pipe - Make pipe r/w MPSAFE, add kern.pipe.mpsafe (disabled by default)
[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 1.50 2008/09/09 04:06:13 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 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/kernel.h>
32 #include <sys/proc.h>
33 #include <sys/fcntl.h>
34 #include <sys/file.h>
35 #include <sys/filedesc.h>
36 #include <sys/filio.h>
37 #include <sys/ttycom.h>
38 #include <sys/stat.h>
39 #include <sys/poll.h>
40 #include <sys/select.h>
41 #include <sys/signalvar.h>
42 #include <sys/sysproto.h>
43 #include <sys/pipe.h>
44 #include <sys/vnode.h>
45 #include <sys/uio.h>
46 #include <sys/event.h>
47 #include <sys/globaldata.h>
48 #include <sys/module.h>
49 #include <sys/malloc.h>
50 #include <sys/sysctl.h>
51 #include <sys/socket.h>
52
53 #include <vm/vm.h>
54 #include <vm/vm_param.h>
55 #include <sys/lock.h>
56 #include <vm/vm_object.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_extern.h>
59 #include <vm/pmap.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_page.h>
62 #include <vm/vm_zone.h>
63
64 #include <sys/file2.h>
65
66 #include <machine/cpufunc.h>
67
68 /*
69  * interfaces to the outside world
70  */
71 static int pipe_read (struct file *fp, struct uio *uio, 
72                 struct ucred *cred, int flags);
73 static int pipe_write (struct file *fp, struct uio *uio, 
74                 struct ucred *cred, int flags);
75 static int pipe_close (struct file *fp);
76 static int pipe_shutdown (struct file *fp, int how);
77 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
78 static int pipe_kqfilter (struct file *fp, struct knote *kn);
79 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
80 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred);
81
82 static struct fileops pipeops = {
83         .fo_read = pipe_read, 
84         .fo_write = pipe_write,
85         .fo_ioctl = pipe_ioctl,
86         .fo_poll = pipe_poll,
87         .fo_kqfilter = pipe_kqfilter,
88         .fo_stat = pipe_stat,
89         .fo_close = pipe_close,
90         .fo_shutdown = pipe_shutdown
91 };
92
93 static void     filt_pipedetach(struct knote *kn);
94 static int      filt_piperead(struct knote *kn, long hint);
95 static int      filt_pipewrite(struct knote *kn, long hint);
96
97 static struct filterops pipe_rfiltops =
98         { 1, NULL, filt_pipedetach, filt_piperead };
99 static struct filterops pipe_wfiltops =
100         { 1, NULL, filt_pipedetach, filt_pipewrite };
101
102 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
103
104 /*
105  * Default pipe buffer size(s), this can be kind-of large now because pipe
106  * space is pageable.  The pipe code will try to maintain locality of
107  * reference for performance reasons, so small amounts of outstanding I/O
108  * will not wipe the cache.
109  */
110 #define MINPIPESIZE (PIPE_SIZE/3)
111 #define MAXPIPESIZE (2*PIPE_SIZE/3)
112
113 /*
114  * Limit the number of "big" pipes
115  */
116 #define LIMITBIGPIPES   64
117 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
118
119 static int pipe_maxbig = LIMITBIGPIPES;
120 static int pipe_maxcache = PIPEQ_MAX_CACHE;
121 static int pipe_bigcount;
122 static int pipe_nbig;
123 static int pipe_bcache_alloc;
124 static int pipe_bkmem_alloc;
125
126 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
127 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
128         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
129 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
130         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
131 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
132         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
133 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
134         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
135 #ifdef SMP
136 static int pipe_mpsafe = 0;
137 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
138         CTLFLAG_RW, &pipe_mpsafe, 0, "");
139 #endif
140 #if !defined(NO_PIPE_SYSCTL_STATS)
141 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
142         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
143 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
144         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
145 #endif
146
147 static void pipeclose (struct pipe *cpipe);
148 static void pipe_free_kmem (struct pipe *cpipe);
149 static int pipe_create (struct pipe **cpipep);
150 static __inline void pipeselwakeup (struct pipe *cpipe);
151 static int pipespace (struct pipe *cpipe, int size);
152
153 static __inline void
154 pipeselwakeup(struct pipe *cpipe)
155 {
156         if (cpipe->pipe_state & PIPE_SEL) {
157                 get_mplock();
158                 cpipe->pipe_state &= ~PIPE_SEL;
159                 selwakeup(&cpipe->pipe_sel);
160                 rel_mplock();
161         }
162         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
163                 get_mplock();
164                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
165                 rel_mplock();
166         }
167         if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
168                 get_mplock();
169                 KNOTE(&cpipe->pipe_sel.si_note, 0);
170                 rel_mplock();
171         }
172 }
173
174 /*
175  * These routines are called before and after a UIO.  The UIO
176  * may block, causing our held tokens to be lost temporarily.
177  *
178  * We use these routines to serialize reads against other reads
179  * and writes against other writes.
180  *
181  * The read token is held on entry so *ipp does not race.
182  */
183 static __inline int
184 pipe_start_uio(struct pipe *cpipe, u_int *ipp)
185 {
186         int error;
187
188         while (*ipp) {
189                 *ipp = -1;
190                 error = tsleep(ipp, PCATCH, "pipexx", 0);
191                 if (error)
192                         return (error);
193         }
194         *ipp = 1;
195         return (0);
196 }
197
198 static __inline void
199 pipe_end_uio(struct pipe *cpipe, u_int *ipp)
200 {
201         if (*ipp < 0) {
202                 *ipp = 0;
203                 wakeup(ipp);
204         } else {
205                 *ipp = 0;
206         }
207 }
208
209 static __inline void
210 pipe_get_mplock(int *save)
211 {
212 #ifdef SMP
213         if (pipe_mpsafe == 0) {
214                 get_mplock();
215                 *save = 1;
216         } else
217 #endif
218         {
219                 *save = 0;
220         }
221 }
222
223 static __inline void
224 pipe_rel_mplock(int *save)
225 {
226 #ifdef SMP
227         if (*save)
228                 rel_mplock();
229 #endif
230 }
231
232
233 /*
234  * The pipe system call for the DTYPE_PIPE type of pipes
235  *
236  * pipe_ARgs(int dummy)
237  */
238
239 /* ARGSUSED */
240 int
241 sys_pipe(struct pipe_args *uap)
242 {
243         struct thread *td = curthread;
244         struct proc *p = td->td_proc;
245         struct file *rf, *wf;
246         struct pipe *rpipe, *wpipe;
247         int fd1, fd2, error;
248
249         KKASSERT(p);
250
251         rpipe = wpipe = NULL;
252         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
253                 pipeclose(rpipe); 
254                 pipeclose(wpipe); 
255                 return (ENFILE);
256         }
257         
258         error = falloc(p, &rf, &fd1);
259         if (error) {
260                 pipeclose(rpipe);
261                 pipeclose(wpipe);
262                 return (error);
263         }
264         uap->sysmsg_fds[0] = fd1;
265
266         /*
267          * Warning: once we've gotten past allocation of the fd for the
268          * read-side, we can only drop the read side via fdrop() in order
269          * to avoid races against processes which manage to dup() the read
270          * side while we are blocked trying to allocate the write side.
271          */
272         rf->f_type = DTYPE_PIPE;
273         rf->f_flag = FREAD | FWRITE;
274         rf->f_ops = &pipeops;
275         rf->f_data = rpipe;
276         error = falloc(p, &wf, &fd2);
277         if (error) {
278                 fsetfd(p, NULL, fd1);
279                 fdrop(rf);
280                 /* rpipe has been closed by fdrop(). */
281                 pipeclose(wpipe);
282                 return (error);
283         }
284         wf->f_type = DTYPE_PIPE;
285         wf->f_flag = FREAD | FWRITE;
286         wf->f_ops = &pipeops;
287         wf->f_data = wpipe;
288         uap->sysmsg_fds[1] = fd2;
289
290         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
291                                     M_PIPE, M_WAITOK|M_ZERO);
292         wpipe->pipe_slock = rpipe->pipe_slock;
293         rpipe->pipe_peer = wpipe;
294         wpipe->pipe_peer = rpipe;
295         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
296
297         /*
298          * Once activated the peer relationship remains valid until
299          * both sides are closed.
300          */
301         fsetfd(p, rf, fd1);
302         fsetfd(p, wf, fd2);
303         fdrop(rf);
304         fdrop(wf);
305
306         return (0);
307 }
308
309 /*
310  * Allocate kva for pipe circular buffer, the space is pageable
311  * This routine will 'realloc' the size of a pipe safely, if it fails
312  * it will retain the old buffer.
313  * If it fails it will return ENOMEM.
314  */
315 static int
316 pipespace(struct pipe *cpipe, int size)
317 {
318         struct vm_object *object;
319         caddr_t buffer;
320         int npages, error;
321
322         npages = round_page(size) / PAGE_SIZE;
323         object = cpipe->pipe_buffer.object;
324
325         /*
326          * [re]create the object if necessary and reserve space for it
327          * in the kernel_map.  The object and memory are pageable.  On
328          * success, free the old resources before assigning the new
329          * ones.
330          */
331         if (object == NULL || object->size != npages) {
332                 get_mplock();
333                 object = vm_object_allocate(OBJT_DEFAULT, npages);
334                 buffer = (caddr_t)vm_map_min(&kernel_map);
335
336                 error = vm_map_find(&kernel_map, object, 0,
337                                     (vm_offset_t *)&buffer, size,
338                                     1,
339                                     VM_MAPTYPE_NORMAL,
340                                     VM_PROT_ALL, VM_PROT_ALL,
341                                     0);
342
343                 if (error != KERN_SUCCESS) {
344                         vm_object_deallocate(object);
345                         rel_mplock();
346                         return (ENOMEM);
347                 }
348                 pipe_free_kmem(cpipe);
349                 rel_mplock();
350                 cpipe->pipe_buffer.object = object;
351                 cpipe->pipe_buffer.buffer = buffer;
352                 cpipe->pipe_buffer.size = size;
353                 ++pipe_bkmem_alloc;
354         } else {
355                 ++pipe_bcache_alloc;
356         }
357         cpipe->pipe_buffer.rindex = 0;
358         cpipe->pipe_buffer.windex = 0;
359         return (0);
360 }
361
362 /*
363  * Initialize and allocate VM and memory for pipe, pulling the pipe from
364  * our per-cpu cache if possible.  For now make sure it is sized for the
365  * smaller PIPE_SIZE default.
366  */
367 static int
368 pipe_create(struct pipe **cpipep)
369 {
370         globaldata_t gd = mycpu;
371         struct pipe *cpipe;
372         int error;
373
374         if ((cpipe = gd->gd_pipeq) != NULL) {
375                 gd->gd_pipeq = cpipe->pipe_peer;
376                 --gd->gd_pipeqcount;
377                 cpipe->pipe_peer = NULL;
378                 cpipe->pipe_wantwcnt = 0;
379         } else {
380                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
381         }
382         *cpipep = cpipe;
383         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
384                 return (error);
385         vfs_timestamp(&cpipe->pipe_ctime);
386         cpipe->pipe_atime = cpipe->pipe_ctime;
387         cpipe->pipe_mtime = cpipe->pipe_ctime;
388         lwkt_token_init(&cpipe->pipe_rlock);
389         lwkt_token_init(&cpipe->pipe_wlock);
390         return (0);
391 }
392
393 /*
394  * MPALMOSTSAFE (acquires mplock)
395  */
396 static int
397 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
398 {
399         struct pipe *rpipe;
400         int error;
401         int nread = 0;
402         int nbio;
403         u_int size;     /* total bytes available */
404         u_int nsize;    /* total bytes to read */
405         u_int rindex;   /* contiguous bytes available */
406         u_int half_way;
407         lwkt_tokref rlock;
408         lwkt_tokref wlock;
409         int mpsave;
410
411         /*
412          * Degenerate case
413          */
414         if (uio->uio_resid == 0)
415                 return(0);
416
417         /*
418          * Setup locks, calculate nbio
419          */
420         pipe_get_mplock(&mpsave);
421         rpipe = (struct pipe *)fp->f_data;
422         lwkt_gettoken(&rlock, &rpipe->pipe_rlock);
423
424         if (fflags & O_FBLOCKING)
425                 nbio = 0;
426         else if (fflags & O_FNONBLOCKING)
427                 nbio = 1;
428         else if (fp->f_flag & O_NONBLOCK)
429                 nbio = 1;
430         else
431                 nbio = 0;
432
433         /*
434          * Reads are serialized.  Note howeverthat pipe_buffer.buffer and
435          * pipe_buffer.size can change out from under us when the number
436          * of bytes in the buffer are zero due to the write-side doing a
437          * pipespace().
438          */
439         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
440         if (error) {
441                 pipe_rel_mplock(&mpsave);
442                 lwkt_reltoken(&rlock);
443                 return (error);
444         }
445         while (uio->uio_resid) {
446                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
447                 cpu_lfence();
448                 if (size) {
449                         rindex = rpipe->pipe_buffer.rindex &
450                                  (rpipe->pipe_buffer.size - 1);
451                         nsize = size;
452                         if (nsize > rpipe->pipe_buffer.size - rindex)
453                                 nsize = rpipe->pipe_buffer.size - rindex;
454                         if (nsize > (u_int)uio->uio_resid)
455                                 nsize = (u_int)uio->uio_resid;
456
457                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
458                                         nsize, uio);
459                         if (error)
460                                 break;
461                         cpu_mfence();
462                         rpipe->pipe_buffer.rindex += nsize;
463                         nread += nsize;
464
465                         /*
466                          * Shortcut to the loop-up if there is no writer
467                          * waiting or if we have not transitioned across
468                          * the half-way point.
469                          */
470                         half_way = rpipe->pipe_buffer.size >> 1;
471                         if ((rpipe->pipe_state & PIPE_WANTW) == 0 ||
472                             size <= half_way || size - nsize > half_way) {
473                         } {
474                                 continue;
475                         }
476                 }
477
478                 /*
479                  * If the "write-side" was blocked we wake it up.  This code
480                  * is reached either when the buffer is completely emptied
481                  * or if it becomes more then half-empty.
482                  *
483                  * Pipe_state can only be modified if both the rlock and
484                  * wlock are held.
485                  */
486                 if (rpipe->pipe_state & PIPE_WANTW) {
487                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
488                         if (rpipe->pipe_state & PIPE_WANTW) {
489                                 rpipe->pipe_state &= ~PIPE_WANTW;
490                                 lwkt_reltoken(&wlock);
491                                 wakeup(rpipe);
492                         } else {
493                                 lwkt_reltoken(&wlock);
494                         }
495                 }
496
497                 /*
498                  * Pick up our copy loop again if the writer sent data to
499                  * us.
500                  */
501                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
502                 if (size)
503                         continue;
504
505                 /*
506                  * Detect EOF condition, do not set error.
507                  */
508                 if (rpipe->pipe_state & PIPE_REOF)
509                         break;
510
511 #ifdef SMP
512                 /*
513                  * Gravy train if SMP box.  This saves a ton of IPIs and
514                  * allows two cpus to operate in lockstep.
515                  *
516                  * XXX check pipe_wip also?
517                  */
518                 DELAY(1);
519                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
520                 if (size)
521                         continue;
522 #endif
523
524                 /*
525                  * Break if some data was read, or if this was a non-blocking
526                  * read.
527                  */
528                 if (nread > 0)
529                         break;
530
531                 if (nbio) {
532                         error = EAGAIN;
533                         break;
534                 }
535
536                 /*
537                  * Last chance, interlock with WANTR.
538                  */
539                 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
540                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
541                 if (size) {
542                         lwkt_reltoken(&wlock);
543                         continue;
544                 }
545
546                 /*
547                  * If there is no more to read in the pipe, reset its
548                  * pointers to the beginning.  This improves cache hit
549                  * stats.
550                  *
551                  * We need both locks to modify both pointers, and there
552                  * must also not be a write in progress or the uiomove()
553                  * in the write might block and temporarily release
554                  * its wlock, then reacquire and update windex.  We are
555                  * only serialized against reads, not writes.
556                  *
557                  * XXX should we even bother resetting the indices?  It
558                  *     might actually be more cache efficient not to.
559                  */
560                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
561                     rpipe->pipe_wip == 0) {
562                         rpipe->pipe_buffer.rindex = 0;
563                         rpipe->pipe_buffer.windex = 0;
564                 }
565
566                 /*
567                  * Wait for more data.
568                  *
569                  * Pipe_state can only be set if both the rlock and wlock
570                  * are held.
571                  */
572                 rpipe->pipe_state |= PIPE_WANTR;
573                 tsleep_interlock(rpipe);
574                 lwkt_reltoken(&wlock);
575                 error = tsleep(rpipe, PCATCH, "piperd", 0);
576                 if (error)
577                         break;
578         }
579         pipe_end_uio(rpipe, &rpipe->pipe_rip);
580
581         /*
582          * Uptime last access time
583          */
584         if (error == 0 && nread)
585                 vfs_timestamp(&rpipe->pipe_atime);
586
587 #if 0
588         /*
589          * Handle write blocking hysteresis.  size can only increase while
590          * we hold the rlock.
591          *
592          * XXX shouldn't need this any more.  We will wakeup the writer
593          *     when we've drained past half-way.  The worst the writer
594          *     can do is fill the buffer up, not make it smaller, so
595          *     we are guaranteed our half-way test.
596          */
597         if (rpipe->pipe_state & PIPE_WANTW) {
598                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
599                 if (size <= (rpipe->pipe_buffer.size >> 1)) {
600                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
601                         if (rpipe->pipe_state & PIPE_WANTW) {
602                                 rpipe->pipe_state &= ~PIPE_WANTW;
603                                 lwkt_reltoken(&wlock);
604                                 wakeup(rpipe);
605                         } else {
606                                 lwkt_reltoken(&wlock);
607                         }
608                 }
609         }
610 #endif
611         size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
612         lwkt_reltoken(&rlock);
613
614         /*
615          * If enough space is available in buffer then wakeup sel writers?
616          */
617         if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF)
618                 pipeselwakeup(rpipe);
619         pipe_rel_mplock(&mpsave);
620         return (error);
621 }
622
623 /*
624  * MPALMOSTSAFE - acquires mplock
625  */
626 static int
627 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
628 {
629         int error;
630         int orig_resid;
631         int nbio;
632         struct pipe *wpipe, *rpipe;
633         lwkt_tokref rlock;
634         lwkt_tokref wlock;
635         u_int windex;
636         u_int space;
637         u_int wcount;
638         int mpsave;
639
640         pipe_get_mplock(&mpsave);
641
642         /*
643          * Writes go to the peer.  The peer will always exist.
644          */
645         rpipe = (struct pipe *) fp->f_data;
646         wpipe = rpipe->pipe_peer;
647         lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
648         if (wpipe->pipe_state & PIPE_WEOF) {
649                 pipe_rel_mplock(&mpsave);
650                 lwkt_reltoken(&wlock);
651                 return (EPIPE);
652         }
653
654         /*
655          * Degenerate case (EPIPE takes prec)
656          */
657         if (uio->uio_resid == 0) {
658                 pipe_rel_mplock(&mpsave);
659                 lwkt_reltoken(&wlock);
660                 return(0);
661         }
662
663         /*
664          * Writes are serialized (start_uio must be called with wlock)
665          */
666         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
667         if (error) {
668                 pipe_rel_mplock(&mpsave);
669                 lwkt_reltoken(&wlock);
670                 return (error);
671         }
672
673         if (fflags & O_FBLOCKING)
674                 nbio = 0;
675         else if (fflags & O_FNONBLOCKING)
676                 nbio = 1;
677         else if (fp->f_flag & O_NONBLOCK)
678                 nbio = 1;
679         else
680                 nbio = 0;
681
682         /*
683          * If it is advantageous to resize the pipe buffer, do
684          * so.  We are write-serialized so we can block safely.
685          */
686         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
687             (pipe_nbig < pipe_maxbig) &&
688             wpipe->pipe_wantwcnt > 4 &&
689             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
690                 /* 
691                  * Recheck after lock.
692                  */
693                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
694                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
695                     (pipe_nbig < pipe_maxbig) &&
696                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
697                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) {
698                                 ++pipe_bigcount;
699                                 pipe_nbig++;
700                         }
701                 }
702                 lwkt_reltoken(&rlock);
703         }
704
705         orig_resid = uio->uio_resid;
706         wcount = 0;
707
708         while (uio->uio_resid) {
709                 if (wpipe->pipe_state & PIPE_WEOF) {
710                         error = EPIPE;
711                         break;
712                 }
713
714                 windex = wpipe->pipe_buffer.windex &
715                          (wpipe->pipe_buffer.size - 1);
716                 space = wpipe->pipe_buffer.size -
717                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
718                 cpu_lfence();
719
720                 /* Writes of size <= PIPE_BUF must be atomic. */
721                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
722                         space = 0;
723
724                 /* 
725                  * Write to fill, read size handles write hysteresis.  Also
726                  * additional restrictions can cause select-based non-blocking
727                  * writes to spin.
728                  */
729                 if (space > 0) {
730                         u_int segsize;
731
732                         /*
733                          * Transfer size is minimum of uio transfer
734                          * and free space in pipe buffer.
735                          *
736                          * Limit each uiocopy to no more then PIPE_SIZE
737                          * so we can keep the gravy train going on a
738                          * SMP box.  This doubles the performance for
739                          * write sizes > 16K.  Otherwise large writes
740                          * wind up doing an inefficient synchronous
741                          * ping-pong.
742                          */
743                         if (space > (u_int)uio->uio_resid)
744                                 space = (u_int)uio->uio_resid;
745                         if (space > PIPE_SIZE)
746                                 space = PIPE_SIZE;
747
748                         /*
749                          * First segment to transfer is minimum of
750                          * transfer size and contiguous space in
751                          * pipe buffer.  If first segment to transfer
752                          * is less than the transfer size, we've got
753                          * a wraparound in the buffer.
754                          */
755                         segsize = wpipe->pipe_buffer.size - windex;
756                         if (segsize > space)
757                                 segsize = space;
758
759                         /*
760                          * If this is the first loop and the reader is
761                          * blocked, do a preemptive wakeup of the reader.
762                          *
763                          * This works for both SMP and UP.  On SMP the IPI
764                          * latency plus the wlock interlock on the reader
765                          * side is the fastest way to get the reader going.
766                          * (The scheduler will hard loop on lock tokens).
767                          *
768                          * NOTE: We can't clear WANTR here without acquiring
769                          * the rlock, which we don't want to do here!
770                          */
771                         if (wpipe->pipe_state & PIPE_WANTR)
772                                 wakeup(wpipe);
773
774                         /*
775                          * Transfer first segment
776                          */
777                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
778                                         segsize, uio);
779                         cpu_mfence();
780                         wpipe->pipe_buffer.windex += segsize;
781
782                         if (error == 0 && segsize < space) {
783                                 /*
784                                  * Transfer remaining part now, to
785                                  * support atomic writes.  Wraparound
786                                  * happened.
787                                  */
788                                 segsize = space - segsize;
789                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
790                                                 segsize, uio);
791                                 cpu_mfence();
792                                 wpipe->pipe_buffer.windex += segsize;
793                         }
794                         if (error)
795                                 break;
796                         wcount += space;
797                         continue;
798                 }
799
800                 /*
801                  * We need both the rlock and the wlock to interlock against
802                  * the EOF, WANTW, and size checks, and to modify pipe_state.
803                  *
804                  * These are token locks so we do not have to worry about
805                  * deadlocks.
806                  */
807                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
808
809                 /*
810                  * If the "read-side" has been blocked, wake it up now
811                  * and yield to let it drain synchronously rather
812                  * then block.
813                  */
814                 if (wpipe->pipe_state & PIPE_WANTR) {
815                         wpipe->pipe_state &= ~PIPE_WANTR;
816                         wakeup(wpipe);
817                 }
818
819                 /*
820                  * don't block on non-blocking I/O
821                  */
822                 if (nbio) {
823                         lwkt_reltoken(&rlock);
824                         error = EAGAIN;
825                         break;
826                 }
827
828                 /*
829                  * We have no more space and have something to offer,
830                  * wake up select/poll.
831                  */
832                 pipeselwakeup(wpipe);
833
834                 ++wpipe->pipe_wantwcnt; /* don't care about overflow */
835                 wpipe->pipe_state |= PIPE_WANTW;
836                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
837                 lwkt_reltoken(&rlock);
838
839                 /*
840                  * Break out if we errored or the read side wants us to go
841                  * away.
842                  */
843                 if (error)
844                         break;
845                 if (wpipe->pipe_state & PIPE_WEOF) {
846                         error = EPIPE;
847                         break;
848                 }
849         }
850         pipe_end_uio(wpipe, &wpipe->pipe_wip);
851
852         /*
853          * If we have put any characters in the buffer, we wake up
854          * the reader.
855          *
856          * Both rlock and wlock are required to be able to modify pipe_state.
857          */
858         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
859                 if (wpipe->pipe_state & PIPE_WANTR) {
860                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
861                         if (wpipe->pipe_state & PIPE_WANTR) {
862                                 wpipe->pipe_state &= ~PIPE_WANTR;
863                                 lwkt_reltoken(&rlock);
864                                 wakeup(wpipe);
865                         } else {
866                                 lwkt_reltoken(&rlock);
867                         }
868                 }
869         }
870
871         /*
872          * Don't return EPIPE if I/O was successful
873          */
874         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
875             (uio->uio_resid == 0) &&
876             (error == EPIPE)) {
877                 error = 0;
878         }
879
880         if (error == 0)
881                 vfs_timestamp(&wpipe->pipe_mtime);
882
883         /*
884          * We have something to offer,
885          * wake up select/poll.
886          */
887         space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;
888         lwkt_reltoken(&wlock);
889         if (space)
890                 pipeselwakeup(wpipe);
891         pipe_rel_mplock(&mpsave);
892         return (error);
893 }
894
895 /*
896  * MPALMOSTSAFE - acquires mplock
897  *
898  * we implement a very minimal set of ioctls for compatibility with sockets.
899  */
900 int
901 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
902 {
903         struct pipe *mpipe;
904         int error;
905
906         get_mplock();
907         mpipe = (struct pipe *)fp->f_data;
908
909         switch (cmd) {
910         case FIOASYNC:
911                 if (*(int *)data) {
912                         mpipe->pipe_state |= PIPE_ASYNC;
913                 } else {
914                         mpipe->pipe_state &= ~PIPE_ASYNC;
915                 }
916                 error = 0;
917                 break;
918         case FIONREAD:
919                 *(int *)data = mpipe->pipe_buffer.windex -
920                                 mpipe->pipe_buffer.rindex;
921                 error = 0;
922                 break;
923         case FIOSETOWN:
924                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
925                 break;
926         case FIOGETOWN:
927                 *(int *)data = fgetown(mpipe->pipe_sigio);
928                 error = 0;
929                 break;
930         case TIOCSPGRP:
931                 /* This is deprecated, FIOSETOWN should be used instead. */
932                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
933                 break;
934
935         case TIOCGPGRP:
936                 /* This is deprecated, FIOGETOWN should be used instead. */
937                 *(int *)data = -fgetown(mpipe->pipe_sigio);
938                 error = 0;
939                 break;
940         default:
941                 error = ENOTTY;
942                 break;
943         }
944         rel_mplock();
945         return (error);
946 }
947
948 /*
949  * MPALMOSTSAFE - acquires mplock
950  */
951 int
952 pipe_poll(struct file *fp, int events, struct ucred *cred)
953 {
954         struct pipe *rpipe;
955         struct pipe *wpipe;
956         int revents = 0;
957         u_int space;
958         int mpsave;
959
960         pipe_get_mplock(&mpsave);
961         rpipe = (struct pipe *)fp->f_data;
962         wpipe = rpipe->pipe_peer;
963         if (events & (POLLIN | POLLRDNORM)) {
964                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
965                     (rpipe->pipe_state & PIPE_REOF)) {
966                         revents |= events & (POLLIN | POLLRDNORM);
967                 }
968         }
969
970         if (events & (POLLOUT | POLLWRNORM)) {
971                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
972                         revents |= events & (POLLOUT | POLLWRNORM);
973                 } else {
974                         space = wpipe->pipe_buffer.windex -
975                                 wpipe->pipe_buffer.rindex;
976                         space = wpipe->pipe_buffer.size - space;
977                         if (space >= PIPE_BUF)
978                                 revents |= events & (POLLOUT | POLLWRNORM);
979                 }
980         }
981
982         if ((rpipe->pipe_state & PIPE_REOF) ||
983             (wpipe == NULL) ||
984             (wpipe->pipe_state & PIPE_WEOF))
985                 revents |= POLLHUP;
986
987         if (revents == 0) {
988                 if (events & (POLLIN | POLLRDNORM)) {
989                         selrecord(curthread, &rpipe->pipe_sel);
990                         rpipe->pipe_state |= PIPE_SEL;
991                 }
992
993                 if (events & (POLLOUT | POLLWRNORM)) {
994                         selrecord(curthread, &wpipe->pipe_sel);
995                         wpipe->pipe_state |= PIPE_SEL;
996                 }
997         }
998         pipe_rel_mplock(&mpsave);
999         return (revents);
1000 }
1001
1002 /*
1003  * MPALMOSTSAFE - acquires mplock
1004  */
1005 static int
1006 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1007 {
1008         struct pipe *pipe;
1009         int mpsave;
1010
1011         pipe_get_mplock(&mpsave);
1012         pipe = (struct pipe *)fp->f_data;
1013
1014         bzero((caddr_t)ub, sizeof(*ub));
1015         ub->st_mode = S_IFIFO;
1016         ub->st_blksize = pipe->pipe_buffer.size;
1017         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1018         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1019         ub->st_atimespec = pipe->pipe_atime;
1020         ub->st_mtimespec = pipe->pipe_mtime;
1021         ub->st_ctimespec = pipe->pipe_ctime;
1022         /*
1023          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1024          * st_flags, st_gen.
1025          * XXX (st_dev, st_ino) should be unique.
1026          */
1027         pipe_rel_mplock(&mpsave);
1028         return (0);
1029 }
1030
1031 /*
1032  * MPALMOSTSAFE - acquires mplock
1033  */
1034 static int
1035 pipe_close(struct file *fp)
1036 {
1037         struct pipe *cpipe;
1038
1039         get_mplock();
1040         cpipe = (struct pipe *)fp->f_data;
1041         fp->f_ops = &badfileops;
1042         fp->f_data = NULL;
1043         funsetown(cpipe->pipe_sigio);
1044         pipeclose(cpipe);
1045         rel_mplock();
1046         return (0);
1047 }
1048
1049 /*
1050  * Shutdown one or both directions of a full-duplex pipe.
1051  *
1052  * MPALMOSTSAFE - acquires mplock
1053  */
1054 static int
1055 pipe_shutdown(struct file *fp, int how)
1056 {
1057         struct pipe *rpipe;
1058         struct pipe *wpipe;
1059         int error = EPIPE;
1060         lwkt_tokref rpipe_rlock;
1061         lwkt_tokref rpipe_wlock;
1062         lwkt_tokref wpipe_rlock;
1063         lwkt_tokref wpipe_wlock;
1064         int mpsave;
1065
1066         pipe_get_mplock(&mpsave);
1067         rpipe = (struct pipe *)fp->f_data;
1068         wpipe = rpipe->pipe_peer;
1069
1070         /*
1071          * We modify pipe_state on both pipes, which means we need
1072          * all four tokens!
1073          */
1074         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1075         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1076         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1077         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1078
1079         switch(how) {
1080         case SHUT_RDWR:
1081         case SHUT_RD:
1082                 rpipe->pipe_state |= PIPE_REOF;
1083                 wpipe->pipe_state |= PIPE_WEOF;
1084                 if (rpipe->pipe_state & PIPE_WANTR) {
1085                         rpipe->pipe_state &= ~PIPE_WANTR;
1086                         wakeup(rpipe);
1087                 }
1088                 if (wpipe->pipe_state & PIPE_WANTW) {
1089                         wpipe->pipe_state &= ~PIPE_WANTW;
1090                         wakeup(wpipe);
1091                 }
1092                 pipeselwakeup(rpipe);
1093                 error = 0;
1094                 if (how == SHUT_RD)
1095                         break;
1096                 /* fall through */
1097         case SHUT_WR:
1098                 wpipe->pipe_state |= PIPE_WEOF;
1099                 rpipe->pipe_state |= PIPE_REOF;
1100                 if (wpipe->pipe_state & PIPE_WANTW) {
1101                         wpipe->pipe_state &= ~PIPE_WANTW;
1102                         wakeup(wpipe);
1103                 }
1104                 if (rpipe->pipe_state & PIPE_WANTR) {
1105                         rpipe->pipe_state &= ~PIPE_WANTR;
1106                         wakeup(rpipe);
1107                 }
1108                 pipeselwakeup(wpipe);
1109                 error = 0;
1110                 break;
1111         }
1112
1113         lwkt_reltoken(&rpipe_rlock);
1114         lwkt_reltoken(&rpipe_wlock);
1115         lwkt_reltoken(&wpipe_rlock);
1116         lwkt_reltoken(&wpipe_wlock);
1117
1118         pipe_rel_mplock(&mpsave);
1119         return (error);
1120 }
1121
1122 static void
1123 pipe_free_kmem(struct pipe *cpipe)
1124 {
1125         if (cpipe->pipe_buffer.buffer != NULL) {
1126                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1127                         --pipe_nbig;
1128                 kmem_free(&kernel_map,
1129                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1130                         cpipe->pipe_buffer.size);
1131                 cpipe->pipe_buffer.buffer = NULL;
1132                 cpipe->pipe_buffer.object = NULL;
1133         }
1134 }
1135
1136 /*
1137  * Close the pipe.  The slock must be held to interlock against simultanious
1138  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1139  */
1140 static void
1141 pipeclose(struct pipe *cpipe)
1142 {
1143         globaldata_t gd;
1144         struct pipe *ppipe;
1145         lwkt_tokref cpipe_rlock;
1146         lwkt_tokref cpipe_wlock;
1147         lwkt_tokref ppipe_rlock;
1148         lwkt_tokref ppipe_wlock;
1149
1150         if (cpipe == NULL)
1151                 return;
1152
1153         /*
1154          * The slock may not have been allocated yet (close during
1155          * initialization)
1156          *
1157          * We need both the read and write tokens to modify pipe_state.
1158          */
1159         if (cpipe->pipe_slock)
1160                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1161         lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1162         lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1163
1164         /*
1165          * Set our state, wakeup anyone waiting in select, and
1166          * wakeup anyone blocked on our pipe.
1167          */
1168         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1169         pipeselwakeup(cpipe);
1170         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1171                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1172                 wakeup(cpipe);
1173         }
1174
1175         /*
1176          * Disconnect from peer
1177          */
1178         if ((ppipe = cpipe->pipe_peer) != NULL) {
1179                 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1180                 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1181                 ppipe->pipe_state |= PIPE_REOF;
1182                 pipeselwakeup(ppipe);
1183                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1184                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1185                         wakeup(ppipe);
1186                 }
1187                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1188                         get_mplock();
1189                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1190                         rel_mplock();
1191                 }
1192                 lwkt_reltoken(&ppipe_rlock);
1193                 lwkt_reltoken(&ppipe_wlock);
1194         }
1195
1196         /*
1197          * If the peer is also closed we can free resources for both
1198          * sides, otherwise we leave our side intact to deal with any
1199          * races (since we only have the slock).
1200          */
1201         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1202                 cpipe->pipe_peer = NULL;
1203                 ppipe->pipe_peer = NULL;
1204                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1205                 pipeclose(ppipe);
1206                 ppipe = NULL;
1207         }
1208
1209         lwkt_reltoken(&cpipe_rlock);
1210         lwkt_reltoken(&cpipe_wlock);
1211         if (cpipe->pipe_slock)
1212                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1213
1214         /*
1215          * If we disassociated from our peer we can free resources
1216          */
1217         if (ppipe == NULL) {
1218                 gd = mycpu;
1219                 if (cpipe->pipe_slock) {
1220                         kfree(cpipe->pipe_slock, M_PIPE);
1221                         cpipe->pipe_slock = NULL;
1222                 }
1223                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1224                     cpipe->pipe_buffer.size != PIPE_SIZE
1225                 ) {
1226                         pipe_free_kmem(cpipe);
1227                         kfree(cpipe, M_PIPE);
1228                 } else {
1229                         cpipe->pipe_state = 0;
1230                         cpipe->pipe_peer = gd->gd_pipeq;
1231                         gd->gd_pipeq = cpipe;
1232                         ++gd->gd_pipeqcount;
1233                 }
1234         }
1235 }
1236
1237 /*
1238  * MPALMOSTSAFE - acquires mplock
1239  */
1240 static int
1241 pipe_kqfilter(struct file *fp, struct knote *kn)
1242 {
1243         struct pipe *cpipe;
1244
1245         get_mplock();
1246         cpipe = (struct pipe *)kn->kn_fp->f_data;
1247
1248         switch (kn->kn_filter) {
1249         case EVFILT_READ:
1250                 kn->kn_fop = &pipe_rfiltops;
1251                 break;
1252         case EVFILT_WRITE:
1253                 kn->kn_fop = &pipe_wfiltops;
1254                 cpipe = cpipe->pipe_peer;
1255                 if (cpipe == NULL) {
1256                         /* other end of pipe has been closed */
1257                         rel_mplock();
1258                         return (EPIPE);
1259                 }
1260                 break;
1261         default:
1262                 return (1);
1263         }
1264         kn->kn_hook = (caddr_t)cpipe;
1265
1266         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1267         rel_mplock();
1268         return (0);
1269 }
1270
1271 static void
1272 filt_pipedetach(struct knote *kn)
1273 {
1274         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1275
1276         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1277 }
1278
1279 /*ARGSUSED*/
1280 static int
1281 filt_piperead(struct knote *kn, long hint)
1282 {
1283         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1284
1285         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1286
1287         /* XXX RACE */
1288         if (rpipe->pipe_state & PIPE_REOF) {
1289                 kn->kn_flags |= EV_EOF; 
1290                 return (1);
1291         }
1292         return (kn->kn_data > 0);
1293 }
1294
1295 /*ARGSUSED*/
1296 static int
1297 filt_pipewrite(struct knote *kn, long hint)
1298 {
1299         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1300         struct pipe *wpipe = rpipe->pipe_peer;
1301         u_int32_t space;
1302
1303         /* XXX RACE */
1304         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1305                 kn->kn_data = 0;
1306                 kn->kn_flags |= EV_EOF; 
1307                 return (1);
1308         }
1309         space = wpipe->pipe_buffer.windex -
1310                 wpipe->pipe_buffer.rindex;
1311         space = wpipe->pipe_buffer.size - space;
1312         kn->kn_data = space;
1313         return (kn->kn_data >= PIPE_BUF);
1314 }