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 unrhdr64 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         new_unrhdr64(&pipeino_unr, 1);
 261         pipedev_ino = devfs_alloc_cdp_inode();
 262         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 263 }
 264
 265 static int
 266 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 267 {
 268         struct pipepair *pp;
 269         struct pipe *rpipe, *wpipe;
 270
 271         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 272
 273         pp = (struct pipepair *)mem;
 274
 275         /*
 276          * We zero both pipe endpoints to make sure all the kmem pointers
 277          * are NULL, flag fields are zero'd, etc.  We timestamp both
 278          * endpoints with the same time.
 279          */
 280         rpipe = &pp->pp_rpipe;
 281         bzero(rpipe, sizeof(*rpipe));
 282         vfs_timestamp(&rpipe->pipe_ctime);
 283         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 284
 285         wpipe = &pp->pp_wpipe;
 286         bzero(wpipe, sizeof(*wpipe));
 287         wpipe->pipe_ctime = rpipe->pipe_ctime;
 288         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 289
 290         rpipe->pipe_peer = wpipe;
 291         rpipe->pipe_pair = pp;
 292         wpipe->pipe_peer = rpipe;
 293         wpipe->pipe_pair = pp;
 294
 295         /*
 296          * Mark both endpoints as present; they will later get free'd
 297          * one at a time.  When both are free'd, then the whole pair
 298          * is released.
 299          */
 300         rpipe->pipe_present = PIPE_ACTIVE;
 301         wpipe->pipe_present = PIPE_ACTIVE;
 302
 303         /*
 304          * Eventually, the MAC Framework may initialize the label
 305          * in ctor or init, but for now we do it elswhere to avoid
 306          * blocking in ctor or init.
 307          */
 308         pp->pp_label = NULL;
 309
 310         return (0);
 311 }
 312
 313 static int
 314 pipe_zone_init(void *mem, int size, int flags)
 315 {
 316         struct pipepair *pp;
 317
 318         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 319
 320         pp = (struct pipepair *)mem;
 321
 322         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 323         return (0);
 324 }
 325
 326 static void
 327 pipe_zone_fini(void *mem, int size)
 328 {
 329         struct pipepair *pp;
 330
 331         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 332
 333         pp = (struct pipepair *)mem;
 334
 335         mtx_destroy(&pp->pp_mtx);
 336 }
 337
 338 static void
 339 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 340 {
 341         struct pipepair *pp;
 342         struct pipe *rpipe, *wpipe;
 343
 344         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 345 #ifdef MAC
 346         /*
 347          * The MAC label is shared between the connected endpoints.  As a
 348          * result mac_pipe_init() and mac_pipe_create() are called once
 349          * for the pair, and not on the endpoints.
 350          */
 351         mac_pipe_init(pp);
 352         mac_pipe_create(td->td_ucred, pp);
 353 #endif
 354         rpipe = &pp->pp_rpipe;
 355         wpipe = &pp->pp_wpipe;
 356
 357         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 358         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 359
 360         /* Only the forward direction pipe is backed by default */
 361         pipe_create(rpipe, 1);
 362         pipe_create(wpipe, 0);
 363
 364         rpipe->pipe_state |= PIPE_DIRECTOK;
 365         wpipe->pipe_state |= PIPE_DIRECTOK;
 366 }
 367
 368 void
 369 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 370 {
 371         struct pipepair *pp;
 372
 373         pipe_paircreate(td, &pp);
 374         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 375         *ppipe = &pp->pp_rpipe;
 376 }
 377
 378 void
 379 pipe_dtor(struct pipe *dpipe)
 380 {
 381         struct pipe *peer;
 382
 383         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 384         funsetown(&dpipe->pipe_sigio);
 385         pipeclose(dpipe);
 386         if (peer != NULL) {
 387                 funsetown(&peer->pipe_sigio);
 388                 pipeclose(peer);
 389         }
 390 }
 391
 392 /*
 393  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 394  * the zone pick up the pieces via pipeclose().
 395  */
 396 int
 397 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 398     struct filecaps *fcaps2)
 399 {
 400         struct file *rf, *wf;
 401         struct pipe *rpipe, *wpipe;
 402         struct pipepair *pp;
 403         int fd, fflags, error;
 404
 405         pipe_paircreate(td, &pp);
 406         rpipe = &pp->pp_rpipe;
 407         wpipe = &pp->pp_wpipe;
 408         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 409         if (error) {
 410                 pipeclose(rpipe);
 411                 pipeclose(wpipe);
 412                 return (error);
 413         }
 414         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 415         fildes[0] = fd;
 416
 417         fflags = FREAD | FWRITE;
 418         if ((flags & O_NONBLOCK) != 0)
 419                 fflags |= FNONBLOCK;
 420
 421         /*
 422          * Warning: once we've gotten past allocation of the fd for the
 423          * read-side, we can only drop the read side via fdrop() in order
 424          * to avoid races against processes which manage to dup() the read
 425          * side while we are blocked trying to allocate the write side.
 426          */
 427         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 428         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 429         if (error) {
 430                 fdclose(td, rf, fildes[0]);
 431                 fdrop(rf, td);
 432                 /* rpipe has been closed by fdrop(). */
 433                 pipeclose(wpipe);
 434                 return (error);
 435         }
 436         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 437         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 438         fdrop(wf, td);
 439         fildes[1] = fd;
 440         fdrop(rf, td);
 441
 442         return (0);
 443 }
 444
 445 #ifdef COMPAT_FREEBSD10
 446 /* ARGSUSED */
 447 int
 448 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 449 {
 450         int error;
 451         int fildes[2];
 452
 453         error = kern_pipe(td, fildes, 0, NULL, NULL);
 454         if (error)
 455                 return (error);
 456
 457         td->td_retval[0] = fildes[0];
 458         td->td_retval[1] = fildes[1];
 459
 460         return (0);
 461 }
 462 #endif
 463
 464 int
 465 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 466 {
 467         int error, fildes[2];
 468
 469         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 470                 return (EINVAL);
 471         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 472         if (error)
 473                 return (error);
 474         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 475         if (error) {
 476                 (void)kern_close(td, fildes[0]);
 477                 (void)kern_close(td, fildes[1]);
 478         }
 479         return (error);
 480 }
 481
 482 /*
 483  * Allocate kva for pipe circular buffer, the space is pageable
 484  * This routine will 'realloc' the size of a pipe safely, if it fails
 485  * it will retain the old buffer.
 486  * If it fails it will return ENOMEM.
 487  */
 488 static int
 489 pipespace_new(struct pipe *cpipe, int size)
 490 {
 491         caddr_t buffer;
 492         int error, cnt, firstseg;
 493         static int curfail = 0;
 494         static struct timeval lastfail;
 495
 496         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 497         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 498                 ("pipespace: resize of direct writes not allowed"));
 499 retry:
 500         cnt = cpipe->pipe_buffer.cnt;
 501         if (cnt > size)
 502                 size = cnt;
 503
 504         size = round_page(size);
 505         buffer = (caddr_t) vm_map_min(pipe_map);
 506
 507         error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
 508             VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
 509         if (error != KERN_SUCCESS) {
 510                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 511                         (size > SMALL_PIPE_SIZE)) {
 512                         size = SMALL_PIPE_SIZE;
 513                         pipefragretry++;
 514                         goto retry;
 515                 }
 516                 if (cpipe->pipe_buffer.buffer == NULL) {
 517                         pipeallocfail++;
 518                         if (ppsratecheck(&lastfail, &curfail, 1))
 519                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 520                 } else {
 521                         piperesizefail++;
 522                 }
 523                 return (ENOMEM);
 524         }
 525
 526         /* copy data, then free old resources if we're resizing */
 527         if (cnt > 0) {
 528                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 529                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 530                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 531                                 buffer, firstseg);
 532                         if ((cnt - firstseg) > 0)
 533                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 534                                         cpipe->pipe_buffer.in);
 535                 } else {
 536                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 537                                 buffer, cnt);
 538                 }
 539         }
 540         pipe_free_kmem(cpipe);
 541         cpipe->pipe_buffer.buffer = buffer;
 542         cpipe->pipe_buffer.size = size;
 543         cpipe->pipe_buffer.in = cnt;
 544         cpipe->pipe_buffer.out = 0;
 545         cpipe->pipe_buffer.cnt = cnt;
 546         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 547         return (0);
 548 }
 549
 550 /*
 551  * Wrapper for pipespace_new() that performs locking assertions.
 552  */
 553 static int
 554 pipespace(struct pipe *cpipe, int size)
 555 {
 556
 557         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 558                 ("Unlocked pipe passed to pipespace"));
 559         return (pipespace_new(cpipe, size));
 560 }
 561
 562 /*
 563  * lock a pipe for I/O, blocking other access
 564  */
 565 static __inline int
 566 pipelock(struct pipe *cpipe, int catch)
 567 {
 568         int error;
 569
 570         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 571         while (cpipe->pipe_state & PIPE_LOCKFL) {
 572                 cpipe->pipe_state |= PIPE_LWANT;
 573                 error = msleep(cpipe, PIPE_MTX(cpipe),
 574                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 575                     "pipelk", 0);
 576                 if (error != 0)
 577                         return (error);
 578         }
 579         cpipe->pipe_state |= PIPE_LOCKFL;
 580         return (0);
 581 }
 582
 583 /*
 584  * unlock a pipe I/O lock
 585  */
 586 static __inline void
 587 pipeunlock(struct pipe *cpipe)
 588 {
 589
 590         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 591         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 592                 ("Unlocked pipe passed to pipeunlock"));
 593         cpipe->pipe_state &= ~PIPE_LOCKFL;
 594         if (cpipe->pipe_state & PIPE_LWANT) {
 595                 cpipe->pipe_state &= ~PIPE_LWANT;
 596                 wakeup(cpipe);
 597         }
 598 }
 599
 600 void
 601 pipeselwakeup(struct pipe *cpipe)
 602 {
 603
 604         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 605         if (cpipe->pipe_state & PIPE_SEL) {
 606                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 607                 if (!SEL_WAITING(&cpipe->pipe_sel))
 608                         cpipe->pipe_state &= ~PIPE_SEL;
 609         }
 610         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 611                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 612         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 613 }
 614
 615 /*
 616  * Initialize and allocate VM and memory for pipe.  The structure
 617  * will start out zero'd from the ctor, so we just manage the kmem.
 618  */
 619 static void
 620 pipe_create(struct pipe *pipe, int backing)
 621 {
 622
 623         if (backing) {
 624                 /*
 625                  * Note that these functions can fail if pipe map is exhausted
 626                  * (as a result of too many pipes created), but we ignore the
 627                  * error as it is not fatal and could be provoked by
 628                  * unprivileged users. The only consequence is worse performance
 629                  * with given pipe.
 630                  */
 631                 if (amountpipekva > maxpipekva / 2)
 632                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 633                 else
 634                         (void)pipespace_new(pipe, PIPE_SIZE);
 635         }
 636
 637         pipe->pipe_ino = alloc_unr64(&pipeino_unr);
 638 }
 639
 640 /* ARGSUSED */
 641 static int
 642 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 643     int flags, struct thread *td)
 644 {
 645         struct pipe *rpipe;
 646         int error;
 647         int nread = 0;
 648         int size;
 649
 650         rpipe = fp->f_data;
 651         PIPE_LOCK(rpipe);
 652         ++rpipe->pipe_busy;
 653         error = pipelock(rpipe, 1);
 654         if (error)
 655                 goto unlocked_error;
 656
 657 #ifdef MAC
 658         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 659         if (error)
 660                 goto locked_error;
 661 #endif
 662         if (amountpipekva > (3 * maxpipekva) / 4) {
 663                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 664                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 665                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 666                         (piperesizeallowed == 1)) {
 667                         PIPE_UNLOCK(rpipe);
 668                         pipespace(rpipe, SMALL_PIPE_SIZE);
 669                         PIPE_LOCK(rpipe);
 670                 }
 671         }
 672
 673         while (uio->uio_resid) {
 674                 /*
 675                  * normal pipe buffer receive
 676                  */
 677                 if (rpipe->pipe_buffer.cnt > 0) {
 678                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 679                         if (size > rpipe->pipe_buffer.cnt)
 680                                 size = rpipe->pipe_buffer.cnt;
 681                         if (size > uio->uio_resid)
 682                                 size = uio->uio_resid;
 683
 684                         PIPE_UNLOCK(rpipe);
 685                         error = uiomove(
 686                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 687                             size, uio);
 688                         PIPE_LOCK(rpipe);
 689                         if (error)
 690                                 break;
 691
 692                         rpipe->pipe_buffer.out += size;
 693                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 694                                 rpipe->pipe_buffer.out = 0;
 695
 696                         rpipe->pipe_buffer.cnt -= size;
 697
 698                         /*
 699                          * If there is no more to read in the pipe, reset
 700                          * its pointers to the beginning.  This improves
 701                          * cache hit stats.
 702                          */
 703                         if (rpipe->pipe_buffer.cnt == 0) {
 704                                 rpipe->pipe_buffer.in = 0;
 705                                 rpipe->pipe_buffer.out = 0;
 706                         }
 707                         nread += size;
 708 #ifndef PIPE_NODIRECT
 709                 /*
 710                  * Direct copy, bypassing a kernel buffer.
 711                  */
 712                 } else if ((size = rpipe->pipe_map.cnt) != 0) {
 713                         if (size > uio->uio_resid)
 714                                 size = (u_int) uio->uio_resid;
 715                         PIPE_UNLOCK(rpipe);
 716                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 717                             rpipe->pipe_map.pos, size, uio);
 718                         PIPE_LOCK(rpipe);
 719                         if (error)
 720                                 break;
 721                         nread += size;
 722                         rpipe->pipe_map.pos += size;
 723                         rpipe->pipe_map.cnt -= size;
 724                         if (rpipe->pipe_map.cnt == 0) {
 725                                 rpipe->pipe_state &= ~PIPE_WANTW;
 726                                 wakeup(rpipe);
 727                         }
 728 #endif
 729                 } else {
 730                         /*
 731                          * detect EOF condition
 732                          * read returns 0 on EOF, no need to set error
 733                          */
 734                         if (rpipe->pipe_state & PIPE_EOF)
 735                                 break;
 736
 737                         /*
 738                          * If the "write-side" has been blocked, wake it up now.
 739                          */
 740                         if (rpipe->pipe_state & PIPE_WANTW) {
 741                                 rpipe->pipe_state &= ~PIPE_WANTW;
 742                                 wakeup(rpipe);
 743                         }
 744
 745                         /*
 746                          * Break if some data was read.
 747                          */
 748                         if (nread > 0)
 749                                 break;
 750
 751                         /*
 752                          * Unlock the pipe buffer for our remaining processing.
 753                          * We will either break out with an error or we will
 754                          * sleep and relock to loop.
 755                          */
 756                         pipeunlock(rpipe);
 757
 758                         /*
 759                          * Handle non-blocking mode operation or
 760                          * wait for more data.
 761                          */
 762                         if (fp->f_flag & FNONBLOCK) {
 763                                 error = EAGAIN;
 764                         } else {
 765                                 rpipe->pipe_state |= PIPE_WANTR;
 766                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 767                                     PRIBIO | PCATCH,
 768                                     "piperd", 0)) == 0)
 769                                         error = pipelock(rpipe, 1);
 770                         }
 771                         if (error)
 772                                 goto unlocked_error;
 773                 }
 774         }
 775 #ifdef MAC
 776 locked_error:
 777 #endif
 778         pipeunlock(rpipe);
 779
 780         /* XXX: should probably do this before getting any locks. */
 781         if (error == 0)
 782                 vfs_timestamp(&rpipe->pipe_atime);
 783 unlocked_error:
 784         --rpipe->pipe_busy;
 785
 786         /*
 787          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 788          */
 789         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 790                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 791                 wakeup(rpipe);
 792         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 793                 /*
 794                  * Handle write blocking hysteresis.
 795                  */
 796                 if (rpipe->pipe_state & PIPE_WANTW) {
 797                         rpipe->pipe_state &= ~PIPE_WANTW;
 798                         wakeup(rpipe);
 799                 }
 800         }
 801
 802         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 803                 pipeselwakeup(rpipe);
 804
 805         PIPE_UNLOCK(rpipe);
 806         return (error);
 807 }
 808
 809 #ifndef PIPE_NODIRECT
 810 /*
 811  * Map the sending processes' buffer into kernel space and wire it.
 812  * This is similar to a physical write operation.
 813  */
 814 static int
 815 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 816 {
 817         u_int size;
 818         int i;
 819
 820         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 821         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
 822             ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
 823         KASSERT(wpipe->pipe_map.cnt == 0,
 824             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 825
 826         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 827                 size = wpipe->pipe_buffer.size;
 828         else
 829                 size = uio->uio_iov->iov_len;
 830
 831         wpipe->pipe_state |= PIPE_DIRECTW;
 832         PIPE_UNLOCK(wpipe);
 833         i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 834             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 835             wpipe->pipe_map.ms, PIPENPAGES);
 836         PIPE_LOCK(wpipe);
 837         if (i < 0) {
 838                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 839                 return (EFAULT);
 840         }
 841
 842         wpipe->pipe_map.npages = i;
 843         wpipe->pipe_map.pos =
 844             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 845         wpipe->pipe_map.cnt = size;
 846
 847         uio->uio_iov->iov_len -= size;
 848         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 849         if (uio->uio_iov->iov_len == 0)
 850                 uio->uio_iov++;
 851         uio->uio_resid -= size;
 852         uio->uio_offset += size;
 853         return (0);
 854 }
 855
 856 /*
 857  * Unwire the process buffer.
 858  */
 859 static void
 860 pipe_destroy_write_buffer(struct pipe *wpipe)
 861 {
 862
 863         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 864         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 865             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 866         KASSERT(wpipe->pipe_map.cnt == 0,
 867             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 868
 869         wpipe->pipe_state &= ~PIPE_DIRECTW;
 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         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 889             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 890
 891         size = wpipe->pipe_map.cnt;
 892         pos = wpipe->pipe_map.pos;
 893         wpipe->pipe_map.cnt = 0;
 894
 895         wpipe->pipe_buffer.in = size;
 896         wpipe->pipe_buffer.out = 0;
 897         wpipe->pipe_buffer.cnt = size;
 898
 899         PIPE_UNLOCK(wpipe);
 900         iov.iov_base = wpipe->pipe_buffer.buffer;
 901         iov.iov_len = size;
 902         uio.uio_iov = &iov;
 903         uio.uio_iovcnt = 1;
 904         uio.uio_offset = 0;
 905         uio.uio_resid = size;
 906         uio.uio_segflg = UIO_SYSSPACE;
 907         uio.uio_rw = UIO_READ;
 908         uio.uio_td = curthread;
 909         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 910         PIPE_LOCK(wpipe);
 911         pipe_destroy_write_buffer(wpipe);
 912 }
 913
 914 /*
 915  * This implements the pipe buffer write mechanism.  Note that only
 916  * a direct write OR a normal pipe write can be pending at any given time.
 917  * If there are any characters in the pipe buffer, the direct write will
 918  * be deferred until the receiving process grabs all of the bytes from
 919  * the pipe buffer.  Then the direct mapping write is set-up.
 920  */
 921 static int
 922 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 923 {
 924         int error;
 925
 926 retry:
 927         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 928         error = pipelock(wpipe, 1);
 929         if (error != 0)
 930                 goto error1;
 931         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 932                 error = EPIPE;
 933                 pipeunlock(wpipe);
 934                 goto error1;
 935         }
 936         if (wpipe->pipe_state & PIPE_DIRECTW) {
 937                 if (wpipe->pipe_state & PIPE_WANTR) {
 938                         wpipe->pipe_state &= ~PIPE_WANTR;
 939                         wakeup(wpipe);
 940                 }
 941                 pipeselwakeup(wpipe);
 942                 wpipe->pipe_state |= PIPE_WANTW;
 943                 pipeunlock(wpipe);
 944                 error = msleep(wpipe, PIPE_MTX(wpipe),
 945                     PRIBIO | PCATCH, "pipdww", 0);
 946                 if (error)
 947                         goto error1;
 948                 else
 949                         goto retry;
 950         }
 951         if (wpipe->pipe_buffer.cnt > 0) {
 952                 if (wpipe->pipe_state & PIPE_WANTR) {
 953                         wpipe->pipe_state &= ~PIPE_WANTR;
 954                         wakeup(wpipe);
 955                 }
 956                 pipeselwakeup(wpipe);
 957                 wpipe->pipe_state |= PIPE_WANTW;
 958                 pipeunlock(wpipe);
 959                 error = msleep(wpipe, PIPE_MTX(wpipe),
 960                     PRIBIO | PCATCH, "pipdwc", 0);
 961                 if (error)
 962                         goto error1;
 963                 else
 964                         goto retry;
 965         }
 966
 967         error = pipe_build_write_buffer(wpipe, uio);
 968         if (error) {
 969                 pipeunlock(wpipe);
 970                 goto error1;
 971         }
 972
 973         while (wpipe->pipe_map.cnt != 0 &&
 974             (wpipe->pipe_state & PIPE_EOF) == 0) {
 975                 if (wpipe->pipe_state & PIPE_WANTR) {
 976                         wpipe->pipe_state &= ~PIPE_WANTR;
 977                         wakeup(wpipe);
 978                 }
 979                 pipeselwakeup(wpipe);
 980                 wpipe->pipe_state |= PIPE_WANTW;
 981                 pipeunlock(wpipe);
 982                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 983                     "pipdwt", 0);
 984                 pipelock(wpipe, 0);
 985                 if (error != 0)
 986                         break;
 987         }
 988
 989         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 990                 wpipe->pipe_map.cnt = 0;
 991                 pipe_destroy_write_buffer(wpipe);
 992                 pipeselwakeup(wpipe);
 993                 error = EPIPE;
 994         } else if (error == EINTR || error == ERESTART) {
 995                 pipe_clone_write_buffer(wpipe);
 996         } else {
 997                 pipe_destroy_write_buffer(wpipe);
 998         }
 999         pipeunlock(wpipe);
1000         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1001             ("pipe %p leaked PIPE_DIRECTW", wpipe));
1002         return (error);
1003
1004 error1:
1005         wakeup(wpipe);
1006         return (error);
1007 }
1008 #endif
1009
1010 static int
1011 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1012     int flags, struct thread *td)
1013 {
1014         int error = 0;
1015         int desiredsize;
1016         ssize_t orig_resid;
1017         struct pipe *wpipe, *rpipe;
1018
1019         rpipe = fp->f_data;
1020         wpipe = PIPE_PEER(rpipe);
1021         PIPE_LOCK(rpipe);
1022         error = pipelock(wpipe, 1);
1023         if (error) {
1024                 PIPE_UNLOCK(rpipe);
1025                 return (error);
1026         }
1027         /*
1028          * detect loss of pipe read side, issue SIGPIPE if lost.
1029          */
1030         if (wpipe->pipe_present != PIPE_ACTIVE ||
1031             (wpipe->pipe_state & PIPE_EOF)) {
1032                 pipeunlock(wpipe);
1033                 PIPE_UNLOCK(rpipe);
1034                 return (EPIPE);
1035         }
1036 #ifdef MAC
1037         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1038         if (error) {
1039                 pipeunlock(wpipe);
1040                 PIPE_UNLOCK(rpipe);
1041                 return (error);
1042         }
1043 #endif
1044         ++wpipe->pipe_busy;
1045
1046         /* Choose a larger size if it's advantageous */
1047         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1048         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1049                 if (piperesizeallowed != 1)
1050                         break;
1051                 if (amountpipekva > maxpipekva / 2)
1052                         break;
1053                 if (desiredsize == BIG_PIPE_SIZE)
1054                         break;
1055                 desiredsize = desiredsize * 2;
1056         }
1057
1058         /* Choose a smaller size if we're in a OOM situation */
1059         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1060                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1061                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1062                 (piperesizeallowed == 1))
1063                 desiredsize = SMALL_PIPE_SIZE;
1064
1065         /* Resize if the above determined that a new size was necessary */
1066         if ((desiredsize != wpipe->pipe_buffer.size) &&
1067                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1068                 PIPE_UNLOCK(wpipe);
1069                 pipespace(wpipe, desiredsize);
1070                 PIPE_LOCK(wpipe);
1071         }
1072         if (wpipe->pipe_buffer.size == 0) {
1073                 /*
1074                  * This can only happen for reverse direction use of pipes
1075                  * in a complete OOM situation.
1076                  */
1077                 error = ENOMEM;
1078                 --wpipe->pipe_busy;
1079                 pipeunlock(wpipe);
1080                 PIPE_UNLOCK(wpipe);
1081                 return (error);
1082         }
1083
1084         pipeunlock(wpipe);
1085
1086         orig_resid = uio->uio_resid;
1087
1088         while (uio->uio_resid) {
1089                 int space;
1090
1091                 pipelock(wpipe, 0);
1092                 if (wpipe->pipe_state & PIPE_EOF) {
1093                         pipeunlock(wpipe);
1094                         error = EPIPE;
1095                         break;
1096                 }
1097 #ifndef PIPE_NODIRECT
1098                 /*
1099                  * If the transfer is large, we can gain performance if
1100                  * we do process-to-process copies directly.
1101                  * If the write is non-blocking, we don't use the
1102                  * direct write mechanism.
1103                  *
1104                  * The direct write mechanism will detect the reader going
1105                  * away on us.
1106                  */
1107                 if (uio->uio_segflg == UIO_USERSPACE &&
1108                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1109                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1110                     (fp->f_flag & FNONBLOCK) == 0) {
1111                         pipeunlock(wpipe);
1112                         error = pipe_direct_write(wpipe, uio);
1113                         if (error)
1114                                 break;
1115                         continue;
1116                 }
1117 #endif
1118
1119                 /*
1120                  * Pipe buffered writes cannot be coincidental with
1121                  * direct writes.  We wait until the currently executing
1122                  * direct write is completed before we start filling the
1123                  * pipe buffer.  We break out if a signal occurs or the
1124                  * reader goes away.
1125                  */
1126                 if (wpipe->pipe_map.cnt != 0) {
1127                         if (wpipe->pipe_state & PIPE_WANTR) {
1128                                 wpipe->pipe_state &= ~PIPE_WANTR;
1129                                 wakeup(wpipe);
1130                         }
1131                         pipeselwakeup(wpipe);
1132                         wpipe->pipe_state |= PIPE_WANTW;
1133                         pipeunlock(wpipe);
1134                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1135                             "pipbww", 0);
1136                         if (error)
1137                                 break;
1138                         else
1139                                 continue;
1140                 }
1141
1142                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1143
1144                 /* Writes of size <= PIPE_BUF must be atomic. */
1145                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1146                         space = 0;
1147
1148                 if (space > 0) {
1149                         int size;       /* Transfer size */
1150                         int segsize;    /* first segment to transfer */
1151
1152                         /*
1153                          * Transfer size is minimum of uio transfer
1154                          * and free space in pipe buffer.
1155                          */
1156                         if (space > uio->uio_resid)
1157                                 size = uio->uio_resid;
1158                         else
1159                                 size = space;
1160                         /*
1161                          * First segment to transfer is minimum of
1162                          * transfer size and contiguous space in
1163                          * pipe buffer.  If first segment to transfer
1164                          * is less than the transfer size, we've got
1165                          * a wraparound in the buffer.
1166                          */
1167                         segsize = wpipe->pipe_buffer.size -
1168                                 wpipe->pipe_buffer.in;
1169                         if (segsize > size)
1170                                 segsize = size;
1171
1172                         /* Transfer first segment */
1173
1174                         PIPE_UNLOCK(rpipe);
1175                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1176                                         segsize, uio);
1177                         PIPE_LOCK(rpipe);
1178
1179                         if (error == 0 && segsize < size) {
1180                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1181                                         wpipe->pipe_buffer.size,
1182                                         ("Pipe buffer wraparound disappeared"));
1183                                 /*
1184                                  * Transfer remaining part now, to
1185                                  * support atomic writes.  Wraparound
1186                                  * happened.
1187                                  */
1188
1189                                 PIPE_UNLOCK(rpipe);
1190                                 error = uiomove(
1191                                     &wpipe->pipe_buffer.buffer[0],
1192                                     size - segsize, uio);
1193                                 PIPE_LOCK(rpipe);
1194                         }
1195                         if (error == 0) {
1196                                 wpipe->pipe_buffer.in += size;
1197                                 if (wpipe->pipe_buffer.in >=
1198                                     wpipe->pipe_buffer.size) {
1199                                         KASSERT(wpipe->pipe_buffer.in ==
1200                                                 size - segsize +
1201                                                 wpipe->pipe_buffer.size,
1202                                                 ("Expected wraparound bad"));
1203                                         wpipe->pipe_buffer.in = size - segsize;
1204                                 }
1205
1206                                 wpipe->pipe_buffer.cnt += size;
1207                                 KASSERT(wpipe->pipe_buffer.cnt <=
1208                                         wpipe->pipe_buffer.size,
1209                                         ("Pipe buffer overflow"));
1210                         }
1211                         pipeunlock(wpipe);
1212                         if (error != 0)
1213                                 break;
1214                 } else {
1215                         /*
1216                          * If the "read-side" has been blocked, wake it up now.
1217                          */
1218                         if (wpipe->pipe_state & PIPE_WANTR) {
1219                                 wpipe->pipe_state &= ~PIPE_WANTR;
1220                                 wakeup(wpipe);
1221                         }
1222
1223                         /*
1224                          * don't block on non-blocking I/O
1225                          */
1226                         if (fp->f_flag & FNONBLOCK) {
1227                                 error = EAGAIN;
1228                                 pipeunlock(wpipe);
1229                                 break;
1230                         }
1231
1232                         /*
1233                          * We have no more space and have something to offer,
1234                          * wake up select/poll.
1235                          */
1236                         pipeselwakeup(wpipe);
1237
1238                         wpipe->pipe_state |= PIPE_WANTW;
1239                         pipeunlock(wpipe);
1240                         error = msleep(wpipe, PIPE_MTX(rpipe),
1241                             PRIBIO | PCATCH, "pipewr", 0);
1242                         if (error != 0)
1243                                 break;
1244                 }
1245         }
1246
1247         pipelock(wpipe, 0);
1248         --wpipe->pipe_busy;
1249
1250         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1251                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1252                 wakeup(wpipe);
1253         } else if (wpipe->pipe_buffer.cnt > 0) {
1254                 /*
1255                  * If we have put any characters in the buffer, we wake up
1256                  * the reader.
1257                  */
1258                 if (wpipe->pipe_state & PIPE_WANTR) {
1259                         wpipe->pipe_state &= ~PIPE_WANTR;
1260                         wakeup(wpipe);
1261                 }
1262         }
1263
1264         /*
1265          * Don't return EPIPE if any byte was written.
1266          * EINTR and other interrupts are handled by generic I/O layer.
1267          * Do not pretend that I/O succeeded for obvious user error
1268          * like EFAULT.
1269          */
1270         if (uio->uio_resid != orig_resid && error == EPIPE)
1271                 error = 0;
1272
1273         if (error == 0)
1274                 vfs_timestamp(&wpipe->pipe_mtime);
1275
1276         /*
1277          * We have something to offer,
1278          * wake up select/poll.
1279          */
1280         if (wpipe->pipe_buffer.cnt)
1281                 pipeselwakeup(wpipe);
1282
1283         pipeunlock(wpipe);
1284         PIPE_UNLOCK(rpipe);
1285         return (error);
1286 }
1287
1288 /* ARGSUSED */
1289 static int
1290 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1291     struct thread *td)
1292 {
1293         struct pipe *cpipe;
1294         int error;
1295
1296         cpipe = fp->f_data;
1297         if (cpipe->pipe_state & PIPE_NAMED)
1298                 error = vnops.fo_truncate(fp, length, active_cred, td);
1299         else
1300                 error = invfo_truncate(fp, length, active_cred, td);
1301         return (error);
1302 }
1303
1304 /*
1305  * we implement a very minimal set of ioctls for compatibility with sockets.
1306  */
1307 static int
1308 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1309     struct thread *td)
1310 {
1311         struct pipe *mpipe = fp->f_data;
1312         int error;
1313
1314         PIPE_LOCK(mpipe);
1315
1316 #ifdef MAC
1317         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1318         if (error) {
1319                 PIPE_UNLOCK(mpipe);
1320                 return (error);
1321         }
1322 #endif
1323
1324         error = 0;
1325         switch (cmd) {
1326
1327         case FIONBIO:
1328                 break;
1329
1330         case FIOASYNC:
1331                 if (*(int *)data) {
1332                         mpipe->pipe_state |= PIPE_ASYNC;
1333                 } else {
1334                         mpipe->pipe_state &= ~PIPE_ASYNC;
1335                 }
1336                 break;
1337
1338         case FIONREAD:
1339                 if (!(fp->f_flag & FREAD)) {
1340                         *(int *)data = 0;
1341                         PIPE_UNLOCK(mpipe);
1342                         return (0);
1343                 }
1344                 if (mpipe->pipe_map.cnt != 0)
1345                         *(int *)data = mpipe->pipe_map.cnt;
1346                 else
1347                         *(int *)data = mpipe->pipe_buffer.cnt;
1348                 break;
1349
1350         case FIOSETOWN:
1351                 PIPE_UNLOCK(mpipe);
1352                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1353                 goto out_unlocked;
1354
1355         case FIOGETOWN:
1356                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1357                 break;
1358
1359         /* This is deprecated, FIOSETOWN should be used instead. */
1360         case TIOCSPGRP:
1361                 PIPE_UNLOCK(mpipe);
1362                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1363                 goto out_unlocked;
1364
1365         /* This is deprecated, FIOGETOWN should be used instead. */
1366         case TIOCGPGRP:
1367                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1368                 break;
1369
1370         default:
1371                 error = ENOTTY;
1372                 break;
1373         }
1374         PIPE_UNLOCK(mpipe);
1375 out_unlocked:
1376         return (error);
1377 }
1378
1379 static int
1380 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1381     struct thread *td)
1382 {
1383         struct pipe *rpipe;
1384         struct pipe *wpipe;
1385         int levents, revents;
1386 #ifdef MAC
1387         int error;
1388 #endif
1389
1390         revents = 0;
1391         rpipe = fp->f_data;
1392         wpipe = PIPE_PEER(rpipe);
1393         PIPE_LOCK(rpipe);
1394 #ifdef MAC
1395         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1396         if (error)
1397                 goto locked_error;
1398 #endif
1399         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1400                 if (rpipe->pipe_map.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1401                         revents |= events & (POLLIN | POLLRDNORM);
1402
1403         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1404                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1405                     (wpipe->pipe_state & PIPE_EOF) ||
1406                     ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1407                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1408                          wpipe->pipe_buffer.size == 0)))
1409                         revents |= events & (POLLOUT | POLLWRNORM);
1410
1411         levents = events &
1412             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1413         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1414             fp->f_pipegen == rpipe->pipe_wgen)
1415                 events |= POLLINIGNEOF;
1416
1417         if ((events & POLLINIGNEOF) == 0) {
1418                 if (rpipe->pipe_state & PIPE_EOF) {
1419                         revents |= (events & (POLLIN | POLLRDNORM));
1420                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1421                             (wpipe->pipe_state & PIPE_EOF))
1422                                 revents |= POLLHUP;
1423                 }
1424         }
1425
1426         if (revents == 0) {
1427                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1428                         selrecord(td, &rpipe->pipe_sel);
1429                         if (SEL_WAITING(&rpipe->pipe_sel))
1430                                 rpipe->pipe_state |= PIPE_SEL;
1431                 }
1432
1433                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1434                         selrecord(td, &wpipe->pipe_sel);
1435                         if (SEL_WAITING(&wpipe->pipe_sel))
1436                                 wpipe->pipe_state |= PIPE_SEL;
1437                 }
1438         }
1439 #ifdef MAC
1440 locked_error:
1441 #endif
1442         PIPE_UNLOCK(rpipe);
1443
1444         return (revents);
1445 }
1446
1447 /*
1448  * We shouldn't need locks here as we're doing a read and this should
1449  * be a natural race.
1450  */
1451 static int
1452 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1453     struct thread *td)
1454 {
1455         struct pipe *pipe;
1456 #ifdef MAC
1457         int error;
1458 #endif
1459
1460         pipe = fp->f_data;
1461         PIPE_LOCK(pipe);
1462 #ifdef MAC
1463         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1464         if (error) {
1465                 PIPE_UNLOCK(pipe);
1466                 return (error);
1467         }
1468 #endif
1469
1470         /* For named pipes ask the underlying filesystem. */
1471         if (pipe->pipe_state & PIPE_NAMED) {
1472                 PIPE_UNLOCK(pipe);
1473                 return (vnops.fo_stat(fp, ub, active_cred, td));
1474         }
1475
1476         PIPE_UNLOCK(pipe);
1477
1478         bzero(ub, sizeof(*ub));
1479         ub->st_mode = S_IFIFO;
1480         ub->st_blksize = PAGE_SIZE;
1481         if (pipe->pipe_map.cnt != 0)
1482                 ub->st_size = pipe->pipe_map.cnt;
1483         else
1484                 ub->st_size = pipe->pipe_buffer.cnt;
1485         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1486         ub->st_atim = pipe->pipe_atime;
1487         ub->st_mtim = pipe->pipe_mtime;
1488         ub->st_ctim = pipe->pipe_ctime;
1489         ub->st_uid = fp->f_cred->cr_uid;
1490         ub->st_gid = fp->f_cred->cr_gid;
1491         ub->st_dev = pipedev_ino;
1492         ub->st_ino = pipe->pipe_ino;
1493         /*
1494          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1495          */
1496         return (0);
1497 }
1498
1499 /* ARGSUSED */
1500 static int
1501 pipe_close(struct file *fp, struct thread *td)
1502 {
1503
1504         if (fp->f_vnode != NULL)
1505                 return vnops.fo_close(fp, td);
1506         fp->f_ops = &badfileops;
1507         pipe_dtor(fp->f_data);
1508         fp->f_data = NULL;
1509         return (0);
1510 }
1511
1512 static int
1513 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1514 {
1515         struct pipe *cpipe;
1516         int error;
1517
1518         cpipe = fp->f_data;
1519         if (cpipe->pipe_state & PIPE_NAMED)
1520                 error = vn_chmod(fp, mode, active_cred, td);
1521         else
1522                 error = invfo_chmod(fp, mode, active_cred, td);
1523         return (error);
1524 }
1525
1526 static int
1527 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1528     struct thread *td)
1529 {
1530         struct pipe *cpipe;
1531         int error;
1532
1533         cpipe = fp->f_data;
1534         if (cpipe->pipe_state & PIPE_NAMED)
1535                 error = vn_chown(fp, uid, gid, active_cred, td);
1536         else
1537                 error = invfo_chown(fp, uid, gid, active_cred, td);
1538         return (error);
1539 }
1540
1541 static int
1542 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1543 {
1544         struct pipe *pi;
1545
1546         if (fp->f_type == DTYPE_FIFO)
1547                 return (vn_fill_kinfo(fp, kif, fdp));
1548         kif->kf_type = KF_TYPE_PIPE;
1549         pi = fp->f_data;
1550         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1551         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1552         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1553         return (0);
1554 }
1555
1556 static void
1557 pipe_free_kmem(struct pipe *cpipe)
1558 {
1559
1560         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1561             ("pipe_free_kmem: pipe mutex locked"));
1562
1563         if (cpipe->pipe_buffer.buffer != NULL) {
1564                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1565                 vm_map_remove(pipe_map,
1566                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1567                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1568                 cpipe->pipe_buffer.buffer = NULL;
1569         }
1570 #ifndef PIPE_NODIRECT
1571         {
1572                 cpipe->pipe_map.cnt = 0;
1573                 cpipe->pipe_map.pos = 0;
1574                 cpipe->pipe_map.npages = 0;
1575         }
1576 #endif
1577 }
1578
1579 /*
1580  * shutdown the pipe
1581  */
1582 static void
1583 pipeclose(struct pipe *cpipe)
1584 {
1585         struct pipepair *pp;
1586         struct pipe *ppipe;
1587
1588         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1589
1590         PIPE_LOCK(cpipe);
1591         pipelock(cpipe, 0);
1592         pp = cpipe->pipe_pair;
1593
1594         pipeselwakeup(cpipe);
1595
1596         /*
1597          * If the other side is blocked, wake it up saying that
1598          * we want to close it down.
1599          */
1600         cpipe->pipe_state |= PIPE_EOF;
1601         while (cpipe->pipe_busy) {
1602                 wakeup(cpipe);
1603                 cpipe->pipe_state |= PIPE_WANT;
1604                 pipeunlock(cpipe);
1605                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1606                 pipelock(cpipe, 0);
1607         }
1608
1609
1610         /*
1611          * Disconnect from peer, if any.
1612          */
1613         ppipe = cpipe->pipe_peer;
1614         if (ppipe->pipe_present == PIPE_ACTIVE) {
1615                 pipeselwakeup(ppipe);
1616
1617                 ppipe->pipe_state |= PIPE_EOF;
1618                 wakeup(ppipe);
1619                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1620         }
1621
1622         /*
1623          * Mark this endpoint as free.  Release kmem resources.  We
1624          * don't mark this endpoint as unused until we've finished
1625          * doing that, or the pipe might disappear out from under
1626          * us.
1627          */
1628         PIPE_UNLOCK(cpipe);
1629         pipe_free_kmem(cpipe);
1630         PIPE_LOCK(cpipe);
1631         cpipe->pipe_present = PIPE_CLOSING;
1632         pipeunlock(cpipe);
1633
1634         /*
1635          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1636          * PIPE_FINALIZED, that allows other end to free the
1637          * pipe_pair, only after the knotes are completely dismantled.
1638          */
1639         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1640         cpipe->pipe_present = PIPE_FINALIZED;
1641         seldrain(&cpipe->pipe_sel);
1642         knlist_destroy(&cpipe->pipe_sel.si_note);
1643
1644         /*
1645          * If both endpoints are now closed, release the memory for the
1646          * pipe pair.  If not, unlock.
1647          */
1648         if (ppipe->pipe_present == PIPE_FINALIZED) {
1649                 PIPE_UNLOCK(cpipe);
1650 #ifdef MAC
1651                 mac_pipe_destroy(pp);
1652 #endif
1653                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1654         } else
1655                 PIPE_UNLOCK(cpipe);
1656 }
1657
1658 /*ARGSUSED*/
1659 static int
1660 pipe_kqfilter(struct file *fp, struct knote *kn)
1661 {
1662         struct pipe *cpipe;
1663
1664         /*
1665          * If a filter is requested that is not supported by this file
1666          * descriptor, don't return an error, but also don't ever generate an
1667          * event.
1668          */
1669         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1670                 kn->kn_fop = &pipe_nfiltops;
1671                 return (0);
1672         }
1673         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1674                 kn->kn_fop = &pipe_nfiltops;
1675                 return (0);
1676         }
1677         cpipe = fp->f_data;
1678         PIPE_LOCK(cpipe);
1679         switch (kn->kn_filter) {
1680         case EVFILT_READ:
1681                 kn->kn_fop = &pipe_rfiltops;
1682                 break;
1683         case EVFILT_WRITE:
1684                 kn->kn_fop = &pipe_wfiltops;
1685                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1686                         /* other end of pipe has been closed */
1687                         PIPE_UNLOCK(cpipe);
1688                         return (EPIPE);
1689                 }
1690                 cpipe = PIPE_PEER(cpipe);
1691                 break;
1692         default:
1693                 PIPE_UNLOCK(cpipe);
1694                 return (EINVAL);
1695         }
1696
1697         kn->kn_hook = cpipe;
1698         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1699         PIPE_UNLOCK(cpipe);
1700         return (0);
1701 }
1702
1703 static void
1704 filt_pipedetach(struct knote *kn)
1705 {
1706         struct pipe *cpipe = kn->kn_hook;
1707
1708         PIPE_LOCK(cpipe);
1709         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1710         PIPE_UNLOCK(cpipe);
1711 }
1712
1713 /*ARGSUSED*/
1714 static int
1715 filt_piperead(struct knote *kn, long hint)
1716 {
1717         struct pipe *rpipe = kn->kn_hook;
1718         struct pipe *wpipe = rpipe->pipe_peer;
1719         int ret;
1720
1721         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1722         kn->kn_data = rpipe->pipe_buffer.cnt;
1723         if (kn->kn_data == 0)
1724                 kn->kn_data = rpipe->pipe_map.cnt;
1725
1726         if ((rpipe->pipe_state & PIPE_EOF) ||
1727             wpipe->pipe_present != PIPE_ACTIVE ||
1728             (wpipe->pipe_state & PIPE_EOF)) {
1729                 kn->kn_flags |= EV_EOF;
1730                 return (1);
1731         }
1732         ret = kn->kn_data > 0;
1733         return ret;
1734 }
1735
1736 /*ARGSUSED*/
1737 static int
1738 filt_pipewrite(struct knote *kn, long hint)
1739 {
1740         struct pipe *wpipe;
1741
1742         /*
1743          * If this end of the pipe is closed, the knote was removed from the
1744          * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1745          */
1746         wpipe = kn->kn_hook;
1747         if (wpipe->pipe_present != PIPE_ACTIVE ||
1748             (wpipe->pipe_state & PIPE_EOF)) {
1749                 kn->kn_data = 0;
1750                 kn->kn_flags |= EV_EOF;
1751                 return (1);
1752         }
1753         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1754         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1755             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1756         if (wpipe->pipe_state & PIPE_DIRECTW)
1757                 kn->kn_data = 0;
1758
1759         return (kn->kn_data >= PIPE_BUF);
1760 }
1761
1762 static void
1763 filt_pipedetach_notsup(struct knote *kn)
1764 {
1765
1766 }
1767
1768 static int
1769 filt_pipenotsup(struct knote *kn, long hint)
1770 {
1771
1772         return (0);
1773 }