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         int notify_writer;
 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         notify_writer = 0;
 446         while (uio->uio_resid) {
 447                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 448                 cpu_lfence();
 449                 if (size) {
 450                         rindex = rpipe->pipe_buffer.rindex &
 451                                  (rpipe->pipe_buffer.size - 1);
 452                         nsize = size;
 453                         if (nsize > rpipe->pipe_buffer.size - rindex)
 454                                 nsize = rpipe->pipe_buffer.size - rindex;
 455                         if (nsize > (u_int)uio->uio_resid)
 456                                 nsize = (u_int)uio->uio_resid;
 457
 458                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 459                                         nsize, uio);
 460                         if (error)
 461                                 break;
 462                         cpu_mfence();
 463                         rpipe->pipe_buffer.rindex += nsize;
 464                         nread += nsize;
 465
 466                         /*
 467                          * If the FIFO has not been drained past the half-way
 468                          * mark then just continue and do not try to notify
 469                          * the writer yet.
 470                          */
 471                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 472                                 notify_writer = 0;
 473                                 continue;
 474                         }
 475
 476                         /*
 477                          * If the FIFO has been drained past the half-way
 478                          * mark we have to check the writer at some point,
 479                          * but for now continue if the writer is not yet
 480                          * blocked.
 481                          */
 482                         notify_writer = 1;
 483                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 484                                 continue;
 485                 }
 486
 487                 /*
 488                  * If the "write-side" was blocked we wake it up.  This code
 489                  * is reached either when the buffer is completely emptied
 490                  * or if it becomes more then half-empty.
 491                  *
 492                  * Pipe_state can only be modified if both the rlock and
 493                  * wlock are held.
 494                  */
 495                 if (rpipe->pipe_state & PIPE_WANTW) {
 496                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 497                         if (rpipe->pipe_state & PIPE_WANTW) {
 498                                 notify_writer = 0;
 499                                 rpipe->pipe_state &= ~PIPE_WANTW;
 500                                 lwkt_reltoken(&wlock);
 501                                 wakeup(rpipe);
 502                         } else {
 503                                 lwkt_reltoken(&wlock);
 504                         }
 505                 }
 506
 507                 /*
 508                  * Pick up our copy loop again if the writer sent data to
 509                  * us.
 510                  */
 511                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 512                 if (size)
 513                         continue;
 514
 515                 /*
 516                  * Detect EOF condition, do not set error.
 517                  */
 518                 if (rpipe->pipe_state & PIPE_REOF)
 519                         break;
 520
 521 #ifdef SMP
 522                 /*
 523                  * Gravy train if SMP box.  This saves a ton of IPIs and
 524                  * allows two cpus to operate in lockstep.
 525                  *
 526                  * XXX check pipe_wip also?
 527                  */
 528                 DELAY(1);
 529                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 530                 if (size)
 531                         continue;
 532 #endif
 533
 534                 /*
 535                  * Break if some data was read, or if this was a non-blocking
 536                  * read.
 537                  */
 538                 if (nread > 0)
 539                         break;
 540
 541                 if (nbio) {
 542                         error = EAGAIN;
 543                         break;
 544                 }
 545
 546                 /*
 547                  * Last chance, interlock with WANTR.
 548                  */
 549                 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 550                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 551                 if (size) {
 552                         lwkt_reltoken(&wlock);
 553                         continue;
 554                 }
 555
 556                 /*
 557                  * If there is no more to read in the pipe, reset its
 558                  * pointers to the beginning.  This improves cache hit
 559                  * stats.
 560                  *
 561                  * We need both locks to modify both pointers, and there
 562                  * must also not be a write in progress or the uiomove()
 563                  * in the write might block and temporarily release
 564                  * its wlock, then reacquire and update windex.  We are
 565                  * only serialized against reads, not writes.
 566                  *
 567                  * XXX should we even bother resetting the indices?  It
 568                  *     might actually be more cache efficient not to.
 569                  */
 570                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 571                     rpipe->pipe_wip == 0) {
 572                         rpipe->pipe_buffer.rindex = 0;
 573                         rpipe->pipe_buffer.windex = 0;
 574                 }
 575
 576                 /*
 577                  * Wait for more data.
 578                  *
 579                  * Pipe_state can only be set if both the rlock and wlock
 580                  * are held.
 581                  */
 582                 rpipe->pipe_state |= PIPE_WANTR;
 583                 tsleep_interlock(rpipe);
 584                 lwkt_reltoken(&wlock);
 585                 error = tsleep(rpipe, PCATCH, "piperd", 0);
 586                 if (error)
 587                         break;
 588         }
 589         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 590
 591         /*
 592          * Uptime last access time
 593          */
 594         if (error == 0 && nread)
 595                 vfs_timestamp(&rpipe->pipe_atime);
 596
 597         /*
 598          * If we drained the FIFO more then half way then handle
 599          * write blocking hysteresis.
 600          *
 601          * Note that PIPE_WANTW cannot be set by the writer without
 602          * it holding both rlock and wlock, so we can test it
 603          * while holding just rlock.
 604          */
 605         if (notify_writer) {
 606                 if (rpipe->pipe_state & PIPE_WANTW) {
 607                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 608                         if (rpipe->pipe_state & PIPE_WANTW) {
 609                                 rpipe->pipe_state &= ~PIPE_WANTW;
 610                                 lwkt_reltoken(&wlock);
 611                                 wakeup(rpipe);
 612                         } else {
 613                                 lwkt_reltoken(&wlock);
 614                         }
 615                 }
 616         }
 617         size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 618         lwkt_reltoken(&rlock);
 619
 620         /*
 621          * If enough space is available in buffer then wakeup sel writers?
 622          */
 623         if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF)
 624                 pipeselwakeup(rpipe);
 625         pipe_rel_mplock(&mpsave);
 626         return (error);
 627 }
 628
 629 /*
 630  * MPALMOSTSAFE - acquires mplock
 631  */
 632 static int
 633 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 634 {
 635         int error;
 636         int orig_resid;
 637         int nbio;
 638         struct pipe *wpipe, *rpipe;
 639         lwkt_tokref rlock;
 640         lwkt_tokref wlock;
 641         u_int windex;
 642         u_int space;
 643         u_int wcount;
 644         int mpsave;
 645
 646         pipe_get_mplock(&mpsave);
 647
 648         /*
 649          * Writes go to the peer.  The peer will always exist.
 650          */
 651         rpipe = (struct pipe *) fp->f_data;
 652         wpipe = rpipe->pipe_peer;
 653         lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
 654         if (wpipe->pipe_state & PIPE_WEOF) {
 655                 pipe_rel_mplock(&mpsave);
 656                 lwkt_reltoken(&wlock);
 657                 return (EPIPE);
 658         }
 659
 660         /*
 661          * Degenerate case (EPIPE takes prec)
 662          */
 663         if (uio->uio_resid == 0) {
 664                 pipe_rel_mplock(&mpsave);
 665                 lwkt_reltoken(&wlock);
 666                 return(0);
 667         }
 668
 669         /*
 670          * Writes are serialized (start_uio must be called with wlock)
 671          */
 672         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 673         if (error) {
 674                 pipe_rel_mplock(&mpsave);
 675                 lwkt_reltoken(&wlock);
 676                 return (error);
 677         }
 678
 679         if (fflags & O_FBLOCKING)
 680                 nbio = 0;
 681         else if (fflags & O_FNONBLOCKING)
 682                 nbio = 1;
 683         else if (fp->f_flag & O_NONBLOCK)
 684                 nbio = 1;
 685         else
 686                 nbio = 0;
 687
 688         /*
 689          * If it is advantageous to resize the pipe buffer, do
 690          * so.  We are write-serialized so we can block safely.
 691          */
 692         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 693             (pipe_nbig < pipe_maxbig) &&
 694             wpipe->pipe_wantwcnt > 4 &&
 695             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 696                 /*
 697                  * Recheck after lock.
 698                  */
 699                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 700                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 701                     (pipe_nbig < pipe_maxbig) &&
 702                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 703                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) {
 704                                 ++pipe_bigcount;
 705                                 pipe_nbig++;
 706                         }
 707                 }
 708                 lwkt_reltoken(&rlock);
 709         }
 710
 711         orig_resid = uio->uio_resid;
 712         wcount = 0;
 713
 714         while (uio->uio_resid) {
 715                 if (wpipe->pipe_state & PIPE_WEOF) {
 716                         error = EPIPE;
 717                         break;
 718                 }
 719
 720                 windex = wpipe->pipe_buffer.windex &
 721                          (wpipe->pipe_buffer.size - 1);
 722                 space = wpipe->pipe_buffer.size -
 723                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 724                 cpu_lfence();
 725
 726                 /* Writes of size <= PIPE_BUF must be atomic. */
 727                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 728                         space = 0;
 729
 730                 /*
 731                  * Write to fill, read size handles write hysteresis.  Also
 732                  * additional restrictions can cause select-based non-blocking
 733                  * writes to spin.
 734                  */
 735                 if (space > 0) {
 736                         u_int segsize;
 737
 738                         /*
 739                          * Transfer size is minimum of uio transfer
 740                          * and free space in pipe buffer.
 741                          *
 742                          * Limit each uiocopy to no more then PIPE_SIZE
 743                          * so we can keep the gravy train going on a
 744                          * SMP box.  This doubles the performance for
 745                          * write sizes > 16K.  Otherwise large writes
 746                          * wind up doing an inefficient synchronous
 747                          * ping-pong.
 748                          */
 749                         if (space > (u_int)uio->uio_resid)
 750                                 space = (u_int)uio->uio_resid;
 751                         if (space > PIPE_SIZE)
 752                                 space = PIPE_SIZE;
 753
 754                         /*
 755                          * First segment to transfer is minimum of
 756                          * transfer size and contiguous space in
 757                          * pipe buffer.  If first segment to transfer
 758                          * is less than the transfer size, we've got
 759                          * a wraparound in the buffer.
 760                          */
 761                         segsize = wpipe->pipe_buffer.size - windex;
 762                         if (segsize > space)
 763                                 segsize = space;
 764
 765                         /*
 766                          * If this is the first loop and the reader is
 767                          * blocked, do a preemptive wakeup of the reader.
 768                          *
 769                          * This works for both SMP and UP.  On SMP the IPI
 770                          * latency plus the wlock interlock on the reader
 771                          * side is the fastest way to get the reader going.
 772                          * (The scheduler will hard loop on lock tokens).
 773                          *
 774                          * NOTE: We can't clear WANTR here without acquiring
 775                          * the rlock, which we don't want to do here!
 776                          */
 777                         if (wpipe->pipe_state & PIPE_WANTR)
 778                                 wakeup(wpipe);
 779
 780                         /*
 781                          * Transfer first segment
 782                          */
 783                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 784                                         segsize, uio);
 785                         cpu_mfence();
 786                         wpipe->pipe_buffer.windex += segsize;
 787
 788                         if (error == 0 && segsize < space) {
 789                                 /*
 790                                  * Transfer remaining part now, to
 791                                  * support atomic writes.  Wraparound
 792                                  * happened.
 793                                  */
 794                                 segsize = space - segsize;
 795                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 796                                                 segsize, uio);
 797                                 cpu_mfence();
 798                                 wpipe->pipe_buffer.windex += segsize;
 799                         }
 800                         if (error)
 801                                 break;
 802                         wcount += space;
 803                         continue;
 804                 }
 805
 806                 /*
 807                  * We need both the rlock and the wlock to interlock against
 808                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 809                  *
 810                  * These are token locks so we do not have to worry about
 811                  * deadlocks.
 812                  */
 813                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 814
 815                 /*
 816                  * If the "read-side" has been blocked, wake it up now
 817                  * and yield to let it drain synchronously rather
 818                  * then block.
 819                  */
 820                 if (wpipe->pipe_state & PIPE_WANTR) {
 821                         wpipe->pipe_state &= ~PIPE_WANTR;
 822                         wakeup(wpipe);
 823                 }
 824
 825                 /*
 826                  * don't block on non-blocking I/O
 827                  */
 828                 if (nbio) {
 829                         lwkt_reltoken(&rlock);
 830                         error = EAGAIN;
 831                         break;
 832                 }
 833
 834                 /*
 835                  * re-test whether we have to block in the writer after
 836                  * acquiring both locks, in case the reader opened up
 837                  * some space.
 838                  */
 839                 space = wpipe->pipe_buffer.size -
 840                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 841                 cpu_lfence();
 842                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 843                         space = 0;
 844
 845                 /*
 846                  * We have no more space and have something to offer,
 847                  * wake up select/poll.
 848                  */
 849                 if (space == 0) {
 850                         pipeselwakeup(wpipe);
 851                         ++wpipe->pipe_wantwcnt;
 852                         wpipe->pipe_state |= PIPE_WANTW;
 853                         error = tsleep(wpipe, PCATCH, "pipewr", 0);
 854                 }
 855                 lwkt_reltoken(&rlock);
 856
 857                 /*
 858                  * Break out if we errored or the read side wants us to go
 859                  * away.
 860                  */
 861                 if (error)
 862                         break;
 863                 if (wpipe->pipe_state & PIPE_WEOF) {
 864                         error = EPIPE;
 865                         break;
 866                 }
 867         }
 868         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 869
 870         /*
 871          * If we have put any characters in the buffer, we wake up
 872          * the reader.
 873          *
 874          * Both rlock and wlock are required to be able to modify pipe_state.
 875          */
 876         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 877                 if (wpipe->pipe_state & PIPE_WANTR) {
 878                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 879                         if (wpipe->pipe_state & PIPE_WANTR) {
 880                                 wpipe->pipe_state &= ~PIPE_WANTR;
 881                                 lwkt_reltoken(&rlock);
 882                                 wakeup(wpipe);
 883                         } else {
 884                                 lwkt_reltoken(&rlock);
 885                         }
 886                 }
 887         }
 888
 889         /*
 890          * Don't return EPIPE if I/O was successful
 891          */
 892         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 893             (uio->uio_resid == 0) &&
 894             (error == EPIPE)) {
 895                 error = 0;
 896         }
 897
 898         if (error == 0)
 899                 vfs_timestamp(&wpipe->pipe_mtime);
 900
 901         /*
 902          * We have something to offer,
 903          * wake up select/poll.
 904          */
 905         space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;
 906         lwkt_reltoken(&wlock);
 907         if (space)
 908                 pipeselwakeup(wpipe);
 909         pipe_rel_mplock(&mpsave);
 910         return (error);
 911 }
 912
 913 /*
 914  * MPALMOSTSAFE - acquires mplock
 915  *
 916  * we implement a very minimal set of ioctls for compatibility with sockets.
 917  */
 918 int
 919 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
 920 {
 921         struct pipe *mpipe;
 922         int error;
 923
 924         get_mplock();
 925         mpipe = (struct pipe *)fp->f_data;
 926
 927         switch (cmd) {
 928         case FIOASYNC:
 929                 if (*(int *)data) {
 930                         mpipe->pipe_state |= PIPE_ASYNC;
 931                 } else {
 932                         mpipe->pipe_state &= ~PIPE_ASYNC;
 933                 }
 934                 error = 0;
 935                 break;
 936         case FIONREAD:
 937                 *(int *)data = mpipe->pipe_buffer.windex -
 938                                 mpipe->pipe_buffer.rindex;
 939                 error = 0;
 940                 break;
 941         case FIOSETOWN:
 942                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
 943                 break;
 944         case FIOGETOWN:
 945                 *(int *)data = fgetown(mpipe->pipe_sigio);
 946                 error = 0;
 947                 break;
 948         case TIOCSPGRP:
 949                 /* This is deprecated, FIOSETOWN should be used instead. */
 950                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
 951                 break;
 952
 953         case TIOCGPGRP:
 954                 /* This is deprecated, FIOGETOWN should be used instead. */
 955                 *(int *)data = -fgetown(mpipe->pipe_sigio);
 956                 error = 0;
 957                 break;
 958         default:
 959                 error = ENOTTY;
 960                 break;
 961         }
 962         rel_mplock();
 963         return (error);
 964 }
 965
 966 /*
 967  * MPALMOSTSAFE - acquires mplock
 968  */
 969 int
 970 pipe_poll(struct file *fp, int events, struct ucred *cred)
 971 {
 972         struct pipe *rpipe;
 973         struct pipe *wpipe;
 974         int revents = 0;
 975         u_int space;
 976         int mpsave;
 977
 978         pipe_get_mplock(&mpsave);
 979         rpipe = (struct pipe *)fp->f_data;
 980         wpipe = rpipe->pipe_peer;
 981         if (events & (POLLIN | POLLRDNORM)) {
 982                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
 983                     (rpipe->pipe_state & PIPE_REOF)) {
 984                         revents |= events & (POLLIN | POLLRDNORM);
 985                 }
 986         }
 987
 988         if (events & (POLLOUT | POLLWRNORM)) {
 989                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
 990                         revents |= events & (POLLOUT | POLLWRNORM);
 991                 } else {
 992                         space = wpipe->pipe_buffer.windex -
 993                                 wpipe->pipe_buffer.rindex;
 994                         space = wpipe->pipe_buffer.size - space;
 995                         if (space >= PIPE_BUF)
 996                                 revents |= events & (POLLOUT | POLLWRNORM);
 997                 }
 998         }
 999
1000         if ((rpipe->pipe_state & PIPE_REOF) ||
1001             (wpipe == NULL) ||
1002             (wpipe->pipe_state & PIPE_WEOF))
1003                 revents |= POLLHUP;
1004
1005         if (revents == 0) {
1006                 if (events & (POLLIN | POLLRDNORM)) {
1007                         selrecord(curthread, &rpipe->pipe_sel);
1008                         rpipe->pipe_state |= PIPE_SEL;
1009                 }
1010
1011                 if (events & (POLLOUT | POLLWRNORM)) {
1012                         selrecord(curthread, &wpipe->pipe_sel);
1013                         wpipe->pipe_state |= PIPE_SEL;
1014                 }
1015         }
1016         pipe_rel_mplock(&mpsave);
1017         return (revents);
1018 }
1019
1020 /*
1021  * MPALMOSTSAFE - acquires mplock
1022  */
1023 static int
1024 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1025 {
1026         struct pipe *pipe;
1027         int mpsave;
1028
1029         pipe_get_mplock(&mpsave);
1030         pipe = (struct pipe *)fp->f_data;
1031
1032         bzero((caddr_t)ub, sizeof(*ub));
1033         ub->st_mode = S_IFIFO;
1034         ub->st_blksize = pipe->pipe_buffer.size;
1035         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1036         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1037         ub->st_atimespec = pipe->pipe_atime;
1038         ub->st_mtimespec = pipe->pipe_mtime;
1039         ub->st_ctimespec = pipe->pipe_ctime;
1040         /*
1041          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1042          * st_flags, st_gen.
1043          * XXX (st_dev, st_ino) should be unique.
1044          */
1045         pipe_rel_mplock(&mpsave);
1046         return (0);
1047 }
1048
1049 /*
1050  * MPALMOSTSAFE - acquires mplock
1051  */
1052 static int
1053 pipe_close(struct file *fp)
1054 {
1055         struct pipe *cpipe;
1056
1057         get_mplock();
1058         cpipe = (struct pipe *)fp->f_data;
1059         fp->f_ops = &badfileops;
1060         fp->f_data = NULL;
1061         funsetown(cpipe->pipe_sigio);
1062         pipeclose(cpipe);
1063         rel_mplock();
1064         return (0);
1065 }
1066
1067 /*
1068  * Shutdown one or both directions of a full-duplex pipe.
1069  *
1070  * MPALMOSTSAFE - acquires mplock
1071  */
1072 static int
1073 pipe_shutdown(struct file *fp, int how)
1074 {
1075         struct pipe *rpipe;
1076         struct pipe *wpipe;
1077         int error = EPIPE;
1078         lwkt_tokref rpipe_rlock;
1079         lwkt_tokref rpipe_wlock;
1080         lwkt_tokref wpipe_rlock;
1081         lwkt_tokref wpipe_wlock;
1082         int mpsave;
1083
1084         pipe_get_mplock(&mpsave);
1085         rpipe = (struct pipe *)fp->f_data;
1086         wpipe = rpipe->pipe_peer;
1087
1088         /*
1089          * We modify pipe_state on both pipes, which means we need
1090          * all four tokens!
1091          */
1092         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1093         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1094         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1095         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1096
1097         switch(how) {
1098         case SHUT_RDWR:
1099         case SHUT_RD:
1100                 rpipe->pipe_state |= PIPE_REOF;
1101                 wpipe->pipe_state |= PIPE_WEOF;
1102                 if (rpipe->pipe_state & PIPE_WANTR) {
1103                         rpipe->pipe_state &= ~PIPE_WANTR;
1104                         wakeup(rpipe);
1105                 }
1106                 if (wpipe->pipe_state & PIPE_WANTW) {
1107                         wpipe->pipe_state &= ~PIPE_WANTW;
1108                         wakeup(wpipe);
1109                 }
1110                 pipeselwakeup(rpipe);
1111                 error = 0;
1112                 if (how == SHUT_RD)
1113                         break;
1114                 /* fall through */
1115         case SHUT_WR:
1116                 wpipe->pipe_state |= PIPE_WEOF;
1117                 rpipe->pipe_state |= PIPE_REOF;
1118                 if (wpipe->pipe_state & PIPE_WANTW) {
1119                         wpipe->pipe_state &= ~PIPE_WANTW;
1120                         wakeup(wpipe);
1121                 }
1122                 if (rpipe->pipe_state & PIPE_WANTR) {
1123                         rpipe->pipe_state &= ~PIPE_WANTR;
1124                         wakeup(rpipe);
1125                 }
1126                 pipeselwakeup(wpipe);
1127                 error = 0;
1128                 break;
1129         }
1130
1131         lwkt_reltoken(&rpipe_rlock);
1132         lwkt_reltoken(&rpipe_wlock);
1133         lwkt_reltoken(&wpipe_rlock);
1134         lwkt_reltoken(&wpipe_wlock);
1135
1136         pipe_rel_mplock(&mpsave);
1137         return (error);
1138 }
1139
1140 static void
1141 pipe_free_kmem(struct pipe *cpipe)
1142 {
1143         if (cpipe->pipe_buffer.buffer != NULL) {
1144                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1145                         --pipe_nbig;
1146                 kmem_free(&kernel_map,
1147                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1148                         cpipe->pipe_buffer.size);
1149                 cpipe->pipe_buffer.buffer = NULL;
1150                 cpipe->pipe_buffer.object = NULL;
1151         }
1152 }
1153
1154 /*
1155  * Close the pipe.  The slock must be held to interlock against simultanious
1156  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1157  */
1158 static void
1159 pipeclose(struct pipe *cpipe)
1160 {
1161         globaldata_t gd;
1162         struct pipe *ppipe;
1163         lwkt_tokref cpipe_rlock;
1164         lwkt_tokref cpipe_wlock;
1165         lwkt_tokref ppipe_rlock;
1166         lwkt_tokref ppipe_wlock;
1167
1168         if (cpipe == NULL)
1169                 return;
1170
1171         /*
1172          * The slock may not have been allocated yet (close during
1173          * initialization)
1174          *
1175          * We need both the read and write tokens to modify pipe_state.
1176          */
1177         if (cpipe->pipe_slock)
1178                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1179         lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1180         lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1181
1182         /*
1183          * Set our state, wakeup anyone waiting in select, and
1184          * wakeup anyone blocked on our pipe.
1185          */
1186         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1187         pipeselwakeup(cpipe);
1188         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1189                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1190                 wakeup(cpipe);
1191         }
1192
1193         /*
1194          * Disconnect from peer
1195          */
1196         if ((ppipe = cpipe->pipe_peer) != NULL) {
1197                 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1198                 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1199                 ppipe->pipe_state |= PIPE_REOF;
1200                 pipeselwakeup(ppipe);
1201                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1202                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1203                         wakeup(ppipe);
1204                 }
1205                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1206                         get_mplock();
1207                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1208                         rel_mplock();
1209                 }
1210                 lwkt_reltoken(&ppipe_rlock);
1211                 lwkt_reltoken(&ppipe_wlock);
1212         }
1213
1214         /*
1215          * If the peer is also closed we can free resources for both
1216          * sides, otherwise we leave our side intact to deal with any
1217          * races (since we only have the slock).
1218          */
1219         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1220                 cpipe->pipe_peer = NULL;
1221                 ppipe->pipe_peer = NULL;
1222                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1223                 pipeclose(ppipe);
1224                 ppipe = NULL;
1225         }
1226
1227         lwkt_reltoken(&cpipe_rlock);
1228         lwkt_reltoken(&cpipe_wlock);
1229         if (cpipe->pipe_slock)
1230                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1231
1232         /*
1233          * If we disassociated from our peer we can free resources
1234          */
1235         if (ppipe == NULL) {
1236                 gd = mycpu;
1237                 if (cpipe->pipe_slock) {
1238                         kfree(cpipe->pipe_slock, M_PIPE);
1239                         cpipe->pipe_slock = NULL;
1240                 }
1241                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1242                     cpipe->pipe_buffer.size != PIPE_SIZE
1243                 ) {
1244                         pipe_free_kmem(cpipe);
1245                         kfree(cpipe, M_PIPE);
1246                 } else {
1247                         cpipe->pipe_state = 0;
1248                         cpipe->pipe_peer = gd->gd_pipeq;
1249                         gd->gd_pipeq = cpipe;
1250                         ++gd->gd_pipeqcount;
1251                 }
1252         }
1253 }
1254
1255 /*
1256  * MPALMOSTSAFE - acquires mplock
1257  */
1258 static int
1259 pipe_kqfilter(struct file *fp, struct knote *kn)
1260 {
1261         struct pipe *cpipe;
1262
1263         get_mplock();
1264         cpipe = (struct pipe *)kn->kn_fp->f_data;
1265
1266         switch (kn->kn_filter) {
1267         case EVFILT_READ:
1268                 kn->kn_fop = &pipe_rfiltops;
1269                 break;
1270         case EVFILT_WRITE:
1271                 kn->kn_fop = &pipe_wfiltops;
1272                 cpipe = cpipe->pipe_peer;
1273                 if (cpipe == NULL) {
1274                         /* other end of pipe has been closed */
1275                         rel_mplock();
1276                         return (EPIPE);
1277                 }
1278                 break;
1279         default:
1280                 return (1);
1281         }
1282         kn->kn_hook = (caddr_t)cpipe;
1283
1284         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1285         rel_mplock();
1286         return (0);
1287 }
1288
1289 static void
1290 filt_pipedetach(struct knote *kn)
1291 {
1292         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1293
1294         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1295 }
1296
1297 /*ARGSUSED*/
1298 static int
1299 filt_piperead(struct knote *kn, long hint)
1300 {
1301         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1302
1303         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1304
1305         /* XXX RACE */
1306         if (rpipe->pipe_state & PIPE_REOF) {
1307                 kn->kn_flags |= EV_EOF;
1308                 return (1);
1309         }
1310         return (kn->kn_data > 0);
1311 }
1312
1313 /*ARGSUSED*/
1314 static int
1315 filt_pipewrite(struct knote *kn, long hint)
1316 {
1317         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1318         struct pipe *wpipe = rpipe->pipe_peer;
1319         u_int32_t space;
1320
1321         /* XXX RACE */
1322         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1323                 kn->kn_data = 0;
1324                 kn->kn_flags |= EV_EOF;
1325                 return (1);
1326         }
1327         space = wpipe->pipe_buffer.windex -
1328                 wpipe->pipe_buffer.rindex;
1329         space = wpipe->pipe_buffer.size - space;
1330         kn->kn_data = space;
1331         return (kn->kn_data >= PIPE_BUF);
1332 }