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 filedesc *fdp;
410 struct file *rf, *wf;
411 struct pipe *rpipe, *wpipe;
413 int fd, fflags, error;
415 fdp = td->td_proc->p_fd;
416 pipe_paircreate(td, &pp);
417 rpipe = &pp->pp_rpipe;
418 wpipe = &pp->pp_wpipe;
419 error = falloc(td, &rf, &fd, flags);
425 /* An extra reference on `rf' has been held for us by falloc(). */
428 fflags = FREAD | FWRITE;
429 if ((flags & O_NONBLOCK) != 0)
433 * Warning: once we've gotten past allocation of the fd for the
434 * read-side, we can only drop the read side via fdrop() in order
435 * to avoid races against processes which manage to dup() the read
436 * side while we are blocked trying to allocate the write side.
438 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
439 error = falloc(td, &wf, &fd, flags);
441 fdclose(fdp, rf, fildes[0], td);
443 /* rpipe has been closed by fdrop(). */
447 /* An extra reference on `wf' has been held for us by falloc(). */
448 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
458 sys_pipe(struct thread *td, struct pipe_args *uap)
463 error = kern_pipe(td, fildes);
467 td->td_retval[0] = fildes[0];
468 td->td_retval[1] = fildes[1];
474 sys_pipe2(struct thread *td, struct pipe2_args *uap)
476 int error, fildes[2];
478 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
480 error = kern_pipe2(td, fildes, uap->flags);
483 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
485 (void)kern_close(td, fildes[0]);
486 (void)kern_close(td, fildes[1]);
492 * Allocate kva for pipe circular buffer, the space is pageable
493 * This routine will 'realloc' the size of a pipe safely, if it fails
494 * it will retain the old buffer.
495 * If it fails it will return ENOMEM.
498 pipespace_new(cpipe, size)
503 int error, cnt, firstseg;
504 static int curfail = 0;
505 static struct timeval lastfail;
507 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
508 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
509 ("pipespace: resize of direct writes not allowed"));
511 cnt = cpipe->pipe_buffer.cnt;
515 size = round_page(size);
516 buffer = (caddr_t) vm_map_min(pipe_map);
518 error = vm_map_find(pipe_map, NULL, 0,
519 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
520 VM_PROT_ALL, VM_PROT_ALL, 0);
521 if (error != KERN_SUCCESS) {
522 if ((cpipe->pipe_buffer.buffer == NULL) &&
523 (size > SMALL_PIPE_SIZE)) {
524 size = SMALL_PIPE_SIZE;
528 if (cpipe->pipe_buffer.buffer == NULL) {
530 if (ppsratecheck(&lastfail, &curfail, 1))
531 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
538 /* copy data, then free old resources if we're resizing */
540 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
541 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
542 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
544 if ((cnt - firstseg) > 0)
545 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
546 cpipe->pipe_buffer.in);
548 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
552 pipe_free_kmem(cpipe);
553 cpipe->pipe_buffer.buffer = buffer;
554 cpipe->pipe_buffer.size = size;
555 cpipe->pipe_buffer.in = cnt;
556 cpipe->pipe_buffer.out = 0;
557 cpipe->pipe_buffer.cnt = cnt;
558 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
563 * Wrapper for pipespace_new() that performs locking assertions.
566 pipespace(cpipe, size)
571 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
572 ("Unlocked pipe passed to pipespace"));
573 return (pipespace_new(cpipe, size));
577 * lock a pipe for I/O, blocking other access
580 pipelock(cpipe, catch)
586 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
587 while (cpipe->pipe_state & PIPE_LOCKFL) {
588 cpipe->pipe_state |= PIPE_LWANT;
589 error = msleep(cpipe, PIPE_MTX(cpipe),
590 catch ? (PRIBIO | PCATCH) : PRIBIO,
595 cpipe->pipe_state |= PIPE_LOCKFL;
600 * unlock a pipe I/O lock
607 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
608 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
609 ("Unlocked pipe passed to pipeunlock"));
610 cpipe->pipe_state &= ~PIPE_LOCKFL;
611 if (cpipe->pipe_state & PIPE_LWANT) {
612 cpipe->pipe_state &= ~PIPE_LWANT;
622 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
623 if (cpipe->pipe_state & PIPE_SEL) {
624 selwakeuppri(&cpipe->pipe_sel, PSOCK);
625 if (!SEL_WAITING(&cpipe->pipe_sel))
626 cpipe->pipe_state &= ~PIPE_SEL;
628 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
629 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
630 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
634 * Initialize and allocate VM and memory for pipe. The structure
635 * will start out zero'd from the ctor, so we just manage the kmem.
638 pipe_create(pipe, backing)
645 * Note that these functions can fail if pipe map is exhausted
646 * (as a result of too many pipes created), but we ignore the
647 * error as it is not fatal and could be provoked by
648 * unprivileged users. The only consequence is worse performance
651 if (amountpipekva > maxpipekva / 2)
652 (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
654 (void)pipespace_new(pipe, PIPE_SIZE);
662 pipe_read(fp, uio, active_cred, flags, td)
665 struct ucred *active_cred;
677 error = pipelock(rpipe, 1);
682 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
686 if (amountpipekva > (3 * maxpipekva) / 4) {
687 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
688 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
689 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
690 (piperesizeallowed == 1)) {
692 pipespace(rpipe, SMALL_PIPE_SIZE);
697 while (uio->uio_resid) {
699 * normal pipe buffer receive
701 if (rpipe->pipe_buffer.cnt > 0) {
702 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
703 if (size > rpipe->pipe_buffer.cnt)
704 size = rpipe->pipe_buffer.cnt;
705 if (size > uio->uio_resid)
706 size = uio->uio_resid;
710 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
716 rpipe->pipe_buffer.out += size;
717 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
718 rpipe->pipe_buffer.out = 0;
720 rpipe->pipe_buffer.cnt -= size;
723 * If there is no more to read in the pipe, reset
724 * its pointers to the beginning. This improves
727 if (rpipe->pipe_buffer.cnt == 0) {
728 rpipe->pipe_buffer.in = 0;
729 rpipe->pipe_buffer.out = 0;
732 #ifndef PIPE_NODIRECT
734 * Direct copy, bypassing a kernel buffer.
736 } else if ((size = rpipe->pipe_map.cnt) &&
737 (rpipe->pipe_state & PIPE_DIRECTW)) {
738 if (size > uio->uio_resid)
739 size = (u_int) uio->uio_resid;
742 error = uiomove_fromphys(rpipe->pipe_map.ms,
743 rpipe->pipe_map.pos, size, uio);
748 rpipe->pipe_map.pos += size;
749 rpipe->pipe_map.cnt -= size;
750 if (rpipe->pipe_map.cnt == 0) {
751 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
757 * detect EOF condition
758 * read returns 0 on EOF, no need to set error
760 if (rpipe->pipe_state & PIPE_EOF)
764 * If the "write-side" has been blocked, wake it up now.
766 if (rpipe->pipe_state & PIPE_WANTW) {
767 rpipe->pipe_state &= ~PIPE_WANTW;
772 * Break if some data was read.
778 * Unlock the pipe buffer for our remaining processing.
779 * We will either break out with an error or we will
780 * sleep and relock to loop.
785 * Handle non-blocking mode operation or
786 * wait for more data.
788 if (fp->f_flag & FNONBLOCK) {
791 rpipe->pipe_state |= PIPE_WANTR;
792 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
795 error = pipelock(rpipe, 1);
806 /* XXX: should probably do this before getting any locks. */
808 vfs_timestamp(&rpipe->pipe_atime);
813 * PIPE_WANT processing only makes sense if pipe_busy is 0.
815 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
816 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
818 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
820 * Handle write blocking hysteresis.
822 if (rpipe->pipe_state & PIPE_WANTW) {
823 rpipe->pipe_state &= ~PIPE_WANTW;
828 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
829 pipeselwakeup(rpipe);
835 #ifndef PIPE_NODIRECT
837 * Map the sending processes' buffer into kernel space and wire it.
838 * This is similar to a physical write operation.
841 pipe_build_write_buffer(wpipe, uio)
848 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
849 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
850 ("Clone attempt on non-direct write pipe!"));
852 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
853 size = wpipe->pipe_buffer.size;
855 size = uio->uio_iov->iov_len;
857 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
858 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
859 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
863 * set up the control block
865 wpipe->pipe_map.npages = i;
866 wpipe->pipe_map.pos =
867 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
868 wpipe->pipe_map.cnt = size;
871 * and update the uio data
874 uio->uio_iov->iov_len -= size;
875 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
876 if (uio->uio_iov->iov_len == 0)
878 uio->uio_resid -= size;
879 uio->uio_offset += size;
884 * unmap and unwire the process buffer
887 pipe_destroy_write_buffer(wpipe)
891 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
892 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
893 wpipe->pipe_map.npages = 0;
897 * In the case of a signal, the writing process might go away. This
898 * code copies the data into the circular buffer so that the source
899 * pages can be freed without loss of data.
902 pipe_clone_write_buffer(wpipe)
910 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
911 size = wpipe->pipe_map.cnt;
912 pos = wpipe->pipe_map.pos;
914 wpipe->pipe_buffer.in = size;
915 wpipe->pipe_buffer.out = 0;
916 wpipe->pipe_buffer.cnt = size;
917 wpipe->pipe_state &= ~PIPE_DIRECTW;
920 iov.iov_base = wpipe->pipe_buffer.buffer;
925 uio.uio_resid = size;
926 uio.uio_segflg = UIO_SYSSPACE;
927 uio.uio_rw = UIO_READ;
928 uio.uio_td = curthread;
929 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
931 pipe_destroy_write_buffer(wpipe);
935 * This implements the pipe buffer write mechanism. Note that only
936 * a direct write OR a normal pipe write can be pending at any given time.
937 * If there are any characters in the pipe buffer, the direct write will
938 * be deferred until the receiving process grabs all of the bytes from
939 * the pipe buffer. Then the direct mapping write is set-up.
942 pipe_direct_write(wpipe, uio)
949 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
950 error = pipelock(wpipe, 1);
951 if (wpipe->pipe_state & PIPE_EOF)
957 while (wpipe->pipe_state & PIPE_DIRECTW) {
958 if (wpipe->pipe_state & PIPE_WANTR) {
959 wpipe->pipe_state &= ~PIPE_WANTR;
962 pipeselwakeup(wpipe);
963 wpipe->pipe_state |= PIPE_WANTW;
965 error = msleep(wpipe, PIPE_MTX(wpipe),
966 PRIBIO | PCATCH, "pipdww", 0);
972 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
973 if (wpipe->pipe_buffer.cnt > 0) {
974 if (wpipe->pipe_state & PIPE_WANTR) {
975 wpipe->pipe_state &= ~PIPE_WANTR;
978 pipeselwakeup(wpipe);
979 wpipe->pipe_state |= PIPE_WANTW;
981 error = msleep(wpipe, PIPE_MTX(wpipe),
982 PRIBIO | PCATCH, "pipdwc", 0);
989 wpipe->pipe_state |= PIPE_DIRECTW;
992 error = pipe_build_write_buffer(wpipe, uio);
995 wpipe->pipe_state &= ~PIPE_DIRECTW;
1001 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1002 if (wpipe->pipe_state & PIPE_EOF) {
1003 pipe_destroy_write_buffer(wpipe);
1004 pipeselwakeup(wpipe);
1009 if (wpipe->pipe_state & PIPE_WANTR) {
1010 wpipe->pipe_state &= ~PIPE_WANTR;
1013 pipeselwakeup(wpipe);
1014 wpipe->pipe_state |= PIPE_WANTW;
1016 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1021 if (wpipe->pipe_state & PIPE_EOF)
1023 if (wpipe->pipe_state & PIPE_DIRECTW) {
1025 * this bit of trickery substitutes a kernel buffer for
1026 * the process that might be going away.
1028 pipe_clone_write_buffer(wpipe);
1030 pipe_destroy_write_buffer(wpipe);
1042 pipe_write(fp, uio, active_cred, flags, td)
1045 struct ucred *active_cred;
1052 struct pipe *wpipe, *rpipe;
1055 wpipe = PIPE_PEER(rpipe);
1057 error = pipelock(wpipe, 1);
1063 * detect loss of pipe read side, issue SIGPIPE if lost.
1065 if (wpipe->pipe_present != PIPE_ACTIVE ||
1066 (wpipe->pipe_state & PIPE_EOF)) {
1072 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1081 /* Choose a larger size if it's advantageous */
1082 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1083 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1084 if (piperesizeallowed != 1)
1086 if (amountpipekva > maxpipekva / 2)
1088 if (desiredsize == BIG_PIPE_SIZE)
1090 desiredsize = desiredsize * 2;
1093 /* Choose a smaller size if we're in a OOM situation */
1094 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1095 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1096 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1097 (piperesizeallowed == 1))
1098 desiredsize = SMALL_PIPE_SIZE;
1100 /* Resize if the above determined that a new size was necessary */
1101 if ((desiredsize != wpipe->pipe_buffer.size) &&
1102 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1104 pipespace(wpipe, desiredsize);
1107 if (wpipe->pipe_buffer.size == 0) {
1109 * This can only happen for reverse direction use of pipes
1110 * in a complete OOM situation.
1121 orig_resid = uio->uio_resid;
1123 while (uio->uio_resid) {
1127 if (wpipe->pipe_state & PIPE_EOF) {
1132 #ifndef PIPE_NODIRECT
1134 * If the transfer is large, we can gain performance if
1135 * we do process-to-process copies directly.
1136 * If the write is non-blocking, we don't use the
1137 * direct write mechanism.
1139 * The direct write mechanism will detect the reader going
1142 if (uio->uio_segflg == UIO_USERSPACE &&
1143 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1144 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1145 (fp->f_flag & FNONBLOCK) == 0) {
1147 error = pipe_direct_write(wpipe, uio);
1155 * Pipe buffered writes cannot be coincidental with
1156 * direct writes. We wait until the currently executing
1157 * direct write is completed before we start filling the
1158 * pipe buffer. We break out if a signal occurs or the
1161 if (wpipe->pipe_state & PIPE_DIRECTW) {
1162 if (wpipe->pipe_state & PIPE_WANTR) {
1163 wpipe->pipe_state &= ~PIPE_WANTR;
1166 pipeselwakeup(wpipe);
1167 wpipe->pipe_state |= PIPE_WANTW;
1169 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1177 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1179 /* Writes of size <= PIPE_BUF must be atomic. */
1180 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1184 int size; /* Transfer size */
1185 int segsize; /* first segment to transfer */
1188 * Transfer size is minimum of uio transfer
1189 * and free space in pipe buffer.
1191 if (space > uio->uio_resid)
1192 size = uio->uio_resid;
1196 * First segment to transfer is minimum of
1197 * transfer size and contiguous space in
1198 * pipe buffer. If first segment to transfer
1199 * is less than the transfer size, we've got
1200 * a wraparound in the buffer.
1202 segsize = wpipe->pipe_buffer.size -
1203 wpipe->pipe_buffer.in;
1207 /* Transfer first segment */
1210 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1214 if (error == 0 && segsize < size) {
1215 KASSERT(wpipe->pipe_buffer.in + segsize ==
1216 wpipe->pipe_buffer.size,
1217 ("Pipe buffer wraparound disappeared"));
1219 * Transfer remaining part now, to
1220 * support atomic writes. Wraparound
1226 &wpipe->pipe_buffer.buffer[0],
1227 size - segsize, uio);
1231 wpipe->pipe_buffer.in += size;
1232 if (wpipe->pipe_buffer.in >=
1233 wpipe->pipe_buffer.size) {
1234 KASSERT(wpipe->pipe_buffer.in ==
1236 wpipe->pipe_buffer.size,
1237 ("Expected wraparound bad"));
1238 wpipe->pipe_buffer.in = size - segsize;
1241 wpipe->pipe_buffer.cnt += size;
1242 KASSERT(wpipe->pipe_buffer.cnt <=
1243 wpipe->pipe_buffer.size,
1244 ("Pipe buffer overflow"));
1251 * If the "read-side" has been blocked, wake it up now.
1253 if (wpipe->pipe_state & PIPE_WANTR) {
1254 wpipe->pipe_state &= ~PIPE_WANTR;
1259 * don't block on non-blocking I/O
1261 if (fp->f_flag & FNONBLOCK) {
1268 * We have no more space and have something to offer,
1269 * wake up select/poll.
1271 pipeselwakeup(wpipe);
1273 wpipe->pipe_state |= PIPE_WANTW;
1275 error = msleep(wpipe, PIPE_MTX(rpipe),
1276 PRIBIO | PCATCH, "pipewr", 0);
1285 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1286 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1288 } else if (wpipe->pipe_buffer.cnt > 0) {
1290 * If we have put any characters in the buffer, we wake up
1293 if (wpipe->pipe_state & PIPE_WANTR) {
1294 wpipe->pipe_state &= ~PIPE_WANTR;
1300 * Don't return EPIPE if any byte was written.
1301 * EINTR and other interrupts are handled by generic I/O layer.
1302 * Do not pretend that I/O succeeded for obvious user error
1305 if (uio->uio_resid != orig_resid && error == EPIPE)
1309 vfs_timestamp(&wpipe->pipe_mtime);
1312 * We have something to offer,
1313 * wake up select/poll.
1315 if (wpipe->pipe_buffer.cnt)
1316 pipeselwakeup(wpipe);
1325 pipe_truncate(fp, length, active_cred, td)
1328 struct ucred *active_cred;
1335 if (cpipe->pipe_state & PIPE_NAMED)
1336 error = vnops.fo_truncate(fp, length, active_cred, td);
1338 error = invfo_truncate(fp, length, active_cred, td);
1343 * we implement a very minimal set of ioctls for compatibility with sockets.
1346 pipe_ioctl(fp, cmd, data, active_cred, td)
1350 struct ucred *active_cred;
1353 struct pipe *mpipe = fp->f_data;
1359 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1374 mpipe->pipe_state |= PIPE_ASYNC;
1376 mpipe->pipe_state &= ~PIPE_ASYNC;
1381 if (!(fp->f_flag & FREAD)) {
1386 if (mpipe->pipe_state & PIPE_DIRECTW)
1387 *(int *)data = mpipe->pipe_map.cnt;
1389 *(int *)data = mpipe->pipe_buffer.cnt;
1394 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1398 *(int *)data = fgetown(&mpipe->pipe_sigio);
1401 /* This is deprecated, FIOSETOWN should be used instead. */
1404 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1407 /* This is deprecated, FIOGETOWN should be used instead. */
1409 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1422 pipe_poll(fp, events, active_cred, td)
1425 struct ucred *active_cred;
1430 int levents, revents;
1437 wpipe = PIPE_PEER(rpipe);
1440 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1444 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1445 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1446 (rpipe->pipe_buffer.cnt > 0))
1447 revents |= events & (POLLIN | POLLRDNORM);
1449 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1450 if (wpipe->pipe_present != PIPE_ACTIVE ||
1451 (wpipe->pipe_state & PIPE_EOF) ||
1452 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1453 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1454 wpipe->pipe_buffer.size == 0)))
1455 revents |= events & (POLLOUT | POLLWRNORM);
1458 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1459 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1460 fp->f_seqcount == rpipe->pipe_wgen)
1461 events |= POLLINIGNEOF;
1463 if ((events & POLLINIGNEOF) == 0) {
1464 if (rpipe->pipe_state & PIPE_EOF) {
1465 revents |= (events & (POLLIN | POLLRDNORM));
1466 if (wpipe->pipe_present != PIPE_ACTIVE ||
1467 (wpipe->pipe_state & PIPE_EOF))
1473 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1474 selrecord(td, &rpipe->pipe_sel);
1475 if (SEL_WAITING(&rpipe->pipe_sel))
1476 rpipe->pipe_state |= PIPE_SEL;
1479 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1480 selrecord(td, &wpipe->pipe_sel);
1481 if (SEL_WAITING(&wpipe->pipe_sel))
1482 wpipe->pipe_state |= PIPE_SEL;
1494 * We shouldn't need locks here as we're doing a read and this should
1495 * be a natural race.
1498 pipe_stat(fp, ub, active_cred, td)
1501 struct ucred *active_cred;
1513 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1520 /* For named pipes ask the underlying filesystem. */
1521 if (pipe->pipe_state & PIPE_NAMED) {
1523 return (vnops.fo_stat(fp, ub, active_cred, td));
1527 * Lazily allocate an inode number for the pipe. Most pipe
1528 * users do not call fstat(2) on the pipe, which means that
1529 * postponing the inode allocation until it is must be
1530 * returned to userland is useful. If alloc_unr failed,
1531 * assign st_ino zero instead of returning an error.
1532 * Special pipe_ino values:
1533 * -1 - not yet initialized;
1534 * 0 - alloc_unr failed, return 0 as st_ino forever.
1536 if (pipe->pipe_ino == (ino_t)-1) {
1537 new_unr = alloc_unr(pipeino_unr);
1539 pipe->pipe_ino = new_unr;
1545 bzero(ub, sizeof(*ub));
1546 ub->st_mode = S_IFIFO;
1547 ub->st_blksize = PAGE_SIZE;
1548 if (pipe->pipe_state & PIPE_DIRECTW)
1549 ub->st_size = pipe->pipe_map.cnt;
1551 ub->st_size = pipe->pipe_buffer.cnt;
1552 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1553 ub->st_atim = pipe->pipe_atime;
1554 ub->st_mtim = pipe->pipe_mtime;
1555 ub->st_ctim = pipe->pipe_ctime;
1556 ub->st_uid = fp->f_cred->cr_uid;
1557 ub->st_gid = fp->f_cred->cr_gid;
1558 ub->st_dev = pipedev_ino;
1559 ub->st_ino = pipe->pipe_ino;
1561 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1573 if (fp->f_vnode != NULL)
1574 return vnops.fo_close(fp, td);
1575 fp->f_ops = &badfileops;
1576 pipe_dtor(fp->f_data);
1582 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1588 if (cpipe->pipe_state & PIPE_NAMED)
1589 error = vn_chmod(fp, mode, active_cred, td);
1591 error = invfo_chmod(fp, mode, active_cred, td);
1596 pipe_chown(fp, uid, gid, active_cred, td)
1600 struct ucred *active_cred;
1607 if (cpipe->pipe_state & PIPE_NAMED)
1608 error = vn_chown(fp, uid, gid, active_cred, td);
1610 error = invfo_chown(fp, uid, gid, active_cred, td);
1615 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1619 if (fp->f_type == DTYPE_FIFO)
1620 return (vn_fill_kinfo(fp, kif, fdp));
1621 kif->kf_type = KF_TYPE_PIPE;
1623 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1624 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1625 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1630 pipe_free_kmem(cpipe)
1634 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1635 ("pipe_free_kmem: pipe mutex locked"));
1637 if (cpipe->pipe_buffer.buffer != NULL) {
1638 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1639 vm_map_remove(pipe_map,
1640 (vm_offset_t)cpipe->pipe_buffer.buffer,
1641 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1642 cpipe->pipe_buffer.buffer = NULL;
1644 #ifndef PIPE_NODIRECT
1646 cpipe->pipe_map.cnt = 0;
1647 cpipe->pipe_map.pos = 0;
1648 cpipe->pipe_map.npages = 0;
1660 struct pipepair *pp;
1663 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1667 pp = cpipe->pipe_pair;
1669 pipeselwakeup(cpipe);
1672 * If the other side is blocked, wake it up saying that
1673 * we want to close it down.
1675 cpipe->pipe_state |= PIPE_EOF;
1676 while (cpipe->pipe_busy) {
1678 cpipe->pipe_state |= PIPE_WANT;
1680 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1686 * Disconnect from peer, if any.
1688 ppipe = cpipe->pipe_peer;
1689 if (ppipe->pipe_present == PIPE_ACTIVE) {
1690 pipeselwakeup(ppipe);
1692 ppipe->pipe_state |= PIPE_EOF;
1694 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1698 * Mark this endpoint as free. Release kmem resources. We
1699 * don't mark this endpoint as unused until we've finished
1700 * doing that, or the pipe might disappear out from under
1704 pipe_free_kmem(cpipe);
1706 cpipe->pipe_present = PIPE_CLOSING;
1710 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1711 * PIPE_FINALIZED, that allows other end to free the
1712 * pipe_pair, only after the knotes are completely dismantled.
1714 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1715 cpipe->pipe_present = PIPE_FINALIZED;
1716 seldrain(&cpipe->pipe_sel);
1717 knlist_destroy(&cpipe->pipe_sel.si_note);
1720 * If both endpoints are now closed, release the memory for the
1721 * pipe pair. If not, unlock.
1723 if (ppipe->pipe_present == PIPE_FINALIZED) {
1726 mac_pipe_destroy(pp);
1728 uma_zfree(pipe_zone, cpipe->pipe_pair);
1735 pipe_kqfilter(struct file *fp, struct knote *kn)
1740 * If a filter is requested that is not supported by this file
1741 * descriptor, don't return an error, but also don't ever generate an
1744 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1745 kn->kn_fop = &pipe_nfiltops;
1748 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1749 kn->kn_fop = &pipe_nfiltops;
1754 switch (kn->kn_filter) {
1756 kn->kn_fop = &pipe_rfiltops;
1759 kn->kn_fop = &pipe_wfiltops;
1760 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1761 /* other end of pipe has been closed */
1765 cpipe = PIPE_PEER(cpipe);
1772 kn->kn_hook = cpipe;
1773 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1779 filt_pipedetach(struct knote *kn)
1781 struct pipe *cpipe = kn->kn_hook;
1784 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1790 filt_piperead(struct knote *kn, long hint)
1792 struct pipe *rpipe = kn->kn_hook;
1793 struct pipe *wpipe = rpipe->pipe_peer;
1796 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1797 kn->kn_data = rpipe->pipe_buffer.cnt;
1798 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1799 kn->kn_data = rpipe->pipe_map.cnt;
1801 if ((rpipe->pipe_state & PIPE_EOF) ||
1802 wpipe->pipe_present != PIPE_ACTIVE ||
1803 (wpipe->pipe_state & PIPE_EOF)) {
1804 kn->kn_flags |= EV_EOF;
1807 ret = kn->kn_data > 0;
1813 filt_pipewrite(struct knote *kn, long hint)
1817 wpipe = kn->kn_hook;
1818 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1819 if (wpipe->pipe_present != PIPE_ACTIVE ||
1820 (wpipe->pipe_state & PIPE_EOF)) {
1822 kn->kn_flags |= EV_EOF;
1825 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1826 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1827 if (wpipe->pipe_state & PIPE_DIRECTW)
1830 return (kn->kn_data >= PIPE_BUF);
1834 filt_pipedetach_notsup(struct knote *kn)
1840 filt_pipenotsup(struct knote *kn, long hint)