sys/kern/sys_pipe.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-4-Clause
   3  *
   4  * Copyright (c) 1996 John S. Dyson
   5  * Copyright (c) 2012 Giovanni Trematerra
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice immediately at the beginning of the file, without modification,
  13  *    this list of conditions, and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Absolutely no warranty of function or purpose is made by the author
  18  *    John S. Dyson.
  19  * 4. Modifications may be freely made to this file if the above conditions
  20  *    are met.
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  36  * memory, and the receiving process copies directly from these pinned pages
  37  * in the sending process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.
  50  *
  51  * In order to limit the resource use of pipes, two sysctls exist:
  52  *
  53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  54  * address space available to us in pipe_map. This value is normally
  55  * autotuned, but may also be loader tuned.
  56  *
  57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  58  * memory in use by pipes.
  59  *
  60  * Based on how large pipekva is relative to maxpipekva, the following
  61  * will happen:
  62  *
  63  * 0% - 50%:
  64  *     New pipes are given 16K of memory backing, pipes may dynamically
  65  *     grow to as large as 64K where needed.
  66  * 50% - 75%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes may NOT grow.
  69  * 75% - 100%:
  70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  71  *     existing pipes will be shrunk down to 4K whenever possible.
  72  *
  73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  75  * resize which MUST occur for reverse-direction pipes when they are
  76  * first used.
  77  *
  78  * Additional information about the current state of pipes may be obtained
  79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  80  * and kern.ipc.piperesizefail.
  81  *
  82  * Locking rules:  There are two locks present here:  A mutex, used via
  83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  84  * the flag, as mutexes can not persist over uiomove.  The mutex
  85  * exists only to guard access to the flag, and is not in itself a
  86  * locking mechanism.  Also note that there is only a single mutex for
  87  * both directions of a pipe.
  88  *
  89  * As pipelock() may have to sleep before it can acquire the flag, it
  90  * is important to reread all data after a call to pipelock(); everything
  91  * in the structure may have changed.
  92  */
  93
  94 #include <sys/cdefs.h>
  95 __FBSDID("$FreeBSD$");
  96
  97 #include <sys/param.h>
  98 #include <sys/systm.h>
  99 #include <sys/conf.h>
 100 #include <sys/fcntl.h>
 101 #include <sys/file.h>
 102 #include <sys/filedesc.h>
 103 #include <sys/filio.h>
 104 #include <sys/kernel.h>
 105 #include <sys/lock.h>
 106 #include <sys/mutex.h>
 107 #include <sys/ttycom.h>
 108 #include <sys/stat.h>
 109 #include <sys/malloc.h>
 110 #include <sys/poll.h>
 111 #include <sys/selinfo.h>
 112 #include <sys/signalvar.h>
 113 #include <sys/syscallsubr.h>
 114 #include <sys/sysctl.h>
 115 #include <sys/sysproto.h>
 116 #include <sys/pipe.h>
 117 #include <sys/proc.h>
 118 #include <sys/vnode.h>
 119 #include <sys/uio.h>
 120 #include <sys/user.h>
 121 #include <sys/event.h>
 122
 123 #include <security/mac/mac_framework.h>
 124
 125 #include <vm/vm.h>
 126 #include <vm/vm_param.h>
 127 #include <vm/vm_object.h>
 128 #include <vm/vm_kern.h>
 129 #include <vm/vm_extern.h>
 130 #include <vm/pmap.h>
 131 #include <vm/vm_map.h>
 132 #include <vm/vm_page.h>
 133 #include <vm/uma.h>
 134
 135 /*
 136  * Use this define if you want to disable *fancy* VM things.  Expect an
 137  * approx 30% decrease in transfer rate.  This could be useful for
 138  * NetBSD or OpenBSD.
 139  */
 140 /* #define PIPE_NODIRECT */
 141
 142 #define PIPE_PEER(pipe) \
 143         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 144
 145 /*
 146  * interfaces to the outside world
 147  */
 148 static fo_rdwr_t        pipe_read;
 149 static fo_rdwr_t        pipe_write;
 150 static fo_truncate_t    pipe_truncate;
 151 static fo_ioctl_t       pipe_ioctl;
 152 static fo_poll_t        pipe_poll;
 153 static fo_kqfilter_t    pipe_kqfilter;
 154 static fo_stat_t        pipe_stat;
 155 static fo_close_t       pipe_close;
 156 static fo_chmod_t       pipe_chmod;
 157 static fo_chown_t       pipe_chown;
 158 static fo_fill_kinfo_t  pipe_fill_kinfo;
 159
 160 struct fileops pipeops = {
 161         .fo_read = pipe_read,
 162         .fo_write = pipe_write,
 163         .fo_truncate = pipe_truncate,
 164         .fo_ioctl = pipe_ioctl,
 165         .fo_poll = pipe_poll,
 166         .fo_kqfilter = pipe_kqfilter,
 167         .fo_stat = pipe_stat,
 168         .fo_close = pipe_close,
 169         .fo_chmod = pipe_chmod,
 170         .fo_chown = pipe_chown,
 171         .fo_sendfile = invfo_sendfile,
 172         .fo_fill_kinfo = pipe_fill_kinfo,
 173         .fo_flags = DFLAG_PASSABLE
 174 };
 175
 176 static void     filt_pipedetach(struct knote *kn);
 177 static void     filt_pipedetach_notsup(struct knote *kn);
 178 static int      filt_pipenotsup(struct knote *kn, long hint);
 179 static int      filt_piperead(struct knote *kn, long hint);
 180 static int      filt_pipewrite(struct knote *kn, long hint);
 181
 182 static struct filterops pipe_nfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach_notsup,
 185         .f_event = filt_pipenotsup
 186 };
 187 static struct filterops pipe_rfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_piperead
 191 };
 192 static struct filterops pipe_wfiltops = {
 193         .f_isfd = 1,
 194         .f_detach = filt_pipedetach,
 195         .f_event = filt_pipewrite
 196 };
 197
 198 /*
 199  * Default pipe buffer size(s), this can be kind-of large now because pipe
 200  * space is pageable.  The pipe code will try to maintain locality of
 201  * reference for performance reasons, so small amounts of outstanding I/O
 202  * will not wipe the cache.
 203  */
 204 #define MINPIPESIZE (PIPE_SIZE/3)
 205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 206
 207 static long amountpipekva;
 208 static int pipefragretry;
 209 static int pipeallocfail;
 210 static int piperesizefail;
 211 static int piperesizeallowed = 1;
 212
 213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 214            &maxpipekva, 0, "Pipe KVA limit");
 215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 216            &amountpipekva, 0, "Pipe KVA usage");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 218           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 220           &pipeallocfail, 0, "Pipe allocation failures");
 221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 222           &piperesizefail, 0, "Pipe resize failures");
 223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 224           &piperesizeallowed, 0, "Pipe resizing allowed");
 225
 226 static void pipeinit(void *dummy __unused);
 227 static void pipeclose(struct pipe *cpipe);
 228 static void pipe_free_kmem(struct pipe *cpipe);
 229 static void pipe_create(struct pipe *pipe, int backing);
 230 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 231 static __inline int pipelock(struct pipe *cpipe, int catch);
 232 static __inline void pipeunlock(struct pipe *cpipe);
 233 #ifndef PIPE_NODIRECT
 234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 237 static void pipe_clone_write_buffer(struct pipe *wpipe);
 238 #endif
 239 static int pipespace(struct pipe *cpipe, int size);
 240 static int pipespace_new(struct pipe *cpipe, int size);
 241
 242 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 243 static int      pipe_zone_init(void *mem, int size, int flags);
 244 static void     pipe_zone_fini(void *mem, int size);
 245
 246 static uma_zone_t pipe_zone;
 247 static struct unrhdr *pipeino_unr;
 248 static dev_t pipedev_ino;
 249
 250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 251
 252 static void
 253 pipeinit(void *dummy __unused)
 254 {
 255
 256         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 257             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 258             UMA_ALIGN_PTR, 0);
 259         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 260         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 261         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 262         pipedev_ino = devfs_alloc_cdp_inode();
 263         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 264 }
 265
 266 static int
 267 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 268 {
 269         struct pipepair *pp;
 270         struct pipe *rpipe, *wpipe;
 271
 272         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 273
 274         pp = (struct pipepair *)mem;
 275
 276         /*
 277          * We zero both pipe endpoints to make sure all the kmem pointers
 278          * are NULL, flag fields are zero'd, etc.  We timestamp both
 279          * endpoints with the same time.
 280          */
 281         rpipe = &pp->pp_rpipe;
 282         bzero(rpipe, sizeof(*rpipe));
 283         vfs_timestamp(&rpipe->pipe_ctime);
 284         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 285
 286         wpipe = &pp->pp_wpipe;
 287         bzero(wpipe, sizeof(*wpipe));
 288         wpipe->pipe_ctime = rpipe->pipe_ctime;
 289         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 290
 291         rpipe->pipe_peer = wpipe;
 292         rpipe->pipe_pair = pp;
 293         wpipe->pipe_peer = rpipe;
 294         wpipe->pipe_pair = pp;
 295
 296         /*
 297          * Mark both endpoints as present; they will later get free'd
 298          * one at a time.  When both are free'd, then the whole pair
 299          * is released.
 300          */
 301         rpipe->pipe_present = PIPE_ACTIVE;
 302         wpipe->pipe_present = PIPE_ACTIVE;
 303
 304         /*
 305          * Eventually, the MAC Framework may initialize the label
 306          * in ctor or init, but for now we do it elswhere to avoid
 307          * blocking in ctor or init.
 308          */
 309         pp->pp_label = NULL;
 310
 311         return (0);
 312 }
 313
 314 static int
 315 pipe_zone_init(void *mem, int size, int flags)
 316 {
 317         struct pipepair *pp;
 318
 319         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 320
 321         pp = (struct pipepair *)mem;
 322
 323         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 324         return (0);
 325 }
 326
 327 static void
 328 pipe_zone_fini(void *mem, int size)
 329 {
 330         struct pipepair *pp;
 331
 332         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 333
 334         pp = (struct pipepair *)mem;
 335
 336         mtx_destroy(&pp->pp_mtx);
 337 }
 338
 339 static void
 340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 341 {
 342         struct pipepair *pp;
 343         struct pipe *rpipe, *wpipe;
 344
 345         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 346 #ifdef MAC
 347         /*
 348          * The MAC label is shared between the connected endpoints.  As a
 349          * result mac_pipe_init() and mac_pipe_create() are called once
 350          * for the pair, and not on the endpoints.
 351          */
 352         mac_pipe_init(pp);
 353         mac_pipe_create(td->td_ucred, pp);
 354 #endif
 355         rpipe = &pp->pp_rpipe;
 356         wpipe = &pp->pp_wpipe;
 357
 358         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 359         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 360
 361         /* Only the forward direction pipe is backed by default */
 362         pipe_create(rpipe, 1);
 363         pipe_create(wpipe, 0);
 364
 365         rpipe->pipe_state |= PIPE_DIRECTOK;
 366         wpipe->pipe_state |= PIPE_DIRECTOK;
 367 }
 368
 369 void
 370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 371 {
 372         struct pipepair *pp;
 373
 374         pipe_paircreate(td, &pp);
 375         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 376         *ppipe = &pp->pp_rpipe;
 377 }
 378
 379 void
 380 pipe_dtor(struct pipe *dpipe)
 381 {
 382         struct pipe *peer;
 383         ino_t ino;
 384
 385         ino = dpipe->pipe_ino;
 386         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 387         funsetown(&dpipe->pipe_sigio);
 388         pipeclose(dpipe);
 389         if (peer != NULL) {
 390                 funsetown(&peer->pipe_sigio);
 391                 pipeclose(peer);
 392         }
 393         if (ino != 0 && ino != (ino_t)-1)
 394                 free_unr(pipeino_unr, ino);
 395 }
 396
 397 /*
 398  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 399  * the zone pick up the pieces via pipeclose().
 400  */
 401 int
 402 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 403     struct filecaps *fcaps2)
 404 {
 405         struct file *rf, *wf;
 406         struct pipe *rpipe, *wpipe;
 407         struct pipepair *pp;
 408         int fd, fflags, error;
 409
 410         pipe_paircreate(td, &pp);
 411         rpipe = &pp->pp_rpipe;
 412         wpipe = &pp->pp_wpipe;
 413         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 414         if (error) {
 415                 pipeclose(rpipe);
 416                 pipeclose(wpipe);
 417                 return (error);
 418         }
 419         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 420         fildes[0] = fd;
 421
 422         fflags = FREAD | FWRITE;
 423         if ((flags & O_NONBLOCK) != 0)
 424                 fflags |= FNONBLOCK;
 425
 426         /*
 427          * Warning: once we've gotten past allocation of the fd for the
 428          * read-side, we can only drop the read side via fdrop() in order
 429          * to avoid races against processes which manage to dup() the read
 430          * side while we are blocked trying to allocate the write side.
 431          */
 432         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 433         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 434         if (error) {
 435                 fdclose(td, rf, fildes[0]);
 436                 fdrop(rf, td);
 437                 /* rpipe has been closed by fdrop(). */
 438                 pipeclose(wpipe);
 439                 return (error);
 440         }
 441         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 442         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 443         fdrop(wf, td);
 444         fildes[1] = fd;
 445         fdrop(rf, td);
 446
 447         return (0);
 448 }
 449
 450 #ifdef COMPAT_FREEBSD10
 451 /* ARGSUSED */
 452 int
 453 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 454 {
 455         int error;
 456         int fildes[2];
 457
 458         error = kern_pipe(td, fildes, 0, NULL, NULL);
 459         if (error)
 460                 return (error);
 461
 462         td->td_retval[0] = fildes[0];
 463         td->td_retval[1] = fildes[1];
 464
 465         return (0);
 466 }
 467 #endif
 468
 469 int
 470 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 471 {
 472         int error, fildes[2];
 473
 474         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 475                 return (EINVAL);
 476         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 477         if (error)
 478                 return (error);
 479         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 480         if (error) {
 481                 (void)kern_close(td, fildes[0]);
 482                 (void)kern_close(td, fildes[1]);
 483         }
 484         return (error);
 485 }
 486
 487 /*
 488  * Allocate kva for pipe circular buffer, the space is pageable
 489  * This routine will 'realloc' the size of a pipe safely, if it fails
 490  * it will retain the old buffer.
 491  * If it fails it will return ENOMEM.
 492  */
 493 static int
 494 pipespace_new(struct pipe *cpipe, int size)
 495 {
 496         caddr_t buffer;
 497         int error, cnt, firstseg;
 498         static int curfail = 0;
 499         static struct timeval lastfail;
 500
 501         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 502         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 503                 ("pipespace: resize of direct writes not allowed"));
 504 retry:
 505         cnt = cpipe->pipe_buffer.cnt;
 506         if (cnt > size)
 507                 size = cnt;
 508
 509         size = round_page(size);
 510         buffer = (caddr_t) vm_map_min(pipe_map);
 511
 512         error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
 513             VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
 514         if (error != KERN_SUCCESS) {
 515                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 516                         (size > SMALL_PIPE_SIZE)) {
 517                         size = SMALL_PIPE_SIZE;
 518                         pipefragretry++;
 519                         goto retry;
 520                 }
 521                 if (cpipe->pipe_buffer.buffer == NULL) {
 522                         pipeallocfail++;
 523                         if (ppsratecheck(&lastfail, &curfail, 1))
 524                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 525                 } else {
 526                         piperesizefail++;
 527                 }
 528                 return (ENOMEM);
 529         }
 530
 531         /* copy data, then free old resources if we're resizing */
 532         if (cnt > 0) {
 533                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 534                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 535                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 536                                 buffer, firstseg);
 537                         if ((cnt - firstseg) > 0)
 538                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 539                                         cpipe->pipe_buffer.in);
 540                 } else {
 541                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 542                                 buffer, cnt);
 543                 }
 544         }
 545         pipe_free_kmem(cpipe);
 546         cpipe->pipe_buffer.buffer = buffer;
 547         cpipe->pipe_buffer.size = size;
 548         cpipe->pipe_buffer.in = cnt;
 549         cpipe->pipe_buffer.out = 0;
 550         cpipe->pipe_buffer.cnt = cnt;
 551         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 552         return (0);
 553 }
 554
 555 /*
 556  * Wrapper for pipespace_new() that performs locking assertions.
 557  */
 558 static int
 559 pipespace(struct pipe *cpipe, int size)
 560 {
 561
 562         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 563                 ("Unlocked pipe passed to pipespace"));
 564         return (pipespace_new(cpipe, size));
 565 }
 566
 567 /*
 568  * lock a pipe for I/O, blocking other access
 569  */
 570 static __inline int
 571 pipelock(struct pipe *cpipe, int catch)
 572 {
 573         int error;
 574
 575         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 576         while (cpipe->pipe_state & PIPE_LOCKFL) {
 577                 cpipe->pipe_state |= PIPE_LWANT;
 578                 error = msleep(cpipe, PIPE_MTX(cpipe),
 579                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 580                     "pipelk", 0);
 581                 if (error != 0)
 582                         return (error);
 583         }
 584         cpipe->pipe_state |= PIPE_LOCKFL;
 585         return (0);
 586 }
 587
 588 /*
 589  * unlock a pipe I/O lock
 590  */
 591 static __inline void
 592 pipeunlock(struct pipe *cpipe)
 593 {
 594
 595         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 596         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 597                 ("Unlocked pipe passed to pipeunlock"));
 598         cpipe->pipe_state &= ~PIPE_LOCKFL;
 599         if (cpipe->pipe_state & PIPE_LWANT) {
 600                 cpipe->pipe_state &= ~PIPE_LWANT;
 601                 wakeup(cpipe);
 602         }
 603 }
 604
 605 void
 606 pipeselwakeup(struct pipe *cpipe)
 607 {
 608
 609         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 610         if (cpipe->pipe_state & PIPE_SEL) {
 611                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 612                 if (!SEL_WAITING(&cpipe->pipe_sel))
 613                         cpipe->pipe_state &= ~PIPE_SEL;
 614         }
 615         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 616                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 617         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 618 }
 619
 620 /*
 621  * Initialize and allocate VM and memory for pipe.  The structure
 622  * will start out zero'd from the ctor, so we just manage the kmem.
 623  */
 624 static void
 625 pipe_create(struct pipe *pipe, int backing)
 626 {
 627
 628         if (backing) {
 629                 /*
 630                  * Note that these functions can fail if pipe map is exhausted
 631                  * (as a result of too many pipes created), but we ignore the
 632                  * error as it is not fatal and could be provoked by
 633                  * unprivileged users. The only consequence is worse performance
 634                  * with given pipe.
 635                  */
 636                 if (amountpipekva > maxpipekva / 2)
 637                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 638                 else
 639                         (void)pipespace_new(pipe, PIPE_SIZE);
 640         }
 641
 642         pipe->pipe_ino = -1;
 643 }
 644
 645 /* ARGSUSED */
 646 static int
 647 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 648     int flags, struct thread *td)
 649 {
 650         struct pipe *rpipe;
 651         int error;
 652         int nread = 0;
 653         int size;
 654
 655         rpipe = fp->f_data;
 656         PIPE_LOCK(rpipe);
 657         ++rpipe->pipe_busy;
 658         error = pipelock(rpipe, 1);
 659         if (error)
 660                 goto unlocked_error;
 661
 662 #ifdef MAC
 663         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 664         if (error)
 665                 goto locked_error;
 666 #endif
 667         if (amountpipekva > (3 * maxpipekva) / 4) {
 668                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 669                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 670                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 671                         (piperesizeallowed == 1)) {
 672                         PIPE_UNLOCK(rpipe);
 673                         pipespace(rpipe, SMALL_PIPE_SIZE);
 674                         PIPE_LOCK(rpipe);
 675                 }
 676         }
 677
 678         while (uio->uio_resid) {
 679                 /*
 680                  * normal pipe buffer receive
 681                  */
 682                 if (rpipe->pipe_buffer.cnt > 0) {
 683                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 684                         if (size > rpipe->pipe_buffer.cnt)
 685                                 size = rpipe->pipe_buffer.cnt;
 686                         if (size > uio->uio_resid)
 687                                 size = uio->uio_resid;
 688
 689                         PIPE_UNLOCK(rpipe);
 690                         error = uiomove(
 691                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 692                             size, uio);
 693                         PIPE_LOCK(rpipe);
 694                         if (error)
 695                                 break;
 696
 697                         rpipe->pipe_buffer.out += size;
 698                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 699                                 rpipe->pipe_buffer.out = 0;
 700
 701                         rpipe->pipe_buffer.cnt -= size;
 702
 703                         /*
 704                          * If there is no more to read in the pipe, reset
 705                          * its pointers to the beginning.  This improves
 706                          * cache hit stats.
 707                          */
 708                         if (rpipe->pipe_buffer.cnt == 0) {
 709                                 rpipe->pipe_buffer.in = 0;
 710                                 rpipe->pipe_buffer.out = 0;
 711                         }
 712                         nread += size;
 713 #ifndef PIPE_NODIRECT
 714                 /*
 715                  * Direct copy, bypassing a kernel buffer.
 716                  */
 717                 } else if ((size = rpipe->pipe_map.cnt) &&
 718                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 719                         if (size > uio->uio_resid)
 720                                 size = (u_int) uio->uio_resid;
 721
 722                         PIPE_UNLOCK(rpipe);
 723                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 724                             rpipe->pipe_map.pos, size, uio);
 725                         PIPE_LOCK(rpipe);
 726                         if (error)
 727                                 break;
 728                         nread += size;
 729                         rpipe->pipe_map.pos += size;
 730                         rpipe->pipe_map.cnt -= size;
 731                         if (rpipe->pipe_map.cnt == 0) {
 732                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 733                                 wakeup(rpipe);
 734                         }
 735 #endif
 736                 } else {
 737                         /*
 738                          * detect EOF condition
 739                          * read returns 0 on EOF, no need to set error
 740                          */
 741                         if (rpipe->pipe_state & PIPE_EOF)
 742                                 break;
 743
 744                         /*
 745                          * If the "write-side" has been blocked, wake it up now.
 746                          */
 747                         if (rpipe->pipe_state & PIPE_WANTW) {
 748                                 rpipe->pipe_state &= ~PIPE_WANTW;
 749                                 wakeup(rpipe);
 750                         }
 751
 752                         /*
 753                          * Break if some data was read.
 754                          */
 755                         if (nread > 0)
 756                                 break;
 757
 758                         /*
 759                          * Unlock the pipe buffer for our remaining processing.
 760                          * We will either break out with an error or we will
 761                          * sleep and relock to loop.
 762                          */
 763                         pipeunlock(rpipe);
 764
 765                         /*
 766                          * Handle non-blocking mode operation or
 767                          * wait for more data.
 768                          */
 769                         if (fp->f_flag & FNONBLOCK) {
 770                                 error = EAGAIN;
 771                         } else {
 772                                 rpipe->pipe_state |= PIPE_WANTR;
 773                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 774                                     PRIBIO | PCATCH,
 775                                     "piperd", 0)) == 0)
 776                                         error = pipelock(rpipe, 1);
 777                         }
 778                         if (error)
 779                                 goto unlocked_error;
 780                 }
 781         }
 782 #ifdef MAC
 783 locked_error:
 784 #endif
 785         pipeunlock(rpipe);
 786
 787         /* XXX: should probably do this before getting any locks. */
 788         if (error == 0)
 789                 vfs_timestamp(&rpipe->pipe_atime);
 790 unlocked_error:
 791         --rpipe->pipe_busy;
 792
 793         /*
 794          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 795          */
 796         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 797                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 798                 wakeup(rpipe);
 799         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 800                 /*
 801                  * Handle write blocking hysteresis.
 802                  */
 803                 if (rpipe->pipe_state & PIPE_WANTW) {
 804                         rpipe->pipe_state &= ~PIPE_WANTW;
 805                         wakeup(rpipe);
 806                 }
 807         }
 808
 809         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 810                 pipeselwakeup(rpipe);
 811
 812         PIPE_UNLOCK(rpipe);
 813         return (error);
 814 }
 815
 816 #ifndef PIPE_NODIRECT
 817 /*
 818  * Map the sending processes' buffer into kernel space and wire it.
 819  * This is similar to a physical write operation.
 820  */
 821 static int
 822 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 823 {
 824         u_int size;
 825         int i;
 826
 827         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 828         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 829                 ("Clone attempt on non-direct write pipe!"));
 830
 831         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 832                 size = wpipe->pipe_buffer.size;
 833         else
 834                 size = uio->uio_iov->iov_len;
 835
 836         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 837             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 838             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 839                 return (EFAULT);
 840
 841 /*
 842  * set up the control block
 843  */
 844         wpipe->pipe_map.npages = i;
 845         wpipe->pipe_map.pos =
 846             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 847         wpipe->pipe_map.cnt = size;
 848
 849 /*
 850  * and update the uio data
 851  */
 852
 853         uio->uio_iov->iov_len -= size;
 854         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 855         if (uio->uio_iov->iov_len == 0)
 856                 uio->uio_iov++;
 857         uio->uio_resid -= size;
 858         uio->uio_offset += size;
 859         return (0);
 860 }
 861
 862 /*
 863  * unmap and unwire the process buffer
 864  */
 865 static void
 866 pipe_destroy_write_buffer(struct pipe *wpipe)
 867 {
 868
 869         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 870         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 871         wpipe->pipe_map.npages = 0;
 872 }
 873
 874 /*
 875  * In the case of a signal, the writing process might go away.  This
 876  * code copies the data into the circular buffer so that the source
 877  * pages can be freed without loss of data.
 878  */
 879 static void
 880 pipe_clone_write_buffer(struct pipe *wpipe)
 881 {
 882         struct uio uio;
 883         struct iovec iov;
 884         int size;
 885         int pos;
 886
 887         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 888         size = wpipe->pipe_map.cnt;
 889         pos = wpipe->pipe_map.pos;
 890
 891         wpipe->pipe_buffer.in = size;
 892         wpipe->pipe_buffer.out = 0;
 893         wpipe->pipe_buffer.cnt = size;
 894         wpipe->pipe_state &= ~PIPE_DIRECTW;
 895
 896         PIPE_UNLOCK(wpipe);
 897         iov.iov_base = wpipe->pipe_buffer.buffer;
 898         iov.iov_len = size;
 899         uio.uio_iov = &iov;
 900         uio.uio_iovcnt = 1;
 901         uio.uio_offset = 0;
 902         uio.uio_resid = size;
 903         uio.uio_segflg = UIO_SYSSPACE;
 904         uio.uio_rw = UIO_READ;
 905         uio.uio_td = curthread;
 906         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 907         PIPE_LOCK(wpipe);
 908         pipe_destroy_write_buffer(wpipe);
 909 }
 910
 911 /*
 912  * This implements the pipe buffer write mechanism.  Note that only
 913  * a direct write OR a normal pipe write can be pending at any given time.
 914  * If there are any characters in the pipe buffer, the direct write will
 915  * be deferred until the receiving process grabs all of the bytes from
 916  * the pipe buffer.  Then the direct mapping write is set-up.
 917  */
 918 static int
 919 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 920 {
 921         int error;
 922
 923 retry:
 924         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 925         error = pipelock(wpipe, 1);
 926         if (error != 0)
 927                 goto error1;
 928         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 929                 error = EPIPE;
 930                 pipeunlock(wpipe);
 931                 goto error1;
 932         }
 933         while (wpipe->pipe_state & PIPE_DIRECTW) {
 934                 if (wpipe->pipe_state & PIPE_WANTR) {
 935                         wpipe->pipe_state &= ~PIPE_WANTR;
 936                         wakeup(wpipe);
 937                 }
 938                 pipeselwakeup(wpipe);
 939                 wpipe->pipe_state |= PIPE_WANTW;
 940                 pipeunlock(wpipe);
 941                 error = msleep(wpipe, PIPE_MTX(wpipe),
 942                     PRIBIO | PCATCH, "pipdww", 0);
 943                 if (error)
 944                         goto error1;
 945                 else
 946                         goto retry;
 947         }
 948         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 949         if (wpipe->pipe_buffer.cnt > 0) {
 950                 if (wpipe->pipe_state & PIPE_WANTR) {
 951                         wpipe->pipe_state &= ~PIPE_WANTR;
 952                         wakeup(wpipe);
 953                 }
 954                 pipeselwakeup(wpipe);
 955                 wpipe->pipe_state |= PIPE_WANTW;
 956                 pipeunlock(wpipe);
 957                 error = msleep(wpipe, PIPE_MTX(wpipe),
 958                     PRIBIO | PCATCH, "pipdwc", 0);
 959                 if (error)
 960                         goto error1;
 961                 else
 962                         goto retry;
 963         }
 964
 965         wpipe->pipe_state |= PIPE_DIRECTW;
 966
 967         PIPE_UNLOCK(wpipe);
 968         error = pipe_build_write_buffer(wpipe, uio);
 969         PIPE_LOCK(wpipe);
 970         if (error) {
 971                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 972                 pipeunlock(wpipe);
 973                 goto error1;
 974         }
 975
 976         error = 0;
 977         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 978                 if (wpipe->pipe_state & PIPE_EOF) {
 979                         pipe_destroy_write_buffer(wpipe);
 980                         pipeselwakeup(wpipe);
 981                         pipeunlock(wpipe);
 982                         error = EPIPE;
 983                         goto error1;
 984                 }
 985                 if (wpipe->pipe_state & PIPE_WANTR) {
 986                         wpipe->pipe_state &= ~PIPE_WANTR;
 987                         wakeup(wpipe);
 988                 }
 989                 pipeselwakeup(wpipe);
 990                 wpipe->pipe_state |= PIPE_WANTW;
 991                 pipeunlock(wpipe);
 992                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 993                     "pipdwt", 0);
 994                 pipelock(wpipe, 0);
 995         }
 996
 997         if (wpipe->pipe_state & PIPE_EOF)
 998                 error = EPIPE;
 999         if (wpipe->pipe_state & PIPE_DIRECTW) {
1000                 /*
1001                  * this bit of trickery substitutes a kernel buffer for
1002                  * the process that might be going away.
1003                  */
1004                 pipe_clone_write_buffer(wpipe);
1005         } else {
1006                 pipe_destroy_write_buffer(wpipe);
1007         }
1008         pipeunlock(wpipe);
1009         return (error);
1010
1011 error1:
1012         wakeup(wpipe);
1013         return (error);
1014 }
1015 #endif
1016
1017 static int
1018 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1019     int flags, struct thread *td)
1020 {
1021         int error = 0;
1022         int desiredsize;
1023         ssize_t orig_resid;
1024         struct pipe *wpipe, *rpipe;
1025
1026         rpipe = fp->f_data;
1027         wpipe = PIPE_PEER(rpipe);
1028         PIPE_LOCK(rpipe);
1029         error = pipelock(wpipe, 1);
1030         if (error) {
1031                 PIPE_UNLOCK(rpipe);
1032                 return (error);
1033         }
1034         /*
1035          * detect loss of pipe read side, issue SIGPIPE if lost.
1036          */
1037         if (wpipe->pipe_present != PIPE_ACTIVE ||
1038             (wpipe->pipe_state & PIPE_EOF)) {
1039                 pipeunlock(wpipe);
1040                 PIPE_UNLOCK(rpipe);
1041                 return (EPIPE);
1042         }
1043 #ifdef MAC
1044         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1045         if (error) {
1046                 pipeunlock(wpipe);
1047                 PIPE_UNLOCK(rpipe);
1048                 return (error);
1049         }
1050 #endif
1051         ++wpipe->pipe_busy;
1052
1053         /* Choose a larger size if it's advantageous */
1054         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1055         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1056                 if (piperesizeallowed != 1)
1057                         break;
1058                 if (amountpipekva > maxpipekva / 2)
1059                         break;
1060                 if (desiredsize == BIG_PIPE_SIZE)
1061                         break;
1062                 desiredsize = desiredsize * 2;
1063         }
1064
1065         /* Choose a smaller size if we're in a OOM situation */
1066         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1067                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1068                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1069                 (piperesizeallowed == 1))
1070                 desiredsize = SMALL_PIPE_SIZE;
1071
1072         /* Resize if the above determined that a new size was necessary */
1073         if ((desiredsize != wpipe->pipe_buffer.size) &&
1074                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1075                 PIPE_UNLOCK(wpipe);
1076                 pipespace(wpipe, desiredsize);
1077                 PIPE_LOCK(wpipe);
1078         }
1079         if (wpipe->pipe_buffer.size == 0) {
1080                 /*
1081                  * This can only happen for reverse direction use of pipes
1082                  * in a complete OOM situation.
1083                  */
1084                 error = ENOMEM;
1085                 --wpipe->pipe_busy;
1086                 pipeunlock(wpipe);
1087                 PIPE_UNLOCK(wpipe);
1088                 return (error);
1089         }
1090
1091         pipeunlock(wpipe);
1092
1093         orig_resid = uio->uio_resid;
1094
1095         while (uio->uio_resid) {
1096                 int space;
1097
1098                 pipelock(wpipe, 0);
1099                 if (wpipe->pipe_state & PIPE_EOF) {
1100                         pipeunlock(wpipe);
1101                         error = EPIPE;
1102                         break;
1103                 }
1104 #ifndef PIPE_NODIRECT
1105                 /*
1106                  * If the transfer is large, we can gain performance if
1107                  * we do process-to-process copies directly.
1108                  * If the write is non-blocking, we don't use the
1109                  * direct write mechanism.
1110                  *
1111                  * The direct write mechanism will detect the reader going
1112                  * away on us.
1113                  */
1114                 if (uio->uio_segflg == UIO_USERSPACE &&
1115                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1116                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1117                     (fp->f_flag & FNONBLOCK) == 0) {
1118                         pipeunlock(wpipe);
1119                         error = pipe_direct_write(wpipe, uio);
1120                         if (error)
1121                                 break;
1122                         continue;
1123                 }
1124 #endif
1125
1126                 /*
1127                  * Pipe buffered writes cannot be coincidental with
1128                  * direct writes.  We wait until the currently executing
1129                  * direct write is completed before we start filling the
1130                  * pipe buffer.  We break out if a signal occurs or the
1131                  * reader goes away.
1132                  */
1133                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1134                         if (wpipe->pipe_state & PIPE_WANTR) {
1135                                 wpipe->pipe_state &= ~PIPE_WANTR;
1136                                 wakeup(wpipe);
1137                         }
1138                         pipeselwakeup(wpipe);
1139                         wpipe->pipe_state |= PIPE_WANTW;
1140                         pipeunlock(wpipe);
1141                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1142                             "pipbww", 0);
1143                         if (error)
1144                                 break;
1145                         else
1146                                 continue;
1147                 }
1148
1149                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1150
1151                 /* Writes of size <= PIPE_BUF must be atomic. */
1152                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1153                         space = 0;
1154
1155                 if (space > 0) {
1156                         int size;       /* Transfer size */
1157                         int segsize;    /* first segment to transfer */
1158
1159                         /*
1160                          * Transfer size is minimum of uio transfer
1161                          * and free space in pipe buffer.
1162                          */
1163                         if (space > uio->uio_resid)
1164                                 size = uio->uio_resid;
1165                         else
1166                                 size = space;
1167                         /*
1168                          * First segment to transfer is minimum of
1169                          * transfer size and contiguous space in
1170                          * pipe buffer.  If first segment to transfer
1171                          * is less than the transfer size, we've got
1172                          * a wraparound in the buffer.
1173                          */
1174                         segsize = wpipe->pipe_buffer.size -
1175                                 wpipe->pipe_buffer.in;
1176                         if (segsize > size)
1177                                 segsize = size;
1178
1179                         /* Transfer first segment */
1180
1181                         PIPE_UNLOCK(rpipe);
1182                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1183                                         segsize, uio);
1184                         PIPE_LOCK(rpipe);
1185
1186                         if (error == 0 && segsize < size) {
1187                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1188                                         wpipe->pipe_buffer.size,
1189                                         ("Pipe buffer wraparound disappeared"));
1190                                 /*
1191                                  * Transfer remaining part now, to
1192                                  * support atomic writes.  Wraparound
1193                                  * happened.
1194                                  */
1195
1196                                 PIPE_UNLOCK(rpipe);
1197                                 error = uiomove(
1198                                     &wpipe->pipe_buffer.buffer[0],
1199                                     size - segsize, uio);
1200                                 PIPE_LOCK(rpipe);
1201                         }
1202                         if (error == 0) {
1203                                 wpipe->pipe_buffer.in += size;
1204                                 if (wpipe->pipe_buffer.in >=
1205                                     wpipe->pipe_buffer.size) {
1206                                         KASSERT(wpipe->pipe_buffer.in ==
1207                                                 size - segsize +
1208                                                 wpipe->pipe_buffer.size,
1209                                                 ("Expected wraparound bad"));
1210                                         wpipe->pipe_buffer.in = size - segsize;
1211                                 }
1212
1213                                 wpipe->pipe_buffer.cnt += size;
1214                                 KASSERT(wpipe->pipe_buffer.cnt <=
1215                                         wpipe->pipe_buffer.size,
1216                                         ("Pipe buffer overflow"));
1217                         }
1218                         pipeunlock(wpipe);
1219                         if (error != 0)
1220                                 break;
1221                 } else {
1222                         /*
1223                          * If the "read-side" has been blocked, wake it up now.
1224                          */
1225                         if (wpipe->pipe_state & PIPE_WANTR) {
1226                                 wpipe->pipe_state &= ~PIPE_WANTR;
1227                                 wakeup(wpipe);
1228                         }
1229
1230                         /*
1231                          * don't block on non-blocking I/O
1232                          */
1233                         if (fp->f_flag & FNONBLOCK) {
1234                                 error = EAGAIN;
1235                                 pipeunlock(wpipe);
1236                                 break;
1237                         }
1238
1239                         /*
1240                          * We have no more space and have something to offer,
1241                          * wake up select/poll.
1242                          */
1243                         pipeselwakeup(wpipe);
1244
1245                         wpipe->pipe_state |= PIPE_WANTW;
1246                         pipeunlock(wpipe);
1247                         error = msleep(wpipe, PIPE_MTX(rpipe),
1248                             PRIBIO | PCATCH, "pipewr", 0);
1249                         if (error != 0)
1250                                 break;
1251                 }
1252         }
1253
1254         pipelock(wpipe, 0);
1255         --wpipe->pipe_busy;
1256
1257         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1258                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1259                 wakeup(wpipe);
1260         } else if (wpipe->pipe_buffer.cnt > 0) {
1261                 /*
1262                  * If we have put any characters in the buffer, we wake up
1263                  * the reader.
1264                  */
1265                 if (wpipe->pipe_state & PIPE_WANTR) {
1266                         wpipe->pipe_state &= ~PIPE_WANTR;
1267                         wakeup(wpipe);
1268                 }
1269         }
1270
1271         /*
1272          * Don't return EPIPE if any byte was written.
1273          * EINTR and other interrupts are handled by generic I/O layer.
1274          * Do not pretend that I/O succeeded for obvious user error
1275          * like EFAULT.
1276          */
1277         if (uio->uio_resid != orig_resid && error == EPIPE)
1278                 error = 0;
1279
1280         if (error == 0)
1281                 vfs_timestamp(&wpipe->pipe_mtime);
1282
1283         /*
1284          * We have something to offer,
1285          * wake up select/poll.
1286          */
1287         if (wpipe->pipe_buffer.cnt)
1288                 pipeselwakeup(wpipe);
1289
1290         pipeunlock(wpipe);
1291         PIPE_UNLOCK(rpipe);
1292         return (error);
1293 }
1294
1295 /* ARGSUSED */
1296 static int
1297 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1298     struct thread *td)
1299 {
1300         struct pipe *cpipe;
1301         int error;
1302
1303         cpipe = fp->f_data;
1304         if (cpipe->pipe_state & PIPE_NAMED)
1305                 error = vnops.fo_truncate(fp, length, active_cred, td);
1306         else
1307                 error = invfo_truncate(fp, length, active_cred, td);
1308         return (error);
1309 }
1310
1311 /*
1312  * we implement a very minimal set of ioctls for compatibility with sockets.
1313  */
1314 static int
1315 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1316     struct thread *td)
1317 {
1318         struct pipe *mpipe = fp->f_data;
1319         int error;
1320
1321         PIPE_LOCK(mpipe);
1322
1323 #ifdef MAC
1324         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1325         if (error) {
1326                 PIPE_UNLOCK(mpipe);
1327                 return (error);
1328         }
1329 #endif
1330
1331         error = 0;
1332         switch (cmd) {
1333
1334         case FIONBIO:
1335                 break;
1336
1337         case FIOASYNC:
1338                 if (*(int *)data) {
1339                         mpipe->pipe_state |= PIPE_ASYNC;
1340                 } else {
1341                         mpipe->pipe_state &= ~PIPE_ASYNC;
1342                 }
1343                 break;
1344
1345         case FIONREAD:
1346                 if (!(fp->f_flag & FREAD)) {
1347                         *(int *)data = 0;
1348                         PIPE_UNLOCK(mpipe);
1349                         return (0);
1350                 }
1351                 if (mpipe->pipe_state & PIPE_DIRECTW)
1352                         *(int *)data = mpipe->pipe_map.cnt;
1353                 else
1354                         *(int *)data = mpipe->pipe_buffer.cnt;
1355                 break;
1356
1357         case FIOSETOWN:
1358                 PIPE_UNLOCK(mpipe);
1359                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1360                 goto out_unlocked;
1361
1362         case FIOGETOWN:
1363                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1364                 break;
1365
1366         /* This is deprecated, FIOSETOWN should be used instead. */
1367         case TIOCSPGRP:
1368                 PIPE_UNLOCK(mpipe);
1369                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1370                 goto out_unlocked;
1371
1372         /* This is deprecated, FIOGETOWN should be used instead. */
1373         case TIOCGPGRP:
1374                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1375                 break;
1376
1377         default:
1378                 error = ENOTTY;
1379                 break;
1380         }
1381         PIPE_UNLOCK(mpipe);
1382 out_unlocked:
1383         return (error);
1384 }
1385
1386 static int
1387 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1388     struct thread *td)
1389 {
1390         struct pipe *rpipe;
1391         struct pipe *wpipe;
1392         int levents, revents;
1393 #ifdef MAC
1394         int error;
1395 #endif
1396
1397         revents = 0;
1398         rpipe = fp->f_data;
1399         wpipe = PIPE_PEER(rpipe);
1400         PIPE_LOCK(rpipe);
1401 #ifdef MAC
1402         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1403         if (error)
1404                 goto locked_error;
1405 #endif
1406         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1407                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1408                     (rpipe->pipe_buffer.cnt > 0))
1409                         revents |= events & (POLLIN | POLLRDNORM);
1410
1411         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1412                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1413                     (wpipe->pipe_state & PIPE_EOF) ||
1414                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1415                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1416                          wpipe->pipe_buffer.size == 0)))
1417                         revents |= events & (POLLOUT | POLLWRNORM);
1418
1419         levents = events &
1420             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1421         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1422             fp->f_seqcount == rpipe->pipe_wgen)
1423                 events |= POLLINIGNEOF;
1424
1425         if ((events & POLLINIGNEOF) == 0) {
1426                 if (rpipe->pipe_state & PIPE_EOF) {
1427                         revents |= (events & (POLLIN | POLLRDNORM));
1428                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1429                             (wpipe->pipe_state & PIPE_EOF))
1430                                 revents |= POLLHUP;
1431                 }
1432         }
1433
1434         if (revents == 0) {
1435                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1436                         selrecord(td, &rpipe->pipe_sel);
1437                         if (SEL_WAITING(&rpipe->pipe_sel))
1438                                 rpipe->pipe_state |= PIPE_SEL;
1439                 }
1440
1441                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1442                         selrecord(td, &wpipe->pipe_sel);
1443                         if (SEL_WAITING(&wpipe->pipe_sel))
1444                                 wpipe->pipe_state |= PIPE_SEL;
1445                 }
1446         }
1447 #ifdef MAC
1448 locked_error:
1449 #endif
1450         PIPE_UNLOCK(rpipe);
1451
1452         return (revents);
1453 }
1454
1455 /*
1456  * We shouldn't need locks here as we're doing a read and this should
1457  * be a natural race.
1458  */
1459 static int
1460 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1461     struct thread *td)
1462 {
1463         struct pipe *pipe;
1464         int new_unr;
1465 #ifdef MAC
1466         int error;
1467 #endif
1468
1469         pipe = fp->f_data;
1470         PIPE_LOCK(pipe);
1471 #ifdef MAC
1472         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1473         if (error) {
1474                 PIPE_UNLOCK(pipe);
1475                 return (error);
1476         }
1477 #endif
1478
1479         /* For named pipes ask the underlying filesystem. */
1480         if (pipe->pipe_state & PIPE_NAMED) {
1481                 PIPE_UNLOCK(pipe);
1482                 return (vnops.fo_stat(fp, ub, active_cred, td));
1483         }
1484
1485         /*
1486          * Lazily allocate an inode number for the pipe.  Most pipe
1487          * users do not call fstat(2) on the pipe, which means that
1488          * postponing the inode allocation until it is must be
1489          * returned to userland is useful.  If alloc_unr failed,
1490          * assign st_ino zero instead of returning an error.
1491          * Special pipe_ino values:
1492          *  -1 - not yet initialized;
1493          *  0  - alloc_unr failed, return 0 as st_ino forever.
1494          */
1495         if (pipe->pipe_ino == (ino_t)-1) {
1496                 new_unr = alloc_unr(pipeino_unr);
1497                 if (new_unr != -1)
1498                         pipe->pipe_ino = new_unr;
1499                 else
1500                         pipe->pipe_ino = 0;
1501         }
1502         PIPE_UNLOCK(pipe);
1503
1504         bzero(ub, sizeof(*ub));
1505         ub->st_mode = S_IFIFO;
1506         ub->st_blksize = PAGE_SIZE;
1507         if (pipe->pipe_state & PIPE_DIRECTW)
1508                 ub->st_size = pipe->pipe_map.cnt;
1509         else
1510                 ub->st_size = pipe->pipe_buffer.cnt;
1511         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1512         ub->st_atim = pipe->pipe_atime;
1513         ub->st_mtim = pipe->pipe_mtime;
1514         ub->st_ctim = pipe->pipe_ctime;
1515         ub->st_uid = fp->f_cred->cr_uid;
1516         ub->st_gid = fp->f_cred->cr_gid;
1517         ub->st_dev = pipedev_ino;
1518         ub->st_ino = pipe->pipe_ino;
1519         /*
1520          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1521          */
1522         return (0);
1523 }
1524
1525 /* ARGSUSED */
1526 static int
1527 pipe_close(struct file *fp, struct thread *td)
1528 {
1529
1530         if (fp->f_vnode != NULL)
1531                 return vnops.fo_close(fp, td);
1532         fp->f_ops = &badfileops;
1533         pipe_dtor(fp->f_data);
1534         fp->f_data = NULL;
1535         return (0);
1536 }
1537
1538 static int
1539 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1540 {
1541         struct pipe *cpipe;
1542         int error;
1543
1544         cpipe = fp->f_data;
1545         if (cpipe->pipe_state & PIPE_NAMED)
1546                 error = vn_chmod(fp, mode, active_cred, td);
1547         else
1548                 error = invfo_chmod(fp, mode, active_cred, td);
1549         return (error);
1550 }
1551
1552 static int
1553 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1554     struct thread *td)
1555 {
1556         struct pipe *cpipe;
1557         int error;
1558
1559         cpipe = fp->f_data;
1560         if (cpipe->pipe_state & PIPE_NAMED)
1561                 error = vn_chown(fp, uid, gid, active_cred, td);
1562         else
1563                 error = invfo_chown(fp, uid, gid, active_cred, td);
1564         return (error);
1565 }
1566
1567 static int
1568 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1569 {
1570         struct pipe *pi;
1571
1572         if (fp->f_type == DTYPE_FIFO)
1573                 return (vn_fill_kinfo(fp, kif, fdp));
1574         kif->kf_type = KF_TYPE_PIPE;
1575         pi = fp->f_data;
1576         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1577         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1578         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1579         return (0);
1580 }
1581
1582 static void
1583 pipe_free_kmem(struct pipe *cpipe)
1584 {
1585
1586         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1587             ("pipe_free_kmem: pipe mutex locked"));
1588
1589         if (cpipe->pipe_buffer.buffer != NULL) {
1590                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1591                 vm_map_remove(pipe_map,
1592                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1593                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1594                 cpipe->pipe_buffer.buffer = NULL;
1595         }
1596 #ifndef PIPE_NODIRECT
1597         {
1598                 cpipe->pipe_map.cnt = 0;
1599                 cpipe->pipe_map.pos = 0;
1600                 cpipe->pipe_map.npages = 0;
1601         }
1602 #endif
1603 }
1604
1605 /*
1606  * shutdown the pipe
1607  */
1608 static void
1609 pipeclose(struct pipe *cpipe)
1610 {
1611         struct pipepair *pp;
1612         struct pipe *ppipe;
1613
1614         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1615
1616         PIPE_LOCK(cpipe);
1617         pipelock(cpipe, 0);
1618         pp = cpipe->pipe_pair;
1619
1620         pipeselwakeup(cpipe);
1621
1622         /*
1623          * If the other side is blocked, wake it up saying that
1624          * we want to close it down.
1625          */
1626         cpipe->pipe_state |= PIPE_EOF;
1627         while (cpipe->pipe_busy) {
1628                 wakeup(cpipe);
1629                 cpipe->pipe_state |= PIPE_WANT;
1630                 pipeunlock(cpipe);
1631                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1632                 pipelock(cpipe, 0);
1633         }
1634
1635
1636         /*
1637          * Disconnect from peer, if any.
1638          */
1639         ppipe = cpipe->pipe_peer;
1640         if (ppipe->pipe_present == PIPE_ACTIVE) {
1641                 pipeselwakeup(ppipe);
1642
1643                 ppipe->pipe_state |= PIPE_EOF;
1644                 wakeup(ppipe);
1645                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1646         }
1647
1648         /*
1649          * Mark this endpoint as free.  Release kmem resources.  We
1650          * don't mark this endpoint as unused until we've finished
1651          * doing that, or the pipe might disappear out from under
1652          * us.
1653          */
1654         PIPE_UNLOCK(cpipe);
1655         pipe_free_kmem(cpipe);
1656         PIPE_LOCK(cpipe);
1657         cpipe->pipe_present = PIPE_CLOSING;
1658         pipeunlock(cpipe);
1659
1660         /*
1661          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1662          * PIPE_FINALIZED, that allows other end to free the
1663          * pipe_pair, only after the knotes are completely dismantled.
1664          */
1665         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1666         cpipe->pipe_present = PIPE_FINALIZED;
1667         seldrain(&cpipe->pipe_sel);
1668         knlist_destroy(&cpipe->pipe_sel.si_note);
1669
1670         /*
1671          * If both endpoints are now closed, release the memory for the
1672          * pipe pair.  If not, unlock.
1673          */
1674         if (ppipe->pipe_present == PIPE_FINALIZED) {
1675                 PIPE_UNLOCK(cpipe);
1676 #ifdef MAC
1677                 mac_pipe_destroy(pp);
1678 #endif
1679                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1680         } else
1681                 PIPE_UNLOCK(cpipe);
1682 }
1683
1684 /*ARGSUSED*/
1685 static int
1686 pipe_kqfilter(struct file *fp, struct knote *kn)
1687 {
1688         struct pipe *cpipe;
1689
1690         /*
1691          * If a filter is requested that is not supported by this file
1692          * descriptor, don't return an error, but also don't ever generate an
1693          * event.
1694          */
1695         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1696                 kn->kn_fop = &pipe_nfiltops;
1697                 return (0);
1698         }
1699         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1700                 kn->kn_fop = &pipe_nfiltops;
1701                 return (0);
1702         }
1703         cpipe = fp->f_data;
1704         PIPE_LOCK(cpipe);
1705         switch (kn->kn_filter) {
1706         case EVFILT_READ:
1707                 kn->kn_fop = &pipe_rfiltops;
1708                 break;
1709         case EVFILT_WRITE:
1710                 kn->kn_fop = &pipe_wfiltops;
1711                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1712                         /* other end of pipe has been closed */
1713                         PIPE_UNLOCK(cpipe);
1714                         return (EPIPE);
1715                 }
1716                 cpipe = PIPE_PEER(cpipe);
1717                 break;
1718         default:
1719                 PIPE_UNLOCK(cpipe);
1720                 return (EINVAL);
1721         }
1722
1723         kn->kn_hook = cpipe;
1724         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1725         PIPE_UNLOCK(cpipe);
1726         return (0);
1727 }
1728
1729 static void
1730 filt_pipedetach(struct knote *kn)
1731 {
1732         struct pipe *cpipe = kn->kn_hook;
1733
1734         PIPE_LOCK(cpipe);
1735         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1736         PIPE_UNLOCK(cpipe);
1737 }
1738
1739 /*ARGSUSED*/
1740 static int
1741 filt_piperead(struct knote *kn, long hint)
1742 {
1743         struct pipe *rpipe = kn->kn_hook;
1744         struct pipe *wpipe = rpipe->pipe_peer;
1745         int ret;
1746
1747         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1748         kn->kn_data = rpipe->pipe_buffer.cnt;
1749         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1750                 kn->kn_data = rpipe->pipe_map.cnt;
1751
1752         if ((rpipe->pipe_state & PIPE_EOF) ||
1753             wpipe->pipe_present != PIPE_ACTIVE ||
1754             (wpipe->pipe_state & PIPE_EOF)) {
1755                 kn->kn_flags |= EV_EOF;
1756                 return (1);
1757         }
1758         ret = kn->kn_data > 0;
1759         return ret;
1760 }
1761
1762 /*ARGSUSED*/
1763 static int
1764 filt_pipewrite(struct knote *kn, long hint)
1765 {
1766         struct pipe *wpipe;
1767
1768         wpipe = kn->kn_hook;
1769         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1770         if (wpipe->pipe_present != PIPE_ACTIVE ||
1771             (wpipe->pipe_state & PIPE_EOF)) {
1772                 kn->kn_data = 0;
1773                 kn->kn_flags |= EV_EOF;
1774                 return (1);
1775         }
1776         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1777             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1778         if (wpipe->pipe_state & PIPE_DIRECTW)
1779                 kn->kn_data = 0;
1780
1781         return (kn->kn_data >= PIPE_BUF);
1782 }
1783
1784 static void
1785 filt_pipedetach_notsup(struct knote *kn)
1786 {
1787
1788 }
1789
1790 static int
1791 filt_pipenotsup(struct knote *kn, long hint)
1792 {
1793
1794         return (0);
1795 }