ~PIPE_WANTW;
823 			wakeup(rpipe);
824 		}
825 	}
826 
827 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
828 		pipeselwakeup(rpipe);
829 
830 	PIPE_UNLOCK(rpipe);
831 	return (error);
832 }
833 
834 #ifndef PIPE_NODIRECT
835 /*
836  * Map the sending processes' buffer into kernel space and wire it.
837  * This is similar to a physical write operation.
838  */
839 static int
840 pipe_build_write_buffer(wpipe, uio)
841 	struct pipe *wpipe;
842 	struct uio *uio;
843 {
844 	u_int size;
845 	int i;
846 
847 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
848 	KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
849 		("Clone attempt on non-direct write pipe!"));
850 
851 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
852                 size = wpipe->pipe_buffer.size;
853 	else
854                 size = uio->uio_iov->iov_len;
855 
856 	if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
857 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
858 	    wpipe->pipe_map.ms, PIPENPAGES)) < 0)
859 		return (EFAULT);
860 
861 /*
862  * set up the control block
863  */
864 	wpipe->pipe_map.npages = i;
865 	wpipe->pipe_map.pos =
866 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
867 	wpipe->pipe_map.cnt = size;
868 
869 /*
870  * and update the uio data
871  */
872 
873 	uio->uio_iov->iov_len -= size;
874 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
875 	if (uio->uio_iov->iov_len == 0)
876 		uio->uio_iov++;
877 	uio->uio_resid -= size;
878 	uio->uio_offset += size;
879 	return (0);
880 }
881 
882 /*
883  * unmap and unwire the process buffer
884  */
885 static void
886 pipe_destroy_write_buffer(wpipe)
887 	struct pipe *wpipe;
888 {
889 
890 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
891 	vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
892 	wpipe->pipe_map.npages = 0;
893 }
894 
895 /*
896  * In the case of a signal, the writing process might go away.  This
897  * code copies the data into the circular buffer so that the source
898  * pages can be freed without loss of data.
899  */
900 static void
901 pipe_clone_write_buffer(wpipe)
902 	struct pipe *wpipe;
903 {
904 	struct uio uio;
905 	struct iovec iov;
906 	int size;
907 	int pos;
908 
909 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
910 	size = wpipe->pipe_map.cnt;
911 	pos = wpipe->pipe_map.pos;
912 
913 	wpipe->pipe_buffer.in = size;
914 	wpipe->pipe_buffer.out = 0;
915 	wpipe->pipe_buffer.cnt = size;
916 	wpipe->pipe_state &= ~PIPE_DIRECTW;
917 
918 	PIPE_UNLOCK(wpipe);
919 	iov.iov_base = wpipe->pipe_buffer.buffer;
920 	iov.iov_len = size;
921 	uio.uio_iov = &iov;
922 	uio.uio_iovcnt = 1;
923 	uio.uio_offset = 0;
924 	uio.uio_resid = size;
925 	uio.uio_segflg = UIO_SYSSPACE;
926 	uio.uio_rw = UIO_READ;
927 	uio.uio_td = curthread;
928 	uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
929 	PIPE_LOCK(wpipe);
930 	pipe_destroy_write_buffer(wpipe);
931 }
932 
933 /*
934  * This implements the pipe buffer write mechanism.  Note that only
935  * a direct write OR a normal pipe write can be pending at any given time.
936  * If there are any characters in the pipe buffer, the direct write will
937  * be deferred until the receiving process grabs all of the bytes from
938  * the pipe buffer.  Then the direct mapping write is set-up.
939  */
940 static int
941 pipe_direct_write(wpipe, uio)
942 	struct pipe *wpipe;
943 	struct uio *uio;
944 {
945 	int error;
946 
947 retry:
948 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
949 	error = pipelock(wpipe, 1);
950 	if (wpipe->pipe_state & PIPE_EOF)
951 		error = EPIPE;
952 	if (error) {
953 		pipeunlock(wpipe);
954 		goto error1;
955 	}
956 	while (wpipe->pipe_state & PIPE_DIRECTW) {
957 		if (wpipe->pipe_state & PIPE_WANTR) {
958 			wpipe->pipe_state &= ~PIPE_WANTR;
959 			wakeup(wpipe);
960 		}
961 		pipeselwakeup(wpipe);
962 		wpipe->pipe_state |= PIPE_WANTW;
963 		pipeunlock(wpipe);
964 		error = msleep(wpipe, PIPE_MTX(wpipe),
965 		    PRIBIO | PCATCH, "pipdww", 0);
966 		if (error)
967 			goto error1;
968 		else
969 			goto retry;
970 	}
971 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
972 	if (wpipe->pipe_buffer.cnt > 0) {
973 		if (wpipe->pipe_state & PIPE_WANTR) {
974 			wpipe->pipe_state &= ~PIPE_WANTR;
975 			wakeup(wpipe);
976 		}
977 		pipeselwakeup(wpipe);
978 		wpipe->pipe_state |= PIPE_WANTW;
979 		pipeunlock(wpipe);
980 		error = msleep(wpipe, PIPE_MTX(wpipe),
981 		    PRIBIO | PCATCH, "pipdwc", 0);
982 		if (error)
983 			goto error1;
984 		else
985 			goto retry;
986 	}
987 
988 	wpipe->pipe_state |= PIPE_DIRECTW;
989 
990 	PIPE_UNLOCK(wpipe);
991 	error = pipe_build_write_buffer(wpipe, uio);
992 	PIPE_LOCK(wpipe);
993 	if (error) {
994 		wpipe->pipe_state &= ~PIPE_DIRECTW;
995 		pipeunlock(wpipe);
996 		goto error1;
997 	}
998 
999 	error = 0;
1000 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1001 		if (wpipe->pipe_state & PIPE_EOF) {
1002 			pipe_destroy_write_buffer(wpipe);
1003 			pipeselwakeup(wpipe);
1004 			pipeunlock(wpipe);
1005 			error = EPIPE;
1006 			goto error1;
1007 		}
1008 		if (wpipe->pipe_state & PIPE_WANTR) {
1009 			wpipe->pipe_state &= ~PIPE_WANTR;
1010 			wakeup(wpipe);
1011 		}
1012 		pipeselwakeup(wpipe);
1013 		wpipe->pipe_state |= PIPE_WANTW;
1014 		pipeunlock(wpipe);
1015 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1016 		    "pipdwt", 0);
1017 		pipelock(wpipe, 0);
1018 	}
1019 
1020 	if (wpipe->pipe_state & PIPE_EOF)
1021 		error = EPIPE;
1022 	if (wpipe->pipe_state & PIPE_DIRECTW) {
1023 		/*
1024 		 * this bit of trickery substitutes a kernel buffer for
1025 		 * the process that might be going away.
1026 		 */
1027 		pipe_clone_write_buffer(wpipe);
1028 	} else {
1029 		pipe_destroy_write_buffer(wpipe);
1030 	}
1031 	pipeunlock(wpipe);
1032 	return (error);
1033 
1034 error1:
1035 	wakeup(wpipe);
1036 	return (error);
1037 }
1038 #endif
1039 
1040 static int
1041 pipe_write(fp, uio, active_cred, flags, td)
1042 	struct file *fp;
1043 	struct uio *uio;
1044 	struct ucred *active_cred;
1045 	struct thread *td;
1046 	int flags;
1047 {
1048 	int error = 0;
1049 	int desiredsize;
1050 	ssize_t orig_resid;
1051 	struct pipe *wpipe, *rpipe;
1052 
1053 	rpipe = fp->f_data;
1054 	wpipe = PIPE_PEER(rpipe);
1055 	PIPE_LOCK(rpipe);
1056 	error = pipelock(wpipe, 1);
1057 	if (error) {
1058 		PIPE_UNLOCK(rpipe);
1059 		return (error);
1060 	}
1061 	/*
1062 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1063 	 */
1064 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1065 	    (wpipe->pipe_state & PIPE_EOF)) {
1066 		pipeunlock(wpipe);
1067 		PIPE_UNLOCK(rpipe);
1068 		return (EPIPE);
1069 	}
1070 #ifdef MAC
1071 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1072 	if (error) {
1073 		pipeunlock(wpipe);
1074 		PIPE_UNLOCK(rpipe);
1075 		return (error);
1076 	}
1077 #endif
1078 	++wpipe->pipe_busy;
1079 
1080 	/* Choose a larger size if it's advantageous */
1081 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1082 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1083 		if (piperesizeallowed != 1)
1084 			break;
1085 		if (amountpipekva > maxpipekva / 2)
1086 			break;
1087 		if (desiredsize == BIG_PIPE_SIZE)
1088 			break;
1089 		desiredsize = desiredsize * 2;
1090 	}
1091 
1092 	/* Choose a smaller size if we're in a OOM situation */
1093 	if ((amountpipekva > (3 * maxpipekva) / 4) &&
1094 		(wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1095 		(wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1096 		(piperesizeallowed == 1))
1097 		desiredsize = SMALL_PIPE_SIZE;
1098 
1099 	/* Resize if the above determined that a new size was necessary */
1100 	if ((desiredsize != wpipe->pipe_buffer.size) &&
1101 		((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1102 		PIPE_UNLOCK(wpipe);
1103 		pipespace(wpipe, desiredsize);
1104 		PIPE_LOCK(wpipe);
1105 	}
1106 	if (wpipe->pipe_buffer.size == 0) {
1107 		/*
1108 		 * This can only happen for reverse direction use of pipes
1109 		 * in a complete OOM situation.
1110 		 */
1111 		error = ENOMEM;
1112 		--wpipe->pipe_busy;
1113 		pipeunlock(wpipe);
1114 		PIPE_UNLOCK(wpipe);
1115 		return (error);
1116 	}
1117 
1118 	pipeunlock(wpipe);
1119 
1120 	orig_resid = uio->uio_resid;
1121 
1122 	while (uio->uio_resid) {
1123 		int space;
1124 
1125 		pipelock(wpipe, 0);
1126 		if (wpipe->pipe_state & PIPE_EOF) {
1127 			pipeunlock(wpipe);
1128 			error = EPIPE;
1129 			break;
1130 		}
1131 #ifndef PIPE_NODIRECT
1132 		/*
1133 		 * If the transfer is large, we can gain performance if
1134 		 * we do process-to-process copies directly.
1135 		 * If the write is non-blocking, we don't use the
1136 		 * direct write mechanism.
1137 		 *
1138 		 * The direct write mechanism will detect the reader going
1139 		 * away on us.
1140 		 */
1141 		if (uio->uio_segflg == UIO_USERSPACE &&
1142 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1143 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1144 		    (fp->f_flag & FNONBLOCK) == 0) {
1145 			pipeunlock(wpipe);
1146 			error = pipe_direct_write(wpipe, uio);
1147 			if (error)
1148 				break;
1149 			continue;
1150 		}
1151 #endif
1152 
1153 		/*
1154 		 * Pipe buffered writes cannot be coincidental with
1155 		 * direct writes.  We wait until the currently executing
1156 		 * direct write is completed before we start filling the
1157 		 * pipe buffer.  We break out if a signal occurs or the
1158 		 * reader goes away.
1159 		 */
1160 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1161 			if (wpipe->pipe_state & PIPE_WANTR) {
1162 				wpipe->pipe_state &= ~PIPE_WANTR;
1163 				wakeup(wpipe);
1164 			}
1165 			pipeselwakeup(wpipe);
1166 			wpipe->pipe_state |= PIPE_WANTW;
1167 			pipeunlock(wpipe);
1168 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1169 			    "pipbww", 0);
1170 			if (error)
1171 				break;
1172 			else
1173 				continue;
1174 		}
1175 
1176 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1177 
1178 		/* Writes of size <= PIPE_BUF must be atomic. */
1179 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1180 			space = 0;
1181 
1182 		if (space > 0) {
1183 			int size;	/* Transfer size */
1184 			int segsize;	/* first segment to transfer */
1185 
1186 			/*
1187 			 * Transfer size is minimum of uio transfer
1188 			 * and free space in pipe buffer.
1189 			 */
1190 			if (space > uio->uio_resid)
1191 				size = uio->uio_resid;
1192 			else
1193 				size = space;
1194 			/*
1195 			 * First segment to transfer is minimum of
1196 			 * transfer size and contiguous space in
1197 			 * pipe buffer.  If first segment to transfer
1198 			 * is less than the transfer size, we've got
1199 			 * a wraparound in the buffer.
1200 			 */
1201 			segsize = wpipe->pipe_buffer.size -
1202 				wpipe->pipe_buffer.in;
1203 			if (segsize > size)
1204 				segsize = size;
1205 
1206 			/* Transfer first segment */
1207 
1208 			PIPE_UNLOCK(rpipe);
1209 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1210 					segsize, uio);
1211 			PIPE_LOCK(rpipe);
1212 
1213 			if (error == 0 && segsize < size) {
1214 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1215 					wpipe->pipe_buffer.size,
1216 					("Pipe buffer wraparound disappeared"));
1217 				/*
1218 				 * Transfer remaining part now, to
1219 				 * support atomic writes.  Wraparound
1220 				 * happened.
1221 				 */
1222 
1223 				PIPE_UNLOCK(rpipe);
1224 				error = uiomove(
1225 				    &wpipe->pipe_buffer.buffer[0],
1226 				    size - segsize, uio);
1227 				PIPE_LOCK(rpipe);
1228 			}
1229 			if (error == 0) {
1230 				wpipe->pipe_buffer.in += size;
1231 				if (wpipe->pipe_buffer.in >=
1232 				    wpipe->pipe_buffer.size) {
1233 					KASSERT(wpipe->pipe_buffer.in ==
1234 						size - segsize +
1235 						wpipe->pipe_buffer.size,
1236 						("Expected wraparound bad"));
1237 					wpipe->pipe_buffer.in = size - segsize;
1238 				}
1239 
1240 				wpipe->pipe_buffer.cnt += size;
1241 				KASSERT(wpipe->pipe_buffer.cnt <=
1242 					wpipe->pipe_buffer.size,
1243 					("Pipe buffer overflow"));
1244 			}
1245 			pipeunlock(wpipe);
1246 			if (error != 0)
1247 				break;
1248 		} else {
1249 			/*
1250 			 * If the "read-side" has been blocked, wake it up now.
1251 			 */
1252 			if (wpipe->pipe_state & PIPE_WANTR) {
1253 				wpipe->pipe_state &= ~PIPE_WANTR;
1254 				wakeup(wpipe);
1255 			}
1256 
1257 			/*
1258 			 * don't block on non-blocking I/O
1259 			 */
1260 			if (fp->f_flag & FNONBLOCK) {
1261 				error = EAGAIN;
1262 				pipeunlock(wpipe);
1263 				break;
1264 			}
1265 
1266 			/*
1267 			 * We have no more space and have something to offer,
1268 			 * wake up select/poll.
1269 			 */
1270 			pipeselwakeup(wpipe);
1271 
1272 			wpipe->pipe_state |= PIPE_WANTW;
1273 			pipeunlock(wpipe);
1274 			error = msleep(wpipe, PIPE_MTX(rpipe),
1275 			    PRIBIO | PCATCH, "pipewr", 0);
1276 			if (error != 0)
1277 				break;
1278 		}
1279 	}
1280 
1281 	pipelock(wpipe, 0);
1282 	--wpipe->pipe_busy;
1283 
1284 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1285 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1286 		wakeup(wpipe);
1287 	} else if (wpipe->pipe_buffer.cnt > 0) {
1288 		/*
1289 		 * If we have put any characters in the buffer, we wake up
1290 		 * the reader.
1291 		 */
1292 		if (wpipe->pipe_state & PIPE_WANTR) {
1293 			wpipe->pipe_state &= ~PIPE_WANTR;
1294 			wakeup(wpipe);
1295 		}
1296 	}
1297 
1298 	/*
1299 	 * Don't return EPIPE if any byte was written.
1300 	 * EINTR and other interrupts are handled by generic I/O layer.
1301 	 * Do not pretend that I/O succeeded for obvious user error
1302 	 * like EFAULT.
1303 	 */
1304 	if (uio->uio_resid != orig_resid && error == EPIPE)
1305 		error = 0;
1306 
1307 	if (error == 0)
1308 		vfs_timestamp(&wpipe->pipe_mtime);
1309 
1310 	/*
1311 	 * We have something to offer,
1312 	 * wake up select/poll.
1313 	 */
1314 	if (wpipe->pipe_buffer.cnt)
1315 		pipeselwakeup(wpipe);
1316 
1317 	pipeunlock(wpipe);
1318 	PIPE_UNLOCK(rpipe);
1319 	return (error);
1320 }
1321 
1322 /* ARGSUSED */
1323 static int
1324 pipe_truncate(fp, length, active_cred, td)
1325 	struct file *fp;
1326 	off_t length;
1327 	struct ucred *active_cred;
1328 	struct thread *td;
1329 {
1330 	struct pipe *cpipe;
1331 	int error;
1332 
1333 	cpipe = fp->f_data;
1334 	if (cpipe->pipe_state & PIPE_NAMED)
1335 		error = vnops.fo_truncate(fp, length, active_cred, td);
1336 	else
1337 		error = invfo_truncate(fp, length, active_cred, td);
1338 	return (error);
1339 }
1340 
1341 /*
1342  * we implement a very minimal set of ioctls for compatibility with sockets.
1343  */
1344 static int
1345 pipe_ioctl(fp, cmd, data, active_cred, td)
1346 	struct file *fp;
1347 	u_long cmd;
1348 	void *data;
1349 	struct ucred *active_cred;
1350 	struct thread *td;
1351 {
1352 	struct pipe *mpipe = fp->f_data;
1353 	int error;
1354 
1355 	PIPE_LOCK(mpipe);
1356 
1357 #ifdef MAC
1358 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1359 	if (error) {
1360 		PIPE_UNLOCK(mpipe);
1361 		return (error);
1362 	}
1363 #endif
1364 
1365 	error = 0;
1366 	switch (cmd) {
1367 
1368 	case FIONBIO:
1369 		break;
1370 
1371 	case FIOASYNC:
1372 		if (*(int *)data) {
1373 			mpipe->pipe_state |= PIPE_ASYNC;
1374 		} else {
1375 			mpipe->pipe_state &= ~PIPE_ASYNC;
1376 		}
1377 		break;
1378 
1379 	case FIONREAD:
1380 		if (!(fp->f_flag & FREAD)) {
1381 			*(int *)data = 0;
1382 			PIPE_UNLOCK(mpipe);
1383 			return (0);
1384 		}
1385 		if (mpipe->pipe_state & PIPE_DIRECTW)
1386 			*(int *)data = mpipe->pipe_map.cnt;
1387 		else
1388 			*(int *)data = mpipe->pipe_buffer.cnt;
1389 		break;
1390 
1391 	case FIOSETOWN:
1392 		PIPE_UNLOCK(mpipe);
1393 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1394 		goto out_unlocked;
1395 
1396 	case FIOGETOWN:
1397 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1398 		break;
1399 
1400 	/* This is deprecated, FIOSETOWN should be used instead. */
1401 	case TIOCSPGRP:
1402 		PIPE_UNLOCK(mpipe);
1403 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1404 		goto out_unlocked;
1405 
1406 	/* This is deprecated, FIOGETOWN should be used instead. */
1407 	case TIOCGPGRP:
1408 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1409 		break;
1410 
1411 	default:
1412 		error = ENOTTY;
1413 		break;
1414 	}
1415 	PIPE_UNLOCK(mpipe);
1416 out_unlocked:
1417 	return (error);
1418 }
1419 
1420 static int
1421 pipe_poll(fp, events, active_cred, td)
1422 	struct file *fp;
1423 	int events;
1424 	struct ucred *active_cred;
1425 	struct thread *td;
1426 {
1427 	struct pipe *rpipe;
1428 	struct pipe *wpipe;
1429 	int levents, revents;
1430 #ifdef MAC
1431 	int error;
1432 #endif
1433 
1434 	revents = 0;
1435 	rpipe = fp->f_data;
1436 	wpipe = PIPE_PEER(rpipe);
1437 	PIPE_LOCK(rpipe);
1438 #ifdef MAC
1439 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1440 	if (error)
1441 		goto locked_error;
1442 #endif
1443 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1444 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1445 		    (rpipe->pipe_buffer.cnt > 0))
1446 			revents |= events & (POLLIN | POLLRDNORM);
1447 
1448 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1449 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1450 		    (wpipe->pipe_state & PIPE_EOF) ||
1451 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1452 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1453 			 wpipe->pipe_buffer.size == 0)))
1454 			revents |= events & (POLLOUT | POLLWRNORM);
1455 
1456 	levents = events &
1457 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1458 	if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1459 	    fp->f_seqcount == rpipe->pipe_wgen)
1460 		events |= POLLINIGNEOF;
1461 
1462 	if ((events & POLLINIGNEOF) == 0) {
1463 		if (rpipe->pipe_state & PIPE_EOF) {
1464 			revents |= (events & (POLLIN | POLLRDNORM));
1465 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1466 			    (wpipe->pipe_state & PIPE_EOF))
1467 				revents |= POLLHUP;
1468 		}
1469 	}
1470 
1471 	if (revents == 0) {
1472 		if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1473 			selrecord(td, &rpipe->pipe_sel);
1474 			if (SEL_WAITING(&rpipe->pipe_sel))
1475 				rpipe->pipe_state |= PIPE_SEL;
1476 		}
1477 
1478 		if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1479 			selrecord(td, &wpipe->pipe_sel);
1480 			if (SEL_WAITING(&wpipe->pipe_sel))
1481 				wpipe->pipe_state |= PIPE_SEL;
1482 		}
1483 	}
1484 #ifdef MAC
1485 locked_error:
1486 #endif
1487 	PIPE_UNLOCK(rpipe);
1488 
1489 	return (revents);
1490 }
1491 
1492 /*
1493  * We shouldn't need locks here as we're doing a read and this should
1494  * be a natural race.
1495  */
1496 static int
1497 pipe_stat(fp, ub, active_cred, td)
1498 	struct file *fp;
1499 	struct stat *ub;
1500 	struct ucred *active_cred;
1501 	struct thread *td;
1502 {
1503 	struct pipe *pipe;
1504 	int new_unr;
1505 #ifdef MAC
1506 	int error;
1507 #endif
1508 
1509 	pipe = fp->f_data;
1510 	PIPE_LOCK(pipe);
1511 #ifdef MAC
1512 	error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1513 	if (error) {
1514 		PIPE_UNLOCK(pipe);
1515 		return (error);
1516 	}
1517 #endif
1518 
1519 	/* For named pipes ask the underlying filesystem. */
1520 	if (pipe->pipe_state & PIPE_NAMED) {
1521 		PIPE_UNLOCK(pipe);
1522 		return (vnops.fo_stat(fp, ub, active_cred, td));
1523 	}
1524 
1525 	/*
1526 	 * Lazily allocate an inode number for the pipe.  Most pipe
1527 	 * users do not call fstat(2) on the pipe, which means that
1528 	 * postponing the inode allocation until it is must be
1529 	 * returned to userland is useful.  If alloc_unr failed,
1530 	 * assign st_ino zero instead of returning an error.
1531 	 * Special pipe_ino values:
1532 	 *  -1 - not yet initialized;
1533 	 *  0  - alloc_unr failed, return 0 as st_ino forever.
1534 	 */
1535 	if (pipe->pipe_ino == (ino_t)-1) {
1536 		new_unr = alloc_unr(pipeino_unr);
1537 		if (new_unr != -1)
1538 			pipe->pipe_ino = new_unr;
1539 		else
1540 			pipe->pipe_ino = 0;
1541 	}
1542 	PIPE_UNLOCK(pipe);
1543 
1544 	bzero(ub, sizeof(*ub));
1545 	ub->st_mode = S_IFIFO;
1546 	ub->st_blksize = PAGE_SIZE;
1547 	if (pipe->pipe_state & PIPE_DIRECTW)
1548 		ub->st_size = pipe->pipe_map.cnt;
1549 	else
1550 		ub->st_size = pipe->pipe_buffer.cnt;
1551 	ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1552 	ub->st_atim = pipe->pipe_atime;
1553 	ub->st_mtim = pipe->pipe_mtime;
1554 	ub->st_ctim = pipe->pipe_ctime;
1555 	ub->st_uid = fp->f_cred->cr_uid;
1556 	ub->st_gid = fp->f_cred->cr_gid;
1557 	ub->st_dev = pipedev_ino;
1558 	ub->st_ino = pipe->pipe_ino;
1559 	/*
1560 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1561 	 */
1562 	return (0);
1563 }
1564 
1565 /* ARGSUSED */
1566 static int
1567 pipe_close(fp, td)
1568 	struct file *fp;
1569 	struct thread *td;
1570 {
1571 
1572 	if (fp->f_vnode != NULL)
1573 		return vnops.fo_close(fp, td);
1574 	fp->f_ops = &badfileops;
1575 	pipe_dtor(fp->f_data);
1576 	fp->f_data = NULL;
1577 	return (0);
1578 }
1579 
1580 static int
1581 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1582 {
1583 	struct pipe *cpipe;
1584 	int error;
1585 
1586 	cpipe = fp->f_data;
1587 	if (cpipe->pipe_state & PIPE_NAMED)
1588 		error = vn_chmod(fp, mode, active_cred, td);
1589 	else
1590 		error = invfo_chmod(fp, mode, active_cred, td);
1591 	return (error);
1592 }
1593 
1594 static int
1595 pipe_chown(fp, uid, gid, active_cred, td)
1596 	struct file *fp;
1597 	uid_t uid;
1598 	gid_t gid;
1599 	struct ucred *active_cred;
1600 	struct thread *td;
1601 {
1602 	struct pipe *cpipe;
1603 	int error;
1604 
1605 	cpipe = fp->f_data;
1606 	if (cpipe->pipe_state & PIPE_NAMED)
1607 		error = vn_chown(fp, uid, gid, active_cred, td);
1608 	else
1609 		error = invfo_chown(fp, uid, gid, active_cred, td);
1610 	return (error);
1611 }
1612 
1613 static int
1614 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1615 {
1616 	struct pipe *pi;
1617 
1618 	if (fp->f_type == DTYPE_FIFO)
1619 		return (vn_fill_kinfo(fp, kif, fdp));
1620 	kif->kf_type = KF_TYPE_PIPE;
1621 	pi = fp->f_data;
1622 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1623 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1624 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1625 	return (0);
1626 }
1627 
1628 static void
1629 pipe_free_kmem(cpipe)
1630 	struct pipe *cpipe;
1631 {
1632 
1633 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1634 	    ("pipe_free_kmem: pipe mutex locked"));
1635 
1636 	if (cpipe->pipe_buffer.buffer != NULL) {
1637 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1638 		vm_map_remove(pipe_map,
1639 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1640 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1641 		cpipe->pipe_buffer.buffer = NULL;
1642 	}
1643 #ifndef PIPE_NODIRECT
1644 	{
1645 		cpipe->pipe_map.cnt = 0;
1646 		cpipe->pipe_map.pos = 0;
1647 		cpipe->pipe_map.npages = 0;
1648 	}
1649 #endif
1650 }
1651 
1652 /*
1653  * shutdown the pipe
1654  */
1655 static void
1656 pipeclose(cpipe)
1657 	struct pipe *cpipe;
1658 {
1659 	struct pipepair *pp;
1660 	struct pipe *ppipe;
1661 
1662 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1663 
1664 	PIPE_LOCK(cpipe);
1665 	pipelock(cpipe, 0);
1666 	pp = cpipe->pipe_pair;
1667 
1668 	pipeselwakeup(cpipe);
1669 
1670 	/*
1671 	 * If the other side is blocked, wake it up saying that
1672 	 * we want to close it down.
1673 	 */
1674 	cpipe->pipe_state |= PIPE_EOF;
1675 	while (cpipe->pipe_busy) {
1676 		wakeup(cpipe);
1677 		cpipe->pipe_state |= PIPE_WANT;
1678 		pipeunlock(cpipe);
1679 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1680 		pipelock(cpipe, 0);
1681 	}
1682 
1683 
1684 	/*
1685 	 * Disconnect from peer, if any.
1686 	 */
1687 	ppipe = cpipe->pipe_peer;
1688 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1689 		pipeselwakeup(ppipe);
1690 
1691 		ppipe->pipe_state |= PIPE_EOF;
1692 		wakeup(ppipe);
1693 		KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1694 	}
1695 
1696 	/*
1697 	 * Mark this endpoint as free.  Release kmem resources.  We
1698 	 * don't mark this endpoint as unused until we've finished
1699 	 * doing that, or the pipe might disappear out from under
1700 	 * us.
1701 	 */
1702 	PIPE_UNLOCK(cpipe);
1703 	pipe_free_kmem(cpipe);
1704 	PIPE_LOCK(cpipe);
1705 	cpipe->pipe_present = PIPE_CLOSING;
1706 	pipeunlock(cpipe);
1707 
1708 	/*
1709 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1710 	 * PIPE_FINALIZED, that allows other end to free the
1711 	 * pipe_pair, only after the knotes are completely dismantled.
1712 	 */
1713 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1714 	cpipe->pipe_present = PIPE_FINALIZED;
1715 	seldrain(&cpipe->pipe_sel);
1716 	knlist_destroy(&cpipe->pipe_sel.si_note);
1717 
1718 	/*
1719 	 * If both endpoints are now closed, release the memory for the
1720 	 * pipe pair.  If not, unlock.
1721 	 */
1722 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1723 		PIPE_UNLOCK(cpipe);
1724 #ifdef MAC
1725 		mac_pipe_destroy(pp);
1726 #endif
1727 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1728 	} else
1729 		PIPE_UNLOCK(cpipe);
1730 }
1731 
1732 /*ARGSUSED*/
1733 static int
1734 pipe_kqfilter(struct file *fp, struct knote *kn)
1735 {
1736 	struct pipe *cpipe;
1737 
1738 	/*
1739 	 * If a filter is requested that is not supported by this file
1740 	 * descriptor, don't return an error, but also don't ever generate an
1741 	 * event.
1742 	 */
1743 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1744 		kn->kn_fop = &pipe_nfiltops;
1745 		return (0);
1746 	}
1747 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1748 		kn->kn_fop = &pipe_nfiltops;
1749 		return (0);
1750 	}
1751 	cpipe = fp->f_data;
1752 	PIPE_LOCK(cpipe);
1753 	switch (kn->kn_filter) {
1754 	case EVFILT_READ:
1755 		kn->kn_fop = &pipe_rfiltops;
1756 		break;
1757 	case EVFILT_WRITE:
1758 		kn->kn_fop = &pipe_wfiltops;
1759 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1760 			/* other end of pipe has been closed */
1761 			PIPE_UNLOCK(cpipe);
1762 			return (EPIPE);
1763 		}
1764 		cpipe = PIPE_PEER(cpipe);
1765 		break;
1766 	default:
1767 		PIPE_UNLOCK(cpipe);
1768 		return (EINVAL);
1769 	}
1770 
1771 	kn->kn_hook = cpipe;
1772 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1773 	PIPE_UNLOCK(cpipe);
1774 	return (0);
1775 }
1776 
1777 static void
1778 filt_pipedetach(struct knote *kn)
1779 {
1780 	struct pipe *cpipe = kn->kn_hook;
1781 
1782 	PIPE_LOCK(cpipe);
1783 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1784 	PIPE_UNLOCK(cpipe);
1785 }
1786 
1787 /*ARGSUSED*/
1788 static int
1789 filt_piperead(struct knote *kn, long hint)
1790 {
1791 	struct pipe *rpipe = kn->kn_hook;
1792 	struct pipe *wpipe = rpipe->pipe_peer;
1793 	int ret;
1794 
1795 	PIPE_LOCK(rpipe);
1796 	kn->kn_data = rpipe->pipe_buffer.cnt;
1797 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1798 		kn->kn_data = rpipe->pipe_map.cnt;
1799 
1800 	if ((rpipe->pipe_state & PIPE_EOF) ||
1801 	    wpipe->pipe_present != PIPE_ACTIVE ||
1802 	    (wpipe->pipe_state & PIPE_EOF)) {
1803 		kn->kn_flags |= EV_EOF;
1804 		PIPE_UNLOCK(rpipe);
1805 		return (1);
1806 	}
1807 	ret = kn->kn_data > 0;
1808 	PIPE_UNLOCK(rpipe);
1809 	return ret;
1810 }
1811 
1812 /*ARGSUSED*/
1813 static int
1814 filt_pipewrite(struct knote *kn, long hint)
1815 {
1816 	struct pipe *wpipe;
1817 
1818 	wpipe = kn->kn_hook;
1819 	PIPE_LOCK(wpipe);
1820 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1821 	    (wpipe->pipe_state & PIPE_EOF)) {
1822 		kn->kn_data = 0;
1823 		kn->kn_flags |= EV_EOF;
1824 		PIPE_UNLOCK(wpipe);
1825 		return (1);
1826 	}
1827 	kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1828 	    (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1829 	if (wpipe->pipe_state & PIPE_DIRECTW)
1830 		kn->kn_data = 0;
1831 
1832 	PIPE_UNLOCK(wpipe);
1833 	return (kn->kn_data >= PIPE_BUF);
1834 }
1835 
1836 static void
1837 filt_pipedetach_notsup(struct knote *kn)
1838 {
1839 
1840 }
1841 
1842 static int
1843 filt_pipenotsup(struct knote *kn, long hint)
1844 {
1845 
1846 	return (0);
1847 }
1848