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 #include <sys/mplock2.h>
  67
  68 #include <machine/cpufunc.h>
  69
  70 /*
  71  * interfaces to the outside world
  72  */
  73 static int pipe_read (struct file *fp, struct uio *uio,
  74                 struct ucred *cred, int flags);
  75 static int pipe_write (struct file *fp, struct uio *uio,
  76                 struct ucred *cred, int flags);
  77 static int pipe_close (struct file *fp);
  78 static int pipe_shutdown (struct file *fp, int how);
  79 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
  80 static int pipe_kqfilter (struct file *fp, struct knote *kn);
  81 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
  82 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data,
  83                 struct ucred *cred, struct sysmsg *msg);
  84
  85 static struct fileops pipeops = {
  86         .fo_read = pipe_read,
  87         .fo_write = pipe_write,
  88         .fo_ioctl = pipe_ioctl,
  89         .fo_poll = pipe_poll,
  90         .fo_kqfilter = pipe_kqfilter,
  91         .fo_stat = pipe_stat,
  92         .fo_close = pipe_close,
  93         .fo_shutdown = pipe_shutdown
  94 };
  95
  96 static void     filt_pipedetach(struct knote *kn);
  97 static int      filt_piperead(struct knote *kn, long hint);
  98 static int      filt_pipewrite(struct knote *kn, long hint);
  99
 100 static struct filterops pipe_rfiltops =
 101         { 1, NULL, filt_pipedetach, filt_piperead };
 102 static struct filterops pipe_wfiltops =
 103         { 1, NULL, filt_pipedetach, filt_pipewrite };
 104
 105 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 106
 107 /*
 108  * Default pipe buffer size(s), this can be kind-of large now because pipe
 109  * space is pageable.  The pipe code will try to maintain locality of
 110  * reference for performance reasons, so small amounts of outstanding I/O
 111  * will not wipe the cache.
 112  */
 113 #define MINPIPESIZE (PIPE_SIZE/3)
 114 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 115
 116 /*
 117  * Limit the number of "big" pipes
 118  */
 119 #define LIMITBIGPIPES   64
 120 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 121
 122 static int pipe_maxbig = LIMITBIGPIPES;
 123 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 124 static int pipe_bigcount;
 125 static int pipe_nbig;
 126 static int pipe_bcache_alloc;
 127 static int pipe_bkmem_alloc;
 128 static int pipe_rblocked_count;
 129 static int pipe_wblocked_count;
 130
 131 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 132 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 133         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 134 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 135         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 136 SYSCTL_INT(_kern_pipe, OID_AUTO, rblocked,
 137         CTLFLAG_RW, &pipe_rblocked_count, 0, "number of times pipe expanded");
 138 SYSCTL_INT(_kern_pipe, OID_AUTO, wblocked,
 139         CTLFLAG_RW, &pipe_wblocked_count, 0, "number of times pipe expanded");
 140 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 141         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 142 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 143         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 144 #ifdef SMP
 145 static int pipe_delay = 5000;   /* 5uS default */
 146 SYSCTL_INT(_kern_pipe, OID_AUTO, delay,
 147         CTLFLAG_RW, &pipe_delay, 0, "SMP delay optimization in ns");
 148 static int pipe_mpsafe = 1;
 149 SYSCTL_INT(_kern_pipe, OID_AUTO, mpsafe,
 150         CTLFLAG_RW, &pipe_mpsafe, 0, "");
 151 #endif
 152 #if !defined(NO_PIPE_SYSCTL_STATS)
 153 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 154         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 155 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 156         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 157 #endif
 158
 159 static void pipeclose (struct pipe *cpipe);
 160 static void pipe_free_kmem (struct pipe *cpipe);
 161 static int pipe_create (struct pipe **cpipep);
 162 static __inline void pipeselwakeup (struct pipe *cpipe);
 163 static int pipespace (struct pipe *cpipe, int size);
 164
 165 static __inline int
 166 pipeseltest(struct pipe *cpipe)
 167 {
 168         return ((cpipe->pipe_state & PIPE_SEL) ||
 169                 ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) ||
 170                 SLIST_FIRST(&cpipe->pipe_sel.si_note));
 171 }
 172
 173 static __inline void
 174 pipeselwakeup(struct pipe *cpipe)
 175 {
 176         if (cpipe->pipe_state & PIPE_SEL) {
 177                 get_mplock();
 178                 cpipe->pipe_state &= ~PIPE_SEL;
 179                 selwakeup(&cpipe->pipe_sel);
 180                 rel_mplock();
 181         }
 182         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio) {
 183                 get_mplock();
 184                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 185                 rel_mplock();
 186         }
 187         if (SLIST_FIRST(&cpipe->pipe_sel.si_note)) {
 188                 get_mplock();
 189                 KNOTE(&cpipe->pipe_sel.si_note, 0);
 190                 rel_mplock();
 191         }
 192 }
 193
 194 /*
 195  * These routines are called before and after a UIO.  The UIO
 196  * may block, causing our held tokens to be lost temporarily.
 197  *
 198  * We use these routines to serialize reads against other reads
 199  * and writes against other writes.
 200  *
 201  * The read token is held on entry so *ipp does not race.
 202  */
 203 static __inline int
 204 pipe_start_uio(struct pipe *cpipe, int *ipp)
 205 {
 206         int error;
 207
 208         while (*ipp) {
 209                 *ipp = -1;
 210                 error = tsleep(ipp, PCATCH, "pipexx", 0);
 211                 if (error)
 212                         return (error);
 213         }
 214         *ipp = 1;
 215         return (0);
 216 }
 217
 218 static __inline void
 219 pipe_end_uio(struct pipe *cpipe, int *ipp)
 220 {
 221         if (*ipp < 0) {
 222                 *ipp = 0;
 223                 wakeup(ipp);
 224         } else {
 225                 KKASSERT(*ipp > 0);
 226                 *ipp = 0;
 227         }
 228 }
 229
 230 static __inline void
 231 pipe_get_mplock(int *save)
 232 {
 233 #ifdef SMP
 234         if (pipe_mpsafe == 0) {
 235                 get_mplock();
 236                 *save = 1;
 237         } else
 238 #endif
 239         {
 240                 *save = 0;
 241         }
 242 }
 243
 244 static __inline void
 245 pipe_rel_mplock(int *save)
 246 {
 247 #ifdef SMP
 248         if (*save)
 249                 rel_mplock();
 250 #endif
 251 }
 252
 253
 254 /*
 255  * The pipe system call for the DTYPE_PIPE type of pipes
 256  *
 257  * pipe_args(int dummy)
 258  *
 259  * MPSAFE
 260  */
 261 int
 262 sys_pipe(struct pipe_args *uap)
 263 {
 264         struct thread *td = curthread;
 265         struct filedesc *fdp = td->td_proc->p_fd;
 266         struct file *rf, *wf;
 267         struct pipe *rpipe, *wpipe;
 268         int fd1, fd2, error;
 269
 270         rpipe = wpipe = NULL;
 271         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 272                 pipeclose(rpipe);
 273                 pipeclose(wpipe);
 274                 return (ENFILE);
 275         }
 276
 277         error = falloc(td->td_lwp, &rf, &fd1);
 278         if (error) {
 279                 pipeclose(rpipe);
 280                 pipeclose(wpipe);
 281                 return (error);
 282         }
 283         uap->sysmsg_fds[0] = fd1;
 284
 285         /*
 286          * Warning: once we've gotten past allocation of the fd for the
 287          * read-side, we can only drop the read side via fdrop() in order
 288          * to avoid races against processes which manage to dup() the read
 289          * side while we are blocked trying to allocate the write side.
 290          */
 291         rf->f_type = DTYPE_PIPE;
 292         rf->f_flag = FREAD | FWRITE;
 293         rf->f_ops = &pipeops;
 294         rf->f_data = rpipe;
 295         error = falloc(td->td_lwp, &wf, &fd2);
 296         if (error) {
 297                 fsetfd(fdp, NULL, fd1);
 298                 fdrop(rf);
 299                 /* rpipe has been closed by fdrop(). */
 300                 pipeclose(wpipe);
 301                 return (error);
 302         }
 303         wf->f_type = DTYPE_PIPE;
 304         wf->f_flag = FREAD | FWRITE;
 305         wf->f_ops = &pipeops;
 306         wf->f_data = wpipe;
 307         uap->sysmsg_fds[1] = fd2;
 308
 309         rpipe->pipe_slock = kmalloc(sizeof(struct lock),
 310                                     M_PIPE, M_WAITOK|M_ZERO);
 311         wpipe->pipe_slock = rpipe->pipe_slock;
 312         rpipe->pipe_peer = wpipe;
 313         wpipe->pipe_peer = rpipe;
 314         lockinit(rpipe->pipe_slock, "pipecl", 0, 0);
 315
 316         /*
 317          * Once activated the peer relationship remains valid until
 318          * both sides are closed.
 319          */
 320         fsetfd(fdp, rf, fd1);
 321         fsetfd(fdp, wf, fd2);
 322         fdrop(rf);
 323         fdrop(wf);
 324
 325         return (0);
 326 }
 327
 328 /*
 329  * Allocate kva for pipe circular buffer, the space is pageable
 330  * This routine will 'realloc' the size of a pipe safely, if it fails
 331  * it will retain the old buffer.
 332  * If it fails it will return ENOMEM.
 333  */
 334 static int
 335 pipespace(struct pipe *cpipe, int size)
 336 {
 337         struct vm_object *object;
 338         caddr_t buffer;
 339         int npages, error;
 340
 341         npages = round_page(size) / PAGE_SIZE;
 342         object = cpipe->pipe_buffer.object;
 343
 344         /*
 345          * [re]create the object if necessary and reserve space for it
 346          * in the kernel_map.  The object and memory are pageable.  On
 347          * success, free the old resources before assigning the new
 348          * ones.
 349          */
 350         if (object == NULL || object->size != npages) {
 351                 get_mplock();
 352                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 353                 buffer = (caddr_t)vm_map_min(&kernel_map);
 354
 355                 error = vm_map_find(&kernel_map, object, 0,
 356                                     (vm_offset_t *)&buffer,
 357                                     size, PAGE_SIZE,
 358                                     1, VM_MAPTYPE_NORMAL,
 359                                     VM_PROT_ALL, VM_PROT_ALL,
 360                                     0);
 361
 362                 if (error != KERN_SUCCESS) {
 363                         vm_object_deallocate(object);
 364                         rel_mplock();
 365                         return (ENOMEM);
 366                 }
 367                 pipe_free_kmem(cpipe);
 368                 rel_mplock();
 369                 cpipe->pipe_buffer.object = object;
 370                 cpipe->pipe_buffer.buffer = buffer;
 371                 cpipe->pipe_buffer.size = size;
 372                 ++pipe_bkmem_alloc;
 373         } else {
 374                 ++pipe_bcache_alloc;
 375         }
 376         cpipe->pipe_buffer.rindex = 0;
 377         cpipe->pipe_buffer.windex = 0;
 378         return (0);
 379 }
 380
 381 /*
 382  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 383  * our per-cpu cache if possible.  For now make sure it is sized for the
 384  * smaller PIPE_SIZE default.
 385  */
 386 static int
 387 pipe_create(struct pipe **cpipep)
 388 {
 389         globaldata_t gd = mycpu;
 390         struct pipe *cpipe;
 391         int error;
 392
 393         if ((cpipe = gd->gd_pipeq) != NULL) {
 394                 gd->gd_pipeq = cpipe->pipe_peer;
 395                 --gd->gd_pipeqcount;
 396                 cpipe->pipe_peer = NULL;
 397                 cpipe->pipe_wantwcnt = 0;
 398         } else {
 399                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 400         }
 401         *cpipep = cpipe;
 402         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 403                 return (error);
 404         vfs_timestamp(&cpipe->pipe_ctime);
 405         cpipe->pipe_atime = cpipe->pipe_ctime;
 406         cpipe->pipe_mtime = cpipe->pipe_ctime;
 407         lwkt_token_init(&cpipe->pipe_rlock, 1);
 408         lwkt_token_init(&cpipe->pipe_wlock, 1);
 409         return (0);
 410 }
 411
 412 /*
 413  * MPALMOSTSAFE (acquires mplock)
 414  */
 415 static int
 416 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 417 {
 418         struct pipe *rpipe;
 419         int error;
 420         size_t nread = 0;
 421         int nbio;
 422         u_int size;     /* total bytes available */
 423         u_int nsize;    /* total bytes to read */
 424         u_int rindex;   /* contiguous bytes available */
 425         int notify_writer;
 426         int mpsave;
 427         int bigread;
 428         int bigcount;
 429
 430         if (uio->uio_resid == 0)
 431                 return(0);
 432
 433         /*
 434          * Setup locks, calculate nbio
 435          */
 436         pipe_get_mplock(&mpsave);
 437         rpipe = (struct pipe *)fp->f_data;
 438         lwkt_gettoken(&rpipe->pipe_rlock);
 439
 440         if (fflags & O_FBLOCKING)
 441                 nbio = 0;
 442         else if (fflags & O_FNONBLOCKING)
 443                 nbio = 1;
 444         else if (fp->f_flag & O_NONBLOCK)
 445                 nbio = 1;
 446         else
 447                 nbio = 0;
 448
 449         /*
 450          * Reads are serialized.  Note howeverthat pipe_buffer.buffer and
 451          * pipe_buffer.size can change out from under us when the number
 452          * of bytes in the buffer are zero due to the write-side doing a
 453          * pipespace().
 454          */
 455         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 456         if (error) {
 457                 pipe_rel_mplock(&mpsave);
 458                 lwkt_reltoken(&rpipe->pipe_rlock);
 459                 return (error);
 460         }
 461         notify_writer = 0;
 462
 463         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 464         bigcount = 10;
 465
 466         while (uio->uio_resid) {
 467                 /*
 468                  * Don't hog the cpu.
 469                  */
 470                 if (bigread && --bigcount == 0) {
 471                         lwkt_user_yield();
 472                         bigcount = 10;
 473                         if (CURSIG(curthread->td_lwp)) {
 474                                 error = EINTR;
 475                                 break;
 476                         }
 477                 }
 478
 479                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 480                 cpu_lfence();
 481                 if (size) {
 482                         rindex = rpipe->pipe_buffer.rindex &
 483                                  (rpipe->pipe_buffer.size - 1);
 484                         nsize = size;
 485                         if (nsize > rpipe->pipe_buffer.size - rindex)
 486                                 nsize = rpipe->pipe_buffer.size - rindex;
 487                         nsize = szmin(nsize, uio->uio_resid);
 488
 489                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 490                                         nsize, uio);
 491                         if (error)
 492                                 break;
 493                         cpu_mfence();
 494                         rpipe->pipe_buffer.rindex += nsize;
 495                         nread += nsize;
 496
 497                         /*
 498                          * If the FIFO is still over half full just continue
 499                          * and do not try to notify the writer yet.
 500                          */
 501                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 502                                 notify_writer = 0;
 503                                 continue;
 504                         }
 505
 506                         /*
 507                          * When the FIFO is less then half full notify any
 508                          * waiting writer.  WANTW can be checked while
 509                          * holding just the rlock.
 510                          */
 511                         notify_writer = 1;
 512                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 513                                 continue;
 514                 }
 515
 516                 /*
 517                  * If the "write-side" was blocked we wake it up.  This code
 518                  * is reached either when the buffer is completely emptied
 519                  * or if it becomes more then half-empty.
 520                  *
 521                  * Pipe_state can only be modified if both the rlock and
 522                  * wlock are held.
 523                  */
 524                 if (rpipe->pipe_state & PIPE_WANTW) {
 525                         lwkt_gettoken(&rpipe->pipe_wlock);
 526                         if (rpipe->pipe_state & PIPE_WANTW) {
 527                                 notify_writer = 0;
 528                                 rpipe->pipe_state &= ~PIPE_WANTW;
 529                                 lwkt_reltoken(&rpipe->pipe_wlock);
 530                                 wakeup(rpipe);
 531                         } else {
 532                                 lwkt_reltoken(&rpipe->pipe_wlock);
 533                         }
 534                 }
 535
 536                 /*
 537                  * Pick up our copy loop again if the writer sent data to
 538                  * us while we were messing around.
 539                  *
 540                  * On a SMP box poll up to pipe_delay nanoseconds for new
 541                  * data.  Typically a value of 2000 to 4000 is sufficient
 542                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 543                  * is used for synchronous communications with small packets,
 544                  * and 8000 or so (8uS) will pipeline large buffer xfers
 545                  * between cpus over a pipe.
 546                  *
 547                  * For synchronous communications a hit means doing a
 548                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 549                  * where as miss requiring a tsleep/wakeup sequence
 550                  * will take 7uS or more.
 551                  */
 552                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 553                         continue;
 554
 555 #if defined(SMP) && defined(_RDTSC_SUPPORTED_)
 556                 if (pipe_delay) {
 557                         int64_t tsc_target;
 558                         int good = 0;
 559
 560                         tsc_target = tsc_get_target(pipe_delay);
 561                         while (tsc_test_target(tsc_target) == 0) {
 562                                 if (rpipe->pipe_buffer.windex !=
 563                                     rpipe->pipe_buffer.rindex) {
 564                                         good = 1;
 565                                         break;
 566                                 }
 567                         }
 568                         if (good)
 569                                 continue;
 570                 }
 571 #endif
 572
 573                 /*
 574                  * Detect EOF condition, do not set error.
 575                  */
 576                 if (rpipe->pipe_state & PIPE_REOF)
 577                         break;
 578
 579                 /*
 580                  * Break if some data was read, or if this was a non-blocking
 581                  * read.
 582                  */
 583                 if (nread > 0)
 584                         break;
 585
 586                 if (nbio) {
 587                         error = EAGAIN;
 588                         break;
 589                 }
 590
 591                 /*
 592                  * Last chance, interlock with WANTR.
 593                  */
 594                 lwkt_gettoken(&rpipe->pipe_wlock);
 595                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 596                 if (size) {
 597                         lwkt_reltoken(&rpipe->pipe_wlock);
 598                         continue;
 599                 }
 600
 601                 /*
 602                  * Retest EOF - acquiring a new token can temporarily release
 603                  * tokens already held.
 604                  */
 605                 if (rpipe->pipe_state & PIPE_REOF) {
 606                         lwkt_reltoken(&rpipe->pipe_wlock);
 607                         break;
 608                 }
 609
 610                 /*
 611                  * If there is no more to read in the pipe, reset its
 612                  * pointers to the beginning.  This improves cache hit
 613                  * stats.
 614                  *
 615                  * We need both locks to modify both pointers, and there
 616                  * must also not be a write in progress or the uiomove()
 617                  * in the write might block and temporarily release
 618                  * its wlock, then reacquire and update windex.  We are
 619                  * only serialized against reads, not writes.
 620                  *
 621                  * XXX should we even bother resetting the indices?  It
 622                  *     might actually be more cache efficient not to.
 623                  */
 624                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 625                     rpipe->pipe_wip == 0) {
 626                         rpipe->pipe_buffer.rindex = 0;
 627                         rpipe->pipe_buffer.windex = 0;
 628                 }
 629
 630                 /*
 631                  * Wait for more data.
 632                  *
 633                  * Pipe_state can only be set if both the rlock and wlock
 634                  * are held.
 635                  */
 636                 rpipe->pipe_state |= PIPE_WANTR;
 637                 tsleep_interlock(rpipe, PCATCH);
 638                 lwkt_reltoken(&rpipe->pipe_wlock);
 639                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 640                 ++pipe_rblocked_count;
 641                 if (error)
 642                         break;
 643         }
 644         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 645
 646         /*
 647          * Uptime last access time
 648          */
 649         if (error == 0 && nread)
 650                 vfs_timestamp(&rpipe->pipe_atime);
 651
 652         /*
 653          * If we drained the FIFO more then half way then handle
 654          * write blocking hysteresis.
 655          *
 656          * Note that PIPE_WANTW cannot be set by the writer without
 657          * it holding both rlock and wlock, so we can test it
 658          * while holding just rlock.
 659          */
 660         if (notify_writer) {
 661                 if (rpipe->pipe_state & PIPE_WANTW) {
 662                         lwkt_gettoken(&rpipe->pipe_wlock);
 663                         if (rpipe->pipe_state & PIPE_WANTW) {
 664                                 rpipe->pipe_state &= ~PIPE_WANTW;
 665                                 lwkt_reltoken(&rpipe->pipe_wlock);
 666                                 wakeup(rpipe);
 667                         } else {
 668                                 lwkt_reltoken(&rpipe->pipe_wlock);
 669                         }
 670                 }
 671                 if (pipeseltest(rpipe)) {
 672                         lwkt_gettoken(&rpipe->pipe_wlock);
 673                         pipeselwakeup(rpipe);
 674                         lwkt_reltoken(&rpipe->pipe_wlock);
 675                 }
 676         }
 677         /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
 678         lwkt_reltoken(&rpipe->pipe_rlock);
 679
 680         pipe_rel_mplock(&mpsave);
 681         return (error);
 682 }
 683
 684 /*
 685  * MPALMOSTSAFE - acquires mplock
 686  */
 687 static int
 688 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 689 {
 690         int error;
 691         int orig_resid;
 692         int nbio;
 693         struct pipe *wpipe, *rpipe;
 694         u_int windex;
 695         u_int space;
 696         u_int wcount;
 697         int mpsave;
 698         int bigwrite;
 699         int bigcount;
 700
 701         pipe_get_mplock(&mpsave);
 702
 703         /*
 704          * Writes go to the peer.  The peer will always exist.
 705          */
 706         rpipe = (struct pipe *) fp->f_data;
 707         wpipe = rpipe->pipe_peer;
 708         lwkt_gettoken(&wpipe->pipe_wlock);
 709         if (wpipe->pipe_state & PIPE_WEOF) {
 710                 pipe_rel_mplock(&mpsave);
 711                 lwkt_reltoken(&wpipe->pipe_wlock);
 712                 return (EPIPE);
 713         }
 714
 715         /*
 716          * Degenerate case (EPIPE takes prec)
 717          */
 718         if (uio->uio_resid == 0) {
 719                 pipe_rel_mplock(&mpsave);
 720                 lwkt_reltoken(&wpipe->pipe_wlock);
 721                 return(0);
 722         }
 723
 724         /*
 725          * Writes are serialized (start_uio must be called with wlock)
 726          */
 727         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 728         if (error) {
 729                 pipe_rel_mplock(&mpsave);
 730                 lwkt_reltoken(&wpipe->pipe_wlock);
 731                 return (error);
 732         }
 733
 734         if (fflags & O_FBLOCKING)
 735                 nbio = 0;
 736         else if (fflags & O_FNONBLOCKING)
 737                 nbio = 1;
 738         else if (fp->f_flag & O_NONBLOCK)
 739                 nbio = 1;
 740         else
 741                 nbio = 0;
 742
 743         /*
 744          * If it is advantageous to resize the pipe buffer, do
 745          * so.  We are write-serialized so we can block safely.
 746          */
 747         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 748             (pipe_nbig < pipe_maxbig) &&
 749             wpipe->pipe_wantwcnt > 4 &&
 750             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 751                 /*
 752                  * Recheck after lock.
 753                  */
 754                 lwkt_gettoken(&wpipe->pipe_rlock);
 755                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 756                     (pipe_nbig < pipe_maxbig) &&
 757                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 758                         atomic_add_int(&pipe_nbig, 1);
 759                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 760                                 ++pipe_bigcount;
 761                         else
 762                                 atomic_subtract_int(&pipe_nbig, 1);
 763                 }
 764                 lwkt_reltoken(&wpipe->pipe_rlock);
 765         }
 766
 767         orig_resid = uio->uio_resid;
 768         wcount = 0;
 769
 770         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 771         bigcount = 10;
 772
 773         while (uio->uio_resid) {
 774                 if (wpipe->pipe_state & PIPE_WEOF) {
 775                         error = EPIPE;
 776                         break;
 777                 }
 778
 779                 /*
 780                  * Don't hog the cpu.
 781                  */
 782                 if (bigwrite && --bigcount == 0) {
 783                         lwkt_user_yield();
 784                         bigcount = 10;
 785                         if (CURSIG(curthread->td_lwp)) {
 786                                 error = EINTR;
 787                                 break;
 788                         }
 789                 }
 790
 791                 windex = wpipe->pipe_buffer.windex &
 792                          (wpipe->pipe_buffer.size - 1);
 793                 space = wpipe->pipe_buffer.size -
 794                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 795                 cpu_lfence();
 796
 797                 /* Writes of size <= PIPE_BUF must be atomic. */
 798                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 799                         space = 0;
 800
 801                 /*
 802                  * Write to fill, read size handles write hysteresis.  Also
 803                  * additional restrictions can cause select-based non-blocking
 804                  * writes to spin.
 805                  */
 806                 if (space > 0) {
 807                         u_int segsize;
 808
 809                         /*
 810                          * Transfer size is minimum of uio transfer
 811                          * and free space in pipe buffer.
 812                          *
 813                          * Limit each uiocopy to no more then PIPE_SIZE
 814                          * so we can keep the gravy train going on a
 815                          * SMP box.  This doubles the performance for
 816                          * write sizes > 16K.  Otherwise large writes
 817                          * wind up doing an inefficient synchronous
 818                          * ping-pong.
 819                          */
 820                         space = szmin(space, uio->uio_resid);
 821                         if (space > PIPE_SIZE)
 822                                 space = PIPE_SIZE;
 823
 824                         /*
 825                          * First segment to transfer is minimum of
 826                          * transfer size and contiguous space in
 827                          * pipe buffer.  If first segment to transfer
 828                          * is less than the transfer size, we've got
 829                          * a wraparound in the buffer.
 830                          */
 831                         segsize = wpipe->pipe_buffer.size - windex;
 832                         if (segsize > space)
 833                                 segsize = space;
 834
 835 #ifdef SMP
 836                         /*
 837                          * If this is the first loop and the reader is
 838                          * blocked, do a preemptive wakeup of the reader.
 839                          *
 840                          * On SMP the IPI latency plus the wlock interlock
 841                          * on the reader side is the fastest way to get the
 842                          * reader going.  (The scheduler will hard loop on
 843                          * lock tokens).
 844                          *
 845                          * NOTE: We can't clear WANTR here without acquiring
 846                          * the rlock, which we don't want to do here!
 847                          */
 848                         if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1)
 849                                 wakeup(wpipe);
 850 #endif
 851
 852                         /*
 853                          * Transfer segment, which may include a wrap-around.
 854                          * Update windex to account for both all in one go
 855                          * so the reader can read() the data atomically.
 856                          */
 857                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 858                                         segsize, uio);
 859                         if (error == 0 && segsize < space) {
 860                                 segsize = space - segsize;
 861                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 862                                                 segsize, uio);
 863                         }
 864                         if (error)
 865                                 break;
 866                         cpu_mfence();
 867                         wpipe->pipe_buffer.windex += space;
 868                         wcount += space;
 869                         continue;
 870                 }
 871
 872                 /*
 873                  * We need both the rlock and the wlock to interlock against
 874                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 875                  *
 876                  * These are token locks so we do not have to worry about
 877                  * deadlocks.
 878                  */
 879                 lwkt_gettoken(&wpipe->pipe_rlock);
 880
 881                 /*
 882                  * If the "read-side" has been blocked, wake it up now
 883                  * and yield to let it drain synchronously rather
 884                  * then block.
 885                  */
 886                 if (wpipe->pipe_state & PIPE_WANTR) {
 887                         wpipe->pipe_state &= ~PIPE_WANTR;
 888                         wakeup(wpipe);
 889                 }
 890
 891                 /*
 892                  * don't block on non-blocking I/O
 893                  */
 894                 if (nbio) {
 895                         lwkt_reltoken(&wpipe->pipe_rlock);
 896                         error = EAGAIN;
 897                         break;
 898                 }
 899
 900                 /*
 901                  * re-test whether we have to block in the writer after
 902                  * acquiring both locks, in case the reader opened up
 903                  * some space.
 904                  */
 905                 space = wpipe->pipe_buffer.size -
 906                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 907                 cpu_lfence();
 908                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 909                         space = 0;
 910
 911                 /*
 912                  * Retest EOF - acquiring a new token can temporarily release
 913                  * tokens already held.
 914                  */
 915                 if (wpipe->pipe_state & PIPE_WEOF) {
 916                         lwkt_reltoken(&wpipe->pipe_rlock);
 917                         error = EPIPE;
 918                         break;
 919                 }
 920
 921                 /*
 922                  * We have no more space and have something to offer,
 923                  * wake up select/poll.
 924                  */
 925                 if (space == 0) {
 926                         wpipe->pipe_state |= PIPE_WANTW;
 927                         ++wpipe->pipe_wantwcnt;
 928                         pipeselwakeup(wpipe);
 929                         if (wpipe->pipe_state & PIPE_WANTW)
 930                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 931                         ++pipe_wblocked_count;
 932                 }
 933                 lwkt_reltoken(&wpipe->pipe_rlock);
 934
 935                 /*
 936                  * Break out if we errored or the read side wants us to go
 937                  * away.
 938                  */
 939                 if (error)
 940                         break;
 941                 if (wpipe->pipe_state & PIPE_WEOF) {
 942                         error = EPIPE;
 943                         break;
 944                 }
 945         }
 946         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 947
 948         /*
 949          * If we have put any characters in the buffer, we wake up
 950          * the reader.
 951          *
 952          * Both rlock and wlock are required to be able to modify pipe_state.
 953          */
 954         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 955                 if (wpipe->pipe_state & PIPE_WANTR) {
 956                         lwkt_gettoken(&wpipe->pipe_rlock);
 957                         if (wpipe->pipe_state & PIPE_WANTR) {
 958                                 wpipe->pipe_state &= ~PIPE_WANTR;
 959                                 lwkt_reltoken(&wpipe->pipe_rlock);
 960                                 wakeup(wpipe);
 961                         } else {
 962                                 lwkt_reltoken(&wpipe->pipe_rlock);
 963                         }
 964                 }
 965                 if (pipeseltest(wpipe)) {
 966                         lwkt_gettoken(&wpipe->pipe_rlock);
 967                         pipeselwakeup(wpipe);
 968                         lwkt_reltoken(&wpipe->pipe_rlock);
 969                 }
 970         }
 971
 972         /*
 973          * Don't return EPIPE if I/O was successful
 974          */
 975         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 976             (uio->uio_resid == 0) &&
 977             (error == EPIPE)) {
 978                 error = 0;
 979         }
 980
 981         if (error == 0)
 982                 vfs_timestamp(&wpipe->pipe_mtime);
 983
 984         /*
 985          * We have something to offer,
 986          * wake up select/poll.
 987          */
 988         /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
 989         lwkt_reltoken(&wpipe->pipe_wlock);
 990         pipe_rel_mplock(&mpsave);
 991         return (error);
 992 }
 993
 994 /*
 995  * MPALMOSTSAFE - acquires mplock
 996  *
 997  * we implement a very minimal set of ioctls for compatibility with sockets.
 998  */
 999 int
1000 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
1001            struct ucred *cred, struct sysmsg *msg)
1002 {
1003         struct pipe *mpipe;
1004         int error;
1005         int mpsave;
1006
1007         pipe_get_mplock(&mpsave);
1008         mpipe = (struct pipe *)fp->f_data;
1009
1010         lwkt_gettoken(&mpipe->pipe_rlock);
1011         lwkt_gettoken(&mpipe->pipe_wlock);
1012
1013         switch (cmd) {
1014         case FIOASYNC:
1015                 if (*(int *)data) {
1016                         mpipe->pipe_state |= PIPE_ASYNC;
1017                 } else {
1018                         mpipe->pipe_state &= ~PIPE_ASYNC;
1019                 }
1020                 error = 0;
1021                 break;
1022         case FIONREAD:
1023                 *(int *)data = mpipe->pipe_buffer.windex -
1024                                 mpipe->pipe_buffer.rindex;
1025                 error = 0;
1026                 break;
1027         case FIOSETOWN:
1028                 get_mplock();
1029                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1030                 rel_mplock();
1031                 break;
1032         case FIOGETOWN:
1033                 *(int *)data = fgetown(mpipe->pipe_sigio);
1034                 error = 0;
1035                 break;
1036         case TIOCSPGRP:
1037                 /* This is deprecated, FIOSETOWN should be used instead. */
1038                 get_mplock();
1039                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1040                 rel_mplock();
1041                 break;
1042
1043         case TIOCGPGRP:
1044                 /* This is deprecated, FIOGETOWN should be used instead. */
1045                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1046                 error = 0;
1047                 break;
1048         default:
1049                 error = ENOTTY;
1050                 break;
1051         }
1052         lwkt_reltoken(&mpipe->pipe_wlock);
1053         lwkt_reltoken(&mpipe->pipe_rlock);
1054         pipe_rel_mplock(&mpsave);
1055
1056         return (error);
1057 }
1058
1059 /*
1060  * MPALMOSTSAFE - acquires mplock
1061  *
1062  * poll for events (helper)
1063  */
1064 static int
1065 pipe_poll_events(struct pipe *rpipe, struct pipe *wpipe, int events)
1066 {
1067         int revents = 0;
1068         u_int space;
1069
1070         if (events & (POLLIN | POLLRDNORM)) {
1071                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
1072                     (rpipe->pipe_state & PIPE_REOF)) {
1073                         revents |= events & (POLLIN | POLLRDNORM);
1074                 }
1075         }
1076
1077         if (events & (POLLOUT | POLLWRNORM)) {
1078                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
1079                         revents |= events & (POLLOUT | POLLWRNORM);
1080                 } else {
1081                         space = wpipe->pipe_buffer.windex -
1082                                 wpipe->pipe_buffer.rindex;
1083                         space = wpipe->pipe_buffer.size - space;
1084                         if (space >= PIPE_BUF)
1085                                 revents |= events & (POLLOUT | POLLWRNORM);
1086                 }
1087         }
1088
1089         if ((rpipe->pipe_state & PIPE_REOF) ||
1090             (wpipe == NULL) ||
1091             (wpipe->pipe_state & PIPE_WEOF)) {
1092                 revents |= POLLHUP;
1093         }
1094         return (revents);
1095 }
1096
1097 /*
1098  * Poll for events from file pointer.
1099  */
1100 int
1101 pipe_poll(struct file *fp, int events, struct ucred *cred)
1102 {
1103         struct pipe *rpipe;
1104         struct pipe *wpipe;
1105         int revents = 0;
1106         int mpsave;
1107
1108         pipe_get_mplock(&mpsave);
1109         rpipe = (struct pipe *)fp->f_data;
1110         wpipe = rpipe->pipe_peer;
1111
1112         revents = pipe_poll_events(rpipe, wpipe, events);
1113         if (revents == 0) {
1114                 if (events & (POLLIN | POLLRDNORM)) {
1115                         lwkt_gettoken(&rpipe->pipe_rlock);
1116                         lwkt_gettoken(&rpipe->pipe_wlock);
1117                 }
1118                 if (events & (POLLOUT | POLLWRNORM)) {
1119                         lwkt_gettoken(&wpipe->pipe_rlock);
1120                         lwkt_gettoken(&wpipe->pipe_wlock);
1121                 }
1122                 revents = pipe_poll_events(rpipe, wpipe, events);
1123                 if (revents == 0) {
1124                         if (events & (POLLIN | POLLRDNORM)) {
1125                                 selrecord(curthread, &rpipe->pipe_sel);
1126                                 rpipe->pipe_state |= PIPE_SEL;
1127                         }
1128
1129                         if (events & (POLLOUT | POLLWRNORM)) {
1130                                 selrecord(curthread, &wpipe->pipe_sel);
1131                                 wpipe->pipe_state |= PIPE_SEL;
1132                         }
1133                 }
1134                 if (events & (POLLOUT | POLLWRNORM)) {
1135                         lwkt_reltoken(&wpipe->pipe_wlock);
1136                         lwkt_reltoken(&wpipe->pipe_rlock);
1137                 }
1138                 if (events & (POLLIN | POLLRDNORM)) {
1139                         lwkt_reltoken(&rpipe->pipe_wlock);
1140                         lwkt_reltoken(&rpipe->pipe_rlock);
1141                 }
1142         }
1143         pipe_rel_mplock(&mpsave);
1144         return (revents);
1145 }
1146
1147 /*
1148  * MPSAFE
1149  */
1150 static int
1151 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1152 {
1153         struct pipe *pipe;
1154         int mpsave;
1155
1156         pipe_get_mplock(&mpsave);
1157         pipe = (struct pipe *)fp->f_data;
1158
1159         bzero((caddr_t)ub, sizeof(*ub));
1160         ub->st_mode = S_IFIFO;
1161         ub->st_blksize = pipe->pipe_buffer.size;
1162         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1163         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1164         ub->st_atimespec = pipe->pipe_atime;
1165         ub->st_mtimespec = pipe->pipe_mtime;
1166         ub->st_ctimespec = pipe->pipe_ctime;
1167         /*
1168          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1169          * st_flags, st_gen.
1170          * XXX (st_dev, st_ino) should be unique.
1171          */
1172         pipe_rel_mplock(&mpsave);
1173         return (0);
1174 }
1175
1176 /*
1177  * MPALMOSTSAFE - acquires mplock
1178  */
1179 static int
1180 pipe_close(struct file *fp)
1181 {
1182         struct pipe *cpipe;
1183
1184         get_mplock();
1185         cpipe = (struct pipe *)fp->f_data;
1186         fp->f_ops = &badfileops;
1187         fp->f_data = NULL;
1188         funsetown(cpipe->pipe_sigio);
1189         pipeclose(cpipe);
1190         rel_mplock();
1191         return (0);
1192 }
1193
1194 /*
1195  * Shutdown one or both directions of a full-duplex pipe.
1196  *
1197  * MPALMOSTSAFE - acquires mplock
1198  */
1199 static int
1200 pipe_shutdown(struct file *fp, int how)
1201 {
1202         struct pipe *rpipe;
1203         struct pipe *wpipe;
1204         int error = EPIPE;
1205         int mpsave;
1206
1207         pipe_get_mplock(&mpsave);
1208         rpipe = (struct pipe *)fp->f_data;
1209         wpipe = rpipe->pipe_peer;
1210
1211         /*
1212          * We modify pipe_state on both pipes, which means we need
1213          * all four tokens!
1214          */
1215         lwkt_gettoken(&rpipe->pipe_rlock);
1216         lwkt_gettoken(&rpipe->pipe_wlock);
1217         lwkt_gettoken(&wpipe->pipe_rlock);
1218         lwkt_gettoken(&wpipe->pipe_wlock);
1219
1220         switch(how) {
1221         case SHUT_RDWR:
1222         case SHUT_RD:
1223                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1224                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1225                 if (rpipe->pipe_state & PIPE_WANTR) {
1226                         rpipe->pipe_state &= ~PIPE_WANTR;
1227                         wakeup(rpipe);
1228                 }
1229                 if (rpipe->pipe_state & PIPE_WANTW) {
1230                         rpipe->pipe_state &= ~PIPE_WANTW;
1231                         wakeup(rpipe);
1232                 }
1233                 error = 0;
1234                 if (how == SHUT_RD)
1235                         break;
1236                 /* fall through */
1237         case SHUT_WR:
1238                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1239                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1240                 if (wpipe->pipe_state & PIPE_WANTR) {
1241                         wpipe->pipe_state &= ~PIPE_WANTR;
1242                         wakeup(wpipe);
1243                 }
1244                 if (wpipe->pipe_state & PIPE_WANTW) {
1245                         wpipe->pipe_state &= ~PIPE_WANTW;
1246                         wakeup(wpipe);
1247                 }
1248                 error = 0;
1249                 break;
1250         }
1251         pipeselwakeup(rpipe);
1252         pipeselwakeup(wpipe);
1253
1254         lwkt_reltoken(&wpipe->pipe_wlock);
1255         lwkt_reltoken(&wpipe->pipe_rlock);
1256         lwkt_reltoken(&rpipe->pipe_wlock);
1257         lwkt_reltoken(&rpipe->pipe_rlock);
1258
1259         pipe_rel_mplock(&mpsave);
1260         return (error);
1261 }
1262
1263 static void
1264 pipe_free_kmem(struct pipe *cpipe)
1265 {
1266         if (cpipe->pipe_buffer.buffer != NULL) {
1267                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1268                         atomic_subtract_int(&pipe_nbig, 1);
1269                 kmem_free(&kernel_map,
1270                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1271                         cpipe->pipe_buffer.size);
1272                 cpipe->pipe_buffer.buffer = NULL;
1273                 cpipe->pipe_buffer.object = NULL;
1274         }
1275 }
1276
1277 /*
1278  * Close the pipe.  The slock must be held to interlock against simultanious
1279  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1280  */
1281 static void
1282 pipeclose(struct pipe *cpipe)
1283 {
1284         globaldata_t gd;
1285         struct pipe *ppipe;
1286
1287         if (cpipe == NULL)
1288                 return;
1289
1290         /*
1291          * The slock may not have been allocated yet (close during
1292          * initialization)
1293          *
1294          * We need both the read and write tokens to modify pipe_state.
1295          */
1296         if (cpipe->pipe_slock)
1297                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1298         lwkt_gettoken(&cpipe->pipe_rlock);
1299         lwkt_gettoken(&cpipe->pipe_wlock);
1300
1301         /*
1302          * Set our state, wakeup anyone waiting in select, and
1303          * wakeup anyone blocked on our pipe.
1304          */
1305         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1306         pipeselwakeup(cpipe);
1307         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1308                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1309                 wakeup(cpipe);
1310         }
1311
1312         /*
1313          * Disconnect from peer.
1314          */
1315         if ((ppipe = cpipe->pipe_peer) != NULL) {
1316                 lwkt_gettoken(&ppipe->pipe_rlock);
1317                 lwkt_gettoken(&ppipe->pipe_wlock);
1318                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1319                 pipeselwakeup(ppipe);
1320                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1321                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1322                         wakeup(ppipe);
1323                 }
1324                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1325                         get_mplock();
1326                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1327                         rel_mplock();
1328                 }
1329                 lwkt_reltoken(&ppipe->pipe_wlock);
1330                 lwkt_reltoken(&ppipe->pipe_rlock);
1331         }
1332
1333         /*
1334          * If the peer is also closed we can free resources for both
1335          * sides, otherwise we leave our side intact to deal with any
1336          * races (since we only have the slock).
1337          */
1338         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1339                 cpipe->pipe_peer = NULL;
1340                 ppipe->pipe_peer = NULL;
1341                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1342                 pipeclose(ppipe);
1343                 ppipe = NULL;
1344         }
1345
1346         lwkt_reltoken(&cpipe->pipe_wlock);
1347         lwkt_reltoken(&cpipe->pipe_rlock);
1348         if (cpipe->pipe_slock)
1349                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1350
1351         /*
1352          * If we disassociated from our peer we can free resources
1353          */
1354         if (ppipe == NULL) {
1355                 gd = mycpu;
1356                 if (cpipe->pipe_slock) {
1357                         kfree(cpipe->pipe_slock, M_PIPE);
1358                         cpipe->pipe_slock = NULL;
1359                 }
1360                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1361                     cpipe->pipe_buffer.size != PIPE_SIZE
1362                 ) {
1363                         pipe_free_kmem(cpipe);
1364                         kfree(cpipe, M_PIPE);
1365                 } else {
1366                         cpipe->pipe_state = 0;
1367                         cpipe->pipe_peer = gd->gd_pipeq;
1368                         gd->gd_pipeq = cpipe;
1369                         ++gd->gd_pipeqcount;
1370                 }
1371         }
1372 }
1373
1374 /*
1375  * MPALMOSTSAFE - acquires mplock
1376  */
1377 static int
1378 pipe_kqfilter(struct file *fp, struct knote *kn)
1379 {
1380         struct pipe *cpipe;
1381
1382         get_mplock();
1383         cpipe = (struct pipe *)kn->kn_fp->f_data;
1384
1385         switch (kn->kn_filter) {
1386         case EVFILT_READ:
1387                 kn->kn_fop = &pipe_rfiltops;
1388                 break;
1389         case EVFILT_WRITE:
1390                 kn->kn_fop = &pipe_wfiltops;
1391                 cpipe = cpipe->pipe_peer;
1392                 if (cpipe == NULL) {
1393                         /* other end of pipe has been closed */
1394                         rel_mplock();
1395                         return (EPIPE);
1396                 }
1397                 break;
1398         default:
1399                 rel_mplock();
1400                 return (EPERM);
1401         }
1402         kn->kn_hook = (caddr_t)cpipe;
1403
1404         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1405         rel_mplock();
1406         return (0);
1407 }
1408
1409 static void
1410 filt_pipedetach(struct knote *kn)
1411 {
1412         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1413
1414         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1415 }
1416
1417 /*ARGSUSED*/
1418 static int
1419 filt_piperead(struct knote *kn, long hint)
1420 {
1421         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1422
1423         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1424
1425         /* XXX RACE */
1426         if (rpipe->pipe_state & PIPE_REOF) {
1427                 kn->kn_flags |= EV_EOF;
1428                 return (1);
1429         }
1430         return (kn->kn_data > 0);
1431 }
1432
1433 /*ARGSUSED*/
1434 static int
1435 filt_pipewrite(struct knote *kn, long hint)
1436 {
1437         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1438         struct pipe *wpipe = rpipe->pipe_peer;
1439         u_int32_t space;
1440
1441         /* XXX RACE */
1442         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1443                 kn->kn_data = 0;
1444                 kn->kn_flags |= EV_EOF;
1445                 return (1);
1446         }
1447         space = wpipe->pipe_buffer.windex -
1448                 wpipe->pipe_buffer.rindex;
1449         space = wpipe->pipe_buffer.size - space;
1450         kn->kn_data = space;
1451         return (kn->kn_data >= PIPE_BUF);
1452 }