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 #include <sys/kern_syscall.h>
  50 #include <sys/lock.h>
  51 #include <sys/mutex.h>
  52
  53 #include <vm/vm.h>
  54 #include <vm/vm_param.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/mutex2.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         { FILTEROP_ISFD|FILTEROP_MPSAFE, NULL, filt_pipedetach, filt_piperead };
  99 static struct filterops pipe_wfiltops =
 100         { FILTEROP_ISFD|FILTEROP_MPSAFE, 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 static struct mtx *pipe_gdlocks;
 128
 129 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 130 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 131         CTLFLAG_RD, &pipe_nbig, 0, "number of big pipes allocated");
 132 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 133         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 134 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked,
 135         CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
 136 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
 137         CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
 138 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 139         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 140 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 141         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 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 #if !defined(NO_PIPE_SYSCTL_STATS)
 146 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 147         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 148 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 149         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 150 #endif
 151
 152 /*
 153  * Auto-size pipe cache to reduce kmem allocations and frees.
 154  */
 155 static
 156 void
 157 pipeinit(void *dummy)
 158 {
 159         size_t mbytes = kmem_lim_size();
 160         int n;
 161
 162         if (pipe_maxbig == LIMITBIGPIPES) {
 163                 if (mbytes >= 7 * 1024)
 164                         pipe_maxbig *= 2;
 165                 if (mbytes >= 15 * 1024)
 166                         pipe_maxbig *= 2;
 167         }
 168         if (pipe_maxcache == PIPEQ_MAX_CACHE) {
 169                 if (mbytes >= 7 * 1024)
 170                         pipe_maxcache *= 2;
 171                 if (mbytes >= 15 * 1024)
 172                         pipe_maxcache *= 2;
 173         }
 174         pipe_gdlocks = kmalloc(sizeof(*pipe_gdlocks) * ncpus,
 175                              M_PIPE, M_WAITOK | M_ZERO);
 176         for (n = 0; n < ncpus; ++n)
 177                 mtx_init(&pipe_gdlocks[n], "pipekm");
 178 }
 179 SYSINIT(kmem, SI_BOOT2_MACHDEP, SI_ORDER_ANY, pipeinit, NULL);
 180
 181 static void pipeclose (struct pipe *cpipe);
 182 static void pipe_free_kmem (struct pipe *cpipe);
 183 static int pipe_create (struct pipe **cpipep);
 184 static int pipespace (struct pipe *cpipe, int size);
 185
 186 static __inline void
 187 pipewakeup(struct pipe *cpipe, int dosigio)
 188 {
 189         if (dosigio && (cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 190                 lwkt_gettoken(&sigio_token);
 191                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 192                 lwkt_reltoken(&sigio_token);
 193         }
 194         KNOTE(&cpipe->pipe_kq.ki_note, 0);
 195 }
 196
 197 /*
 198  * These routines are called before and after a UIO.  The UIO
 199  * may block, causing our held tokens to be lost temporarily.
 200  *
 201  * We use these routines to serialize reads against other reads
 202  * and writes against other writes.
 203  *
 204  * The read token is held on entry so *ipp does not race.
 205  */
 206 static __inline int
 207 pipe_start_uio(struct pipe *cpipe, int *ipp)
 208 {
 209         int error;
 210
 211         while (*ipp) {
 212                 *ipp = -1;
 213                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 214                 if (error)
 215                         return (error);
 216         }
 217         *ipp = 1;
 218         return (0);
 219 }
 220
 221 static __inline void
 222 pipe_end_uio(struct pipe *cpipe, int *ipp)
 223 {
 224         if (*ipp < 0) {
 225                 *ipp = 0;
 226                 wakeup(ipp);
 227         } else {
 228                 KKASSERT(*ipp > 0);
 229                 *ipp = 0;
 230         }
 231 }
 232
 233 /*
 234  * The pipe system call for the DTYPE_PIPE type of pipes
 235  *
 236  * pipe_args(int dummy)
 237  *
 238  * MPSAFE
 239  */
 240 int
 241 sys_pipe(struct pipe_args *uap)
 242 {
 243         return kern_pipe(uap->sysmsg_fds, 0);
 244 }
 245
 246 int
 247 sys_pipe2(struct pipe2_args *uap)
 248 {
 249         return kern_pipe(uap->sysmsg_fds, uap->flags);
 250 }
 251
 252 int
 253 kern_pipe(long *fds, int flags)
 254 {
 255         struct thread *td = curthread;
 256         struct filedesc *fdp = td->td_proc->p_fd;
 257         struct file *rf, *wf;
 258         struct pipe *rpipe, *wpipe;
 259         int fd1, fd2, error;
 260
 261         rpipe = wpipe = NULL;
 262         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 263                 pipeclose(rpipe);
 264                 pipeclose(wpipe);
 265                 return (ENFILE);
 266         }
 267
 268         error = falloc(td->td_lwp, &rf, &fd1);
 269         if (error) {
 270                 pipeclose(rpipe);
 271                 pipeclose(wpipe);
 272                 return (error);
 273         }
 274         fds[0] = fd1;
 275
 276         /*
 277          * Warning: once we've gotten past allocation of the fd for the
 278          * read-side, we can only drop the read side via fdrop() in order
 279          * to avoid races against processes which manage to dup() the read
 280          * side while we are blocked trying to allocate the write side.
 281          */
 282         rf->f_type = DTYPE_PIPE;
 283         rf->f_flag = FREAD | FWRITE;
 284         rf->f_ops = &pipeops;
 285         rf->f_data = rpipe;
 286         if (flags & O_NONBLOCK)
 287                 rf->f_flag |= O_NONBLOCK;
 288         if (flags & O_CLOEXEC)
 289                 fdp->fd_files[fd1].fileflags |= UF_EXCLOSE;
 290
 291         error = falloc(td->td_lwp, &wf, &fd2);
 292         if (error) {
 293                 fsetfd(fdp, NULL, fd1);
 294                 fdrop(rf);
 295                 /* rpipe has been closed by fdrop(). */
 296                 pipeclose(wpipe);
 297                 return (error);
 298         }
 299         wf->f_type = DTYPE_PIPE;
 300         wf->f_flag = FREAD | FWRITE;
 301         wf->f_ops = &pipeops;
 302         wf->f_data = wpipe;
 303         if (flags & O_NONBLOCK)
 304                 wf->f_flag |= O_NONBLOCK;
 305         if (flags & O_CLOEXEC)
 306                 fdp->fd_files[fd2].fileflags |= UF_EXCLOSE;
 307
 308         fds[1] = fd2;
 309
 310         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 311                                     M_PIPE, M_WAITOK|M_ZERO);
 312         wpipe->pipe_slock = rpipe->pipe_slock;
 313         rpipe->pipe_peer = wpipe;
 314         wpipe->pipe_peer = rpipe;
 315         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 316
 317         /*
 318          * Once activated the peer relationship remains valid until
 319          * both sides are closed.
 320          */
 321         fsetfd(fdp, rf, fd1);
 322         fsetfd(fdp, wf, fd2);
 323         fdrop(rf);
 324         fdrop(wf);
 325
 326         return (0);
 327 }
 328
 329 /*
 330  * Allocate kva for pipe circular buffer, the space is pageable
 331  * This routine will 'realloc' the size of a pipe safely, if it fails
 332  * it will retain the old buffer.
 333  * If it fails it will return ENOMEM.
 334  */
 335 static int
 336 pipespace(struct pipe *cpipe, int size)
 337 {
 338         struct vm_object *object;
 339         caddr_t buffer;
 340         int npages, error;
 341
 342         npages = round_page(size) / PAGE_SIZE;
 343         object = cpipe->pipe_buffer.object;
 344
 345         /*
 346          * [re]create the object if necessary and reserve space for it
 347          * in the kernel_map.  The object and memory are pageable.  On
 348          * success, free the old resources before assigning the new
 349          * ones.
 350          */
 351         if (object == NULL || object->size != npages) {
 352                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 353                 buffer = (caddr_t)vm_map_min(&kernel_map);
 354
 355                 error = vm_map_find(&kernel_map, object, NULL,
 356                                     0, (vm_offset_t *)&buffer, size,
 357                                     PAGE_SIZE, TRUE,
 358                                     VM_MAPTYPE_NORMAL, VM_SUBSYS_PIPE,
 359                                     VM_PROT_ALL, VM_PROT_ALL, 0);
 360
 361                 if (error != KERN_SUCCESS) {
 362                         vm_object_deallocate(object);
 363                         return (ENOMEM);
 364                 }
 365                 pipe_free_kmem(cpipe);
 366                 cpipe->pipe_buffer.object = object;
 367                 cpipe->pipe_buffer.buffer = buffer;
 368                 cpipe->pipe_buffer.size = size;
 369                 ++pipe_bkmem_alloc;
 370         } else {
 371                 ++pipe_bcache_alloc;
 372         }
 373         cpipe->pipe_buffer.rindex = 0;
 374         cpipe->pipe_buffer.windex = 0;
 375         return (0);
 376 }
 377
 378 /*
 379  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 380  * our per-cpu cache if possible.  For now make sure it is sized for the
 381  * smaller PIPE_SIZE default.
 382  */
 383 static int
 384 pipe_create(struct pipe **cpipep)
 385 {
 386         globaldata_t gd = mycpu;
 387         struct pipe *cpipe;
 388         int error;
 389
 390         if ((cpipe = gd->gd_pipeq) != NULL) {
 391                 gd->gd_pipeq = cpipe->pipe_peer;
 392                 --gd->gd_pipeqcount;
 393                 cpipe->pipe_peer = NULL;
 394                 cpipe->pipe_wantwcnt = 0;
 395         } else {
 396                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 397         }
 398         *cpipep = cpipe;
 399         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 400                 return (error);
 401         vfs_timestamp(&cpipe->pipe_ctime);
 402         cpipe->pipe_atime = cpipe->pipe_ctime;
 403         cpipe->pipe_mtime = cpipe->pipe_ctime;
 404         lwkt_token_init(&cpipe->pipe_rlock, "piper");
 405         lwkt_token_init(&cpipe->pipe_wlock, "pipew");
 406         return (0);
 407 }
 408
 409 static int
 410 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 411 {
 412         struct pipe *rpipe;
 413         struct pipe *wpipe;
 414         int error;
 415         size_t nread = 0;
 416         int nbio;
 417         u_int size;     /* total bytes available */
 418         u_int nsize;    /* total bytes to read */
 419         u_int rindex;   /* contiguous bytes available */
 420         int notify_writer;
 421         int bigread;
 422         int bigcount;
 423
 424         atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
 425
 426         if (uio->uio_resid == 0)
 427                 return(0);
 428
 429         /*
 430          * Setup locks, calculate nbio
 431          */
 432         rpipe = (struct pipe *)fp->f_data;
 433         wpipe = rpipe->pipe_peer;
 434         lwkt_gettoken(&rpipe->pipe_rlock);
 435
 436         if (fflags & O_FBLOCKING)
 437                 nbio = 0;
 438         else if (fflags & O_FNONBLOCKING)
 439                 nbio = 1;
 440         else if (fp->f_flag & O_NONBLOCK)
 441                 nbio = 1;
 442         else
 443                 nbio = 0;
 444
 445         /*
 446          * Reads are serialized.  Note however that pipe_buffer.buffer and
 447          * pipe_buffer.size can change out from under us when the number
 448          * of bytes in the buffer are zero due to the write-side doing a
 449          * pipespace().
 450          */
 451         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 452         if (error) {
 453                 lwkt_reltoken(&rpipe->pipe_rlock);
 454                 return (error);
 455         }
 456         notify_writer = 0;
 457
 458         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 459         bigcount = 10;
 460
 461         while (uio->uio_resid) {
 462                 /*
 463                  * Don't hog the cpu.
 464                  */
 465                 if (bigread && --bigcount == 0) {
 466                         lwkt_user_yield();
 467                         bigcount = 10;
 468                         if (CURSIG(curthread->td_lwp)) {
 469                                 error = EINTR;
 470                                 break;
 471                         }
 472                 }
 473
 474                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 475                 cpu_lfence();
 476                 if (size) {
 477                         rindex = rpipe->pipe_buffer.rindex &
 478                                  (rpipe->pipe_buffer.size - 1);
 479                         nsize = size;
 480                         if (nsize > rpipe->pipe_buffer.size - rindex)
 481                                 nsize = rpipe->pipe_buffer.size - rindex;
 482                         nsize = szmin(nsize, uio->uio_resid);
 483
 484                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 485                                         nsize, uio);
 486                         if (error)
 487                                 break;
 488                         cpu_mfence();
 489                         rpipe->pipe_buffer.rindex += nsize;
 490                         nread += nsize;
 491
 492                         /*
 493                          * If the FIFO is still over half full just continue
 494                          * and do not try to notify the writer yet.
 495                          */
 496                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 497                                 notify_writer = 0;
 498                                 continue;
 499                         }
 500
 501                         /*
 502                          * When the FIFO is less then half full notify any
 503                          * waiting writer.  WANTW can be checked while
 504                          * holding just the rlock.
 505                          */
 506                         notify_writer = 1;
 507                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 508                                 continue;
 509                 }
 510
 511                 /*
 512                  * If the "write-side" was blocked we wake it up.  This code
 513                  * is reached either when the buffer is completely emptied
 514                  * or if it becomes more then half-empty.
 515                  *
 516                  * Pipe_state can only be modified if both the rlock and
 517                  * wlock are held.
 518                  */
 519                 if (rpipe->pipe_state & PIPE_WANTW) {
 520                         lwkt_gettoken(&rpipe->pipe_wlock);
 521                         if (rpipe->pipe_state & PIPE_WANTW) {
 522                                 rpipe->pipe_state &= ~PIPE_WANTW;
 523                                 lwkt_reltoken(&rpipe->pipe_wlock);
 524                                 wakeup(rpipe);
 525                         } else {
 526                                 lwkt_reltoken(&rpipe->pipe_wlock);
 527                         }
 528                 }
 529
 530                 /*
 531                  * Pick up our copy loop again if the writer sent data to
 532                  * us while we were messing around.
 533                  *
 534                  * On a SMP box poll up to pipe_delay nanoseconds for new
 535                  * data.  Typically a value of 2000 to 4000 is sufficient
 536                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 537                  * is used for synchronous communications with small packets,
 538                  * and 8000 or so (8uS) will pipeline large buffer xfers
 539                  * between cpus over a pipe.
 540                  *
 541                  * For synchronous communications a hit means doing a
 542                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 543                  * where as miss requiring a tsleep/wakeup sequence
 544                  * will take 7uS or more.
 545                  */
 546                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 547                         continue;
 548
 549 #ifdef _RDTSC_SUPPORTED_
 550                 if (pipe_delay) {
 551                         int64_t tsc_target;
 552                         int good = 0;
 553
 554                         tsc_target = tsc_get_target(pipe_delay);
 555                         while (tsc_test_target(tsc_target) == 0) {
 556                                 if (rpipe->pipe_buffer.windex !=
 557                                     rpipe->pipe_buffer.rindex) {
 558                                         good = 1;
 559                                         break;
 560                                 }
 561                         }
 562                         if (good)
 563                                 continue;
 564                 }
 565 #endif
 566
 567                 /*
 568                  * Detect EOF condition, do not set error.
 569                  */
 570                 if (rpipe->pipe_state & PIPE_REOF)
 571                         break;
 572
 573                 /*
 574                  * Break if some data was read, or if this was a non-blocking
 575                  * read.
 576                  */
 577                 if (nread > 0)
 578                         break;
 579
 580                 if (nbio) {
 581                         error = EAGAIN;
 582                         break;
 583                 }
 584
 585                 /*
 586                  * Last chance, interlock with WANTR.
 587                  */
 588                 lwkt_gettoken(&rpipe->pipe_wlock);
 589                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 590                 if (size) {
 591                         lwkt_reltoken(&rpipe->pipe_wlock);
 592                         continue;
 593                 }
 594
 595                 /*
 596                  * Retest EOF - acquiring a new token can temporarily release
 597                  * tokens already held.
 598                  */
 599                 if (rpipe->pipe_state & PIPE_REOF) {
 600                         lwkt_reltoken(&rpipe->pipe_wlock);
 601                         break;
 602                 }
 603
 604                 /*
 605                  * If there is no more to read in the pipe, reset its
 606                  * pointers to the beginning.  This improves cache hit
 607                  * stats.
 608                  *
 609                  * We need both locks to modify both pointers, and there
 610                  * must also not be a write in progress or the uiomove()
 611                  * in the write might block and temporarily release
 612                  * its wlock, then reacquire and update windex.  We are
 613                  * only serialized against reads, not writes.
 614                  *
 615                  * XXX should we even bother resetting the indices?  It
 616                  *     might actually be more cache efficient not to.
 617                  */
 618                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 619                     rpipe->pipe_wip == 0) {
 620                         rpipe->pipe_buffer.rindex = 0;
 621                         rpipe->pipe_buffer.windex = 0;
 622                 }
 623
 624                 /*
 625                  * Wait for more data.
 626                  *
 627                  * Pipe_state can only be set if both the rlock and wlock
 628                  * are held.
 629                  */
 630                 rpipe->pipe_state |= PIPE_WANTR;
 631                 tsleep_interlock(rpipe, PCATCH);
 632                 lwkt_reltoken(&rpipe->pipe_wlock);
 633                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 634                 ++pipe_rblocked_count;
 635                 if (error)
 636                         break;
 637         }
 638         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 639
 640         /*
 641          * Uptime last access time
 642          */
 643         if (error == 0 && nread)
 644                 vfs_timestamp(&rpipe->pipe_atime);
 645
 646         /*
 647          * If we drained the FIFO more then half way then handle
 648          * write blocking hysteresis.
 649          *
 650          * Note that PIPE_WANTW cannot be set by the writer without
 651          * it holding both rlock and wlock, so we can test it
 652          * while holding just rlock.
 653          */
 654         if (notify_writer) {
 655                 /*
 656                  * Synchronous blocking is done on the pipe involved
 657                  */
 658                 if (rpipe->pipe_state & PIPE_WANTW) {
 659                         lwkt_gettoken(&rpipe->pipe_wlock);
 660                         if (rpipe->pipe_state & PIPE_WANTW) {
 661                                 rpipe->pipe_state &= ~PIPE_WANTW;
 662                                 lwkt_reltoken(&rpipe->pipe_wlock);
 663                                 wakeup(rpipe);
 664                         } else {
 665                                 lwkt_reltoken(&rpipe->pipe_wlock);
 666                         }
 667                 }
 668
 669                 /*
 670                  * But we may also have to deal with a kqueue which is
 671                  * stored on the same pipe as its descriptor, so a
 672                  * EVFILT_WRITE event waiting for our side to drain will
 673                  * be on the other side.
 674                  */
 675                 lwkt_gettoken(&wpipe->pipe_wlock);
 676                 pipewakeup(wpipe, 0);
 677                 lwkt_reltoken(&wpipe->pipe_wlock);
 678         }
 679         /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
 680         lwkt_reltoken(&rpipe->pipe_rlock);
 681
 682         return (error);
 683 }
 684
 685 static int
 686 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 687 {
 688         int error;
 689         int orig_resid;
 690         int nbio;
 691         struct pipe *wpipe;
 692         struct pipe *rpipe;
 693         u_int windex;
 694         u_int space;
 695         u_int wcount;
 696         int bigwrite;
 697         int bigcount;
 698
 699         /*
 700          * Writes go to the peer.  The peer will always exist.
 701          */
 702         rpipe = (struct pipe *) fp->f_data;
 703         wpipe = rpipe->pipe_peer;
 704         lwkt_gettoken(&wpipe->pipe_wlock);
 705         if (wpipe->pipe_state & PIPE_WEOF) {
 706                 lwkt_reltoken(&wpipe->pipe_wlock);
 707                 return (EPIPE);
 708         }
 709
 710         /*
 711          * Degenerate case (EPIPE takes prec)
 712          */
 713         if (uio->uio_resid == 0) {
 714                 lwkt_reltoken(&wpipe->pipe_wlock);
 715                 return(0);
 716         }
 717
 718         /*
 719          * Writes are serialized (start_uio must be called with wlock)
 720          */
 721         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 722         if (error) {
 723                 lwkt_reltoken(&wpipe->pipe_wlock);
 724                 return (error);
 725         }
 726
 727         if (fflags & O_FBLOCKING)
 728                 nbio = 0;
 729         else if (fflags & O_FNONBLOCKING)
 730                 nbio = 1;
 731         else if (fp->f_flag & O_NONBLOCK)
 732                 nbio = 1;
 733         else
 734                 nbio = 0;
 735
 736         /*
 737          * If it is advantageous to resize the pipe buffer, do
 738          * so.  We are write-serialized so we can block safely.
 739          */
 740         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 741             (pipe_nbig < pipe_maxbig) &&
 742             wpipe->pipe_wantwcnt > 4 &&
 743             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 744                 /*
 745                  * Recheck after lock.
 746                  */
 747                 lwkt_gettoken(&wpipe->pipe_rlock);
 748                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 749                     (pipe_nbig < pipe_maxbig) &&
 750                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 751                         atomic_add_int(&pipe_nbig, 1);
 752                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 753                                 ++pipe_bigcount;
 754                         else
 755                                 atomic_subtract_int(&pipe_nbig, 1);
 756                 }
 757                 lwkt_reltoken(&wpipe->pipe_rlock);
 758         }
 759
 760         orig_resid = uio->uio_resid;
 761         wcount = 0;
 762
 763         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 764         bigcount = 10;
 765
 766         while (uio->uio_resid) {
 767                 if (wpipe->pipe_state & PIPE_WEOF) {
 768                         error = EPIPE;
 769                         break;
 770                 }
 771
 772                 /*
 773                  * Don't hog the cpu.
 774                  */
 775                 if (bigwrite && --bigcount == 0) {
 776                         lwkt_user_yield();
 777                         bigcount = 10;
 778                         if (CURSIG(curthread->td_lwp)) {
 779                                 error = EINTR;
 780                                 break;
 781                         }
 782                 }
 783
 784                 windex = wpipe->pipe_buffer.windex &
 785                          (wpipe->pipe_buffer.size - 1);
 786                 space = wpipe->pipe_buffer.size -
 787                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 788                 cpu_lfence();
 789
 790                 /* Writes of size <= PIPE_BUF must be atomic. */
 791                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 792                         space = 0;
 793
 794                 /*
 795                  * Write to fill, read size handles write hysteresis.  Also
 796                  * additional restrictions can cause select-based non-blocking
 797                  * writes to spin.
 798                  */
 799                 if (space > 0) {
 800                         u_int segsize;
 801
 802                         /*
 803                          * Transfer size is minimum of uio transfer
 804                          * and free space in pipe buffer.
 805                          *
 806                          * Limit each uiocopy to no more then PIPE_SIZE
 807                          * so we can keep the gravy train going on a
 808                          * SMP box.  This doubles the performance for
 809                          * write sizes > 16K.  Otherwise large writes
 810                          * wind up doing an inefficient synchronous
 811                          * ping-pong.
 812                          */
 813                         space = szmin(space, uio->uio_resid);
 814                         if (space > PIPE_SIZE)
 815                                 space = PIPE_SIZE;
 816
 817                         /*
 818                          * First segment to transfer is minimum of
 819                          * transfer size and contiguous space in
 820                          * pipe buffer.  If first segment to transfer
 821                          * is less than the transfer size, we've got
 822                          * a wraparound in the buffer.
 823                          */
 824                         segsize = wpipe->pipe_buffer.size - windex;
 825                         if (segsize > space)
 826                                 segsize = space;
 827
 828                         /*
 829                          * If this is the first loop and the reader is
 830                          * blocked, do a preemptive wakeup of the reader.
 831                          *
 832                          * On SMP the IPI latency plus the wlock interlock
 833                          * on the reader side is the fastest way to get the
 834                          * reader going.  (The scheduler will hard loop on
 835                          * lock tokens).
 836                          *
 837                          * NOTE: We can't clear WANTR here without acquiring
 838                          * the rlock, which we don't want to do here!
 839                          */
 840                         if ((wpipe->pipe_state & PIPE_WANTR))
 841                                 wakeup(wpipe);
 842
 843                         /*
 844                          * Transfer segment, which may include a wrap-around.
 845                          * Update windex to account for both all in one go
 846                          * so the reader can read() the data atomically.
 847                          */
 848                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 849                                         segsize, uio);
 850                         if (error == 0 && segsize < space) {
 851                                 segsize = space - segsize;
 852                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 853                                                 segsize, uio);
 854                         }
 855                         if (error)
 856                                 break;
 857                         cpu_mfence();
 858                         wpipe->pipe_buffer.windex += space;
 859                         wcount += space;
 860                         continue;
 861                 }
 862
 863                 /*
 864                  * We need both the rlock and the wlock to interlock against
 865                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 866                  *
 867                  * These are token locks so we do not have to worry about
 868                  * deadlocks.
 869                  */
 870                 lwkt_gettoken(&wpipe->pipe_rlock);
 871
 872                 /*
 873                  * If the "read-side" has been blocked, wake it up now
 874                  * and yield to let it drain synchronously rather
 875                  * then block.
 876                  */
 877                 if (wpipe->pipe_state & PIPE_WANTR) {
 878                         wpipe->pipe_state &= ~PIPE_WANTR;
 879                         wakeup(wpipe);
 880                 }
 881
 882                 /*
 883                  * don't block on non-blocking I/O
 884                  */
 885                 if (nbio) {
 886                         lwkt_reltoken(&wpipe->pipe_rlock);
 887                         error = EAGAIN;
 888                         break;
 889                 }
 890
 891                 /*
 892                  * re-test whether we have to block in the writer after
 893                  * acquiring both locks, in case the reader opened up
 894                  * some space.
 895                  */
 896                 space = wpipe->pipe_buffer.size -
 897                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 898                 cpu_lfence();
 899                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 900                         space = 0;
 901
 902                 /*
 903                  * Retest EOF - acquiring a new token can temporarily release
 904                  * tokens already held.
 905                  */
 906                 if (wpipe->pipe_state & PIPE_WEOF) {
 907                         lwkt_reltoken(&wpipe->pipe_rlock);
 908                         error = EPIPE;
 909                         break;
 910                 }
 911
 912                 /*
 913                  * We have no more space and have something to offer,
 914                  * wake up select/poll/kq.
 915                  */
 916                 if (space == 0) {
 917                         wpipe->pipe_state |= PIPE_WANTW;
 918                         ++wpipe->pipe_wantwcnt;
 919                         pipewakeup(wpipe, 1);
 920                         if (wpipe->pipe_state & PIPE_WANTW)
 921                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 922                         ++pipe_wblocked_count;
 923                 }
 924                 lwkt_reltoken(&wpipe->pipe_rlock);
 925
 926                 /*
 927                  * Break out if we errored or the read side wants us to go
 928                  * away.
 929                  */
 930                 if (error)
 931                         break;
 932                 if (wpipe->pipe_state & PIPE_WEOF) {
 933                         error = EPIPE;
 934                         break;
 935                 }
 936         }
 937         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 938
 939         /*
 940          * If we have put any characters in the buffer, we wake up
 941          * the reader.
 942          *
 943          * Both rlock and wlock are required to be able to modify pipe_state.
 944          */
 945         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 946                 if (wpipe->pipe_state & PIPE_WANTR) {
 947                         lwkt_gettoken(&wpipe->pipe_rlock);
 948                         if (wpipe->pipe_state & PIPE_WANTR) {
 949                                 wpipe->pipe_state &= ~PIPE_WANTR;
 950                                 lwkt_reltoken(&wpipe->pipe_rlock);
 951                                 wakeup(wpipe);
 952                         } else {
 953                                 lwkt_reltoken(&wpipe->pipe_rlock);
 954                         }
 955                 }
 956                 lwkt_gettoken(&wpipe->pipe_rlock);
 957                 pipewakeup(wpipe, 1);
 958                 lwkt_reltoken(&wpipe->pipe_rlock);
 959         }
 960
 961         /*
 962          * Don't return EPIPE if I/O was successful
 963          */
 964         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 965             (uio->uio_resid == 0) &&
 966             (error == EPIPE)) {
 967                 error = 0;
 968         }
 969
 970         if (error == 0)
 971                 vfs_timestamp(&wpipe->pipe_mtime);
 972
 973         /*
 974          * We have something to offer,
 975          * wake up select/poll/kq.
 976          */
 977         /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
 978         lwkt_reltoken(&wpipe->pipe_wlock);
 979         return (error);
 980 }
 981
 982 /*
 983  * we implement a very minimal set of ioctls for compatibility with sockets.
 984  */
 985 static int
 986 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
 987            struct ucred *cred, struct sysmsg *msg)
 988 {
 989         struct pipe *mpipe;
 990         int error;
 991
 992         mpipe = (struct pipe *)fp->f_data;
 993
 994         lwkt_gettoken(&mpipe->pipe_rlock);
 995         lwkt_gettoken(&mpipe->pipe_wlock);
 996
 997         switch (cmd) {
 998         case FIOASYNC:
 999                 if (*(int *)data) {
1000                         mpipe->pipe_state |= PIPE_ASYNC;
1001                 } else {
1002                         mpipe->pipe_state &= ~PIPE_ASYNC;
1003                 }
1004                 error = 0;
1005                 break;
1006         case FIONREAD:
1007                 *(int *)data = mpipe->pipe_buffer.windex -
1008                                 mpipe->pipe_buffer.rindex;
1009                 error = 0;
1010                 break;
1011         case FIOSETOWN:
1012                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1013                 break;
1014         case FIOGETOWN:
1015                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1016                 error = 0;
1017                 break;
1018         case TIOCSPGRP:
1019                 /* This is deprecated, FIOSETOWN should be used instead. */
1020                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1021                 break;
1022
1023         case TIOCGPGRP:
1024                 /* This is deprecated, FIOGETOWN should be used instead. */
1025                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1026                 error = 0;
1027                 break;
1028         default:
1029                 error = ENOTTY;
1030                 break;
1031         }
1032         lwkt_reltoken(&mpipe->pipe_wlock);
1033         lwkt_reltoken(&mpipe->pipe_rlock);
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
1046         pipe = (struct pipe *)fp->f_data;
1047
1048         bzero((caddr_t)ub, sizeof(*ub));
1049         ub->st_mode = S_IFIFO;
1050         ub->st_blksize = pipe->pipe_buffer.size;
1051         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1052         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1053         ub->st_atimespec = pipe->pipe_atime;
1054         ub->st_mtimespec = pipe->pipe_mtime;
1055         ub->st_ctimespec = pipe->pipe_ctime;
1056         /*
1057          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1058          * st_flags, st_gen.
1059          * XXX (st_dev, st_ino) should be unique.
1060          */
1061         return (0);
1062 }
1063
1064 static int
1065 pipe_close(struct file *fp)
1066 {
1067         struct pipe *cpipe;
1068
1069         cpipe = (struct pipe *)fp->f_data;
1070         fp->f_ops = &badfileops;
1071         fp->f_data = NULL;
1072         funsetown(&cpipe->pipe_sigio);
1073         pipeclose(cpipe);
1074         return (0);
1075 }
1076
1077 /*
1078  * Shutdown one or both directions of a full-duplex pipe.
1079  */
1080 static int
1081 pipe_shutdown(struct file *fp, int how)
1082 {
1083         struct pipe *rpipe;
1084         struct pipe *wpipe;
1085         int error = EPIPE;
1086
1087         rpipe = (struct pipe *)fp->f_data;
1088         wpipe = rpipe->pipe_peer;
1089
1090         /*
1091          * We modify pipe_state on both pipes, which means we need
1092          * all four tokens!
1093          */
1094         lwkt_gettoken(&rpipe->pipe_rlock);
1095         lwkt_gettoken(&rpipe->pipe_wlock);
1096         lwkt_gettoken(&wpipe->pipe_rlock);
1097         lwkt_gettoken(&wpipe->pipe_wlock);
1098
1099         switch(how) {
1100         case SHUT_RDWR:
1101         case SHUT_RD:
1102                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1103                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1104                 if (rpipe->pipe_state & PIPE_WANTR) {
1105                         rpipe->pipe_state &= ~PIPE_WANTR;
1106                         wakeup(rpipe);
1107                 }
1108                 if (rpipe->pipe_state & PIPE_WANTW) {
1109                         rpipe->pipe_state &= ~PIPE_WANTW;
1110                         wakeup(rpipe);
1111                 }
1112                 error = 0;
1113                 if (how == SHUT_RD)
1114                         break;
1115                 /* fall through */
1116         case SHUT_WR:
1117                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1118                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1119                 if (wpipe->pipe_state & PIPE_WANTR) {
1120                         wpipe->pipe_state &= ~PIPE_WANTR;
1121                         wakeup(wpipe);
1122                 }
1123                 if (wpipe->pipe_state & PIPE_WANTW) {
1124                         wpipe->pipe_state &= ~PIPE_WANTW;
1125                         wakeup(wpipe);
1126                 }
1127                 error = 0;
1128                 break;
1129         }
1130         pipewakeup(rpipe, 1);
1131         pipewakeup(wpipe, 1);
1132
1133         lwkt_reltoken(&wpipe->pipe_wlock);
1134         lwkt_reltoken(&wpipe->pipe_rlock);
1135         lwkt_reltoken(&rpipe->pipe_wlock);
1136         lwkt_reltoken(&rpipe->pipe_rlock);
1137
1138         return (error);
1139 }
1140
1141 /*
1142  * Destroy the pipe buffer.
1143  */
1144 static void
1145 pipe_free_kmem(struct pipe *cpipe)
1146 {
1147         if (cpipe->pipe_buffer.buffer != NULL) {
1148                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1149                         atomic_subtract_int(&pipe_nbig, 1);
1150                 kmem_free(&kernel_map,
1151                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1152                         cpipe->pipe_buffer.size);
1153                 cpipe->pipe_buffer.buffer = NULL;
1154                 cpipe->pipe_buffer.object = NULL;
1155         }
1156 }
1157
1158 /*
1159  * Close the pipe.  The slock must be held to interlock against simultanious
1160  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1161  */
1162 static void
1163 pipeclose(struct pipe *cpipe)
1164 {
1165         globaldata_t gd;
1166         struct pipe *ppipe;
1167
1168         if (cpipe == NULL)
1169                 return;
1170
1171         /*
1172          * The slock may not have been allocated yet (close during
1173          * initialization)
1174          *
1175          * We need both the read and write tokens to modify pipe_state.
1176          */
1177         if (cpipe->pipe_slock)
1178                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1179         lwkt_gettoken(&cpipe->pipe_rlock);
1180         lwkt_gettoken(&cpipe->pipe_wlock);
1181
1182         /*
1183          * Set our state, wakeup anyone waiting in select/poll/kq, and
1184          * wakeup anyone blocked on our pipe.
1185          */
1186         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1187         pipewakeup(cpipe, 1);
1188         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1189                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1190                 wakeup(cpipe);
1191         }
1192
1193         /*
1194          * Disconnect from peer.
1195          */
1196         if ((ppipe = cpipe->pipe_peer) != NULL) {
1197                 lwkt_gettoken(&ppipe->pipe_rlock);
1198                 lwkt_gettoken(&ppipe->pipe_wlock);
1199                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1200                 pipewakeup(ppipe, 1);
1201                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1202                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1203                         wakeup(ppipe);
1204                 }
1205                 if (SLIST_FIRST(&ppipe->pipe_kq.ki_note))
1206                         KNOTE(&ppipe->pipe_kq.ki_note, 0);
1207                 lwkt_reltoken(&ppipe->pipe_wlock);
1208                 lwkt_reltoken(&ppipe->pipe_rlock);
1209         }
1210
1211         /*
1212          * If the peer is also closed we can free resources for both
1213          * sides, otherwise we leave our side intact to deal with any
1214          * races (since we only have the slock).
1215          */
1216         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1217                 cpipe->pipe_peer = NULL;
1218                 ppipe->pipe_peer = NULL;
1219                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1220                 pipeclose(ppipe);
1221                 ppipe = NULL;
1222         }
1223
1224         lwkt_reltoken(&cpipe->pipe_wlock);
1225         lwkt_reltoken(&cpipe->pipe_rlock);
1226         if (cpipe->pipe_slock)
1227                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1228
1229         /*
1230          * If we disassociated from our peer we can free resources.  We
1231          * maintain a pcpu cache to improve performance, so the actual
1232          * tear-down case is limited to bulk situations.
1233          *
1234          * However, the bulk tear-down case can cause intense contention
1235          * on the kernel_map when, e.g. hundreds to hundreds of thousands
1236          * of processes are killed at the same time.  To deal with this we
1237          * use a pcpu mutex to maintain concurrency but also limit the
1238          * number of threads banging on the map and pmap.
1239          *
1240          * We use the mtx mechanism instead of the lockmgr mechanism because
1241          * the mtx mechanism utilizes a queued design which will not break
1242          * down in the face of thousands to hundreds of thousands of
1243          * processes trying to free pipes simultaneously.  The lockmgr
1244          * mechanism will wind up waking them all up each time a lock
1245          * cycles.
1246          */
1247         if (ppipe == NULL) {
1248                 gd = mycpu;
1249                 if (cpipe->pipe_slock) {
1250                         kfree(cpipe->pipe_slock, M_PIPE);
1251                         cpipe->pipe_slock = NULL;
1252                 }
1253                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1254                     cpipe->pipe_buffer.size != PIPE_SIZE
1255                 ) {
1256                         mtx_lock(&pipe_gdlocks[gd->gd_cpuid]);
1257                         pipe_free_kmem(cpipe);
1258                         mtx_unlock(&pipe_gdlocks[gd->gd_cpuid]);
1259                         kfree(cpipe, M_PIPE);
1260                 } else {
1261                         cpipe->pipe_state = 0;
1262                         cpipe->pipe_peer = gd->gd_pipeq;
1263                         gd->gd_pipeq = cpipe;
1264                         ++gd->gd_pipeqcount;
1265                 }
1266         }
1267 }
1268
1269 static int
1270 pipe_kqfilter(struct file *fp, struct knote *kn)
1271 {
1272         struct pipe *cpipe;
1273
1274         cpipe = (struct pipe *)kn->kn_fp->f_data;
1275
1276         switch (kn->kn_filter) {
1277         case EVFILT_READ:
1278                 kn->kn_fop = &pipe_rfiltops;
1279                 break;
1280         case EVFILT_WRITE:
1281                 kn->kn_fop = &pipe_wfiltops;
1282                 if (cpipe->pipe_peer == NULL) {
1283                         /* other end of pipe has been closed */
1284                         return (EPIPE);
1285                 }
1286                 break;
1287         default:
1288                 return (EOPNOTSUPP);
1289         }
1290         kn->kn_hook = (caddr_t)cpipe;
1291
1292         knote_insert(&cpipe->pipe_kq.ki_note, kn);
1293
1294         return (0);
1295 }
1296
1297 static void
1298 filt_pipedetach(struct knote *kn)
1299 {
1300         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1301
1302         knote_remove(&cpipe->pipe_kq.ki_note, kn);
1303 }
1304
1305 /*ARGSUSED*/
1306 static int
1307 filt_piperead(struct knote *kn, long hint)
1308 {
1309         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1310         int ready = 0;
1311
1312         lwkt_gettoken(&rpipe->pipe_rlock);
1313         lwkt_gettoken(&rpipe->pipe_wlock);
1314
1315         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1316
1317         if (rpipe->pipe_state & PIPE_REOF) {
1318                 /*
1319                  * Only set NODATA if all data has been exhausted
1320                  */
1321                 if (kn->kn_data == 0)
1322                         kn->kn_flags |= EV_NODATA;
1323                 kn->kn_flags |= EV_EOF;
1324                 ready = 1;
1325         }
1326
1327         lwkt_reltoken(&rpipe->pipe_wlock);
1328         lwkt_reltoken(&rpipe->pipe_rlock);
1329
1330         if (!ready)
1331                 ready = kn->kn_data > 0;
1332
1333         return (ready);
1334 }
1335
1336 /*ARGSUSED*/
1337 static int
1338 filt_pipewrite(struct knote *kn, long hint)
1339 {
1340         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1341         struct pipe *wpipe = rpipe->pipe_peer;
1342         int ready = 0;
1343
1344         kn->kn_data = 0;
1345         if (wpipe == NULL) {
1346                 kn->kn_flags |= (EV_EOF | EV_NODATA);
1347                 return (1);
1348         }
1349
1350         lwkt_gettoken(&wpipe->pipe_rlock);
1351         lwkt_gettoken(&wpipe->pipe_wlock);
1352
1353         if (wpipe->pipe_state & PIPE_WEOF) {
1354                 kn->kn_flags |= (EV_EOF | EV_NODATA);
1355                 ready = 1;
1356         }
1357
1358         if (!ready)
1359                 kn->kn_data = wpipe->pipe_buffer.size -
1360                               (wpipe->pipe_buffer.windex -
1361                                wpipe->pipe_buffer.rindex);
1362
1363         lwkt_reltoken(&wpipe->pipe_wlock);
1364         lwkt_reltoken(&wpipe->pipe_rlock);
1365
1366         if (!ready)
1367                 ready = kn->kn_data >= PIPE_BUF;
1368
1369         return (ready);
1370 }