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, struct ucred *cred);
  82
  83 static struct fileops pipeops = {
  84         .fo_read = pipe_read,
  85         .fo_write = pipe_write,
  86         .fo_ioctl = pipe_ioctl,
  87         .fo_poll = pipe_poll,
  88         .fo_kqfilter = pipe_kqfilter,
  89         .fo_stat = pipe_stat,
  90         .fo_close = pipe_close,
  91         .fo_shutdown = pipe_shutdown
  92 };
  93
  94 static void     filt_pipedetach(struct knote *kn);
  95 static int      filt_piperead(struct knote *kn, long hint);
  96 static int      filt_pipewrite(struct knote *kn, long hint);
  97
  98 static struct filterops pipe_rfiltops =
  99         { 1, NULL, filt_pipedetach, filt_piperead };
 100 static struct filterops pipe_wfiltops =
 101         { 1, NULL, filt_pipedetach, filt_pipewrite };
 102
 103 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 104
 105 /*
 106  * Default pipe buffer size(s), this can be kind-of large now because pipe
 107  * space is pageable.  The pipe code will try to maintain locality of
 108  * reference for performance reasons, so small amounts of outstanding I/O
 109  * will not wipe the cache.
 110  */
 111 #define MINPIPESIZE (PIPE_SIZE/3)
 112 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 113
 114 /*
 115  * Limit the number of "big" pipes
 116  */
 117 #define LIMITBIGPIPES   64
 118 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 119
 120 static int pipe_maxbig = LIMITBIGPIPES;
 121 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 122 static int pipe_bigcount;
 123 static int pipe_nbig;
 124 static int pipe_bcache_alloc;
 125 static int pipe_bkmem_alloc;
 126 static int pipe_rblocked_count;
 127 static int pipe_wblocked_count;
 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, "numer 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 #ifdef SMP
 143 static int pipe_delay = 5000;   /* 5uS default */
 144 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
 145         CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
 146 static int pipe_mpsafe = 1;
 147 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
 148         CTLFLAG_RW, &pipe_mpsafe, 0, "");
 149 #endif
 150 #if !defined(NO_PIPE_SYSCTL_STATS)
 151 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 152         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 153 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 154         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 155 #endif
 156
 157 static void pipeclose (struct pipe *cpipe);
 158 static void pipe_free_kmem (struct pipe *cpipe);
 159 static int pipe_create (struct pipe **cpipep);
 160 static __inline void pipeselwakeup (struct pipe *cpipe);
 161 static int pipespace (struct pipe *cpipe, int size);
 162
 163 static __inline void
 164 pipeselwakeup(struct pipe *cpipe)
 165 {
 166         if (cpipe->pipe_state & PIPE_SEL) {
 167                 get_mplock();
 168                 cpipe->pipe_state &= ~PIPE_SEL;
 169                 selwakeup(&cpipe->pipe_sel);
 170                 rel_mplock();
 171         }
 172         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 173                 get_mplock();
 174                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 175                 rel_mplock();
 176         }
 177         if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
 178                 get_mplock();
 179                 KNOTE(&cpipe->pipe_sel.si_note, 0);
 180                 rel_mplock();
 181         }
 182 }
 183
 184 /*
 185  * These routines are called before and after a UIO.  The UIO
 186  * may block, causing our held tokens to be lost temporarily.
 187  *
 188  * We use these routines to serialize reads against other reads
 189  * and writes against other writes.
 190  *
 191  * The read token is held on entry so *ipp does not race.
 192  */
 193 static __inline int
 194 pipe_start_uio(struct pipe *cpipe, int *ipp)
 195 {
 196         int error;
 197
 198         while (*ipp) {
 199                 *ipp = -1;
 200                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 201                 if (error)
 202                         return (error);
 203         }
 204         *ipp = 1;
 205         return (0);
 206 }
 207
 208 static __inline void
 209 pipe_end_uio(struct pipe *cpipe, int *ipp)
 210 {
 211         if (*ipp < 0) {
 212                 *ipp = 0;
 213                 wakeup(ipp);
 214         } else {
 215                 KKASSERT(*ipp > 0);
 216                 *ipp = 0;
 217         }
 218 }
 219
 220 static __inline void
 221 pipe_get_mplock(int *save)
 222 {
 223 #ifdef SMP
 224         if (pipe_mpsafe == 0) {
 225                 get_mplock();
 226                 *save = 1;
 227         } else
 228 #endif
 229         {
 230                 *save = 0;
 231         }
 232 }
 233
 234 static __inline void
 235 pipe_rel_mplock(int *save)
 236 {
 237 #ifdef SMP
 238         if (*save)
 239                 rel_mplock();
 240 #endif
 241 }
 242
 243
 244 /*
 245  * The pipe system call for the DTYPE_PIPE type of pipes
 246  *
 247  * pipe_ARgs(int dummy)
 248  */
 249
 250 /* ARGSUSED */
 251 int
 252 sys_pipe(struct pipe_args *uap)
 253 {
 254         struct thread *td = curthread;
 255         struct proc *p = td->td_proc;
 256         struct file *rf, *wf;
 257         struct pipe *rpipe, *wpipe;
 258         int fd1, fd2, error;
 259
 260         KKASSERT(p);
 261
 262         rpipe = wpipe = NULL;
 263         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 264                 pipeclose(rpipe);
 265                 pipeclose(wpipe);
 266                 return (ENFILE);
 267         }
 268
 269         error = falloc(p, &rf, &fd1);
 270         if (error) {
 271                 pipeclose(rpipe);
 272                 pipeclose(wpipe);
 273                 return (error);
 274         }
 275         uap->sysmsg_fds[0] = fd1;
 276
 277         /*
 278          * Warning: once we've gotten past allocation of the fd for the
 279          * read-side, we can only drop the read side via fdrop() in order
 280          * to avoid races against processes which manage to dup() the read
 281          * side while we are blocked trying to allocate the write side.
 282          */
 283         rf->f_type = DTYPE_PIPE;
 284         rf->f_flag = FREAD | FWRITE;
 285         rf->f_ops = &pipeops;
 286         rf->f_data = rpipe;
 287         error = falloc(p, &wf, &fd2);
 288         if (error) {
 289                 fsetfd(p, NULL, fd1);
 290                 fdrop(rf);
 291                 /* rpipe has been closed by fdrop(). */
 292                 pipeclose(wpipe);
 293                 return (error);
 294         }
 295         wf->f_type = DTYPE_PIPE;
 296         wf->f_flag = FREAD | FWRITE;
 297         wf->f_ops = &pipeops;
 298         wf->f_data = wpipe;
 299         uap->sysmsg_fds[1] = fd2;
 300
 301         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 302                                     M_PIPE, M_WAITOK|M_ZERO);
 303         wpipe->pipe_slock = rpipe->pipe_slock;
 304         rpipe->pipe_peer = wpipe;
 305         wpipe->pipe_peer = rpipe;
 306         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 307
 308         /*
 309          * Once activated the peer relationship remains valid until
 310          * both sides are closed.
 311          */
 312         fsetfd(p, rf, fd1);
 313         fsetfd(p, wf, fd2);
 314         fdrop(rf);
 315         fdrop(wf);
 316
 317         return (0);
 318 }
 319
 320 /*
 321  * Allocate kva for pipe circular buffer, the space is pageable
 322  * This routine will 'realloc' the size of a pipe safely, if it fails
 323  * it will retain the old buffer.
 324  * If it fails it will return ENOMEM.
 325  */
 326 static int
 327 pipespace(struct pipe *cpipe, int size)
 328 {
 329         struct vm_object *object;
 330         caddr_t buffer;
 331         int npages, error;
 332
 333         npages = round_page(size) / PAGE_SIZE;
 334         object = cpipe->pipe_buffer.object;
 335
 336         /*
 337          * [re]create the object if necessary and reserve space for it
 338          * in the kernel_map.  The object and memory are pageable.  On
 339          * success, free the old resources before assigning the new
 340          * ones.
 341          */
 342         if (object == NULL || object->size != npages) {
 343                 get_mplock();
 344                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 345                 buffer = (caddr_t)vm_map_min(&kernel_map);
 346
 347                 error = vm_map_find(&kernel_map, object, 0,
 348                                     (vm_offset_t *)&buffer, size,
 349                                     1,
 350                                     VM_MAPTYPE_NORMAL,
 351                                     VM_PROT_ALL, VM_PROT_ALL,
 352                                     0);
 353
 354                 if (error != KERN_SUCCESS) {
 355                         vm_object_deallocate(object);
 356                         rel_mplock();
 357                         return (ENOMEM);
 358                 }
 359                 pipe_free_kmem(cpipe);
 360                 rel_mplock();
 361                 cpipe->pipe_buffer.object = object;
 362                 cpipe->pipe_buffer.buffer = buffer;
 363                 cpipe->pipe_buffer.size = size;
 364                 ++pipe_bkmem_alloc;
 365         } else {
 366                 ++pipe_bcache_alloc;
 367         }
 368         cpipe->pipe_buffer.rindex = 0;
 369         cpipe->pipe_buffer.windex = 0;
 370         return (0);
 371 }
 372
 373 /*
 374  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 375  * our per-cpu cache if possible.  For now make sure it is sized for the
 376  * smaller PIPE_SIZE default.
 377  */
 378 static int
 379 pipe_create(struct pipe **cpipep)
 380 {
 381         globaldata_t gd = mycpu;
 382         struct pipe *cpipe;
 383         int error;
 384
 385         if ((cpipe = gd->gd_pipeq) != NULL) {
 386                 gd->gd_pipeq = cpipe->pipe_peer;
 387                 --gd->gd_pipeqcount;
 388                 cpipe->pipe_peer = NULL;
 389                 cpipe->pipe_wantwcnt = 0;
 390         } else {
 391                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 392         }
 393         *cpipep = cpipe;
 394         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 395                 return (error);
 396         vfs_timestamp(&cpipe->pipe_ctime);
 397         cpipe->pipe_atime = cpipe->pipe_ctime;
 398         cpipe->pipe_mtime = cpipe->pipe_ctime;
 399         lwkt_token_init(&cpipe->pipe_rlock);
 400         lwkt_token_init(&cpipe->pipe_wlock);
 401         return (0);
 402 }
 403
 404 /*
 405  * MPALMOSTSAFE (acquires mplock)
 406  */
 407 static int
 408 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 409 {
 410         struct pipe *rpipe;
 411         int error;
 412         size_t nread = 0;
 413         int nbio;
 414         u_int size;     /* total bytes available */
 415         u_int nsize;    /* total bytes to read */
 416         u_int rindex;   /* contiguous bytes available */
 417         int notify_writer;
 418         lwkt_tokref rlock;
 419         lwkt_tokref wlock;
 420         int mpsave;
 421         int bigread;
 422         int bigcount;
 423
 424         if (uio->uio_resid == 0)
 425                 return(0);
 426
 427         /*
 428          * Setup locks, calculate nbio
 429          */
 430         pipe_get_mplock(&mpsave);
 431         rpipe = (struct pipe *)fp->f_data;
 432         lwkt_gettoken(&rlock, &rpipe->pipe_rlock);
 433
 434         if (fflags & O_FBLOCKING)
 435                 nbio = 0;
 436         else if (fflags & O_FNONBLOCKING)
 437                 nbio = 1;
 438         else if (fp->f_flag & O_NONBLOCK)
 439                 nbio = 1;
 440         else
 441                 nbio = 0;
 442
 443         /*
 444          * Reads are serialized.  Note howeverthat pipe_buffer.buffer and
 445          * pipe_buffer.size can change out from under us when the number
 446          * of bytes in the buffer are zero due to the write-side doing a
 447          * pipespace().
 448          */
 449         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 450         if (error) {
 451                 pipe_rel_mplock(&mpsave);
 452                 lwkt_reltoken(&rlock);
 453                 return (error);
 454         }
 455         notify_writer = 0;
 456
 457         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 458         bigcount = 10;
 459
 460         while (uio->uio_resid) {
 461                 /*
 462                  * Don't hog the cpu.
 463                  */
 464                 if (bigread && --bigcount == 0) {
 465                         lwkt_user_yield();
 466                         bigcount = 10;
 467                         if (CURSIG(curthread->td_lwp)) {
 468                                 error = EINTR;
 469                                 break;
 470                         }
 471                 }
 472
 473                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 474                 cpu_lfence();
 475                 if (size) {
 476                         rindex = rpipe->pipe_buffer.rindex &
 477                                  (rpipe->pipe_buffer.size - 1);
 478                         nsize = size;
 479                         if (nsize > rpipe->pipe_buffer.size - rindex)
 480                                 nsize = rpipe->pipe_buffer.size - rindex;
 481                         nsize = szmin(nsize, uio->uio_resid);
 482
 483                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 484                                         nsize, uio);
 485                         if (error)
 486                                 break;
 487                         cpu_mfence();
 488                         rpipe->pipe_buffer.rindex += nsize;
 489                         nread += nsize;
 490
 491                         /*
 492                          * If the FIFO is still over half full just continue
 493                          * and do not try to notify the writer yet.
 494                          */
 495                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 496                                 notify_writer = 0;
 497                                 continue;
 498                         }
 499
 500                         /*
 501                          * When the FIFO is less then half full notify any
 502                          * waiting writer.  WANTW can be checked while
 503                          * holding just the rlock.
 504                          */
 505                         notify_writer = 1;
 506                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 507                                 continue;
 508                 }
 509
 510                 /*
 511                  * If the "write-side" was blocked we wake it up.  This code
 512                  * is reached either when the buffer is completely emptied
 513                  * or if it becomes more then half-empty.
 514                  *
 515                  * Pipe_state can only be modified if both the rlock and
 516                  * wlock are held.
 517                  */
 518                 if (rpipe->pipe_state & PIPE_WANTW) {
 519                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 520                         if (rpipe->pipe_state & PIPE_WANTW) {
 521                                 notify_writer = 0;
 522                                 rpipe->pipe_state &= ~PIPE_WANTW;
 523                                 lwkt_reltoken(&wlock);
 524                                 wakeup(rpipe);
 525                         } else {
 526                                 lwkt_reltoken(&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 #if defined(SMP) && defined(_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(&wlock, &rpipe->pipe_wlock);
 589                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 590                 if (size) {
 591                         lwkt_reltoken(&wlock);
 592                         continue;
 593                 }
 594
 595                 /*
 596                  * If there is no more to read in the pipe, reset its
 597                  * pointers to the beginning.  This improves cache hit
 598                  * stats.
 599                  *
 600                  * We need both locks to modify both pointers, and there
 601                  * must also not be a write in progress or the uiomove()
 602                  * in the write might block and temporarily release
 603                  * its wlock, then reacquire and update windex.  We are
 604                  * only serialized against reads, not writes.
 605                  *
 606                  * XXX should we even bother resetting the indices?  It
 607                  *     might actually be more cache efficient not to.
 608                  */
 609                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 610                     rpipe->pipe_wip == 0) {
 611                         rpipe->pipe_buffer.rindex = 0;
 612                         rpipe->pipe_buffer.windex = 0;
 613                 }
 614
 615                 /*
 616                  * Wait for more data.
 617                  *
 618                  * Pipe_state can only be set if both the rlock and wlock
 619                  * are held.
 620                  */
 621                 rpipe->pipe_state |= PIPE_WANTR;
 622                 tsleep_interlock(rpipe, PCATCH);
 623                 lwkt_reltoken(&wlock);
 624                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 625                 ++pipe_rblocked_count;
 626                 if (error)
 627                         break;
 628         }
 629         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 630
 631         /*
 632          * Uptime last access time
 633          */
 634         if (error == 0 && nread)
 635                 vfs_timestamp(&rpipe->pipe_atime);
 636
 637         /*
 638          * If we drained the FIFO more then half way then handle
 639          * write blocking hysteresis.
 640          *
 641          * Note that PIPE_WANTW cannot be set by the writer without
 642          * it holding both rlock and wlock, so we can test it
 643          * while holding just rlock.
 644          */
 645         if (notify_writer) {
 646                 if (rpipe->pipe_state & PIPE_WANTW) {
 647                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 648                         if (rpipe->pipe_state & PIPE_WANTW) {
 649                                 rpipe->pipe_state &= ~PIPE_WANTW;
 650                                 lwkt_reltoken(&wlock);
 651                                 wakeup(rpipe);
 652                         } else {
 653                                 lwkt_reltoken(&wlock);
 654                         }
 655                 }
 656         }
 657         size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 658         lwkt_reltoken(&rlock);
 659
 660         /*
 661          * If enough space is available in buffer then wakeup sel writers?
 662          */
 663         if ((rpipe->pipe_buffer.size - size) >= PIPE_BUF)
 664                 pipeselwakeup(rpipe);
 665         pipe_rel_mplock(&mpsave);
 666         return (error);
 667 }
 668
 669 /*
 670  * MPALMOSTSAFE - acquires mplock
 671  */
 672 static int
 673 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 674 {
 675         int error;
 676         int orig_resid;
 677         int nbio;
 678         struct pipe *wpipe, *rpipe;
 679         lwkt_tokref rlock;
 680         lwkt_tokref wlock;
 681         u_int windex;
 682         u_int space;
 683         u_int wcount;
 684         int mpsave;
 685         int bigwrite;
 686         int bigcount;
 687
 688         pipe_get_mplock(&mpsave);
 689
 690         /*
 691          * Writes go to the peer.  The peer will always exist.
 692          */
 693         rpipe = (struct pipe *) fp->f_data;
 694         wpipe = rpipe->pipe_peer;
 695         lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
 696         if (wpipe->pipe_state & PIPE_WEOF) {
 697                 pipe_rel_mplock(&mpsave);
 698                 lwkt_reltoken(&wlock);
 699                 return (EPIPE);
 700         }
 701
 702         /*
 703          * Degenerate case (EPIPE takes prec)
 704          */
 705         if (uio->uio_resid == 0) {
 706                 pipe_rel_mplock(&mpsave);
 707                 lwkt_reltoken(&wlock);
 708                 return(0);
 709         }
 710
 711         /*
 712          * Writes are serialized (start_uio must be called with wlock)
 713          */
 714         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 715         if (error) {
 716                 pipe_rel_mplock(&mpsave);
 717                 lwkt_reltoken(&wlock);
 718                 return (error);
 719         }
 720
 721         if (fflags & O_FBLOCKING)
 722                 nbio = 0;
 723         else if (fflags & O_FNONBLOCKING)
 724                 nbio = 1;
 725         else if (fp->f_flag & O_NONBLOCK)
 726                 nbio = 1;
 727         else
 728                 nbio = 0;
 729
 730         /*
 731          * If it is advantageous to resize the pipe buffer, do
 732          * so.  We are write-serialized so we can block safely.
 733          */
 734         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 735             (pipe_nbig < pipe_maxbig) &&
 736             wpipe->pipe_wantwcnt > 4 &&
 737             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 738                 /*
 739                  * Recheck after lock.
 740                  */
 741                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 742                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 743                     (pipe_nbig < pipe_maxbig) &&
 744                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 745                         atomic_add_int(&pipe_nbig, 1);
 746                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 747                                 ++pipe_bigcount;
 748                         else
 749                                 atomic_subtract_int(&pipe_nbig, 1);
 750                 }
 751                 lwkt_reltoken(&rlock);
 752         }
 753
 754         orig_resid = uio->uio_resid;
 755         wcount = 0;
 756
 757         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 758         bigcount = 10;
 759
 760         while (uio->uio_resid) {
 761                 if (wpipe->pipe_state & PIPE_WEOF) {
 762                         error = EPIPE;
 763                         break;
 764                 }
 765
 766                 /*
 767                  * Don't hog the cpu.
 768                  */
 769                 if (bigwrite && --bigcount == 0) {
 770                         lwkt_user_yield();
 771                         bigcount = 10;
 772                         if (CURSIG(curthread->td_lwp)) {
 773                                 error = EINTR;
 774                                 break;
 775                         }
 776                 }
 777
 778                 windex = wpipe->pipe_buffer.windex &
 779                          (wpipe->pipe_buffer.size - 1);
 780                 space = wpipe->pipe_buffer.size -
 781                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 782                 cpu_lfence();
 783
 784                 /* Writes of size <= PIPE_BUF must be atomic. */
 785                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 786                         space = 0;
 787
 788                 /*
 789                  * Write to fill, read size handles write hysteresis.  Also
 790                  * additional restrictions can cause select-based non-blocking
 791                  * writes to spin.
 792                  */
 793                 if (space > 0) {
 794                         u_int segsize;
 795
 796                         /*
 797                          * Transfer size is minimum of uio transfer
 798                          * and free space in pipe buffer.
 799                          *
 800                          * Limit each uiocopy to no more then PIPE_SIZE
 801                          * so we can keep the gravy train going on a
 802                          * SMP box.  This doubles the performance for
 803                          * write sizes > 16K.  Otherwise large writes
 804                          * wind up doing an inefficient synchronous
 805                          * ping-pong.
 806                          */
 807                         space = szmin(space, uio->uio_resid);
 808                         if (space > PIPE_SIZE)
 809                                 space = PIPE_SIZE;
 810
 811                         /*
 812                          * First segment to transfer is minimum of
 813                          * transfer size and contiguous space in
 814                          * pipe buffer.  If first segment to transfer
 815                          * is less than the transfer size, we've got
 816                          * a wraparound in the buffer.
 817                          */
 818                         segsize = wpipe->pipe_buffer.size - windex;
 819                         if (segsize > space)
 820                                 segsize = space;
 821
 822 #ifdef SMP
 823                         /*
 824                          * If this is the first loop and the reader is
 825                          * blocked, do a preemptive wakeup of the reader.
 826                          *
 827                          * On SMP the IPI latency plus the wlock interlock
 828                          * on the reader side is the fastest way to get the
 829                          * reader going.  (The scheduler will hard loop on
 830                          * lock tokens).
 831                          *
 832                          * NOTE: We can't clear WANTR here without acquiring
 833                          * the rlock, which we don't want to do here!
 834                          */
 835                         if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1)
 836                                 wakeup(wpipe);
 837 #endif
 838
 839                         /*
 840                          * Transfer segment, which may include a wrap-around.
 841                          * Update windex to account for both all in one go
 842                          * so the reader can read() the data atomically.
 843                          */
 844                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 845                                         segsize, uio);
 846                         if (error == 0 && segsize < space) {
 847                                 segsize = space - segsize;
 848                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 849                                                 segsize, uio);
 850                         }
 851                         if (error)
 852                                 break;
 853                         cpu_mfence();
 854                         wpipe->pipe_buffer.windex += space;
 855                         wcount += space;
 856                         continue;
 857                 }
 858
 859                 /*
 860                  * We need both the rlock and the wlock to interlock against
 861                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 862                  *
 863                  * These are token locks so we do not have to worry about
 864                  * deadlocks.
 865                  */
 866                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 867
 868                 /*
 869                  * If the "read-side" has been blocked, wake it up now
 870                  * and yield to let it drain synchronously rather
 871                  * then block.
 872                  */
 873                 if (wpipe->pipe_state & PIPE_WANTR) {
 874                         wpipe->pipe_state &= ~PIPE_WANTR;
 875                         wakeup(wpipe);
 876                 }
 877
 878                 /*
 879                  * don't block on non-blocking I/O
 880                  */
 881                 if (nbio) {
 882                         lwkt_reltoken(&rlock);
 883                         error = EAGAIN;
 884                         break;
 885                 }
 886
 887                 /*
 888                  * re-test whether we have to block in the writer after
 889                  * acquiring both locks, in case the reader opened up
 890                  * some space.
 891                  */
 892                 space = wpipe->pipe_buffer.size -
 893                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 894                 cpu_lfence();
 895                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 896                         space = 0;
 897
 898                 /*
 899                  * We have no more space and have something to offer,
 900                  * wake up select/poll.
 901                  */
 902                 if (space == 0) {
 903                         wpipe->pipe_state |= PIPE_WANTW;
 904                         ++wpipe->pipe_wantwcnt;
 905                         pipeselwakeup(wpipe);
 906                         if (wpipe->pipe_state & PIPE_WANTW)
 907                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 908                         ++pipe_wblocked_count;
 909                 }
 910                 lwkt_reltoken(&rlock);
 911
 912                 /*
 913                  * Break out if we errored or the read side wants us to go
 914                  * away.
 915                  */
 916                 if (error)
 917                         break;
 918                 if (wpipe->pipe_state & PIPE_WEOF) {
 919                         error = EPIPE;
 920                         break;
 921                 }
 922         }
 923         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 924
 925         /*
 926          * If we have put any characters in the buffer, we wake up
 927          * the reader.
 928          *
 929          * Both rlock and wlock are required to be able to modify pipe_state.
 930          */
 931         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 932                 if (wpipe->pipe_state & PIPE_WANTR) {
 933                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 934                         if (wpipe->pipe_state & PIPE_WANTR) {
 935                                 wpipe->pipe_state &= ~PIPE_WANTR;
 936                                 lwkt_reltoken(&rlock);
 937                                 wakeup(wpipe);
 938                         } else {
 939                                 lwkt_reltoken(&rlock);
 940                         }
 941                 }
 942         }
 943
 944         /*
 945          * Don't return EPIPE if I/O was successful
 946          */
 947         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 948             (uio->uio_resid == 0) &&
 949             (error == EPIPE)) {
 950                 error = 0;
 951         }
 952
 953         if (error == 0)
 954                 vfs_timestamp(&wpipe->pipe_mtime);
 955
 956         /*
 957          * We have something to offer,
 958          * wake up select/poll.
 959          */
 960         space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;
 961         lwkt_reltoken(&wlock);
 962         if (space)
 963                 pipeselwakeup(wpipe);
 964         pipe_rel_mplock(&mpsave);
 965         return (error);
 966 }
 967
 968 /*
 969  * MPALMOSTSAFE - acquires mplock
 970  *
 971  * we implement a very minimal set of ioctls for compatibility with sockets.
 972  */
 973 int
 974 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
 975 {
 976         struct pipe *mpipe;
 977         lwkt_tokref rlock;
 978         lwkt_tokref wlock;
 979         int error;
 980         int mpsave;
 981
 982         pipe_get_mplock(&mpsave);
 983         mpipe = (struct pipe *)fp->f_data;
 984
 985         lwkt_gettoken(&rlock, &mpipe->pipe_rlock);
 986         lwkt_gettoken(&wlock, &mpipe->pipe_wlock);
 987
 988         switch (cmd) {
 989         case FIOASYNC:
 990                 if (*(int *)data) {
 991                         mpipe->pipe_state |= PIPE_ASYNC;
 992                 } else {
 993                         mpipe->pipe_state &= ~PIPE_ASYNC;
 994                 }
 995                 error = 0;
 996                 break;
 997         case FIONREAD:
 998                 *(int *)data = mpipe->pipe_buffer.windex -
 999                                 mpipe->pipe_buffer.rindex;
1000                 error = 0;
1001                 break;
1002         case FIOSETOWN:
1003                 get_mplock();
1004                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1005                 rel_mplock();
1006                 break;
1007         case FIOGETOWN:
1008                 *(int *)data = fgetown(mpipe->pipe_sigio);
1009                 error = 0;
1010                 break;
1011         case TIOCSPGRP:
1012                 /* This is deprecated, FIOSETOWN should be used instead. */
1013                 get_mplock();
1014                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1015                 rel_mplock();
1016                 break;
1017
1018         case TIOCGPGRP:
1019                 /* This is deprecated, FIOGETOWN should be used instead. */
1020                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1021                 error = 0;
1022                 break;
1023         default:
1024                 error = ENOTTY;
1025                 break;
1026         }
1027         lwkt_reltoken(&rlock);
1028         lwkt_reltoken(&wlock);
1029         pipe_rel_mplock(&mpsave);
1030
1031         return (error);
1032 }
1033
1034 /*
1035  * MPALMOSTSAFE - acquires mplock
1036  */
1037 int
1038 pipe_poll(struct file *fp, int events, struct ucred *cred)
1039 {
1040         struct pipe *rpipe;
1041         struct pipe *wpipe;
1042         int revents = 0;
1043         u_int space;
1044         int mpsave;
1045
1046         pipe_get_mplock(&mpsave);
1047         rpipe = (struct pipe *)fp->f_data;
1048         wpipe = rpipe->pipe_peer;
1049         if (events & (POLLIN | POLLRDNORM)) {
1050                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
1051                     (rpipe->pipe_state & PIPE_REOF)) {
1052                         revents |= events & (POLLIN | POLLRDNORM);
1053                 }
1054         }
1055
1056         if (events & (POLLOUT | POLLWRNORM)) {
1057                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
1058                         revents |= events & (POLLOUT | POLLWRNORM);
1059                 } else {
1060                         space = wpipe->pipe_buffer.windex -
1061                                 wpipe->pipe_buffer.rindex;
1062                         space = wpipe->pipe_buffer.size - space;
1063                         if (space >= PIPE_BUF)
1064                                 revents |= events & (POLLOUT | POLLWRNORM);
1065                 }
1066         }
1067
1068         if ((rpipe->pipe_state & PIPE_REOF) ||
1069             (wpipe == NULL) ||
1070             (wpipe->pipe_state & PIPE_WEOF))
1071                 revents |= POLLHUP;
1072
1073         if (revents == 0) {
1074                 if (events & (POLLIN | POLLRDNORM)) {
1075                         selrecord(curthread, &rpipe->pipe_sel);
1076                         rpipe->pipe_state |= PIPE_SEL;
1077                 }
1078
1079                 if (events & (POLLOUT | POLLWRNORM)) {
1080                         selrecord(curthread, &wpipe->pipe_sel);
1081                         wpipe->pipe_state |= PIPE_SEL;
1082                 }
1083         }
1084         pipe_rel_mplock(&mpsave);
1085         return (revents);
1086 }
1087
1088 /*
1089  * MPSAFE
1090  */
1091 static int
1092 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1093 {
1094         struct pipe *pipe;
1095         int mpsave;
1096
1097         pipe_get_mplock(&mpsave);
1098         pipe = (struct pipe *)fp->f_data;
1099
1100         bzero((caddr_t)ub, sizeof(*ub));
1101         ub->st_mode = S_IFIFO;
1102         ub->st_blksize = pipe->pipe_buffer.size;
1103         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1104         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1105         ub->st_atimespec = pipe->pipe_atime;
1106         ub->st_mtimespec = pipe->pipe_mtime;
1107         ub->st_ctimespec = pipe->pipe_ctime;
1108         /*
1109          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1110          * st_flags, st_gen.
1111          * XXX (st_dev, st_ino) should be unique.
1112          */
1113         pipe_rel_mplock(&mpsave);
1114         return (0);
1115 }
1116
1117 /*
1118  * MPALMOSTSAFE - acquires mplock
1119  */
1120 static int
1121 pipe_close(struct file *fp)
1122 {
1123         struct pipe *cpipe;
1124
1125         get_mplock();
1126         cpipe = (struct pipe *)fp->f_data;
1127         fp->f_ops = &badfileops;
1128         fp->f_data = NULL;
1129         funsetown(cpipe->pipe_sigio);
1130         pipeclose(cpipe);
1131         rel_mplock();
1132         return (0);
1133 }
1134
1135 /*
1136  * Shutdown one or both directions of a full-duplex pipe.
1137  *
1138  * MPALMOSTSAFE - acquires mplock
1139  */
1140 static int
1141 pipe_shutdown(struct file *fp, int how)
1142 {
1143         struct pipe *rpipe;
1144         struct pipe *wpipe;
1145         int error = EPIPE;
1146         lwkt_tokref rpipe_rlock;
1147         lwkt_tokref rpipe_wlock;
1148         lwkt_tokref wpipe_rlock;
1149         lwkt_tokref wpipe_wlock;
1150         int mpsave;
1151
1152         pipe_get_mplock(&mpsave);
1153         rpipe = (struct pipe *)fp->f_data;
1154         wpipe = rpipe->pipe_peer;
1155
1156         /*
1157          * We modify pipe_state on both pipes, which means we need
1158          * all four tokens!
1159          */
1160         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1161         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1162         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1163         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1164
1165         switch(how) {
1166         case SHUT_RDWR:
1167         case SHUT_RD:
1168                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1169                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1170                 if (rpipe->pipe_state & PIPE_WANTR) {
1171                         rpipe->pipe_state &= ~PIPE_WANTR;
1172                         wakeup(rpipe);
1173                 }
1174                 if (rpipe->pipe_state & PIPE_WANTW) {
1175                         rpipe->pipe_state &= ~PIPE_WANTW;
1176                         wakeup(rpipe);
1177                 }
1178                 error = 0;
1179                 if (how == SHUT_RD)
1180                         break;
1181                 /* fall through */
1182         case SHUT_WR:
1183                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1184                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1185                 if (wpipe->pipe_state & PIPE_WANTR) {
1186                         wpipe->pipe_state &= ~PIPE_WANTR;
1187                         wakeup(wpipe);
1188                 }
1189                 if (wpipe->pipe_state & PIPE_WANTW) {
1190                         wpipe->pipe_state &= ~PIPE_WANTW;
1191                         wakeup(wpipe);
1192                 }
1193                 error = 0;
1194                 break;
1195         }
1196         pipeselwakeup(rpipe);
1197         pipeselwakeup(wpipe);
1198
1199         lwkt_reltoken(&rpipe_rlock);
1200         lwkt_reltoken(&rpipe_wlock);
1201         lwkt_reltoken(&wpipe_rlock);
1202         lwkt_reltoken(&wpipe_wlock);
1203
1204         pipe_rel_mplock(&mpsave);
1205         return (error);
1206 }
1207
1208 static void
1209 pipe_free_kmem(struct pipe *cpipe)
1210 {
1211         if (cpipe->pipe_buffer.buffer != NULL) {
1212                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1213                         atomic_subtract_int(&pipe_nbig, 1);
1214                 kmem_free(&kernel_map,
1215                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1216                         cpipe->pipe_buffer.size);
1217                 cpipe->pipe_buffer.buffer = NULL;
1218                 cpipe->pipe_buffer.object = NULL;
1219         }
1220 }
1221
1222 /*
1223  * Close the pipe.  The slock must be held to interlock against simultanious
1224  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1225  */
1226 static void
1227 pipeclose(struct pipe *cpipe)
1228 {
1229         globaldata_t gd;
1230         struct pipe *ppipe;
1231         lwkt_tokref cpipe_rlock;
1232         lwkt_tokref cpipe_wlock;
1233         lwkt_tokref ppipe_rlock;
1234         lwkt_tokref ppipe_wlock;
1235
1236         if (cpipe == NULL)
1237                 return;
1238
1239         /*
1240          * The slock may not have been allocated yet (close during
1241          * initialization)
1242          *
1243          * We need both the read and write tokens to modify pipe_state.
1244          */
1245         if (cpipe->pipe_slock)
1246                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1247         lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1248         lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1249
1250         /*
1251          * Set our state, wakeup anyone waiting in select, and
1252          * wakeup anyone blocked on our pipe.
1253          */
1254         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1255         pipeselwakeup(cpipe);
1256         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1257                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1258                 wakeup(cpipe);
1259         }
1260
1261         /*
1262          * Disconnect from peer.
1263          */
1264         if ((ppipe = cpipe->pipe_peer) != NULL) {
1265                 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1266                 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1267                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1268                 pipeselwakeup(ppipe);
1269                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1270                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1271                         wakeup(ppipe);
1272                 }
1273                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1274                         get_mplock();
1275                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1276                         rel_mplock();
1277                 }
1278                 lwkt_reltoken(&ppipe_rlock);
1279                 lwkt_reltoken(&ppipe_wlock);
1280         }
1281
1282         /*
1283          * If the peer is also closed we can free resources for both
1284          * sides, otherwise we leave our side intact to deal with any
1285          * races (since we only have the slock).
1286          */
1287         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1288                 cpipe->pipe_peer = NULL;
1289                 ppipe->pipe_peer = NULL;
1290                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1291                 pipeclose(ppipe);
1292                 ppipe = NULL;
1293         }
1294
1295         lwkt_reltoken(&cpipe_rlock);
1296         lwkt_reltoken(&cpipe_wlock);
1297         if (cpipe->pipe_slock)
1298                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1299
1300         /*
1301          * If we disassociated from our peer we can free resources
1302          */
1303         if (ppipe == NULL) {
1304                 gd = mycpu;
1305                 if (cpipe->pipe_slock) {
1306                         kfree(cpipe->pipe_slock, M_PIPE);
1307                         cpipe->pipe_slock = NULL;
1308                 }
1309                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1310                     cpipe->pipe_buffer.size != PIPE_SIZE
1311                 ) {
1312                         pipe_free_kmem(cpipe);
1313                         kfree(cpipe, M_PIPE);
1314                 } else {
1315                         cpipe->pipe_state = 0;
1316                         cpipe->pipe_peer = gd->gd_pipeq;
1317                         gd->gd_pipeq = cpipe;
1318                         ++gd->gd_pipeqcount;
1319                 }
1320         }
1321 }
1322
1323 /*
1324  * MPALMOSTSAFE - acquires mplock
1325  */
1326 static int
1327 pipe_kqfilter(struct file *fp, struct knote *kn)
1328 {
1329         struct pipe *cpipe;
1330
1331         get_mplock();
1332         cpipe = (struct pipe *)kn->kn_fp->f_data;
1333
1334         switch (kn->kn_filter) {
1335         case EVFILT_READ:
1336                 kn->kn_fop = &pipe_rfiltops;
1337                 break;
1338         case EVFILT_WRITE:
1339                 kn->kn_fop = &pipe_wfiltops;
1340                 cpipe = cpipe->pipe_peer;
1341                 if (cpipe == NULL) {
1342                         /* other end of pipe has been closed */
1343                         rel_mplock();
1344                         return (EPIPE);
1345                 }
1346                 break;
1347         default:
1348                 return (1);
1349         }
1350         kn->kn_hook = (caddr_t)cpipe;
1351
1352         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1353         rel_mplock();
1354         return (0);
1355 }
1356
1357 static void
1358 filt_pipedetach(struct knote *kn)
1359 {
1360         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1361
1362         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1363 }
1364
1365 /*ARGSUSED*/
1366 static int
1367 filt_piperead(struct knote *kn, long hint)
1368 {
1369         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1370
1371         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1372
1373         /* XXX RACE */
1374         if (rpipe->pipe_state & PIPE_REOF) {
1375                 kn->kn_flags |= EV_EOF;
1376                 return (1);
1377         }
1378         return (kn->kn_data > 0);
1379 }
1380
1381 /*ARGSUSED*/
1382 static int
1383 filt_pipewrite(struct knote *kn, long hint)
1384 {
1385         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1386         struct pipe *wpipe = rpipe->pipe_peer;
1387         u_int32_t space;
1388
1389         /* XXX RACE */
1390         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1391                 kn->kn_data = 0;
1392                 kn->kn_flags |= EV_EOF;
1393                 return (1);
1394         }
1395         space = wpipe->pipe_buffer.windex -
1396                 wpipe->pipe_buffer.rindex;
1397         space = wpipe->pipe_buffer.size - space;
1398         kn->kn_data = space;
1399         return (kn->kn_data >= PIPE_BUF);
1400 }