xref: /freebsd/sys/kern/sys_pipe.c (revision 2be1a816b9ff69588e55be0a84cbe2a31efc0f2f)
1 /*-
2  * Copyright (c) 1996 John S. Dyson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice immediately at the beginning of the file, without modification,
10  *    this list of conditions, and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. Absolutely no warranty of function or purpose is made by the author
15  *    John S. Dyson.
16  * 4. Modifications may be freely made to this file if the above conditions
17  *    are met.
18  */
19 
20 /*
21  * This file contains a high-performance replacement for the socket-based
22  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
23  * all features of sockets, but does do everything that pipes normally
24  * do.
25  */
26 
27 /*
28  * This code has two modes of operation, a small write mode and a large
29  * write mode.  The small write mode acts like conventional pipes with
30  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
31  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
32  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
33  * the receiving process can copy it directly from the pages in the sending
34  * process.
35  *
36  * If the sending process receives a signal, it is possible that it will
37  * go away, and certainly its address space can change, because control
38  * is returned back to the user-mode side.  In that case, the pipe code
39  * arranges to copy the buffer supplied by the user process, to a pageable
40  * kernel buffer, and the receiving process will grab the data from the
41  * pageable kernel buffer.  Since signals don't happen all that often,
42  * the copy operation is normally eliminated.
43  *
44  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
45  * happen for small transfers so that the system will not spend all of
46  * its time context switching.
47  *
48  * In order to limit the resource use of pipes, two sysctls exist:
49  *
50  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
51  * address space available to us in pipe_map. This value is normally
52  * autotuned, but may also be loader tuned.
53  *
54  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
55  * memory in use by pipes.
56  *
57  * Based on how large pipekva is relative to maxpipekva, the following
58  * will happen:
59  *
60  * 0% - 50%:
61  *     New pipes are given 16K of memory backing, pipes may dynamically
62  *     grow to as large as 64K where needed.
63  * 50% - 75%:
64  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
65  *     existing pipes may NOT grow.
66  * 75% - 100%:
67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
68  *     existing pipes will be shrunk down to 4K whenever possible.
69  *
70  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
71  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
72  * resize which MUST occur for reverse-direction pipes when they are
73  * first used.
74  *
75  * Additional information about the current state of pipes may be obtained
76  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
77  * and kern.ipc.piperesizefail.
78  *
79  * Locking rules:  There are two locks present here:  A mutex, used via
80  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
81  * the flag, as mutexes can not persist over uiomove.  The mutex
82  * exists only to guard access to the flag, and is not in itself a
83  * locking mechanism.  Also note that there is only a single mutex for
84  * both directions of a pipe.
85  *
86  * As pipelock() may have to sleep before it can acquire the flag, it
87  * is important to reread all data after a call to pipelock(); everything
88  * in the structure may have changed.
89  */
90 
91 #include <sys/cdefs.h>
92 __FBSDID("$FreeBSD$");
93 
94 #include "opt_mac.h"
95 
96 #include <sys/param.h>
97 #include <sys/systm.h>
98 #include <sys/fcntl.h>
99 #include <sys/file.h>
100 #include <sys/filedesc.h>
101 #include <sys/filio.h>
102 #include <sys/kernel.h>
103 #include <sys/lock.h>
104 #include <sys/mutex.h>
105 #include <sys/ttycom.h>
106 #include <sys/stat.h>
107 #include <sys/malloc.h>
108 #include <sys/poll.h>
109 #include <sys/selinfo.h>
110 #include <sys/signalvar.h>
111 #include <sys/sysctl.h>
112 #include <sys/sysproto.h>
113 #include <sys/pipe.h>
114 #include <sys/proc.h>
115 #include <sys/vnode.h>
116 #include <sys/uio.h>
117 #include <sys/event.h>
118 
119 #include <security/mac/mac_framework.h>
120 
121 #include <vm/vm.h>
122 #include <vm/vm_param.h>
123 #include <vm/vm_object.h>
124 #include <vm/vm_kern.h>
125 #include <vm/vm_extern.h>
126 #include <vm/pmap.h>
127 #include <vm/vm_map.h>
128 #include <vm/vm_page.h>
129 #include <vm/uma.h>
130 
131 /*
132  * Use this define if you want to disable *fancy* VM things.  Expect an
133  * approx 30% decrease in transfer rate.  This could be useful for
134  * NetBSD or OpenBSD.
135  */
136 /* #define PIPE_NODIRECT */
137 
138 /*
139  * interfaces to the outside world
140  */
141 static fo_rdwr_t	pipe_read;
142 static fo_rdwr_t	pipe_write;
143 static fo_truncate_t	pipe_truncate;
144 static fo_ioctl_t	pipe_ioctl;
145 static fo_poll_t	pipe_poll;
146 static fo_kqfilter_t	pipe_kqfilter;
147 static fo_stat_t	pipe_stat;
148 static fo_close_t	pipe_close;
149 
150 static struct fileops pipeops = {
151 	.fo_read = pipe_read,
152 	.fo_write = pipe_write,
153 	.fo_truncate = pipe_truncate,
154 	.fo_ioctl = pipe_ioctl,
155 	.fo_poll = pipe_poll,
156 	.fo_kqfilter = pipe_kqfilter,
157 	.fo_stat = pipe_stat,
158 	.fo_close = pipe_close,
159 	.fo_flags = DFLAG_PASSABLE
160 };
161 
162 static void	filt_pipedetach(struct knote *kn);
163 static int	filt_piperead(struct knote *kn, long hint);
164 static int	filt_pipewrite(struct knote *kn, long hint);
165 
166 static struct filterops pipe_rfiltops =
167 	{ 1, NULL, filt_pipedetach, filt_piperead };
168 static struct filterops pipe_wfiltops =
169 	{ 1, NULL, filt_pipedetach, filt_pipewrite };
170 
171 /*
172  * Default pipe buffer size(s), this can be kind-of large now because pipe
173  * space is pageable.  The pipe code will try to maintain locality of
174  * reference for performance reasons, so small amounts of outstanding I/O
175  * will not wipe the cache.
176  */
177 #define MINPIPESIZE (PIPE_SIZE/3)
178 #define MAXPIPESIZE (2*PIPE_SIZE/3)
179 
180 static int amountpipekva;
181 static int pipefragretry;
182 static int pipeallocfail;
183 static int piperesizefail;
184 static int piperesizeallowed = 1;
185 
186 SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
187 	   &maxpipekva, 0, "Pipe KVA limit");
188 SYSCTL_INT(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
189 	   &amountpipekva, 0, "Pipe KVA usage");
190 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
191 	  &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
192 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
193 	  &pipeallocfail, 0, "Pipe allocation failures");
194 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
195 	  &piperesizefail, 0, "Pipe resize failures");
196 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
197 	  &piperesizeallowed, 0, "Pipe resizing allowed");
198 
199 static void pipeinit(void *dummy __unused);
200 static void pipeclose(struct pipe *cpipe);
201 static void pipe_free_kmem(struct pipe *cpipe);
202 static int pipe_create(struct pipe *pipe, int backing);
203 static __inline int pipelock(struct pipe *cpipe, int catch);
204 static __inline void pipeunlock(struct pipe *cpipe);
205 static __inline void pipeselwakeup(struct pipe *cpipe);
206 #ifndef PIPE_NODIRECT
207 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
208 static void pipe_destroy_write_buffer(struct pipe *wpipe);
209 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
210 static void pipe_clone_write_buffer(struct pipe *wpipe);
211 #endif
212 static int pipespace(struct pipe *cpipe, int size);
213 static int pipespace_new(struct pipe *cpipe, int size);
214 
215 static int	pipe_zone_ctor(void *mem, int size, void *arg, int flags);
216 static int	pipe_zone_init(void *mem, int size, int flags);
217 static void	pipe_zone_fini(void *mem, int size);
218 
219 static uma_zone_t pipe_zone;
220 
221 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
222 
223 static void
224 pipeinit(void *dummy __unused)
225 {
226 
227 	pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
228 	    pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
229 	    UMA_ALIGN_PTR, 0);
230 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
231 }
232 
233 static int
234 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
235 {
236 	struct pipepair *pp;
237 	struct pipe *rpipe, *wpipe;
238 
239 	KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
240 
241 	pp = (struct pipepair *)mem;
242 
243 	/*
244 	 * We zero both pipe endpoints to make sure all the kmem pointers
245 	 * are NULL, flag fields are zero'd, etc.  We timestamp both
246 	 * endpoints with the same time.
247 	 */
248 	rpipe = &pp->pp_rpipe;
249 	bzero(rpipe, sizeof(*rpipe));
250 	vfs_timestamp(&rpipe->pipe_ctime);
251 	rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
252 
253 	wpipe = &pp->pp_wpipe;
254 	bzero(wpipe, sizeof(*wpipe));
255 	wpipe->pipe_ctime = rpipe->pipe_ctime;
256 	wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
257 
258 	rpipe->pipe_peer = wpipe;
259 	rpipe->pipe_pair = pp;
260 	wpipe->pipe_peer = rpipe;
261 	wpipe->pipe_pair = pp;
262 
263 	/*
264 	 * Mark both endpoints as present; they will later get free'd
265 	 * one at a time.  When both are free'd, then the whole pair
266 	 * is released.
267 	 */
268 	rpipe->pipe_present = 1;
269 	wpipe->pipe_present = 1;
270 
271 	/*
272 	 * Eventually, the MAC Framework may initialize the label
273 	 * in ctor or init, but for now we do it elswhere to avoid
274 	 * blocking in ctor or init.
275 	 */
276 	pp->pp_label = NULL;
277 
278 	return (0);
279 }
280 
281 static int
282 pipe_zone_init(void *mem, int size, int flags)
283 {
284 	struct pipepair *pp;
285 
286 	KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
287 
288 	pp = (struct pipepair *)mem;
289 
290 	mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
291 	return (0);
292 }
293 
294 static void
295 pipe_zone_fini(void *mem, int size)
296 {
297 	struct pipepair *pp;
298 
299 	KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
300 
301 	pp = (struct pipepair *)mem;
302 
303 	mtx_destroy(&pp->pp_mtx);
304 }
305 
306 /*
307  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
308  * the zone pick up the pieces via pipeclose().
309  */
310 /* ARGSUSED */
311 int
312 pipe(td, uap)
313 	struct thread *td;
314 	struct pipe_args /* {
315 		int	dummy;
316 	} */ *uap;
317 {
318 	struct filedesc *fdp = td->td_proc->p_fd;
319 	struct file *rf, *wf;
320 	struct pipepair *pp;
321 	struct pipe *rpipe, *wpipe;
322 	int fd, error;
323 
324 	pp = uma_zalloc(pipe_zone, M_WAITOK);
325 #ifdef MAC
326 	/*
327 	 * The MAC label is shared between the connected endpoints.  As a
328 	 * result mac_pipe_init() and mac_pipe_create() are called once
329 	 * for the pair, and not on the endpoints.
330 	 */
331 	mac_pipe_init(pp);
332 	mac_pipe_create(td->td_ucred, pp);
333 #endif
334 	rpipe = &pp->pp_rpipe;
335 	wpipe = &pp->pp_wpipe;
336 
337 	knlist_init(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe), NULL, NULL,
338 	    NULL);
339 	knlist_init(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe), NULL, NULL,
340 	    NULL);
341 
342 	/* Only the forward direction pipe is backed by default */
343 	if ((error = pipe_create(rpipe, 1)) != 0 ||
344 	    (error = pipe_create(wpipe, 0)) != 0) {
345 		pipeclose(rpipe);
346 		pipeclose(wpipe);
347 		return (error);
348 	}
349 
350 	rpipe->pipe_state |= PIPE_DIRECTOK;
351 	wpipe->pipe_state |= PIPE_DIRECTOK;
352 
353 	error = falloc(td, &rf, &fd);
354 	if (error) {
355 		pipeclose(rpipe);
356 		pipeclose(wpipe);
357 		return (error);
358 	}
359 	/* An extra reference on `rf' has been held for us by falloc(). */
360 	td->td_retval[0] = fd;
361 
362 	/*
363 	 * Warning: once we've gotten past allocation of the fd for the
364 	 * read-side, we can only drop the read side via fdrop() in order
365 	 * to avoid races against processes which manage to dup() the read
366 	 * side while we are blocked trying to allocate the write side.
367 	 */
368 	finit(rf, FREAD | FWRITE, DTYPE_PIPE, rpipe, &pipeops);
369 	error = falloc(td, &wf, &fd);
370 	if (error) {
371 		fdclose(fdp, rf, td->td_retval[0], td);
372 		fdrop(rf, td);
373 		/* rpipe has been closed by fdrop(). */
374 		pipeclose(wpipe);
375 		return (error);
376 	}
377 	/* An extra reference on `wf' has been held for us by falloc(). */
378 	finit(wf, FREAD | FWRITE, DTYPE_PIPE, wpipe, &pipeops);
379 	fdrop(wf, td);
380 	td->td_retval[1] = fd;
381 	fdrop(rf, td);
382 
383 	return (0);
384 }
385 
386 /*
387  * Allocate kva for pipe circular buffer, the space is pageable
388  * This routine will 'realloc' the size of a pipe safely, if it fails
389  * it will retain the old buffer.
390  * If it fails it will return ENOMEM.
391  */
392 static int
393 pipespace_new(cpipe, size)
394 	struct pipe *cpipe;
395 	int size;
396 {
397 	caddr_t buffer;
398 	int error, cnt, firstseg;
399 	static int curfail = 0;
400 	static struct timeval lastfail;
401 
402 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
403 	KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
404 		("pipespace: resize of direct writes not allowed"));
405 retry:
406 	cnt = cpipe->pipe_buffer.cnt;
407 	if (cnt > size)
408 		size = cnt;
409 
410 	size = round_page(size);
411 	buffer = (caddr_t) vm_map_min(pipe_map);
412 
413 	error = vm_map_find(pipe_map, NULL, 0,
414 		(vm_offset_t *) &buffer, size, 1,
415 		VM_PROT_ALL, VM_PROT_ALL, 0);
416 	if (error != KERN_SUCCESS) {
417 		if ((cpipe->pipe_buffer.buffer == NULL) &&
418 			(size > SMALL_PIPE_SIZE)) {
419 			size = SMALL_PIPE_SIZE;
420 			pipefragretry++;
421 			goto retry;
422 		}
423 		if (cpipe->pipe_buffer.buffer == NULL) {
424 			pipeallocfail++;
425 			if (ppsratecheck(&lastfail, &curfail, 1))
426 				printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
427 		} else {
428 			piperesizefail++;
429 		}
430 		return (ENOMEM);
431 	}
432 
433 	/* copy data, then free old resources if we're resizing */
434 	if (cnt > 0) {
435 		if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
436 			firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
437 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
438 				buffer, firstseg);
439 			if ((cnt - firstseg) > 0)
440 				bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
441 					cpipe->pipe_buffer.in);
442 		} else {
443 			bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
444 				buffer, cnt);
445 		}
446 	}
447 	pipe_free_kmem(cpipe);
448 	cpipe->pipe_buffer.buffer = buffer;
449 	cpipe->pipe_buffer.size = size;
450 	cpipe->pipe_buffer.in = cnt;
451 	cpipe->pipe_buffer.out = 0;
452 	cpipe->pipe_buffer.cnt = cnt;
453 	atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size);
454 	return (0);
455 }
456 
457 /*
458  * Wrapper for pipespace_new() that performs locking assertions.
459  */
460 static int
461 pipespace(cpipe, size)
462 	struct pipe *cpipe;
463 	int size;
464 {
465 
466 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
467 		("Unlocked pipe passed to pipespace"));
468 	return (pipespace_new(cpipe, size));
469 }
470 
471 /*
472  * lock a pipe for I/O, blocking other access
473  */
474 static __inline int
475 pipelock(cpipe, catch)
476 	struct pipe *cpipe;
477 	int catch;
478 {
479 	int error;
480 
481 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
482 	while (cpipe->pipe_state & PIPE_LOCKFL) {
483 		cpipe->pipe_state |= PIPE_LWANT;
484 		error = msleep(cpipe, PIPE_MTX(cpipe),
485 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
486 		    "pipelk", 0);
487 		if (error != 0)
488 			return (error);
489 	}
490 	cpipe->pipe_state |= PIPE_LOCKFL;
491 	return (0);
492 }
493 
494 /*
495  * unlock a pipe I/O lock
496  */
497 static __inline void
498 pipeunlock(cpipe)
499 	struct pipe *cpipe;
500 {
501 
502 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
503 	KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
504 		("Unlocked pipe passed to pipeunlock"));
505 	cpipe->pipe_state &= ~PIPE_LOCKFL;
506 	if (cpipe->pipe_state & PIPE_LWANT) {
507 		cpipe->pipe_state &= ~PIPE_LWANT;
508 		wakeup(cpipe);
509 	}
510 }
511 
512 static __inline void
513 pipeselwakeup(cpipe)
514 	struct pipe *cpipe;
515 {
516 
517 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
518 	if (cpipe->pipe_state & PIPE_SEL) {
519 		selwakeuppri(&cpipe->pipe_sel, PSOCK);
520 		if (!SEL_WAITING(&cpipe->pipe_sel))
521 			cpipe->pipe_state &= ~PIPE_SEL;
522 	}
523 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
524 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
525 	KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
526 }
527 
528 /*
529  * Initialize and allocate VM and memory for pipe.  The structure
530  * will start out zero'd from the ctor, so we just manage the kmem.
531  */
532 static int
533 pipe_create(pipe, backing)
534 	struct pipe *pipe;
535 	int backing;
536 {
537 	int error;
538 
539 	if (backing) {
540 		if (amountpipekva > maxpipekva / 2)
541 			error = pipespace_new(pipe, SMALL_PIPE_SIZE);
542 		else
543 			error = pipespace_new(pipe, PIPE_SIZE);
544 	} else {
545 		/* If we're not backing this pipe, no need to do anything. */
546 		error = 0;
547 	}
548 	return (error);
549 }
550 
551 /* ARGSUSED */
552 static int
553 pipe_read(fp, uio, active_cred, flags, td)
554 	struct file *fp;
555 	struct uio *uio;
556 	struct ucred *active_cred;
557 	struct thread *td;
558 	int flags;
559 {
560 	struct pipe *rpipe = fp->f_data;
561 	int error;
562 	int nread = 0;
563 	u_int size;
564 
565 	PIPE_LOCK(rpipe);
566 	++rpipe->pipe_busy;
567 	error = pipelock(rpipe, 1);
568 	if (error)
569 		goto unlocked_error;
570 
571 #ifdef MAC
572 	error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
573 	if (error)
574 		goto locked_error;
575 #endif
576 	if (amountpipekva > (3 * maxpipekva) / 4) {
577 		if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
578 			(rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
579 			(rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
580 			(piperesizeallowed == 1)) {
581 			PIPE_UNLOCK(rpipe);
582 			pipespace(rpipe, SMALL_PIPE_SIZE);
583 			PIPE_LOCK(rpipe);
584 		}
585 	}
586 
587 	while (uio->uio_resid) {
588 		/*
589 		 * normal pipe buffer receive
590 		 */
591 		if (rpipe->pipe_buffer.cnt > 0) {
592 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
593 			if (size > rpipe->pipe_buffer.cnt)
594 				size = rpipe->pipe_buffer.cnt;
595 			if (size > (u_int) uio->uio_resid)
596 				size = (u_int) uio->uio_resid;
597 
598 			PIPE_UNLOCK(rpipe);
599 			error = uiomove(
600 			    &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
601 			    size, uio);
602 			PIPE_LOCK(rpipe);
603 			if (error)
604 				break;
605 
606 			rpipe->pipe_buffer.out += size;
607 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
608 				rpipe->pipe_buffer.out = 0;
609 
610 			rpipe->pipe_buffer.cnt -= size;
611 
612 			/*
613 			 * If there is no more to read in the pipe, reset
614 			 * its pointers to the beginning.  This improves
615 			 * cache hit stats.
616 			 */
617 			if (rpipe->pipe_buffer.cnt == 0) {
618 				rpipe->pipe_buffer.in = 0;
619 				rpipe->pipe_buffer.out = 0;
620 			}
621 			nread += size;
622 #ifndef PIPE_NODIRECT
623 		/*
624 		 * Direct copy, bypassing a kernel buffer.
625 		 */
626 		} else if ((size = rpipe->pipe_map.cnt) &&
627 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
628 			if (size > (u_int) uio->uio_resid)
629 				size = (u_int) uio->uio_resid;
630 
631 			PIPE_UNLOCK(rpipe);
632 			error = uiomove_fromphys(rpipe->pipe_map.ms,
633 			    rpipe->pipe_map.pos, size, uio);
634 			PIPE_LOCK(rpipe);
635 			if (error)
636 				break;
637 			nread += size;
638 			rpipe->pipe_map.pos += size;
639 			rpipe->pipe_map.cnt -= size;
640 			if (rpipe->pipe_map.cnt == 0) {
641 				rpipe->pipe_state &= ~PIPE_DIRECTW;
642 				wakeup(rpipe);
643 			}
644 #endif
645 		} else {
646 			/*
647 			 * detect EOF condition
648 			 * read returns 0 on EOF, no need to set error
649 			 */
650 			if (rpipe->pipe_state & PIPE_EOF)
651 				break;
652 
653 			/*
654 			 * If the "write-side" has been blocked, wake it up now.
655 			 */
656 			if (rpipe->pipe_state & PIPE_WANTW) {
657 				rpipe->pipe_state &= ~PIPE_WANTW;
658 				wakeup(rpipe);
659 			}
660 
661 			/*
662 			 * Break if some data was read.
663 			 */
664 			if (nread > 0)
665 				break;
666 
667 			/*
668 			 * Unlock the pipe buffer for our remaining processing.
669 			 * We will either break out with an error or we will
670 			 * sleep and relock to loop.
671 			 */
672 			pipeunlock(rpipe);
673 
674 			/*
675 			 * Handle non-blocking mode operation or
676 			 * wait for more data.
677 			 */
678 			if (fp->f_flag & FNONBLOCK) {
679 				error = EAGAIN;
680 			} else {
681 				rpipe->pipe_state |= PIPE_WANTR;
682 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
683 				    PRIBIO | PCATCH,
684 				    "piperd", 0)) == 0)
685 					error = pipelock(rpipe, 1);
686 			}
687 			if (error)
688 				goto unlocked_error;
689 		}
690 	}
691 #ifdef MAC
692 locked_error:
693 #endif
694 	pipeunlock(rpipe);
695 
696 	/* XXX: should probably do this before getting any locks. */
697 	if (error == 0)
698 		vfs_timestamp(&rpipe->pipe_atime);
699 unlocked_error:
700 	--rpipe->pipe_busy;
701 
702 	/*
703 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
704 	 */
705 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
706 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
707 		wakeup(rpipe);
708 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
709 		/*
710 		 * Handle write blocking hysteresis.
711 		 */
712 		if (rpipe->pipe_state & PIPE_WANTW) {
713 			rpipe->pipe_state &= ~PIPE_WANTW;
714 			wakeup(rpipe);
715 		}
716 	}
717 
718 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
719 		pipeselwakeup(rpipe);
720 
721 	PIPE_UNLOCK(rpipe);
722 	return (error);
723 }
724 
725 #ifndef PIPE_NODIRECT
726 /*
727  * Map the sending processes' buffer into kernel space and wire it.
728  * This is similar to a physical write operation.
729  */
730 static int
731 pipe_build_write_buffer(wpipe, uio)
732 	struct pipe *wpipe;
733 	struct uio *uio;
734 {
735 	pmap_t pmap;
736 	u_int size;
737 	int i, j;
738 	vm_offset_t addr, endaddr;
739 
740 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
741 	KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
742 		("Clone attempt on non-direct write pipe!"));
743 
744 	size = (u_int) uio->uio_iov->iov_len;
745 	if (size > wpipe->pipe_buffer.size)
746 		size = wpipe->pipe_buffer.size;
747 
748 	pmap = vmspace_pmap(curproc->p_vmspace);
749 	endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
750 	addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
751 	for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
752 		/*
753 		 * vm_fault_quick() can sleep.  Consequently,
754 		 * vm_page_lock_queue() and vm_page_unlock_queue()
755 		 * should not be performed outside of this loop.
756 		 */
757 	race:
758 		if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) {
759 			vm_page_lock_queues();
760 			for (j = 0; j < i; j++)
761 				vm_page_unhold(wpipe->pipe_map.ms[j]);
762 			vm_page_unlock_queues();
763 			return (EFAULT);
764 		}
765 		wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr,
766 		    VM_PROT_READ);
767 		if (wpipe->pipe_map.ms[i] == NULL)
768 			goto race;
769 	}
770 
771 /*
772  * set up the control block
773  */
774 	wpipe->pipe_map.npages = i;
775 	wpipe->pipe_map.pos =
776 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
777 	wpipe->pipe_map.cnt = size;
778 
779 /*
780  * and update the uio data
781  */
782 
783 	uio->uio_iov->iov_len -= size;
784 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
785 	if (uio->uio_iov->iov_len == 0)
786 		uio->uio_iov++;
787 	uio->uio_resid -= size;
788 	uio->uio_offset += size;
789 	return (0);
790 }
791 
792 /*
793  * unmap and unwire the process buffer
794  */
795 static void
796 pipe_destroy_write_buffer(wpipe)
797 	struct pipe *wpipe;
798 {
799 	int i;
800 
801 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
802 	vm_page_lock_queues();
803 	for (i = 0; i < wpipe->pipe_map.npages; i++) {
804 		vm_page_unhold(wpipe->pipe_map.ms[i]);
805 	}
806 	vm_page_unlock_queues();
807 	wpipe->pipe_map.npages = 0;
808 }
809 
810 /*
811  * In the case of a signal, the writing process might go away.  This
812  * code copies the data into the circular buffer so that the source
813  * pages can be freed without loss of data.
814  */
815 static void
816 pipe_clone_write_buffer(wpipe)
817 	struct pipe *wpipe;
818 {
819 	struct uio uio;
820 	struct iovec iov;
821 	int size;
822 	int pos;
823 
824 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
825 	size = wpipe->pipe_map.cnt;
826 	pos = wpipe->pipe_map.pos;
827 
828 	wpipe->pipe_buffer.in = size;
829 	wpipe->pipe_buffer.out = 0;
830 	wpipe->pipe_buffer.cnt = size;
831 	wpipe->pipe_state &= ~PIPE_DIRECTW;
832 
833 	PIPE_UNLOCK(wpipe);
834 	iov.iov_base = wpipe->pipe_buffer.buffer;
835 	iov.iov_len = size;
836 	uio.uio_iov = &iov;
837 	uio.uio_iovcnt = 1;
838 	uio.uio_offset = 0;
839 	uio.uio_resid = size;
840 	uio.uio_segflg = UIO_SYSSPACE;
841 	uio.uio_rw = UIO_READ;
842 	uio.uio_td = curthread;
843 	uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
844 	PIPE_LOCK(wpipe);
845 	pipe_destroy_write_buffer(wpipe);
846 }
847 
848 /*
849  * This implements the pipe buffer write mechanism.  Note that only
850  * a direct write OR a normal pipe write can be pending at any given time.
851  * If there are any characters in the pipe buffer, the direct write will
852  * be deferred until the receiving process grabs all of the bytes from
853  * the pipe buffer.  Then the direct mapping write is set-up.
854  */
855 static int
856 pipe_direct_write(wpipe, uio)
857 	struct pipe *wpipe;
858 	struct uio *uio;
859 {
860 	int error;
861 
862 retry:
863 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
864 	error = pipelock(wpipe, 1);
865 	if (wpipe->pipe_state & PIPE_EOF)
866 		error = EPIPE;
867 	if (error) {
868 		pipeunlock(wpipe);
869 		goto error1;
870 	}
871 	while (wpipe->pipe_state & PIPE_DIRECTW) {
872 		if (wpipe->pipe_state & PIPE_WANTR) {
873 			wpipe->pipe_state &= ~PIPE_WANTR;
874 			wakeup(wpipe);
875 		}
876 		pipeselwakeup(wpipe);
877 		wpipe->pipe_state |= PIPE_WANTW;
878 		pipeunlock(wpipe);
879 		error = msleep(wpipe, PIPE_MTX(wpipe),
880 		    PRIBIO | PCATCH, "pipdww", 0);
881 		if (error)
882 			goto error1;
883 		else
884 			goto retry;
885 	}
886 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
887 	if (wpipe->pipe_buffer.cnt > 0) {
888 		if (wpipe->pipe_state & PIPE_WANTR) {
889 			wpipe->pipe_state &= ~PIPE_WANTR;
890 			wakeup(wpipe);
891 		}
892 		pipeselwakeup(wpipe);
893 		wpipe->pipe_state |= PIPE_WANTW;
894 		pipeunlock(wpipe);
895 		error = msleep(wpipe, PIPE_MTX(wpipe),
896 		    PRIBIO | PCATCH, "pipdwc", 0);
897 		if (error)
898 			goto error1;
899 		else
900 			goto retry;
901 	}
902 
903 	wpipe->pipe_state |= PIPE_DIRECTW;
904 
905 	PIPE_UNLOCK(wpipe);
906 	error = pipe_build_write_buffer(wpipe, uio);
907 	PIPE_LOCK(wpipe);
908 	if (error) {
909 		wpipe->pipe_state &= ~PIPE_DIRECTW;
910 		pipeunlock(wpipe);
911 		goto error1;
912 	}
913 
914 	error = 0;
915 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
916 		if (wpipe->pipe_state & PIPE_EOF) {
917 			pipe_destroy_write_buffer(wpipe);
918 			pipeselwakeup(wpipe);
919 			pipeunlock(wpipe);
920 			error = EPIPE;
921 			goto error1;
922 		}
923 		if (wpipe->pipe_state & PIPE_WANTR) {
924 			wpipe->pipe_state &= ~PIPE_WANTR;
925 			wakeup(wpipe);
926 		}
927 		pipeselwakeup(wpipe);
928 		pipeunlock(wpipe);
929 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
930 		    "pipdwt", 0);
931 		pipelock(wpipe, 0);
932 	}
933 
934 	if (wpipe->pipe_state & PIPE_EOF)
935 		error = EPIPE;
936 	if (wpipe->pipe_state & PIPE_DIRECTW) {
937 		/*
938 		 * this bit of trickery substitutes a kernel buffer for
939 		 * the process that might be going away.
940 		 */
941 		pipe_clone_write_buffer(wpipe);
942 	} else {
943 		pipe_destroy_write_buffer(wpipe);
944 	}
945 	pipeunlock(wpipe);
946 	return (error);
947 
948 error1:
949 	wakeup(wpipe);
950 	return (error);
951 }
952 #endif
953 
954 static int
955 pipe_write(fp, uio, active_cred, flags, td)
956 	struct file *fp;
957 	struct uio *uio;
958 	struct ucred *active_cred;
959 	struct thread *td;
960 	int flags;
961 {
962 	int error = 0;
963 	int desiredsize, orig_resid;
964 	struct pipe *wpipe, *rpipe;
965 
966 	rpipe = fp->f_data;
967 	wpipe = rpipe->pipe_peer;
968 
969 	PIPE_LOCK(rpipe);
970 	error = pipelock(wpipe, 1);
971 	if (error) {
972 		PIPE_UNLOCK(rpipe);
973 		return (error);
974 	}
975 	/*
976 	 * detect loss of pipe read side, issue SIGPIPE if lost.
977 	 */
978 	if ((!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
979 		pipeunlock(wpipe);
980 		PIPE_UNLOCK(rpipe);
981 		return (EPIPE);
982 	}
983 #ifdef MAC
984 	error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
985 	if (error) {
986 		pipeunlock(wpipe);
987 		PIPE_UNLOCK(rpipe);
988 		return (error);
989 	}
990 #endif
991 	++wpipe->pipe_busy;
992 
993 	/* Choose a larger size if it's advantageous */
994 	desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
995 	while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
996 		if (piperesizeallowed != 1)
997 			break;
998 		if (amountpipekva > maxpipekva / 2)
999 			break;
1000 		if (desiredsize == BIG_PIPE_SIZE)
1001 			break;
1002 		desiredsize = desiredsize * 2;
1003 	}
1004 
1005 	/* Choose a smaller size if we're in a OOM situation */
1006 	if ((amountpipekva > (3 * maxpipekva) / 4) &&
1007 		(wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1008 		(wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1009 		(piperesizeallowed == 1))
1010 		desiredsize = SMALL_PIPE_SIZE;
1011 
1012 	/* Resize if the above determined that a new size was necessary */
1013 	if ((desiredsize != wpipe->pipe_buffer.size) &&
1014 		((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1015 		PIPE_UNLOCK(wpipe);
1016 		pipespace(wpipe, desiredsize);
1017 		PIPE_LOCK(wpipe);
1018 	}
1019 	if (wpipe->pipe_buffer.size == 0) {
1020 		/*
1021 		 * This can only happen for reverse direction use of pipes
1022 		 * in a complete OOM situation.
1023 		 */
1024 		error = ENOMEM;
1025 		--wpipe->pipe_busy;
1026 		pipeunlock(wpipe);
1027 		PIPE_UNLOCK(wpipe);
1028 		return (error);
1029 	}
1030 
1031 	pipeunlock(wpipe);
1032 
1033 	orig_resid = uio->uio_resid;
1034 
1035 	while (uio->uio_resid) {
1036 		int space;
1037 
1038 		pipelock(wpipe, 0);
1039 		if (wpipe->pipe_state & PIPE_EOF) {
1040 			pipeunlock(wpipe);
1041 			error = EPIPE;
1042 			break;
1043 		}
1044 #ifndef PIPE_NODIRECT
1045 		/*
1046 		 * If the transfer is large, we can gain performance if
1047 		 * we do process-to-process copies directly.
1048 		 * If the write is non-blocking, we don't use the
1049 		 * direct write mechanism.
1050 		 *
1051 		 * The direct write mechanism will detect the reader going
1052 		 * away on us.
1053 		 */
1054 		if (uio->uio_segflg == UIO_USERSPACE &&
1055 		    uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1056 		    wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1057 		    (fp->f_flag & FNONBLOCK) == 0) {
1058 			pipeunlock(wpipe);
1059 			error = pipe_direct_write(wpipe, uio);
1060 			if (error)
1061 				break;
1062 			continue;
1063 		}
1064 #endif
1065 
1066 		/*
1067 		 * Pipe buffered writes cannot be coincidental with
1068 		 * direct writes.  We wait until the currently executing
1069 		 * direct write is completed before we start filling the
1070 		 * pipe buffer.  We break out if a signal occurs or the
1071 		 * reader goes away.
1072 		 */
1073 		if (wpipe->pipe_state & PIPE_DIRECTW) {
1074 			if (wpipe->pipe_state & PIPE_WANTR) {
1075 				wpipe->pipe_state &= ~PIPE_WANTR;
1076 				wakeup(wpipe);
1077 			}
1078 			pipeselwakeup(wpipe);
1079 			wpipe->pipe_state |= PIPE_WANTW;
1080 			pipeunlock(wpipe);
1081 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1082 			    "pipbww", 0);
1083 			if (error)
1084 				break;
1085 			else
1086 				continue;
1087 		}
1088 
1089 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1090 
1091 		/* Writes of size <= PIPE_BUF must be atomic. */
1092 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1093 			space = 0;
1094 
1095 		if (space > 0) {
1096 			int size;	/* Transfer size */
1097 			int segsize;	/* first segment to transfer */
1098 
1099 			/*
1100 			 * Transfer size is minimum of uio transfer
1101 			 * and free space in pipe buffer.
1102 			 */
1103 			if (space > uio->uio_resid)
1104 				size = uio->uio_resid;
1105 			else
1106 				size = space;
1107 			/*
1108 			 * First segment to transfer is minimum of
1109 			 * transfer size and contiguous space in
1110 			 * pipe buffer.  If first segment to transfer
1111 			 * is less than the transfer size, we've got
1112 			 * a wraparound in the buffer.
1113 			 */
1114 			segsize = wpipe->pipe_buffer.size -
1115 				wpipe->pipe_buffer.in;
1116 			if (segsize > size)
1117 				segsize = size;
1118 
1119 			/* Transfer first segment */
1120 
1121 			PIPE_UNLOCK(rpipe);
1122 			error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1123 					segsize, uio);
1124 			PIPE_LOCK(rpipe);
1125 
1126 			if (error == 0 && segsize < size) {
1127 				KASSERT(wpipe->pipe_buffer.in + segsize ==
1128 					wpipe->pipe_buffer.size,
1129 					("Pipe buffer wraparound disappeared"));
1130 				/*
1131 				 * Transfer remaining part now, to
1132 				 * support atomic writes.  Wraparound
1133 				 * happened.
1134 				 */
1135 
1136 				PIPE_UNLOCK(rpipe);
1137 				error = uiomove(
1138 				    &wpipe->pipe_buffer.buffer[0],
1139 				    size - segsize, uio);
1140 				PIPE_LOCK(rpipe);
1141 			}
1142 			if (error == 0) {
1143 				wpipe->pipe_buffer.in += size;
1144 				if (wpipe->pipe_buffer.in >=
1145 				    wpipe->pipe_buffer.size) {
1146 					KASSERT(wpipe->pipe_buffer.in ==
1147 						size - segsize +
1148 						wpipe->pipe_buffer.size,
1149 						("Expected wraparound bad"));
1150 					wpipe->pipe_buffer.in = size - segsize;
1151 				}
1152 
1153 				wpipe->pipe_buffer.cnt += size;
1154 				KASSERT(wpipe->pipe_buffer.cnt <=
1155 					wpipe->pipe_buffer.size,
1156 					("Pipe buffer overflow"));
1157 			}
1158 			pipeunlock(wpipe);
1159 			if (error != 0)
1160 				break;
1161 		} else {
1162 			/*
1163 			 * If the "read-side" has been blocked, wake it up now.
1164 			 */
1165 			if (wpipe->pipe_state & PIPE_WANTR) {
1166 				wpipe->pipe_state &= ~PIPE_WANTR;
1167 				wakeup(wpipe);
1168 			}
1169 
1170 			/*
1171 			 * don't block on non-blocking I/O
1172 			 */
1173 			if (fp->f_flag & FNONBLOCK) {
1174 				error = EAGAIN;
1175 				pipeunlock(wpipe);
1176 				break;
1177 			}
1178 
1179 			/*
1180 			 * We have no more space and have something to offer,
1181 			 * wake up select/poll.
1182 			 */
1183 			pipeselwakeup(wpipe);
1184 
1185 			wpipe->pipe_state |= PIPE_WANTW;
1186 			pipeunlock(wpipe);
1187 			error = msleep(wpipe, PIPE_MTX(rpipe),
1188 			    PRIBIO | PCATCH, "pipewr", 0);
1189 			if (error != 0)
1190 				break;
1191 		}
1192 	}
1193 
1194 	pipelock(wpipe, 0);
1195 	--wpipe->pipe_busy;
1196 
1197 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1198 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1199 		wakeup(wpipe);
1200 	} else if (wpipe->pipe_buffer.cnt > 0) {
1201 		/*
1202 		 * If we have put any characters in the buffer, we wake up
1203 		 * the reader.
1204 		 */
1205 		if (wpipe->pipe_state & PIPE_WANTR) {
1206 			wpipe->pipe_state &= ~PIPE_WANTR;
1207 			wakeup(wpipe);
1208 		}
1209 	}
1210 
1211 	/*
1212 	 * Don't return EPIPE if I/O was successful
1213 	 */
1214 	if ((wpipe->pipe_buffer.cnt == 0) &&
1215 	    (uio->uio_resid == 0) &&
1216 	    (error == EPIPE)) {
1217 		error = 0;
1218 	}
1219 
1220 	if (error == 0)
1221 		vfs_timestamp(&wpipe->pipe_mtime);
1222 
1223 	/*
1224 	 * We have something to offer,
1225 	 * wake up select/poll.
1226 	 */
1227 	if (wpipe->pipe_buffer.cnt)
1228 		pipeselwakeup(wpipe);
1229 
1230 	pipeunlock(wpipe);
1231 	PIPE_UNLOCK(rpipe);
1232 	return (error);
1233 }
1234 
1235 /* ARGSUSED */
1236 static int
1237 pipe_truncate(fp, length, active_cred, td)
1238 	struct file *fp;
1239 	off_t length;
1240 	struct ucred *active_cred;
1241 	struct thread *td;
1242 {
1243 
1244 	return (EINVAL);
1245 }
1246 
1247 /*
1248  * we implement a very minimal set of ioctls for compatibility with sockets.
1249  */
1250 static int
1251 pipe_ioctl(fp, cmd, data, active_cred, td)
1252 	struct file *fp;
1253 	u_long cmd;
1254 	void *data;
1255 	struct ucred *active_cred;
1256 	struct thread *td;
1257 {
1258 	struct pipe *mpipe = fp->f_data;
1259 	int error;
1260 
1261 	PIPE_LOCK(mpipe);
1262 
1263 #ifdef MAC
1264 	error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1265 	if (error) {
1266 		PIPE_UNLOCK(mpipe);
1267 		return (error);
1268 	}
1269 #endif
1270 
1271 	error = 0;
1272 	switch (cmd) {
1273 
1274 	case FIONBIO:
1275 		break;
1276 
1277 	case FIOASYNC:
1278 		if (*(int *)data) {
1279 			mpipe->pipe_state |= PIPE_ASYNC;
1280 		} else {
1281 			mpipe->pipe_state &= ~PIPE_ASYNC;
1282 		}
1283 		break;
1284 
1285 	case FIONREAD:
1286 		if (mpipe->pipe_state & PIPE_DIRECTW)
1287 			*(int *)data = mpipe->pipe_map.cnt;
1288 		else
1289 			*(int *)data = mpipe->pipe_buffer.cnt;
1290 		break;
1291 
1292 	case FIOSETOWN:
1293 		PIPE_UNLOCK(mpipe);
1294 		error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1295 		goto out_unlocked;
1296 
1297 	case FIOGETOWN:
1298 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1299 		break;
1300 
1301 	/* This is deprecated, FIOSETOWN should be used instead. */
1302 	case TIOCSPGRP:
1303 		PIPE_UNLOCK(mpipe);
1304 		error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1305 		goto out_unlocked;
1306 
1307 	/* This is deprecated, FIOGETOWN should be used instead. */
1308 	case TIOCGPGRP:
1309 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1310 		break;
1311 
1312 	default:
1313 		error = ENOTTY;
1314 		break;
1315 	}
1316 	PIPE_UNLOCK(mpipe);
1317 out_unlocked:
1318 	return (error);
1319 }
1320 
1321 static int
1322 pipe_poll(fp, events, active_cred, td)
1323 	struct file *fp;
1324 	int events;
1325 	struct ucred *active_cred;
1326 	struct thread *td;
1327 {
1328 	struct pipe *rpipe = fp->f_data;
1329 	struct pipe *wpipe;
1330 	int revents = 0;
1331 #ifdef MAC
1332 	int error;
1333 #endif
1334 
1335 	wpipe = rpipe->pipe_peer;
1336 	PIPE_LOCK(rpipe);
1337 #ifdef MAC
1338 	error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1339 	if (error)
1340 		goto locked_error;
1341 #endif
1342 	if (events & (POLLIN | POLLRDNORM))
1343 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1344 		    (rpipe->pipe_buffer.cnt > 0) ||
1345 		    (rpipe->pipe_state & PIPE_EOF))
1346 			revents |= events & (POLLIN | POLLRDNORM);
1347 
1348 	if (events & (POLLOUT | POLLWRNORM))
1349 		if (!wpipe->pipe_present || (wpipe->pipe_state & PIPE_EOF) ||
1350 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1351 		     (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1352 			revents |= events & (POLLOUT | POLLWRNORM);
1353 
1354 	if ((rpipe->pipe_state & PIPE_EOF) ||
1355 	    (!wpipe->pipe_present) ||
1356 	    (wpipe->pipe_state & PIPE_EOF))
1357 		revents |= POLLHUP;
1358 
1359 	if (revents == 0) {
1360 		if (events & (POLLIN | POLLRDNORM)) {
1361 			selrecord(td, &rpipe->pipe_sel);
1362 			if (SEL_WAITING(&rpipe->pipe_sel))
1363 				rpipe->pipe_state |= PIPE_SEL;
1364 		}
1365 
1366 		if (events & (POLLOUT | POLLWRNORM)) {
1367 			selrecord(td, &wpipe->pipe_sel);
1368 			if (SEL_WAITING(&wpipe->pipe_sel))
1369 				wpipe->pipe_state |= PIPE_SEL;
1370 		}
1371 	}
1372 #ifdef MAC
1373 locked_error:
1374 #endif
1375 	PIPE_UNLOCK(rpipe);
1376 
1377 	return (revents);
1378 }
1379 
1380 /*
1381  * We shouldn't need locks here as we're doing a read and this should
1382  * be a natural race.
1383  */
1384 static int
1385 pipe_stat(fp, ub, active_cred, td)
1386 	struct file *fp;
1387 	struct stat *ub;
1388 	struct ucred *active_cred;
1389 	struct thread *td;
1390 {
1391 	struct pipe *pipe = fp->f_data;
1392 #ifdef MAC
1393 	int error;
1394 
1395 	PIPE_LOCK(pipe);
1396 	error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1397 	PIPE_UNLOCK(pipe);
1398 	if (error)
1399 		return (error);
1400 #endif
1401 	bzero(ub, sizeof(*ub));
1402 	ub->st_mode = S_IFIFO;
1403 	ub->st_blksize = PAGE_SIZE;
1404 	if (pipe->pipe_state & PIPE_DIRECTW)
1405 		ub->st_size = pipe->pipe_map.cnt;
1406 	else
1407 		ub->st_size = pipe->pipe_buffer.cnt;
1408 	ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1409 	ub->st_atimespec = pipe->pipe_atime;
1410 	ub->st_mtimespec = pipe->pipe_mtime;
1411 	ub->st_ctimespec = pipe->pipe_ctime;
1412 	ub->st_uid = fp->f_cred->cr_uid;
1413 	ub->st_gid = fp->f_cred->cr_gid;
1414 	/*
1415 	 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1416 	 * XXX (st_dev, st_ino) should be unique.
1417 	 */
1418 	return (0);
1419 }
1420 
1421 /* ARGSUSED */
1422 static int
1423 pipe_close(fp, td)
1424 	struct file *fp;
1425 	struct thread *td;
1426 {
1427 	struct pipe *cpipe = fp->f_data;
1428 
1429 	fp->f_ops = &badfileops;
1430 	fp->f_data = NULL;
1431 	funsetown(&cpipe->pipe_sigio);
1432 	pipeclose(cpipe);
1433 	return (0);
1434 }
1435 
1436 static void
1437 pipe_free_kmem(cpipe)
1438 	struct pipe *cpipe;
1439 {
1440 
1441 	KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1442 	    ("pipe_free_kmem: pipe mutex locked"));
1443 
1444 	if (cpipe->pipe_buffer.buffer != NULL) {
1445 		atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size);
1446 		vm_map_remove(pipe_map,
1447 		    (vm_offset_t)cpipe->pipe_buffer.buffer,
1448 		    (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1449 		cpipe->pipe_buffer.buffer = NULL;
1450 	}
1451 #ifndef PIPE_NODIRECT
1452 	{
1453 		cpipe->pipe_map.cnt = 0;
1454 		cpipe->pipe_map.pos = 0;
1455 		cpipe->pipe_map.npages = 0;
1456 	}
1457 #endif
1458 }
1459 
1460 /*
1461  * shutdown the pipe
1462  */
1463 static void
1464 pipeclose(cpipe)
1465 	struct pipe *cpipe;
1466 {
1467 	struct pipepair *pp;
1468 	struct pipe *ppipe;
1469 
1470 	KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1471 
1472 	PIPE_LOCK(cpipe);
1473 	pipelock(cpipe, 0);
1474 	pp = cpipe->pipe_pair;
1475 
1476 	pipeselwakeup(cpipe);
1477 
1478 	/*
1479 	 * If the other side is blocked, wake it up saying that
1480 	 * we want to close it down.
1481 	 */
1482 	cpipe->pipe_state |= PIPE_EOF;
1483 	while (cpipe->pipe_busy) {
1484 		wakeup(cpipe);
1485 		cpipe->pipe_state |= PIPE_WANT;
1486 		pipeunlock(cpipe);
1487 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1488 		pipelock(cpipe, 0);
1489 	}
1490 
1491 
1492 	/*
1493 	 * Disconnect from peer, if any.
1494 	 */
1495 	ppipe = cpipe->pipe_peer;
1496 	if (ppipe->pipe_present != 0) {
1497 		pipeselwakeup(ppipe);
1498 
1499 		ppipe->pipe_state |= PIPE_EOF;
1500 		wakeup(ppipe);
1501 		KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1502 	}
1503 
1504 	/*
1505 	 * Mark this endpoint as free.  Release kmem resources.  We
1506 	 * don't mark this endpoint as unused until we've finished
1507 	 * doing that, or the pipe might disappear out from under
1508 	 * us.
1509 	 */
1510 	PIPE_UNLOCK(cpipe);
1511 	pipe_free_kmem(cpipe);
1512 	PIPE_LOCK(cpipe);
1513 	cpipe->pipe_present = 0;
1514 	pipeunlock(cpipe);
1515 	knlist_clear(&cpipe->pipe_sel.si_note, 1);
1516 	knlist_destroy(&cpipe->pipe_sel.si_note);
1517 
1518 	/*
1519 	 * If both endpoints are now closed, release the memory for the
1520 	 * pipe pair.  If not, unlock.
1521 	 */
1522 	if (ppipe->pipe_present == 0) {
1523 		PIPE_UNLOCK(cpipe);
1524 #ifdef MAC
1525 		mac_pipe_destroy(pp);
1526 #endif
1527 		uma_zfree(pipe_zone, cpipe->pipe_pair);
1528 	} else
1529 		PIPE_UNLOCK(cpipe);
1530 }
1531 
1532 /*ARGSUSED*/
1533 static int
1534 pipe_kqfilter(struct file *fp, struct knote *kn)
1535 {
1536 	struct pipe *cpipe;
1537 
1538 	cpipe = kn->kn_fp->f_data;
1539 	PIPE_LOCK(cpipe);
1540 	switch (kn->kn_filter) {
1541 	case EVFILT_READ:
1542 		kn->kn_fop = &pipe_rfiltops;
1543 		break;
1544 	case EVFILT_WRITE:
1545 		kn->kn_fop = &pipe_wfiltops;
1546 		if (!cpipe->pipe_peer->pipe_present) {
1547 			/* other end of pipe has been closed */
1548 			PIPE_UNLOCK(cpipe);
1549 			return (EPIPE);
1550 		}
1551 		cpipe = cpipe->pipe_peer;
1552 		break;
1553 	default:
1554 		PIPE_UNLOCK(cpipe);
1555 		return (EINVAL);
1556 	}
1557 
1558 	knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1559 	PIPE_UNLOCK(cpipe);
1560 	return (0);
1561 }
1562 
1563 static void
1564 filt_pipedetach(struct knote *kn)
1565 {
1566 	struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1567 
1568 	PIPE_LOCK(cpipe);
1569 	if (kn->kn_filter == EVFILT_WRITE) {
1570 		if (!cpipe->pipe_peer->pipe_present) {
1571 			PIPE_UNLOCK(cpipe);
1572 			return;
1573 		}
1574 		cpipe = cpipe->pipe_peer;
1575 	}
1576 	knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1577 	PIPE_UNLOCK(cpipe);
1578 }
1579 
1580 /*ARGSUSED*/
1581 static int
1582 filt_piperead(struct knote *kn, long hint)
1583 {
1584 	struct pipe *rpipe = kn->kn_fp->f_data;
1585 	struct pipe *wpipe = rpipe->pipe_peer;
1586 	int ret;
1587 
1588 	PIPE_LOCK(rpipe);
1589 	kn->kn_data = rpipe->pipe_buffer.cnt;
1590 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1591 		kn->kn_data = rpipe->pipe_map.cnt;
1592 
1593 	if ((rpipe->pipe_state & PIPE_EOF) ||
1594 	    (!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
1595 		kn->kn_flags |= EV_EOF;
1596 		PIPE_UNLOCK(rpipe);
1597 		return (1);
1598 	}
1599 	ret = kn->kn_data > 0;
1600 	PIPE_UNLOCK(rpipe);
1601 	return ret;
1602 }
1603 
1604 /*ARGSUSED*/
1605 static int
1606 filt_pipewrite(struct knote *kn, long hint)
1607 {
1608 	struct pipe *rpipe = kn->kn_fp->f_data;
1609 	struct pipe *wpipe = rpipe->pipe_peer;
1610 
1611 	PIPE_LOCK(rpipe);
1612 	if ((!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
1613 		kn->kn_data = 0;
1614 		kn->kn_flags |= EV_EOF;
1615 		PIPE_UNLOCK(rpipe);
1616 		return (1);
1617 	}
1618 	kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1619 	if (wpipe->pipe_state & PIPE_DIRECTW)
1620 		kn->kn_data = 0;
1621 
1622 	PIPE_UNLOCK(rpipe);
1623 	return (kn->kn_data >= PIPE_BUF);
1624 }
1625