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/select.h>
  40 #include <sys/signalvar.h>
  41 #include <sys/sysproto.h>
  42 #include <sys/pipe.h>
  43 #include <sys/vnode.h>
  44 #include <sys/uio.h>
  45 #include <sys/event.h>
  46 #include <sys/globaldata.h>
  47 #include <sys/module.h>
  48 #include <sys/malloc.h>
  49 #include <sys/sysctl.h>
  50 #include <sys/socket.h>
  51
  52 #include <vm/vm.h>
  53 #include <vm/vm_param.h>
  54 #include <sys/lock.h>
  55 #include <vm/vm_object.h>
  56 #include <vm/vm_kern.h>
  57 #include <vm/vm_extern.h>
  58 #include <vm/pmap.h>
  59 #include <vm/vm_map.h>
  60 #include <vm/vm_page.h>
  61 #include <vm/vm_zone.h>
  62
  63 #include <sys/file2.h>
  64 #include <sys/signal2.h>
  65 #include <sys/mplock2.h>
  66
  67 #include <machine/cpufunc.h>
  68
  69 /*
  70  * interfaces to the outside world
  71  */
  72 static int pipe_read (struct file *fp, struct uio *uio,
  73                 struct ucred *cred, int flags);
  74 static int pipe_write (struct file *fp, struct uio *uio,
  75                 struct ucred *cred, int flags);
  76 static int pipe_close (struct file *fp);
  77 static int pipe_shutdown (struct file *fp, int how);
  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,
  81                 struct ucred *cred, struct sysmsg *msg);
  82
  83 static struct fileops pipeops = {
  84         .fo_read = pipe_read,
  85         .fo_write = pipe_write,
  86         .fo_ioctl = pipe_ioctl,
  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 static int pipe_rblocked_count;
 126 static int pipe_wblocked_count;
 127
 128 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 129 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 130         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 131 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 132         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 133 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked,
 134         CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
 135 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
 136         CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
 137 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 138         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 139 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 140         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 141 #ifdef SMP
 142 static int pipe_delay = 5000;   /* 5uS default */
 143 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
 144         CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
 145 static int pipe_mpsafe = 1;
 146 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
 147         CTLFLAG_RW, &pipe_mpsafe, 0, "");
 148 #endif
 149 #if !defined(NO_PIPE_SYSCTL_STATS)
 150 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 151         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 152 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 153         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 154 #endif
 155
 156 static void pipeclose (struct pipe *cpipe);
 157 static void pipe_free_kmem (struct pipe *cpipe);
 158 static int pipe_create (struct pipe **cpipep);
 159 static __inline void pipeselwakeup (struct pipe *cpipe);
 160 static int pipespace (struct pipe *cpipe, int size);
 161
 162 static __inline void
 163 pipeselwakeup(struct pipe *cpipe)
 164 {
 165         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 166                 get_mplock();
 167                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 168                 rel_mplock();
 169         }
 170         if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
 171                 get_mplock();
 172                 KNOTE(&cpipe->pipe_sel.si_note, 0);
 173                 rel_mplock();
 174         }
 175 }
 176
 177 /*
 178  * These routines are called before and after a UIO.  The UIO
 179  * may block, causing our held tokens to be lost temporarily.
 180  *
 181  * We use these routines to serialize reads against other reads
 182  * and writes against other writes.
 183  *
 184  * The read token is held on entry so *ipp does not race.
 185  */
 186 static __inline int
 187 pipe_start_uio(struct pipe *cpipe, int *ipp)
 188 {
 189         int error;
 190
 191         while (*ipp) {
 192                 *ipp = -1;
 193                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 194                 if (error)
 195                         return (error);
 196         }
 197         *ipp = 1;
 198         return (0);
 199 }
 200
 201 static __inline void
 202 pipe_end_uio(struct pipe *cpipe, int *ipp)
 203 {
 204         if (*ipp < 0) {
 205                 *ipp = 0;
 206                 wakeup(ipp);
 207         } else {
 208                 KKASSERT(*ipp > 0);
 209                 *ipp = 0;
 210         }
 211 }
 212
 213 static __inline void
 214 pipe_get_mplock(int *save)
 215 {
 216 #ifdef SMP
 217         if (pipe_mpsafe == 0) {
 218                 get_mplock();
 219                 *save = 1;
 220         } else
 221 #endif
 222         {
 223                 *save = 0;
 224         }
 225 }
 226
 227 static __inline void
 228 pipe_rel_mplock(int *save)
 229 {
 230 #ifdef SMP
 231         if (*save)
 232                 rel_mplock();
 233 #endif
 234 }
 235
 236
 237 /*
 238  * The pipe system call for the DTYPE_PIPE type of pipes
 239  *
 240  * pipe_args(int dummy)
 241  *
 242  * MPSAFE
 243  */
 244 int
 245 sys_pipe(struct pipe_args *uap)
 246 {
 247         struct thread *td = curthread;
 248         struct filedesc *fdp = td->td_proc->p_fd;
 249         struct file *rf, *wf;
 250         struct pipe *rpipe, *wpipe;
 251         int fd1, fd2, error;
 252
 253         rpipe = wpipe = NULL;
 254         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 255                 pipeclose(rpipe);
 256                 pipeclose(wpipe);
 257                 return (ENFILE);
 258         }
 259
 260         error = falloc(td->td_lwp, &rf, &fd1);
 261         if (error) {
 262                 pipeclose(rpipe);
 263                 pipeclose(wpipe);
 264                 return (error);
 265         }
 266         uap->sysmsg_fds[0] = fd1;
 267
 268         /*
 269          * Warning: once we've gotten past allocation of the fd for the
 270          * read-side, we can only drop the read side via fdrop() in order
 271          * to avoid races against processes which manage to dup() the read
 272          * side while we are blocked trying to allocate the write side.
 273          */
 274         rf->f_type = DTYPE_PIPE;
 275         rf->f_flag = FREAD | FWRITE;
 276         rf->f_ops = &pipeops;
 277         rf->f_data = rpipe;
 278         error = falloc(td->td_lwp, &wf, &fd2);
 279         if (error) {
 280                 fsetfd(fdp, NULL, fd1);
 281                 fdrop(rf);
 282                 /* rpipe has been closed by fdrop(). */
 283                 pipeclose(wpipe);
 284                 return (error);
 285         }
 286         wf->f_type = DTYPE_PIPE;
 287         wf->f_flag = FREAD | FWRITE;
 288         wf->f_ops = &pipeops;
 289         wf->f_data = wpipe;
 290         uap->sysmsg_fds[1] = fd2;
 291
 292         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 293                                     M_PIPE, M_WAITOK|M_ZERO);
 294         wpipe->pipe_slock = rpipe->pipe_slock;
 295         rpipe->pipe_peer = wpipe;
 296         wpipe->pipe_peer = rpipe;
 297         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 298
 299         /*
 300          * Once activated the peer relationship remains valid until
 301          * both sides are closed.
 302          */
 303         fsetfd(fdp, rf, fd1);
 304         fsetfd(fdp, wf, fd2);
 305         fdrop(rf);
 306         fdrop(wf);
 307
 308         return (0);
 309 }
 310
 311 /*
 312  * Allocate kva for pipe circular buffer, the space is pageable
 313  * This routine will 'realloc' the size of a pipe safely, if it fails
 314  * it will retain the old buffer.
 315  * If it fails it will return ENOMEM.
 316  */
 317 static int
 318 pipespace(struct pipe *cpipe, int size)
 319 {
 320         struct vm_object *object;
 321         caddr_t buffer;
 322         int npages, error;
 323
 324         npages = round_page(size) / PAGE_SIZE;
 325         object = cpipe->pipe_buffer.object;
 326
 327         /*
 328          * [re]create the object if necessary and reserve space for it
 329          * in the kernel_map.  The object and memory are pageable.  On
 330          * success, free the old resources before assigning the new
 331          * ones.
 332          */
 333         if (object == NULL || object->size != npages) {
 334                 get_mplock();
 335                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 336                 buffer = (caddr_t)vm_map_min(&kernel_map);
 337
 338                 error = vm_map_find(&kernel_map, object, 0,
 339                                     (vm_offset_t *)&buffer,
 340                                     size, PAGE_SIZE,
 341                                     1, VM_MAPTYPE_NORMAL,
 342                                     VM_PROT_ALL, VM_PROT_ALL,
 343                                     0);
 344
 345                 if (error != KERN_SUCCESS) {
 346                         vm_object_deallocate(object);
 347                         rel_mplock();
 348                         return (ENOMEM);
 349                 }
 350                 pipe_free_kmem(cpipe);
 351                 rel_mplock();
 352                 cpipe->pipe_buffer.object = object;
 353                 cpipe->pipe_buffer.buffer = buffer;
 354                 cpipe->pipe_buffer.size = size;
 355                 ++pipe_bkmem_alloc;
 356         } else {
 357                 ++pipe_bcache_alloc;
 358         }
 359         cpipe->pipe_buffer.rindex = 0;
 360         cpipe->pipe_buffer.windex = 0;
 361         return (0);
 362 }
 363
 364 /*
 365  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 366  * our per-cpu cache if possible.  For now make sure it is sized for the
 367  * smaller PIPE_SIZE default.
 368  */
 369 static int
 370 pipe_create(struct pipe **cpipep)
 371 {
 372         globaldata_t gd = mycpu;
 373         struct pipe *cpipe;
 374         int error;
 375
 376         if ((cpipe = gd->gd_pipeq) != NULL) {
 377                 gd->gd_pipeq = cpipe->pipe_peer;
 378                 --gd->gd_pipeqcount;
 379                 cpipe->pipe_peer = NULL;
 380                 cpipe->pipe_wantwcnt = 0;
 381         } else {
 382                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 383         }
 384         *cpipep = cpipe;
 385         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 386                 return (error);
 387         vfs_timestamp(&cpipe->pipe_ctime);
 388         cpipe->pipe_atime = cpipe->pipe_ctime;
 389         cpipe->pipe_mtime = cpipe->pipe_ctime;
 390         lwkt_token_init(&cpipe->pipe_rlock, 1);
 391         lwkt_token_init(&cpipe->pipe_wlock, 1);
 392         return (0);
 393 }
 394
 395 /*
 396  * MPALMOSTSAFE (acquires mplock)
 397  */
 398 static int
 399 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 400 {
 401         struct pipe *rpipe;
 402         int error;
 403         size_t nread = 0;
 404         int nbio;
 405         u_int size;     /* total bytes available */
 406         u_int nsize;    /* total bytes to read */
 407         u_int rindex;   /* contiguous bytes available */
 408         int notify_writer;
 409         int mpsave;
 410         int bigread;
 411         int bigcount;
 412
 413         if (uio->uio_resid == 0)
 414                 return(0);
 415
 416         /*
 417          * Setup locks, calculate nbio
 418          */
 419         pipe_get_mplock(&mpsave);
 420         rpipe = (struct pipe *)fp->f_data;
 421         lwkt_gettoken(&rpipe->pipe_rlock);
 422
 423         if (fflags & O_FBLOCKING)
 424                 nbio = 0;
 425         else if (fflags & O_FNONBLOCKING)
 426                 nbio = 1;
 427         else if (fp->f_flag & O_NONBLOCK)
 428                 nbio = 1;
 429         else
 430                 nbio = 0;
 431
 432         /*
 433          * Reads are serialized.  Note however that pipe_buffer.buffer and
 434          * pipe_buffer.size can change out from under us when the number
 435          * of bytes in the buffer are zero due to the write-side doing a
 436          * pipespace().
 437          */
 438         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 439         if (error) {
 440                 pipe_rel_mplock(&mpsave);
 441                 lwkt_reltoken(&rpipe->pipe_rlock);
 442                 return (error);
 443         }
 444         notify_writer = 0;
 445
 446         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 447         bigcount = 10;
 448
 449         while (uio->uio_resid) {
 450                 /*
 451                  * Don't hog the cpu.
 452                  */
 453                 if (bigread && --bigcount == 0) {
 454                         lwkt_user_yield();
 455                         bigcount = 10;
 456                         if (CURSIG(curthread->td_lwp)) {
 457                                 error = EINTR;
 458                                 break;
 459                         }
 460                 }
 461
 462                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 463                 cpu_lfence();
 464                 if (size) {
 465                         rindex = rpipe->pipe_buffer.rindex &
 466                                  (rpipe->pipe_buffer.size - 1);
 467                         nsize = size;
 468                         if (nsize > rpipe->pipe_buffer.size - rindex)
 469                                 nsize = rpipe->pipe_buffer.size - rindex;
 470                         nsize = szmin(nsize, uio->uio_resid);
 471
 472                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 473                                         nsize, uio);
 474                         if (error)
 475                                 break;
 476                         cpu_mfence();
 477                         rpipe->pipe_buffer.rindex += nsize;
 478                         nread += nsize;
 479
 480                         /*
 481                          * If the FIFO is still over half full just continue
 482                          * and do not try to notify the writer yet.
 483                          */
 484                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 485                                 notify_writer = 0;
 486                                 continue;
 487                         }
 488
 489                         /*
 490                          * When the FIFO is less then half full notify any
 491                          * waiting writer.  WANTW can be checked while
 492                          * holding just the rlock.
 493                          */
 494                         notify_writer = 1;
 495                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 496                                 continue;
 497                 }
 498
 499                 /*
 500                  * If the "write-side" was blocked we wake it up.  This code
 501                  * is reached either when the buffer is completely emptied
 502                  * or if it becomes more then half-empty.
 503                  *
 504                  * Pipe_state can only be modified if both the rlock and
 505                  * wlock are held.
 506                  */
 507                 if (rpipe->pipe_state & PIPE_WANTW) {
 508                         lwkt_gettoken(&rpipe->pipe_wlock);
 509                         if (rpipe->pipe_state & PIPE_WANTW) {
 510                                 notify_writer = 0;
 511                                 rpipe->pipe_state &= ~PIPE_WANTW;
 512                                 lwkt_reltoken(&rpipe->pipe_wlock);
 513                                 wakeup(rpipe);
 514                         } else {
 515                                 lwkt_reltoken(&rpipe->pipe_wlock);
 516                         }
 517                 }
 518
 519                 /*
 520                  * Pick up our copy loop again if the writer sent data to
 521                  * us while we were messing around.
 522                  *
 523                  * On a SMP box poll up to pipe_delay nanoseconds for new
 524                  * data.  Typically a value of 2000 to 4000 is sufficient
 525                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 526                  * is used for synchronous communications with small packets,
 527                  * and 8000 or so (8uS) will pipeline large buffer xfers
 528                  * between cpus over a pipe.
 529                  *
 530                  * For synchronous communications a hit means doing a
 531                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 532                  * where as miss requiring a tsleep/wakeup sequence
 533                  * will take 7uS or more.
 534                  */
 535                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 536                         continue;
 537
 538 #if defined(SMP) && defined(_RDTSC_SUPPORTED_)
 539                 if (pipe_delay) {
 540                         int64_t tsc_target;
 541                         int good = 0;
 542
 543                         tsc_target = tsc_get_target(pipe_delay);
 544                         while (tsc_test_target(tsc_target) == 0) {
 545                                 if (rpipe->pipe_buffer.windex !=
 546                                     rpipe->pipe_buffer.rindex) {
 547                                         good = 1;
 548                                         break;
 549                                 }
 550                         }
 551                         if (good)
 552                                 continue;
 553                 }
 554 #endif
 555
 556                 /*
 557                  * Detect EOF condition, do not set error.
 558                  */
 559                 if (rpipe->pipe_state & PIPE_REOF)
 560                         break;
 561
 562                 /*
 563                  * Break if some data was read, or if this was a non-blocking
 564                  * read.
 565                  */
 566                 if (nread > 0)
 567                         break;
 568
 569                 if (nbio) {
 570                         error = EAGAIN;
 571                         break;
 572                 }
 573
 574                 /*
 575                  * Last chance, interlock with WANTR.
 576                  */
 577                 lwkt_gettoken(&rpipe->pipe_wlock);
 578                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 579                 if (size) {
 580                         lwkt_reltoken(&rpipe->pipe_wlock);
 581                         continue;
 582                 }
 583
 584                 /*
 585                  * Retest EOF - acquiring a new token can temporarily release
 586                  * tokens already held.
 587                  */
 588                 if (rpipe->pipe_state & PIPE_REOF) {
 589                         lwkt_reltoken(&rpipe->pipe_wlock);
 590                         break;
 591                 }
 592
 593                 /*
 594                  * If there is no more to read in the pipe, reset its
 595                  * pointers to the beginning.  This improves cache hit
 596                  * stats.
 597                  *
 598                  * We need both locks to modify both pointers, and there
 599                  * must also not be a write in progress or the uiomove()
 600                  * in the write might block and temporarily release
 601                  * its wlock, then reacquire and update windex.  We are
 602                  * only serialized against reads, not writes.
 603                  *
 604                  * XXX should we even bother resetting the indices?  It
 605                  *     might actually be more cache efficient not to.
 606                  */
 607                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 608                     rpipe->pipe_wip == 0) {
 609                         rpipe->pipe_buffer.rindex = 0;
 610                         rpipe->pipe_buffer.windex = 0;
 611                 }
 612
 613                 /*
 614                  * Wait for more data.
 615                  *
 616                  * Pipe_state can only be set if both the rlock and wlock
 617                  * are held.
 618                  */
 619                 rpipe->pipe_state |= PIPE_WANTR;
 620                 tsleep_interlock(rpipe, PCATCH);
 621                 lwkt_reltoken(&rpipe->pipe_wlock);
 622                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 623                 ++pipe_rblocked_count;
 624                 if (error)
 625                         break;
 626         }
 627         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 628
 629         /*
 630          * Uptime last access time
 631          */
 632         if (error == 0 && nread)
 633                 vfs_timestamp(&rpipe->pipe_atime);
 634
 635         /*
 636          * If we drained the FIFO more then half way then handle
 637          * write blocking hysteresis.
 638          *
 639          * Note that PIPE_WANTW cannot be set by the writer without
 640          * it holding both rlock and wlock, so we can test it
 641          * while holding just rlock.
 642          */
 643         if (notify_writer) {
 644                 if (rpipe->pipe_state & PIPE_WANTW) {
 645                         lwkt_gettoken(&rpipe->pipe_wlock);
 646                         if (rpipe->pipe_state & PIPE_WANTW) {
 647                                 rpipe->pipe_state &= ~PIPE_WANTW;
 648                                 lwkt_reltoken(&rpipe->pipe_wlock);
 649                                 wakeup(rpipe);
 650                         } else {
 651                                 lwkt_reltoken(&rpipe->pipe_wlock);
 652                         }
 653                 }
 654                 lwkt_gettoken(&rpipe->pipe_wlock);
 655                 pipeselwakeup(rpipe);
 656                 lwkt_reltoken(&rpipe->pipe_wlock);
 657         }
 658         /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
 659         lwkt_reltoken(&rpipe->pipe_rlock);
 660
 661         pipe_rel_mplock(&mpsave);
 662         return (error);
 663 }
 664
 665 /*
 666  * MPALMOSTSAFE - acquires mplock
 667  */
 668 static int
 669 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 670 {
 671         int error;
 672         int orig_resid;
 673         int nbio;
 674         struct pipe *wpipe, *rpipe;
 675         u_int windex;
 676         u_int space;
 677         u_int wcount;
 678         int mpsave;
 679         int bigwrite;
 680         int bigcount;
 681
 682         pipe_get_mplock(&mpsave);
 683
 684         /*
 685          * Writes go to the peer.  The peer will always exist.
 686          */
 687         rpipe = (struct pipe *) fp->f_data;
 688         wpipe = rpipe->pipe_peer;
 689         lwkt_gettoken(&wpipe->pipe_wlock);
 690         if (wpipe->pipe_state & PIPE_WEOF) {
 691                 pipe_rel_mplock(&mpsave);
 692                 lwkt_reltoken(&wpipe->pipe_wlock);
 693                 return (EPIPE);
 694         }
 695
 696         /*
 697          * Degenerate case (EPIPE takes prec)
 698          */
 699         if (uio->uio_resid == 0) {
 700                 pipe_rel_mplock(&mpsave);
 701                 lwkt_reltoken(&wpipe->pipe_wlock);
 702                 return(0);
 703         }
 704
 705         /*
 706          * Writes are serialized (start_uio must be called with wlock)
 707          */
 708         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 709         if (error) {
 710                 pipe_rel_mplock(&mpsave);
 711                 lwkt_reltoken(&wpipe->pipe_wlock);
 712                 return (error);
 713         }
 714
 715         if (fflags & O_FBLOCKING)
 716                 nbio = 0;
 717         else if (fflags & O_FNONBLOCKING)
 718                 nbio = 1;
 719         else if (fp->f_flag & O_NONBLOCK)
 720                 nbio = 1;
 721         else
 722                 nbio = 0;
 723
 724         /*
 725          * If it is advantageous to resize the pipe buffer, do
 726          * so.  We are write-serialized so we can block safely.
 727          */
 728         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 729             (pipe_nbig < pipe_maxbig) &&
 730             wpipe->pipe_wantwcnt > 4 &&
 731             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 732                 /*
 733                  * Recheck after lock.
 734                  */
 735                 lwkt_gettoken(&wpipe->pipe_rlock);
 736                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 737                     (pipe_nbig < pipe_maxbig) &&
 738                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 739                         atomic_add_int(&pipe_nbig, 1);
 740                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 741                                 ++pipe_bigcount;
 742                         else
 743                                 atomic_subtract_int(&pipe_nbig, 1);
 744                 }
 745                 lwkt_reltoken(&wpipe->pipe_rlock);
 746         }
 747
 748         orig_resid = uio->uio_resid;
 749         wcount = 0;
 750
 751         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 752         bigcount = 10;
 753
 754         while (uio->uio_resid) {
 755                 if (wpipe->pipe_state & PIPE_WEOF) {
 756                         error = EPIPE;
 757                         break;
 758                 }
 759
 760                 /*
 761                  * Don't hog the cpu.
 762                  */
 763                 if (bigwrite && --bigcount == 0) {
 764                         lwkt_user_yield();
 765                         bigcount = 10;
 766                         if (CURSIG(curthread->td_lwp)) {
 767                                 error = EINTR;
 768                                 break;
 769                         }
 770                 }
 771
 772                 windex = wpipe->pipe_buffer.windex &
 773                          (wpipe->pipe_buffer.size - 1);
 774                 space = wpipe->pipe_buffer.size -
 775                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 776                 cpu_lfence();
 777
 778                 /* Writes of size <= PIPE_BUF must be atomic. */
 779                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 780                         space = 0;
 781
 782                 /*
 783                  * Write to fill, read size handles write hysteresis.  Also
 784                  * additional restrictions can cause select-based non-blocking
 785                  * writes to spin.
 786                  */
 787                 if (space > 0) {
 788                         u_int segsize;
 789
 790                         /*
 791                          * Transfer size is minimum of uio transfer
 792                          * and free space in pipe buffer.
 793                          *
 794                          * Limit each uiocopy to no more then PIPE_SIZE
 795                          * so we can keep the gravy train going on a
 796                          * SMP box.  This doubles the performance for
 797                          * write sizes > 16K.  Otherwise large writes
 798                          * wind up doing an inefficient synchronous
 799                          * ping-pong.
 800                          */
 801                         space = szmin(space, uio->uio_resid);
 802                         if (space > PIPE_SIZE)
 803                                 space = PIPE_SIZE;
 804
 805                         /*
 806                          * First segment to transfer is minimum of
 807                          * transfer size and contiguous space in
 808                          * pipe buffer.  If first segment to transfer
 809                          * is less than the transfer size, we've got
 810                          * a wraparound in the buffer.
 811                          */
 812                         segsize = wpipe->pipe_buffer.size - windex;
 813                         if (segsize > space)
 814                                 segsize = space;
 815
 816 #ifdef SMP
 817                         /*
 818                          * If this is the first loop and the reader is
 819                          * blocked, do a preemptive wakeup of the reader.
 820                          *
 821                          * On SMP the IPI latency plus the wlock interlock
 822                          * on the reader side is the fastest way to get the
 823                          * reader going.  (The scheduler will hard loop on
 824                          * lock tokens).
 825                          *
 826                          * NOTE: We can't clear WANTR here without acquiring
 827                          * the rlock, which we don't want to do here!
 828                          */
 829                         if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1)
 830                                 wakeup(wpipe);
 831 #endif
 832
 833                         /*
 834                          * Transfer segment, which may include a wrap-around.
 835                          * Update windex to account for both all in one go
 836                          * so the reader can read() the data atomically.
 837                          */
 838                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 839                                         segsize, uio);
 840                         if (error == 0 && segsize < space) {
 841                                 segsize = space - segsize;
 842                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 843                                                 segsize, uio);
 844                         }
 845                         if (error)
 846                                 break;
 847                         cpu_mfence();
 848                         wpipe->pipe_buffer.windex += space;
 849                         wcount += space;
 850                         continue;
 851                 }
 852
 853                 /*
 854                  * We need both the rlock and the wlock to interlock against
 855                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 856                  *
 857                  * These are token locks so we do not have to worry about
 858                  * deadlocks.
 859                  */
 860                 lwkt_gettoken(&wpipe->pipe_rlock);
 861
 862                 /*
 863                  * If the "read-side" has been blocked, wake it up now
 864                  * and yield to let it drain synchronously rather
 865                  * then block.
 866                  */
 867                 if (wpipe->pipe_state & PIPE_WANTR) {
 868                         wpipe->pipe_state &= ~PIPE_WANTR;
 869                         wakeup(wpipe);
 870                 }
 871
 872                 /*
 873                  * don't block on non-blocking I/O
 874                  */
 875                 if (nbio) {
 876                         lwkt_reltoken(&wpipe->pipe_rlock);
 877                         error = EAGAIN;
 878                         break;
 879                 }
 880
 881                 /*
 882                  * re-test whether we have to block in the writer after
 883                  * acquiring both locks, in case the reader opened up
 884                  * some space.
 885                  */
 886                 space = wpipe->pipe_buffer.size -
 887                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 888                 cpu_lfence();
 889                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 890                         space = 0;
 891
 892                 /*
 893                  * Retest EOF - acquiring a new token can temporarily release
 894                  * tokens already held.
 895                  */
 896                 if (wpipe->pipe_state & PIPE_WEOF) {
 897                         lwkt_reltoken(&wpipe->pipe_rlock);
 898                         error = EPIPE;
 899                         break;
 900                 }
 901
 902                 /*
 903                  * We have no more space and have something to offer,
 904                  * wake up select/poll.
 905                  */
 906                 if (space == 0) {
 907                         wpipe->pipe_state |= PIPE_WANTW;
 908                         ++wpipe->pipe_wantwcnt;
 909                         pipeselwakeup(wpipe);
 910                         if (wpipe->pipe_state & PIPE_WANTW)
 911                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 912                         ++pipe_wblocked_count;
 913                 }
 914                 lwkt_reltoken(&wpipe->pipe_rlock);
 915
 916                 /*
 917                  * Break out if we errored or the read side wants us to go
 918                  * away.
 919                  */
 920                 if (error)
 921                         break;
 922                 if (wpipe->pipe_state & PIPE_WEOF) {
 923                         error = EPIPE;
 924                         break;
 925                 }
 926         }
 927         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 928
 929         /*
 930          * If we have put any characters in the buffer, we wake up
 931          * the reader.
 932          *
 933          * Both rlock and wlock are required to be able to modify pipe_state.
 934          */
 935         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 936                 if (wpipe->pipe_state & PIPE_WANTR) {
 937                         lwkt_gettoken(&wpipe->pipe_rlock);
 938                         if (wpipe->pipe_state & PIPE_WANTR) {
 939                                 wpipe->pipe_state &= ~PIPE_WANTR;
 940                                 lwkt_reltoken(&wpipe->pipe_rlock);
 941                                 wakeup(wpipe);
 942                         } else {
 943                                 lwkt_reltoken(&wpipe->pipe_rlock);
 944                         }
 945                 }
 946                 lwkt_gettoken(&wpipe->pipe_rlock);
 947                 pipeselwakeup(wpipe);
 948                 lwkt_reltoken(&wpipe->pipe_rlock);
 949         }
 950
 951         /*
 952          * Don't return EPIPE if I/O was successful
 953          */
 954         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 955             (uio->uio_resid == 0) &&
 956             (error == EPIPE)) {
 957                 error = 0;
 958         }
 959
 960         if (error == 0)
 961                 vfs_timestamp(&wpipe->pipe_mtime);
 962
 963         /*
 964          * We have something to offer,
 965          * wake up select/poll.
 966          */
 967         /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
 968         lwkt_reltoken(&wpipe->pipe_wlock);
 969         pipe_rel_mplock(&mpsave);
 970         return (error);
 971 }
 972
 973 /*
 974  * MPALMOSTSAFE - acquires mplock
 975  *
 976  * we implement a very minimal set of ioctls for compatibility with sockets.
 977  */
 978 int
 979 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
 980            struct ucred *cred, struct sysmsg *msg)
 981 {
 982         struct pipe *mpipe;
 983         int error;
 984         int mpsave;
 985
 986         pipe_get_mplock(&mpsave);
 987         mpipe = (struct pipe *)fp->f_data;
 988
 989         lwkt_gettoken(&mpipe->pipe_rlock);
 990         lwkt_gettoken(&mpipe->pipe_wlock);
 991
 992         switch (cmd) {
 993         case FIOASYNC:
 994                 if (*(int *)data) {
 995                         mpipe->pipe_state |= PIPE_ASYNC;
 996                 } else {
 997                         mpipe->pipe_state &= ~PIPE_ASYNC;
 998                 }
 999                 error = 0;
1000                 break;
1001         case FIONREAD:
1002                 *(int *)data = mpipe->pipe_buffer.windex -
1003                                 mpipe->pipe_buffer.rindex;
1004                 error = 0;
1005                 break;
1006         case FIOSETOWN:
1007                 get_mplock();
1008                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1009                 rel_mplock();
1010                 break;
1011         case FIOGETOWN:
1012                 *(int *)data = fgetown(mpipe->pipe_sigio);
1013                 error = 0;
1014                 break;
1015         case TIOCSPGRP:
1016                 /* This is deprecated, FIOSETOWN should be used instead. */
1017                 get_mplock();
1018                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1019                 rel_mplock();
1020                 break;
1021
1022         case TIOCGPGRP:
1023                 /* This is deprecated, FIOGETOWN should be used instead. */
1024                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1025                 error = 0;
1026                 break;
1027         default:
1028                 error = ENOTTY;
1029                 break;
1030         }
1031         lwkt_reltoken(&mpipe->pipe_wlock);
1032         lwkt_reltoken(&mpipe->pipe_rlock);
1033         pipe_rel_mplock(&mpsave);
1034
1035         return (error);
1036 }
1037
1038 /*
1039  * MPSAFE
1040  */
1041 static int
1042 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1043 {
1044         struct pipe *pipe;
1045         int mpsave;
1046
1047         pipe_get_mplock(&mpsave);
1048         pipe = (struct pipe *)fp->f_data;
1049
1050         bzero((caddr_t)ub, sizeof(*ub));
1051         ub->st_mode = S_IFIFO;
1052         ub->st_blksize = pipe->pipe_buffer.size;
1053         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1054         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1055         ub->st_atimespec = pipe->pipe_atime;
1056         ub->st_mtimespec = pipe->pipe_mtime;
1057         ub->st_ctimespec = pipe->pipe_ctime;
1058         /*
1059          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1060          * st_flags, st_gen.
1061          * XXX (st_dev, st_ino) should be unique.
1062          */
1063         pipe_rel_mplock(&mpsave);
1064         return (0);
1065 }
1066
1067 /*
1068  * MPALMOSTSAFE - acquires mplock
1069  */
1070 static int
1071 pipe_close(struct file *fp)
1072 {
1073         struct pipe *cpipe;
1074
1075         get_mplock();
1076         cpipe = (struct pipe *)fp->f_data;
1077         fp->f_ops = &badfileops;
1078         fp->f_data = NULL;
1079         funsetown(cpipe->pipe_sigio);
1080         pipeclose(cpipe);
1081         rel_mplock();
1082         return (0);
1083 }
1084
1085 /*
1086  * Shutdown one or both directions of a full-duplex pipe.
1087  *
1088  * MPALMOSTSAFE - acquires mplock
1089  */
1090 static int
1091 pipe_shutdown(struct file *fp, int how)
1092 {
1093         struct pipe *rpipe;
1094         struct pipe *wpipe;
1095         int error = EPIPE;
1096         int mpsave;
1097
1098         pipe_get_mplock(&mpsave);
1099         rpipe = (struct pipe *)fp->f_data;
1100         wpipe = rpipe->pipe_peer;
1101
1102         /*
1103          * We modify pipe_state on both pipes, which means we need
1104          * all four tokens!
1105          */
1106         lwkt_gettoken(&rpipe->pipe_rlock);
1107         lwkt_gettoken(&rpipe->pipe_wlock);
1108         lwkt_gettoken(&wpipe->pipe_rlock);
1109         lwkt_gettoken(&wpipe->pipe_wlock);
1110
1111         switch(how) {
1112         case SHUT_RDWR:
1113         case SHUT_RD:
1114                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1115                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1116                 if (rpipe->pipe_state & PIPE_WANTR) {
1117                         rpipe->pipe_state &= ~PIPE_WANTR;
1118                         wakeup(rpipe);
1119                 }
1120                 if (rpipe->pipe_state & PIPE_WANTW) {
1121                         rpipe->pipe_state &= ~PIPE_WANTW;
1122                         wakeup(rpipe);
1123                 }
1124                 error = 0;
1125                 if (how == SHUT_RD)
1126                         break;
1127                 /* fall through */
1128         case SHUT_WR:
1129                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1130                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1131                 if (wpipe->pipe_state & PIPE_WANTR) {
1132                         wpipe->pipe_state &= ~PIPE_WANTR;
1133                         wakeup(wpipe);
1134                 }
1135                 if (wpipe->pipe_state & PIPE_WANTW) {
1136                         wpipe->pipe_state &= ~PIPE_WANTW;
1137                         wakeup(wpipe);
1138                 }
1139                 error = 0;
1140                 break;
1141         }
1142         pipeselwakeup(rpipe);
1143         pipeselwakeup(wpipe);
1144
1145         lwkt_reltoken(&wpipe->pipe_wlock);
1146         lwkt_reltoken(&wpipe->pipe_rlock);
1147         lwkt_reltoken(&rpipe->pipe_wlock);
1148         lwkt_reltoken(&rpipe->pipe_rlock);
1149
1150         pipe_rel_mplock(&mpsave);
1151         return (error);
1152 }
1153
1154 static void
1155 pipe_free_kmem(struct pipe *cpipe)
1156 {
1157         if (cpipe->pipe_buffer.buffer != NULL) {
1158                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1159                         atomic_subtract_int(&pipe_nbig, 1);
1160                 kmem_free(&kernel_map,
1161                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1162                         cpipe->pipe_buffer.size);
1163                 cpipe->pipe_buffer.buffer = NULL;
1164                 cpipe->pipe_buffer.object = NULL;
1165         }
1166 }
1167
1168 /*
1169  * Close the pipe.  The slock must be held to interlock against simultanious
1170  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1171  */
1172 static void
1173 pipeclose(struct pipe *cpipe)
1174 {
1175         globaldata_t gd;
1176         struct pipe *ppipe;
1177
1178         if (cpipe == NULL)
1179                 return;
1180
1181         /*
1182          * The slock may not have been allocated yet (close during
1183          * initialization)
1184          *
1185          * We need both the read and write tokens to modify pipe_state.
1186          */
1187         if (cpipe->pipe_slock)
1188                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1189         lwkt_gettoken(&cpipe->pipe_rlock);
1190         lwkt_gettoken(&cpipe->pipe_wlock);
1191
1192         /*
1193          * Set our state, wakeup anyone waiting in select, and
1194          * wakeup anyone blocked on our pipe.
1195          */
1196         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1197         pipeselwakeup(cpipe);
1198         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1199                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1200                 wakeup(cpipe);
1201         }
1202
1203         /*
1204          * Disconnect from peer.
1205          */
1206         if ((ppipe = cpipe->pipe_peer) != NULL) {
1207                 lwkt_gettoken(&ppipe->pipe_rlock);
1208                 lwkt_gettoken(&ppipe->pipe_wlock);
1209                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1210                 pipeselwakeup(ppipe);
1211                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1212                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1213                         wakeup(ppipe);
1214                 }
1215                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1216                         get_mplock();
1217                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1218                         rel_mplock();
1219                 }
1220                 lwkt_reltoken(&ppipe->pipe_wlock);
1221                 lwkt_reltoken(&ppipe->pipe_rlock);
1222         }
1223
1224         /*
1225          * If the peer is also closed we can free resources for both
1226          * sides, otherwise we leave our side intact to deal with any
1227          * races (since we only have the slock).
1228          */
1229         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1230                 cpipe->pipe_peer = NULL;
1231                 ppipe->pipe_peer = NULL;
1232                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1233                 pipeclose(ppipe);
1234                 ppipe = NULL;
1235         }
1236
1237         lwkt_reltoken(&cpipe->pipe_wlock);
1238         lwkt_reltoken(&cpipe->pipe_rlock);
1239         if (cpipe->pipe_slock)
1240                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1241
1242         /*
1243          * If we disassociated from our peer we can free resources
1244          */
1245         if (ppipe == NULL) {
1246                 gd = mycpu;
1247                 if (cpipe->pipe_slock) {
1248                         kfree(cpipe->pipe_slock, M_PIPE);
1249                         cpipe->pipe_slock = NULL;
1250                 }
1251                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1252                     cpipe->pipe_buffer.size != PIPE_SIZE
1253                 ) {
1254                         pipe_free_kmem(cpipe);
1255                         kfree(cpipe, M_PIPE);
1256                 } else {
1257                         cpipe->pipe_state = 0;
1258                         cpipe->pipe_peer = gd->gd_pipeq;
1259                         gd->gd_pipeq = cpipe;
1260                         ++gd->gd_pipeqcount;
1261                 }
1262         }
1263 }
1264
1265 /*
1266  * MPALMOSTSAFE - acquires mplock
1267  */
1268 static int
1269 pipe_kqfilter(struct file *fp, struct knote *kn)
1270 {
1271         struct pipe *cpipe;
1272
1273         cpipe = (struct pipe *)kn->kn_fp->f_data;
1274
1275         switch (kn->kn_filter) {
1276         case EVFILT_READ:
1277                 kn->kn_fop = &pipe_rfiltops;
1278                 break;
1279         case EVFILT_WRITE:
1280                 kn->kn_fop = &pipe_wfiltops;
1281                 if (cpipe->pipe_peer == NULL) {
1282                         /* other end of pipe has been closed */
1283                         rel_mplock();
1284                         return (EPIPE);
1285                 }
1286                 break;
1287         default:
1288                 return (EOPNOTSUPP);
1289         }
1290         kn->kn_hook = (caddr_t)cpipe;
1291
1292         crit_enter();
1293         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1294         crit_exit();
1295
1296         return (0);
1297 }
1298
1299 static void
1300 filt_pipedetach(struct knote *kn)
1301 {
1302         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1303
1304         crit_enter();
1305         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1306         crit_exit();
1307 }
1308
1309 /*ARGSUSED*/
1310 static int
1311 filt_piperead(struct knote *kn, long hint)
1312 {
1313         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1314
1315         /* XXX RACE */
1316         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1317         if (rpipe->pipe_state & PIPE_REOF) {
1318                 kn->kn_flags |= EV_EOF;
1319                 return (1);
1320         }
1321
1322         return (kn->kn_data > 0);
1323 }
1324
1325 /*ARGSUSED*/
1326 static int
1327 filt_pipewrite(struct knote *kn, long hint)
1328 {
1329         struct pipe *wpipe = (struct pipe *)kn->kn_fp->f_data;
1330         u_int32_t space;
1331
1332         /* XXX RACE */
1333         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1334                 kn->kn_data = 0;
1335                 kn->kn_flags |= EV_EOF;
1336                 return (1);
1337         }
1338         space = wpipe->pipe_buffer.windex -
1339                 wpipe->pipe_buffer.rindex;
1340         space = wpipe->pipe_buffer.size - space;
1341
1342         kn->kn_data = space;
1343         return (kn->kn_data >= PIPE_BUF);
1344 }