2 * Copyright (c) 1996 John S. Dyson
3 * Copyright (c) 2012 Giovanni Trematerra
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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
17 * 4. Modifications may be freely made to this file if the above conditions
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
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.
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.
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.
49 * In order to limit the resource use of pipes, two sysctls exist:
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.
55 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56 * memory in use by pipes.
58 * Based on how large pipekva is relative to maxpipekva, the following
62 * New pipes are given 16K of memory backing, pipes may dynamically
63 * grow to as large as 64K where needed.
65 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
66 * existing pipes may NOT grow.
68 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
69 * existing pipes will be shrunk down to 4K whenever possible.
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
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.
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.
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.
92 #include <sys/cdefs.h>
93 __FBSDID("$FreeBSD$");
95 #include <sys/param.h>
96 #include <sys/systm.h>
98 #include <sys/fcntl.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/syscallsubr.h>
112 #include <sys/sysctl.h>
113 #include <sys/sysproto.h>
114 #include <sys/pipe.h>
115 #include <sys/proc.h>
116 #include <sys/vnode.h>
118 #include <sys/user.h>
119 #include <sys/event.h>
121 #include <security/mac/mac_framework.h>
124 #include <vm/vm_param.h>
125 #include <vm/vm_object.h>
126 #include <vm/vm_kern.h>
127 #include <vm/vm_extern.h>
129 #include <vm/vm_map.h>
130 #include <vm/vm_page.h>
134 * Use this define if you want to disable *fancy* VM things. Expect an
135 * approx 30% decrease in transfer rate. This could be useful for
138 /* #define PIPE_NODIRECT */
140 #define PIPE_PEER(pipe) \
141 (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
144 * interfaces to the outside world
146 static fo_rdwr_t pipe_read;
147 static fo_rdwr_t pipe_write;
148 static fo_truncate_t pipe_truncate;
149 static fo_ioctl_t pipe_ioctl;
150 static fo_poll_t pipe_poll;
151 static fo_kqfilter_t pipe_kqfilter;
152 static fo_stat_t pipe_stat;
153 static fo_close_t pipe_close;
154 static fo_chmod_t pipe_chmod;
155 static fo_chown_t pipe_chown;
156 static fo_fill_kinfo_t pipe_fill_kinfo;
158 struct fileops pipeops = {
159 .fo_read = pipe_read,
160 .fo_write = pipe_write,
161 .fo_truncate = pipe_truncate,
162 .fo_ioctl = pipe_ioctl,
163 .fo_poll = pipe_poll,
164 .fo_kqfilter = pipe_kqfilter,
165 .fo_stat = pipe_stat,
166 .fo_close = pipe_close,
167 .fo_chmod = pipe_chmod,
168 .fo_chown = pipe_chown,
169 .fo_sendfile = invfo_sendfile,
170 .fo_fill_kinfo = pipe_fill_kinfo,
171 .fo_flags = DFLAG_PASSABLE
174 static void filt_pipedetach(struct knote *kn);
175 static void filt_pipedetach_notsup(struct knote *kn);
176 static int filt_pipenotsup(struct knote *kn, long hint);
177 static int filt_piperead(struct knote *kn, long hint);
178 static int filt_pipewrite(struct knote *kn, long hint);
180 static struct filterops pipe_nfiltops = {
182 .f_detach = filt_pipedetach_notsup,
183 .f_event = filt_pipenotsup
185 static struct filterops pipe_rfiltops = {
187 .f_detach = filt_pipedetach,
188 .f_event = filt_piperead
190 static struct filterops pipe_wfiltops = {
192 .f_detach = filt_pipedetach,
193 .f_event = filt_pipewrite
197 * Default pipe buffer size(s), this can be kind-of large now because pipe
198 * space is pageable. The pipe code will try to maintain locality of
199 * reference for performance reasons, so small amounts of outstanding I/O
200 * will not wipe the cache.
202 #define MINPIPESIZE (PIPE_SIZE/3)
203 #define MAXPIPESIZE (2*PIPE_SIZE/3)
205 static long amountpipekva;
206 static int pipefragretry;
207 static int pipeallocfail;
208 static int piperesizefail;
209 static int piperesizeallowed = 1;
211 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
212 &maxpipekva, 0, "Pipe KVA limit");
213 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
214 &amountpipekva, 0, "Pipe KVA usage");
215 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
216 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
218 &pipeallocfail, 0, "Pipe allocation failures");
219 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
220 &piperesizefail, 0, "Pipe resize failures");
221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
222 &piperesizeallowed, 0, "Pipe resizing allowed");
224 static void pipeinit(void *dummy __unused);
225 static void pipeclose(struct pipe *cpipe);
226 static void pipe_free_kmem(struct pipe *cpipe);
227 static void pipe_create(struct pipe *pipe, int backing);
228 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
229 static __inline int pipelock(struct pipe *cpipe, int catch);
230 static __inline void pipeunlock(struct pipe *cpipe);
231 #ifndef PIPE_NODIRECT
232 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
233 static void pipe_destroy_write_buffer(struct pipe *wpipe);
234 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
235 static void pipe_clone_write_buffer(struct pipe *wpipe);
237 static int pipespace(struct pipe *cpipe, int size);
238 static int pipespace_new(struct pipe *cpipe, int size);
240 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
241 static int pipe_zone_init(void *mem, int size, int flags);
242 static void pipe_zone_fini(void *mem, int size);
244 static uma_zone_t pipe_zone;
245 static struct unrhdr *pipeino_unr;
246 static dev_t pipedev_ino;
248 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
251 pipeinit(void *dummy __unused)
254 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
255 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
257 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
258 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
259 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
260 pipedev_ino = devfs_alloc_cdp_inode();
261 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
265 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
268 struct pipe *rpipe, *wpipe;
270 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
272 pp = (struct pipepair *)mem;
275 * We zero both pipe endpoints to make sure all the kmem pointers
276 * are NULL, flag fields are zero'd, etc. We timestamp both
277 * endpoints with the same time.
279 rpipe = &pp->pp_rpipe;
280 bzero(rpipe, sizeof(*rpipe));
281 vfs_timestamp(&rpipe->pipe_ctime);
282 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
284 wpipe = &pp->pp_wpipe;
285 bzero(wpipe, sizeof(*wpipe));
286 wpipe->pipe_ctime = rpipe->pipe_ctime;
287 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
289 rpipe->pipe_peer = wpipe;
290 rpipe->pipe_pair = pp;
291 wpipe->pipe_peer = rpipe;
292 wpipe->pipe_pair = pp;
295 * Mark both endpoints as present; they will later get free'd
296 * one at a time. When both are free'd, then the whole pair
299 rpipe->pipe_present = PIPE_ACTIVE;
300 wpipe->pipe_present = PIPE_ACTIVE;
303 * Eventually, the MAC Framework may initialize the label
304 * in ctor or init, but for now we do it elswhere to avoid
305 * blocking in ctor or init.
313 pipe_zone_init(void *mem, int size, int flags)
317 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
319 pp = (struct pipepair *)mem;
321 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
326 pipe_zone_fini(void *mem, int size)
330 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
332 pp = (struct pipepair *)mem;
334 mtx_destroy(&pp->pp_mtx);
338 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
341 struct pipe *rpipe, *wpipe;
343 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
346 * The MAC label is shared between the connected endpoints. As a
347 * result mac_pipe_init() and mac_pipe_create() are called once
348 * for the pair, and not on the endpoints.
351 mac_pipe_create(td->td_ucred, pp);
353 rpipe = &pp->pp_rpipe;
354 wpipe = &pp->pp_wpipe;
356 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
357 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
359 /* Only the forward direction pipe is backed by default */
360 pipe_create(rpipe, 1);
361 pipe_create(wpipe, 0);
363 rpipe->pipe_state |= PIPE_DIRECTOK;
364 wpipe->pipe_state |= PIPE_DIRECTOK;
368 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
372 pipe_paircreate(td, &pp);
373 pp->pp_rpipe.pipe_state |= PIPE_NAMED;
374 *ppipe = &pp->pp_rpipe;
378 pipe_dtor(struct pipe *dpipe)
383 ino = dpipe->pipe_ino;
384 peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
385 funsetown(&dpipe->pipe_sigio);
388 funsetown(&peer->pipe_sigio);
391 if (ino != 0 && ino != (ino_t)-1)
392 free_unr(pipeino_unr, ino);
396 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
397 * the zone pick up the pieces via pipeclose().
400 kern_pipe(struct thread *td, int fildes[2])
403 return (kern_pipe2(td, fildes, 0));
407 kern_pipe2(struct thread *td, int fildes[2], int flags)
409 struct file *rf, *wf;
410 struct pipe *rpipe, *wpipe;
412 int fd, fflags, error;
414 pipe_paircreate(td, &pp);
415 rpipe = &pp->pp_rpipe;
416 wpipe = &pp->pp_wpipe;
417 error = falloc(td, &rf, &fd, flags);
423 /* An extra reference on `rf' has been held for us by falloc(). */
426 fflags = FREAD | FWRITE;
427 if ((flags & O_NONBLOCK) != 0)
431 * Warning: once we've gotten past allocation of the fd for the
432 * read-side, we can only drop the read side via fdrop() in order
433 * to avoid races against processes which manage to dup() the read
434 * side while we are blocked trying to allocate the write side.
436 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
437 error = falloc(td, &wf, &fd, flags);
439 fdclose(td, rf, fildes[0]);
441 /* rpipe has been closed by fdrop(). */
445 /* An extra reference on `wf' has been held for us by falloc(). */
446 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
456 sys_pipe(struct thread *td, struct pipe_args *uap)
461 error = kern_pipe(td, fildes);
465 td->td_retval[0] = fildes[0];
466 td->td_retval[1] = fildes[1];
472 sys_pipe2(struct thread *td, struct pipe2_args *uap)
474 int error, fildes[2];
476 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
478 error = kern_pipe2(td, fildes, uap->flags);
481 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
483 (void)kern_close(td, fildes[0]);
484 (void)kern_close(td, fildes[1]);
490 * Allocate kva for pipe circular buffer, the space is pageable
491 * This routine will 'realloc' the size of a pipe safely, if it fails
492 * it will retain the old buffer.
493 * If it fails it will return ENOMEM.
496 pipespace_new(cpipe, size)
501 int error, cnt, firstseg;
502 static int curfail = 0;
503 static struct timeval lastfail;
505 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
506 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
507 ("pipespace: resize of direct writes not allowed"));
509 cnt = cpipe->pipe_buffer.cnt;
513 size = round_page(size);
514 buffer = (caddr_t) vm_map_min(pipe_map);
516 error = vm_map_find(pipe_map, NULL, 0,
517 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
518 VM_PROT_ALL, VM_PROT_ALL, 0);
519 if (error != KERN_SUCCESS) {
520 if ((cpipe->pipe_buffer.buffer == NULL) &&
521 (size > SMALL_PIPE_SIZE)) {
522 size = SMALL_PIPE_SIZE;
526 if (cpipe->pipe_buffer.buffer == NULL) {
528 if (ppsratecheck(&lastfail, &curfail, 1))
529 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
536 /* copy data, then free old resources if we're resizing */
538 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
539 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
540 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
542 if ((cnt - firstseg) > 0)
543 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
544 cpipe->pipe_buffer.in);
546 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
550 pipe_free_kmem(cpipe);
551 cpipe->pipe_buffer.buffer = buffer;
552 cpipe->pipe_buffer.size = size;
553 cpipe->pipe_buffer.in = cnt;
554 cpipe->pipe_buffer.out = 0;
555 cpipe->pipe_buffer.cnt = cnt;
556 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
561 * Wrapper for pipespace_new() that performs locking assertions.
564 pipespace(cpipe, size)
569 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
570 ("Unlocked pipe passed to pipespace"));
571 return (pipespace_new(cpipe, size));
575 * lock a pipe for I/O, blocking other access
578 pipelock(cpipe, catch)
584 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
585 while (cpipe->pipe_state & PIPE_LOCKFL) {
586 cpipe->pipe_state |= PIPE_LWANT;
587 error = msleep(cpipe, PIPE_MTX(cpipe),
588 catch ? (PRIBIO | PCATCH) : PRIBIO,
593 cpipe->pipe_state |= PIPE_LOCKFL;
598 * unlock a pipe I/O lock
605 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
606 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
607 ("Unlocked pipe passed to pipeunlock"));
608 cpipe->pipe_state &= ~PIPE_LOCKFL;
609 if (cpipe->pipe_state & PIPE_LWANT) {
610 cpipe->pipe_state &= ~PIPE_LWANT;
620 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
621 if (cpipe->pipe_state & PIPE_SEL) {
622 selwakeuppri(&cpipe->pipe_sel, PSOCK);
623 if (!SEL_WAITING(&cpipe->pipe_sel))
624 cpipe->pipe_state &= ~PIPE_SEL;
626 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
627 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
628 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
632 * Initialize and allocate VM and memory for pipe. The structure
633 * will start out zero'd from the ctor, so we just manage the kmem.
636 pipe_create(pipe, backing)
643 * Note that these functions can fail if pipe map is exhausted
644 * (as a result of too many pipes created), but we ignore the
645 * error as it is not fatal and could be provoked by
646 * unprivileged users. The only consequence is worse performance
649 if (amountpipekva > maxpipekva / 2)
650 (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
652 (void)pipespace_new(pipe, PIPE_SIZE);
660 pipe_read(fp, uio, active_cred, flags, td)
663 struct ucred *active_cred;
675 error = pipelock(rpipe, 1);
680 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
684 if (amountpipekva > (3 * maxpipekva) / 4) {
685 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
686 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
687 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
688 (piperesizeallowed == 1)) {
690 pipespace(rpipe, SMALL_PIPE_SIZE);
695 while (uio->uio_resid) {
697 * normal pipe buffer receive
699 if (rpipe->pipe_buffer.cnt > 0) {
700 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
701 if (size > rpipe->pipe_buffer.cnt)
702 size = rpipe->pipe_buffer.cnt;
703 if (size > uio->uio_resid)
704 size = uio->uio_resid;
708 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
714 rpipe->pipe_buffer.out += size;
715 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
716 rpipe->pipe_buffer.out = 0;
718 rpipe->pipe_buffer.cnt -= size;
721 * If there is no more to read in the pipe, reset
722 * its pointers to the beginning. This improves
725 if (rpipe->pipe_buffer.cnt == 0) {
726 rpipe->pipe_buffer.in = 0;
727 rpipe->pipe_buffer.out = 0;
730 #ifndef PIPE_NODIRECT
732 * Direct copy, bypassing a kernel buffer.
734 } else if ((size = rpipe->pipe_map.cnt) &&
735 (rpipe->pipe_state & PIPE_DIRECTW)) {
736 if (size > uio->uio_resid)
737 size = (u_int) uio->uio_resid;
740 error = uiomove_fromphys(rpipe->pipe_map.ms,
741 rpipe->pipe_map.pos, size, uio);
746 rpipe->pipe_map.pos += size;
747 rpipe->pipe_map.cnt -= size;
748 if (rpipe->pipe_map.cnt == 0) {
749 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
755 * detect EOF condition
756 * read returns 0 on EOF, no need to set error
758 if (rpipe->pipe_state & PIPE_EOF)
762 * If the "write-side" has been blocked, wake it up now.
764 if (rpipe->pipe_state & PIPE_WANTW) {
765 rpipe->pipe_state &= ~PIPE_WANTW;
770 * Break if some data was read.
776 * Unlock the pipe buffer for our remaining processing.
777 * We will either break out with an error or we will
778 * sleep and relock to loop.
783 * Handle non-blocking mode operation or
784 * wait for more data.
786 if (fp->f_flag & FNONBLOCK) {
789 rpipe->pipe_state |= PIPE_WANTR;
790 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
793 error = pipelock(rpipe, 1);
804 /* XXX: should probably do this before getting any locks. */
806 vfs_timestamp(&rpipe->pipe_atime);
811 * PIPE_WANT processing only makes sense if pipe_busy is 0.
813 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
814 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
816 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
818 * Handle write blocking hysteresis.
820 if (rpipe->pipe_state & PIPE_WANTW) {
821 rpipe->pipe_state &= ~PIPE_WANTW;
826 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
827 pipeselwakeup(rpipe);
833 #ifndef PIPE_NODIRECT
835 * Map the sending processes' buffer into kernel space and wire it.
836 * This is similar to a physical write operation.
839 pipe_build_write_buffer(wpipe, uio)
846 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
847 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
848 ("Clone attempt on non-direct write pipe!"));
850 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
851 size = wpipe->pipe_buffer.size;
853 size = uio->uio_iov->iov_len;
855 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
856 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
857 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
861 * set up the control block
863 wpipe->pipe_map.npages = i;
864 wpipe->pipe_map.pos =
865 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
866 wpipe->pipe_map.cnt = size;
869 * and update the uio data
872 uio->uio_iov->iov_len -= size;
873 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
874 if (uio->uio_iov->iov_len == 0)
876 uio->uio_resid -= size;
877 uio->uio_offset += size;
882 * unmap and unwire the process buffer
885 pipe_destroy_write_buffer(wpipe)
889 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
890 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
891 wpipe->pipe_map.npages = 0;
895 * In the case of a signal, the writing process might go away. This
896 * code copies the data into the circular buffer so that the source
897 * pages can be freed without loss of data.
900 pipe_clone_write_buffer(wpipe)
908 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
909 size = wpipe->pipe_map.cnt;
910 pos = wpipe->pipe_map.pos;
912 wpipe->pipe_buffer.in = size;
913 wpipe->pipe_buffer.out = 0;
914 wpipe->pipe_buffer.cnt = size;
915 wpipe->pipe_state &= ~PIPE_DIRECTW;
918 iov.iov_base = wpipe->pipe_buffer.buffer;
923 uio.uio_resid = size;
924 uio.uio_segflg = UIO_SYSSPACE;
925 uio.uio_rw = UIO_READ;
926 uio.uio_td = curthread;
927 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
929 pipe_destroy_write_buffer(wpipe);
933 * This implements the pipe buffer write mechanism. Note that only
934 * a direct write OR a normal pipe write can be pending at any given time.
935 * If there are any characters in the pipe buffer, the direct write will
936 * be deferred until the receiving process grabs all of the bytes from
937 * the pipe buffer. Then the direct mapping write is set-up.
940 pipe_direct_write(wpipe, uio)
947 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
948 error = pipelock(wpipe, 1);
949 if (wpipe->pipe_state & PIPE_EOF)
955 while (wpipe->pipe_state & PIPE_DIRECTW) {
956 if (wpipe->pipe_state & PIPE_WANTR) {
957 wpipe->pipe_state &= ~PIPE_WANTR;
960 pipeselwakeup(wpipe);
961 wpipe->pipe_state |= PIPE_WANTW;
963 error = msleep(wpipe, PIPE_MTX(wpipe),
964 PRIBIO | PCATCH, "pipdww", 0);
970 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
971 if (wpipe->pipe_buffer.cnt > 0) {
972 if (wpipe->pipe_state & PIPE_WANTR) {
973 wpipe->pipe_state &= ~PIPE_WANTR;
976 pipeselwakeup(wpipe);
977 wpipe->pipe_state |= PIPE_WANTW;
979 error = msleep(wpipe, PIPE_MTX(wpipe),
980 PRIBIO | PCATCH, "pipdwc", 0);
987 wpipe->pipe_state |= PIPE_DIRECTW;
990 error = pipe_build_write_buffer(wpipe, uio);
993 wpipe->pipe_state &= ~PIPE_DIRECTW;
999 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1000 if (wpipe->pipe_state & PIPE_EOF) {
1001 pipe_destroy_write_buffer(wpipe);
1002 pipeselwakeup(wpipe);
1007 if (wpipe->pipe_state & PIPE_WANTR) {
1008 wpipe->pipe_state &= ~PIPE_WANTR;
1011 pipeselwakeup(wpipe);
1012 wpipe->pipe_state |= PIPE_WANTW;
1014 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1019 if (wpipe->pipe_state & PIPE_EOF)
1021 if (wpipe->pipe_state & PIPE_DIRECTW) {
1023 * this bit of trickery substitutes a kernel buffer for
1024 * the process that might be going away.
1026 pipe_clone_write_buffer(wpipe);
1028 pipe_destroy_write_buffer(wpipe);
1040 pipe_write(fp, uio, active_cred, flags, td)
1043 struct ucred *active_cred;
1050 struct pipe *wpipe, *rpipe;
1053 wpipe = PIPE_PEER(rpipe);
1055 error = pipelock(wpipe, 1);
1061 * detect loss of pipe read side, issue SIGPIPE if lost.
1063 if (wpipe->pipe_present != PIPE_ACTIVE ||
1064 (wpipe->pipe_state & PIPE_EOF)) {
1070 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1079 /* Choose a larger size if it's advantageous */
1080 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1081 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1082 if (piperesizeallowed != 1)
1084 if (amountpipekva > maxpipekva / 2)
1086 if (desiredsize == BIG_PIPE_SIZE)
1088 desiredsize = desiredsize * 2;
1091 /* Choose a smaller size if we're in a OOM situation */
1092 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1093 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1094 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1095 (piperesizeallowed == 1))
1096 desiredsize = SMALL_PIPE_SIZE;
1098 /* Resize if the above determined that a new size was necessary */
1099 if ((desiredsize != wpipe->pipe_buffer.size) &&
1100 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1102 pipespace(wpipe, desiredsize);
1105 if (wpipe->pipe_buffer.size == 0) {
1107 * This can only happen for reverse direction use of pipes
1108 * in a complete OOM situation.
1119 orig_resid = uio->uio_resid;
1121 while (uio->uio_resid) {
1125 if (wpipe->pipe_state & PIPE_EOF) {
1130 #ifndef PIPE_NODIRECT
1132 * If the transfer is large, we can gain performance if
1133 * we do process-to-process copies directly.
1134 * If the write is non-blocking, we don't use the
1135 * direct write mechanism.
1137 * The direct write mechanism will detect the reader going
1140 if (uio->uio_segflg == UIO_USERSPACE &&
1141 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1142 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1143 (fp->f_flag & FNONBLOCK) == 0) {
1145 error = pipe_direct_write(wpipe, uio);
1153 * Pipe buffered writes cannot be coincidental with
1154 * direct writes. We wait until the currently executing
1155 * direct write is completed before we start filling the
1156 * pipe buffer. We break out if a signal occurs or the
1159 if (wpipe->pipe_state & PIPE_DIRECTW) {
1160 if (wpipe->pipe_state & PIPE_WANTR) {
1161 wpipe->pipe_state &= ~PIPE_WANTR;
1164 pipeselwakeup(wpipe);
1165 wpipe->pipe_state |= PIPE_WANTW;
1167 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1175 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1177 /* Writes of size <= PIPE_BUF must be atomic. */
1178 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1182 int size; /* Transfer size */
1183 int segsize; /* first segment to transfer */
1186 * Transfer size is minimum of uio transfer
1187 * and free space in pipe buffer.
1189 if (space > uio->uio_resid)
1190 size = uio->uio_resid;
1194 * First segment to transfer is minimum of
1195 * transfer size and contiguous space in
1196 * pipe buffer. If first segment to transfer
1197 * is less than the transfer size, we've got
1198 * a wraparound in the buffer.
1200 segsize = wpipe->pipe_buffer.size -
1201 wpipe->pipe_buffer.in;
1205 /* Transfer first segment */
1208 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1212 if (error == 0 && segsize < size) {
1213 KASSERT(wpipe->pipe_buffer.in + segsize ==
1214 wpipe->pipe_buffer.size,
1215 ("Pipe buffer wraparound disappeared"));
1217 * Transfer remaining part now, to
1218 * support atomic writes. Wraparound
1224 &wpipe->pipe_buffer.buffer[0],
1225 size - segsize, uio);
1229 wpipe->pipe_buffer.in += size;
1230 if (wpipe->pipe_buffer.in >=
1231 wpipe->pipe_buffer.size) {
1232 KASSERT(wpipe->pipe_buffer.in ==
1234 wpipe->pipe_buffer.size,
1235 ("Expected wraparound bad"));
1236 wpipe->pipe_buffer.in = size - segsize;
1239 wpipe->pipe_buffer.cnt += size;
1240 KASSERT(wpipe->pipe_buffer.cnt <=
1241 wpipe->pipe_buffer.size,
1242 ("Pipe buffer overflow"));
1249 * If the "read-side" has been blocked, wake it up now.
1251 if (wpipe->pipe_state & PIPE_WANTR) {
1252 wpipe->pipe_state &= ~PIPE_WANTR;
1257 * don't block on non-blocking I/O
1259 if (fp->f_flag & FNONBLOCK) {
1266 * We have no more space and have something to offer,
1267 * wake up select/poll.
1269 pipeselwakeup(wpipe);
1271 wpipe->pipe_state |= PIPE_WANTW;
1273 error = msleep(wpipe, PIPE_MTX(rpipe),
1274 PRIBIO | PCATCH, "pipewr", 0);
1283 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1284 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1286 } else if (wpipe->pipe_buffer.cnt > 0) {
1288 * If we have put any characters in the buffer, we wake up
1291 if (wpipe->pipe_state & PIPE_WANTR) {
1292 wpipe->pipe_state &= ~PIPE_WANTR;
1298 * Don't return EPIPE if any byte was written.
1299 * EINTR and other interrupts are handled by generic I/O layer.
1300 * Do not pretend that I/O succeeded for obvious user error
1303 if (uio->uio_resid != orig_resid && error == EPIPE)
1307 vfs_timestamp(&wpipe->pipe_mtime);
1310 * We have something to offer,
1311 * wake up select/poll.
1313 if (wpipe->pipe_buffer.cnt)
1314 pipeselwakeup(wpipe);
1323 pipe_truncate(fp, length, active_cred, td)
1326 struct ucred *active_cred;
1333 if (cpipe->pipe_state & PIPE_NAMED)
1334 error = vnops.fo_truncate(fp, length, active_cred, td);
1336 error = invfo_truncate(fp, length, active_cred, td);
1341 * we implement a very minimal set of ioctls for compatibility with sockets.
1344 pipe_ioctl(fp, cmd, data, active_cred, td)
1348 struct ucred *active_cred;
1351 struct pipe *mpipe = fp->f_data;
1357 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1372 mpipe->pipe_state |= PIPE_ASYNC;
1374 mpipe->pipe_state &= ~PIPE_ASYNC;
1379 if (!(fp->f_flag & FREAD)) {
1384 if (mpipe->pipe_state & PIPE_DIRECTW)
1385 *(int *)data = mpipe->pipe_map.cnt;
1387 *(int *)data = mpipe->pipe_buffer.cnt;
1392 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1396 *(int *)data = fgetown(&mpipe->pipe_sigio);
1399 /* This is deprecated, FIOSETOWN should be used instead. */
1402 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1405 /* This is deprecated, FIOGETOWN should be used instead. */
1407 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1420 pipe_poll(fp, events, active_cred, td)
1423 struct ucred *active_cred;
1428 int levents, revents;
1435 wpipe = PIPE_PEER(rpipe);
1438 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1442 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1443 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1444 (rpipe->pipe_buffer.cnt > 0))
1445 revents |= events & (POLLIN | POLLRDNORM);
1447 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1448 if (wpipe->pipe_present != PIPE_ACTIVE ||
1449 (wpipe->pipe_state & PIPE_EOF) ||
1450 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1451 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1452 wpipe->pipe_buffer.size == 0)))
1453 revents |= events & (POLLOUT | POLLWRNORM);
1456 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1457 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1458 fp->f_seqcount == rpipe->pipe_wgen)
1459 events |= POLLINIGNEOF;
1461 if ((events & POLLINIGNEOF) == 0) {
1462 if (rpipe->pipe_state & PIPE_EOF) {
1463 revents |= (events & (POLLIN | POLLRDNORM));
1464 if (wpipe->pipe_present != PIPE_ACTIVE ||
1465 (wpipe->pipe_state & PIPE_EOF))
1471 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1472 selrecord(td, &rpipe->pipe_sel);
1473 if (SEL_WAITING(&rpipe->pipe_sel))
1474 rpipe->pipe_state |= PIPE_SEL;
1477 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1478 selrecord(td, &wpipe->pipe_sel);
1479 if (SEL_WAITING(&wpipe->pipe_sel))
1480 wpipe->pipe_state |= PIPE_SEL;
1492 * We shouldn't need locks here as we're doing a read and this should
1493 * be a natural race.
1496 pipe_stat(fp, ub, active_cred, td)
1499 struct ucred *active_cred;
1511 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1518 /* For named pipes ask the underlying filesystem. */
1519 if (pipe->pipe_state & PIPE_NAMED) {
1521 return (vnops.fo_stat(fp, ub, active_cred, td));
1525 * Lazily allocate an inode number for the pipe. Most pipe
1526 * users do not call fstat(2) on the pipe, which means that
1527 * postponing the inode allocation until it is must be
1528 * returned to userland is useful. If alloc_unr failed,
1529 * assign st_ino zero instead of returning an error.
1530 * Special pipe_ino values:
1531 * -1 - not yet initialized;
1532 * 0 - alloc_unr failed, return 0 as st_ino forever.
1534 if (pipe->pipe_ino == (ino_t)-1) {
1535 new_unr = alloc_unr(pipeino_unr);
1537 pipe->pipe_ino = new_unr;
1543 bzero(ub, sizeof(*ub));
1544 ub->st_mode = S_IFIFO;
1545 ub->st_blksize = PAGE_SIZE;
1546 if (pipe->pipe_state & PIPE_DIRECTW)
1547 ub->st_size = pipe->pipe_map.cnt;
1549 ub->st_size = pipe->pipe_buffer.cnt;
1550 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1551 ub->st_atim = pipe->pipe_atime;
1552 ub->st_mtim = pipe->pipe_mtime;
1553 ub->st_ctim = pipe->pipe_ctime;
1554 ub->st_uid = fp->f_cred->cr_uid;
1555 ub->st_gid = fp->f_cred->cr_gid;
1556 ub->st_dev = pipedev_ino;
1557 ub->st_ino = pipe->pipe_ino;
1559 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1571 if (fp->f_vnode != NULL)
1572 return vnops.fo_close(fp, td);
1573 fp->f_ops = &badfileops;
1574 pipe_dtor(fp->f_data);
1580 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1586 if (cpipe->pipe_state & PIPE_NAMED)
1587 error = vn_chmod(fp, mode, active_cred, td);
1589 error = invfo_chmod(fp, mode, active_cred, td);
1594 pipe_chown(fp, uid, gid, active_cred, td)
1598 struct ucred *active_cred;
1605 if (cpipe->pipe_state & PIPE_NAMED)
1606 error = vn_chown(fp, uid, gid, active_cred, td);
1608 error = invfo_chown(fp, uid, gid, active_cred, td);
1613 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1617 if (fp->f_type == DTYPE_FIFO)
1618 return (vn_fill_kinfo(fp, kif, fdp));
1619 kif->kf_type = KF_TYPE_PIPE;
1621 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1622 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1623 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1628 pipe_free_kmem(cpipe)
1632 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1633 ("pipe_free_kmem: pipe mutex locked"));
1635 if (cpipe->pipe_buffer.buffer != NULL) {
1636 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1637 vm_map_remove(pipe_map,
1638 (vm_offset_t)cpipe->pipe_buffer.buffer,
1639 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1640 cpipe->pipe_buffer.buffer = NULL;
1642 #ifndef PIPE_NODIRECT
1644 cpipe->pipe_map.cnt = 0;
1645 cpipe->pipe_map.pos = 0;
1646 cpipe->pipe_map.npages = 0;
1658 struct pipepair *pp;
1661 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1665 pp = cpipe->pipe_pair;
1667 pipeselwakeup(cpipe);
1670 * If the other side is blocked, wake it up saying that
1671 * we want to close it down.
1673 cpipe->pipe_state |= PIPE_EOF;
1674 while (cpipe->pipe_busy) {
1676 cpipe->pipe_state |= PIPE_WANT;
1678 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1684 * Disconnect from peer, if any.
1686 ppipe = cpipe->pipe_peer;
1687 if (ppipe->pipe_present == PIPE_ACTIVE) {
1688 pipeselwakeup(ppipe);
1690 ppipe->pipe_state |= PIPE_EOF;
1692 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1696 * Mark this endpoint as free. Release kmem resources. We
1697 * don't mark this endpoint as unused until we've finished
1698 * doing that, or the pipe might disappear out from under
1702 pipe_free_kmem(cpipe);
1704 cpipe->pipe_present = PIPE_CLOSING;
1708 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1709 * PIPE_FINALIZED, that allows other end to free the
1710 * pipe_pair, only after the knotes are completely dismantled.
1712 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1713 cpipe->pipe_present = PIPE_FINALIZED;
1714 seldrain(&cpipe->pipe_sel);
1715 knlist_destroy(&cpipe->pipe_sel.si_note);
1718 * If both endpoints are now closed, release the memory for the
1719 * pipe pair. If not, unlock.
1721 if (ppipe->pipe_present == PIPE_FINALIZED) {
1724 mac_pipe_destroy(pp);
1726 uma_zfree(pipe_zone, cpipe->pipe_pair);
1733 pipe_kqfilter(struct file *fp, struct knote *kn)
1738 * If a filter is requested that is not supported by this file
1739 * descriptor, don't return an error, but also don't ever generate an
1742 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1743 kn->kn_fop = &pipe_nfiltops;
1746 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1747 kn->kn_fop = &pipe_nfiltops;
1752 switch (kn->kn_filter) {
1754 kn->kn_fop = &pipe_rfiltops;
1757 kn->kn_fop = &pipe_wfiltops;
1758 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1759 /* other end of pipe has been closed */
1763 cpipe = PIPE_PEER(cpipe);
1770 kn->kn_hook = cpipe;
1771 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1777 filt_pipedetach(struct knote *kn)
1779 struct pipe *cpipe = kn->kn_hook;
1782 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1788 filt_piperead(struct knote *kn, long hint)
1790 struct pipe *rpipe = kn->kn_hook;
1791 struct pipe *wpipe = rpipe->pipe_peer;
1794 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1795 kn->kn_data = rpipe->pipe_buffer.cnt;
1796 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1797 kn->kn_data = rpipe->pipe_map.cnt;
1799 if ((rpipe->pipe_state & PIPE_EOF) ||
1800 wpipe->pipe_present != PIPE_ACTIVE ||
1801 (wpipe->pipe_state & PIPE_EOF)) {
1802 kn->kn_flags |= EV_EOF;
1805 ret = kn->kn_data > 0;
1811 filt_pipewrite(struct knote *kn, long hint)
1815 wpipe = kn->kn_hook;
1816 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1817 if (wpipe->pipe_present != PIPE_ACTIVE ||
1818 (wpipe->pipe_state & PIPE_EOF)) {
1820 kn->kn_flags |= EV_EOF;
1823 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1824 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1825 if (wpipe->pipe_state & PIPE_DIRECTW)
1828 return (kn->kn_data >= PIPE_BUF);
1832 filt_pipedetach_notsup(struct knote *kn)
1838 filt_pipenotsup(struct knote *kn, long hint)