xref: /freebsd/sys/kern/sys_pipe.c (revision 729362425c09cf6b362366aabc6fb547eee8035a)
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  * $FreeBSD$
20  */
21 
22 /*
23  * This file contains a high-performance replacement for the socket-based
24  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
25  * all features of sockets, but does do everything that pipes normally
26  * do.
27  */
28 
29 /*
30  * This code has two modes of operation, a small write mode and a large
31  * write mode.  The small write mode acts like conventional pipes with
32  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
33  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
34  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
35  * the receiving process can copy it directly from the pages in the sending
36  * process.
37  *
38  * If the sending process receives a signal, it is possible that it will
39  * go away, and certainly its address space can change, because control
40  * is returned back to the user-mode side.  In that case, the pipe code
41  * arranges to copy the buffer supplied by the user process, to a pageable
42  * kernel buffer, and the receiving process will grab the data from the
43  * pageable kernel buffer.  Since signals don't happen all that often,
44  * the copy operation is normally eliminated.
45  *
46  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
47  * happen for small transfers so that the system will not spend all of
48  * its time context switching.  PIPE_SIZE is constrained by the
49  * amount of kernel virtual memory.
50  */
51 
52 #include "opt_mac.h"
53 
54 #include <sys/param.h>
55 #include <sys/systm.h>
56 #include <sys/fcntl.h>
57 #include <sys/file.h>
58 #include <sys/filedesc.h>
59 #include <sys/filio.h>
60 #include <sys/kernel.h>
61 #include <sys/lock.h>
62 #include <sys/mac.h>
63 #include <sys/mutex.h>
64 #include <sys/ttycom.h>
65 #include <sys/stat.h>
66 #include <sys/malloc.h>
67 #include <sys/poll.h>
68 #include <sys/selinfo.h>
69 #include <sys/signalvar.h>
70 #include <sys/sysproto.h>
71 #include <sys/pipe.h>
72 #include <sys/proc.h>
73 #include <sys/vnode.h>
74 #include <sys/uio.h>
75 #include <sys/event.h>
76 
77 #include <vm/vm.h>
78 #include <vm/vm_param.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_kern.h>
81 #include <vm/vm_extern.h>
82 #include <vm/pmap.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_page.h>
85 #include <vm/uma.h>
86 
87 /*
88  * Use this define if you want to disable *fancy* VM things.  Expect an
89  * approx 30% decrease in transfer rate.  This could be useful for
90  * NetBSD or OpenBSD.
91  */
92 /* #define PIPE_NODIRECT */
93 
94 /*
95  * interfaces to the outside world
96  */
97 static fo_rdwr_t	pipe_read;
98 static fo_rdwr_t	pipe_write;
99 static fo_ioctl_t	pipe_ioctl;
100 static fo_poll_t	pipe_poll;
101 static fo_kqfilter_t	pipe_kqfilter;
102 static fo_stat_t	pipe_stat;
103 static fo_close_t	pipe_close;
104 
105 static struct fileops pipeops = {
106 	pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
107 	pipe_stat, pipe_close, DFLAG_PASSABLE
108 };
109 
110 static void	filt_pipedetach(struct knote *kn);
111 static int	filt_piperead(struct knote *kn, long hint);
112 static int	filt_pipewrite(struct knote *kn, long hint);
113 
114 static struct filterops pipe_rfiltops =
115 	{ 1, NULL, filt_pipedetach, filt_piperead };
116 static struct filterops pipe_wfiltops =
117 	{ 1, NULL, filt_pipedetach, filt_pipewrite };
118 
119 #define PIPE_GET_GIANT(pipe)						\
120 	do {								\
121 		KASSERT(((pipe)->pipe_state & PIPE_LOCKFL) != 0,	\
122 		    ("%s:%d PIPE_GET_GIANT: line pipe not locked",	\
123 		     __FILE__, __LINE__));				\
124 		PIPE_UNLOCK(pipe);					\
125 		mtx_lock(&Giant);					\
126 	} while (0)
127 
128 #define PIPE_DROP_GIANT(pipe)						\
129 	do {								\
130 		mtx_unlock(&Giant);					\
131 		PIPE_LOCK(pipe);					\
132 	} while (0)
133 
134 /*
135  * Default pipe buffer size(s), this can be kind-of large now because pipe
136  * space is pageable.  The pipe code will try to maintain locality of
137  * reference for performance reasons, so small amounts of outstanding I/O
138  * will not wipe the cache.
139  */
140 #define MINPIPESIZE (PIPE_SIZE/3)
141 #define MAXPIPESIZE (2*PIPE_SIZE/3)
142 
143 /*
144  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
145  * is there so that on large systems, we don't exhaust it.
146  */
147 #define MAXPIPEKVA (8*1024*1024)
148 
149 /*
150  * Limit for direct transfers, we cannot, of course limit
151  * the amount of kva for pipes in general though.
152  */
153 #define LIMITPIPEKVA (16*1024*1024)
154 
155 /*
156  * Limit the number of "big" pipes
157  */
158 #define LIMITBIGPIPES	32
159 static int nbigpipe;
160 
161 static int amountpipekva;
162 
163 static void pipeinit(void *dummy __unused);
164 static void pipeclose(struct pipe *cpipe);
165 static void pipe_free_kmem(struct pipe *cpipe);
166 static int pipe_create(struct pipe **cpipep);
167 static __inline int pipelock(struct pipe *cpipe, int catch);
168 static __inline void pipeunlock(struct pipe *cpipe);
169 static __inline void pipeselwakeup(struct pipe *cpipe);
170 #ifndef PIPE_NODIRECT
171 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
172 static void pipe_destroy_write_buffer(struct pipe *wpipe);
173 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
174 static void pipe_clone_write_buffer(struct pipe *wpipe);
175 #endif
176 static int pipespace(struct pipe *cpipe, int size);
177 
178 static uma_zone_t pipe_zone;
179 
180 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
181 
182 static void
183 pipeinit(void *dummy __unused)
184 {
185 	pipe_zone = uma_zcreate("PIPE", sizeof(struct pipe), NULL,
186 	    NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
187 }
188 
189 /*
190  * The pipe system call for the DTYPE_PIPE type of pipes
191  */
192 
193 /* ARGSUSED */
194 int
195 pipe(td, uap)
196 	struct thread *td;
197 	struct pipe_args /* {
198 		int	dummy;
199 	} */ *uap;
200 {
201 	struct filedesc *fdp = td->td_proc->p_fd;
202 	struct file *rf, *wf;
203 	struct pipe *rpipe, *wpipe;
204 	struct mtx *pmtx;
205 	int fd, error;
206 
207 	KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
208 
209 	pmtx = malloc(sizeof(*pmtx), M_TEMP, M_WAITOK | M_ZERO);
210 
211 	rpipe = wpipe = NULL;
212 	if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
213 		pipeclose(rpipe);
214 		pipeclose(wpipe);
215 		free(pmtx, M_TEMP);
216 		return (ENFILE);
217 	}
218 
219 	rpipe->pipe_state |= PIPE_DIRECTOK;
220 	wpipe->pipe_state |= PIPE_DIRECTOK;
221 
222 	error = falloc(td, &rf, &fd);
223 	if (error) {
224 		pipeclose(rpipe);
225 		pipeclose(wpipe);
226 		free(pmtx, M_TEMP);
227 		return (error);
228 	}
229 	fhold(rf);
230 	td->td_retval[0] = fd;
231 
232 	/*
233 	 * Warning: once we've gotten past allocation of the fd for the
234 	 * read-side, we can only drop the read side via fdrop() in order
235 	 * to avoid races against processes which manage to dup() the read
236 	 * side while we are blocked trying to allocate the write side.
237 	 */
238 	FILE_LOCK(rf);
239 	rf->f_flag = FREAD | FWRITE;
240 	rf->f_type = DTYPE_PIPE;
241 	rf->f_data = rpipe;
242 	rf->f_ops = &pipeops;
243 	FILE_UNLOCK(rf);
244 	error = falloc(td, &wf, &fd);
245 	if (error) {
246 		FILEDESC_LOCK(fdp);
247 		if (fdp->fd_ofiles[td->td_retval[0]] == rf) {
248 			fdp->fd_ofiles[td->td_retval[0]] = NULL;
249 			FILEDESC_UNLOCK(fdp);
250 			fdrop(rf, td);
251 		} else
252 			FILEDESC_UNLOCK(fdp);
253 		fdrop(rf, td);
254 		/* rpipe has been closed by fdrop(). */
255 		pipeclose(wpipe);
256 		free(pmtx, M_TEMP);
257 		return (error);
258 	}
259 	FILE_LOCK(wf);
260 	wf->f_flag = FREAD | FWRITE;
261 	wf->f_type = DTYPE_PIPE;
262 	wf->f_data = wpipe;
263 	wf->f_ops = &pipeops;
264 	FILE_UNLOCK(wf);
265 	td->td_retval[1] = fd;
266 	rpipe->pipe_peer = wpipe;
267 	wpipe->pipe_peer = rpipe;
268 #ifdef MAC
269 	/*
270 	 * struct pipe represents a pipe endpoint.  The MAC label is shared
271 	 * between the connected endpoints.  As a result mac_init_pipe() and
272 	 * mac_create_pipe() should only be called on one of the endpoints
273 	 * after they have been connected.
274 	 */
275 	mac_init_pipe(rpipe);
276 	mac_create_pipe(td->td_ucred, rpipe);
277 #endif
278 	mtx_init(pmtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
279 	rpipe->pipe_mtxp = wpipe->pipe_mtxp = pmtx;
280 	fdrop(rf, td);
281 
282 	return (0);
283 }
284 
285 /*
286  * Allocate kva for pipe circular buffer, the space is pageable
287  * This routine will 'realloc' the size of a pipe safely, if it fails
288  * it will retain the old buffer.
289  * If it fails it will return ENOMEM.
290  */
291 static int
292 pipespace(cpipe, size)
293 	struct pipe *cpipe;
294 	int size;
295 {
296 	struct vm_object *object;
297 	caddr_t buffer;
298 	int npages, error;
299 
300 	GIANT_REQUIRED;
301 	KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)),
302 	       ("pipespace: pipe mutex locked"));
303 
304 	npages = round_page(size)/PAGE_SIZE;
305 	/*
306 	 * Create an object, I don't like the idea of paging to/from
307 	 * kernel_object.
308 	 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
309 	 */
310 	object = vm_object_allocate(OBJT_DEFAULT, npages);
311 	buffer = (caddr_t) vm_map_min(kernel_map);
312 
313 	/*
314 	 * Insert the object into the kernel map, and allocate kva for it.
315 	 * The map entry is, by default, pageable.
316 	 * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
317 	 */
318 	error = vm_map_find(kernel_map, object, 0,
319 		(vm_offset_t *) &buffer, size, 1,
320 		VM_PROT_ALL, VM_PROT_ALL, 0);
321 
322 	if (error != KERN_SUCCESS) {
323 		vm_object_deallocate(object);
324 		return (ENOMEM);
325 	}
326 
327 	/* free old resources if we're resizing */
328 	pipe_free_kmem(cpipe);
329 	cpipe->pipe_buffer.object = object;
330 	cpipe->pipe_buffer.buffer = buffer;
331 	cpipe->pipe_buffer.size = size;
332 	cpipe->pipe_buffer.in = 0;
333 	cpipe->pipe_buffer.out = 0;
334 	cpipe->pipe_buffer.cnt = 0;
335 	atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size);
336 	return (0);
337 }
338 
339 /*
340  * initialize and allocate VM and memory for pipe
341  */
342 static int
343 pipe_create(cpipep)
344 	struct pipe **cpipep;
345 {
346 	struct pipe *cpipe;
347 	int error;
348 
349 	*cpipep = uma_zalloc(pipe_zone, M_WAITOK);
350 	if (*cpipep == NULL)
351 		return (ENOMEM);
352 
353 	cpipe = *cpipep;
354 
355 	/* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */
356 	cpipe->pipe_buffer.object = NULL;
357 #ifndef PIPE_NODIRECT
358 	cpipe->pipe_map.kva = 0;
359 #endif
360 	/*
361 	 * protect so pipeclose() doesn't follow a junk pointer
362 	 * if pipespace() fails.
363 	 */
364 	bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel));
365 	cpipe->pipe_state = 0;
366 	cpipe->pipe_peer = NULL;
367 	cpipe->pipe_busy = 0;
368 
369 #ifndef PIPE_NODIRECT
370 	/*
371 	 * pipe data structure initializations to support direct pipe I/O
372 	 */
373 	cpipe->pipe_map.cnt = 0;
374 	cpipe->pipe_map.kva = 0;
375 	cpipe->pipe_map.pos = 0;
376 	cpipe->pipe_map.npages = 0;
377 	/* cpipe->pipe_map.ms[] = invalid */
378 #endif
379 
380 	cpipe->pipe_mtxp = NULL;	/* avoid pipespace assertion */
381 	error = pipespace(cpipe, PIPE_SIZE);
382 	if (error)
383 		return (error);
384 
385 	vfs_timestamp(&cpipe->pipe_ctime);
386 	cpipe->pipe_atime = cpipe->pipe_ctime;
387 	cpipe->pipe_mtime = cpipe->pipe_ctime;
388 
389 	return (0);
390 }
391 
392 
393 /*
394  * lock a pipe for I/O, blocking other access
395  */
396 static __inline int
397 pipelock(cpipe, catch)
398 	struct pipe *cpipe;
399 	int catch;
400 {
401 	int error;
402 
403 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
404 	while (cpipe->pipe_state & PIPE_LOCKFL) {
405 		cpipe->pipe_state |= PIPE_LWANT;
406 		error = msleep(cpipe, PIPE_MTX(cpipe),
407 		    catch ? (PRIBIO | PCATCH) : PRIBIO,
408 		    "pipelk", 0);
409 		if (error != 0)
410 			return (error);
411 	}
412 	cpipe->pipe_state |= PIPE_LOCKFL;
413 	return (0);
414 }
415 
416 /*
417  * unlock a pipe I/O lock
418  */
419 static __inline void
420 pipeunlock(cpipe)
421 	struct pipe *cpipe;
422 {
423 
424 	PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
425 	cpipe->pipe_state &= ~PIPE_LOCKFL;
426 	if (cpipe->pipe_state & PIPE_LWANT) {
427 		cpipe->pipe_state &= ~PIPE_LWANT;
428 		wakeup(cpipe);
429 	}
430 }
431 
432 static __inline void
433 pipeselwakeup(cpipe)
434 	struct pipe *cpipe;
435 {
436 
437 	if (cpipe->pipe_state & PIPE_SEL) {
438 		cpipe->pipe_state &= ~PIPE_SEL;
439 		selwakeup(&cpipe->pipe_sel);
440 	}
441 	if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
442 		pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
443 	KNOTE(&cpipe->pipe_sel.si_note, 0);
444 }
445 
446 /* ARGSUSED */
447 static int
448 pipe_read(fp, uio, active_cred, flags, td)
449 	struct file *fp;
450 	struct uio *uio;
451 	struct ucred *active_cred;
452 	struct thread *td;
453 	int flags;
454 {
455 	struct pipe *rpipe = fp->f_data;
456 	int error;
457 	int nread = 0;
458 	u_int size;
459 
460 	PIPE_LOCK(rpipe);
461 	++rpipe->pipe_busy;
462 	error = pipelock(rpipe, 1);
463 	if (error)
464 		goto unlocked_error;
465 
466 #ifdef MAC
467 	error = mac_check_pipe_read(active_cred, rpipe);
468 	if (error)
469 		goto locked_error;
470 #endif
471 
472 	while (uio->uio_resid) {
473 		/*
474 		 * normal pipe buffer receive
475 		 */
476 		if (rpipe->pipe_buffer.cnt > 0) {
477 			size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
478 			if (size > rpipe->pipe_buffer.cnt)
479 				size = rpipe->pipe_buffer.cnt;
480 			if (size > (u_int) uio->uio_resid)
481 				size = (u_int) uio->uio_resid;
482 
483 			PIPE_UNLOCK(rpipe);
484 			error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
485 					size, uio);
486 			PIPE_LOCK(rpipe);
487 			if (error)
488 				break;
489 
490 			rpipe->pipe_buffer.out += size;
491 			if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
492 				rpipe->pipe_buffer.out = 0;
493 
494 			rpipe->pipe_buffer.cnt -= size;
495 
496 			/*
497 			 * If there is no more to read in the pipe, reset
498 			 * its pointers to the beginning.  This improves
499 			 * cache hit stats.
500 			 */
501 			if (rpipe->pipe_buffer.cnt == 0) {
502 				rpipe->pipe_buffer.in = 0;
503 				rpipe->pipe_buffer.out = 0;
504 			}
505 			nread += size;
506 #ifndef PIPE_NODIRECT
507 		/*
508 		 * Direct copy, bypassing a kernel buffer.
509 		 */
510 		} else if ((size = rpipe->pipe_map.cnt) &&
511 			   (rpipe->pipe_state & PIPE_DIRECTW)) {
512 			caddr_t	va;
513 			if (size > (u_int) uio->uio_resid)
514 				size = (u_int) uio->uio_resid;
515 
516 			va = (caddr_t) rpipe->pipe_map.kva +
517 			    rpipe->pipe_map.pos;
518 			PIPE_UNLOCK(rpipe);
519 			error = uiomove(va, size, uio);
520 			PIPE_LOCK(rpipe);
521 			if (error)
522 				break;
523 			nread += size;
524 			rpipe->pipe_map.pos += size;
525 			rpipe->pipe_map.cnt -= size;
526 			if (rpipe->pipe_map.cnt == 0) {
527 				rpipe->pipe_state &= ~PIPE_DIRECTW;
528 				wakeup(rpipe);
529 			}
530 #endif
531 		} else {
532 			/*
533 			 * detect EOF condition
534 			 * read returns 0 on EOF, no need to set error
535 			 */
536 			if (rpipe->pipe_state & PIPE_EOF)
537 				break;
538 
539 			/*
540 			 * If the "write-side" has been blocked, wake it up now.
541 			 */
542 			if (rpipe->pipe_state & PIPE_WANTW) {
543 				rpipe->pipe_state &= ~PIPE_WANTW;
544 				wakeup(rpipe);
545 			}
546 
547 			/*
548 			 * Break if some data was read.
549 			 */
550 			if (nread > 0)
551 				break;
552 
553 			/*
554 			 * Unlock the pipe buffer for our remaining processing.  We
555 			 * will either break out with an error or we will sleep and
556 			 * relock to loop.
557 			 */
558 			pipeunlock(rpipe);
559 
560 			/*
561 			 * Handle non-blocking mode operation or
562 			 * wait for more data.
563 			 */
564 			if (fp->f_flag & FNONBLOCK) {
565 				error = EAGAIN;
566 			} else {
567 				rpipe->pipe_state |= PIPE_WANTR;
568 				if ((error = msleep(rpipe, PIPE_MTX(rpipe),
569 				    PRIBIO | PCATCH,
570 				    "piperd", 0)) == 0)
571 					error = pipelock(rpipe, 1);
572 			}
573 			if (error)
574 				goto unlocked_error;
575 		}
576 	}
577 #ifdef MAC
578 locked_error:
579 #endif
580 	pipeunlock(rpipe);
581 
582 	/* XXX: should probably do this before getting any locks. */
583 	if (error == 0)
584 		vfs_timestamp(&rpipe->pipe_atime);
585 unlocked_error:
586 	--rpipe->pipe_busy;
587 
588 	/*
589 	 * PIPE_WANT processing only makes sense if pipe_busy is 0.
590 	 */
591 	if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
592 		rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
593 		wakeup(rpipe);
594 	} else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
595 		/*
596 		 * Handle write blocking hysteresis.
597 		 */
598 		if (rpipe->pipe_state & PIPE_WANTW) {
599 			rpipe->pipe_state &= ~PIPE_WANTW;
600 			wakeup(rpipe);
601 		}
602 	}
603 
604 	if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
605 		pipeselwakeup(rpipe);
606 
607 	PIPE_UNLOCK(rpipe);
608 	return (error);
609 }
610 
611 #ifndef PIPE_NODIRECT
612 /*
613  * Map the sending processes' buffer into kernel space and wire it.
614  * This is similar to a physical write operation.
615  */
616 static int
617 pipe_build_write_buffer(wpipe, uio)
618 	struct pipe *wpipe;
619 	struct uio *uio;
620 {
621 	u_int size;
622 	int i;
623 	vm_offset_t addr, endaddr;
624 	vm_paddr_t paddr;
625 
626 	GIANT_REQUIRED;
627 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
628 
629 	size = (u_int) uio->uio_iov->iov_len;
630 	if (size > wpipe->pipe_buffer.size)
631 		size = wpipe->pipe_buffer.size;
632 
633 	endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
634 	addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
635 	for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
636 		vm_page_t m;
637 
638 		/*
639 		 * vm_fault_quick() can sleep.  Consequently,
640 		 * vm_page_lock_queue() and vm_page_unlock_queue()
641 		 * should not be performed outside of this loop.
642 		 */
643 		if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
644 		    (paddr = pmap_extract(vmspace_pmap(curproc->p_vmspace),
645 		     addr)) == 0) {
646 			int j;
647 
648 			vm_page_lock_queues();
649 			for (j = 0; j < i; j++)
650 				vm_page_unwire(wpipe->pipe_map.ms[j], 1);
651 			vm_page_unlock_queues();
652 			return (EFAULT);
653 		}
654 
655 		m = PHYS_TO_VM_PAGE(paddr);
656 		vm_page_lock_queues();
657 		vm_page_wire(m);
658 		vm_page_unlock_queues();
659 		wpipe->pipe_map.ms[i] = m;
660 	}
661 
662 /*
663  * set up the control block
664  */
665 	wpipe->pipe_map.npages = i;
666 	wpipe->pipe_map.pos =
667 	    ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
668 	wpipe->pipe_map.cnt = size;
669 
670 /*
671  * and map the buffer
672  */
673 	if (wpipe->pipe_map.kva == 0) {
674 		/*
675 		 * We need to allocate space for an extra page because the
676 		 * address range might (will) span pages at times.
677 		 */
678 		wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
679 			wpipe->pipe_buffer.size + PAGE_SIZE);
680 		atomic_add_int(&amountpipekva,
681 		    wpipe->pipe_buffer.size + PAGE_SIZE);
682 	}
683 	pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
684 		wpipe->pipe_map.npages);
685 
686 /*
687  * and update the uio data
688  */
689 
690 	uio->uio_iov->iov_len -= size;
691 	uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
692 	if (uio->uio_iov->iov_len == 0)
693 		uio->uio_iov++;
694 	uio->uio_resid -= size;
695 	uio->uio_offset += size;
696 	return (0);
697 }
698 
699 /*
700  * unmap and unwire the process buffer
701  */
702 static void
703 pipe_destroy_write_buffer(wpipe)
704 	struct pipe *wpipe;
705 {
706 	int i;
707 
708 	GIANT_REQUIRED;
709 	PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
710 
711 	if (wpipe->pipe_map.kva) {
712 		pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
713 
714 		if (amountpipekva > MAXPIPEKVA) {
715 			vm_offset_t kva = wpipe->pipe_map.kva;
716 			wpipe->pipe_map.kva = 0;
717 			kmem_free(kernel_map, kva,
718 				wpipe->pipe_buffer.size + PAGE_SIZE);
719 			atomic_subtract_int(&amountpipekva,
720 			    wpipe->pipe_buffer.size + PAGE_SIZE);
721 		}
722 	}
723 	vm_page_lock_queues();
724 	for (i = 0; i < wpipe->pipe_map.npages; i++)
725 		vm_page_unwire(wpipe->pipe_map.ms[i], 1);
726 	vm_page_unlock_queues();
727 	wpipe->pipe_map.npages = 0;
728 }
729 
730 /*
731  * In the case of a signal, the writing process might go away.  This
732  * code copies the data into the circular buffer so that the source
733  * pages can be freed without loss of data.
734  */
735 static void
736 pipe_clone_write_buffer(wpipe)
737 	struct pipe *wpipe;
738 {
739 	int size;
740 	int pos;
741 
742 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
743 	size = wpipe->pipe_map.cnt;
744 	pos = wpipe->pipe_map.pos;
745 
746 	wpipe->pipe_buffer.in = size;
747 	wpipe->pipe_buffer.out = 0;
748 	wpipe->pipe_buffer.cnt = size;
749 	wpipe->pipe_state &= ~PIPE_DIRECTW;
750 
751 	PIPE_GET_GIANT(wpipe);
752 	bcopy((caddr_t) wpipe->pipe_map.kva + pos,
753 	    wpipe->pipe_buffer.buffer, size);
754 	pipe_destroy_write_buffer(wpipe);
755 	PIPE_DROP_GIANT(wpipe);
756 }
757 
758 /*
759  * This implements the pipe buffer write mechanism.  Note that only
760  * a direct write OR a normal pipe write can be pending at any given time.
761  * If there are any characters in the pipe buffer, the direct write will
762  * be deferred until the receiving process grabs all of the bytes from
763  * the pipe buffer.  Then the direct mapping write is set-up.
764  */
765 static int
766 pipe_direct_write(wpipe, uio)
767 	struct pipe *wpipe;
768 	struct uio *uio;
769 {
770 	int error;
771 
772 retry:
773 	PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
774 	while (wpipe->pipe_state & PIPE_DIRECTW) {
775 		if (wpipe->pipe_state & PIPE_WANTR) {
776 			wpipe->pipe_state &= ~PIPE_WANTR;
777 			wakeup(wpipe);
778 		}
779 		wpipe->pipe_state |= PIPE_WANTW;
780 		error = msleep(wpipe, PIPE_MTX(wpipe),
781 		    PRIBIO | PCATCH, "pipdww", 0);
782 		if (error)
783 			goto error1;
784 		if (wpipe->pipe_state & PIPE_EOF) {
785 			error = EPIPE;
786 			goto error1;
787 		}
788 	}
789 	wpipe->pipe_map.cnt = 0;	/* transfer not ready yet */
790 	if (wpipe->pipe_buffer.cnt > 0) {
791 		if (wpipe->pipe_state & PIPE_WANTR) {
792 			wpipe->pipe_state &= ~PIPE_WANTR;
793 			wakeup(wpipe);
794 		}
795 
796 		wpipe->pipe_state |= PIPE_WANTW;
797 		error = msleep(wpipe, PIPE_MTX(wpipe),
798 		    PRIBIO | PCATCH, "pipdwc", 0);
799 		if (error)
800 			goto error1;
801 		if (wpipe->pipe_state & PIPE_EOF) {
802 			error = EPIPE;
803 			goto error1;
804 		}
805 		goto retry;
806 	}
807 
808 	wpipe->pipe_state |= PIPE_DIRECTW;
809 
810 	pipelock(wpipe, 0);
811 	PIPE_GET_GIANT(wpipe);
812 	error = pipe_build_write_buffer(wpipe, uio);
813 	PIPE_DROP_GIANT(wpipe);
814 	pipeunlock(wpipe);
815 	if (error) {
816 		wpipe->pipe_state &= ~PIPE_DIRECTW;
817 		goto error1;
818 	}
819 
820 	error = 0;
821 	while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
822 		if (wpipe->pipe_state & PIPE_EOF) {
823 			pipelock(wpipe, 0);
824 			PIPE_GET_GIANT(wpipe);
825 			pipe_destroy_write_buffer(wpipe);
826 			PIPE_DROP_GIANT(wpipe);
827 			pipeunlock(wpipe);
828 			pipeselwakeup(wpipe);
829 			error = EPIPE;
830 			goto error1;
831 		}
832 		if (wpipe->pipe_state & PIPE_WANTR) {
833 			wpipe->pipe_state &= ~PIPE_WANTR;
834 			wakeup(wpipe);
835 		}
836 		pipeselwakeup(wpipe);
837 		error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
838 		    "pipdwt", 0);
839 	}
840 
841 	pipelock(wpipe,0);
842 	if (wpipe->pipe_state & PIPE_DIRECTW) {
843 		/*
844 		 * this bit of trickery substitutes a kernel buffer for
845 		 * the process that might be going away.
846 		 */
847 		pipe_clone_write_buffer(wpipe);
848 	} else {
849 		PIPE_GET_GIANT(wpipe);
850 		pipe_destroy_write_buffer(wpipe);
851 		PIPE_DROP_GIANT(wpipe);
852 	}
853 	pipeunlock(wpipe);
854 	return (error);
855 
856 error1:
857 	wakeup(wpipe);
858 	return (error);
859 }
860 #endif
861 
862 static int
863 pipe_write(fp, uio, active_cred, flags, td)
864 	struct file *fp;
865 	struct uio *uio;
866 	struct ucred *active_cred;
867 	struct thread *td;
868 	int flags;
869 {
870 	int error = 0;
871 	int orig_resid;
872 	struct pipe *wpipe, *rpipe;
873 
874 	rpipe = fp->f_data;
875 	wpipe = rpipe->pipe_peer;
876 
877 	PIPE_LOCK(rpipe);
878 	/*
879 	 * detect loss of pipe read side, issue SIGPIPE if lost.
880 	 */
881 	if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
882 		PIPE_UNLOCK(rpipe);
883 		return (EPIPE);
884 	}
885 #ifdef MAC
886 	error = mac_check_pipe_write(active_cred, wpipe);
887 	if (error) {
888 		PIPE_UNLOCK(rpipe);
889 		return (error);
890 	}
891 #endif
892 	++wpipe->pipe_busy;
893 
894 	/*
895 	 * If it is advantageous to resize the pipe buffer, do
896 	 * so.
897 	 */
898 	if ((uio->uio_resid > PIPE_SIZE) &&
899 		(nbigpipe < LIMITBIGPIPES) &&
900 		(wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
901 		(wpipe->pipe_buffer.size <= PIPE_SIZE) &&
902 		(wpipe->pipe_buffer.cnt == 0)) {
903 
904 		if ((error = pipelock(wpipe, 1)) == 0) {
905 			PIPE_GET_GIANT(wpipe);
906 			if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
907 				nbigpipe++;
908 			PIPE_DROP_GIANT(wpipe);
909 			pipeunlock(wpipe);
910 		}
911 	}
912 
913 	/*
914 	 * If an early error occured unbusy and return, waking up any pending
915 	 * readers.
916 	 */
917 	if (error) {
918 		--wpipe->pipe_busy;
919 		if ((wpipe->pipe_busy == 0) &&
920 		    (wpipe->pipe_state & PIPE_WANT)) {
921 			wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
922 			wakeup(wpipe);
923 		}
924 		PIPE_UNLOCK(rpipe);
925 		return(error);
926 	}
927 
928 	orig_resid = uio->uio_resid;
929 
930 	while (uio->uio_resid) {
931 		int space;
932 
933 #ifndef PIPE_NODIRECT
934 		/*
935 		 * If the transfer is large, we can gain performance if
936 		 * we do process-to-process copies directly.
937 		 * If the write is non-blocking, we don't use the
938 		 * direct write mechanism.
939 		 *
940 		 * The direct write mechanism will detect the reader going
941 		 * away on us.
942 		 */
943 		if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
944 		    (fp->f_flag & FNONBLOCK) == 0 &&
945 			(wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) &&
946 			(uio->uio_iov->iov_len >= PIPE_MINDIRECT)) {
947 			error = pipe_direct_write(wpipe, uio);
948 			if (error)
949 				break;
950 			continue;
951 		}
952 #endif
953 
954 		/*
955 		 * Pipe buffered writes cannot be coincidental with
956 		 * direct writes.  We wait until the currently executing
957 		 * direct write is completed before we start filling the
958 		 * pipe buffer.  We break out if a signal occurs or the
959 		 * reader goes away.
960 		 */
961 	retrywrite:
962 		while (wpipe->pipe_state & PIPE_DIRECTW) {
963 			if (wpipe->pipe_state & PIPE_WANTR) {
964 				wpipe->pipe_state &= ~PIPE_WANTR;
965 				wakeup(wpipe);
966 			}
967 			error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
968 			    "pipbww", 0);
969 			if (wpipe->pipe_state & PIPE_EOF)
970 				break;
971 			if (error)
972 				break;
973 		}
974 		if (wpipe->pipe_state & PIPE_EOF) {
975 			error = EPIPE;
976 			break;
977 		}
978 
979 		space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
980 
981 		/* Writes of size <= PIPE_BUF must be atomic. */
982 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
983 			space = 0;
984 
985 		if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
986 			if ((error = pipelock(wpipe,1)) == 0) {
987 				int size;	/* Transfer size */
988 				int segsize;	/* first segment to transfer */
989 
990 				/*
991 				 * It is possible for a direct write to
992 				 * slip in on us... handle it here...
993 				 */
994 				if (wpipe->pipe_state & PIPE_DIRECTW) {
995 					pipeunlock(wpipe);
996 					goto retrywrite;
997 				}
998 				/*
999 				 * If a process blocked in uiomove, our
1000 				 * value for space might be bad.
1001 				 *
1002 				 * XXX will we be ok if the reader has gone
1003 				 * away here?
1004 				 */
1005 				if (space > wpipe->pipe_buffer.size -
1006 				    wpipe->pipe_buffer.cnt) {
1007 					pipeunlock(wpipe);
1008 					goto retrywrite;
1009 				}
1010 
1011 				/*
1012 				 * Transfer size is minimum of uio transfer
1013 				 * and free space in pipe buffer.
1014 				 */
1015 				if (space > uio->uio_resid)
1016 					size = uio->uio_resid;
1017 				else
1018 					size = space;
1019 				/*
1020 				 * First segment to transfer is minimum of
1021 				 * transfer size and contiguous space in
1022 				 * pipe buffer.  If first segment to transfer
1023 				 * is less than the transfer size, we've got
1024 				 * a wraparound in the buffer.
1025 				 */
1026 				segsize = wpipe->pipe_buffer.size -
1027 					wpipe->pipe_buffer.in;
1028 				if (segsize > size)
1029 					segsize = size;
1030 
1031 				/* Transfer first segment */
1032 
1033 				PIPE_UNLOCK(rpipe);
1034 				error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1035 						segsize, uio);
1036 				PIPE_LOCK(rpipe);
1037 
1038 				if (error == 0 && segsize < size) {
1039 					/*
1040 					 * Transfer remaining part now, to
1041 					 * support atomic writes.  Wraparound
1042 					 * happened.
1043 					 */
1044 					if (wpipe->pipe_buffer.in + segsize !=
1045 					    wpipe->pipe_buffer.size)
1046 						panic("Expected pipe buffer wraparound disappeared");
1047 
1048 					PIPE_UNLOCK(rpipe);
1049 					error = uiomove(&wpipe->pipe_buffer.buffer[0],
1050 							size - segsize, uio);
1051 					PIPE_LOCK(rpipe);
1052 				}
1053 				if (error == 0) {
1054 					wpipe->pipe_buffer.in += size;
1055 					if (wpipe->pipe_buffer.in >=
1056 					    wpipe->pipe_buffer.size) {
1057 						if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1058 							panic("Expected wraparound bad");
1059 						wpipe->pipe_buffer.in = size - segsize;
1060 					}
1061 
1062 					wpipe->pipe_buffer.cnt += size;
1063 					if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1064 						panic("Pipe buffer overflow");
1065 
1066 				}
1067 				pipeunlock(wpipe);
1068 			}
1069 			if (error)
1070 				break;
1071 
1072 		} else {
1073 			/*
1074 			 * If the "read-side" has been blocked, wake it up now.
1075 			 */
1076 			if (wpipe->pipe_state & PIPE_WANTR) {
1077 				wpipe->pipe_state &= ~PIPE_WANTR;
1078 				wakeup(wpipe);
1079 			}
1080 
1081 			/*
1082 			 * don't block on non-blocking I/O
1083 			 */
1084 			if (fp->f_flag & FNONBLOCK) {
1085 				error = EAGAIN;
1086 				break;
1087 			}
1088 
1089 			/*
1090 			 * We have no more space and have something to offer,
1091 			 * wake up select/poll.
1092 			 */
1093 			pipeselwakeup(wpipe);
1094 
1095 			wpipe->pipe_state |= PIPE_WANTW;
1096 			error = msleep(wpipe, PIPE_MTX(rpipe),
1097 			    PRIBIO | PCATCH, "pipewr", 0);
1098 			if (error != 0)
1099 				break;
1100 			/*
1101 			 * If read side wants to go away, we just issue a signal
1102 			 * to ourselves.
1103 			 */
1104 			if (wpipe->pipe_state & PIPE_EOF) {
1105 				error = EPIPE;
1106 				break;
1107 			}
1108 		}
1109 	}
1110 
1111 	--wpipe->pipe_busy;
1112 
1113 	if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1114 		wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1115 		wakeup(wpipe);
1116 	} else if (wpipe->pipe_buffer.cnt > 0) {
1117 		/*
1118 		 * If we have put any characters in the buffer, we wake up
1119 		 * the reader.
1120 		 */
1121 		if (wpipe->pipe_state & PIPE_WANTR) {
1122 			wpipe->pipe_state &= ~PIPE_WANTR;
1123 			wakeup(wpipe);
1124 		}
1125 	}
1126 
1127 	/*
1128 	 * Don't return EPIPE if I/O was successful
1129 	 */
1130 	if ((wpipe->pipe_buffer.cnt == 0) &&
1131 	    (uio->uio_resid == 0) &&
1132 	    (error == EPIPE)) {
1133 		error = 0;
1134 	}
1135 
1136 	if (error == 0)
1137 		vfs_timestamp(&wpipe->pipe_mtime);
1138 
1139 	/*
1140 	 * We have something to offer,
1141 	 * wake up select/poll.
1142 	 */
1143 	if (wpipe->pipe_buffer.cnt)
1144 		pipeselwakeup(wpipe);
1145 
1146 	PIPE_UNLOCK(rpipe);
1147 	return (error);
1148 }
1149 
1150 /*
1151  * we implement a very minimal set of ioctls for compatibility with sockets.
1152  */
1153 static int
1154 pipe_ioctl(fp, cmd, data, active_cred, td)
1155 	struct file *fp;
1156 	u_long cmd;
1157 	void *data;
1158 	struct ucred *active_cred;
1159 	struct thread *td;
1160 {
1161 	struct pipe *mpipe = fp->f_data;
1162 #ifdef MAC
1163 	int error;
1164 #endif
1165 
1166 	PIPE_LOCK(mpipe);
1167 
1168 #ifdef MAC
1169 	error = mac_check_pipe_ioctl(active_cred, mpipe, cmd, data);
1170 	if (error)
1171 		return (error);
1172 #endif
1173 
1174 	switch (cmd) {
1175 
1176 	case FIONBIO:
1177 		PIPE_UNLOCK(mpipe);
1178 		return (0);
1179 
1180 	case FIOASYNC:
1181 		if (*(int *)data) {
1182 			mpipe->pipe_state |= PIPE_ASYNC;
1183 		} else {
1184 			mpipe->pipe_state &= ~PIPE_ASYNC;
1185 		}
1186 		PIPE_UNLOCK(mpipe);
1187 		return (0);
1188 
1189 	case FIONREAD:
1190 		if (mpipe->pipe_state & PIPE_DIRECTW)
1191 			*(int *)data = mpipe->pipe_map.cnt;
1192 		else
1193 			*(int *)data = mpipe->pipe_buffer.cnt;
1194 		PIPE_UNLOCK(mpipe);
1195 		return (0);
1196 
1197 	case FIOSETOWN:
1198 		PIPE_UNLOCK(mpipe);
1199 		return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1200 
1201 	case FIOGETOWN:
1202 		PIPE_UNLOCK(mpipe);
1203 		*(int *)data = fgetown(&mpipe->pipe_sigio);
1204 		return (0);
1205 
1206 	/* This is deprecated, FIOSETOWN should be used instead. */
1207 	case TIOCSPGRP:
1208 		PIPE_UNLOCK(mpipe);
1209 		return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1210 
1211 	/* This is deprecated, FIOGETOWN should be used instead. */
1212 	case TIOCGPGRP:
1213 		PIPE_UNLOCK(mpipe);
1214 		*(int *)data = -fgetown(&mpipe->pipe_sigio);
1215 		return (0);
1216 
1217 	}
1218 	PIPE_UNLOCK(mpipe);
1219 	return (ENOTTY);
1220 }
1221 
1222 static int
1223 pipe_poll(fp, events, active_cred, td)
1224 	struct file *fp;
1225 	int events;
1226 	struct ucred *active_cred;
1227 	struct thread *td;
1228 {
1229 	struct pipe *rpipe = fp->f_data;
1230 	struct pipe *wpipe;
1231 	int revents = 0;
1232 #ifdef MAC
1233 	int error;
1234 #endif
1235 
1236 	wpipe = rpipe->pipe_peer;
1237 	PIPE_LOCK(rpipe);
1238 #ifdef MAC
1239 	error = mac_check_pipe_poll(active_cred, rpipe);
1240 	if (error)
1241 		goto locked_error;
1242 #endif
1243 	if (events & (POLLIN | POLLRDNORM))
1244 		if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1245 		    (rpipe->pipe_buffer.cnt > 0) ||
1246 		    (rpipe->pipe_state & PIPE_EOF))
1247 			revents |= events & (POLLIN | POLLRDNORM);
1248 
1249 	if (events & (POLLOUT | POLLWRNORM))
1250 		if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1251 		    (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1252 		     (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1253 			revents |= events & (POLLOUT | POLLWRNORM);
1254 
1255 	if ((rpipe->pipe_state & PIPE_EOF) ||
1256 	    (wpipe == NULL) ||
1257 	    (wpipe->pipe_state & PIPE_EOF))
1258 		revents |= POLLHUP;
1259 
1260 	if (revents == 0) {
1261 		if (events & (POLLIN | POLLRDNORM)) {
1262 			selrecord(td, &rpipe->pipe_sel);
1263 			rpipe->pipe_state |= PIPE_SEL;
1264 		}
1265 
1266 		if (events & (POLLOUT | POLLWRNORM)) {
1267 			selrecord(td, &wpipe->pipe_sel);
1268 			wpipe->pipe_state |= PIPE_SEL;
1269 		}
1270 	}
1271 #ifdef MAC
1272 locked_error:
1273 #endif
1274 	PIPE_UNLOCK(rpipe);
1275 
1276 	return (revents);
1277 }
1278 
1279 /*
1280  * We shouldn't need locks here as we're doing a read and this should
1281  * be a natural race.
1282  */
1283 static int
1284 pipe_stat(fp, ub, active_cred, td)
1285 	struct file *fp;
1286 	struct stat *ub;
1287 	struct ucred *active_cred;
1288 	struct thread *td;
1289 {
1290 	struct pipe *pipe = fp->f_data;
1291 #ifdef MAC
1292 	int error;
1293 
1294 	PIPE_LOCK(pipe);
1295 	error = mac_check_pipe_stat(active_cred, pipe);
1296 	PIPE_UNLOCK(pipe);
1297 	if (error)
1298 		return (error);
1299 #endif
1300 	bzero(ub, sizeof(*ub));
1301 	ub->st_mode = S_IFIFO;
1302 	ub->st_blksize = pipe->pipe_buffer.size;
1303 	ub->st_size = pipe->pipe_buffer.cnt;
1304 	ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1305 	ub->st_atimespec = pipe->pipe_atime;
1306 	ub->st_mtimespec = pipe->pipe_mtime;
1307 	ub->st_ctimespec = pipe->pipe_ctime;
1308 	ub->st_uid = fp->f_cred->cr_uid;
1309 	ub->st_gid = fp->f_cred->cr_gid;
1310 	/*
1311 	 * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1312 	 * XXX (st_dev, st_ino) should be unique.
1313 	 */
1314 	return (0);
1315 }
1316 
1317 /* ARGSUSED */
1318 static int
1319 pipe_close(fp, td)
1320 	struct file *fp;
1321 	struct thread *td;
1322 {
1323 	struct pipe *cpipe = fp->f_data;
1324 
1325 	fp->f_ops = &badfileops;
1326 	fp->f_data = NULL;
1327 	funsetown(&cpipe->pipe_sigio);
1328 	pipeclose(cpipe);
1329 	return (0);
1330 }
1331 
1332 static void
1333 pipe_free_kmem(cpipe)
1334 	struct pipe *cpipe;
1335 {
1336 
1337 	GIANT_REQUIRED;
1338 	KASSERT(cpipe->pipe_mtxp == NULL || !mtx_owned(PIPE_MTX(cpipe)),
1339 	       ("pipespace: pipe mutex locked"));
1340 
1341 	if (cpipe->pipe_buffer.buffer != NULL) {
1342 		if (cpipe->pipe_buffer.size > PIPE_SIZE)
1343 			--nbigpipe;
1344 		atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size);
1345 		kmem_free(kernel_map,
1346 			(vm_offset_t)cpipe->pipe_buffer.buffer,
1347 			cpipe->pipe_buffer.size);
1348 		cpipe->pipe_buffer.buffer = NULL;
1349 	}
1350 #ifndef PIPE_NODIRECT
1351 	if (cpipe->pipe_map.kva != 0) {
1352 		atomic_subtract_int(&amountpipekva,
1353 		    cpipe->pipe_buffer.size + PAGE_SIZE);
1354 		kmem_free(kernel_map,
1355 			cpipe->pipe_map.kva,
1356 			cpipe->pipe_buffer.size + PAGE_SIZE);
1357 		cpipe->pipe_map.cnt = 0;
1358 		cpipe->pipe_map.kva = 0;
1359 		cpipe->pipe_map.pos = 0;
1360 		cpipe->pipe_map.npages = 0;
1361 	}
1362 #endif
1363 }
1364 
1365 /*
1366  * shutdown the pipe
1367  */
1368 static void
1369 pipeclose(cpipe)
1370 	struct pipe *cpipe;
1371 {
1372 	struct pipe *ppipe;
1373 	int hadpeer;
1374 
1375 	if (cpipe == NULL)
1376 		return;
1377 
1378 	hadpeer = 0;
1379 
1380 	/* partially created pipes won't have a valid mutex. */
1381 	if (PIPE_MTX(cpipe) != NULL)
1382 		PIPE_LOCK(cpipe);
1383 
1384 	pipeselwakeup(cpipe);
1385 
1386 	/*
1387 	 * If the other side is blocked, wake it up saying that
1388 	 * we want to close it down.
1389 	 */
1390 	while (cpipe->pipe_busy) {
1391 		wakeup(cpipe);
1392 		cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1393 		msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1394 	}
1395 
1396 #ifdef MAC
1397 	if (cpipe->pipe_label != NULL && cpipe->pipe_peer == NULL)
1398 		mac_destroy_pipe(cpipe);
1399 #endif
1400 
1401 	/*
1402 	 * Disconnect from peer
1403 	 */
1404 	if ((ppipe = cpipe->pipe_peer) != NULL) {
1405 		hadpeer++;
1406 		pipeselwakeup(ppipe);
1407 
1408 		ppipe->pipe_state |= PIPE_EOF;
1409 		wakeup(ppipe);
1410 		KNOTE(&ppipe->pipe_sel.si_note, 0);
1411 		ppipe->pipe_peer = NULL;
1412 	}
1413 	/*
1414 	 * free resources
1415 	 */
1416 	if (PIPE_MTX(cpipe) != NULL) {
1417 		PIPE_UNLOCK(cpipe);
1418 		if (!hadpeer) {
1419 			mtx_destroy(PIPE_MTX(cpipe));
1420 			free(PIPE_MTX(cpipe), M_TEMP);
1421 		}
1422 	}
1423 	mtx_lock(&Giant);
1424 	pipe_free_kmem(cpipe);
1425 	uma_zfree(pipe_zone, cpipe);
1426 	mtx_unlock(&Giant);
1427 }
1428 
1429 /*ARGSUSED*/
1430 static int
1431 pipe_kqfilter(struct file *fp, struct knote *kn)
1432 {
1433 	struct pipe *cpipe;
1434 
1435 	cpipe = kn->kn_fp->f_data;
1436 	switch (kn->kn_filter) {
1437 	case EVFILT_READ:
1438 		kn->kn_fop = &pipe_rfiltops;
1439 		break;
1440 	case EVFILT_WRITE:
1441 		kn->kn_fop = &pipe_wfiltops;
1442 		cpipe = cpipe->pipe_peer;
1443 		if (cpipe == NULL)
1444 			/* other end of pipe has been closed */
1445 			return (EBADF);
1446 		break;
1447 	default:
1448 		return (1);
1449 	}
1450 	kn->kn_hook = cpipe;
1451 
1452 	PIPE_LOCK(cpipe);
1453 	SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1454 	PIPE_UNLOCK(cpipe);
1455 	return (0);
1456 }
1457 
1458 static void
1459 filt_pipedetach(struct knote *kn)
1460 {
1461 	struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1462 
1463 	PIPE_LOCK(cpipe);
1464 	SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1465 	PIPE_UNLOCK(cpipe);
1466 }
1467 
1468 /*ARGSUSED*/
1469 static int
1470 filt_piperead(struct knote *kn, long hint)
1471 {
1472 	struct pipe *rpipe = kn->kn_fp->f_data;
1473 	struct pipe *wpipe = rpipe->pipe_peer;
1474 
1475 	PIPE_LOCK(rpipe);
1476 	kn->kn_data = rpipe->pipe_buffer.cnt;
1477 	if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1478 		kn->kn_data = rpipe->pipe_map.cnt;
1479 
1480 	if ((rpipe->pipe_state & PIPE_EOF) ||
1481 	    (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1482 		kn->kn_flags |= EV_EOF;
1483 		PIPE_UNLOCK(rpipe);
1484 		return (1);
1485 	}
1486 	PIPE_UNLOCK(rpipe);
1487 	return (kn->kn_data > 0);
1488 }
1489 
1490 /*ARGSUSED*/
1491 static int
1492 filt_pipewrite(struct knote *kn, long hint)
1493 {
1494 	struct pipe *rpipe = kn->kn_fp->f_data;
1495 	struct pipe *wpipe = rpipe->pipe_peer;
1496 
1497 	PIPE_LOCK(rpipe);
1498 	if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1499 		kn->kn_data = 0;
1500 		kn->kn_flags |= EV_EOF;
1501 		PIPE_UNLOCK(rpipe);
1502 		return (1);
1503 	}
1504 	kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1505 	if (wpipe->pipe_state & PIPE_DIRECTW)
1506 		kn->kn_data = 0;
1507 
1508 	PIPE_UNLOCK(rpipe);
1509 	return (kn->kn_data >= PIPE_BUF);
1510 }
1511