sys/kern/sys_pipe.c

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