e5971f86bffe6a439c36a7f777c19e0025b0c7ff
[dragonfly.git] / sys / kern / sys_pipe.c
1 /*
2  * Copyright (c) 1996 John S. Dyson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
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
15  *    John S. Dyson.
16  * 4. Modifications may be freely made to this file if the above conditions
17  *    are met.
18  *
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 1.25 2004/11/12 00:09:24 dillon Exp $
21  */
22
23 /*
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
27  * do.
28  */
29
30 /*
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
37  * process.
38  *
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.
46  *
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.
51  */
52
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
56 #include <sys/proc.h>
57 #include <sys/fcntl.h>
58 #include <sys/file.h>
59 #include <sys/filedesc.h>
60 #include <sys/filio.h>
61 #include <sys/ttycom.h>
62 #include <sys/stat.h>
63 #include <sys/poll.h>
64 #include <sys/select.h>
65 #include <sys/signalvar.h>
66 #include <sys/sysproto.h>
67 #include <sys/pipe.h>
68 #include <sys/vnode.h>
69 #include <sys/uio.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>
75
76 #include <vm/vm.h>
77 #include <vm/vm_param.h>
78 #include <sys/lock.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_kern.h>
81 #include <vm/vm_extern.h>
82 #include <vm/pmap.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_zone.h>
86
87 #include <sys/file2.h>
88
89 #include <machine/cpufunc.h>
90
91 /*
92  * interfaces to the outside world
93  */
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,
100                 struct thread *td);
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);
104
105 static struct fileops pipeops = {
106         NULL,   /* port */
107         NULL,   /* clone */
108         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
109         pipe_stat, pipe_close
110 };
111
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);
115
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 };
120
121 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
122
123 /*
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.
128  */
129 #define MINPIPESIZE (PIPE_SIZE/3)
130 #define MAXPIPESIZE (2*PIPE_SIZE/3)
131
132 /*
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.
135  */
136 #define MAXPIPEKVA (8*1024*1024)
137
138 /*
139  * Limit for direct transfers, we cannot, of course limit
140  * the amount of kva for pipes in general though.
141  */
142 #define LIMITPIPEKVA (16*1024*1024)
143
144 /*
145  * Limit the number of "big" pipes
146  */
147 #define LIMITBIGPIPES   32
148 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
149
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 */
158
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");
176 #endif
177
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);
188 #endif
189 static int pipespace (struct pipe *cpipe, int size);
190
191 /*
192  * The pipe system call for the DTYPE_PIPE type of pipes
193  *
194  * pipe_ARgs(int dummy)
195  */
196
197 /* ARGSUSED */
198 int
199 pipe(struct pipe_args *uap)
200 {
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;
206         int fd1, fd2, error;
207
208         KKASSERT(p);
209         fdp = p->p_fd;
210
211         rpipe = wpipe = NULL;
212         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
213                 pipeclose(rpipe); 
214                 pipeclose(wpipe); 
215                 return (ENFILE);
216         }
217         
218         rpipe->pipe_state |= PIPE_DIRECTOK;
219         wpipe->pipe_state |= PIPE_DIRECTOK;
220
221         /*
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
224          * pipe is created.
225          *
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
229          * moment.
230          */
231         wpipe->pipe_feature = PIPE_COPY;
232         if (pipe_dwrite_enable) {
233                 switch(pipe_dwrite_sfbuf) {
234                 case 0:
235                         wpipe->pipe_feature = PIPE_KMEM;
236                         break;
237                 case 1:
238                         wpipe->pipe_feature = PIPE_SFBUF1;
239                         break;
240                 case 2:
241                 case 3:
242                         wpipe->pipe_feature = PIPE_SFBUF2;
243                         break;
244                 }
245         }
246         rpipe->pipe_feature = wpipe->pipe_feature;
247
248         error = falloc(p, &rf, &fd1);
249         if (error) {
250                 pipeclose(rpipe);
251                 pipeclose(wpipe);
252                 return (error);
253         }
254         uap->sysmsg_fds[0] = fd1;
255
256         /*
257          * Warning: once we've gotten past allocation of the fd for the
258          * read-side, we can only drop the read side via fdrop() in order
259          * to avoid races against processes which manage to dup() the read
260          * side while we are blocked trying to allocate the write side.
261          */
262         rf->f_flag = FREAD | FWRITE;
263         rf->f_type = DTYPE_PIPE;
264         rf->f_data = (caddr_t)rpipe;
265         rf->f_ops = &pipeops;
266         error = falloc(p, &wf, &fd2);
267         if (error) {
268                 if (fdp->fd_ofiles[fd1] == rf) {
269                         fdp->fd_ofiles[fd1] = NULL;
270                         fdrop(rf, td);
271                 }
272                 fdrop(rf, td);
273                 /* rpipe has been closed by fdrop(). */
274                 pipeclose(wpipe);
275                 return (error);
276         }
277         wf->f_flag = FREAD | FWRITE;
278         wf->f_type = DTYPE_PIPE;
279         wf->f_data = (caddr_t)wpipe;
280         wf->f_ops = &pipeops;
281         uap->sysmsg_fds[1] = fd2;
282
283         rpipe->pipe_peer = wpipe;
284         wpipe->pipe_peer = rpipe;
285         fdrop(rf, td);
286         fdrop(wf, td);
287
288         return (0);
289 }
290
291 /*
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.
296  */
297 static int
298 pipespace(struct pipe *cpipe, int size)
299 {
300         struct vm_object *object;
301         caddr_t buffer;
302         int npages, error;
303
304         npages = round_page(size) / PAGE_SIZE;
305         object = cpipe->pipe_buffer.object;
306
307         /*
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
311          * ones.
312          */
313         if (object == NULL || object->size != npages) {
314                 object = vm_object_allocate(OBJT_DEFAULT, npages);
315                 buffer = (caddr_t) vm_map_min(kernel_map);
316
317                 error = vm_map_find(kernel_map, object, 0,
318                         (vm_offset_t *) &buffer, size, 1,
319                         VM_PROT_ALL, VM_PROT_ALL, 0);
320
321                 if (error != KERN_SUCCESS) {
322                         vm_object_deallocate(object);
323                         return (ENOMEM);
324                 }
325                 pipe_free_kmem(cpipe);
326                 cpipe->pipe_buffer.object = object;
327                 cpipe->pipe_buffer.buffer = buffer;
328                 cpipe->pipe_buffer.size = size;
329                 ++pipe_bkmem_alloc;
330         } else {
331                 ++pipe_bcache_alloc;
332         }
333         cpipe->pipe_buffer.in = 0;
334         cpipe->pipe_buffer.out = 0;
335         cpipe->pipe_buffer.cnt = 0;
336         return (0);
337 }
338
339 /*
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.
343  */
344 static int
345 pipe_create(cpipep)
346         struct pipe **cpipep;
347 {
348         globaldata_t gd = mycpu;
349         struct pipe *cpipe;
350         int error;
351
352         if ((cpipe = gd->gd_pipeq) != NULL) {
353                 gd->gd_pipeq = cpipe->pipe_peer;
354                 --gd->gd_pipeqcount;
355                 cpipe->pipe_peer = NULL;
356         } else {
357                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
358         }
359         *cpipep = cpipe;
360         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
361                 return (error);
362         vfs_timestamp(&cpipe->pipe_ctime);
363         cpipe->pipe_atime = cpipe->pipe_ctime;
364         cpipe->pipe_mtime = cpipe->pipe_ctime;
365         return (0);
366 }
367
368
369 /*
370  * lock a pipe for I/O, blocking other access
371  */
372 static __inline int
373 pipelock(cpipe, catch)
374         struct pipe *cpipe;
375         int catch;
376 {
377         int error;
378
379         while (cpipe->pipe_state & PIPE_LOCK) {
380                 cpipe->pipe_state |= PIPE_LWANT;
381                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
382                 if (error != 0) 
383                         return (error);
384         }
385         cpipe->pipe_state |= PIPE_LOCK;
386         return (0);
387 }
388
389 /*
390  * unlock a pipe I/O lock
391  */
392 static __inline void
393 pipeunlock(cpipe)
394         struct pipe *cpipe;
395 {
396
397         cpipe->pipe_state &= ~PIPE_LOCK;
398         if (cpipe->pipe_state & PIPE_LWANT) {
399                 cpipe->pipe_state &= ~PIPE_LWANT;
400                 wakeup(cpipe);
401         }
402 }
403
404 static __inline void
405 pipeselwakeup(cpipe)
406         struct pipe *cpipe;
407 {
408
409         if (cpipe->pipe_state & PIPE_SEL) {
410                 cpipe->pipe_state &= ~PIPE_SEL;
411                 selwakeup(&cpipe->pipe_sel);
412         }
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);
416 }
417
418 /* ARGSUSED */
419 static int
420 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
421         int flags, struct thread *td)
422 {
423         struct pipe *rpipe = (struct pipe *) fp->f_data;
424         int error;
425         int nread = 0;
426         u_int size;
427
428         ++rpipe->pipe_busy;
429         error = pipelock(rpipe, 1);
430         if (error)
431                 goto unlocked_error;
432
433         while (uio->uio_resid) {
434                 caddr_t va;
435
436                 if (rpipe->pipe_buffer.cnt > 0) {
437                         /*
438                          * normal pipe buffer receive
439                          */
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;
445
446                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
447                                         size, uio);
448                         if (error)
449                                 break;
450
451                         rpipe->pipe_buffer.out += size;
452                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
453                                 rpipe->pipe_buffer.out = 0;
454
455                         rpipe->pipe_buffer.cnt -= size;
456
457                         /*
458                          * If there is no more to read in the pipe, reset
459                          * its pointers to the beginning.  This improves
460                          * cache hit stats.
461                          */
462                         if (rpipe->pipe_buffer.cnt == 0) {
463                                 rpipe->pipe_buffer.in = 0;
464                                 rpipe->pipe_buffer.out = 0;
465                         }
466                         nread += size;
467 #ifndef PIPE_NODIRECT
468                 } else if (rpipe->pipe_kva &&
469                            rpipe->pipe_feature == PIPE_KMEM &&
470                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) 
471                                == PIPE_DIRECTW
472                 ) {
473                         /*
474                          * Direct copy using source-side kva mapping
475                          */
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);
481                         if (error)
482                                 break;
483                         nread += size;
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;
489                                 wakeup(rpipe);
490                         }
491                 } else if (rpipe->pipe_map.xio_bytes &&
492                            rpipe->pipe_kva &&
493                            rpipe->pipe_feature == PIPE_SFBUF2 &&
494                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) 
495                                == PIPE_DIRECTW
496                 ) {
497                         /*
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.
504                          */
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);
516                         if (error)
517                                 break;
518                         nread += size;
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;
524                                 wakeup(rpipe);
525                         }
526                 } else if (rpipe->pipe_map.xio_bytes &&
527                            rpipe->pipe_feature == PIPE_SFBUF1 &&
528                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) 
529                                 == PIPE_DIRECTW
530                 ) {
531                         /*
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.
538                          */
539                         error = xio_uio_copy(&rpipe->pipe_map, uio, &size);
540                         if (error)
541                                 break;
542                         nread += size;
543                         if (rpipe->pipe_map.xio_bytes == 0) {
544                                 rpipe->pipe_state |= PIPE_DIRECTIP;
545                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
546                                 wakeup(rpipe);
547                         }
548 #endif
549                 } else {
550                         /*
551                          * detect EOF condition
552                          * read returns 0 on EOF, no need to set error
553                          */
554                         if (rpipe->pipe_state & PIPE_EOF)
555                                 break;
556
557                         /*
558                          * If the "write-side" has been blocked, wake it up now.
559                          */
560                         if (rpipe->pipe_state & PIPE_WANTW) {
561                                 rpipe->pipe_state &= ~PIPE_WANTW;
562                                 wakeup(rpipe);
563                         }
564
565                         /*
566                          * Break if some data was read.
567                          */
568                         if (nread > 0)
569                                 break;
570
571                         /*
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.
575                          */
576                         pipeunlock(rpipe);
577
578                         /*
579                          * Handle non-blocking mode operation or
580                          * wait for more data.
581                          */
582                         if (fp->f_flag & FNONBLOCK) {
583                                 error = EAGAIN;
584                         } else {
585                                 rpipe->pipe_state |= PIPE_WANTR;
586                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
587                                     "piperd", 0)) == 0) {
588                                         error = pipelock(rpipe, 1);
589                                 }
590                         }
591                         if (error)
592                                 goto unlocked_error;
593                 }
594         }
595         pipeunlock(rpipe);
596
597         if (error == 0)
598                 vfs_timestamp(&rpipe->pipe_atime);
599 unlocked_error:
600         --rpipe->pipe_busy;
601
602         /*
603          * PIPE_WANT processing only makes sense if pipe_busy is 0.
604          */
605         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
606                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
607                 wakeup(rpipe);
608         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
609                 /*
610                  * Handle write blocking hysteresis.
611                  */
612                 if (rpipe->pipe_state & PIPE_WANTW) {
613                         rpipe->pipe_state &= ~PIPE_WANTW;
614                         wakeup(rpipe);
615                 }
616         }
617
618         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
619                 pipeselwakeup(rpipe);
620         return (error);
621 }
622
623 #ifndef PIPE_NODIRECT
624 /*
625  * Map the sending processes' buffer into kernel space and wire it.
626  * This is similar to a physical write operation.
627  */
628 static int
629 pipe_build_write_buffer(wpipe, uio)
630         struct pipe *wpipe;
631         struct uio *uio;
632 {
633         int error;
634         u_int size;
635
636         size = (u_int) uio->uio_iov->iov_len;
637         if (size > wpipe->pipe_buffer.size)
638                 size = wpipe->pipe_buffer.size;
639
640         error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base, 
641                                 size, XIOF_READ);
642         if (error)
643                 return(error);
644
645         /*
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.
648          */
649         switch(wpipe->pipe_feature) {
650         case PIPE_KMEM:
651         case PIPE_SFBUF2:
652                 if (wpipe->pipe_kva == NULL) {
653                         wpipe->pipe_kva = 
654                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
655                         wpipe->pipe_kvamask = 0;
656                 }
657                 if (wpipe->pipe_feature == PIPE_KMEM) {
658                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
659                                     wpipe->pipe_map.xio_npages);
660                 }
661                 break;
662         default:
663                 break;
664         }
665
666         /*
667          * And update the uio data.  The XIO might have loaded fewer bytes
668          * then requested so reload 'size'.
669          */
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)
674                 uio->uio_iov++;
675         uio->uio_resid -= size;
676         uio->uio_offset += size;
677         return (0);
678 }
679
680 /*
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.
684  */
685 static void
686 pipe_clone_write_buffer(wpipe)
687         struct pipe *wpipe;
688 {
689         int size;
690
691         size = wpipe->pipe_map.xio_bytes;
692
693         KKASSERT(size <= wpipe->pipe_buffer.size);
694
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);
699
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;
706         }
707 }
708
709 /*
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.
715  */
716 static int
717 pipe_direct_write(wpipe, uio)
718         struct pipe *wpipe;
719         struct uio *uio;
720 {
721         int error;
722
723 retry:
724         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
725                 if (wpipe->pipe_state & PIPE_WANTR) {
726                         wpipe->pipe_state &= ~PIPE_WANTR;
727                         wakeup(wpipe);
728                 }
729                 wpipe->pipe_state |= PIPE_WANTW;
730                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
731                 if (error)
732                         goto error2;
733                 if (wpipe->pipe_state & PIPE_EOF) {
734                         error = EPIPE;
735                         goto error2;
736                 }
737         }
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;
742                         wakeup(wpipe);
743                 }
744                         
745                 wpipe->pipe_state |= PIPE_WANTW;
746                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
747                 if (error)
748                         goto error2;
749                 if (wpipe->pipe_state & PIPE_EOF) {
750                         error = EPIPE;
751                         goto error2;
752                 }
753                 goto retry;
754         }
755
756         /*
757          * Build our direct-write buffer
758          */
759         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
760         error = pipe_build_write_buffer(wpipe, uio);
761         if (error)
762                 goto error1;
763         wpipe->pipe_state &= ~PIPE_DIRECTIP;
764
765         /*
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().
769          */
770         error = 0;
771         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
772                 if (wpipe->pipe_state & PIPE_EOF) {
773                         pipelock(wpipe, 0);
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;
779                         }
780                         pipeunlock(wpipe);
781                         pipeselwakeup(wpipe);
782                         error = EPIPE;
783                         goto error1;
784                 }
785                 if (wpipe->pipe_state & PIPE_WANTR) {
786                         wpipe->pipe_state &= ~PIPE_WANTR;
787                         wakeup(wpipe);
788                 }
789                 pipeselwakeup(wpipe);
790                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
791         }
792         pipelock(wpipe,0);
793         if (wpipe->pipe_state & PIPE_DIRECTW) {
794                 /*
795                  * this bit of trickery substitutes a kernel buffer for
796                  * the process that might be going away.
797                  */
798                 pipe_clone_write_buffer(wpipe);
799                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
800         } else {
801                 /*
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
806                  * caching benefit. 
807                  *
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.
811                  */
812                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
813                 xio_release(&wpipe->pipe_map);
814                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
815         }
816         pipeunlock(wpipe);
817         return (error);
818
819         /*
820          * Direct-write error, clear the direct write flags.
821          */
822 error1:
823         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
824         /* fallthrough */
825
826         /*
827          * General error, wakeup the other side if it happens to be sleeping.
828          */
829 error2:
830         wakeup(wpipe);
831         return (error);
832 }
833 #endif
834         
835 static int
836 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
837         int flags, struct thread *td)
838 {
839         int error = 0;
840         int orig_resid;
841         struct pipe *wpipe, *rpipe;
842
843         rpipe = (struct pipe *) fp->f_data;
844         wpipe = rpipe->pipe_peer;
845
846         /*
847          * detect loss of pipe read side, issue SIGPIPE if lost.
848          */
849         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
850                 return (EPIPE);
851         }
852         ++wpipe->pipe_busy;
853
854         /*
855          * If it is advantageous to resize the pipe buffer, do
856          * so.
857          */
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)) {
863
864                 if ((error = pipelock(wpipe,1)) == 0) {
865                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
866                                 pipe_nbig++;
867                         pipeunlock(wpipe);
868                 }
869         }
870
871         /*
872          * If an early error occured unbusy and return, waking up any pending
873          * readers.
874          */
875         if (error) {
876                 --wpipe->pipe_busy;
877                 if ((wpipe->pipe_busy == 0) && 
878                     (wpipe->pipe_state & PIPE_WANT)) {
879                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
880                         wakeup(wpipe);
881                 }
882                 return(error);
883         }
884                 
885         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
886
887         orig_resid = uio->uio_resid;
888
889         while (uio->uio_resid) {
890                 int space;
891
892 #ifndef PIPE_NODIRECT
893                 /*
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.
898                  *
899                  * The direct write mechanism will detect the reader going
900                  * away on us.
901                  */
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);
907                         if (error)
908                                 break;
909                         continue;
910                 }
911 #endif
912
913                 /*
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
918                  * reader goes away.
919                  */
920         retrywrite:
921                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
922                         if (wpipe->pipe_state & PIPE_WANTR) {
923                                 wpipe->pipe_state &= ~PIPE_WANTR;
924                                 wakeup(wpipe);
925                         }
926                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
927                         if (wpipe->pipe_state & PIPE_EOF)
928                                 break;
929                         if (error)
930                                 break;
931                 }
932                 if (wpipe->pipe_state & PIPE_EOF) {
933                         error = EPIPE;
934                         break;
935                 }
936
937                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
938
939                 /* Writes of size <= PIPE_BUF must be atomic. */
940                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
941                         space = 0;
942
943                 /* 
944                  * Write to fill, read size handles write hysteresis.  Also
945                  * additional restrictions can cause select-based non-blocking
946                  * writes to spin.
947                  */
948                 if (space > 0) {
949                         if ((error = pipelock(wpipe,1)) == 0) {
950                                 int size;       /* Transfer size */
951                                 int segsize;    /* first segment to transfer */
952
953                                 /*
954                                  * It is possible for a direct write to
955                                  * slip in on us... handle it here...
956                                  */
957                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
958                                         pipeunlock(wpipe);
959                                         goto retrywrite;
960                                 }
961                                 /* 
962                                  * If a process blocked in uiomove, our
963                                  * value for space might be bad.
964                                  *
965                                  * XXX will we be ok if the reader has gone
966                                  * away here?
967                                  */
968                                 if (space > wpipe->pipe_buffer.size - 
969                                     wpipe->pipe_buffer.cnt) {
970                                         pipeunlock(wpipe);
971                                         goto retrywrite;
972                                 }
973
974                                 /*
975                                  * Transfer size is minimum of uio transfer
976                                  * and free space in pipe buffer.
977                                  */
978                                 if (space > uio->uio_resid)
979                                         size = uio->uio_resid;
980                                 else
981                                         size = space;
982                                 /*
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.
988                                  */
989                                 segsize = wpipe->pipe_buffer.size - 
990                                         wpipe->pipe_buffer.in;
991                                 if (segsize > size)
992                                         segsize = size;
993                                 
994                                 /* Transfer first segment */
995
996                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 
997                                                 segsize, uio);
998                                 
999                                 if (error == 0 && segsize < size) {
1000                                         /* 
1001                                          * Transfer remaining part now, to
1002                                          * support atomic writes.  Wraparound
1003                                          * happened.
1004                                          */
1005                                         if (wpipe->pipe_buffer.in + segsize != 
1006                                             wpipe->pipe_buffer.size)
1007                                                 panic("Expected pipe buffer wraparound disappeared");
1008                                                 
1009                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1010                                                         size - segsize, uio);
1011                                 }
1012                                 if (error == 0) {
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;
1019                                         }
1020                                 
1021                                         wpipe->pipe_buffer.cnt += size;
1022                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1023                                                 panic("Pipe buffer overflow");
1024                                 
1025                                 }
1026                                 pipeunlock(wpipe);
1027                         }
1028                         if (error)
1029                                 break;
1030
1031                 } else {
1032                         /*
1033                          * If the "read-side" has been blocked, wake it up now
1034                          * and yield to let it drain synchronously rather
1035                          * then block.
1036                          */
1037                         if (wpipe->pipe_state & PIPE_WANTR) {
1038                                 wpipe->pipe_state &= ~PIPE_WANTR;
1039                                 wakeup(wpipe);
1040                         }
1041
1042                         /*
1043                          * don't block on non-blocking I/O
1044                          */
1045                         if (fp->f_flag & FNONBLOCK) {
1046                                 error = EAGAIN;
1047                                 break;
1048                         }
1049
1050                         /*
1051                          * We have no more space and have something to offer,
1052                          * wake up select/poll.
1053                          */
1054                         pipeselwakeup(wpipe);
1055
1056                         wpipe->pipe_state |= PIPE_WANTW;
1057                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1058                         if (error != 0)
1059                                 break;
1060                         /*
1061                          * If read side wants to go away, we just issue a signal
1062                          * to ourselves.
1063                          */
1064                         if (wpipe->pipe_state & PIPE_EOF) {
1065                                 error = EPIPE;
1066                                 break;
1067                         }       
1068                 }
1069         }
1070
1071         --wpipe->pipe_busy;
1072
1073         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1074                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1075                 wakeup(wpipe);
1076         } else if (wpipe->pipe_buffer.cnt > 0) {
1077                 /*
1078                  * If we have put any characters in the buffer, we wake up
1079                  * the reader.
1080                  */
1081                 if (wpipe->pipe_state & PIPE_WANTR) {
1082                         wpipe->pipe_state &= ~PIPE_WANTR;
1083                         wakeup(wpipe);
1084                 }
1085         }
1086
1087         /*
1088          * Don't return EPIPE if I/O was successful
1089          */
1090         if ((wpipe->pipe_buffer.cnt == 0) &&
1091             (uio->uio_resid == 0) &&
1092             (error == EPIPE)) {
1093                 error = 0;
1094         }
1095
1096         if (error == 0)
1097                 vfs_timestamp(&wpipe->pipe_mtime);
1098
1099         /*
1100          * We have something to offer,
1101          * wake up select/poll.
1102          */
1103         if (wpipe->pipe_buffer.cnt)
1104                 pipeselwakeup(wpipe);
1105
1106         return (error);
1107 }
1108
1109 /*
1110  * we implement a very minimal set of ioctls for compatibility with sockets.
1111  */
1112 int
1113 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1114 {
1115         struct pipe *mpipe = (struct pipe *)fp->f_data;
1116
1117         switch (cmd) {
1118
1119         case FIONBIO:
1120                 return (0);
1121
1122         case FIOASYNC:
1123                 if (*(int *)data) {
1124                         mpipe->pipe_state |= PIPE_ASYNC;
1125                 } else {
1126                         mpipe->pipe_state &= ~PIPE_ASYNC;
1127                 }
1128                 return (0);
1129
1130         case FIONREAD:
1131                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1132                         *(int *)data = mpipe->pipe_map.xio_bytes;
1133                 } else {
1134                         *(int *)data = mpipe->pipe_buffer.cnt;
1135                 }
1136                 return (0);
1137
1138         case FIOSETOWN:
1139                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1140
1141         case FIOGETOWN:
1142                 *(int *)data = fgetown(mpipe->pipe_sigio);
1143                 return (0);
1144
1145         /* This is deprecated, FIOSETOWN should be used instead. */
1146         case TIOCSPGRP:
1147                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1148
1149         /* This is deprecated, FIOGETOWN should be used instead. */
1150         case TIOCGPGRP:
1151                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1152                 return (0);
1153
1154         }
1155         return (ENOTTY);
1156 }
1157
1158 int
1159 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1160 {
1161         struct pipe *rpipe = (struct pipe *)fp->f_data;
1162         struct pipe *wpipe;
1163         int revents = 0;
1164
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);
1171
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);
1177
1178         if ((rpipe->pipe_state & PIPE_EOF) ||
1179             (wpipe == NULL) ||
1180             (wpipe->pipe_state & PIPE_EOF))
1181                 revents |= POLLHUP;
1182
1183         if (revents == 0) {
1184                 if (events & (POLLIN | POLLRDNORM)) {
1185                         selrecord(td, &rpipe->pipe_sel);
1186                         rpipe->pipe_state |= PIPE_SEL;
1187                 }
1188
1189                 if (events & (POLLOUT | POLLWRNORM)) {
1190                         selrecord(td, &wpipe->pipe_sel);
1191                         wpipe->pipe_state |= PIPE_SEL;
1192                 }
1193         }
1194
1195         return (revents);
1196 }
1197
1198 static int
1199 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1200 {
1201         struct pipe *pipe = (struct pipe *)fp->f_data;
1202
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         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW))
1208                 ub->st_size = pipe->pipe_map.xio_bytes;
1209         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1210         ub->st_atimespec = pipe->pipe_atime;
1211         ub->st_mtimespec = pipe->pipe_mtime;
1212         ub->st_ctimespec = pipe->pipe_ctime;
1213         /*
1214          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1215          * st_flags, st_gen.
1216          * XXX (st_dev, st_ino) should be unique.
1217          */
1218         return (0);
1219 }
1220
1221 /* ARGSUSED */
1222 static int
1223 pipe_close(struct file *fp, struct thread *td)
1224 {
1225         struct pipe *cpipe = (struct pipe *)fp->f_data;
1226
1227         fp->f_ops = &badfileops;
1228         fp->f_data = NULL;
1229         funsetown(cpipe->pipe_sigio);
1230         pipeclose(cpipe);
1231         return (0);
1232 }
1233
1234 static void
1235 pipe_free_kmem(struct pipe *cpipe)
1236 {
1237         if (cpipe->pipe_buffer.buffer != NULL) {
1238                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1239                         --pipe_nbig;
1240                 kmem_free(kernel_map,
1241                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1242                         cpipe->pipe_buffer.size);
1243                 cpipe->pipe_buffer.buffer = NULL;
1244                 cpipe->pipe_buffer.object = NULL;
1245         }
1246 #ifndef PIPE_NODIRECT
1247         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1248                 cpipe->pipe_map.xio_offset == 0 &&
1249                 cpipe->pipe_map.xio_npages == 0);
1250 #endif
1251 }
1252
1253 /*
1254  * shutdown the pipe
1255  */
1256 static void
1257 pipeclose(struct pipe *cpipe)
1258 {
1259         globaldata_t gd;
1260         struct pipe *ppipe;
1261
1262         if (cpipe == NULL)
1263                 return;
1264
1265         pipeselwakeup(cpipe);
1266
1267         /*
1268          * If the other side is blocked, wake it up saying that
1269          * we want to close it down.
1270          */
1271         while (cpipe->pipe_busy) {
1272                 wakeup(cpipe);
1273                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1274                 tsleep(cpipe, 0, "pipecl", 0);
1275         }
1276
1277         /*
1278          * Disconnect from peer
1279          */
1280         if ((ppipe = cpipe->pipe_peer) != NULL) {
1281                 pipeselwakeup(ppipe);
1282
1283                 ppipe->pipe_state |= PIPE_EOF;
1284                 wakeup(ppipe);
1285                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1286                 ppipe->pipe_peer = NULL;
1287         }
1288
1289         if (cpipe->pipe_kva) {
1290                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1291                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1292                 cpipe->pipe_kva = NULL;
1293         }
1294
1295         /*
1296          * free or cache resources
1297          */
1298         gd = mycpu;
1299         if (gd->gd_pipeqcount >= pipe_maxcache ||
1300             cpipe->pipe_buffer.size != PIPE_SIZE
1301         ) {
1302                 pipe_free_kmem(cpipe);
1303                 free(cpipe, M_PIPE);
1304         } else {
1305                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1306                         cpipe->pipe_map.xio_bytes == 0 &&
1307                         cpipe->pipe_map.xio_offset == 0);
1308                 cpipe->pipe_state = 0;
1309                 cpipe->pipe_busy = 0;
1310                 cpipe->pipe_peer = gd->gd_pipeq;
1311                 gd->gd_pipeq = cpipe;
1312                 ++gd->gd_pipeqcount;
1313         }
1314 }
1315
1316 /*ARGSUSED*/
1317 static int
1318 pipe_kqfilter(struct file *fp, struct knote *kn)
1319 {
1320         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1321
1322         switch (kn->kn_filter) {
1323         case EVFILT_READ:
1324                 kn->kn_fop = &pipe_rfiltops;
1325                 break;
1326         case EVFILT_WRITE:
1327                 kn->kn_fop = &pipe_wfiltops;
1328                 cpipe = cpipe->pipe_peer;
1329                 if (cpipe == NULL)
1330                         /* other end of pipe has been closed */
1331                         return (EPIPE);
1332                 break;
1333         default:
1334                 return (1);
1335         }
1336         kn->kn_hook = (caddr_t)cpipe;
1337
1338         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1339         return (0);
1340 }
1341
1342 static void
1343 filt_pipedetach(struct knote *kn)
1344 {
1345         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1346
1347         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1348 }
1349
1350 /*ARGSUSED*/
1351 static int
1352 filt_piperead(struct knote *kn, long hint)
1353 {
1354         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1355         struct pipe *wpipe = rpipe->pipe_peer;
1356
1357         kn->kn_data = rpipe->pipe_buffer.cnt;
1358         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1359                 kn->kn_data = rpipe->pipe_map.xio_bytes;
1360
1361         if ((rpipe->pipe_state & PIPE_EOF) ||
1362             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1363                 kn->kn_flags |= EV_EOF; 
1364                 return (1);
1365         }
1366         return (kn->kn_data > 0);
1367 }
1368
1369 /*ARGSUSED*/
1370 static int
1371 filt_pipewrite(struct knote *kn, long hint)
1372 {
1373         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1374         struct pipe *wpipe = rpipe->pipe_peer;
1375
1376         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1377                 kn->kn_data = 0;
1378                 kn->kn_flags |= EV_EOF; 
1379                 return (1);
1380         }
1381         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1382         if (wpipe->pipe_state & PIPE_DIRECTW)
1383                 kn->kn_data = 0;
1384
1385         return (kn->kn_data >= PIPE_BUF);
1386 }