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
  20 /*
  21  * This file contains a high-performance replacement for the socket-based
  22  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  23  * all features of sockets, but does do everything that pipes normally
  24  * do.
  25  */
  26
  27 /*
  28  * This code has two modes of operation, a small write mode and a large
  29  * write mode.  The small write mode acts like conventional pipes with
  30  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  31  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  32  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
  33  * the receiving process can copy it directly from the pages in the sending
  34  * process.
  35  *
  36  * If the sending process receives a signal, it is possible that it will
  37  * go away, and certainly its address space can change, because control
  38  * is returned back to the user-mode side.  In that case, the pipe code
  39  * arranges to copy the buffer supplied by the user process, to a pageable
  40  * kernel buffer, and the receiving process will grab the data from the
  41  * pageable kernel buffer.  Since signals don't happen all that often,
  42  * the copy operation is normally eliminated.
  43  *
  44  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  45  * happen for small transfers so that the system will not spend all of
  46  * its time context switching.
  47  *
  48  * In order to limit the resource use of pipes, two sysctls exist:
  49  *
  50  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  51  * address space available to us in pipe_map. This value is normally
  52  * autotuned, but may also be loader tuned.
  53  *
  54  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  55  * memory in use by pipes.
  56  *
  57  * Based on how large pipekva is relative to maxpipekva, the following
  58  * will happen:
  59  *
  60  * 0% - 50%:
  61  *     New pipes are given 16K of memory backing, pipes may dynamically
  62  *     grow to as large as 64K where needed.
  63  * 50% - 75%:
  64  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  65  *     existing pipes may NOT grow.
  66  * 75% - 100%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes will be shrunk down to 4K whenever possible.
  69  *
  70  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  71  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  72  * resize which MUST occur for reverse-direction pipes when they are
  73  * first used.
  74  *
  75  * Additional information about the current state of pipes may be obtained
  76  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  77  * and kern.ipc.piperesizefail.
  78  *
  79  * Locking rules:  There are two locks present here:  A mutex, used via
  80  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  81  * the flag, as mutexes can not persist over uiomove.  The mutex
  82  * exists only to guard access to the flag, and is not in itself a
  83  * locking mechanism.  Also note that there is only a single mutex for
  84  * both directions of a pipe.
  85  *
  86  * As pipelock() may have to sleep before it can acquire the flag, it
  87  * is important to reread all data after a call to pipelock(); everything
  88  * in the structure may have changed.
  89  */
  90
  91 #include <sys/cdefs.h>
  92 __FBSDID("$FreeBSD$");
  93
  94 #include "opt_mac.h"
  95
  96 #include <sys/param.h>
  97 #include <sys/systm.h>
  98 #include <sys/fcntl.h>
  99 #include <sys/file.h>
 100 #include <sys/filedesc.h>
 101 #include <sys/filio.h>
 102 #include <sys/kernel.h>
 103 #include <sys/lock.h>
 104 #include <sys/mutex.h>
 105 #include <sys/ttycom.h>
 106 #include <sys/stat.h>
 107 #include <sys/malloc.h>
 108 #include <sys/poll.h>
 109 #include <sys/selinfo.h>
 110 #include <sys/signalvar.h>
 111 #include <sys/sysctl.h>
 112 #include <sys/sysproto.h>
 113 #include <sys/pipe.h>
 114 #include <sys/proc.h>
 115 #include <sys/vnode.h>
 116 #include <sys/uio.h>
 117 #include <sys/event.h>
 118
 119 #include <security/mac/mac_framework.h>
 120
 121 #include <vm/vm.h>
 122 #include <vm/vm_param.h>
 123 #include <vm/vm_object.h>
 124 #include <vm/vm_kern.h>
 125 #include <vm/vm_extern.h>
 126 #include <vm/pmap.h>
 127 #include <vm/vm_map.h>
 128 #include <vm/vm_page.h>
 129 #include <vm/uma.h>
 130
 131 /*
 132  * Use this define if you want to disable *fancy* VM things.  Expect an
 133  * approx 30% decrease in transfer rate.  This could be useful for
 134  * NetBSD or OpenBSD.
 135  */
 136 /* #define PIPE_NODIRECT */
 137
 138 /*
 139  * interfaces to the outside world
 140  */
 141 static fo_rdwr_t        pipe_read;
 142 static fo_rdwr_t        pipe_write;
 143 static fo_ioctl_t       pipe_ioctl;
 144 static fo_poll_t        pipe_poll;
 145 static fo_kqfilter_t    pipe_kqfilter;
 146 static fo_stat_t        pipe_stat;
 147 static fo_close_t       pipe_close;
 148
 149 static struct fileops pipeops = {
 150         .fo_read = pipe_read,
 151         .fo_write = pipe_write,
 152         .fo_ioctl = pipe_ioctl,
 153         .fo_poll = pipe_poll,
 154         .fo_kqfilter = pipe_kqfilter,
 155         .fo_stat = pipe_stat,
 156         .fo_close = pipe_close,
 157         .fo_flags = DFLAG_PASSABLE
 158 };
 159
 160 static void     filt_pipedetach(struct knote *kn);
 161 static int      filt_piperead(struct knote *kn, long hint);
 162 static int      filt_pipewrite(struct knote *kn, long hint);
 163
 164 static struct filterops pipe_rfiltops =
 165         { 1, NULL, filt_pipedetach, filt_piperead };
 166 static struct filterops pipe_wfiltops =
 167         { 1, NULL, filt_pipedetach, filt_pipewrite };
 168
 169 /*
 170  * Default pipe buffer size(s), this can be kind-of large now because pipe
 171  * space is pageable.  The pipe code will try to maintain locality of
 172  * reference for performance reasons, so small amounts of outstanding I/O
 173  * will not wipe the cache.
 174  */
 175 #define MINPIPESIZE (PIPE_SIZE/3)
 176 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 177
 178 static int amountpipes;
 179 static int amountpipekva;
 180 static int pipefragretry;
 181 static int pipeallocfail;
 182 static int piperesizefail;
 183 static int piperesizeallowed = 1;
 184
 185 SYSCTL_INT(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 186            &maxpipekva, 0, "Pipe KVA limit");
 187 SYSCTL_INT(_kern_ipc, OID_AUTO, pipes, CTLFLAG_RD,
 188            &amountpipes, 0, "Current # of pipes");
 189 SYSCTL_INT(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 190            &amountpipekva, 0, "Pipe KVA usage");
 191 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 192           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 193 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 194           &pipeallocfail, 0, "Pipe allocation failures");
 195 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 196           &piperesizefail, 0, "Pipe resize failures");
 197 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 198           &piperesizeallowed, 0, "Pipe resizing allowed");
 199
 200 static void pipeinit(void *dummy __unused);
 201 static void pipeclose(struct pipe *cpipe);
 202 static void pipe_free_kmem(struct pipe *cpipe);
 203 static int pipe_create(struct pipe *pipe, int backing);
 204 static __inline int pipelock(struct pipe *cpipe, int catch);
 205 static __inline void pipeunlock(struct pipe *cpipe);
 206 static __inline void pipeselwakeup(struct pipe *cpipe);
 207 #ifndef PIPE_NODIRECT
 208 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 209 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 210 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 211 static void pipe_clone_write_buffer(struct pipe *wpipe);
 212 #endif
 213 static int pipespace(struct pipe *cpipe, int size);
 214 static int pipespace_new(struct pipe *cpipe, int size);
 215
 216 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 217 static void     pipe_zone_dtor(void *mem, int size, void *arg);
 218 static int      pipe_zone_init(void *mem, int size, int flags);
 219 static void     pipe_zone_fini(void *mem, int size);
 220
 221 static uma_zone_t pipe_zone;
 222
 223 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 224
 225 static void
 226 pipeinit(void *dummy __unused)
 227 {
 228
 229         pipe_zone = uma_zcreate("PIPE", sizeof(struct pipepair),
 230             pipe_zone_ctor, pipe_zone_dtor, pipe_zone_init, pipe_zone_fini,
 231             UMA_ALIGN_PTR, 0);
 232         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 233 }
 234
 235 static int
 236 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 237 {
 238         struct pipepair *pp;
 239         struct pipe *rpipe, *wpipe;
 240
 241         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 242
 243         pp = (struct pipepair *)mem;
 244
 245         /*
 246          * We zero both pipe endpoints to make sure all the kmem pointers
 247          * are NULL, flag fields are zero'd, etc.  We timestamp both
 248          * endpoints with the same time.
 249          */
 250         rpipe = &pp->pp_rpipe;
 251         bzero(rpipe, sizeof(*rpipe));
 252         vfs_timestamp(&rpipe->pipe_ctime);
 253         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 254
 255         wpipe = &pp->pp_wpipe;
 256         bzero(wpipe, sizeof(*wpipe));
 257         wpipe->pipe_ctime = rpipe->pipe_ctime;
 258         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 259
 260         rpipe->pipe_peer = wpipe;
 261         rpipe->pipe_pair = pp;
 262         wpipe->pipe_peer = rpipe;
 263         wpipe->pipe_pair = pp;
 264
 265         /*
 266          * Mark both endpoints as present; they will later get free'd
 267          * one at a time.  When both are free'd, then the whole pair
 268          * is released.
 269          */
 270         rpipe->pipe_present = 1;
 271         wpipe->pipe_present = 1;
 272
 273         /*
 274          * Eventually, the MAC Framework may initialize the label
 275          * in ctor or init, but for now we do it elswhere to avoid
 276          * blocking in ctor or init.
 277          */
 278         pp->pp_label = NULL;
 279
 280         atomic_add_int(&amountpipes, 2);
 281         return (0);
 282 }
 283
 284 static void
 285 pipe_zone_dtor(void *mem, int size, void *arg)
 286 {
 287         struct pipepair *pp;
 288
 289         KASSERT(size == sizeof(*pp), ("pipe_zone_dtor: wrong size"));
 290
 291         pp = (struct pipepair *)mem;
 292
 293         atomic_subtract_int(&amountpipes, 2);
 294 }
 295
 296 static int
 297 pipe_zone_init(void *mem, int size, int flags)
 298 {
 299         struct pipepair *pp;
 300
 301         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 302
 303         pp = (struct pipepair *)mem;
 304
 305         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 306         return (0);
 307 }
 308
 309 static void
 310 pipe_zone_fini(void *mem, int size)
 311 {
 312         struct pipepair *pp;
 313
 314         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 315
 316         pp = (struct pipepair *)mem;
 317
 318         mtx_destroy(&pp->pp_mtx);
 319 }
 320
 321 /*
 322  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail,
 323  * let the zone pick up the pieces via pipeclose().
 324  */
 325
 326 /* ARGSUSED */
 327 int
 328 pipe(td, uap)
 329         struct thread *td;
 330         struct pipe_args /* {
 331                 int     dummy;
 332         } */ *uap;
 333 {
 334         struct filedesc *fdp = td->td_proc->p_fd;
 335         struct file *rf, *wf;
 336         struct pipepair *pp;
 337         struct pipe *rpipe, *wpipe;
 338         int fd, error;
 339
 340         pp = uma_zalloc(pipe_zone, M_WAITOK);
 341 #ifdef MAC
 342         /*
 343          * The MAC label is shared between the connected endpoints.  As a
 344          * result mac_init_pipe() and mac_create_pipe() are called once
 345          * for the pair, and not on the endpoints.
 346          */
 347         mac_init_pipe(pp);
 348         mac_create_pipe(td->td_ucred, pp);
 349 #endif
 350         rpipe = &pp->pp_rpipe;
 351         wpipe = &pp->pp_wpipe;
 352
 353         knlist_init(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe), NULL, NULL,
 354             NULL);
 355         knlist_init(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe), NULL, NULL,
 356             NULL);
 357
 358         /* Only the forward direction pipe is backed by default */
 359         if ((error = pipe_create(rpipe, 1)) != 0 ||
 360             (error = pipe_create(wpipe, 0)) != 0) {
 361                 pipeclose(rpipe);
 362                 pipeclose(wpipe);
 363                 return (error);
 364         }
 365
 366         rpipe->pipe_state |= PIPE_DIRECTOK;
 367         wpipe->pipe_state |= PIPE_DIRECTOK;
 368
 369         error = falloc(td, &rf, &fd);
 370         if (error) {
 371                 pipeclose(rpipe);
 372                 pipeclose(wpipe);
 373                 return (error);
 374         }
 375         /* An extra reference on `rf' has been held for us by falloc(). */
 376         td->td_retval[0] = fd;
 377
 378         /*
 379          * Warning: once we've gotten past allocation of the fd for the
 380          * read-side, we can only drop the read side via fdrop() in order
 381          * to avoid races against processes which manage to dup() the read
 382          * side while we are blocked trying to allocate the write side.
 383          */
 384         FILE_LOCK(rf);
 385         rf->f_flag = FREAD | FWRITE;
 386         rf->f_type = DTYPE_PIPE;
 387         rf->f_data = rpipe;
 388         rf->f_ops = &pipeops;
 389         FILE_UNLOCK(rf);
 390         error = falloc(td, &wf, &fd);
 391         if (error) {
 392                 fdclose(fdp, rf, td->td_retval[0], td);
 393                 fdrop(rf, td);
 394                 /* rpipe has been closed by fdrop(). */
 395                 pipeclose(wpipe);
 396                 return (error);
 397         }
 398         /* An extra reference on `wf' has been held for us by falloc(). */
 399         FILE_LOCK(wf);
 400         wf->f_flag = FREAD | FWRITE;
 401         wf->f_type = DTYPE_PIPE;
 402         wf->f_data = wpipe;
 403         wf->f_ops = &pipeops;
 404         FILE_UNLOCK(wf);
 405         fdrop(wf, td);
 406         td->td_retval[1] = fd;
 407         fdrop(rf, td);
 408
 409         return (0);
 410 }
 411
 412 /*
 413  * Allocate kva for pipe circular buffer, the space is pageable
 414  * This routine will 'realloc' the size of a pipe safely, if it fails
 415  * it will retain the old buffer.
 416  * If it fails it will return ENOMEM.
 417  */
 418 static int
 419 pipespace_new(cpipe, size)
 420         struct pipe *cpipe;
 421         int size;
 422 {
 423         caddr_t buffer;
 424         int error, cnt, firstseg;
 425         static int curfail = 0;
 426         static struct timeval lastfail;
 427
 428         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 429         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 430                 ("pipespace: resize of direct writes not allowed"));
 431 retry:
 432         cnt = cpipe->pipe_buffer.cnt;
 433         if (cnt > size)
 434                 size = cnt;
 435
 436         size = round_page(size);
 437         buffer = (caddr_t) vm_map_min(pipe_map);
 438
 439         error = vm_map_find(pipe_map, NULL, 0,
 440                 (vm_offset_t *) &buffer, size, 1,
 441                 VM_PROT_ALL, VM_PROT_ALL, 0);
 442         if (error != KERN_SUCCESS) {
 443                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 444                         (size > SMALL_PIPE_SIZE)) {
 445                         size = SMALL_PIPE_SIZE;
 446                         pipefragretry++;
 447                         goto retry;
 448                 }
 449                 if (cpipe->pipe_buffer.buffer == NULL) {
 450                         pipeallocfail++;
 451                         if (ppsratecheck(&lastfail, &curfail, 1))
 452                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 453                 } else {
 454                         piperesizefail++;
 455                 }
 456                 return (ENOMEM);
 457         }
 458
 459         /* copy data, then free old resources if we're resizing */
 460         if (cnt > 0) {
 461                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 462                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 463                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 464                                 buffer, firstseg);
 465                         if ((cnt - firstseg) > 0)
 466                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 467                                         cpipe->pipe_buffer.in);
 468                 } else {
 469                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 470                                 buffer, cnt);
 471                 }
 472         }
 473         pipe_free_kmem(cpipe);
 474         cpipe->pipe_buffer.buffer = buffer;
 475         cpipe->pipe_buffer.size = size;
 476         cpipe->pipe_buffer.in = cnt;
 477         cpipe->pipe_buffer.out = 0;
 478         cpipe->pipe_buffer.cnt = cnt;
 479         atomic_add_int(&amountpipekva, cpipe->pipe_buffer.size);
 480         return (0);
 481 }
 482
 483 /*
 484  * Wrapper for pipespace_new() that performs locking assertions.
 485  */
 486 static int
 487 pipespace(cpipe, size)
 488         struct pipe *cpipe;
 489         int size;
 490 {
 491
 492         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 493                 ("Unlocked pipe passed to pipespace"));
 494         return (pipespace_new(cpipe, size));
 495 }
 496
 497 /*
 498  * lock a pipe for I/O, blocking other access
 499  */
 500 static __inline int
 501 pipelock(cpipe, catch)
 502         struct pipe *cpipe;
 503         int catch;
 504 {
 505         int error;
 506
 507         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 508         while (cpipe->pipe_state & PIPE_LOCKFL) {
 509                 cpipe->pipe_state |= PIPE_LWANT;
 510                 error = msleep(cpipe, PIPE_MTX(cpipe),
 511                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 512                     "pipelk", 0);
 513                 if (error != 0)
 514                         return (error);
 515         }
 516         cpipe->pipe_state |= PIPE_LOCKFL;
 517         return (0);
 518 }
 519
 520 /*
 521  * unlock a pipe I/O lock
 522  */
 523 static __inline void
 524 pipeunlock(cpipe)
 525         struct pipe *cpipe;
 526 {
 527
 528         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 529         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 530                 ("Unlocked pipe passed to pipeunlock"));
 531         cpipe->pipe_state &= ~PIPE_LOCKFL;
 532         if (cpipe->pipe_state & PIPE_LWANT) {
 533                 cpipe->pipe_state &= ~PIPE_LWANT;
 534                 wakeup(cpipe);
 535         }
 536 }
 537
 538 static __inline void
 539 pipeselwakeup(cpipe)
 540         struct pipe *cpipe;
 541 {
 542
 543         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 544         if (cpipe->pipe_state & PIPE_SEL) {
 545                 cpipe->pipe_state &= ~PIPE_SEL;
 546                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 547         }
 548         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 549                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 550         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 551 }
 552
 553 /*
 554  * Initialize and allocate VM and memory for pipe.  The structure
 555  * will start out zero'd from the ctor, so we just manage the kmem.
 556  */
 557 static int
 558 pipe_create(pipe, backing)
 559         struct pipe *pipe;
 560         int backing;
 561 {
 562         int error;
 563
 564         if (backing) {
 565                 if (amountpipekva > maxpipekva / 2)
 566                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 567                 else
 568                         error = pipespace_new(pipe, PIPE_SIZE);
 569         } else {
 570                 /* If we're not backing this pipe, no need to do anything. */
 571                 error = 0;
 572         }
 573         return (error);
 574 }
 575
 576 /* ARGSUSED */
 577 static int
 578 pipe_read(fp, uio, active_cred, flags, td)
 579         struct file *fp;
 580         struct uio *uio;
 581         struct ucred *active_cred;
 582         struct thread *td;
 583         int flags;
 584 {
 585         struct pipe *rpipe = fp->f_data;
 586         int error;
 587         int nread = 0;
 588         u_int size;
 589
 590         PIPE_LOCK(rpipe);
 591         ++rpipe->pipe_busy;
 592         error = pipelock(rpipe, 1);
 593         if (error)
 594                 goto unlocked_error;
 595
 596 #ifdef MAC
 597         error = mac_check_pipe_read(active_cred, rpipe->pipe_pair);
 598         if (error)
 599                 goto locked_error;
 600 #endif
 601         if (amountpipekva > (3 * maxpipekva) / 4) {
 602                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 603                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 604                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 605                         (piperesizeallowed == 1)) {
 606                         PIPE_UNLOCK(rpipe);
 607                         pipespace(rpipe, SMALL_PIPE_SIZE);
 608                         PIPE_LOCK(rpipe);
 609                 }
 610         }
 611
 612         while (uio->uio_resid) {
 613                 /*
 614                  * normal pipe buffer receive
 615                  */
 616                 if (rpipe->pipe_buffer.cnt > 0) {
 617                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 618                         if (size > rpipe->pipe_buffer.cnt)
 619                                 size = rpipe->pipe_buffer.cnt;
 620                         if (size > (u_int) uio->uio_resid)
 621                                 size = (u_int) uio->uio_resid;
 622
 623                         PIPE_UNLOCK(rpipe);
 624                         error = uiomove(
 625                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 626                             size, uio);
 627                         PIPE_LOCK(rpipe);
 628                         if (error)
 629                                 break;
 630
 631                         rpipe->pipe_buffer.out += size;
 632                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 633                                 rpipe->pipe_buffer.out = 0;
 634
 635                         rpipe->pipe_buffer.cnt -= size;
 636
 637                         /*
 638                          * If there is no more to read in the pipe, reset
 639                          * its pointers to the beginning.  This improves
 640                          * cache hit stats.
 641                          */
 642                         if (rpipe->pipe_buffer.cnt == 0) {
 643                                 rpipe->pipe_buffer.in = 0;
 644                                 rpipe->pipe_buffer.out = 0;
 645                         }
 646                         nread += size;
 647 #ifndef PIPE_NODIRECT
 648                 /*
 649                  * Direct copy, bypassing a kernel buffer.
 650                  */
 651                 } else if ((size = rpipe->pipe_map.cnt) &&
 652                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 653                         if (size > (u_int) uio->uio_resid)
 654                                 size = (u_int) uio->uio_resid;
 655
 656                         PIPE_UNLOCK(rpipe);
 657                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 658                             rpipe->pipe_map.pos, size, uio);
 659                         PIPE_LOCK(rpipe);
 660                         if (error)
 661                                 break;
 662                         nread += size;
 663                         rpipe->pipe_map.pos += size;
 664                         rpipe->pipe_map.cnt -= size;
 665                         if (rpipe->pipe_map.cnt == 0) {
 666                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 667                                 wakeup(rpipe);
 668                         }
 669 #endif
 670                 } else {
 671                         /*
 672                          * detect EOF condition
 673                          * read returns 0 on EOF, no need to set error
 674                          */
 675                         if (rpipe->pipe_state & PIPE_EOF)
 676                                 break;
 677
 678                         /*
 679                          * If the "write-side" has been blocked, wake it up now.
 680                          */
 681                         if (rpipe->pipe_state & PIPE_WANTW) {
 682                                 rpipe->pipe_state &= ~PIPE_WANTW;
 683                                 wakeup(rpipe);
 684                         }
 685
 686                         /*
 687                          * Break if some data was read.
 688                          */
 689                         if (nread > 0)
 690                                 break;
 691
 692                         /*
 693                          * Unlock the pipe buffer for our remaining processing.
 694                          * We will either break out with an error or we will
 695                          * sleep and relock to loop.
 696                          */
 697                         pipeunlock(rpipe);
 698
 699                         /*
 700                          * Handle non-blocking mode operation or
 701                          * wait for more data.
 702                          */
 703                         if (fp->f_flag & FNONBLOCK) {
 704                                 error = EAGAIN;
 705                         } else {
 706                                 rpipe->pipe_state |= PIPE_WANTR;
 707                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 708                                     PRIBIO | PCATCH,
 709                                     "piperd", 0)) == 0)
 710                                         error = pipelock(rpipe, 1);
 711                         }
 712                         if (error)
 713                                 goto unlocked_error;
 714                 }
 715         }
 716 #ifdef MAC
 717 locked_error:
 718 #endif
 719         pipeunlock(rpipe);
 720
 721         /* XXX: should probably do this before getting any locks. */
 722         if (error == 0)
 723                 vfs_timestamp(&rpipe->pipe_atime);
 724 unlocked_error:
 725         --rpipe->pipe_busy;
 726
 727         /*
 728          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 729          */
 730         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 731                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 732                 wakeup(rpipe);
 733         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 734                 /*
 735                  * Handle write blocking hysteresis.
 736                  */
 737                 if (rpipe->pipe_state & PIPE_WANTW) {
 738                         rpipe->pipe_state &= ~PIPE_WANTW;
 739                         wakeup(rpipe);
 740                 }
 741         }
 742
 743         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 744                 pipeselwakeup(rpipe);
 745
 746         PIPE_UNLOCK(rpipe);
 747         return (error);
 748 }
 749
 750 #ifndef PIPE_NODIRECT
 751 /*
 752  * Map the sending processes' buffer into kernel space and wire it.
 753  * This is similar to a physical write operation.
 754  */
 755 static int
 756 pipe_build_write_buffer(wpipe, uio)
 757         struct pipe *wpipe;
 758         struct uio *uio;
 759 {
 760         pmap_t pmap;
 761         u_int size;
 762         int i, j;
 763         vm_offset_t addr, endaddr;
 764
 765         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 766         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 767                 ("Clone attempt on non-direct write pipe!"));
 768
 769         size = (u_int) uio->uio_iov->iov_len;
 770         if (size > wpipe->pipe_buffer.size)
 771                 size = wpipe->pipe_buffer.size;
 772
 773         pmap = vmspace_pmap(curproc->p_vmspace);
 774         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 775         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 776         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 777                 /*
 778                  * vm_fault_quick() can sleep.  Consequently,
 779                  * vm_page_lock_queue() and vm_page_unlock_queue()
 780                  * should not be performed outside of this loop.
 781                  */
 782         race:
 783                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) {
 784                         vm_page_lock_queues();
 785                         for (j = 0; j < i; j++)
 786                                 vm_page_unhold(wpipe->pipe_map.ms[j]);
 787                         vm_page_unlock_queues();
 788                         return (EFAULT);
 789                 }
 790                 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr,
 791                     VM_PROT_READ);
 792                 if (wpipe->pipe_map.ms[i] == NULL)
 793                         goto race;
 794         }
 795
 796 /*
 797  * set up the control block
 798  */
 799         wpipe->pipe_map.npages = i;
 800         wpipe->pipe_map.pos =
 801             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 802         wpipe->pipe_map.cnt = size;
 803
 804 /*
 805  * and update the uio data
 806  */
 807
 808         uio->uio_iov->iov_len -= size;
 809         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 810         if (uio->uio_iov->iov_len == 0)
 811                 uio->uio_iov++;
 812         uio->uio_resid -= size;
 813         uio->uio_offset += size;
 814         return (0);
 815 }
 816
 817 /*
 818  * unmap and unwire the process buffer
 819  */
 820 static void
 821 pipe_destroy_write_buffer(wpipe)
 822         struct pipe *wpipe;
 823 {
 824         int i;
 825
 826         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 827         vm_page_lock_queues();
 828         for (i = 0; i < wpipe->pipe_map.npages; i++) {
 829                 vm_page_unhold(wpipe->pipe_map.ms[i]);
 830         }
 831         vm_page_unlock_queues();
 832         wpipe->pipe_map.npages = 0;
 833 }
 834
 835 /*
 836  * In the case of a signal, the writing process might go away.  This
 837  * code copies the data into the circular buffer so that the source
 838  * pages can be freed without loss of data.
 839  */
 840 static void
 841 pipe_clone_write_buffer(wpipe)
 842         struct pipe *wpipe;
 843 {
 844         struct uio uio;
 845         struct iovec iov;
 846         int size;
 847         int pos;
 848
 849         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 850         size = wpipe->pipe_map.cnt;
 851         pos = wpipe->pipe_map.pos;
 852
 853         wpipe->pipe_buffer.in = size;
 854         wpipe->pipe_buffer.out = 0;
 855         wpipe->pipe_buffer.cnt = size;
 856         wpipe->pipe_state &= ~PIPE_DIRECTW;
 857
 858         PIPE_UNLOCK(wpipe);
 859         iov.iov_base = wpipe->pipe_buffer.buffer;
 860         iov.iov_len = size;
 861         uio.uio_iov = &iov;
 862         uio.uio_iovcnt = 1;
 863         uio.uio_offset = 0;
 864         uio.uio_resid = size;
 865         uio.uio_segflg = UIO_SYSSPACE;
 866         uio.uio_rw = UIO_READ;
 867         uio.uio_td = curthread;
 868         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 869         PIPE_LOCK(wpipe);
 870         pipe_destroy_write_buffer(wpipe);
 871 }
 872
 873 /*
 874  * This implements the pipe buffer write mechanism.  Note that only
 875  * a direct write OR a normal pipe write can be pending at any given time.
 876  * If there are any characters in the pipe buffer, the direct write will
 877  * be deferred until the receiving process grabs all of the bytes from
 878  * the pipe buffer.  Then the direct mapping write is set-up.
 879  */
 880 static int
 881 pipe_direct_write(wpipe, uio)
 882         struct pipe *wpipe;
 883         struct uio *uio;
 884 {
 885         int error;
 886
 887 retry:
 888         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 889         error = pipelock(wpipe, 1);
 890         if (wpipe->pipe_state & PIPE_EOF)
 891                 error = EPIPE;
 892         if (error) {
 893                 pipeunlock(wpipe);
 894                 goto error1;
 895         }
 896         while (wpipe->pipe_state & PIPE_DIRECTW) {
 897                 if (wpipe->pipe_state & PIPE_WANTR) {
 898                         wpipe->pipe_state &= ~PIPE_WANTR;
 899                         wakeup(wpipe);
 900                 }
 901                 wpipe->pipe_state |= PIPE_WANTW;
 902                 pipeunlock(wpipe);
 903                 error = msleep(wpipe, PIPE_MTX(wpipe),
 904                     PRIBIO | PCATCH, "pipdww", 0);
 905                 if (error)
 906                         goto error1;
 907                 else
 908                         goto retry;
 909         }
 910         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 911         if (wpipe->pipe_buffer.cnt > 0) {
 912                 if (wpipe->pipe_state & PIPE_WANTR) {
 913                         wpipe->pipe_state &= ~PIPE_WANTR;
 914                         wakeup(wpipe);
 915                 }
 916                 wpipe->pipe_state |= PIPE_WANTW;
 917                 pipeunlock(wpipe);
 918                 error = msleep(wpipe, PIPE_MTX(wpipe),
 919                     PRIBIO | PCATCH, "pipdwc", 0);
 920                 if (error)
 921                         goto error1;
 922                 else
 923                         goto retry;
 924         }
 925
 926         wpipe->pipe_state |= PIPE_DIRECTW;
 927
 928         PIPE_UNLOCK(wpipe);
 929         error = pipe_build_write_buffer(wpipe, uio);
 930         PIPE_LOCK(wpipe);
 931         if (error) {
 932                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 933                 pipeunlock(wpipe);
 934                 goto error1;
 935         }
 936
 937         error = 0;
 938         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 939                 if (wpipe->pipe_state & PIPE_EOF) {
 940                         pipe_destroy_write_buffer(wpipe);
 941                         pipeselwakeup(wpipe);
 942                         pipeunlock(wpipe);
 943                         error = EPIPE;
 944                         goto error1;
 945                 }
 946                 if (wpipe->pipe_state & PIPE_WANTR) {
 947                         wpipe->pipe_state &= ~PIPE_WANTR;
 948                         wakeup(wpipe);
 949                 }
 950                 pipeselwakeup(wpipe);
 951                 pipeunlock(wpipe);
 952                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 953                     "pipdwt", 0);
 954                 pipelock(wpipe, 0);
 955         }
 956
 957         if (wpipe->pipe_state & PIPE_EOF)
 958                 error = EPIPE;
 959         if (wpipe->pipe_state & PIPE_DIRECTW) {
 960                 /*
 961                  * this bit of trickery substitutes a kernel buffer for
 962                  * the process that might be going away.
 963                  */
 964                 pipe_clone_write_buffer(wpipe);
 965         } else {
 966                 pipe_destroy_write_buffer(wpipe);
 967         }
 968         pipeunlock(wpipe);
 969         return (error);
 970
 971 error1:
 972         wakeup(wpipe);
 973         return (error);
 974 }
 975 #endif
 976
 977 static int
 978 pipe_write(fp, uio, active_cred, flags, td)
 979         struct file *fp;
 980         struct uio *uio;
 981         struct ucred *active_cred;
 982         struct thread *td;
 983         int flags;
 984 {
 985         int error = 0;
 986         int desiredsize, orig_resid;
 987         struct pipe *wpipe, *rpipe;
 988
 989         rpipe = fp->f_data;
 990         wpipe = rpipe->pipe_peer;
 991
 992         PIPE_LOCK(rpipe);
 993         error = pipelock(wpipe, 1);
 994         if (error) {
 995                 PIPE_UNLOCK(rpipe);
 996                 return (error);
 997         }
 998         /*
 999          * detect loss of pipe read side, issue SIGPIPE if lost.
1000          */
1001         if ((!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
1002                 pipeunlock(wpipe);
1003                 PIPE_UNLOCK(rpipe);
1004                 return (EPIPE);
1005         }
1006 #ifdef MAC
1007         error = mac_check_pipe_write(active_cred, wpipe->pipe_pair);
1008         if (error) {
1009                 pipeunlock(wpipe);
1010                 PIPE_UNLOCK(rpipe);
1011                 return (error);
1012         }
1013 #endif
1014         ++wpipe->pipe_busy;
1015
1016         /* Choose a larger size if it's advantageous */
1017         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1018         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1019                 if (piperesizeallowed != 1)
1020                         break;
1021                 if (amountpipekva > maxpipekva / 2)
1022                         break;
1023                 if (desiredsize == BIG_PIPE_SIZE)
1024                         break;
1025                 desiredsize = desiredsize * 2;
1026         }
1027
1028         /* Choose a smaller size if we're in a OOM situation */
1029         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1030                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1031                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1032                 (piperesizeallowed == 1))
1033                 desiredsize = SMALL_PIPE_SIZE;
1034
1035         /* Resize if the above determined that a new size was necessary */
1036         if ((desiredsize != wpipe->pipe_buffer.size) &&
1037                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1038                 PIPE_UNLOCK(wpipe);
1039                 pipespace(wpipe, desiredsize);
1040                 PIPE_LOCK(wpipe);
1041         }
1042         if (wpipe->pipe_buffer.size == 0) {
1043                 /*
1044                  * This can only happen for reverse direction use of pipes
1045                  * in a complete OOM situation.
1046                  */
1047                 error = ENOMEM;
1048                 --wpipe->pipe_busy;
1049                 pipeunlock(wpipe);
1050                 PIPE_UNLOCK(wpipe);
1051                 return (error);
1052         }
1053
1054         pipeunlock(wpipe);
1055
1056         orig_resid = uio->uio_resid;
1057
1058         while (uio->uio_resid) {
1059                 int space;
1060
1061                 pipelock(wpipe, 0);
1062                 if (wpipe->pipe_state & PIPE_EOF) {
1063                         pipeunlock(wpipe);
1064                         error = EPIPE;
1065                         break;
1066                 }
1067 #ifndef PIPE_NODIRECT
1068                 /*
1069                  * If the transfer is large, we can gain performance if
1070                  * we do process-to-process copies directly.
1071                  * If the write is non-blocking, we don't use the
1072                  * direct write mechanism.
1073                  *
1074                  * The direct write mechanism will detect the reader going
1075                  * away on us.
1076                  */
1077                 if (uio->uio_segflg == UIO_USERSPACE &&
1078                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1079                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1080                     (fp->f_flag & FNONBLOCK) == 0) {
1081                         pipeunlock(wpipe);
1082                         error = pipe_direct_write(wpipe, uio);
1083                         if (error)
1084                                 break;
1085                         continue;
1086                 }
1087 #endif
1088
1089                 /*
1090                  * Pipe buffered writes cannot be coincidental with
1091                  * direct writes.  We wait until the currently executing
1092                  * direct write is completed before we start filling the
1093                  * pipe buffer.  We break out if a signal occurs or the
1094                  * reader goes away.
1095                  */
1096                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1097                         if (wpipe->pipe_state & PIPE_WANTR) {
1098                                 wpipe->pipe_state &= ~PIPE_WANTR;
1099                                 wakeup(wpipe);
1100                         }
1101                         pipeunlock(wpipe);
1102                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1103                             "pipbww", 0);
1104                         if (error)
1105                                 break;
1106                         else
1107                                 continue;
1108                 }
1109
1110                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1111
1112                 /* Writes of size <= PIPE_BUF must be atomic. */
1113                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1114                         space = 0;
1115
1116                 if (space > 0) {
1117                         int size;       /* Transfer size */
1118                         int segsize;    /* first segment to transfer */
1119
1120                         /*
1121                          * Transfer size is minimum of uio transfer
1122                          * and free space in pipe buffer.
1123                          */
1124                         if (space > uio->uio_resid)
1125                                 size = uio->uio_resid;
1126                         else
1127                                 size = space;
1128                         /*
1129                          * First segment to transfer is minimum of
1130                          * transfer size and contiguous space in
1131                          * pipe buffer.  If first segment to transfer
1132                          * is less than the transfer size, we've got
1133                          * a wraparound in the buffer.
1134                          */
1135                         segsize = wpipe->pipe_buffer.size -
1136                                 wpipe->pipe_buffer.in;
1137                         if (segsize > size)
1138                                 segsize = size;
1139
1140                         /* Transfer first segment */
1141
1142                         PIPE_UNLOCK(rpipe);
1143                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1144                                         segsize, uio);
1145                         PIPE_LOCK(rpipe);
1146
1147                         if (error == 0 && segsize < size) {
1148                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1149                                         wpipe->pipe_buffer.size,
1150                                         ("Pipe buffer wraparound disappeared"));
1151                                 /*
1152                                  * Transfer remaining part now, to
1153                                  * support atomic writes.  Wraparound
1154                                  * happened.
1155                                  */
1156
1157                                 PIPE_UNLOCK(rpipe);
1158                                 error = uiomove(
1159                                     &wpipe->pipe_buffer.buffer[0],
1160                                     size - segsize, uio);
1161                                 PIPE_LOCK(rpipe);
1162                         }
1163                         if (error == 0) {
1164                                 wpipe->pipe_buffer.in += size;
1165                                 if (wpipe->pipe_buffer.in >=
1166                                     wpipe->pipe_buffer.size) {
1167                                         KASSERT(wpipe->pipe_buffer.in ==
1168                                                 size - segsize +
1169                                                 wpipe->pipe_buffer.size,
1170                                                 ("Expected wraparound bad"));
1171                                         wpipe->pipe_buffer.in = size - segsize;
1172                                 }
1173
1174                                 wpipe->pipe_buffer.cnt += size;
1175                                 KASSERT(wpipe->pipe_buffer.cnt <=
1176                                         wpipe->pipe_buffer.size,
1177                                         ("Pipe buffer overflow"));
1178                         }
1179                         pipeunlock(wpipe);
1180                         if (error != 0)
1181                                 break;
1182                 } else {
1183                         /*
1184                          * If the "read-side" has been blocked, wake it up now.
1185                          */
1186                         if (wpipe->pipe_state & PIPE_WANTR) {
1187                                 wpipe->pipe_state &= ~PIPE_WANTR;
1188                                 wakeup(wpipe);
1189                         }
1190
1191                         /*
1192                          * don't block on non-blocking I/O
1193                          */
1194                         if (fp->f_flag & FNONBLOCK) {
1195                                 error = EAGAIN;
1196                                 pipeunlock(wpipe);
1197                                 break;
1198                         }
1199
1200                         /*
1201                          * We have no more space and have something to offer,
1202                          * wake up select/poll.
1203                          */
1204                         pipeselwakeup(wpipe);
1205
1206                         wpipe->pipe_state |= PIPE_WANTW;
1207                         pipeunlock(wpipe);
1208                         error = msleep(wpipe, PIPE_MTX(rpipe),
1209                             PRIBIO | PCATCH, "pipewr", 0);
1210                         if (error != 0)
1211                                 break;
1212                 }
1213         }
1214
1215         pipelock(wpipe, 0);
1216         --wpipe->pipe_busy;
1217
1218         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1219                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1220                 wakeup(wpipe);
1221         } else if (wpipe->pipe_buffer.cnt > 0) {
1222                 /*
1223                  * If we have put any characters in the buffer, we wake up
1224                  * the reader.
1225                  */
1226                 if (wpipe->pipe_state & PIPE_WANTR) {
1227                         wpipe->pipe_state &= ~PIPE_WANTR;
1228                         wakeup(wpipe);
1229                 }
1230         }
1231
1232         /*
1233          * Don't return EPIPE if I/O was successful
1234          */
1235         if ((wpipe->pipe_buffer.cnt == 0) &&
1236             (uio->uio_resid == 0) &&
1237             (error == EPIPE)) {
1238                 error = 0;
1239         }
1240
1241         if (error == 0)
1242                 vfs_timestamp(&wpipe->pipe_mtime);
1243
1244         /*
1245          * We have something to offer,
1246          * wake up select/poll.
1247          */
1248         if (wpipe->pipe_buffer.cnt)
1249                 pipeselwakeup(wpipe);
1250
1251         pipeunlock(wpipe);
1252         PIPE_UNLOCK(rpipe);
1253         return (error);
1254 }
1255
1256 /*
1257  * we implement a very minimal set of ioctls for compatibility with sockets.
1258  */
1259 static int
1260 pipe_ioctl(fp, cmd, data, active_cred, td)
1261         struct file *fp;
1262         u_long cmd;
1263         void *data;
1264         struct ucred *active_cred;
1265         struct thread *td;
1266 {
1267         struct pipe *mpipe = fp->f_data;
1268         int error;
1269
1270         PIPE_LOCK(mpipe);
1271
1272 #ifdef MAC
1273         error = mac_check_pipe_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1274         if (error) {
1275                 PIPE_UNLOCK(mpipe);
1276                 return (error);
1277         }
1278 #endif
1279
1280         error = 0;
1281         switch (cmd) {
1282
1283         case FIONBIO:
1284                 break;
1285
1286         case FIOASYNC:
1287                 if (*(int *)data) {
1288                         mpipe->pipe_state |= PIPE_ASYNC;
1289                 } else {
1290                         mpipe->pipe_state &= ~PIPE_ASYNC;
1291                 }
1292                 break;
1293
1294         case FIONREAD:
1295                 if (mpipe->pipe_state & PIPE_DIRECTW)
1296                         *(int *)data = mpipe->pipe_map.cnt;
1297                 else
1298                         *(int *)data = mpipe->pipe_buffer.cnt;
1299                 break;
1300
1301         case FIOSETOWN:
1302                 PIPE_UNLOCK(mpipe);
1303                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1304                 goto out_unlocked;
1305
1306         case FIOGETOWN:
1307                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1308                 break;
1309
1310         /* This is deprecated, FIOSETOWN should be used instead. */
1311         case TIOCSPGRP:
1312                 PIPE_UNLOCK(mpipe);
1313                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1314                 goto out_unlocked;
1315
1316         /* This is deprecated, FIOGETOWN should be used instead. */
1317         case TIOCGPGRP:
1318                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1319                 break;
1320
1321         default:
1322                 error = ENOTTY;
1323                 break;
1324         }
1325         PIPE_UNLOCK(mpipe);
1326 out_unlocked:
1327         return (error);
1328 }
1329
1330 static int
1331 pipe_poll(fp, events, active_cred, td)
1332         struct file *fp;
1333         int events;
1334         struct ucred *active_cred;
1335         struct thread *td;
1336 {
1337         struct pipe *rpipe = fp->f_data;
1338         struct pipe *wpipe;
1339         int revents = 0;
1340 #ifdef MAC
1341         int error;
1342 #endif
1343
1344         wpipe = rpipe->pipe_peer;
1345         PIPE_LOCK(rpipe);
1346 #ifdef MAC
1347         error = mac_check_pipe_poll(active_cred, rpipe->pipe_pair);
1348         if (error)
1349                 goto locked_error;
1350 #endif
1351         if (events & (POLLIN | POLLRDNORM))
1352                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1353                     (rpipe->pipe_buffer.cnt > 0) ||
1354                     (rpipe->pipe_state & PIPE_EOF))
1355                         revents |= events & (POLLIN | POLLRDNORM);
1356
1357         if (events & (POLLOUT | POLLWRNORM))
1358                 if (!wpipe->pipe_present || (wpipe->pipe_state & PIPE_EOF) ||
1359                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1360                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1361                         revents |= events & (POLLOUT | POLLWRNORM);
1362
1363         if ((rpipe->pipe_state & PIPE_EOF) ||
1364             (!wpipe->pipe_present) ||
1365             (wpipe->pipe_state & PIPE_EOF))
1366                 revents |= POLLHUP;
1367
1368         if (revents == 0) {
1369                 if (events & (POLLIN | POLLRDNORM)) {
1370                         selrecord(td, &rpipe->pipe_sel);
1371                         rpipe->pipe_state |= PIPE_SEL;
1372                 }
1373
1374                 if (events & (POLLOUT | POLLWRNORM)) {
1375                         selrecord(td, &wpipe->pipe_sel);
1376                         wpipe->pipe_state |= PIPE_SEL;
1377                 }
1378         }
1379 #ifdef MAC
1380 locked_error:
1381 #endif
1382         PIPE_UNLOCK(rpipe);
1383
1384         return (revents);
1385 }
1386
1387 /*
1388  * We shouldn't need locks here as we're doing a read and this should
1389  * be a natural race.
1390  */
1391 static int
1392 pipe_stat(fp, ub, active_cred, td)
1393         struct file *fp;
1394         struct stat *ub;
1395         struct ucred *active_cred;
1396         struct thread *td;
1397 {
1398         struct pipe *pipe = fp->f_data;
1399 #ifdef MAC
1400         int error;
1401
1402         PIPE_LOCK(pipe);
1403         error = mac_check_pipe_stat(active_cred, pipe->pipe_pair);
1404         PIPE_UNLOCK(pipe);
1405         if (error)
1406                 return (error);
1407 #endif
1408         bzero(ub, sizeof(*ub));
1409         ub->st_mode = S_IFIFO;
1410         ub->st_blksize = PAGE_SIZE;
1411         if (pipe->pipe_state & PIPE_DIRECTW)
1412                 ub->st_size = pipe->pipe_map.cnt;
1413         else
1414                 ub->st_size = pipe->pipe_buffer.cnt;
1415         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1416         ub->st_atimespec = pipe->pipe_atime;
1417         ub->st_mtimespec = pipe->pipe_mtime;
1418         ub->st_ctimespec = pipe->pipe_ctime;
1419         ub->st_uid = fp->f_cred->cr_uid;
1420         ub->st_gid = fp->f_cred->cr_gid;
1421         /*
1422          * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1423          * XXX (st_dev, st_ino) should be unique.
1424          */
1425         return (0);
1426 }
1427
1428 /* ARGSUSED */
1429 static int
1430 pipe_close(fp, td)
1431         struct file *fp;
1432         struct thread *td;
1433 {
1434         struct pipe *cpipe = fp->f_data;
1435
1436         fp->f_ops = &badfileops;
1437         fp->f_data = NULL;
1438         funsetown(&cpipe->pipe_sigio);
1439         pipeclose(cpipe);
1440         return (0);
1441 }
1442
1443 static void
1444 pipe_free_kmem(cpipe)
1445         struct pipe *cpipe;
1446 {
1447
1448         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1449             ("pipe_free_kmem: pipe mutex locked"));
1450
1451         if (cpipe->pipe_buffer.buffer != NULL) {
1452                 atomic_subtract_int(&amountpipekva, cpipe->pipe_buffer.size);
1453                 vm_map_remove(pipe_map,
1454                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1455                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1456                 cpipe->pipe_buffer.buffer = NULL;
1457         }
1458 #ifndef PIPE_NODIRECT
1459         {
1460                 cpipe->pipe_map.cnt = 0;
1461                 cpipe->pipe_map.pos = 0;
1462                 cpipe->pipe_map.npages = 0;
1463         }
1464 #endif
1465 }
1466
1467 /*
1468  * shutdown the pipe
1469  */
1470 static void
1471 pipeclose(cpipe)
1472         struct pipe *cpipe;
1473 {
1474         struct pipepair *pp;
1475         struct pipe *ppipe;
1476
1477         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1478
1479         PIPE_LOCK(cpipe);
1480         pipelock(cpipe, 0);
1481         pp = cpipe->pipe_pair;
1482
1483         pipeselwakeup(cpipe);
1484
1485         /*
1486          * If the other side is blocked, wake it up saying that
1487          * we want to close it down.
1488          */
1489         cpipe->pipe_state |= PIPE_EOF;
1490         while (cpipe->pipe_busy) {
1491                 wakeup(cpipe);
1492                 cpipe->pipe_state |= PIPE_WANT;
1493                 pipeunlock(cpipe);
1494                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1495                 pipelock(cpipe, 0);
1496         }
1497
1498
1499         /*
1500          * Disconnect from peer, if any.
1501          */
1502         ppipe = cpipe->pipe_peer;
1503         if (ppipe->pipe_present != 0) {
1504                 pipeselwakeup(ppipe);
1505
1506                 ppipe->pipe_state |= PIPE_EOF;
1507                 wakeup(ppipe);
1508                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1509         }
1510
1511         /*
1512          * Mark this endpoint as free.  Release kmem resources.  We
1513          * don't mark this endpoint as unused until we've finished
1514          * doing that, or the pipe might disappear out from under
1515          * us.
1516          */
1517         PIPE_UNLOCK(cpipe);
1518         pipe_free_kmem(cpipe);
1519         PIPE_LOCK(cpipe);
1520         cpipe->pipe_present = 0;
1521         pipeunlock(cpipe);
1522         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1523         knlist_destroy(&cpipe->pipe_sel.si_note);
1524
1525         /*
1526          * If both endpoints are now closed, release the memory for the
1527          * pipe pair.  If not, unlock.
1528          */
1529         if (ppipe->pipe_present == 0) {
1530                 PIPE_UNLOCK(cpipe);
1531 #ifdef MAC
1532                 mac_destroy_pipe(pp);
1533 #endif
1534                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1535         } else
1536                 PIPE_UNLOCK(cpipe);
1537 }
1538
1539 /*ARGSUSED*/
1540 static int
1541 pipe_kqfilter(struct file *fp, struct knote *kn)
1542 {
1543         struct pipe *cpipe;
1544
1545         cpipe = kn->kn_fp->f_data;
1546         PIPE_LOCK(cpipe);
1547         switch (kn->kn_filter) {
1548         case EVFILT_READ:
1549                 kn->kn_fop = &pipe_rfiltops;
1550                 break;
1551         case EVFILT_WRITE:
1552                 kn->kn_fop = &pipe_wfiltops;
1553                 if (!cpipe->pipe_peer->pipe_present) {
1554                         /* other end of pipe has been closed */
1555                         PIPE_UNLOCK(cpipe);
1556                         return (EPIPE);
1557                 }
1558                 cpipe = cpipe->pipe_peer;
1559                 break;
1560         default:
1561                 PIPE_UNLOCK(cpipe);
1562                 return (EINVAL);
1563         }
1564
1565         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1566         PIPE_UNLOCK(cpipe);
1567         return (0);
1568 }
1569
1570 static void
1571 filt_pipedetach(struct knote *kn)
1572 {
1573         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1574
1575         PIPE_LOCK(cpipe);
1576         if (kn->kn_filter == EVFILT_WRITE) {
1577                 if (!cpipe->pipe_peer->pipe_present) {
1578                         PIPE_UNLOCK(cpipe);
1579                         return;
1580                 }
1581                 cpipe = cpipe->pipe_peer;
1582         }
1583         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1584         PIPE_UNLOCK(cpipe);
1585 }
1586
1587 /*ARGSUSED*/
1588 static int
1589 filt_piperead(struct knote *kn, long hint)
1590 {
1591         struct pipe *rpipe = kn->kn_fp->f_data;
1592         struct pipe *wpipe = rpipe->pipe_peer;
1593         int ret;
1594
1595         PIPE_LOCK(rpipe);
1596         kn->kn_data = rpipe->pipe_buffer.cnt;
1597         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1598                 kn->kn_data = rpipe->pipe_map.cnt;
1599
1600         if ((rpipe->pipe_state & PIPE_EOF) ||
1601             (!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
1602                 kn->kn_flags |= EV_EOF;
1603                 PIPE_UNLOCK(rpipe);
1604                 return (1);
1605         }
1606         ret = kn->kn_data > 0;
1607         PIPE_UNLOCK(rpipe);
1608         return ret;
1609 }
1610
1611 /*ARGSUSED*/
1612 static int
1613 filt_pipewrite(struct knote *kn, long hint)
1614 {
1615         struct pipe *rpipe = kn->kn_fp->f_data;
1616         struct pipe *wpipe = rpipe->pipe_peer;
1617
1618         PIPE_LOCK(rpipe);
1619         if ((!wpipe->pipe_present) || (wpipe->pipe_state & PIPE_EOF)) {
1620                 kn->kn_data = 0;
1621                 kn->kn_flags |= EV_EOF;
1622                 PIPE_UNLOCK(rpipe);
1623                 return (1);
1624         }
1625         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1626         if (wpipe->pipe_state & PIPE_DIRECTW)
1627                 kn->kn_data = 0;
1628
1629         PIPE_UNLOCK(rpipe);
1630         return (kn->kn_data >= PIPE_BUF);
1631 }