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_type & PIPE_TYPE_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 int pipe_create(struct pipe *pipe, bool backing);
230 static int 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 static void pipe_timestamp(struct timespec *tsp);
234 #ifndef PIPE_NODIRECT
235 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
236 static void pipe_destroy_write_buffer(struct pipe *wpipe);
237 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
238 static void pipe_clone_write_buffer(struct pipe *wpipe);
240 static int pipespace(struct pipe *cpipe, int size);
241 static int pipespace_new(struct pipe *cpipe, int size);
243 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
244 static int pipe_zone_init(void *mem, int size, int flags);
245 static void pipe_zone_fini(void *mem, int size);
247 static uma_zone_t pipe_zone;
248 static struct unrhdr64 pipeino_unr;
249 static dev_t pipedev_ino;
251 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
254 pipeinit(void *dummy __unused)
257 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
258 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
260 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
261 new_unrhdr64(&pipeino_unr, 1);
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 pipe_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;
346 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
349 * The MAC label is shared between the connected endpoints. As a
350 * result mac_pipe_init() and mac_pipe_create() are called once
351 * for the pair, and not on the endpoints.
354 mac_pipe_create(td->td_ucred, pp);
356 rpipe = &pp->pp_rpipe;
357 wpipe = &pp->pp_wpipe;
359 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
360 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
363 * Only the forward direction pipe is backed by big buffer by
366 error = pipe_create(rpipe, true);
369 error = pipe_create(wpipe, false);
372 * This cleanup leaves the pipe inode number for rpipe
373 * still allocated, but never used. We do not free
374 * inode numbers for opened pipes, which is required
375 * for correctness because numbers must be unique.
376 * But also it avoids any memory use by the unr
377 * allocator, so stashing away the transient inode
378 * number is reasonable.
380 pipe_free_kmem(rpipe);
384 rpipe->pipe_state |= PIPE_DIRECTOK;
385 wpipe->pipe_state |= PIPE_DIRECTOK;
389 knlist_destroy(&rpipe->pipe_sel.si_note);
390 knlist_destroy(&wpipe->pipe_sel.si_note);
392 mac_pipe_destroy(pp);
398 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
403 error = pipe_paircreate(td, &pp);
406 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
407 *ppipe = &pp->pp_rpipe;
412 pipe_dtor(struct pipe *dpipe)
416 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
417 funsetown(&dpipe->pipe_sigio);
420 funsetown(&peer->pipe_sigio);
428 * This used to be vfs_timestamp but the higher precision is unnecessary and
429 * can very negatively affect performance in virtualized environments (e.g., on
430 * vms running on amd64 when using the rdtscp instruction).
433 pipe_timestamp(struct timespec *tsp)
440 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
441 * the zone pick up the pieces via pipeclose().
444 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
445 struct filecaps *fcaps2)
447 struct file *rf, *wf;
448 struct pipe *rpipe, *wpipe;
450 int fd, fflags, error;
452 error = pipe_paircreate(td, &pp);
455 rpipe = &pp->pp_rpipe;
456 wpipe = &pp->pp_wpipe;
457 error = falloc_caps(td, &rf, &fd, flags, fcaps1);
463 /* An extra reference on `rf' has been held for us by falloc_caps(). */
466 fflags = FREAD | FWRITE;
467 if ((flags & O_NONBLOCK) != 0)
471 * Warning: once we've gotten past allocation of the fd for the
472 * read-side, we can only drop the read side via fdrop() in order
473 * to avoid races against processes which manage to dup() the read
474 * side while we are blocked trying to allocate the write side.
476 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
477 error = falloc_caps(td, &wf, &fd, flags, fcaps2);
479 fdclose(td, rf, fildes[0]);
481 /* rpipe has been closed by fdrop(). */
485 /* An extra reference on `wf' has been held for us by falloc_caps(). */
486 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
494 #ifdef COMPAT_FREEBSD10
497 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
502 error = kern_pipe(td, fildes, 0, NULL, NULL);
506 td->td_retval[0] = fildes[0];
507 td->td_retval[1] = fildes[1];
514 sys_pipe2(struct thread *td, struct pipe2_args *uap)
516 int error, fildes[2];
518 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
520 error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
523 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
525 (void)kern_close(td, fildes[0]);
526 (void)kern_close(td, fildes[1]);
532 * Allocate kva for pipe circular buffer, the space is pageable
533 * This routine will 'realloc' the size of a pipe safely, if it fails
534 * it will retain the old buffer.
535 * If it fails it will return ENOMEM.
538 pipespace_new(struct pipe *cpipe, int size)
541 int error, cnt, firstseg;
542 static int curfail = 0;
543 static struct timeval lastfail;
545 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
546 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
547 ("pipespace: resize of direct writes not allowed"));
549 cnt = cpipe->pipe_buffer.cnt;
553 size = round_page(size);
554 buffer = (caddr_t) vm_map_min(pipe_map);
556 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
557 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
558 if (error != KERN_SUCCESS) {
559 if (cpipe->pipe_buffer.buffer == NULL &&
560 size > SMALL_PIPE_SIZE) {
561 size = SMALL_PIPE_SIZE;
565 if (cpipe->pipe_buffer.buffer == NULL) {
567 if (ppsratecheck(&lastfail, &curfail, 1))
568 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
575 /* copy data, then free old resources if we're resizing */
577 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
578 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
579 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
581 if ((cnt - firstseg) > 0)
582 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
583 cpipe->pipe_buffer.in);
585 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
589 pipe_free_kmem(cpipe);
590 cpipe->pipe_buffer.buffer = buffer;
591 cpipe->pipe_buffer.size = size;
592 cpipe->pipe_buffer.in = cnt;
593 cpipe->pipe_buffer.out = 0;
594 cpipe->pipe_buffer.cnt = cnt;
595 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
600 * Wrapper for pipespace_new() that performs locking assertions.
603 pipespace(struct pipe *cpipe, int size)
606 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
607 ("Unlocked pipe passed to pipespace"));
608 return (pipespace_new(cpipe, size));
612 * lock a pipe for I/O, blocking other access
615 pipelock(struct pipe *cpipe, int catch)
619 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
624 while (cpipe->pipe_state & PIPE_LOCKFL) {
625 KASSERT(cpipe->pipe_waiters >= 0,
626 ("%s: bad waiter count %d", __func__,
627 cpipe->pipe_waiters));
628 cpipe->pipe_waiters++;
629 error = msleep(cpipe, PIPE_MTX(cpipe),
631 cpipe->pipe_waiters--;
635 cpipe->pipe_state |= PIPE_LOCKFL;
640 * unlock a pipe I/O lock
643 pipeunlock(struct pipe *cpipe)
646 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
647 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
648 ("Unlocked pipe passed to pipeunlock"));
649 KASSERT(cpipe->pipe_waiters >= 0,
650 ("%s: bad waiter count %d", __func__,
651 cpipe->pipe_waiters));
652 cpipe->pipe_state &= ~PIPE_LOCKFL;
653 if (cpipe->pipe_waiters > 0) {
659 pipeselwakeup(struct pipe *cpipe)
662 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
663 if (cpipe->pipe_state & PIPE_SEL) {
664 selwakeuppri(&cpipe->pipe_sel, PSOCK);
665 if (!SEL_WAITING(&cpipe->pipe_sel))
666 cpipe->pipe_state &= ~PIPE_SEL;
668 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
669 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
670 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
674 * Initialize and allocate VM and memory for pipe. The structure
675 * will start out zero'd from the ctor, so we just manage the kmem.
678 pipe_create(struct pipe *pipe, bool large_backing)
682 error = pipespace_new(pipe, !large_backing || amountpipekva >
683 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
685 pipe->pipe_ino = alloc_unr64(&pipeino_unr);
691 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
692 int flags, struct thread *td)
702 error = pipelock(rpipe, 1);
707 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
711 if (amountpipekva > (3 * maxpipekva) / 4) {
712 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
713 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
714 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
715 piperesizeallowed == 1) {
717 pipespace(rpipe, SMALL_PIPE_SIZE);
722 while (uio->uio_resid) {
724 * normal pipe buffer receive
726 if (rpipe->pipe_buffer.cnt > 0) {
727 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
728 if (size > rpipe->pipe_buffer.cnt)
729 size = rpipe->pipe_buffer.cnt;
730 if (size > uio->uio_resid)
731 size = uio->uio_resid;
735 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
741 rpipe->pipe_buffer.out += size;
742 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
743 rpipe->pipe_buffer.out = 0;
745 rpipe->pipe_buffer.cnt -= size;
748 * If there is no more to read in the pipe, reset
749 * its pointers to the beginning. This improves
752 if (rpipe->pipe_buffer.cnt == 0) {
753 rpipe->pipe_buffer.in = 0;
754 rpipe->pipe_buffer.out = 0;
757 #ifndef PIPE_NODIRECT
759 * Direct copy, bypassing a kernel buffer.
761 } else if ((size = rpipe->pipe_pages.cnt) != 0) {
762 if (size > uio->uio_resid)
763 size = (u_int) uio->uio_resid;
765 error = uiomove_fromphys(rpipe->pipe_pages.ms,
766 rpipe->pipe_pages.pos, size, uio);
771 rpipe->pipe_pages.pos += size;
772 rpipe->pipe_pages.cnt -= size;
773 if (rpipe->pipe_pages.cnt == 0) {
774 rpipe->pipe_state &= ~PIPE_WANTW;
780 * detect EOF condition
781 * read returns 0 on EOF, no need to set error
783 if (rpipe->pipe_state & PIPE_EOF)
787 * If the "write-side" has been blocked, wake it up now.
789 if (rpipe->pipe_state & PIPE_WANTW) {
790 rpipe->pipe_state &= ~PIPE_WANTW;
795 * Break if some data was read.
801 * Unlock the pipe buffer for our remaining processing.
802 * We will either break out with an error or we will
803 * sleep and relock to loop.
808 * Handle non-blocking mode operation or
809 * wait for more data.
811 if (fp->f_flag & FNONBLOCK) {
814 rpipe->pipe_state |= PIPE_WANTR;
815 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
818 error = pipelock(rpipe, 1);
829 /* XXX: should probably do this before getting any locks. */
831 pipe_timestamp(&rpipe->pipe_atime);
836 * PIPE_WANT processing only makes sense if pipe_busy is 0.
838 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
839 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
841 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
843 * Handle write blocking hysteresis.
845 if (rpipe->pipe_state & PIPE_WANTW) {
846 rpipe->pipe_state &= ~PIPE_WANTW;
852 * Only wake up writers if there was actually something read.
853 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
856 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
857 pipeselwakeup(rpipe);
861 td->td_ru.ru_msgrcv++;
865 #ifndef PIPE_NODIRECT
867 * Map the sending processes' buffer into kernel space and wire it.
868 * This is similar to a physical write operation.
871 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
876 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
877 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
878 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
879 KASSERT(wpipe->pipe_pages.cnt == 0,
880 ("%s: pipe map for %p contains residual data", __func__, wpipe));
882 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
883 size = wpipe->pipe_buffer.size;
885 size = uio->uio_iov->iov_len;
887 wpipe->pipe_state |= PIPE_DIRECTW;
889 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
890 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
891 wpipe->pipe_pages.ms, PIPENPAGES);
894 wpipe->pipe_state &= ~PIPE_DIRECTW;
898 wpipe->pipe_pages.npages = i;
899 wpipe->pipe_pages.pos =
900 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
901 wpipe->pipe_pages.cnt = size;
903 uio->uio_iov->iov_len -= size;
904 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
905 if (uio->uio_iov->iov_len == 0)
907 uio->uio_resid -= size;
908 uio->uio_offset += size;
913 * Unwire the process buffer.
916 pipe_destroy_write_buffer(struct pipe *wpipe)
919 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
920 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
921 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
922 KASSERT(wpipe->pipe_pages.cnt == 0,
923 ("%s: pipe map for %p contains residual data", __func__, wpipe));
925 wpipe->pipe_state &= ~PIPE_DIRECTW;
926 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
927 wpipe->pipe_pages.npages = 0;
931 * In the case of a signal, the writing process might go away. This
932 * code copies the data into the circular buffer so that the source
933 * pages can be freed without loss of data.
936 pipe_clone_write_buffer(struct pipe *wpipe)
943 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
944 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
945 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
947 size = wpipe->pipe_pages.cnt;
948 pos = wpipe->pipe_pages.pos;
949 wpipe->pipe_pages.cnt = 0;
951 wpipe->pipe_buffer.in = size;
952 wpipe->pipe_buffer.out = 0;
953 wpipe->pipe_buffer.cnt = size;
956 iov.iov_base = wpipe->pipe_buffer.buffer;
961 uio.uio_resid = size;
962 uio.uio_segflg = UIO_SYSSPACE;
963 uio.uio_rw = UIO_READ;
964 uio.uio_td = curthread;
965 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
967 pipe_destroy_write_buffer(wpipe);
971 * This implements the pipe buffer write mechanism. Note that only
972 * a direct write OR a normal pipe write can be pending at any given time.
973 * If there are any characters in the pipe buffer, the direct write will
974 * be deferred until the receiving process grabs all of the bytes from
975 * the pipe buffer. Then the direct mapping write is set-up.
978 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
983 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
984 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
988 if (wpipe->pipe_state & PIPE_DIRECTW) {
989 if (wpipe->pipe_state & PIPE_WANTR) {
990 wpipe->pipe_state &= ~PIPE_WANTR;
993 pipeselwakeup(wpipe);
994 wpipe->pipe_state |= PIPE_WANTW;
996 error = msleep(wpipe, PIPE_MTX(wpipe),
997 PRIBIO | PCATCH, "pipdww", 0);
1003 if (wpipe->pipe_buffer.cnt > 0) {
1004 if (wpipe->pipe_state & PIPE_WANTR) {
1005 wpipe->pipe_state &= ~PIPE_WANTR;
1008 pipeselwakeup(wpipe);
1009 wpipe->pipe_state |= PIPE_WANTW;
1011 error = msleep(wpipe, PIPE_MTX(wpipe),
1012 PRIBIO | PCATCH, "pipdwc", 0);
1019 error = pipe_build_write_buffer(wpipe, uio);
1024 while (wpipe->pipe_pages.cnt != 0 &&
1025 (wpipe->pipe_state & PIPE_EOF) == 0) {
1026 if (wpipe->pipe_state & PIPE_WANTR) {
1027 wpipe->pipe_state &= ~PIPE_WANTR;
1030 pipeselwakeup(wpipe);
1031 wpipe->pipe_state |= PIPE_WANTW;
1033 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1040 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1041 wpipe->pipe_pages.cnt = 0;
1042 pipe_destroy_write_buffer(wpipe);
1043 pipeselwakeup(wpipe);
1045 } else if (error == EINTR || error == ERESTART) {
1046 pipe_clone_write_buffer(wpipe);
1048 pipe_destroy_write_buffer(wpipe);
1050 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1051 ("pipe %p leaked PIPE_DIRECTW", wpipe));
1061 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1062 int flags, struct thread *td)
1064 struct pipe *wpipe, *rpipe;
1066 int desiredsize, error;
1069 wpipe = PIPE_PEER(rpipe);
1071 error = pipelock(wpipe, 1);
1077 * detect loss of pipe read side, issue SIGPIPE if lost.
1079 if (wpipe->pipe_present != PIPE_ACTIVE ||
1080 (wpipe->pipe_state & PIPE_EOF)) {
1086 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1095 /* Choose a larger size if it's advantageous */
1096 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1097 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1098 if (piperesizeallowed != 1)
1100 if (amountpipekva > maxpipekva / 2)
1102 if (desiredsize == BIG_PIPE_SIZE)
1104 desiredsize = desiredsize * 2;
1107 /* Choose a smaller size if we're in a OOM situation */
1108 if (amountpipekva > (3 * maxpipekva) / 4 &&
1109 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1110 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1111 piperesizeallowed == 1)
1112 desiredsize = SMALL_PIPE_SIZE;
1114 /* Resize if the above determined that a new size was necessary */
1115 if (desiredsize != wpipe->pipe_buffer.size &&
1116 (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1118 pipespace(wpipe, desiredsize);
1121 MPASS(wpipe->pipe_buffer.size != 0);
1123 orig_resid = uio->uio_resid;
1125 while (uio->uio_resid) {
1128 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) {
1146 error = pipe_direct_write(wpipe, uio);
1154 * Pipe buffered writes cannot be coincidental with
1155 * direct writes. We wait until the currently executing
1156 * direct write is completed before we start filling the
1157 * pipe buffer. We break out if a signal occurs or the
1160 if (wpipe->pipe_pages.cnt != 0) {
1161 if (wpipe->pipe_state & PIPE_WANTR) {
1162 wpipe->pipe_state &= ~PIPE_WANTR;
1165 pipeselwakeup(wpipe);
1166 wpipe->pipe_state |= PIPE_WANTW;
1168 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1176 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1178 /* Writes of size <= PIPE_BUF must be atomic. */
1179 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1183 int size; /* Transfer size */
1184 int segsize; /* first segment to transfer */
1187 * Transfer size is minimum of uio transfer
1188 * and free space in pipe buffer.
1190 if (space > uio->uio_resid)
1191 size = uio->uio_resid;
1195 * First segment to transfer is minimum of
1196 * transfer size and contiguous space in
1197 * pipe buffer. If first segment to transfer
1198 * is less than the transfer size, we've got
1199 * a wraparound in the buffer.
1201 segsize = wpipe->pipe_buffer.size -
1202 wpipe->pipe_buffer.in;
1206 /* Transfer first segment */
1209 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1213 if (error == 0 && segsize < size) {
1214 KASSERT(wpipe->pipe_buffer.in + segsize ==
1215 wpipe->pipe_buffer.size,
1216 ("Pipe buffer wraparound disappeared"));
1218 * Transfer remaining part now, to
1219 * support atomic writes. Wraparound
1225 &wpipe->pipe_buffer.buffer[0],
1226 size - segsize, uio);
1230 wpipe->pipe_buffer.in += size;
1231 if (wpipe->pipe_buffer.in >=
1232 wpipe->pipe_buffer.size) {
1233 KASSERT(wpipe->pipe_buffer.in ==
1235 wpipe->pipe_buffer.size,
1236 ("Expected wraparound bad"));
1237 wpipe->pipe_buffer.in = size - segsize;
1240 wpipe->pipe_buffer.cnt += size;
1241 KASSERT(wpipe->pipe_buffer.cnt <=
1242 wpipe->pipe_buffer.size,
1243 ("Pipe buffer overflow"));
1250 * If the "read-side" has been blocked, wake it up now.
1252 if (wpipe->pipe_state & PIPE_WANTR) {
1253 wpipe->pipe_state &= ~PIPE_WANTR;
1258 * don't block on non-blocking I/O
1260 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);
1284 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1285 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1287 } else if (wpipe->pipe_buffer.cnt > 0) {
1289 * If we have put any characters in the buffer, we wake up
1292 if (wpipe->pipe_state & PIPE_WANTR) {
1293 wpipe->pipe_state &= ~PIPE_WANTR;
1299 * Don't return EPIPE if any byte was written.
1300 * EINTR and other interrupts are handled by generic I/O layer.
1301 * Do not pretend that I/O succeeded for obvious user error
1304 if (uio->uio_resid != orig_resid && error == EPIPE)
1308 pipe_timestamp(&wpipe->pipe_mtime);
1311 * We have something to offer,
1312 * wake up select/poll.
1314 if (wpipe->pipe_buffer.cnt)
1315 pipeselwakeup(wpipe);
1319 if (uio->uio_resid != orig_resid)
1320 td->td_ru.ru_msgsnd++;
1326 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1333 if (cpipe->pipe_type & PIPE_TYPE_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(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1347 struct pipe *mpipe = fp->f_data;
1353 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1367 mpipe->pipe_state |= PIPE_ASYNC;
1369 mpipe->pipe_state &= ~PIPE_ASYNC;
1374 if (!(fp->f_flag & FREAD)) {
1379 if (mpipe->pipe_pages.cnt != 0)
1380 *(int *)data = mpipe->pipe_pages.cnt;
1382 *(int *)data = mpipe->pipe_buffer.cnt;
1387 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1391 *(int *)data = fgetown(&mpipe->pipe_sigio);
1394 /* This is deprecated, FIOSETOWN should be used instead. */
1397 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1400 /* This is deprecated, FIOGETOWN should be used instead. */
1402 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1415 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1420 int levents, revents;
1427 wpipe = PIPE_PEER(rpipe);
1430 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1434 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1435 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1436 revents |= events & (POLLIN | POLLRDNORM);
1438 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1439 if (wpipe->pipe_present != PIPE_ACTIVE ||
1440 (wpipe->pipe_state & PIPE_EOF) ||
1441 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1442 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1443 wpipe->pipe_buffer.size == 0)))
1444 revents |= events & (POLLOUT | POLLWRNORM);
1447 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1448 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1449 fp->f_pipegen == rpipe->pipe_wgen)
1450 events |= POLLINIGNEOF;
1452 if ((events & POLLINIGNEOF) == 0) {
1453 if (rpipe->pipe_state & PIPE_EOF) {
1454 if (fp->f_flag & FREAD)
1455 revents |= (events & (POLLIN | POLLRDNORM));
1456 if (wpipe->pipe_present != PIPE_ACTIVE ||
1457 (wpipe->pipe_state & PIPE_EOF))
1464 * Add ourselves regardless of eventmask as we have to return
1465 * POLLHUP even if it was not asked for.
1467 if ((fp->f_flag & FREAD) != 0) {
1468 selrecord(td, &rpipe->pipe_sel);
1469 if (SEL_WAITING(&rpipe->pipe_sel))
1470 rpipe->pipe_state |= PIPE_SEL;
1473 if ((fp->f_flag & FWRITE) != 0 &&
1474 wpipe->pipe_present == PIPE_ACTIVE) {
1475 selrecord(td, &wpipe->pipe_sel);
1476 if (SEL_WAITING(&wpipe->pipe_sel))
1477 wpipe->pipe_state |= PIPE_SEL;
1489 * We shouldn't need locks here as we're doing a read and this should
1490 * be a natural race.
1493 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1503 if (mac_pipe_check_stat_enabled()) {
1505 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1513 /* For named pipes ask the underlying filesystem. */
1514 if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1515 return (vnops.fo_stat(fp, ub, active_cred, td));
1518 bzero(ub, sizeof(*ub));
1519 ub->st_mode = S_IFIFO;
1520 ub->st_blksize = PAGE_SIZE;
1521 if (pipe->pipe_pages.cnt != 0)
1522 ub->st_size = pipe->pipe_pages.cnt;
1524 ub->st_size = pipe->pipe_buffer.cnt;
1525 ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1526 ub->st_atim = pipe->pipe_atime;
1527 ub->st_mtim = pipe->pipe_mtime;
1528 ub->st_ctim = pipe->pipe_ctime;
1529 ub->st_uid = fp->f_cred->cr_uid;
1530 ub->st_gid = fp->f_cred->cr_gid;
1531 ub->st_dev = pipedev_ino;
1532 ub->st_ino = pipe->pipe_ino;
1534 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1541 pipe_close(struct file *fp, struct thread *td)
1544 if (fp->f_vnode != NULL)
1545 return vnops.fo_close(fp, td);
1546 fp->f_ops = &badfileops;
1547 pipe_dtor(fp->f_data);
1553 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1559 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1560 error = vn_chmod(fp, mode, active_cred, td);
1562 error = invfo_chmod(fp, mode, active_cred, td);
1567 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1574 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1575 error = vn_chown(fp, uid, gid, active_cred, td);
1577 error = invfo_chown(fp, uid, gid, active_cred, td);
1582 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1586 if (fp->f_type == DTYPE_FIFO)
1587 return (vn_fill_kinfo(fp, kif, fdp));
1588 kif->kf_type = KF_TYPE_PIPE;
1590 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1591 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1592 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1597 pipe_free_kmem(struct pipe *cpipe)
1600 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1601 ("pipe_free_kmem: pipe mutex locked"));
1603 if (cpipe->pipe_buffer.buffer != NULL) {
1604 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1605 vm_map_remove(pipe_map,
1606 (vm_offset_t)cpipe->pipe_buffer.buffer,
1607 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1608 cpipe->pipe_buffer.buffer = NULL;
1610 #ifndef PIPE_NODIRECT
1612 cpipe->pipe_pages.cnt = 0;
1613 cpipe->pipe_pages.pos = 0;
1614 cpipe->pipe_pages.npages = 0;
1623 pipeclose(struct pipe *cpipe)
1625 struct pipepair *pp;
1628 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1632 pp = cpipe->pipe_pair;
1635 * If the other side is blocked, wake it up saying that
1636 * we want to close it down.
1638 cpipe->pipe_state |= PIPE_EOF;
1639 while (cpipe->pipe_busy) {
1641 cpipe->pipe_state |= PIPE_WANT;
1643 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1647 pipeselwakeup(cpipe);
1650 * Disconnect from peer, if any.
1652 ppipe = cpipe->pipe_peer;
1653 if (ppipe->pipe_present == PIPE_ACTIVE) {
1654 ppipe->pipe_state |= PIPE_EOF;
1656 pipeselwakeup(ppipe);
1660 * Mark this endpoint as free. Release kmem resources. We
1661 * don't mark this endpoint as unused until we've finished
1662 * doing that, or the pipe might disappear out from under
1666 pipe_free_kmem(cpipe);
1668 cpipe->pipe_present = PIPE_CLOSING;
1672 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1673 * PIPE_FINALIZED, that allows other end to free the
1674 * pipe_pair, only after the knotes are completely dismantled.
1676 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1677 cpipe->pipe_present = PIPE_FINALIZED;
1678 seldrain(&cpipe->pipe_sel);
1679 knlist_destroy(&cpipe->pipe_sel.si_note);
1682 * If both endpoints are now closed, release the memory for the
1683 * pipe pair. If not, unlock.
1685 if (ppipe->pipe_present == PIPE_FINALIZED) {
1688 mac_pipe_destroy(pp);
1690 uma_zfree(pipe_zone, cpipe->pipe_pair);
1697 pipe_kqfilter(struct file *fp, struct knote *kn)
1702 * If a filter is requested that is not supported by this file
1703 * descriptor, don't return an error, but also don't ever generate an
1706 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1707 kn->kn_fop = &pipe_nfiltops;
1710 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1711 kn->kn_fop = &pipe_nfiltops;
1716 switch (kn->kn_filter) {
1718 kn->kn_fop = &pipe_rfiltops;
1721 kn->kn_fop = &pipe_wfiltops;
1722 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1723 /* other end of pipe has been closed */
1727 cpipe = PIPE_PEER(cpipe);
1734 kn->kn_hook = cpipe;
1735 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1741 filt_pipedetach(struct knote *kn)
1743 struct pipe *cpipe = kn->kn_hook;
1746 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1752 filt_piperead(struct knote *kn, long hint)
1754 struct file *fp = kn->kn_fp;
1755 struct pipe *rpipe = kn->kn_hook;
1757 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1758 kn->kn_data = rpipe->pipe_buffer.cnt;
1759 if (kn->kn_data == 0)
1760 kn->kn_data = rpipe->pipe_pages.cnt;
1762 if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1763 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1764 fp->f_pipegen != rpipe->pipe_wgen)) {
1765 kn->kn_flags |= EV_EOF;
1768 kn->kn_flags &= ~EV_EOF;
1769 return (kn->kn_data > 0);
1774 filt_pipewrite(struct knote *kn, long hint)
1776 struct pipe *wpipe = kn->kn_hook;
1779 * If this end of the pipe is closed, the knote was removed from the
1780 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1782 if (wpipe->pipe_present == PIPE_ACTIVE ||
1783 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1784 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1786 if (wpipe->pipe_state & PIPE_DIRECTW) {
1788 } else if (wpipe->pipe_buffer.size > 0) {
1789 kn->kn_data = wpipe->pipe_buffer.size -
1790 wpipe->pipe_buffer.cnt;
1792 kn->kn_data = PIPE_BUF;
1796 if (wpipe->pipe_present != PIPE_ACTIVE ||
1797 (wpipe->pipe_state & PIPE_EOF)) {
1798 kn->kn_flags |= EV_EOF;
1801 kn->kn_flags &= ~EV_EOF;
1802 return (kn->kn_data >= PIPE_BUF);
1806 filt_pipedetach_notsup(struct knote *kn)
1812 filt_pipenotsup(struct knote *kn, long hint)