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 1.23 2004/05/21 14:26:28 hmp 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
31 * This code has two modes of operation, a small write mode and a large
32 * write mode. The small write mode acts like conventional pipes with
33 * a kernel buffer. If the buffer is less than PIPE_MINDIRECT, then the
34 * "normal" pipe buffering is done. If the buffer is between PIPE_MINDIRECT
35 * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
36 * the receiving process can copy it directly from the pages in the sending
39 * If the sending process receives a signal, it is possible that it will
40 * go away, and certainly its address space can change, because control
41 * is returned back to the user-mode side. In that case, the pipe code
42 * arranges to copy the buffer supplied by the user process, to a pageable
43 * kernel buffer, and the receiving process will grab the data from the
44 * pageable kernel buffer. Since signals don't happen all that often,
45 * the copy operation is normally eliminated.
47 * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
48 * happen for small transfers so that the system will not spend all of
49 * its time context switching. PIPE_SIZE is constrained by the
50 * amount of kernel virtual memory.
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
57 #include <sys/fcntl.h>
59 #include <sys/filedesc.h>
60 #include <sys/filio.h>
61 #include <sys/ttycom.h>
64 #include <sys/select.h>
65 #include <sys/signalvar.h>
66 #include <sys/sysproto.h>
68 #include <sys/vnode.h>
70 #include <sys/event.h>
71 #include <sys/globaldata.h>
72 #include <sys/module.h>
73 #include <sys/malloc.h>
74 #include <sys/sysctl.h>
77 #include <vm/vm_param.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_kern.h>
81 #include <vm/vm_extern.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_zone.h>
87 #include <sys/file2.h>
89 #include <machine/cpufunc.h>
92 * interfaces to the outside world
94 static int pipe_read (struct file *fp, struct uio *uio,
95 struct ucred *cred, int flags, struct thread *td);
96 static int pipe_write (struct file *fp, struct uio *uio,
97 struct ucred *cred, int flags, struct thread *td);
98 static int pipe_close (struct file *fp, struct thread *td);
99 static int pipe_poll (struct file *fp, int events, struct ucred *cred,
101 static int pipe_kqfilter (struct file *fp, struct knote *kn);
102 static int pipe_stat (struct file *fp, struct stat *sb, struct thread *td);
103 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct thread *td);
105 static struct fileops pipeops = {
108 pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
109 pipe_stat, pipe_close
112 static void filt_pipedetach(struct knote *kn);
113 static int filt_piperead(struct knote *kn, long hint);
114 static int filt_pipewrite(struct knote *kn, long hint);
116 static struct filterops pipe_rfiltops =
117 { 1, NULL, filt_pipedetach, filt_piperead };
118 static struct filterops pipe_wfiltops =
119 { 1, NULL, filt_pipedetach, filt_pipewrite };
121 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
124 * Default pipe buffer size(s), this can be kind-of large now because pipe
125 * space is pageable. The pipe code will try to maintain locality of
126 * reference for performance reasons, so small amounts of outstanding I/O
127 * will not wipe the cache.
129 #define MINPIPESIZE (PIPE_SIZE/3)
130 #define MAXPIPESIZE (2*PIPE_SIZE/3)
133 * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
134 * is there so that on large systems, we don't exhaust it.
136 #define MAXPIPEKVA (8*1024*1024)
139 * Limit for direct transfers, we cannot, of course limit
140 * the amount of kva for pipes in general though.
142 #define LIMITPIPEKVA (16*1024*1024)
145 * Limit the number of "big" pipes
147 #define LIMITBIGPIPES 32
148 #define PIPEQ_MAX_CACHE 16 /* per-cpu pipe structure cache */
150 static int pipe_maxbig = LIMITBIGPIPES;
151 static int pipe_maxcache = PIPEQ_MAX_CACHE;
152 static int pipe_nbig;
153 static int pipe_bcache_alloc;
154 static int pipe_bkmem_alloc;
155 static int pipe_dwrite_enable = 1; /* 0:copy, 1:kmem/sfbuf 2:force */
156 static int pipe_dwrite_sfbuf = 1; /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
157 /* 3:sfbuf_dmap w/ forced invlpg */
159 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
160 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
161 CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
162 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
163 CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
164 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
165 CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
166 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
167 CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
168 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
169 CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
170 "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
171 #if !defined(NO_PIPE_SYSCTL_STATS)
172 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
173 CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
174 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
175 CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
178 static void pipeclose (struct pipe *cpipe);
179 static void pipe_free_kmem (struct pipe *cpipe);
180 static int pipe_create (struct pipe **cpipep);
181 static __inline int pipelock (struct pipe *cpipe, int catch);
182 static __inline void pipeunlock (struct pipe *cpipe);
183 static __inline void pipeselwakeup (struct pipe *cpipe);
184 #ifndef PIPE_NODIRECT
185 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
186 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
187 static void pipe_clone_write_buffer (struct pipe *wpipe);
189 static int pipespace (struct pipe *cpipe, int size);
192 * The pipe system call for the DTYPE_PIPE type of pipes
194 * pipe_ARgs(int dummy)
199 pipe(struct pipe_args *uap)
201 struct thread *td = curthread;
202 struct proc *p = td->td_proc;
203 struct filedesc *fdp;
204 struct file *rf, *wf;
205 struct pipe *rpipe, *wpipe;
211 rpipe = wpipe = NULL;
212 if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
218 rpipe->pipe_state |= PIPE_DIRECTOK;
219 wpipe->pipe_state |= PIPE_DIRECTOK;
222 * Select the direct-map features to use for this pipe. Since the
223 * sysctl's can change on the fly we record the settings when the
226 * Generally speaking the system will default to what we consider
227 * to be the best-balanced and most stable option. Right now this
228 * is SFBUF1. Modes 2 and 3 are considered experiemental at the
231 wpipe->pipe_feature = PIPE_COPY;
232 if (pipe_dwrite_enable) {
233 switch(pipe_dwrite_sfbuf) {
235 wpipe->pipe_feature = PIPE_KMEM;
238 wpipe->pipe_feature = PIPE_SFBUF1;
242 wpipe->pipe_feature = PIPE_SFBUF2;
246 rpipe->pipe_feature = wpipe->pipe_feature;
248 error = falloc(p, &rf, &fd1);
255 uap->sysmsg_fds[0] = fd1;
258 * Warning: once we've gotten past allocation of the fd for the
259 * read-side, we can only drop the read side via fdrop() in order
260 * to avoid races against processes which manage to dup() the read
261 * side while we are blocked trying to allocate the write side.
263 rf->f_flag = FREAD | FWRITE;
264 rf->f_type = DTYPE_PIPE;
265 rf->f_data = (caddr_t)rpipe;
266 rf->f_ops = &pipeops;
267 error = falloc(p, &wf, &fd2);
269 if (fdp->fd_ofiles[fd1] == rf) {
270 fdp->fd_ofiles[fd1] = NULL;
274 /* rpipe has been closed by fdrop(). */
278 wf->f_flag = FREAD | FWRITE;
279 wf->f_type = DTYPE_PIPE;
280 wf->f_data = (caddr_t)wpipe;
281 wf->f_ops = &pipeops;
282 uap->sysmsg_fds[1] = fd2;
284 rpipe->pipe_peer = wpipe;
285 wpipe->pipe_peer = rpipe;
292 * Allocate kva for pipe circular buffer, the space is pageable
293 * This routine will 'realloc' the size of a pipe safely, if it fails
294 * it will retain the old buffer.
295 * If it fails it will return ENOMEM.
298 pipespace(struct pipe *cpipe, int size)
300 struct vm_object *object;
304 npages = round_page(size) / PAGE_SIZE;
305 object = cpipe->pipe_buffer.object;
308 * [re]create the object if necessary and reserve space for it
309 * in the kernel_map. The object and memory are pageable. On
310 * success, free the old resources before assigning the new
313 if (object == NULL || object->size != npages) {
314 object = vm_object_allocate(OBJT_DEFAULT, npages);
315 buffer = (caddr_t) vm_map_min(kernel_map);
317 error = vm_map_find(kernel_map, object, 0,
318 (vm_offset_t *) &buffer, size, 1,
319 VM_PROT_ALL, VM_PROT_ALL, 0);
321 if (error != KERN_SUCCESS) {
322 vm_object_deallocate(object);
325 pipe_free_kmem(cpipe);
326 cpipe->pipe_buffer.object = object;
327 cpipe->pipe_buffer.buffer = buffer;
328 cpipe->pipe_buffer.size = size;
333 cpipe->pipe_buffer.in = 0;
334 cpipe->pipe_buffer.out = 0;
335 cpipe->pipe_buffer.cnt = 0;
340 * Initialize and allocate VM and memory for pipe, pulling the pipe from
341 * our per-cpu cache if possible. For now make sure it is sized for the
342 * smaller PIPE_SIZE default.
346 struct pipe **cpipep;
348 globaldata_t gd = mycpu;
352 if ((cpipe = gd->gd_pipeq) != NULL) {
353 gd->gd_pipeq = cpipe->pipe_peer;
355 cpipe->pipe_peer = NULL;
357 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
360 if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
362 vfs_timestamp(&cpipe->pipe_ctime);
363 cpipe->pipe_atime = cpipe->pipe_ctime;
364 cpipe->pipe_mtime = cpipe->pipe_ctime;
370 * lock a pipe for I/O, blocking other access
373 pipelock(cpipe, catch)
379 while (cpipe->pipe_state & PIPE_LOCK) {
380 cpipe->pipe_state |= PIPE_LWANT;
381 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
385 cpipe->pipe_state |= PIPE_LOCK;
390 * unlock a pipe I/O lock
397 cpipe->pipe_state &= ~PIPE_LOCK;
398 if (cpipe->pipe_state & PIPE_LWANT) {
399 cpipe->pipe_state &= ~PIPE_LWANT;
409 if (cpipe->pipe_state & PIPE_SEL) {
410 cpipe->pipe_state &= ~PIPE_SEL;
411 selwakeup(&cpipe->pipe_sel);
413 if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
414 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
415 KNOTE(&cpipe->pipe_sel.si_note, 0);
420 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
421 int flags, struct thread *td)
423 struct pipe *rpipe = (struct pipe *) fp->f_data;
429 error = pipelock(rpipe, 1);
433 while (uio->uio_resid) {
436 if (rpipe->pipe_buffer.cnt > 0) {
438 * normal pipe buffer receive
440 size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
441 if (size > rpipe->pipe_buffer.cnt)
442 size = rpipe->pipe_buffer.cnt;
443 if (size > (u_int) uio->uio_resid)
444 size = (u_int) uio->uio_resid;
446 error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
451 rpipe->pipe_buffer.out += size;
452 if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
453 rpipe->pipe_buffer.out = 0;
455 rpipe->pipe_buffer.cnt -= size;
458 * If there is no more to read in the pipe, reset
459 * its pointers to the beginning. This improves
462 if (rpipe->pipe_buffer.cnt == 0) {
463 rpipe->pipe_buffer.in = 0;
464 rpipe->pipe_buffer.out = 0;
467 #ifndef PIPE_NODIRECT
468 } else if (rpipe->pipe_kva &&
469 rpipe->pipe_feature == PIPE_KMEM &&
470 (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
474 * Direct copy using source-side kva mapping
476 size = rpipe->pipe_map.xio_bytes;
477 if (size > (u_int)uio->uio_resid)
478 size = (u_int)uio->uio_resid;
479 va = (caddr_t)rpipe->pipe_kva + rpipe->pipe_map.xio_offset;
480 error = uiomove(va, size, uio);
484 rpipe->pipe_map.xio_offset += size;
485 rpipe->pipe_map.xio_bytes -= size;
486 if (rpipe->pipe_map.xio_bytes == 0) {
487 rpipe->pipe_state |= PIPE_DIRECTIP;
488 rpipe->pipe_state &= ~PIPE_DIRECTW;
491 } else if (rpipe->pipe_map.xio_bytes &&
493 rpipe->pipe_feature == PIPE_SFBUF2 &&
494 (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
498 * Direct copy, bypassing a kernel buffer. We cannot
499 * mess with the direct-write buffer until
500 * PIPE_DIRECTIP is cleared. In order to prevent
501 * the pipe_write code from racing itself in
502 * direct_write, we set DIRECTIP when we clear
503 * DIRECTW after we have exhausted the buffer.
505 if (pipe_dwrite_sfbuf == 3)
506 rpipe->pipe_kvamask = 0;
507 pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
508 rpipe->pipe_map.xio_npages,
509 &rpipe->pipe_kvamask);
510 size = rpipe->pipe_map.xio_bytes;
511 if (size > (u_int)uio->uio_resid)
512 size = (u_int)uio->uio_resid;
513 va = (caddr_t)rpipe->pipe_kva +
514 rpipe->pipe_map.xio_offset;
515 error = uiomove(va, size, uio);
519 rpipe->pipe_map.xio_offset += size;
520 rpipe->pipe_map.xio_bytes -= size;
521 if (rpipe->pipe_map.xio_bytes == 0) {
522 rpipe->pipe_state |= PIPE_DIRECTIP;
523 rpipe->pipe_state &= ~PIPE_DIRECTW;
526 } else if (rpipe->pipe_map.xio_bytes &&
527 rpipe->pipe_feature == PIPE_SFBUF1 &&
528 (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
532 * Direct copy, bypassing a kernel buffer. We cannot
533 * mess with the direct-write buffer until
534 * PIPE_DIRECTIP is cleared. In order to prevent
535 * the pipe_write code from racing itself in
536 * direct_write, we set DIRECTIP when we clear
537 * DIRECTW after we have exhausted the buffer.
539 error = xio_uio_copy(&rpipe->pipe_map, uio, &size);
543 if (rpipe->pipe_map.xio_bytes == 0) {
544 rpipe->pipe_state |= PIPE_DIRECTIP;
545 rpipe->pipe_state &= ~PIPE_DIRECTW;
551 * detect EOF condition
552 * read returns 0 on EOF, no need to set error
554 if (rpipe->pipe_state & PIPE_EOF)
558 * If the "write-side" has been blocked, wake it up now.
560 if (rpipe->pipe_state & PIPE_WANTW) {
561 rpipe->pipe_state &= ~PIPE_WANTW;
566 * Break if some data was read.
572 * Unlock the pipe buffer for our remaining
573 * processing. We will either break out with an
574 * error or we will sleep and relock to loop.
579 * Handle non-blocking mode operation or
580 * wait for more data.
582 if (fp->f_flag & FNONBLOCK) {
585 rpipe->pipe_state |= PIPE_WANTR;
586 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
587 "piperd", 0)) == 0) {
588 error = pipelock(rpipe, 1);
598 vfs_timestamp(&rpipe->pipe_atime);
603 * PIPE_WANT processing only makes sense if pipe_busy is 0.
605 if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
606 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
608 } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
610 * Handle write blocking hysteresis.
612 if (rpipe->pipe_state & PIPE_WANTW) {
613 rpipe->pipe_state &= ~PIPE_WANTW;
618 if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
619 pipeselwakeup(rpipe);
623 #ifndef PIPE_NODIRECT
625 * Map the sending processes' buffer into kernel space and wire it.
626 * This is similar to a physical write operation.
629 pipe_build_write_buffer(wpipe, uio)
636 size = (u_int) uio->uio_iov->iov_len;
637 if (size > wpipe->pipe_buffer.size)
638 size = wpipe->pipe_buffer.size;
640 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
646 * Create a kernel map for KMEM and SFBUF2 copy modes. SFBUF2 will
647 * map the pages on the target while KMEM maps the pages now.
649 switch(wpipe->pipe_feature) {
652 if (wpipe->pipe_kva == NULL) {
654 kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
655 wpipe->pipe_kvamask = 0;
657 if (wpipe->pipe_feature == PIPE_KMEM) {
658 pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
659 wpipe->pipe_map.xio_npages);
667 * And update the uio data. The XIO might have loaded fewer bytes
668 * then requested so reload 'size'.
670 size = wpipe->pipe_map.xio_bytes;
671 uio->uio_iov->iov_len -= size;
672 uio->uio_iov->iov_base += size;
673 if (uio->uio_iov->iov_len == 0)
675 uio->uio_resid -= size;
676 uio->uio_offset += size;
681 * In the case of a signal, the writing process might go away. This
682 * code copies the data into the circular buffer so that the source
683 * pages can be freed without loss of data.
686 pipe_clone_write_buffer(wpipe)
691 size = wpipe->pipe_map.xio_bytes;
693 KKASSERT(size <= wpipe->pipe_buffer.size);
695 wpipe->pipe_buffer.in = size;
696 wpipe->pipe_buffer.out = 0;
697 wpipe->pipe_buffer.cnt = size;
698 wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
700 xio_copy_xtok(&wpipe->pipe_map, wpipe->pipe_buffer.buffer, size);
701 xio_release(&wpipe->pipe_map);
702 if (wpipe->pipe_kva) {
703 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
704 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
705 wpipe->pipe_kva = NULL;
710 * This implements the pipe buffer write mechanism. Note that only
711 * a direct write OR a normal pipe write can be pending at any given time.
712 * If there are any characters in the pipe buffer, the direct write will
713 * be deferred until the receiving process grabs all of the bytes from
714 * the pipe buffer. Then the direct mapping write is set-up.
717 pipe_direct_write(wpipe, uio)
724 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
725 if (wpipe->pipe_state & PIPE_WANTR) {
726 wpipe->pipe_state &= ~PIPE_WANTR;
729 wpipe->pipe_state |= PIPE_WANTW;
730 error = tsleep(wpipe, PCATCH, "pipdww", 0);
733 if (wpipe->pipe_state & PIPE_EOF) {
738 KKASSERT(wpipe->pipe_map.xio_bytes == 0);
739 if (wpipe->pipe_buffer.cnt > 0) {
740 if (wpipe->pipe_state & PIPE_WANTR) {
741 wpipe->pipe_state &= ~PIPE_WANTR;
745 wpipe->pipe_state |= PIPE_WANTW;
746 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
749 if (wpipe->pipe_state & PIPE_EOF) {
757 * Build our direct-write buffer
759 wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
760 error = pipe_build_write_buffer(wpipe, uio);
763 wpipe->pipe_state &= ~PIPE_DIRECTIP;
766 * Wait until the receiver has snarfed the data. Since we are likely
767 * going to sleep we optimize the case and yield synchronously,
768 * possibly avoiding the tsleep().
771 while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
772 if (wpipe->pipe_state & PIPE_EOF) {
774 xio_release(&wpipe->pipe_map);
775 if (wpipe->pipe_kva) {
776 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
777 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
778 wpipe->pipe_kva = NULL;
781 pipeselwakeup(wpipe);
785 if (wpipe->pipe_state & PIPE_WANTR) {
786 wpipe->pipe_state &= ~PIPE_WANTR;
789 pipeselwakeup(wpipe);
790 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
793 if (wpipe->pipe_state & PIPE_DIRECTW) {
795 * this bit of trickery substitutes a kernel buffer for
796 * the process that might be going away.
798 pipe_clone_write_buffer(wpipe);
799 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
802 * note: The pipe_kva mapping is not qremove'd here. For
803 * legacy PIPE_KMEM mode this constitutes an improvement
804 * over the original FreeBSD-4 algorithm. For PIPE_SFBUF2
805 * mode the kva mapping must not be removed to get the
808 * For testing purposes we will give the original algorithm
809 * the benefit of the doubt 'what it could have been', and
810 * keep the optimization.
812 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
813 xio_release(&wpipe->pipe_map);
814 wpipe->pipe_state &= ~PIPE_DIRECTIP;
820 * Direct-write error, clear the direct write flags.
823 wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
827 * General error, wakeup the other side if it happens to be sleeping.
836 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
837 int flags, struct thread *td)
841 struct pipe *wpipe, *rpipe;
843 rpipe = (struct pipe *) fp->f_data;
844 wpipe = rpipe->pipe_peer;
847 * detect loss of pipe read side, issue SIGPIPE if lost.
849 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
855 * If it is advantageous to resize the pipe buffer, do
858 if ((uio->uio_resid > PIPE_SIZE) &&
859 (pipe_nbig < pipe_maxbig) &&
860 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
861 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
862 (wpipe->pipe_buffer.cnt == 0)) {
864 if ((error = pipelock(wpipe,1)) == 0) {
865 if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
872 * If an early error occured unbusy and return, waking up any pending
877 if ((wpipe->pipe_busy == 0) &&
878 (wpipe->pipe_state & PIPE_WANT)) {
879 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
885 KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
887 orig_resid = uio->uio_resid;
889 while (uio->uio_resid) {
892 #ifndef PIPE_NODIRECT
894 * If the transfer is large, we can gain performance if
895 * we do process-to-process copies directly.
896 * If the write is non-blocking, we don't use the
897 * direct write mechanism.
899 * The direct write mechanism will detect the reader going
902 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
903 pipe_dwrite_enable > 1) &&
904 (fp->f_flag & FNONBLOCK) == 0 &&
905 pipe_dwrite_enable) {
906 error = pipe_direct_write( wpipe, uio);
914 * Pipe buffered writes cannot be coincidental with
915 * direct writes. We wait until the currently executing
916 * direct write is completed before we start filling the
917 * pipe buffer. We break out if a signal occurs or the
921 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
922 if (wpipe->pipe_state & PIPE_WANTR) {
923 wpipe->pipe_state &= ~PIPE_WANTR;
926 error = tsleep(wpipe, PCATCH, "pipbww", 0);
927 if (wpipe->pipe_state & PIPE_EOF)
932 if (wpipe->pipe_state & PIPE_EOF) {
937 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
939 /* Writes of size <= PIPE_BUF must be atomic. */
940 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
944 * Write to fill, read size handles write hysteresis. Also
945 * additional restrictions can cause select-based non-blocking
949 if ((error = pipelock(wpipe,1)) == 0) {
950 int size; /* Transfer size */
951 int segsize; /* first segment to transfer */
954 * It is possible for a direct write to
955 * slip in on us... handle it here...
957 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
962 * If a process blocked in uiomove, our
963 * value for space might be bad.
965 * XXX will we be ok if the reader has gone
968 if (space > wpipe->pipe_buffer.size -
969 wpipe->pipe_buffer.cnt) {
975 * Transfer size is minimum of uio transfer
976 * and free space in pipe buffer.
978 if (space > uio->uio_resid)
979 size = uio->uio_resid;
983 * First segment to transfer is minimum of
984 * transfer size and contiguous space in
985 * pipe buffer. If first segment to transfer
986 * is less than the transfer size, we've got
987 * a wraparound in the buffer.
989 segsize = wpipe->pipe_buffer.size -
990 wpipe->pipe_buffer.in;
994 /* Transfer first segment */
996 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
999 if (error == 0 && segsize < size) {
1001 * Transfer remaining part now, to
1002 * support atomic writes. Wraparound
1005 if (wpipe->pipe_buffer.in + segsize !=
1006 wpipe->pipe_buffer.size)
1007 panic("Expected pipe buffer wraparound disappeared");
1009 error = uiomove(&wpipe->pipe_buffer.buffer[0],
1010 size - segsize, uio);
1013 wpipe->pipe_buffer.in += size;
1014 if (wpipe->pipe_buffer.in >=
1015 wpipe->pipe_buffer.size) {
1016 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1017 panic("Expected wraparound bad");
1018 wpipe->pipe_buffer.in = size - segsize;
1021 wpipe->pipe_buffer.cnt += size;
1022 if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1023 panic("Pipe buffer overflow");
1033 * If the "read-side" has been blocked, wake it up now
1034 * and yield to let it drain synchronously rather
1037 if (wpipe->pipe_state & PIPE_WANTR) {
1038 wpipe->pipe_state &= ~PIPE_WANTR;
1043 * don't block on non-blocking I/O
1045 if (fp->f_flag & FNONBLOCK) {
1051 * We have no more space and have something to offer,
1052 * wake up select/poll.
1054 pipeselwakeup(wpipe);
1056 wpipe->pipe_state |= PIPE_WANTW;
1057 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1061 * If read side wants to go away, we just issue a signal
1064 if (wpipe->pipe_state & PIPE_EOF) {
1073 if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1074 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1076 } else if (wpipe->pipe_buffer.cnt > 0) {
1078 * If we have put any characters in the buffer, we wake up
1081 if (wpipe->pipe_state & PIPE_WANTR) {
1082 wpipe->pipe_state &= ~PIPE_WANTR;
1088 * Don't return EPIPE if I/O was successful
1090 if ((wpipe->pipe_buffer.cnt == 0) &&
1091 (uio->uio_resid == 0) &&
1097 vfs_timestamp(&wpipe->pipe_mtime);
1100 * We have something to offer,
1101 * wake up select/poll.
1103 if (wpipe->pipe_buffer.cnt)
1104 pipeselwakeup(wpipe);
1110 * we implement a very minimal set of ioctls for compatibility with sockets.
1113 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1115 struct pipe *mpipe = (struct pipe *)fp->f_data;
1124 mpipe->pipe_state |= PIPE_ASYNC;
1126 mpipe->pipe_state &= ~PIPE_ASYNC;
1131 if (mpipe->pipe_state & PIPE_DIRECTW) {
1132 *(int *)data = mpipe->pipe_map.xio_bytes;
1134 *(int *)data = mpipe->pipe_buffer.cnt;
1139 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1142 *(int *)data = fgetown(mpipe->pipe_sigio);
1145 /* This is deprecated, FIOSETOWN should be used instead. */
1147 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1149 /* This is deprecated, FIOGETOWN should be used instead. */
1151 *(int *)data = -fgetown(mpipe->pipe_sigio);
1159 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1161 struct pipe *rpipe = (struct pipe *)fp->f_data;
1165 wpipe = rpipe->pipe_peer;
1166 if (events & (POLLIN | POLLRDNORM))
1167 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1168 (rpipe->pipe_buffer.cnt > 0) ||
1169 (rpipe->pipe_state & PIPE_EOF))
1170 revents |= events & (POLLIN | POLLRDNORM);
1172 if (events & (POLLOUT | POLLWRNORM))
1173 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1174 (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1175 (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1176 revents |= events & (POLLOUT | POLLWRNORM);
1178 if ((rpipe->pipe_state & PIPE_EOF) ||
1180 (wpipe->pipe_state & PIPE_EOF))
1184 if (events & (POLLIN | POLLRDNORM)) {
1185 selrecord(td, &rpipe->pipe_sel);
1186 rpipe->pipe_state |= PIPE_SEL;
1189 if (events & (POLLOUT | POLLWRNORM)) {
1190 selrecord(td, &wpipe->pipe_sel);
1191 wpipe->pipe_state |= PIPE_SEL;
1199 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1201 struct pipe *pipe = (struct pipe *)fp->f_data;
1203 bzero((caddr_t)ub, sizeof(*ub));
1204 ub->st_mode = S_IFIFO;
1205 ub->st_blksize = pipe->pipe_buffer.size;
1206 ub->st_size = pipe->pipe_buffer.cnt;
1207 ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1208 ub->st_atimespec = pipe->pipe_atime;
1209 ub->st_mtimespec = pipe->pipe_mtime;
1210 ub->st_ctimespec = pipe->pipe_ctime;
1212 * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1214 * XXX (st_dev, st_ino) should be unique.
1221 pipe_close(struct file *fp, struct thread *td)
1223 struct pipe *cpipe = (struct pipe *)fp->f_data;
1225 fp->f_ops = &badfileops;
1227 funsetown(cpipe->pipe_sigio);
1233 pipe_free_kmem(struct pipe *cpipe)
1235 if (cpipe->pipe_buffer.buffer != NULL) {
1236 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1238 kmem_free(kernel_map,
1239 (vm_offset_t)cpipe->pipe_buffer.buffer,
1240 cpipe->pipe_buffer.size);
1241 cpipe->pipe_buffer.buffer = NULL;
1242 cpipe->pipe_buffer.object = NULL;
1244 #ifndef PIPE_NODIRECT
1245 KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1246 cpipe->pipe_map.xio_offset == 0 &&
1247 cpipe->pipe_map.xio_npages == 0);
1255 pipeclose(struct pipe *cpipe)
1263 pipeselwakeup(cpipe);
1266 * If the other side is blocked, wake it up saying that
1267 * we want to close it down.
1269 while (cpipe->pipe_busy) {
1271 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1272 tsleep(cpipe, 0, "pipecl", 0);
1276 * Disconnect from peer
1278 if ((ppipe = cpipe->pipe_peer) != NULL) {
1279 pipeselwakeup(ppipe);
1281 ppipe->pipe_state |= PIPE_EOF;
1283 KNOTE(&ppipe->pipe_sel.si_note, 0);
1284 ppipe->pipe_peer = NULL;
1287 if (cpipe->pipe_kva) {
1288 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1289 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1290 cpipe->pipe_kva = NULL;
1294 * free or cache resources
1297 if (gd->gd_pipeqcount >= pipe_maxcache ||
1298 cpipe->pipe_buffer.size != PIPE_SIZE
1300 pipe_free_kmem(cpipe);
1301 free(cpipe, M_PIPE);
1303 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1304 cpipe->pipe_map.xio_bytes == 0 &&
1305 cpipe->pipe_map.xio_offset == 0);
1306 cpipe->pipe_state = 0;
1307 cpipe->pipe_busy = 0;
1308 cpipe->pipe_peer = gd->gd_pipeq;
1309 gd->gd_pipeq = cpipe;
1310 ++gd->gd_pipeqcount;
1316 pipe_kqfilter(struct file *fp, struct knote *kn)
1318 struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1320 switch (kn->kn_filter) {
1322 kn->kn_fop = &pipe_rfiltops;
1325 kn->kn_fop = &pipe_wfiltops;
1326 cpipe = cpipe->pipe_peer;
1328 /* other end of pipe has been closed */
1334 kn->kn_hook = (caddr_t)cpipe;
1336 SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1341 filt_pipedetach(struct knote *kn)
1343 struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1345 SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1350 filt_piperead(struct knote *kn, long hint)
1352 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1353 struct pipe *wpipe = rpipe->pipe_peer;
1355 kn->kn_data = rpipe->pipe_buffer.cnt;
1356 if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1357 kn->kn_data = rpipe->pipe_map.xio_bytes;
1359 if ((rpipe->pipe_state & PIPE_EOF) ||
1360 (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1361 kn->kn_flags |= EV_EOF;
1364 return (kn->kn_data > 0);
1369 filt_pipewrite(struct knote *kn, long hint)
1371 struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1372 struct pipe *wpipe = rpipe->pipe_peer;
1374 if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1376 kn->kn_flags |= EV_EOF;
1379 kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1380 if (wpipe->pipe_state & PIPE_DIRECTW)
1383 return (kn->kn_data >= PIPE_BUF);