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;
212 static long pipe_mindirect = PIPE_MINDIRECT;
214 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
215 &maxpipekva, 0, "Pipe KVA limit");
216 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
217 &amountpipekva, 0, "Pipe KVA usage");
218 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
219 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
220 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
221 &pipeallocfail, 0, "Pipe allocation failures");
222 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
223 &piperesizefail, 0, "Pipe resize failures");
224 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
225 &piperesizeallowed, 0, "Pipe resizing allowed");
227 static void pipeinit(void *dummy __unused);
228 static void pipeclose(struct pipe *cpipe);
229 static void pipe_free_kmem(struct pipe *cpipe);
230 static int pipe_create(struct pipe *pipe, bool backing);
231 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
232 static __inline int pipelock(struct pipe *cpipe, int catch);
233 static __inline void pipeunlock(struct pipe *cpipe);
234 static void pipe_timestamp(struct timespec *tsp);
235 #ifndef PIPE_NODIRECT
236 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
237 static void pipe_destroy_write_buffer(struct pipe *wpipe);
238 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
239 static void pipe_clone_write_buffer(struct pipe *wpipe);
241 static int pipespace(struct pipe *cpipe, int size);
242 static int pipespace_new(struct pipe *cpipe, int size);
244 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
245 static int pipe_zone_init(void *mem, int size, int flags);
246 static void pipe_zone_fini(void *mem, int size);
248 static uma_zone_t pipe_zone;
249 static struct unrhdr64 pipeino_unr;
250 static dev_t pipedev_ino;
252 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
255 pipeinit(void *dummy __unused)
258 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
259 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
261 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
262 new_unrhdr64(&pipeino_unr, 1);
263 pipedev_ino = devfs_alloc_cdp_inode();
264 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
268 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS)
271 long tmp_pipe_mindirect = pipe_mindirect;
273 error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req);
274 if (error != 0 || req->newptr == NULL)
278 * Don't allow pipe_mindirect to be set so low that we violate
279 * atomicity requirements.
281 if (tmp_pipe_mindirect <= PIPE_BUF)
283 pipe_mindirect = tmp_pipe_mindirect;
286 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW,
287 &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L",
288 "Minimum write size triggering VM optimization");
291 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
294 struct pipe *rpipe, *wpipe;
296 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
298 pp = (struct pipepair *)mem;
301 * We zero both pipe endpoints to make sure all the kmem pointers
302 * are NULL, flag fields are zero'd, etc. We timestamp both
303 * endpoints with the same time.
305 rpipe = &pp->pp_rpipe;
306 bzero(rpipe, sizeof(*rpipe));
307 pipe_timestamp(&rpipe->pipe_ctime);
308 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
310 wpipe = &pp->pp_wpipe;
311 bzero(wpipe, sizeof(*wpipe));
312 wpipe->pipe_ctime = rpipe->pipe_ctime;
313 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
315 rpipe->pipe_peer = wpipe;
316 rpipe->pipe_pair = pp;
317 wpipe->pipe_peer = rpipe;
318 wpipe->pipe_pair = pp;
321 * Mark both endpoints as present; they will later get free'd
322 * one at a time. When both are free'd, then the whole pair
325 rpipe->pipe_present = PIPE_ACTIVE;
326 wpipe->pipe_present = PIPE_ACTIVE;
329 * Eventually, the MAC Framework may initialize the label
330 * in ctor or init, but for now we do it elswhere to avoid
331 * blocking in ctor or init.
339 pipe_zone_init(void *mem, int size, int flags)
343 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
345 pp = (struct pipepair *)mem;
347 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
352 pipe_zone_fini(void *mem, int size)
356 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
358 pp = (struct pipepair *)mem;
360 mtx_destroy(&pp->pp_mtx);
364 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
367 struct pipe *rpipe, *wpipe;
370 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
373 * The MAC label is shared between the connected endpoints. As a
374 * result mac_pipe_init() and mac_pipe_create() are called once
375 * for the pair, and not on the endpoints.
378 mac_pipe_create(td->td_ucred, pp);
380 rpipe = &pp->pp_rpipe;
381 wpipe = &pp->pp_wpipe;
383 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
384 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
387 * Only the forward direction pipe is backed by big buffer by
390 error = pipe_create(rpipe, true);
393 error = pipe_create(wpipe, false);
396 * This cleanup leaves the pipe inode number for rpipe
397 * still allocated, but never used. We do not free
398 * inode numbers for opened pipes, which is required
399 * for correctness because numbers must be unique.
400 * But also it avoids any memory use by the unr
401 * allocator, so stashing away the transient inode
402 * number is reasonable.
404 pipe_free_kmem(rpipe);
408 rpipe->pipe_state |= PIPE_DIRECTOK;
409 wpipe->pipe_state |= PIPE_DIRECTOK;
413 knlist_destroy(&rpipe->pipe_sel.si_note);
414 knlist_destroy(&wpipe->pipe_sel.si_note);
416 mac_pipe_destroy(pp);
418 uma_zfree(pipe_zone, pp);
423 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
428 error = pipe_paircreate(td, &pp);
431 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
432 *ppipe = &pp->pp_rpipe;
437 pipe_dtor(struct pipe *dpipe)
441 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
442 funsetown(&dpipe->pipe_sigio);
445 funsetown(&peer->pipe_sigio);
453 * This used to be vfs_timestamp but the higher precision is unnecessary and
454 * can very negatively affect performance in virtualized environments (e.g., on
455 * vms running on amd64 when using the rdtscp instruction).
458 pipe_timestamp(struct timespec *tsp)
465 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
466 * the zone pick up the pieces via pipeclose().
469 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
470 struct filecaps *fcaps2)
472 struct file *rf, *wf;
473 struct pipe *rpipe, *wpipe;
475 int fd, fflags, error;
477 error = pipe_paircreate(td, &pp);
480 rpipe = &pp->pp_rpipe;
481 wpipe = &pp->pp_wpipe;
482 error = falloc_caps(td, &rf, &fd, flags, fcaps1);
488 /* An extra reference on `rf' has been held for us by falloc_caps(). */
491 fflags = FREAD | FWRITE;
492 if ((flags & O_NONBLOCK) != 0)
496 * Warning: once we've gotten past allocation of the fd for the
497 * read-side, we can only drop the read side via fdrop() in order
498 * to avoid races against processes which manage to dup() the read
499 * side while we are blocked trying to allocate the write side.
501 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
502 error = falloc_caps(td, &wf, &fd, flags, fcaps2);
504 fdclose(td, rf, fildes[0]);
506 /* rpipe has been closed by fdrop(). */
510 /* An extra reference on `wf' has been held for us by falloc_caps(). */
511 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
519 #ifdef COMPAT_FREEBSD10
522 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
527 error = kern_pipe(td, fildes, 0, NULL, NULL);
531 td->td_retval[0] = fildes[0];
532 td->td_retval[1] = fildes[1];
539 sys_pipe2(struct thread *td, struct pipe2_args *uap)
541 int error, fildes[2];
543 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
545 error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
548 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
550 (void)kern_close(td, fildes[0]);
551 (void)kern_close(td, fildes[1]);
557 * Allocate kva for pipe circular buffer, the space is pageable
558 * This routine will 'realloc' the size of a pipe safely, if it fails
559 * it will retain the old buffer.
560 * If it fails it will return ENOMEM.
563 pipespace_new(struct pipe *cpipe, int size)
566 int error, cnt, firstseg;
567 static int curfail = 0;
568 static struct timeval lastfail;
570 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
571 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
572 ("pipespace: resize of direct writes not allowed"));
574 cnt = cpipe->pipe_buffer.cnt;
578 size = round_page(size);
579 buffer = (caddr_t) vm_map_min(pipe_map);
581 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
582 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
583 if (error != KERN_SUCCESS) {
584 if (cpipe->pipe_buffer.buffer == NULL &&
585 size > SMALL_PIPE_SIZE) {
586 size = SMALL_PIPE_SIZE;
590 if (cpipe->pipe_buffer.buffer == NULL) {
592 if (ppsratecheck(&lastfail, &curfail, 1))
593 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
600 /* copy data, then free old resources if we're resizing */
602 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
603 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
604 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
606 if ((cnt - firstseg) > 0)
607 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
608 cpipe->pipe_buffer.in);
610 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
614 pipe_free_kmem(cpipe);
615 cpipe->pipe_buffer.buffer = buffer;
616 cpipe->pipe_buffer.size = size;
617 cpipe->pipe_buffer.in = cnt;
618 cpipe->pipe_buffer.out = 0;
619 cpipe->pipe_buffer.cnt = cnt;
620 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
625 * Wrapper for pipespace_new() that performs locking assertions.
628 pipespace(struct pipe *cpipe, int size)
631 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
632 ("Unlocked pipe passed to pipespace"));
633 return (pipespace_new(cpipe, size));
637 * lock a pipe for I/O, blocking other access
640 pipelock(struct pipe *cpipe, int catch)
644 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
649 while (cpipe->pipe_state & PIPE_LOCKFL) {
650 KASSERT(cpipe->pipe_waiters >= 0,
651 ("%s: bad waiter count %d", __func__,
652 cpipe->pipe_waiters));
653 cpipe->pipe_waiters++;
654 error = msleep(cpipe, PIPE_MTX(cpipe),
656 cpipe->pipe_waiters--;
660 cpipe->pipe_state |= PIPE_LOCKFL;
665 * unlock a pipe I/O lock
668 pipeunlock(struct pipe *cpipe)
671 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
672 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
673 ("Unlocked pipe passed to pipeunlock"));
674 KASSERT(cpipe->pipe_waiters >= 0,
675 ("%s: bad waiter count %d", __func__,
676 cpipe->pipe_waiters));
677 cpipe->pipe_state &= ~PIPE_LOCKFL;
678 if (cpipe->pipe_waiters > 0) {
684 pipeselwakeup(struct pipe *cpipe)
687 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
688 if (cpipe->pipe_state & PIPE_SEL) {
689 selwakeuppri(&cpipe->pipe_sel, PSOCK);
690 if (!SEL_WAITING(&cpipe->pipe_sel))
691 cpipe->pipe_state &= ~PIPE_SEL;
693 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
694 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
695 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
699 * Initialize and allocate VM and memory for pipe. The structure
700 * will start out zero'd from the ctor, so we just manage the kmem.
703 pipe_create(struct pipe *pipe, bool large_backing)
707 error = pipespace_new(pipe, !large_backing || amountpipekva >
708 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
710 pipe->pipe_ino = alloc_unr64(&pipeino_unr);
716 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
717 int flags, struct thread *td)
727 error = pipelock(rpipe, 1);
732 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
736 if (amountpipekva > (3 * maxpipekva) / 4) {
737 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
738 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
739 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
740 piperesizeallowed == 1) {
742 pipespace(rpipe, SMALL_PIPE_SIZE);
747 while (uio->uio_resid) {
749 * normal pipe buffer receive
751 if (rpipe->pipe_buffer.cnt > 0) {
752 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
753 if (size > rpipe->pipe_buffer.cnt)
754 size = rpipe->pipe_buffer.cnt;
755 if (size > uio->uio_resid)
756 size = uio->uio_resid;
760 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
766 rpipe->pipe_buffer.out += size;
767 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
768 rpipe->pipe_buffer.out = 0;
770 rpipe->pipe_buffer.cnt -= size;
773 * If there is no more to read in the pipe, reset
774 * its pointers to the beginning. This improves
777 if (rpipe->pipe_buffer.cnt == 0) {
778 rpipe->pipe_buffer.in = 0;
779 rpipe->pipe_buffer.out = 0;
782 #ifndef PIPE_NODIRECT
784 * Direct copy, bypassing a kernel buffer.
786 } else if ((size = rpipe->pipe_pages.cnt) != 0) {
787 if (size > uio->uio_resid)
788 size = (u_int) uio->uio_resid;
790 error = uiomove_fromphys(rpipe->pipe_pages.ms,
791 rpipe->pipe_pages.pos, size, uio);
796 rpipe->pipe_pages.pos += size;
797 rpipe->pipe_pages.cnt -= size;
798 if (rpipe->pipe_pages.cnt == 0) {
799 rpipe->pipe_state &= ~PIPE_WANTW;
805 * detect EOF condition
806 * read returns 0 on EOF, no need to set error
808 if (rpipe->pipe_state & PIPE_EOF)
812 * If the "write-side" has been blocked, wake it up now.
814 if (rpipe->pipe_state & PIPE_WANTW) {
815 rpipe->pipe_state &= ~PIPE_WANTW;
820 * Break if some data was read.
826 * Unlock the pipe buffer for our remaining processing.
827 * We will either break out with an error or we will
828 * sleep and relock to loop.
833 * Handle non-blocking mode operation or
834 * wait for more data.
836 if (fp->f_flag & FNONBLOCK) {
839 rpipe->pipe_state |= PIPE_WANTR;
840 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
843 error = pipelock(rpipe, 1);
854 /* XXX: should probably do this before getting any locks. */
856 pipe_timestamp(&rpipe->pipe_atime);
861 * PIPE_WANT processing only makes sense if pipe_busy is 0.
863 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
864 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
866 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
868 * Handle write blocking hysteresis.
870 if (rpipe->pipe_state & PIPE_WANTW) {
871 rpipe->pipe_state &= ~PIPE_WANTW;
877 * Only wake up writers if there was actually something read.
878 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
881 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
882 pipeselwakeup(rpipe);
886 td->td_ru.ru_msgrcv++;
890 #ifndef PIPE_NODIRECT
892 * Map the sending processes' buffer into kernel space and wire it.
893 * This is similar to a physical write operation.
896 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
901 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
902 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
903 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
904 KASSERT(wpipe->pipe_pages.cnt == 0,
905 ("%s: pipe map for %p contains residual data", __func__, wpipe));
907 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
908 size = wpipe->pipe_buffer.size;
910 size = uio->uio_iov->iov_len;
912 wpipe->pipe_state |= PIPE_DIRECTW;
914 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
915 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
916 wpipe->pipe_pages.ms, PIPENPAGES);
919 wpipe->pipe_state &= ~PIPE_DIRECTW;
923 wpipe->pipe_pages.npages = i;
924 wpipe->pipe_pages.pos =
925 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
926 wpipe->pipe_pages.cnt = size;
928 uio->uio_iov->iov_len -= size;
929 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
930 if (uio->uio_iov->iov_len == 0)
932 uio->uio_resid -= size;
933 uio->uio_offset += size;
938 * Unwire the process buffer.
941 pipe_destroy_write_buffer(struct pipe *wpipe)
944 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
945 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
946 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
947 KASSERT(wpipe->pipe_pages.cnt == 0,
948 ("%s: pipe map for %p contains residual data", __func__, wpipe));
950 wpipe->pipe_state &= ~PIPE_DIRECTW;
951 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
952 wpipe->pipe_pages.npages = 0;
956 * In the case of a signal, the writing process might go away. This
957 * code copies the data into the circular buffer so that the source
958 * pages can be freed without loss of data.
961 pipe_clone_write_buffer(struct pipe *wpipe)
968 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
969 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
970 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
972 size = wpipe->pipe_pages.cnt;
973 pos = wpipe->pipe_pages.pos;
974 wpipe->pipe_pages.cnt = 0;
976 wpipe->pipe_buffer.in = size;
977 wpipe->pipe_buffer.out = 0;
978 wpipe->pipe_buffer.cnt = size;
981 iov.iov_base = wpipe->pipe_buffer.buffer;
986 uio.uio_resid = size;
987 uio.uio_segflg = UIO_SYSSPACE;
988 uio.uio_rw = UIO_READ;
989 uio.uio_td = curthread;
990 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
992 pipe_destroy_write_buffer(wpipe);
996 * This implements the pipe buffer write mechanism. Note that only
997 * a direct write OR a normal pipe write can be pending at any given time.
998 * If there are any characters in the pipe buffer, the direct write will
999 * be deferred until the receiving process grabs all of the bytes from
1000 * the pipe buffer. Then the direct mapping write is set-up.
1003 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
1008 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1009 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1013 if (wpipe->pipe_state & PIPE_DIRECTW) {
1014 if (wpipe->pipe_state & PIPE_WANTR) {
1015 wpipe->pipe_state &= ~PIPE_WANTR;
1018 pipeselwakeup(wpipe);
1019 wpipe->pipe_state |= PIPE_WANTW;
1021 error = msleep(wpipe, PIPE_MTX(wpipe),
1022 PRIBIO | PCATCH, "pipdww", 0);
1028 if (wpipe->pipe_buffer.cnt > 0) {
1029 if (wpipe->pipe_state & PIPE_WANTR) {
1030 wpipe->pipe_state &= ~PIPE_WANTR;
1033 pipeselwakeup(wpipe);
1034 wpipe->pipe_state |= PIPE_WANTW;
1036 error = msleep(wpipe, PIPE_MTX(wpipe),
1037 PRIBIO | PCATCH, "pipdwc", 0);
1044 error = pipe_build_write_buffer(wpipe, uio);
1049 while (wpipe->pipe_pages.cnt != 0 &&
1050 (wpipe->pipe_state & PIPE_EOF) == 0) {
1051 if (wpipe->pipe_state & PIPE_WANTR) {
1052 wpipe->pipe_state &= ~PIPE_WANTR;
1055 pipeselwakeup(wpipe);
1056 wpipe->pipe_state |= PIPE_WANTW;
1058 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1065 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1066 wpipe->pipe_pages.cnt = 0;
1067 pipe_destroy_write_buffer(wpipe);
1068 pipeselwakeup(wpipe);
1070 } else if (error == EINTR || error == ERESTART) {
1071 pipe_clone_write_buffer(wpipe);
1073 pipe_destroy_write_buffer(wpipe);
1075 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1076 ("pipe %p leaked PIPE_DIRECTW", wpipe));
1086 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1087 int flags, struct thread *td)
1089 struct pipe *wpipe, *rpipe;
1091 int desiredsize, error;
1094 wpipe = PIPE_PEER(rpipe);
1096 error = pipelock(wpipe, 1);
1102 * detect loss of pipe read side, issue SIGPIPE if lost.
1104 if (wpipe->pipe_present != PIPE_ACTIVE ||
1105 (wpipe->pipe_state & PIPE_EOF)) {
1111 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1120 /* Choose a larger size if it's advantageous */
1121 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1122 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1123 if (piperesizeallowed != 1)
1125 if (amountpipekva > maxpipekva / 2)
1127 if (desiredsize == BIG_PIPE_SIZE)
1129 desiredsize = desiredsize * 2;
1132 /* Choose a smaller size if we're in a OOM situation */
1133 if (amountpipekva > (3 * maxpipekva) / 4 &&
1134 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1135 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1136 piperesizeallowed == 1)
1137 desiredsize = SMALL_PIPE_SIZE;
1139 /* Resize if the above determined that a new size was necessary */
1140 if (desiredsize != wpipe->pipe_buffer.size &&
1141 (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1143 pipespace(wpipe, desiredsize);
1146 MPASS(wpipe->pipe_buffer.size != 0);
1148 orig_resid = uio->uio_resid;
1150 while (uio->uio_resid) {
1153 if (wpipe->pipe_state & PIPE_EOF) {
1157 #ifndef PIPE_NODIRECT
1159 * If the transfer is large, we can gain performance if
1160 * we do process-to-process copies directly.
1161 * If the write is non-blocking, we don't use the
1162 * direct write mechanism.
1164 * The direct write mechanism will detect the reader going
1167 if (uio->uio_segflg == UIO_USERSPACE &&
1168 uio->uio_iov->iov_len >= pipe_mindirect &&
1169 wpipe->pipe_buffer.size >= pipe_mindirect &&
1170 (fp->f_flag & FNONBLOCK) == 0) {
1171 error = pipe_direct_write(wpipe, uio);
1179 * Pipe buffered writes cannot be coincidental with
1180 * direct writes. We wait until the currently executing
1181 * direct write is completed before we start filling the
1182 * pipe buffer. We break out if a signal occurs or the
1185 if (wpipe->pipe_pages.cnt != 0) {
1186 if (wpipe->pipe_state & PIPE_WANTR) {
1187 wpipe->pipe_state &= ~PIPE_WANTR;
1190 pipeselwakeup(wpipe);
1191 wpipe->pipe_state |= PIPE_WANTW;
1193 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1201 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1203 /* Writes of size <= PIPE_BUF must be atomic. */
1204 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1208 int size; /* Transfer size */
1209 int segsize; /* first segment to transfer */
1212 * Transfer size is minimum of uio transfer
1213 * and free space in pipe buffer.
1215 if (space > uio->uio_resid)
1216 size = uio->uio_resid;
1220 * First segment to transfer is minimum of
1221 * transfer size and contiguous space in
1222 * pipe buffer. If first segment to transfer
1223 * is less than the transfer size, we've got
1224 * a wraparound in the buffer.
1226 segsize = wpipe->pipe_buffer.size -
1227 wpipe->pipe_buffer.in;
1231 /* Transfer first segment */
1234 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1238 if (error == 0 && segsize < size) {
1239 KASSERT(wpipe->pipe_buffer.in + segsize ==
1240 wpipe->pipe_buffer.size,
1241 ("Pipe buffer wraparound disappeared"));
1243 * Transfer remaining part now, to
1244 * support atomic writes. Wraparound
1250 &wpipe->pipe_buffer.buffer[0],
1251 size - segsize, uio);
1255 wpipe->pipe_buffer.in += size;
1256 if (wpipe->pipe_buffer.in >=
1257 wpipe->pipe_buffer.size) {
1258 KASSERT(wpipe->pipe_buffer.in ==
1260 wpipe->pipe_buffer.size,
1261 ("Expected wraparound bad"));
1262 wpipe->pipe_buffer.in = size - segsize;
1265 wpipe->pipe_buffer.cnt += size;
1266 KASSERT(wpipe->pipe_buffer.cnt <=
1267 wpipe->pipe_buffer.size,
1268 ("Pipe buffer overflow"));
1275 * If the "read-side" has been blocked, wake it up now.
1277 if (wpipe->pipe_state & PIPE_WANTR) {
1278 wpipe->pipe_state &= ~PIPE_WANTR;
1283 * don't block on non-blocking I/O
1285 if (fp->f_flag & FNONBLOCK) {
1291 * We have no more space and have something to offer,
1292 * wake up select/poll.
1294 pipeselwakeup(wpipe);
1296 wpipe->pipe_state |= PIPE_WANTW;
1298 error = msleep(wpipe, PIPE_MTX(rpipe),
1299 PRIBIO | PCATCH, "pipewr", 0);
1309 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1310 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1312 } else if (wpipe->pipe_buffer.cnt > 0) {
1314 * If we have put any characters in the buffer, we wake up
1317 if (wpipe->pipe_state & PIPE_WANTR) {
1318 wpipe->pipe_state &= ~PIPE_WANTR;
1324 * Don't return EPIPE if any byte was written.
1325 * EINTR and other interrupts are handled by generic I/O layer.
1326 * Do not pretend that I/O succeeded for obvious user error
1329 if (uio->uio_resid != orig_resid && error == EPIPE)
1333 pipe_timestamp(&wpipe->pipe_mtime);
1336 * We have something to offer,
1337 * wake up select/poll.
1339 if (wpipe->pipe_buffer.cnt)
1340 pipeselwakeup(wpipe);
1344 if (uio->uio_resid != orig_resid)
1345 td->td_ru.ru_msgsnd++;
1351 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1358 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1359 error = vnops.fo_truncate(fp, length, active_cred, td);
1361 error = invfo_truncate(fp, length, active_cred, td);
1366 * we implement a very minimal set of ioctls for compatibility with sockets.
1369 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1372 struct pipe *mpipe = fp->f_data;
1378 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1392 mpipe->pipe_state |= PIPE_ASYNC;
1394 mpipe->pipe_state &= ~PIPE_ASYNC;
1399 if (!(fp->f_flag & FREAD)) {
1404 if (mpipe->pipe_pages.cnt != 0)
1405 *(int *)data = mpipe->pipe_pages.cnt;
1407 *(int *)data = mpipe->pipe_buffer.cnt;
1412 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1416 *(int *)data = fgetown(&mpipe->pipe_sigio);
1419 /* This is deprecated, FIOSETOWN should be used instead. */
1422 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1425 /* This is deprecated, FIOGETOWN should be used instead. */
1427 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1440 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1445 int levents, revents;
1452 wpipe = PIPE_PEER(rpipe);
1455 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1459 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1460 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1461 revents |= events & (POLLIN | POLLRDNORM);
1463 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1464 if (wpipe->pipe_present != PIPE_ACTIVE ||
1465 (wpipe->pipe_state & PIPE_EOF) ||
1466 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1467 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1468 wpipe->pipe_buffer.size == 0)))
1469 revents |= events & (POLLOUT | POLLWRNORM);
1472 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1473 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1474 fp->f_pipegen == rpipe->pipe_wgen)
1475 events |= POLLINIGNEOF;
1477 if ((events & POLLINIGNEOF) == 0) {
1478 if (rpipe->pipe_state & PIPE_EOF) {
1479 if (fp->f_flag & FREAD)
1480 revents |= (events & (POLLIN | POLLRDNORM));
1481 if (wpipe->pipe_present != PIPE_ACTIVE ||
1482 (wpipe->pipe_state & PIPE_EOF))
1489 * Add ourselves regardless of eventmask as we have to return
1490 * POLLHUP even if it was not asked for.
1492 if ((fp->f_flag & FREAD) != 0) {
1493 selrecord(td, &rpipe->pipe_sel);
1494 if (SEL_WAITING(&rpipe->pipe_sel))
1495 rpipe->pipe_state |= PIPE_SEL;
1498 if ((fp->f_flag & FWRITE) != 0 &&
1499 wpipe->pipe_present == PIPE_ACTIVE) {
1500 selrecord(td, &wpipe->pipe_sel);
1501 if (SEL_WAITING(&wpipe->pipe_sel))
1502 wpipe->pipe_state |= PIPE_SEL;
1514 * We shouldn't need locks here as we're doing a read and this should
1515 * be a natural race.
1518 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred)
1527 if (mac_pipe_check_stat_enabled()) {
1529 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1537 /* For named pipes ask the underlying filesystem. */
1538 if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1539 return (vnops.fo_stat(fp, ub, active_cred));
1542 bzero(ub, sizeof(*ub));
1543 ub->st_mode = S_IFIFO;
1544 ub->st_blksize = PAGE_SIZE;
1545 if (pipe->pipe_pages.cnt != 0)
1546 ub->st_size = pipe->pipe_pages.cnt;
1548 ub->st_size = pipe->pipe_buffer.cnt;
1549 ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1550 ub->st_atim = pipe->pipe_atime;
1551 ub->st_mtim = pipe->pipe_mtime;
1552 ub->st_ctim = pipe->pipe_ctime;
1553 ub->st_uid = fp->f_cred->cr_uid;
1554 ub->st_gid = fp->f_cred->cr_gid;
1555 ub->st_dev = pipedev_ino;
1556 ub->st_ino = pipe->pipe_ino;
1558 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1565 pipe_close(struct file *fp, struct thread *td)
1568 if (fp->f_vnode != NULL)
1569 return vnops.fo_close(fp, td);
1570 fp->f_ops = &badfileops;
1571 pipe_dtor(fp->f_data);
1577 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1583 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1584 error = vn_chmod(fp, mode, active_cred, td);
1586 error = invfo_chmod(fp, mode, active_cred, td);
1591 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1598 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1599 error = vn_chown(fp, uid, gid, active_cred, td);
1601 error = invfo_chown(fp, uid, gid, active_cred, td);
1606 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1610 if (fp->f_type == DTYPE_FIFO)
1611 return (vn_fill_kinfo(fp, kif, fdp));
1612 kif->kf_type = KF_TYPE_PIPE;
1614 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1615 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1616 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1621 pipe_free_kmem(struct pipe *cpipe)
1624 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1625 ("pipe_free_kmem: pipe mutex locked"));
1627 if (cpipe->pipe_buffer.buffer != NULL) {
1628 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1629 vm_map_remove(pipe_map,
1630 (vm_offset_t)cpipe->pipe_buffer.buffer,
1631 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1632 cpipe->pipe_buffer.buffer = NULL;
1634 #ifndef PIPE_NODIRECT
1636 cpipe->pipe_pages.cnt = 0;
1637 cpipe->pipe_pages.pos = 0;
1638 cpipe->pipe_pages.npages = 0;
1647 pipeclose(struct pipe *cpipe)
1650 struct pipepair *pp;
1654 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1659 pp = cpipe->pipe_pair;
1663 * If the other side is blocked, wake it up saying that
1664 * we want to close it down.
1666 cpipe->pipe_state |= PIPE_EOF;
1667 while (cpipe->pipe_busy) {
1669 cpipe->pipe_state |= PIPE_WANT;
1671 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1675 pipeselwakeup(cpipe);
1678 * Disconnect from peer, if any.
1680 ppipe = cpipe->pipe_peer;
1681 if (ppipe->pipe_present == PIPE_ACTIVE) {
1682 ppipe->pipe_state |= PIPE_EOF;
1684 pipeselwakeup(ppipe);
1688 * Mark this endpoint as free. Release kmem resources. We
1689 * don't mark this endpoint as unused until we've finished
1690 * doing that, or the pipe might disappear out from under
1694 pipe_free_kmem(cpipe);
1696 cpipe->pipe_present = PIPE_CLOSING;
1700 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1701 * PIPE_FINALIZED, that allows other end to free the
1702 * pipe_pair, only after the knotes are completely dismantled.
1704 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1705 cpipe->pipe_present = PIPE_FINALIZED;
1706 seldrain(&cpipe->pipe_sel);
1707 knlist_destroy(&cpipe->pipe_sel.si_note);
1710 * If both endpoints are now closed, release the memory for the
1711 * pipe pair. If not, unlock.
1713 if (ppipe->pipe_present == PIPE_FINALIZED) {
1716 mac_pipe_destroy(pp);
1718 uma_zfree(pipe_zone, cpipe->pipe_pair);
1725 pipe_kqfilter(struct file *fp, struct knote *kn)
1730 * If a filter is requested that is not supported by this file
1731 * descriptor, don't return an error, but also don't ever generate an
1734 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1735 kn->kn_fop = &pipe_nfiltops;
1738 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1739 kn->kn_fop = &pipe_nfiltops;
1744 switch (kn->kn_filter) {
1746 kn->kn_fop = &pipe_rfiltops;
1749 kn->kn_fop = &pipe_wfiltops;
1750 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1751 /* other end of pipe has been closed */
1755 cpipe = PIPE_PEER(cpipe);
1758 if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1760 return (vnops.fo_kqfilter(fp, kn));
1766 kn->kn_hook = cpipe;
1767 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1773 filt_pipedetach(struct knote *kn)
1775 struct pipe *cpipe = kn->kn_hook;
1778 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1784 filt_piperead(struct knote *kn, long hint)
1786 struct file *fp = kn->kn_fp;
1787 struct pipe *rpipe = kn->kn_hook;
1789 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1790 kn->kn_data = rpipe->pipe_buffer.cnt;
1791 if (kn->kn_data == 0)
1792 kn->kn_data = rpipe->pipe_pages.cnt;
1794 if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1795 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1796 fp->f_pipegen != rpipe->pipe_wgen)) {
1797 kn->kn_flags |= EV_EOF;
1800 kn->kn_flags &= ~EV_EOF;
1801 return (kn->kn_data > 0);
1806 filt_pipewrite(struct knote *kn, long hint)
1808 struct pipe *wpipe = kn->kn_hook;
1811 * If this end of the pipe is closed, the knote was removed from the
1812 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1814 if (wpipe->pipe_present == PIPE_ACTIVE ||
1815 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1816 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1818 if (wpipe->pipe_state & PIPE_DIRECTW) {
1820 } else if (wpipe->pipe_buffer.size > 0) {
1821 kn->kn_data = wpipe->pipe_buffer.size -
1822 wpipe->pipe_buffer.cnt;
1824 kn->kn_data = PIPE_BUF;
1828 if (wpipe->pipe_present != PIPE_ACTIVE ||
1829 (wpipe->pipe_state & PIPE_EOF)) {
1830 kn->kn_flags |= EV_EOF;
1833 kn->kn_flags &= ~EV_EOF;
1834 return (kn->kn_data >= PIPE_BUF);
1838 filt_pipedetach_notsup(struct knote *kn)
1844 filt_pipenotsup(struct knote *kn, long hint)