sys/kern/sys_pipe.c

   1 /*
   2  * Copyright (c) 1996 John S. Dyson
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice immediately at the beginning of the file, without modification,
  10  *    this list of conditions, and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  * 3. Absolutely no warranty of function or purpose is made by the author
  15  *    John S. Dyson.
  16  * 4. Modifications may be freely made to this file if the above conditions
  17  *    are met.
  18  *
  19  * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.60.2.13 2002/08/05 15:05:15 des Exp $
  20  * $DragonFly: src/sys/kern/sys_pipe.c,v 1.50 2008/09/09 04:06:13 dillon Exp $
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/kernel.h>
  32 #include <sys/proc.h>
  33 #include <sys/fcntl.h>
  34 #include <sys/file.h>
  35 #include <sys/filedesc.h>
  36 #include <sys/filio.h>
  37 #include <sys/ttycom.h>
  38 #include <sys/stat.h>
  39 #include <sys/poll.h>
  40 #include <sys/select.h>
  41 #include <sys/signalvar.h>
  42 #include <sys/sysproto.h>
  43 #include <sys/pipe.h>
  44 #include <sys/vnode.h>
  45 #include <sys/uio.h>
  46 #include <sys/event.h>
  47 #include <sys/globaldata.h>
  48 #include <sys/module.h>
  49 #include <sys/malloc.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/socket.h>
  52
  53 #include <vm/vm.h>
  54 #include <vm/vm_param.h>
  55 #include <sys/lock.h>
  56 #include <vm/vm_object.h>
  57 #include <vm/vm_kern.h>
  58 #include <vm/vm_extern.h>
  59 #include <vm/pmap.h>
  60 #include <vm/vm_map.h>
  61 #include <vm/vm_page.h>
  62 #include <vm/vm_zone.h>
  63
  64 #include <sys/file2.h>
  65 #include <sys/signal2.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_poll (struct file *fp, int events, struct ucred *cred);
  79 static int pipe_kqfilter (struct file *fp, struct knote *kn);
  80 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
  81 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data,
  82                 struct ucred *cred, struct sysmsg *msg);
  83
  84 static struct fileops pipeops = {
  85         .fo_read = pipe_read,
  86         .fo_write = pipe_write,
  87         .fo_ioctl = pipe_ioctl,
  88         .fo_poll = pipe_poll,
  89         .fo_kqfilter = pipe_kqfilter,
  90         .fo_stat = pipe_stat,
  91         .fo_close = pipe_close,
  92         .fo_shutdown = pipe_shutdown
  93 };
  94
  95 static void     filt_pipedetach(struct knote *kn);
  96 static int      filt_piperead(struct knote *kn, long hint);
  97 static int      filt_pipewrite(struct knote *kn, long hint);
  98
  99 static struct filterops pipe_rfiltops =
 100         { 1, NULL, filt_pipedetach, filt_piperead };
 101 static struct filterops pipe_wfiltops =
 102         { 1, NULL, filt_pipedetach, filt_pipewrite };
 103
 104 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 105
 106 /*
 107  * Default pipe buffer size(s), this can be kind-of large now because pipe
 108  * space is pageable.  The pipe code will try to maintain locality of
 109  * reference for performance reasons, so small amounts of outstanding I/O
 110  * will not wipe the cache.
 111  */
 112 #define MINPIPESIZE (PIPE_SIZE/3)
 113 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 114
 115 /*
 116  * Limit the number of "big" pipes
 117  */
 118 #define LIMITBIGPIPES   64
 119 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 120
 121 static int pipe_maxbig = LIMITBIGPIPES;
 122 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 123 static int pipe_bigcount;
 124 static int pipe_nbig;
 125 static int pipe_bcache_alloc;
 126 static int pipe_bkmem_alloc;
 127 static int pipe_rblocked_count;
 128 static int pipe_wblocked_count;
 129
 130 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 131 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 132         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 133 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 134         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 135 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked,
 136         CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
 137 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
 138         CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
 139 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 140         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 141 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 142         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 143 #ifdef SMP
 144 static int pipe_delay = 5000;   /* 5uS default */
 145 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
 146         CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
 147 static int pipe_mpsafe = 1;
 148 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
 149         CTLFLAG_RW, &pipe_mpsafe, 0, "");
 150 #endif
 151 #if !defined(NO_PIPE_SYSCTL_STATS)
 152 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 153         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 154 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 155         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 156 #endif
 157
 158 static void pipeclose (struct pipe *cpipe);
 159 static void pipe_free_kmem (struct pipe *cpipe);
 160 static int pipe_create (struct pipe **cpipep);
 161 static __inline void pipeselwakeup (struct pipe *cpipe);
 162 static int pipespace (struct pipe *cpipe, int size);
 163
 164 static __inline void
 165 pipeselwakeup(struct pipe *cpipe)
 166 {
 167         if (cpipe->pipe_state & PIPE_SEL) {
 168                 get_mplock();
 169                 cpipe->pipe_state &= ~PIPE_SEL;
 170                 selwakeup(&cpipe->pipe_sel);
 171                 rel_mplock();
 172         }
 173         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 174                 get_mplock();
 175                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 176                 rel_mplock();
 177         }
 178         if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
 179                 get_mplock();
 180                 KNOTE(&cpipe->pipe_sel.si_note, 0);
 181                 rel_mplock();
 182         }
 183 }
 184
 185 /*
 186  * These routines are called before and after a UIO.  The UIO
 187  * may block, causing our held tokens to be lost temporarily.
 188  *
 189  * We use these routines to serialize reads against other reads
 190  * and writes against other writes.
 191  *
 192  * The read token is held on entry so *ipp does not race.
 193  */
 194 static __inline int
 195 pipe_start_uio(struct pipe *cpipe, int *ipp)
 196 {
 197         int error;
 198
 199         while (*ipp) {
 200                 *ipp = -1;
 201                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 202                 if (error)
 203                         return (error);
 204         }
 205         *ipp = 1;
 206         return (0);
 207 }
 208
 209 static __inline void
 210 pipe_end_uio(struct pipe *cpipe, int *ipp)
 211 {
 212         if (*ipp < 0) {
 213                 *ipp = 0;
 214                 wakeup(ipp);
 215         } else {
 216                 KKASSERT(*ipp > 0);
 217                 *ipp = 0;
 218         }
 219 }
 220
 221 static __inline void
 222 pipe_get_mplock(int *save)
 223 {
 224 #ifdef SMP
 225         if (pipe_mpsafe == 0) {
 226                 get_mplock();
 227                 *save = 1;
 228         } else
 229 #endif
 230         {
 231                 *save = 0;
 232         }
 233 }
 234
 235 static __inline void
 236 pipe_rel_mplock(int *save)
 237 {
 238 #ifdef SMP
 239         if (*save)
 240                 rel_mplock();
 241 #endif
 242 }
 243
 244
 245 /*
 246  * The pipe system call for the DTYPE_PIPE type of pipes
 247  *
 248  * pipe_ARgs(int dummy)
 249  */
 250
 251 /* ARGSUSED */
 252 int
 253 sys_pipe(struct pipe_args *uap)
 254 {
 255         struct thread *td = curthread;
 256         struct proc *p = td->td_proc;
 257         struct file *rf, *wf;
 258         struct pipe *rpipe, *wpipe;
 259         int fd1, fd2, error;
 260
 261         KKASSERT(p);
 262
 263         rpipe = wpipe = NULL;
 264         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 265                 pipeclose(rpipe);
 266                 pipeclose(wpipe);
 267                 return (ENFILE);
 268         }
 269
 270         error = falloc(p, &rf, &fd1);
 271         if (error) {
 272                 pipeclose(rpipe);
 273                 pipeclose(wpipe);
 274                 return (error);
 275         }
 276         uap->sysmsg_fds[0] = fd1;
 277
 278         /*
 279          * Warning: once we've gotten past allocation of the fd for the
 280          * read-side, we can only drop the read side via fdrop() in order
 281          * to avoid races against processes which manage to dup() the read
 282          * side while we are blocked trying to allocate the write side.
 283          */
 284         rf->f_type = DTYPE_PIPE;
 285         rf->f_flag = FREAD | FWRITE;
 286         rf->f_ops = &pipeops;
 287         rf->f_data = rpipe;
 288         error = falloc(p, &wf, &fd2);
 289         if (error) {
 290                 fsetfd(p, NULL, fd1);
 291                 fdrop(rf);
 292                 /* rpipe has been closed by fdrop(). */
 293                 pipeclose(wpipe);
 294                 return (error);
 295         }
 296         wf->f_type = DTYPE_PIPE;
 297         wf->f_flag = FREAD | FWRITE;
 298         wf->f_ops = &pipeops;
 299         wf->f_data = wpipe;
 300         uap->sysmsg_fds[1] = fd2;
 301
 302         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 303                                     M_PIPE, M_WAITOK|M_ZERO);
 304         wpipe->pipe_slock = rpipe->pipe_slock;
 305         rpipe->pipe_peer = wpipe;
 306         wpipe->pipe_peer = rpipe;
 307         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 308
 309         /*
 310          * Once activated the peer relationship remains valid until
 311          * both sides are closed.
 312          */
 313         fsetfd(p, rf, fd1);
 314         fsetfd(p, wf, fd2);
 315         fdrop(rf);
 316         fdrop(wf);
 317
 318         return (0);
 319 }
 320
 321 /*
 322  * Allocate kva for pipe circular buffer, the space is pageable
 323  * This routine will 'realloc' the size of a pipe safely, if it fails
 324  * it will retain the old buffer.
 325  * If it fails it will return ENOMEM.
 326  */
 327 static int
 328 pipespace(struct pipe *cpipe, int size)
 329 {
 330         struct vm_object *object;
 331         caddr_t buffer;
 332         int npages, error;
 333
 334         npages = round_page(size) / PAGE_SIZE;
 335         object = cpipe->pipe_buffer.object;
 336
 337         /*
 338          * [re]create the object if necessary and reserve space for it
 339          * in the kernel_map.  The object and memory are pageable.  On
 340          * success, free the old resources before assigning the new
 341          * ones.
 342          */
 343         if (object == NULL || object->size != npages) {
 344                 get_mplock();
 345                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 346                 buffer = (caddr_t)vm_map_min(&kernel_map);
 347
 348                 error = vm_map_find(&kernel_map, object, 0,
 349                                     (vm_offset_t *)&buffer, size,
 350                                     1,
 351                                     VM_MAPTYPE_NORMAL,
 352                                     VM_PROT_ALL, VM_PROT_ALL,
 353                                     0);
 354
 355                 if (error != KERN_SUCCESS) {
 356                         vm_object_deallocate(object);
 357                         rel_mplock();
 358                         return (ENOMEM);
 359                 }
 360                 pipe_free_kmem(cpipe);
 361                 rel_mplock();
 362                 cpipe->pipe_buffer.object = object;
 363                 cpipe->pipe_buffer.buffer = buffer;
 364                 cpipe->pipe_buffer.size = size;
 365                 ++pipe_bkmem_alloc;
 366         } else {
 367                 ++pipe_bcache_alloc;
 368         }
 369         cpipe->pipe_buffer.rindex = 0;
 370         cpipe->pipe_buffer.windex = 0;
 371         return (0);
 372 }
 373
 374 /*
 375  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 376  * our per-cpu cache if possible.  For now make sure it is sized for the
 377  * smaller PIPE_SIZE default.
 378  */
 379 static int
 380 pipe_create(struct pipe **cpipep)
 381 {
 382         globaldata_t gd = mycpu;
 383         struct pipe *cpipe;
 384         int error;
 385
 386         if ((cpipe = gd->gd_pipeq) != NULL) {
 387                 gd->gd_pipeq = cpipe->pipe_peer;
 388                 --gd->gd_pipeqcount;
 389                 cpipe->pipe_peer = NULL;
 390                 cpipe->pipe_wantwcnt = 0;
 391         } else {
 392                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 393         }
 394         *cpipep = cpipe;
 395         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 396                 return (error);
 397         vfs_timestamp(&cpipe->pipe_ctime);
 398         cpipe->pipe_atime = cpipe->pipe_ctime;
 399         cpipe->pipe_mtime = cpipe->pipe_ctime;
 400         lwkt_token_init(&cpipe->pipe_rlock);
 401         lwkt_token_init(&cpipe->pipe_wlock);
 402         return (0);
 403 }
 404
 405 /*
 406  * MPALMOSTSAFE (acquires mplock)
 407  */
 408 static int
 409 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 410 {
 411         struct pipe *rpipe;
 412         int error;
 413         size_t nread = 0;
 414         int nbio;
 415         u_int size;     /* total bytes available */
 416         u_int nsize;    /* total bytes to read */
 417         u_int rindex;   /* contiguous bytes available */
 418         int notify_writer;
 419         lwkt_tokref rlock;
 420         lwkt_tokref wlock;
 421         int mpsave;
 422         int bigread;
 423         int bigcount;
 424
 425         if (uio->uio_resid == 0)
 426                 return(0);
 427
 428         /*
 429          * Setup locks, calculate nbio
 430          */
 431         pipe_get_mplock(&mpsave);
 432         rpipe = (struct pipe *)fp->f_data;
 433         lwkt_gettoken(&rlock, &rpipe->pipe_rlock);
 434
 435         if (fflags & O_FBLOCKING)
 436                 nbio = 0;
 437         else if (fflags & O_FNONBLOCKING)
 438                 nbio = 1;
 439         else if (fp->f_flag & O_NONBLOCK)
 440                 nbio = 1;
 441         else
 442                 nbio = 0;
 443
 444         /*
 445          * Reads are serialized.  Note howeverthat pipe_buffer.buffer and
 446          * pipe_buffer.size can change out from under us when the number
 447          * of bytes in the buffer are zero due to the write-side doing a
 448          * pipespace().
 449          */
 450         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 451         if (error) {
 452                 pipe_rel_mplock(&mpsave);
 453                 lwkt_reltoken(&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(&wlock, &rpipe->pipe_wlock);
 521                         if (rpipe->pipe_state & PIPE_WANTW) {
 522                                 notify_writer = 0;
 523                                 rpipe->pipe_state &= ~PIPE_WANTW;
 524                                 lwkt_reltoken(&wlock);
 525                                 wakeup(rpipe);
 526                         } else {
 527                                 lwkt_reltoken(&wlock);
 528                         }
 529                 }
 530
 531                 /*
 532                  * Pick up our copy loop again if the writer sent data to
 533                  * us while we were messing around.
 534                  *
 535                  * On a SMP box poll up to pipe_delay nanoseconds for new
 536                  * data.  Typically a value of 2000 to 4000 is sufficient
 537                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 538                  * is used for synchronous communications with small packets,
 539                  * and 8000 or so (8uS) will pipeline large buffer xfers
 540                  * between cpus over a pipe.
 541                  *
 542                  * For synchronous communications a hit means doing a
 543                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 544                  * where as miss requiring a tsleep/wakeup sequence
 545                  * will take 7uS or more.
 546                  */
 547                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 548                         continue;
 549
 550 #if defined(SMP) && defined(_RDTSC_SUPPORTED_)
 551                 if (pipe_delay) {
 552                         int64_t tsc_target;
 553                         int good = 0;
 554
 555                         tsc_target = tsc_get_target(pipe_delay);
 556                         while (tsc_test_target(tsc_target) == 0) {
 557                                 if (rpipe->pipe_buffer.windex !=
 558                                     rpipe->pipe_buffer.rindex) {
 559                                         good = 1;
 560                                         break;
 561                                 }
 562                         }
 563                         if (good)
 564                                 continue;
 565                 }
 566 #endif
 567
 568                 /*
 569                  * Detect EOF condition, do not set error.
 570                  */
 571                 if (rpipe->pipe_state & PIPE_REOF)
 572                         break;
 573
 574                 /*
 575                  * Break if some data was read, or if this was a non-blocking
 576                  * read.
 577                  */
 578                 if (nread > 0)
 579                         break;
 580
 581                 if (nbio) {
 582                         error = EAGAIN;
 583                         break;
 584                 }
 585
 586                 /*
 587                  * Last chance, interlock with WANTR.
 588                  */
 589                 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 590                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 591                 if (size) {
 592                         lwkt_reltoken(&wlock);
 593                         continue;
 594                 }
 595
 596                 /*
 597                  * If there is no more to read in the pipe, reset its
 598                  * pointers to the beginning.  This improves cache hit
 599                  * stats.
 600                  *
 601                  * We need both locks to modify both pointers, and there
 602                  * must also not be a write in progress or the uiomove()
 603                  * in the write might block and temporarily release
 604                  * its wlock, then reacquire and update windex.  We are
 605                  * only serialized against reads, not writes.
 606                  *
 607                  * XXX should we even bother resetting the indices?  It
 608                  *     might actually be more cache efficient not to.
 609                  */
 610                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 611                     rpipe->pipe_wip == 0) {
 612                         rpipe->pipe_buffer.rindex = 0;
 613                         rpipe->pipe_buffer.windex = 0;
 614                 }
 615
 616                 /*
 617                  * Wait for more data.
 618                  *
 619                  * Pipe_state can only be set if both the rlock and wlock
 620                  * are held.
 621                  */
 622                 rpipe->pipe_state |= PIPE_WANTR;
 623                 tsleep_interlock(rpipe, PCATCH);
 624                 lwkt_reltoken(&wlock);
 625                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 626                 ++pipe_rblocked_count;
 627                 if (error)
 628                         break;
 629         }
 630         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 631
 632         /*
 633          * Uptime last access time
 634          */
 635         if (error == 0 && nread)
 636                 vfs_timestamp(&rpipe->pipe_atime);
 637
 638         /*
 639          * If we drained the FIFO more then half way then handle
 640          * write blocking hysteresis.
 641          *
 642          * Note that PIPE_WANTW cannot be set by the writer without
 643          * it holding both rlock and wlock, so we can test it
 644          * while holding just rlock.
 645          */
 646         if (notify_writer) {
 647                 if (rpipe->pipe_state & PIPE_WANTW) {
 648                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 649                         if (rpipe->pipe_state & PIPE_WANTW) {
 650                                 rpipe->pipe_state &= ~PIPE_WANTW;
 651                                 lwkt_reltoken(&wlock);
 652                                 wakeup(rpipe);
 653                         } else {
 654                                 lwkt_reltoken(&wlock);
 655                         }
 656                 }
 657         }
 658         size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 659         lwkt_reltoken(&rlock);
 660
 661         /*
 662          * If enough space is available in buffer then wakeup sel writers?
 663          */
 664         if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF)
 665                 pipeselwakeup(rpipe);
 666         pipe_rel_mplock(&mpsave);
 667         return (error);
 668 }
 669
 670 /*
 671  * MPALMOSTSAFE - acquires mplock
 672  */
 673 static int
 674 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 675 {
 676         int error;
 677         int orig_resid;
 678         int nbio;
 679         struct pipe *wpipe, *rpipe;
 680         lwkt_tokref rlock;
 681         lwkt_tokref wlock;
 682         u_int windex;
 683         u_int space;
 684         u_int wcount;
 685         int mpsave;
 686         int bigwrite;
 687         int bigcount;
 688
 689         pipe_get_mplock(&mpsave);
 690
 691         /*
 692          * Writes go to the peer.  The peer will always exist.
 693          */
 694         rpipe = (struct pipe *) fp->f_data;
 695         wpipe = rpipe->pipe_peer;
 696         lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
 697         if (wpipe->pipe_state & PIPE_WEOF) {
 698                 pipe_rel_mplock(&mpsave);
 699                 lwkt_reltoken(&wlock);
 700                 return (EPIPE);
 701         }
 702
 703         /*
 704          * Degenerate case (EPIPE takes prec)
 705          */
 706         if (uio->uio_resid == 0) {
 707                 pipe_rel_mplock(&mpsave);
 708                 lwkt_reltoken(&wlock);
 709                 return(0);
 710         }
 711
 712         /*
 713          * Writes are serialized (start_uio must be called with wlock)
 714          */
 715         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 716         if (error) {
 717                 pipe_rel_mplock(&mpsave);
 718                 lwkt_reltoken(&wlock);
 719                 return (error);
 720         }
 721
 722         if (fflags & O_FBLOCKING)
 723                 nbio = 0;
 724         else if (fflags & O_FNONBLOCKING)
 725                 nbio = 1;
 726         else if (fp->f_flag & O_NONBLOCK)
 727                 nbio = 1;
 728         else
 729                 nbio = 0;
 730
 731         /*
 732          * If it is advantageous to resize the pipe buffer, do
 733          * so.  We are write-serialized so we can block safely.
 734          */
 735         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 736             (pipe_nbig < pipe_maxbig) &&
 737             wpipe->pipe_wantwcnt > 4 &&
 738             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 739                 /*
 740                  * Recheck after lock.
 741                  */
 742                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 743                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 744                     (pipe_nbig < pipe_maxbig) &&
 745                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 746                         atomic_add_int(&pipe_nbig, 1);
 747                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 748                                 ++pipe_bigcount;
 749                         else
 750                                 atomic_subtract_int(&pipe_nbig, 1);
 751                 }
 752                 lwkt_reltoken(&rlock);
 753         }
 754
 755         orig_resid = uio->uio_resid;
 756         wcount = 0;
 757
 758         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 759         bigcount = 10;
 760
 761         while (uio->uio_resid) {
 762                 if (wpipe->pipe_state & PIPE_WEOF) {
 763                         error = EPIPE;
 764                         break;
 765                 }
 766
 767                 /*
 768                  * Don't hog the cpu.
 769                  */
 770                 if (bigwrite && --bigcount == 0) {
 771                         lwkt_user_yield();
 772                         bigcount = 10;
 773                         if (CURSIG(curthread->td_lwp)) {
 774                                 error = EINTR;
 775                                 break;
 776                         }
 777                 }
 778
 779                 windex = wpipe->pipe_buffer.windex &
 780                          (wpipe->pipe_buffer.size - 1);
 781                 space = wpipe->pipe_buffer.size -
 782                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 783                 cpu_lfence();
 784
 785                 /* Writes of size <= PIPE_BUF must be atomic. */
 786                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 787                         space = 0;
 788
 789                 /*
 790                  * Write to fill, read size handles write hysteresis.  Also
 791                  * additional restrictions can cause select-based non-blocking
 792                  * writes to spin.
 793                  */
 794                 if (space > 0) {
 795                         u_int segsize;
 796
 797                         /*
 798                          * Transfer size is minimum of uio transfer
 799                          * and free space in pipe buffer.
 800                          *
 801                          * Limit each uiocopy to no more then PIPE_SIZE
 802                          * so we can keep the gravy train going on a
 803                          * SMP box.  This doubles the performance for
 804                          * write sizes > 16K.  Otherwise large writes
 805                          * wind up doing an inefficient synchronous
 806                          * ping-pong.
 807                          */
 808                         space = szmin(space, uio->uio_resid);
 809                         if (space > PIPE_SIZE)
 810                                 space = PIPE_SIZE;
 811
 812                         /*
 813                          * First segment to transfer is minimum of
 814                          * transfer size and contiguous space in
 815                          * pipe buffer.  If first segment to transfer
 816                          * is less than the transfer size, we've got
 817                          * a wraparound in the buffer.
 818                          */
 819                         segsize = wpipe->pipe_buffer.size - windex;
 820                         if (segsize > space)
 821                                 segsize = space;
 822
 823 #ifdef SMP
 824                         /*
 825                          * If this is the first loop and the reader is
 826                          * blocked, do a preemptive wakeup of the reader.
 827                          *
 828                          * On SMP the IPI latency plus the wlock interlock
 829                          * on the reader side is the fastest way to get the
 830                          * reader going.  (The scheduler will hard loop on
 831                          * lock tokens).
 832                          *
 833                          * NOTE: We can't clear WANTR here without acquiring
 834                          * the rlock, which we don't want to do here!
 835                          */
 836                         if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1)
 837                                 wakeup(wpipe);
 838 #endif
 839
 840                         /*
 841                          * Transfer segment, which may include a wrap-around.
 842                          * Update windex to account for both all in one go
 843                          * so the reader can read() the data atomically.
 844                          */
 845                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 846                                         segsize, uio);
 847                         if (error == 0 && segsize < space) {
 848                                 segsize = space - segsize;
 849                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 850                                                 segsize, uio);
 851                         }
 852                         if (error)
 853                                 break;
 854                         cpu_mfence();
 855                         wpipe->pipe_buffer.windex += space;
 856                         wcount += space;
 857                         continue;
 858                 }
 859
 860                 /*
 861                  * We need both the rlock and the wlock to interlock against
 862                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 863                  *
 864                  * These are token locks so we do not have to worry about
 865                  * deadlocks.
 866                  */
 867                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 868
 869                 /*
 870                  * If the "read-side" has been blocked, wake it up now
 871                  * and yield to let it drain synchronously rather
 872                  * then block.
 873                  */
 874                 if (wpipe->pipe_state & PIPE_WANTR) {
 875                         wpipe->pipe_state &= ~PIPE_WANTR;
 876                         wakeup(wpipe);
 877                 }
 878
 879                 /*
 880                  * don't block on non-blocking I/O
 881                  */
 882                 if (nbio) {
 883                         lwkt_reltoken(&rlock);
 884                         error = EAGAIN;
 885                         break;
 886                 }
 887
 888                 /*
 889                  * re-test whether we have to block in the writer after
 890                  * acquiring both locks, in case the reader opened up
 891                  * some space.
 892                  */
 893                 space = wpipe->pipe_buffer.size -
 894                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 895                 cpu_lfence();
 896                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 897                         space = 0;
 898
 899                 /*
 900                  * We have no more space and have something to offer,
 901                  * wake up select/poll.
 902                  */
 903                 if (space == 0) {
 904                         wpipe->pipe_state |= PIPE_WANTW;
 905                         ++wpipe->pipe_wantwcnt;
 906                         pipeselwakeup(wpipe);
 907                         if (wpipe->pipe_state & PIPE_WANTW)
 908                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 909                         ++pipe_wblocked_count;
 910                 }
 911                 lwkt_reltoken(&rlock);
 912
 913                 /*
 914                  * Break out if we errored or the read side wants us to go
 915                  * away.
 916                  */
 917                 if (error)
 918                         break;
 919                 if (wpipe->pipe_state & PIPE_WEOF) {
 920                         error = EPIPE;
 921                         break;
 922                 }
 923         }
 924         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 925
 926         /*
 927          * If we have put any characters in the buffer, we wake up
 928          * the reader.
 929          *
 930          * Both rlock and wlock are required to be able to modify pipe_state.
 931          */
 932         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 933                 if (wpipe->pipe_state & PIPE_WANTR) {
 934                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 935                         if (wpipe->pipe_state & PIPE_WANTR) {
 936                                 wpipe->pipe_state &= ~PIPE_WANTR;
 937                                 lwkt_reltoken(&rlock);
 938                                 wakeup(wpipe);
 939                         } else {
 940                                 lwkt_reltoken(&rlock);
 941                         }
 942                 }
 943         }
 944
 945         /*
 946          * Don't return EPIPE if I/O was successful
 947          */
 948         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 949             (uio->uio_resid == 0) &&
 950             (error == EPIPE)) {
 951                 error = 0;
 952         }
 953
 954         if (error == 0)
 955                 vfs_timestamp(&wpipe->pipe_mtime);
 956
 957         /*
 958          * We have something to offer,
 959          * wake up select/poll.
 960          */
 961         space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;
 962         lwkt_reltoken(&wlock);
 963         if (space)
 964                 pipeselwakeup(wpipe);
 965         pipe_rel_mplock(&mpsave);
 966         return (error);
 967 }
 968
 969 /*
 970  * MPALMOSTSAFE - acquires mplock
 971  *
 972  * we implement a very minimal set of ioctls for compatibility with sockets.
 973  */
 974 int
 975 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
 976            struct ucred *cred, struct sysmsg *msg)
 977 {
 978         struct pipe *mpipe;
 979         lwkt_tokref rlock;
 980         lwkt_tokref wlock;
 981         int error;
 982         int mpsave;
 983
 984         pipe_get_mplock(&mpsave);
 985         mpipe = (struct pipe *)fp->f_data;
 986
 987         lwkt_gettoken(&rlock, &mpipe->pipe_rlock);
 988         lwkt_gettoken(&wlock, &mpipe->pipe_wlock);
 989
 990         switch (cmd) {
 991         case FIOASYNC:
 992                 if (*(int *)data) {
 993                         mpipe->pipe_state |= PIPE_ASYNC;
 994                 } else {
 995                         mpipe->pipe_state &= ~PIPE_ASYNC;
 996                 }
 997                 error = 0;
 998                 break;
 999         case FIONREAD:
1000                 *(int *)data = mpipe->pipe_buffer.windex -
1001                                 mpipe->pipe_buffer.rindex;
1002                 error = 0;
1003                 break;
1004         case FIOSETOWN:
1005                 get_mplock();
1006                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1007                 rel_mplock();
1008                 break;
1009         case FIOGETOWN:
1010                 *(int *)data = fgetown(mpipe->pipe_sigio);
1011                 error = 0;
1012                 break;
1013         case TIOCSPGRP:
1014                 /* This is deprecated, FIOSETOWN should be used instead. */
1015                 get_mplock();
1016                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1017                 rel_mplock();
1018                 break;
1019
1020         case TIOCGPGRP:
1021                 /* This is deprecated, FIOGETOWN should be used instead. */
1022                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1023                 error = 0;
1024                 break;
1025         default:
1026                 error = ENOTTY;
1027                 break;
1028         }
1029         lwkt_reltoken(&rlock);
1030         lwkt_reltoken(&wlock);
1031         pipe_rel_mplock(&mpsave);
1032
1033         return (error);
1034 }
1035
1036 /*
1037  * MPALMOSTSAFE - acquires mplock
1038  */
1039 int
1040 pipe_poll(struct file *fp, int events, struct ucred *cred)
1041 {
1042         struct pipe *rpipe;
1043         struct pipe *wpipe;
1044         int revents = 0;
1045         u_int space;
1046         int mpsave;
1047
1048         pipe_get_mplock(&mpsave);
1049         rpipe = (struct pipe *)fp->f_data;
1050         wpipe = rpipe->pipe_peer;
1051         if (events & (POLLIN | POLLRDNORM)) {
1052                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
1053                     (rpipe->pipe_state & PIPE_REOF)) {
1054                         revents |= events & (POLLIN | POLLRDNORM);
1055                 }
1056         }
1057
1058         if (events & (POLLOUT | POLLWRNORM)) {
1059                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
1060                         revents |= events & (POLLOUT | POLLWRNORM);
1061                 } else {
1062                         space = wpipe->pipe_buffer.windex -
1063                                 wpipe->pipe_buffer.rindex;
1064                         space = wpipe->pipe_buffer.size - space;
1065                         if (space >= PIPE_BUF)
1066                                 revents |= events & (POLLOUT | POLLWRNORM);
1067                 }
1068         }
1069
1070         if ((rpipe->pipe_state & PIPE_REOF) ||
1071             (wpipe == NULL) ||
1072             (wpipe->pipe_state & PIPE_WEOF))
1073                 revents |= POLLHUP;
1074
1075         if (revents == 0) {
1076                 if (events & (POLLIN | POLLRDNORM)) {
1077                         selrecord(curthread, &rpipe->pipe_sel);
1078                         rpipe->pipe_state |= PIPE_SEL;
1079                 }
1080
1081                 if (events & (POLLOUT | POLLWRNORM)) {
1082                         selrecord(curthread, &wpipe->pipe_sel);
1083                         wpipe->pipe_state |= PIPE_SEL;
1084                 }
1085         }
1086         pipe_rel_mplock(&mpsave);
1087         return (revents);
1088 }
1089
1090 /*
1091  * MPSAFE
1092  */
1093 static int
1094 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1095 {
1096         struct pipe *pipe;
1097         int mpsave;
1098
1099         pipe_get_mplock(&mpsave);
1100         pipe = (struct pipe *)fp->f_data;
1101
1102         bzero((caddr_t)ub, sizeof(*ub));
1103         ub->st_mode = S_IFIFO;
1104         ub->st_blksize = pipe->pipe_buffer.size;
1105         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1106         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1107         ub->st_atimespec = pipe->pipe_atime;
1108         ub->st_mtimespec = pipe->pipe_mtime;
1109         ub->st_ctimespec = pipe->pipe_ctime;
1110         /*
1111          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1112          * st_flags, st_gen.
1113          * XXX (st_dev, st_ino) should be unique.
1114          */
1115         pipe_rel_mplock(&mpsave);
1116         return (0);
1117 }
1118
1119 /*
1120  * MPALMOSTSAFE - acquires mplock
1121  */
1122 static int
1123 pipe_close(struct file *fp)
1124 {
1125         struct pipe *cpipe;
1126
1127         get_mplock();
1128         cpipe = (struct pipe *)fp->f_data;
1129         fp->f_ops = &badfileops;
1130         fp->f_data = NULL;
1131         funsetown(cpipe->pipe_sigio);
1132         pipeclose(cpipe);
1133         rel_mplock();
1134         return (0);
1135 }
1136
1137 /*
1138  * Shutdown one or both directions of a full-duplex pipe.
1139  *
1140  * MPALMOSTSAFE - acquires mplock
1141  */
1142 static int
1143 pipe_shutdown(struct file *fp, int how)
1144 {
1145         struct pipe *rpipe;
1146         struct pipe *wpipe;
1147         int error = EPIPE;
1148         lwkt_tokref rpipe_rlock;
1149         lwkt_tokref rpipe_wlock;
1150         lwkt_tokref wpipe_rlock;
1151         lwkt_tokref wpipe_wlock;
1152         int mpsave;
1153
1154         pipe_get_mplock(&mpsave);
1155         rpipe = (struct pipe *)fp->f_data;
1156         wpipe = rpipe->pipe_peer;
1157
1158         /*
1159          * We modify pipe_state on both pipes, which means we need
1160          * all four tokens!
1161          */
1162         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1163         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1164         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1165         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1166
1167         switch(how) {
1168         case SHUT_RDWR:
1169         case SHUT_RD:
1170                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1171                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1172                 if (rpipe->pipe_state & PIPE_WANTR) {
1173                         rpipe->pipe_state &= ~PIPE_WANTR;
1174                         wakeup(rpipe);
1175                 }
1176                 if (rpipe->pipe_state & PIPE_WANTW) {
1177                         rpipe->pipe_state &= ~PIPE_WANTW;
1178                         wakeup(rpipe);
1179                 }
1180                 error = 0;
1181                 if (how == SHUT_RD)
1182                         break;
1183                 /* fall through */
1184         case SHUT_WR:
1185                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1186                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1187                 if (wpipe->pipe_state & PIPE_WANTR) {
1188                         wpipe->pipe_state &= ~PIPE_WANTR;
1189                         wakeup(wpipe);
1190                 }
1191                 if (wpipe->pipe_state & PIPE_WANTW) {
1192                         wpipe->pipe_state &= ~PIPE_WANTW;
1193                         wakeup(wpipe);
1194                 }
1195                 error = 0;
1196                 break;
1197         }
1198         pipeselwakeup(rpipe);
1199         pipeselwakeup(wpipe);
1200
1201         lwkt_reltoken(&rpipe_rlock);
1202         lwkt_reltoken(&rpipe_wlock);
1203         lwkt_reltoken(&wpipe_rlock);
1204         lwkt_reltoken(&wpipe_wlock);
1205
1206         pipe_rel_mplock(&mpsave);
1207         return (error);
1208 }
1209
1210 static void
1211 pipe_free_kmem(struct pipe *cpipe)
1212 {
1213         if (cpipe->pipe_buffer.buffer != NULL) {
1214                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1215                         atomic_subtract_int(&pipe_nbig, 1);
1216                 kmem_free(&kernel_map,
1217                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1218                         cpipe->pipe_buffer.size);
1219                 cpipe->pipe_buffer.buffer = NULL;
1220                 cpipe->pipe_buffer.object = NULL;
1221         }
1222 }
1223
1224 /*
1225  * Close the pipe.  The slock must be held to interlock against simultanious
1226  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1227  */
1228 static void
1229 pipeclose(struct pipe *cpipe)
1230 {
1231         globaldata_t gd;
1232         struct pipe *ppipe;
1233         lwkt_tokref cpipe_rlock;
1234         lwkt_tokref cpipe_wlock;
1235         lwkt_tokref ppipe_rlock;
1236         lwkt_tokref ppipe_wlock;
1237
1238         if (cpipe == NULL)
1239                 return;
1240
1241         /*
1242          * The slock may not have been allocated yet (close during
1243          * initialization)
1244          *
1245          * We need both the read and write tokens to modify pipe_state.
1246          */
1247         if (cpipe->pipe_slock)
1248                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1249         lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1250         lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1251
1252         /*
1253          * Set our state, wakeup anyone waiting in select, and
1254          * wakeup anyone blocked on our pipe.
1255          */
1256         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1257         pipeselwakeup(cpipe);
1258         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1259                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1260                 wakeup(cpipe);
1261         }
1262
1263         /*
1264          * Disconnect from peer.
1265          */
1266         if ((ppipe = cpipe->pipe_peer) != NULL) {
1267                 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1268                 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1269                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1270                 pipeselwakeup(ppipe);
1271                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1272                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1273                         wakeup(ppipe);
1274                 }
1275                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1276                         get_mplock();
1277                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1278                         rel_mplock();
1279                 }
1280                 lwkt_reltoken(&ppipe_rlock);
1281                 lwkt_reltoken(&ppipe_wlock);
1282         }
1283
1284         /*
1285          * If the peer is also closed we can free resources for both
1286          * sides, otherwise we leave our side intact to deal with any
1287          * races (since we only have the slock).
1288          */
1289         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1290                 cpipe->pipe_peer = NULL;
1291                 ppipe->pipe_peer = NULL;
1292                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1293                 pipeclose(ppipe);
1294                 ppipe = NULL;
1295         }
1296
1297         lwkt_reltoken(&cpipe_rlock);
1298         lwkt_reltoken(&cpipe_wlock);
1299         if (cpipe->pipe_slock)
1300                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1301
1302         /*
1303          * If we disassociated from our peer we can free resources
1304          */
1305         if (ppipe == NULL) {
1306                 gd = mycpu;
1307                 if (cpipe->pipe_slock) {
1308                         kfree(cpipe->pipe_slock, M_PIPE);
1309                         cpipe->pipe_slock = NULL;
1310                 }
1311                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1312                     cpipe->pipe_buffer.size != PIPE_SIZE
1313                 ) {
1314                         pipe_free_kmem(cpipe);
1315                         kfree(cpipe, M_PIPE);
1316                 } else {
1317                         cpipe->pipe_state = 0;
1318                         cpipe->pipe_peer = gd->gd_pipeq;
1319                         gd->gd_pipeq = cpipe;
1320                         ++gd->gd_pipeqcount;
1321                 }
1322         }
1323 }
1324
1325 /*
1326  * MPALMOSTSAFE - acquires mplock
1327  */
1328 static int
1329 pipe_kqfilter(struct file *fp, struct knote *kn)
1330 {
1331         struct pipe *cpipe;
1332
1333         get_mplock();
1334         cpipe = (struct pipe *)kn->kn_fp->f_data;
1335
1336         switch (kn->kn_filter) {
1337         case EVFILT_READ:
1338                 kn->kn_fop = &pipe_rfiltops;
1339                 break;
1340         case EVFILT_WRITE:
1341                 kn->kn_fop = &pipe_wfiltops;
1342                 cpipe = cpipe->pipe_peer;
1343                 if (cpipe == NULL) {
1344                         /* other end of pipe has been closed */
1345                         rel_mplock();
1346                         return (EPIPE);
1347                 }
1348                 break;
1349         default:
1350                 return (1);
1351         }
1352         kn->kn_hook = (caddr_t)cpipe;
1353
1354         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1355         rel_mplock();
1356         return (0);
1357 }
1358
1359 static void
1360 filt_pipedetach(struct knote *kn)
1361 {
1362         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1363
1364         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1365 }
1366
1367 /*ARGSUSED*/
1368 static int
1369 filt_piperead(struct knote *kn, long hint)
1370 {
1371         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1372
1373         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1374
1375         /* XXX RACE */
1376         if (rpipe->pipe_state & PIPE_REOF) {
1377                 kn->kn_flags |= EV_EOF;
1378                 return (1);
1379         }
1380         return (kn->kn_data > 0);
1381 }
1382
1383 /*ARGSUSED*/
1384 static int
1385 filt_pipewrite(struct knote *kn, long hint)
1386 {
1387         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1388         struct pipe *wpipe = rpipe->pipe_peer;
1389         u_int32_t space;
1390
1391         /* XXX RACE */
1392         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1393                 kn->kn_data = 0;
1394                 kn->kn_flags |= EV_EOF;
1395                 return (1);
1396         }
1397         space = wpipe->pipe_buffer.windex -
1398                 wpipe->pipe_buffer.rindex;
1399         space = wpipe->pipe_buffer.size - space;
1400         kn->kn_data = space;
1401         return (kn->kn_data >= PIPE_BUF);
1402 }