2 * SPDX-License-Identifier: BSD-4-Clause
4 * Copyright (c) 1996 John S. Dyson
5 * Copyright (c) 2012 Giovanni Trematerra
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice immediately at the beginning of the file, without modification,
13 * this list of conditions, and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Absolutely no warranty of function or purpose is made by the author
19 * 4. Modifications may be freely made to this file if the above conditions
24 * This file contains a high-performance replacement for the socket-based
25 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support
26 * all features of sockets, but does do everything that pipes normally
31 * This code has two modes of operation, a small write mode and a large
32 * write mode. The small write mode acts like conventional pipes with
33 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the
34 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT
35 * and PIPE_SIZE in size, the sending process pins the underlying pages in
36 * memory, and the receiving process copies directly from these pinned pages
37 * in the sending process.
39 * If the sending process receives a signal, it is possible that it will
40 * go away, and certainly its address space can change, because control
41 * is returned back to the user-mode side. In that case, the pipe code
42 * arranges to copy the buffer supplied by the user process, to a pageable
43 * kernel buffer, and the receiving process will grab the data from the
44 * pageable kernel buffer. Since signals don't happen all that often,
45 * the copy operation is normally eliminated.
47 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
48 * happen for small transfers so that the system will not spend all of
49 * its time context switching.
51 * In order to limit the resource use of pipes, two sysctls exist:
53 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
54 * address space available to us in pipe_map. This value is normally
55 * autotuned, but may also be loader tuned.
57 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
58 * memory in use by pipes.
60 * Based on how large pipekva is relative to maxpipekva, the following
64 * New pipes are given 16K of memory backing, pipes may dynamically
65 * grow to as large as 64K where needed.
67 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
68 * existing pipes may NOT grow.
70 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
71 * existing pipes will be shrunk down to 4K whenever possible.
73 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If
74 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
75 * resize which MUST occur for reverse-direction pipes when they are
78 * Additional information about the current state of pipes may be obtained
79 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
80 * and kern.ipc.piperesizefail.
82 * Locking rules: There are two locks present here: A mutex, used via
83 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via
84 * the flag, as mutexes can not persist over uiomove. The mutex
85 * exists only to guard access to the flag, and is not in itself a
86 * locking mechanism. Also note that there is only a single mutex for
87 * both directions of a pipe.
89 * As pipelock() may have to sleep before it can acquire the flag, it
90 * is important to reread all data after a call to pipelock(); everything
91 * in the structure may have changed.
94 #include <sys/cdefs.h>
95 __FBSDID("$FreeBSD$");
97 #include <sys/param.h>
98 #include <sys/systm.h>
100 #include <sys/fcntl.h>
101 #include <sys/file.h>
102 #include <sys/filedesc.h>
103 #include <sys/filio.h>
104 #include <sys/kernel.h>
105 #include <sys/lock.h>
106 #include <sys/mutex.h>
107 #include <sys/ttycom.h>
108 #include <sys/stat.h>
109 #include <sys/malloc.h>
110 #include <sys/poll.h>
111 #include <sys/selinfo.h>
112 #include <sys/signalvar.h>
113 #include <sys/syscallsubr.h>
114 #include <sys/sysctl.h>
115 #include <sys/sysproto.h>
116 #include <sys/pipe.h>
117 #include <sys/proc.h>
118 #include <sys/vnode.h>
120 #include <sys/user.h>
121 #include <sys/event.h>
123 #include <security/mac/mac_framework.h>
126 #include <vm/vm_param.h>
127 #include <vm/vm_object.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_extern.h>
131 #include <vm/vm_map.h>
132 #include <vm/vm_page.h>
136 * Use this define if you want to disable *fancy* VM things. Expect an
137 * approx 30% decrease in transfer rate. This could be useful for
140 /* #define PIPE_NODIRECT */
142 #define PIPE_PEER(pipe) \
143 (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
146 * interfaces to the outside world
148 static fo_rdwr_t pipe_read;
149 static fo_rdwr_t pipe_write;
150 static fo_truncate_t pipe_truncate;
151 static fo_ioctl_t pipe_ioctl;
152 static fo_poll_t pipe_poll;
153 static fo_kqfilter_t pipe_kqfilter;
154 static fo_stat_t pipe_stat;
155 static fo_close_t pipe_close;
156 static fo_chmod_t pipe_chmod;
157 static fo_chown_t pipe_chown;
158 static fo_fill_kinfo_t pipe_fill_kinfo;
160 struct fileops pipeops = {
161 .fo_read = pipe_read,
162 .fo_write = pipe_write,
163 .fo_truncate = pipe_truncate,
164 .fo_ioctl = pipe_ioctl,
165 .fo_poll = pipe_poll,
166 .fo_kqfilter = pipe_kqfilter,
167 .fo_stat = pipe_stat,
168 .fo_close = pipe_close,
169 .fo_chmod = pipe_chmod,
170 .fo_chown = pipe_chown,
171 .fo_sendfile = invfo_sendfile,
172 .fo_fill_kinfo = pipe_fill_kinfo,
173 .fo_flags = DFLAG_PASSABLE
176 static void filt_pipedetach(struct knote *kn);
177 static void filt_pipedetach_notsup(struct knote *kn);
178 static int filt_pipenotsup(struct knote *kn, long hint);
179 static int filt_piperead(struct knote *kn, long hint);
180 static int filt_pipewrite(struct knote *kn, long hint);
182 static struct filterops pipe_nfiltops = {
184 .f_detach = filt_pipedetach_notsup,
185 .f_event = filt_pipenotsup
187 static struct filterops pipe_rfiltops = {
189 .f_detach = filt_pipedetach,
190 .f_event = filt_piperead
192 static struct filterops pipe_wfiltops = {
194 .f_detach = filt_pipedetach,
195 .f_event = filt_pipewrite
199 * Default pipe buffer size(s), this can be kind-of large now because pipe
200 * space is pageable. The pipe code will try to maintain locality of
201 * reference for performance reasons, so small amounts of outstanding I/O
202 * will not wipe the cache.
204 #define MINPIPESIZE (PIPE_SIZE/3)
205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
207 static long amountpipekva;
208 static int pipefragretry;
209 static int pipeallocfail;
210 static int piperesizefail;
211 static int piperesizeallowed = 1;
213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
214 &maxpipekva, 0, "Pipe KVA limit");
215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
216 &amountpipekva, 0, "Pipe KVA usage");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
218 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
220 &pipeallocfail, 0, "Pipe allocation failures");
221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
222 &piperesizefail, 0, "Pipe resize failures");
223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
224 &piperesizeallowed, 0, "Pipe resizing allowed");
226 static void pipeinit(void *dummy __unused);
227 static void pipeclose(struct pipe *cpipe);
228 static void pipe_free_kmem(struct pipe *cpipe);
229 static void pipe_create(struct pipe *pipe, int backing);
230 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
231 static __inline int pipelock(struct pipe *cpipe, int catch);
232 static __inline void pipeunlock(struct pipe *cpipe);
233 #ifndef PIPE_NODIRECT
234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
237 static void pipe_clone_write_buffer(struct pipe *wpipe);
239 static int pipespace(struct pipe *cpipe, int size);
240 static int pipespace_new(struct pipe *cpipe, int size);
242 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
243 static int pipe_zone_init(void *mem, int size, int flags);
244 static void pipe_zone_fini(void *mem, int size);
246 static uma_zone_t pipe_zone;
247 static struct unrhdr *pipeino_unr;
248 static dev_t pipedev_ino;
250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
253 pipeinit(void *dummy __unused)
256 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
257 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
259 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
260 pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
261 KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
262 pipedev_ino = devfs_alloc_cdp_inode();
263 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
267 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
270 struct pipe *rpipe, *wpipe;
272 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
274 pp = (struct pipepair *)mem;
277 * We zero both pipe endpoints to make sure all the kmem pointers
278 * are NULL, flag fields are zero'd, etc. We timestamp both
279 * endpoints with the same time.
281 rpipe = &pp->pp_rpipe;
282 bzero(rpipe, sizeof(*rpipe));
283 vfs_timestamp(&rpipe->pipe_ctime);
284 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
286 wpipe = &pp->pp_wpipe;
287 bzero(wpipe, sizeof(*wpipe));
288 wpipe->pipe_ctime = rpipe->pipe_ctime;
289 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
291 rpipe->pipe_peer = wpipe;
292 rpipe->pipe_pair = pp;
293 wpipe->pipe_peer = rpipe;
294 wpipe->pipe_pair = pp;
297 * Mark both endpoints as present; they will later get free'd
298 * one at a time. When both are free'd, then the whole pair
301 rpipe->pipe_present = PIPE_ACTIVE;
302 wpipe->pipe_present = PIPE_ACTIVE;
305 * Eventually, the MAC Framework may initialize the label
306 * in ctor or init, but for now we do it elswhere to avoid
307 * blocking in ctor or init.
315 pipe_zone_init(void *mem, int size, int flags)
319 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
321 pp = (struct pipepair *)mem;
323 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
328 pipe_zone_fini(void *mem, int size)
332 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
334 pp = (struct pipepair *)mem;
336 mtx_destroy(&pp->pp_mtx);
340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
343 struct pipe *rpipe, *wpipe;
345 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
348 * The MAC label is shared between the connected endpoints. As a
349 * result mac_pipe_init() and mac_pipe_create() are called once
350 * for the pair, and not on the endpoints.
353 mac_pipe_create(td->td_ucred, pp);
355 rpipe = &pp->pp_rpipe;
356 wpipe = &pp->pp_wpipe;
358 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
359 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
361 /* Only the forward direction pipe is backed by default */
362 pipe_create(rpipe, 1);
363 pipe_create(wpipe, 0);
365 rpipe->pipe_state |= PIPE_DIRECTOK;
366 wpipe->pipe_state |= PIPE_DIRECTOK;
370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
374 pipe_paircreate(td, &pp);
375 pp->pp_rpipe.pipe_state |= PIPE_NAMED;
376 *ppipe = &pp->pp_rpipe;
380 pipe_dtor(struct pipe *dpipe)
385 ino = dpipe->pipe_ino;
386 peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
387 funsetown(&dpipe->pipe_sigio);
390 funsetown(&peer->pipe_sigio);
393 if (ino != 0 && ino != (ino_t)-1)
394 free_unr(pipeino_unr, ino);
398 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
399 * the zone pick up the pieces via pipeclose().
402 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
403 struct filecaps *fcaps2)
405 struct file *rf, *wf;
406 struct pipe *rpipe, *wpipe;
408 int fd, fflags, error;
410 pipe_paircreate(td, &pp);
411 rpipe = &pp->pp_rpipe;
412 wpipe = &pp->pp_wpipe;
413 error = falloc_caps(td, &rf, &fd, flags, fcaps1);
419 /* An extra reference on `rf' has been held for us by falloc_caps(). */
422 fflags = FREAD | FWRITE;
423 if ((flags & O_NONBLOCK) != 0)
427 * Warning: once we've gotten past allocation of the fd for the
428 * read-side, we can only drop the read side via fdrop() in order
429 * to avoid races against processes which manage to dup() the read
430 * side while we are blocked trying to allocate the write side.
432 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
433 error = falloc_caps(td, &wf, &fd, flags, fcaps2);
435 fdclose(td, rf, fildes[0]);
437 /* rpipe has been closed by fdrop(). */
441 /* An extra reference on `wf' has been held for us by falloc_caps(). */
442 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
450 #ifdef COMPAT_FREEBSD10
453 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
458 error = kern_pipe(td, fildes, 0, NULL, NULL);
462 td->td_retval[0] = fildes[0];
463 td->td_retval[1] = fildes[1];
470 sys_pipe2(struct thread *td, struct pipe2_args *uap)
472 int error, fildes[2];
474 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
476 error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
479 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
481 (void)kern_close(td, fildes[0]);
482 (void)kern_close(td, fildes[1]);
488 * Allocate kva for pipe circular buffer, the space is pageable
489 * This routine will 'realloc' the size of a pipe safely, if it fails
490 * it will retain the old buffer.
491 * If it fails it will return ENOMEM.
494 pipespace_new(struct pipe *cpipe, int size)
497 int error, cnt, firstseg;
498 static int curfail = 0;
499 static struct timeval lastfail;
501 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
502 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
503 ("pipespace: resize of direct writes not allowed"));
505 cnt = cpipe->pipe_buffer.cnt;
509 size = round_page(size);
510 buffer = (caddr_t) vm_map_min(pipe_map);
512 error = vm_map_find(pipe_map, NULL, 0,
513 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
514 VM_PROT_ALL, VM_PROT_ALL, 0);
515 if (error != KERN_SUCCESS) {
516 if ((cpipe->pipe_buffer.buffer == NULL) &&
517 (size > SMALL_PIPE_SIZE)) {
518 size = SMALL_PIPE_SIZE;
522 if (cpipe->pipe_buffer.buffer == NULL) {
524 if (ppsratecheck(&lastfail, &curfail, 1))
525 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
532 /* copy data, then free old resources if we're resizing */
534 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
535 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
536 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
538 if ((cnt - firstseg) > 0)
539 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
540 cpipe->pipe_buffer.in);
542 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
546 pipe_free_kmem(cpipe);
547 cpipe->pipe_buffer.buffer = buffer;
548 cpipe->pipe_buffer.size = size;
549 cpipe->pipe_buffer.in = cnt;
550 cpipe->pipe_buffer.out = 0;
551 cpipe->pipe_buffer.cnt = cnt;
552 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
557 * Wrapper for pipespace_new() that performs locking assertions.
560 pipespace(struct pipe *cpipe, int size)
563 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
564 ("Unlocked pipe passed to pipespace"));
565 return (pipespace_new(cpipe, size));
569 * lock a pipe for I/O, blocking other access
572 pipelock(struct pipe *cpipe, int catch)
576 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
577 while (cpipe->pipe_state & PIPE_LOCKFL) {
578 cpipe->pipe_state |= PIPE_LWANT;
579 error = msleep(cpipe, PIPE_MTX(cpipe),
580 catch ? (PRIBIO | PCATCH) : PRIBIO,
585 cpipe->pipe_state |= PIPE_LOCKFL;
590 * unlock a pipe I/O lock
593 pipeunlock(struct pipe *cpipe)
596 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
597 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
598 ("Unlocked pipe passed to pipeunlock"));
599 cpipe->pipe_state &= ~PIPE_LOCKFL;
600 if (cpipe->pipe_state & PIPE_LWANT) {
601 cpipe->pipe_state &= ~PIPE_LWANT;
607 pipeselwakeup(struct pipe *cpipe)
610 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
611 if (cpipe->pipe_state & PIPE_SEL) {
612 selwakeuppri(&cpipe->pipe_sel, PSOCK);
613 if (!SEL_WAITING(&cpipe->pipe_sel))
614 cpipe->pipe_state &= ~PIPE_SEL;
616 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
617 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
618 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
622 * Initialize and allocate VM and memory for pipe. The structure
623 * will start out zero'd from the ctor, so we just manage the kmem.
626 pipe_create(struct pipe *pipe, int backing)
631 * Note that these functions can fail if pipe map is exhausted
632 * (as a result of too many pipes created), but we ignore the
633 * error as it is not fatal and could be provoked by
634 * unprivileged users. The only consequence is worse performance
637 if (amountpipekva > maxpipekva / 2)
638 (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
640 (void)pipespace_new(pipe, PIPE_SIZE);
648 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
649 int flags, struct thread *td)
659 error = pipelock(rpipe, 1);
664 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
668 if (amountpipekva > (3 * maxpipekva) / 4) {
669 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
670 (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
671 (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
672 (piperesizeallowed == 1)) {
674 pipespace(rpipe, SMALL_PIPE_SIZE);
679 while (uio->uio_resid) {
681 * normal pipe buffer receive
683 if (rpipe->pipe_buffer.cnt > 0) {
684 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
685 if (size > rpipe->pipe_buffer.cnt)
686 size = rpipe->pipe_buffer.cnt;
687 if (size > uio->uio_resid)
688 size = uio->uio_resid;
692 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
698 rpipe->pipe_buffer.out += size;
699 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
700 rpipe->pipe_buffer.out = 0;
702 rpipe->pipe_buffer.cnt -= size;
705 * If there is no more to read in the pipe, reset
706 * its pointers to the beginning. This improves
709 if (rpipe->pipe_buffer.cnt == 0) {
710 rpipe->pipe_buffer.in = 0;
711 rpipe->pipe_buffer.out = 0;
714 #ifndef PIPE_NODIRECT
716 * Direct copy, bypassing a kernel buffer.
718 } else if ((size = rpipe->pipe_map.cnt) &&
719 (rpipe->pipe_state & PIPE_DIRECTW)) {
720 if (size > uio->uio_resid)
721 size = (u_int) uio->uio_resid;
724 error = uiomove_fromphys(rpipe->pipe_map.ms,
725 rpipe->pipe_map.pos, size, uio);
730 rpipe->pipe_map.pos += size;
731 rpipe->pipe_map.cnt -= size;
732 if (rpipe->pipe_map.cnt == 0) {
733 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
739 * detect EOF condition
740 * read returns 0 on EOF, no need to set error
742 if (rpipe->pipe_state & PIPE_EOF)
746 * If the "write-side" has been blocked, wake it up now.
748 if (rpipe->pipe_state & PIPE_WANTW) {
749 rpipe->pipe_state &= ~PIPE_WANTW;
754 * Break if some data was read.
760 * Unlock the pipe buffer for our remaining processing.
761 * We will either break out with an error or we will
762 * sleep and relock to loop.
767 * Handle non-blocking mode operation or
768 * wait for more data.
770 if (fp->f_flag & FNONBLOCK) {
773 rpipe->pipe_state |= PIPE_WANTR;
774 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
777 error = pipelock(rpipe, 1);
788 /* XXX: should probably do this before getting any locks. */
790 vfs_timestamp(&rpipe->pipe_atime);
795 * PIPE_WANT processing only makes sense if pipe_busy is 0.
797 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
798 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
800 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
802 * Handle write blocking hysteresis.
804 if (rpipe->pipe_state & PIPE_WANTW) {
805 rpipe->pipe_state &= ~PIPE_WANTW;
810 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
811 pipeselwakeup(rpipe);
817 #ifndef PIPE_NODIRECT
819 * Map the sending processes' buffer into kernel space and wire it.
820 * This is similar to a physical write operation.
823 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
828 PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
829 KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
830 ("Clone attempt on non-direct write pipe!"));
832 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
833 size = wpipe->pipe_buffer.size;
835 size = uio->uio_iov->iov_len;
837 if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
838 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
839 wpipe->pipe_map.ms, PIPENPAGES)) < 0)
843 * set up the control block
845 wpipe->pipe_map.npages = i;
846 wpipe->pipe_map.pos =
847 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
848 wpipe->pipe_map.cnt = size;
851 * and update the uio data
854 uio->uio_iov->iov_len -= size;
855 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
856 if (uio->uio_iov->iov_len == 0)
858 uio->uio_resid -= size;
859 uio->uio_offset += size;
864 * unmap and unwire the process buffer
867 pipe_destroy_write_buffer(struct pipe *wpipe)
870 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
871 vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
872 wpipe->pipe_map.npages = 0;
876 * In the case of a signal, the writing process might go away. This
877 * code copies the data into the circular buffer so that the source
878 * pages can be freed without loss of data.
881 pipe_clone_write_buffer(struct pipe *wpipe)
888 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
889 size = wpipe->pipe_map.cnt;
890 pos = wpipe->pipe_map.pos;
892 wpipe->pipe_buffer.in = size;
893 wpipe->pipe_buffer.out = 0;
894 wpipe->pipe_buffer.cnt = size;
895 wpipe->pipe_state &= ~PIPE_DIRECTW;
898 iov.iov_base = wpipe->pipe_buffer.buffer;
903 uio.uio_resid = size;
904 uio.uio_segflg = UIO_SYSSPACE;
905 uio.uio_rw = UIO_READ;
906 uio.uio_td = curthread;
907 uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
909 pipe_destroy_write_buffer(wpipe);
913 * This implements the pipe buffer write mechanism. Note that only
914 * a direct write OR a normal pipe write can be pending at any given time.
915 * If there are any characters in the pipe buffer, the direct write will
916 * be deferred until the receiving process grabs all of the bytes from
917 * the pipe buffer. Then the direct mapping write is set-up.
920 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
925 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
926 error = pipelock(wpipe, 1);
929 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
934 while (wpipe->pipe_state & PIPE_DIRECTW) {
935 if (wpipe->pipe_state & PIPE_WANTR) {
936 wpipe->pipe_state &= ~PIPE_WANTR;
939 pipeselwakeup(wpipe);
940 wpipe->pipe_state |= PIPE_WANTW;
942 error = msleep(wpipe, PIPE_MTX(wpipe),
943 PRIBIO | PCATCH, "pipdww", 0);
949 wpipe->pipe_map.cnt = 0; /* transfer not ready yet */
950 if (wpipe->pipe_buffer.cnt > 0) {
951 if (wpipe->pipe_state & PIPE_WANTR) {
952 wpipe->pipe_state &= ~PIPE_WANTR;
955 pipeselwakeup(wpipe);
956 wpipe->pipe_state |= PIPE_WANTW;
958 error = msleep(wpipe, PIPE_MTX(wpipe),
959 PRIBIO | PCATCH, "pipdwc", 0);
966 wpipe->pipe_state |= PIPE_DIRECTW;
969 error = pipe_build_write_buffer(wpipe, uio);
972 wpipe->pipe_state &= ~PIPE_DIRECTW;
978 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
979 if (wpipe->pipe_state & PIPE_EOF) {
980 pipe_destroy_write_buffer(wpipe);
981 pipeselwakeup(wpipe);
986 if (wpipe->pipe_state & PIPE_WANTR) {
987 wpipe->pipe_state &= ~PIPE_WANTR;
990 pipeselwakeup(wpipe);
991 wpipe->pipe_state |= PIPE_WANTW;
993 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
998 if (wpipe->pipe_state & PIPE_EOF)
1000 if (wpipe->pipe_state & PIPE_DIRECTW) {
1002 * this bit of trickery substitutes a kernel buffer for
1003 * the process that might be going away.
1005 pipe_clone_write_buffer(wpipe);
1007 pipe_destroy_write_buffer(wpipe);
1019 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1020 int flags, struct thread *td)
1025 struct pipe *wpipe, *rpipe;
1028 wpipe = PIPE_PEER(rpipe);
1030 error = pipelock(wpipe, 1);
1036 * detect loss of pipe read side, issue SIGPIPE if lost.
1038 if (wpipe->pipe_present != PIPE_ACTIVE ||
1039 (wpipe->pipe_state & PIPE_EOF)) {
1045 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1054 /* Choose a larger size if it's advantageous */
1055 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1056 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1057 if (piperesizeallowed != 1)
1059 if (amountpipekva > maxpipekva / 2)
1061 if (desiredsize == BIG_PIPE_SIZE)
1063 desiredsize = desiredsize * 2;
1066 /* Choose a smaller size if we're in a OOM situation */
1067 if ((amountpipekva > (3 * maxpipekva) / 4) &&
1068 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1069 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1070 (piperesizeallowed == 1))
1071 desiredsize = SMALL_PIPE_SIZE;
1073 /* Resize if the above determined that a new size was necessary */
1074 if ((desiredsize != wpipe->pipe_buffer.size) &&
1075 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1077 pipespace(wpipe, desiredsize);
1080 if (wpipe->pipe_buffer.size == 0) {
1082 * This can only happen for reverse direction use of pipes
1083 * in a complete OOM situation.
1094 orig_resid = uio->uio_resid;
1096 while (uio->uio_resid) {
1100 if (wpipe->pipe_state & PIPE_EOF) {
1105 #ifndef PIPE_NODIRECT
1107 * If the transfer is large, we can gain performance if
1108 * we do process-to-process copies directly.
1109 * If the write is non-blocking, we don't use the
1110 * direct write mechanism.
1112 * The direct write mechanism will detect the reader going
1115 if (uio->uio_segflg == UIO_USERSPACE &&
1116 uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1117 wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1118 (fp->f_flag & FNONBLOCK) == 0) {
1120 error = pipe_direct_write(wpipe, uio);
1128 * Pipe buffered writes cannot be coincidental with
1129 * direct writes. We wait until the currently executing
1130 * direct write is completed before we start filling the
1131 * pipe buffer. We break out if a signal occurs or the
1134 if (wpipe->pipe_state & PIPE_DIRECTW) {
1135 if (wpipe->pipe_state & PIPE_WANTR) {
1136 wpipe->pipe_state &= ~PIPE_WANTR;
1139 pipeselwakeup(wpipe);
1140 wpipe->pipe_state |= PIPE_WANTW;
1142 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1150 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1152 /* Writes of size <= PIPE_BUF must be atomic. */
1153 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1157 int size; /* Transfer size */
1158 int segsize; /* first segment to transfer */
1161 * Transfer size is minimum of uio transfer
1162 * and free space in pipe buffer.
1164 if (space > uio->uio_resid)
1165 size = uio->uio_resid;
1169 * First segment to transfer is minimum of
1170 * transfer size and contiguous space in
1171 * pipe buffer. If first segment to transfer
1172 * is less than the transfer size, we've got
1173 * a wraparound in the buffer.
1175 segsize = wpipe->pipe_buffer.size -
1176 wpipe->pipe_buffer.in;
1180 /* Transfer first segment */
1183 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1187 if (error == 0 && segsize < size) {
1188 KASSERT(wpipe->pipe_buffer.in + segsize ==
1189 wpipe->pipe_buffer.size,
1190 ("Pipe buffer wraparound disappeared"));
1192 * Transfer remaining part now, to
1193 * support atomic writes. Wraparound
1199 &wpipe->pipe_buffer.buffer[0],
1200 size - segsize, uio);
1204 wpipe->pipe_buffer.in += size;
1205 if (wpipe->pipe_buffer.in >=
1206 wpipe->pipe_buffer.size) {
1207 KASSERT(wpipe->pipe_buffer.in ==
1209 wpipe->pipe_buffer.size,
1210 ("Expected wraparound bad"));
1211 wpipe->pipe_buffer.in = size - segsize;
1214 wpipe->pipe_buffer.cnt += size;
1215 KASSERT(wpipe->pipe_buffer.cnt <=
1216 wpipe->pipe_buffer.size,
1217 ("Pipe buffer overflow"));
1224 * If the "read-side" has been blocked, wake it up now.
1226 if (wpipe->pipe_state & PIPE_WANTR) {
1227 wpipe->pipe_state &= ~PIPE_WANTR;
1232 * don't block on non-blocking I/O
1234 if (fp->f_flag & FNONBLOCK) {
1241 * We have no more space and have something to offer,
1242 * wake up select/poll.
1244 pipeselwakeup(wpipe);
1246 wpipe->pipe_state |= PIPE_WANTW;
1248 error = msleep(wpipe, PIPE_MTX(rpipe),
1249 PRIBIO | PCATCH, "pipewr", 0);
1258 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1259 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1261 } else if (wpipe->pipe_buffer.cnt > 0) {
1263 * If we have put any characters in the buffer, we wake up
1266 if (wpipe->pipe_state & PIPE_WANTR) {
1267 wpipe->pipe_state &= ~PIPE_WANTR;
1273 * Don't return EPIPE if any byte was written.
1274 * EINTR and other interrupts are handled by generic I/O layer.
1275 * Do not pretend that I/O succeeded for obvious user error
1278 if (uio->uio_resid != orig_resid && error == EPIPE)
1282 vfs_timestamp(&wpipe->pipe_mtime);
1285 * We have something to offer,
1286 * wake up select/poll.
1288 if (wpipe->pipe_buffer.cnt)
1289 pipeselwakeup(wpipe);
1298 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1305 if (cpipe->pipe_state & PIPE_NAMED)
1306 error = vnops.fo_truncate(fp, length, active_cred, td);
1308 error = invfo_truncate(fp, length, active_cred, td);
1313 * we implement a very minimal set of ioctls for compatibility with sockets.
1316 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1319 struct pipe *mpipe = fp->f_data;
1325 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1340 mpipe->pipe_state |= PIPE_ASYNC;
1342 mpipe->pipe_state &= ~PIPE_ASYNC;
1347 if (!(fp->f_flag & FREAD)) {
1352 if (mpipe->pipe_state & PIPE_DIRECTW)
1353 *(int *)data = mpipe->pipe_map.cnt;
1355 *(int *)data = mpipe->pipe_buffer.cnt;
1360 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1364 *(int *)data = fgetown(&mpipe->pipe_sigio);
1367 /* This is deprecated, FIOSETOWN should be used instead. */
1370 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1373 /* This is deprecated, FIOGETOWN should be used instead. */
1375 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1388 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1393 int levents, revents;
1400 wpipe = PIPE_PEER(rpipe);
1403 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1407 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1408 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1409 (rpipe->pipe_buffer.cnt > 0))
1410 revents |= events & (POLLIN | POLLRDNORM);
1412 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1413 if (wpipe->pipe_present != PIPE_ACTIVE ||
1414 (wpipe->pipe_state & PIPE_EOF) ||
1415 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1416 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1417 wpipe->pipe_buffer.size == 0)))
1418 revents |= events & (POLLOUT | POLLWRNORM);
1421 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1422 if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1423 fp->f_seqcount == rpipe->pipe_wgen)
1424 events |= POLLINIGNEOF;
1426 if ((events & POLLINIGNEOF) == 0) {
1427 if (rpipe->pipe_state & PIPE_EOF) {
1428 revents |= (events & (POLLIN | POLLRDNORM));
1429 if (wpipe->pipe_present != PIPE_ACTIVE ||
1430 (wpipe->pipe_state & PIPE_EOF))
1436 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1437 selrecord(td, &rpipe->pipe_sel);
1438 if (SEL_WAITING(&rpipe->pipe_sel))
1439 rpipe->pipe_state |= PIPE_SEL;
1442 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1443 selrecord(td, &wpipe->pipe_sel);
1444 if (SEL_WAITING(&wpipe->pipe_sel))
1445 wpipe->pipe_state |= PIPE_SEL;
1457 * We shouldn't need locks here as we're doing a read and this should
1458 * be a natural race.
1461 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1473 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1480 /* For named pipes ask the underlying filesystem. */
1481 if (pipe->pipe_state & PIPE_NAMED) {
1483 return (vnops.fo_stat(fp, ub, active_cred, td));
1487 * Lazily allocate an inode number for the pipe. Most pipe
1488 * users do not call fstat(2) on the pipe, which means that
1489 * postponing the inode allocation until it is must be
1490 * returned to userland is useful. If alloc_unr failed,
1491 * assign st_ino zero instead of returning an error.
1492 * Special pipe_ino values:
1493 * -1 - not yet initialized;
1494 * 0 - alloc_unr failed, return 0 as st_ino forever.
1496 if (pipe->pipe_ino == (ino_t)-1) {
1497 new_unr = alloc_unr(pipeino_unr);
1499 pipe->pipe_ino = new_unr;
1505 bzero(ub, sizeof(*ub));
1506 ub->st_mode = S_IFIFO;
1507 ub->st_blksize = PAGE_SIZE;
1508 if (pipe->pipe_state & PIPE_DIRECTW)
1509 ub->st_size = pipe->pipe_map.cnt;
1511 ub->st_size = pipe->pipe_buffer.cnt;
1512 ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1513 ub->st_atim = pipe->pipe_atime;
1514 ub->st_mtim = pipe->pipe_mtime;
1515 ub->st_ctim = pipe->pipe_ctime;
1516 ub->st_uid = fp->f_cred->cr_uid;
1517 ub->st_gid = fp->f_cred->cr_gid;
1518 ub->st_dev = pipedev_ino;
1519 ub->st_ino = pipe->pipe_ino;
1521 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1528 pipe_close(struct file *fp, struct thread *td)
1531 if (fp->f_vnode != NULL)
1532 return vnops.fo_close(fp, td);
1533 fp->f_ops = &badfileops;
1534 pipe_dtor(fp->f_data);
1540 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1546 if (cpipe->pipe_state & PIPE_NAMED)
1547 error = vn_chmod(fp, mode, active_cred, td);
1549 error = invfo_chmod(fp, mode, active_cred, td);
1554 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1561 if (cpipe->pipe_state & PIPE_NAMED)
1562 error = vn_chown(fp, uid, gid, active_cred, td);
1564 error = invfo_chown(fp, uid, gid, active_cred, td);
1569 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1573 if (fp->f_type == DTYPE_FIFO)
1574 return (vn_fill_kinfo(fp, kif, fdp));
1575 kif->kf_type = KF_TYPE_PIPE;
1577 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1578 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1579 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1584 pipe_free_kmem(struct pipe *cpipe)
1587 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1588 ("pipe_free_kmem: pipe mutex locked"));
1590 if (cpipe->pipe_buffer.buffer != NULL) {
1591 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1592 vm_map_remove(pipe_map,
1593 (vm_offset_t)cpipe->pipe_buffer.buffer,
1594 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1595 cpipe->pipe_buffer.buffer = NULL;
1597 #ifndef PIPE_NODIRECT
1599 cpipe->pipe_map.cnt = 0;
1600 cpipe->pipe_map.pos = 0;
1601 cpipe->pipe_map.npages = 0;
1610 pipeclose(struct pipe *cpipe)
1612 struct pipepair *pp;
1615 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1619 pp = cpipe->pipe_pair;
1621 pipeselwakeup(cpipe);
1624 * If the other side is blocked, wake it up saying that
1625 * we want to close it down.
1627 cpipe->pipe_state |= PIPE_EOF;
1628 while (cpipe->pipe_busy) {
1630 cpipe->pipe_state |= PIPE_WANT;
1632 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1638 * Disconnect from peer, if any.
1640 ppipe = cpipe->pipe_peer;
1641 if (ppipe->pipe_present == PIPE_ACTIVE) {
1642 pipeselwakeup(ppipe);
1644 ppipe->pipe_state |= PIPE_EOF;
1646 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1650 * Mark this endpoint as free. Release kmem resources. We
1651 * don't mark this endpoint as unused until we've finished
1652 * doing that, or the pipe might disappear out from under
1656 pipe_free_kmem(cpipe);
1658 cpipe->pipe_present = PIPE_CLOSING;
1662 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1663 * PIPE_FINALIZED, that allows other end to free the
1664 * pipe_pair, only after the knotes are completely dismantled.
1666 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1667 cpipe->pipe_present = PIPE_FINALIZED;
1668 seldrain(&cpipe->pipe_sel);
1669 knlist_destroy(&cpipe->pipe_sel.si_note);
1672 * If both endpoints are now closed, release the memory for the
1673 * pipe pair. If not, unlock.
1675 if (ppipe->pipe_present == PIPE_FINALIZED) {
1678 mac_pipe_destroy(pp);
1680 uma_zfree(pipe_zone, cpipe->pipe_pair);
1687 pipe_kqfilter(struct file *fp, struct knote *kn)
1692 * If a filter is requested that is not supported by this file
1693 * descriptor, don't return an error, but also don't ever generate an
1696 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1697 kn->kn_fop = &pipe_nfiltops;
1700 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1701 kn->kn_fop = &pipe_nfiltops;
1706 switch (kn->kn_filter) {
1708 kn->kn_fop = &pipe_rfiltops;
1711 kn->kn_fop = &pipe_wfiltops;
1712 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1713 /* other end of pipe has been closed */
1717 cpipe = PIPE_PEER(cpipe);
1724 kn->kn_hook = cpipe;
1725 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1731 filt_pipedetach(struct knote *kn)
1733 struct pipe *cpipe = kn->kn_hook;
1736 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1742 filt_piperead(struct knote *kn, long hint)
1744 struct pipe *rpipe = kn->kn_hook;
1745 struct pipe *wpipe = rpipe->pipe_peer;
1748 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1749 kn->kn_data = rpipe->pipe_buffer.cnt;
1750 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1751 kn->kn_data = rpipe->pipe_map.cnt;
1753 if ((rpipe->pipe_state & PIPE_EOF) ||
1754 wpipe->pipe_present != PIPE_ACTIVE ||
1755 (wpipe->pipe_state & PIPE_EOF)) {
1756 kn->kn_flags |= EV_EOF;
1759 ret = kn->kn_data > 0;
1765 filt_pipewrite(struct knote *kn, long hint)
1769 wpipe = kn->kn_hook;
1770 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1771 if (wpipe->pipe_present != PIPE_ACTIVE ||
1772 (wpipe->pipe_state & PIPE_EOF)) {
1774 kn->kn_flags |= EV_EOF;
1777 kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1778 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1779 if (wpipe->pipe_state & PIPE_DIRECTW)
1782 return (kn->kn_data >= PIPE_BUF);
1786 filt_pipedetach_notsup(struct knote *kn)
1792 filt_pipenotsup(struct knote *kn, long hint)