sys/kern/sys_pipe.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-4-Clause
   3  *
   4  * Copyright (c) 1996 John S. Dyson
   5  * Copyright (c) 2012 Giovanni Trematerra
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice immediately at the beginning of the file, without modification,
  13  *    this list of conditions, and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Absolutely no warranty of function or purpose is made by the author
  18  *    John S. Dyson.
  19  * 4. Modifications may be freely made to this file if the above conditions
  20  *    are met.
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  36  * memory, and the receiving process copies directly from these pinned pages
  37  * in the sending process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.
  50  *
  51  * In order to limit the resource use of pipes, two sysctls exist:
  52  *
  53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  54  * address space available to us in pipe_map. This value is normally
  55  * autotuned, but may also be loader tuned.
  56  *
  57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  58  * memory in use by pipes.
  59  *
  60  * Based on how large pipekva is relative to maxpipekva, the following
  61  * will happen:
  62  *
  63  * 0% - 50%:
  64  *     New pipes are given 16K of memory backing, pipes may dynamically
  65  *     grow to as large as 64K where needed.
  66  * 50% - 75%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes may NOT grow.
  69  * 75% - 100%:
  70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  71  *     existing pipes will be shrunk down to 4K whenever possible.
  72  *
  73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  75  * resize which MUST occur for reverse-direction pipes when they are
  76  * first used.
  77  *
  78  * Additional information about the current state of pipes may be obtained
  79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  80  * and kern.ipc.piperesizefail.
  81  *
  82  * Locking rules:  There are two locks present here:  A mutex, used via
  83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  84  * the flag, as mutexes can not persist over uiomove.  The mutex
  85  * exists only to guard access to the flag, and is not in itself a
  86  * locking mechanism.  Also note that there is only a single mutex for
  87  * both directions of a pipe.
  88  *
  89  * As pipelock() may have to sleep before it can acquire the flag, it
  90  * is important to reread all data after a call to pipelock(); everything
  91  * in the structure may have changed.
  92  */
  93
  94 #include <sys/cdefs.h>
  95 __FBSDID("$FreeBSD$");
  96
  97 #include <sys/param.h>
  98 #include <sys/systm.h>
  99 #include <sys/conf.h>
 100 #include <sys/fcntl.h>
 101 #include <sys/file.h>
 102 #include <sys/filedesc.h>
 103 #include <sys/filio.h>
 104 #include <sys/kernel.h>
 105 #include <sys/lock.h>
 106 #include <sys/mutex.h>
 107 #include <sys/ttycom.h>
 108 #include <sys/stat.h>
 109 #include <sys/malloc.h>
 110 #include <sys/poll.h>
 111 #include <sys/selinfo.h>
 112 #include <sys/signalvar.h>
 113 #include <sys/syscallsubr.h>
 114 #include <sys/sysctl.h>
 115 #include <sys/sysproto.h>
 116 #include <sys/pipe.h>
 117 #include <sys/proc.h>
 118 #include <sys/vnode.h>
 119 #include <sys/uio.h>
 120 #include <sys/user.h>
 121 #include <sys/event.h>
 122
 123 #include <security/mac/mac_framework.h>
 124
 125 #include <vm/vm.h>
 126 #include <vm/vm_param.h>
 127 #include <vm/vm_object.h>
 128 #include <vm/vm_kern.h>
 129 #include <vm/vm_extern.h>
 130 #include <vm/pmap.h>
 131 #include <vm/vm_map.h>
 132 #include <vm/vm_page.h>
 133 #include <vm/uma.h>
 134
 135 /*
 136  * Use this define if you want to disable *fancy* VM things.  Expect an
 137  * approx 30% decrease in transfer rate.  This could be useful for
 138  * NetBSD or OpenBSD.
 139  */
 140 /* #define PIPE_NODIRECT */
 141
 142 #define PIPE_PEER(pipe) \
 143         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 144
 145 /*
 146  * interfaces to the outside world
 147  */
 148 static fo_rdwr_t        pipe_read;
 149 static fo_rdwr_t        pipe_write;
 150 static fo_truncate_t    pipe_truncate;
 151 static fo_ioctl_t       pipe_ioctl;
 152 static fo_poll_t        pipe_poll;
 153 static fo_kqfilter_t    pipe_kqfilter;
 154 static fo_stat_t        pipe_stat;
 155 static fo_close_t       pipe_close;
 156 static fo_chmod_t       pipe_chmod;
 157 static fo_chown_t       pipe_chown;
 158 static fo_fill_kinfo_t  pipe_fill_kinfo;
 159
 160 struct fileops pipeops = {
 161         .fo_read = pipe_read,
 162         .fo_write = pipe_write,
 163         .fo_truncate = pipe_truncate,
 164         .fo_ioctl = pipe_ioctl,
 165         .fo_poll = pipe_poll,
 166         .fo_kqfilter = pipe_kqfilter,
 167         .fo_stat = pipe_stat,
 168         .fo_close = pipe_close,
 169         .fo_chmod = pipe_chmod,
 170         .fo_chown = pipe_chown,
 171         .fo_sendfile = invfo_sendfile,
 172         .fo_fill_kinfo = pipe_fill_kinfo,
 173         .fo_flags = DFLAG_PASSABLE
 174 };
 175
 176 static void     filt_pipedetach(struct knote *kn);
 177 static void     filt_pipedetach_notsup(struct knote *kn);
 178 static int      filt_pipenotsup(struct knote *kn, long hint);
 179 static int      filt_piperead(struct knote *kn, long hint);
 180 static int      filt_pipewrite(struct knote *kn, long hint);
 181
 182 static struct filterops pipe_nfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach_notsup,
 185         .f_event = filt_pipenotsup
 186 };
 187 static struct filterops pipe_rfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_piperead
 191 };
 192 static struct filterops pipe_wfiltops = {
 193         .f_isfd = 1,
 194         .f_detach = filt_pipedetach,
 195         .f_event = filt_pipewrite
 196 };
 197
 198 /*
 199  * Default pipe buffer size(s), this can be kind-of large now because pipe
 200  * space is pageable.  The pipe code will try to maintain locality of
 201  * reference for performance reasons, so small amounts of outstanding I/O
 202  * will not wipe the cache.
 203  */
 204 #define MINPIPESIZE (PIPE_SIZE/3)
 205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 206
 207 static long amountpipekva;
 208 static int pipefragretry;
 209 static int pipeallocfail;
 210 static int piperesizefail;
 211 static int piperesizeallowed = 1;
 212
 213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 214            &maxpipekva, 0, "Pipe KVA limit");
 215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 216            &amountpipekva, 0, "Pipe KVA usage");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 218           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 220           &pipeallocfail, 0, "Pipe allocation failures");
 221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 222           &piperesizefail, 0, "Pipe resize failures");
 223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 224           &piperesizeallowed, 0, "Pipe resizing allowed");
 225
 226 static void pipeinit(void *dummy __unused);
 227 static void pipeclose(struct pipe *cpipe);
 228 static void pipe_free_kmem(struct pipe *cpipe);
 229 static void pipe_create(struct pipe *pipe, int backing);
 230 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 231 static __inline int pipelock(struct pipe *cpipe, int catch);
 232 static __inline void pipeunlock(struct pipe *cpipe);
 233 #ifndef PIPE_NODIRECT
 234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 237 static void pipe_clone_write_buffer(struct pipe *wpipe);
 238 #endif
 239 static int pipespace(struct pipe *cpipe, int size);
 240 static int pipespace_new(struct pipe *cpipe, int size);
 241
 242 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 243 static int      pipe_zone_init(void *mem, int size, int flags);
 244 static void     pipe_zone_fini(void *mem, int size);
 245
 246 static uma_zone_t pipe_zone;
 247 static struct unrhdr64 pipeino_unr;
 248 static dev_t pipedev_ino;
 249
 250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 251
 252 static void
 253 pipeinit(void *dummy __unused)
 254 {
 255
 256         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 257             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 258             UMA_ALIGN_PTR, 0);
 259         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 260         new_unrhdr64(&pipeino_unr, 1);
 261         pipedev_ino = devfs_alloc_cdp_inode();
 262         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 263 }
 264
 265 static int
 266 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 267 {
 268         struct pipepair *pp;
 269         struct pipe *rpipe, *wpipe;
 270
 271         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 272
 273         pp = (struct pipepair *)mem;
 274
 275         /*
 276          * We zero both pipe endpoints to make sure all the kmem pointers
 277          * are NULL, flag fields are zero'd, etc.  We timestamp both
 278          * endpoints with the same time.
 279          */
 280         rpipe = &pp->pp_rpipe;
 281         bzero(rpipe, sizeof(*rpipe));
 282         vfs_timestamp(&rpipe->pipe_ctime);
 283         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 284
 285         wpipe = &pp->pp_wpipe;
 286         bzero(wpipe, sizeof(*wpipe));
 287         wpipe->pipe_ctime = rpipe->pipe_ctime;
 288         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 289
 290         rpipe->pipe_peer = wpipe;
 291         rpipe->pipe_pair = pp;
 292         wpipe->pipe_peer = rpipe;
 293         wpipe->pipe_pair = pp;
 294
 295         /*
 296          * Mark both endpoints as present; they will later get free'd
 297          * one at a time.  When both are free'd, then the whole pair
 298          * is released.
 299          */
 300         rpipe->pipe_present = PIPE_ACTIVE;
 301         wpipe->pipe_present = PIPE_ACTIVE;
 302
 303         /*
 304          * Eventually, the MAC Framework may initialize the label
 305          * in ctor or init, but for now we do it elswhere to avoid
 306          * blocking in ctor or init.
 307          */
 308         pp->pp_label = NULL;
 309
 310         return (0);
 311 }
 312
 313 static int
 314 pipe_zone_init(void *mem, int size, int flags)
 315 {
 316         struct pipepair *pp;
 317
 318         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 319
 320         pp = (struct pipepair *)mem;
 321
 322         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 323         return (0);
 324 }
 325
 326 static void
 327 pipe_zone_fini(void *mem, int size)
 328 {
 329         struct pipepair *pp;
 330
 331         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 332
 333         pp = (struct pipepair *)mem;
 334
 335         mtx_destroy(&pp->pp_mtx);
 336 }
 337
 338 static void
 339 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 340 {
 341         struct pipepair *pp;
 342         struct pipe *rpipe, *wpipe;
 343
 344         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 345 #ifdef MAC
 346         /*
 347          * The MAC label is shared between the connected endpoints.  As a
 348          * result mac_pipe_init() and mac_pipe_create() are called once
 349          * for the pair, and not on the endpoints.
 350          */
 351         mac_pipe_init(pp);
 352         mac_pipe_create(td->td_ucred, pp);
 353 #endif
 354         rpipe = &pp->pp_rpipe;
 355         wpipe = &pp->pp_wpipe;
 356
 357         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 358         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 359
 360         /* Only the forward direction pipe is backed by default */
 361         pipe_create(rpipe, 1);
 362         pipe_create(wpipe, 0);
 363
 364         rpipe->pipe_state |= PIPE_DIRECTOK;
 365         wpipe->pipe_state |= PIPE_DIRECTOK;
 366 }
 367
 368 void
 369 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 370 {
 371         struct pipepair *pp;
 372
 373         pipe_paircreate(td, &pp);
 374         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 375         *ppipe = &pp->pp_rpipe;
 376 }
 377
 378 void
 379 pipe_dtor(struct pipe *dpipe)
 380 {
 381         struct pipe *peer;
 382         ino_t ino;
 383
 384         ino = dpipe->pipe_ino;
 385         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 386         funsetown(&dpipe->pipe_sigio);
 387         pipeclose(dpipe);
 388         if (peer != NULL) {
 389                 funsetown(&peer->pipe_sigio);
 390                 pipeclose(peer);
 391         }
 392 }
 393
 394 /*
 395  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 396  * the zone pick up the pieces via pipeclose().
 397  */
 398 int
 399 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 400     struct filecaps *fcaps2)
 401 {
 402         struct file *rf, *wf;
 403         struct pipe *rpipe, *wpipe;
 404         struct pipepair *pp;
 405         int fd, fflags, error;
 406
 407         pipe_paircreate(td, &pp);
 408         rpipe = &pp->pp_rpipe;
 409         wpipe = &pp->pp_wpipe;
 410         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 411         if (error) {
 412                 pipeclose(rpipe);
 413                 pipeclose(wpipe);
 414                 return (error);
 415         }
 416         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 417         fildes[0] = fd;
 418
 419         fflags = FREAD | FWRITE;
 420         if ((flags & O_NONBLOCK) != 0)
 421                 fflags |= FNONBLOCK;
 422
 423         /*
 424          * Warning: once we've gotten past allocation of the fd for the
 425          * read-side, we can only drop the read side via fdrop() in order
 426          * to avoid races against processes which manage to dup() the read
 427          * side while we are blocked trying to allocate the write side.
 428          */
 429         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 430         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 431         if (error) {
 432                 fdclose(td, rf, fildes[0]);
 433                 fdrop(rf, td);
 434                 /* rpipe has been closed by fdrop(). */
 435                 pipeclose(wpipe);
 436                 return (error);
 437         }
 438         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 439         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 440         fdrop(wf, td);
 441         fildes[1] = fd;
 442         fdrop(rf, td);
 443
 444         return (0);
 445 }
 446
 447 #ifdef COMPAT_FREEBSD10
 448 /* ARGSUSED */
 449 int
 450 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 451 {
 452         int error;
 453         int fildes[2];
 454
 455         error = kern_pipe(td, fildes, 0, NULL, NULL);
 456         if (error)
 457                 return (error);
 458
 459         td->td_retval[0] = fildes[0];
 460         td->td_retval[1] = fildes[1];
 461
 462         return (0);
 463 }
 464 #endif
 465
 466 int
 467 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 468 {
 469         int error, fildes[2];
 470
 471         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 472                 return (EINVAL);
 473         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 474         if (error)
 475                 return (error);
 476         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 477         if (error) {
 478                 (void)kern_close(td, fildes[0]);
 479                 (void)kern_close(td, fildes[1]);
 480         }
 481         return (error);
 482 }
 483
 484 /*
 485  * Allocate kva for pipe circular buffer, the space is pageable
 486  * This routine will 'realloc' the size of a pipe safely, if it fails
 487  * it will retain the old buffer.
 488  * If it fails it will return ENOMEM.
 489  */
 490 static int
 491 pipespace_new(struct pipe *cpipe, int size)
 492 {
 493         caddr_t buffer;
 494         int error, cnt, firstseg;
 495         static int curfail = 0;
 496         static struct timeval lastfail;
 497
 498         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 499         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 500                 ("pipespace: resize of direct writes not allowed"));
 501 retry:
 502         cnt = cpipe->pipe_buffer.cnt;
 503         if (cnt > size)
 504                 size = cnt;
 505
 506         size = round_page(size);
 507         buffer = (caddr_t) vm_map_min(pipe_map);
 508
 509         error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
 510             VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
 511         if (error != KERN_SUCCESS) {
 512                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 513                         (size > SMALL_PIPE_SIZE)) {
 514                         size = SMALL_PIPE_SIZE;
 515                         pipefragretry++;
 516                         goto retry;
 517                 }
 518                 if (cpipe->pipe_buffer.buffer == NULL) {
 519                         pipeallocfail++;
 520                         if (ppsratecheck(&lastfail, &curfail, 1))
 521                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 522                 } else {
 523                         piperesizefail++;
 524                 }
 525                 return (ENOMEM);
 526         }
 527
 528         /* copy data, then free old resources if we're resizing */
 529         if (cnt > 0) {
 530                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 531                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 532                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 533                                 buffer, firstseg);
 534                         if ((cnt - firstseg) > 0)
 535                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 536                                         cpipe->pipe_buffer.in);
 537                 } else {
 538                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 539                                 buffer, cnt);
 540                 }
 541         }
 542         pipe_free_kmem(cpipe);
 543         cpipe->pipe_buffer.buffer = buffer;
 544         cpipe->pipe_buffer.size = size;
 545         cpipe->pipe_buffer.in = cnt;
 546         cpipe->pipe_buffer.out = 0;
 547         cpipe->pipe_buffer.cnt = cnt;
 548         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 549         return (0);
 550 }
 551
 552 /*
 553  * Wrapper for pipespace_new() that performs locking assertions.
 554  */
 555 static int
 556 pipespace(struct pipe *cpipe, int size)
 557 {
 558
 559         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 560                 ("Unlocked pipe passed to pipespace"));
 561         return (pipespace_new(cpipe, size));
 562 }
 563
 564 /*
 565  * lock a pipe for I/O, blocking other access
 566  */
 567 static __inline int
 568 pipelock(struct pipe *cpipe, int catch)
 569 {
 570         int error;
 571
 572         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 573         while (cpipe->pipe_state & PIPE_LOCKFL) {
 574                 cpipe->pipe_state |= PIPE_LWANT;
 575                 error = msleep(cpipe, PIPE_MTX(cpipe),
 576                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 577                     "pipelk", 0);
 578                 if (error != 0)
 579                         return (error);
 580         }
 581         cpipe->pipe_state |= PIPE_LOCKFL;
 582         return (0);
 583 }
 584
 585 /*
 586  * unlock a pipe I/O lock
 587  */
 588 static __inline void
 589 pipeunlock(struct pipe *cpipe)
 590 {
 591
 592         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 593         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 594                 ("Unlocked pipe passed to pipeunlock"));
 595         cpipe->pipe_state &= ~PIPE_LOCKFL;
 596         if (cpipe->pipe_state & PIPE_LWANT) {
 597                 cpipe->pipe_state &= ~PIPE_LWANT;
 598                 wakeup(cpipe);
 599         }
 600 }
 601
 602 void
 603 pipeselwakeup(struct pipe *cpipe)
 604 {
 605
 606         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 607         if (cpipe->pipe_state & PIPE_SEL) {
 608                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 609                 if (!SEL_WAITING(&cpipe->pipe_sel))
 610                         cpipe->pipe_state &= ~PIPE_SEL;
 611         }
 612         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 613                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 614         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 615 }
 616
 617 /*
 618  * Initialize and allocate VM and memory for pipe.  The structure
 619  * will start out zero'd from the ctor, so we just manage the kmem.
 620  */
 621 static void
 622 pipe_create(struct pipe *pipe, int backing)
 623 {
 624
 625         if (backing) {
 626                 /*
 627                  * Note that these functions can fail if pipe map is exhausted
 628                  * (as a result of too many pipes created), but we ignore the
 629                  * error as it is not fatal and could be provoked by
 630                  * unprivileged users. The only consequence is worse performance
 631                  * with given pipe.
 632                  */
 633                 if (amountpipekva > maxpipekva / 2)
 634                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 635                 else
 636                         (void)pipespace_new(pipe, PIPE_SIZE);
 637         }
 638
 639         pipe->pipe_ino = alloc_unr64(&pipeino_unr);
 640 }
 641
 642 /* ARGSUSED */
 643 static int
 644 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 645     int flags, struct thread *td)
 646 {
 647         struct pipe *rpipe;
 648         int error;
 649         int nread = 0;
 650         int size;
 651
 652         rpipe = fp->f_data;
 653         PIPE_LOCK(rpipe);
 654         ++rpipe->pipe_busy;
 655         error = pipelock(rpipe, 1);
 656         if (error)
 657                 goto unlocked_error;
 658
 659 #ifdef MAC
 660         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 661         if (error)
 662                 goto locked_error;
 663 #endif
 664         if (amountpipekva > (3 * maxpipekva) / 4) {
 665                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 666                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 667                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 668                         (piperesizeallowed == 1)) {
 669                         PIPE_UNLOCK(rpipe);
 670                         pipespace(rpipe, SMALL_PIPE_SIZE);
 671                         PIPE_LOCK(rpipe);
 672                 }
 673         }
 674
 675         while (uio->uio_resid) {
 676                 /*
 677                  * normal pipe buffer receive
 678                  */
 679                 if (rpipe->pipe_buffer.cnt > 0) {
 680                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 681                         if (size > rpipe->pipe_buffer.cnt)
 682                                 size = rpipe->pipe_buffer.cnt;
 683                         if (size > uio->uio_resid)
 684                                 size = uio->uio_resid;
 685
 686                         PIPE_UNLOCK(rpipe);
 687                         error = uiomove(
 688                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 689                             size, uio);
 690                         PIPE_LOCK(rpipe);
 691                         if (error)
 692                                 break;
 693
 694                         rpipe->pipe_buffer.out += size;
 695                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 696                                 rpipe->pipe_buffer.out = 0;
 697
 698                         rpipe->pipe_buffer.cnt -= size;
 699
 700                         /*
 701                          * If there is no more to read in the pipe, reset
 702                          * its pointers to the beginning.  This improves
 703                          * cache hit stats.
 704                          */
 705                         if (rpipe->pipe_buffer.cnt == 0) {
 706                                 rpipe->pipe_buffer.in = 0;
 707                                 rpipe->pipe_buffer.out = 0;
 708                         }
 709                         nread += size;
 710 #ifndef PIPE_NODIRECT
 711                 /*
 712                  * Direct copy, bypassing a kernel buffer.
 713                  */
 714                 } else if ((size = rpipe->pipe_map.cnt) &&
 715                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 716                         if (size > uio->uio_resid)
 717                                 size = (u_int) uio->uio_resid;
 718
 719                         PIPE_UNLOCK(rpipe);
 720                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 721                             rpipe->pipe_map.pos, size, uio);
 722                         PIPE_LOCK(rpipe);
 723                         if (error)
 724                                 break;
 725                         nread += size;
 726                         rpipe->pipe_map.pos += size;
 727                         rpipe->pipe_map.cnt -= size;
 728                         if (rpipe->pipe_map.cnt == 0) {
 729                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 730                                 wakeup(rpipe);
 731                         }
 732 #endif
 733                 } else {
 734                         /*
 735                          * detect EOF condition
 736                          * read returns 0 on EOF, no need to set error
 737                          */
 738                         if (rpipe->pipe_state & PIPE_EOF)
 739                                 break;
 740
 741                         /*
 742                          * If the "write-side" has been blocked, wake it up now.
 743                          */
 744                         if (rpipe->pipe_state & PIPE_WANTW) {
 745                                 rpipe->pipe_state &= ~PIPE_WANTW;
 746                                 wakeup(rpipe);
 747                         }
 748
 749                         /*
 750                          * Break if some data was read.
 751                          */
 752                         if (nread > 0)
 753                                 break;
 754
 755                         /*
 756                          * Unlock the pipe buffer for our remaining processing.
 757                          * We will either break out with an error or we will
 758                          * sleep and relock to loop.
 759                          */
 760                         pipeunlock(rpipe);
 761
 762                         /*
 763                          * Handle non-blocking mode operation or
 764                          * wait for more data.
 765                          */
 766                         if (fp->f_flag & FNONBLOCK) {
 767                                 error = EAGAIN;
 768                         } else {
 769                                 rpipe->pipe_state |= PIPE_WANTR;
 770                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 771                                     PRIBIO | PCATCH,
 772                                     "piperd", 0)) == 0)
 773                                         error = pipelock(rpipe, 1);
 774                         }
 775                         if (error)
 776                                 goto unlocked_error;
 777                 }
 778         }
 779 #ifdef MAC
 780 locked_error:
 781 #endif
 782         pipeunlock(rpipe);
 783
 784         /* XXX: should probably do this before getting any locks. */
 785         if (error == 0)
 786                 vfs_timestamp(&rpipe->pipe_atime);
 787 unlocked_error:
 788         --rpipe->pipe_busy;
 789
 790         /*
 791          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 792          */
 793         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 794                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 795                 wakeup(rpipe);
 796         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 797                 /*
 798                  * Handle write blocking hysteresis.
 799                  */
 800                 if (rpipe->pipe_state & PIPE_WANTW) {
 801                         rpipe->pipe_state &= ~PIPE_WANTW;
 802                         wakeup(rpipe);
 803                 }
 804         }
 805
 806         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 807                 pipeselwakeup(rpipe);
 808
 809         PIPE_UNLOCK(rpipe);
 810         return (error);
 811 }
 812
 813 #ifndef PIPE_NODIRECT
 814 /*
 815  * Map the sending processes' buffer into kernel space and wire it.
 816  * This is similar to a physical write operation.
 817  */
 818 static int
 819 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 820 {
 821         u_int size;
 822         int i;
 823
 824         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 825         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 826                 ("Clone attempt on non-direct write pipe!"));
 827
 828         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 829                 size = wpipe->pipe_buffer.size;
 830         else
 831                 size = uio->uio_iov->iov_len;
 832
 833         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 834             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 835             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 836                 return (EFAULT);
 837
 838 /*
 839  * set up the control block
 840  */
 841         wpipe->pipe_map.npages = i;
 842         wpipe->pipe_map.pos =
 843             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 844         wpipe->pipe_map.cnt = size;
 845
 846 /*
 847  * and update the uio data
 848  */
 849
 850         uio->uio_iov->iov_len -= size;
 851         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 852         if (uio->uio_iov->iov_len == 0)
 853                 uio->uio_iov++;
 854         uio->uio_resid -= size;
 855         uio->uio_offset += size;
 856         return (0);
 857 }
 858
 859 /*
 860  * unmap and unwire the process buffer
 861  */
 862 static void
 863 pipe_destroy_write_buffer(struct pipe *wpipe)
 864 {
 865
 866         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 867         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 868         wpipe->pipe_map.npages = 0;
 869 }
 870
 871 /*
 872  * In the case of a signal, the writing process might go away.  This
 873  * code copies the data into the circular buffer so that the source
 874  * pages can be freed without loss of data.
 875  */
 876 static void
 877 pipe_clone_write_buffer(struct pipe *wpipe)
 878 {
 879         struct uio uio;
 880         struct iovec iov;
 881         int size;
 882         int pos;
 883
 884         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 885         size = wpipe->pipe_map.cnt;
 886         pos = wpipe->pipe_map.pos;
 887
 888         wpipe->pipe_buffer.in = size;
 889         wpipe->pipe_buffer.out = 0;
 890         wpipe->pipe_buffer.cnt = size;
 891         wpipe->pipe_state &= ~PIPE_DIRECTW;
 892
 893         PIPE_UNLOCK(wpipe);
 894         iov.iov_base = wpipe->pipe_buffer.buffer;
 895         iov.iov_len = size;
 896         uio.uio_iov = &iov;
 897         uio.uio_iovcnt = 1;
 898         uio.uio_offset = 0;
 899         uio.uio_resid = size;
 900         uio.uio_segflg = UIO_SYSSPACE;
 901         uio.uio_rw = UIO_READ;
 902         uio.uio_td = curthread;
 903         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 904         PIPE_LOCK(wpipe);
 905         pipe_destroy_write_buffer(wpipe);
 906 }
 907
 908 /*
 909  * This implements the pipe buffer write mechanism.  Note that only
 910  * a direct write OR a normal pipe write can be pending at any given time.
 911  * If there are any characters in the pipe buffer, the direct write will
 912  * be deferred until the receiving process grabs all of the bytes from
 913  * the pipe buffer.  Then the direct mapping write is set-up.
 914  */
 915 static int
 916 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 917 {
 918         int error;
 919
 920 retry:
 921         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 922         error = pipelock(wpipe, 1);
 923         if (error != 0)
 924                 goto error1;
 925         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 926                 error = EPIPE;
 927                 pipeunlock(wpipe);
 928                 goto error1;
 929         }
 930         while (wpipe->pipe_state & PIPE_DIRECTW) {
 931                 if (wpipe->pipe_state & PIPE_WANTR) {
 932                         wpipe->pipe_state &= ~PIPE_WANTR;
 933                         wakeup(wpipe);
 934                 }
 935                 pipeselwakeup(wpipe);
 936                 wpipe->pipe_state |= PIPE_WANTW;
 937                 pipeunlock(wpipe);
 938                 error = msleep(wpipe, PIPE_MTX(wpipe),
 939                     PRIBIO | PCATCH, "pipdww", 0);
 940                 if (error)
 941                         goto error1;
 942                 else
 943                         goto retry;
 944         }
 945         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 946         if (wpipe->pipe_buffer.cnt > 0) {
 947                 if (wpipe->pipe_state & PIPE_WANTR) {
 948                         wpipe->pipe_state &= ~PIPE_WANTR;
 949                         wakeup(wpipe);
 950                 }
 951                 pipeselwakeup(wpipe);
 952                 wpipe->pipe_state |= PIPE_WANTW;
 953                 pipeunlock(wpipe);
 954                 error = msleep(wpipe, PIPE_MTX(wpipe),
 955                     PRIBIO | PCATCH, "pipdwc", 0);
 956                 if (error)
 957                         goto error1;
 958                 else
 959                         goto retry;
 960         }
 961
 962         wpipe->pipe_state |= PIPE_DIRECTW;
 963
 964         PIPE_UNLOCK(wpipe);
 965         error = pipe_build_write_buffer(wpipe, uio);
 966         PIPE_LOCK(wpipe);
 967         if (error) {
 968                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 969                 pipeunlock(wpipe);
 970                 goto error1;
 971         }
 972
 973         error = 0;
 974         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 975                 if (wpipe->pipe_state & PIPE_EOF) {
 976                         pipe_destroy_write_buffer(wpipe);
 977                         pipeselwakeup(wpipe);
 978                         pipeunlock(wpipe);
 979                         error = EPIPE;
 980                         goto error1;
 981                 }
 982                 if (wpipe->pipe_state & PIPE_WANTR) {
 983                         wpipe->pipe_state &= ~PIPE_WANTR;
 984                         wakeup(wpipe);
 985                 }
 986                 pipeselwakeup(wpipe);
 987                 wpipe->pipe_state |= PIPE_WANTW;
 988                 pipeunlock(wpipe);
 989                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 990                     "pipdwt", 0);
 991                 pipelock(wpipe, 0);
 992         }
 993
 994         if (wpipe->pipe_state & PIPE_EOF)
 995                 error = EPIPE;
 996         if (wpipe->pipe_state & PIPE_DIRECTW) {
 997                 /*
 998                  * this bit of trickery substitutes a kernel buffer for
 999                  * the process that might be going away.
1000                  */
1001                 pipe_clone_write_buffer(wpipe);
1002         } else {
1003                 pipe_destroy_write_buffer(wpipe);
1004         }
1005         pipeunlock(wpipe);
1006         return (error);
1007
1008 error1:
1009         wakeup(wpipe);
1010         return (error);
1011 }
1012 #endif
1013
1014 static int
1015 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1016     int flags, struct thread *td)
1017 {
1018         int error = 0;
1019         int desiredsize;
1020         ssize_t orig_resid;
1021         struct pipe *wpipe, *rpipe;
1022
1023         rpipe = fp->f_data;
1024         wpipe = PIPE_PEER(rpipe);
1025         PIPE_LOCK(rpipe);
1026         error = pipelock(wpipe, 1);
1027         if (error) {
1028                 PIPE_UNLOCK(rpipe);
1029                 return (error);
1030         }
1031         /*
1032          * detect loss of pipe read side, issue SIGPIPE if lost.
1033          */
1034         if (wpipe->pipe_present != PIPE_ACTIVE ||
1035             (wpipe->pipe_state & PIPE_EOF)) {
1036                 pipeunlock(wpipe);
1037                 PIPE_UNLOCK(rpipe);
1038                 return (EPIPE);
1039         }
1040 #ifdef MAC
1041         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1042         if (error) {
1043                 pipeunlock(wpipe);
1044                 PIPE_UNLOCK(rpipe);
1045                 return (error);
1046         }
1047 #endif
1048         ++wpipe->pipe_busy;
1049
1050         /* Choose a larger size if it's advantageous */
1051         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1052         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1053                 if (piperesizeallowed != 1)
1054                         break;
1055                 if (amountpipekva > maxpipekva / 2)
1056                         break;
1057                 if (desiredsize == BIG_PIPE_SIZE)
1058                         break;
1059                 desiredsize = desiredsize * 2;
1060         }
1061
1062         /* Choose a smaller size if we're in a OOM situation */
1063         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1064                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1065                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1066                 (piperesizeallowed == 1))
1067                 desiredsize = SMALL_PIPE_SIZE;
1068
1069         /* Resize if the above determined that a new size was necessary */
1070         if ((desiredsize != wpipe->pipe_buffer.size) &&
1071                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1072                 PIPE_UNLOCK(wpipe);
1073                 pipespace(wpipe, desiredsize);
1074                 PIPE_LOCK(wpipe);
1075         }
1076         if (wpipe->pipe_buffer.size == 0) {
1077                 /*
1078                  * This can only happen for reverse direction use of pipes
1079                  * in a complete OOM situation.
1080                  */
1081                 error = ENOMEM;
1082                 --wpipe->pipe_busy;
1083                 pipeunlock(wpipe);
1084                 PIPE_UNLOCK(wpipe);
1085                 return (error);
1086         }
1087
1088         pipeunlock(wpipe);
1089
1090         orig_resid = uio->uio_resid;
1091
1092         while (uio->uio_resid) {
1093                 int space;
1094
1095                 pipelock(wpipe, 0);
1096                 if (wpipe->pipe_state & PIPE_EOF) {
1097                         pipeunlock(wpipe);
1098                         error = EPIPE;
1099                         break;
1100                 }
1101 #ifndef PIPE_NODIRECT
1102                 /*
1103                  * If the transfer is large, we can gain performance if
1104                  * we do process-to-process copies directly.
1105                  * If the write is non-blocking, we don't use the
1106                  * direct write mechanism.
1107                  *
1108                  * The direct write mechanism will detect the reader going
1109                  * away on us.
1110                  */
1111                 if (uio->uio_segflg == UIO_USERSPACE &&
1112                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1113                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1114                     (fp->f_flag & FNONBLOCK) == 0) {
1115                         pipeunlock(wpipe);
1116                         error = pipe_direct_write(wpipe, uio);
1117                         if (error)
1118                                 break;
1119                         continue;
1120                 }
1121 #endif
1122
1123                 /*
1124                  * Pipe buffered writes cannot be coincidental with
1125                  * direct writes.  We wait until the currently executing
1126                  * direct write is completed before we start filling the
1127                  * pipe buffer.  We break out if a signal occurs or the
1128                  * reader goes away.
1129                  */
1130                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1131                         if (wpipe->pipe_state & PIPE_WANTR) {
1132                                 wpipe->pipe_state &= ~PIPE_WANTR;
1133                                 wakeup(wpipe);
1134                         }
1135                         pipeselwakeup(wpipe);
1136                         wpipe->pipe_state |= PIPE_WANTW;
1137                         pipeunlock(wpipe);
1138                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1139                             "pipbww", 0);
1140                         if (error)
1141                                 break;
1142                         else
1143                                 continue;
1144                 }
1145
1146                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1147
1148                 /* Writes of size <= PIPE_BUF must be atomic. */
1149                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1150                         space = 0;
1151
1152                 if (space > 0) {
1153                         int size;       /* Transfer size */
1154                         int segsize;    /* first segment to transfer */
1155
1156                         /*
1157                          * Transfer size is minimum of uio transfer
1158                          * and free space in pipe buffer.
1159                          */
1160                         if (space > uio->uio_resid)
1161                                 size = uio->uio_resid;
1162                         else
1163                                 size = space;
1164                         /*
1165                          * First segment to transfer is minimum of
1166                          * transfer size and contiguous space in
1167                          * pipe buffer.  If first segment to transfer
1168                          * is less than the transfer size, we've got
1169                          * a wraparound in the buffer.
1170                          */
1171                         segsize = wpipe->pipe_buffer.size -
1172                                 wpipe->pipe_buffer.in;
1173                         if (segsize > size)
1174                                 segsize = size;
1175
1176                         /* Transfer first segment */
1177
1178                         PIPE_UNLOCK(rpipe);
1179                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1180                                         segsize, uio);
1181                         PIPE_LOCK(rpipe);
1182
1183                         if (error == 0 && segsize < size) {
1184                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1185                                         wpipe->pipe_buffer.size,
1186                                         ("Pipe buffer wraparound disappeared"));
1187                                 /*
1188                                  * Transfer remaining part now, to
1189                                  * support atomic writes.  Wraparound
1190                                  * happened.
1191                                  */
1192
1193                                 PIPE_UNLOCK(rpipe);
1194                                 error = uiomove(
1195                                     &wpipe->pipe_buffer.buffer[0],
1196                                     size - segsize, uio);
1197                                 PIPE_LOCK(rpipe);
1198                         }
1199                         if (error == 0) {
1200                                 wpipe->pipe_buffer.in += size;
1201                                 if (wpipe->pipe_buffer.in >=
1202                                     wpipe->pipe_buffer.size) {
1203                                         KASSERT(wpipe->pipe_buffer.in ==
1204                                                 size - segsize +
1205                                                 wpipe->pipe_buffer.size,
1206                                                 ("Expected wraparound bad"));
1207                                         wpipe->pipe_buffer.in = size - segsize;
1208                                 }
1209
1210                                 wpipe->pipe_buffer.cnt += size;
1211                                 KASSERT(wpipe->pipe_buffer.cnt <=
1212                                         wpipe->pipe_buffer.size,
1213                                         ("Pipe buffer overflow"));
1214                         }
1215                         pipeunlock(wpipe);
1216                         if (error != 0)
1217                                 break;
1218                 } else {
1219                         /*
1220                          * If the "read-side" has been blocked, wake it up now.
1221                          */
1222                         if (wpipe->pipe_state & PIPE_WANTR) {
1223                                 wpipe->pipe_state &= ~PIPE_WANTR;
1224                                 wakeup(wpipe);
1225                         }
1226
1227                         /*
1228                          * don't block on non-blocking I/O
1229                          */
1230                         if (fp->f_flag & FNONBLOCK) {
1231                                 error = EAGAIN;
1232                                 pipeunlock(wpipe);
1233                                 break;
1234                         }
1235
1236                         /*
1237                          * We have no more space and have something to offer,
1238                          * wake up select/poll.
1239                          */
1240                         pipeselwakeup(wpipe);
1241
1242                         wpipe->pipe_state |= PIPE_WANTW;
1243                         pipeunlock(wpipe);
1244                         error = msleep(wpipe, PIPE_MTX(rpipe),
1245                             PRIBIO | PCATCH, "pipewr", 0);
1246                         if (error != 0)
1247                                 break;
1248                 }
1249         }
1250
1251         pipelock(wpipe, 0);
1252         --wpipe->pipe_busy;
1253
1254         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1255                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1256                 wakeup(wpipe);
1257         } else if (wpipe->pipe_buffer.cnt > 0) {
1258                 /*
1259                  * If we have put any characters in the buffer, we wake up
1260                  * the reader.
1261                  */
1262                 if (wpipe->pipe_state & PIPE_WANTR) {
1263                         wpipe->pipe_state &= ~PIPE_WANTR;
1264                         wakeup(wpipe);
1265                 }
1266         }
1267
1268         /*
1269          * Don't return EPIPE if any byte was written.
1270          * EINTR and other interrupts are handled by generic I/O layer.
1271          * Do not pretend that I/O succeeded for obvious user error
1272          * like EFAULT.
1273          */
1274         if (uio->uio_resid != orig_resid && error == EPIPE)
1275                 error = 0;
1276
1277         if (error == 0)
1278                 vfs_timestamp(&wpipe->pipe_mtime);
1279
1280         /*
1281          * We have something to offer,
1282          * wake up select/poll.
1283          */
1284         if (wpipe->pipe_buffer.cnt)
1285                 pipeselwakeup(wpipe);
1286
1287         pipeunlock(wpipe);
1288         PIPE_UNLOCK(rpipe);
1289         return (error);
1290 }
1291
1292 /* ARGSUSED */
1293 static int
1294 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1295     struct thread *td)
1296 {
1297         struct pipe *cpipe;
1298         int error;
1299
1300         cpipe = fp->f_data;
1301         if (cpipe->pipe_state & PIPE_NAMED)
1302                 error = vnops.fo_truncate(fp, length, active_cred, td);
1303         else
1304                 error = invfo_truncate(fp, length, active_cred, td);
1305         return (error);
1306 }
1307
1308 /*
1309  * we implement a very minimal set of ioctls for compatibility with sockets.
1310  */
1311 static int
1312 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1313     struct thread *td)
1314 {
1315         struct pipe *mpipe = fp->f_data;
1316         int error;
1317
1318         PIPE_LOCK(mpipe);
1319
1320 #ifdef MAC
1321         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1322         if (error) {
1323                 PIPE_UNLOCK(mpipe);
1324                 return (error);
1325         }
1326 #endif
1327
1328         error = 0;
1329         switch (cmd) {
1330
1331         case FIONBIO:
1332                 break;
1333
1334         case FIOASYNC:
1335                 if (*(int *)data) {
1336                         mpipe->pipe_state |= PIPE_ASYNC;
1337                 } else {
1338                         mpipe->pipe_state &= ~PIPE_ASYNC;
1339                 }
1340                 break;
1341
1342         case FIONREAD:
1343                 if (!(fp->f_flag & FREAD)) {
1344                         *(int *)data = 0;
1345                         PIPE_UNLOCK(mpipe);
1346                         return (0);
1347                 }
1348                 if (mpipe->pipe_state & PIPE_DIRECTW)
1349                         *(int *)data = mpipe->pipe_map.cnt;
1350                 else
1351                         *(int *)data = mpipe->pipe_buffer.cnt;
1352                 break;
1353
1354         case FIOSETOWN:
1355                 PIPE_UNLOCK(mpipe);
1356                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1357                 goto out_unlocked;
1358
1359         case FIOGETOWN:
1360                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1361                 break;
1362
1363         /* This is deprecated, FIOSETOWN should be used instead. */
1364         case TIOCSPGRP:
1365                 PIPE_UNLOCK(mpipe);
1366                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1367                 goto out_unlocked;
1368
1369         /* This is deprecated, FIOGETOWN should be used instead. */
1370         case TIOCGPGRP:
1371                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1372                 break;
1373
1374         default:
1375                 error = ENOTTY;
1376                 break;
1377         }
1378         PIPE_UNLOCK(mpipe);
1379 out_unlocked:
1380         return (error);
1381 }
1382
1383 static int
1384 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1385     struct thread *td)
1386 {
1387         struct pipe *rpipe;
1388         struct pipe *wpipe;
1389         int levents, revents;
1390 #ifdef MAC
1391         int error;
1392 #endif
1393
1394         revents = 0;
1395         rpipe = fp->f_data;
1396         wpipe = PIPE_PEER(rpipe);
1397         PIPE_LOCK(rpipe);
1398 #ifdef MAC
1399         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1400         if (error)
1401                 goto locked_error;
1402 #endif
1403         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1404                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1405                     (rpipe->pipe_buffer.cnt > 0))
1406                         revents |= events & (POLLIN | POLLRDNORM);
1407
1408         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1409                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1410                     (wpipe->pipe_state & PIPE_EOF) ||
1411                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1412                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1413                          wpipe->pipe_buffer.size == 0)))
1414                         revents |= events & (POLLOUT | POLLWRNORM);
1415
1416         levents = events &
1417             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1418         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1419             fp->f_seqcount == rpipe->pipe_wgen)
1420                 events |= POLLINIGNEOF;
1421
1422         if ((events & POLLINIGNEOF) == 0) {
1423                 if (rpipe->pipe_state & PIPE_EOF) {
1424                         revents |= (events & (POLLIN | POLLRDNORM));
1425                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1426                             (wpipe->pipe_state & PIPE_EOF))
1427                                 revents |= POLLHUP;
1428                 }
1429         }
1430
1431         if (revents == 0) {
1432                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1433                         selrecord(td, &rpipe->pipe_sel);
1434                         if (SEL_WAITING(&rpipe->pipe_sel))
1435                                 rpipe->pipe_state |= PIPE_SEL;
1436                 }
1437
1438                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1439                         selrecord(td, &wpipe->pipe_sel);
1440                         if (SEL_WAITING(&wpipe->pipe_sel))
1441                                 wpipe->pipe_state |= PIPE_SEL;
1442                 }
1443         }
1444 #ifdef MAC
1445 locked_error:
1446 #endif
1447         PIPE_UNLOCK(rpipe);
1448
1449         return (revents);
1450 }
1451
1452 /*
1453  * We shouldn't need locks here as we're doing a read and this should
1454  * be a natural race.
1455  */
1456 static int
1457 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1458     struct thread *td)
1459 {
1460         struct pipe *pipe;
1461 #ifdef MAC
1462         int error;
1463 #endif
1464
1465         pipe = fp->f_data;
1466         PIPE_LOCK(pipe);
1467 #ifdef MAC
1468         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1469         if (error) {
1470                 PIPE_UNLOCK(pipe);
1471                 return (error);
1472         }
1473 #endif
1474
1475         /* For named pipes ask the underlying filesystem. */
1476         if (pipe->pipe_state & PIPE_NAMED) {
1477                 PIPE_UNLOCK(pipe);
1478                 return (vnops.fo_stat(fp, ub, active_cred, td));
1479         }
1480
1481         PIPE_UNLOCK(pipe);
1482
1483         bzero(ub, sizeof(*ub));
1484         ub->st_mode = S_IFIFO;
1485         ub->st_blksize = PAGE_SIZE;
1486         if (pipe->pipe_state & PIPE_DIRECTW)
1487                 ub->st_size = pipe->pipe_map.cnt;
1488         else
1489                 ub->st_size = pipe->pipe_buffer.cnt;
1490         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1491         ub->st_atim = pipe->pipe_atime;
1492         ub->st_mtim = pipe->pipe_mtime;
1493         ub->st_ctim = pipe->pipe_ctime;
1494         ub->st_uid = fp->f_cred->cr_uid;
1495         ub->st_gid = fp->f_cred->cr_gid;
1496         ub->st_dev = pipedev_ino;
1497         ub->st_ino = pipe->pipe_ino;
1498         /*
1499          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1500          */
1501         return (0);
1502 }
1503
1504 /* ARGSUSED */
1505 static int
1506 pipe_close(struct file *fp, struct thread *td)
1507 {
1508
1509         if (fp->f_vnode != NULL)
1510                 return vnops.fo_close(fp, td);
1511         fp->f_ops = &badfileops;
1512         pipe_dtor(fp->f_data);
1513         fp->f_data = NULL;
1514         return (0);
1515 }
1516
1517 static int
1518 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1519 {
1520         struct pipe *cpipe;
1521         int error;
1522
1523         cpipe = fp->f_data;
1524         if (cpipe->pipe_state & PIPE_NAMED)
1525                 error = vn_chmod(fp, mode, active_cred, td);
1526         else
1527                 error = invfo_chmod(fp, mode, active_cred, td);
1528         return (error);
1529 }
1530
1531 static int
1532 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1533     struct thread *td)
1534 {
1535         struct pipe *cpipe;
1536         int error;
1537
1538         cpipe = fp->f_data;
1539         if (cpipe->pipe_state & PIPE_NAMED)
1540                 error = vn_chown(fp, uid, gid, active_cred, td);
1541         else
1542                 error = invfo_chown(fp, uid, gid, active_cred, td);
1543         return (error);
1544 }
1545
1546 static int
1547 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1548 {
1549         struct pipe *pi;
1550
1551         if (fp->f_type == DTYPE_FIFO)
1552                 return (vn_fill_kinfo(fp, kif, fdp));
1553         kif->kf_type = KF_TYPE_PIPE;
1554         pi = fp->f_data;
1555         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1556         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1557         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1558         return (0);
1559 }
1560
1561 static void
1562 pipe_free_kmem(struct pipe *cpipe)
1563 {
1564
1565         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1566             ("pipe_free_kmem: pipe mutex locked"));
1567
1568         if (cpipe->pipe_buffer.buffer != NULL) {
1569                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1570                 vm_map_remove(pipe_map,
1571                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1572                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1573                 cpipe->pipe_buffer.buffer = NULL;
1574         }
1575 #ifndef PIPE_NODIRECT
1576         {
1577                 cpipe->pipe_map.cnt = 0;
1578                 cpipe->pipe_map.pos = 0;
1579                 cpipe->pipe_map.npages = 0;
1580         }
1581 #endif
1582 }
1583
1584 /*
1585  * shutdown the pipe
1586  */
1587 static void
1588 pipeclose(struct pipe *cpipe)
1589 {
1590         struct pipepair *pp;
1591         struct pipe *ppipe;
1592
1593         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1594
1595         PIPE_LOCK(cpipe);
1596         pipelock(cpipe, 0);
1597         pp = cpipe->pipe_pair;
1598
1599         pipeselwakeup(cpipe);
1600
1601         /*
1602          * If the other side is blocked, wake it up saying that
1603          * we want to close it down.
1604          */
1605         cpipe->pipe_state |= PIPE_EOF;
1606         while (cpipe->pipe_busy) {
1607                 wakeup(cpipe);
1608                 cpipe->pipe_state |= PIPE_WANT;
1609                 pipeunlock(cpipe);
1610                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1611                 pipelock(cpipe, 0);
1612         }
1613
1614
1615         /*
1616          * Disconnect from peer, if any.
1617          */
1618         ppipe = cpipe->pipe_peer;
1619         if (ppipe->pipe_present == PIPE_ACTIVE) {
1620                 pipeselwakeup(ppipe);
1621
1622                 ppipe->pipe_state |= PIPE_EOF;
1623                 wakeup(ppipe);
1624                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1625         }
1626
1627         /*
1628          * Mark this endpoint as free.  Release kmem resources.  We
1629          * don't mark this endpoint as unused until we've finished
1630          * doing that, or the pipe might disappear out from under
1631          * us.
1632          */
1633         PIPE_UNLOCK(cpipe);
1634         pipe_free_kmem(cpipe);
1635         PIPE_LOCK(cpipe);
1636         cpipe->pipe_present = PIPE_CLOSING;
1637         pipeunlock(cpipe);
1638
1639         /*
1640          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1641          * PIPE_FINALIZED, that allows other end to free the
1642          * pipe_pair, only after the knotes are completely dismantled.
1643          */
1644         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1645         cpipe->pipe_present = PIPE_FINALIZED;
1646         seldrain(&cpipe->pipe_sel);
1647         knlist_destroy(&cpipe->pipe_sel.si_note);
1648
1649         /*
1650          * If both endpoints are now closed, release the memory for the
1651          * pipe pair.  If not, unlock.
1652          */
1653         if (ppipe->pipe_present == PIPE_FINALIZED) {
1654                 PIPE_UNLOCK(cpipe);
1655 #ifdef MAC
1656                 mac_pipe_destroy(pp);
1657 #endif
1658                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1659         } else
1660                 PIPE_UNLOCK(cpipe);
1661 }
1662
1663 /*ARGSUSED*/
1664 static int
1665 pipe_kqfilter(struct file *fp, struct knote *kn)
1666 {
1667         struct pipe *cpipe;
1668
1669         /*
1670          * If a filter is requested that is not supported by this file
1671          * descriptor, don't return an error, but also don't ever generate an
1672          * event.
1673          */
1674         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1675                 kn->kn_fop = &pipe_nfiltops;
1676                 return (0);
1677         }
1678         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1679                 kn->kn_fop = &pipe_nfiltops;
1680                 return (0);
1681         }
1682         cpipe = fp->f_data;
1683         PIPE_LOCK(cpipe);
1684         switch (kn->kn_filter) {
1685         case EVFILT_READ:
1686                 kn->kn_fop = &pipe_rfiltops;
1687                 break;
1688         case EVFILT_WRITE:
1689                 kn->kn_fop = &pipe_wfiltops;
1690                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1691                         /* other end of pipe has been closed */
1692                         PIPE_UNLOCK(cpipe);
1693                         return (EPIPE);
1694                 }
1695                 cpipe = PIPE_PEER(cpipe);
1696                 break;
1697         default:
1698                 PIPE_UNLOCK(cpipe);
1699                 return (EINVAL);
1700         }
1701
1702         kn->kn_hook = cpipe;
1703         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1704         PIPE_UNLOCK(cpipe);
1705         return (0);
1706 }
1707
1708 static void
1709 filt_pipedetach(struct knote *kn)
1710 {
1711         struct pipe *cpipe = kn->kn_hook;
1712
1713         PIPE_LOCK(cpipe);
1714         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1715         PIPE_UNLOCK(cpipe);
1716 }
1717
1718 /*ARGSUSED*/
1719 static int
1720 filt_piperead(struct knote *kn, long hint)
1721 {
1722         struct pipe *rpipe = kn->kn_hook;
1723         struct pipe *wpipe = rpipe->pipe_peer;
1724         int ret;
1725
1726         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1727         kn->kn_data = rpipe->pipe_buffer.cnt;
1728         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1729                 kn->kn_data = rpipe->pipe_map.cnt;
1730
1731         if ((rpipe->pipe_state & PIPE_EOF) ||
1732             wpipe->pipe_present != PIPE_ACTIVE ||
1733             (wpipe->pipe_state & PIPE_EOF)) {
1734                 kn->kn_flags |= EV_EOF;
1735                 return (1);
1736         }
1737         ret = kn->kn_data > 0;
1738         return ret;
1739 }
1740
1741 /*ARGSUSED*/
1742 static int
1743 filt_pipewrite(struct knote *kn, long hint)
1744 {
1745         struct pipe *wpipe;
1746
1747         wpipe = kn->kn_hook;
1748         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1749         if (wpipe->pipe_present != PIPE_ACTIVE ||
1750             (wpipe->pipe_state & PIPE_EOF)) {
1751                 kn->kn_data = 0;
1752                 kn->kn_flags |= EV_EOF;
1753                 return (1);
1754         }
1755         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1756             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1757         if (wpipe->pipe_state & PIPE_DIRECTW)
1758                 kn->kn_data = 0;
1759
1760         return (kn->kn_data >= PIPE_BUF);
1761 }
1762
1763 static void
1764 filt_pipedetach_notsup(struct knote *kn)
1765 {
1766
1767 }
1768
1769 static int
1770 filt_pipenotsup(struct knote *kn, long hint)
1771 {
1772
1773         return (0);
1774 }