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, size,
 357                                     1,
 358                                     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);
 408         lwkt_token_init(&cpipe->pipe_wlock);
 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         lwkt_tokref rlock;
 427         lwkt_tokref wlock;
 428         int mpsave;
 429         int bigread;
 430         int bigcount;
 431
 432         if (uio->uio_resid == 0)
 433                 return(0);
 434
 435         /*
 436          * Setup locks, calculate nbio
 437          */
 438         pipe_get_mplock(&mpsave);
 439         rpipe = (struct pipe *)fp->f_data;
 440         lwkt_gettoken(&rlock, &rpipe->pipe_rlock);
 441
 442         if (fflags & O_FBLOCKING)
 443                 nbio = 0;
 444         else if (fflags & O_FNONBLOCKING)
 445                 nbio = 1;
 446         else if (fp->f_flag & O_NONBLOCK)
 447                 nbio = 1;
 448         else
 449                 nbio = 0;
 450
 451         /*
 452          * Reads are serialized.  Note howeverthat pipe_buffer.buffer and
 453          * pipe_buffer.size can change out from under us when the number
 454          * of bytes in the buffer are zero due to the write-side doing a
 455          * pipespace().
 456          */
 457         error = pipe_start_uio(rpipe, &rpipe->pipe_rip);
 458         if (error) {
 459                 pipe_rel_mplock(&mpsave);
 460                 lwkt_reltoken(&rlock);
 461                 return (error);
 462         }
 463         notify_writer = 0;
 464
 465         bigread = (uio->uio_resid > 10 * 1024 * 1024);
 466         bigcount = 10;
 467
 468         while (uio->uio_resid) {
 469                 /*
 470                  * Don't hog the cpu.
 471                  */
 472                 if (bigread && --bigcount == 0) {
 473                         lwkt_user_yield();
 474                         bigcount = 10;
 475                         if (CURSIG(curthread->td_lwp)) {
 476                                 error = EINTR;
 477                                 break;
 478                         }
 479                 }
 480
 481                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 482                 cpu_lfence();
 483                 if (size) {
 484                         rindex = rpipe->pipe_buffer.rindex &
 485                                  (rpipe->pipe_buffer.size - 1);
 486                         nsize = size;
 487                         if (nsize > rpipe->pipe_buffer.size - rindex)
 488                                 nsize = rpipe->pipe_buffer.size - rindex;
 489                         nsize = szmin(nsize, uio->uio_resid);
 490
 491                         error = uiomove(&rpipe->pipe_buffer.buffer[rindex],
 492                                         nsize, uio);
 493                         if (error)
 494                                 break;
 495                         cpu_mfence();
 496                         rpipe->pipe_buffer.rindex += nsize;
 497                         nread += nsize;
 498
 499                         /*
 500                          * If the FIFO is still over half full just continue
 501                          * and do not try to notify the writer yet.
 502                          */
 503                         if (size - nsize >= (rpipe->pipe_buffer.size >> 1)) {
 504                                 notify_writer = 0;
 505                                 continue;
 506                         }
 507
 508                         /*
 509                          * When the FIFO is less then half full notify any
 510                          * waiting writer.  WANTW can be checked while
 511                          * holding just the rlock.
 512                          */
 513                         notify_writer = 1;
 514                         if ((rpipe->pipe_state & PIPE_WANTW) == 0)
 515                                 continue;
 516                 }
 517
 518                 /*
 519                  * If the "write-side" was blocked we wake it up.  This code
 520                  * is reached either when the buffer is completely emptied
 521                  * or if it becomes more then half-empty.
 522                  *
 523                  * Pipe_state can only be modified if both the rlock and
 524                  * wlock are held.
 525                  */
 526                 if (rpipe->pipe_state & PIPE_WANTW) {
 527                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 528                         if (rpipe->pipe_state & PIPE_WANTW) {
 529                                 notify_writer = 0;
 530                                 rpipe->pipe_state &= ~PIPE_WANTW;
 531                                 lwkt_reltoken(&wlock);
 532                                 wakeup(rpipe);
 533                         } else {
 534                                 lwkt_reltoken(&wlock);
 535                         }
 536                 }
 537
 538                 /*
 539                  * Pick up our copy loop again if the writer sent data to
 540                  * us while we were messing around.
 541                  *
 542                  * On a SMP box poll up to pipe_delay nanoseconds for new
 543                  * data.  Typically a value of 2000 to 4000 is sufficient
 544                  * to eradicate most IPIs/tsleeps/wakeups when a pipe
 545                  * is used for synchronous communications with small packets,
 546                  * and 8000 or so (8uS) will pipeline large buffer xfers
 547                  * between cpus over a pipe.
 548                  *
 549                  * For synchronous communications a hit means doing a
 550                  * full Awrite-Bread-Bwrite-Aread cycle in less then 2uS,
 551                  * where as miss requiring a tsleep/wakeup sequence
 552                  * will take 7uS or more.
 553                  */
 554                 if (rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex)
 555                         continue;
 556
 557 #if defined(SMP) && defined(_RDTSC_SUPPORTED_)
 558                 if (pipe_delay) {
 559                         int64_t tsc_target;
 560                         int good = 0;
 561
 562                         tsc_target = tsc_get_target(pipe_delay);
 563                         while (tsc_test_target(tsc_target) == 0) {
 564                                 if (rpipe->pipe_buffer.windex !=
 565                                     rpipe->pipe_buffer.rindex) {
 566                                         good = 1;
 567                                         break;
 568                                 }
 569                         }
 570                         if (good)
 571                                 continue;
 572                 }
 573 #endif
 574
 575                 /*
 576                  * Detect EOF condition, do not set error.
 577                  */
 578                 if (rpipe->pipe_state & PIPE_REOF)
 579                         break;
 580
 581                 /*
 582                  * Break if some data was read, or if this was a non-blocking
 583                  * read.
 584                  */
 585                 if (nread > 0)
 586                         break;
 587
 588                 if (nbio) {
 589                         error = EAGAIN;
 590                         break;
 591                 }
 592
 593                 /*
 594                  * Last chance, interlock with WANTR.
 595                  */
 596                 lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 597                 size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
 598                 if (size) {
 599                         lwkt_reltoken(&wlock);
 600                         continue;
 601                 }
 602
 603                 /*
 604                  * Retest EOF - acquiring a new token can temporarily release
 605                  * tokens already held.
 606                  */
 607                 if (rpipe->pipe_state & PIPE_REOF) {
 608                         lwkt_reltoken(&wlock);
 609                         break;
 610                 }
 611
 612                 /*
 613                  * If there is no more to read in the pipe, reset its
 614                  * pointers to the beginning.  This improves cache hit
 615                  * stats.
 616                  *
 617                  * We need both locks to modify both pointers, and there
 618                  * must also not be a write in progress or the uiomove()
 619                  * in the write might block and temporarily release
 620                  * its wlock, then reacquire and update windex.  We are
 621                  * only serialized against reads, not writes.
 622                  *
 623                  * XXX should we even bother resetting the indices?  It
 624                  *     might actually be more cache efficient not to.
 625                  */
 626                 if (rpipe->pipe_buffer.rindex == rpipe->pipe_buffer.windex &&
 627                     rpipe->pipe_wip == 0) {
 628                         rpipe->pipe_buffer.rindex = 0;
 629                         rpipe->pipe_buffer.windex = 0;
 630                 }
 631
 632                 /*
 633                  * Wait for more data.
 634                  *
 635                  * Pipe_state can only be set if both the rlock and wlock
 636                  * are held.
 637                  */
 638                 rpipe->pipe_state |= PIPE_WANTR;
 639                 tsleep_interlock(rpipe, PCATCH);
 640                 lwkt_reltoken(&wlock);
 641                 error = tsleep(rpipe, PCATCH | PINTERLOCKED, "piperd", 0);
 642                 ++pipe_rblocked_count;
 643                 if (error)
 644                         break;
 645         }
 646         pipe_end_uio(rpipe, &rpipe->pipe_rip);
 647
 648         /*
 649          * Uptime last access time
 650          */
 651         if (error == 0 && nread)
 652                 vfs_timestamp(&rpipe->pipe_atime);
 653
 654         /*
 655          * If we drained the FIFO more then half way then handle
 656          * write blocking hysteresis.
 657          *
 658          * Note that PIPE_WANTW cannot be set by the writer without
 659          * it holding both rlock and wlock, so we can test it
 660          * while holding just rlock.
 661          */
 662         if (notify_writer) {
 663                 if (rpipe->pipe_state & PIPE_WANTW) {
 664                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 665                         if (rpipe->pipe_state & PIPE_WANTW) {
 666                                 rpipe->pipe_state &= ~PIPE_WANTW;
 667                                 lwkt_reltoken(&wlock);
 668                                 wakeup(rpipe);
 669                         } else {
 670                                 lwkt_reltoken(&wlock);
 671                         }
 672                 }
 673                 if (pipeseltest(rpipe)) {
 674                         lwkt_gettoken(&wlock, &rpipe->pipe_wlock);
 675                         pipeselwakeup(rpipe);
 676                         lwkt_reltoken(&wlock);
 677                 }
 678         }
 679         /*size = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;*/
 680         lwkt_reltoken(&rlock);
 681
 682         pipe_rel_mplock(&mpsave);
 683         return (error);
 684 }
 685
 686 /*
 687  * MPALMOSTSAFE - acquires mplock
 688  */
 689 static int
 690 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 691 {
 692         int error;
 693         int orig_resid;
 694         int nbio;
 695         struct pipe *wpipe, *rpipe;
 696         lwkt_tokref rlock;
 697         lwkt_tokref wlock;
 698         u_int windex;
 699         u_int space;
 700         u_int wcount;
 701         int mpsave;
 702         int bigwrite;
 703         int bigcount;
 704
 705         pipe_get_mplock(&mpsave);
 706
 707         /*
 708          * Writes go to the peer.  The peer will always exist.
 709          */
 710         rpipe = (struct pipe *) fp->f_data;
 711         wpipe = rpipe->pipe_peer;
 712         lwkt_gettoken(&wlock, &wpipe->pipe_wlock);
 713         if (wpipe->pipe_state & PIPE_WEOF) {
 714                 pipe_rel_mplock(&mpsave);
 715                 lwkt_reltoken(&wlock);
 716                 return (EPIPE);
 717         }
 718
 719         /*
 720          * Degenerate case (EPIPE takes prec)
 721          */
 722         if (uio->uio_resid == 0) {
 723                 pipe_rel_mplock(&mpsave);
 724                 lwkt_reltoken(&wlock);
 725                 return(0);
 726         }
 727
 728         /*
 729          * Writes are serialized (start_uio must be called with wlock)
 730          */
 731         error = pipe_start_uio(wpipe, &wpipe->pipe_wip);
 732         if (error) {
 733                 pipe_rel_mplock(&mpsave);
 734                 lwkt_reltoken(&wlock);
 735                 return (error);
 736         }
 737
 738         if (fflags & O_FBLOCKING)
 739                 nbio = 0;
 740         else if (fflags & O_FNONBLOCKING)
 741                 nbio = 1;
 742         else if (fp->f_flag & O_NONBLOCK)
 743                 nbio = 1;
 744         else
 745                 nbio = 0;
 746
 747         /*
 748          * If it is advantageous to resize the pipe buffer, do
 749          * so.  We are write-serialized so we can block safely.
 750          */
 751         if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 752             (pipe_nbig < pipe_maxbig) &&
 753             wpipe->pipe_wantwcnt > 4 &&
 754             (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 755                 /*
 756                  * Recheck after lock.
 757                  */
 758                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 759                 if ((wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 760                     (pipe_nbig < pipe_maxbig) &&
 761                     (wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex)) {
 762                         atomic_add_int(&pipe_nbig, 1);
 763                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 764                                 ++pipe_bigcount;
 765                         else
 766                                 atomic_subtract_int(&pipe_nbig, 1);
 767                 }
 768                 lwkt_reltoken(&rlock);
 769         }
 770
 771         orig_resid = uio->uio_resid;
 772         wcount = 0;
 773
 774         bigwrite = (uio->uio_resid > 10 * 1024 * 1024);
 775         bigcount = 10;
 776
 777         while (uio->uio_resid) {
 778                 if (wpipe->pipe_state & PIPE_WEOF) {
 779                         error = EPIPE;
 780                         break;
 781                 }
 782
 783                 /*
 784                  * Don't hog the cpu.
 785                  */
 786                 if (bigwrite && --bigcount == 0) {
 787                         lwkt_user_yield();
 788                         bigcount = 10;
 789                         if (CURSIG(curthread->td_lwp)) {
 790                                 error = EINTR;
 791                                 break;
 792                         }
 793                 }
 794
 795                 windex = wpipe->pipe_buffer.windex &
 796                          (wpipe->pipe_buffer.size - 1);
 797                 space = wpipe->pipe_buffer.size -
 798                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 799                 cpu_lfence();
 800
 801                 /* Writes of size <= PIPE_BUF must be atomic. */
 802                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 803                         space = 0;
 804
 805                 /*
 806                  * Write to fill, read size handles write hysteresis.  Also
 807                  * additional restrictions can cause select-based non-blocking
 808                  * writes to spin.
 809                  */
 810                 if (space > 0) {
 811                         u_int segsize;
 812
 813                         /*
 814                          * Transfer size is minimum of uio transfer
 815                          * and free space in pipe buffer.
 816                          *
 817                          * Limit each uiocopy to no more then PIPE_SIZE
 818                          * so we can keep the gravy train going on a
 819                          * SMP box.  This doubles the performance for
 820                          * write sizes > 16K.  Otherwise large writes
 821                          * wind up doing an inefficient synchronous
 822                          * ping-pong.
 823                          */
 824                         space = szmin(space, uio->uio_resid);
 825                         if (space > PIPE_SIZE)
 826                                 space = PIPE_SIZE;
 827
 828                         /*
 829                          * First segment to transfer is minimum of
 830                          * transfer size and contiguous space in
 831                          * pipe buffer.  If first segment to transfer
 832                          * is less than the transfer size, we've got
 833                          * a wraparound in the buffer.
 834                          */
 835                         segsize = wpipe->pipe_buffer.size - windex;
 836                         if (segsize > space)
 837                                 segsize = space;
 838
 839 #ifdef SMP
 840                         /*
 841                          * If this is the first loop and the reader is
 842                          * blocked, do a preemptive wakeup of the reader.
 843                          *
 844                          * On SMP the IPI latency plus the wlock interlock
 845                          * on the reader side is the fastest way to get the
 846                          * reader going.  (The scheduler will hard loop on
 847                          * lock tokens).
 848                          *
 849                          * NOTE: We can't clear WANTR here without acquiring
 850                          * the rlock, which we don't want to do here!
 851                          */
 852                         if ((wpipe->pipe_state & PIPE_WANTR) && pipe_mpsafe > 1)
 853                                 wakeup(wpipe);
 854 #endif
 855
 856                         /*
 857                          * Transfer segment, which may include a wrap-around.
 858                          * Update windex to account for both all in one go
 859                          * so the reader can read() the data atomically.
 860                          */
 861                         error = uiomove(&wpipe->pipe_buffer.buffer[windex],
 862                                         segsize, uio);
 863                         if (error == 0 && segsize < space) {
 864                                 segsize = space - segsize;
 865                                 error = uiomove(&wpipe->pipe_buffer.buffer[0],
 866                                                 segsize, uio);
 867                         }
 868                         if (error)
 869                                 break;
 870                         cpu_mfence();
 871                         wpipe->pipe_buffer.windex += space;
 872                         wcount += space;
 873                         continue;
 874                 }
 875
 876                 /*
 877                  * We need both the rlock and the wlock to interlock against
 878                  * the EOF, WANTW, and size checks, and to modify pipe_state.
 879                  *
 880                  * These are token locks so we do not have to worry about
 881                  * deadlocks.
 882                  */
 883                 lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 884
 885                 /*
 886                  * If the "read-side" has been blocked, wake it up now
 887                  * and yield to let it drain synchronously rather
 888                  * then block.
 889                  */
 890                 if (wpipe->pipe_state & PIPE_WANTR) {
 891                         wpipe->pipe_state &= ~PIPE_WANTR;
 892                         wakeup(wpipe);
 893                 }
 894
 895                 /*
 896                  * don't block on non-blocking I/O
 897                  */
 898                 if (nbio) {
 899                         lwkt_reltoken(&rlock);
 900                         error = EAGAIN;
 901                         break;
 902                 }
 903
 904                 /*
 905                  * re-test whether we have to block in the writer after
 906                  * acquiring both locks, in case the reader opened up
 907                  * some space.
 908                  */
 909                 space = wpipe->pipe_buffer.size -
 910                         (wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex);
 911                 cpu_lfence();
 912                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 913                         space = 0;
 914
 915                 /*
 916                  * Retest EOF - acquiring a new token can temporarily release
 917                  * tokens already held.
 918                  */
 919                 if (wpipe->pipe_state & PIPE_WEOF) {
 920                         lwkt_reltoken(&rlock);
 921                         error = EPIPE;
 922                         break;
 923                 }
 924
 925                 /*
 926                  * We have no more space and have something to offer,
 927                  * wake up select/poll.
 928                  */
 929                 if (space == 0) {
 930                         wpipe->pipe_state |= PIPE_WANTW;
 931                         ++wpipe->pipe_wantwcnt;
 932                         pipeselwakeup(wpipe);
 933                         if (wpipe->pipe_state & PIPE_WANTW)
 934                                 error = tsleep(wpipe, PCATCH, "pipewr", 0);
 935                         ++pipe_wblocked_count;
 936                 }
 937                 lwkt_reltoken(&rlock);
 938
 939                 /*
 940                  * Break out if we errored or the read side wants us to go
 941                  * away.
 942                  */
 943                 if (error)
 944                         break;
 945                 if (wpipe->pipe_state & PIPE_WEOF) {
 946                         error = EPIPE;
 947                         break;
 948                 }
 949         }
 950         pipe_end_uio(wpipe, &wpipe->pipe_wip);
 951
 952         /*
 953          * If we have put any characters in the buffer, we wake up
 954          * the reader.
 955          *
 956          * Both rlock and wlock are required to be able to modify pipe_state.
 957          */
 958         if (wpipe->pipe_buffer.windex != wpipe->pipe_buffer.rindex) {
 959                 if (wpipe->pipe_state & PIPE_WANTR) {
 960                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 961                         if (wpipe->pipe_state & PIPE_WANTR) {
 962                                 wpipe->pipe_state &= ~PIPE_WANTR;
 963                                 lwkt_reltoken(&rlock);
 964                                 wakeup(wpipe);
 965                         } else {
 966                                 lwkt_reltoken(&rlock);
 967                         }
 968                 }
 969                 if (pipeseltest(wpipe)) {
 970                         lwkt_gettoken(&rlock, &wpipe->pipe_rlock);
 971                         pipeselwakeup(wpipe);
 972                         lwkt_reltoken(&rlock);
 973                 }
 974         }
 975
 976         /*
 977          * Don't return EPIPE if I/O was successful
 978          */
 979         if ((wpipe->pipe_buffer.rindex == wpipe->pipe_buffer.windex) &&
 980             (uio->uio_resid == 0) &&
 981             (error == EPIPE)) {
 982                 error = 0;
 983         }
 984
 985         if (error == 0)
 986                 vfs_timestamp(&wpipe->pipe_mtime);
 987
 988         /*
 989          * We have something to offer,
 990          * wake up select/poll.
 991          */
 992         /*space = wpipe->pipe_buffer.windex - wpipe->pipe_buffer.rindex;*/
 993         lwkt_reltoken(&wlock);
 994         pipe_rel_mplock(&mpsave);
 995         return (error);
 996 }
 997
 998 /*
 999  * MPALMOSTSAFE - acquires mplock
1000  *
1001  * we implement a very minimal set of ioctls for compatibility with sockets.
1002  */
1003 int
1004 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data,
1005            struct ucred *cred, struct sysmsg *msg)
1006 {
1007         struct pipe *mpipe;
1008         lwkt_tokref rlock;
1009         lwkt_tokref wlock;
1010         int error;
1011         int mpsave;
1012
1013         pipe_get_mplock(&mpsave);
1014         mpipe = (struct pipe *)fp->f_data;
1015
1016         lwkt_gettoken(&rlock, &mpipe->pipe_rlock);
1017         lwkt_gettoken(&wlock, &mpipe->pipe_wlock);
1018
1019         switch (cmd) {
1020         case FIOASYNC:
1021                 if (*(int *)data) {
1022                         mpipe->pipe_state |= PIPE_ASYNC;
1023                 } else {
1024                         mpipe->pipe_state &= ~PIPE_ASYNC;
1025                 }
1026                 error = 0;
1027                 break;
1028         case FIONREAD:
1029                 *(int *)data = mpipe->pipe_buffer.windex -
1030                                 mpipe->pipe_buffer.rindex;
1031                 error = 0;
1032                 break;
1033         case FIOSETOWN:
1034                 get_mplock();
1035                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1036                 rel_mplock();
1037                 break;
1038         case FIOGETOWN:
1039                 *(int *)data = fgetown(mpipe->pipe_sigio);
1040                 error = 0;
1041                 break;
1042         case TIOCSPGRP:
1043                 /* This is deprecated, FIOSETOWN should be used instead. */
1044                 get_mplock();
1045                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1046                 rel_mplock();
1047                 break;
1048
1049         case TIOCGPGRP:
1050                 /* This is deprecated, FIOGETOWN should be used instead. */
1051                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1052                 error = 0;
1053                 break;
1054         default:
1055                 error = ENOTTY;
1056                 break;
1057         }
1058         lwkt_reltoken(&rlock);
1059         lwkt_reltoken(&wlock);
1060         pipe_rel_mplock(&mpsave);
1061
1062         return (error);
1063 }
1064
1065 /*
1066  * MPALMOSTSAFE - acquires mplock
1067  *
1068  * poll for events (helper)
1069  */
1070 static int
1071 pipe_poll_events(struct pipe *rpipe, struct pipe *wpipe, int events)
1072 {
1073         int revents = 0;
1074         u_int space;
1075
1076         if (events & (POLLIN | POLLRDNORM)) {
1077                 if ((rpipe->pipe_buffer.windex != rpipe->pipe_buffer.rindex) ||
1078                     (rpipe->pipe_state & PIPE_REOF)) {
1079                         revents |= events & (POLLIN | POLLRDNORM);
1080                 }
1081         }
1082
1083         if (events & (POLLOUT | POLLWRNORM)) {
1084                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_WEOF)) {
1085                         revents |= events & (POLLOUT | POLLWRNORM);
1086                 } else {
1087                         space = wpipe->pipe_buffer.windex -
1088                                 wpipe->pipe_buffer.rindex;
1089                         space = wpipe->pipe_buffer.size - space;
1090                         if (space >= PIPE_BUF)
1091                                 revents |= events & (POLLOUT | POLLWRNORM);
1092                 }
1093         }
1094
1095         if ((rpipe->pipe_state & PIPE_REOF) ||
1096             (wpipe == NULL) ||
1097             (wpipe->pipe_state & PIPE_WEOF)) {
1098                 revents |= POLLHUP;
1099         }
1100         return (revents);
1101 }
1102
1103 /*
1104  * Poll for events from file pointer.
1105  */
1106 int
1107 pipe_poll(struct file *fp, int events, struct ucred *cred)
1108 {
1109         lwkt_tokref rpipe_rlock;
1110         lwkt_tokref rpipe_wlock;
1111         lwkt_tokref wpipe_rlock;
1112         lwkt_tokref wpipe_wlock;
1113         struct pipe *rpipe;
1114         struct pipe *wpipe;
1115         int revents = 0;
1116         int mpsave;
1117
1118         pipe_get_mplock(&mpsave);
1119         rpipe = (struct pipe *)fp->f_data;
1120         wpipe = rpipe->pipe_peer;
1121
1122         revents = pipe_poll_events(rpipe, wpipe, events);
1123         if (revents == 0) {
1124                 if (events & (POLLIN | POLLRDNORM)) {
1125                         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1126                         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1127                 }
1128                 if (events & (POLLOUT | POLLWRNORM)) {
1129                         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1130                         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1131                 }
1132                 revents = pipe_poll_events(rpipe, wpipe, events);
1133                 if (revents == 0) {
1134                         if (events & (POLLIN | POLLRDNORM)) {
1135                                 selrecord(curthread, &rpipe->pipe_sel);
1136                                 rpipe->pipe_state |= PIPE_SEL;
1137                         }
1138
1139                         if (events & (POLLOUT | POLLWRNORM)) {
1140                                 selrecord(curthread, &wpipe->pipe_sel);
1141                                 wpipe->pipe_state |= PIPE_SEL;
1142                         }
1143                 }
1144                 if (events & (POLLIN | POLLRDNORM)) {
1145                         lwkt_reltoken(&rpipe_rlock);
1146                         lwkt_reltoken(&rpipe_wlock);
1147                 }
1148                 if (events & (POLLOUT | POLLWRNORM)) {
1149                         lwkt_reltoken(&wpipe_rlock);
1150                         lwkt_reltoken(&wpipe_wlock);
1151                 }
1152         }
1153         pipe_rel_mplock(&mpsave);
1154         return (revents);
1155 }
1156
1157 /*
1158  * MPSAFE
1159  */
1160 static int
1161 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1162 {
1163         struct pipe *pipe;
1164         int mpsave;
1165
1166         pipe_get_mplock(&mpsave);
1167         pipe = (struct pipe *)fp->f_data;
1168
1169         bzero((caddr_t)ub, sizeof(*ub));
1170         ub->st_mode = S_IFIFO;
1171         ub->st_blksize = pipe->pipe_buffer.size;
1172         ub->st_size = pipe->pipe_buffer.windex - pipe->pipe_buffer.rindex;
1173         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1174         ub->st_atimespec = pipe->pipe_atime;
1175         ub->st_mtimespec = pipe->pipe_mtime;
1176         ub->st_ctimespec = pipe->pipe_ctime;
1177         /*
1178          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1179          * st_flags, st_gen.
1180          * XXX (st_dev, st_ino) should be unique.
1181          */
1182         pipe_rel_mplock(&mpsave);
1183         return (0);
1184 }
1185
1186 /*
1187  * MPALMOSTSAFE - acquires mplock
1188  */
1189 static int
1190 pipe_close(struct file *fp)
1191 {
1192         struct pipe *cpipe;
1193
1194         get_mplock();
1195         cpipe = (struct pipe *)fp->f_data;
1196         fp->f_ops = &badfileops;
1197         fp->f_data = NULL;
1198         funsetown(cpipe->pipe_sigio);
1199         pipeclose(cpipe);
1200         rel_mplock();
1201         return (0);
1202 }
1203
1204 /*
1205  * Shutdown one or both directions of a full-duplex pipe.
1206  *
1207  * MPALMOSTSAFE - acquires mplock
1208  */
1209 static int
1210 pipe_shutdown(struct file *fp, int how)
1211 {
1212         struct pipe *rpipe;
1213         struct pipe *wpipe;
1214         int error = EPIPE;
1215         lwkt_tokref rpipe_rlock;
1216         lwkt_tokref rpipe_wlock;
1217         lwkt_tokref wpipe_rlock;
1218         lwkt_tokref wpipe_wlock;
1219         int mpsave;
1220
1221         pipe_get_mplock(&mpsave);
1222         rpipe = (struct pipe *)fp->f_data;
1223         wpipe = rpipe->pipe_peer;
1224
1225         /*
1226          * We modify pipe_state on both pipes, which means we need
1227          * all four tokens!
1228          */
1229         lwkt_gettoken(&rpipe_rlock, &rpipe->pipe_rlock);
1230         lwkt_gettoken(&rpipe_wlock, &rpipe->pipe_wlock);
1231         lwkt_gettoken(&wpipe_rlock, &wpipe->pipe_rlock);
1232         lwkt_gettoken(&wpipe_wlock, &wpipe->pipe_wlock);
1233
1234         switch(how) {
1235         case SHUT_RDWR:
1236         case SHUT_RD:
1237                 rpipe->pipe_state |= PIPE_REOF;         /* my reads */
1238                 rpipe->pipe_state |= PIPE_WEOF;         /* peer writes */
1239                 if (rpipe->pipe_state & PIPE_WANTR) {
1240                         rpipe->pipe_state &= ~PIPE_WANTR;
1241                         wakeup(rpipe);
1242                 }
1243                 if (rpipe->pipe_state & PIPE_WANTW) {
1244                         rpipe->pipe_state &= ~PIPE_WANTW;
1245                         wakeup(rpipe);
1246                 }
1247                 error = 0;
1248                 if (how == SHUT_RD)
1249                         break;
1250                 /* fall through */
1251         case SHUT_WR:
1252                 wpipe->pipe_state |= PIPE_REOF;         /* peer reads */
1253                 wpipe->pipe_state |= PIPE_WEOF;         /* my writes */
1254                 if (wpipe->pipe_state & PIPE_WANTR) {
1255                         wpipe->pipe_state &= ~PIPE_WANTR;
1256                         wakeup(wpipe);
1257                 }
1258                 if (wpipe->pipe_state & PIPE_WANTW) {
1259                         wpipe->pipe_state &= ~PIPE_WANTW;
1260                         wakeup(wpipe);
1261                 }
1262                 error = 0;
1263                 break;
1264         }
1265         pipeselwakeup(rpipe);
1266         pipeselwakeup(wpipe);
1267
1268         lwkt_reltoken(&rpipe_rlock);
1269         lwkt_reltoken(&rpipe_wlock);
1270         lwkt_reltoken(&wpipe_rlock);
1271         lwkt_reltoken(&wpipe_wlock);
1272
1273         pipe_rel_mplock(&mpsave);
1274         return (error);
1275 }
1276
1277 static void
1278 pipe_free_kmem(struct pipe *cpipe)
1279 {
1280         if (cpipe->pipe_buffer.buffer != NULL) {
1281                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1282                         atomic_subtract_int(&pipe_nbig, 1);
1283                 kmem_free(&kernel_map,
1284                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1285                         cpipe->pipe_buffer.size);
1286                 cpipe->pipe_buffer.buffer = NULL;
1287                 cpipe->pipe_buffer.object = NULL;
1288         }
1289 }
1290
1291 /*
1292  * Close the pipe.  The slock must be held to interlock against simultanious
1293  * closes.  The rlock and wlock must be held to adjust the pipe_state.
1294  */
1295 static void
1296 pipeclose(struct pipe *cpipe)
1297 {
1298         globaldata_t gd;
1299         struct pipe *ppipe;
1300         lwkt_tokref cpipe_rlock;
1301         lwkt_tokref cpipe_wlock;
1302         lwkt_tokref ppipe_rlock;
1303         lwkt_tokref ppipe_wlock;
1304
1305         if (cpipe == NULL)
1306                 return;
1307
1308         /*
1309          * The slock may not have been allocated yet (close during
1310          * initialization)
1311          *
1312          * We need both the read and write tokens to modify pipe_state.
1313          */
1314         if (cpipe->pipe_slock)
1315                 lockmgr(cpipe->pipe_slock, LK_EXCLUSIVE);
1316         lwkt_gettoken(&cpipe_rlock, &cpipe->pipe_rlock);
1317         lwkt_gettoken(&cpipe_wlock, &cpipe->pipe_wlock);
1318
1319         /*
1320          * Set our state, wakeup anyone waiting in select, and
1321          * wakeup anyone blocked on our pipe.
1322          */
1323         cpipe->pipe_state |= PIPE_CLOSED | PIPE_REOF | PIPE_WEOF;
1324         pipeselwakeup(cpipe);
1325         if (cpipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1326                 cpipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1327                 wakeup(cpipe);
1328         }
1329
1330         /*
1331          * Disconnect from peer.
1332          */
1333         if ((ppipe = cpipe->pipe_peer) != NULL) {
1334                 lwkt_gettoken(&ppipe_rlock, &ppipe->pipe_rlock);
1335                 lwkt_gettoken(&ppipe_wlock, &ppipe->pipe_wlock);
1336                 ppipe->pipe_state |= PIPE_REOF | PIPE_WEOF;
1337                 pipeselwakeup(ppipe);
1338                 if (ppipe->pipe_state & (PIPE_WANTR | PIPE_WANTW)) {
1339                         ppipe->pipe_state &= ~(PIPE_WANTR | PIPE_WANTW);
1340                         wakeup(ppipe);
1341                 }
1342                 if (SLIST_FIRST(&ppipe->pipe_sel.si_note)) {
1343                         get_mplock();
1344                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1345                         rel_mplock();
1346                 }
1347                 lwkt_reltoken(&ppipe_rlock);
1348                 lwkt_reltoken(&ppipe_wlock);
1349         }
1350
1351         /*
1352          * If the peer is also closed we can free resources for both
1353          * sides, otherwise we leave our side intact to deal with any
1354          * races (since we only have the slock).
1355          */
1356         if (ppipe && (ppipe->pipe_state & PIPE_CLOSED)) {
1357                 cpipe->pipe_peer = NULL;
1358                 ppipe->pipe_peer = NULL;
1359                 ppipe->pipe_slock = NULL;       /* we will free the slock */
1360                 pipeclose(ppipe);
1361                 ppipe = NULL;
1362         }
1363
1364         lwkt_reltoken(&cpipe_rlock);
1365         lwkt_reltoken(&cpipe_wlock);
1366         if (cpipe->pipe_slock)
1367                 lockmgr(cpipe->pipe_slock, LK_RELEASE);
1368
1369         /*
1370          * If we disassociated from our peer we can free resources
1371          */
1372         if (ppipe == NULL) {
1373                 gd = mycpu;
1374                 if (cpipe->pipe_slock) {
1375                         kfree(cpipe->pipe_slock, M_PIPE);
1376                         cpipe->pipe_slock = NULL;
1377                 }
1378                 if (gd->gd_pipeqcount >= pipe_maxcache ||
1379                     cpipe->pipe_buffer.size != PIPE_SIZE
1380                 ) {
1381                         pipe_free_kmem(cpipe);
1382                         kfree(cpipe, M_PIPE);
1383                 } else {
1384                         cpipe->pipe_state = 0;
1385                         cpipe->pipe_peer = gd->gd_pipeq;
1386                         gd->gd_pipeq = cpipe;
1387                         ++gd->gd_pipeqcount;
1388                 }
1389         }
1390 }
1391
1392 /*
1393  * MPALMOSTSAFE - acquires mplock
1394  */
1395 static int
1396 pipe_kqfilter(struct file *fp, struct knote *kn)
1397 {
1398         struct pipe *cpipe;
1399
1400         get_mplock();
1401         cpipe = (struct pipe *)kn->kn_fp->f_data;
1402
1403         switch (kn->kn_filter) {
1404         case EVFILT_READ:
1405                 kn->kn_fop = &pipe_rfiltops;
1406                 break;
1407         case EVFILT_WRITE:
1408                 kn->kn_fop = &pipe_wfiltops;
1409                 cpipe = cpipe->pipe_peer;
1410                 if (cpipe == NULL) {
1411                         /* other end of pipe has been closed */
1412                         rel_mplock();
1413                         return (EPIPE);
1414                 }
1415                 break;
1416         default:
1417                 return (1);
1418         }
1419         kn->kn_hook = (caddr_t)cpipe;
1420
1421         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1422         rel_mplock();
1423         return (0);
1424 }
1425
1426 static void
1427 filt_pipedetach(struct knote *kn)
1428 {
1429         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1430
1431         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1432 }
1433
1434 /*ARGSUSED*/
1435 static int
1436 filt_piperead(struct knote *kn, long hint)
1437 {
1438         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1439
1440         kn->kn_data = rpipe->pipe_buffer.windex - rpipe->pipe_buffer.rindex;
1441
1442         /* XXX RACE */
1443         if (rpipe->pipe_state & PIPE_REOF) {
1444                 kn->kn_flags |= EV_EOF;
1445                 return (1);
1446         }
1447         return (kn->kn_data > 0);
1448 }
1449
1450 /*ARGSUSED*/
1451 static int
1452 filt_pipewrite(struct knote *kn, long hint)
1453 {
1454         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1455         struct pipe *wpipe = rpipe->pipe_peer;
1456         u_int32_t space;
1457
1458         /* XXX RACE */
1459         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_WEOF)) {
1460                 kn->kn_data = 0;
1461                 kn->kn_flags |= EV_EOF;
1462                 return (1);
1463         }
1464         space = wpipe->pipe_buffer.windex -
1465                 wpipe->pipe_buffer.rindex;
1466         space = wpipe->pipe_buffer.size - space;
1467         kn->kn_data = space;
1468         return (kn->kn_data >= PIPE_BUF);
1469 }