sys/kern/sys_pipe.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-4-Clause
   3  *
   4  * Copyright (c) 1996 John S. Dyson
   5  * Copyright (c) 2012 Giovanni Trematerra
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice immediately at the beginning of the file, without modification,
  13  *    this list of conditions, and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Absolutely no warranty of function or purpose is made by the author
  18  *    John S. Dyson.
  19  * 4. Modifications may be freely made to this file if the above conditions
  20  *    are met.
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  36  * memory, and the receiving process copies directly from these pinned pages
  37  * in the sending process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.
  50  *
  51  * In order to limit the resource use of pipes, two sysctls exist:
  52  *
  53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  54  * address space available to us in pipe_map. This value is normally
  55  * autotuned, but may also be loader tuned.
  56  *
  57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  58  * memory in use by pipes.
  59  *
  60  * Based on how large pipekva is relative to maxpipekva, the following
  61  * will happen:
  62  *
  63  * 0% - 50%:
  64  *     New pipes are given 16K of memory backing, pipes may dynamically
  65  *     grow to as large as 64K where needed.
  66  * 50% - 75%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes may NOT grow.
  69  * 75% - 100%:
  70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  71  *     existing pipes will be shrunk down to 4K whenever possible.
  72  *
  73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  75  * resize which MUST occur for reverse-direction pipes when they are
  76  * first used.
  77  *
  78  * Additional information about the current state of pipes may be obtained
  79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  80  * and kern.ipc.piperesizefail.
  81  *
  82  * Locking rules:  There are two locks present here:  A mutex, used via
  83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  84  * the flag, as mutexes can not persist over uiomove.  The mutex
  85  * exists only to guard access to the flag, and is not in itself a
  86  * locking mechanism.  Also note that there is only a single mutex for
  87  * both directions of a pipe.
  88  *
  89  * As pipelock() may have to sleep before it can acquire the flag, it
  90  * is important to reread all data after a call to pipelock(); everything
  91  * in the structure may have changed.
  92  */
  93
  94 #include <sys/cdefs.h>
  95 __FBSDID("$FreeBSD$");
  96
  97 #include <sys/param.h>
  98 #include <sys/systm.h>
  99 #include <sys/conf.h>
 100 #include <sys/fcntl.h>
 101 #include <sys/file.h>
 102 #include <sys/filedesc.h>
 103 #include <sys/filio.h>
 104 #include <sys/kernel.h>
 105 #include <sys/lock.h>
 106 #include <sys/mutex.h>
 107 #include <sys/ttycom.h>
 108 #include <sys/stat.h>
 109 #include <sys/malloc.h>
 110 #include <sys/poll.h>
 111 #include <sys/selinfo.h>
 112 #include <sys/signalvar.h>
 113 #include <sys/syscallsubr.h>
 114 #include <sys/sysctl.h>
 115 #include <sys/sysproto.h>
 116 #include <sys/pipe.h>
 117 #include <sys/proc.h>
 118 #include <sys/vnode.h>
 119 #include <sys/uio.h>
 120 #include <sys/user.h>
 121 #include <sys/event.h>
 122
 123 #include <security/mac/mac_framework.h>
 124
 125 #include <vm/vm.h>
 126 #include <vm/vm_param.h>
 127 #include <vm/vm_object.h>
 128 #include <vm/vm_kern.h>
 129 #include <vm/vm_extern.h>
 130 #include <vm/pmap.h>
 131 #include <vm/vm_map.h>
 132 #include <vm/vm_page.h>
 133 #include <vm/uma.h>
 134
 135 /*
 136  * Use this define if you want to disable *fancy* VM things.  Expect an
 137  * approx 30% decrease in transfer rate.  This could be useful for
 138  * NetBSD or OpenBSD.
 139  */
 140 /* #define PIPE_NODIRECT */
 141
 142 #define PIPE_PEER(pipe) \
 143         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 144
 145 /*
 146  * interfaces to the outside world
 147  */
 148 static fo_rdwr_t        pipe_read;
 149 static fo_rdwr_t        pipe_write;
 150 static fo_truncate_t    pipe_truncate;
 151 static fo_ioctl_t       pipe_ioctl;
 152 static fo_poll_t        pipe_poll;
 153 static fo_kqfilter_t    pipe_kqfilter;
 154 static fo_stat_t        pipe_stat;
 155 static fo_close_t       pipe_close;
 156 static fo_chmod_t       pipe_chmod;
 157 static fo_chown_t       pipe_chown;
 158 static fo_fill_kinfo_t  pipe_fill_kinfo;
 159
 160 struct fileops pipeops = {
 161         .fo_read = pipe_read,
 162         .fo_write = pipe_write,
 163         .fo_truncate = pipe_truncate,
 164         .fo_ioctl = pipe_ioctl,
 165         .fo_poll = pipe_poll,
 166         .fo_kqfilter = pipe_kqfilter,
 167         .fo_stat = pipe_stat,
 168         .fo_close = pipe_close,
 169         .fo_chmod = pipe_chmod,
 170         .fo_chown = pipe_chown,
 171         .fo_sendfile = invfo_sendfile,
 172         .fo_fill_kinfo = pipe_fill_kinfo,
 173         .fo_flags = DFLAG_PASSABLE
 174 };
 175
 176 static void     filt_pipedetach(struct knote *kn);
 177 static void     filt_pipedetach_notsup(struct knote *kn);
 178 static int      filt_pipenotsup(struct knote *kn, long hint);
 179 static int      filt_piperead(struct knote *kn, long hint);
 180 static int      filt_pipewrite(struct knote *kn, long hint);
 181
 182 static struct filterops pipe_nfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach_notsup,
 185         .f_event = filt_pipenotsup
 186 };
 187 static struct filterops pipe_rfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_piperead
 191 };
 192 static struct filterops pipe_wfiltops = {
 193         .f_isfd = 1,
 194         .f_detach = filt_pipedetach,
 195         .f_event = filt_pipewrite
 196 };
 197
 198 /*
 199  * Default pipe buffer size(s), this can be kind-of large now because pipe
 200  * space is pageable.  The pipe code will try to maintain locality of
 201  * reference for performance reasons, so small amounts of outstanding I/O
 202  * will not wipe the cache.
 203  */
 204 #define MINPIPESIZE (PIPE_SIZE/3)
 205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 206
 207 static long amountpipekva;
 208 static int pipefragretry;
 209 static int pipeallocfail;
 210 static int piperesizefail;
 211 static int piperesizeallowed = 1;
 212
 213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 214            &maxpipekva, 0, "Pipe KVA limit");
 215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 216            &amountpipekva, 0, "Pipe KVA usage");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 218           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 220           &pipeallocfail, 0, "Pipe allocation failures");
 221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 222           &piperesizefail, 0, "Pipe resize failures");
 223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 224           &piperesizeallowed, 0, "Pipe resizing allowed");
 225
 226 static void pipeinit(void *dummy __unused);
 227 static void pipeclose(struct pipe *cpipe);
 228 static void pipe_free_kmem(struct pipe *cpipe);
 229 static void pipe_create(struct pipe *pipe, int backing);
 230 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 231 static __inline int pipelock(struct pipe *cpipe, int catch);
 232 static __inline void pipeunlock(struct pipe *cpipe);
 233 #ifndef PIPE_NODIRECT
 234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 237 static void pipe_clone_write_buffer(struct pipe *wpipe);
 238 #endif
 239 static int pipespace(struct pipe *cpipe, int size);
 240 static int pipespace_new(struct pipe *cpipe, int size);
 241
 242 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 243 static int      pipe_zone_init(void *mem, int size, int flags);
 244 static void     pipe_zone_fini(void *mem, int size);
 245
 246 static uma_zone_t pipe_zone;
 247 static struct unrhdr *pipeino_unr;
 248 static dev_t pipedev_ino;
 249
 250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 251
 252 static void
 253 pipeinit(void *dummy __unused)
 254 {
 255
 256         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 257             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 258             UMA_ALIGN_PTR, 0);
 259         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 260         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 261         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 262         pipedev_ino = devfs_alloc_cdp_inode();
 263         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 264 }
 265
 266 static int
 267 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 268 {
 269         struct pipepair *pp;
 270         struct pipe *rpipe, *wpipe;
 271
 272         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 273
 274         pp = (struct pipepair *)mem;
 275
 276         /*
 277          * We zero both pipe endpoints to make sure all the kmem pointers
 278          * are NULL, flag fields are zero'd, etc.  We timestamp both
 279          * endpoints with the same time.
 280          */
 281         rpipe = &pp->pp_rpipe;
 282         bzero(rpipe, sizeof(*rpipe));
 283         vfs_timestamp(&rpipe->pipe_ctime);
 284         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 285
 286         wpipe = &pp->pp_wpipe;
 287         bzero(wpipe, sizeof(*wpipe));
 288         wpipe->pipe_ctime = rpipe->pipe_ctime;
 289         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 290
 291         rpipe->pipe_peer = wpipe;
 292         rpipe->pipe_pair = pp;
 293         wpipe->pipe_peer = rpipe;
 294         wpipe->pipe_pair = pp;
 295
 296         /*
 297          * Mark both endpoints as present; they will later get free'd
 298          * one at a time.  When both are free'd, then the whole pair
 299          * is released.
 300          */
 301         rpipe->pipe_present = PIPE_ACTIVE;
 302         wpipe->pipe_present = PIPE_ACTIVE;
 303
 304         /*
 305          * Eventually, the MAC Framework may initialize the label
 306          * in ctor or init, but for now we do it elswhere to avoid
 307          * blocking in ctor or init.
 308          */
 309         pp->pp_label = NULL;
 310
 311         return (0);
 312 }
 313
 314 static int
 315 pipe_zone_init(void *mem, int size, int flags)
 316 {
 317         struct pipepair *pp;
 318
 319         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 320
 321         pp = (struct pipepair *)mem;
 322
 323         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 324         return (0);
 325 }
 326
 327 static void
 328 pipe_zone_fini(void *mem, int size)
 329 {
 330         struct pipepair *pp;
 331
 332         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 333
 334         pp = (struct pipepair *)mem;
 335
 336         mtx_destroy(&pp->pp_mtx);
 337 }
 338
 339 static void
 340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 341 {
 342         struct pipepair *pp;
 343         struct pipe *rpipe, *wpipe;
 344
 345         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 346 #ifdef MAC
 347         /*
 348          * The MAC label is shared between the connected endpoints.  As a
 349          * result mac_pipe_init() and mac_pipe_create() are called once
 350          * for the pair, and not on the endpoints.
 351          */
 352         mac_pipe_init(pp);
 353         mac_pipe_create(td->td_ucred, pp);
 354 #endif
 355         rpipe = &pp->pp_rpipe;
 356         wpipe = &pp->pp_wpipe;
 357
 358         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 359         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 360
 361         /* Only the forward direction pipe is backed by default */
 362         pipe_create(rpipe, 1);
 363         pipe_create(wpipe, 0);
 364
 365         rpipe->pipe_state |= PIPE_DIRECTOK;
 366         wpipe->pipe_state |= PIPE_DIRECTOK;
 367 }
 368
 369 void
 370 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 371 {
 372         struct pipepair *pp;
 373
 374         pipe_paircreate(td, &pp);
 375         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 376         *ppipe = &pp->pp_rpipe;
 377 }
 378
 379 void
 380 pipe_dtor(struct pipe *dpipe)
 381 {
 382         struct pipe *peer;
 383         ino_t ino;
 384
 385         ino = dpipe->pipe_ino;
 386         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 387         funsetown(&dpipe->pipe_sigio);
 388         pipeclose(dpipe);
 389         if (peer != NULL) {
 390                 funsetown(&peer->pipe_sigio);
 391                 pipeclose(peer);
 392         }
 393         if (ino != 0 && ino != (ino_t)-1)
 394                 free_unr(pipeino_unr, ino);
 395 }
 396
 397 /*
 398  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 399  * the zone pick up the pieces via pipeclose().
 400  */
 401 int
 402 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 403     struct filecaps *fcaps2)
 404 {
 405         struct file *rf, *wf;
 406         struct pipe *rpipe, *wpipe;
 407         struct pipepair *pp;
 408         int fd, fflags, error;
 409
 410         pipe_paircreate(td, &pp);
 411         rpipe = &pp->pp_rpipe;
 412         wpipe = &pp->pp_wpipe;
 413         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 414         if (error) {
 415                 pipeclose(rpipe);
 416                 pipeclose(wpipe);
 417                 return (error);
 418         }
 419         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 420         fildes[0] = fd;
 421
 422         fflags = FREAD | FWRITE;
 423         if ((flags & O_NONBLOCK) != 0)
 424                 fflags |= FNONBLOCK;
 425
 426         /*
 427          * Warning: once we've gotten past allocation of the fd for the
 428          * read-side, we can only drop the read side via fdrop() in order
 429          * to avoid races against processes which manage to dup() the read
 430          * side while we are blocked trying to allocate the write side.
 431          */
 432         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 433         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 434         if (error) {
 435                 fdclose(td, rf, fildes[0]);
 436                 fdrop(rf, td);
 437                 /* rpipe has been closed by fdrop(). */
 438                 pipeclose(wpipe);
 439                 return (error);
 440         }
 441         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 442         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 443         fdrop(wf, td);
 444         fildes[1] = fd;
 445         fdrop(rf, td);
 446
 447         return (0);
 448 }
 449
 450 #ifdef COMPAT_FREEBSD10
 451 /* ARGSUSED */
 452 int
 453 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 454 {
 455         int error;
 456         int fildes[2];
 457
 458         error = kern_pipe(td, fildes, 0, NULL, NULL);
 459         if (error)
 460                 return (error);
 461
 462         td->td_retval[0] = fildes[0];
 463         td->td_retval[1] = fildes[1];
 464
 465         return (0);
 466 }
 467 #endif
 468
 469 int
 470 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 471 {
 472         int error, fildes[2];
 473
 474         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 475                 return (EINVAL);
 476         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 477         if (error)
 478                 return (error);
 479         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 480         if (error) {
 481                 (void)kern_close(td, fildes[0]);
 482                 (void)kern_close(td, fildes[1]);
 483         }
 484         return (error);
 485 }
 486
 487 /*
 488  * Allocate kva for pipe circular buffer, the space is pageable
 489  * This routine will 'realloc' the size of a pipe safely, if it fails
 490  * it will retain the old buffer.
 491  * If it fails it will return ENOMEM.
 492  */
 493 static int
 494 pipespace_new(struct pipe *cpipe, int size)
 495 {
 496         caddr_t buffer;
 497         int error, cnt, firstseg;
 498         static int curfail = 0;
 499         static struct timeval lastfail;
 500
 501         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 502         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 503                 ("pipespace: resize of direct writes not allowed"));
 504 retry:
 505         cnt = cpipe->pipe_buffer.cnt;
 506         if (cnt > size)
 507                 size = cnt;
 508
 509         size = round_page(size);
 510         buffer = (caddr_t) vm_map_min(pipe_map);
 511
 512         error = vm_map_find(pipe_map, NULL, 0,
 513                 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
 514                 VM_PROT_ALL, VM_PROT_ALL, 0);
 515         if (error != KERN_SUCCESS) {
 516                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 517                         (size > SMALL_PIPE_SIZE)) {
 518                         size = SMALL_PIPE_SIZE;
 519                         pipefragretry++;
 520                         goto retry;
 521                 }
 522                 if (cpipe->pipe_buffer.buffer == NULL) {
 523                         pipeallocfail++;
 524                         if (ppsratecheck(&lastfail, &curfail, 1))
 525                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 526                 } else {
 527                         piperesizefail++;
 528                 }
 529                 return (ENOMEM);
 530         }
 531
 532         /* copy data, then free old resources if we're resizing */
 533         if (cnt > 0) {
 534                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 535                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 536                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 537                                 buffer, firstseg);
 538                         if ((cnt - firstseg) > 0)
 539                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 540                                         cpipe->pipe_buffer.in);
 541                 } else {
 542                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 543                                 buffer, cnt);
 544                 }
 545         }
 546         pipe_free_kmem(cpipe);
 547         cpipe->pipe_buffer.buffer = buffer;
 548         cpipe->pipe_buffer.size = size;
 549         cpipe->pipe_buffer.in = cnt;
 550         cpipe->pipe_buffer.out = 0;
 551         cpipe->pipe_buffer.cnt = cnt;
 552         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 553         return (0);
 554 }
 555
 556 /*
 557  * Wrapper for pipespace_new() that performs locking assertions.
 558  */
 559 static int
 560 pipespace(struct pipe *cpipe, int size)
 561 {
 562
 563         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 564                 ("Unlocked pipe passed to pipespace"));
 565         return (pipespace_new(cpipe, size));
 566 }
 567
 568 /*
 569  * lock a pipe for I/O, blocking other access
 570  */
 571 static __inline int
 572 pipelock(struct pipe *cpipe, int catch)
 573 {
 574         int error;
 575
 576         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 577         while (cpipe->pipe_state & PIPE_LOCKFL) {
 578                 cpipe->pipe_state |= PIPE_LWANT;
 579                 error = msleep(cpipe, PIPE_MTX(cpipe),
 580                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 581                     "pipelk", 0);
 582                 if (error != 0)
 583                         return (error);
 584         }
 585         cpipe->pipe_state |= PIPE_LOCKFL;
 586         return (0);
 587 }
 588
 589 /*
 590  * unlock a pipe I/O lock
 591  */
 592 static __inline void
 593 pipeunlock(struct pipe *cpipe)
 594 {
 595
 596         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 597         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 598                 ("Unlocked pipe passed to pipeunlock"));
 599         cpipe->pipe_state &= ~PIPE_LOCKFL;
 600         if (cpipe->pipe_state & PIPE_LWANT) {
 601                 cpipe->pipe_state &= ~PIPE_LWANT;
 602                 wakeup(cpipe);
 603         }
 604 }
 605
 606 void
 607 pipeselwakeup(struct pipe *cpipe)
 608 {
 609
 610         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 611         if (cpipe->pipe_state & PIPE_SEL) {
 612                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 613                 if (!SEL_WAITING(&cpipe->pipe_sel))
 614                         cpipe->pipe_state &= ~PIPE_SEL;
 615         }
 616         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 617                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 618         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 619 }
 620
 621 /*
 622  * Initialize and allocate VM and memory for pipe.  The structure
 623  * will start out zero'd from the ctor, so we just manage the kmem.
 624  */
 625 static void
 626 pipe_create(struct pipe *pipe, int backing)
 627 {
 628
 629         if (backing) {
 630                 /*
 631                  * Note that these functions can fail if pipe map is exhausted
 632                  * (as a result of too many pipes created), but we ignore the
 633                  * error as it is not fatal and could be provoked by
 634                  * unprivileged users. The only consequence is worse performance
 635                  * with given pipe.
 636                  */
 637                 if (amountpipekva > maxpipekva / 2)
 638                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 639                 else
 640                         (void)pipespace_new(pipe, PIPE_SIZE);
 641         }
 642
 643         pipe->pipe_ino = -1;
 644 }
 645
 646 /* ARGSUSED */
 647 static int
 648 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 649     int flags, struct thread *td)
 650 {
 651         struct pipe *rpipe;
 652         int error;
 653         int nread = 0;
 654         int size;
 655
 656         rpipe = fp->f_data;
 657         PIPE_LOCK(rpipe);
 658         ++rpipe->pipe_busy;
 659         error = pipelock(rpipe, 1);
 660         if (error)
 661                 goto unlocked_error;
 662
 663 #ifdef MAC
 664         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 665         if (error)
 666                 goto locked_error;
 667 #endif
 668         if (amountpipekva > (3 * maxpipekva) / 4) {
 669                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 670                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 671                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 672                         (piperesizeallowed == 1)) {
 673                         PIPE_UNLOCK(rpipe);
 674                         pipespace(rpipe, SMALL_PIPE_SIZE);
 675                         PIPE_LOCK(rpipe);
 676                 }
 677         }
 678
 679         while (uio->uio_resid) {
 680                 /*
 681                  * normal pipe buffer receive
 682                  */
 683                 if (rpipe->pipe_buffer.cnt > 0) {
 684                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 685                         if (size > rpipe->pipe_buffer.cnt)
 686                                 size = rpipe->pipe_buffer.cnt;
 687                         if (size > uio->uio_resid)
 688                                 size = uio->uio_resid;
 689
 690                         PIPE_UNLOCK(rpipe);
 691                         error = uiomove(
 692                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 693                             size, uio);
 694                         PIPE_LOCK(rpipe);
 695                         if (error)
 696                                 break;
 697
 698                         rpipe->pipe_buffer.out += size;
 699                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 700                                 rpipe->pipe_buffer.out = 0;
 701
 702                         rpipe->pipe_buffer.cnt -= size;
 703
 704                         /*
 705                          * If there is no more to read in the pipe, reset
 706                          * its pointers to the beginning.  This improves
 707                          * cache hit stats.
 708                          */
 709                         if (rpipe->pipe_buffer.cnt == 0) {
 710                                 rpipe->pipe_buffer.in = 0;
 711                                 rpipe->pipe_buffer.out = 0;
 712                         }
 713                         nread += size;
 714 #ifndef PIPE_NODIRECT
 715                 /*
 716                  * Direct copy, bypassing a kernel buffer.
 717                  */
 718                 } else if ((size = rpipe->pipe_map.cnt) &&
 719                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 720                         if (size > uio->uio_resid)
 721                                 size = (u_int) uio->uio_resid;
 722
 723                         PIPE_UNLOCK(rpipe);
 724                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 725                             rpipe->pipe_map.pos, size, uio);
 726                         PIPE_LOCK(rpipe);
 727                         if (error)
 728                                 break;
 729                         nread += size;
 730                         rpipe->pipe_map.pos += size;
 731                         rpipe->pipe_map.cnt -= size;
 732                         if (rpipe->pipe_map.cnt == 0) {
 733                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 734                                 wakeup(rpipe);
 735                         }
 736 #endif
 737                 } else {
 738                         /*
 739                          * detect EOF condition
 740                          * read returns 0 on EOF, no need to set error
 741                          */
 742                         if (rpipe->pipe_state & PIPE_EOF)
 743                                 break;
 744
 745                         /*
 746                          * If the "write-side" has been blocked, wake it up now.
 747                          */
 748                         if (rpipe->pipe_state & PIPE_WANTW) {
 749                                 rpipe->pipe_state &= ~PIPE_WANTW;
 750                                 wakeup(rpipe);
 751                         }
 752
 753                         /*
 754                          * Break if some data was read.
 755                          */
 756                         if (nread > 0)
 757                                 break;
 758
 759                         /*
 760                          * Unlock the pipe buffer for our remaining processing.
 761                          * We will either break out with an error or we will
 762                          * sleep and relock to loop.
 763                          */
 764                         pipeunlock(rpipe);
 765
 766                         /*
 767                          * Handle non-blocking mode operation or
 768                          * wait for more data.
 769                          */
 770                         if (fp->f_flag & FNONBLOCK) {
 771                                 error = EAGAIN;
 772                         } else {
 773                                 rpipe->pipe_state |= PIPE_WANTR;
 774                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 775                                     PRIBIO | PCATCH,
 776                                     "piperd", 0)) == 0)
 777                                         error = pipelock(rpipe, 1);
 778                         }
 779                         if (error)
 780                                 goto unlocked_error;
 781                 }
 782         }
 783 #ifdef MAC
 784 locked_error:
 785 #endif
 786         pipeunlock(rpipe);
 787
 788         /* XXX: should probably do this before getting any locks. */
 789         if (error == 0)
 790                 vfs_timestamp(&rpipe->pipe_atime);
 791 unlocked_error:
 792         --rpipe->pipe_busy;
 793
 794         /*
 795          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 796          */
 797         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 798                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 799                 wakeup(rpipe);
 800         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 801                 /*
 802                  * Handle write blocking hysteresis.
 803                  */
 804                 if (rpipe->pipe_state & PIPE_WANTW) {
 805                         rpipe->pipe_state &= ~PIPE_WANTW;
 806                         wakeup(rpipe);
 807                 }
 808         }
 809
 810         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 811                 pipeselwakeup(rpipe);
 812
 813         PIPE_UNLOCK(rpipe);
 814         return (error);
 815 }
 816
 817 #ifndef PIPE_NODIRECT
 818 /*
 819  * Map the sending processes' buffer into kernel space and wire it.
 820  * This is similar to a physical write operation.
 821  */
 822 static int
 823 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 824 {
 825         u_int size;
 826         int i;
 827
 828         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 829         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 830                 ("Clone attempt on non-direct write pipe!"));
 831
 832         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 833                 size = wpipe->pipe_buffer.size;
 834         else
 835                 size = uio->uio_iov->iov_len;
 836
 837         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 838             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 839             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 840                 return (EFAULT);
 841
 842 /*
 843  * set up the control block
 844  */
 845         wpipe->pipe_map.npages = i;
 846         wpipe->pipe_map.pos =
 847             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 848         wpipe->pipe_map.cnt = size;
 849
 850 /*
 851  * and update the uio data
 852  */
 853
 854         uio->uio_iov->iov_len -= size;
 855         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 856         if (uio->uio_iov->iov_len == 0)
 857                 uio->uio_iov++;
 858         uio->uio_resid -= size;
 859         uio->uio_offset += size;
 860         return (0);
 861 }
 862
 863 /*
 864  * unmap and unwire the process buffer
 865  */
 866 static void
 867 pipe_destroy_write_buffer(struct pipe *wpipe)
 868 {
 869
 870         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 871         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 872         wpipe->pipe_map.npages = 0;
 873 }
 874
 875 /*
 876  * In the case of a signal, the writing process might go away.  This
 877  * code copies the data into the circular buffer so that the source
 878  * pages can be freed without loss of data.
 879  */
 880 static void
 881 pipe_clone_write_buffer(struct pipe *wpipe)
 882 {
 883         struct uio uio;
 884         struct iovec iov;
 885         int size;
 886         int pos;
 887
 888         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 889         size = wpipe->pipe_map.cnt;
 890         pos = wpipe->pipe_map.pos;
 891
 892         wpipe->pipe_buffer.in = size;
 893         wpipe->pipe_buffer.out = 0;
 894         wpipe->pipe_buffer.cnt = size;
 895         wpipe->pipe_state &= ~PIPE_DIRECTW;
 896
 897         PIPE_UNLOCK(wpipe);
 898         iov.iov_base = wpipe->pipe_buffer.buffer;
 899         iov.iov_len = size;
 900         uio.uio_iov = &iov;
 901         uio.uio_iovcnt = 1;
 902         uio.uio_offset = 0;
 903         uio.uio_resid = size;
 904         uio.uio_segflg = UIO_SYSSPACE;
 905         uio.uio_rw = UIO_READ;
 906         uio.uio_td = curthread;
 907         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 908         PIPE_LOCK(wpipe);
 909         pipe_destroy_write_buffer(wpipe);
 910 }
 911
 912 /*
 913  * This implements the pipe buffer write mechanism.  Note that only
 914  * a direct write OR a normal pipe write can be pending at any given time.
 915  * If there are any characters in the pipe buffer, the direct write will
 916  * be deferred until the receiving process grabs all of the bytes from
 917  * the pipe buffer.  Then the direct mapping write is set-up.
 918  */
 919 static int
 920 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 921 {
 922         int error;
 923
 924 retry:
 925         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 926         error = pipelock(wpipe, 1);
 927         if (error != 0)
 928                 goto error1;
 929         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 930                 error = EPIPE;
 931                 pipeunlock(wpipe);
 932                 goto error1;
 933         }
 934         while (wpipe->pipe_state & PIPE_DIRECTW) {
 935                 if (wpipe->pipe_state & PIPE_WANTR) {
 936                         wpipe->pipe_state &= ~PIPE_WANTR;
 937                         wakeup(wpipe);
 938                 }
 939                 pipeselwakeup(wpipe);
 940                 wpipe->pipe_state |= PIPE_WANTW;
 941                 pipeunlock(wpipe);
 942                 error = msleep(wpipe, PIPE_MTX(wpipe),
 943                     PRIBIO | PCATCH, "pipdww", 0);
 944                 if (error)
 945                         goto error1;
 946                 else
 947                         goto retry;
 948         }
 949         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 950         if (wpipe->pipe_buffer.cnt > 0) {
 951                 if (wpipe->pipe_state & PIPE_WANTR) {
 952                         wpipe->pipe_state &= ~PIPE_WANTR;
 953                         wakeup(wpipe);
 954                 }
 955                 pipeselwakeup(wpipe);
 956                 wpipe->pipe_state |= PIPE_WANTW;
 957                 pipeunlock(wpipe);
 958                 error = msleep(wpipe, PIPE_MTX(wpipe),
 959                     PRIBIO | PCATCH, "pipdwc", 0);
 960                 if (error)
 961                         goto error1;
 962                 else
 963                         goto retry;
 964         }
 965
 966         wpipe->pipe_state |= PIPE_DIRECTW;
 967
 968         PIPE_UNLOCK(wpipe);
 969         error = pipe_build_write_buffer(wpipe, uio);
 970         PIPE_LOCK(wpipe);
 971         if (error) {
 972                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 973                 pipeunlock(wpipe);
 974                 goto error1;
 975         }
 976
 977         error = 0;
 978         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 979                 if (wpipe->pipe_state & PIPE_EOF) {
 980                         pipe_destroy_write_buffer(wpipe);
 981                         pipeselwakeup(wpipe);
 982                         pipeunlock(wpipe);
 983                         error = EPIPE;
 984                         goto error1;
 985                 }
 986                 if (wpipe->pipe_state & PIPE_WANTR) {
 987                         wpipe->pipe_state &= ~PIPE_WANTR;
 988                         wakeup(wpipe);
 989                 }
 990                 pipeselwakeup(wpipe);
 991                 wpipe->pipe_state |= PIPE_WANTW;
 992                 pipeunlock(wpipe);
 993                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 994                     "pipdwt", 0);
 995                 pipelock(wpipe, 0);
 996         }
 997
 998         if (wpipe->pipe_state & PIPE_EOF)
 999                 error = EPIPE;
1000         if (wpipe->pipe_state & PIPE_DIRECTW) {
1001                 /*
1002                  * this bit of trickery substitutes a kernel buffer for
1003                  * the process that might be going away.
1004                  */
1005                 pipe_clone_write_buffer(wpipe);
1006         } else {
1007                 pipe_destroy_write_buffer(wpipe);
1008         }
1009         pipeunlock(wpipe);
1010         return (error);
1011
1012 error1:
1013         wakeup(wpipe);
1014         return (error);
1015 }
1016 #endif
1017
1018 static int
1019 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1020     int flags, struct thread *td)
1021 {
1022         int error = 0;
1023         int desiredsize;
1024         ssize_t orig_resid;
1025         struct pipe *wpipe, *rpipe;
1026
1027         rpipe = fp->f_data;
1028         wpipe = PIPE_PEER(rpipe);
1029         PIPE_LOCK(rpipe);
1030         error = pipelock(wpipe, 1);
1031         if (error) {
1032                 PIPE_UNLOCK(rpipe);
1033                 return (error);
1034         }
1035         /*
1036          * detect loss of pipe read side, issue SIGPIPE if lost.
1037          */
1038         if (wpipe->pipe_present != PIPE_ACTIVE ||
1039             (wpipe->pipe_state & PIPE_EOF)) {
1040                 pipeunlock(wpipe);
1041                 PIPE_UNLOCK(rpipe);
1042                 return (EPIPE);
1043         }
1044 #ifdef MAC
1045         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1046         if (error) {
1047                 pipeunlock(wpipe);
1048                 PIPE_UNLOCK(rpipe);
1049                 return (error);
1050         }
1051 #endif
1052         ++wpipe->pipe_busy;
1053
1054         /* Choose a larger size if it's advantageous */
1055         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1056         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1057                 if (piperesizeallowed != 1)
1058                         break;
1059                 if (amountpipekva > maxpipekva / 2)
1060                         break;
1061                 if (desiredsize == BIG_PIPE_SIZE)
1062                         break;
1063                 desiredsize = desiredsize * 2;
1064         }
1065
1066         /* Choose a smaller size if we're in a OOM situation */
1067         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1068                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1069                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1070                 (piperesizeallowed == 1))
1071                 desiredsize = SMALL_PIPE_SIZE;
1072
1073         /* Resize if the above determined that a new size was necessary */
1074         if ((desiredsize != wpipe->pipe_buffer.size) &&
1075                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1076                 PIPE_UNLOCK(wpipe);
1077                 pipespace(wpipe, desiredsize);
1078                 PIPE_LOCK(wpipe);
1079         }
1080         if (wpipe->pipe_buffer.size == 0) {
1081                 /*
1082                  * This can only happen for reverse direction use of pipes
1083                  * in a complete OOM situation.
1084                  */
1085                 error = ENOMEM;
1086                 --wpipe->pipe_busy;
1087                 pipeunlock(wpipe);
1088                 PIPE_UNLOCK(wpipe);
1089                 return (error);
1090         }
1091
1092         pipeunlock(wpipe);
1093
1094         orig_resid = uio->uio_resid;
1095
1096         while (uio->uio_resid) {
1097                 int space;
1098
1099                 pipelock(wpipe, 0);
1100                 if (wpipe->pipe_state & PIPE_EOF) {
1101                         pipeunlock(wpipe);
1102                         error = EPIPE;
1103                         break;
1104                 }
1105 #ifndef PIPE_NODIRECT
1106                 /*
1107                  * If the transfer is large, we can gain performance if
1108                  * we do process-to-process copies directly.
1109                  * If the write is non-blocking, we don't use the
1110                  * direct write mechanism.
1111                  *
1112                  * The direct write mechanism will detect the reader going
1113                  * away on us.
1114                  */
1115                 if (uio->uio_segflg == UIO_USERSPACE &&
1116                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1117                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1118                     (fp->f_flag & FNONBLOCK) == 0) {
1119                         pipeunlock(wpipe);
1120                         error = pipe_direct_write(wpipe, uio);
1121                         if (error)
1122                                 break;
1123                         continue;
1124                 }
1125 #endif
1126
1127                 /*
1128                  * Pipe buffered writes cannot be coincidental with
1129                  * direct writes.  We wait until the currently executing
1130                  * direct write is completed before we start filling the
1131                  * pipe buffer.  We break out if a signal occurs or the
1132                  * reader goes away.
1133                  */
1134                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1135                         if (wpipe->pipe_state & PIPE_WANTR) {
1136                                 wpipe->pipe_state &= ~PIPE_WANTR;
1137                                 wakeup(wpipe);
1138                         }
1139                         pipeselwakeup(wpipe);
1140                         wpipe->pipe_state |= PIPE_WANTW;
1141                         pipeunlock(wpipe);
1142                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1143                             "pipbww", 0);
1144                         if (error)
1145                                 break;
1146                         else
1147                                 continue;
1148                 }
1149
1150                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1151
1152                 /* Writes of size <= PIPE_BUF must be atomic. */
1153                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1154                         space = 0;
1155
1156                 if (space > 0) {
1157                         int size;       /* Transfer size */
1158                         int segsize;    /* first segment to transfer */
1159
1160                         /*
1161                          * Transfer size is minimum of uio transfer
1162                          * and free space in pipe buffer.
1163                          */
1164                         if (space > uio->uio_resid)
1165                                 size = uio->uio_resid;
1166                         else
1167                                 size = space;
1168                         /*
1169                          * First segment to transfer is minimum of
1170                          * transfer size and contiguous space in
1171                          * pipe buffer.  If first segment to transfer
1172                          * is less than the transfer size, we've got
1173                          * a wraparound in the buffer.
1174                          */
1175                         segsize = wpipe->pipe_buffer.size -
1176                                 wpipe->pipe_buffer.in;
1177                         if (segsize > size)
1178                                 segsize = size;
1179
1180                         /* Transfer first segment */
1181
1182                         PIPE_UNLOCK(rpipe);
1183                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1184                                         segsize, uio);
1185                         PIPE_LOCK(rpipe);
1186
1187                         if (error == 0 && segsize < size) {
1188                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1189                                         wpipe->pipe_buffer.size,
1190                                         ("Pipe buffer wraparound disappeared"));
1191                                 /*
1192                                  * Transfer remaining part now, to
1193                                  * support atomic writes.  Wraparound
1194                                  * happened.
1195                                  */
1196
1197                                 PIPE_UNLOCK(rpipe);
1198                                 error = uiomove(
1199                                     &wpipe->pipe_buffer.buffer[0],
1200                                     size - segsize, uio);
1201                                 PIPE_LOCK(rpipe);
1202                         }
1203                         if (error == 0) {
1204                                 wpipe->pipe_buffer.in += size;
1205                                 if (wpipe->pipe_buffer.in >=
1206                                     wpipe->pipe_buffer.size) {
1207                                         KASSERT(wpipe->pipe_buffer.in ==
1208                                                 size - segsize +
1209                                                 wpipe->pipe_buffer.size,
1210                                                 ("Expected wraparound bad"));
1211                                         wpipe->pipe_buffer.in = size - segsize;
1212                                 }
1213
1214                                 wpipe->pipe_buffer.cnt += size;
1215                                 KASSERT(wpipe->pipe_buffer.cnt <=
1216                                         wpipe->pipe_buffer.size,
1217                                         ("Pipe buffer overflow"));
1218                         }
1219                         pipeunlock(wpipe);
1220                         if (error != 0)
1221                                 break;
1222                 } else {
1223                         /*
1224                          * If the "read-side" has been blocked, wake it up now.
1225                          */
1226                         if (wpipe->pipe_state & PIPE_WANTR) {
1227                                 wpipe->pipe_state &= ~PIPE_WANTR;
1228                                 wakeup(wpipe);
1229                         }
1230
1231                         /*
1232                          * don't block on non-blocking I/O
1233                          */
1234                         if (fp->f_flag & FNONBLOCK) {
1235                                 error = EAGAIN;
1236                                 pipeunlock(wpipe);
1237                                 break;
1238                         }
1239
1240                         /*
1241                          * We have no more space and have something to offer,
1242                          * wake up select/poll.
1243                          */
1244                         pipeselwakeup(wpipe);
1245
1246                         wpipe->pipe_state |= PIPE_WANTW;
1247                         pipeunlock(wpipe);
1248                         error = msleep(wpipe, PIPE_MTX(rpipe),
1249                             PRIBIO | PCATCH, "pipewr", 0);
1250                         if (error != 0)
1251                                 break;
1252                 }
1253         }
1254
1255         pipelock(wpipe, 0);
1256         --wpipe->pipe_busy;
1257
1258         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1259                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1260                 wakeup(wpipe);
1261         } else if (wpipe->pipe_buffer.cnt > 0) {
1262                 /*
1263                  * If we have put any characters in the buffer, we wake up
1264                  * the reader.
1265                  */
1266                 if (wpipe->pipe_state & PIPE_WANTR) {
1267                         wpipe->pipe_state &= ~PIPE_WANTR;
1268                         wakeup(wpipe);
1269                 }
1270         }
1271
1272         /*
1273          * Don't return EPIPE if any byte was written.
1274          * EINTR and other interrupts are handled by generic I/O layer.
1275          * Do not pretend that I/O succeeded for obvious user error
1276          * like EFAULT.
1277          */
1278         if (uio->uio_resid != orig_resid && error == EPIPE)
1279                 error = 0;
1280
1281         if (error == 0)
1282                 vfs_timestamp(&wpipe->pipe_mtime);
1283
1284         /*
1285          * We have something to offer,
1286          * wake up select/poll.
1287          */
1288         if (wpipe->pipe_buffer.cnt)
1289                 pipeselwakeup(wpipe);
1290
1291         pipeunlock(wpipe);
1292         PIPE_UNLOCK(rpipe);
1293         return (error);
1294 }
1295
1296 /* ARGSUSED */
1297 static int
1298 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1299     struct thread *td)
1300 {
1301         struct pipe *cpipe;
1302         int error;
1303
1304         cpipe = fp->f_data;
1305         if (cpipe->pipe_state & PIPE_NAMED)
1306                 error = vnops.fo_truncate(fp, length, active_cred, td);
1307         else
1308                 error = invfo_truncate(fp, length, active_cred, td);
1309         return (error);
1310 }
1311
1312 /*
1313  * we implement a very minimal set of ioctls for compatibility with sockets.
1314  */
1315 static int
1316 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1317     struct thread *td)
1318 {
1319         struct pipe *mpipe = fp->f_data;
1320         int error;
1321
1322         PIPE_LOCK(mpipe);
1323
1324 #ifdef MAC
1325         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1326         if (error) {
1327                 PIPE_UNLOCK(mpipe);
1328                 return (error);
1329         }
1330 #endif
1331
1332         error = 0;
1333         switch (cmd) {
1334
1335         case FIONBIO:
1336                 break;
1337
1338         case FIOASYNC:
1339                 if (*(int *)data) {
1340                         mpipe->pipe_state |= PIPE_ASYNC;
1341                 } else {
1342                         mpipe->pipe_state &= ~PIPE_ASYNC;
1343                 }
1344                 break;
1345
1346         case FIONREAD:
1347                 if (!(fp->f_flag & FREAD)) {
1348                         *(int *)data = 0;
1349                         PIPE_UNLOCK(mpipe);
1350                         return (0);
1351                 }
1352                 if (mpipe->pipe_state & PIPE_DIRECTW)
1353                         *(int *)data = mpipe->pipe_map.cnt;
1354                 else
1355                         *(int *)data = mpipe->pipe_buffer.cnt;
1356                 break;
1357
1358         case FIOSETOWN:
1359                 PIPE_UNLOCK(mpipe);
1360                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1361                 goto out_unlocked;
1362
1363         case FIOGETOWN:
1364                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1365                 break;
1366
1367         /* This is deprecated, FIOSETOWN should be used instead. */
1368         case TIOCSPGRP:
1369                 PIPE_UNLOCK(mpipe);
1370                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1371                 goto out_unlocked;
1372
1373         /* This is deprecated, FIOGETOWN should be used instead. */
1374         case TIOCGPGRP:
1375                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1376                 break;
1377
1378         default:
1379                 error = ENOTTY;
1380                 break;
1381         }
1382         PIPE_UNLOCK(mpipe);
1383 out_unlocked:
1384         return (error);
1385 }
1386
1387 static int
1388 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1389     struct thread *td)
1390 {
1391         struct pipe *rpipe;
1392         struct pipe *wpipe;
1393         int levents, revents;
1394 #ifdef MAC
1395         int error;
1396 #endif
1397
1398         revents = 0;
1399         rpipe = fp->f_data;
1400         wpipe = PIPE_PEER(rpipe);
1401         PIPE_LOCK(rpipe);
1402 #ifdef MAC
1403         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1404         if (error)
1405                 goto locked_error;
1406 #endif
1407         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1408                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1409                     (rpipe->pipe_buffer.cnt > 0))
1410                         revents |= events & (POLLIN | POLLRDNORM);
1411
1412         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1413                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1414                     (wpipe->pipe_state & PIPE_EOF) ||
1415                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1416                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1417                          wpipe->pipe_buffer.size == 0)))
1418                         revents |= events & (POLLOUT | POLLWRNORM);
1419
1420         levents = events &
1421             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1422         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1423             fp->f_seqcount == rpipe->pipe_wgen)
1424                 events |= POLLINIGNEOF;
1425
1426         if ((events & POLLINIGNEOF) == 0) {
1427                 if (rpipe->pipe_state & PIPE_EOF) {
1428                         revents |= (events & (POLLIN | POLLRDNORM));
1429                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1430                             (wpipe->pipe_state & PIPE_EOF))
1431                                 revents |= POLLHUP;
1432                 }
1433         }
1434
1435         if (revents == 0) {
1436                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1437                         selrecord(td, &rpipe->pipe_sel);
1438                         if (SEL_WAITING(&rpipe->pipe_sel))
1439                                 rpipe->pipe_state |= PIPE_SEL;
1440                 }
1441
1442                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1443                         selrecord(td, &wpipe->pipe_sel);
1444                         if (SEL_WAITING(&wpipe->pipe_sel))
1445                                 wpipe->pipe_state |= PIPE_SEL;
1446                 }
1447         }
1448 #ifdef MAC
1449 locked_error:
1450 #endif
1451         PIPE_UNLOCK(rpipe);
1452
1453         return (revents);
1454 }
1455
1456 /*
1457  * We shouldn't need locks here as we're doing a read and this should
1458  * be a natural race.
1459  */
1460 static int
1461 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1462     struct thread *td)
1463 {
1464         struct pipe *pipe;
1465         int new_unr;
1466 #ifdef MAC
1467         int error;
1468 #endif
1469
1470         pipe = fp->f_data;
1471         PIPE_LOCK(pipe);
1472 #ifdef MAC
1473         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1474         if (error) {
1475                 PIPE_UNLOCK(pipe);
1476                 return (error);
1477         }
1478 #endif
1479
1480         /* For named pipes ask the underlying filesystem. */
1481         if (pipe->pipe_state & PIPE_NAMED) {
1482                 PIPE_UNLOCK(pipe);
1483                 return (vnops.fo_stat(fp, ub, active_cred, td));
1484         }
1485
1486         /*
1487          * Lazily allocate an inode number for the pipe.  Most pipe
1488          * users do not call fstat(2) on the pipe, which means that
1489          * postponing the inode allocation until it is must be
1490          * returned to userland is useful.  If alloc_unr failed,
1491          * assign st_ino zero instead of returning an error.
1492          * Special pipe_ino values:
1493          *  -1 - not yet initialized;
1494          *  0  - alloc_unr failed, return 0 as st_ino forever.
1495          */
1496         if (pipe->pipe_ino == (ino_t)-1) {
1497                 new_unr = alloc_unr(pipeino_unr);
1498                 if (new_unr != -1)
1499                         pipe->pipe_ino = new_unr;
1500                 else
1501                         pipe->pipe_ino = 0;
1502         }
1503         PIPE_UNLOCK(pipe);
1504
1505         bzero(ub, sizeof(*ub));
1506         ub->st_mode = S_IFIFO;
1507         ub->st_blksize = PAGE_SIZE;
1508         if (pipe->pipe_state & PIPE_DIRECTW)
1509                 ub->st_size = pipe->pipe_map.cnt;
1510         else
1511                 ub->st_size = pipe->pipe_buffer.cnt;
1512         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1513         ub->st_atim = pipe->pipe_atime;
1514         ub->st_mtim = pipe->pipe_mtime;
1515         ub->st_ctim = pipe->pipe_ctime;
1516         ub->st_uid = fp->f_cred->cr_uid;
1517         ub->st_gid = fp->f_cred->cr_gid;
1518         ub->st_dev = pipedev_ino;
1519         ub->st_ino = pipe->pipe_ino;
1520         /*
1521          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1522          */
1523         return (0);
1524 }
1525
1526 /* ARGSUSED */
1527 static int
1528 pipe_close(struct file *fp, struct thread *td)
1529 {
1530
1531         if (fp->f_vnode != NULL)
1532                 return vnops.fo_close(fp, td);
1533         fp->f_ops = &badfileops;
1534         pipe_dtor(fp->f_data);
1535         fp->f_data = NULL;
1536         return (0);
1537 }
1538
1539 static int
1540 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1541 {
1542         struct pipe *cpipe;
1543         int error;
1544
1545         cpipe = fp->f_data;
1546         if (cpipe->pipe_state & PIPE_NAMED)
1547                 error = vn_chmod(fp, mode, active_cred, td);
1548         else
1549                 error = invfo_chmod(fp, mode, active_cred, td);
1550         return (error);
1551 }
1552
1553 static int
1554 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1555     struct thread *td)
1556 {
1557         struct pipe *cpipe;
1558         int error;
1559
1560         cpipe = fp->f_data;
1561         if (cpipe->pipe_state & PIPE_NAMED)
1562                 error = vn_chown(fp, uid, gid, active_cred, td);
1563         else
1564                 error = invfo_chown(fp, uid, gid, active_cred, td);
1565         return (error);
1566 }
1567
1568 static int
1569 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1570 {
1571         struct pipe *pi;
1572
1573         if (fp->f_type == DTYPE_FIFO)
1574                 return (vn_fill_kinfo(fp, kif, fdp));
1575         kif->kf_type = KF_TYPE_PIPE;
1576         pi = fp->f_data;
1577         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1578         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1579         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1580         return (0);
1581 }
1582
1583 static void
1584 pipe_free_kmem(struct pipe *cpipe)
1585 {
1586
1587         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1588             ("pipe_free_kmem: pipe mutex locked"));
1589
1590         if (cpipe->pipe_buffer.buffer != NULL) {
1591                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1592                 vm_map_remove(pipe_map,
1593                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1594                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1595                 cpipe->pipe_buffer.buffer = NULL;
1596         }
1597 #ifndef PIPE_NODIRECT
1598         {
1599                 cpipe->pipe_map.cnt = 0;
1600                 cpipe->pipe_map.pos = 0;
1601                 cpipe->pipe_map.npages = 0;
1602         }
1603 #endif
1604 }
1605
1606 /*
1607  * shutdown the pipe
1608  */
1609 static void
1610 pipeclose(struct pipe *cpipe)
1611 {
1612         struct pipepair *pp;
1613         struct pipe *ppipe;
1614
1615         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1616
1617         PIPE_LOCK(cpipe);
1618         pipelock(cpipe, 0);
1619         pp = cpipe->pipe_pair;
1620
1621         pipeselwakeup(cpipe);
1622
1623         /*
1624          * If the other side is blocked, wake it up saying that
1625          * we want to close it down.
1626          */
1627         cpipe->pipe_state |= PIPE_EOF;
1628         while (cpipe->pipe_busy) {
1629                 wakeup(cpipe);
1630                 cpipe->pipe_state |= PIPE_WANT;
1631                 pipeunlock(cpipe);
1632                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1633                 pipelock(cpipe, 0);
1634         }
1635
1636
1637         /*
1638          * Disconnect from peer, if any.
1639          */
1640         ppipe = cpipe->pipe_peer;
1641         if (ppipe->pipe_present == PIPE_ACTIVE) {
1642                 pipeselwakeup(ppipe);
1643
1644                 ppipe->pipe_state |= PIPE_EOF;
1645                 wakeup(ppipe);
1646                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1647         }
1648
1649         /*
1650          * Mark this endpoint as free.  Release kmem resources.  We
1651          * don't mark this endpoint as unused until we've finished
1652          * doing that, or the pipe might disappear out from under
1653          * us.
1654          */
1655         PIPE_UNLOCK(cpipe);
1656         pipe_free_kmem(cpipe);
1657         PIPE_LOCK(cpipe);
1658         cpipe->pipe_present = PIPE_CLOSING;
1659         pipeunlock(cpipe);
1660
1661         /*
1662          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1663          * PIPE_FINALIZED, that allows other end to free the
1664          * pipe_pair, only after the knotes are completely dismantled.
1665          */
1666         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1667         cpipe->pipe_present = PIPE_FINALIZED;
1668         seldrain(&cpipe->pipe_sel);
1669         knlist_destroy(&cpipe->pipe_sel.si_note);
1670
1671         /*
1672          * If both endpoints are now closed, release the memory for the
1673          * pipe pair.  If not, unlock.
1674          */
1675         if (ppipe->pipe_present == PIPE_FINALIZED) {
1676                 PIPE_UNLOCK(cpipe);
1677 #ifdef MAC
1678                 mac_pipe_destroy(pp);
1679 #endif
1680                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1681         } else
1682                 PIPE_UNLOCK(cpipe);
1683 }
1684
1685 /*ARGSUSED*/
1686 static int
1687 pipe_kqfilter(struct file *fp, struct knote *kn)
1688 {
1689         struct pipe *cpipe;
1690
1691         /*
1692          * If a filter is requested that is not supported by this file
1693          * descriptor, don't return an error, but also don't ever generate an
1694          * event.
1695          */
1696         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1697                 kn->kn_fop = &pipe_nfiltops;
1698                 return (0);
1699         }
1700         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1701                 kn->kn_fop = &pipe_nfiltops;
1702                 return (0);
1703         }
1704         cpipe = fp->f_data;
1705         PIPE_LOCK(cpipe);
1706         switch (kn->kn_filter) {
1707         case EVFILT_READ:
1708                 kn->kn_fop = &pipe_rfiltops;
1709                 break;
1710         case EVFILT_WRITE:
1711                 kn->kn_fop = &pipe_wfiltops;
1712                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1713                         /* other end of pipe has been closed */
1714                         PIPE_UNLOCK(cpipe);
1715                         return (EPIPE);
1716                 }
1717                 cpipe = PIPE_PEER(cpipe);
1718                 break;
1719         default:
1720                 PIPE_UNLOCK(cpipe);
1721                 return (EINVAL);
1722         }
1723
1724         kn->kn_hook = cpipe;
1725         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1726         PIPE_UNLOCK(cpipe);
1727         return (0);
1728 }
1729
1730 static void
1731 filt_pipedetach(struct knote *kn)
1732 {
1733         struct pipe *cpipe = kn->kn_hook;
1734
1735         PIPE_LOCK(cpipe);
1736         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1737         PIPE_UNLOCK(cpipe);
1738 }
1739
1740 /*ARGSUSED*/
1741 static int
1742 filt_piperead(struct knote *kn, long hint)
1743 {
1744         struct pipe *rpipe = kn->kn_hook;
1745         struct pipe *wpipe = rpipe->pipe_peer;
1746         int ret;
1747
1748         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1749         kn->kn_data = rpipe->pipe_buffer.cnt;
1750         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1751                 kn->kn_data = rpipe->pipe_map.cnt;
1752
1753         if ((rpipe->pipe_state & PIPE_EOF) ||
1754             wpipe->pipe_present != PIPE_ACTIVE ||
1755             (wpipe->pipe_state & PIPE_EOF)) {
1756                 kn->kn_flags |= EV_EOF;
1757                 return (1);
1758         }
1759         ret = kn->kn_data > 0;
1760         return ret;
1761 }
1762
1763 /*ARGSUSED*/
1764 static int
1765 filt_pipewrite(struct knote *kn, long hint)
1766 {
1767         struct pipe *wpipe;
1768
1769         wpipe = kn->kn_hook;
1770         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1771         if (wpipe->pipe_present != PIPE_ACTIVE ||
1772             (wpipe->pipe_state & PIPE_EOF)) {
1773                 kn->kn_data = 0;
1774                 kn->kn_flags |= EV_EOF;
1775                 return (1);
1776         }
1777         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1778             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1779         if (wpipe->pipe_state & PIPE_DIRECTW)
1780                 kn->kn_data = 0;
1781
1782         return (kn->kn_data >= PIPE_BUF);
1783 }
1784
1785 static void
1786 filt_pipedetach_notsup(struct knote *kn)
1787 {
1788
1789 }
1790
1791 static int
1792 filt_pipenotsup(struct knote *kn, long hint)
1793 {
1794
1795         return (0);
1796 }