Clean up the XIO API and structure. XIO no longer tries to 'track' partial
[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 1.26 2005/03/01 23:35:14 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                                 rpipe->pipe_buffer.out;
478                         if (size > (u_int)uio->uio_resid)
479                                 size = (u_int)uio->uio_resid;
480                         va = (caddr_t)rpipe->pipe_kva + 
481                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
482                         error = uiomove(va, size, uio);
483                         if (error)
484                                 break;
485                         nread += size;
486                         rpipe->pipe_buffer.out += size;
487                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
488                                 rpipe->pipe_state |= PIPE_DIRECTIP;
489                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
490                                 wakeup(rpipe);
491                         }
492                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
493                            rpipe->pipe_kva &&
494                            rpipe->pipe_feature == PIPE_SFBUF2 &&
495                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) 
496                                == PIPE_DIRECTW
497                 ) {
498                         /*
499                          * Direct copy, bypassing a kernel buffer.  We cannot
500                          * mess with the direct-write buffer until
501                          * PIPE_DIRECTIP is cleared.  In order to prevent 
502                          * the pipe_write code from racing itself in
503                          * direct_write, we set DIRECTIP when we clear
504                          * DIRECTW after we have exhausted the buffer.
505                          */
506                         if (pipe_dwrite_sfbuf == 3)
507                                 rpipe->pipe_kvamask = 0;
508                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
509                                     rpipe->pipe_map.xio_npages,
510                                     &rpipe->pipe_kvamask);
511                         size = rpipe->pipe_map.xio_bytes - 
512                                 rpipe->pipe_buffer.out;
513                         if (size > (u_int)uio->uio_resid)
514                                 size = (u_int)uio->uio_resid;
515                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
516                         error = uiomove(va, size, uio);
517                         if (error)
518                                 break;
519                         nread += size;
520                         rpipe->pipe_buffer.out += size;
521                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
522                                 rpipe->pipe_state |= PIPE_DIRECTIP;
523                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
524                                 wakeup(rpipe);
525                         }
526                 } else if (rpipe->pipe_buffer.out != 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, rpipe->pipe_buffer.out, uio, &size);
540                         if (error)
541                                 break;
542                         nread += size;
543                         rpipe->pipe_buffer.out += size;
544                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
545                                 rpipe->pipe_state |= PIPE_DIRECTIP;
546                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
547                                 wakeup(rpipe);
548                         }
549 #endif
550                 } else {
551                         /*
552                          * detect EOF condition
553                          * read returns 0 on EOF, no need to set error
554                          */
555                         if (rpipe->pipe_state & PIPE_EOF)
556                                 break;
557
558                         /*
559                          * If the "write-side" has been blocked, wake it up now.
560                          */
561                         if (rpipe->pipe_state & PIPE_WANTW) {
562                                 rpipe->pipe_state &= ~PIPE_WANTW;
563                                 wakeup(rpipe);
564                         }
565
566                         /*
567                          * Break if some data was read.
568                          */
569                         if (nread > 0)
570                                 break;
571
572                         /*
573                          * Unlock the pipe buffer for our remaining
574                          * processing.  We will either break out with an
575                          * error or we will sleep and relock to loop.
576                          */
577                         pipeunlock(rpipe);
578
579                         /*
580                          * Handle non-blocking mode operation or
581                          * wait for more data.
582                          */
583                         if (fp->f_flag & FNONBLOCK) {
584                                 error = EAGAIN;
585                         } else {
586                                 rpipe->pipe_state |= PIPE_WANTR;
587                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
588                                     "piperd", 0)) == 0) {
589                                         error = pipelock(rpipe, 1);
590                                 }
591                         }
592                         if (error)
593                                 goto unlocked_error;
594                 }
595         }
596         pipeunlock(rpipe);
597
598         if (error == 0)
599                 vfs_timestamp(&rpipe->pipe_atime);
600 unlocked_error:
601         --rpipe->pipe_busy;
602
603         /*
604          * PIPE_WANT processing only makes sense if pipe_busy is 0.
605          */
606         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
607                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
608                 wakeup(rpipe);
609         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
610                 /*
611                  * Handle write blocking hysteresis.
612                  */
613                 if (rpipe->pipe_state & PIPE_WANTW) {
614                         rpipe->pipe_state &= ~PIPE_WANTW;
615                         wakeup(rpipe);
616                 }
617         }
618
619         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
620                 pipeselwakeup(rpipe);
621         return (error);
622 }
623
624 #ifndef PIPE_NODIRECT
625 /*
626  * Map the sending processes' buffer into kernel space and wire it.
627  * This is similar to a physical write operation.
628  */
629 static int
630 pipe_build_write_buffer(wpipe, uio)
631         struct pipe *wpipe;
632         struct uio *uio;
633 {
634         int error;
635         u_int size;
636
637         size = (u_int) uio->uio_iov->iov_len;
638         if (size > wpipe->pipe_buffer.size)
639                 size = wpipe->pipe_buffer.size;
640
641         error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base, 
642                                 size, XIOF_READ);
643         wpipe->pipe_buffer.out = 0;
644         if (error)
645                 return(error);
646
647         /*
648          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
649          * map the pages on the target while KMEM maps the pages now.
650          */
651         switch(wpipe->pipe_feature) {
652         case PIPE_KMEM:
653         case PIPE_SFBUF2:
654                 if (wpipe->pipe_kva == NULL) {
655                         wpipe->pipe_kva = 
656                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
657                         wpipe->pipe_kvamask = 0;
658                 }
659                 if (wpipe->pipe_feature == PIPE_KMEM) {
660                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
661                                     wpipe->pipe_map.xio_npages);
662                 }
663                 break;
664         default:
665                 break;
666         }
667
668         /*
669          * And update the uio data.  The XIO might have loaded fewer bytes
670          * then requested so reload 'size'.
671          */
672         size = wpipe->pipe_map.xio_bytes;
673         uio->uio_iov->iov_len -= size;
674         uio->uio_iov->iov_base += size;
675         if (uio->uio_iov->iov_len == 0)
676                 uio->uio_iov++;
677         uio->uio_resid -= size;
678         uio->uio_offset += size;
679         return (0);
680 }
681
682 /*
683  * In the case of a signal, the writing process might go away.  This
684  * code copies the data into the circular buffer so that the source
685  * pages can be freed without loss of data.
686  *
687  * Note that in direct mode pipe_buffer.out is used to track the
688  * XIO offset.  We are converting the direct mode into buffered mode
689  * which changes the meaning of pipe_buffer.out.
690  */
691 static void
692 pipe_clone_write_buffer(wpipe)
693         struct pipe *wpipe;
694 {
695         int size;
696         int offset;
697
698         offset = wpipe->pipe_buffer.out;
699         size = wpipe->pipe_map.xio_bytes - offset;
700
701         KKASSERT(size <= wpipe->pipe_buffer.size);
702
703         wpipe->pipe_buffer.in = size;
704         wpipe->pipe_buffer.out = 0;
705         wpipe->pipe_buffer.cnt = size;
706         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
707
708         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
709         xio_release(&wpipe->pipe_map);
710         if (wpipe->pipe_kva) {
711                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
712                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
713                 wpipe->pipe_kva = NULL;
714         }
715 }
716
717 /*
718  * This implements the pipe buffer write mechanism.  Note that only
719  * a direct write OR a normal pipe write can be pending at any given time.
720  * If there are any characters in the pipe buffer, the direct write will
721  * be deferred until the receiving process grabs all of the bytes from
722  * the pipe buffer.  Then the direct mapping write is set-up.
723  */
724 static int
725 pipe_direct_write(wpipe, uio)
726         struct pipe *wpipe;
727         struct uio *uio;
728 {
729         int error;
730
731 retry:
732         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
733                 if (wpipe->pipe_state & PIPE_WANTR) {
734                         wpipe->pipe_state &= ~PIPE_WANTR;
735                         wakeup(wpipe);
736                 }
737                 wpipe->pipe_state |= PIPE_WANTW;
738                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
739                 if (error)
740                         goto error2;
741                 if (wpipe->pipe_state & PIPE_EOF) {
742                         error = EPIPE;
743                         goto error2;
744                 }
745         }
746         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
747         if (wpipe->pipe_buffer.cnt > 0) {
748                 if (wpipe->pipe_state & PIPE_WANTR) {
749                         wpipe->pipe_state &= ~PIPE_WANTR;
750                         wakeup(wpipe);
751                 }
752                         
753                 wpipe->pipe_state |= PIPE_WANTW;
754                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
755                 if (error)
756                         goto error2;
757                 if (wpipe->pipe_state & PIPE_EOF) {
758                         error = EPIPE;
759                         goto error2;
760                 }
761                 goto retry;
762         }
763
764         /*
765          * Build our direct-write buffer
766          */
767         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
768         error = pipe_build_write_buffer(wpipe, uio);
769         if (error)
770                 goto error1;
771         wpipe->pipe_state &= ~PIPE_DIRECTIP;
772
773         /*
774          * Wait until the receiver has snarfed the data.  Since we are likely
775          * going to sleep we optimize the case and yield synchronously,
776          * possibly avoiding the tsleep().
777          */
778         error = 0;
779         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
780                 if (wpipe->pipe_state & PIPE_EOF) {
781                         pipelock(wpipe, 0);
782                         xio_release(&wpipe->pipe_map);
783                         if (wpipe->pipe_kva) {
784                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
785                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
786                                 wpipe->pipe_kva = NULL;
787                         }
788                         pipeunlock(wpipe);
789                         pipeselwakeup(wpipe);
790                         error = EPIPE;
791                         goto error1;
792                 }
793                 if (wpipe->pipe_state & PIPE_WANTR) {
794                         wpipe->pipe_state &= ~PIPE_WANTR;
795                         wakeup(wpipe);
796                 }
797                 pipeselwakeup(wpipe);
798                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
799         }
800         pipelock(wpipe,0);
801         if (wpipe->pipe_state & PIPE_DIRECTW) {
802                 /*
803                  * this bit of trickery substitutes a kernel buffer for
804                  * the process that might be going away.
805                  */
806                 pipe_clone_write_buffer(wpipe);
807                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
808         } else {
809                 /*
810                  * note: The pipe_kva mapping is not qremove'd here.  For
811                  * legacy PIPE_KMEM mode this constitutes an improvement
812                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
813                  * mode the kva mapping must not be removed to get the
814                  * caching benefit. 
815                  *
816                  * For testing purposes we will give the original algorithm
817                  * the benefit of the doubt 'what it could have been', and
818                  * keep the optimization.
819                  */
820                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
821                 xio_release(&wpipe->pipe_map);
822                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
823         }
824         pipeunlock(wpipe);
825         return (error);
826
827         /*
828          * Direct-write error, clear the direct write flags.
829          */
830 error1:
831         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
832         /* fallthrough */
833
834         /*
835          * General error, wakeup the other side if it happens to be sleeping.
836          */
837 error2:
838         wakeup(wpipe);
839         return (error);
840 }
841 #endif
842         
843 static int
844 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
845         int flags, struct thread *td)
846 {
847         int error = 0;
848         int orig_resid;
849         struct pipe *wpipe, *rpipe;
850
851         rpipe = (struct pipe *) fp->f_data;
852         wpipe = rpipe->pipe_peer;
853
854         /*
855          * detect loss of pipe read side, issue SIGPIPE if lost.
856          */
857         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
858                 return (EPIPE);
859         }
860         ++wpipe->pipe_busy;
861
862         /*
863          * If it is advantageous to resize the pipe buffer, do
864          * so.
865          */
866         if ((uio->uio_resid > PIPE_SIZE) &&
867                 (pipe_nbig < pipe_maxbig) &&
868                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
869                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
870                 (wpipe->pipe_buffer.cnt == 0)) {
871
872                 if ((error = pipelock(wpipe,1)) == 0) {
873                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
874                                 pipe_nbig++;
875                         pipeunlock(wpipe);
876                 }
877         }
878
879         /*
880          * If an early error occured unbusy and return, waking up any pending
881          * readers.
882          */
883         if (error) {
884                 --wpipe->pipe_busy;
885                 if ((wpipe->pipe_busy == 0) && 
886                     (wpipe->pipe_state & PIPE_WANT)) {
887                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
888                         wakeup(wpipe);
889                 }
890                 return(error);
891         }
892                 
893         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
894
895         orig_resid = uio->uio_resid;
896
897         while (uio->uio_resid) {
898                 int space;
899
900 #ifndef PIPE_NODIRECT
901                 /*
902                  * If the transfer is large, we can gain performance if
903                  * we do process-to-process copies directly.
904                  * If the write is non-blocking, we don't use the
905                  * direct write mechanism.
906                  *
907                  * The direct write mechanism will detect the reader going
908                  * away on us.
909                  */
910                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
911                     pipe_dwrite_enable > 1) &&
912                     (fp->f_flag & FNONBLOCK) == 0 &&
913                     pipe_dwrite_enable) {
914                         error = pipe_direct_write( wpipe, uio);
915                         if (error)
916                                 break;
917                         continue;
918                 }
919 #endif
920
921                 /*
922                  * Pipe buffered writes cannot be coincidental with
923                  * direct writes.  We wait until the currently executing
924                  * direct write is completed before we start filling the
925                  * pipe buffer.  We break out if a signal occurs or the
926                  * reader goes away.
927                  */
928         retrywrite:
929                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
930                         if (wpipe->pipe_state & PIPE_WANTR) {
931                                 wpipe->pipe_state &= ~PIPE_WANTR;
932                                 wakeup(wpipe);
933                         }
934                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
935                         if (wpipe->pipe_state & PIPE_EOF)
936                                 break;
937                         if (error)
938                                 break;
939                 }
940                 if (wpipe->pipe_state & PIPE_EOF) {
941                         error = EPIPE;
942                         break;
943                 }
944
945                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
946
947                 /* Writes of size <= PIPE_BUF must be atomic. */
948                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
949                         space = 0;
950
951                 /* 
952                  * Write to fill, read size handles write hysteresis.  Also
953                  * additional restrictions can cause select-based non-blocking
954                  * writes to spin.
955                  */
956                 if (space > 0) {
957                         if ((error = pipelock(wpipe,1)) == 0) {
958                                 int size;       /* Transfer size */
959                                 int segsize;    /* first segment to transfer */
960
961                                 /*
962                                  * It is possible for a direct write to
963                                  * slip in on us... handle it here...
964                                  */
965                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
966                                         pipeunlock(wpipe);
967                                         goto retrywrite;
968                                 }
969                                 /* 
970                                  * If a process blocked in uiomove, our
971                                  * value for space might be bad.
972                                  *
973                                  * XXX will we be ok if the reader has gone
974                                  * away here?
975                                  */
976                                 if (space > wpipe->pipe_buffer.size - 
977                                     wpipe->pipe_buffer.cnt) {
978                                         pipeunlock(wpipe);
979                                         goto retrywrite;
980                                 }
981
982                                 /*
983                                  * Transfer size is minimum of uio transfer
984                                  * and free space in pipe buffer.
985                                  */
986                                 if (space > uio->uio_resid)
987                                         size = uio->uio_resid;
988                                 else
989                                         size = space;
990                                 /*
991                                  * First segment to transfer is minimum of 
992                                  * transfer size and contiguous space in
993                                  * pipe buffer.  If first segment to transfer
994                                  * is less than the transfer size, we've got
995                                  * a wraparound in the buffer.
996                                  */
997                                 segsize = wpipe->pipe_buffer.size - 
998                                         wpipe->pipe_buffer.in;
999                                 if (segsize > size)
1000                                         segsize = size;
1001                                 
1002                                 /* Transfer first segment */
1003
1004                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in], 
1005                                                 segsize, uio);
1006                                 
1007                                 if (error == 0 && segsize < size) {
1008                                         /* 
1009                                          * Transfer remaining part now, to
1010                                          * support atomic writes.  Wraparound
1011                                          * happened.
1012                                          */
1013                                         if (wpipe->pipe_buffer.in + segsize != 
1014                                             wpipe->pipe_buffer.size)
1015                                                 panic("Expected pipe buffer wraparound disappeared");
1016                                                 
1017                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1018                                                         size - segsize, uio);
1019                                 }
1020                                 if (error == 0) {
1021                                         wpipe->pipe_buffer.in += size;
1022                                         if (wpipe->pipe_buffer.in >=
1023                                             wpipe->pipe_buffer.size) {
1024                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1025                                                         panic("Expected wraparound bad");
1026                                                 wpipe->pipe_buffer.in = size - segsize;
1027                                         }
1028                                 
1029                                         wpipe->pipe_buffer.cnt += size;
1030                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1031                                                 panic("Pipe buffer overflow");
1032                                 
1033                                 }
1034                                 pipeunlock(wpipe);
1035                         }
1036                         if (error)
1037                                 break;
1038
1039                 } else {
1040                         /*
1041                          * If the "read-side" has been blocked, wake it up now
1042                          * and yield to let it drain synchronously rather
1043                          * then block.
1044                          */
1045                         if (wpipe->pipe_state & PIPE_WANTR) {
1046                                 wpipe->pipe_state &= ~PIPE_WANTR;
1047                                 wakeup(wpipe);
1048                         }
1049
1050                         /*
1051                          * don't block on non-blocking I/O
1052                          */
1053                         if (fp->f_flag & FNONBLOCK) {
1054                                 error = EAGAIN;
1055                                 break;
1056                         }
1057
1058                         /*
1059                          * We have no more space and have something to offer,
1060                          * wake up select/poll.
1061                          */
1062                         pipeselwakeup(wpipe);
1063
1064                         wpipe->pipe_state |= PIPE_WANTW;
1065                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1066                         if (error != 0)
1067                                 break;
1068                         /*
1069                          * If read side wants to go away, we just issue a signal
1070                          * to ourselves.
1071                          */
1072                         if (wpipe->pipe_state & PIPE_EOF) {
1073                                 error = EPIPE;
1074                                 break;
1075                         }       
1076                 }
1077         }
1078
1079         --wpipe->pipe_busy;
1080
1081         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1082                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1083                 wakeup(wpipe);
1084         } else if (wpipe->pipe_buffer.cnt > 0) {
1085                 /*
1086                  * If we have put any characters in the buffer, we wake up
1087                  * the reader.
1088                  */
1089                 if (wpipe->pipe_state & PIPE_WANTR) {
1090                         wpipe->pipe_state &= ~PIPE_WANTR;
1091                         wakeup(wpipe);
1092                 }
1093         }
1094
1095         /*
1096          * Don't return EPIPE if I/O was successful
1097          */
1098         if ((wpipe->pipe_buffer.cnt == 0) &&
1099             (uio->uio_resid == 0) &&
1100             (error == EPIPE)) {
1101                 error = 0;
1102         }
1103
1104         if (error == 0)
1105                 vfs_timestamp(&wpipe->pipe_mtime);
1106
1107         /*
1108          * We have something to offer,
1109          * wake up select/poll.
1110          */
1111         if (wpipe->pipe_buffer.cnt)
1112                 pipeselwakeup(wpipe);
1113
1114         return (error);
1115 }
1116
1117 /*
1118  * we implement a very minimal set of ioctls for compatibility with sockets.
1119  */
1120 int
1121 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1122 {
1123         struct pipe *mpipe = (struct pipe *)fp->f_data;
1124
1125         switch (cmd) {
1126
1127         case FIONBIO:
1128                 return (0);
1129
1130         case FIOASYNC:
1131                 if (*(int *)data) {
1132                         mpipe->pipe_state |= PIPE_ASYNC;
1133                 } else {
1134                         mpipe->pipe_state &= ~PIPE_ASYNC;
1135                 }
1136                 return (0);
1137
1138         case FIONREAD:
1139                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1140                         *(int *)data = mpipe->pipe_map.xio_bytes -
1141                                         mpipe->pipe_buffer.out;
1142                 } else {
1143                         *(int *)data = mpipe->pipe_buffer.cnt;
1144                 }
1145                 return (0);
1146
1147         case FIOSETOWN:
1148                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1149
1150         case FIOGETOWN:
1151                 *(int *)data = fgetown(mpipe->pipe_sigio);
1152                 return (0);
1153
1154         /* This is deprecated, FIOSETOWN should be used instead. */
1155         case TIOCSPGRP:
1156                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1157
1158         /* This is deprecated, FIOGETOWN should be used instead. */
1159         case TIOCGPGRP:
1160                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1161                 return (0);
1162
1163         }
1164         return (ENOTTY);
1165 }
1166
1167 int
1168 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1169 {
1170         struct pipe *rpipe = (struct pipe *)fp->f_data;
1171         struct pipe *wpipe;
1172         int revents = 0;
1173
1174         wpipe = rpipe->pipe_peer;
1175         if (events & (POLLIN | POLLRDNORM))
1176                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1177                     (rpipe->pipe_buffer.cnt > 0) ||
1178                     (rpipe->pipe_state & PIPE_EOF))
1179                         revents |= events & (POLLIN | POLLRDNORM);
1180
1181         if (events & (POLLOUT | POLLWRNORM))
1182                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1183                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1184                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1185                         revents |= events & (POLLOUT | POLLWRNORM);
1186
1187         if ((rpipe->pipe_state & PIPE_EOF) ||
1188             (wpipe == NULL) ||
1189             (wpipe->pipe_state & PIPE_EOF))
1190                 revents |= POLLHUP;
1191
1192         if (revents == 0) {
1193                 if (events & (POLLIN | POLLRDNORM)) {
1194                         selrecord(td, &rpipe->pipe_sel);
1195                         rpipe->pipe_state |= PIPE_SEL;
1196                 }
1197
1198                 if (events & (POLLOUT | POLLWRNORM)) {
1199                         selrecord(td, &wpipe->pipe_sel);
1200                         wpipe->pipe_state |= PIPE_SEL;
1201                 }
1202         }
1203
1204         return (revents);
1205 }
1206
1207 static int
1208 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1209 {
1210         struct pipe *pipe = (struct pipe *)fp->f_data;
1211
1212         bzero((caddr_t)ub, sizeof(*ub));
1213         ub->st_mode = S_IFIFO;
1214         ub->st_blksize = pipe->pipe_buffer.size;
1215         ub->st_size = pipe->pipe_buffer.cnt;
1216         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1217                 ub->st_size = pipe->pipe_map.xio_bytes -
1218                                 pipe->pipe_buffer.out;
1219         }
1220         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1221         ub->st_atimespec = pipe->pipe_atime;
1222         ub->st_mtimespec = pipe->pipe_mtime;
1223         ub->st_ctimespec = pipe->pipe_ctime;
1224         /*
1225          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1226          * st_flags, st_gen.
1227          * XXX (st_dev, st_ino) should be unique.
1228          */
1229         return (0);
1230 }
1231
1232 /* ARGSUSED */
1233 static int
1234 pipe_close(struct file *fp, struct thread *td)
1235 {
1236         struct pipe *cpipe = (struct pipe *)fp->f_data;
1237
1238         fp->f_ops = &badfileops;
1239         fp->f_data = NULL;
1240         funsetown(cpipe->pipe_sigio);
1241         pipeclose(cpipe);
1242         return (0);
1243 }
1244
1245 static void
1246 pipe_free_kmem(struct pipe *cpipe)
1247 {
1248         if (cpipe->pipe_buffer.buffer != NULL) {
1249                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1250                         --pipe_nbig;
1251                 kmem_free(kernel_map,
1252                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1253                         cpipe->pipe_buffer.size);
1254                 cpipe->pipe_buffer.buffer = NULL;
1255                 cpipe->pipe_buffer.object = NULL;
1256         }
1257 #ifndef PIPE_NODIRECT
1258         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1259                 cpipe->pipe_map.xio_offset == 0 &&
1260                 cpipe->pipe_map.xio_npages == 0);
1261 #endif
1262 }
1263
1264 /*
1265  * shutdown the pipe
1266  */
1267 static void
1268 pipeclose(struct pipe *cpipe)
1269 {
1270         globaldata_t gd;
1271         struct pipe *ppipe;
1272
1273         if (cpipe == NULL)
1274                 return;
1275
1276         pipeselwakeup(cpipe);
1277
1278         /*
1279          * If the other side is blocked, wake it up saying that
1280          * we want to close it down.
1281          */
1282         while (cpipe->pipe_busy) {
1283                 wakeup(cpipe);
1284                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1285                 tsleep(cpipe, 0, "pipecl", 0);
1286         }
1287
1288         /*
1289          * Disconnect from peer
1290          */
1291         if ((ppipe = cpipe->pipe_peer) != NULL) {
1292                 pipeselwakeup(ppipe);
1293
1294                 ppipe->pipe_state |= PIPE_EOF;
1295                 wakeup(ppipe);
1296                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1297                 ppipe->pipe_peer = NULL;
1298         }
1299
1300         if (cpipe->pipe_kva) {
1301                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1302                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1303                 cpipe->pipe_kva = NULL;
1304         }
1305
1306         /*
1307          * free or cache resources
1308          */
1309         gd = mycpu;
1310         if (gd->gd_pipeqcount >= pipe_maxcache ||
1311             cpipe->pipe_buffer.size != PIPE_SIZE
1312         ) {
1313                 pipe_free_kmem(cpipe);
1314                 free(cpipe, M_PIPE);
1315         } else {
1316                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1317                         cpipe->pipe_map.xio_bytes == 0 &&
1318                         cpipe->pipe_map.xio_offset == 0);
1319                 cpipe->pipe_state = 0;
1320                 cpipe->pipe_busy = 0;
1321                 cpipe->pipe_peer = gd->gd_pipeq;
1322                 gd->gd_pipeq = cpipe;
1323                 ++gd->gd_pipeqcount;
1324         }
1325 }
1326
1327 /*ARGSUSED*/
1328 static int
1329 pipe_kqfilter(struct file *fp, struct knote *kn)
1330 {
1331         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1332
1333         switch (kn->kn_filter) {
1334         case EVFILT_READ:
1335                 kn->kn_fop = &pipe_rfiltops;
1336                 break;
1337         case EVFILT_WRITE:
1338                 kn->kn_fop = &pipe_wfiltops;
1339                 cpipe = cpipe->pipe_peer;
1340                 if (cpipe == NULL)
1341                         /* other end of pipe has been closed */
1342                         return (EPIPE);
1343                 break;
1344         default:
1345                 return (1);
1346         }
1347         kn->kn_hook = (caddr_t)cpipe;
1348
1349         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1350         return (0);
1351 }
1352
1353 static void
1354 filt_pipedetach(struct knote *kn)
1355 {
1356         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1357
1358         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1359 }
1360
1361 /*ARGSUSED*/
1362 static int
1363 filt_piperead(struct knote *kn, long hint)
1364 {
1365         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1366         struct pipe *wpipe = rpipe->pipe_peer;
1367
1368         kn->kn_data = rpipe->pipe_buffer.cnt;
1369         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1370                 kn->kn_data = rpipe->pipe_map.xio_bytes - 
1371                                 rpipe->pipe_buffer.out;
1372         }
1373
1374         if ((rpipe->pipe_state & PIPE_EOF) ||
1375             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1376                 kn->kn_flags |= EV_EOF; 
1377                 return (1);
1378         }
1379         return (kn->kn_data > 0);
1380 }
1381
1382 /*ARGSUSED*/
1383 static int
1384 filt_pipewrite(struct knote *kn, long hint)
1385 {
1386         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1387         struct pipe *wpipe = rpipe->pipe_peer;
1388
1389         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1390                 kn->kn_data = 0;
1391                 kn->kn_flags |= EV_EOF; 
1392                 return (1);
1393         }
1394         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1395         if (wpipe->pipe_state & PIPE_DIRECTW)
1396                 kn->kn_data = 0;
1397
1398         return (kn->kn_data >= PIPE_BUF);
1399 }