sys/kern/sys_pipe.c

   1 /*-
   2  * Copyright (c) 1996 John S. Dyson
   3  * Copyright (c) 2012 Giovanni Trematerra
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice immediately at the beginning of the file, without modification,
  11  *    this list of conditions, and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Absolutely no warranty of function or purpose is made by the author
  16  *    John S. Dyson.
  17  * 4. Modifications may be freely made to this file if the above conditions
  18  *    are met.
  19  */
  20
  21 /*
  22  * This file contains a high-performance replacement for the socket-based
  23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  24  * all features of sockets, but does do everything that pipes normally
  25  * do.
  26  */
  27
  28 /*
  29  * This code has two modes of operation, a small write mode and a large
  30  * write mode.  The small write mode acts like conventional pipes with
  31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  34  * memory, and the receiving process copies directly from these pinned pages
  35  * in the sending process.
  36  *
  37  * If the sending process receives a signal, it is possible that it will
  38  * go away, and certainly its address space can change, because control
  39  * is returned back to the user-mode side.  In that case, the pipe code
  40  * arranges to copy the buffer supplied by the user process, to a pageable
  41  * kernel buffer, and the receiving process will grab the data from the
  42  * pageable kernel buffer.  Since signals don't happen all that often,
  43  * the copy operation is normally eliminated.
  44  *
  45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  46  * happen for small transfers so that the system will not spend all of
  47  * its time context switching.
  48  *
  49  * In order to limit the resource use of pipes, two sysctls exist:
  50  *
  51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  52  * address space available to us in pipe_map. This value is normally
  53  * autotuned, but may also be loader tuned.
  54  *
  55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  56  * memory in use by pipes.
  57  *
  58  * Based on how large pipekva is relative to maxpipekva, the following
  59  * will happen:
  60  *
  61  * 0% - 50%:
  62  *     New pipes are given 16K of memory backing, pipes may dynamically
  63  *     grow to as large as 64K where needed.
  64  * 50% - 75%:
  65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  66  *     existing pipes may NOT grow.
  67  * 75% - 100%:
  68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  69  *     existing pipes will be shrunk down to 4K whenever possible.
  70  *
  71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  73  * resize which MUST occur for reverse-direction pipes when they are
  74  * first used.
  75  *
  76  * Additional information about the current state of pipes may be obtained
  77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  78  * and kern.ipc.piperesizefail.
  79  *
  80  * Locking rules:  There are two locks present here:  A mutex, used via
  81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  82  * the flag, as mutexes can not persist over uiomove.  The mutex
  83  * exists only to guard access to the flag, and is not in itself a
  84  * locking mechanism.  Also note that there is only a single mutex for
  85  * both directions of a pipe.
  86  *
  87  * As pipelock() may have to sleep before it can acquire the flag, it
  88  * is important to reread all data after a call to pipelock(); everything
  89  * in the structure may have changed.
  90  */
  91
  92 #include <sys/cdefs.h>
  93 __FBSDID("$FreeBSD$");
  94
  95 #include <sys/param.h>
  96 #include <sys/systm.h>
  97 #include <sys/conf.h>
  98 #include <sys/fcntl.h>
  99 #include <sys/file.h>
 100 #include <sys/filedesc.h>
 101 #include <sys/filio.h>
 102 #include <sys/kernel.h>
 103 #include <sys/lock.h>
 104 #include <sys/mutex.h>
 105 #include <sys/ttycom.h>
 106 #include <sys/stat.h>
 107 #include <sys/malloc.h>
 108 #include <sys/poll.h>
 109 #include <sys/selinfo.h>
 110 #include <sys/signalvar.h>
 111 #include <sys/syscallsubr.h>
 112 #include <sys/sysctl.h>
 113 #include <sys/sysproto.h>
 114 #include <sys/pipe.h>
 115 #include <sys/proc.h>
 116 #include <sys/vnode.h>
 117 #include <sys/uio.h>
 118 #include <sys/user.h>
 119 #include <sys/event.h>
 120
 121 #include <security/mac/mac_framework.h>
 122
 123 #include <vm/vm.h>
 124 #include <vm/vm_param.h>
 125 #include <vm/vm_object.h>
 126 #include <vm/vm_kern.h>
 127 #include <vm/vm_extern.h>
 128 #include <vm/pmap.h>
 129 #include <vm/vm_map.h>
 130 #include <vm/vm_page.h>
 131 #include <vm/uma.h>
 132
 133 /*
 134  * Use this define if you want to disable *fancy* VM things.  Expect an
 135  * approx 30% decrease in transfer rate.  This could be useful for
 136  * NetBSD or OpenBSD.
 137  */
 138 /* #define PIPE_NODIRECT */
 139
 140 #define PIPE_PEER(pipe) \
 141         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 142
 143 /*
 144  * interfaces to the outside world
 145  */
 146 static fo_rdwr_t        pipe_read;
 147 static fo_rdwr_t        pipe_write;
 148 static fo_truncate_t    pipe_truncate;
 149 static fo_ioctl_t       pipe_ioctl;
 150 static fo_poll_t        pipe_poll;
 151 static fo_kqfilter_t    pipe_kqfilter;
 152 static fo_stat_t        pipe_stat;
 153 static fo_close_t       pipe_close;
 154 static fo_chmod_t       pipe_chmod;
 155 static fo_chown_t       pipe_chown;
 156 static fo_fill_kinfo_t  pipe_fill_kinfo;
 157
 158 struct fileops pipeops = {
 159         .fo_read = pipe_read,
 160         .fo_write = pipe_write,
 161         .fo_truncate = pipe_truncate,
 162         .fo_ioctl = pipe_ioctl,
 163         .fo_poll = pipe_poll,
 164         .fo_kqfilter = pipe_kqfilter,
 165         .fo_stat = pipe_stat,
 166         .fo_close = pipe_close,
 167         .fo_chmod = pipe_chmod,
 168         .fo_chown = pipe_chown,
 169         .fo_sendfile = invfo_sendfile,
 170         .fo_fill_kinfo = pipe_fill_kinfo,
 171         .fo_flags = DFLAG_PASSABLE
 172 };
 173
 174 static void     filt_pipedetach(struct knote *kn);
 175 static void     filt_pipedetach_notsup(struct knote *kn);
 176 static int      filt_pipenotsup(struct knote *kn, long hint);
 177 static int      filt_piperead(struct knote *kn, long hint);
 178 static int      filt_pipewrite(struct knote *kn, long hint);
 179
 180 static struct filterops pipe_nfiltops = {
 181         .f_isfd = 1,
 182         .f_detach = filt_pipedetach_notsup,
 183         .f_event = filt_pipenotsup
 184 };
 185 static struct filterops pipe_rfiltops = {
 186         .f_isfd = 1,
 187         .f_detach = filt_pipedetach,
 188         .f_event = filt_piperead
 189 };
 190 static struct filterops pipe_wfiltops = {
 191         .f_isfd = 1,
 192         .f_detach = filt_pipedetach,
 193         .f_event = filt_pipewrite
 194 };
 195
 196 /*
 197  * Default pipe buffer size(s), this can be kind-of large now because pipe
 198  * space is pageable.  The pipe code will try to maintain locality of
 199  * reference for performance reasons, so small amounts of outstanding I/O
 200  * will not wipe the cache.
 201  */
 202 #define MINPIPESIZE (PIPE_SIZE/3)
 203 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 204
 205 static long amountpipekva;
 206 static int pipefragretry;
 207 static int pipeallocfail;
 208 static int piperesizefail;
 209 static int piperesizeallowed = 1;
 210
 211 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 212            &maxpipekva, 0, "Pipe KVA limit");
 213 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 214            &amountpipekva, 0, "Pipe KVA usage");
 215 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 216           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 218           &pipeallocfail, 0, "Pipe allocation failures");
 219 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 220           &piperesizefail, 0, "Pipe resize failures");
 221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 222           &piperesizeallowed, 0, "Pipe resizing allowed");
 223
 224 static void pipeinit(void *dummy __unused);
 225 static void pipeclose(struct pipe *cpipe);
 226 static void pipe_free_kmem(struct pipe *cpipe);
 227 static void pipe_create(struct pipe *pipe, int backing);
 228 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 229 static __inline int pipelock(struct pipe *cpipe, int catch);
 230 static __inline void pipeunlock(struct pipe *cpipe);
 231 #ifndef PIPE_NODIRECT
 232 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 233 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 234 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 235 static void pipe_clone_write_buffer(struct pipe *wpipe);
 236 #endif
 237 static int pipespace(struct pipe *cpipe, int size);
 238 static int pipespace_new(struct pipe *cpipe, int size);
 239
 240 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 241 static int      pipe_zone_init(void *mem, int size, int flags);
 242 static void     pipe_zone_fini(void *mem, int size);
 243
 244 static uma_zone_t pipe_zone;
 245 static struct unrhdr *pipeino_unr;
 246 static dev_t pipedev_ino;
 247
 248 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 249
 250 static void
 251 pipeinit(void *dummy __unused)
 252 {
 253
 254         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 255             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 256             UMA_ALIGN_PTR, 0);
 257         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 258         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 259         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 260         pipedev_ino = devfs_alloc_cdp_inode();
 261         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 262 }
 263
 264 static int
 265 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 266 {
 267         struct pipepair *pp;
 268         struct pipe *rpipe, *wpipe;
 269
 270         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 271
 272         pp = (struct pipepair *)mem;
 273
 274         /*
 275          * We zero both pipe endpoints to make sure all the kmem pointers
 276          * are NULL, flag fields are zero'd, etc.  We timestamp both
 277          * endpoints with the same time.
 278          */
 279         rpipe = &pp->pp_rpipe;
 280         bzero(rpipe, sizeof(*rpipe));
 281         vfs_timestamp(&rpipe->pipe_ctime);
 282         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 283
 284         wpipe = &pp->pp_wpipe;
 285         bzero(wpipe, sizeof(*wpipe));
 286         wpipe->pipe_ctime = rpipe->pipe_ctime;
 287         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 288
 289         rpipe->pipe_peer = wpipe;
 290         rpipe->pipe_pair = pp;
 291         wpipe->pipe_peer = rpipe;
 292         wpipe->pipe_pair = pp;
 293
 294         /*
 295          * Mark both endpoints as present; they will later get free'd
 296          * one at a time.  When both are free'd, then the whole pair
 297          * is released.
 298          */
 299         rpipe->pipe_present = PIPE_ACTIVE;
 300         wpipe->pipe_present = PIPE_ACTIVE;
 301
 302         /*
 303          * Eventually, the MAC Framework may initialize the label
 304          * in ctor or init, but for now we do it elswhere to avoid
 305          * blocking in ctor or init.
 306          */
 307         pp->pp_label = NULL;
 308
 309         return (0);
 310 }
 311
 312 static int
 313 pipe_zone_init(void *mem, int size, int flags)
 314 {
 315         struct pipepair *pp;
 316
 317         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 318
 319         pp = (struct pipepair *)mem;
 320
 321         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 322         return (0);
 323 }
 324
 325 static void
 326 pipe_zone_fini(void *mem, int size)
 327 {
 328         struct pipepair *pp;
 329
 330         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 331
 332         pp = (struct pipepair *)mem;
 333
 334         mtx_destroy(&pp->pp_mtx);
 335 }
 336
 337 static void
 338 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 339 {
 340         struct pipepair *pp;
 341         struct pipe *rpipe, *wpipe;
 342
 343         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 344 #ifdef MAC
 345         /*
 346          * The MAC label is shared between the connected endpoints.  As a
 347          * result mac_pipe_init() and mac_pipe_create() are called once
 348          * for the pair, and not on the endpoints.
 349          */
 350         mac_pipe_init(pp);
 351         mac_pipe_create(td->td_ucred, pp);
 352 #endif
 353         rpipe = &pp->pp_rpipe;
 354         wpipe = &pp->pp_wpipe;
 355
 356         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 357         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 358
 359         /* Only the forward direction pipe is backed by default */
 360         pipe_create(rpipe, 1);
 361         pipe_create(wpipe, 0);
 362
 363         rpipe->pipe_state |= PIPE_DIRECTOK;
 364         wpipe->pipe_state |= PIPE_DIRECTOK;
 365 }
 366
 367 void
 368 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 369 {
 370         struct pipepair *pp;
 371
 372         pipe_paircreate(td, &pp);
 373         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 374         *ppipe = &pp->pp_rpipe;
 375 }
 376
 377 void
 378 pipe_dtor(struct pipe *dpipe)
 379 {
 380         struct pipe *peer;
 381         ino_t ino;
 382
 383         ino = dpipe->pipe_ino;
 384         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 385         funsetown(&dpipe->pipe_sigio);
 386         pipeclose(dpipe);
 387         if (peer != NULL) {
 388                 funsetown(&peer->pipe_sigio);
 389                 pipeclose(peer);
 390         }
 391         if (ino != 0 && ino != (ino_t)-1)
 392                 free_unr(pipeino_unr, ino);
 393 }
 394
 395 /*
 396  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 397  * the zone pick up the pieces via pipeclose().
 398  */
 399 int
 400 kern_pipe(struct thread *td, int fildes[2])
 401 {
 402
 403         return (kern_pipe2(td, fildes, 0));
 404 }
 405
 406 int
 407 kern_pipe2(struct thread *td, int fildes[2], int flags)
 408 {
 409         struct file *rf, *wf;
 410         struct pipe *rpipe, *wpipe;
 411         struct pipepair *pp;
 412         int fd, fflags, error;
 413
 414         pipe_paircreate(td, &pp);
 415         rpipe = &pp->pp_rpipe;
 416         wpipe = &pp->pp_wpipe;
 417         error = falloc(td, &rf, &fd, flags);
 418         if (error) {
 419                 pipeclose(rpipe);
 420                 pipeclose(wpipe);
 421                 return (error);
 422         }
 423         /* An extra reference on `rf' has been held for us by falloc(). */
 424         fildes[0] = fd;
 425
 426         fflags = FREAD | FWRITE;
 427         if ((flags & O_NONBLOCK) != 0)
 428                 fflags |= FNONBLOCK;
 429
 430         /*
 431          * Warning: once we've gotten past allocation of the fd for the
 432          * read-side, we can only drop the read side via fdrop() in order
 433          * to avoid races against processes which manage to dup() the read
 434          * side while we are blocked trying to allocate the write side.
 435          */
 436         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 437         error = falloc(td, &wf, &fd, flags);
 438         if (error) {
 439                 fdclose(td, rf, fildes[0]);
 440                 fdrop(rf, td);
 441                 /* rpipe has been closed by fdrop(). */
 442                 pipeclose(wpipe);
 443                 return (error);
 444         }
 445         /* An extra reference on `wf' has been held for us by falloc(). */
 446         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 447         fdrop(wf, td);
 448         fildes[1] = fd;
 449         fdrop(rf, td);
 450
 451         return (0);
 452 }
 453
 454 /* ARGSUSED */
 455 int
 456 sys_pipe(struct thread *td, struct pipe_args *uap)
 457 {
 458         int error;
 459         int fildes[2];
 460
 461         error = kern_pipe(td, fildes);
 462         if (error)
 463                 return (error);
 464
 465         td->td_retval[0] = fildes[0];
 466         td->td_retval[1] = fildes[1];
 467
 468         return (0);
 469 }
 470
 471 int
 472 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 473 {
 474         int error, fildes[2];
 475
 476         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 477                 return (EINVAL);
 478         error = kern_pipe2(td, fildes, uap->flags);
 479         if (error)
 480                 return (error);
 481         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 482         if (error) {
 483                 (void)kern_close(td, fildes[0]);
 484                 (void)kern_close(td, fildes[1]);
 485         }
 486         return (error);
 487 }
 488
 489 /*
 490  * Allocate kva for pipe circular buffer, the space is pageable
 491  * This routine will 'realloc' the size of a pipe safely, if it fails
 492  * it will retain the old buffer.
 493  * If it fails it will return ENOMEM.
 494  */
 495 static int
 496 pipespace_new(cpipe, size)
 497         struct pipe *cpipe;
 498         int size;
 499 {
 500         caddr_t buffer;
 501         int error, cnt, firstseg;
 502         static int curfail = 0;
 503         static struct timeval lastfail;
 504
 505         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 506         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 507                 ("pipespace: resize of direct writes not allowed"));
 508 retry:
 509         cnt = cpipe->pipe_buffer.cnt;
 510         if (cnt > size)
 511                 size = cnt;
 512
 513         size = round_page(size);
 514         buffer = (caddr_t) vm_map_min(pipe_map);
 515
 516         error = vm_map_find(pipe_map, NULL, 0,
 517                 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
 518                 VM_PROT_ALL, VM_PROT_ALL, 0);
 519         if (error != KERN_SUCCESS) {
 520                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 521                         (size > SMALL_PIPE_SIZE)) {
 522                         size = SMALL_PIPE_SIZE;
 523                         pipefragretry++;
 524                         goto retry;
 525                 }
 526                 if (cpipe->pipe_buffer.buffer == NULL) {
 527                         pipeallocfail++;
 528                         if (ppsratecheck(&lastfail, &curfail, 1))
 529                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 530                 } else {
 531                         piperesizefail++;
 532                 }
 533                 return (ENOMEM);
 534         }
 535
 536         /* copy data, then free old resources if we're resizing */
 537         if (cnt > 0) {
 538                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 539                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 540                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 541                                 buffer, firstseg);
 542                         if ((cnt - firstseg) > 0)
 543                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 544                                         cpipe->pipe_buffer.in);
 545                 } else {
 546                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 547                                 buffer, cnt);
 548                 }
 549         }
 550         pipe_free_kmem(cpipe);
 551         cpipe->pipe_buffer.buffer = buffer;
 552         cpipe->pipe_buffer.size = size;
 553         cpipe->pipe_buffer.in = cnt;
 554         cpipe->pipe_buffer.out = 0;
 555         cpipe->pipe_buffer.cnt = cnt;
 556         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 557         return (0);
 558 }
 559
 560 /*
 561  * Wrapper for pipespace_new() that performs locking assertions.
 562  */
 563 static int
 564 pipespace(cpipe, size)
 565         struct pipe *cpipe;
 566         int size;
 567 {
 568
 569         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 570                 ("Unlocked pipe passed to pipespace"));
 571         return (pipespace_new(cpipe, size));
 572 }
 573
 574 /*
 575  * lock a pipe for I/O, blocking other access
 576  */
 577 static __inline int
 578 pipelock(cpipe, catch)
 579         struct pipe *cpipe;
 580         int catch;
 581 {
 582         int error;
 583
 584         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 585         while (cpipe->pipe_state & PIPE_LOCKFL) {
 586                 cpipe->pipe_state |= PIPE_LWANT;
 587                 error = msleep(cpipe, PIPE_MTX(cpipe),
 588                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 589                     "pipelk", 0);
 590                 if (error != 0)
 591                         return (error);
 592         }
 593         cpipe->pipe_state |= PIPE_LOCKFL;
 594         return (0);
 595 }
 596
 597 /*
 598  * unlock a pipe I/O lock
 599  */
 600 static __inline void
 601 pipeunlock(cpipe)
 602         struct pipe *cpipe;
 603 {
 604
 605         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 606         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 607                 ("Unlocked pipe passed to pipeunlock"));
 608         cpipe->pipe_state &= ~PIPE_LOCKFL;
 609         if (cpipe->pipe_state & PIPE_LWANT) {
 610                 cpipe->pipe_state &= ~PIPE_LWANT;
 611                 wakeup(cpipe);
 612         }
 613 }
 614
 615 void
 616 pipeselwakeup(cpipe)
 617         struct pipe *cpipe;
 618 {
 619
 620         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 621         if (cpipe->pipe_state & PIPE_SEL) {
 622                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 623                 if (!SEL_WAITING(&cpipe->pipe_sel))
 624                         cpipe->pipe_state &= ~PIPE_SEL;
 625         }
 626         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 627                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 628         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 629 }
 630
 631 /*
 632  * Initialize and allocate VM and memory for pipe.  The structure
 633  * will start out zero'd from the ctor, so we just manage the kmem.
 634  */
 635 static void
 636 pipe_create(pipe, backing)
 637         struct pipe *pipe;
 638         int backing;
 639 {
 640
 641         if (backing) {
 642                 /*
 643                  * Note that these functions can fail if pipe map is exhausted
 644                  * (as a result of too many pipes created), but we ignore the
 645                  * error as it is not fatal and could be provoked by
 646                  * unprivileged users. The only consequence is worse performance
 647                  * with given pipe.
 648                  */
 649                 if (amountpipekva > maxpipekva / 2)
 650                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 651                 else
 652                         (void)pipespace_new(pipe, PIPE_SIZE);
 653         }
 654
 655         pipe->pipe_ino = -1;
 656 }
 657
 658 /* ARGSUSED */
 659 static int
 660 pipe_read(fp, uio, active_cred, flags, td)
 661         struct file *fp;
 662         struct uio *uio;
 663         struct ucred *active_cred;
 664         struct thread *td;
 665         int flags;
 666 {
 667         struct pipe *rpipe;
 668         int error;
 669         int nread = 0;
 670         int size;
 671
 672         rpipe = fp->f_data;
 673         PIPE_LOCK(rpipe);
 674         ++rpipe->pipe_busy;
 675         error = pipelock(rpipe, 1);
 676         if (error)
 677                 goto unlocked_error;
 678
 679 #ifdef MAC
 680         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 681         if (error)
 682                 goto locked_error;
 683 #endif
 684         if (amountpipekva > (3 * maxpipekva) / 4) {
 685                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 686                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 687                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 688                         (piperesizeallowed == 1)) {
 689                         PIPE_UNLOCK(rpipe);
 690                         pipespace(rpipe, SMALL_PIPE_SIZE);
 691                         PIPE_LOCK(rpipe);
 692                 }
 693         }
 694
 695         while (uio->uio_resid) {
 696                 /*
 697                  * normal pipe buffer receive
 698                  */
 699                 if (rpipe->pipe_buffer.cnt > 0) {
 700                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 701                         if (size > rpipe->pipe_buffer.cnt)
 702                                 size = rpipe->pipe_buffer.cnt;
 703                         if (size > uio->uio_resid)
 704                                 size = uio->uio_resid;
 705
 706                         PIPE_UNLOCK(rpipe);
 707                         error = uiomove(
 708                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 709                             size, uio);
 710                         PIPE_LOCK(rpipe);
 711                         if (error)
 712                                 break;
 713
 714                         rpipe->pipe_buffer.out += size;
 715                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 716                                 rpipe->pipe_buffer.out = 0;
 717
 718                         rpipe->pipe_buffer.cnt -= size;
 719
 720                         /*
 721                          * If there is no more to read in the pipe, reset
 722                          * its pointers to the beginning.  This improves
 723                          * cache hit stats.
 724                          */
 725                         if (rpipe->pipe_buffer.cnt == 0) {
 726                                 rpipe->pipe_buffer.in = 0;
 727                                 rpipe->pipe_buffer.out = 0;
 728                         }
 729                         nread += size;
 730 #ifndef PIPE_NODIRECT
 731                 /*
 732                  * Direct copy, bypassing a kernel buffer.
 733                  */
 734                 } else if ((size = rpipe->pipe_map.cnt) &&
 735                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 736                         if (size > uio->uio_resid)
 737                                 size = (u_int) uio->uio_resid;
 738
 739                         PIPE_UNLOCK(rpipe);
 740                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 741                             rpipe->pipe_map.pos, size, uio);
 742                         PIPE_LOCK(rpipe);
 743                         if (error)
 744                                 break;
 745                         nread += size;
 746                         rpipe->pipe_map.pos += size;
 747                         rpipe->pipe_map.cnt -= size;
 748                         if (rpipe->pipe_map.cnt == 0) {
 749                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 750                                 wakeup(rpipe);
 751                         }
 752 #endif
 753                 } else {
 754                         /*
 755                          * detect EOF condition
 756                          * read returns 0 on EOF, no need to set error
 757                          */
 758                         if (rpipe->pipe_state & PIPE_EOF)
 759                                 break;
 760
 761                         /*
 762                          * If the "write-side" has been blocked, wake it up now.
 763                          */
 764                         if (rpipe->pipe_state & PIPE_WANTW) {
 765                                 rpipe->pipe_state &= ~PIPE_WANTW;
 766                                 wakeup(rpipe);
 767                         }
 768
 769                         /*
 770                          * Break if some data was read.
 771                          */
 772                         if (nread > 0)
 773                                 break;
 774
 775                         /*
 776                          * Unlock the pipe buffer for our remaining processing.
 777                          * We will either break out with an error or we will
 778                          * sleep and relock to loop.
 779                          */
 780                         pipeunlock(rpipe);
 781
 782                         /*
 783                          * Handle non-blocking mode operation or
 784                          * wait for more data.
 785                          */
 786                         if (fp->f_flag & FNONBLOCK) {
 787                                 error = EAGAIN;
 788                         } else {
 789                                 rpipe->pipe_state |= PIPE_WANTR;
 790                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 791                                     PRIBIO | PCATCH,
 792                                     "piperd", 0)) == 0)
 793                                         error = pipelock(rpipe, 1);
 794                         }
 795                         if (error)
 796                                 goto unlocked_error;
 797                 }
 798         }
 799 #ifdef MAC
 800 locked_error:
 801 #endif
 802         pipeunlock(rpipe);
 803
 804         /* XXX: should probably do this before getting any locks. */
 805         if (error == 0)
 806                 vfs_timestamp(&rpipe->pipe_atime);
 807 unlocked_error:
 808         --rpipe->pipe_busy;
 809
 810         /*
 811          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 812          */
 813         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 814                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 815                 wakeup(rpipe);
 816         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 817                 /*
 818                  * Handle write blocking hysteresis.
 819                  */
 820                 if (rpipe->pipe_state & PIPE_WANTW) {
 821                         rpipe->pipe_state &= ~PIPE_WANTW;
 822                         wakeup(rpipe);
 823                 }
 824         }
 825
 826         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 827                 pipeselwakeup(rpipe);
 828
 829         PIPE_UNLOCK(rpipe);
 830         return (error);
 831 }
 832
 833 #ifndef PIPE_NODIRECT
 834 /*
 835  * Map the sending processes' buffer into kernel space and wire it.
 836  * This is similar to a physical write operation.
 837  */
 838 static int
 839 pipe_build_write_buffer(wpipe, uio)
 840         struct pipe *wpipe;
 841         struct uio *uio;
 842 {
 843         u_int size;
 844         int i;
 845
 846         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 847         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 848                 ("Clone attempt on non-direct write pipe!"));
 849
 850         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 851                 size = wpipe->pipe_buffer.size;
 852         else
 853                 size = uio->uio_iov->iov_len;
 854
 855         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 856             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 857             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 858                 return (EFAULT);
 859
 860 /*
 861  * set up the control block
 862  */
 863         wpipe->pipe_map.npages = i;
 864         wpipe->pipe_map.pos =
 865             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 866         wpipe->pipe_map.cnt = size;
 867
 868 /*
 869  * and update the uio data
 870  */
 871
 872         uio->uio_iov->iov_len -= size;
 873         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 874         if (uio->uio_iov->iov_len == 0)
 875                 uio->uio_iov++;
 876         uio->uio_resid -= size;
 877         uio->uio_offset += size;
 878         return (0);
 879 }
 880
 881 /*
 882  * unmap and unwire the process buffer
 883  */
 884 static void
 885 pipe_destroy_write_buffer(wpipe)
 886         struct pipe *wpipe;
 887 {
 888
 889         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 890         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 891         wpipe->pipe_map.npages = 0;
 892 }
 893
 894 /*
 895  * In the case of a signal, the writing process might go away.  This
 896  * code copies the data into the circular buffer so that the source
 897  * pages can be freed without loss of data.
 898  */
 899 static void
 900 pipe_clone_write_buffer(wpipe)
 901         struct pipe *wpipe;
 902 {
 903         struct uio uio;
 904         struct iovec iov;
 905         int size;
 906         int pos;
 907
 908         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 909         size = wpipe->pipe_map.cnt;
 910         pos = wpipe->pipe_map.pos;
 911
 912         wpipe->pipe_buffer.in = size;
 913         wpipe->pipe_buffer.out = 0;
 914         wpipe->pipe_buffer.cnt = size;
 915         wpipe->pipe_state &= ~PIPE_DIRECTW;
 916
 917         PIPE_UNLOCK(wpipe);
 918         iov.iov_base = wpipe->pipe_buffer.buffer;
 919         iov.iov_len = size;
 920         uio.uio_iov = &iov;
 921         uio.uio_iovcnt = 1;
 922         uio.uio_offset = 0;
 923         uio.uio_resid = size;
 924         uio.uio_segflg = UIO_SYSSPACE;
 925         uio.uio_rw = UIO_READ;
 926         uio.uio_td = curthread;
 927         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 928         PIPE_LOCK(wpipe);
 929         pipe_destroy_write_buffer(wpipe);
 930 }
 931
 932 /*
 933  * This implements the pipe buffer write mechanism.  Note that only
 934  * a direct write OR a normal pipe write can be pending at any given time.
 935  * If there are any characters in the pipe buffer, the direct write will
 936  * be deferred until the receiving process grabs all of the bytes from
 937  * the pipe buffer.  Then the direct mapping write is set-up.
 938  */
 939 static int
 940 pipe_direct_write(wpipe, uio)
 941         struct pipe *wpipe;
 942         struct uio *uio;
 943 {
 944         int error;
 945
 946 retry:
 947         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 948         error = pipelock(wpipe, 1);
 949         if (wpipe->pipe_state & PIPE_EOF)
 950                 error = EPIPE;
 951         if (error) {
 952                 pipeunlock(wpipe);
 953                 goto error1;
 954         }
 955         while (wpipe->pipe_state & PIPE_DIRECTW) {
 956                 if (wpipe->pipe_state & PIPE_WANTR) {
 957                         wpipe->pipe_state &= ~PIPE_WANTR;
 958                         wakeup(wpipe);
 959                 }
 960                 pipeselwakeup(wpipe);
 961                 wpipe->pipe_state |= PIPE_WANTW;
 962                 pipeunlock(wpipe);
 963                 error = msleep(wpipe, PIPE_MTX(wpipe),
 964                     PRIBIO | PCATCH, "pipdww", 0);
 965                 if (error)
 966                         goto error1;
 967                 else
 968                         goto retry;
 969         }
 970         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 971         if (wpipe->pipe_buffer.cnt > 0) {
 972                 if (wpipe->pipe_state & PIPE_WANTR) {
 973                         wpipe->pipe_state &= ~PIPE_WANTR;
 974                         wakeup(wpipe);
 975                 }
 976                 pipeselwakeup(wpipe);
 977                 wpipe->pipe_state |= PIPE_WANTW;
 978                 pipeunlock(wpipe);
 979                 error = msleep(wpipe, PIPE_MTX(wpipe),
 980                     PRIBIO | PCATCH, "pipdwc", 0);
 981                 if (error)
 982                         goto error1;
 983                 else
 984                         goto retry;
 985         }
 986
 987         wpipe->pipe_state |= PIPE_DIRECTW;
 988
 989         PIPE_UNLOCK(wpipe);
 990         error = pipe_build_write_buffer(wpipe, uio);
 991         PIPE_LOCK(wpipe);
 992         if (error) {
 993                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 994                 pipeunlock(wpipe);
 995                 goto error1;
 996         }
 997
 998         error = 0;
 999         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1000                 if (wpipe->pipe_state & PIPE_EOF) {
1001                         pipe_destroy_write_buffer(wpipe);
1002                         pipeselwakeup(wpipe);
1003                         pipeunlock(wpipe);
1004                         error = EPIPE;
1005                         goto error1;
1006                 }
1007                 if (wpipe->pipe_state & PIPE_WANTR) {
1008                         wpipe->pipe_state &= ~PIPE_WANTR;
1009                         wakeup(wpipe);
1010                 }
1011                 pipeselwakeup(wpipe);
1012                 wpipe->pipe_state |= PIPE_WANTW;
1013                 pipeunlock(wpipe);
1014                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1015                     "pipdwt", 0);
1016                 pipelock(wpipe, 0);
1017         }
1018
1019         if (wpipe->pipe_state & PIPE_EOF)
1020                 error = EPIPE;
1021         if (wpipe->pipe_state & PIPE_DIRECTW) {
1022                 /*
1023                  * this bit of trickery substitutes a kernel buffer for
1024                  * the process that might be going away.
1025                  */
1026                 pipe_clone_write_buffer(wpipe);
1027         } else {
1028                 pipe_destroy_write_buffer(wpipe);
1029         }
1030         pipeunlock(wpipe);
1031         return (error);
1032
1033 error1:
1034         wakeup(wpipe);
1035         return (error);
1036 }
1037 #endif
1038
1039 static int
1040 pipe_write(fp, uio, active_cred, flags, td)
1041         struct file *fp;
1042         struct uio *uio;
1043         struct ucred *active_cred;
1044         struct thread *td;
1045         int flags;
1046 {
1047         int error = 0;
1048         int desiredsize;
1049         ssize_t orig_resid;
1050         struct pipe *wpipe, *rpipe;
1051
1052         rpipe = fp->f_data;
1053         wpipe = PIPE_PEER(rpipe);
1054         PIPE_LOCK(rpipe);
1055         error = pipelock(wpipe, 1);
1056         if (error) {
1057                 PIPE_UNLOCK(rpipe);
1058                 return (error);
1059         }
1060         /*
1061          * detect loss of pipe read side, issue SIGPIPE if lost.
1062          */
1063         if (wpipe->pipe_present != PIPE_ACTIVE ||
1064             (wpipe->pipe_state & PIPE_EOF)) {
1065                 pipeunlock(wpipe);
1066                 PIPE_UNLOCK(rpipe);
1067                 return (EPIPE);
1068         }
1069 #ifdef MAC
1070         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1071         if (error) {
1072                 pipeunlock(wpipe);
1073                 PIPE_UNLOCK(rpipe);
1074                 return (error);
1075         }
1076 #endif
1077         ++wpipe->pipe_busy;
1078
1079         /* Choose a larger size if it's advantageous */
1080         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1081         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1082                 if (piperesizeallowed != 1)
1083                         break;
1084                 if (amountpipekva > maxpipekva / 2)
1085                         break;
1086                 if (desiredsize == BIG_PIPE_SIZE)
1087                         break;
1088                 desiredsize = desiredsize * 2;
1089         }
1090
1091         /* Choose a smaller size if we're in a OOM situation */
1092         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1093                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1094                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1095                 (piperesizeallowed == 1))
1096                 desiredsize = SMALL_PIPE_SIZE;
1097
1098         /* Resize if the above determined that a new size was necessary */
1099         if ((desiredsize != wpipe->pipe_buffer.size) &&
1100                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1101                 PIPE_UNLOCK(wpipe);
1102                 pipespace(wpipe, desiredsize);
1103                 PIPE_LOCK(wpipe);
1104         }
1105         if (wpipe->pipe_buffer.size == 0) {
1106                 /*
1107                  * This can only happen for reverse direction use of pipes
1108                  * in a complete OOM situation.
1109                  */
1110                 error = ENOMEM;
1111                 --wpipe->pipe_busy;
1112                 pipeunlock(wpipe);
1113                 PIPE_UNLOCK(wpipe);
1114                 return (error);
1115         }
1116
1117         pipeunlock(wpipe);
1118
1119         orig_resid = uio->uio_resid;
1120
1121         while (uio->uio_resid) {
1122                 int space;
1123
1124                 pipelock(wpipe, 0);
1125                 if (wpipe->pipe_state & PIPE_EOF) {
1126                         pipeunlock(wpipe);
1127                         error = EPIPE;
1128                         break;
1129                 }
1130 #ifndef PIPE_NODIRECT
1131                 /*
1132                  * If the transfer is large, we can gain performance if
1133                  * we do process-to-process copies directly.
1134                  * If the write is non-blocking, we don't use the
1135                  * direct write mechanism.
1136                  *
1137                  * The direct write mechanism will detect the reader going
1138                  * away on us.
1139                  */
1140                 if (uio->uio_segflg == UIO_USERSPACE &&
1141                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1142                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1143                     (fp->f_flag & FNONBLOCK) == 0) {
1144                         pipeunlock(wpipe);
1145                         error = pipe_direct_write(wpipe, uio);
1146                         if (error)
1147                                 break;
1148                         continue;
1149                 }
1150 #endif
1151
1152                 /*
1153                  * Pipe buffered writes cannot be coincidental with
1154                  * direct writes.  We wait until the currently executing
1155                  * direct write is completed before we start filling the
1156                  * pipe buffer.  We break out if a signal occurs or the
1157                  * reader goes away.
1158                  */
1159                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1160                         if (wpipe->pipe_state & PIPE_WANTR) {
1161                                 wpipe->pipe_state &= ~PIPE_WANTR;
1162                                 wakeup(wpipe);
1163                         }
1164                         pipeselwakeup(wpipe);
1165                         wpipe->pipe_state |= PIPE_WANTW;
1166                         pipeunlock(wpipe);
1167                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1168                             "pipbww", 0);
1169                         if (error)
1170                                 break;
1171                         else
1172                                 continue;
1173                 }
1174
1175                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1176
1177                 /* Writes of size <= PIPE_BUF must be atomic. */
1178                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1179                         space = 0;
1180
1181                 if (space > 0) {
1182                         int size;       /* Transfer size */
1183                         int segsize;    /* first segment to transfer */
1184
1185                         /*
1186                          * Transfer size is minimum of uio transfer
1187                          * and free space in pipe buffer.
1188                          */
1189                         if (space > uio->uio_resid)
1190                                 size = uio->uio_resid;
1191                         else
1192                                 size = space;
1193                         /*
1194                          * First segment to transfer is minimum of
1195                          * transfer size and contiguous space in
1196                          * pipe buffer.  If first segment to transfer
1197                          * is less than the transfer size, we've got
1198                          * a wraparound in the buffer.
1199                          */
1200                         segsize = wpipe->pipe_buffer.size -
1201                                 wpipe->pipe_buffer.in;
1202                         if (segsize > size)
1203                                 segsize = size;
1204
1205                         /* Transfer first segment */
1206
1207                         PIPE_UNLOCK(rpipe);
1208                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1209                                         segsize, uio);
1210                         PIPE_LOCK(rpipe);
1211
1212                         if (error == 0 && segsize < size) {
1213                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1214                                         wpipe->pipe_buffer.size,
1215                                         ("Pipe buffer wraparound disappeared"));
1216                                 /*
1217                                  * Transfer remaining part now, to
1218                                  * support atomic writes.  Wraparound
1219                                  * happened.
1220                                  */
1221
1222                                 PIPE_UNLOCK(rpipe);
1223                                 error = uiomove(
1224                                     &wpipe->pipe_buffer.buffer[0],
1225                                     size - segsize, uio);
1226                                 PIPE_LOCK(rpipe);
1227                         }
1228                         if (error == 0) {
1229                                 wpipe->pipe_buffer.in += size;
1230                                 if (wpipe->pipe_buffer.in >=
1231                                     wpipe->pipe_buffer.size) {
1232                                         KASSERT(wpipe->pipe_buffer.in ==
1233                                                 size - segsize +
1234                                                 wpipe->pipe_buffer.size,
1235                                                 ("Expected wraparound bad"));
1236                                         wpipe->pipe_buffer.in = size - segsize;
1237                                 }
1238
1239                                 wpipe->pipe_buffer.cnt += size;
1240                                 KASSERT(wpipe->pipe_buffer.cnt <=
1241                                         wpipe->pipe_buffer.size,
1242                                         ("Pipe buffer overflow"));
1243                         }
1244                         pipeunlock(wpipe);
1245                         if (error != 0)
1246                                 break;
1247                 } else {
1248                         /*
1249                          * If the "read-side" has been blocked, wake it up now.
1250                          */
1251                         if (wpipe->pipe_state & PIPE_WANTR) {
1252                                 wpipe->pipe_state &= ~PIPE_WANTR;
1253                                 wakeup(wpipe);
1254                         }
1255
1256                         /*
1257                          * don't block on non-blocking I/O
1258                          */
1259                         if (fp->f_flag & FNONBLOCK) {
1260                                 error = EAGAIN;
1261                                 pipeunlock(wpipe);
1262                                 break;
1263                         }
1264
1265                         /*
1266                          * We have no more space and have something to offer,
1267                          * wake up select/poll.
1268                          */
1269                         pipeselwakeup(wpipe);
1270
1271                         wpipe->pipe_state |= PIPE_WANTW;
1272                         pipeunlock(wpipe);
1273                         error = msleep(wpipe, PIPE_MTX(rpipe),
1274                             PRIBIO | PCATCH, "pipewr", 0);
1275                         if (error != 0)
1276                                 break;
1277                 }
1278         }
1279
1280         pipelock(wpipe, 0);
1281         --wpipe->pipe_busy;
1282
1283         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1284                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1285                 wakeup(wpipe);
1286         } else if (wpipe->pipe_buffer.cnt > 0) {
1287                 /*
1288                  * If we have put any characters in the buffer, we wake up
1289                  * the reader.
1290                  */
1291                 if (wpipe->pipe_state & PIPE_WANTR) {
1292                         wpipe->pipe_state &= ~PIPE_WANTR;
1293                         wakeup(wpipe);
1294                 }
1295         }
1296
1297         /*
1298          * Don't return EPIPE if any byte was written.
1299          * EINTR and other interrupts are handled by generic I/O layer.
1300          * Do not pretend that I/O succeeded for obvious user error
1301          * like EFAULT.
1302          */
1303         if (uio->uio_resid != orig_resid && error == EPIPE)
1304                 error = 0;
1305
1306         if (error == 0)
1307                 vfs_timestamp(&wpipe->pipe_mtime);
1308
1309         /*
1310          * We have something to offer,
1311          * wake up select/poll.
1312          */
1313         if (wpipe->pipe_buffer.cnt)
1314                 pipeselwakeup(wpipe);
1315
1316         pipeunlock(wpipe);
1317         PIPE_UNLOCK(rpipe);
1318         return (error);
1319 }
1320
1321 /* ARGSUSED */
1322 static int
1323 pipe_truncate(fp, length, active_cred, td)
1324         struct file *fp;
1325         off_t length;
1326         struct ucred *active_cred;
1327         struct thread *td;
1328 {
1329         struct pipe *cpipe;
1330         int error;
1331
1332         cpipe = fp->f_data;
1333         if (cpipe->pipe_state & PIPE_NAMED)
1334                 error = vnops.fo_truncate(fp, length, active_cred, td);
1335         else
1336                 error = invfo_truncate(fp, length, active_cred, td);
1337         return (error);
1338 }
1339
1340 /*
1341  * we implement a very minimal set of ioctls for compatibility with sockets.
1342  */
1343 static int
1344 pipe_ioctl(fp, cmd, data, active_cred, td)
1345         struct file *fp;
1346         u_long cmd;
1347         void *data;
1348         struct ucred *active_cred;
1349         struct thread *td;
1350 {
1351         struct pipe *mpipe = fp->f_data;
1352         int error;
1353
1354         PIPE_LOCK(mpipe);
1355
1356 #ifdef MAC
1357         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1358         if (error) {
1359                 PIPE_UNLOCK(mpipe);
1360                 return (error);
1361         }
1362 #endif
1363
1364         error = 0;
1365         switch (cmd) {
1366
1367         case FIONBIO:
1368                 break;
1369
1370         case FIOASYNC:
1371                 if (*(int *)data) {
1372                         mpipe->pipe_state |= PIPE_ASYNC;
1373                 } else {
1374                         mpipe->pipe_state &= ~PIPE_ASYNC;
1375                 }
1376                 break;
1377
1378         case FIONREAD:
1379                 if (!(fp->f_flag & FREAD)) {
1380                         *(int *)data = 0;
1381                         PIPE_UNLOCK(mpipe);
1382                         return (0);
1383                 }
1384                 if (mpipe->pipe_state & PIPE_DIRECTW)
1385                         *(int *)data = mpipe->pipe_map.cnt;
1386                 else
1387                         *(int *)data = mpipe->pipe_buffer.cnt;
1388                 break;
1389
1390         case FIOSETOWN:
1391                 PIPE_UNLOCK(mpipe);
1392                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1393                 goto out_unlocked;
1394
1395         case FIOGETOWN:
1396                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1397                 break;
1398
1399         /* This is deprecated, FIOSETOWN should be used instead. */
1400         case TIOCSPGRP:
1401                 PIPE_UNLOCK(mpipe);
1402                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1403                 goto out_unlocked;
1404
1405         /* This is deprecated, FIOGETOWN should be used instead. */
1406         case TIOCGPGRP:
1407                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1408                 break;
1409
1410         default:
1411                 error = ENOTTY;
1412                 break;
1413         }
1414         PIPE_UNLOCK(mpipe);
1415 out_unlocked:
1416         return (error);
1417 }
1418
1419 static int
1420 pipe_poll(fp, events, active_cred, td)
1421         struct file *fp;
1422         int events;
1423         struct ucred *active_cred;
1424         struct thread *td;
1425 {
1426         struct pipe *rpipe;
1427         struct pipe *wpipe;
1428         int levents, revents;
1429 #ifdef MAC
1430         int error;
1431 #endif
1432
1433         revents = 0;
1434         rpipe = fp->f_data;
1435         wpipe = PIPE_PEER(rpipe);
1436         PIPE_LOCK(rpipe);
1437 #ifdef MAC
1438         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1439         if (error)
1440                 goto locked_error;
1441 #endif
1442         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1443                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1444                     (rpipe->pipe_buffer.cnt > 0))
1445                         revents |= events & (POLLIN | POLLRDNORM);
1446
1447         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1448                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1449                     (wpipe->pipe_state & PIPE_EOF) ||
1450                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1451                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1452                          wpipe->pipe_buffer.size == 0)))
1453                         revents |= events & (POLLOUT | POLLWRNORM);
1454
1455         levents = events &
1456             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1457         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1458             fp->f_seqcount == rpipe->pipe_wgen)
1459                 events |= POLLINIGNEOF;
1460
1461         if ((events & POLLINIGNEOF) == 0) {
1462                 if (rpipe->pipe_state & PIPE_EOF) {
1463                         revents |= (events & (POLLIN | POLLRDNORM));
1464                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1465                             (wpipe->pipe_state & PIPE_EOF))
1466                                 revents |= POLLHUP;
1467                 }
1468         }
1469
1470         if (revents == 0) {
1471                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1472                         selrecord(td, &rpipe->pipe_sel);
1473                         if (SEL_WAITING(&rpipe->pipe_sel))
1474                                 rpipe->pipe_state |= PIPE_SEL;
1475                 }
1476
1477                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1478                         selrecord(td, &wpipe->pipe_sel);
1479                         if (SEL_WAITING(&wpipe->pipe_sel))
1480                                 wpipe->pipe_state |= PIPE_SEL;
1481                 }
1482         }
1483 #ifdef MAC
1484 locked_error:
1485 #endif
1486         PIPE_UNLOCK(rpipe);
1487
1488         return (revents);
1489 }
1490
1491 /*
1492  * We shouldn't need locks here as we're doing a read and this should
1493  * be a natural race.
1494  */
1495 static int
1496 pipe_stat(fp, ub, active_cred, td)
1497         struct file *fp;
1498         struct stat *ub;
1499         struct ucred *active_cred;
1500         struct thread *td;
1501 {
1502         struct pipe *pipe;
1503         int new_unr;
1504 #ifdef MAC
1505         int error;
1506 #endif
1507
1508         pipe = fp->f_data;
1509         PIPE_LOCK(pipe);
1510 #ifdef MAC
1511         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1512         if (error) {
1513                 PIPE_UNLOCK(pipe);
1514                 return (error);
1515         }
1516 #endif
1517
1518         /* For named pipes ask the underlying filesystem. */
1519         if (pipe->pipe_state & PIPE_NAMED) {
1520                 PIPE_UNLOCK(pipe);
1521                 return (vnops.fo_stat(fp, ub, active_cred, td));
1522         }
1523
1524         /*
1525          * Lazily allocate an inode number for the pipe.  Most pipe
1526          * users do not call fstat(2) on the pipe, which means that
1527          * postponing the inode allocation until it is must be
1528          * returned to userland is useful.  If alloc_unr failed,
1529          * assign st_ino zero instead of returning an error.
1530          * Special pipe_ino values:
1531          *  -1 - not yet initialized;
1532          *  0  - alloc_unr failed, return 0 as st_ino forever.
1533          */
1534         if (pipe->pipe_ino == (ino_t)-1) {
1535                 new_unr = alloc_unr(pipeino_unr);
1536                 if (new_unr != -1)
1537                         pipe->pipe_ino = new_unr;
1538                 else
1539                         pipe->pipe_ino = 0;
1540         }
1541         PIPE_UNLOCK(pipe);
1542
1543         bzero(ub, sizeof(*ub));
1544         ub->st_mode = S_IFIFO;
1545         ub->st_blksize = PAGE_SIZE;
1546         if (pipe->pipe_state & PIPE_DIRECTW)
1547                 ub->st_size = pipe->pipe_map.cnt;
1548         else
1549                 ub->st_size = pipe->pipe_buffer.cnt;
1550         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1551         ub->st_atim = pipe->pipe_atime;
1552         ub->st_mtim = pipe->pipe_mtime;
1553         ub->st_ctim = pipe->pipe_ctime;
1554         ub->st_uid = fp->f_cred->cr_uid;
1555         ub->st_gid = fp->f_cred->cr_gid;
1556         ub->st_dev = pipedev_ino;
1557         ub->st_ino = pipe->pipe_ino;
1558         /*
1559          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1560          */
1561         return (0);
1562 }
1563
1564 /* ARGSUSED */
1565 static int
1566 pipe_close(fp, td)
1567         struct file *fp;
1568         struct thread *td;
1569 {
1570
1571         if (fp->f_vnode != NULL)
1572                 return vnops.fo_close(fp, td);
1573         fp->f_ops = &badfileops;
1574         pipe_dtor(fp->f_data);
1575         fp->f_data = NULL;
1576         return (0);
1577 }
1578
1579 static int
1580 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1581 {
1582         struct pipe *cpipe;
1583         int error;
1584
1585         cpipe = fp->f_data;
1586         if (cpipe->pipe_state & PIPE_NAMED)
1587                 error = vn_chmod(fp, mode, active_cred, td);
1588         else
1589                 error = invfo_chmod(fp, mode, active_cred, td);
1590         return (error);
1591 }
1592
1593 static int
1594 pipe_chown(fp, uid, gid, active_cred, td)
1595         struct file *fp;
1596         uid_t uid;
1597         gid_t gid;
1598         struct ucred *active_cred;
1599         struct thread *td;
1600 {
1601         struct pipe *cpipe;
1602         int error;
1603
1604         cpipe = fp->f_data;
1605         if (cpipe->pipe_state & PIPE_NAMED)
1606                 error = vn_chown(fp, uid, gid, active_cred, td);
1607         else
1608                 error = invfo_chown(fp, uid, gid, active_cred, td);
1609         return (error);
1610 }
1611
1612 static int
1613 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1614 {
1615         struct pipe *pi;
1616
1617         if (fp->f_type == DTYPE_FIFO)
1618                 return (vn_fill_kinfo(fp, kif, fdp));
1619         kif->kf_type = KF_TYPE_PIPE;
1620         pi = fp->f_data;
1621         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1622         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1623         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1624         return (0);
1625 }
1626
1627 static void
1628 pipe_free_kmem(cpipe)
1629         struct pipe *cpipe;
1630 {
1631
1632         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1633             ("pipe_free_kmem: pipe mutex locked"));
1634
1635         if (cpipe->pipe_buffer.buffer != NULL) {
1636                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1637                 vm_map_remove(pipe_map,
1638                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1639                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1640                 cpipe->pipe_buffer.buffer = NULL;
1641         }
1642 #ifndef PIPE_NODIRECT
1643         {
1644                 cpipe->pipe_map.cnt = 0;
1645                 cpipe->pipe_map.pos = 0;
1646                 cpipe->pipe_map.npages = 0;
1647         }
1648 #endif
1649 }
1650
1651 /*
1652  * shutdown the pipe
1653  */
1654 static void
1655 pipeclose(cpipe)
1656         struct pipe *cpipe;
1657 {
1658         struct pipepair *pp;
1659         struct pipe *ppipe;
1660
1661         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1662
1663         PIPE_LOCK(cpipe);
1664         pipelock(cpipe, 0);
1665         pp = cpipe->pipe_pair;
1666
1667         pipeselwakeup(cpipe);
1668
1669         /*
1670          * If the other side is blocked, wake it up saying that
1671          * we want to close it down.
1672          */
1673         cpipe->pipe_state |= PIPE_EOF;
1674         while (cpipe->pipe_busy) {
1675                 wakeup(cpipe);
1676                 cpipe->pipe_state |= PIPE_WANT;
1677                 pipeunlock(cpipe);
1678                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1679                 pipelock(cpipe, 0);
1680         }
1681
1682
1683         /*
1684          * Disconnect from peer, if any.
1685          */
1686         ppipe = cpipe->pipe_peer;
1687         if (ppipe->pipe_present == PIPE_ACTIVE) {
1688                 pipeselwakeup(ppipe);
1689
1690                 ppipe->pipe_state |= PIPE_EOF;
1691                 wakeup(ppipe);
1692                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1693         }
1694
1695         /*
1696          * Mark this endpoint as free.  Release kmem resources.  We
1697          * don't mark this endpoint as unused until we've finished
1698          * doing that, or the pipe might disappear out from under
1699          * us.
1700          */
1701         PIPE_UNLOCK(cpipe);
1702         pipe_free_kmem(cpipe);
1703         PIPE_LOCK(cpipe);
1704         cpipe->pipe_present = PIPE_CLOSING;
1705         pipeunlock(cpipe);
1706
1707         /*
1708          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1709          * PIPE_FINALIZED, that allows other end to free the
1710          * pipe_pair, only after the knotes are completely dismantled.
1711          */
1712         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1713         cpipe->pipe_present = PIPE_FINALIZED;
1714         seldrain(&cpipe->pipe_sel);
1715         knlist_destroy(&cpipe->pipe_sel.si_note);
1716
1717         /*
1718          * If both endpoints are now closed, release the memory for the
1719          * pipe pair.  If not, unlock.
1720          */
1721         if (ppipe->pipe_present == PIPE_FINALIZED) {
1722                 PIPE_UNLOCK(cpipe);
1723 #ifdef MAC
1724                 mac_pipe_destroy(pp);
1725 #endif
1726                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1727         } else
1728                 PIPE_UNLOCK(cpipe);
1729 }
1730
1731 /*ARGSUSED*/
1732 static int
1733 pipe_kqfilter(struct file *fp, struct knote *kn)
1734 {
1735         struct pipe *cpipe;
1736
1737         /*
1738          * If a filter is requested that is not supported by this file
1739          * descriptor, don't return an error, but also don't ever generate an
1740          * event.
1741          */
1742         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1743                 kn->kn_fop = &pipe_nfiltops;
1744                 return (0);
1745         }
1746         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1747                 kn->kn_fop = &pipe_nfiltops;
1748                 return (0);
1749         }
1750         cpipe = fp->f_data;
1751         PIPE_LOCK(cpipe);
1752         switch (kn->kn_filter) {
1753         case EVFILT_READ:
1754                 kn->kn_fop = &pipe_rfiltops;
1755                 break;
1756         case EVFILT_WRITE:
1757                 kn->kn_fop = &pipe_wfiltops;
1758                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1759                         /* other end of pipe has been closed */
1760                         PIPE_UNLOCK(cpipe);
1761                         return (EPIPE);
1762                 }
1763                 cpipe = PIPE_PEER(cpipe);
1764                 break;
1765         default:
1766                 PIPE_UNLOCK(cpipe);
1767                 return (EINVAL);
1768         }
1769
1770         kn->kn_hook = cpipe;
1771         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1772         PIPE_UNLOCK(cpipe);
1773         return (0);
1774 }
1775
1776 static void
1777 filt_pipedetach(struct knote *kn)
1778 {
1779         struct pipe *cpipe = kn->kn_hook;
1780
1781         PIPE_LOCK(cpipe);
1782         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1783         PIPE_UNLOCK(cpipe);
1784 }
1785
1786 /*ARGSUSED*/
1787 static int
1788 filt_piperead(struct knote *kn, long hint)
1789 {
1790         struct pipe *rpipe = kn->kn_hook;
1791         struct pipe *wpipe = rpipe->pipe_peer;
1792         int ret;
1793
1794         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1795         kn->kn_data = rpipe->pipe_buffer.cnt;
1796         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1797                 kn->kn_data = rpipe->pipe_map.cnt;
1798
1799         if ((rpipe->pipe_state & PIPE_EOF) ||
1800             wpipe->pipe_present != PIPE_ACTIVE ||
1801             (wpipe->pipe_state & PIPE_EOF)) {
1802                 kn->kn_flags |= EV_EOF;
1803                 return (1);
1804         }
1805         ret = kn->kn_data > 0;
1806         return ret;
1807 }
1808
1809 /*ARGSUSED*/
1810 static int
1811 filt_pipewrite(struct knote *kn, long hint)
1812 {
1813         struct pipe *wpipe;
1814
1815         wpipe = kn->kn_hook;
1816         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1817         if (wpipe->pipe_present != PIPE_ACTIVE ||
1818             (wpipe->pipe_state & PIPE_EOF)) {
1819                 kn->kn_data = 0;
1820                 kn->kn_flags |= EV_EOF;
1821                 return (1);
1822         }
1823         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1824             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1825         if (wpipe->pipe_state & PIPE_DIRECTW)
1826                 kn->kn_data = 0;
1827
1828         return (kn->kn_data >= PIPE_BUF);
1829 }
1830
1831 static void
1832 filt_pipedetach_notsup(struct knote *kn)
1833 {
1834
1835 }
1836
1837 static int
1838 filt_pipenotsup(struct knote *kn, long hint)
1839 {
1840
1841         return (0);
1842 }