xref: /freebsd/sys/kern/sys_pipe.c (revision 0957b409)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1996 John S. Dyson
5  * Copyright (c) 2012 Giovanni Trematerra
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice immediately at the beginning of the file, without modification,
13  *    this list of conditions, and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Absolutely no warranty of function or purpose is made by the author
18  *    John S. Dyson.
19  * 4. Modifications may be freely made to this file if the above conditions
20  *    are met.
21  */
22 
23 /*
24  * This file contains a high-performance replacement for the socket-based
25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
26  * all features of sockets, but does do everything that pipes normally
27  * do.
28  */
29 
30 /*
31  * This code has two modes of operation, a small write mode and a large
32  * write mode.  The small write mode acts like conventional pipes with
33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
36  * memory, and the receiving process copies directly from these pinned pages
37  * in the sending process.
38  *
39  * If the sending process receives a signal, it is possible that it will
40  * go away, and certainly its address space can change, because control
41  * is returned back to the user-mode side.  In that case, the pipe code
42  * arranges to copy the buffer supplied by the user process, to a pageable
43  * kernel buffer, and the receiving process will grab the data from the
44  * pageable kernel buffer.  Since signals don't happen all that often,
45  * the copy operation is normally eliminated.
46  *
47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
48  * happen for small transfers so that the system will not spend all of
49  * its time context switching.
50  *
51  * In order to limit the resource use of pipes, two sysctls exist:
52  *
53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
54  * address space available to us in pipe_map. This value is normally
55  * autotuned, but may also be loader tuned.
56  *
57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
58  * memory in use by pipes.
59  *
60  * Based on how large pipekva is relative to maxpipekva, the following
61  * will happen:
62  *
63  * 0% - 50%:
64  *     New pipes are given 16K of memory backing, pipes may dynamically
65  *     grow to as large as 64K where needed.
66  * 50% - 75%:
67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
68  *     existing pipes may NOT grow.
69  * 75% - 100%:
70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
71  *     existing pipes will be shrunk down to 4K whenever possible.
72  *
73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
75  * resize which MUST occur for reverse-direction pipes when they are
76  * first used.
77  *
78  * Additional information about the current state of pipes may be obtained
79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
80  * and kern.ipc.piperesizefail.
81  *
82  * Locking rules:  There are two locks present here:  A mutex, used via
83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
84  * the flag, as mutexes can not persist over uiomove.  The mutex
85  * exists only to guard access to the flag, and is not in itself a
86  * locking mechanism.  Also note that there is only a single mutex for
87  * both directions of a pipe.
88  *
89  * As pipelock() may have to sleep before it can acquire the flag, it
90  * is important to reread all data after a call to pipelock(); everything
91  * in the structure may have changed.
92  */
93 
94 #include <sys/cdefs.h>
95 __FBSDID("$FreeBSD$");
96 
97 #include <sys/param.h>
98 #include <sys/systm.h>
99 #include <sys/conf.h>
100 #include <sys/fcntl.h>
101 #include <sys/file.h>
102 #include <sys/filedesc.h>
103 #include <sys/filio.h>
104 #include <sys/kernel.h>
105 #include <sys/lock.h>
106 #include <sys/mutex.h>
107 #include <sys/ttycom.h>
108 #include <sys/stat.h>
109 #include <sys/malloc.h>
110 #include <sys/poll.h>
111 #include <sys/selinfo.h>
112 #include <sys/signalvar.h>
113 #include <sys/syscallsubr.h>
114 #include <sys/sysctl.h>
115 #include <sys/sysproto.h>
116 #include <sys/pipe.h>
117 #include <sys/proc.h>
118 #include <sys/vnode.h>
119 #include <sys/uio.h>
120 #include <sys/user.h>
121 #include <sys/event.h>
122 
123 #include <security/mac/mac_framework.h>
124 
125 #include <vm/vm.h>
126 #include <vm/vm_param.h>
127 #include <vm/vm_object.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_extern.h>
130 #include <vm/pmap.h>
131 #include <vm/vm_map.h>
132 #include <vm/vm_page.h>
133 #include <vm/uma.h>
134 
135 /*
136  * Use this define if you want to disable *fancy* VM things.  Expect an
137  * approx 30% decrease in transfer rate.  This could be useful for
138  * NetBSD or OpenBSD.
139  */
140 /* #define PIPE_NODIRECT */
141 
142 #define PIPE_PEER(pipe)	\
143 	(((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
144 
145 /*
146  * interfaces to the outside world
147  */
148 static fo_rdwr_t	pipe_read;
149 static fo_rdwr_t	pipe_write;
150 static fo_truncate_t	pipe_truncate;
151 static fo_ioctl_t	pipe_ioctl;
152 static fo_poll_t	pipe_poll;
153 static fo_kqfilter_t	pipe_kqfilter;
154 static fo_stat_t	pipe_stat;
155 static fo_close_t	pipe_close;
156 static fo_chmod_t	pipe_chmod;
157 static fo_chown_t	pipe_chown;
158 static fo_fill_kinfo_t	pipe_fill_kinfo;
159 
160 struct fileops pipeops = {
161 	.fo_read = pipe_read,
162 	.fo_write = pipe_write,
163 	.fo_truncate = pipe_truncate,
164 	.fo_ioctl = pipe_ioctl,
165 	.fo_poll = pipe_poll,
166 	.fo_kqfilter = pipe_kqfilter,
167 	.fo_stat = pipe_stat,
168 	.fo_close = pipe_close,
169 	.fo_chmod = pipe_chmod,
170 	.fo_chown = pipe_chown,
171 	.fo_sendfile = invfo_sendfile,
172 	.fo_fill_kinfo = pipe_fill_kinfo,
173 	.fo_flags = DFLAG_PASSABLE
174 };
175 
176 static void	filt_pipedetach(struct knote *kn);
177 static void	filt_pipedetach_notsup(struct knote *kn);
178 static int	filt_pipenotsup(struct knote *kn, long hint);
179 static int	filt_piperead(struct knote *kn, long hint);
180 static int	filt_pipewrite(struct knote *kn, long hint);
181 
182 static struct filterops pipe_nfiltops = {
183 	.f_isfd = 1,
184 	.f_detach = filt_pipedetach_notsup,
185 	.f_event = filt_pipenotsup
186 };
187 static struct filterops pipe_rfiltops = {
188 	.f_isfd = 1,
189 	.f_detach = filt_pipedetach,
190 	.f_event = filt_piperead
191 };
192 static struct filterops pipe_wfiltops = {
193 	.f_isfd = 1,
194 	.f_detach = filt_pipedetach,
195 	.f_event = filt_pipewrite
196 };
197 
198 /*
199  * Default pipe buffer size(s), this can be kind-of large now because pipe
200  * space is pageable.  The pipe code will try to maintain locality of
201  * reference for performance reasons, so small amounts of outstanding I/O
202  * will not wipe the cache.
203  */
204 #define MINPIPESIZE (PIPE_SIZE/3)
205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
206 
207 static long amountpipekva;
208 static int pipefragretry;
209 static int pipeallocfail;
210 static int piperesizefail;
211 static int piperesizeallowed = 1;
212 
213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
214 	   &maxpipekva, 0, "Pipe KVA limit");
215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
216 	   &amountpipekva, 0, "Pipe KVA usage");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
218 	  &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
220 	  &pipeallocfail, 0, "Pipe allocation failures");
221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
222 	  &piperesizefail, 0, "Pipe resize failures");
223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
224 	  &piperesizeallowed, 0, "Pipe resizing allowed");
225 
226 static void pipeinit(void *dummy __unused);
227 static void pipeclose(struct pipe *cpipe);
228 static void pipe_free_kmem(struct pipe *cpipe);
229 static void pipe_create(struct pipe *pipe, int backing);
230 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
231 static __inline int pipelock(struct pipe *cpipe, int catch);
232 static __inline void pipeunlock(struct pipe *cpipe);
233 #ifndef PIPE_NODIRECT
234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
237 static void pipe_clone_write_buffer(struct pipe *wpipe);
238 #endif
239 static int pipespace(struct pipe *cpipe, int size);
240 static int pipespace_new(struct pipe *cpipe, int size);
241 
242 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
243 static int	pipe_zone_init(void *mem, int size, int flags);
244 static void	pipe_zone_fini(void *mem, int size);
245 
246 static uma_zone_t pipe_zone;
247 static struct unrhdr64 pipeino_unr;
248 static dev_t pipedev_ino;
249 
250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
251 
252 static void
253 pipeinit(void *dummy __unused)
254 {
255 
256 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
257 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
258 	    UMA_ALIGN_PTR, 0);
259 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
260 	new_unrhdr64(&pipeino_unr, 1);
261 	pipedev_ino = devfs_alloc_cdp_inode();
262 	KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
263 }
264 
265 static int
266 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
267 {
268 	struct pipepair *pp;
269 	struct pipe *rpipe, *wpipe;
270 
271 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
272 
273 	pp = (struct pipepair *)mem;
274 
275 	/*
276 	 * We zero both pipe endpoints to make sure all the kmem pointers
277 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
278 	 * endpoints with the same time.
279 	 */
280 	rpipe = &pp->pp_rpipe;
281 	bzero(rpipe, sizeof(*rpipe));
282 	vfs_timestamp(&rpipe->pipe_ctime);
283 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
284 
285 	wpipe = &pp->pp_wpipe;
286 	bzero(wpipe, sizeof(*wpipe));
287 	wpipe->pipe_ctime = rpipe->pipe_ctime;
288 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
289 
290 	rpipe->pipe_peer = wpipe;
291 	rpipe->pipe_pair = pp;
292 	wpipe->pipe_peer = rpipe;
293 	wpipe->pipe_pair = pp;
294 
295 	/*
296 	 * Mark both endpoints as present; they will later get free'd
297 	 * one at a time.  When both are free'd, then the whole pair
298 	 * is released.
299 	 */
300 	rpipe->pipe_present = PIPE_ACTIVE;
301 	wpipe->pipe_present = PIPE_ACTIVE;
302 
303 	/*
304 	 * Eventually, the MAC Framework may initialize the label
305 	 * in ctor or init, but for now we do it elswhere to avoid
306 	 * blocking in ctor or init.
307 	 */
308 	pp->pp_label = NULL;
309 
310 	return (0);
311 }
312 
313 static int
314 pipe_zone_init(void *mem, int size, int flags)
315 {
316 	struct pipepair *pp;
317 
318 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
319 
320 	pp = (struct pipepair *)mem;
321 
322 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
323 	return (0);
324 }
325 
326 static void
327 pipe_zone_fini(void *mem, int size)
328 {
329 	struct pipepair *pp;
330 
331 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
332 
333 	pp = (struct pipepair *)mem;
334 
335 	mtx_destroy(&pp->pp_mtx);
336 }
337 
338 static void
339 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
340 {
341 	struct pipepair *pp;
342 	struct pipe *rpipe, *wpipe;
343 
344 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
345 #ifdef MAC
346 	/*
347 	 * The MAC label is shared between the connected endpoints.  As a
348 	 * result mac_pipe_init() and mac_pipe_create() are called once
349 	 * for the pair, and not on the endpoints.
350 	 */
351 	mac_pipe_init(pp);
352 	mac_pipe_create(td->td_ucred, pp);
353 #endif
354 	rpipe = &pp->pp_rpipe;
355 	wpipe = &pp->pp_wpipe;
356 
357 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
358 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
359 
360 	/* Only the forward direction pipe is backed by default */
361 	pipe_create(rpipe, 1);
362 	pipe_create(wpipe, 0);
363 
364 	rpipe->pipe_state |= PIPE_DIRECTOK;
365 	wpipe->pipe_state |= PIPE_DIRECTOK;
366 }
367 
368 void
369 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
370 {
371 	struct pipepair *pp;
372 
373 	pipe_paircreate(td, &pp);
374 	pp->pp_rpipe.pipe_state |= PIPE_NAMED;
375 	*ppipe = &pp->pp_rpipe;
376 }
377 
378 void
379 pipe_dtor(struct pipe *dpipe)
380 {
381 	struct pipe *peer;
382 
383 	peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
384 	funsetown(&dpipe->pipe_sigio);
385 	pipeclose(dpipe);
386 	if (peer != NULL) {
387 		funsetown(&peer->pipe_sigio);
388 		pipeclose(peer);
389 	}
390 }
391 
392 /*
393  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
394  * the zone pick up the pieces via pipeclose().
395  */
396 int
397 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
398     struct filecaps *fcaps2)
399 {
400 	struct file *rf, *wf;
401 	struct pipe *rpipe, *wpipe;
402 	struct pipepair *pp;
403 	int fd, fflags, error;
404 
405 	pipe_paircreate(td, &pp);
406 	rpipe = &pp->pp_rpipe;
407 	wpipe = &pp->pp_wpipe;
408 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
409 	if (error) {
410 		pipeclose(rpipe);
411 		pipeclose(wpipe);
412 		return (error);
413 	}
414 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
415 	fildes[0] = fd;
416 
417 	fflags = FREAD | FWRITE;
418 	if ((flags & O_NONBLOCK) != 0)
419 		fflags |= FNONBLOCK;
420 
421 	/*
422 	 * Warning: once we've gotten past allocation of the fd for the
423 	 * read-side, we can only drop the read side via fdrop() in order
424 	 * to avoid races against processes which manage to dup() the read
425 	 * side while we are blocked trying to allocate the write side.
426 	 */
427 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
428 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
429 	if (error) {
430 		fdclose(td, rf, fildes[0]);
431 		fdrop(rf, td);
432 		/* rpipe has been closed by fdrop(). */
433 		pipeclose(wpipe);
434 		return (error);
435 	}
436 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
437 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
438 	fdrop(wf, td);
439 	fildes[1] = fd;
440 	fdrop(rf, td);
441 
442 	return (0);
443 }
444 
445 #ifdef COMPAT_FREEBSD10
446 /* ARGSUSED */
447 int
448 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
449 {
450 	int error;
451 	int fildes[2];
452 
453 	error = kern_pipe(td, fildes, 0, NULL, NULL);
454 	if (error)
455 		return (error);
456 
457 	td->td_retval[0] = fildes[0];
458 	td->td_retval[1] = fildes[1];
459 
460 	return (0);
461 }
462 #endif
463 
464 int
465 sys_pipe2(struct thread *td, struct pipe2_args *uap)
466 {
467 	int error, fildes[2];
468 
469 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
470 		return (EINVAL);
471 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
472 	if (error)
473 		return (error);
474 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
475 	if (error) {
476 		(void)kern_close(td, fildes[0]);
477 		(void)kern_close(td, fildes[1]);
478 	}
479 	return (error);
480 }
481 
482 /*
483  * Allocate kva for pipe circular buffer, the space is pageable
484  * This routine will 'realloc' the size of a pipe safely, if it fails
485  * it will retain the old buffer.
486  * If it fails it will return ENOMEM.
487  */
488 static int
489 pipespace_new(struct pipe *cpipe, int size)
490 {
491 	caddr_t buffer;
492 	int error, cnt, firstseg;
493 	static int curfail = 0;
494 	static struct timeval lastfail;
495 
496 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
497 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
498 		("pipespace: resize of direct writes not allowed"));
499 retry:
500 	cnt = cpipe->pipe_buffer.cnt;
501 	if (cnt > size)
502 		size = cnt;
503 
504 	size = round_page(size);
505 	buffer = (caddr_t) vm_map_min(pipe_map);
506 
507 	error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
508 	    VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
509 	if (error != KERN_SUCCESS) {
510 		if ((cpipe->pipe_buffer.buffer == NULL) &&
511 			(size > SMALL_PIPE_SIZE)) {
512 			size = SMALL_PIPE_SIZE;
513 			pipefragretry++;
514 			goto retry;
515 		}
516 		if (cpipe->pipe_buffer.buffer == NULL) {
517 			pipeallocfail++;
518 			if (ppsratecheck(&lastfail, &curfail, 1))
519 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
520 		} else {
521 			piperesizefail++;
522 		}
523 		return (ENOMEM);
524 	}
525 
526 	/* copy data, then free old resources if we're resizing */
527 	if (cnt > 0) {
528 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
529 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
530 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
531 				buffer, firstseg);
532 			if ((cnt - firstseg) > 0)
533 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
534 					cpipe->pipe_buffer.in);
535 		} else {
536 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
537 				buffer, cnt);
538 		}
539 	}
540 	pipe_free_kmem(cpipe);
541 	cpipe->pipe_buffer.buffer = buffer;
542 	cpipe->pipe_buffer.size = size;
543 	cpipe->pipe_buffer.in = cnt;
544 	cpipe->pipe_buffer.out = 0;
545 	cpipe->pipe_buffer.cnt = cnt;
546 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
547 	return (0);
548 }
549 
550 /*
551  * Wrapper for pipespace_new() that performs locking assertions.
552  */
553 static int
554 pipespace(struct pipe *cpipe, int size)
555 {
556 
557 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
558 		("Unlocked pipe passed to pipespace"));
559 	return (pipespace_new(cpipe, size));
560 }
561 
562 /*
563  * lock a pipe for I/O, blocking other access
564  */
565 static __inline int
566 pipelock(struct pipe *cpipe, int catch)
567 {
568 	int error;
569 
570 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
571 	while (cpipe->pipe_state & PIPE_LOCKFL) {
572 		cpipe->pipe_state |= PIPE_LWANT;
573 		error = msleep(cpipe, PIPE_MTX(cpipe),
574 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
575 		    "pipelk", 0);
576 		if (error != 0)
577 			return (error);
578 	}
579 	cpipe->pipe_state |= PIPE_LOCKFL;
580 	return (0);
581 }
582 
583 /*
584  * unlock a pipe I/O lock
585  */
586 static __inline void
587 pipeunlock(struct pipe *cpipe)
588 {
589 
590 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
591 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
592 		("Unlocked pipe passed to pipeunlock"));
593 	cpipe->pipe_state &= ~PIPE_LOCKFL;
594 	if (cpipe->pipe_state & PIPE_LWANT) {
595 		cpipe->pipe_state &= ~PIPE_LWANT;
596 		wakeup(cpipe);
597 	}
598 }
599 
600 void
601 pipeselwakeup(struct pipe *cpipe)
602 {
603 
604 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
605 	if (cpipe->pipe_state & PIPE_SEL) {
606 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
607 		if (!SEL_WAITING(&cpipe->pipe_sel))
608 			cpipe->pipe_state &= ~PIPE_SEL;
609 	}
610 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
611 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
612 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
613 }
614 
615 /*
616  * Initialize and allocate VM and memory for pipe.  The structure
617  * will start out zero'd from the ctor, so we just manage the kmem.
618  */
619 static void
620 pipe_create(struct pipe *pipe, int backing)
621 {
622 
623 	if (backing) {
624 		/*
625 		 * Note that these functions can fail if pipe map is exhausted
626 		 * (as a result of too many pipes created), but we ignore the
627 		 * error as it is not fatal and could be provoked by
628 		 * unprivileged users. The only consequence is worse performance
629 		 * with given pipe.
630 		 */
631 		if (amountpipekva > maxpipekva / 2)
632 			(void)pipespace_new(pipe, SMALL_PIPE_SIZE);
633 		else
634 			(void)pipespace_new(pipe, PIPE_SIZE);
635 	}
636 
637 	pipe->pipe_ino = alloc_unr64(&pipeino_unr);
638 }
639 
640 /* ARGSUSED */
641 static int
642 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
643     int flags, struct thread *td)
644 {
645 	struct pipe *rpipe;
646 	int error;
647 	int nread = 0;
648 	int size;
649 
650 	rpipe = fp->f_data;
651 	PIPE_LOCK(rpipe);
652 	++rpipe->pipe_busy;
653 	error = pipelock(rpipe, 1);
654 	if (error)
655 		goto unlocked_error;
656 
657 #ifdef MAC
658 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
659 	if (error)
660 		goto locked_error;
661 #endif
662 	if (amountpipekva > (3 * maxpipekva) / 4) {
663 		if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
664 			(rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
665 			(rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
666 			(piperesizeallowed == 1)) {
667 			PIPE_UNLOCK(rpipe);
668 			pipespace(rpipe, SMALL_PIPE_SIZE);
669 			PIPE_LOCK(rpipe);
670 		}
671 	}
672 
673 	while (uio->uio_resid) {
674 		/*
675 		 * normal pipe buffer receive
676 		 */
677 		if (rpipe->pipe_buffer.cnt > 0) {
678 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
679 			if (size > rpipe->pipe_buffer.cnt)
680 				size = rpipe->pipe_buffer.cnt;
681 			if (size > uio->uio_resid)
682 				size = uio->uio_resid;
683 
684 			PIPE_UNLOCK(rpipe);
685 			error = uiomove(
686 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
687 			    size, uio);
688 			PIPE_LOCK(rpipe);
689 			if (error)
690 				break;
691 
692 			rpipe->pipe_buffer.out += size;
693 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
694 				rpipe->pipe_buffer.out = 0;
695 
696 			rpipe->pipe_buffer.cnt -= size;
697 
698 			/*
699 			 * If there is no more to read in the pipe, reset
700 			 * its pointers to the beginning.  This improves
701 			 * cache hit stats.
702 			 */
703 			if (rpipe->pipe_buffer.cnt == 0) {
704 				rpipe->pipe_buffer.in = 0;
705 				rpipe->pipe_buffer.out = 0;
706 			}
707 			nread += size;
708 #ifndef PIPE_NODIRECT
709 		/*
710 		 * Direct copy, bypassing a kernel buffer.
711 		 */
712 		} else if ((size = rpipe->pipe_map.cnt) &&
713 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
714 			if (size > uio->uio_resid)
715 				size = (u_int) uio->uio_resid;
716 
717 			PIPE_UNLOCK(rpipe);
718 			error = uiomove_fromphys(rpipe->pipe_map.ms,
719 			    rpipe->pipe_map.pos, size, uio);
720 			PIPE_LOCK(rpipe);
721 			if (error)
722 				break;
723 			nread += size;
724 			rpipe->pipe_map.pos += size;
725 			rpipe->pipe_map.cnt -= size;
726 			if (rpipe->pipe_map.cnt == 0) {
727 				rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
728 				wakeup(rpipe);
729 			}
730 #endif
731 		} else {
732 			/*
733 			 * detect EOF condition
734 			 * read returns 0 on EOF, no need to set error
735 			 */
736 			if (rpipe->pipe_state & PIPE_EOF)
737 				break;
738 
739 			/*
740 			 * If the "write-side" has been blocked, wake it up now.
741 			 */
742 			if (rpipe->pipe_state & PIPE_WANTW) {
743 				rpipe->pipe_state &= ~PIPE_WANTW;
744 				wakeup(rpipe);
745 			}
746 
747 			/*
748 			 * Break if some data was read.
749 			 */
750 			if (nread > 0)
751 				break;
752 
753 			/*
754 			 * Unlock the pipe buffer for our remaining processing.
755 			 * We will either break out with an error or we will
756 			 * sleep and relock to loop.
757 			 */
758 			pipeunlock(rpipe);
759 
760 			/*
761 			 * Handle non-blocking mode operation or
762 			 * wait for more data.
763 			 */
764 			if (fp->f_flag & FNONBLOCK) {
765 				error = EAGAIN;
766 			} else {
767 				rpipe->pipe_state |= PIPE_WANTR;
768 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
769 				    PRIBIO | PCATCH,
770 				    "piperd", 0)) == 0)
771 					error = pipelock(rpipe, 1);
772 			}
773 			if (error)
774 				goto unlocked_error;
775 		}
776 	}
777 #ifdef MAC
778 locked_error:
779 #endif
780 	pipeunlock(rpipe);
781 
782 	/* XXX: should probably do this before getting any locks. */
783 	if (error == 0)
784 		vfs_timestamp(&rpipe->pipe_atime);
785 unlocked_error:
786 	--rpipe->pipe_busy;
787 
788 	/*
789 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
790 	 */
791 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
792 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
793 		wakeup(rpipe);
794 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
795 		/*
796 		 * Handle write blocking hysteresis.
797 		 */
798 		if (rpipe->pipe_state & PIPE_WANTW) {
799 			rpipe->pipe_state &= ~PIPE_WANTW;
800 			wakeup(rpipe);
801 		}
802 	}
803 
804 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
805 		pipeselwakeup(rpipe);
806 
807 	PIPE_UNLOCK(rpipe);
808 	return (error);
809 }
810 
811 #ifndef PIPE_NODIRECT
812 /*
813  * Map the sending processes' buffer into kernel space and wire it.
814  * This is similar to a physical write operation.
815  */
816 static int
817 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
818 {
819 	u_int size;
820 	int i;
821 
822 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
823 	KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
824 		("Clone attempt on non-direct write pipe!"));
825 
826 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
827                 size = wpipe->pipe_buffer.size;
828 	else
829                 size = uio->uio_iov->iov_len;
830 
831 	if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
832 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
833 	    wpipe->pipe_map.ms, PIPENPAGES)) < 0)
834 		return (EFAULT);
835 
836 /*
837  * set up the control block
838  */
839 	wpipe->pipe_map.npages = i;
840 	wpipe->pipe_map.pos =
841 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
842 	wpipe->pipe_map.cnt = size;
843 
844 /*
845  * and update the uio data
846  */
847 
848 	uio->uio_iov->iov_len -= size;
849 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
850 	if (uio->uio_iov->iov_len == 0)
851 		uio->uio_iov++;
852 	uio->uio_resid -= size;
853 	uio->uio_offset += size;
854 	return (0);
855 }
856 
857 /*
858  * unmap and unwire the process buffer
859  */
860 static void
861 pipe_destroy_write_buffer(struct pipe *wpipe)
862 {
863 
864 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
865 	vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
866 	wpipe->pipe_map.npages = 0;
867 }
868 
869 /*
870  * In the case of a signal, the writing process might go away.  This
871  * code copies the data into the circular buffer so that the source
872  * pages can be freed without loss of data.
873  */
874 static void
875 pipe_clone_write_buffer(struct pipe *wpipe)
876 {
877 	struct uio uio;
878 	struct iovec iov;
879 	int size;
880 	int pos;
881 
882 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
883 	size = wpipe->pipe_map.cnt;
884 	pos = wpipe->pipe_map.pos;
885 
886 	wpipe->pipe_buffer.in = size;
887 	wpipe->pipe_buffer.out = 0;
888 	wpipe->pipe_buffer.cnt = size;
889 	wpipe->pipe_state &= ~PIPE_DIRECTW;
890 
891 	PIPE_UNLOCK(wpipe);
892 	iov.iov_base = wpipe->pipe_buffer.buffer;
893 	iov.iov_len = size;
894 	uio.uio_iov = &iov;
895 	uio.uio_iovcnt = 1;
896 	uio.uio_offset = 0;
897 	uio.uio_resid = size;
898 	uio.uio_segflg = UIO_SYSSPACE;
899 	uio.uio_rw = UIO_READ;
900 	uio.uio_td = curthread;
901 	uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
902 	PIPE_LOCK(wpipe);
903 	pipe_destroy_write_buffer(wpipe);
904 }
905 
906 /*
907  * This implements the pipe buffer write mechanism.  Note that only
908  * a direct write OR a normal pipe write can be pending at any given time.
909  * If there are any characters in the pipe buffer, the direct write will
910  * be deferred until the receiving process grabs all of the bytes from
911  * the pipe buffer.  Then the direct mapping write is set-up.
912  */
913 static int
914 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
915 {
916 	int error;
917 
918 retry:
919 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
920 	error = pipelock(wpipe, 1);
921 	if (error != 0)
922 		goto error1;
923 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
924 		error = EPIPE;
925 		pipeunlock(wpipe);
926 		goto error1;
927 	}
928 	while (wpipe->pipe_state & PIPE_DIRECTW) {
929 		if (wpipe->pipe_state & PIPE_WANTR) {
930 			wpipe->pipe_state &= ~PIPE_WANTR;
931 			wakeup(wpipe);
932 		}
933 		pipeselwakeup(wpipe);
934 		wpipe->pipe_state |= PIPE_WANTW;
935 		pipeunlock(wpipe);
936 		error = msleep(wpipe, PIPE_MTX(wpipe),
937 		    PRIBIO | PCATCH, "pipdww", 0);
938 		if (error)
939 			goto error1;
940 		else
941 			goto retry;
942 	}
943 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
944 	if (wpipe->pipe_buffer.cnt > 0) {
945 		if (wpipe->pipe_state & PIPE_WANTR) {
946 			wpipe->pipe_state &= ~PIPE_WANTR;
947 			wakeup(wpipe);
948 		}
949 		pipeselwakeup(wpipe);
950 		wpipe->pipe_state |= PIPE_WANTW;
951 		pipeunlock(wpipe);
952 		error = msleep(wpipe, PIPE_MTX(wpipe),
953 		    PRIBIO | PCATCH, "pipdwc", 0);
954 		if (error)
955 			goto error1;
956 		else
957 			goto retry;
958 	}
959 
960 	wpipe->pipe_state |= PIPE_DIRECTW;
961 
962 	PIPE_UNLOCK(wpipe);
963 	error = pipe_build_write_buffer(wpipe, uio);
964 	PIPE_LOCK(wpipe);
965 	if (error) {
966 		wpipe->pipe_state &= ~PIPE_DIRECTW;
967 		pipeunlock(wpipe);
968 		goto error1;
969 	}
970 
971 	error = 0;
972 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
973 		if (wpipe->pipe_state & PIPE_EOF) {
974 			pipe_destroy_write_buffer(wpipe);
975 			pipeselwakeup(wpipe);
976 			pipeunlock(wpipe);
977 			error = EPIPE;
978 			goto error1;
979 		}
980 		if (wpipe->pipe_state & PIPE_WANTR) {
981 			wpipe->pipe_state &= ~PIPE_WANTR;
982 			wakeup(wpipe);
983 		}
984 		pipeselwakeup(wpipe);
985 		wpipe->pipe_state |= PIPE_WANTW;
986 		pipeunlock(wpipe);
987 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
988 		    "pipdwt", 0);
989 		pipelock(wpipe, 0);
990 	}
991 
992 	if (wpipe->pipe_state & PIPE_EOF)
993 		error = EPIPE;
994 	if (wpipe->pipe_state & PIPE_DIRECTW) {
995 		/*
996 		 * this bit of trickery substitutes a kernel buffer for
997 		 * the process that might be going away.
998 		 */
999 		pipe_clone_write_buffer(wpipe);
1000 	} else {
1001 		pipe_destroy_write_buffer(wpipe);
1002 	}
1003 	pipeunlock(wpipe);
1004 	return (error);
1005 
1006 error1:
1007 	wakeup(wpipe);
1008 	return (error);
1009 }
1010 #endif
1011 
1012 static int
1013 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1014     int flags, struct thread *td)
1015 {
1016 	int error = 0;
1017 	int desiredsize;
1018 	ssize_t orig_resid;
1019 	struct pipe *wpipe, *rpipe;
1020 
1021 	rpipe = fp->f_data;
1022 	wpipe = PIPE_PEER(rpipe);
1023 	PIPE_LOCK(rpipe);
1024 	error = pipelock(wpipe, 1);
1025 	if (error) {
1026 		PIPE_UNLOCK(rpipe);
1027 		return (error);
1028 	}
1029 	/*
1030 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1031 	 */
1032 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1033 	    (wpipe->pipe_state & PIPE_EOF)) {
1034 		pipeunlock(wpipe);
1035 		PIPE_UNLOCK(rpipe);
1036 		return (EPIPE);
1037 	}
1038 #ifdef MAC
1039 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1040 	if (error) {
1041 		pipeunlock(wpipe);
1042 		PIPE_UNLOCK(rpipe);
1043 		return (error);
1044 	}
1045 #endif
1046 	++wpipe->pipe_busy;
1047 
1048 	/* Choose a larger size if it's advantageous */
1049 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1050 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1051 		if (piperesizeallowed != 1)
1052 			break;
1053 		if (amountpipekva > maxpipekva / 2)
1054 			break;
1055 		if (desiredsize == BIG_PIPE_SIZE)
1056 			break;
1057 		desiredsize = desiredsize * 2;
1058 	}
1059 
1060 	/* Choose a smaller size if we're in a OOM situation */
1061 	if ((amountpipekva > (3 * maxpipekva) / 4) &&
1062 		(wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1063 		(wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1064 		(piperesizeallowed == 1))
1065 		desiredsize = SMALL_PIPE_SIZE;
1066 
1067 	/* Resize if the above determined that a new size was necessary */
1068 	if ((desiredsize != wpipe->pipe_buffer.size) &&
1069 		((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1070 		PIPE_UNLOCK(wpipe);
1071 		pipespace(wpipe, desiredsize);
1072 		PIPE_LOCK(wpipe);
1073 	}
1074 	if (wpipe->pipe_buffer.size == 0) {
1075 		/*
1076 		 * This can only happen for reverse direction use of pipes
1077 		 * in a complete OOM situation.
1078 		 */
1079 		error = ENOMEM;
1080 		--wpipe->pipe_busy;
1081 		pipeunlock(wpipe);
1082 		PIPE_UNLOCK(wpipe);
1083 		return (error);
1084 	}
1085 
1086 	pipeunlock(wpipe);
1087 
1088 	orig_resid = uio->uio_resid;
1089 
1090 	while (uio->uio_resid) {
1091 		int space;
1092 
1093 		pipelock(wpipe, 0);
1094 		if (wpipe->pipe_state & PIPE_EOF) {
1095 			pipeunlock(wpipe);
1096 			error = EPIPE;
1097 			break;
1098 		}
1099 #ifndef PIPE_NODIRECT
1100 		/*
1101 		 * If the transfer is large, we can gain performance if
1102 		 * we do process-to-process copies directly.
1103 		 * If the write is non-blocking, we don't use the
1104 		 * direct write mechanism.
1105 		 *
1106 		 * The direct write mechanism will detect the reader going
1107 		 * away on us.
1108 		 */
1109 		if (uio->uio_segflg == UIO_USERSPACE &&
1110 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1111 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1112 		    (fp->f_flag & FNONBLOCK) == 0) {
1113 			pipeunlock(wpipe);
1114 			error = pipe_direct_write(wpipe, uio);
1115 			if (error)
1116 				break;
1117 			continue;
1118 		}
1119 #endif
1120 
1121 		/*
1122 		 * Pipe buffered writes cannot be coincidental with
1123 		 * direct writes.  We wait until the currently executing
1124 		 * direct write is completed before we start filling the
1125 		 * pipe buffer.  We break out if a signal occurs or the
1126 		 * reader goes away.
1127 		 */
1128 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1129 			if (wpipe->pipe_state & PIPE_WANTR) {
1130 				wpipe->pipe_state &= ~PIPE_WANTR;
1131 				wakeup(wpipe);
1132 			}
1133 			pipeselwakeup(wpipe);
1134 			wpipe->pipe_state |= PIPE_WANTW;
1135 			pipeunlock(wpipe);
1136 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1137 			    "pipbww", 0);
1138 			if (error)
1139 				break;
1140 			else
1141 				continue;
1142 		}
1143 
1144 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1145 
1146 		/* Writes of size <= PIPE_BUF must be atomic. */
1147 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1148 			space = 0;
1149 
1150 		if (space > 0) {
1151 			int size;	/* Transfer size */
1152 			int segsize;	/* first segment to transfer */
1153 
1154 			/*
1155 			 * Transfer size is minimum of uio transfer
1156 			 * and free space in pipe buffer.
1157 			 */
1158 			if (space > uio->uio_resid)
1159 				size = uio->uio_resid;
1160 			else
1161 				size = space;
1162 			/*
1163 			 * First segment to transfer is minimum of
1164 			 * transfer size and contiguous space in
1165 			 * pipe buffer.  If first segment to transfer
1166 			 * is less than the transfer size, we've got
1167 			 * a wraparound in the buffer.
1168 			 */
1169 			segsize = wpipe->pipe_buffer.size -
1170 				wpipe->pipe_buffer.in;
1171 			if (segsize > size)
1172 				segsize = size;
1173 
1174 			/* Transfer first segment */
1175 
1176 			PIPE_UNLOCK(rpipe);
1177 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1178 					segsize, uio);
1179 			PIPE_LOCK(rpipe);
1180 
1181 			if (error == 0 && segsize < size) {
1182 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1183 					wpipe->pipe_buffer.size,
1184 					("Pipe buffer wraparound disappeared"));
1185 				/*
1186 				 * Transfer remaining part now, to
1187 				 * support atomic writes.  Wraparound
1188 				 * happened.
1189 				 */
1190 
1191 				PIPE_UNLOCK(rpipe);
1192 				error = uiomove(
1193 				    &wpipe->pipe_buffer.buffer[0],
1194 				    size - segsize, uio);
1195 				PIPE_LOCK(rpipe);
1196 			}
1197 			if (error == 0) {
1198 				wpipe->pipe_buffer.in += size;
1199 				if (wpipe->pipe_buffer.in >=
1200 				    wpipe->pipe_buffer.size) {
1201 					KASSERT(wpipe->pipe_buffer.in ==
1202 						size - segsize +
1203 						wpipe->pipe_buffer.size,
1204 						("Expected wraparound bad"));
1205 					wpipe->pipe_buffer.in = size - segsize;
1206 				}
1207 
1208 				wpipe->pipe_buffer.cnt += size;
1209 				KASSERT(wpipe->pipe_buffer.cnt <=
1210 					wpipe->pipe_buffer.size,
1211 					("Pipe buffer overflow"));
1212 			}
1213 			pipeunlock(wpipe);
1214 			if (error != 0)
1215 				break;
1216 		} else {
1217 			/*
1218 			 * If the "read-side" has been blocked, wake it up now.
1219 			 */
1220 			if (wpipe->pipe_state & PIPE_WANTR) {
1221 				wpipe->pipe_state &= ~PIPE_WANTR;
1222 				wakeup(wpipe);
1223 			}
1224 
1225 			/*
1226 			 * don't block on non-blocking I/O
1227 			 */
1228 			if (fp->f_flag & FNONBLOCK) {
1229 				error = EAGAIN;
1230 				pipeunlock(wpipe);
1231 				break;
1232 			}
1233 
1234 			/*
1235 			 * We have no more space and have something to offer,
1236 			 * wake up select/poll.
1237 			 */
1238 			pipeselwakeup(wpipe);
1239 
1240 			wpipe->pipe_state |= PIPE_WANTW;
1241 			pipeunlock(wpipe);
1242 			error = msleep(wpipe, PIPE_MTX(rpipe),
1243 			    PRIBIO | PCATCH, "pipewr", 0);
1244 			if (error != 0)
1245 				break;
1246 		}
1247 	}
1248 
1249 	pipelock(wpipe, 0);
1250 	--wpipe->pipe_busy;
1251 
1252 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1253 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1254 		wakeup(wpipe);
1255 	} else if (wpipe->pipe_buffer.cnt > 0) {
1256 		/*
1257 		 * If we have put any characters in the buffer, we wake up
1258 		 * the reader.
1259 		 */
1260 		if (wpipe->pipe_state & PIPE_WANTR) {
1261 			wpipe->pipe_state &= ~PIPE_WANTR;
1262 			wakeup(wpipe);
1263 		}
1264 	}
1265 
1266 	/*
1267 	 * Don't return EPIPE if any byte was written.
1268 	 * EINTR and other interrupts are handled by generic I/O layer.
1269 	 * Do not pretend that I/O succeeded for obvious user error
1270 	 * like EFAULT.
1271 	 */
1272 	if (uio->uio_resid != orig_resid && error == EPIPE)
1273 		error = 0;
1274 
1275 	if (error == 0)
1276 		vfs_timestamp(&wpipe->pipe_mtime);
1277 
1278 	/*
1279 	 * We have something to offer,
1280 	 * wake up select/poll.
1281 	 */
1282 	if (wpipe->pipe_buffer.cnt)
1283 		pipeselwakeup(wpipe);
1284 
1285 	pipeunlock(wpipe);
1286 	PIPE_UNLOCK(rpipe);
1287 	return (error);
1288 }
1289 
1290 /* ARGSUSED */
1291 static int
1292 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1293     struct thread *td)
1294 {
1295 	struct pipe *cpipe;
1296 	int error;
1297 
1298 	cpipe = fp->f_data;
1299 	if (cpipe->pipe_state & PIPE_NAMED)
1300 		error = vnops.fo_truncate(fp, length, active_cred, td);
1301 	else
1302 		error = invfo_truncate(fp, length, active_cred, td);
1303 	return (error);
1304 }
1305 
1306 /*
1307  * we implement a very minimal set of ioctls for compatibility with sockets.
1308  */
1309 static int
1310 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1311     struct thread *td)
1312 {
1313 	struct pipe *mpipe = fp->f_data;
1314 	int error;
1315 
1316 	PIPE_LOCK(mpipe);
1317 
1318 #ifdef MAC
1319 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1320 	if (error) {
1321 		PIPE_UNLOCK(mpipe);
1322 		return (error);
1323 	}
1324 #endif
1325 
1326 	error = 0;
1327 	switch (cmd) {
1328 
1329 	case FIONBIO:
1330 		break;
1331 
1332 	case FIOASYNC:
1333 		if (*(int *)data) {
1334 			mpipe->pipe_state |= PIPE_ASYNC;
1335 		} else {
1336 			mpipe->pipe_state &= ~PIPE_ASYNC;
1337 		}
1338 		break;
1339 
1340 	case FIONREAD:
1341 		if (!(fp->f_flag & FREAD)) {
1342 			*(int *)data = 0;
1343 			PIPE_UNLOCK(mpipe);
1344 			return (0);
1345 		}
1346 		if (mpipe->pipe_state & PIPE_DIRECTW)
1347 			*(int *)data = mpipe->pipe_map.cnt;
1348 		else
1349 			*(int *)data = mpipe->pipe_buffer.cnt;
1350 		break;
1351 
1352 	case FIOSETOWN:
1353 		PIPE_UNLOCK(mpipe);
1354 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1355 		goto out_unlocked;
1356 
1357 	case FIOGETOWN:
1358 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1359 		break;
1360 
1361 	/* This is deprecated, FIOSETOWN should be used instead. */
1362 	case TIOCSPGRP:
1363 		PIPE_UNLOCK(mpipe);
1364 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1365 		goto out_unlocked;
1366 
1367 	/* This is deprecated, FIOGETOWN should be used instead. */
1368 	case TIOCGPGRP:
1369 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1370 		break;
1371 
1372 	default:
1373 		error = ENOTTY;
1374 		break;
1375 	}
1376 	PIPE_UNLOCK(mpipe);
1377 out_unlocked:
1378 	return (error);
1379 }
1380 
1381 static int
1382 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1383     struct thread *td)
1384 {
1385 	struct pipe *rpipe;
1386 	struct pipe *wpipe;
1387 	int levents, revents;
1388 #ifdef MAC
1389 	int error;
1390 #endif
1391 
1392 	revents = 0;
1393 	rpipe = fp->f_data;
1394 	wpipe = PIPE_PEER(rpipe);
1395 	PIPE_LOCK(rpipe);
1396 #ifdef MAC
1397 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1398 	if (error)
1399 		goto locked_error;
1400 #endif
1401 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1402 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1403 		    (rpipe->pipe_buffer.cnt > 0))
1404 			revents |= events & (POLLIN | POLLRDNORM);
1405 
1406 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1407 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1408 		    (wpipe->pipe_state & PIPE_EOF) ||
1409 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1410 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1411 			 wpipe->pipe_buffer.size == 0)))
1412 			revents |= events & (POLLOUT | POLLWRNORM);
1413 
1414 	levents = events &
1415 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1416 	if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1417 	    fp->f_seqcount == rpipe->pipe_wgen)
1418 		events |= POLLINIGNEOF;
1419 
1420 	if ((events & POLLINIGNEOF) == 0) {
1421 		if (rpipe->pipe_state & PIPE_EOF) {
1422 			revents |= (events & (POLLIN | POLLRDNORM));
1423 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1424 			    (wpipe->pipe_state & PIPE_EOF))
1425 				revents |= POLLHUP;
1426 		}
1427 	}
1428 
1429 	if (revents == 0) {
1430 		if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1431 			selrecord(td, &rpipe->pipe_sel);
1432 			if (SEL_WAITING(&rpipe->pipe_sel))
1433 				rpipe->pipe_state |= PIPE_SEL;
1434 		}
1435 
1436 		if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1437 			selrecord(td, &wpipe->pipe_sel);
1438 			if (SEL_WAITING(&wpipe->pipe_sel))
1439 				wpipe->pipe_state |= PIPE_SEL;
1440 		}
1441 	}
1442 #ifdef MAC
1443 locked_error:
1444 #endif
1445 	PIPE_UNLOCK(rpipe);
1446 
1447 	return (revents);
1448 }
1449 
1450 /*
1451  * We shouldn't need locks here as we're doing a read and this should
1452  * be a natural race.
1453  */
1454 static int
1455 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1456     struct thread *td)
1457 {
1458 	struct pipe *pipe;
1459 #ifdef MAC
1460 	int error;
1461 #endif
1462 
1463 	pipe = fp->f_data;
1464 	PIPE_LOCK(pipe);
1465 #ifdef MAC
1466 	error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1467 	if (error) {
1468 		PIPE_UNLOCK(pipe);
1469 		return (error);
1470 	}
1471 #endif
1472 
1473 	/* For named pipes ask the underlying filesystem. */
1474 	if (pipe->pipe_state & PIPE_NAMED) {
1475 		PIPE_UNLOCK(pipe);
1476 		return (vnops.fo_stat(fp, ub, active_cred, td));
1477 	}
1478 
1479 	PIPE_UNLOCK(pipe);
1480 
1481 	bzero(ub, sizeof(*ub));
1482 	ub->st_mode = S_IFIFO;
1483 	ub->st_blksize = PAGE_SIZE;
1484 	if (pipe->pipe_state & PIPE_DIRECTW)
1485 		ub->st_size = pipe->pipe_map.cnt;
1486 	else
1487 		ub->st_size = pipe->pipe_buffer.cnt;
1488 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1489 	ub->st_atim = pipe->pipe_atime;
1490 	ub->st_mtim = pipe->pipe_mtime;
1491 	ub->st_ctim = pipe->pipe_ctime;
1492 	ub->st_uid = fp->f_cred->cr_uid;
1493 	ub->st_gid = fp->f_cred->cr_gid;
1494 	ub->st_dev = pipedev_ino;
1495 	ub->st_ino = pipe->pipe_ino;
1496 	/*
1497 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1498 	 */
1499 	return (0);
1500 }
1501 
1502 /* ARGSUSED */
1503 static int
1504 pipe_close(struct file *fp, struct thread *td)
1505 {
1506 
1507 	if (fp->f_vnode != NULL)
1508 		return vnops.fo_close(fp, td);
1509 	fp->f_ops = &badfileops;
1510 	pipe_dtor(fp->f_data);
1511 	fp->f_data = NULL;
1512 	return (0);
1513 }
1514 
1515 static int
1516 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1517 {
1518 	struct pipe *cpipe;
1519 	int error;
1520 
1521 	cpipe = fp->f_data;
1522 	if (cpipe->pipe_state & PIPE_NAMED)
1523 		error = vn_chmod(fp, mode, active_cred, td);
1524 	else
1525 		error = invfo_chmod(fp, mode, active_cred, td);
1526 	return (error);
1527 }
1528 
1529 static int
1530 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1531     struct thread *td)
1532 {
1533 	struct pipe *cpipe;
1534 	int error;
1535 
1536 	cpipe = fp->f_data;
1537 	if (cpipe->pipe_state & PIPE_NAMED)
1538 		error = vn_chown(fp, uid, gid, active_cred, td);
1539 	else
1540 		error = invfo_chown(fp, uid, gid, active_cred, td);
1541 	return (error);
1542 }
1543 
1544 static int
1545 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1546 {
1547 	struct pipe *pi;
1548 
1549 	if (fp->f_type == DTYPE_FIFO)
1550 		return (vn_fill_kinfo(fp, kif, fdp));
1551 	kif->kf_type = KF_TYPE_PIPE;
1552 	pi = fp->f_data;
1553 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1554 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1555 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1556 	return (0);
1557 }
1558 
1559 static void
1560 pipe_free_kmem(struct pipe *cpipe)
1561 {
1562 
1563 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1564 	    ("pipe_free_kmem: pipe mutex locked"));
1565 
1566 	if (cpipe->pipe_buffer.buffer != NULL) {
1567 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1568 		vm_map_remove(pipe_map,
1569 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1570 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1571 		cpipe->pipe_buffer.buffer = NULL;
1572 	}
1573 #ifndef PIPE_NODIRECT
1574 	{
1575 		cpipe->pipe_map.cnt = 0;
1576 		cpipe->pipe_map.pos = 0;
1577 		cpipe->pipe_map.npages = 0;
1578 	}
1579 #endif
1580 }
1581 
1582 /*
1583  * shutdown the pipe
1584  */
1585 static void
1586 pipeclose(struct pipe *cpipe)
1587 {
1588 	struct pipepair *pp;
1589 	struct pipe *ppipe;
1590 
1591 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1592 
1593 	PIPE_LOCK(cpipe);
1594 	pipelock(cpipe, 0);
1595 	pp = cpipe->pipe_pair;
1596 
1597 	pipeselwakeup(cpipe);
1598 
1599 	/*
1600 	 * If the other side is blocked, wake it up saying that
1601 	 * we want to close it down.
1602 	 */
1603 	cpipe->pipe_state |= PIPE_EOF;
1604 	while (cpipe->pipe_busy) {
1605 		wakeup(cpipe);
1606 		cpipe->pipe_state |= PIPE_WANT;
1607 		pipeunlock(cpipe);
1608 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1609 		pipelock(cpipe, 0);
1610 	}
1611 
1612 
1613 	/*
1614 	 * Disconnect from peer, if any.
1615 	 */
1616 	ppipe = cpipe->pipe_peer;
1617 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1618 		pipeselwakeup(ppipe);
1619 
1620 		ppipe->pipe_state |= PIPE_EOF;
1621 		wakeup(ppipe);
1622 		KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1623 	}
1624 
1625 	/*
1626 	 * Mark this endpoint as free.  Release kmem resources.  We
1627 	 * don't mark this endpoint as unused until we've finished
1628 	 * doing that, or the pipe might disappear out from under
1629 	 * us.
1630 	 */
1631 	PIPE_UNLOCK(cpipe);
1632 	pipe_free_kmem(cpipe);
1633 	PIPE_LOCK(cpipe);
1634 	cpipe->pipe_present = PIPE_CLOSING;
1635 	pipeunlock(cpipe);
1636 
1637 	/*
1638 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1639 	 * PIPE_FINALIZED, that allows other end to free the
1640 	 * pipe_pair, only after the knotes are completely dismantled.
1641 	 */
1642 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1643 	cpipe->pipe_present = PIPE_FINALIZED;
1644 	seldrain(&cpipe->pipe_sel);
1645 	knlist_destroy(&cpipe->pipe_sel.si_note);
1646 
1647 	/*
1648 	 * If both endpoints are now closed, release the memory for the
1649 	 * pipe pair.  If not, unlock.
1650 	 */
1651 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1652 		PIPE_UNLOCK(cpipe);
1653 #ifdef MAC
1654 		mac_pipe_destroy(pp);
1655 #endif
1656 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1657 	} else
1658 		PIPE_UNLOCK(cpipe);
1659 }
1660 
1661 /*ARGSUSED*/
1662 static int
1663 pipe_kqfilter(struct file *fp, struct knote *kn)
1664 {
1665 	struct pipe *cpipe;
1666 
1667 	/*
1668 	 * If a filter is requested that is not supported by this file
1669 	 * descriptor, don't return an error, but also don't ever generate an
1670 	 * event.
1671 	 */
1672 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1673 		kn->kn_fop = &pipe_nfiltops;
1674 		return (0);
1675 	}
1676 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1677 		kn->kn_fop = &pipe_nfiltops;
1678 		return (0);
1679 	}
1680 	cpipe = fp->f_data;
1681 	PIPE_LOCK(cpipe);
1682 	switch (kn->kn_filter) {
1683 	case EVFILT_READ:
1684 		kn->kn_fop = &pipe_rfiltops;
1685 		break;
1686 	case EVFILT_WRITE:
1687 		kn->kn_fop = &pipe_wfiltops;
1688 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1689 			/* other end of pipe has been closed */
1690 			PIPE_UNLOCK(cpipe);
1691 			return (EPIPE);
1692 		}
1693 		cpipe = PIPE_PEER(cpipe);
1694 		break;
1695 	default:
1696 		PIPE_UNLOCK(cpipe);
1697 		return (EINVAL);
1698 	}
1699 
1700 	kn->kn_hook = cpipe;
1701 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1702 	PIPE_UNLOCK(cpipe);
1703 	return (0);
1704 }
1705 
1706 static void
1707 filt_pipedetach(struct knote *kn)
1708 {
1709 	struct pipe *cpipe = kn->kn_hook;
1710 
1711 	PIPE_LOCK(cpipe);
1712 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1713 	PIPE_UNLOCK(cpipe);
1714 }
1715 
1716 /*ARGSUSED*/
1717 static int
1718 filt_piperead(struct knote *kn, long hint)
1719 {
1720 	struct pipe *rpipe = kn->kn_hook;
1721 	struct pipe *wpipe = rpipe->pipe_peer;
1722 	int ret;
1723 
1724 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1725 	kn->kn_data = rpipe->pipe_buffer.cnt;
1726 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1727 		kn->kn_data = rpipe->pipe_map.cnt;
1728 
1729 	if ((rpipe->pipe_state & PIPE_EOF) ||
1730 	    wpipe->pipe_present != PIPE_ACTIVE ||
1731 	    (wpipe->pipe_state & PIPE_EOF)) {
1732 		kn->kn_flags |= EV_EOF;
1733 		return (1);
1734 	}
1735 	ret = kn->kn_data > 0;
1736 	return ret;
1737 }
1738 
1739 /*ARGSUSED*/
1740 static int
1741 filt_pipewrite(struct knote *kn, long hint)
1742 {
1743 	struct pipe *wpipe;
1744 
1745 	/*
1746 	 * If this end of the pipe is closed, the knote was removed from the
1747 	 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1748 	 */
1749 	wpipe = kn->kn_hook;
1750 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1751 	    (wpipe->pipe_state & PIPE_EOF)) {
1752 		kn->kn_data = 0;
1753 		kn->kn_flags |= EV_EOF;
1754 		return (1);
1755 	}
1756 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1757 	kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1758 	    (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1759 	if (wpipe->pipe_state & PIPE_DIRECTW)
1760 		kn->kn_data = 0;
1761 
1762 	return (kn->kn_data >= PIPE_BUF);
1763 }
1764 
1765 static void
1766 filt_pipedetach_notsup(struct knote *kn)
1767 {
1768 
1769 }
1770 
1771 static int
1772 filt_pipenotsup(struct knote *kn, long hint)
1773 {
1774 
1775 	return (0);
1776 }
1777