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>
39 #include <sys/signalvar.h>
40 #include <sys/sysproto.h>
42 #include <sys/vnode.h>
44 #include <sys/event.h>
45 #include <sys/globaldata.h>
46 #include <sys/module.h>
47 #include <sys/malloc.h>
48 #include <sys/sysctl.h>
49 #include <sys/socket.h>
52 #include <vm/vm_param.h>
54 #include <vm/vm_object.h>
55 #include <vm/vm_kern.h>
56 #include <vm/vm_extern.h>
58 #include <vm/vm_map.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_zone.h>
62 #include <sys/file2.h>
63 #include <sys/signal2.h>
65 #include <machine/cpufunc.h>
68 * interfaces to the outside world
70 static int pipe_read (struct file *fp, struct uio *uio,
71 struct ucred *cred, int flags);
72 static int pipe_write (struct file *fp, struct uio *uio,
73 struct ucred *cred, int flags);
74 static int pipe_close (struct file *fp);
75 static int pipe_shutdown (struct file *fp, int how);
76 static int pipe_kqfilter (struct file *fp, struct knote *kn);
77 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
78 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data,
79 struct ucred *cred, struct sysmsg *msg);
81 static struct fileops pipeops = {
83 .fo_write = pipe_write,
84 .fo_ioctl = pipe_ioctl,
85 .fo_kqfilter = pipe_kqfilter,
87 .fo_close = pipe_close,
88 .fo_shutdown = pipe_shutdown
91 static void filt_pipedetach(struct knote *kn);
92 static int filt_piperead(struct knote *kn, long hint);
93 static int filt_pipewrite(struct knote *kn, long hint);
95 static struct filterops pipe_rfiltops =
96 { FILTEROP_ISFD, NULL, filt_pipedetach, filt_piperead };
97 static struct filterops pipe_wfiltops =
98 { FILTEROP_ISFD, NULL, filt_pipedetach, filt_pipewrite };
100 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
103 * Default pipe buffer size(s), this can be kind-of large now because pipe
104 * space is pageable. The pipe code will try to maintain locality of
105 * reference for performance reasons, so small amounts of outstanding I/O
106 * will not wipe the cache.
108 #define MINPIPESIZE (PIPE_SIZE/3)
109 #define MAXPIPESIZE (2*PIPE_SIZE/3)
112 * Limit the number of "big" pipes
114 #define LIMITBIGPIPES 64
115 #define PIPEQ_MAX_CACHE 16 /* per-cpu pipe structure cache */
117 static int pipe_maxbig = LIMITBIGPIPES;
118 static int pipe_maxcache = PIPEQ_MAX_CACHE;
119 static int pipe_bigcount;
120 static int pipe_nbig;
121 static int pipe_bcache_alloc;
122 static int pipe_bkmem_alloc;
123 static int pipe_rblocked_count;
124 static int pipe_wblocked_count;
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, rblocked,
132 CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
133 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
134 CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
135 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
136 CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
137 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
138 CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
140 static int pipe_delay = 5000; /* 5uS default */
141 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
142 CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
144 #if !defined(NO_PIPE_SYSCTL_STATS)
145 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
146 CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
147 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
148 CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
151 static void pipeclose (struct pipe *cpipe);
152 static void pipe_free_kmem (struct pipe *cpipe);
153 static int pipe_create (struct pipe **cpipep);
154 static int pipespace (struct pipe *cpipe, int size);
157 pipewakeup(struct pipe *cpipe, int dosigio)
159 if (dosigio && (cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
160 lwkt_gettoken(&proc_token);
161 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
162 lwkt_reltoken(&proc_token);
164 KNOTE(&cpipe->pipe_kq.ki_note, 0);
168 * These routines are called before and after a UIO. The UIO
169 * may block, causing our held tokens to be lost temporarily.
171 * We use these routines to serialize reads against other reads
172 * and writes against other writes.
174 * The read token is held on entry so *ipp does not race.
177 pipe_start_uio(struct pipe *cpipe, int *ipp)
183 error = tsleep(ipp, PCATCH, "pipexx", 0);
192 pipe_end_uio(struct pipe *cpipe, int *ipp)
204 * The pipe system call for the DTYPE_PIPE type of pipes
206 * pipe_args(int dummy)
211 sys_pipe(struct pipe_args *uap)
213 struct thread *td = curthread;
214 struct filedesc *fdp = td->td_proc->p_fd;
215 struct file *rf, *wf;
216 struct pipe *rpipe, *wpipe;
219 rpipe = wpipe = NULL;
220 if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
226 error = falloc(td->td_lwp, &rf, &fd1);
232 uap->sysmsg_fds[0] = fd1;
235 * Warning: once we've gotten past allocation of the fd for the
236 * read-side, we can only drop the read side via fdrop() in order
237 * to avoid races against processes which manage to dup() the read
238 * side while we are blocked trying to allocate the write side.
240 rf->f_type = DTYPE_PIPE;
241 rf->f_flag = FREAD | FWRITE;
242 rf->f_ops = &pipeops;
244 error = falloc(td->td_lwp, &wf, &fd2);
246 fsetfd(fdp, NULL, fd1);
248 /* rpipe has been closed by fdrop(). */
252 wf->f_type = DTYPE_PIPE;
253 wf->f_flag = FREAD | FWRITE;
254 wf->f_ops = &pipeops;
256 uap->sysmsg_fds[1] = fd2;
258 rpipe->pipe_slock = kmalloc(sizeof(struct lock),
259 M_PIPE, M_WAITOK|M_ZERO);
260 wpipe->pipe_slock = rpipe->pipe_slock;
261 rpipe->pipe_peer = wpipe;
262 wpipe->pipe_peer = rpipe;
263 lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
266 * Once activated the peer relationship remains valid until
267 * both sides are closed.
269 fsetfd(fdp, rf, fd1);
270 fsetfd(fdp, wf, fd2);
278 * Allocate kva for pipe circular buffer, the space is pageable
279 * This routine will 'realloc' the size of a pipe safely, if it fails
280 * it will retain the old buffer.
281 * If it fails it will return ENOMEM.
284 pipespace(struct pipe *cpipe, int size)
286 struct vm_object *object;
290 npages = round_page(size) / PAGE_SIZE;
291 object = cpipe->pipe_buffer.object;
294 * [re]create the object if necessary and reserve space for it
295 * in the kernel_map. The object and memory are pageable. On
296 * success, free the old resources before assigning the new
299 if (object == NULL || object->size != npages) {
300 object = vm_object_allocate(OBJT_DEFAULT, npages);
301 buffer = (caddr_t)vm_map_min(&kernel_map);
303 error = vm_map_find(&kernel_map, object, 0,
304 (vm_offset_t *)&buffer,
306 1, VM_MAPTYPE_NORMAL,
307 VM_PROT_ALL, VM_PROT_ALL,
310 if (error != KERN_SUCCESS) {
311 vm_object_deallocate(object);
314 pipe_free_kmem(cpipe);
315 cpipe->pipe_buffer.object = object;
316 cpipe->pipe_buffer.buffer = buffer;
317 cpipe->pipe_buffer.size = size;
322 cpipe->pipe_buffer.rindex = 0;
323 cpipe->pipe_buffer.windex = 0;
328 * Initialize and allocate VM and memory for pipe, pulling the pipe from
329 * our per-cpu cache if possible. For now make sure it is sized for the
330 * smaller PIPE_SIZE default.
333 pipe_create(struct pipe **cpipep)
335 globaldata_t gd = mycpu;
339 if ((cpipe = gd->gd_pipeq) != NULL) {
340 gd->gd_pipeq = cpipe->pipe_peer;
342 cpipe->pipe_peer = NULL;
343 cpipe->pipe_wantwcnt = 0;
345 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
348 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
350 vfs_timestamp(&cpipe->pipe_ctime);
351 cpipe->pipe_atime = cpipe->pipe_ctime;
352 cpipe->pipe_mtime = cpipe->pipe_ctime;
353 lwkt_token_init(&cpipe->pipe_rlock, 1, "piper");
354 lwkt_token_init(&cpipe->pipe_wlock, 1, "pipew");
359 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
366 u_int size; /* total bytes available */
367 u_int nsize; /* total bytes to read */
368 u_int rindex; /* contiguous bytes available */
373 if (uio->uio_resid == 0)
377 * Setup locks, calculate nbio
379 rpipe = (struct pipe *)fp->f_data;
380 wpipe = rpipe->pipe_peer;
381 lwkt_gettoken(&rpipe->pipe_rlock);
383 if (fflags & O_FBLOCKING)
385 else if (fflags & O_FNONBLOCKING)
387 else if (fp->f_flag & O_NONBLOCK)
393 * Reads are serialized. Note however that pipe_buffer.buffer and
394 * pipe_buffer.size can change out from under us when the number
395 * of bytes in the buffer are zero due to the write-side doing a
398 error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
400 lwkt_reltoken(&rpipe->pipe_rlock);
405 bigread = (uio->uio_resid > 10 * 1024 * 1024);
408 while (uio->uio_resid) {
412 if (bigread && --bigcount == 0) {
415 if (CURSIG(curthread->td_lwp)) {
421 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
424 rindex = rpipe->pipe_buffer.rindex &
425 (rpipe->pipe_buffer.size - 1);
427 if (nsize > rpipe->pipe_buffer.size - rindex)
428 nsize = rpipe->pipe_buffer.size - rindex;
429 nsize = szmin(nsize, uio->uio_resid);
431 error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
436 rpipe->pipe_buffer.rindex += nsize;
440 * If the FIFO is still over half full just continue
441 * and do not try to notify the writer yet.
443 if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
449 * When the FIFO is less then half full notify any
450 * waiting writer. WANTW can be checked while
451 * holding just the rlock.
454 if ((rpipe->pipe_state & PIPE_WANTW) == 0)
459 * If the "write-side" was blocked we wake it up. This code
460 * is reached either when the buffer is completely emptied
461 * or if it becomes more then half-empty.
463 * Pipe_state can only be modified if both the rlock and
466 if (rpipe->pipe_state & PIPE_WANTW) {
467 lwkt_gettoken(&rpipe->pipe_wlock);
468 if (rpipe->pipe_state & PIPE_WANTW) {
469 rpipe->pipe_state &= ~PIPE_WANTW;
470 lwkt_reltoken(&rpipe->pipe_wlock);
473 lwkt_reltoken(&rpipe->pipe_wlock);
478 * Pick up our copy loop again if the writer sent data to
479 * us while we were messing around.
481 * On a SMP box poll up to pipe_delay nanoseconds for new
482 * data. Typically a value of 2000 to 4000 is sufficient
483 * to eradicate most IPIs/tsleeps/wakeups when a pipe
484 * is used for synchronous communications with small packets,
485 * and 8000 or so (8uS) will pipeline large buffer xfers
486 * between cpus over a pipe.
488 * For synchronous communications a hit means doing a
489 * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
490 * where as miss requiring a tsleep/wakeup sequence
491 * will take 7uS or more.
493 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
496 #if defined(SMP) && defined(_RDTSC_SUPPORTED_)
501 tsc_target = tsc_get_target(pipe_delay);
502 while (tsc_test_target(tsc_target) == 0) {
503 if (rpipe->pipe_buffer.windex !=
504 rpipe->pipe_buffer.rindex) {
515 * Detect EOF condition, do not set error.
517 if (rpipe->pipe_state & PIPE_REOF)
521 * Break if some data was read, or if this was a non-blocking
533 * Last chance, interlock with WANTR.
535 lwkt_gettoken(&rpipe->pipe_wlock);
536 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
538 lwkt_reltoken(&rpipe->pipe_wlock);
543 * Retest EOF - acquiring a new token can temporarily release
544 * tokens already held.
546 if (rpipe->pipe_state & PIPE_REOF) {
547 lwkt_reltoken(&rpipe->pipe_wlock);
552 * If there is no more to read in the pipe, reset its
553 * pointers to the beginning. This improves cache hit
556 * We need both locks to modify both pointers, and there
557 * must also not be a write in progress or the uiomove()
558 * in the write might block and temporarily release
559 * its wlock, then reacquire and update windex. We are
560 * only serialized against reads, not writes.
562 * XXX should we even bother resetting the indices? It
563 * might actually be more cache efficient not to.
565 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
566 rpipe->pipe_wip == 0) {
567 rpipe->pipe_buffer.rindex = 0;
568 rpipe->pipe_buffer.windex = 0;
572 * Wait for more data.
574 * Pipe_state can only be set if both the rlock and wlock
577 rpipe->pipe_state |= PIPE_WANTR;
578 tsleep_interlock(rpipe, PCATCH);
579 lwkt_reltoken(&rpipe->pipe_wlock);
580 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
581 ++pipe_rblocked_count;
585 pipe_end_uio(rpipe, &rpipe->pipe_rip);
588 * Uptime last access time
590 if (error == 0 && nread)
591 vfs_timestamp(&rpipe->pipe_atime);
594 * If we drained the FIFO more then half way then handle
595 * write blocking hysteresis.
597 * Note that PIPE_WANTW cannot be set by the writer without
598 * it holding both rlock and wlock, so we can test it
599 * while holding just rlock.
603 * Synchronous blocking is done on the pipe involved
605 if (rpipe->pipe_state & PIPE_WANTW) {
606 lwkt_gettoken(&rpipe->pipe_wlock);
607 if (rpipe->pipe_state & PIPE_WANTW) {
608 rpipe->pipe_state &= ~PIPE_WANTW;
609 lwkt_reltoken(&rpipe->pipe_wlock);
612 lwkt_reltoken(&rpipe->pipe_wlock);
617 * But we may also have to deal with a kqueue which is
618 * stored on the same pipe as its descriptor, so a
619 * EVFILT_WRITE event waiting for our side to drain will
620 * be on the other side.
622 lwkt_gettoken(&wpipe->pipe_wlock);
623 pipewakeup(wpipe, 0);
624 lwkt_reltoken(&wpipe->pipe_wlock);
626 /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
627 lwkt_reltoken(&rpipe->pipe_rlock);
633 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
647 * Writes go to the peer. The peer will always exist.
649 rpipe = (struct pipe *) fp->f_data;
650 wpipe = rpipe->pipe_peer;
651 lwkt_gettoken(&wpipe->pipe_wlock);
652 if (wpipe->pipe_state & PIPE_WEOF) {
653 lwkt_reltoken(&wpipe->pipe_wlock);
658 * Degenerate case (EPIPE takes prec)
660 if (uio->uio_resid == 0) {
661 lwkt_reltoken(&wpipe->pipe_wlock);
666 * Writes are serialized (start_uio must be called with wlock)
668 error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
670 lwkt_reltoken(&wpipe->pipe_wlock);
674 if (fflags & O_FBLOCKING)
676 else if (fflags & O_FNONBLOCKING)
678 else if (fp->f_flag & O_NONBLOCK)
684 * If it is advantageous to resize the pipe buffer, do
685 * so. We are write-serialized so we can block safely.
687 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
688 (pipe_nbig < pipe_maxbig) &&
689 wpipe->pipe_wantwcnt > 4 &&
690 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
692 * Recheck after lock.
694 lwkt_gettoken(&wpipe->pipe_rlock);
695 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
696 (pipe_nbig < pipe_maxbig) &&
697 (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
698 atomic_add_int(&pipe_nbig, 1);
699 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
702 atomic_subtract_int(&pipe_nbig, 1);
704 lwkt_reltoken(&wpipe->pipe_rlock);
707 orig_resid = uio->uio_resid;
710 bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
713 while (uio->uio_resid) {
714 if (wpipe->pipe_state & PIPE_WEOF) {
722 if (bigwrite && --bigcount == 0) {
725 if (CURSIG(curthread->td_lwp)) {
731 windex = wpipe->pipe_buffer.windex &
732 (wpipe->pipe_buffer.size - 1);
733 space = wpipe->pipe_buffer.size -
734 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
737 /* Writes of size <= PIPE_BUF must be atomic. */
738 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
742 * Write to fill, read size handles write hysteresis. Also
743 * additional restrictions can cause select-based non-blocking
750 * Transfer size is minimum of uio transfer
751 * and free space in pipe buffer.
753 * Limit each uiocopy to no more then PIPE_SIZE
754 * so we can keep the gravy train going on a
755 * SMP box. This doubles the performance for
756 * write sizes > 16K. Otherwise large writes
757 * wind up doing an inefficient synchronous
760 space = szmin(space, uio->uio_resid);
761 if (space > PIPE_SIZE)
765 * First segment to transfer is minimum of
766 * transfer size and contiguous space in
767 * pipe buffer. If first segment to transfer
768 * is less than the transfer size, we've got
769 * a wraparound in the buffer.
771 segsize = wpipe->pipe_buffer.size - windex;
777 * If this is the first loop and the reader is
778 * blocked, do a preemptive wakeup of the reader.
780 * On SMP the IPI latency plus the wlock interlock
781 * on the reader side is the fastest way to get the
782 * reader going. (The scheduler will hard loop on
785 * NOTE: We can't clear WANTR here without acquiring
786 * the rlock, which we don't want to do here!
788 if ((wpipe->pipe_state & PIPE_WANTR))
793 * Transfer segment, which may include a wrap-around.
794 * Update windex to account for both all in one go
795 * so the reader can read() the data atomically.
797 error = uiomove(&wpipe->pipe_buffer.buffer[windex],
799 if (error == 0 && segsize < space) {
800 segsize = space - segsize;
801 error = uiomove(&wpipe->pipe_buffer.buffer[0],
807 wpipe->pipe_buffer.windex += space;
813 * We need both the rlock and the wlock to interlock against
814 * the EOF, WANTW, and size checks, and to modify pipe_state.
816 * These are token locks so we do not have to worry about
819 lwkt_gettoken(&wpipe->pipe_rlock);
822 * If the "read-side" has been blocked, wake it up now
823 * and yield to let it drain synchronously rather
826 if (wpipe->pipe_state & PIPE_WANTR) {
827 wpipe->pipe_state &= ~PIPE_WANTR;
832 * don't block on non-blocking I/O
835 lwkt_reltoken(&wpipe->pipe_rlock);
841 * re-test whether we have to block in the writer after
842 * acquiring both locks, in case the reader opened up
845 space = wpipe->pipe_buffer.size -
846 (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
848 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
852 * Retest EOF - acquiring a new token can temporarily release
853 * tokens already held.
855 if (wpipe->pipe_state & PIPE_WEOF) {
856 lwkt_reltoken(&wpipe->pipe_rlock);
862 * We have no more space and have something to offer,
863 * wake up select/poll/kq.
866 wpipe->pipe_state |= PIPE_WANTW;
867 ++wpipe->pipe_wantwcnt;
868 pipewakeup(wpipe, 1);
869 if (wpipe->pipe_state & PIPE_WANTW)
870 error = tsleep(wpipe, PCATCH, "pipewr", 0);
871 ++pipe_wblocked_count;
873 lwkt_reltoken(&wpipe->pipe_rlock);
876 * Break out if we errored or the read side wants us to go
881 if (wpipe->pipe_state & PIPE_WEOF) {
886 pipe_end_uio(wpipe, &wpipe->pipe_wip);
889 * If we have put any characters in the buffer, we wake up
892 * Both rlock and wlock are required to be able to modify pipe_state.
894 if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
895 if (wpipe->pipe_state & PIPE_WANTR) {
896 lwkt_gettoken(&wpipe->pipe_rlock);
897 if (wpipe->pipe_state & PIPE_WANTR) {
898 wpipe->pipe_state &= ~PIPE_WANTR;
899 lwkt_reltoken(&wpipe->pipe_rlock);
902 lwkt_reltoken(&wpipe->pipe_rlock);
905 lwkt_gettoken(&wpipe->pipe_rlock);
906 pipewakeup(wpipe, 1);
907 lwkt_reltoken(&wpipe->pipe_rlock);
911 * Don't return EPIPE if I/O was successful
913 if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
914 (uio->uio_resid == 0) &&
920 vfs_timestamp(&wpipe->pipe_mtime);
923 * We have something to offer,
924 * wake up select/poll/kq.
926 /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
927 lwkt_reltoken(&wpipe->pipe_wlock);
932 * we implement a very minimal set of ioctls for compatibility with sockets.
935 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
936 struct ucred *cred, struct sysmsg *msg)
941 mpipe = (struct pipe *)fp->f_data;
943 lwkt_gettoken(&mpipe->pipe_rlock);
944 lwkt_gettoken(&mpipe->pipe_wlock);
949 mpipe->pipe_state |= PIPE_ASYNC;
951 mpipe->pipe_state &= ~PIPE_ASYNC;
956 *(int *)data = mpipe->pipe_buffer.windex -
957 mpipe->pipe_buffer.rindex;
961 lwkt_gettoken(&proc_token);
962 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
963 lwkt_reltoken(&proc_token);
966 *(int *)data = fgetown(mpipe->pipe_sigio);
970 /* This is deprecated, FIOSETOWN should be used instead. */
971 lwkt_gettoken(&proc_token);
972 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
973 lwkt_reltoken(&proc_token);
977 /* This is deprecated, FIOGETOWN should be used instead. */
978 *(int *)data = -fgetown(mpipe->pipe_sigio);
985 lwkt_reltoken(&mpipe->pipe_wlock);
986 lwkt_reltoken(&mpipe->pipe_rlock);
995 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
999 pipe = (struct pipe *)fp->f_data;
1001 bzero((caddr_t)ub, sizeof(*ub));
1002 ub->st_mode = S_IFIFO;
1003 ub->st_blksize = pipe->pipe_buffer.size;
1004 ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1005 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1006 ub->st_atimespec = pipe->pipe_atime;
1007 ub->st_mtimespec = pipe->pipe_mtime;
1008 ub->st_ctimespec = pipe->pipe_ctime;
1010 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1012 * XXX (st_dev, st_ino) should be unique.
1018 pipe_close(struct file *fp)
1022 cpipe = (struct pipe *)fp->f_data;
1023 fp->f_ops = &badfileops;
1025 lwkt_gettoken(&proc_token);
1026 funsetown(cpipe->pipe_sigio);
1027 lwkt_reltoken(&proc_token);
1033 * Shutdown one or both directions of a full-duplex pipe.
1036 pipe_shutdown(struct file *fp, int how)
1042 rpipe = (struct pipe *)fp->f_data;
1043 wpipe = rpipe->pipe_peer;
1046 * We modify pipe_state on both pipes, which means we need
1049 lwkt_gettoken(&rpipe->pipe_rlock);
1050 lwkt_gettoken(&rpipe->pipe_wlock);
1051 lwkt_gettoken(&wpipe->pipe_rlock);
1052 lwkt_gettoken(&wpipe->pipe_wlock);
1057 rpipe->pipe_state |= PIPE_REOF; /* my reads */
1058 rpipe->pipe_state |= PIPE_WEOF; /* peer writes */
1059 if (rpipe->pipe_state & PIPE_WANTR) {
1060 rpipe->pipe_state &= ~PIPE_WANTR;
1063 if (rpipe->pipe_state & PIPE_WANTW) {
1064 rpipe->pipe_state &= ~PIPE_WANTW;
1072 wpipe->pipe_state |= PIPE_REOF; /* peer reads */
1073 wpipe->pipe_state |= PIPE_WEOF; /* my writes */
1074 if (wpipe->pipe_state & PIPE_WANTR) {
1075 wpipe->pipe_state &= ~PIPE_WANTR;
1078 if (wpipe->pipe_state & PIPE_WANTW) {
1079 wpipe->pipe_state &= ~PIPE_WANTW;
1085 pipewakeup(rpipe, 1);
1086 pipewakeup(wpipe, 1);
1088 lwkt_reltoken(&wpipe->pipe_wlock);
1089 lwkt_reltoken(&wpipe->pipe_rlock);
1090 lwkt_reltoken(&rpipe->pipe_wlock);
1091 lwkt_reltoken(&rpipe->pipe_rlock);
1097 pipe_free_kmem(struct pipe *cpipe)
1099 if (cpipe->pipe_buffer.buffer != NULL) {
1100 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1101 atomic_subtract_int(&pipe_nbig, 1);
1102 kmem_free(&kernel_map,
1103 (vm_offset_t)cpipe->pipe_buffer.buffer,
1104 cpipe->pipe_buffer.size);
1105 cpipe->pipe_buffer.buffer = NULL;
1106 cpipe->pipe_buffer.object = NULL;
1111 * Close the pipe. The slock must be held to interlock against simultanious
1112 * closes. The rlock and wlock must be held to adjust the pipe_state.
1115 pipeclose(struct pipe *cpipe)
1124 * The slock may not have been allocated yet (close during
1127 * We need both the read and write tokens to modify pipe_state.
1129 if (cpipe->pipe_slock)
1130 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1131 lwkt_gettoken(&cpipe->pipe_rlock);
1132 lwkt_gettoken(&cpipe->pipe_wlock);
1135 * Set our state, wakeup anyone waiting in select/poll/kq, and
1136 * wakeup anyone blocked on our pipe.
1138 cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1139 pipewakeup(cpipe, 1);
1140 if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1141 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1146 * Disconnect from peer.
1148 if ((ppipe = cpipe->pipe_peer) != NULL) {
1149 lwkt_gettoken(&ppipe->pipe_rlock);
1150 lwkt_gettoken(&ppipe->pipe_wlock);
1151 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1152 pipewakeup(ppipe, 1);
1153 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1154 ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1157 if (SLIST_FIRST(&ppipe->pipe_kq.ki_note))
1158 KNOTE(&ppipe->pipe_kq.ki_note, 0);
1159 lwkt_reltoken(&ppipe->pipe_wlock);
1160 lwkt_reltoken(&ppipe->pipe_rlock);
1164 * If the peer is also closed we can free resources for both
1165 * sides, otherwise we leave our side intact to deal with any
1166 * races (since we only have the slock).
1168 if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1169 cpipe->pipe_peer = NULL;
1170 ppipe->pipe_peer = NULL;
1171 ppipe->pipe_slock = NULL; /* we will free the slock */
1176 lwkt_reltoken(&cpipe->pipe_wlock);
1177 lwkt_reltoken(&cpipe->pipe_rlock);
1178 if (cpipe->pipe_slock)
1179 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1182 * If we disassociated from our peer we can free resources
1184 if (ppipe == NULL) {
1186 if (cpipe->pipe_slock) {
1187 kfree(cpipe->pipe_slock, M_PIPE);
1188 cpipe->pipe_slock = NULL;
1190 if (gd->gd_pipeqcount >= pipe_maxcache ||
1191 cpipe->pipe_buffer.size != PIPE_SIZE
1193 pipe_free_kmem(cpipe);
1194 kfree(cpipe, M_PIPE);
1196 cpipe->pipe_state = 0;
1197 cpipe->pipe_peer = gd->gd_pipeq;
1198 gd->gd_pipeq = cpipe;
1199 ++gd->gd_pipeqcount;
1205 pipe_kqfilter(struct file *fp, struct knote *kn)
1209 cpipe = (struct pipe *)kn->kn_fp->f_data;
1211 switch (kn->kn_filter) {
1213 kn->kn_fop = &pipe_rfiltops;
1216 kn->kn_fop = &pipe_wfiltops;
1217 if (cpipe->pipe_peer == NULL) {
1218 /* other end of pipe has been closed */
1223 return (EOPNOTSUPP);
1225 kn->kn_hook = (caddr_t)cpipe;
1227 knote_insert(&cpipe->pipe_kq.ki_note, kn);
1233 filt_pipedetach(struct knote *kn)
1235 struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1237 knote_remove(&cpipe->pipe_kq.ki_note, kn);
1242 filt_piperead(struct knote *kn, long hint)
1244 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1247 lwkt_gettoken(&rpipe->pipe_rlock);
1248 lwkt_gettoken(&rpipe->pipe_wlock);
1250 kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1253 * Only set EOF if all data has been exhausted
1255 if ((rpipe->pipe_state & PIPE_REOF) && kn->kn_data == 0) {
1256 kn->kn_flags |= EV_EOF;
1260 lwkt_reltoken(&rpipe->pipe_wlock);
1261 lwkt_reltoken(&rpipe->pipe_rlock);
1264 ready = kn->kn_data > 0;
1271 filt_pipewrite(struct knote *kn, long hint)
1273 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1274 struct pipe *wpipe = rpipe->pipe_peer;
1278 if (wpipe == NULL) {
1279 kn->kn_flags |= EV_EOF;
1283 lwkt_gettoken(&wpipe->pipe_rlock);
1284 lwkt_gettoken(&wpipe->pipe_wlock);
1286 if (wpipe->pipe_state & PIPE_WEOF) {
1287 kn->kn_flags |= EV_EOF;
1292 kn->kn_data = wpipe->pipe_buffer.size -
1293 (wpipe->pipe_buffer.windex -
1294 wpipe->pipe_buffer.rindex);
1296 lwkt_reltoken(&wpipe->pipe_wlock);
1297 lwkt_reltoken(&wpipe->pipe_rlock);
1300 ready = kn->kn_data >= PIPE_BUF;