xref: /freebsd/sys/kern/sys_pipe.c (revision 21a4258d89a4e27632cfd87e5ad6e8538a6e77a2)
1 /*-
2  * Copyright (c) 1996 John S. Dyson
3  * Copyright (c) 2012 Giovanni Trematerra
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice immediately at the beginning of the file, without modification,
11  *    this list of conditions, and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Absolutely no warranty of function or purpose is made by the author
16  *    John S. Dyson.
17  * 4. Modifications may be freely made to this file if the above conditions
18  *    are met.
19  */
20 
21 /*
22  * This file contains a high-performance replacement for the socket-based
23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
24  * all features of sockets, but does do everything that pipes normally
25  * do.
26  */
27 
28 /*
29  * This code has two modes of operation, a small write mode and a large
30  * write mode.  The small write mode acts like conventional pipes with
31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
34  * memory, and the receiving process copies directly from these pinned pages
35  * in the sending process.
36  *
37  * If the sending process receives a signal, it is possible that it will
38  * go away, and certainly its address space can change, because control
39  * is returned back to the user-mode side.  In that case, the pipe code
40  * arranges to copy the buffer supplied by the user process, to a pageable
41  * kernel buffer, and the receiving process will grab the data from the
42  * pageable kernel buffer.  Since signals don't happen all that often,
43  * the copy operation is normally eliminated.
44  *
45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
46  * happen for small transfers so that the system will not spend all of
47  * its time context switching.
48  *
49  * In order to limit the resource use of pipes, two sysctls exist:
50  *
51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
52  * address space available to us in pipe_map. This value is normally
53  * autotuned, but may also be loader tuned.
54  *
55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56  * memory in use by pipes.
57  *
58  * Based on how large pipekva is relative to maxpipekva, the following
59  * will happen:
60  *
61  * 0% - 50%:
62  *     New pipes are given 16K of memory backing, pipes may dynamically
63  *     grow to as large as 64K where needed.
64  * 50% - 75%:
65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
66  *     existing pipes may NOT grow.
67  * 75% - 100%:
68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
69  *     existing pipes will be shrunk down to 4K whenever possible.
70  *
71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
73  * resize which MUST occur for reverse-direction pipes when they are
74  * first used.
75  *
76  * Additional information about the current state of pipes may be obtained
77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
78  * and kern.ipc.piperesizefail.
79  *
80  * Locking rules:  There are two locks present here:  A mutex, used via
81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
82  * the flag, as mutexes can not persist over uiomove.  The mutex
83  * exists only to guard access to the flag, and is not in itself a
84  * locking mechanism.  Also note that there is only a single mutex for
85  * both directions of a pipe.
86  *
87  * As pipelock() may have to sleep before it can acquire the flag, it
88  * is important to reread all data after a call to pipelock(); everything
89  * in the structure may have changed.
90  */
91 
92 #include "opt_compat.h"
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 unrhdr *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 	pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
261 	KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
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 	vfs_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 void
340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
341 {
342 	struct pipepair *pp;
343 	struct pipe *rpipe, *wpipe;
344 
345 	*p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
346 #ifdef MAC
347 	/*
348 	 * The MAC label is shared between the connected endpoints.  As a
349 	 * result mac_pipe_init() and mac_pipe_create() are called once
350 	 * for the pair, and not on the endpoints.
351 	 */
352 	mac_pipe_init(pp);
353 	mac_pipe_create(td->td_ucred, pp);
354 #endif
355 	rpipe = &pp->pp_rpipe;
356 	wpipe = &pp->pp_wpipe;
357 
358 	knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
359 	knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
360 
361 	/* Only the forward direction pipe is backed by default */
362 	pipe_create(rpipe, 1);
363 	pipe_create(wpipe, 0);
364 
365 	rpipe->pipe_state |= PIPE_DIRECTOK;
366 	wpipe->pipe_state |= PIPE_DIRECTOK;
367 }
368 
369 void
370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
371 {
372 	struct pipepair *pp;
373 
374 	pipe_paircreate(td, &pp);
375 	pp->pp_rpipe.pipe_state |= PIPE_NAMED;
376 	*ppipe = &pp->pp_rpipe;
377 }
378 
379 void
380 pipe_dtor(struct pipe *dpipe)
381 {
382 	struct pipe *peer;
383 	ino_t ino;
384 
385 	ino = dpipe->pipe_ino;
386 	peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
387 	funsetown(&dpipe->pipe_sigio);
388 	pipeclose(dpipe);
389 	if (peer != NULL) {
390 		funsetown(&peer->pipe_sigio);
391 		pipeclose(peer);
392 	}
393 	if (ino != 0 && ino != (ino_t)-1)
394 		free_unr(pipeino_unr, ino);
395 }
396 
397 /*
398  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
399  * the zone pick up the pieces via pipeclose().
400  */
401 int
402 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
403     struct filecaps *fcaps2)
404 {
405 	struct file *rf, *wf;
406 	struct pipe *rpipe, *wpipe;
407 	struct pipepair *pp;
408 	int fd, fflags, error;
409 
410 	pipe_paircreate(td, &pp);
411 	rpipe = &pp->pp_rpipe;
412 	wpipe = &pp->pp_wpipe;
413 	error = falloc_caps(td, &rf, &fd, flags, fcaps1);
414 	if (error) {
415 		pipeclose(rpipe);
416 		pipeclose(wpipe);
417 		return (error);
418 	}
419 	/* An extra reference on `rf' has been held for us by falloc_caps(). */
420 	fildes[0] = fd;
421 
422 	fflags = FREAD | FWRITE;
423 	if ((flags & O_NONBLOCK) != 0)
424 		fflags |= FNONBLOCK;
425 
426 	/*
427 	 * Warning: once we've gotten past allocation of the fd for the
428 	 * read-side, we can only drop the read side via fdrop() in order
429 	 * to avoid races against processes which manage to dup() the read
430 	 * side while we are blocked trying to allocate the write side.
431 	 */
432 	finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
433 	error = falloc_caps(td, &wf, &fd, flags, fcaps2);
434 	if (error) {
435 		fdclose(td, rf, fildes[0]);
436 		fdrop(rf, td);
437 		/* rpipe has been closed by fdrop(). */
438 		pipeclose(wpipe);
439 		return (error);
440 	}
441 	/* An extra reference on `wf' has been held for us by falloc_caps(). */
442 	finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
443 	fdrop(wf, td);
444 	fildes[1] = fd;
445 	fdrop(rf, td);
446 
447 	return (0);
448 }
449 
450 #ifdef COMPAT_FREEBSD10
451 /* ARGSUSED */
452 int
453 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
454 {
455 	int error;
456 	int fildes[2];
457 
458 	error = kern_pipe(td, fildes, 0, NULL, NULL);
459 	if (error)
460 		return (error);
461 
462 	td->td_retval[0] = fildes[0];
463 	td->td_retval[1] = fildes[1];
464 
465 	return (0);
466 }
467 #endif
468 
469 int
470 sys_pipe2(struct thread *td, struct pipe2_args *uap)
471 {
472 	int error, fildes[2];
473 
474 	if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
475 		return (EINVAL);
476 	error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
477 	if (error)
478 		return (error);
479 	error = copyout(fildes, uap->fildes, 2 * sizeof(int));
480 	if (error) {
481 		(void)kern_close(td, fildes[0]);
482 		(void)kern_close(td, fildes[1]);
483 	}
484 	return (error);
485 }
486 
487 /*
488  * Allocate kva for pipe circular buffer, the space is pageable
489  * This routine will 'realloc' the size of a pipe safely, if it fails
490  * it will retain the old buffer.
491  * If it fails it will return ENOMEM.
492  */
493 static int
494 pipespace_new(cpipe, size)
495 	struct pipe *cpipe;
496 	int size;
497 {
498 	caddr_t buffer;
499 	int error, cnt, firstseg;
500 	static int curfail = 0;
501 	static struct timeval lastfail;
502 
503 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
504 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
505 		("pipespace: resize of direct writes not allowed"));
506 retry:
507 	cnt = cpipe->pipe_buffer.cnt;
508 	if (cnt > size)
509 		size = cnt;
510 
511 	size = round_page(size);
512 	buffer = (caddr_t) vm_map_min(pipe_map);
513 
514 	error = vm_map_find(pipe_map, NULL, 0,
515 		(vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
516 		VM_PROT_ALL, VM_PROT_ALL, 0);
517 	if (error != KERN_SUCCESS) {
518 		if ((cpipe->pipe_buffer.buffer == NULL) &&
519 			(size > SMALL_PIPE_SIZE)) {
520 			size = SMALL_PIPE_SIZE;
521 			pipefragretry++;
522 			goto retry;
523 		}
524 		if (cpipe->pipe_buffer.buffer == NULL) {
525 			pipeallocfail++;
526 			if (ppsratecheck(&lastfail, &curfail, 1))
527 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
528 		} else {
529 			piperesizefail++;
530 		}
531 		return (ENOMEM);
532 	}
533 
534 	/* copy data, then free old resources if we're resizing */
535 	if (cnt > 0) {
536 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
537 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
538 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
539 				buffer, firstseg);
540 			if ((cnt - firstseg) > 0)
541 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
542 					cpipe->pipe_buffer.in);
543 		} else {
544 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
545 				buffer, cnt);
546 		}
547 	}
548 	pipe_free_kmem(cpipe);
549 	cpipe->pipe_buffer.buffer = buffer;
550 	cpipe->pipe_buffer.size = size;
551 	cpipe->pipe_buffer.in = cnt;
552 	cpipe->pipe_buffer.out = 0;
553 	cpipe->pipe_buffer.cnt = cnt;
554 	atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
555 	return (0);
556 }
557 
558 /*
559  * Wrapper for pipespace_new() that performs locking assertions.
560  */
561 static int
562 pipespace(cpipe, size)
563 	struct pipe *cpipe;
564 	int size;
565 {
566 
567 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
568 		("Unlocked pipe passed to pipespace"));
569 	return (pipespace_new(cpipe, size));
570 }
571 
572 /*
573  * lock a pipe for I/O, blocking other access
574  */
575 static __inline int
576 pipelock(cpipe, catch)
577 	struct pipe *cpipe;
578 	int catch;
579 {
580 	int error;
581 
582 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
583 	while (cpipe->pipe_state & PIPE_LOCKFL) {
584 		cpipe->pipe_state |= PIPE_LWANT;
585 		error = msleep(cpipe, PIPE_MTX(cpipe),
586 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
587 		    "pipelk", 0);
588 		if (error != 0)
589 			return (error);
590 	}
591 	cpipe->pipe_state |= PIPE_LOCKFL;
592 	return (0);
593 }
594 
595 /*
596  * unlock a pipe I/O lock
597  */
598 static __inline void
599 pipeunlock(cpipe)
600 	struct pipe *cpipe;
601 {
602 
603 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
604 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
605 		("Unlocked pipe passed to pipeunlock"));
606 	cpipe->pipe_state &= ~PIPE_LOCKFL;
607 	if (cpipe->pipe_state & PIPE_LWANT) {
608 		cpipe->pipe_state &= ~PIPE_LWANT;
609 		wakeup(cpipe);
610 	}
611 }
612 
613 void
614 pipeselwakeup(cpipe)
615 	struct pipe *cpipe;
616 {
617 
618 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
619 	if (cpipe->pipe_state & PIPE_SEL) {
620 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
621 		if (!SEL_WAITING(&cpipe->pipe_sel))
622 			cpipe->pipe_state &= ~PIPE_SEL;
623 	}
624 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
625 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
626 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
627 }
628 
629 /*
630  * Initialize and allocate VM and memory for pipe.  The structure
631  * will start out zero'd from the ctor, so we just manage the kmem.
632  */
633 static void
634 pipe_create(pipe, backing)
635 	struct pipe *pipe;
636 	int backing;
637 {
638 
639 	if (backing) {
640 		/*
641 		 * Note that these functions can fail if pipe map is exhausted
642 		 * (as a result of too many pipes created), but we ignore the
643 		 * error as it is not fatal and could be provoked by
644 		 * unprivileged users. The only consequence is worse performance
645 		 * with given pipe.
646 		 */
647 		if (amountpipekva > maxpipekva / 2)
648 			(void)pipespace_new(pipe, SMALL_PIPE_SIZE);
649 		else
650 			(void)pipespace_new(pipe, PIPE_SIZE);
651 	}
652 
653 	pipe->pipe_ino = -1;
654 }
655 
656 /* ARGSUSED */
657 static int
658 pipe_read(fp, uio, active_cred, flags, td)
659 	struct file *fp;
660 	struct uio *uio;
661 	struct ucred *active_cred;
662 	struct thread *td;
663 	int flags;
664 {
665 	struct pipe *rpipe;
666 	int error;
667 	int nread = 0;
668 	int size;
669 
670 	rpipe = fp->f_data;
671 	PIPE_LOCK(rpipe);
672 	++rpipe->pipe_busy;
673 	error = pipelock(rpipe, 1);
674 	if (error)
675 		goto unlocked_error;
676 
677 #ifdef MAC
678 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
679 	if (error)
680 		goto locked_error;
681 #endif
682 	if (amountpipekva > (3 * maxpipekva) / 4) {
683 		if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
684 			(rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
685 			(rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
686 			(piperesizeallowed == 1)) {
687 			PIPE_UNLOCK(rpipe);
688 			pipespace(rpipe, SMALL_PIPE_SIZE);
689 			PIPE_LOCK(rpipe);
690 		}
691 	}
692 
693 	while (uio->uio_resid) {
694 		/*
695 		 * normal pipe buffer receive
696 		 */
697 		if (rpipe->pipe_buffer.cnt > 0) {
698 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
699 			if (size > rpipe->pipe_buffer.cnt)
700 				size = rpipe->pipe_buffer.cnt;
701 			if (size > uio->uio_resid)
702 				size = uio->uio_resid;
703 
704 			PIPE_UNLOCK(rpipe);
705 			error = uiomove(
706 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
707 			    size, uio);
708 			PIPE_LOCK(rpipe);
709 			if (error)
710 				break;
711 
712 			rpipe->pipe_buffer.out += size;
713 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
714 				rpipe->pipe_buffer.out = 0;
715 
716 			rpipe->pipe_buffer.cnt -= size;
717 
718 			/*
719 			 * If there is no more to read in the pipe, reset
720 			 * its pointers to the beginning.  This improves
721 			 * cache hit stats.
722 			 */
723 			if (rpipe->pipe_buffer.cnt == 0) {
724 				rpipe->pipe_buffer.in = 0;
725 				rpipe->pipe_buffer.out = 0;
726 			}
727 			nread += size;
728 #ifndef PIPE_NODIRECT
729 		/*
730 		 * Direct copy, bypassing a kernel buffer.
731 		 */
732 		} else if ((size = rpipe->pipe_map.cnt) &&
733 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
734 			if (size > uio->uio_resid)
735 				size = (u_int) uio->uio_resid;
736 
737 			PIPE_UNLOCK(rpipe);
738 			error = uiomove_fromphys(rpipe->pipe_map.ms,
739 			    rpipe->pipe_map.pos, size, uio);
740 			PIPE_LOCK(rpipe);
741 			if (error)
742 				break;
743 			nread += size;
744 			rpipe->pipe_map.pos += size;
745 			rpipe->pipe_map.cnt -= size;
746 			if (rpipe->pipe_map.cnt == 0) {
747 				rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
748 				wakeup(rpipe);
749 			}
750 #endif
751 		} else {
752 			/*
753 			 * detect EOF condition
754 			 * read returns 0 on EOF, no need to set error
755 			 */
756 			if (rpipe->pipe_state & PIPE_EOF)
757 				break;
758 
759 			/*
760 			 * If the "write-side" has been blocked, wake it up now.
761 			 */
762 			if (rpipe->pipe_state & PIPE_WANTW) {
763 				rpipe->pipe_state &= ~PIPE_WANTW;
764 				wakeup(rpipe);
765 			}
766 
767 			/*
768 			 * Break if some data was read.
769 			 */
770 			if (nread > 0)
771 				break;
772 
773 			/*
774 			 * Unlock the pipe buffer for our remaining processing.
775 			 * We will either break out with an error or we will
776 			 * sleep and relock to loop.
777 			 */
778 			pipeunlock(rpipe);
779 
780 			/*
781 			 * Handle non-blocking mode operation or
782 			 * wait for more data.
783 			 */
784 			if (fp->f_flag & FNONBLOCK) {
785 				error = EAGAIN;
786 			} else {
787 				rpipe->pipe_state |= PIPE_WANTR;
788 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
789 				    PRIBIO | PCATCH,
790 				    "piperd", 0)) == 0)
791 					error = pipelock(rpipe, 1);
792 			}
793 			if (error)
794 				goto unlocked_error;
795 		}
796 	}
797 #ifdef MAC
798 locked_error:
799 #endif
800 	pipeunlock(rpipe);
801 
802 	/* XXX: should probably do this before getting any locks. */
803 	if (error == 0)
804 		vfs_timestamp(&rpipe->pipe_atime);
805 unlocked_error:
806 	--rpipe->pipe_busy;
807 
808 	/*
809 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
810 	 */
811 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
812 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
813 		wakeup(rpipe);
814 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
815 		/*
816 		 * Handle write blocking hysteresis.
817 		 */
818 		if (rpipe->pipe_state & PIPE_WANTW) {
819 			rpipe->pipe_state &= ~PIPE_WANTW;
820 			wakeup(rpipe);
821 		}
822 	}
823 
824 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
825 		pipeselwakeup(rpipe);
826 
827 	PIPE_UNLOCK(rpipe);
828 	return (error);
829 }
830 
831 #ifndef PIPE_NODIRECT
832 /*
833  * Map the sending processes' buffer into kernel space and wire it.
834  * This is similar to a physical write operation.
835  */
836 static int
837 pipe_build_write_buffer(wpipe, uio)
838 	struct pipe *wpipe;
839 	struct uio *uio;
840 {
841 	u_int size;
842 	int i;
843 
844 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
845 	KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
846 		("Clone attempt on non-direct write pipe!"));
847 
848 	if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
849                 size = wpipe->pipe_buffer.size;
850 	else
851                 size = uio->uio_iov->iov_len;
852 
853 	if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
854 	    (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
855 	    wpipe->pipe_map.ms, PIPENPAGES)) < 0)
856 		return (EFAULT);
857 
858 /*
859  * set up the control block
860  */
861 	wpipe->pipe_map.npages = i;
862 	wpipe->pipe_map.pos =
863 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
864 	wpipe->pipe_map.cnt = size;
865 
866 /*
867  * and update the uio data
868  */
869 
870 	uio->uio_iov->iov_len -= size;
871 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
872 	if (uio->uio_iov->iov_len == 0)
873 		uio->uio_iov++;
874 	uio->uio_resid -= size;
875 	uio->uio_offset += size;
876 	return (0);
877 }
878 
879 /*
880  * unmap and unwire the process buffer
881  */
882 static void
883 pipe_destroy_write_buffer(wpipe)
884 	struct pipe *wpipe;
885 {
886 
887 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
888 	vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
889 	wpipe->pipe_map.npages = 0;
890 }
891 
892 /*
893  * In the case of a signal, the writing process might go away.  This
894  * code copies the data into the circular buffer so that the source
895  * pages can be freed without loss of data.
896  */
897 static void
898 pipe_clone_write_buffer(wpipe)
899 	struct pipe *wpipe;
900 {
901 	struct uio uio;
902 	struct iovec iov;
903 	int size;
904 	int pos;
905 
906 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
907 	size = wpipe->pipe_map.cnt;
908 	pos = wpipe->pipe_map.pos;
909 
910 	wpipe->pipe_buffer.in = size;
911 	wpipe->pipe_buffer.out = 0;
912 	wpipe->pipe_buffer.cnt = size;
913 	wpipe->pipe_state &= ~PIPE_DIRECTW;
914 
915 	PIPE_UNLOCK(wpipe);
916 	iov.iov_base = wpipe->pipe_buffer.buffer;
917 	iov.iov_len = size;
918 	uio.uio_iov = &iov;
919 	uio.uio_iovcnt = 1;
920 	uio.uio_offset = 0;
921 	uio.uio_resid = size;
922 	uio.uio_segflg = UIO_SYSSPACE;
923 	uio.uio_rw = UIO_READ;
924 	uio.uio_td = curthread;
925 	uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
926 	PIPE_LOCK(wpipe);
927 	pipe_destroy_write_buffer(wpipe);
928 }
929 
930 /*
931  * This implements the pipe buffer write mechanism.  Note that only
932  * a direct write OR a normal pipe write can be pending at any given time.
933  * If there are any characters in the pipe buffer, the direct write will
934  * be deferred until the receiving process grabs all of the bytes from
935  * the pipe buffer.  Then the direct mapping write is set-up.
936  */
937 static int
938 pipe_direct_write(wpipe, uio)
939 	struct pipe *wpipe;
940 	struct uio *uio;
941 {
942 	int error;
943 
944 retry:
945 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
946 	error = pipelock(wpipe, 1);
947 	if (error != 0)
948 		goto error1;
949 	if ((wpipe->pipe_state & PIPE_EOF) != 0) {
950 		error = EPIPE;
951 		pipeunlock(wpipe);
952 		goto error1;
953 	}
954 	while (wpipe->pipe_state & PIPE_DIRECTW) {
955 		if (wpipe->pipe_state & PIPE_WANTR) {
956 			wpipe->pipe_state &= ~PIPE_WANTR;
957 			wakeup(wpipe);
958 		}
959 		pipeselwakeup(wpipe);
960 		wpipe->pipe_state |= PIPE_WANTW;
961 		pipeunlock(wpipe);
962 		error = msleep(wpipe, PIPE_MTX(wpipe),
963 		    PRIBIO | PCATCH, "pipdww", 0);
964 		if (error)
965 			goto error1;
966 		else
967 			goto retry;
968 	}
969 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
970 	if (wpipe->pipe_buffer.cnt > 0) {
971 		if (wpipe->pipe_state & PIPE_WANTR) {
972 			wpipe->pipe_state &= ~PIPE_WANTR;
973 			wakeup(wpipe);
974 		}
975 		pipeselwakeup(wpipe);
976 		wpipe->pipe_state |= PIPE_WANTW;
977 		pipeunlock(wpipe);
978 		error = msleep(wpipe, PIPE_MTX(wpipe),
979 		    PRIBIO | PCATCH, "pipdwc", 0);
980 		if (error)
981 			goto error1;
982 		else
983 			goto retry;
984 	}
985 
986 	wpipe->pipe_state |= PIPE_DIRECTW;
987 
988 	PIPE_UNLOCK(wpipe);
989 	error = pipe_build_write_buffer(wpipe, uio);
990 	PIPE_LOCK(wpipe);
991 	if (error) {
992 		wpipe->pipe_state &= ~PIPE_DIRECTW;
993 		pipeunlock(wpipe);
994 		goto error1;
995 	}
996 
997 	error = 0;
998 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
999 		if (wpipe->pipe_state & PIPE_EOF) {
1000 			pipe_destroy_write_buffer(wpipe);
1001 			pipeselwakeup(wpipe);
1002 			pipeunlock(wpipe);
1003 			error = EPIPE;
1004 			goto error1;
1005 		}
1006 		if (wpipe->pipe_state & PIPE_WANTR) {
1007 			wpipe->pipe_state &= ~PIPE_WANTR;
1008 			wakeup(wpipe);
1009 		}
1010 		pipeselwakeup(wpipe);
1011 		wpipe->pipe_state |= PIPE_WANTW;
1012 		pipeunlock(wpipe);
1013 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1014 		    "pipdwt", 0);
1015 		pipelock(wpipe, 0);
1016 	}
1017 
1018 	if (wpipe->pipe_state & PIPE_EOF)
1019 		error = EPIPE;
1020 	if (wpipe->pipe_state & PIPE_DIRECTW) {
1021 		/*
1022 		 * this bit of trickery substitutes a kernel buffer for
1023 		 * the process that might be going away.
1024 		 */
1025 		pipe_clone_write_buffer(wpipe);
1026 	} else {
1027 		pipe_destroy_write_buffer(wpipe);
1028 	}
1029 	pipeunlock(wpipe);
1030 	return (error);
1031 
1032 error1:
1033 	wakeup(wpipe);
1034 	return (error);
1035 }
1036 #endif
1037 
1038 static int
1039 pipe_write(fp, uio, active_cred, flags, td)
1040 	struct file *fp;
1041 	struct uio *uio;
1042 	struct ucred *active_cred;
1043 	struct thread *td;
1044 	int flags;
1045 {
1046 	int error = 0;
1047 	int desiredsize;
1048 	ssize_t orig_resid;
1049 	struct pipe *wpipe, *rpipe;
1050 
1051 	rpipe = fp->f_data;
1052 	wpipe = PIPE_PEER(rpipe);
1053 	PIPE_LOCK(rpipe);
1054 	error = pipelock(wpipe, 1);
1055 	if (error) {
1056 		PIPE_UNLOCK(rpipe);
1057 		return (error);
1058 	}
1059 	/*
1060 	 * detect loss of pipe read side, issue SIGPIPE if lost.
1061 	 */
1062 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1063 	    (wpipe->pipe_state & PIPE_EOF)) {
1064 		pipeunlock(wpipe);
1065 		PIPE_UNLOCK(rpipe);
1066 		return (EPIPE);
1067 	}
1068 #ifdef MAC
1069 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1070 	if (error) {
1071 		pipeunlock(wpipe);
1072 		PIPE_UNLOCK(rpipe);
1073 		return (error);
1074 	}
1075 #endif
1076 	++wpipe->pipe_busy;
1077 
1078 	/* Choose a larger size if it's advantageous */
1079 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1080 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1081 		if (piperesizeallowed != 1)
1082 			break;
1083 		if (amountpipekva > maxpipekva / 2)
1084 			break;
1085 		if (desiredsize == BIG_PIPE_SIZE)
1086 			break;
1087 		desiredsize = desiredsize * 2;
1088 	}
1089 
1090 	/* Choose a smaller size if we're in a OOM situation */
1091 	if ((amountpipekva > (3 * maxpipekva) / 4) &&
1092 		(wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1093 		(wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1094 		(piperesizeallowed == 1))
1095 		desiredsize = SMALL_PIPE_SIZE;
1096 
1097 	/* Resize if the above determined that a new size was necessary */
1098 	if ((desiredsize != wpipe->pipe_buffer.size) &&
1099 		((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1100 		PIPE_UNLOCK(wpipe);
1101 		pipespace(wpipe, desiredsize);
1102 		PIPE_LOCK(wpipe);
1103 	}
1104 	if (wpipe->pipe_buffer.size == 0) {
1105 		/*
1106 		 * This can only happen for reverse direction use of pipes
1107 		 * in a complete OOM situation.
1108 		 */
1109 		error = ENOMEM;
1110 		--wpipe->pipe_busy;
1111 		pipeunlock(wpipe);
1112 		PIPE_UNLOCK(wpipe);
1113 		return (error);
1114 	}
1115 
1116 	pipeunlock(wpipe);
1117 
1118 	orig_resid = uio->uio_resid;
1119 
1120 	while (uio->uio_resid) {
1121 		int space;
1122 
1123 		pipelock(wpipe, 0);
1124 		if (wpipe->pipe_state & PIPE_EOF) {
1125 			pipeunlock(wpipe);
1126 			error = EPIPE;
1127 			break;
1128 		}
1129 #ifndef PIPE_NODIRECT
1130 		/*
1131 		 * If the transfer is large, we can gain performance if
1132 		 * we do process-to-process copies directly.
1133 		 * If the write is non-blocking, we don't use the
1134 		 * direct write mechanism.
1135 		 *
1136 		 * The direct write mechanism will detect the reader going
1137 		 * away on us.
1138 		 */
1139 		if (uio->uio_segflg == UIO_USERSPACE &&
1140 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1141 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1142 		    (fp->f_flag & FNONBLOCK) == 0) {
1143 			pipeunlock(wpipe);
1144 			error = pipe_direct_write(wpipe, uio);
1145 			if (error)
1146 				break;
1147 			continue;
1148 		}
1149 #endif
1150 
1151 		/*
1152 		 * Pipe buffered writes cannot be coincidental with
1153 		 * direct writes.  We wait until the currently executing
1154 		 * direct write is completed before we start filling the
1155 		 * pipe buffer.  We break out if a signal occurs or the
1156 		 * reader goes away.
1157 		 */
1158 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1159 			if (wpipe->pipe_state & PIPE_WANTR) {
1160 				wpipe->pipe_state &= ~PIPE_WANTR;
1161 				wakeup(wpipe);
1162 			}
1163 			pipeselwakeup(wpipe);
1164 			wpipe->pipe_state |= PIPE_WANTW;
1165 			pipeunlock(wpipe);
1166 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1167 			    "pipbww", 0);
1168 			if (error)
1169 				break;
1170 			else
1171 				continue;
1172 		}
1173 
1174 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1175 
1176 		/* Writes of size <= PIPE_BUF must be atomic. */
1177 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1178 			space = 0;
1179 
1180 		if (space > 0) {
1181 			int size;	/* Transfer size */
1182 			int segsize;	/* first segment to transfer */
1183 
1184 			/*
1185 			 * Transfer size is minimum of uio transfer
1186 			 * and free space in pipe buffer.
1187 			 */
1188 			if (space > uio->uio_resid)
1189 				size = uio->uio_resid;
1190 			else
1191 				size = space;
1192 			/*
1193 			 * First segment to transfer is minimum of
1194 			 * transfer size and contiguous space in
1195 			 * pipe buffer.  If first segment to transfer
1196 			 * is less than the transfer size, we've got
1197 			 * a wraparound in the buffer.
1198 			 */
1199 			segsize = wpipe->pipe_buffer.size -
1200 				wpipe->pipe_buffer.in;
1201 			if (segsize > size)
1202 				segsize = size;
1203 
1204 			/* Transfer first segment */
1205 
1206 			PIPE_UNLOCK(rpipe);
1207 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1208 					segsize, uio);
1209 			PIPE_LOCK(rpipe);
1210 
1211 			if (error == 0 && segsize < size) {
1212 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1213 					wpipe->pipe_buffer.size,
1214 					("Pipe buffer wraparound disappeared"));
1215 				/*
1216 				 * Transfer remaining part now, to
1217 				 * support atomic writes.  Wraparound
1218 				 * happened.
1219 				 */
1220 
1221 				PIPE_UNLOCK(rpipe);
1222 				error = uiomove(
1223 				    &wpipe->pipe_buffer.buffer[0],
1224 				    size - segsize, uio);
1225 				PIPE_LOCK(rpipe);
1226 			}
1227 			if (error == 0) {
1228 				wpipe->pipe_buffer.in += size;
1229 				if (wpipe->pipe_buffer.in >=
1230 				    wpipe->pipe_buffer.size) {
1231 					KASSERT(wpipe->pipe_buffer.in ==
1232 						size - segsize +
1233 						wpipe->pipe_buffer.size,
1234 						("Expected wraparound bad"));
1235 					wpipe->pipe_buffer.in = size - segsize;
1236 				}
1237 
1238 				wpipe->pipe_buffer.cnt += size;
1239 				KASSERT(wpipe->pipe_buffer.cnt <=
1240 					wpipe->pipe_buffer.size,
1241 					("Pipe buffer overflow"));
1242 			}
1243 			pipeunlock(wpipe);
1244 			if (error != 0)
1245 				break;
1246 		} else {
1247 			/*
1248 			 * If the "read-side" has been blocked, wake it up now.
1249 			 */
1250 			if (wpipe->pipe_state & PIPE_WANTR) {
1251 				wpipe->pipe_state &= ~PIPE_WANTR;
1252 				wakeup(wpipe);
1253 			}
1254 
1255 			/*
1256 			 * don't block on non-blocking I/O
1257 			 */
1258 			if (fp->f_flag & FNONBLOCK) {
1259 				error = EAGAIN;
1260 				pipeunlock(wpipe);
1261 				break;
1262 			}
1263 
1264 			/*
1265 			 * We have no more space and have something to offer,
1266 			 * wake up select/poll.
1267 			 */
1268 			pipeselwakeup(wpipe);
1269 
1270 			wpipe->pipe_state |= PIPE_WANTW;
1271 			pipeunlock(wpipe);
1272 			error = msleep(wpipe, PIPE_MTX(rpipe),
1273 			    PRIBIO | PCATCH, "pipewr", 0);
1274 			if (error != 0)
1275 				break;
1276 		}
1277 	}
1278 
1279 	pipelock(wpipe, 0);
1280 	--wpipe->pipe_busy;
1281 
1282 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1283 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1284 		wakeup(wpipe);
1285 	} else if (wpipe->pipe_buffer.cnt > 0) {
1286 		/*
1287 		 * If we have put any characters in the buffer, we wake up
1288 		 * the reader.
1289 		 */
1290 		if (wpipe->pipe_state & PIPE_WANTR) {
1291 			wpipe->pipe_state &= ~PIPE_WANTR;
1292 			wakeup(wpipe);
1293 		}
1294 	}
1295 
1296 	/*
1297 	 * Don't return EPIPE if any byte was written.
1298 	 * EINTR and other interrupts are handled by generic I/O layer.
1299 	 * Do not pretend that I/O succeeded for obvious user error
1300 	 * like EFAULT.
1301 	 */
1302 	if (uio->uio_resid != orig_resid && error == EPIPE)
1303 		error = 0;
1304 
1305 	if (error == 0)
1306 		vfs_timestamp(&wpipe->pipe_mtime);
1307 
1308 	/*
1309 	 * We have something to offer,
1310 	 * wake up select/poll.
1311 	 */
1312 	if (wpipe->pipe_buffer.cnt)
1313 		pipeselwakeup(wpipe);
1314 
1315 	pipeunlock(wpipe);
1316 	PIPE_UNLOCK(rpipe);
1317 	return (error);
1318 }
1319 
1320 /* ARGSUSED */
1321 static int
1322 pipe_truncate(fp, length, active_cred, td)
1323 	struct file *fp;
1324 	off_t length;
1325 	struct ucred *active_cred;
1326 	struct thread *td;
1327 {
1328 	struct pipe *cpipe;
1329 	int error;
1330 
1331 	cpipe = fp->f_data;
1332 	if (cpipe->pipe_state & PIPE_NAMED)
1333 		error = vnops.fo_truncate(fp, length, active_cred, td);
1334 	else
1335 		error = invfo_truncate(fp, length, active_cred, td);
1336 	return (error);
1337 }
1338 
1339 /*
1340  * we implement a very minimal set of ioctls for compatibility with sockets.
1341  */
1342 static int
1343 pipe_ioctl(fp, cmd, data, active_cred, td)
1344 	struct file *fp;
1345 	u_long cmd;
1346 	void *data;
1347 	struct ucred *active_cred;
1348 	struct thread *td;
1349 {
1350 	struct pipe *mpipe = fp->f_data;
1351 	int error;
1352 
1353 	PIPE_LOCK(mpipe);
1354 
1355 #ifdef MAC
1356 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1357 	if (error) {
1358 		PIPE_UNLOCK(mpipe);
1359 		return (error);
1360 	}
1361 #endif
1362 
1363 	error = 0;
1364 	switch (cmd) {
1365 
1366 	case FIONBIO:
1367 		break;
1368 
1369 	case FIOASYNC:
1370 		if (*(int *)data) {
1371 			mpipe->pipe_state |= PIPE_ASYNC;
1372 		} else {
1373 			mpipe->pipe_state &= ~PIPE_ASYNC;
1374 		}
1375 		break;
1376 
1377 	case FIONREAD:
1378 		if (!(fp->f_flag & FREAD)) {
1379 			*(int *)data = 0;
1380 			PIPE_UNLOCK(mpipe);
1381 			return (0);
1382 		}
1383 		if (mpipe->pipe_state & PIPE_DIRECTW)
1384 			*(int *)data = mpipe->pipe_map.cnt;
1385 		else
1386 			*(int *)data = mpipe->pipe_buffer.cnt;
1387 		break;
1388 
1389 	case FIOSETOWN:
1390 		PIPE_UNLOCK(mpipe);
1391 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1392 		goto out_unlocked;
1393 
1394 	case FIOGETOWN:
1395 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1396 		break;
1397 
1398 	/* This is deprecated, FIOSETOWN should be used instead. */
1399 	case TIOCSPGRP:
1400 		PIPE_UNLOCK(mpipe);
1401 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1402 		goto out_unlocked;
1403 
1404 	/* This is deprecated, FIOGETOWN should be used instead. */
1405 	case TIOCGPGRP:
1406 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1407 		break;
1408 
1409 	default:
1410 		error = ENOTTY;
1411 		break;
1412 	}
1413 	PIPE_UNLOCK(mpipe);
1414 out_unlocked:
1415 	return (error);
1416 }
1417 
1418 static int
1419 pipe_poll(fp, events, active_cred, td)
1420 	struct file *fp;
1421 	int events;
1422 	struct ucred *active_cred;
1423 	struct thread *td;
1424 {
1425 	struct pipe *rpipe;
1426 	struct pipe *wpipe;
1427 	int levents, revents;
1428 #ifdef MAC
1429 	int error;
1430 #endif
1431 
1432 	revents = 0;
1433 	rpipe = fp->f_data;
1434 	wpipe = PIPE_PEER(rpipe);
1435 	PIPE_LOCK(rpipe);
1436 #ifdef MAC
1437 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1438 	if (error)
1439 		goto locked_error;
1440 #endif
1441 	if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1442 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1443 		    (rpipe->pipe_buffer.cnt > 0))
1444 			revents |= events & (POLLIN | POLLRDNORM);
1445 
1446 	if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1447 		if (wpipe->pipe_present != PIPE_ACTIVE ||
1448 		    (wpipe->pipe_state & PIPE_EOF) ||
1449 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1450 		     ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1451 			 wpipe->pipe_buffer.size == 0)))
1452 			revents |= events & (POLLOUT | POLLWRNORM);
1453 
1454 	levents = events &
1455 	    (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1456 	if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1457 	    fp->f_seqcount == rpipe->pipe_wgen)
1458 		events |= POLLINIGNEOF;
1459 
1460 	if ((events & POLLINIGNEOF) == 0) {
1461 		if (rpipe->pipe_state & PIPE_EOF) {
1462 			revents |= (events & (POLLIN | POLLRDNORM));
1463 			if (wpipe->pipe_present != PIPE_ACTIVE ||
1464 			    (wpipe->pipe_state & PIPE_EOF))
1465 				revents |= POLLHUP;
1466 		}
1467 	}
1468 
1469 	if (revents == 0) {
1470 		if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1471 			selrecord(td, &rpipe->pipe_sel);
1472 			if (SEL_WAITING(&rpipe->pipe_sel))
1473 				rpipe->pipe_state |= PIPE_SEL;
1474 		}
1475 
1476 		if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1477 			selrecord(td, &wpipe->pipe_sel);
1478 			if (SEL_WAITING(&wpipe->pipe_sel))
1479 				wpipe->pipe_state |= PIPE_SEL;
1480 		}
1481 	}
1482 #ifdef MAC
1483 locked_error:
1484 #endif
1485 	PIPE_UNLOCK(rpipe);
1486 
1487 	return (revents);
1488 }
1489 
1490 /*
1491  * We shouldn't need locks here as we're doing a read and this should
1492  * be a natural race.
1493  */
1494 static int
1495 pipe_stat(fp, ub, active_cred, td)
1496 	struct file *fp;
1497 	struct stat *ub;
1498 	struct ucred *active_cred;
1499 	struct thread *td;
1500 {
1501 	struct pipe *pipe;
1502 	int new_unr;
1503 #ifdef MAC
1504 	int error;
1505 #endif
1506 
1507 	pipe = fp->f_data;
1508 	PIPE_LOCK(pipe);
1509 #ifdef MAC
1510 	error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1511 	if (error) {
1512 		PIPE_UNLOCK(pipe);
1513 		return (error);
1514 	}
1515 #endif
1516 
1517 	/* For named pipes ask the underlying filesystem. */
1518 	if (pipe->pipe_state & PIPE_NAMED) {
1519 		PIPE_UNLOCK(pipe);
1520 		return (vnops.fo_stat(fp, ub, active_cred, td));
1521 	}
1522 
1523 	/*
1524 	 * Lazily allocate an inode number for the pipe.  Most pipe
1525 	 * users do not call fstat(2) on the pipe, which means that
1526 	 * postponing the inode allocation until it is must be
1527 	 * returned to userland is useful.  If alloc_unr failed,
1528 	 * assign st_ino zero instead of returning an error.
1529 	 * Special pipe_ino values:
1530 	 *  -1 - not yet initialized;
1531 	 *  0  - alloc_unr failed, return 0 as st_ino forever.
1532 	 */
1533 	if (pipe->pipe_ino == (ino_t)-1) {
1534 		new_unr = alloc_unr(pipeino_unr);
1535 		if (new_unr != -1)
1536 			pipe->pipe_ino = new_unr;
1537 		else
1538 			pipe->pipe_ino = 0;
1539 	}
1540 	PIPE_UNLOCK(pipe);
1541 
1542 	bzero(ub, sizeof(*ub));
1543 	ub->st_mode = S_IFIFO;
1544 	ub->st_blksize = PAGE_SIZE;
1545 	if (pipe->pipe_state & PIPE_DIRECTW)
1546 		ub->st_size = pipe->pipe_map.cnt;
1547 	else
1548 		ub->st_size = pipe->pipe_buffer.cnt;
1549 	ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1550 	ub->st_atim = pipe->pipe_atime;
1551 	ub->st_mtim = pipe->pipe_mtime;
1552 	ub->st_ctim = pipe->pipe_ctime;
1553 	ub->st_uid = fp->f_cred->cr_uid;
1554 	ub->st_gid = fp->f_cred->cr_gid;
1555 	ub->st_dev = pipedev_ino;
1556 	ub->st_ino = pipe->pipe_ino;
1557 	/*
1558 	 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1559 	 */
1560 	return (0);
1561 }
1562 
1563 /* ARGSUSED */
1564 static int
1565 pipe_close(fp, td)
1566 	struct file *fp;
1567 	struct thread *td;
1568 {
1569 
1570 	if (fp->f_vnode != NULL)
1571 		return vnops.fo_close(fp, td);
1572 	fp->f_ops = &badfileops;
1573 	pipe_dtor(fp->f_data);
1574 	fp->f_data = NULL;
1575 	return (0);
1576 }
1577 
1578 static int
1579 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1580 {
1581 	struct pipe *cpipe;
1582 	int error;
1583 
1584 	cpipe = fp->f_data;
1585 	if (cpipe->pipe_state & PIPE_NAMED)
1586 		error = vn_chmod(fp, mode, active_cred, td);
1587 	else
1588 		error = invfo_chmod(fp, mode, active_cred, td);
1589 	return (error);
1590 }
1591 
1592 static int
1593 pipe_chown(fp, uid, gid, active_cred, td)
1594 	struct file *fp;
1595 	uid_t uid;
1596 	gid_t gid;
1597 	struct ucred *active_cred;
1598 	struct thread *td;
1599 {
1600 	struct pipe *cpipe;
1601 	int error;
1602 
1603 	cpipe = fp->f_data;
1604 	if (cpipe->pipe_state & PIPE_NAMED)
1605 		error = vn_chown(fp, uid, gid, active_cred, td);
1606 	else
1607 		error = invfo_chown(fp, uid, gid, active_cred, td);
1608 	return (error);
1609 }
1610 
1611 static int
1612 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1613 {
1614 	struct pipe *pi;
1615 
1616 	if (fp->f_type == DTYPE_FIFO)
1617 		return (vn_fill_kinfo(fp, kif, fdp));
1618 	kif->kf_type = KF_TYPE_PIPE;
1619 	pi = fp->f_data;
1620 	kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1621 	kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1622 	kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1623 	return (0);
1624 }
1625 
1626 static void
1627 pipe_free_kmem(cpipe)
1628 	struct pipe *cpipe;
1629 {
1630 
1631 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1632 	    ("pipe_free_kmem: pipe mutex locked"));
1633 
1634 	if (cpipe->pipe_buffer.buffer != NULL) {
1635 		atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1636 		vm_map_remove(pipe_map,
1637 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1638 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1639 		cpipe->pipe_buffer.buffer = NULL;
1640 	}
1641 #ifndef PIPE_NODIRECT
1642 	{
1643 		cpipe->pipe_map.cnt = 0;
1644 		cpipe->pipe_map.pos = 0;
1645 		cpipe->pipe_map.npages = 0;
1646 	}
1647 #endif
1648 }
1649 
1650 /*
1651  * shutdown the pipe
1652  */
1653 static void
1654 pipeclose(cpipe)
1655 	struct pipe *cpipe;
1656 {
1657 	struct pipepair *pp;
1658 	struct pipe *ppipe;
1659 
1660 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1661 
1662 	PIPE_LOCK(cpipe);
1663 	pipelock(cpipe, 0);
1664 	pp = cpipe->pipe_pair;
1665 
1666 	pipeselwakeup(cpipe);
1667 
1668 	/*
1669 	 * If the other side is blocked, wake it up saying that
1670 	 * we want to close it down.
1671 	 */
1672 	cpipe->pipe_state |= PIPE_EOF;
1673 	while (cpipe->pipe_busy) {
1674 		wakeup(cpipe);
1675 		cpipe->pipe_state |= PIPE_WANT;
1676 		pipeunlock(cpipe);
1677 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1678 		pipelock(cpipe, 0);
1679 	}
1680 
1681 
1682 	/*
1683 	 * Disconnect from peer, if any.
1684 	 */
1685 	ppipe = cpipe->pipe_peer;
1686 	if (ppipe->pipe_present == PIPE_ACTIVE) {
1687 		pipeselwakeup(ppipe);
1688 
1689 		ppipe->pipe_state |= PIPE_EOF;
1690 		wakeup(ppipe);
1691 		KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1692 	}
1693 
1694 	/*
1695 	 * Mark this endpoint as free.  Release kmem resources.  We
1696 	 * don't mark this endpoint as unused until we've finished
1697 	 * doing that, or the pipe might disappear out from under
1698 	 * us.
1699 	 */
1700 	PIPE_UNLOCK(cpipe);
1701 	pipe_free_kmem(cpipe);
1702 	PIPE_LOCK(cpipe);
1703 	cpipe->pipe_present = PIPE_CLOSING;
1704 	pipeunlock(cpipe);
1705 
1706 	/*
1707 	 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1708 	 * PIPE_FINALIZED, that allows other end to free the
1709 	 * pipe_pair, only after the knotes are completely dismantled.
1710 	 */
1711 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1712 	cpipe->pipe_present = PIPE_FINALIZED;
1713 	seldrain(&cpipe->pipe_sel);
1714 	knlist_destroy(&cpipe->pipe_sel.si_note);
1715 
1716 	/*
1717 	 * If both endpoints are now closed, release the memory for the
1718 	 * pipe pair.  If not, unlock.
1719 	 */
1720 	if (ppipe->pipe_present == PIPE_FINALIZED) {
1721 		PIPE_UNLOCK(cpipe);
1722 #ifdef MAC
1723 		mac_pipe_destroy(pp);
1724 #endif
1725 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1726 	} else
1727 		PIPE_UNLOCK(cpipe);
1728 }
1729 
1730 /*ARGSUSED*/
1731 static int
1732 pipe_kqfilter(struct file *fp, struct knote *kn)
1733 {
1734 	struct pipe *cpipe;
1735 
1736 	/*
1737 	 * If a filter is requested that is not supported by this file
1738 	 * descriptor, don't return an error, but also don't ever generate an
1739 	 * event.
1740 	 */
1741 	if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1742 		kn->kn_fop = &pipe_nfiltops;
1743 		return (0);
1744 	}
1745 	if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1746 		kn->kn_fop = &pipe_nfiltops;
1747 		return (0);
1748 	}
1749 	cpipe = fp->f_data;
1750 	PIPE_LOCK(cpipe);
1751 	switch (kn->kn_filter) {
1752 	case EVFILT_READ:
1753 		kn->kn_fop = &pipe_rfiltops;
1754 		break;
1755 	case EVFILT_WRITE:
1756 		kn->kn_fop = &pipe_wfiltops;
1757 		if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1758 			/* other end of pipe has been closed */
1759 			PIPE_UNLOCK(cpipe);
1760 			return (EPIPE);
1761 		}
1762 		cpipe = PIPE_PEER(cpipe);
1763 		break;
1764 	default:
1765 		PIPE_UNLOCK(cpipe);
1766 		return (EINVAL);
1767 	}
1768 
1769 	kn->kn_hook = cpipe;
1770 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1771 	PIPE_UNLOCK(cpipe);
1772 	return (0);
1773 }
1774 
1775 static void
1776 filt_pipedetach(struct knote *kn)
1777 {
1778 	struct pipe *cpipe = kn->kn_hook;
1779 
1780 	PIPE_LOCK(cpipe);
1781 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1782 	PIPE_UNLOCK(cpipe);
1783 }
1784 
1785 /*ARGSUSED*/
1786 static int
1787 filt_piperead(struct knote *kn, long hint)
1788 {
1789 	struct pipe *rpipe = kn->kn_hook;
1790 	struct pipe *wpipe = rpipe->pipe_peer;
1791 	int ret;
1792 
1793 	PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1794 	kn->kn_data = rpipe->pipe_buffer.cnt;
1795 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1796 		kn->kn_data = rpipe->pipe_map.cnt;
1797 
1798 	if ((rpipe->pipe_state & PIPE_EOF) ||
1799 	    wpipe->pipe_present != PIPE_ACTIVE ||
1800 	    (wpipe->pipe_state & PIPE_EOF)) {
1801 		kn->kn_flags |= EV_EOF;
1802 		return (1);
1803 	}
1804 	ret = kn->kn_data > 0;
1805 	return ret;
1806 }
1807 
1808 /*ARGSUSED*/
1809 static int
1810 filt_pipewrite(struct knote *kn, long hint)
1811 {
1812 	struct pipe *wpipe;
1813 
1814 	wpipe = kn->kn_hook;
1815 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1816 	if (wpipe->pipe_present != PIPE_ACTIVE ||
1817 	    (wpipe->pipe_state & PIPE_EOF)) {
1818 		kn->kn_data = 0;
1819 		kn->kn_flags |= EV_EOF;
1820 		return (1);
1821 	}
1822 	kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1823 	    (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1824 	if (wpipe->pipe_state & PIPE_DIRECTW)
1825 		kn->kn_data = 0;
1826 
1827 	return (kn->kn_data >= PIPE_BUF);
1828 }
1829 
1830 static void
1831 filt_pipedetach_notsup(struct knote *kn)
1832 {
1833 
1834 }
1835 
1836 static int
1837 filt_pipenotsup(struct knote *kn, long hint)
1838 {
1839 
1840 	return (0);
1841 }
1842