sys/kern/sys_pipe.c

   1 /*-
   2  * Copyright (c) 1996 John S. Dyson
   3  * Copyright (c) 2012 Giovanni Trematerra
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice immediately at the beginning of the file, without modification,
  11  *    this list of conditions, and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Absolutely no warranty of function or purpose is made by the author
  16  *    John S. Dyson.
  17  * 4. Modifications may be freely made to this file if the above conditions
  18  *    are met.
  19  */
  20
  21 /*
  22  * This file contains a high-performance replacement for the socket-based
  23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  24  * all features of sockets, but does do everything that pipes normally
  25  * do.
  26  */
  27
  28 /*
  29  * This code has two modes of operation, a small write mode and a large
  30  * write mode.  The small write mode acts like conventional pipes with
  31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  34  * memory, and the receiving process copies directly from these pinned pages
  35  * in the sending process.
  36  *
  37  * If the sending process receives a signal, it is possible that it will
  38  * go away, and certainly its address space can change, because control
  39  * is returned back to the user-mode side.  In that case, the pipe code
  40  * arranges to copy the buffer supplied by the user process, to a pageable
  41  * kernel buffer, and the receiving process will grab the data from the
  42  * pageable kernel buffer.  Since signals don't happen all that often,
  43  * the copy operation is normally eliminated.
  44  *
  45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  46  * happen for small transfers so that the system will not spend all of
  47  * its time context switching.
  48  *
  49  * In order to limit the resource use of pipes, two sysctls exist:
  50  *
  51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  52  * address space available to us in pipe_map. This value is normally
  53  * autotuned, but may also be loader tuned.
  54  *
  55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  56  * memory in use by pipes.
  57  *
  58  * Based on how large pipekva is relative to maxpipekva, the following
  59  * will happen:
  60  *
  61  * 0% - 50%:
  62  *     New pipes are given 16K of memory backing, pipes may dynamically
  63  *     grow to as large as 64K where needed.
  64  * 50% - 75%:
  65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  66  *     existing pipes may NOT grow.
  67  * 75% - 100%:
  68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  69  *     existing pipes will be shrunk down to 4K whenever possible.
  70  *
  71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  73  * resize which MUST occur for reverse-direction pipes when they are
  74  * first used.
  75  *
  76  * Additional information about the current state of pipes may be obtained
  77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  78  * and kern.ipc.piperesizefail.
  79  *
  80  * Locking rules:  There are two locks present here:  A mutex, used via
  81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  82  * the flag, as mutexes can not persist over uiomove.  The mutex
  83  * exists only to guard access to the flag, and is not in itself a
  84  * locking mechanism.  Also note that there is only a single mutex for
  85  * both directions of a pipe.
  86  *
  87  * As pipelock() may have to sleep before it can acquire the flag, it
  88  * is important to reread all data after a call to pipelock(); everything
  89  * in the structure may have changed.
  90  */
  91
  92 #include <sys/cdefs.h>
  93 __FBSDID("$FreeBSD$");
  94
  95 #include <sys/param.h>
  96 #include <sys/systm.h>
  97 #include <sys/conf.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/syscallsubr.h>
 112 #include <sys/sysctl.h>
 113 #include <sys/sysproto.h>
 114 #include <sys/pipe.h>
 115 #include <sys/proc.h>
 116 #include <sys/vnode.h>
 117 #include <sys/uio.h>
 118 #include <sys/event.h>
 119
 120 #include <security/mac/mac_framework.h>
 121
 122 #include <vm/vm.h>
 123 #include <vm/vm_param.h>
 124 #include <vm/vm_object.h>
 125 #include <vm/vm_kern.h>
 126 #include <vm/vm_extern.h>
 127 #include <vm/pmap.h>
 128 #include <vm/vm_map.h>
 129 #include <vm/vm_page.h>
 130 #include <vm/uma.h>
 131
 132 /*
 133  * Use this define if you want to disable *fancy* VM things.  Expect an
 134  * approx 30% decrease in transfer rate.  This could be useful for
 135  * NetBSD or OpenBSD.
 136  */
 137 /* #define PIPE_NODIRECT */
 138
 139 #define PIPE_PEER(pipe) \
 140         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 141
 142 /*
 143  * interfaces to the outside world
 144  */
 145 static fo_rdwr_t        pipe_read;
 146 static fo_rdwr_t        pipe_write;
 147 static fo_truncate_t    pipe_truncate;
 148 static fo_ioctl_t       pipe_ioctl;
 149 static fo_poll_t        pipe_poll;
 150 static fo_kqfilter_t    pipe_kqfilter;
 151 static fo_stat_t        pipe_stat;
 152 static fo_close_t       pipe_close;
 153 static fo_chmod_t       pipe_chmod;
 154 static fo_chown_t       pipe_chown;
 155
 156 struct fileops pipeops = {
 157         .fo_read = pipe_read,
 158         .fo_write = pipe_write,
 159         .fo_truncate = pipe_truncate,
 160         .fo_ioctl = pipe_ioctl,
 161         .fo_poll = pipe_poll,
 162         .fo_kqfilter = pipe_kqfilter,
 163         .fo_stat = pipe_stat,
 164         .fo_close = pipe_close,
 165         .fo_chmod = pipe_chmod,
 166         .fo_chown = pipe_chown,
 167         .fo_flags = DFLAG_PASSABLE
 168 };
 169
 170 static void     filt_pipedetach(struct knote *kn);
 171 static void     filt_pipedetach_notsup(struct knote *kn);
 172 static int      filt_pipenotsup(struct knote *kn, long hint);
 173 static int      filt_piperead(struct knote *kn, long hint);
 174 static int      filt_pipewrite(struct knote *kn, long hint);
 175
 176 static struct filterops pipe_nfiltops = {
 177         .f_isfd = 1,
 178         .f_detach = filt_pipedetach_notsup,
 179         .f_event = filt_pipenotsup
 180 };
 181 static struct filterops pipe_rfiltops = {
 182         .f_isfd = 1,
 183         .f_detach = filt_pipedetach,
 184         .f_event = filt_piperead
 185 };
 186 static struct filterops pipe_wfiltops = {
 187         .f_isfd = 1,
 188         .f_detach = filt_pipedetach,
 189         .f_event = filt_pipewrite
 190 };
 191
 192 /*
 193  * Default pipe buffer size(s), this can be kind-of large now because pipe
 194  * space is pageable.  The pipe code will try to maintain locality of
 195  * reference for performance reasons, so small amounts of outstanding I/O
 196  * will not wipe the cache.
 197  */
 198 #define MINPIPESIZE (PIPE_SIZE/3)
 199 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 200
 201 static long amountpipekva;
 202 static int pipefragretry;
 203 static int pipeallocfail;
 204 static int piperesizefail;
 205 static int piperesizeallowed = 1;
 206
 207 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 208            &maxpipekva, 0, "Pipe KVA limit");
 209 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 210            &amountpipekva, 0, "Pipe KVA usage");
 211 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 212           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 213 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 214           &pipeallocfail, 0, "Pipe allocation failures");
 215 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 216           &piperesizefail, 0, "Pipe resize failures");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 218           &piperesizeallowed, 0, "Pipe resizing allowed");
 219
 220 static void pipeinit(void *dummy __unused);
 221 static void pipeclose(struct pipe *cpipe);
 222 static void pipe_free_kmem(struct pipe *cpipe);
 223 static int pipe_create(struct pipe *pipe, int backing);
 224 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 225 static __inline int pipelock(struct pipe *cpipe, int catch);
 226 static __inline void pipeunlock(struct pipe *cpipe);
 227 #ifndef PIPE_NODIRECT
 228 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 229 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 230 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 231 static void pipe_clone_write_buffer(struct pipe *wpipe);
 232 #endif
 233 static int pipespace(struct pipe *cpipe, int size);
 234 static int pipespace_new(struct pipe *cpipe, int size);
 235
 236 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 237 static int      pipe_zone_init(void *mem, int size, int flags);
 238 static void     pipe_zone_fini(void *mem, int size);
 239
 240 static uma_zone_t pipe_zone;
 241 static struct unrhdr *pipeino_unr;
 242 static dev_t pipedev_ino;
 243
 244 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 245
 246 static void
 247 pipeinit(void *dummy __unused)
 248 {
 249
 250         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 251             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 252             UMA_ALIGN_PTR, 0);
 253         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 254         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 255         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 256         pipedev_ino = devfs_alloc_cdp_inode();
 257         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 258 }
 259
 260 static int
 261 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 262 {
 263         struct pipepair *pp;
 264         struct pipe *rpipe, *wpipe;
 265
 266         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 267
 268         pp = (struct pipepair *)mem;
 269
 270         /*
 271          * We zero both pipe endpoints to make sure all the kmem pointers
 272          * are NULL, flag fields are zero'd, etc.  We timestamp both
 273          * endpoints with the same time.
 274          */
 275         rpipe = &pp->pp_rpipe;
 276         bzero(rpipe, sizeof(*rpipe));
 277         vfs_timestamp(&rpipe->pipe_ctime);
 278         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 279
 280         wpipe = &pp->pp_wpipe;
 281         bzero(wpipe, sizeof(*wpipe));
 282         wpipe->pipe_ctime = rpipe->pipe_ctime;
 283         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 284
 285         rpipe->pipe_peer = wpipe;
 286         rpipe->pipe_pair = pp;
 287         wpipe->pipe_peer = rpipe;
 288         wpipe->pipe_pair = pp;
 289
 290         /*
 291          * Mark both endpoints as present; they will later get free'd
 292          * one at a time.  When both are free'd, then the whole pair
 293          * is released.
 294          */
 295         rpipe->pipe_present = PIPE_ACTIVE;
 296         wpipe->pipe_present = PIPE_ACTIVE;
 297
 298         /*
 299          * Eventually, the MAC Framework may initialize the label
 300          * in ctor or init, but for now we do it elswhere to avoid
 301          * blocking in ctor or init.
 302          */
 303         pp->pp_label = NULL;
 304
 305         return (0);
 306 }
 307
 308 static int
 309 pipe_zone_init(void *mem, int size, int flags)
 310 {
 311         struct pipepair *pp;
 312
 313         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 314
 315         pp = (struct pipepair *)mem;
 316
 317         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 318         return (0);
 319 }
 320
 321 static void
 322 pipe_zone_fini(void *mem, int size)
 323 {
 324         struct pipepair *pp;
 325
 326         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 327
 328         pp = (struct pipepair *)mem;
 329
 330         mtx_destroy(&pp->pp_mtx);
 331 }
 332
 333 static int
 334 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 335 {
 336         struct pipepair *pp;
 337         struct pipe *rpipe, *wpipe;
 338         int error;
 339
 340         *p_pp = 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_pipe_init() and mac_pipe_create() are called once
 345          * for the pair, and not on the endpoints.
 346          */
 347         mac_pipe_init(pp);
 348         mac_pipe_create(td->td_ucred, pp);
 349 #endif
 350         rpipe = &pp->pp_rpipe;
 351         wpipe = &pp->pp_wpipe;
 352
 353         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 354         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 355
 356         /* Only the forward direction pipe is backed by default */
 357         if ((error = pipe_create(rpipe, 1)) != 0 ||
 358             (error = pipe_create(wpipe, 0)) != 0) {
 359                 pipeclose(rpipe);
 360                 pipeclose(wpipe);
 361                 return (error);
 362         }
 363
 364         rpipe->pipe_state |= PIPE_DIRECTOK;
 365         wpipe->pipe_state |= PIPE_DIRECTOK;
 366         return (0);
 367 }
 368
 369 int
 370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 371 {
 372         struct pipepair *pp;
 373         int error;
 374
 375         error = pipe_paircreate(td, &pp);
 376         if (error != 0)
 377                 return (error);
 378         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 379         *ppipe = &pp->pp_rpipe;
 380         return (0);
 381 }
 382
 383 void
 384 pipe_dtor(struct pipe *dpipe)
 385 {
 386         ino_t ino;
 387
 388         ino = dpipe->pipe_ino;
 389         funsetown(&dpipe->pipe_sigio);
 390         pipeclose(dpipe);
 391         if (dpipe->pipe_state & PIPE_NAMED) {
 392                 dpipe = dpipe->pipe_peer;
 393                 funsetown(&dpipe->pipe_sigio);
 394                 pipeclose(dpipe);
 395         }
 396         if (ino != 0 && ino != (ino_t)-1)
 397                 free_unr(pipeino_unr, ino);
 398 }
 399
 400 /*
 401  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 402  * the zone pick up the pieces via pipeclose().
 403  */
 404 int
 405 kern_pipe(struct thread *td, int fildes[2])
 406 {
 407
 408         return (kern_pipe2(td, fildes, 0));
 409 }
 410
 411 int
 412 kern_pipe2(struct thread *td, int fildes[2], int flags)
 413 {
 414         struct filedesc *fdp;
 415         struct file *rf, *wf;
 416         struct pipe *rpipe, *wpipe;
 417         struct pipepair *pp;
 418         int fd, fflags, error;
 419
 420         fdp = td->td_proc->p_fd;
 421         error = pipe_paircreate(td, &pp);
 422         if (error != 0)
 423                 return (error);
 424         rpipe = &pp->pp_rpipe;
 425         wpipe = &pp->pp_wpipe;
 426         error = falloc(td, &rf, &fd, flags);
 427         if (error) {
 428                 pipeclose(rpipe);
 429                 pipeclose(wpipe);
 430                 return (error);
 431         }
 432         /* An extra reference on `rf' has been held for us by falloc(). */
 433         fildes[0] = fd;
 434
 435         fflags = FREAD | FWRITE;
 436         if ((flags & O_NONBLOCK) != 0)
 437                 fflags |= FNONBLOCK;
 438
 439         /*
 440          * Warning: once we've gotten past allocation of the fd for the
 441          * read-side, we can only drop the read side via fdrop() in order
 442          * to avoid races against processes which manage to dup() the read
 443          * side while we are blocked trying to allocate the write side.
 444          */
 445         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 446         error = falloc(td, &wf, &fd, flags);
 447         if (error) {
 448                 fdclose(fdp, rf, fildes[0], td);
 449                 fdrop(rf, td);
 450                 /* rpipe has been closed by fdrop(). */
 451                 pipeclose(wpipe);
 452                 return (error);
 453         }
 454         /* An extra reference on `wf' has been held for us by falloc(). */
 455         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 456         fdrop(wf, td);
 457         fildes[1] = fd;
 458         fdrop(rf, td);
 459
 460         return (0);
 461 }
 462
 463 /* ARGSUSED */
 464 int
 465 sys_pipe(struct thread *td, struct pipe_args *uap)
 466 {
 467         int error;
 468         int fildes[2];
 469
 470         error = kern_pipe(td, fildes);
 471         if (error)
 472                 return (error);
 473
 474         td->td_retval[0] = fildes[0];
 475         td->td_retval[1] = fildes[1];
 476
 477         return (0);
 478 }
 479
 480 int
 481 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 482 {
 483         int error, fildes[2];
 484
 485         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 486                 return (EINVAL);
 487         error = kern_pipe2(td, fildes, uap->flags);
 488         if (error)
 489                 return (error);
 490         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 491         if (error) {
 492                 (void)kern_close(td, fildes[0]);
 493                 (void)kern_close(td, fildes[1]);
 494         }
 495         return (error);
 496 }
 497
 498 /*
 499  * Allocate kva for pipe circular buffer, the space is pageable
 500  * This routine will 'realloc' the size of a pipe safely, if it fails
 501  * it will retain the old buffer.
 502  * If it fails it will return ENOMEM.
 503  */
 504 static int
 505 pipespace_new(cpipe, size)
 506         struct pipe *cpipe;
 507         int size;
 508 {
 509         caddr_t buffer;
 510         int error, cnt, firstseg;
 511         static int curfail = 0;
 512         static struct timeval lastfail;
 513
 514         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 515         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 516                 ("pipespace: resize of direct writes not allowed"));
 517 retry:
 518         cnt = cpipe->pipe_buffer.cnt;
 519         if (cnt > size)
 520                 size = cnt;
 521
 522         size = round_page(size);
 523         buffer = (caddr_t) vm_map_min(pipe_map);
 524
 525         error = vm_map_find(pipe_map, NULL, 0,
 526                 (vm_offset_t *) &buffer, size, 1,
 527                 VM_PROT_ALL, VM_PROT_ALL, 0);
 528         if (error != KERN_SUCCESS) {
 529                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 530                         (size > SMALL_PIPE_SIZE)) {
 531                         size = SMALL_PIPE_SIZE;
 532                         pipefragretry++;
 533                         goto retry;
 534                 }
 535                 if (cpipe->pipe_buffer.buffer == NULL) {
 536                         pipeallocfail++;
 537                         if (ppsratecheck(&lastfail, &curfail, 1))
 538                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 539                 } else {
 540                         piperesizefail++;
 541                 }
 542                 return (ENOMEM);
 543         }
 544
 545         /* copy data, then free old resources if we're resizing */
 546         if (cnt > 0) {
 547                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 548                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 549                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 550                                 buffer, firstseg);
 551                         if ((cnt - firstseg) > 0)
 552                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 553                                         cpipe->pipe_buffer.in);
 554                 } else {
 555                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 556                                 buffer, cnt);
 557                 }
 558         }
 559         pipe_free_kmem(cpipe);
 560         cpipe->pipe_buffer.buffer = buffer;
 561         cpipe->pipe_buffer.size = size;
 562         cpipe->pipe_buffer.in = cnt;
 563         cpipe->pipe_buffer.out = 0;
 564         cpipe->pipe_buffer.cnt = cnt;
 565         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 566         return (0);
 567 }
 568
 569 /*
 570  * Wrapper for pipespace_new() that performs locking assertions.
 571  */
 572 static int
 573 pipespace(cpipe, size)
 574         struct pipe *cpipe;
 575         int size;
 576 {
 577
 578         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 579                 ("Unlocked pipe passed to pipespace"));
 580         return (pipespace_new(cpipe, size));
 581 }
 582
 583 /*
 584  * lock a pipe for I/O, blocking other access
 585  */
 586 static __inline int
 587 pipelock(cpipe, catch)
 588         struct pipe *cpipe;
 589         int catch;
 590 {
 591         int error;
 592
 593         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 594         while (cpipe->pipe_state & PIPE_LOCKFL) {
 595                 cpipe->pipe_state |= PIPE_LWANT;
 596                 error = msleep(cpipe, PIPE_MTX(cpipe),
 597                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 598                     "pipelk", 0);
 599                 if (error != 0)
 600                         return (error);
 601         }
 602         cpipe->pipe_state |= PIPE_LOCKFL;
 603         return (0);
 604 }
 605
 606 /*
 607  * unlock a pipe I/O lock
 608  */
 609 static __inline void
 610 pipeunlock(cpipe)
 611         struct pipe *cpipe;
 612 {
 613
 614         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 615         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 616                 ("Unlocked pipe passed to pipeunlock"));
 617         cpipe->pipe_state &= ~PIPE_LOCKFL;
 618         if (cpipe->pipe_state & PIPE_LWANT) {
 619                 cpipe->pipe_state &= ~PIPE_LWANT;
 620                 wakeup(cpipe);
 621         }
 622 }
 623
 624 void
 625 pipeselwakeup(cpipe)
 626         struct pipe *cpipe;
 627 {
 628
 629         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 630         if (cpipe->pipe_state & PIPE_SEL) {
 631                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 632                 if (!SEL_WAITING(&cpipe->pipe_sel))
 633                         cpipe->pipe_state &= ~PIPE_SEL;
 634         }
 635         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 636                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 637         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 638 }
 639
 640 /*
 641  * Initialize and allocate VM and memory for pipe.  The structure
 642  * will start out zero'd from the ctor, so we just manage the kmem.
 643  */
 644 static int
 645 pipe_create(pipe, backing)
 646         struct pipe *pipe;
 647         int backing;
 648 {
 649         int error;
 650
 651         if (backing) {
 652                 if (amountpipekva > maxpipekva / 2)
 653                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 654                 else
 655                         error = pipespace_new(pipe, PIPE_SIZE);
 656         } else {
 657                 /* If we're not backing this pipe, no need to do anything. */
 658                 error = 0;
 659         }
 660         pipe->pipe_ino = -1;
 661         return (error);
 662 }
 663
 664 /* ARGSUSED */
 665 static int
 666 pipe_read(fp, uio, active_cred, flags, td)
 667         struct file *fp;
 668         struct uio *uio;
 669         struct ucred *active_cred;
 670         struct thread *td;
 671         int flags;
 672 {
 673         struct pipe *rpipe;
 674         int error;
 675         int nread = 0;
 676         int size;
 677
 678         rpipe = fp->f_data;
 679         PIPE_LOCK(rpipe);
 680         ++rpipe->pipe_busy;
 681         error = pipelock(rpipe, 1);
 682         if (error)
 683                 goto unlocked_error;
 684
 685 #ifdef MAC
 686         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 687         if (error)
 688                 goto locked_error;
 689 #endif
 690         if (amountpipekva > (3 * maxpipekva) / 4) {
 691                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 692                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 693                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 694                         (piperesizeallowed == 1)) {
 695                         PIPE_UNLOCK(rpipe);
 696                         pipespace(rpipe, SMALL_PIPE_SIZE);
 697                         PIPE_LOCK(rpipe);
 698                 }
 699         }
 700
 701         while (uio->uio_resid) {
 702                 /*
 703                  * normal pipe buffer receive
 704                  */
 705                 if (rpipe->pipe_buffer.cnt > 0) {
 706                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 707                         if (size > rpipe->pipe_buffer.cnt)
 708                                 size = rpipe->pipe_buffer.cnt;
 709                         if (size > uio->uio_resid)
 710                                 size = uio->uio_resid;
 711
 712                         PIPE_UNLOCK(rpipe);
 713                         error = uiomove(
 714                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 715                             size, uio);
 716                         PIPE_LOCK(rpipe);
 717                         if (error)
 718                                 break;
 719
 720                         rpipe->pipe_buffer.out += size;
 721                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 722                                 rpipe->pipe_buffer.out = 0;
 723
 724                         rpipe->pipe_buffer.cnt -= size;
 725
 726                         /*
 727                          * If there is no more to read in the pipe, reset
 728                          * its pointers to the beginning.  This improves
 729                          * cache hit stats.
 730                          */
 731                         if (rpipe->pipe_buffer.cnt == 0) {
 732                                 rpipe->pipe_buffer.in = 0;
 733                                 rpipe->pipe_buffer.out = 0;
 734                         }
 735                         nread += size;
 736 #ifndef PIPE_NODIRECT
 737                 /*
 738                  * Direct copy, bypassing a kernel buffer.
 739                  */
 740                 } else if ((size = rpipe->pipe_map.cnt) &&
 741                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 742                         if (size > uio->uio_resid)
 743                                 size = (u_int) uio->uio_resid;
 744
 745                         PIPE_UNLOCK(rpipe);
 746                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 747                             rpipe->pipe_map.pos, size, uio);
 748                         PIPE_LOCK(rpipe);
 749                         if (error)
 750                                 break;
 751                         nread += size;
 752                         rpipe->pipe_map.pos += size;
 753                         rpipe->pipe_map.cnt -= size;
 754                         if (rpipe->pipe_map.cnt == 0) {
 755                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 756                                 wakeup(rpipe);
 757                         }
 758 #endif
 759                 } else {
 760                         /*
 761                          * detect EOF condition
 762                          * read returns 0 on EOF, no need to set error
 763                          */
 764                         if (rpipe->pipe_state & PIPE_EOF)
 765                                 break;
 766
 767                         /*
 768                          * If the "write-side" has been blocked, wake it up now.
 769                          */
 770                         if (rpipe->pipe_state & PIPE_WANTW) {
 771                                 rpipe->pipe_state &= ~PIPE_WANTW;
 772                                 wakeup(rpipe);
 773                         }
 774
 775                         /*
 776                          * Break if some data was read.
 777                          */
 778                         if (nread > 0)
 779                                 break;
 780
 781                         /*
 782                          * Unlock the pipe buffer for our remaining processing.
 783                          * We will either break out with an error or we will
 784                          * sleep and relock to loop.
 785                          */
 786                         pipeunlock(rpipe);
 787
 788                         /*
 789                          * Handle non-blocking mode operation or
 790                          * wait for more data.
 791                          */
 792                         if (fp->f_flag & FNONBLOCK) {
 793                                 error = EAGAIN;
 794                         } else {
 795                                 rpipe->pipe_state |= PIPE_WANTR;
 796                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 797                                     PRIBIO | PCATCH,
 798                                     "piperd", 0)) == 0)
 799                                         error = pipelock(rpipe, 1);
 800                         }
 801                         if (error)
 802                                 goto unlocked_error;
 803                 }
 804         }
 805 #ifdef MAC
 806 locked_error:
 807 #endif
 808         pipeunlock(rpipe);
 809
 810         /* XXX: should probably do this before getting any locks. */
 811         if (error == 0)
 812                 vfs_timestamp(&rpipe->pipe_atime);
 813 unlocked_error:
 814         --rpipe->pipe_busy;
 815
 816         /*
 817          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 818          */
 819         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 820                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 821                 wakeup(rpipe);
 822         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 823                 /*
 824                  * Handle write blocking hysteresis.
 825                  */
 826                 if (rpipe->pipe_state & PIPE_WANTW) {
 827                         rpipe->pipe_state &= ~PIPE_WANTW;
 828                         wakeup(rpipe);
 829                 }
 830         }
 831
 832         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 833                 pipeselwakeup(rpipe);
 834
 835         PIPE_UNLOCK(rpipe);
 836         return (error);
 837 }
 838
 839 #ifndef PIPE_NODIRECT
 840 /*
 841  * Map the sending processes' buffer into kernel space and wire it.
 842  * This is similar to a physical write operation.
 843  */
 844 static int
 845 pipe_build_write_buffer(wpipe, uio)
 846         struct pipe *wpipe;
 847         struct uio *uio;
 848 {
 849         u_int size;
 850         int i;
 851
 852         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 853         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 854                 ("Clone attempt on non-direct write pipe!"));
 855
 856         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 857                 size = wpipe->pipe_buffer.size;
 858         else
 859                 size = uio->uio_iov->iov_len;
 860
 861         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 862             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 863             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 864                 return (EFAULT);
 865
 866 /*
 867  * set up the control block
 868  */
 869         wpipe->pipe_map.npages = i;
 870         wpipe->pipe_map.pos =
 871             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 872         wpipe->pipe_map.cnt = size;
 873
 874 /*
 875  * and update the uio data
 876  */
 877
 878         uio->uio_iov->iov_len -= size;
 879         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 880         if (uio->uio_iov->iov_len == 0)
 881                 uio->uio_iov++;
 882         uio->uio_resid -= size;
 883         uio->uio_offset += size;
 884         return (0);
 885 }
 886
 887 /*
 888  * unmap and unwire the process buffer
 889  */
 890 static void
 891 pipe_destroy_write_buffer(wpipe)
 892         struct pipe *wpipe;
 893 {
 894
 895         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 896         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 897         wpipe->pipe_map.npages = 0;
 898 }
 899
 900 /*
 901  * In the case of a signal, the writing process might go away.  This
 902  * code copies the data into the circular buffer so that the source
 903  * pages can be freed without loss of data.
 904  */
 905 static void
 906 pipe_clone_write_buffer(wpipe)
 907         struct pipe *wpipe;
 908 {
 909         struct uio uio;
 910         struct iovec iov;
 911         int size;
 912         int pos;
 913
 914         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 915         size = wpipe->pipe_map.cnt;
 916         pos = wpipe->pipe_map.pos;
 917
 918         wpipe->pipe_buffer.in = size;
 919         wpipe->pipe_buffer.out = 0;
 920         wpipe->pipe_buffer.cnt = size;
 921         wpipe->pipe_state &= ~PIPE_DIRECTW;
 922
 923         PIPE_UNLOCK(wpipe);
 924         iov.iov_base = wpipe->pipe_buffer.buffer;
 925         iov.iov_len = size;
 926         uio.uio_iov = &iov;
 927         uio.uio_iovcnt = 1;
 928         uio.uio_offset = 0;
 929         uio.uio_resid = size;
 930         uio.uio_segflg = UIO_SYSSPACE;
 931         uio.uio_rw = UIO_READ;
 932         uio.uio_td = curthread;
 933         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 934         PIPE_LOCK(wpipe);
 935         pipe_destroy_write_buffer(wpipe);
 936 }
 937
 938 /*
 939  * This implements the pipe buffer write mechanism.  Note that only
 940  * a direct write OR a normal pipe write can be pending at any given time.
 941  * If there are any characters in the pipe buffer, the direct write will
 942  * be deferred until the receiving process grabs all of the bytes from
 943  * the pipe buffer.  Then the direct mapping write is set-up.
 944  */
 945 static int
 946 pipe_direct_write(wpipe, uio)
 947         struct pipe *wpipe;
 948         struct uio *uio;
 949 {
 950         int error;
 951
 952 retry:
 953         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 954         error = pipelock(wpipe, 1);
 955         if (wpipe->pipe_state & PIPE_EOF)
 956                 error = EPIPE;
 957         if (error) {
 958                 pipeunlock(wpipe);
 959                 goto error1;
 960         }
 961         while (wpipe->pipe_state & PIPE_DIRECTW) {
 962                 if (wpipe->pipe_state & PIPE_WANTR) {
 963                         wpipe->pipe_state &= ~PIPE_WANTR;
 964                         wakeup(wpipe);
 965                 }
 966                 pipeselwakeup(wpipe);
 967                 wpipe->pipe_state |= PIPE_WANTW;
 968                 pipeunlock(wpipe);
 969                 error = msleep(wpipe, PIPE_MTX(wpipe),
 970                     PRIBIO | PCATCH, "pipdww", 0);
 971                 if (error)
 972                         goto error1;
 973                 else
 974                         goto retry;
 975         }
 976         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 977         if (wpipe->pipe_buffer.cnt > 0) {
 978                 if (wpipe->pipe_state & PIPE_WANTR) {
 979                         wpipe->pipe_state &= ~PIPE_WANTR;
 980                         wakeup(wpipe);
 981                 }
 982                 pipeselwakeup(wpipe);
 983                 wpipe->pipe_state |= PIPE_WANTW;
 984                 pipeunlock(wpipe);
 985                 error = msleep(wpipe, PIPE_MTX(wpipe),
 986                     PRIBIO | PCATCH, "pipdwc", 0);
 987                 if (error)
 988                         goto error1;
 989                 else
 990                         goto retry;
 991         }
 992
 993         wpipe->pipe_state |= PIPE_DIRECTW;
 994
 995         PIPE_UNLOCK(wpipe);
 996         error = pipe_build_write_buffer(wpipe, uio);
 997         PIPE_LOCK(wpipe);
 998         if (error) {
 999                 wpipe->pipe_state &= ~PIPE_DIRECTW;
1000                 pipeunlock(wpipe);
1001                 goto error1;
1002         }
1003
1004         error = 0;
1005         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1006                 if (wpipe->pipe_state & PIPE_EOF) {
1007                         pipe_destroy_write_buffer(wpipe);
1008                         pipeselwakeup(wpipe);
1009                         pipeunlock(wpipe);
1010                         error = EPIPE;
1011                         goto error1;
1012                 }
1013                 if (wpipe->pipe_state & PIPE_WANTR) {
1014                         wpipe->pipe_state &= ~PIPE_WANTR;
1015                         wakeup(wpipe);
1016                 }
1017                 pipeselwakeup(wpipe);
1018                 wpipe->pipe_state |= PIPE_WANTW;
1019                 pipeunlock(wpipe);
1020                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1021                     "pipdwt", 0);
1022                 pipelock(wpipe, 0);
1023         }
1024
1025         if (wpipe->pipe_state & PIPE_EOF)
1026                 error = EPIPE;
1027         if (wpipe->pipe_state & PIPE_DIRECTW) {
1028                 /*
1029                  * this bit of trickery substitutes a kernel buffer for
1030                  * the process that might be going away.
1031                  */
1032                 pipe_clone_write_buffer(wpipe);
1033         } else {
1034                 pipe_destroy_write_buffer(wpipe);
1035         }
1036         pipeunlock(wpipe);
1037         return (error);
1038
1039 error1:
1040         wakeup(wpipe);
1041         return (error);
1042 }
1043 #endif
1044
1045 static int
1046 pipe_write(fp, uio, active_cred, flags, td)
1047         struct file *fp;
1048         struct uio *uio;
1049         struct ucred *active_cred;
1050         struct thread *td;
1051         int flags;
1052 {
1053         int error = 0;
1054         int desiredsize;
1055         ssize_t orig_resid;
1056         struct pipe *wpipe, *rpipe;
1057
1058         rpipe = fp->f_data;
1059         wpipe = PIPE_PEER(rpipe);
1060         PIPE_LOCK(rpipe);
1061         error = pipelock(wpipe, 1);
1062         if (error) {
1063                 PIPE_UNLOCK(rpipe);
1064                 return (error);
1065         }
1066         /*
1067          * detect loss of pipe read side, issue SIGPIPE if lost.
1068          */
1069         if (wpipe->pipe_present != PIPE_ACTIVE ||
1070             (wpipe->pipe_state & PIPE_EOF)) {
1071                 pipeunlock(wpipe);
1072                 PIPE_UNLOCK(rpipe);
1073                 return (EPIPE);
1074         }
1075 #ifdef MAC
1076         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1077         if (error) {
1078                 pipeunlock(wpipe);
1079                 PIPE_UNLOCK(rpipe);
1080                 return (error);
1081         }
1082 #endif
1083         ++wpipe->pipe_busy;
1084
1085         /* Choose a larger size if it's advantageous */
1086         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1087         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1088                 if (piperesizeallowed != 1)
1089                         break;
1090                 if (amountpipekva > maxpipekva / 2)
1091                         break;
1092                 if (desiredsize == BIG_PIPE_SIZE)
1093                         break;
1094                 desiredsize = desiredsize * 2;
1095         }
1096
1097         /* Choose a smaller size if we're in a OOM situation */
1098         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1099                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1100                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1101                 (piperesizeallowed == 1))
1102                 desiredsize = SMALL_PIPE_SIZE;
1103
1104         /* Resize if the above determined that a new size was necessary */
1105         if ((desiredsize != wpipe->pipe_buffer.size) &&
1106                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1107                 PIPE_UNLOCK(wpipe);
1108                 pipespace(wpipe, desiredsize);
1109                 PIPE_LOCK(wpipe);
1110         }
1111         if (wpipe->pipe_buffer.size == 0) {
1112                 /*
1113                  * This can only happen for reverse direction use of pipes
1114                  * in a complete OOM situation.
1115                  */
1116                 error = ENOMEM;
1117                 --wpipe->pipe_busy;
1118                 pipeunlock(wpipe);
1119                 PIPE_UNLOCK(wpipe);
1120                 return (error);
1121         }
1122
1123         pipeunlock(wpipe);
1124
1125         orig_resid = uio->uio_resid;
1126
1127         while (uio->uio_resid) {
1128                 int space;
1129
1130                 pipelock(wpipe, 0);
1131                 if (wpipe->pipe_state & PIPE_EOF) {
1132                         pipeunlock(wpipe);
1133                         error = EPIPE;
1134                         break;
1135                 }
1136 #ifndef PIPE_NODIRECT
1137                 /*
1138                  * If the transfer is large, we can gain performance if
1139                  * we do process-to-process copies directly.
1140                  * If the write is non-blocking, we don't use the
1141                  * direct write mechanism.
1142                  *
1143                  * The direct write mechanism will detect the reader going
1144                  * away on us.
1145                  */
1146                 if (uio->uio_segflg == UIO_USERSPACE &&
1147                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1148                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1149                     (fp->f_flag & FNONBLOCK) == 0) {
1150                         pipeunlock(wpipe);
1151                         error = pipe_direct_write(wpipe, uio);
1152                         if (error)
1153                                 break;
1154                         continue;
1155                 }
1156 #endif
1157
1158                 /*
1159                  * Pipe buffered writes cannot be coincidental with
1160                  * direct writes.  We wait until the currently executing
1161                  * direct write is completed before we start filling the
1162                  * pipe buffer.  We break out if a signal occurs or the
1163                  * reader goes away.
1164                  */
1165                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1166                         if (wpipe->pipe_state & PIPE_WANTR) {
1167                                 wpipe->pipe_state &= ~PIPE_WANTR;
1168                                 wakeup(wpipe);
1169                         }
1170                         pipeselwakeup(wpipe);
1171                         wpipe->pipe_state |= PIPE_WANTW;
1172                         pipeunlock(wpipe);
1173                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1174                             "pipbww", 0);
1175                         if (error)
1176                                 break;
1177                         else
1178                                 continue;
1179                 }
1180
1181                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1182
1183                 /* Writes of size <= PIPE_BUF must be atomic. */
1184                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1185                         space = 0;
1186
1187                 if (space > 0) {
1188                         int size;       /* Transfer size */
1189                         int segsize;    /* first segment to transfer */
1190
1191                         /*
1192                          * Transfer size is minimum of uio transfer
1193                          * and free space in pipe buffer.
1194                          */
1195                         if (space > uio->uio_resid)
1196                                 size = uio->uio_resid;
1197                         else
1198                                 size = space;
1199                         /*
1200                          * First segment to transfer is minimum of
1201                          * transfer size and contiguous space in
1202                          * pipe buffer.  If first segment to transfer
1203                          * is less than the transfer size, we've got
1204                          * a wraparound in the buffer.
1205                          */
1206                         segsize = wpipe->pipe_buffer.size -
1207                                 wpipe->pipe_buffer.in;
1208                         if (segsize > size)
1209                                 segsize = size;
1210
1211                         /* Transfer first segment */
1212
1213                         PIPE_UNLOCK(rpipe);
1214                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1215                                         segsize, uio);
1216                         PIPE_LOCK(rpipe);
1217
1218                         if (error == 0 && segsize < size) {
1219                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1220                                         wpipe->pipe_buffer.size,
1221                                         ("Pipe buffer wraparound disappeared"));
1222                                 /*
1223                                  * Transfer remaining part now, to
1224                                  * support atomic writes.  Wraparound
1225                                  * happened.
1226                                  */
1227
1228                                 PIPE_UNLOCK(rpipe);
1229                                 error = uiomove(
1230                                     &wpipe->pipe_buffer.buffer[0],
1231                                     size - segsize, uio);
1232                                 PIPE_LOCK(rpipe);
1233                         }
1234                         if (error == 0) {
1235                                 wpipe->pipe_buffer.in += size;
1236                                 if (wpipe->pipe_buffer.in >=
1237                                     wpipe->pipe_buffer.size) {
1238                                         KASSERT(wpipe->pipe_buffer.in ==
1239                                                 size - segsize +
1240                                                 wpipe->pipe_buffer.size,
1241                                                 ("Expected wraparound bad"));
1242                                         wpipe->pipe_buffer.in = size - segsize;
1243                                 }
1244
1245                                 wpipe->pipe_buffer.cnt += size;
1246                                 KASSERT(wpipe->pipe_buffer.cnt <=
1247                                         wpipe->pipe_buffer.size,
1248                                         ("Pipe buffer overflow"));
1249                         }
1250                         pipeunlock(wpipe);
1251                         if (error != 0)
1252                                 break;
1253                 } else {
1254                         /*
1255                          * If the "read-side" has been blocked, wake it up now.
1256                          */
1257                         if (wpipe->pipe_state & PIPE_WANTR) {
1258                                 wpipe->pipe_state &= ~PIPE_WANTR;
1259                                 wakeup(wpipe);
1260                         }
1261
1262                         /*
1263                          * don't block on non-blocking I/O
1264                          */
1265                         if (fp->f_flag & FNONBLOCK) {
1266                                 error = EAGAIN;
1267                                 pipeunlock(wpipe);
1268                                 break;
1269                         }
1270
1271                         /*
1272                          * We have no more space and have something to offer,
1273                          * wake up select/poll.
1274                          */
1275                         pipeselwakeup(wpipe);
1276
1277                         wpipe->pipe_state |= PIPE_WANTW;
1278                         pipeunlock(wpipe);
1279                         error = msleep(wpipe, PIPE_MTX(rpipe),
1280                             PRIBIO | PCATCH, "pipewr", 0);
1281                         if (error != 0)
1282                                 break;
1283                 }
1284         }
1285
1286         pipelock(wpipe, 0);
1287         --wpipe->pipe_busy;
1288
1289         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1290                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1291                 wakeup(wpipe);
1292         } else if (wpipe->pipe_buffer.cnt > 0) {
1293                 /*
1294                  * If we have put any characters in the buffer, we wake up
1295                  * the reader.
1296                  */
1297                 if (wpipe->pipe_state & PIPE_WANTR) {
1298                         wpipe->pipe_state &= ~PIPE_WANTR;
1299                         wakeup(wpipe);
1300                 }
1301         }
1302
1303         /*
1304          * Don't return EPIPE if I/O was successful
1305          */
1306         if ((wpipe->pipe_buffer.cnt == 0) &&
1307             (uio->uio_resid == 0) &&
1308             (error == EPIPE)) {
1309                 error = 0;
1310         }
1311
1312         if (error == 0)
1313                 vfs_timestamp(&wpipe->pipe_mtime);
1314
1315         /*
1316          * We have something to offer,
1317          * wake up select/poll.
1318          */
1319         if (wpipe->pipe_buffer.cnt)
1320                 pipeselwakeup(wpipe);
1321
1322         pipeunlock(wpipe);
1323         PIPE_UNLOCK(rpipe);
1324         return (error);
1325 }
1326
1327 /* ARGSUSED */
1328 static int
1329 pipe_truncate(fp, length, active_cred, td)
1330         struct file *fp;
1331         off_t length;
1332         struct ucred *active_cred;
1333         struct thread *td;
1334 {
1335
1336         /* For named pipes call the vnode operation. */
1337         if (fp->f_vnode != NULL)
1338                 return (vnops.fo_truncate(fp, length, active_cred, td));
1339         return (EINVAL);
1340 }
1341
1342 /*
1343  * we implement a very minimal set of ioctls for compatibility with sockets.
1344  */
1345 static int
1346 pipe_ioctl(fp, cmd, data, active_cred, td)
1347         struct file *fp;
1348         u_long cmd;
1349         void *data;
1350         struct ucred *active_cred;
1351         struct thread *td;
1352 {
1353         struct pipe *mpipe = fp->f_data;
1354         int error;
1355
1356         PIPE_LOCK(mpipe);
1357
1358 #ifdef MAC
1359         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1360         if (error) {
1361                 PIPE_UNLOCK(mpipe);
1362                 return (error);
1363         }
1364 #endif
1365
1366         error = 0;
1367         switch (cmd) {
1368
1369         case FIONBIO:
1370                 break;
1371
1372         case FIOASYNC:
1373                 if (*(int *)data) {
1374                         mpipe->pipe_state |= PIPE_ASYNC;
1375                 } else {
1376                         mpipe->pipe_state &= ~PIPE_ASYNC;
1377                 }
1378                 break;
1379
1380         case FIONREAD:
1381                 if (!(fp->f_flag & FREAD)) {
1382                         *(int *)data = 0;
1383                         PIPE_UNLOCK(mpipe);
1384                         return (0);
1385                 }
1386                 if (mpipe->pipe_state & PIPE_DIRECTW)
1387                         *(int *)data = mpipe->pipe_map.cnt;
1388                 else
1389                         *(int *)data = mpipe->pipe_buffer.cnt;
1390                 break;
1391
1392         case FIOSETOWN:
1393                 PIPE_UNLOCK(mpipe);
1394                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1395                 goto out_unlocked;
1396
1397         case FIOGETOWN:
1398                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1399                 break;
1400
1401         /* This is deprecated, FIOSETOWN should be used instead. */
1402         case TIOCSPGRP:
1403                 PIPE_UNLOCK(mpipe);
1404                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1405                 goto out_unlocked;
1406
1407         /* This is deprecated, FIOGETOWN should be used instead. */
1408         case TIOCGPGRP:
1409                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1410                 break;
1411
1412         default:
1413                 error = ENOTTY;
1414                 break;
1415         }
1416         PIPE_UNLOCK(mpipe);
1417 out_unlocked:
1418         return (error);
1419 }
1420
1421 static int
1422 pipe_poll(fp, events, active_cred, td)
1423         struct file *fp;
1424         int events;
1425         struct ucred *active_cred;
1426         struct thread *td;
1427 {
1428         struct pipe *rpipe;
1429         struct pipe *wpipe;
1430         int levents, revents;
1431 #ifdef MAC
1432         int error;
1433 #endif
1434
1435         revents = 0;
1436         rpipe = fp->f_data;
1437         wpipe = PIPE_PEER(rpipe);
1438         PIPE_LOCK(rpipe);
1439 #ifdef MAC
1440         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1441         if (error)
1442                 goto locked_error;
1443 #endif
1444         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1445                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1446                     (rpipe->pipe_buffer.cnt > 0))
1447                         revents |= events & (POLLIN | POLLRDNORM);
1448
1449         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1450                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1451                     (wpipe->pipe_state & PIPE_EOF) ||
1452                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1453                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1454                          wpipe->pipe_buffer.size == 0)))
1455                         revents |= events & (POLLOUT | POLLWRNORM);
1456
1457         levents = events &
1458             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1459         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1460             fp->f_seqcount == rpipe->pipe_wgen)
1461                 events |= POLLINIGNEOF;
1462
1463         if ((events & POLLINIGNEOF) == 0) {
1464                 if (rpipe->pipe_state & PIPE_EOF) {
1465                         revents |= (events & (POLLIN | POLLRDNORM));
1466                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1467                             (wpipe->pipe_state & PIPE_EOF))
1468                                 revents |= POLLHUP;
1469                 }
1470         }
1471
1472         if (revents == 0) {
1473                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1474                         selrecord(td, &rpipe->pipe_sel);
1475                         if (SEL_WAITING(&rpipe->pipe_sel))
1476                                 rpipe->pipe_state |= PIPE_SEL;
1477                 }
1478
1479                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1480                         selrecord(td, &wpipe->pipe_sel);
1481                         if (SEL_WAITING(&wpipe->pipe_sel))
1482                                 wpipe->pipe_state |= PIPE_SEL;
1483                 }
1484         }
1485 #ifdef MAC
1486 locked_error:
1487 #endif
1488         PIPE_UNLOCK(rpipe);
1489
1490         return (revents);
1491 }
1492
1493 /*
1494  * We shouldn't need locks here as we're doing a read and this should
1495  * be a natural race.
1496  */
1497 static int
1498 pipe_stat(fp, ub, active_cred, td)
1499         struct file *fp;
1500         struct stat *ub;
1501         struct ucred *active_cred;
1502         struct thread *td;
1503 {
1504         struct pipe *pipe;
1505         int new_unr;
1506 #ifdef MAC
1507         int error;
1508 #endif
1509
1510         pipe = fp->f_data;
1511         PIPE_LOCK(pipe);
1512 #ifdef MAC
1513         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1514         if (error) {
1515                 PIPE_UNLOCK(pipe);
1516                 return (error);
1517         }
1518 #endif
1519
1520         /* For named pipes ask the underlying filesystem. */
1521         if (pipe->pipe_state & PIPE_NAMED) {
1522                 PIPE_UNLOCK(pipe);
1523                 return (vnops.fo_stat(fp, ub, active_cred, td));
1524         }
1525
1526         /*
1527          * Lazily allocate an inode number for the pipe.  Most pipe
1528          * users do not call fstat(2) on the pipe, which means that
1529          * postponing the inode allocation until it is must be
1530          * returned to userland is useful.  If alloc_unr failed,
1531          * assign st_ino zero instead of returning an error.
1532          * Special pipe_ino values:
1533          *  -1 - not yet initialized;
1534          *  0  - alloc_unr failed, return 0 as st_ino forever.
1535          */
1536         if (pipe->pipe_ino == (ino_t)-1) {
1537                 new_unr = alloc_unr(pipeino_unr);
1538                 if (new_unr != -1)
1539                         pipe->pipe_ino = new_unr;
1540                 else
1541                         pipe->pipe_ino = 0;
1542         }
1543         PIPE_UNLOCK(pipe);
1544
1545         bzero(ub, sizeof(*ub));
1546         ub->st_mode = S_IFIFO;
1547         ub->st_blksize = PAGE_SIZE;
1548         if (pipe->pipe_state & PIPE_DIRECTW)
1549                 ub->st_size = pipe->pipe_map.cnt;
1550         else
1551                 ub->st_size = pipe->pipe_buffer.cnt;
1552         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1553         ub->st_atim = pipe->pipe_atime;
1554         ub->st_mtim = pipe->pipe_mtime;
1555         ub->st_ctim = pipe->pipe_ctime;
1556         ub->st_uid = fp->f_cred->cr_uid;
1557         ub->st_gid = fp->f_cred->cr_gid;
1558         ub->st_dev = pipedev_ino;
1559         ub->st_ino = pipe->pipe_ino;
1560         /*
1561          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1562          */
1563         return (0);
1564 }
1565
1566 /* ARGSUSED */
1567 static int
1568 pipe_close(fp, td)
1569         struct file *fp;
1570         struct thread *td;
1571 {
1572
1573         if (fp->f_vnode != NULL)
1574                 return vnops.fo_close(fp, td);
1575         fp->f_ops = &badfileops;
1576         pipe_dtor(fp->f_data);
1577         fp->f_data = NULL;
1578         return (0);
1579 }
1580
1581 static int
1582 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1583 {
1584         struct pipe *cpipe;
1585         int error;
1586
1587         cpipe = fp->f_data;
1588         if (cpipe->pipe_state & PIPE_NAMED)
1589                 error = vn_chmod(fp, mode, active_cred, td);
1590         else
1591                 error = invfo_chmod(fp, mode, active_cred, td);
1592         return (error);
1593 }
1594
1595 static int
1596 pipe_chown(fp, uid, gid, active_cred, td)
1597         struct file *fp;
1598         uid_t uid;
1599         gid_t gid;
1600         struct ucred *active_cred;
1601         struct thread *td;
1602 {
1603         struct pipe *cpipe;
1604         int error;
1605
1606         cpipe = fp->f_data;
1607         if (cpipe->pipe_state & PIPE_NAMED)
1608                 error = vn_chown(fp, uid, gid, active_cred, td);
1609         else
1610                 error = invfo_chown(fp, uid, gid, active_cred, td);
1611         return (error);
1612 }
1613
1614 static void
1615 pipe_free_kmem(cpipe)
1616         struct pipe *cpipe;
1617 {
1618
1619         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1620             ("pipe_free_kmem: pipe mutex locked"));
1621
1622         if (cpipe->pipe_buffer.buffer != NULL) {
1623                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1624                 vm_map_remove(pipe_map,
1625                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1626                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1627                 cpipe->pipe_buffer.buffer = NULL;
1628         }
1629 #ifndef PIPE_NODIRECT
1630         {
1631                 cpipe->pipe_map.cnt = 0;
1632                 cpipe->pipe_map.pos = 0;
1633                 cpipe->pipe_map.npages = 0;
1634         }
1635 #endif
1636 }
1637
1638 /*
1639  * shutdown the pipe
1640  */
1641 static void
1642 pipeclose(cpipe)
1643         struct pipe *cpipe;
1644 {
1645         struct pipepair *pp;
1646         struct pipe *ppipe;
1647
1648         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1649
1650         PIPE_LOCK(cpipe);
1651         pipelock(cpipe, 0);
1652         pp = cpipe->pipe_pair;
1653
1654         pipeselwakeup(cpipe);
1655
1656         /*
1657          * If the other side is blocked, wake it up saying that
1658          * we want to close it down.
1659          */
1660         cpipe->pipe_state |= PIPE_EOF;
1661         while (cpipe->pipe_busy) {
1662                 wakeup(cpipe);
1663                 cpipe->pipe_state |= PIPE_WANT;
1664                 pipeunlock(cpipe);
1665                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1666                 pipelock(cpipe, 0);
1667         }
1668
1669
1670         /*
1671          * Disconnect from peer, if any.
1672          */
1673         ppipe = cpipe->pipe_peer;
1674         if (ppipe->pipe_present == PIPE_ACTIVE) {
1675                 pipeselwakeup(ppipe);
1676
1677                 ppipe->pipe_state |= PIPE_EOF;
1678                 wakeup(ppipe);
1679                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1680         }
1681
1682         /*
1683          * Mark this endpoint as free.  Release kmem resources.  We
1684          * don't mark this endpoint as unused until we've finished
1685          * doing that, or the pipe might disappear out from under
1686          * us.
1687          */
1688         PIPE_UNLOCK(cpipe);
1689         pipe_free_kmem(cpipe);
1690         PIPE_LOCK(cpipe);
1691         cpipe->pipe_present = PIPE_CLOSING;
1692         pipeunlock(cpipe);
1693
1694         /*
1695          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1696          * PIPE_FINALIZED, that allows other end to free the
1697          * pipe_pair, only after the knotes are completely dismantled.
1698          */
1699         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1700         cpipe->pipe_present = PIPE_FINALIZED;
1701         seldrain(&cpipe->pipe_sel);
1702         knlist_destroy(&cpipe->pipe_sel.si_note);
1703
1704         /*
1705          * If both endpoints are now closed, release the memory for the
1706          * pipe pair.  If not, unlock.
1707          */
1708         if (ppipe->pipe_present == PIPE_FINALIZED) {
1709                 PIPE_UNLOCK(cpipe);
1710 #ifdef MAC
1711                 mac_pipe_destroy(pp);
1712 #endif
1713                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1714         } else
1715                 PIPE_UNLOCK(cpipe);
1716 }
1717
1718 /*ARGSUSED*/
1719 static int
1720 pipe_kqfilter(struct file *fp, struct knote *kn)
1721 {
1722         struct pipe *cpipe;
1723
1724         /*
1725          * If a filter is requested that is not supported by this file
1726          * descriptor, don't return an error, but also don't ever generate an
1727          * event.
1728          */
1729         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1730                 kn->kn_fop = &pipe_nfiltops;
1731                 return (0);
1732         }
1733         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1734                 kn->kn_fop = &pipe_nfiltops;
1735                 return (0);
1736         }
1737         cpipe = fp->f_data;
1738         PIPE_LOCK(cpipe);
1739         switch (kn->kn_filter) {
1740         case EVFILT_READ:
1741                 kn->kn_fop = &pipe_rfiltops;
1742                 break;
1743         case EVFILT_WRITE:
1744                 kn->kn_fop = &pipe_wfiltops;
1745                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1746                         /* other end of pipe has been closed */
1747                         PIPE_UNLOCK(cpipe);
1748                         return (EPIPE);
1749                 }
1750                 cpipe = PIPE_PEER(cpipe);
1751                 break;
1752         default:
1753                 PIPE_UNLOCK(cpipe);
1754                 return (EINVAL);
1755         }
1756
1757         kn->kn_hook = cpipe;
1758         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1759         PIPE_UNLOCK(cpipe);
1760         return (0);
1761 }
1762
1763 static void
1764 filt_pipedetach(struct knote *kn)
1765 {
1766         struct pipe *cpipe = kn->kn_hook;
1767
1768         PIPE_LOCK(cpipe);
1769         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1770         PIPE_UNLOCK(cpipe);
1771 }
1772
1773 /*ARGSUSED*/
1774 static int
1775 filt_piperead(struct knote *kn, long hint)
1776 {
1777         struct pipe *rpipe = kn->kn_hook;
1778         struct pipe *wpipe = rpipe->pipe_peer;
1779         int ret;
1780
1781         PIPE_LOCK(rpipe);
1782         kn->kn_data = rpipe->pipe_buffer.cnt;
1783         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1784                 kn->kn_data = rpipe->pipe_map.cnt;
1785
1786         if ((rpipe->pipe_state & PIPE_EOF) ||
1787             wpipe->pipe_present != PIPE_ACTIVE ||
1788             (wpipe->pipe_state & PIPE_EOF)) {
1789                 kn->kn_flags |= EV_EOF;
1790                 PIPE_UNLOCK(rpipe);
1791                 return (1);
1792         }
1793         ret = kn->kn_data > 0;
1794         PIPE_UNLOCK(rpipe);
1795         return ret;
1796 }
1797
1798 /*ARGSUSED*/
1799 static int
1800 filt_pipewrite(struct knote *kn, long hint)
1801 {
1802         struct pipe *wpipe;
1803
1804         wpipe = kn->kn_hook;
1805         PIPE_LOCK(wpipe);
1806         if (wpipe->pipe_present != PIPE_ACTIVE ||
1807             (wpipe->pipe_state & PIPE_EOF)) {
1808                 kn->kn_data = 0;
1809                 kn->kn_flags |= EV_EOF;
1810                 PIPE_UNLOCK(wpipe);
1811                 return (1);
1812         }
1813         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1814             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1815         if (wpipe->pipe_state & PIPE_DIRECTW)
1816                 kn->kn_data = 0;
1817
1818         PIPE_UNLOCK(wpipe);
1819         return (kn->kn_data >= PIPE_BUF);
1820 }
1821
1822 static void
1823 filt_pipedetach_notsup(struct knote *kn)
1824 {
1825
1826 }
1827
1828 static int
1829 filt_pipenotsup(struct knote *kn, long hint)
1830 {
1831
1832         return (0);
1833 }