2 * Copyright (c) 1996 John S. Dyson
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice immediately at the beginning of the file, without modification,
10 * this list of conditions, and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. Absolutely no warranty of function or purpose is made by the author
16 * 4. Modifications may be freely made to this file if the above conditions
19 * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.60.2.13 2002/08/05 15:05:15 des Exp $
20 * $DragonFly: src/sys/kern/sys_pipe.c,v 1.50 2008/09/09 04:06:13 dillon Exp $
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
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
33 #include <sys/fcntl.h>
35 #include <sys/filedesc.h>
36 #include <sys/filio.h>
37 #include <sys/ttycom.h>
40 #include <sys/select.h>
41 #include <sys/signalvar.h>
42 #include <sys/sysproto.h>
44 #include <sys/vnode.h>
46 #include <sys/event.h>
47 #include <sys/globaldata.h>
48 #include <sys/module.h>
49 #include <sys/malloc.h>
50 #include <sys/sysctl.h>
51 #include <sys/socket.h>
54 #include <vm/vm_param.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_extern.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_page.h>
62 #include <vm/vm_zone.h>
64 #include <sys/file2.h>
66 #include <machine/cpufunc.h>
69 * interfaces to the outside world
71 static int pipe_read (struct file *fp, struct uio *uio,
72 struct ucred *cred, int flags);
73 static int pipe_write (struct file *fp, struct uio *uio,
74 struct ucred *cred, int flags);
75 static int pipe_close (struct file *fp);
76 static int pipe_shutdown (struct file *fp, int how);
77 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
78 static int pipe_kqfilter (struct file *fp, struct knote *kn);
79 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
80 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred);
82 static struct fileops pipeops = {
84 .fo_write = pipe_write,
85 .fo_ioctl = pipe_ioctl,
87 .fo_kqfilter = pipe_kqfilter,
89 .fo_close = pipe_close,
90 .fo_shutdown = pipe_shutdown
93 static void filt_pipedetach(struct knote *kn);
94 static int filt_piperead(struct knote *kn, long hint);
95 static int filt_pipewrite(struct knote *kn, long hint);
97 static struct filterops pipe_rfiltops =
98 { 1, NULL, filt_pipedetach, filt_piperead };
99 static struct filterops pipe_wfiltops =
100 { 1, NULL, filt_pipedetach, filt_pipewrite };
102 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
105 * Default pipe buffer size(s), this can be kind-of large now because pipe
106 * space is pageable. The pipe code will try to maintain locality of
107 * reference for performance reasons, so small amounts of outstanding I/O
108 * will not wipe the cache.
110 #define MINPIPESIZE (PIPE_SIZE/3)
111 #define MAXPIPESIZE (2*PIPE_SIZE/3)
114 * Limit the number of "big" pipes
116 #define LIMITBIGPIPES 64
117 #define PIPEQ_MAX_CACHE 16 /* per-cpu pipe structure cache */
119 static int pipe_maxbig = LIMITBIGPIPES;
120 static int pipe_maxcache = PIPEQ_MAX_CACHE;
121 static int pipe_bigcount;
122 static int pipe_nbig;
123 static int pipe_bcache_alloc;
124 static int pipe_bkmem_alloc;
126 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
127 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
128 CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
129 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
130 CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
131 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
132 CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
133 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
134 CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
136 static int pipe_mpsafe = 0;
137 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
138 CTLFLAG_RW, &pipe_mpsafe, 0, "");
140 #if !defined(NO_PIPE_SYSCTL_STATS)
141 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
142 CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
143 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
144 CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
147 static void pipeclose (struct pipe *cpipe);
148 static void pipe_free_kmem (struct pipe *cpipe);
149 static int pipe_create (struct pipe **cpipep);
150 static __inline void pipeselwakeup (struct pipe *cpipe);
151 static int pipespace (struct pipe *cpipe, int size);
154 pipeselwakeup(struct pipe *cpipe)
156 if (cpipe->pipe_state & PIPE_SEL) {
158 cpipe->pipe_state &= ~PIPE_SEL;
159 selwakeup(&cpipe->pipe_sel);
162 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
164 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
167 if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
169 KNOTE(&cpipe->pipe_sel.si_note, 0);
175 * These routines are called before and after a UIO. The UIO
176 * may block, causing our held tokens to be lost temporarily.
178 * We use these routines to serialize reads against other reads
179 * and writes against other writes.
181 * The read token is held on entry so *ipp does not race.
184 pipe_start_uio(struct pipe *cpipe, u_int *ipp)
190 error = tsleep(ipp, PCATCH, "pipexx", 0);
199 pipe_end_uio(struct pipe *cpipe, u_int *ipp)
210 pipe_get_mplock(int *save)
213 if (pipe_mpsafe == 0) {
224 pipe_rel_mplock(int *save)
234 * The pipe system call for the DTYPE_PIPE type of pipes
236 * pipe_ARgs(int dummy)
241 sys_pipe(struct pipe_args *uap)
243 struct thread *td = curthread;
244 struct proc *p = td->td_proc;
245 struct file *rf, *wf;
246 struct pipe *rpipe, *wpipe;
251 rpipe = wpipe = NULL;
252 if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
258 error = falloc(p, &rf, &fd1);
264 uap->sysmsg_fds[0] = fd1;
267 * Warning: once we've gotten past allocation of the fd for the
268 * read-side, we can only drop the read side via fdrop() in order
269 * to avoid races against processes which manage to dup() the read
270 * side while we are blocked trying to allocate the write side.
272 rf->f_type = DTYPE_PIPE;
273 rf->f_flag = FREAD | FWRITE;
274 rf->f_ops = &pipeops;
276 error = falloc(p, &wf, &fd2);
278 fsetfd(p, NULL, fd1);
280 /* rpipe has been closed by fdrop(). */
284 wf->f_type = DTYPE_PIPE;
285 wf->f_flag = FREAD | FWRITE;
286 wf->f_ops = &pipeops;
288 uap->sysmsg_fds[1] = fd2;
290 rpipe->pipe_slock = kmalloc(sizeof(struct lock),
291 M_PIPE, M_WAITOK|M_ZERO);
292 wpipe->pipe_slock = rpipe->pipe_slock;
293 rpipe->pipe_peer = wpipe;
294 wpipe->pipe_peer = rpipe;
295 lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
298 * Once activated the peer relationship remains valid until
299 * both sides are closed.
310 * Allocate kva for pipe circular buffer, the space is pageable
311 * This routine will 'realloc' the size of a pipe safely, if it fails
312 * it will retain the old buffer.
313 * If it fails it will return ENOMEM.
316 pipespace(struct pipe *cpipe, int size)
318 struct vm_object *object;
322 npages = round_page(size) / PAGE_SIZE;
323 object = cpipe->pipe_buffer.object;
326 * [re]create the object if necessary and reserve space for it
327 * in the kernel_map. The object and memory are pageable. On
328 * success, free the old resources before assigning the new
331 if (object == NULL || object->size != npages) {
333 object = vm_object_allocate(OBJT_DEFAULT, npages);
334 buffer = (caddr_t)vm_map_min(&kernel_map);
336 error = vm_map_find(&kernel_map, object, 0,
337 (vm_offset_t *)&buffer, size,
340 VM_PROT_ALL, VM_PROT_ALL,
343 if (error != KERN_SUCCESS) {
344 vm_object_deallocate(object);
348 pipe_free_kmem(cpipe);
350 cpipe->pipe_buffer.object = object;
351 cpipe->pipe_buffer.buffer = buffer;
352 cpipe->pipe_buffer.size = size;
357 cpipe->pipe_buffer.rindex = 0;
358 cpipe->pipe_buffer.windex = 0;
363 * Initialize and allocate VM and memory for pipe, pulling the pipe from
364 * our per-cpu cache if possible. For now make sure it is sized for the
365 * smaller PIPE_SIZE default.
368 pipe_create(struct pipe **cpipep)
370 globaldata_t gd = mycpu;
374 if ((cpipe = gd->gd_pipeq) != NULL) {
375 gd->gd_pipeq = cpipe->pipe_peer;
377 cpipe->pipe_peer = NULL;
378 cpipe->pipe_wantwcnt = 0;
380 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
383 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
385 vfs_timestamp(&cpipe->pipe_ctime);
386 cpipe->pipe_atime = cpipe->pipe_ctime;
387 cpipe->pipe_mtime = cpipe->pipe_ctime;
388 lwkt_token_init(&cpipe->pipe_rlock);
389 lwkt_token_init(&cpipe->pipe_wlock);
394 * MPALMOSTSAFE (acquires mplock)
397 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
403 u_int size; /* total bytes available */
404 u_int nsize; /* total bytes to read */
405 u_int rindex; /* contiguous bytes available */
414 if (uio->uio_resid == 0)
418 * Setup locks, calculate nbio
420 pipe_get_mplock(&mpsave);
421 rpipe = (struct pipe *)fp->f_data;
422 lwkt_gettoken(&rlock, &rpipe->pipe_rlock);
424 if (fflags & O_FBLOCKING)
426 else if (fflags & O_FNONBLOCKING)
428 else if (fp->f_flag & O_NONBLOCK)
434 * Reads are serialized. Note howeverthat pipe_buffer.buffer and
435 * pipe_buffer.size can change out from under us when the number
436 * of bytes in the buffer are zero due to the write-side doing a
439 error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
441 pipe_rel_mplock(&mpsave);
442 lwkt_reltoken(&rlock);
445 while (uio->uio_resid) {
446 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
449 rindex = rpipe->pipe_buffer.rindex &
450 (rpipe->pipe_buffer.size - 1);
452 if (nsize > rpipe->pipe_buffer.size - rindex)
453 nsize = rpipe->pipe_buffer.size - rindex;
454 if (nsize > (u_int)uio->uio_resid)
455 nsize = (u_int)uio->uio_resid;
457 error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
462 rpipe->pipe_buffer.rindex += nsize;
466 * Shortcut to the loop-up if there is no writer
467 * waiting or if we have not transitioned across
468 * the half-way point.
470 half_way = rpipe->pipe_buffer.size >> 1;
471 if ((rpipe->pipe_state & PIPE_WANTW) == 0 ||
472 size <= half_way || size - nsize > half_way) {
479 * If the "write-side" was blocked we wake it up. This code
480 * is reached either when the buffer is completely emptied
481 * or if it becomes more then half-empty.
483 * Pipe_state can only be modified if both the rlock and
486 if (rpipe->pipe_state & PIPE_WANTW) {
487 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
488 if (rpipe->pipe_state & PIPE_WANTW) {
489 rpipe->pipe_state &= ~PIPE_WANTW;
490 lwkt_reltoken(&wlock);
493 lwkt_reltoken(&wlock);
498 * Pick up our copy loop again if the writer sent data to
501 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
506 * Detect EOF condition, do not set error.
508 if (rpipe->pipe_state & PIPE_REOF)
513 * Gravy train if SMP box. This saves a ton of IPIs and
514 * allows two cpus to operate in lockstep.
516 * XXX check pipe_wip also?
519 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
525 * Break if some data was read, or if this was a non-blocking
537 * Last chance, interlock with WANTR.
539 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
540 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
542 lwkt_reltoken(&wlock);
547 * If there is no more to read in the pipe, reset its
548 * pointers to the beginning. This improves cache hit
551 * We need both locks to modify both pointers, and there
552 * must also not be a write in progress or the uiomove()
553 * in the write might block and temporarily release
554 * its wlock, then reacquire and update windex. We are
555 * only serialized against reads, not writes.
557 * XXX should we even bother resetting the indices? It
558 * might actually be more cache efficient not to.
560 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
561 rpipe->pipe_wip == 0) {
562 rpipe->pipe_buffer.rindex = 0;
563 rpipe->pipe_buffer.windex = 0;
567 * Wait for more data.
569 * Pipe_state can only be set if both the rlock and wlock
572 rpipe->pipe_state |= PIPE_WANTR;
573 tsleep_interlock(rpipe);
574 lwkt_reltoken(&wlock);
575 error = tsleep(rpipe, PCATCH, "piperd", 0);
579 pipe_end_uio(rpipe, &rpipe->pipe_rip);
582 * Uptime last access time
584 if (error == 0 && nread)
585 vfs_timestamp(&rpipe->pipe_atime);
589 * Handle write blocking hysteresis. size can only increase while
592 * XXX shouldn't need this any more. We will wakeup the writer
593 * when we've drained past half-way. The worst the writer
594 * can do is fill the buffer up, not make it smaller, so
595 * we are guaranteed our half-way test.
597 if (rpipe->pipe_state & PIPE_WANTW) {
598 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
599 if (size <= (rpipe->pipe_buffer.size >> 1)) {
600 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
601 if (rpipe->pipe_state & PIPE_WANTW) {
602 rpipe->pipe_state &= ~PIPE_WANTW;
603 lwkt_reltoken(&wlock);
606 lwkt_reltoken(&wlock);
611 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
612 lwkt_reltoken(&rlock);
615 * If enough space is available in buffer then wakeup sel writers?
617 if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF)
618 pipeselwakeup(rpipe);
619 pipe_rel_mplock(&mpsave);
624 * MPALMOSTSAFE - acquires mplock
627 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
632 struct pipe *wpipe, *rpipe;
640 pipe_get_mplock(&mpsave);
643 * Writes go to the peer. The peer will always exist.
645 rpipe = (struct pipe *) fp->f_data;
646 wpipe = rpipe->pipe_peer;
647 lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
648 if (wpipe->pipe_state & PIPE_WEOF) {
649 pipe_rel_mplock(&mpsave);
650 lwkt_reltoken(&wlock);
655 * Degenerate case (EPIPE takes prec)
657 if (uio->uio_resid == 0) {
658 pipe_rel_mplock(&mpsave);
659 lwkt_reltoken(&wlock);
664 * Writes are serialized (start_uio must be called with wlock)
666 error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
668 pipe_rel_mplock(&mpsave);
669 lwkt_reltoken(&wlock);
673 if (fflags & O_FBLOCKING)
675 else if (fflags & O_FNONBLOCKING)
677 else if (fp->f_flag & O_NONBLOCK)
683 * If it is advantageous to resize the pipe buffer, do
684 * so. We are write-serialized so we can block safely.
686 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
687 (pipe_nbig < pipe_maxbig) &&
688 wpipe->pipe_wantwcnt > 4 &&
689 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
691 * Recheck after lock.
693 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
694 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
695 (pipe_nbig < pipe_maxbig) &&
696 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
697 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) {
702 lwkt_reltoken(&rlock);
705 orig_resid = uio->uio_resid;
708 while (uio->uio_resid) {
709 if (wpipe->pipe_state & PIPE_WEOF) {
714 windex = wpipe->pipe_buffer.windex &
715 (wpipe->pipe_buffer.size - 1);
716 space = wpipe->pipe_buffer.size -
717 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
720 /* Writes of size <= PIPE_BUF must be atomic. */
721 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
725 * Write to fill, read size handles write hysteresis. Also
726 * additional restrictions can cause select-based non-blocking
733 * Transfer size is minimum of uio transfer
734 * and free space in pipe buffer.
736 * Limit each uiocopy to no more then PIPE_SIZE
737 * so we can keep the gravy train going on a
738 * SMP box. This doubles the performance for
739 * write sizes > 16K. Otherwise large writes
740 * wind up doing an inefficient synchronous
743 if (space > (u_int)uio->uio_resid)
744 space = (u_int)uio->uio_resid;
745 if (space > PIPE_SIZE)
749 * First segment to transfer is minimum of
750 * transfer size and contiguous space in
751 * pipe buffer. If first segment to transfer
752 * is less than the transfer size, we've got
753 * a wraparound in the buffer.
755 segsize = wpipe->pipe_buffer.size - windex;
760 * If this is the first loop and the reader is
761 * blocked, do a preemptive wakeup of the reader.
763 * This works for both SMP and UP. On SMP the IPI
764 * latency plus the wlock interlock on the reader
765 * side is the fastest way to get the reader going.
766 * (The scheduler will hard loop on lock tokens).
768 * NOTE: We can't clear WANTR here without acquiring
769 * the rlock, which we don't want to do here!
771 if (wpipe->pipe_state & PIPE_WANTR)
775 * Transfer first segment
777 error = uiomove(&wpipe->pipe_buffer.buffer[windex],
780 wpipe->pipe_buffer.windex += segsize;
782 if (error == 0 && segsize < space) {
784 * Transfer remaining part now, to
785 * support atomic writes. Wraparound
788 segsize = space - segsize;
789 error = uiomove(&wpipe->pipe_buffer.buffer[0],
792 wpipe->pipe_buffer.windex += segsize;
801 * We need both the rlock and the wlock to interlock against
802 * the EOF, WANTW, and size checks, and to modify pipe_state.
804 * These are token locks so we do not have to worry about
807 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
810 * If the "read-side" has been blocked, wake it up now
811 * and yield to let it drain synchronously rather
814 if (wpipe->pipe_state & PIPE_WANTR) {
815 wpipe->pipe_state &= ~PIPE_WANTR;
820 * don't block on non-blocking I/O
823 lwkt_reltoken(&rlock);
829 * We have no more space and have something to offer,
830 * wake up select/poll.
832 pipeselwakeup(wpipe);
834 ++wpipe->pipe_wantwcnt; /* don't care about overflow */
835 wpipe->pipe_state |= PIPE_WANTW;
836 error = tsleep(wpipe, PCATCH, "pipewr", 0);
837 lwkt_reltoken(&rlock);
840 * Break out if we errored or the read side wants us to go
845 if (wpipe->pipe_state & PIPE_WEOF) {
850 pipe_end_uio(wpipe, &wpipe->pipe_wip);
853 * If we have put any characters in the buffer, we wake up
856 * Both rlock and wlock are required to be able to modify pipe_state.
858 if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
859 if (wpipe->pipe_state & PIPE_WANTR) {
860 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
861 if (wpipe->pipe_state & PIPE_WANTR) {
862 wpipe->pipe_state &= ~PIPE_WANTR;
863 lwkt_reltoken(&rlock);
866 lwkt_reltoken(&rlock);
872 * Don't return EPIPE if I/O was successful
874 if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
875 (uio->uio_resid == 0) &&
881 vfs_timestamp(&wpipe->pipe_mtime);
884 * We have something to offer,
885 * wake up select/poll.
887 space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;
888 lwkt_reltoken(&wlock);
890 pipeselwakeup(wpipe);
891 pipe_rel_mplock(&mpsave);
896 * MPALMOSTSAFE - acquires mplock
898 * we implement a very minimal set of ioctls for compatibility with sockets.
901 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
907 mpipe = (struct pipe *)fp->f_data;
912 mpipe->pipe_state |= PIPE_ASYNC;
914 mpipe->pipe_state &= ~PIPE_ASYNC;
919 *(int *)data = mpipe->pipe_buffer.windex -
920 mpipe->pipe_buffer.rindex;
924 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
927 *(int *)data = fgetown(mpipe->pipe_sigio);
931 /* This is deprecated, FIOSETOWN should be used instead. */
932 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
936 /* This is deprecated, FIOGETOWN should be used instead. */
937 *(int *)data = -fgetown(mpipe->pipe_sigio);
949 * MPALMOSTSAFE - acquires mplock
952 pipe_poll(struct file *fp, int events, struct ucred *cred)
960 pipe_get_mplock(&mpsave);
961 rpipe = (struct pipe *)fp->f_data;
962 wpipe = rpipe->pipe_peer;
963 if (events & (POLLIN | POLLRDNORM)) {
964 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
965 (rpipe->pipe_state & PIPE_REOF)) {
966 revents |= events & (POLLIN | POLLRDNORM);
970 if (events & (POLLOUT | POLLWRNORM)) {
971 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
972 revents |= events & (POLLOUT | POLLWRNORM);
974 space = wpipe->pipe_buffer.windex -
975 wpipe->pipe_buffer.rindex;
976 space = wpipe->pipe_buffer.size - space;
977 if (space >= PIPE_BUF)
978 revents |= events & (POLLOUT | POLLWRNORM);
982 if ((rpipe->pipe_state & PIPE_REOF) ||
984 (wpipe->pipe_state & PIPE_WEOF))
988 if (events & (POLLIN | POLLRDNORM)) {
989 selrecord(curthread, &rpipe->pipe_sel);
990 rpipe->pipe_state |= PIPE_SEL;
993 if (events & (POLLOUT | POLLWRNORM)) {
994 selrecord(curthread, &wpipe->pipe_sel);
995 wpipe->pipe_state |= PIPE_SEL;
998 pipe_rel_mplock(&mpsave);
1003 * MPALMOSTSAFE - acquires mplock
1006 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1011 pipe_get_mplock(&mpsave);
1012 pipe = (struct pipe *)fp->f_data;
1014 bzero((caddr_t)ub, sizeof(*ub));
1015 ub->st_mode = S_IFIFO;
1016 ub->st_blksize = pipe->pipe_buffer.size;
1017 ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1018 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1019 ub->st_atimespec = pipe->pipe_atime;
1020 ub->st_mtimespec = pipe->pipe_mtime;
1021 ub->st_ctimespec = pipe->pipe_ctime;
1023 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1025 * XXX (st_dev, st_ino) should be unique.
1027 pipe_rel_mplock(&mpsave);
1032 * MPALMOSTSAFE - acquires mplock
1035 pipe_close(struct file *fp)
1040 cpipe = (struct pipe *)fp->f_data;
1041 fp->f_ops = &badfileops;
1043 funsetown(cpipe->pipe_sigio);
1050 * Shutdown one or both directions of a full-duplex pipe.
1052 * MPALMOSTSAFE - acquires mplock
1055 pipe_shutdown(struct file *fp, int how)
1060 lwkt_tokref rpipe_rlock;
1061 lwkt_tokref rpipe_wlock;
1062 lwkt_tokref wpipe_rlock;
1063 lwkt_tokref wpipe_wlock;
1066 pipe_get_mplock(&mpsave);
1067 rpipe = (struct pipe *)fp->f_data;
1068 wpipe = rpipe->pipe_peer;
1071 * We modify pipe_state on both pipes, which means we need
1074 lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1075 lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1076 lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1077 lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1082 rpipe->pipe_state |= PIPE_REOF;
1083 wpipe->pipe_state |= PIPE_WEOF;
1084 if (rpipe->pipe_state & PIPE_WANTR) {
1085 rpipe->pipe_state &= ~PIPE_WANTR;
1088 if (wpipe->pipe_state & PIPE_WANTW) {
1089 wpipe->pipe_state &= ~PIPE_WANTW;
1092 pipeselwakeup(rpipe);
1098 wpipe->pipe_state |= PIPE_WEOF;
1099 rpipe->pipe_state |= PIPE_REOF;
1100 if (wpipe->pipe_state & PIPE_WANTW) {
1101 wpipe->pipe_state &= ~PIPE_WANTW;
1104 if (rpipe->pipe_state & PIPE_WANTR) {
1105 rpipe->pipe_state &= ~PIPE_WANTR;
1108 pipeselwakeup(wpipe);
1113 lwkt_reltoken(&rpipe_rlock);
1114 lwkt_reltoken(&rpipe_wlock);
1115 lwkt_reltoken(&wpipe_rlock);
1116 lwkt_reltoken(&wpipe_wlock);
1118 pipe_rel_mplock(&mpsave);
1123 pipe_free_kmem(struct pipe *cpipe)
1125 if (cpipe->pipe_buffer.buffer != NULL) {
1126 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1128 kmem_free(&kernel_map,
1129 (vm_offset_t)cpipe->pipe_buffer.buffer,
1130 cpipe->pipe_buffer.size);
1131 cpipe->pipe_buffer.buffer = NULL;
1132 cpipe->pipe_buffer.object = NULL;
1137 * Close the pipe. The slock must be held to interlock against simultanious
1138 * closes. The rlock and wlock must be held to adjust the pipe_state.
1141 pipeclose(struct pipe *cpipe)
1145 lwkt_tokref cpipe_rlock;
1146 lwkt_tokref cpipe_wlock;
1147 lwkt_tokref ppipe_rlock;
1148 lwkt_tokref ppipe_wlock;
1154 * The slock may not have been allocated yet (close during
1157 * We need both the read and write tokens to modify pipe_state.
1159 if (cpipe->pipe_slock)
1160 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1161 lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1162 lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1165 * Set our state, wakeup anyone waiting in select, and
1166 * wakeup anyone blocked on our pipe.
1168 cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1169 pipeselwakeup(cpipe);
1170 if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1171 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1176 * Disconnect from peer
1178 if ((ppipe = cpipe->pipe_peer) != NULL) {
1179 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1180 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1181 ppipe->pipe_state |= PIPE_REOF;
1182 pipeselwakeup(ppipe);
1183 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1184 ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1187 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1189 KNOTE(&ppipe->pipe_sel.si_note, 0);
1192 lwkt_reltoken(&ppipe_rlock);
1193 lwkt_reltoken(&ppipe_wlock);
1197 * If the peer is also closed we can free resources for both
1198 * sides, otherwise we leave our side intact to deal with any
1199 * races (since we only have the slock).
1201 if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1202 cpipe->pipe_peer = NULL;
1203 ppipe->pipe_peer = NULL;
1204 ppipe->pipe_slock = NULL; /* we will free the slock */
1209 lwkt_reltoken(&cpipe_rlock);
1210 lwkt_reltoken(&cpipe_wlock);
1211 if (cpipe->pipe_slock)
1212 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1215 * If we disassociated from our peer we can free resources
1217 if (ppipe == NULL) {
1219 if (cpipe->pipe_slock) {
1220 kfree(cpipe->pipe_slock, M_PIPE);
1221 cpipe->pipe_slock = NULL;
1223 if (gd->gd_pipeqcount >= pipe_maxcache ||
1224 cpipe->pipe_buffer.size != PIPE_SIZE
1226 pipe_free_kmem(cpipe);
1227 kfree(cpipe, M_PIPE);
1229 cpipe->pipe_state = 0;
1230 cpipe->pipe_peer = gd->gd_pipeq;
1231 gd->gd_pipeq = cpipe;
1232 ++gd->gd_pipeqcount;
1238 * MPALMOSTSAFE - acquires mplock
1241 pipe_kqfilter(struct file *fp, struct knote *kn)
1246 cpipe = (struct pipe *)kn->kn_fp->f_data;
1248 switch (kn->kn_filter) {
1250 kn->kn_fop = &pipe_rfiltops;
1253 kn->kn_fop = &pipe_wfiltops;
1254 cpipe = cpipe->pipe_peer;
1255 if (cpipe == NULL) {
1256 /* other end of pipe has been closed */
1264 kn->kn_hook = (caddr_t)cpipe;
1266 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1272 filt_pipedetach(struct knote *kn)
1274 struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1276 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1281 filt_piperead(struct knote *kn, long hint)
1283 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1285 kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1288 if (rpipe->pipe_state & PIPE_REOF) {
1289 kn->kn_flags |= EV_EOF;
1292 return (kn->kn_data > 0);
1297 filt_pipewrite(struct knote *kn, long hint)
1299 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1300 struct pipe *wpipe = rpipe->pipe_peer;
1304 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1306 kn->kn_flags |= EV_EOF;
1309 space = wpipe->pipe_buffer.windex -
1310 wpipe->pipe_buffer.rindex;
1311 space = wpipe->pipe_buffer.size - space;
1312 kn->kn_data = space;
1313 return (kn->kn_data >= PIPE_BUF);