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  */
  21
  22 /*
  23  * This file contains a high-performance replacement for the socket-based
  24  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  25  * all features of sockets, but does do everything that pipes normally
  26  * do.
  27  */
  28 #include <sys/param.h>
  29 #include <sys/systm.h>
  30 #include <sys/kernel.h>
  31 #include <sys/proc.h>
  32 #include <sys/fcntl.h>
  33 #include <sys/file.h>
  34 #include <sys/filedesc.h>
  35 #include <sys/filio.h>
  36 #include <sys/ttycom.h>
  37 #include <sys/stat.h>
  38 #include <sys/signalvar.h>
  39 #include <sys/sysproto.h>
  40 #include <sys/pipe.h>
  41 #include <sys/vnode.h>
  42 #include <sys/uio.h>
  43 #include <sys/event.h>
  44 #include <sys/globaldata.h>
  45 #include <sys/module.h>
  46 #include <sys/malloc.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/socket.h>
  49
  50 #include <vm/vm.h>
  51 #include <vm/vm_param.h>
  52 #include <sys/lock.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_kern.h>
  55 #include <vm/vm_extern.h>
  56 #include <vm/pmap.h>
  57 #include <vm/vm_map.h>
  58 #include <vm/vm_page.h>
  59 #include <vm/vm_zone.h>
  60
  61 #include <sys/file2.h>
  62 #include <sys/signal2.h>
  63
  64 #include <machine/cpufunc.h>
  65
  66 /*
  67  * interfaces to the outside world
  68  */
  69 static int pipe_read (struct file *fp, struct uio *uio,
  70                 struct ucred *cred, int flags);
  71 static int pipe_write (struct file *fp, struct uio *uio,
  72                 struct ucred *cred, int flags);
  73 static int pipe_close (struct file *fp);
  74 static int pipe_shutdown (struct file *fp, int how);
  75 static int pipe_kqfilter (struct file *fp, struct knote *kn);
  76 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
  77 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data,
  78                 struct ucred *cred, struct sysmsg *msg);
  79
  80 static struct fileops pipeops = {
  81         .fo_read = pipe_read,
  82         .fo_write = pipe_write,
  83         .fo_ioctl = pipe_ioctl,
  84         .fo_kqfilter = pipe_kqfilter,
  85         .fo_stat = pipe_stat,
  86         .fo_close = pipe_close,
  87         .fo_shutdown = pipe_shutdown
  88 };
  89
  90 static void     filt_pipedetach(struct knote *kn);
  91 static int      filt_piperead(struct knote *kn, long hint);
  92 static int      filt_pipewrite(struct knote *kn, long hint);
  93
  94 static struct filterops pipe_rfiltops =
  95         { FILTEROP_ISFD|FILTEROP_MPSAFE, NULL, filt_pipedetach, filt_piperead };
  96 static struct filterops pipe_wfiltops =
  97         { FILTEROP_ISFD|FILTEROP_MPSAFE, NULL, filt_pipedetach, filt_pipewrite };
  98
  99 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 100
 101 /*
 102  * Default pipe buffer size(s), this can be kind-of large now because pipe
 103  * space is pageable.  The pipe code will try to maintain locality of
 104  * reference for performance reasons, so small amounts of outstanding I/O
 105  * will not wipe the cache.
 106  */
 107 #define MINPIPESIZE (PIPE_SIZE/3)
 108 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 109
 110 /*
 111  * Limit the number of "big" pipes
 112  */
 113 #define LIMITBIGPIPES   64
 114 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 115
 116 static int pipe_maxbig = LIMITBIGPIPES;
 117 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 118 static int pipe_bigcount;
 119 static int pipe_nbig;
 120 static int pipe_bcache_alloc;
 121 static int pipe_bkmem_alloc;
 122 static int pipe_rblocked_count;
 123 static int pipe_wblocked_count;
 124
 125 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 126 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 127         CTLFLAG_RD, &pipe_nbig, 0, "number of big pipes allocated");
 128 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 129         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 130 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked,
 131         CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
 132 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
 133         CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
 134 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 135         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 136 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 137         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 138 static int pipe_delay = 5000;   /* 5uS default */
 139 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
 140         CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
 141 #if !defined(NO_PIPE_SYSCTL_STATS)
 142 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 143         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 144 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 145         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 146 #endif
 147
 148 /*
 149  * Auto-size pipe cache to reduce kmem allocations and frees.
 150  */
 151 static
 152 void
 153 pipeinit(void *dummy)
 154 {
 155         size_t mbytes = kmem_lim_size();
 156
 157         if (pipe_maxbig == LIMITBIGPIPES) {
 158                 if (mbytes >= 7 * 1024)
 159                         pipe_maxbig *= 2;
 160                 if (mbytes >= 15 * 1024)
 161                         pipe_maxbig *= 2;
 162         }
 163         if (pipe_maxcache == PIPEQ_MAX_CACHE) {
 164                 if (mbytes >= 7 * 1024)
 165                         pipe_maxcache *= 2;
 166                 if (mbytes >= 15 * 1024)
 167                         pipe_maxcache *= 2;
 168         }
 169 }
 170 SYSINIT(kmem, SI_BOOT2_MACHDEP, SI_ORDER_ANY, pipeinit, NULL)
 171
 172 static void pipeclose (struct pipe *cpipe);
 173 static void pipe_free_kmem (struct pipe *cpipe);
 174 static int pipe_create (struct pipe **cpipep);
 175 static int pipespace (struct pipe *cpipe, int size);
 176
 177 static __inline void
 178 pipewakeup(struct pipe *cpipe, int dosigio)
 179 {
 180         if (dosigio && (cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 181                 lwkt_gettoken(&sigio_token);
 182                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 183                 lwkt_reltoken(&sigio_token);
 184         }
 185         KNOTE(&cpipe->pipe_kq.ki_note, 0);
 186 }
 187
 188 /*
 189  * These routines are called before and after a UIO.  The UIO
 190  * may block, causing our held tokens to be lost temporarily.
 191  *
 192  * We use these routines to serialize reads against other reads
 193  * and writes against other writes.
 194  *
 195  * The read token is held on entry so *ipp does not race.
 196  */
 197 static __inline int
 198 pipe_start_uio(struct pipe *cpipe, int *ipp)
 199 {
 200         int error;
 201
 202         while (*ipp) {
 203                 *ipp = -1;
 204                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 205                 if (error)
 206                         return (error);
 207         }
 208         *ipp = 1;
 209         return (0);
 210 }
 211
 212 static __inline void
 213 pipe_end_uio(struct pipe *cpipe, int *ipp)
 214 {
 215         if (*ipp < 0) {
 216                 *ipp = 0;
 217                 wakeup(ipp);
 218         } else {
 219                 KKASSERT(*ipp > 0);
 220                 *ipp = 0;
 221         }
 222 }
 223
 224 /*
 225  * The pipe system call for the DTYPE_PIPE type of pipes
 226  *
 227  * pipe_args(int dummy)
 228  *
 229  * MPSAFE
 230  */
 231 int
 232 sys_pipe(struct pipe_args *uap)
 233 {
 234         struct thread *td = curthread;
 235         struct filedesc *fdp = td->td_proc->p_fd;
 236         struct file *rf, *wf;
 237         struct pipe *rpipe, *wpipe;
 238         int fd1, fd2, error;
 239
 240         rpipe = wpipe = NULL;
 241         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 242                 pipeclose(rpipe);
 243                 pipeclose(wpipe);
 244                 return (ENFILE);
 245         }
 246
 247         error = falloc(td->td_lwp, &rf, &fd1);
 248         if (error) {
 249                 pipeclose(rpipe);
 250                 pipeclose(wpipe);
 251                 return (error);
 252         }
 253         uap->sysmsg_fds[0] = fd1;
 254
 255         /*
 256          * Warning: once we've gotten past allocation of the fd for the
 257          * read-side, we can only drop the read side via fdrop() in order
 258          * to avoid races against processes which manage to dup() the read
 259          * side while we are blocked trying to allocate the write side.
 260          */
 261         rf->f_type = DTYPE_PIPE;
 262         rf->f_flag = FREAD | FWRITE;
 263         rf->f_ops = &pipeops;
 264         rf->f_data = rpipe;
 265         error = falloc(td->td_lwp, &wf, &fd2);
 266         if (error) {
 267                 fsetfd(fdp, NULL, fd1);
 268                 fdrop(rf);
 269                 /* rpipe has been closed by fdrop(). */
 270                 pipeclose(wpipe);
 271                 return (error);
 272         }
 273         wf->f_type = DTYPE_PIPE;
 274         wf->f_flag = FREAD | FWRITE;
 275         wf->f_ops = &pipeops;
 276         wf->f_data = wpipe;
 277         uap->sysmsg_fds[1] = fd2;
 278
 279         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 280                                     M_PIPE, M_WAITOK|M_ZERO);
 281         wpipe->pipe_slock = rpipe->pipe_slock;
 282         rpipe->pipe_peer = wpipe;
 283         wpipe->pipe_peer = rpipe;
 284         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 285
 286         /*
 287          * Once activated the peer relationship remains valid until
 288          * both sides are closed.
 289          */
 290         fsetfd(fdp, rf, fd1);
 291         fsetfd(fdp, wf, fd2);
 292         fdrop(rf);
 293         fdrop(wf);
 294
 295         return (0);
 296 }
 297
 298 /*
 299  * Allocate kva for pipe circular buffer, the space is pageable
 300  * This routine will 'realloc' the size of a pipe safely, if it fails
 301  * it will retain the old buffer.
 302  * If it fails it will return ENOMEM.
 303  */
 304 static int
 305 pipespace(struct pipe *cpipe, int size)
 306 {
 307         struct vm_object *object;
 308         caddr_t buffer;
 309         int npages, error;
 310
 311         npages = round_page(size) / PAGE_SIZE;
 312         object = cpipe->pipe_buffer.object;
 313
 314         /*
 315          * [re]create the object if necessary and reserve space for it
 316          * in the kernel_map.  The object and memory are pageable.  On
 317          * success, free the old resources before assigning the new
 318          * ones.
 319          */
 320         if (object == NULL || object->size != npages) {
 321                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 322                 buffer = (caddr_t)vm_map_min(&kernel_map);
 323
 324                 error = vm_map_find(&kernel_map, object, NULL,
 325                                     0, (vm_offset_t *)&buffer, size,
 326                                     PAGE_SIZE,
 327                                     1, VM_MAPTYPE_NORMAL,
 328                                     VM_PROT_ALL, VM_PROT_ALL, 0);
 329
 330                 if (error != KERN_SUCCESS) {
 331                         vm_object_deallocate(object);
 332                         return (ENOMEM);
 333                 }
 334                 pipe_free_kmem(cpipe);
 335                 cpipe->pipe_buffer.object = object;
 336                 cpipe->pipe_buffer.buffer = buffer;
 337                 cpipe->pipe_buffer.size = size;
 338                 ++pipe_bkmem_alloc;
 339         } else {
 340                 ++pipe_bcache_alloc;
 341         }
 342         cpipe->pipe_buffer.rindex = 0;
 343         cpipe->pipe_buffer.windex = 0;
 344         return (0);
 345 }
 346
 347 /*
 348  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 349  * our per-cpu cache if possible.  For now make sure it is sized for the
 350  * smaller PIPE_SIZE default.
 351  */
 352 static int
 353 pipe_create(struct pipe **cpipep)
 354 {
 355         globaldata_t gd = mycpu;
 356         struct pipe *cpipe;
 357         int error;
 358
 359         if ((cpipe = gd->gd_pipeq) != NULL) {
 360                 gd->gd_pipeq = cpipe->pipe_peer;
 361                 --gd->gd_pipeqcount;
 362                 cpipe->pipe_peer = NULL;
 363                 cpipe->pipe_wantwcnt = 0;
 364         } else {
 365                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 366         }
 367         *cpipep = cpipe;
 368         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 369                 return (error);
 370         vfs_timestamp(&cpipe->pipe_ctime);
 371         cpipe->pipe_atime = cpipe->pipe_ctime;
 372         cpipe->pipe_mtime = cpipe->pipe_ctime;
 373         lwkt_token_init(&cpipe->pipe_rlock, "piper");
 374         lwkt_token_init(&cpipe->pipe_wlock, "pipew");
 375         return (0);
 376 }
 377
 378 static int
 379 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 380 {
 381         struct pipe *rpipe;
 382         struct pipe *wpipe;
 383         int error;
 384         size_t nread = 0;
 385         int nbio;
 386         u_int size;     /* total bytes available */
 387         u_int nsize;    /* total bytes to read */
 388         u_int rindex;   /* contiguous bytes available */
 389         int notify_writer;
 390         int bigread;
 391         int bigcount;
 392
 393         atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
 394
 395         if (uio->uio_resid == 0)
 396                 return(0);
 397
 398         /*
 399          * Setup locks, calculate nbio
 400          */
 401         rpipe = (struct pipe *)fp->f_data;
 402         wpipe = rpipe->pipe_peer;
 403         lwkt_gettoken(&rpipe->pipe_rlock);
 404
 405         if (fflags & O_FBLOCKING)
 406                 nbio = 0;
 407         else if (fflags & O_FNONBLOCKING)
 408                 nbio = 1;
 409         else if (fp->f_flag & O_NONBLOCK)
 410                 nbio = 1;
 411         else
 412                 nbio = 0;
 413
 414         /*
 415          * Reads are serialized.  Note however that pipe_buffer.buffer and
 416          * pipe_buffer.size can change out from under us when the number
 417          * of bytes in the buffer are zero due to the write-side doing a
 418          * pipespace().
 419          */
 420         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 421         if (error) {
 422                 lwkt_reltoken(&rpipe->pipe_rlock);
 423                 return (error);
 424         }
 425         notify_writer = 0;
 426
 427         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 428         bigcount = 10;
 429
 430         while (uio->uio_resid) {
 431                 /*
 432                  * Don't hog the cpu.
 433                  */
 434                 if (bigread && --bigcount == 0) {
 435                         lwkt_user_yield();
 436                         bigcount = 10;
 437                         if (CURSIG(curthread->td_lwp)) {
 438                                 error = EINTR;
 439                                 break;
 440                         }
 441                 }
 442
 443                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 444                 cpu_lfence();
 445                 if (size) {
 446                         rindex = rpipe->pipe_buffer.rindex &
 447                                  (rpipe->pipe_buffer.size - 1);
 448                         nsize = size;
 449                         if (nsize > rpipe->pipe_buffer.size - rindex)
 450                                 nsize = rpipe->pipe_buffer.size - rindex;
 451                         nsize = szmin(nsize, uio->uio_resid);
 452
 453                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 454                                         nsize, uio);
 455                         if (error)
 456                                 break;
 457                         cpu_mfence();
 458                         rpipe->pipe_buffer.rindex += nsize;
 459                         nread += nsize;
 460
 461                         /*
 462                          * If the FIFO is still over half full just continue
 463                          * and do not try to notify the writer yet.
 464                          */
 465                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 466                                 notify_writer = 0;
 467                                 continue;
 468                         }
 469
 470                         /*
 471                          * When the FIFO is less then half full notify any
 472                          * waiting writer.  WANTW can be checked while
 473                          * holding just the rlock.
 474                          */
 475                         notify_writer = 1;
 476                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 477                                 continue;
 478                 }
 479
 480                 /*
 481                  * If the "write-side" was blocked we wake it up.  This code
 482                  * is reached either when the buffer is completely emptied
 483                  * or if it becomes more then half-empty.
 484                  *
 485                  * Pipe_state can only be modified if both the rlock and
 486                  * wlock are held.
 487                  */
 488                 if (rpipe->pipe_state & PIPE_WANTW) {
 489                         lwkt_gettoken(&rpipe->pipe_wlock);
 490                         if (rpipe->pipe_state & PIPE_WANTW) {
 491                                 rpipe->pipe_state &= ~PIPE_WANTW;
 492                                 lwkt_reltoken(&rpipe->pipe_wlock);
 493                                 wakeup(rpipe);
 494                         } else {
 495                                 lwkt_reltoken(&rpipe->pipe_wlock);
 496                         }
 497                 }
 498
 499                 /*
 500                  * Pick up our copy loop again if the writer sent data to
 501                  * us while we were messing around.
 502                  *
 503                  * On a SMP box poll up to pipe_delay nanoseconds for new
 504                  * data.  Typically a value of 2000 to 4000 is sufficient
 505                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 506                  * is used for synchronous communications with small packets,
 507                  * and 8000 or so (8uS) will pipeline large buffer xfers
 508                  * between cpus over a pipe.
 509                  *
 510                  * For synchronous communications a hit means doing a
 511                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 512                  * where as miss requiring a tsleep/wakeup sequence
 513                  * will take 7uS or more.
 514                  */
 515                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 516                         continue;
 517
 518 #ifdef _RDTSC_SUPPORTED_
 519                 if (pipe_delay) {
 520                         int64_t tsc_target;
 521                         int good = 0;
 522
 523                         tsc_target = tsc_get_target(pipe_delay);
 524                         while (tsc_test_target(tsc_target) == 0) {
 525                                 if (rpipe->pipe_buffer.windex !=
 526                                     rpipe->pipe_buffer.rindex) {
 527                                         good = 1;
 528                                         break;
 529                                 }
 530                         }
 531                         if (good)
 532                                 continue;
 533                 }
 534 #endif
 535
 536                 /*
 537                  * Detect EOF condition, do not set error.
 538                  */
 539                 if (rpipe->pipe_state & PIPE_REOF)
 540                         break;
 541
 542                 /*
 543                  * Break if some data was read, or if this was a non-blocking
 544                  * read.
 545                  */
 546                 if (nread > 0)
 547                         break;
 548
 549                 if (nbio) {
 550                         error = EAGAIN;
 551                         break;
 552                 }
 553
 554                 /*
 555                  * Last chance, interlock with WANTR.
 556                  */
 557                 lwkt_gettoken(&rpipe->pipe_wlock);
 558                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 559                 if (size) {
 560                         lwkt_reltoken(&rpipe->pipe_wlock);
 561                         continue;
 562                 }
 563
 564                 /*
 565                  * Retest EOF - acquiring a new token can temporarily release
 566                  * tokens already held.
 567                  */
 568                 if (rpipe->pipe_state & PIPE_REOF) {
 569                         lwkt_reltoken(&rpipe->pipe_wlock);
 570                         break;
 571                 }
 572
 573                 /*
 574                  * If there is no more to read in the pipe, reset its
 575                  * pointers to the beginning.  This improves cache hit
 576                  * stats.
 577                  *
 578                  * We need both locks to modify both pointers, and there
 579                  * must also not be a write in progress or the uiomove()
 580                  * in the write might block and temporarily release
 581                  * its wlock, then reacquire and update windex.  We are
 582                  * only serialized against reads, not writes.
 583                  *
 584                  * XXX should we even bother resetting the indices?  It
 585                  *     might actually be more cache efficient not to.
 586                  */
 587                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 588                     rpipe->pipe_wip == 0) {
 589                         rpipe->pipe_buffer.rindex = 0;
 590                         rpipe->pipe_buffer.windex = 0;
 591                 }
 592
 593                 /*
 594                  * Wait for more data.
 595                  *
 596                  * Pipe_state can only be set if both the rlock and wlock
 597                  * are held.
 598                  */
 599                 rpipe->pipe_state |= PIPE_WANTR;
 600                 tsleep_interlock(rpipe, PCATCH);
 601                 lwkt_reltoken(&rpipe->pipe_wlock);
 602                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 603                 ++pipe_rblocked_count;
 604                 if (error)
 605                         break;
 606         }
 607         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 608
 609         /*
 610          * Uptime last access time
 611          */
 612         if (error == 0 && nread)
 613                 vfs_timestamp(&rpipe->pipe_atime);
 614
 615         /*
 616          * If we drained the FIFO more then half way then handle
 617          * write blocking hysteresis.
 618          *
 619          * Note that PIPE_WANTW cannot be set by the writer without
 620          * it holding both rlock and wlock, so we can test it
 621          * while holding just rlock.
 622          */
 623         if (notify_writer) {
 624                 /*
 625                  * Synchronous blocking is done on the pipe involved
 626                  */
 627                 if (rpipe->pipe_state & PIPE_WANTW) {
 628                         lwkt_gettoken(&rpipe->pipe_wlock);
 629                         if (rpipe->pipe_state & PIPE_WANTW) {
 630                                 rpipe->pipe_state &= ~PIPE_WANTW;
 631                                 lwkt_reltoken(&rpipe->pipe_wlock);
 632                                 wakeup(rpipe);
 633                         } else {
 634                                 lwkt_reltoken(&rpipe->pipe_wlock);
 635                         }
 636                 }
 637
 638                 /*
 639                  * But we may also have to deal with a kqueue which is
 640                  * stored on the same pipe as its descriptor, so a
 641                  * EVFILT_WRITE event waiting for our side to drain will
 642                  * be on the other side.
 643                  */
 644                 lwkt_gettoken(&wpipe->pipe_wlock);
 645                 pipewakeup(wpipe, 0);
 646                 lwkt_reltoken(&wpipe->pipe_wlock);
 647         }
 648         /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
 649         lwkt_reltoken(&rpipe->pipe_rlock);
 650
 651         return (error);
 652 }
 653
 654 static int
 655 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 656 {
 657         int error;
 658         int orig_resid;
 659         int nbio;
 660         struct pipe *wpipe;
 661         struct pipe *rpipe;
 662         u_int windex;
 663         u_int space;
 664         u_int wcount;
 665         int bigwrite;
 666         int bigcount;
 667
 668         /*
 669          * Writes go to the peer.  The peer will always exist.
 670          */
 671         rpipe = (struct pipe *) fp->f_data;
 672         wpipe = rpipe->pipe_peer;
 673         lwkt_gettoken(&wpipe->pipe_wlock);
 674         if (wpipe->pipe_state & PIPE_WEOF) {
 675                 lwkt_reltoken(&wpipe->pipe_wlock);
 676                 return (EPIPE);
 677         }
 678
 679         /*
 680          * Degenerate case (EPIPE takes prec)
 681          */
 682         if (uio->uio_resid == 0) {
 683                 lwkt_reltoken(&wpipe->pipe_wlock);
 684                 return(0);
 685         }
 686
 687         /*
 688          * Writes are serialized (start_uio must be called with wlock)
 689          */
 690         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 691         if (error) {
 692                 lwkt_reltoken(&wpipe->pipe_wlock);
 693                 return (error);
 694         }
 695
 696         if (fflags & O_FBLOCKING)
 697                 nbio = 0;
 698         else if (fflags & O_FNONBLOCKING)
 699                 nbio = 1;
 700         else if (fp->f_flag & O_NONBLOCK)
 701                 nbio = 1;
 702         else
 703                 nbio = 0;
 704
 705         /*
 706          * If it is advantageous to resize the pipe buffer, do
 707          * so.  We are write-serialized so we can block safely.
 708          */
 709         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 710             (pipe_nbig < pipe_maxbig) &&
 711             wpipe->pipe_wantwcnt > 4 &&
 712             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 713                 /*
 714                  * Recheck after lock.
 715                  */
 716                 lwkt_gettoken(&wpipe->pipe_rlock);
 717                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 718                     (pipe_nbig < pipe_maxbig) &&
 719                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 720                         atomic_add_int(&pipe_nbig, 1);
 721                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 722                                 ++pipe_bigcount;
 723                         else
 724                                 atomic_subtract_int(&pipe_nbig, 1);
 725                 }
 726                 lwkt_reltoken(&wpipe->pipe_rlock);
 727         }
 728
 729         orig_resid = uio->uio_resid;
 730         wcount = 0;
 731
 732         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 733         bigcount = 10;
 734
 735         while (uio->uio_resid) {
 736                 if (wpipe->pipe_state & PIPE_WEOF) {
 737                         error = EPIPE;
 738                         break;
 739                 }
 740
 741                 /*
 742                  * Don't hog the cpu.
 743                  */
 744                 if (bigwrite && --bigcount == 0) {
 745                         lwkt_user_yield();
 746                         bigcount = 10;
 747                         if (CURSIG(curthread->td_lwp)) {
 748                                 error = EINTR;
 749                                 break;
 750                         }
 751                 }
 752
 753                 windex = wpipe->pipe_buffer.windex &
 754                          (wpipe->pipe_buffer.size - 1);
 755                 space = wpipe->pipe_buffer.size -
 756                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 757                 cpu_lfence();
 758
 759                 /* Writes of size <= PIPE_BUF must be atomic. */
 760                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 761                         space = 0;
 762
 763                 /*
 764                  * Write to fill, read size handles write hysteresis.  Also
 765                  * additional restrictions can cause select-based non-blocking
 766                  * writes to spin.
 767                  */
 768                 if (space > 0) {
 769                         u_int segsize;
 770
 771                         /*
 772                          * Transfer size is minimum of uio transfer
 773                          * and free space in pipe buffer.
 774                          *
 775                          * Limit each uiocopy to no more then PIPE_SIZE
 776                          * so we can keep the gravy train going on a
 777                          * SMP box.  This doubles the performance for
 778                          * write sizes > 16K.  Otherwise large writes
 779                          * wind up doing an inefficient synchronous
 780                          * ping-pong.
 781                          */
 782                         space = szmin(space, uio->uio_resid);
 783                         if (space > PIPE_SIZE)
 784                                 space = PIPE_SIZE;
 785
 786                         /*
 787                          * First segment to transfer is minimum of
 788                          * transfer size and contiguous space in
 789                          * pipe buffer.  If first segment to transfer
 790                          * is less than the transfer size, we've got
 791                          * a wraparound in the buffer.
 792                          */
 793                         segsize = wpipe->pipe_buffer.size - windex;
 794                         if (segsize > space)
 795                                 segsize = space;
 796
 797                         /*
 798                          * If this is the first loop and the reader is
 799                          * blocked, do a preemptive wakeup of the reader.
 800                          *
 801                          * On SMP the IPI latency plus the wlock interlock
 802                          * on the reader side is the fastest way to get the
 803                          * reader going.  (The scheduler will hard loop on
 804                          * lock tokens).
 805                          *
 806                          * NOTE: We can't clear WANTR here without acquiring
 807                          * the rlock, which we don't want to do here!
 808                          */
 809                         if ((wpipe->pipe_state & PIPE_WANTR))
 810                                 wakeup(wpipe);
 811
 812                         /*
 813                          * Transfer segment, which may include a wrap-around.
 814                          * Update windex to account for both all in one go
 815                          * so the reader can read() the data atomically.
 816                          */
 817                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 818                                         segsize, uio);
 819                         if (error == 0 && segsize < space) {
 820                                 segsize = space - segsize;
 821                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 822                                                 segsize, uio);
 823                         }
 824                         if (error)
 825                                 break;
 826                         cpu_mfence();
 827                         wpipe->pipe_buffer.windex += space;
 828                         wcount += space;
 829                         continue;
 830                 }
 831
 832                 /*
 833                  * We need both the rlock and the wlock to interlock against
 834                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 835                  *
 836                  * These are token locks so we do not have to worry about
 837                  * deadlocks.
 838                  */
 839                 lwkt_gettoken(&wpipe->pipe_rlock);
 840
 841                 /*
 842                  * If the "read-side" has been blocked, wake it up now
 843                  * and yield to let it drain synchronously rather
 844                  * then block.
 845                  */
 846                 if (wpipe->pipe_state & PIPE_WANTR) {
 847                         wpipe->pipe_state &= ~PIPE_WANTR;
 848                         wakeup(wpipe);
 849                 }
 850
 851                 /*
 852                  * don't block on non-blocking I/O
 853                  */
 854                 if (nbio) {
 855                         lwkt_reltoken(&wpipe->pipe_rlock);
 856                         error = EAGAIN;
 857                         break;
 858                 }
 859
 860                 /*
 861                  * re-test whether we have to block in the writer after
 862                  * acquiring both locks, in case the reader opened up
 863                  * some space.
 864                  */
 865                 space = wpipe->pipe_buffer.size -
 866                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 867                 cpu_lfence();
 868                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 869                         space = 0;
 870
 871                 /*
 872                  * Retest EOF - acquiring a new token can temporarily release
 873                  * tokens already held.
 874                  */
 875                 if (wpipe->pipe_state & PIPE_WEOF) {
 876                         lwkt_reltoken(&wpipe->pipe_rlock);
 877                         error = EPIPE;
 878                         break;
 879                 }
 880
 881                 /*
 882                  * We have no more space and have something to offer,
 883                  * wake up select/poll/kq.
 884                  */
 885                 if (space == 0) {
 886                         wpipe->pipe_state |= PIPE_WANTW;
 887                         ++wpipe->pipe_wantwcnt;
 888                         pipewakeup(wpipe, 1);
 889                         if (wpipe->pipe_state & PIPE_WANTW)
 890                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 891                         ++pipe_wblocked_count;
 892                 }
 893                 lwkt_reltoken(&wpipe->pipe_rlock);
 894
 895                 /*
 896                  * Break out if we errored or the read side wants us to go
 897                  * away.
 898                  */
 899                 if (error)
 900                         break;
 901                 if (wpipe->pipe_state & PIPE_WEOF) {
 902                         error = EPIPE;
 903                         break;
 904                 }
 905         }
 906         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 907
 908         /*
 909          * If we have put any characters in the buffer, we wake up
 910          * the reader.
 911          *
 912          * Both rlock and wlock are required to be able to modify pipe_state.
 913          */
 914         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 915                 if (wpipe->pipe_state & PIPE_WANTR) {
 916                         lwkt_gettoken(&wpipe->pipe_rlock);
 917                         if (wpipe->pipe_state & PIPE_WANTR) {
 918                                 wpipe->pipe_state &= ~PIPE_WANTR;
 919                                 lwkt_reltoken(&wpipe->pipe_rlock);
 920                                 wakeup(wpipe);
 921                         } else {
 922                                 lwkt_reltoken(&wpipe->pipe_rlock);
 923                         }
 924                 }
 925                 lwkt_gettoken(&wpipe->pipe_rlock);
 926                 pipewakeup(wpipe, 1);
 927                 lwkt_reltoken(&wpipe->pipe_rlock);
 928         }
 929
 930         /*
 931          * Don't return EPIPE if I/O was successful
 932          */
 933         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 934             (uio->uio_resid == 0) &&
 935             (error == EPIPE)) {
 936                 error = 0;
 937         }
 938
 939         if (error == 0)
 940                 vfs_timestamp(&wpipe->pipe_mtime);
 941
 942         /*
 943          * We have something to offer,
 944          * wake up select/poll/kq.
 945          */
 946         /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
 947         lwkt_reltoken(&wpipe->pipe_wlock);
 948         return (error);
 949 }
 950
 951 /*
 952  * we implement a very minimal set of ioctls for compatibility with sockets.
 953  */
 954 int
 955 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
 956            struct ucred *cred, struct sysmsg *msg)
 957 {
 958         struct pipe *mpipe;
 959         int error;
 960
 961         mpipe = (struct pipe *)fp->f_data;
 962
 963         lwkt_gettoken(&mpipe->pipe_rlock);
 964         lwkt_gettoken(&mpipe->pipe_wlock);
 965
 966         switch (cmd) {
 967         case FIOASYNC:
 968                 if (*(int *)data) {
 969                         mpipe->pipe_state |= PIPE_ASYNC;
 970                 } else {
 971                         mpipe->pipe_state &= ~PIPE_ASYNC;
 972                 }
 973                 error = 0;
 974                 break;
 975         case FIONREAD:
 976                 *(int *)data = mpipe->pipe_buffer.windex -
 977                                 mpipe->pipe_buffer.rindex;
 978                 error = 0;
 979                 break;
 980         case FIOSETOWN:
 981                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
 982                 break;
 983         case FIOGETOWN:
 984                 *(int *)data = fgetown(&mpipe->pipe_sigio);
 985                 error = 0;
 986                 break;
 987         case TIOCSPGRP:
 988                 /* This is deprecated, FIOSETOWN should be used instead. */
 989                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
 990                 break;
 991
 992         case TIOCGPGRP:
 993                 /* This is deprecated, FIOGETOWN should be used instead. */
 994                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
 995                 error = 0;
 996                 break;
 997         default:
 998                 error = ENOTTY;
 999                 break;
1000         }
1001         lwkt_reltoken(&mpipe->pipe_wlock);
1002         lwkt_reltoken(&mpipe->pipe_rlock);
1003
1004         return (error);
1005 }
1006
1007 /*
1008  * MPSAFE
1009  */
1010 static int
1011 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1012 {
1013         struct pipe *pipe;
1014
1015         pipe = (struct pipe *)fp->f_data;
1016
1017         bzero((caddr_t)ub, sizeof(*ub));
1018         ub->st_mode = S_IFIFO;
1019         ub->st_blksize = pipe->pipe_buffer.size;
1020         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1021         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1022         ub->st_atimespec = pipe->pipe_atime;
1023         ub->st_mtimespec = pipe->pipe_mtime;
1024         ub->st_ctimespec = pipe->pipe_ctime;
1025         /*
1026          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1027          * st_flags, st_gen.
1028          * XXX (st_dev, st_ino) should be unique.
1029          */
1030         return (0);
1031 }
1032
1033 static int
1034 pipe_close(struct file *fp)
1035 {
1036         struct pipe *cpipe;
1037
1038         cpipe = (struct pipe *)fp->f_data;
1039         fp->f_ops = &badfileops;
1040         fp->f_data = NULL;
1041         funsetown(&cpipe->pipe_sigio);
1042         pipeclose(cpipe);
1043         return (0);
1044 }
1045
1046 /*
1047  * Shutdown one or both directions of a full-duplex pipe.
1048  */
1049 static int
1050 pipe_shutdown(struct file *fp, int how)
1051 {
1052         struct pipe *rpipe;
1053         struct pipe *wpipe;
1054         int error = EPIPE;
1055
1056         rpipe = (struct pipe *)fp->f_data;
1057         wpipe = rpipe->pipe_peer;
1058
1059         /*
1060          * We modify pipe_state on both pipes, which means we need
1061          * all four tokens!
1062          */
1063         lwkt_gettoken(&rpipe->pipe_rlock);
1064         lwkt_gettoken(&rpipe->pipe_wlock);
1065         lwkt_gettoken(&wpipe->pipe_rlock);
1066         lwkt_gettoken(&wpipe->pipe_wlock);
1067
1068         switch(how) {
1069         case SHUT_RDWR:
1070         case SHUT_RD:
1071                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1072                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1073                 if (rpipe->pipe_state & PIPE_WANTR) {
1074                         rpipe->pipe_state &= ~PIPE_WANTR;
1075                         wakeup(rpipe);
1076                 }
1077                 if (rpipe->pipe_state & PIPE_WANTW) {
1078                         rpipe->pipe_state &= ~PIPE_WANTW;
1079                         wakeup(rpipe);
1080                 }
1081                 error = 0;
1082                 if (how == SHUT_RD)
1083                         break;
1084                 /* fall through */
1085         case SHUT_WR:
1086                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1087                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1088                 if (wpipe->pipe_state & PIPE_WANTR) {
1089                         wpipe->pipe_state &= ~PIPE_WANTR;
1090                         wakeup(wpipe);
1091                 }
1092                 if (wpipe->pipe_state & PIPE_WANTW) {
1093                         wpipe->pipe_state &= ~PIPE_WANTW;
1094                         wakeup(wpipe);
1095                 }
1096                 error = 0;
1097                 break;
1098         }
1099         pipewakeup(rpipe, 1);
1100         pipewakeup(wpipe, 1);
1101
1102         lwkt_reltoken(&wpipe->pipe_wlock);
1103         lwkt_reltoken(&wpipe->pipe_rlock);
1104         lwkt_reltoken(&rpipe->pipe_wlock);
1105         lwkt_reltoken(&rpipe->pipe_rlock);
1106
1107         return (error);
1108 }
1109
1110 static void
1111 pipe_free_kmem(struct pipe *cpipe)
1112 {
1113         if (cpipe->pipe_buffer.buffer != NULL) {
1114                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1115                         atomic_subtract_int(&pipe_nbig, 1);
1116                 kmem_free(&kernel_map,
1117                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1118                         cpipe->pipe_buffer.size);
1119                 cpipe->pipe_buffer.buffer = NULL;
1120                 cpipe->pipe_buffer.object = NULL;
1121         }
1122 }
1123
1124 /*
1125  * Close the pipe.  The slock must be held to interlock against simultanious
1126  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1127  */
1128 static void
1129 pipeclose(struct pipe *cpipe)
1130 {
1131         globaldata_t gd;
1132         struct pipe *ppipe;
1133
1134         if (cpipe == NULL)
1135                 return;
1136
1137         /*
1138          * The slock may not have been allocated yet (close during
1139          * initialization)
1140          *
1141          * We need both the read and write tokens to modify pipe_state.
1142          */
1143         if (cpipe->pipe_slock)
1144                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1145         lwkt_gettoken(&cpipe->pipe_rlock);
1146         lwkt_gettoken(&cpipe->pipe_wlock);
1147
1148         /*
1149          * Set our state, wakeup anyone waiting in select/poll/kq, and
1150          * wakeup anyone blocked on our pipe.
1151          */
1152         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1153         pipewakeup(cpipe, 1);
1154         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1155                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1156                 wakeup(cpipe);
1157         }
1158
1159         /*
1160          * Disconnect from peer.
1161          */
1162         if ((ppipe = cpipe->pipe_peer) != NULL) {
1163                 lwkt_gettoken(&ppipe->pipe_rlock);
1164                 lwkt_gettoken(&ppipe->pipe_wlock);
1165                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1166                 pipewakeup(ppipe, 1);
1167                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1168                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1169                         wakeup(ppipe);
1170                 }
1171                 if (SLIST_FIRST(&ppipe->pipe_kq.ki_note))
1172                         KNOTE(&ppipe->pipe_kq.ki_note, 0);
1173                 lwkt_reltoken(&ppipe->pipe_wlock);
1174                 lwkt_reltoken(&ppipe->pipe_rlock);
1175         }
1176
1177         /*
1178          * If the peer is also closed we can free resources for both
1179          * sides, otherwise we leave our side intact to deal with any
1180          * races (since we only have the slock).
1181          */
1182         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1183                 cpipe->pipe_peer = NULL;
1184                 ppipe->pipe_peer = NULL;
1185                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1186                 pipeclose(ppipe);
1187                 ppipe = NULL;
1188         }
1189
1190         lwkt_reltoken(&cpipe->pipe_wlock);
1191         lwkt_reltoken(&cpipe->pipe_rlock);
1192         if (cpipe->pipe_slock)
1193                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1194
1195         /*
1196          * If we disassociated from our peer we can free resources
1197          */
1198         if (ppipe == NULL) {
1199                 gd = mycpu;
1200                 if (cpipe->pipe_slock) {
1201                         kfree(cpipe->pipe_slock, M_PIPE);
1202                         cpipe->pipe_slock = NULL;
1203                 }
1204                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1205                     cpipe->pipe_buffer.size != PIPE_SIZE
1206                 ) {
1207                         pipe_free_kmem(cpipe);
1208                         kfree(cpipe, M_PIPE);
1209                 } else {
1210                         cpipe->pipe_state = 0;
1211                         cpipe->pipe_peer = gd->gd_pipeq;
1212                         gd->gd_pipeq = cpipe;
1213                         ++gd->gd_pipeqcount;
1214                 }
1215         }
1216 }
1217
1218 static int
1219 pipe_kqfilter(struct file *fp, struct knote *kn)
1220 {
1221         struct pipe *cpipe;
1222
1223         cpipe = (struct pipe *)kn->kn_fp->f_data;
1224
1225         switch (kn->kn_filter) {
1226         case EVFILT_READ:
1227                 kn->kn_fop = &pipe_rfiltops;
1228                 break;
1229         case EVFILT_WRITE:
1230                 kn->kn_fop = &pipe_wfiltops;
1231                 if (cpipe->pipe_peer == NULL) {
1232                         /* other end of pipe has been closed */
1233                         return (EPIPE);
1234                 }
1235                 break;
1236         default:
1237                 return (EOPNOTSUPP);
1238         }
1239         kn->kn_hook = (caddr_t)cpipe;
1240
1241         knote_insert(&cpipe->pipe_kq.ki_note, kn);
1242
1243         return (0);
1244 }
1245
1246 static void
1247 filt_pipedetach(struct knote *kn)
1248 {
1249         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1250
1251         knote_remove(&cpipe->pipe_kq.ki_note, kn);
1252 }
1253
1254 /*ARGSUSED*/
1255 static int
1256 filt_piperead(struct knote *kn, long hint)
1257 {
1258         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1259         int ready = 0;
1260
1261         lwkt_gettoken(&rpipe->pipe_rlock);
1262         lwkt_gettoken(&rpipe->pipe_wlock);
1263
1264         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1265
1266         if (rpipe->pipe_state & PIPE_REOF) {
1267                 /*
1268                  * Only set NODATA if all data has been exhausted
1269                  */
1270                 if (kn->kn_data == 0)
1271                         kn->kn_flags |= EV_NODATA;
1272                 kn->kn_flags |= EV_EOF;
1273                 ready = 1;
1274         }
1275
1276         lwkt_reltoken(&rpipe->pipe_wlock);
1277         lwkt_reltoken(&rpipe->pipe_rlock);
1278
1279         if (!ready)
1280                 ready = kn->kn_data > 0;
1281
1282         return (ready);
1283 }
1284
1285 /*ARGSUSED*/
1286 static int
1287 filt_pipewrite(struct knote *kn, long hint)
1288 {
1289         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1290         struct pipe *wpipe = rpipe->pipe_peer;
1291         int ready = 0;
1292
1293         kn->kn_data = 0;
1294         if (wpipe == NULL) {
1295                 kn->kn_flags |= (EV_EOF | EV_NODATA);
1296                 return (1);
1297         }
1298
1299         lwkt_gettoken(&wpipe->pipe_rlock);
1300         lwkt_gettoken(&wpipe->pipe_wlock);
1301
1302         if (wpipe->pipe_state & PIPE_WEOF) {
1303                 kn->kn_flags |= (EV_EOF | EV_NODATA);
1304                 ready = 1;
1305         }
1306
1307         if (!ready)
1308                 kn->kn_data = wpipe->pipe_buffer.size -
1309                               (wpipe->pipe_buffer.windex -
1310                                wpipe->pipe_buffer.rindex);
1311
1312         lwkt_reltoken(&wpipe->pipe_wlock);
1313         lwkt_reltoken(&wpipe->pipe_rlock);
1314
1315         if (!ready)
1316                 ready = kn->kn_data >= PIPE_BUF;
1317
1318         return (ready);
1319 }