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 <sys/param.h>
  95 #include <sys/systm.h>
  96 #include <sys/fcntl.h>
  97 #include <sys/file.h>
  98 #include <sys/filedesc.h>
  99 #include <sys/filio.h>
 100 #include <sys/kernel.h>
 101 #include <sys/lock.h>
 102 #include <sys/mutex.h>
 103 #include <sys/ttycom.h>
 104 #include <sys/stat.h>
 105 #include <sys/malloc.h>
 106 #include <sys/poll.h>
 107 #include <sys/selinfo.h>
 108 #include <sys/signalvar.h>
 109 #include <sys/syscallsubr.h>
 110 #include <sys/sysctl.h>
 111 #include <sys/sysproto.h>
 112 #include <sys/pipe.h>
 113 #include <sys/proc.h>
 114 #include <sys/vnode.h>
 115 #include <sys/uio.h>
 116 #include <sys/event.h>
 117
 118 #include <security/mac/mac_framework.h>
 119
 120 #include <vm/vm.h>
 121 #include <vm/vm_param.h>
 122 #include <vm/vm_object.h>
 123 #include <vm/vm_kern.h>
 124 #include <vm/vm_extern.h>
 125 #include <vm/pmap.h>
 126 #include <vm/vm_map.h>
 127 #include <vm/vm_page.h>
 128 #include <vm/uma.h>
 129
 130 /*
 131  * Use this define if you want to disable *fancy* VM things.  Expect an
 132  * approx 30% decrease in transfer rate.  This could be useful for
 133  * NetBSD or OpenBSD.
 134  */
 135 /* #define PIPE_NODIRECT */
 136
 137 /*
 138  * interfaces to the outside world
 139  */
 140 static fo_rdwr_t        pipe_read;
 141 static fo_rdwr_t        pipe_write;
 142 static fo_truncate_t    pipe_truncate;
 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_truncate = pipe_truncate,
 153         .fo_ioctl = pipe_ioctl,
 154         .fo_poll = pipe_poll,
 155         .fo_kqfilter = pipe_kqfilter,
 156         .fo_stat = pipe_stat,
 157         .fo_close = pipe_close,
 158         .fo_flags = DFLAG_PASSABLE
 159 };
 160
 161 static void     filt_pipedetach(struct knote *kn);
 162 static int      filt_piperead(struct knote *kn, long hint);
 163 static int      filt_pipewrite(struct knote *kn, long hint);
 164
 165 static struct filterops pipe_rfiltops =
 166         { 1, NULL, filt_pipedetach, filt_piperead };
 167 static struct filterops pipe_wfiltops =
 168         { 1, NULL, filt_pipedetach, filt_pipewrite };
 169
 170 /*
 171  * Default pipe buffer size(s), this can be kind-of large now because pipe
 172  * space is pageable.  The pipe code will try to maintain locality of
 173  * reference for performance reasons, so small amounts of outstanding I/O
 174  * will not wipe the cache.
 175  */
 176 #define MINPIPESIZE (PIPE_SIZE/3)
 177 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 178
 179 static long amountpipekva;
 180 static int pipefragretry;
 181 static int pipeallocfail;
 182 static int piperesizefail;
 183 static int piperesizeallowed = 1;
 184
 185 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 186            &maxpipekva, 0, "Pipe KVA limit");
 187 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 188            &amountpipekva, 0, "Pipe KVA usage");
 189 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 190           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 191 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 192           &pipeallocfail, 0, "Pipe allocation failures");
 193 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 194           &piperesizefail, 0, "Pipe resize failures");
 195 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 196           &piperesizeallowed, 0, "Pipe resizing allowed");
 197
 198 static void pipeinit(void *dummy __unused);
 199 static void pipeclose(struct pipe *cpipe);
 200 static void pipe_free_kmem(struct pipe *cpipe);
 201 static int pipe_create(struct pipe *pipe, int backing);
 202 static __inline int pipelock(struct pipe *cpipe, int catch);
 203 static __inline void pipeunlock(struct pipe *cpipe);
 204 static __inline void pipeselwakeup(struct pipe *cpipe);
 205 #ifndef PIPE_NODIRECT
 206 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 207 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 208 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 209 static void pipe_clone_write_buffer(struct pipe *wpipe);
 210 #endif
 211 static int pipespace(struct pipe *cpipe, int size);
 212 static int pipespace_new(struct pipe *cpipe, int size);
 213
 214 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 215 static int      pipe_zone_init(void *mem, int size, int flags);
 216 static void     pipe_zone_fini(void *mem, int size);
 217
 218 static uma_zone_t pipe_zone;
 219
 220 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 221
 222 static void
 223 pipeinit(void *dummy __unused)
 224 {
 225
 226         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 227             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 228             UMA_ALIGN_PTR, 0);
 229         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 230 }
 231
 232 static int
 233 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 234 {
 235         struct pipepair *pp;
 236         struct pipe *rpipe, *wpipe;
 237
 238         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 239
 240         pp = (struct pipepair *)mem;
 241
 242         /*
 243          * We zero both pipe endpoints to make sure all the kmem pointers
 244          * are NULL, flag fields are zero'd, etc.  We timestamp both
 245          * endpoints with the same time.
 246          */
 247         rpipe = &pp->pp_rpipe;
 248         bzero(rpipe, sizeof(*rpipe));
 249         vfs_timestamp(&rpipe->pipe_ctime);
 250         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 251
 252         wpipe = &pp->pp_wpipe;
 253         bzero(wpipe, sizeof(*wpipe));
 254         wpipe->pipe_ctime = rpipe->pipe_ctime;
 255         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 256
 257         rpipe->pipe_peer = wpipe;
 258         rpipe->pipe_pair = pp;
 259         wpipe->pipe_peer = rpipe;
 260         wpipe->pipe_pair = pp;
 261
 262         /*
 263          * Mark both endpoints as present; they will later get free'd
 264          * one at a time.  When both are free'd, then the whole pair
 265          * is released.
 266          */
 267         rpipe->pipe_present = PIPE_ACTIVE;
 268         wpipe->pipe_present = PIPE_ACTIVE;
 269
 270         /*
 271          * Eventually, the MAC Framework may initialize the label
 272          * in ctor or init, but for now we do it elswhere to avoid
 273          * blocking in ctor or init.
 274          */
 275         pp->pp_label = NULL;
 276
 277         return (0);
 278 }
 279
 280 static int
 281 pipe_zone_init(void *mem, int size, int flags)
 282 {
 283         struct pipepair *pp;
 284
 285         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 286
 287         pp = (struct pipepair *)mem;
 288
 289         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 290         return (0);
 291 }
 292
 293 static void
 294 pipe_zone_fini(void *mem, int size)
 295 {
 296         struct pipepair *pp;
 297
 298         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 299
 300         pp = (struct pipepair *)mem;
 301
 302         mtx_destroy(&pp->pp_mtx);
 303 }
 304
 305 /*
 306  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 307  * the zone pick up the pieces via pipeclose().
 308  */
 309 int
 310 kern_pipe(struct thread *td, int fildes[2])
 311 {
 312         struct filedesc *fdp = td->td_proc->p_fd;
 313         struct file *rf, *wf;
 314         struct pipepair *pp;
 315         struct pipe *rpipe, *wpipe;
 316         int fd, error;
 317
 318         pp = uma_zalloc(pipe_zone, M_WAITOK);
 319 #ifdef MAC
 320         /*
 321          * The MAC label is shared between the connected endpoints.  As a
 322          * result mac_pipe_init() and mac_pipe_create() are called once
 323          * for the pair, and not on the endpoints.
 324          */
 325         mac_pipe_init(pp);
 326         mac_pipe_create(td->td_ucred, pp);
 327 #endif
 328         rpipe = &pp->pp_rpipe;
 329         wpipe = &pp->pp_wpipe;
 330
 331         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 332         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 333
 334         /* Only the forward direction pipe is backed by default */
 335         if ((error = pipe_create(rpipe, 1)) != 0 ||
 336             (error = pipe_create(wpipe, 0)) != 0) {
 337                 pipeclose(rpipe);
 338                 pipeclose(wpipe);
 339                 return (error);
 340         }
 341
 342         rpipe->pipe_state |= PIPE_DIRECTOK;
 343         wpipe->pipe_state |= PIPE_DIRECTOK;
 344
 345         error = falloc(td, &rf, &fd);
 346         if (error) {
 347                 pipeclose(rpipe);
 348                 pipeclose(wpipe);
 349                 return (error);
 350         }
 351         /* An extra reference on `rf' has been held for us by falloc(). */
 352         fildes[0] = fd;
 353
 354         /*
 355          * Warning: once we've gotten past allocation of the fd for the
 356          * read-side, we can only drop the read side via fdrop() in order
 357          * to avoid races against processes which manage to dup() the read
 358          * side while we are blocked trying to allocate the write side.
 359          */
 360         finit(rf, FREAD | FWRITE, DTYPE_PIPE, rpipe, &pipeops);
 361         error = falloc(td, &wf, &fd);
 362         if (error) {
 363                 fdclose(fdp, rf, fildes[0], td);
 364                 fdrop(rf, td);
 365                 /* rpipe has been closed by fdrop(). */
 366                 pipeclose(wpipe);
 367                 return (error);
 368         }
 369         /* An extra reference on `wf' has been held for us by falloc(). */
 370         finit(wf, FREAD | FWRITE, DTYPE_PIPE, wpipe, &pipeops);
 371         fdrop(wf, td);
 372         fildes[1] = fd;
 373         fdrop(rf, td);
 374
 375         return (0);
 376 }
 377
 378 /* ARGSUSED */
 379 int
 380 pipe(struct thread *td, struct pipe_args *uap)
 381 {
 382         int error;
 383         int fildes[2];
 384
 385         error = kern_pipe(td, fildes);
 386         if (error)
 387                 return (error);
 388
 389         td->td_retval[0] = fildes[0];
 390         td->td_retval[1] = fildes[1];
 391
 392         return (0);
 393 }
 394
 395 /*
 396  * Allocate kva for pipe circular buffer, the space is pageable
 397  * This routine will 'realloc' the size of a pipe safely, if it fails
 398  * it will retain the old buffer.
 399  * If it fails it will return ENOMEM.
 400  */
 401 static int
 402 pipespace_new(cpipe, size)
 403         struct pipe *cpipe;
 404         int size;
 405 {
 406         caddr_t buffer;
 407         int error, cnt, firstseg;
 408         static int curfail = 0;
 409         static struct timeval lastfail;
 410
 411         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 412         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 413                 ("pipespace: resize of direct writes not allowed"));
 414 retry:
 415         cnt = cpipe->pipe_buffer.cnt;
 416         if (cnt > size)
 417                 size = cnt;
 418
 419         size = round_page(size);
 420         buffer = (caddr_t) vm_map_min(pipe_map);
 421
 422         error = vm_map_find(pipe_map, NULL, 0,
 423                 (vm_offset_t *) &buffer, size, 1,
 424                 VM_PROT_ALL, VM_PROT_ALL, 0);
 425         if (error != KERN_SUCCESS) {
 426                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 427                         (size > SMALL_PIPE_SIZE)) {
 428                         size = SMALL_PIPE_SIZE;
 429                         pipefragretry++;
 430                         goto retry;
 431                 }
 432                 if (cpipe->pipe_buffer.buffer == NULL) {
 433                         pipeallocfail++;
 434                         if (ppsratecheck(&lastfail, &curfail, 1))
 435                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 436                 } else {
 437                         piperesizefail++;
 438                 }
 439                 return (ENOMEM);
 440         }
 441
 442         /* copy data, then free old resources if we're resizing */
 443         if (cnt > 0) {
 444                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 445                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 446                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 447                                 buffer, firstseg);
 448                         if ((cnt - firstseg) > 0)
 449                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 450                                         cpipe->pipe_buffer.in);
 451                 } else {
 452                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 453                                 buffer, cnt);
 454                 }
 455         }
 456         pipe_free_kmem(cpipe);
 457         cpipe->pipe_buffer.buffer = buffer;
 458         cpipe->pipe_buffer.size = size;
 459         cpipe->pipe_buffer.in = cnt;
 460         cpipe->pipe_buffer.out = 0;
 461         cpipe->pipe_buffer.cnt = cnt;
 462         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 463         return (0);
 464 }
 465
 466 /*
 467  * Wrapper for pipespace_new() that performs locking assertions.
 468  */
 469 static int
 470 pipespace(cpipe, size)
 471         struct pipe *cpipe;
 472         int size;
 473 {
 474
 475         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 476                 ("Unlocked pipe passed to pipespace"));
 477         return (pipespace_new(cpipe, size));
 478 }
 479
 480 /*
 481  * lock a pipe for I/O, blocking other access
 482  */
 483 static __inline int
 484 pipelock(cpipe, catch)
 485         struct pipe *cpipe;
 486         int catch;
 487 {
 488         int error;
 489
 490         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 491         while (cpipe->pipe_state & PIPE_LOCKFL) {
 492                 cpipe->pipe_state |= PIPE_LWANT;
 493                 error = msleep(cpipe, PIPE_MTX(cpipe),
 494                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 495                     "pipelk", 0);
 496                 if (error != 0)
 497                         return (error);
 498         }
 499         cpipe->pipe_state |= PIPE_LOCKFL;
 500         return (0);
 501 }
 502
 503 /*
 504  * unlock a pipe I/O lock
 505  */
 506 static __inline void
 507 pipeunlock(cpipe)
 508         struct pipe *cpipe;
 509 {
 510
 511         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 512         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 513                 ("Unlocked pipe passed to pipeunlock"));
 514         cpipe->pipe_state &= ~PIPE_LOCKFL;
 515         if (cpipe->pipe_state & PIPE_LWANT) {
 516                 cpipe->pipe_state &= ~PIPE_LWANT;
 517                 wakeup(cpipe);
 518         }
 519 }
 520
 521 static __inline void
 522 pipeselwakeup(cpipe)
 523         struct pipe *cpipe;
 524 {
 525
 526         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 527         if (cpipe->pipe_state & PIPE_SEL) {
 528                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 529                 if (!SEL_WAITING(&cpipe->pipe_sel))
 530                         cpipe->pipe_state &= ~PIPE_SEL;
 531         }
 532         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 533                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 534         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 535 }
 536
 537 /*
 538  * Initialize and allocate VM and memory for pipe.  The structure
 539  * will start out zero'd from the ctor, so we just manage the kmem.
 540  */
 541 static int
 542 pipe_create(pipe, backing)
 543         struct pipe *pipe;
 544         int backing;
 545 {
 546         int error;
 547
 548         if (backing) {
 549                 if (amountpipekva > maxpipekva / 2)
 550                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 551                 else
 552                         error = pipespace_new(pipe, PIPE_SIZE);
 553         } else {
 554                 /* If we're not backing this pipe, no need to do anything. */
 555                 error = 0;
 556         }
 557         return (error);
 558 }
 559
 560 /* ARGSUSED */
 561 static int
 562 pipe_read(fp, uio, active_cred, flags, td)
 563         struct file *fp;
 564         struct uio *uio;
 565         struct ucred *active_cred;
 566         struct thread *td;
 567         int flags;
 568 {
 569         struct pipe *rpipe = fp->f_data;
 570         int error;
 571         int nread = 0;
 572         u_int size;
 573
 574         PIPE_LOCK(rpipe);
 575         ++rpipe->pipe_busy;
 576         error = pipelock(rpipe, 1);
 577         if (error)
 578                 goto unlocked_error;
 579
 580 #ifdef MAC
 581         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 582         if (error)
 583                 goto locked_error;
 584 #endif
 585         if (amountpipekva > (3 * maxpipekva) / 4) {
 586                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 587                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 588                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 589                         (piperesizeallowed == 1)) {
 590                         PIPE_UNLOCK(rpipe);
 591                         pipespace(rpipe, SMALL_PIPE_SIZE);
 592                         PIPE_LOCK(rpipe);
 593                 }
 594         }
 595
 596         while (uio->uio_resid) {
 597                 /*
 598                  * normal pipe buffer receive
 599                  */
 600                 if (rpipe->pipe_buffer.cnt > 0) {
 601                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 602                         if (size > rpipe->pipe_buffer.cnt)
 603                                 size = rpipe->pipe_buffer.cnt;
 604                         if (size > (u_int) uio->uio_resid)
 605                                 size = (u_int) uio->uio_resid;
 606
 607                         PIPE_UNLOCK(rpipe);
 608                         error = uiomove(
 609                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 610                             size, uio);
 611                         PIPE_LOCK(rpipe);
 612                         if (error)
 613                                 break;
 614
 615                         rpipe->pipe_buffer.out += size;
 616                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 617                                 rpipe->pipe_buffer.out = 0;
 618
 619                         rpipe->pipe_buffer.cnt -= size;
 620
 621                         /*
 622                          * If there is no more to read in the pipe, reset
 623                          * its pointers to the beginning.  This improves
 624                          * cache hit stats.
 625                          */
 626                         if (rpipe->pipe_buffer.cnt == 0) {
 627                                 rpipe->pipe_buffer.in = 0;
 628                                 rpipe->pipe_buffer.out = 0;
 629                         }
 630                         nread += size;
 631 #ifndef PIPE_NODIRECT
 632                 /*
 633                  * Direct copy, bypassing a kernel buffer.
 634                  */
 635                 } else if ((size = rpipe->pipe_map.cnt) &&
 636                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 637                         if (size > (u_int) uio->uio_resid)
 638                                 size = (u_int) uio->uio_resid;
 639
 640                         PIPE_UNLOCK(rpipe);
 641                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 642                             rpipe->pipe_map.pos, size, uio);
 643                         PIPE_LOCK(rpipe);
 644                         if (error)
 645                                 break;
 646                         nread += size;
 647                         rpipe->pipe_map.pos += size;
 648                         rpipe->pipe_map.cnt -= size;
 649                         if (rpipe->pipe_map.cnt == 0) {
 650                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 651                                 wakeup(rpipe);
 652                         }
 653 #endif
 654                 } else {
 655                         /*
 656                          * detect EOF condition
 657                          * read returns 0 on EOF, no need to set error
 658                          */
 659                         if (rpipe->pipe_state & PIPE_EOF)
 660                                 break;
 661
 662                         /*
 663                          * If the "write-side" has been blocked, wake it up now.
 664                          */
 665                         if (rpipe->pipe_state & PIPE_WANTW) {
 666                                 rpipe->pipe_state &= ~PIPE_WANTW;
 667                                 wakeup(rpipe);
 668                         }
 669
 670                         /*
 671                          * Break if some data was read.
 672                          */
 673                         if (nread > 0)
 674                                 break;
 675
 676                         /*
 677                          * Unlock the pipe buffer for our remaining processing.
 678                          * We will either break out with an error or we will
 679                          * sleep and relock to loop.
 680                          */
 681                         pipeunlock(rpipe);
 682
 683                         /*
 684                          * Handle non-blocking mode operation or
 685                          * wait for more data.
 686                          */
 687                         if (fp->f_flag & FNONBLOCK) {
 688                                 error = EAGAIN;
 689                         } else {
 690                                 rpipe->pipe_state |= PIPE_WANTR;
 691                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 692                                     PRIBIO | PCATCH,
 693                                     "piperd", 0)) == 0)
 694                                         error = pipelock(rpipe, 1);
 695                         }
 696                         if (error)
 697                                 goto unlocked_error;
 698                 }
 699         }
 700 #ifdef MAC
 701 locked_error:
 702 #endif
 703         pipeunlock(rpipe);
 704
 705         /* XXX: should probably do this before getting any locks. */
 706         if (error == 0)
 707                 vfs_timestamp(&rpipe->pipe_atime);
 708 unlocked_error:
 709         --rpipe->pipe_busy;
 710
 711         /*
 712          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 713          */
 714         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 715                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 716                 wakeup(rpipe);
 717         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 718                 /*
 719                  * Handle write blocking hysteresis.
 720                  */
 721                 if (rpipe->pipe_state & PIPE_WANTW) {
 722                         rpipe->pipe_state &= ~PIPE_WANTW;
 723                         wakeup(rpipe);
 724                 }
 725         }
 726
 727         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 728                 pipeselwakeup(rpipe);
 729
 730         PIPE_UNLOCK(rpipe);
 731         return (error);
 732 }
 733
 734 #ifndef PIPE_NODIRECT
 735 /*
 736  * Map the sending processes' buffer into kernel space and wire it.
 737  * This is similar to a physical write operation.
 738  */
 739 static int
 740 pipe_build_write_buffer(wpipe, uio)
 741         struct pipe *wpipe;
 742         struct uio *uio;
 743 {
 744         pmap_t pmap;
 745         u_int size;
 746         int i, j;
 747         vm_offset_t addr, endaddr;
 748
 749         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 750         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 751                 ("Clone attempt on non-direct write pipe!"));
 752
 753         size = (u_int) uio->uio_iov->iov_len;
 754         if (size > wpipe->pipe_buffer.size)
 755                 size = wpipe->pipe_buffer.size;
 756
 757         pmap = vmspace_pmap(curproc->p_vmspace);
 758         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 759         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 760         if (endaddr < addr)
 761                 return (EFAULT);
 762         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 763                 /*
 764                  * vm_fault_quick() can sleep.  Consequently,
 765                  * vm_page_lock_queue() and vm_page_unlock_queue()
 766                  * should not be performed outside of this loop.
 767                  */
 768         race:
 769                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0) {
 770                         vm_page_lock_queues();
 771                         for (j = 0; j < i; j++)
 772                                 vm_page_unhold(wpipe->pipe_map.ms[j]);
 773                         vm_page_unlock_queues();
 774                         return (EFAULT);
 775                 }
 776                 wpipe->pipe_map.ms[i] = pmap_extract_and_hold(pmap, addr,
 777                     VM_PROT_READ);
 778                 if (wpipe->pipe_map.ms[i] == NULL)
 779                         goto race;
 780         }
 781
 782 /*
 783  * set up the control block
 784  */
 785         wpipe->pipe_map.npages = i;
 786         wpipe->pipe_map.pos =
 787             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 788         wpipe->pipe_map.cnt = size;
 789
 790 /*
 791  * and update the uio data
 792  */
 793
 794         uio->uio_iov->iov_len -= size;
 795         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 796         if (uio->uio_iov->iov_len == 0)
 797                 uio->uio_iov++;
 798         uio->uio_resid -= size;
 799         uio->uio_offset += size;
 800         return (0);
 801 }
 802
 803 /*
 804  * unmap and unwire the process buffer
 805  */
 806 static void
 807 pipe_destroy_write_buffer(wpipe)
 808         struct pipe *wpipe;
 809 {
 810         int i;
 811
 812         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 813         vm_page_lock_queues();
 814         for (i = 0; i < wpipe->pipe_map.npages; i++) {
 815                 vm_page_unhold(wpipe->pipe_map.ms[i]);
 816         }
 817         vm_page_unlock_queues();
 818         wpipe->pipe_map.npages = 0;
 819 }
 820
 821 /*
 822  * In the case of a signal, the writing process might go away.  This
 823  * code copies the data into the circular buffer so that the source
 824  * pages can be freed without loss of data.
 825  */
 826 static void
 827 pipe_clone_write_buffer(wpipe)
 828         struct pipe *wpipe;
 829 {
 830         struct uio uio;
 831         struct iovec iov;
 832         int size;
 833         int pos;
 834
 835         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 836         size = wpipe->pipe_map.cnt;
 837         pos = wpipe->pipe_map.pos;
 838
 839         wpipe->pipe_buffer.in = size;
 840         wpipe->pipe_buffer.out = 0;
 841         wpipe->pipe_buffer.cnt = size;
 842         wpipe->pipe_state &= ~PIPE_DIRECTW;
 843
 844         PIPE_UNLOCK(wpipe);
 845         iov.iov_base = wpipe->pipe_buffer.buffer;
 846         iov.iov_len = size;
 847         uio.uio_iov = &iov;
 848         uio.uio_iovcnt = 1;
 849         uio.uio_offset = 0;
 850         uio.uio_resid = size;
 851         uio.uio_segflg = UIO_SYSSPACE;
 852         uio.uio_rw = UIO_READ;
 853         uio.uio_td = curthread;
 854         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 855         PIPE_LOCK(wpipe);
 856         pipe_destroy_write_buffer(wpipe);
 857 }
 858
 859 /*
 860  * This implements the pipe buffer write mechanism.  Note that only
 861  * a direct write OR a normal pipe write can be pending at any given time.
 862  * If there are any characters in the pipe buffer, the direct write will
 863  * be deferred until the receiving process grabs all of the bytes from
 864  * the pipe buffer.  Then the direct mapping write is set-up.
 865  */
 866 static int
 867 pipe_direct_write(wpipe, uio)
 868         struct pipe *wpipe;
 869         struct uio *uio;
 870 {
 871         int error;
 872
 873 retry:
 874         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 875         error = pipelock(wpipe, 1);
 876         if (wpipe->pipe_state & PIPE_EOF)
 877                 error = EPIPE;
 878         if (error) {
 879                 pipeunlock(wpipe);
 880                 goto error1;
 881         }
 882         while (wpipe->pipe_state & PIPE_DIRECTW) {
 883                 if (wpipe->pipe_state & PIPE_WANTR) {
 884                         wpipe->pipe_state &= ~PIPE_WANTR;
 885                         wakeup(wpipe);
 886                 }
 887                 pipeselwakeup(wpipe);
 888                 wpipe->pipe_state |= PIPE_WANTW;
 889                 pipeunlock(wpipe);
 890                 error = msleep(wpipe, PIPE_MTX(wpipe),
 891                     PRIBIO | PCATCH, "pipdww", 0);
 892                 if (error)
 893                         goto error1;
 894                 else
 895                         goto retry;
 896         }
 897         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 898         if (wpipe->pipe_buffer.cnt > 0) {
 899                 if (wpipe->pipe_state & PIPE_WANTR) {
 900                         wpipe->pipe_state &= ~PIPE_WANTR;
 901                         wakeup(wpipe);
 902                 }
 903                 pipeselwakeup(wpipe);
 904                 wpipe->pipe_state |= PIPE_WANTW;
 905                 pipeunlock(wpipe);
 906                 error = msleep(wpipe, PIPE_MTX(wpipe),
 907                     PRIBIO | PCATCH, "pipdwc", 0);
 908                 if (error)
 909                         goto error1;
 910                 else
 911                         goto retry;
 912         }
 913
 914         wpipe->pipe_state |= PIPE_DIRECTW;
 915
 916         PIPE_UNLOCK(wpipe);
 917         error = pipe_build_write_buffer(wpipe, uio);
 918         PIPE_LOCK(wpipe);
 919         if (error) {
 920                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 921                 pipeunlock(wpipe);
 922                 goto error1;
 923         }
 924
 925         error = 0;
 926         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 927                 if (wpipe->pipe_state & PIPE_EOF) {
 928                         pipe_destroy_write_buffer(wpipe);
 929                         pipeselwakeup(wpipe);
 930                         pipeunlock(wpipe);
 931                         error = EPIPE;
 932                         goto error1;
 933                 }
 934                 if (wpipe->pipe_state & PIPE_WANTR) {
 935                         wpipe->pipe_state &= ~PIPE_WANTR;
 936                         wakeup(wpipe);
 937                 }
 938                 pipeselwakeup(wpipe);
 939                 pipeunlock(wpipe);
 940                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 941                     "pipdwt", 0);
 942                 pipelock(wpipe, 0);
 943         }
 944
 945         if (wpipe->pipe_state & PIPE_EOF)
 946                 error = EPIPE;
 947         if (wpipe->pipe_state & PIPE_DIRECTW) {
 948                 /*
 949                  * this bit of trickery substitutes a kernel buffer for
 950                  * the process that might be going away.
 951                  */
 952                 pipe_clone_write_buffer(wpipe);
 953         } else {
 954                 pipe_destroy_write_buffer(wpipe);
 955         }
 956         pipeunlock(wpipe);
 957         return (error);
 958
 959 error1:
 960         wakeup(wpipe);
 961         return (error);
 962 }
 963 #endif
 964
 965 static int
 966 pipe_write(fp, uio, active_cred, flags, td)
 967         struct file *fp;
 968         struct uio *uio;
 969         struct ucred *active_cred;
 970         struct thread *td;
 971         int flags;
 972 {
 973         int error = 0;
 974         int desiredsize, orig_resid;
 975         struct pipe *wpipe, *rpipe;
 976
 977         rpipe = fp->f_data;
 978         wpipe = rpipe->pipe_peer;
 979
 980         PIPE_LOCK(rpipe);
 981         error = pipelock(wpipe, 1);
 982         if (error) {
 983                 PIPE_UNLOCK(rpipe);
 984                 return (error);
 985         }
 986         /*
 987          * detect loss of pipe read side, issue SIGPIPE if lost.
 988          */
 989         if (wpipe->pipe_present != PIPE_ACTIVE ||
 990             (wpipe->pipe_state & PIPE_EOF)) {
 991                 pipeunlock(wpipe);
 992                 PIPE_UNLOCK(rpipe);
 993                 return (EPIPE);
 994         }
 995 #ifdef MAC
 996         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
 997         if (error) {
 998                 pipeunlock(wpipe);
 999                 PIPE_UNLOCK(rpipe);
1000                 return (error);
1001         }
1002 #endif
1003         ++wpipe->pipe_busy;
1004
1005         /* Choose a larger size if it's advantageous */
1006         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1007         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1008                 if (piperesizeallowed != 1)
1009                         break;
1010                 if (amountpipekva > maxpipekva / 2)
1011                         break;
1012                 if (desiredsize == BIG_PIPE_SIZE)
1013                         break;
1014                 desiredsize = desiredsize * 2;
1015         }
1016
1017         /* Choose a smaller size if we're in a OOM situation */
1018         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1019                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1020                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1021                 (piperesizeallowed == 1))
1022                 desiredsize = SMALL_PIPE_SIZE;
1023
1024         /* Resize if the above determined that a new size was necessary */
1025         if ((desiredsize != wpipe->pipe_buffer.size) &&
1026                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1027                 PIPE_UNLOCK(wpipe);
1028                 pipespace(wpipe, desiredsize);
1029                 PIPE_LOCK(wpipe);
1030         }
1031         if (wpipe->pipe_buffer.size == 0) {
1032                 /*
1033                  * This can only happen for reverse direction use of pipes
1034                  * in a complete OOM situation.
1035                  */
1036                 error = ENOMEM;
1037                 --wpipe->pipe_busy;
1038                 pipeunlock(wpipe);
1039                 PIPE_UNLOCK(wpipe);
1040                 return (error);
1041         }
1042
1043         pipeunlock(wpipe);
1044
1045         orig_resid = uio->uio_resid;
1046
1047         while (uio->uio_resid) {
1048                 int space;
1049
1050                 pipelock(wpipe, 0);
1051                 if (wpipe->pipe_state & PIPE_EOF) {
1052                         pipeunlock(wpipe);
1053                         error = EPIPE;
1054                         break;
1055                 }
1056 #ifndef PIPE_NODIRECT
1057                 /*
1058                  * If the transfer is large, we can gain performance if
1059                  * we do process-to-process copies directly.
1060                  * If the write is non-blocking, we don't use the
1061                  * direct write mechanism.
1062                  *
1063                  * The direct write mechanism will detect the reader going
1064                  * away on us.
1065                  */
1066                 if (uio->uio_segflg == UIO_USERSPACE &&
1067                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1068                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1069                     (fp->f_flag & FNONBLOCK) == 0) {
1070                         pipeunlock(wpipe);
1071                         error = pipe_direct_write(wpipe, uio);
1072                         if (error)
1073                                 break;
1074                         continue;
1075                 }
1076 #endif
1077
1078                 /*
1079                  * Pipe buffered writes cannot be coincidental with
1080                  * direct writes.  We wait until the currently executing
1081                  * direct write is completed before we start filling the
1082                  * pipe buffer.  We break out if a signal occurs or the
1083                  * reader goes away.
1084                  */
1085                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1086                         if (wpipe->pipe_state & PIPE_WANTR) {
1087                                 wpipe->pipe_state &= ~PIPE_WANTR;
1088                                 wakeup(wpipe);
1089                         }
1090                         pipeselwakeup(wpipe);
1091                         wpipe->pipe_state |= PIPE_WANTW;
1092                         pipeunlock(wpipe);
1093                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1094                             "pipbww", 0);
1095                         if (error)
1096                                 break;
1097                         else
1098                                 continue;
1099                 }
1100
1101                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1102
1103                 /* Writes of size <= PIPE_BUF must be atomic. */
1104                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1105                         space = 0;
1106
1107                 if (space > 0) {
1108                         int size;       /* Transfer size */
1109                         int segsize;    /* first segment to transfer */
1110
1111                         /*
1112                          * Transfer size is minimum of uio transfer
1113                          * and free space in pipe buffer.
1114                          */
1115                         if (space > uio->uio_resid)
1116                                 size = uio->uio_resid;
1117                         else
1118                                 size = space;
1119                         /*
1120                          * First segment to transfer is minimum of
1121                          * transfer size and contiguous space in
1122                          * pipe buffer.  If first segment to transfer
1123                          * is less than the transfer size, we've got
1124                          * a wraparound in the buffer.
1125                          */
1126                         segsize = wpipe->pipe_buffer.size -
1127                                 wpipe->pipe_buffer.in;
1128                         if (segsize > size)
1129                                 segsize = size;
1130
1131                         /* Transfer first segment */
1132
1133                         PIPE_UNLOCK(rpipe);
1134                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1135                                         segsize, uio);
1136                         PIPE_LOCK(rpipe);
1137
1138                         if (error == 0 && segsize < size) {
1139                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1140                                         wpipe->pipe_buffer.size,
1141                                         ("Pipe buffer wraparound disappeared"));
1142                                 /*
1143                                  * Transfer remaining part now, to
1144                                  * support atomic writes.  Wraparound
1145                                  * happened.
1146                                  */
1147
1148                                 PIPE_UNLOCK(rpipe);
1149                                 error = uiomove(
1150                                     &wpipe->pipe_buffer.buffer[0],
1151                                     size - segsize, uio);
1152                                 PIPE_LOCK(rpipe);
1153                         }
1154                         if (error == 0) {
1155                                 wpipe->pipe_buffer.in += size;
1156                                 if (wpipe->pipe_buffer.in >=
1157                                     wpipe->pipe_buffer.size) {
1158                                         KASSERT(wpipe->pipe_buffer.in ==
1159                                                 size - segsize +
1160                                                 wpipe->pipe_buffer.size,
1161                                                 ("Expected wraparound bad"));
1162                                         wpipe->pipe_buffer.in = size - segsize;
1163                                 }
1164
1165                                 wpipe->pipe_buffer.cnt += size;
1166                                 KASSERT(wpipe->pipe_buffer.cnt <=
1167                                         wpipe->pipe_buffer.size,
1168                                         ("Pipe buffer overflow"));
1169                         }
1170                         pipeunlock(wpipe);
1171                         if (error != 0)
1172                                 break;
1173                 } else {
1174                         /*
1175                          * If the "read-side" has been blocked, wake it up now.
1176                          */
1177                         if (wpipe->pipe_state & PIPE_WANTR) {
1178                                 wpipe->pipe_state &= ~PIPE_WANTR;
1179                                 wakeup(wpipe);
1180                         }
1181
1182                         /*
1183                          * don't block on non-blocking I/O
1184                          */
1185                         if (fp->f_flag & FNONBLOCK) {
1186                                 error = EAGAIN;
1187                                 pipeunlock(wpipe);
1188                                 break;
1189                         }
1190
1191                         /*
1192                          * We have no more space and have something to offer,
1193                          * wake up select/poll.
1194                          */
1195                         pipeselwakeup(wpipe);
1196
1197                         wpipe->pipe_state |= PIPE_WANTW;
1198                         pipeunlock(wpipe);
1199                         error = msleep(wpipe, PIPE_MTX(rpipe),
1200                             PRIBIO | PCATCH, "pipewr", 0);
1201                         if (error != 0)
1202                                 break;
1203                 }
1204         }
1205
1206         pipelock(wpipe, 0);
1207         --wpipe->pipe_busy;
1208
1209         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1210                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1211                 wakeup(wpipe);
1212         } else if (wpipe->pipe_buffer.cnt > 0) {
1213                 /*
1214                  * If we have put any characters in the buffer, we wake up
1215                  * the reader.
1216                  */
1217                 if (wpipe->pipe_state & PIPE_WANTR) {
1218                         wpipe->pipe_state &= ~PIPE_WANTR;
1219                         wakeup(wpipe);
1220                 }
1221         }
1222
1223         /*
1224          * Don't return EPIPE if I/O was successful
1225          */
1226         if ((wpipe->pipe_buffer.cnt == 0) &&
1227             (uio->uio_resid == 0) &&
1228             (error == EPIPE)) {
1229                 error = 0;
1230         }
1231
1232         if (error == 0)
1233                 vfs_timestamp(&wpipe->pipe_mtime);
1234
1235         /*
1236          * We have something to offer,
1237          * wake up select/poll.
1238          */
1239         if (wpipe->pipe_buffer.cnt)
1240                 pipeselwakeup(wpipe);
1241
1242         pipeunlock(wpipe);
1243         PIPE_UNLOCK(rpipe);
1244         return (error);
1245 }
1246
1247 /* ARGSUSED */
1248 static int
1249 pipe_truncate(fp, length, active_cred, td)
1250         struct file *fp;
1251         off_t length;
1252         struct ucred *active_cred;
1253         struct thread *td;
1254 {
1255
1256         return (EINVAL);
1257 }
1258
1259 /*
1260  * we implement a very minimal set of ioctls for compatibility with sockets.
1261  */
1262 static int
1263 pipe_ioctl(fp, cmd, data, active_cred, td)
1264         struct file *fp;
1265         u_long cmd;
1266         void *data;
1267         struct ucred *active_cred;
1268         struct thread *td;
1269 {
1270         struct pipe *mpipe = fp->f_data;
1271         int error;
1272
1273         PIPE_LOCK(mpipe);
1274
1275 #ifdef MAC
1276         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1277         if (error) {
1278                 PIPE_UNLOCK(mpipe);
1279                 return (error);
1280         }
1281 #endif
1282
1283         error = 0;
1284         switch (cmd) {
1285
1286         case FIONBIO:
1287                 break;
1288
1289         case FIOASYNC:
1290                 if (*(int *)data) {
1291                         mpipe->pipe_state |= PIPE_ASYNC;
1292                 } else {
1293                         mpipe->pipe_state &= ~PIPE_ASYNC;
1294                 }
1295                 break;
1296
1297         case FIONREAD:
1298                 if (mpipe->pipe_state & PIPE_DIRECTW)
1299                         *(int *)data = mpipe->pipe_map.cnt;
1300                 else
1301                         *(int *)data = mpipe->pipe_buffer.cnt;
1302                 break;
1303
1304         case FIOSETOWN:
1305                 PIPE_UNLOCK(mpipe);
1306                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1307                 goto out_unlocked;
1308
1309         case FIOGETOWN:
1310                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1311                 break;
1312
1313         /* This is deprecated, FIOSETOWN should be used instead. */
1314         case TIOCSPGRP:
1315                 PIPE_UNLOCK(mpipe);
1316                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1317                 goto out_unlocked;
1318
1319         /* This is deprecated, FIOGETOWN should be used instead. */
1320         case TIOCGPGRP:
1321                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1322                 break;
1323
1324         default:
1325                 error = ENOTTY;
1326                 break;
1327         }
1328         PIPE_UNLOCK(mpipe);
1329 out_unlocked:
1330         return (error);
1331 }
1332
1333 static int
1334 pipe_poll(fp, events, active_cred, td)
1335         struct file *fp;
1336         int events;
1337         struct ucred *active_cred;
1338         struct thread *td;
1339 {
1340         struct pipe *rpipe = fp->f_data;
1341         struct pipe *wpipe;
1342         int revents = 0;
1343 #ifdef MAC
1344         int error;
1345 #endif
1346
1347         wpipe = rpipe->pipe_peer;
1348         PIPE_LOCK(rpipe);
1349 #ifdef MAC
1350         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1351         if (error)
1352                 goto locked_error;
1353 #endif
1354         if (events & (POLLIN | POLLRDNORM))
1355                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1356                     (rpipe->pipe_buffer.cnt > 0))
1357                         revents |= events & (POLLIN | POLLRDNORM);
1358
1359         if (events & (POLLOUT | POLLWRNORM))
1360                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1361                     (wpipe->pipe_state & PIPE_EOF) ||
1362                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1363                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1364                         revents |= events & (POLLOUT | POLLWRNORM);
1365
1366         if ((events & POLLINIGNEOF) == 0) {
1367                 if (rpipe->pipe_state & PIPE_EOF) {
1368                         revents |= (events & (POLLIN | POLLRDNORM));
1369                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1370                             (wpipe->pipe_state & PIPE_EOF))
1371                                 revents |= POLLHUP;
1372                 }
1373         }
1374
1375         if (revents == 0) {
1376                 if (events & (POLLIN | POLLRDNORM)) {
1377                         selrecord(td, &rpipe->pipe_sel);
1378                         if (SEL_WAITING(&rpipe->pipe_sel))
1379                                 rpipe->pipe_state |= PIPE_SEL;
1380                 }
1381
1382                 if (events & (POLLOUT | POLLWRNORM)) {
1383                         selrecord(td, &wpipe->pipe_sel);
1384                         if (SEL_WAITING(&wpipe->pipe_sel))
1385                                 wpipe->pipe_state |= PIPE_SEL;
1386                 }
1387         }
1388 #ifdef MAC
1389 locked_error:
1390 #endif
1391         PIPE_UNLOCK(rpipe);
1392
1393         return (revents);
1394 }
1395
1396 /*
1397  * We shouldn't need locks here as we're doing a read and this should
1398  * be a natural race.
1399  */
1400 static int
1401 pipe_stat(fp, ub, active_cred, td)
1402         struct file *fp;
1403         struct stat *ub;
1404         struct ucred *active_cred;
1405         struct thread *td;
1406 {
1407         struct pipe *pipe = fp->f_data;
1408 #ifdef MAC
1409         int error;
1410
1411         PIPE_LOCK(pipe);
1412         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1413         PIPE_UNLOCK(pipe);
1414         if (error)
1415                 return (error);
1416 #endif
1417         bzero(ub, sizeof(*ub));
1418         ub->st_mode = S_IFIFO;
1419         ub->st_blksize = PAGE_SIZE;
1420         if (pipe->pipe_state & PIPE_DIRECTW)
1421                 ub->st_size = pipe->pipe_map.cnt;
1422         else
1423                 ub->st_size = pipe->pipe_buffer.cnt;
1424         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1425         ub->st_atimespec = pipe->pipe_atime;
1426         ub->st_mtimespec = pipe->pipe_mtime;
1427         ub->st_ctimespec = pipe->pipe_ctime;
1428         ub->st_uid = fp->f_cred->cr_uid;
1429         ub->st_gid = fp->f_cred->cr_gid;
1430         /*
1431          * Left as 0: st_dev, st_ino, st_nlink, st_rdev, st_flags, st_gen.
1432          * XXX (st_dev, st_ino) should be unique.
1433          */
1434         return (0);
1435 }
1436
1437 /* ARGSUSED */
1438 static int
1439 pipe_close(fp, td)
1440         struct file *fp;
1441         struct thread *td;
1442 {
1443         struct pipe *cpipe = fp->f_data;
1444
1445         fp->f_ops = &badfileops;
1446         fp->f_data = NULL;
1447         funsetown(&cpipe->pipe_sigio);
1448         pipeclose(cpipe);
1449         return (0);
1450 }
1451
1452 static void
1453 pipe_free_kmem(cpipe)
1454         struct pipe *cpipe;
1455 {
1456
1457         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1458             ("pipe_free_kmem: pipe mutex locked"));
1459
1460         if (cpipe->pipe_buffer.buffer != NULL) {
1461                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1462                 vm_map_remove(pipe_map,
1463                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1464                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1465                 cpipe->pipe_buffer.buffer = NULL;
1466         }
1467 #ifndef PIPE_NODIRECT
1468         {
1469                 cpipe->pipe_map.cnt = 0;
1470                 cpipe->pipe_map.pos = 0;
1471                 cpipe->pipe_map.npages = 0;
1472         }
1473 #endif
1474 }
1475
1476 /*
1477  * shutdown the pipe
1478  */
1479 static void
1480 pipeclose(cpipe)
1481         struct pipe *cpipe;
1482 {
1483         struct pipepair *pp;
1484         struct pipe *ppipe;
1485
1486         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1487
1488         PIPE_LOCK(cpipe);
1489         pipelock(cpipe, 0);
1490         pp = cpipe->pipe_pair;
1491
1492         pipeselwakeup(cpipe);
1493
1494         /*
1495          * If the other side is blocked, wake it up saying that
1496          * we want to close it down.
1497          */
1498         cpipe->pipe_state |= PIPE_EOF;
1499         while (cpipe->pipe_busy) {
1500                 wakeup(cpipe);
1501                 cpipe->pipe_state |= PIPE_WANT;
1502                 pipeunlock(cpipe);
1503                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1504                 pipelock(cpipe, 0);
1505         }
1506
1507
1508         /*
1509          * Disconnect from peer, if any.
1510          */
1511         ppipe = cpipe->pipe_peer;
1512         if (ppipe->pipe_present == PIPE_ACTIVE) {
1513                 pipeselwakeup(ppipe);
1514
1515                 ppipe->pipe_state |= PIPE_EOF;
1516                 wakeup(ppipe);
1517                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1518         }
1519
1520         /*
1521          * Mark this endpoint as free.  Release kmem resources.  We
1522          * don't mark this endpoint as unused until we've finished
1523          * doing that, or the pipe might disappear out from under
1524          * us.
1525          */
1526         PIPE_UNLOCK(cpipe);
1527         pipe_free_kmem(cpipe);
1528         PIPE_LOCK(cpipe);
1529         cpipe->pipe_present = PIPE_CLOSING;
1530         pipeunlock(cpipe);
1531
1532         /*
1533          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1534          * PIPE_FINALIZED, that allows other end to free the
1535          * pipe_pair, only after the knotes are completely dismantled.
1536          */
1537         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1538         cpipe->pipe_present = PIPE_FINALIZED;
1539         knlist_destroy(&cpipe->pipe_sel.si_note);
1540
1541         /*
1542          * If both endpoints are now closed, release the memory for the
1543          * pipe pair.  If not, unlock.
1544          */
1545         if (ppipe->pipe_present == PIPE_FINALIZED) {
1546                 PIPE_UNLOCK(cpipe);
1547 #ifdef MAC
1548                 mac_pipe_destroy(pp);
1549 #endif
1550                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1551         } else
1552                 PIPE_UNLOCK(cpipe);
1553 }
1554
1555 /*ARGSUSED*/
1556 static int
1557 pipe_kqfilter(struct file *fp, struct knote *kn)
1558 {
1559         struct pipe *cpipe;
1560
1561         cpipe = kn->kn_fp->f_data;
1562         PIPE_LOCK(cpipe);
1563         switch (kn->kn_filter) {
1564         case EVFILT_READ:
1565                 kn->kn_fop = &pipe_rfiltops;
1566                 break;
1567         case EVFILT_WRITE:
1568                 kn->kn_fop = &pipe_wfiltops;
1569                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1570                         /* other end of pipe has been closed */
1571                         PIPE_UNLOCK(cpipe);
1572                         return (EPIPE);
1573                 }
1574                 cpipe = cpipe->pipe_peer;
1575                 break;
1576         default:
1577                 PIPE_UNLOCK(cpipe);
1578                 return (EINVAL);
1579         }
1580
1581         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1582         PIPE_UNLOCK(cpipe);
1583         return (0);
1584 }
1585
1586 static void
1587 filt_pipedetach(struct knote *kn)
1588 {
1589         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1590
1591         PIPE_LOCK(cpipe);
1592         if (kn->kn_filter == EVFILT_WRITE)
1593                 cpipe = cpipe->pipe_peer;
1594         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1595         PIPE_UNLOCK(cpipe);
1596 }
1597
1598 /*ARGSUSED*/
1599 static int
1600 filt_piperead(struct knote *kn, long hint)
1601 {
1602         struct pipe *rpipe = kn->kn_fp->f_data;
1603         struct pipe *wpipe = rpipe->pipe_peer;
1604         int ret;
1605
1606         PIPE_LOCK(rpipe);
1607         kn->kn_data = rpipe->pipe_buffer.cnt;
1608         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1609                 kn->kn_data = rpipe->pipe_map.cnt;
1610
1611         if ((rpipe->pipe_state & PIPE_EOF) ||
1612             wpipe->pipe_present != PIPE_ACTIVE ||
1613             (wpipe->pipe_state & PIPE_EOF)) {
1614                 kn->kn_flags |= EV_EOF;
1615                 PIPE_UNLOCK(rpipe);
1616                 return (1);
1617         }
1618         ret = kn->kn_data > 0;
1619         PIPE_UNLOCK(rpipe);
1620         return ret;
1621 }
1622
1623 /*ARGSUSED*/
1624 static int
1625 filt_pipewrite(struct knote *kn, long hint)
1626 {
1627         struct pipe *rpipe = kn->kn_fp->f_data;
1628         struct pipe *wpipe = rpipe->pipe_peer;
1629
1630         PIPE_LOCK(rpipe);
1631         if (wpipe->pipe_present != PIPE_ACTIVE ||
1632             (wpipe->pipe_state & PIPE_EOF)) {
1633                 kn->kn_data = 0;
1634                 kn->kn_flags |= EV_EOF;
1635                 PIPE_UNLOCK(rpipe);
1636                 return (1);
1637         }
1638         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1639         if (wpipe->pipe_state & PIPE_DIRECTW)
1640                 kn->kn_data = 0;
1641
1642         PIPE_UNLOCK(rpipe);
1643         return (kn->kn_data >= PIPE_BUF);
1644 }