2 * Copyright (c) 1996 John S. Dyson
3 * Copyright (c) 2012 Giovanni Trematerra
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice immediately at the beginning of the file, without modification,
11 * this list of conditions, and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Absolutely no warranty of function or purpose is made by the author
17 * 4. Modifications may be freely made to this file if the above conditions
22 * This file contains a high-performance replacement for the socket-based
23 * pipes scheme originally used in FreeBSD/4.4Lite. It does not support
24 * all features of sockets, but does do everything that pipes normally
29 * This code has two modes of operation, a small write mode and a large
30 * write mode. The small write mode acts like conventional pipes with
31 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the
32 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT
33 * and PIPE_SIZE in size, the sending process pins the underlying pages in
34 * memory, and the receiving process copies directly from these pinned pages
35 * in the sending process.
37 * If the sending process receives a signal, it is possible that it will
38 * go away, and certainly its address space can change, because control
39 * is returned back to the user-mode side. In that case, the pipe code
40 * arranges to copy the buffer supplied by the user process, to a pageable
41 * kernel buffer, and the receiving process will grab the data from the
42 * pageable kernel buffer. Since signals don't happen all that often,
43 * the copy operation is normally eliminated.
45 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
46 * happen for small transfers so that the system will not spend all of
47 * its time context switching.
49 * In order to limit the resource use of pipes, two sysctls exist:
51 * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
52 * address space available to us in pipe_map. This value is normally
53 * autotuned, but may also be loader tuned.
55 * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
56 * memory in use by pipes.
58 * Based on how large pipekva is relative to maxpipekva, the following
62 * New pipes are given 16K of memory backing, pipes may dynamically
63 * grow to as large as 64K where needed.
65 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
66 * existing pipes may NOT grow.
68 * New pipes are given 4K (or PAGE_SIZE) of memory backing,
69 * existing pipes will be shrunk down to 4K whenever possible.
71 * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0. If
72 * that is set, the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
73 * resize which MUST occur for reverse-direction pipes when they are
76 * Additional information about the current state of pipes may be obtained
77 * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
78 * and kern.ipc.piperesizefail.
80 * Locking rules: There are two locks present here: A mutex, used via
81 * PIPE_LOCK, and a flag, used via pipelock(). All locking is done via
82 * the flag, as mutexes can not persist over uiomove. The mutex
83 * exists only to guard access to the flag, and is not in itself a
84 * locking mechanism. Also note that there is only a single mutex for
85 * both directions of a pipe.
87 * As pipelock() may have to sleep before it can acquire the flag, it
88 * is important to reread all data after a call to pipelock(); everything
89 * in the structure may have changed.
92 #include <sys/param.h>
93 #include <sys/systm.h>
95 #include <sys/fcntl.h>
97 #include <sys/filedesc.h>
98 #include <sys/filio.h>
99 #include <sys/kernel.h>
100 #include <sys/lock.h>
101 #include <sys/mutex.h>
102 #include <sys/ttycom.h>
103 #include <sys/stat.h>
104 #include <sys/malloc.h>
105 #include <sys/poll.h>
106 #include <sys/selinfo.h>
107 #include <sys/signalvar.h>
108 #include <sys/syscallsubr.h>
109 #include <sys/sysctl.h>
110 #include <sys/sysproto.h>
111 #include <sys/pipe.h>
112 #include <sys/proc.h>
113 #include <sys/vnode.h>
115 #include <sys/user.h>
116 #include <sys/event.h>
118 #include <security/mac/mac_framework.h>
121 #include <vm/vm_param.h>
122 #include <vm/vm_object.h>
123 #include <vm/vm_kern.h>
124 #include <vm/vm_extern.h>
126 #include <vm/vm_map.h>
127 #include <vm/vm_page.h>
131 * Use this define if you want to disable *fancy* VM things. Expect an
132 * approx 30% decrease in transfer rate. This could be useful for
135 /* #define PIPE_NODIRECT */
137 #define PIPE_PEER(pipe) \
138 (((pipe)->pipe_type & PIPE_TYPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
141 * interfaces to the outside world
143 static fo_rdwr_t pipe_read;
144 static fo_rdwr_t pipe_write;
145 static fo_truncate_t pipe_truncate;
146 static fo_ioctl_t pipe_ioctl;
147 static fo_poll_t pipe_poll;
148 static fo_kqfilter_t pipe_kqfilter;
149 static fo_stat_t pipe_stat;
150 static fo_close_t pipe_close;
151 static fo_chmod_t pipe_chmod;
152 static fo_chown_t pipe_chown;
153 static fo_fill_kinfo_t pipe_fill_kinfo;
155 struct fileops pipeops = {
156 .fo_read = pipe_read,
157 .fo_write = pipe_write,
158 .fo_truncate = pipe_truncate,
159 .fo_ioctl = pipe_ioctl,
160 .fo_poll = pipe_poll,
161 .fo_kqfilter = pipe_kqfilter,
162 .fo_stat = pipe_stat,
163 .fo_close = pipe_close,
164 .fo_chmod = pipe_chmod,
165 .fo_chown = pipe_chown,
166 .fo_sendfile = invfo_sendfile,
167 .fo_fill_kinfo = pipe_fill_kinfo,
168 .fo_flags = DFLAG_PASSABLE
171 static void filt_pipedetach(struct knote *kn);
172 static void filt_pipedetach_notsup(struct knote *kn);
173 static int filt_pipenotsup(struct knote *kn, long hint);
174 static int filt_piperead(struct knote *kn, long hint);
175 static int filt_pipewrite(struct knote *kn, long hint);
177 static struct filterops pipe_nfiltops = {
179 .f_detach = filt_pipedetach_notsup,
180 .f_event = filt_pipenotsup
182 static struct filterops pipe_rfiltops = {
184 .f_detach = filt_pipedetach,
185 .f_event = filt_piperead
187 static struct filterops pipe_wfiltops = {
189 .f_detach = filt_pipedetach,
190 .f_event = filt_pipewrite
194 * Default pipe buffer size(s), this can be kind-of large now because pipe
195 * space is pageable. The pipe code will try to maintain locality of
196 * reference for performance reasons, so small amounts of outstanding I/O
197 * will not wipe the cache.
199 #define MINPIPESIZE (PIPE_SIZE/3)
200 #define MAXPIPESIZE (2*PIPE_SIZE/3)
202 static long amountpipekva;
203 static int pipefragretry;
204 static int pipeallocfail;
205 static int piperesizefail;
206 static int piperesizeallowed = 1;
207 static long pipe_mindirect = PIPE_MINDIRECT;
209 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
210 &maxpipekva, 0, "Pipe KVA limit");
211 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
212 &amountpipekva, 0, "Pipe KVA usage");
213 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
214 &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
215 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
216 &pipeallocfail, 0, "Pipe allocation failures");
217 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
218 &piperesizefail, 0, "Pipe resize failures");
219 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
220 &piperesizeallowed, 0, "Pipe resizing allowed");
222 static void pipeinit(void *dummy __unused);
223 static void pipeclose(struct pipe *cpipe);
224 static void pipe_free_kmem(struct pipe *cpipe);
225 static int pipe_create(struct pipe *pipe, bool backing);
226 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
227 static __inline int pipelock(struct pipe *cpipe, int catch);
228 static __inline void pipeunlock(struct pipe *cpipe);
229 static void pipe_timestamp(struct timespec *tsp);
230 #ifndef PIPE_NODIRECT
231 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
232 static void pipe_destroy_write_buffer(struct pipe *wpipe);
233 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
234 static void pipe_clone_write_buffer(struct pipe *wpipe);
236 static int pipespace(struct pipe *cpipe, int size);
237 static int pipespace_new(struct pipe *cpipe, int size);
239 static int pipe_zone_ctor(void *mem, int size, void *arg, int flags);
240 static int pipe_zone_init(void *mem, int size, int flags);
241 static void pipe_zone_fini(void *mem, int size);
243 static uma_zone_t pipe_zone;
244 static struct unrhdr64 pipeino_unr;
245 static dev_t pipedev_ino;
247 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
250 pipeinit(void *dummy __unused)
253 pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
254 pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
256 KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
257 new_unrhdr64(&pipeino_unr, 1);
258 pipedev_ino = devfs_alloc_cdp_inode();
259 KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
263 sysctl_handle_pipe_mindirect(SYSCTL_HANDLER_ARGS)
266 long tmp_pipe_mindirect = pipe_mindirect;
268 error = sysctl_handle_long(oidp, &tmp_pipe_mindirect, arg2, req);
269 if (error != 0 || req->newptr == NULL)
273 * Don't allow pipe_mindirect to be set so low that we violate
274 * atomicity requirements.
276 if (tmp_pipe_mindirect <= PIPE_BUF)
278 pipe_mindirect = tmp_pipe_mindirect;
281 SYSCTL_OID(_kern_ipc, OID_AUTO, pipe_mindirect, CTLTYPE_LONG | CTLFLAG_RW,
282 &pipe_mindirect, 0, sysctl_handle_pipe_mindirect, "L",
283 "Minimum write size triggering VM optimization");
286 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
289 struct pipe *rpipe, *wpipe;
291 KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
293 pp = (struct pipepair *)mem;
296 * We zero both pipe endpoints to make sure all the kmem pointers
297 * are NULL, flag fields are zero'd, etc. We timestamp both
298 * endpoints with the same time.
300 rpipe = &pp->pp_rpipe;
301 bzero(rpipe, sizeof(*rpipe));
302 pipe_timestamp(&rpipe->pipe_ctime);
303 rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
305 wpipe = &pp->pp_wpipe;
306 bzero(wpipe, sizeof(*wpipe));
307 wpipe->pipe_ctime = rpipe->pipe_ctime;
308 wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
310 rpipe->pipe_peer = wpipe;
311 rpipe->pipe_pair = pp;
312 wpipe->pipe_peer = rpipe;
313 wpipe->pipe_pair = pp;
316 * Mark both endpoints as present; they will later get free'd
317 * one at a time. When both are free'd, then the whole pair
320 rpipe->pipe_present = PIPE_ACTIVE;
321 wpipe->pipe_present = PIPE_ACTIVE;
324 * Eventually, the MAC Framework may initialize the label
325 * in ctor or init, but for now we do it elswhere to avoid
326 * blocking in ctor or init.
334 pipe_zone_init(void *mem, int size, int flags)
338 KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
340 pp = (struct pipepair *)mem;
342 mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
347 pipe_zone_fini(void *mem, int size)
351 KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
353 pp = (struct pipepair *)mem;
355 mtx_destroy(&pp->pp_mtx);
359 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
362 struct pipe *rpipe, *wpipe;
365 *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
368 * The MAC label is shared between the connected endpoints. As a
369 * result mac_pipe_init() and mac_pipe_create() are called once
370 * for the pair, and not on the endpoints.
373 mac_pipe_create(td->td_ucred, pp);
375 rpipe = &pp->pp_rpipe;
376 wpipe = &pp->pp_wpipe;
378 knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
379 knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
382 * Only the forward direction pipe is backed by big buffer by
385 error = pipe_create(rpipe, true);
388 error = pipe_create(wpipe, false);
391 * This cleanup leaves the pipe inode number for rpipe
392 * still allocated, but never used. We do not free
393 * inode numbers for opened pipes, which is required
394 * for correctness because numbers must be unique.
395 * But also it avoids any memory use by the unr
396 * allocator, so stashing away the transient inode
397 * number is reasonable.
399 pipe_free_kmem(rpipe);
403 rpipe->pipe_state |= PIPE_DIRECTOK;
404 wpipe->pipe_state |= PIPE_DIRECTOK;
408 knlist_destroy(&rpipe->pipe_sel.si_note);
409 knlist_destroy(&wpipe->pipe_sel.si_note);
411 mac_pipe_destroy(pp);
413 uma_zfree(pipe_zone, pp);
418 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
423 error = pipe_paircreate(td, &pp);
426 pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
427 *ppipe = &pp->pp_rpipe;
432 pipe_dtor(struct pipe *dpipe)
436 peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
437 funsetown(&dpipe->pipe_sigio);
440 funsetown(&peer->pipe_sigio);
448 * This used to be vfs_timestamp but the higher precision is unnecessary and
449 * can very negatively affect performance in virtualized environments (e.g., on
450 * vms running on amd64 when using the rdtscp instruction).
453 pipe_timestamp(struct timespec *tsp)
460 * The pipe system call for the DTYPE_PIPE type of pipes. If we fail, let
461 * the zone pick up the pieces via pipeclose().
464 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
465 struct filecaps *fcaps2)
467 struct file *rf, *wf;
468 struct pipe *rpipe, *wpipe;
470 int fd, fflags, error;
472 error = pipe_paircreate(td, &pp);
475 rpipe = &pp->pp_rpipe;
476 wpipe = &pp->pp_wpipe;
477 error = falloc_caps(td, &rf, &fd, flags, fcaps1);
483 /* An extra reference on `rf' has been held for us by falloc_caps(). */
486 fflags = FREAD | FWRITE;
487 if ((flags & O_NONBLOCK) != 0)
491 * Warning: once we've gotten past allocation of the fd for the
492 * read-side, we can only drop the read side via fdrop() in order
493 * to avoid races against processes which manage to dup() the read
494 * side while we are blocked trying to allocate the write side.
496 finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
497 error = falloc_caps(td, &wf, &fd, flags, fcaps2);
499 fdclose(td, rf, fildes[0]);
501 /* rpipe has been closed by fdrop(). */
505 /* An extra reference on `wf' has been held for us by falloc_caps(). */
506 finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
514 #ifdef COMPAT_FREEBSD10
517 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
522 error = kern_pipe(td, fildes, 0, NULL, NULL);
526 td->td_retval[0] = fildes[0];
527 td->td_retval[1] = fildes[1];
534 sys_pipe2(struct thread *td, struct pipe2_args *uap)
536 int error, fildes[2];
538 if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
540 error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
543 error = copyout(fildes, uap->fildes, 2 * sizeof(int));
545 (void)kern_close(td, fildes[0]);
546 (void)kern_close(td, fildes[1]);
552 * Allocate kva for pipe circular buffer, the space is pageable
553 * This routine will 'realloc' the size of a pipe safely, if it fails
554 * it will retain the old buffer.
555 * If it fails it will return ENOMEM.
558 pipespace_new(struct pipe *cpipe, int size)
561 int error, cnt, firstseg;
562 static int curfail = 0;
563 static struct timeval lastfail;
565 KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
566 KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
567 ("pipespace: resize of direct writes not allowed"));
569 cnt = cpipe->pipe_buffer.cnt;
573 size = round_page(size);
574 buffer = (caddr_t) vm_map_min(pipe_map);
576 error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
577 VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
578 if (error != KERN_SUCCESS) {
579 if (cpipe->pipe_buffer.buffer == NULL &&
580 size > SMALL_PIPE_SIZE) {
581 size = SMALL_PIPE_SIZE;
585 if (cpipe->pipe_buffer.buffer == NULL) {
587 if (ppsratecheck(&lastfail, &curfail, 1))
588 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
595 /* copy data, then free old resources if we're resizing */
597 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
598 firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
599 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
601 if ((cnt - firstseg) > 0)
602 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
603 cpipe->pipe_buffer.in);
605 bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
609 pipe_free_kmem(cpipe);
610 cpipe->pipe_buffer.buffer = buffer;
611 cpipe->pipe_buffer.size = size;
612 cpipe->pipe_buffer.in = cnt;
613 cpipe->pipe_buffer.out = 0;
614 cpipe->pipe_buffer.cnt = cnt;
615 atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
620 * Wrapper for pipespace_new() that performs locking assertions.
623 pipespace(struct pipe *cpipe, int size)
626 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
627 ("Unlocked pipe passed to pipespace"));
628 return (pipespace_new(cpipe, size));
632 * lock a pipe for I/O, blocking other access
635 pipelock(struct pipe *cpipe, int catch)
639 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
644 while (cpipe->pipe_state & PIPE_LOCKFL) {
645 KASSERT(cpipe->pipe_waiters >= 0,
646 ("%s: bad waiter count %d", __func__,
647 cpipe->pipe_waiters));
648 cpipe->pipe_waiters++;
649 error = msleep(&cpipe->pipe_waiters, PIPE_MTX(cpipe), prio,
651 cpipe->pipe_waiters--;
655 cpipe->pipe_state |= PIPE_LOCKFL;
660 * unlock a pipe I/O lock
663 pipeunlock(struct pipe *cpipe)
666 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
667 KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
668 ("Unlocked pipe passed to pipeunlock"));
669 KASSERT(cpipe->pipe_waiters >= 0,
670 ("%s: bad waiter count %d", __func__,
671 cpipe->pipe_waiters));
672 cpipe->pipe_state &= ~PIPE_LOCKFL;
673 if (cpipe->pipe_waiters > 0)
674 wakeup_one(&cpipe->pipe_waiters);
678 pipeselwakeup(struct pipe *cpipe)
681 PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
682 if (cpipe->pipe_state & PIPE_SEL) {
683 selwakeuppri(&cpipe->pipe_sel, PSOCK);
684 if (!SEL_WAITING(&cpipe->pipe_sel))
685 cpipe->pipe_state &= ~PIPE_SEL;
687 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
688 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
689 KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
693 * Initialize and allocate VM and memory for pipe. The structure
694 * will start out zero'd from the ctor, so we just manage the kmem.
697 pipe_create(struct pipe *pipe, bool large_backing)
701 error = pipespace_new(pipe, !large_backing || amountpipekva >
702 maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
704 pipe->pipe_ino = alloc_unr64(&pipeino_unr);
710 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
711 int flags, struct thread *td)
721 * Try to avoid locking the pipe if we have nothing to do.
723 * There are programs which share one pipe amongst multiple processes
724 * and perform non-blocking reads in parallel, even if the pipe is
725 * empty. This in particular is the case with BSD make, which when
726 * spawned with a high -j number can find itself with over half of the
727 * calls failing to find anything.
729 if ((fp->f_flag & FNONBLOCK) != 0 && !mac_pipe_check_read_enabled()) {
730 if (__predict_false(uio->uio_resid == 0))
732 if ((atomic_load_short(&rpipe->pipe_state) & PIPE_EOF) == 0 &&
733 atomic_load_int(&rpipe->pipe_buffer.cnt) == 0 &&
734 atomic_load_int(&rpipe->pipe_pages.cnt) == 0)
740 error = pipelock(rpipe, 1);
745 error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
749 if (amountpipekva > (3 * maxpipekva) / 4) {
750 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
751 rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
752 rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
753 piperesizeallowed == 1) {
755 pipespace(rpipe, SMALL_PIPE_SIZE);
760 while (uio->uio_resid) {
762 * normal pipe buffer receive
764 if (rpipe->pipe_buffer.cnt > 0) {
765 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
766 if (size > rpipe->pipe_buffer.cnt)
767 size = rpipe->pipe_buffer.cnt;
768 if (size > uio->uio_resid)
769 size = uio->uio_resid;
773 &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
779 rpipe->pipe_buffer.out += size;
780 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
781 rpipe->pipe_buffer.out = 0;
783 rpipe->pipe_buffer.cnt -= size;
786 * If there is no more to read in the pipe, reset
787 * its pointers to the beginning. This improves
790 if (rpipe->pipe_buffer.cnt == 0) {
791 rpipe->pipe_buffer.in = 0;
792 rpipe->pipe_buffer.out = 0;
795 #ifndef PIPE_NODIRECT
797 * Direct copy, bypassing a kernel buffer.
799 } else if ((size = rpipe->pipe_pages.cnt) != 0) {
800 if (size > uio->uio_resid)
801 size = (u_int) uio->uio_resid;
803 error = uiomove_fromphys(rpipe->pipe_pages.ms,
804 rpipe->pipe_pages.pos, size, uio);
809 rpipe->pipe_pages.pos += size;
810 rpipe->pipe_pages.cnt -= size;
811 if (rpipe->pipe_pages.cnt == 0) {
812 rpipe->pipe_state &= ~PIPE_WANTW;
818 * detect EOF condition
819 * read returns 0 on EOF, no need to set error
821 if (rpipe->pipe_state & PIPE_EOF)
825 * If the "write-side" has been blocked, wake it up now.
827 if (rpipe->pipe_state & PIPE_WANTW) {
828 rpipe->pipe_state &= ~PIPE_WANTW;
833 * Break if some data was read.
839 * Unlock the pipe buffer for our remaining processing.
840 * We will either break out with an error or we will
841 * sleep and relock to loop.
846 * Handle non-blocking mode operation or
847 * wait for more data.
849 if (fp->f_flag & FNONBLOCK) {
852 rpipe->pipe_state |= PIPE_WANTR;
853 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
856 error = pipelock(rpipe, 1);
867 /* XXX: should probably do this before getting any locks. */
869 pipe_timestamp(&rpipe->pipe_atime);
874 * PIPE_WANT processing only makes sense if pipe_busy is 0.
876 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
877 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
879 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
881 * Handle write blocking hysteresis.
883 if (rpipe->pipe_state & PIPE_WANTW) {
884 rpipe->pipe_state &= ~PIPE_WANTW;
890 * Only wake up writers if there was actually something read.
891 * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
894 rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
895 pipeselwakeup(rpipe);
899 td->td_ru.ru_msgrcv++;
903 #ifndef PIPE_NODIRECT
905 * Map the sending processes' buffer into kernel space and wire it.
906 * This is similar to a physical write operation.
909 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
914 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
915 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
916 ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
917 KASSERT(wpipe->pipe_pages.cnt == 0,
918 ("%s: pipe map for %p contains residual data", __func__, wpipe));
920 if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
921 size = wpipe->pipe_buffer.size;
923 size = uio->uio_iov->iov_len;
925 wpipe->pipe_state |= PIPE_DIRECTW;
927 i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
928 (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
929 wpipe->pipe_pages.ms, PIPENPAGES);
932 wpipe->pipe_state &= ~PIPE_DIRECTW;
936 wpipe->pipe_pages.npages = i;
937 wpipe->pipe_pages.pos =
938 ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
939 wpipe->pipe_pages.cnt = size;
941 uio->uio_iov->iov_len -= size;
942 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
943 if (uio->uio_iov->iov_len == 0)
945 uio->uio_resid -= size;
946 uio->uio_offset += size;
951 * Unwire the process buffer.
954 pipe_destroy_write_buffer(struct pipe *wpipe)
957 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
958 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
959 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
960 KASSERT(wpipe->pipe_pages.cnt == 0,
961 ("%s: pipe map for %p contains residual data", __func__, wpipe));
963 wpipe->pipe_state &= ~PIPE_DIRECTW;
964 vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
965 wpipe->pipe_pages.npages = 0;
969 * In the case of a signal, the writing process might go away. This
970 * code copies the data into the circular buffer so that the source
971 * pages can be freed without loss of data.
974 pipe_clone_write_buffer(struct pipe *wpipe)
981 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
982 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
983 ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
985 size = wpipe->pipe_pages.cnt;
986 pos = wpipe->pipe_pages.pos;
987 wpipe->pipe_pages.cnt = 0;
989 wpipe->pipe_buffer.in = size;
990 wpipe->pipe_buffer.out = 0;
991 wpipe->pipe_buffer.cnt = size;
994 iov.iov_base = wpipe->pipe_buffer.buffer;
999 uio.uio_resid = size;
1000 uio.uio_segflg = UIO_SYSSPACE;
1001 uio.uio_rw = UIO_READ;
1002 uio.uio_td = curthread;
1003 uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
1005 pipe_destroy_write_buffer(wpipe);
1009 * This implements the pipe buffer write mechanism. Note that only
1010 * a direct write OR a normal pipe write can be pending at any given time.
1011 * If there are any characters in the pipe buffer, the direct write will
1012 * be deferred until the receiving process grabs all of the bytes from
1013 * the pipe buffer. Then the direct mapping write is set-up.
1016 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
1021 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1022 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1026 if (wpipe->pipe_state & PIPE_DIRECTW) {
1027 if (wpipe->pipe_state & PIPE_WANTR) {
1028 wpipe->pipe_state &= ~PIPE_WANTR;
1031 pipeselwakeup(wpipe);
1032 wpipe->pipe_state |= PIPE_WANTW;
1034 error = msleep(wpipe, PIPE_MTX(wpipe),
1035 PRIBIO | PCATCH, "pipdww", 0);
1041 if (wpipe->pipe_buffer.cnt > 0) {
1042 if (wpipe->pipe_state & PIPE_WANTR) {
1043 wpipe->pipe_state &= ~PIPE_WANTR;
1046 pipeselwakeup(wpipe);
1047 wpipe->pipe_state |= PIPE_WANTW;
1049 error = msleep(wpipe, PIPE_MTX(wpipe),
1050 PRIBIO | PCATCH, "pipdwc", 0);
1057 error = pipe_build_write_buffer(wpipe, uio);
1062 while (wpipe->pipe_pages.cnt != 0 &&
1063 (wpipe->pipe_state & PIPE_EOF) == 0) {
1064 if (wpipe->pipe_state & PIPE_WANTR) {
1065 wpipe->pipe_state &= ~PIPE_WANTR;
1068 pipeselwakeup(wpipe);
1069 wpipe->pipe_state |= PIPE_WANTW;
1071 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1078 if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1079 wpipe->pipe_pages.cnt = 0;
1080 pipe_destroy_write_buffer(wpipe);
1081 pipeselwakeup(wpipe);
1083 } else if (error == EINTR || error == ERESTART) {
1084 pipe_clone_write_buffer(wpipe);
1086 pipe_destroy_write_buffer(wpipe);
1088 KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1089 ("pipe %p leaked PIPE_DIRECTW", wpipe));
1099 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1100 int flags, struct thread *td)
1102 struct pipe *wpipe, *rpipe;
1104 int desiredsize, error;
1107 wpipe = PIPE_PEER(rpipe);
1109 error = pipelock(wpipe, 1);
1115 * detect loss of pipe read side, issue SIGPIPE if lost.
1117 if (wpipe->pipe_present != PIPE_ACTIVE ||
1118 (wpipe->pipe_state & PIPE_EOF)) {
1124 error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1133 /* Choose a larger size if it's advantageous */
1134 desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1135 while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1136 if (piperesizeallowed != 1)
1138 if (amountpipekva > maxpipekva / 2)
1140 if (desiredsize == BIG_PIPE_SIZE)
1142 desiredsize = desiredsize * 2;
1145 /* Choose a smaller size if we're in a OOM situation */
1146 if (amountpipekva > (3 * maxpipekva) / 4 &&
1147 wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1148 wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1149 piperesizeallowed == 1)
1150 desiredsize = SMALL_PIPE_SIZE;
1152 /* Resize if the above determined that a new size was necessary */
1153 if (desiredsize != wpipe->pipe_buffer.size &&
1154 (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1156 pipespace(wpipe, desiredsize);
1159 MPASS(wpipe->pipe_buffer.size != 0);
1161 orig_resid = uio->uio_resid;
1163 while (uio->uio_resid) {
1166 if (wpipe->pipe_state & PIPE_EOF) {
1170 #ifndef PIPE_NODIRECT
1172 * If the transfer is large, we can gain performance if
1173 * we do process-to-process copies directly.
1174 * If the write is non-blocking, we don't use the
1175 * direct write mechanism.
1177 * The direct write mechanism will detect the reader going
1180 if (uio->uio_segflg == UIO_USERSPACE &&
1181 uio->uio_iov->iov_len >= pipe_mindirect &&
1182 wpipe->pipe_buffer.size >= pipe_mindirect &&
1183 (fp->f_flag & FNONBLOCK) == 0) {
1184 error = pipe_direct_write(wpipe, uio);
1192 * Pipe buffered writes cannot be coincidental with
1193 * direct writes. We wait until the currently executing
1194 * direct write is completed before we start filling the
1195 * pipe buffer. We break out if a signal occurs or the
1198 if (wpipe->pipe_pages.cnt != 0) {
1199 if (wpipe->pipe_state & PIPE_WANTR) {
1200 wpipe->pipe_state &= ~PIPE_WANTR;
1203 pipeselwakeup(wpipe);
1204 wpipe->pipe_state |= PIPE_WANTW;
1206 error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1214 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1216 /* Writes of size <= PIPE_BUF must be atomic. */
1217 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1221 int size; /* Transfer size */
1222 int segsize; /* first segment to transfer */
1225 * Transfer size is minimum of uio transfer
1226 * and free space in pipe buffer.
1228 if (space > uio->uio_resid)
1229 size = uio->uio_resid;
1233 * First segment to transfer is minimum of
1234 * transfer size and contiguous space in
1235 * pipe buffer. If first segment to transfer
1236 * is less than the transfer size, we've got
1237 * a wraparound in the buffer.
1239 segsize = wpipe->pipe_buffer.size -
1240 wpipe->pipe_buffer.in;
1244 /* Transfer first segment */
1247 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1251 if (error == 0 && segsize < size) {
1252 KASSERT(wpipe->pipe_buffer.in + segsize ==
1253 wpipe->pipe_buffer.size,
1254 ("Pipe buffer wraparound disappeared"));
1256 * Transfer remaining part now, to
1257 * support atomic writes. Wraparound
1263 &wpipe->pipe_buffer.buffer[0],
1264 size - segsize, uio);
1268 wpipe->pipe_buffer.in += size;
1269 if (wpipe->pipe_buffer.in >=
1270 wpipe->pipe_buffer.size) {
1271 KASSERT(wpipe->pipe_buffer.in ==
1273 wpipe->pipe_buffer.size,
1274 ("Expected wraparound bad"));
1275 wpipe->pipe_buffer.in = size - segsize;
1278 wpipe->pipe_buffer.cnt += size;
1279 KASSERT(wpipe->pipe_buffer.cnt <=
1280 wpipe->pipe_buffer.size,
1281 ("Pipe buffer overflow"));
1288 * If the "read-side" has been blocked, wake it up now.
1290 if (wpipe->pipe_state & PIPE_WANTR) {
1291 wpipe->pipe_state &= ~PIPE_WANTR;
1296 * don't block on non-blocking I/O
1298 if (fp->f_flag & FNONBLOCK) {
1304 * We have no more space and have something to offer,
1305 * wake up select/poll.
1307 pipeselwakeup(wpipe);
1309 wpipe->pipe_state |= PIPE_WANTW;
1311 error = msleep(wpipe, PIPE_MTX(rpipe),
1312 PRIBIO | PCATCH, "pipewr", 0);
1322 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1323 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1325 } else if (wpipe->pipe_buffer.cnt > 0) {
1327 * If we have put any characters in the buffer, we wake up
1330 if (wpipe->pipe_state & PIPE_WANTR) {
1331 wpipe->pipe_state &= ~PIPE_WANTR;
1337 * Don't return EPIPE if any byte was written.
1338 * EINTR and other interrupts are handled by generic I/O layer.
1339 * Do not pretend that I/O succeeded for obvious user error
1342 if (uio->uio_resid != orig_resid && error == EPIPE)
1346 pipe_timestamp(&wpipe->pipe_mtime);
1349 * We have something to offer,
1350 * wake up select/poll.
1352 if (wpipe->pipe_buffer.cnt)
1353 pipeselwakeup(wpipe);
1357 if (uio->uio_resid != orig_resid)
1358 td->td_ru.ru_msgsnd++;
1364 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1371 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1372 error = vnops.fo_truncate(fp, length, active_cred, td);
1374 error = invfo_truncate(fp, length, active_cred, td);
1379 * we implement a very minimal set of ioctls for compatibility with sockets.
1382 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1385 struct pipe *mpipe = fp->f_data;
1391 error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1405 mpipe->pipe_state |= PIPE_ASYNC;
1407 mpipe->pipe_state &= ~PIPE_ASYNC;
1412 if (!(fp->f_flag & FREAD)) {
1417 if (mpipe->pipe_pages.cnt != 0)
1418 *(int *)data = mpipe->pipe_pages.cnt;
1420 *(int *)data = mpipe->pipe_buffer.cnt;
1425 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1429 *(int *)data = fgetown(&mpipe->pipe_sigio);
1432 /* This is deprecated, FIOSETOWN should be used instead. */
1435 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1438 /* This is deprecated, FIOGETOWN should be used instead. */
1440 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1453 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1458 int levents, revents;
1465 wpipe = PIPE_PEER(rpipe);
1468 error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1472 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1473 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1474 revents |= events & (POLLIN | POLLRDNORM);
1476 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1477 if (wpipe->pipe_present != PIPE_ACTIVE ||
1478 (wpipe->pipe_state & PIPE_EOF) ||
1479 ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1480 ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1481 wpipe->pipe_buffer.size == 0)))
1482 revents |= events & (POLLOUT | POLLWRNORM);
1485 (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1486 if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1487 fp->f_pipegen == rpipe->pipe_wgen)
1488 events |= POLLINIGNEOF;
1490 if ((events & POLLINIGNEOF) == 0) {
1491 if (rpipe->pipe_state & PIPE_EOF) {
1492 if (fp->f_flag & FREAD)
1493 revents |= (events & (POLLIN | POLLRDNORM));
1494 if (wpipe->pipe_present != PIPE_ACTIVE ||
1495 (wpipe->pipe_state & PIPE_EOF))
1502 * Add ourselves regardless of eventmask as we have to return
1503 * POLLHUP even if it was not asked for.
1505 if ((fp->f_flag & FREAD) != 0) {
1506 selrecord(td, &rpipe->pipe_sel);
1507 if (SEL_WAITING(&rpipe->pipe_sel))
1508 rpipe->pipe_state |= PIPE_SEL;
1511 if ((fp->f_flag & FWRITE) != 0 &&
1512 wpipe->pipe_present == PIPE_ACTIVE) {
1513 selrecord(td, &wpipe->pipe_sel);
1514 if (SEL_WAITING(&wpipe->pipe_sel))
1515 wpipe->pipe_state |= PIPE_SEL;
1527 * We shouldn't need locks here as we're doing a read and this should
1528 * be a natural race.
1531 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred)
1540 if (mac_pipe_check_stat_enabled()) {
1542 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1550 /* For named pipes ask the underlying filesystem. */
1551 if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1552 return (vnops.fo_stat(fp, ub, active_cred));
1555 bzero(ub, sizeof(*ub));
1556 ub->st_mode = S_IFIFO;
1557 ub->st_blksize = PAGE_SIZE;
1558 if (pipe->pipe_pages.cnt != 0)
1559 ub->st_size = pipe->pipe_pages.cnt;
1561 ub->st_size = pipe->pipe_buffer.cnt;
1562 ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1563 ub->st_atim = pipe->pipe_atime;
1564 ub->st_mtim = pipe->pipe_mtime;
1565 ub->st_ctim = pipe->pipe_ctime;
1566 ub->st_uid = fp->f_cred->cr_uid;
1567 ub->st_gid = fp->f_cred->cr_gid;
1568 ub->st_dev = pipedev_ino;
1569 ub->st_ino = pipe->pipe_ino;
1571 * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1578 pipe_close(struct file *fp, struct thread *td)
1581 if (fp->f_vnode != NULL)
1582 return vnops.fo_close(fp, td);
1583 fp->f_ops = &badfileops;
1584 pipe_dtor(fp->f_data);
1590 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1596 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1597 error = vn_chmod(fp, mode, active_cred, td);
1599 error = invfo_chmod(fp, mode, active_cred, td);
1604 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1611 if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1612 error = vn_chown(fp, uid, gid, active_cred, td);
1614 error = invfo_chown(fp, uid, gid, active_cred, td);
1619 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1623 if (fp->f_type == DTYPE_FIFO)
1624 return (vn_fill_kinfo(fp, kif, fdp));
1625 kif->kf_type = KF_TYPE_PIPE;
1627 kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1628 kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1629 kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1630 kif->kf_un.kf_pipe.kf_pipe_buffer_in = pi->pipe_buffer.in;
1631 kif->kf_un.kf_pipe.kf_pipe_buffer_out = pi->pipe_buffer.out;
1632 kif->kf_un.kf_pipe.kf_pipe_buffer_size = pi->pipe_buffer.size;
1637 pipe_free_kmem(struct pipe *cpipe)
1640 KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1641 ("pipe_free_kmem: pipe mutex locked"));
1643 if (cpipe->pipe_buffer.buffer != NULL) {
1644 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1645 vm_map_remove(pipe_map,
1646 (vm_offset_t)cpipe->pipe_buffer.buffer,
1647 (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1648 cpipe->pipe_buffer.buffer = NULL;
1650 #ifndef PIPE_NODIRECT
1652 cpipe->pipe_pages.cnt = 0;
1653 cpipe->pipe_pages.pos = 0;
1654 cpipe->pipe_pages.npages = 0;
1663 pipeclose(struct pipe *cpipe)
1666 struct pipepair *pp;
1670 KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1675 pp = cpipe->pipe_pair;
1679 * If the other side is blocked, wake it up saying that
1680 * we want to close it down.
1682 cpipe->pipe_state |= PIPE_EOF;
1683 while (cpipe->pipe_busy) {
1685 cpipe->pipe_state |= PIPE_WANT;
1687 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1691 pipeselwakeup(cpipe);
1694 * Disconnect from peer, if any.
1696 ppipe = cpipe->pipe_peer;
1697 if (ppipe->pipe_present == PIPE_ACTIVE) {
1698 ppipe->pipe_state |= PIPE_EOF;
1700 pipeselwakeup(ppipe);
1704 * Mark this endpoint as free. Release kmem resources. We
1705 * don't mark this endpoint as unused until we've finished
1706 * doing that, or the pipe might disappear out from under
1710 pipe_free_kmem(cpipe);
1712 cpipe->pipe_present = PIPE_CLOSING;
1716 * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1717 * PIPE_FINALIZED, that allows other end to free the
1718 * pipe_pair, only after the knotes are completely dismantled.
1720 knlist_clear(&cpipe->pipe_sel.si_note, 1);
1721 cpipe->pipe_present = PIPE_FINALIZED;
1722 seldrain(&cpipe->pipe_sel);
1723 knlist_destroy(&cpipe->pipe_sel.si_note);
1726 * If both endpoints are now closed, release the memory for the
1727 * pipe pair. If not, unlock.
1729 if (ppipe->pipe_present == PIPE_FINALIZED) {
1732 mac_pipe_destroy(pp);
1734 uma_zfree(pipe_zone, cpipe->pipe_pair);
1741 pipe_kqfilter(struct file *fp, struct knote *kn)
1746 * If a filter is requested that is not supported by this file
1747 * descriptor, don't return an error, but also don't ever generate an
1750 if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1751 kn->kn_fop = &pipe_nfiltops;
1754 if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1755 kn->kn_fop = &pipe_nfiltops;
1760 switch (kn->kn_filter) {
1762 kn->kn_fop = &pipe_rfiltops;
1765 kn->kn_fop = &pipe_wfiltops;
1766 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1767 /* other end of pipe has been closed */
1771 cpipe = PIPE_PEER(cpipe);
1774 if ((cpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1776 return (vnops.fo_kqfilter(fp, kn));
1782 kn->kn_hook = cpipe;
1783 knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1789 filt_pipedetach(struct knote *kn)
1791 struct pipe *cpipe = kn->kn_hook;
1794 knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1800 filt_piperead(struct knote *kn, long hint)
1802 struct file *fp = kn->kn_fp;
1803 struct pipe *rpipe = kn->kn_hook;
1805 PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1806 kn->kn_data = rpipe->pipe_buffer.cnt;
1807 if (kn->kn_data == 0)
1808 kn->kn_data = rpipe->pipe_pages.cnt;
1810 if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1811 ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1812 fp->f_pipegen != rpipe->pipe_wgen)) {
1813 kn->kn_flags |= EV_EOF;
1816 kn->kn_flags &= ~EV_EOF;
1817 return (kn->kn_data > 0);
1822 filt_pipewrite(struct knote *kn, long hint)
1824 struct pipe *wpipe = kn->kn_hook;
1827 * If this end of the pipe is closed, the knote was removed from the
1828 * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1830 if (wpipe->pipe_present == PIPE_ACTIVE ||
1831 (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1832 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1834 if (wpipe->pipe_state & PIPE_DIRECTW) {
1836 } else if (wpipe->pipe_buffer.size > 0) {
1837 kn->kn_data = wpipe->pipe_buffer.size -
1838 wpipe->pipe_buffer.cnt;
1840 kn->kn_data = PIPE_BUF;
1844 if (wpipe->pipe_present != PIPE_ACTIVE ||
1845 (wpipe->pipe_state & PIPE_EOF)) {
1846 kn->kn_flags |= EV_EOF;
1849 kn->kn_flags &= ~EV_EOF;
1850 return (kn->kn_data >= PIPE_BUF);
1854 filt_pipedetach_notsup(struct knote *kn)
1860 filt_pipenotsup(struct knote *kn, long hint)