xref: /freebsd/sys/kern/sys_pipe.c (revision f61e92ca5a23450bc28169bbdd71d7674df98c19)
1 /*-
2  * SPDX-License-Identifier: BSD-4-Clause
3  *
4  * Copyright (c) 1996 John S. Dyson
5  * Copyright (c) 2012 Giovanni Trematerra
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice immediately at the beginning of the file, without modification,
13  *    this list of conditions, and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. Absolutely no warranty of function or purpose is made by the author
18  *    John S. Dyson.
19  * 4. Modifications may be freely made to this file if the above conditions
20  *    are met.
21  */
22 
23 /*
24  * This file contains a high-performance replacement for the socket-based
25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
26  * all features of sockets, but does do everything that pipes normally
27  * do.
28  */
29 
30 /*
31  * This code has two modes of operation, a small write mode and a large
32  * write mode.  The small write mode acts like conventional pipes with
33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
36  * memory, and the receiving process copies directly from these pinned pages
37  * in the sending process.
38  *
39  * If the sending process receives a signal, it is possible that it will
40  * go away, and certainly its address space can change, because control
41  * is returned back to the user-mode side.  In that case, the pipe code
42  * arranges to copy the buffer supplied by the user process, to a pageable
43  * kernel buffer, and the receiving process will grab the data from the
44  * pageable kernel buffer.  Since signals don't happen all that often,
45  * the copy operation is normally eliminated.
46  *
47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
48  * happen for small transfers so that the system will not spend all of
49  * its time context switching.
50  *
51  * In order to limit the resource use of pipes, two sysctls exist:
52  *
53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
54  * address space available to us in pipe_map. This value is normally
55  * autotuned, but may also be loader tuned.
56  *
57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
58  * memory in use by pipes.
59  *
60  * Based on how large pipekva is relative to maxpipekva, the following
61  * will happen:
62  *
63  * 0% - 50%:
64  *     New pipes are given 16K of memory backing, pipes may dynamically
65  *     grow to as large as 64K where needed.
66  * 50% - 75%:
67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
68  *     existing pipes may NOT grow.
69  * 75% - 100%:
70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
71  *     existing pipes will be shrunk down to 4K whenever possible.
72  *
73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
75  * resize which MUST occur for reverse-direction pipes when they are
76  * first used.
77  *
78  * Additional information about the current state of pipes may be obtained
79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
80  * and kern.ipc.piperesizefail.
81  *
82  * Locking rules:  There are two locks present here:  A mutex, used via
83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
84  * the flag, as mutexes can not persist over uiomove.  The mutex
85  * exists only to guard access to the flag, and is not in itself a
86  * locking mechanism.  Also note that there is only a single mutex for
87  * both directions of a pipe.
88  *
89  * As pipelock() may have to sleep before it can acquire the flag, it
90  * is important to reread all data after a call to pipelock(); everything
91  * in the structure may have changed.
92  */
93 
94 #include <sys/cdefs.h>
95 __FBSDID("$FreeBSD$");
96 
97 #include <sys/param.h>
98 #include <sys/systm.h>
99 #include <sys/conf.h>
100 #include <sys/fcntl.h>
101 #include <sys/file.h>
102 #include <sys/filedesc.h>
103 #include <sys/filio.h>
104 #include <sys/kernel.h>
105 #include <sys/lock.h>
106 #include <sys/mutex.h>
107 #include <sys/ttycom.h>
108 #include <sys/stat.h>
109 #include <sys/malloc.h>
110 #include <sys/poll.h>
111 #include <sys/selinfo.h>
112 #include <sys/signalvar.h>
113 #include <sys/syscallsubr.h>
114 #include <sys/sysctl.h>
115 #include <sys/sysproto.h>
116 #include <sys/pipe.h>
117 #include <sys/proc.h>
118 #include <sys/vnode.h>
119 #include <sys/uio.h>
120 #include <sys/user.h>
121 #include <sys/event.h>
122 
123 #include <security/mac/mac_framework.h>
124 
125 #include <vm/vm.h>
126 #include <vm/vm_param.h>
127 #include <vm/vm_object.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_extern.h>
130 #include <vm/pmap.h>
131 #include <vm/vm_map.h>
132 #include <vm/vm_page.h>
133 #include <vm/uma.h>
134 
135 /*
136  * Use this define if you want to disable *fancy* VM things.  Expect an
137  * approx 30% decrease in transfer rate.  This could be useful for
138  * NetBSD or OpenBSD.
139  */
140 /* #define PIPE_NODIRECT */
141 
142 #define PIPE_PEER(pipe)	\
143 	(((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
144 
145 /*
146  * interfaces to the outside world
147  */
148 static fo_rdwr_t	pipe_read;
149 static fo_rdwr_t	pipe_write;
150 static fo_truncate_t	pipe_truncate;
151 static fo_ioctl_t	pipe_ioctl;
152 static fo_poll_t	pipe_poll;
153 static fo_kqfilter_t	pipe_kqfilter;
154 static fo_stat_t	pipe_stat;
155 static fo_close_t	pipe_close;
156 static fo_chmod_t	pipe_chmod;
157 static fo_chown_t	pipe_chown;
158 static fo_fill_kinfo_t	pipe_fill_kinfo;
159 
160 struct fileops pipeops = {
161 	.fo_read = pipe_read,
162 	.fo_write = pipe_write,
163 	.fo_truncate = pipe_truncate,
164 	.fo_ioctl = pipe_ioctl,
165 	.fo_poll = pipe_poll,
166 	.fo_kqfilter = pipe_kqfilter,
167 	.fo_stat = pipe_stat,
168 	.fo_close = pipe_close,
169 	.fo_chmod = pipe_chmod,
170 	.fo_chown = pipe_chown,
171 	.fo_sendfile = invfo_sendfile,
172 	.fo_fill_kinfo = pipe_fill_kinfo,
173 	.fo_flags = DFLAG_PASSABLE
174 };
175 
176 static void	filt_pipedetach(struct knote *kn);
177 static void	filt_pipedetach_notsup(struct knote *kn);
178 static int	filt_pipenotsup(struct knote *kn, long hint);
179 static int	filt_piperead(struct knote *kn, long hint);
180 static int	filt_pipewrite(struct knote *kn, long hint);
181 
182 static struct filterops pipe_nfiltops = {
183 	.f_isfd = 1,
184 	.f_detach = filt_pipedetach_notsup,
185 	.f_event = filt_pipenotsup
186 };
187 static struct filterops pipe_rfiltops = {
188 	.f_isfd = 1,
189 	.f_detach = filt_pipedetach,
190 	.f_event = filt_piperead
191 };
192 static struct filterops pipe_wfiltops = {
193 	.f_isfd = 1,
194 	.f_detach = filt_pipedetach,
195 	.f_event = filt_pipewrite
196 };
197 
198 /*
199  * Default pipe buffer size(s), this can be kind-of large now because pipe
200  * space is pageable.  The pipe code will try to maintain locality of
201  * reference for performance reasons, so small amounts of outstanding I/O
202  * will not wipe the cache.
203  */
204 #define MINPIPESIZE (PIPE_SIZE/3)
205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
206 
207 static long amountpipekva;
208 static int pipefragretry;
209 static int pipeallocfail;
210 static int piperesizefail;
211 static int piperesizeallowed = 1;
212 
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 int pipe_create(struct pipe *pipe, bool backing);
230 static int 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 static void pipe_timestamp(struct timespec *tsp);
234 #ifndef PIPE_NODIRECT
235 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
236 static void pipe_destroy_write_buffer(struct pipe *wpipe);
237 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
238 static void pipe_clone_write_buffer(struct pipe *wpipe);
239 #endif
240 static int pipespace(struct pipe *cpipe, int size);
241 static int pipespace_new(struct pipe *cpipe, int size);
242 
243 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
244 static int	pipe_zone_init(void *mem, int size, int flags);
245 static void	pipe_zone_fini(void *mem, int size);
246 
247 static uma_zone_t pipe_zone;
248 static struct unrhdr64 pipeino_unr;
249 static dev_t pipedev_ino;
250 
251 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
252 
253 static void
254 pipeinit(void *dummy __unused)
255 {
256 
257 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
258 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
259 	    UMA_ALIGN_PTR, 0);
260 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
261 	new_unrhdr64(&pipeino_unr, 1);
262 	pipedev_ino = devfs_alloc_cdp_inode();
263 	KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
264 }
265 
266 static int
267 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
268 {
269 	struct pipepair *pp;
270 	struct pipe *rpipe, *wpipe;
271 
272 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
273 
274 	pp = (struct pipepair *)mem;
275 
276 	/*
277 	 * We zero both pipe endpoints to make sure all the kmem pointers
278 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
279 	 * endpoints with the same time.
280 	 */
281 	rpipe = &pp->pp_rpipe;
282 	bzero(rpipe, sizeof(*rpipe));
283 	pipe_timestamp(&rpipe->pipe_ctime);
284 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
285 
286 	wpipe = &pp->pp_wpipe;
287 	bzero(wpipe, sizeof(*wpipe));
288 	wpipe->pipe_ctime = rpipe->pipe_ctime;
289 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
290 
291 	rpipe->pipe_peer = wpipe;
292 	rpipe->pipe_pair = pp;
293 	wpipe->pipe_peer = rpipe;
294 	wpipe->pipe_pair = pp;
295 
296 	/*
297 	 * Mark both endpoints as present; they will later get free'd
298 	 * one at a time.  When both are free'd, then the whole pair
299 	 * is released.
300 	 */
301 	rpipe->pipe_present = PIPE_ACTIVE;
302 	wpipe->pipe_present = PIPE_ACTIVE;
303 
304 	/*
305 	 * Eventually, the MAC Framework may initialize the label
306 	 * in ctor or init, but for now we do it elswhere to avoid
307 	 * blocking in ctor or init.
308 	 */
309 	pp->pp_label = NULL;
310 
311 	return (0);
312 }
313 
314 static int
315 pipe_zone_init(void *mem, int size, int flags)
316 {
317 	struct pipepair *pp;
318 
319 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
320 
321 	pp = (struct pipepair *)mem;
322 
323 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
324 	return (0);
325 }
326 
327 static void
328 pipe_zone_fini(void *mem, int size)
329 {
330 	struct pipepair *pp;
331 
332 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
333 
334 	pp = (struct pipepair *)mem;
335 
336 	mtx_destroy(&pp->pp_mtx);
337 }
338 
339 static int
340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
341 {
342 	struct pipepair *pp;
343 	struct pipe *rpipe, *wpipe;
344 	int error;
345 
346 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
347 #ifdef MAC
348 	/*
349 	 * The MAC label is shared between the connected endpoints.  As a
350 	 * result mac_pipe_init() and mac_pipe_create() are called once
351 	 * for the pair, and not on the endpoints.
352 	 */
353 	mac_pipe_init(pp);
354 	mac_pipe_create(td->td_ucred, pp);
355 #endif
356 	rpipe = &pp->pp_rpipe;
357 	wpipe = &pp->pp_wpipe;
358 
359 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
360 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
361 
362 	/*
363 	 * Only the forward direction pipe is backed by big buffer by
364 	 * default.
365 	 */
366 	error = pipe_create(rpipe, true);
367 	if (error != 0)
368 		goto fail;
369 	error = pipe_create(wpipe, false);
370 	if (error != 0) {
371 		/*
372 		 * This cleanup leaves the pipe inode number for rpipe
373 		 * still allocated, but never used.  We do not free
374 		 * inode numbers for opened pipes, which is required
375 		 * for correctness because numbers must be unique.
376 		 * But also it avoids any memory use by the unr
377 		 * allocator, so stashing away the transient inode
378 		 * number is reasonable.
379 		 */
380 		pipe_free_kmem(rpipe);
381 		goto fail;
382 	}
383 
384 	rpipe->pipe_state |= PIPE_DIRECTOK;
385 	wpipe->pipe_state |= PIPE_DIRECTOK;
386 	return (0);
387 
388 fail:
389 	knlist_destroy(&rpipe->pipe_sel.si_note);
390 	knlist_destroy(&wpipe->pipe_sel.si_note);
391 #ifdef MAC
392 	mac_pipe_destroy(pp);
393 #endif
394 	return (error);
395 }
396 
397 int
398 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
399 {
400 	struct pipepair *pp;
401 	int error;
402 
403 	error = pipe_paircreate(td, &pp);
404 	if (error != 0)
405 		return (error);
406 	pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
407 	*ppipe = &pp->pp_rpipe;
408 	return (0);
409 }
410 
411 void
412 pipe_dtor(struct pipe *dpipe)
413 {
414 	struct pipe *peer;
415 
416 	peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
417 	funsetown(&dpipe->pipe_sigio);
418 	pipeclose(dpipe);
419 	if (peer != NULL) {
420 		funsetown(&peer->pipe_sigio);
421 		pipeclose(peer);
422 	}
423 }
424 
425 /*
426  * Get a timestamp.
427  *
428  * This used to be vfs_timestamp but the higher precision is unnecessary and
429  * can very negatively affect performance in virtualized environments (e.g., on
430  * vms running on amd64 when using the rdtscp instruction).
431  */
432 static void
433 pipe_timestamp(struct timespec *tsp)
434 {
435 
436 	getnanotime(tsp);
437 }
438 
439 /*
440  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
441  * the zone pick up the pieces via pipeclose().
442  */
443 int
444 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
445     struct filecaps *fcaps2)
446 {
447 	struct file *rf, *wf;
448 	struct pipe *rpipe, *wpipe;
449 	struct pipepair *pp;
450 	int fd, fflags, error;
451 
452 	error = pipe_paircreate(td, &pp);
453 	if (error != 0)
454 		return (error);
455 	rpipe = &pp->pp_rpipe;
456 	wpipe = &pp->pp_wpipe;
457 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
458 	if (error) {
459 		pipeclose(rpipe);
460 		pipeclose(wpipe);
461 		return (error);
462 	}
463 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
464 	fildes[0] = fd;
465 
466 	fflags = FREAD | FWRITE;
467 	if ((flags & O_NONBLOCK) != 0)
468 		fflags |= FNONBLOCK;
469 
470 	/*
471 	 * Warning: once we've gotten past allocation of the fd for the
472 	 * read-side, we can only drop the read side via fdrop() in order
473 	 * to avoid races against processes which manage to dup() the read
474 	 * side while we are blocked trying to allocate the write side.
475 	 */
476 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
477 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
478 	if (error) {
479 		fdclose(td, rf, fildes[0]);
480 		fdrop(rf, td);
481 		/* rpipe has been closed by fdrop(). */
482 		pipeclose(wpipe);
483 		return (error);
484 	}
485 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
486 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
487 	fdrop(wf, td);
488 	fildes[1] = fd;
489 	fdrop(rf, td);
490 
491 	return (0);
492 }
493 
494 #ifdef COMPAT_FREEBSD10
495 /* ARGSUSED */
496 int
497 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
498 {
499 	int error;
500 	int fildes[2];
501 
502 	error = kern_pipe(td, fildes, 0, NULL, NULL);
503 	if (error)
504 		return (error);
505 
506 	td->td_retval[0] = fildes[0];
507 	td->td_retval[1] = fildes[1];
508 
509 	return (0);
510 }
511 #endif
512 
513 int
514 sys_pipe2(struct thread *td, struct pipe2_args *uap)
515 {
516 	int error, fildes[2];
517 
518 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
519 		return (EINVAL);
520 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
521 	if (error)
522 		return (error);
523 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
524 	if (error) {
525 		(void)kern_close(td, fildes[0]);
526 		(void)kern_close(td, fildes[1]);
527 	}
528 	return (error);
529 }
530 
531 /*
532  * Allocate kva for pipe circular buffer, the space is pageable
533  * This routine will 'realloc' the size of a pipe safely, if it fails
534  * it will retain the old buffer.
535  * If it fails it will return ENOMEM.
536  */
537 static int
538 pipespace_new(struct pipe *cpipe, int size)
539 {
540 	caddr_t buffer;
541 	int error, cnt, firstseg;
542 	static int curfail = 0;
543 	static struct timeval lastfail;
544 
545 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
546 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
547 		("pipespace: resize of direct writes not allowed"));
548 retry:
549 	cnt = cpipe->pipe_buffer.cnt;
550 	if (cnt > size)
551 		size = cnt;
552 
553 	size = round_page(size);
554 	buffer = (caddr_t) vm_map_min(pipe_map);
555 
556 	error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
557 	    VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
558 	if (error != KERN_SUCCESS) {
559 		if (cpipe->pipe_buffer.buffer == NULL &&
560 		    size > SMALL_PIPE_SIZE) {
561 			size = SMALL_PIPE_SIZE;
562 			pipefragretry++;
563 			goto retry;
564 		}
565 		if (cpipe->pipe_buffer.buffer == NULL) {
566 			pipeallocfail++;
567 			if (ppsratecheck(&lastfail, &curfail, 1))
568 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
569 		} else {
570 			piperesizefail++;
571 		}
572 		return (ENOMEM);
573 	}
574 
575 	/* copy data, then free old resources if we're resizing */
576 	if (cnt > 0) {
577 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
578 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
579 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
580 				buffer, firstseg);
581 			if ((cnt - firstseg) > 0)
582 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
583 					cpipe->pipe_buffer.in);
584 		} else {
585 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
586 				buffer, cnt);
587 		}
588 	}
589 	pipe_free_kmem(cpipe);
590 	cpipe->pipe_buffer.buffer = buffer;
591 	cpipe->pipe_buffer.size = size;
592 	cpipe->pipe_buffer.in = cnt;
593 	cpipe->pipe_buffer.out = 0;
594 	cpipe->pipe_buffer.cnt = cnt;
595 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
596 	return (0);
597 }
598 
599 /*
600  * Wrapper for pipespace_new() that performs locking assertions.
601  */
602 static int
603 pipespace(struct pipe *cpipe, int size)
604 {
605 
606 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
607 	    ("Unlocked pipe passed to pipespace"));
608 	return (pipespace_new(cpipe, size));
609 }
610 
611 /*
612  * lock a pipe for I/O, blocking other access
613  */
614 static __inline int
615 pipelock(struct pipe *cpipe, int catch)
616 {
617 	int error, prio;
618 
619 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
620 
621 	prio = PRIBIO;
622 	if (catch)
623 		prio |= PCATCH;
624 	while (cpipe->pipe_state & PIPE_LOCKFL) {
625 		KASSERT(cpipe->pipe_waiters >= 0,
626 		    ("%s: bad waiter count %d", __func__,
627 		    cpipe->pipe_waiters));
628 		cpipe->pipe_waiters++;
629 		error = msleep(cpipe, PIPE_MTX(cpipe),
630 		    prio, "pipelk", 0);
631 		cpipe->pipe_waiters--;
632 		if (error != 0)
633 			return (error);
634 	}
635 	cpipe->pipe_state |= PIPE_LOCKFL;
636 	return (0);
637 }
638 
639 /*
640  * unlock a pipe I/O lock
641  */
642 static __inline void
643 pipeunlock(struct pipe *cpipe)
644 {
645 
646 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
647 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
648 		("Unlocked pipe passed to pipeunlock"));
649 	KASSERT(cpipe->pipe_waiters >= 0,
650 	    ("%s: bad waiter count %d", __func__,
651 	    cpipe->pipe_waiters));
652 	cpipe->pipe_state &= ~PIPE_LOCKFL;
653 	if (cpipe->pipe_waiters > 0) {
654 		wakeup_one(cpipe);
655 	}
656 }
657 
658 void
659 pipeselwakeup(struct pipe *cpipe)
660 {
661 
662 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
663 	if (cpipe->pipe_state & PIPE_SEL) {
664 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
665 		if (!SEL_WAITING(&cpipe->pipe_sel))
666 			cpipe->pipe_state &= ~PIPE_SEL;
667 	}
668 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
669 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
670 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
671 }
672 
673 /*
674  * Initialize and allocate VM and memory for pipe.  The structure
675  * will start out zero'd from the ctor, so we just manage the kmem.
676  */
677 static int
678 pipe_create(struct pipe *pipe, bool large_backing)
679 {
680 	int error;
681 
682 	error = pipespace_new(pipe, !large_backing || amountpipekva >
683 	    maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
684 	if (error == 0)
685 		pipe->pipe_ino = alloc_unr64(&pipeino_unr);
686 	return (error);
687 }
688 
689 /* ARGSUSED */
690 static int
691 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
692     int flags, struct thread *td)
693 {
694 	struct pipe *rpipe;
695 	int error;
696 	int nread = 0;
697 	int size;
698 
699 	rpipe = fp->f_data;
700 	PIPE_LOCK(rpipe);
701 	++rpipe->pipe_busy;
702 	error = pipelock(rpipe, 1);
703 	if (error)
704 		goto unlocked_error;
705 
706 #ifdef MAC
707 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
708 	if (error)
709 		goto locked_error;
710 #endif
711 	if (amountpipekva > (3 * maxpipekva) / 4) {
712 		if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
713 		    rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
714 		    rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
715 		    piperesizeallowed == 1) {
716 			PIPE_UNLOCK(rpipe);
717 			pipespace(rpipe, SMALL_PIPE_SIZE);
718 			PIPE_LOCK(rpipe);
719 		}
720 	}
721 
722 	while (uio->uio_resid) {
723 		/*
724 		 * normal pipe buffer receive
725 		 */
726 		if (rpipe->pipe_buffer.cnt > 0) {
727 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
728 			if (size > rpipe->pipe_buffer.cnt)
729 				size = rpipe->pipe_buffer.cnt;
730 			if (size > uio->uio_resid)
731 				size = uio->uio_resid;
732 
733 			PIPE_UNLOCK(rpipe);
734 			error = uiomove(
735 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
736 			    size, uio);
737 			PIPE_LOCK(rpipe);
738 			if (error)
739 				break;
740 
741 			rpipe->pipe_buffer.out += size;
742 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
743 				rpipe->pipe_buffer.out = 0;
744 
745 			rpipe->pipe_buffer.cnt -= size;
746 
747 			/*
748 			 * If there is no more to read in the pipe, reset
749 			 * its pointers to the beginning.  This improves
750 			 * cache hit stats.
751 			 */
752 			if (rpipe->pipe_buffer.cnt == 0) {
753 				rpipe->pipe_buffer.in = 0;
754 				rpipe->pipe_buffer.out = 0;
755 			}
756 			nread += size;
757 #ifndef PIPE_NODIRECT
758 		/*
759 		 * Direct copy, bypassing a kernel buffer.
760 		 */
761 		} else if ((size = rpipe->pipe_pages.cnt) != 0) {
762 			if (size > uio->uio_resid)
763 				size = (u_int) uio->uio_resid;
764 			PIPE_UNLOCK(rpipe);
765 			error = uiomove_fromphys(rpipe->pipe_pages.ms,
766 			    rpipe->pipe_pages.pos, size, uio);
767 			PIPE_LOCK(rpipe);
768 			if (error)
769 				break;
770 			nread += size;
771 			rpipe->pipe_pages.pos += size;
772 			rpipe->pipe_pages.cnt -= size;
773 			if (rpipe->pipe_pages.cnt == 0) {
774 				rpipe->pipe_state &= ~PIPE_WANTW;
775 				wakeup(rpipe);
776 			}
777 #endif
778 		} else {
779 			/*
780 			 * detect EOF condition
781 			 * read returns 0 on EOF, no need to set error
782 			 */
783 			if (rpipe->pipe_state & PIPE_EOF)
784 				break;
785 
786 			/*
787 			 * If the "write-side" has been blocked, wake it up now.
788 			 */
789 			if (rpipe->pipe_state & PIPE_WANTW) {
790 				rpipe->pipe_state &= ~PIPE_WANTW;
791 				wakeup(rpipe);
792 			}
793 
794 			/*
795 			 * Break if some data was read.
796 			 */
797 			if (nread > 0)
798 				break;
799 
800 			/*
801 			 * Unlock the pipe buffer for our remaining processing.
802 			 * We will either break out with an error or we will
803 			 * sleep and relock to loop.
804 			 */
805 			pipeunlock(rpipe);
806 
807 			/*
808 			 * Handle non-blocking mode operation or
809 			 * wait for more data.
810 			 */
811 			if (fp->f_flag & FNONBLOCK) {
812 				error = EAGAIN;
813 			} else {
814 				rpipe->pipe_state |= PIPE_WANTR;
815 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
816 				    PRIBIO | PCATCH,
817 				    "piperd", 0)) == 0)
818 					error = pipelock(rpipe, 1);
819 			}
820 			if (error)
821 				goto unlocked_error;
822 		}
823 	}
824 #ifdef MAC
825 locked_error:
826 #endif
827 	pipeunlock(rpipe);
828 
829 	/* XXX: should probably do this before getting any locks. */
830 	if (error == 0)
831 		pipe_timestamp(&rpipe->pipe_atime);
832 unlocked_error:
833 	--rpipe->pipe_busy;
834 
835 	/*
836 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
837 	 */
838 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
839 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
840 		wakeup(rpipe);
841 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
842 		/*
843 		 * Handle write blocking hysteresis.
844 		 */
845 		if (rpipe->pipe_state & PIPE_WANTW) {
846 			rpipe->pipe_state &= ~PIPE_WANTW;
847 			wakeup(rpipe);
848 		}
849 	}
850 
851 	/*
852 	 * Only wake up writers if there was actually something read.
853 	 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
854 	 */
855 	if (nread > 0 &&
856 	    rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
857 		pipeselwakeup(rpipe);
858 
859 	PIPE_UNLOCK(rpipe);
860 	if (nread > 0)
861 		td->td_ru.ru_msgrcv++;
862 	return (error);
863 }
864 
865 #ifndef PIPE_NODIRECT
866 /*
867  * Map the sending processes' buffer into kernel space and wire it.
868  * This is similar to a physical write operation.
869  */
870 static int
871 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
872 {
873 	u_int size;
874 	int i;
875 
876 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
877 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
878 	    ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
879 	KASSERT(wpipe->pipe_pages.cnt == 0,
880 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
881 
882 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
883                 size = wpipe->pipe_buffer.size;
884 	else
885                 size = uio->uio_iov->iov_len;
886 
887 	wpipe->pipe_state |= PIPE_DIRECTW;
888 	PIPE_UNLOCK(wpipe);
889 	i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
890 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
891 	    wpipe->pipe_pages.ms, PIPENPAGES);
892 	PIPE_LOCK(wpipe);
893 	if (i < 0) {
894 		wpipe->pipe_state &= ~PIPE_DIRECTW;
895 		return (EFAULT);
896 	}
897 
898 	wpipe->pipe_pages.npages = i;
899 	wpipe->pipe_pages.pos =
900 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
901 	wpipe->pipe_pages.cnt = size;
902 
903 	uio->uio_iov->iov_len -= size;
904 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
905 	if (uio->uio_iov->iov_len == 0)
906 		uio->uio_iov++;
907 	uio->uio_resid -= size;
908 	uio->uio_offset += size;
909 	return (0);
910 }
911 
912 /*
913  * Unwire the process buffer.
914  */
915 static void
916 pipe_destroy_write_buffer(struct pipe *wpipe)
917 {
918 
919 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
920 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
921 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
922 	KASSERT(wpipe->pipe_pages.cnt == 0,
923 	    ("%s: pipe map for %p contains residual data", __func__, wpipe));
924 
925 	wpipe->pipe_state &= ~PIPE_DIRECTW;
926 	vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
927 	wpipe->pipe_pages.npages = 0;
928 }
929 
930 /*
931  * In the case of a signal, the writing process might go away.  This
932  * code copies the data into the circular buffer so that the source
933  * pages can be freed without loss of data.
934  */
935 static void
936 pipe_clone_write_buffer(struct pipe *wpipe)
937 {
938 	struct uio uio;
939 	struct iovec iov;
940 	int size;
941 	int pos;
942 
943 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
944 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
945 	    ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
946 
947 	size = wpipe->pipe_pages.cnt;
948 	pos = wpipe->pipe_pages.pos;
949 	wpipe->pipe_pages.cnt = 0;
950 
951 	wpipe->pipe_buffer.in = size;
952 	wpipe->pipe_buffer.out = 0;
953 	wpipe->pipe_buffer.cnt = size;
954 
955 	PIPE_UNLOCK(wpipe);
956 	iov.iov_base = wpipe->pipe_buffer.buffer;
957 	iov.iov_len = size;
958 	uio.uio_iov = &iov;
959 	uio.uio_iovcnt = 1;
960 	uio.uio_offset = 0;
961 	uio.uio_resid = size;
962 	uio.uio_segflg = UIO_SYSSPACE;
963 	uio.uio_rw = UIO_READ;
964 	uio.uio_td = curthread;
965 	uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
966 	PIPE_LOCK(wpipe);
967 	pipe_destroy_write_buffer(wpipe);
968 }
969 
970 /*
971  * This implements the pipe buffer write mechanism.  Note that only
972  * a direct write OR a normal pipe write can be pending at any given time.
973  * If there are any characters in the pipe buffer, the direct write will
974  * be deferred until the receiving process grabs all of the bytes from
975  * the pipe buffer.  Then the direct mapping write is set-up.
976  */
977 static int
978 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
979 {
980 	int error;
981 
982 retry:
983 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
984 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
985 		error = EPIPE;
986 		goto error1;
987 	}
988 	if (wpipe->pipe_state & PIPE_DIRECTW) {
989 		if (wpipe->pipe_state & PIPE_WANTR) {
990 			wpipe->pipe_state &= ~PIPE_WANTR;
991 			wakeup(wpipe);
992 		}
993 		pipeselwakeup(wpipe);
994 		wpipe->pipe_state |= PIPE_WANTW;
995 		pipeunlock(wpipe);
996 		error = msleep(wpipe, PIPE_MTX(wpipe),
997 		    PRIBIO | PCATCH, "pipdww", 0);
998 		pipelock(wpipe, 0);
999 		if (error != 0)
1000 			goto error1;
1001 		goto retry;
1002 	}
1003 	if (wpipe->pipe_buffer.cnt > 0) {
1004 		if (wpipe->pipe_state & PIPE_WANTR) {
1005 			wpipe->pipe_state &= ~PIPE_WANTR;
1006 			wakeup(wpipe);
1007 		}
1008 		pipeselwakeup(wpipe);
1009 		wpipe->pipe_state |= PIPE_WANTW;
1010 		pipeunlock(wpipe);
1011 		error = msleep(wpipe, PIPE_MTX(wpipe),
1012 		    PRIBIO | PCATCH, "pipdwc", 0);
1013 		pipelock(wpipe, 0);
1014 		if (error != 0)
1015 			goto error1;
1016 		goto retry;
1017 	}
1018 
1019 	error = pipe_build_write_buffer(wpipe, uio);
1020 	if (error) {
1021 		goto error1;
1022 	}
1023 
1024 	while (wpipe->pipe_pages.cnt != 0 &&
1025 	    (wpipe->pipe_state & PIPE_EOF) == 0) {
1026 		if (wpipe->pipe_state & PIPE_WANTR) {
1027 			wpipe->pipe_state &= ~PIPE_WANTR;
1028 			wakeup(wpipe);
1029 		}
1030 		pipeselwakeup(wpipe);
1031 		wpipe->pipe_state |= PIPE_WANTW;
1032 		pipeunlock(wpipe);
1033 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1034 		    "pipdwt", 0);
1035 		pipelock(wpipe, 0);
1036 		if (error != 0)
1037 			break;
1038 	}
1039 
1040 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1041 		wpipe->pipe_pages.cnt = 0;
1042 		pipe_destroy_write_buffer(wpipe);
1043 		pipeselwakeup(wpipe);
1044 		error = EPIPE;
1045 	} else if (error == EINTR || error == ERESTART) {
1046 		pipe_clone_write_buffer(wpipe);
1047 	} else {
1048 		pipe_destroy_write_buffer(wpipe);
1049 	}
1050 	KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1051 	    ("pipe %p leaked PIPE_DIRECTW", wpipe));
1052 	return (error);
1053 
1054 error1:
1055 	wakeup(wpipe);
1056 	return (error);
1057 }
1058 #endif
1059 
1060 static int
1061 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1062     int flags, struct thread *td)
1063 {
1064 	struct pipe *wpipe, *rpipe;
1065 	ssize_t orig_resid;
1066 	int desiredsize, error;
1067 
1068 	rpipe = fp->f_data;
1069 	wpipe = PIPE_PEER(rpipe);
1070 	PIPE_LOCK(rpipe);
1071 	error = pipelock(wpipe, 1);
1072 	if (error) {
1073 		PIPE_UNLOCK(rpipe);
1074 		return (error);
1075 	}
1076 	/*
1077 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1078 	 */
1079 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1080 	    (wpipe->pipe_state & PIPE_EOF)) {
1081 		pipeunlock(wpipe);
1082 		PIPE_UNLOCK(rpipe);
1083 		return (EPIPE);
1084 	}
1085 #ifdef MAC
1086 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1087 	if (error) {
1088 		pipeunlock(wpipe);
1089 		PIPE_UNLOCK(rpipe);
1090 		return (error);
1091 	}
1092 #endif
1093 	++wpipe->pipe_busy;
1094 
1095 	/* Choose a larger size if it's advantageous */
1096 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1097 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1098 		if (piperesizeallowed != 1)
1099 			break;
1100 		if (amountpipekva > maxpipekva / 2)
1101 			break;
1102 		if (desiredsize == BIG_PIPE_SIZE)
1103 			break;
1104 		desiredsize = desiredsize * 2;
1105 	}
1106 
1107 	/* Choose a smaller size if we're in a OOM situation */
1108 	if (amountpipekva > (3 * maxpipekva) / 4 &&
1109 	    wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1110 	    wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1111 	    piperesizeallowed == 1)
1112 		desiredsize = SMALL_PIPE_SIZE;
1113 
1114 	/* Resize if the above determined that a new size was necessary */
1115 	if (desiredsize != wpipe->pipe_buffer.size &&
1116 	    (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1117 		PIPE_UNLOCK(wpipe);
1118 		pipespace(wpipe, desiredsize);
1119 		PIPE_LOCK(wpipe);
1120 	}
1121 	MPASS(wpipe->pipe_buffer.size != 0);
1122 
1123 	orig_resid = uio->uio_resid;
1124 
1125 	while (uio->uio_resid) {
1126 		int space;
1127 
1128 		if (wpipe->pipe_state & PIPE_EOF) {
1129 			error = EPIPE;
1130 			break;
1131 		}
1132 #ifndef PIPE_NODIRECT
1133 		/*
1134 		 * If the transfer is large, we can gain performance if
1135 		 * we do process-to-process copies directly.
1136 		 * If the write is non-blocking, we don't use the
1137 		 * direct write mechanism.
1138 		 *
1139 		 * The direct write mechanism will detect the reader going
1140 		 * away on us.
1141 		 */
1142 		if (uio->uio_segflg == UIO_USERSPACE &&
1143 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1144 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1145 		    (fp->f_flag & FNONBLOCK) == 0) {
1146 			error = pipe_direct_write(wpipe, uio);
1147 			if (error != 0)
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_pages.cnt != 0) {
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 			pipelock(wpipe, 0);
1171 			if (error != 0)
1172 				break;
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 			if (error != 0)
1246 				break;
1247 			continue;
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 				break;
1263 			}
1264 
1265 			/*
1266 			 * We have no more space and have something to offer,
1267 			 * wake up select/poll.
1268 			 */
1269 			pipeselwakeup(wpipe);
1270 
1271 			wpipe->pipe_state |= PIPE_WANTW;
1272 			pipeunlock(wpipe);
1273 			error = msleep(wpipe, PIPE_MTX(rpipe),
1274 			    PRIBIO | PCATCH, "pipewr", 0);
1275 			pipelock(wpipe, 0);
1276 			if (error != 0)
1277 				break;
1278 			continue;
1279 		}
1280 	}
1281 
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 		pipe_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 	if (uio->uio_resid != orig_resid)
1320 		td->td_ru.ru_msgsnd++;
1321 	return (error);
1322 }
1323 
1324 /* ARGSUSED */
1325 static int
1326 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1327     struct thread *td)
1328 {
1329 	struct pipe *cpipe;
1330 	int error;
1331 
1332 	cpipe = fp->f_data;
1333 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1334 		error = vnops.fo_truncate(fp, length, active_cred, td);
1335 	else
1336 		error = invfo_truncate(fp, length, active_cred, td);
1337 	return (error);
1338 }
1339 
1340 /*
1341  * we implement a very minimal set of ioctls for compatibility with sockets.
1342  */
1343 static int
1344 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1345     struct thread *td)
1346 {
1347 	struct pipe *mpipe = fp->f_data;
1348 	int error;
1349 
1350 	PIPE_LOCK(mpipe);
1351 
1352 #ifdef MAC
1353 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1354 	if (error) {
1355 		PIPE_UNLOCK(mpipe);
1356 		return (error);
1357 	}
1358 #endif
1359 
1360 	error = 0;
1361 	switch (cmd) {
1362 	case FIONBIO:
1363 		break;
1364 
1365 	case FIOASYNC:
1366 		if (*(int *)data) {
1367 			mpipe->pipe_state |= PIPE_ASYNC;
1368 		} else {
1369 			mpipe->pipe_state &= ~PIPE_ASYNC;
1370 		}
1371 		break;
1372 
1373 	case FIONREAD:
1374 		if (!(fp->f_flag & FREAD)) {
1375 			*(int *)data = 0;
1376 			PIPE_UNLOCK(mpipe);
1377 			return (0);
1378 		}
1379 		if (mpipe->pipe_pages.cnt != 0)
1380 			*(int *)data = mpipe->pipe_pages.cnt;
1381 		else
1382 			*(int *)data = mpipe->pipe_buffer.cnt;
1383 		break;
1384 
1385 	case FIOSETOWN:
1386 		PIPE_UNLOCK(mpipe);
1387 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1388 		goto out_unlocked;
1389 
1390 	case FIOGETOWN:
1391 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1392 		break;
1393 
1394 	/* This is deprecated, FIOSETOWN should be used instead. */
1395 	case TIOCSPGRP:
1396 		PIPE_UNLOCK(mpipe);
1397 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1398 		goto out_unlocked;
1399 
1400 	/* This is deprecated, FIOGETOWN should be used instead. */
1401 	case TIOCGPGRP:
1402 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1403 		break;
1404 
1405 	default:
1406 		error = ENOTTY;
1407 		break;
1408 	}
1409 	PIPE_UNLOCK(mpipe);
1410 out_unlocked:
1411 	return (error);
1412 }
1413 
1414 static int
1415 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1416     struct thread *td)
1417 {
1418 	struct pipe *rpipe;
1419 	struct pipe *wpipe;
1420 	int levents, revents;
1421 #ifdef MAC
1422 	int error;
1423 #endif
1424 
1425 	revents = 0;
1426 	rpipe = fp->f_data;
1427 	wpipe = PIPE_PEER(rpipe);
1428 	PIPE_LOCK(rpipe);
1429 #ifdef MAC
1430 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1431 	if (error)
1432 		goto locked_error;
1433 #endif
1434 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1435 		if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1436 			revents |= events & (POLLIN | POLLRDNORM);
1437 
1438 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1439 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1440 		    (wpipe->pipe_state & PIPE_EOF) ||
1441 		    ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1442 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1443 			 wpipe->pipe_buffer.size == 0)))
1444 			revents |= events & (POLLOUT | POLLWRNORM);
1445 
1446 	levents = events &
1447 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1448 	if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1449 	    fp->f_pipegen == rpipe->pipe_wgen)
1450 		events |= POLLINIGNEOF;
1451 
1452 	if ((events & POLLINIGNEOF) == 0) {
1453 		if (rpipe->pipe_state & PIPE_EOF) {
1454 			if (fp->f_flag & FREAD)
1455 				revents |= (events & (POLLIN | POLLRDNORM));
1456 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1457 			    (wpipe->pipe_state & PIPE_EOF))
1458 				revents |= POLLHUP;
1459 		}
1460 	}
1461 
1462 	if (revents == 0) {
1463 		/*
1464 		 * Add ourselves regardless of eventmask as we have to return
1465 		 * POLLHUP even if it was not asked for.
1466 		 */
1467 		if ((fp->f_flag & FREAD) != 0) {
1468 			selrecord(td, &rpipe->pipe_sel);
1469 			if (SEL_WAITING(&rpipe->pipe_sel))
1470 				rpipe->pipe_state |= PIPE_SEL;
1471 		}
1472 
1473 		if ((fp->f_flag & FWRITE) != 0) {
1474 			selrecord(td, &wpipe->pipe_sel);
1475 			if (SEL_WAITING(&wpipe->pipe_sel))
1476 				wpipe->pipe_state |= PIPE_SEL;
1477 		}
1478 	}
1479 #ifdef MAC
1480 locked_error:
1481 #endif
1482 	PIPE_UNLOCK(rpipe);
1483 
1484 	return (revents);
1485 }
1486 
1487 /*
1488  * We shouldn't need locks here as we're doing a read and this should
1489  * be a natural race.
1490  */
1491 static int
1492 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1493     struct thread *td)
1494 {
1495 	struct pipe *pipe;
1496 #ifdef MAC
1497 	int error;
1498 #endif
1499 
1500 	pipe = fp->f_data;
1501 #ifdef MAC
1502 	if (mac_pipe_check_stat_enabled()) {
1503 		PIPE_LOCK(pipe);
1504 		error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1505 		PIPE_UNLOCK(pipe);
1506 		if (error) {
1507 			return (error);
1508 		}
1509 	}
1510 #endif
1511 
1512 	/* For named pipes ask the underlying filesystem. */
1513 	if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1514 		return (vnops.fo_stat(fp, ub, active_cred, td));
1515 	}
1516 
1517 	bzero(ub, sizeof(*ub));
1518 	ub->st_mode = S_IFIFO;
1519 	ub->st_blksize = PAGE_SIZE;
1520 	if (pipe->pipe_pages.cnt != 0)
1521 		ub->st_size = pipe->pipe_pages.cnt;
1522 	else
1523 		ub->st_size = pipe->pipe_buffer.cnt;
1524 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1525 	ub->st_atim = pipe->pipe_atime;
1526 	ub->st_mtim = pipe->pipe_mtime;
1527 	ub->st_ctim = pipe->pipe_ctime;
1528 	ub->st_uid = fp->f_cred->cr_uid;
1529 	ub->st_gid = fp->f_cred->cr_gid;
1530 	ub->st_dev = pipedev_ino;
1531 	ub->st_ino = pipe->pipe_ino;
1532 	/*
1533 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1534 	 */
1535 	return (0);
1536 }
1537 
1538 /* ARGSUSED */
1539 static int
1540 pipe_close(struct file *fp, struct thread *td)
1541 {
1542 
1543 	if (fp->f_vnode != NULL)
1544 		return vnops.fo_close(fp, td);
1545 	fp->f_ops = &badfileops;
1546 	pipe_dtor(fp->f_data);
1547 	fp->f_data = NULL;
1548 	return (0);
1549 }
1550 
1551 static int
1552 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1553 {
1554 	struct pipe *cpipe;
1555 	int error;
1556 
1557 	cpipe = fp->f_data;
1558 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1559 		error = vn_chmod(fp, mode, active_cred, td);
1560 	else
1561 		error = invfo_chmod(fp, mode, active_cred, td);
1562 	return (error);
1563 }
1564 
1565 static int
1566 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1567     struct thread *td)
1568 {
1569 	struct pipe *cpipe;
1570 	int error;
1571 
1572 	cpipe = fp->f_data;
1573 	if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1574 		error = vn_chown(fp, uid, gid, active_cred, td);
1575 	else
1576 		error = invfo_chown(fp, uid, gid, active_cred, td);
1577 	return (error);
1578 }
1579 
1580 static int
1581 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1582 {
1583 	struct pipe *pi;
1584 
1585 	if (fp->f_type == DTYPE_FIFO)
1586 		return (vn_fill_kinfo(fp, kif, fdp));
1587 	kif->kf_type = KF_TYPE_PIPE;
1588 	pi = fp->f_data;
1589 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1590 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1591 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1592 	return (0);
1593 }
1594 
1595 static void
1596 pipe_free_kmem(struct pipe *cpipe)
1597 {
1598 
1599 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1600 	    ("pipe_free_kmem: pipe mutex locked"));
1601 
1602 	if (cpipe->pipe_buffer.buffer != NULL) {
1603 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1604 		vm_map_remove(pipe_map,
1605 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1606 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1607 		cpipe->pipe_buffer.buffer = NULL;
1608 	}
1609 #ifndef PIPE_NODIRECT
1610 	{
1611 		cpipe->pipe_pages.cnt = 0;
1612 		cpipe->pipe_pages.pos = 0;
1613 		cpipe->pipe_pages.npages = 0;
1614 	}
1615 #endif
1616 }
1617 
1618 /*
1619  * shutdown the pipe
1620  */
1621 static void
1622 pipeclose(struct pipe *cpipe)
1623 {
1624 	struct pipepair *pp;
1625 	struct pipe *ppipe;
1626 
1627 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1628 
1629 	PIPE_LOCK(cpipe);
1630 	pipelock(cpipe, 0);
1631 	pp = cpipe->pipe_pair;
1632 
1633 	/*
1634 	 * If the other side is blocked, wake it up saying that
1635 	 * we want to close it down.
1636 	 */
1637 	cpipe->pipe_state |= PIPE_EOF;
1638 	while (cpipe->pipe_busy) {
1639 		wakeup(cpipe);
1640 		cpipe->pipe_state |= PIPE_WANT;
1641 		pipeunlock(cpipe);
1642 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1643 		pipelock(cpipe, 0);
1644 	}
1645 
1646 	pipeselwakeup(cpipe);
1647 
1648 	/*
1649 	 * Disconnect from peer, if any.
1650 	 */
1651 	ppipe = cpipe->pipe_peer;
1652 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1653 		ppipe->pipe_state |= PIPE_EOF;
1654 		wakeup(ppipe);
1655 		pipeselwakeup(ppipe);
1656 	}
1657 
1658 	/*
1659 	 * Mark this endpoint as free.  Release kmem resources.  We
1660 	 * don't mark this endpoint as unused until we've finished
1661 	 * doing that, or the pipe might disappear out from under
1662 	 * us.
1663 	 */
1664 	PIPE_UNLOCK(cpipe);
1665 	pipe_free_kmem(cpipe);
1666 	PIPE_LOCK(cpipe);
1667 	cpipe->pipe_present = PIPE_CLOSING;
1668 	pipeunlock(cpipe);
1669 
1670 	/*
1671 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1672 	 * PIPE_FINALIZED, that allows other end to free the
1673 	 * pipe_pair, only after the knotes are completely dismantled.
1674 	 */
1675 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1676 	cpipe->pipe_present = PIPE_FINALIZED;
1677 	seldrain(&cpipe->pipe_sel);
1678 	knlist_destroy(&cpipe->pipe_sel.si_note);
1679 
1680 	/*
1681 	 * If both endpoints are now closed, release the memory for the
1682 	 * pipe pair.  If not, unlock.
1683 	 */
1684 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1685 		PIPE_UNLOCK(cpipe);
1686 #ifdef MAC
1687 		mac_pipe_destroy(pp);
1688 #endif
1689 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1690 	} else
1691 		PIPE_UNLOCK(cpipe);
1692 }
1693 
1694 /*ARGSUSED*/
1695 static int
1696 pipe_kqfilter(struct file *fp, struct knote *kn)
1697 {
1698 	struct pipe *cpipe;
1699 
1700 	/*
1701 	 * If a filter is requested that is not supported by this file
1702 	 * descriptor, don't return an error, but also don't ever generate an
1703 	 * event.
1704 	 */
1705 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1706 		kn->kn_fop = &pipe_nfiltops;
1707 		return (0);
1708 	}
1709 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1710 		kn->kn_fop = &pipe_nfiltops;
1711 		return (0);
1712 	}
1713 	cpipe = fp->f_data;
1714 	PIPE_LOCK(cpipe);
1715 	switch (kn->kn_filter) {
1716 	case EVFILT_READ:
1717 		kn->kn_fop = &pipe_rfiltops;
1718 		break;
1719 	case EVFILT_WRITE:
1720 		kn->kn_fop = &pipe_wfiltops;
1721 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1722 			/* other end of pipe has been closed */
1723 			PIPE_UNLOCK(cpipe);
1724 			return (EPIPE);
1725 		}
1726 		cpipe = PIPE_PEER(cpipe);
1727 		break;
1728 	default:
1729 		PIPE_UNLOCK(cpipe);
1730 		return (EINVAL);
1731 	}
1732 
1733 	kn->kn_hook = cpipe;
1734 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1735 	PIPE_UNLOCK(cpipe);
1736 	return (0);
1737 }
1738 
1739 static void
1740 filt_pipedetach(struct knote *kn)
1741 {
1742 	struct pipe *cpipe = kn->kn_hook;
1743 
1744 	PIPE_LOCK(cpipe);
1745 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1746 	PIPE_UNLOCK(cpipe);
1747 }
1748 
1749 /*ARGSUSED*/
1750 static int
1751 filt_piperead(struct knote *kn, long hint)
1752 {
1753 	struct file *fp = kn->kn_fp;
1754 	struct pipe *rpipe = kn->kn_hook;
1755 
1756 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1757 	kn->kn_data = rpipe->pipe_buffer.cnt;
1758 	if (kn->kn_data == 0)
1759 		kn->kn_data = rpipe->pipe_pages.cnt;
1760 
1761 	if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1762 	    ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1763 	    fp->f_pipegen != rpipe->pipe_wgen)) {
1764 		kn->kn_flags |= EV_EOF;
1765 		return (1);
1766 	}
1767 	kn->kn_flags &= ~EV_EOF;
1768 	return (kn->kn_data > 0);
1769 }
1770 
1771 /*ARGSUSED*/
1772 static int
1773 filt_pipewrite(struct knote *kn, long hint)
1774 {
1775 	struct pipe *wpipe = kn->kn_hook;
1776 
1777 	/*
1778 	 * If this end of the pipe is closed, the knote was removed from the
1779 	 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1780 	 */
1781 	if (wpipe->pipe_present == PIPE_ACTIVE ||
1782 	    (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1783 		PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1784 
1785 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1786 			kn->kn_data = 0;
1787 		} else if (wpipe->pipe_buffer.size > 0) {
1788 			kn->kn_data = wpipe->pipe_buffer.size -
1789 			    wpipe->pipe_buffer.cnt;
1790 		} else {
1791 			kn->kn_data = PIPE_BUF;
1792 		}
1793 	}
1794 
1795 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1796 	    (wpipe->pipe_state & PIPE_EOF)) {
1797 		kn->kn_flags |= EV_EOF;
1798 		return (1);
1799 	}
1800 	kn->kn_flags &= ~EV_EOF;
1801 	return (kn->kn_data >= PIPE_BUF);
1802 }
1803 
1804 static void
1805 filt_pipedetach_notsup(struct knote *kn)
1806 {
1807 
1808 }
1809 
1810 static int
1811 filt_pipenotsup(struct knote *kn, long hint)
1812 {
1813 
1814 	return (0);
1815 }
1816