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_type & PIPE_TYPE_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 int pipe_create(struct pipe *pipe, bool backing);
 230 static int 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 static void pipe_timestamp(struct timespec *tsp);
 234 #ifndef PIPE_NODIRECT
 235 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 236 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 237 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 238 static void pipe_clone_write_buffer(struct pipe *wpipe);
 239 #endif
 240 static int pipespace(struct pipe *cpipe, int size);
 241 static int pipespace_new(struct pipe *cpipe, int size);
 242
 243 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 244 static int      pipe_zone_init(void *mem, int size, int flags);
 245 static void     pipe_zone_fini(void *mem, int size);
 246
 247 static uma_zone_t pipe_zone;
 248 static struct unrhdr64 pipeino_unr;
 249 static dev_t pipedev_ino;
 250
 251 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 252
 253 static void
 254 pipeinit(void *dummy __unused)
 255 {
 256
 257         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 258             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 259             UMA_ALIGN_PTR, 0);
 260         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 261         new_unrhdr64(&pipeino_unr, 1);
 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         pipe_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 int
 340 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 341 {
 342         struct pipepair *pp;
 343         struct pipe *rpipe, *wpipe;
 344         int error;
 345
 346         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 347 #ifdef MAC
 348         /*
 349          * The MAC label is shared between the connected endpoints.  As a
 350          * result mac_pipe_init() and mac_pipe_create() are called once
 351          * for the pair, and not on the endpoints.
 352          */
 353         mac_pipe_init(pp);
 354         mac_pipe_create(td->td_ucred, pp);
 355 #endif
 356         rpipe = &pp->pp_rpipe;
 357         wpipe = &pp->pp_wpipe;
 358
 359         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 360         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 361
 362         /*
 363          * Only the forward direction pipe is backed by big buffer by
 364          * default.
 365          */
 366         error = pipe_create(rpipe, true);
 367         if (error != 0)
 368                 goto fail;
 369         error = pipe_create(wpipe, false);
 370         if (error != 0) {
 371                 /*
 372                  * This cleanup leaves the pipe inode number for rpipe
 373                  * still allocated, but never used.  We do not free
 374                  * inode numbers for opened pipes, which is required
 375                  * for correctness because numbers must be unique.
 376                  * But also it avoids any memory use by the unr
 377                  * allocator, so stashing away the transient inode
 378                  * number is reasonable.
 379                  */
 380                 pipe_free_kmem(rpipe);
 381                 goto fail;
 382         }
 383
 384         rpipe->pipe_state |= PIPE_DIRECTOK;
 385         wpipe->pipe_state |= PIPE_DIRECTOK;
 386         return (0);
 387
 388 fail:
 389         knlist_destroy(&rpipe->pipe_sel.si_note);
 390         knlist_destroy(&wpipe->pipe_sel.si_note);
 391 #ifdef MAC
 392         mac_pipe_destroy(pp);
 393 #endif
 394         uma_zfree(pipe_zone, pp);
 395         return (error);
 396 }
 397
 398 int
 399 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 400 {
 401         struct pipepair *pp;
 402         int error;
 403
 404         error = pipe_paircreate(td, &pp);
 405         if (error != 0)
 406                 return (error);
 407         pp->pp_rpipe.pipe_type |= PIPE_TYPE_NAMED;
 408         *ppipe = &pp->pp_rpipe;
 409         return (0);
 410 }
 411
 412 void
 413 pipe_dtor(struct pipe *dpipe)
 414 {
 415         struct pipe *peer;
 416
 417         peer = (dpipe->pipe_type & PIPE_TYPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 418         funsetown(&dpipe->pipe_sigio);
 419         pipeclose(dpipe);
 420         if (peer != NULL) {
 421                 funsetown(&peer->pipe_sigio);
 422                 pipeclose(peer);
 423         }
 424 }
 425
 426 /*
 427  * Get a timestamp.
 428  *
 429  * This used to be vfs_timestamp but the higher precision is unnecessary and
 430  * can very negatively affect performance in virtualized environments (e.g., on
 431  * vms running on amd64 when using the rdtscp instruction).
 432  */
 433 static void
 434 pipe_timestamp(struct timespec *tsp)
 435 {
 436
 437         getnanotime(tsp);
 438 }
 439
 440 /*
 441  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 442  * the zone pick up the pieces via pipeclose().
 443  */
 444 int
 445 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 446     struct filecaps *fcaps2)
 447 {
 448         struct file *rf, *wf;
 449         struct pipe *rpipe, *wpipe;
 450         struct pipepair *pp;
 451         int fd, fflags, error;
 452
 453         error = pipe_paircreate(td, &pp);
 454         if (error != 0)
 455                 return (error);
 456         rpipe = &pp->pp_rpipe;
 457         wpipe = &pp->pp_wpipe;
 458         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 459         if (error) {
 460                 pipeclose(rpipe);
 461                 pipeclose(wpipe);
 462                 return (error);
 463         }
 464         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 465         fildes[0] = fd;
 466
 467         fflags = FREAD | FWRITE;
 468         if ((flags & O_NONBLOCK) != 0)
 469                 fflags |= FNONBLOCK;
 470
 471         /*
 472          * Warning: once we've gotten past allocation of the fd for the
 473          * read-side, we can only drop the read side via fdrop() in order
 474          * to avoid races against processes which manage to dup() the read
 475          * side while we are blocked trying to allocate the write side.
 476          */
 477         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 478         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 479         if (error) {
 480                 fdclose(td, rf, fildes[0]);
 481                 fdrop(rf, td);
 482                 /* rpipe has been closed by fdrop(). */
 483                 pipeclose(wpipe);
 484                 return (error);
 485         }
 486         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 487         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 488         fdrop(wf, td);
 489         fildes[1] = fd;
 490         fdrop(rf, td);
 491
 492         return (0);
 493 }
 494
 495 #ifdef COMPAT_FREEBSD10
 496 /* ARGSUSED */
 497 int
 498 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 499 {
 500         int error;
 501         int fildes[2];
 502
 503         error = kern_pipe(td, fildes, 0, NULL, NULL);
 504         if (error)
 505                 return (error);
 506
 507         td->td_retval[0] = fildes[0];
 508         td->td_retval[1] = fildes[1];
 509
 510         return (0);
 511 }
 512 #endif
 513
 514 int
 515 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 516 {
 517         int error, fildes[2];
 518
 519         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 520                 return (EINVAL);
 521         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 522         if (error)
 523                 return (error);
 524         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 525         if (error) {
 526                 (void)kern_close(td, fildes[0]);
 527                 (void)kern_close(td, fildes[1]);
 528         }
 529         return (error);
 530 }
 531
 532 /*
 533  * Allocate kva for pipe circular buffer, the space is pageable
 534  * This routine will 'realloc' the size of a pipe safely, if it fails
 535  * it will retain the old buffer.
 536  * If it fails it will return ENOMEM.
 537  */
 538 static int
 539 pipespace_new(struct pipe *cpipe, int size)
 540 {
 541         caddr_t buffer;
 542         int error, cnt, firstseg;
 543         static int curfail = 0;
 544         static struct timeval lastfail;
 545
 546         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 547         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 548                 ("pipespace: resize of direct writes not allowed"));
 549 retry:
 550         cnt = cpipe->pipe_buffer.cnt;
 551         if (cnt > size)
 552                 size = cnt;
 553
 554         size = round_page(size);
 555         buffer = (caddr_t) vm_map_min(pipe_map);
 556
 557         error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
 558             VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
 559         if (error != KERN_SUCCESS) {
 560                 if (cpipe->pipe_buffer.buffer == NULL &&
 561                     size > SMALL_PIPE_SIZE) {
 562                         size = SMALL_PIPE_SIZE;
 563                         pipefragretry++;
 564                         goto retry;
 565                 }
 566                 if (cpipe->pipe_buffer.buffer == NULL) {
 567                         pipeallocfail++;
 568                         if (ppsratecheck(&lastfail, &curfail, 1))
 569                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 570                 } else {
 571                         piperesizefail++;
 572                 }
 573                 return (ENOMEM);
 574         }
 575
 576         /* copy data, then free old resources if we're resizing */
 577         if (cnt > 0) {
 578                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 579                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 580                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 581                                 buffer, firstseg);
 582                         if ((cnt - firstseg) > 0)
 583                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 584                                         cpipe->pipe_buffer.in);
 585                 } else {
 586                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 587                                 buffer, cnt);
 588                 }
 589         }
 590         pipe_free_kmem(cpipe);
 591         cpipe->pipe_buffer.buffer = buffer;
 592         cpipe->pipe_buffer.size = size;
 593         cpipe->pipe_buffer.in = cnt;
 594         cpipe->pipe_buffer.out = 0;
 595         cpipe->pipe_buffer.cnt = cnt;
 596         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 597         return (0);
 598 }
 599
 600 /*
 601  * Wrapper for pipespace_new() that performs locking assertions.
 602  */
 603 static int
 604 pipespace(struct pipe *cpipe, int size)
 605 {
 606
 607         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 608             ("Unlocked pipe passed to pipespace"));
 609         return (pipespace_new(cpipe, size));
 610 }
 611
 612 /*
 613  * lock a pipe for I/O, blocking other access
 614  */
 615 static __inline int
 616 pipelock(struct pipe *cpipe, int catch)
 617 {
 618         int error, prio;
 619
 620         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 621
 622         prio = PRIBIO;
 623         if (catch)
 624                 prio |= PCATCH;
 625         while (cpipe->pipe_state & PIPE_LOCKFL) {
 626                 KASSERT(cpipe->pipe_waiters >= 0,
 627                     ("%s: bad waiter count %d", __func__,
 628                     cpipe->pipe_waiters));
 629                 cpipe->pipe_waiters++;
 630                 error = msleep(cpipe, PIPE_MTX(cpipe),
 631                     prio, "pipelk", 0);
 632                 cpipe->pipe_waiters--;
 633                 if (error != 0)
 634                         return (error);
 635         }
 636         cpipe->pipe_state |= PIPE_LOCKFL;
 637         return (0);
 638 }
 639
 640 /*
 641  * unlock a pipe I/O lock
 642  */
 643 static __inline void
 644 pipeunlock(struct pipe *cpipe)
 645 {
 646
 647         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 648         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 649                 ("Unlocked pipe passed to pipeunlock"));
 650         KASSERT(cpipe->pipe_waiters >= 0,
 651             ("%s: bad waiter count %d", __func__,
 652             cpipe->pipe_waiters));
 653         cpipe->pipe_state &= ~PIPE_LOCKFL;
 654         if (cpipe->pipe_waiters > 0) {
 655                 wakeup_one(cpipe);
 656         }
 657 }
 658
 659 void
 660 pipeselwakeup(struct pipe *cpipe)
 661 {
 662
 663         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 664         if (cpipe->pipe_state & PIPE_SEL) {
 665                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 666                 if (!SEL_WAITING(&cpipe->pipe_sel))
 667                         cpipe->pipe_state &= ~PIPE_SEL;
 668         }
 669         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 670                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 671         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 672 }
 673
 674 /*
 675  * Initialize and allocate VM and memory for pipe.  The structure
 676  * will start out zero'd from the ctor, so we just manage the kmem.
 677  */
 678 static int
 679 pipe_create(struct pipe *pipe, bool large_backing)
 680 {
 681         int error;
 682
 683         error = pipespace_new(pipe, !large_backing || amountpipekva >
 684             maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
 685         if (error == 0)
 686                 pipe->pipe_ino = alloc_unr64(&pipeino_unr);
 687         return (error);
 688 }
 689
 690 /* ARGSUSED */
 691 static int
 692 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 693     int flags, struct thread *td)
 694 {
 695         struct pipe *rpipe;
 696         int error;
 697         int nread = 0;
 698         int size;
 699
 700         rpipe = fp->f_data;
 701         PIPE_LOCK(rpipe);
 702         ++rpipe->pipe_busy;
 703         error = pipelock(rpipe, 1);
 704         if (error)
 705                 goto unlocked_error;
 706
 707 #ifdef MAC
 708         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 709         if (error)
 710                 goto locked_error;
 711 #endif
 712         if (amountpipekva > (3 * maxpipekva) / 4) {
 713                 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 714                     rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
 715                     rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
 716                     piperesizeallowed == 1) {
 717                         PIPE_UNLOCK(rpipe);
 718                         pipespace(rpipe, SMALL_PIPE_SIZE);
 719                         PIPE_LOCK(rpipe);
 720                 }
 721         }
 722
 723         while (uio->uio_resid) {
 724                 /*
 725                  * normal pipe buffer receive
 726                  */
 727                 if (rpipe->pipe_buffer.cnt > 0) {
 728                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 729                         if (size > rpipe->pipe_buffer.cnt)
 730                                 size = rpipe->pipe_buffer.cnt;
 731                         if (size > uio->uio_resid)
 732                                 size = uio->uio_resid;
 733
 734                         PIPE_UNLOCK(rpipe);
 735                         error = uiomove(
 736                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 737                             size, uio);
 738                         PIPE_LOCK(rpipe);
 739                         if (error)
 740                                 break;
 741
 742                         rpipe->pipe_buffer.out += size;
 743                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 744                                 rpipe->pipe_buffer.out = 0;
 745
 746                         rpipe->pipe_buffer.cnt -= size;
 747
 748                         /*
 749                          * If there is no more to read in the pipe, reset
 750                          * its pointers to the beginning.  This improves
 751                          * cache hit stats.
 752                          */
 753                         if (rpipe->pipe_buffer.cnt == 0) {
 754                                 rpipe->pipe_buffer.in = 0;
 755                                 rpipe->pipe_buffer.out = 0;
 756                         }
 757                         nread += size;
 758 #ifndef PIPE_NODIRECT
 759                 /*
 760                  * Direct copy, bypassing a kernel buffer.
 761                  */
 762                 } else if ((size = rpipe->pipe_pages.cnt) != 0) {
 763                         if (size > uio->uio_resid)
 764                                 size = (u_int) uio->uio_resid;
 765                         PIPE_UNLOCK(rpipe);
 766                         error = uiomove_fromphys(rpipe->pipe_pages.ms,
 767                             rpipe->pipe_pages.pos, size, uio);
 768                         PIPE_LOCK(rpipe);
 769                         if (error)
 770                                 break;
 771                         nread += size;
 772                         rpipe->pipe_pages.pos += size;
 773                         rpipe->pipe_pages.cnt -= size;
 774                         if (rpipe->pipe_pages.cnt == 0) {
 775                                 rpipe->pipe_state &= ~PIPE_WANTW;
 776                                 wakeup(rpipe);
 777                         }
 778 #endif
 779                 } else {
 780                         /*
 781                          * detect EOF condition
 782                          * read returns 0 on EOF, no need to set error
 783                          */
 784                         if (rpipe->pipe_state & PIPE_EOF)
 785                                 break;
 786
 787                         /*
 788                          * If the "write-side" has been blocked, wake it up now.
 789                          */
 790                         if (rpipe->pipe_state & PIPE_WANTW) {
 791                                 rpipe->pipe_state &= ~PIPE_WANTW;
 792                                 wakeup(rpipe);
 793                         }
 794
 795                         /*
 796                          * Break if some data was read.
 797                          */
 798                         if (nread > 0)
 799                                 break;
 800
 801                         /*
 802                          * Unlock the pipe buffer for our remaining processing.
 803                          * We will either break out with an error or we will
 804                          * sleep and relock to loop.
 805                          */
 806                         pipeunlock(rpipe);
 807
 808                         /*
 809                          * Handle non-blocking mode operation or
 810                          * wait for more data.
 811                          */
 812                         if (fp->f_flag & FNONBLOCK) {
 813                                 error = EAGAIN;
 814                         } else {
 815                                 rpipe->pipe_state |= PIPE_WANTR;
 816                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 817                                     PRIBIO | PCATCH,
 818                                     "piperd", 0)) == 0)
 819                                         error = pipelock(rpipe, 1);
 820                         }
 821                         if (error)
 822                                 goto unlocked_error;
 823                 }
 824         }
 825 #ifdef MAC
 826 locked_error:
 827 #endif
 828         pipeunlock(rpipe);
 829
 830         /* XXX: should probably do this before getting any locks. */
 831         if (error == 0)
 832                 pipe_timestamp(&rpipe->pipe_atime);
 833 unlocked_error:
 834         --rpipe->pipe_busy;
 835
 836         /*
 837          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 838          */
 839         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 840                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 841                 wakeup(rpipe);
 842         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 843                 /*
 844                  * Handle write blocking hysteresis.
 845                  */
 846                 if (rpipe->pipe_state & PIPE_WANTW) {
 847                         rpipe->pipe_state &= ~PIPE_WANTW;
 848                         wakeup(rpipe);
 849                 }
 850         }
 851
 852         /*
 853          * Only wake up writers if there was actually something read.
 854          * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
 855          */
 856         if (nread > 0 &&
 857             rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
 858                 pipeselwakeup(rpipe);
 859
 860         PIPE_UNLOCK(rpipe);
 861         if (nread > 0)
 862                 td->td_ru.ru_msgrcv++;
 863         return (error);
 864 }
 865
 866 #ifndef PIPE_NODIRECT
 867 /*
 868  * Map the sending processes' buffer into kernel space and wire it.
 869  * This is similar to a physical write operation.
 870  */
 871 static int
 872 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 873 {
 874         u_int size;
 875         int i;
 876
 877         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 878         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
 879             ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
 880         KASSERT(wpipe->pipe_pages.cnt == 0,
 881             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 882
 883         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 884                 size = wpipe->pipe_buffer.size;
 885         else
 886                 size = uio->uio_iov->iov_len;
 887
 888         wpipe->pipe_state |= PIPE_DIRECTW;
 889         PIPE_UNLOCK(wpipe);
 890         i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 891             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 892             wpipe->pipe_pages.ms, PIPENPAGES);
 893         PIPE_LOCK(wpipe);
 894         if (i < 0) {
 895                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 896                 return (EFAULT);
 897         }
 898
 899         wpipe->pipe_pages.npages = i;
 900         wpipe->pipe_pages.pos =
 901             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 902         wpipe->pipe_pages.cnt = size;
 903
 904         uio->uio_iov->iov_len -= size;
 905         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 906         if (uio->uio_iov->iov_len == 0)
 907                 uio->uio_iov++;
 908         uio->uio_resid -= size;
 909         uio->uio_offset += size;
 910         return (0);
 911 }
 912
 913 /*
 914  * Unwire the process buffer.
 915  */
 916 static void
 917 pipe_destroy_write_buffer(struct pipe *wpipe)
 918 {
 919
 920         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 921         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 922             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 923         KASSERT(wpipe->pipe_pages.cnt == 0,
 924             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 925
 926         wpipe->pipe_state &= ~PIPE_DIRECTW;
 927         vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
 928         wpipe->pipe_pages.npages = 0;
 929 }
 930
 931 /*
 932  * In the case of a signal, the writing process might go away.  This
 933  * code copies the data into the circular buffer so that the source
 934  * pages can be freed without loss of data.
 935  */
 936 static void
 937 pipe_clone_write_buffer(struct pipe *wpipe)
 938 {
 939         struct uio uio;
 940         struct iovec iov;
 941         int size;
 942         int pos;
 943
 944         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 945         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 946             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 947
 948         size = wpipe->pipe_pages.cnt;
 949         pos = wpipe->pipe_pages.pos;
 950         wpipe->pipe_pages.cnt = 0;
 951
 952         wpipe->pipe_buffer.in = size;
 953         wpipe->pipe_buffer.out = 0;
 954         wpipe->pipe_buffer.cnt = size;
 955
 956         PIPE_UNLOCK(wpipe);
 957         iov.iov_base = wpipe->pipe_buffer.buffer;
 958         iov.iov_len = size;
 959         uio.uio_iov = &iov;
 960         uio.uio_iovcnt = 1;
 961         uio.uio_offset = 0;
 962         uio.uio_resid = size;
 963         uio.uio_segflg = UIO_SYSSPACE;
 964         uio.uio_rw = UIO_READ;
 965         uio.uio_td = curthread;
 966         uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
 967         PIPE_LOCK(wpipe);
 968         pipe_destroy_write_buffer(wpipe);
 969 }
 970
 971 /*
 972  * This implements the pipe buffer write mechanism.  Note that only
 973  * a direct write OR a normal pipe write can be pending at any given time.
 974  * If there are any characters in the pipe buffer, the direct write will
 975  * be deferred until the receiving process grabs all of the bytes from
 976  * the pipe buffer.  Then the direct mapping write is set-up.
 977  */
 978 static int
 979 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 980 {
 981         int error;
 982
 983 retry:
 984         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 985         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 986                 error = EPIPE;
 987                 goto error1;
 988         }
 989         if (wpipe->pipe_state & PIPE_DIRECTW) {
 990                 if (wpipe->pipe_state & PIPE_WANTR) {
 991                         wpipe->pipe_state &= ~PIPE_WANTR;
 992                         wakeup(wpipe);
 993                 }
 994                 pipeselwakeup(wpipe);
 995                 wpipe->pipe_state |= PIPE_WANTW;
 996                 pipeunlock(wpipe);
 997                 error = msleep(wpipe, PIPE_MTX(wpipe),
 998                     PRIBIO | PCATCH, "pipdww", 0);
 999                 pipelock(wpipe, 0);
1000                 if (error != 0)
1001                         goto error1;
1002                 goto retry;
1003         }
1004         if (wpipe->pipe_buffer.cnt > 0) {
1005                 if (wpipe->pipe_state & PIPE_WANTR) {
1006                         wpipe->pipe_state &= ~PIPE_WANTR;
1007                         wakeup(wpipe);
1008                 }
1009                 pipeselwakeup(wpipe);
1010                 wpipe->pipe_state |= PIPE_WANTW;
1011                 pipeunlock(wpipe);
1012                 error = msleep(wpipe, PIPE_MTX(wpipe),
1013                     PRIBIO | PCATCH, "pipdwc", 0);
1014                 pipelock(wpipe, 0);
1015                 if (error != 0)
1016                         goto error1;
1017                 goto retry;
1018         }
1019
1020         error = pipe_build_write_buffer(wpipe, uio);
1021         if (error) {
1022                 goto error1;
1023         }
1024
1025         while (wpipe->pipe_pages.cnt != 0 &&
1026             (wpipe->pipe_state & PIPE_EOF) == 0) {
1027                 if (wpipe->pipe_state & PIPE_WANTR) {
1028                         wpipe->pipe_state &= ~PIPE_WANTR;
1029                         wakeup(wpipe);
1030                 }
1031                 pipeselwakeup(wpipe);
1032                 wpipe->pipe_state |= PIPE_WANTW;
1033                 pipeunlock(wpipe);
1034                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1035                     "pipdwt", 0);
1036                 pipelock(wpipe, 0);
1037                 if (error != 0)
1038                         break;
1039         }
1040
1041         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1042                 wpipe->pipe_pages.cnt = 0;
1043                 pipe_destroy_write_buffer(wpipe);
1044                 pipeselwakeup(wpipe);
1045                 error = EPIPE;
1046         } else if (error == EINTR || error == ERESTART) {
1047                 pipe_clone_write_buffer(wpipe);
1048         } else {
1049                 pipe_destroy_write_buffer(wpipe);
1050         }
1051         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1052             ("pipe %p leaked PIPE_DIRECTW", wpipe));
1053         return (error);
1054
1055 error1:
1056         wakeup(wpipe);
1057         return (error);
1058 }
1059 #endif
1060
1061 static int
1062 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1063     int flags, struct thread *td)
1064 {
1065         struct pipe *wpipe, *rpipe;
1066         ssize_t orig_resid;
1067         int desiredsize, error;
1068
1069         rpipe = fp->f_data;
1070         wpipe = PIPE_PEER(rpipe);
1071         PIPE_LOCK(rpipe);
1072         error = pipelock(wpipe, 1);
1073         if (error) {
1074                 PIPE_UNLOCK(rpipe);
1075                 return (error);
1076         }
1077         /*
1078          * detect loss of pipe read side, issue SIGPIPE if lost.
1079          */
1080         if (wpipe->pipe_present != PIPE_ACTIVE ||
1081             (wpipe->pipe_state & PIPE_EOF)) {
1082                 pipeunlock(wpipe);
1083                 PIPE_UNLOCK(rpipe);
1084                 return (EPIPE);
1085         }
1086 #ifdef MAC
1087         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1088         if (error) {
1089                 pipeunlock(wpipe);
1090                 PIPE_UNLOCK(rpipe);
1091                 return (error);
1092         }
1093 #endif
1094         ++wpipe->pipe_busy;
1095
1096         /* Choose a larger size if it's advantageous */
1097         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1098         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1099                 if (piperesizeallowed != 1)
1100                         break;
1101                 if (amountpipekva > maxpipekva / 2)
1102                         break;
1103                 if (desiredsize == BIG_PIPE_SIZE)
1104                         break;
1105                 desiredsize = desiredsize * 2;
1106         }
1107
1108         /* Choose a smaller size if we're in a OOM situation */
1109         if (amountpipekva > (3 * maxpipekva) / 4 &&
1110             wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1111             wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1112             piperesizeallowed == 1)
1113                 desiredsize = SMALL_PIPE_SIZE;
1114
1115         /* Resize if the above determined that a new size was necessary */
1116         if (desiredsize != wpipe->pipe_buffer.size &&
1117             (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1118                 PIPE_UNLOCK(wpipe);
1119                 pipespace(wpipe, desiredsize);
1120                 PIPE_LOCK(wpipe);
1121         }
1122         MPASS(wpipe->pipe_buffer.size != 0);
1123
1124         orig_resid = uio->uio_resid;
1125
1126         while (uio->uio_resid) {
1127                 int space;
1128
1129                 if (wpipe->pipe_state & PIPE_EOF) {
1130                         error = EPIPE;
1131                         break;
1132                 }
1133 #ifndef PIPE_NODIRECT
1134                 /*
1135                  * If the transfer is large, we can gain performance if
1136                  * we do process-to-process copies directly.
1137                  * If the write is non-blocking, we don't use the
1138                  * direct write mechanism.
1139                  *
1140                  * The direct write mechanism will detect the reader going
1141                  * away on us.
1142                  */
1143                 if (uio->uio_segflg == UIO_USERSPACE &&
1144                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1145                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1146                     (fp->f_flag & FNONBLOCK) == 0) {
1147                         error = pipe_direct_write(wpipe, uio);
1148                         if (error != 0)
1149                                 break;
1150                         continue;
1151                 }
1152 #endif
1153
1154                 /*
1155                  * Pipe buffered writes cannot be coincidental with
1156                  * direct writes.  We wait until the currently executing
1157                  * direct write is completed before we start filling the
1158                  * pipe buffer.  We break out if a signal occurs or the
1159                  * reader goes away.
1160                  */
1161                 if (wpipe->pipe_pages.cnt != 0) {
1162                         if (wpipe->pipe_state & PIPE_WANTR) {
1163                                 wpipe->pipe_state &= ~PIPE_WANTR;
1164                                 wakeup(wpipe);
1165                         }
1166                         pipeselwakeup(wpipe);
1167                         wpipe->pipe_state |= PIPE_WANTW;
1168                         pipeunlock(wpipe);
1169                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1170                             "pipbww", 0);
1171                         pipelock(wpipe, 0);
1172                         if (error != 0)
1173                                 break;
1174                         continue;
1175                 }
1176
1177                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1178
1179                 /* Writes of size <= PIPE_BUF must be atomic. */
1180                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1181                         space = 0;
1182
1183                 if (space > 0) {
1184                         int size;       /* Transfer size */
1185                         int segsize;    /* first segment to transfer */
1186
1187                         /*
1188                          * Transfer size is minimum of uio transfer
1189                          * and free space in pipe buffer.
1190                          */
1191                         if (space > uio->uio_resid)
1192                                 size = uio->uio_resid;
1193                         else
1194                                 size = space;
1195                         /*
1196                          * First segment to transfer is minimum of
1197                          * transfer size and contiguous space in
1198                          * pipe buffer.  If first segment to transfer
1199                          * is less than the transfer size, we've got
1200                          * a wraparound in the buffer.
1201                          */
1202                         segsize = wpipe->pipe_buffer.size -
1203                                 wpipe->pipe_buffer.in;
1204                         if (segsize > size)
1205                                 segsize = size;
1206
1207                         /* Transfer first segment */
1208
1209                         PIPE_UNLOCK(rpipe);
1210                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1211                                         segsize, uio);
1212                         PIPE_LOCK(rpipe);
1213
1214                         if (error == 0 && segsize < size) {
1215                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1216                                         wpipe->pipe_buffer.size,
1217                                         ("Pipe buffer wraparound disappeared"));
1218                                 /*
1219                                  * Transfer remaining part now, to
1220                                  * support atomic writes.  Wraparound
1221                                  * happened.
1222                                  */
1223
1224                                 PIPE_UNLOCK(rpipe);
1225                                 error = uiomove(
1226                                     &wpipe->pipe_buffer.buffer[0],
1227                                     size - segsize, uio);
1228                                 PIPE_LOCK(rpipe);
1229                         }
1230                         if (error == 0) {
1231                                 wpipe->pipe_buffer.in += size;
1232                                 if (wpipe->pipe_buffer.in >=
1233                                     wpipe->pipe_buffer.size) {
1234                                         KASSERT(wpipe->pipe_buffer.in ==
1235                                                 size - segsize +
1236                                                 wpipe->pipe_buffer.size,
1237                                                 ("Expected wraparound bad"));
1238                                         wpipe->pipe_buffer.in = size - segsize;
1239                                 }
1240
1241                                 wpipe->pipe_buffer.cnt += size;
1242                                 KASSERT(wpipe->pipe_buffer.cnt <=
1243                                         wpipe->pipe_buffer.size,
1244                                         ("Pipe buffer overflow"));
1245                         }
1246                         if (error != 0)
1247                                 break;
1248                         continue;
1249                 } else {
1250                         /*
1251                          * If the "read-side" has been blocked, wake it up now.
1252                          */
1253                         if (wpipe->pipe_state & PIPE_WANTR) {
1254                                 wpipe->pipe_state &= ~PIPE_WANTR;
1255                                 wakeup(wpipe);
1256                         }
1257
1258                         /*
1259                          * don't block on non-blocking I/O
1260                          */
1261                         if (fp->f_flag & FNONBLOCK) {
1262                                 error = EAGAIN;
1263                                 break;
1264                         }
1265
1266                         /*
1267                          * We have no more space and have something to offer,
1268                          * wake up select/poll.
1269                          */
1270                         pipeselwakeup(wpipe);
1271
1272                         wpipe->pipe_state |= PIPE_WANTW;
1273                         pipeunlock(wpipe);
1274                         error = msleep(wpipe, PIPE_MTX(rpipe),
1275                             PRIBIO | PCATCH, "pipewr", 0);
1276                         pipelock(wpipe, 0);
1277                         if (error != 0)
1278                                 break;
1279                         continue;
1280                 }
1281         }
1282
1283         --wpipe->pipe_busy;
1284
1285         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1286                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1287                 wakeup(wpipe);
1288         } else if (wpipe->pipe_buffer.cnt > 0) {
1289                 /*
1290                  * If we have put any characters in the buffer, we wake up
1291                  * the reader.
1292                  */
1293                 if (wpipe->pipe_state & PIPE_WANTR) {
1294                         wpipe->pipe_state &= ~PIPE_WANTR;
1295                         wakeup(wpipe);
1296                 }
1297         }
1298
1299         /*
1300          * Don't return EPIPE if any byte was written.
1301          * EINTR and other interrupts are handled by generic I/O layer.
1302          * Do not pretend that I/O succeeded for obvious user error
1303          * like EFAULT.
1304          */
1305         if (uio->uio_resid != orig_resid && error == EPIPE)
1306                 error = 0;
1307
1308         if (error == 0)
1309                 pipe_timestamp(&wpipe->pipe_mtime);
1310
1311         /*
1312          * We have something to offer,
1313          * wake up select/poll.
1314          */
1315         if (wpipe->pipe_buffer.cnt)
1316                 pipeselwakeup(wpipe);
1317
1318         pipeunlock(wpipe);
1319         PIPE_UNLOCK(rpipe);
1320         if (uio->uio_resid != orig_resid)
1321                 td->td_ru.ru_msgsnd++;
1322         return (error);
1323 }
1324
1325 /* ARGSUSED */
1326 static int
1327 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1328     struct thread *td)
1329 {
1330         struct pipe *cpipe;
1331         int error;
1332
1333         cpipe = fp->f_data;
1334         if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1335                 error = vnops.fo_truncate(fp, length, active_cred, td);
1336         else
1337                 error = invfo_truncate(fp, length, active_cred, td);
1338         return (error);
1339 }
1340
1341 /*
1342  * we implement a very minimal set of ioctls for compatibility with sockets.
1343  */
1344 static int
1345 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1346     struct thread *td)
1347 {
1348         struct pipe *mpipe = fp->f_data;
1349         int error;
1350
1351         PIPE_LOCK(mpipe);
1352
1353 #ifdef MAC
1354         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1355         if (error) {
1356                 PIPE_UNLOCK(mpipe);
1357                 return (error);
1358         }
1359 #endif
1360
1361         error = 0;
1362         switch (cmd) {
1363         case FIONBIO:
1364                 break;
1365
1366         case FIOASYNC:
1367                 if (*(int *)data) {
1368                         mpipe->pipe_state |= PIPE_ASYNC;
1369                 } else {
1370                         mpipe->pipe_state &= ~PIPE_ASYNC;
1371                 }
1372                 break;
1373
1374         case FIONREAD:
1375                 if (!(fp->f_flag & FREAD)) {
1376                         *(int *)data = 0;
1377                         PIPE_UNLOCK(mpipe);
1378                         return (0);
1379                 }
1380                 if (mpipe->pipe_pages.cnt != 0)
1381                         *(int *)data = mpipe->pipe_pages.cnt;
1382                 else
1383                         *(int *)data = mpipe->pipe_buffer.cnt;
1384                 break;
1385
1386         case FIOSETOWN:
1387                 PIPE_UNLOCK(mpipe);
1388                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1389                 goto out_unlocked;
1390
1391         case FIOGETOWN:
1392                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1393                 break;
1394
1395         /* This is deprecated, FIOSETOWN should be used instead. */
1396         case TIOCSPGRP:
1397                 PIPE_UNLOCK(mpipe);
1398                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1399                 goto out_unlocked;
1400
1401         /* This is deprecated, FIOGETOWN should be used instead. */
1402         case TIOCGPGRP:
1403                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1404                 break;
1405
1406         default:
1407                 error = ENOTTY;
1408                 break;
1409         }
1410         PIPE_UNLOCK(mpipe);
1411 out_unlocked:
1412         return (error);
1413 }
1414
1415 static int
1416 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1417     struct thread *td)
1418 {
1419         struct pipe *rpipe;
1420         struct pipe *wpipe;
1421         int levents, revents;
1422 #ifdef MAC
1423         int error;
1424 #endif
1425
1426         revents = 0;
1427         rpipe = fp->f_data;
1428         wpipe = PIPE_PEER(rpipe);
1429         PIPE_LOCK(rpipe);
1430 #ifdef MAC
1431         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1432         if (error)
1433                 goto locked_error;
1434 #endif
1435         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1436                 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1437                         revents |= events & (POLLIN | POLLRDNORM);
1438
1439         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1440                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1441                     (wpipe->pipe_state & PIPE_EOF) ||
1442                     ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1443                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1444                          wpipe->pipe_buffer.size == 0)))
1445                         revents |= events & (POLLOUT | POLLWRNORM);
1446
1447         levents = events &
1448             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1449         if (rpipe->pipe_type & PIPE_TYPE_NAMED && fp->f_flag & FREAD && levents &&
1450             fp->f_pipegen == rpipe->pipe_wgen)
1451                 events |= POLLINIGNEOF;
1452
1453         if ((events & POLLINIGNEOF) == 0) {
1454                 if (rpipe->pipe_state & PIPE_EOF) {
1455                         if (fp->f_flag & FREAD)
1456                                 revents |= (events & (POLLIN | POLLRDNORM));
1457                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1458                             (wpipe->pipe_state & PIPE_EOF))
1459                                 revents |= POLLHUP;
1460                 }
1461         }
1462
1463         if (revents == 0) {
1464                 /*
1465                  * Add ourselves regardless of eventmask as we have to return
1466                  * POLLHUP even if it was not asked for.
1467                  */
1468                 if ((fp->f_flag & FREAD) != 0) {
1469                         selrecord(td, &rpipe->pipe_sel);
1470                         if (SEL_WAITING(&rpipe->pipe_sel))
1471                                 rpipe->pipe_state |= PIPE_SEL;
1472                 }
1473
1474                 if ((fp->f_flag & FWRITE) != 0 &&
1475                     wpipe->pipe_present == PIPE_ACTIVE) {
1476                         selrecord(td, &wpipe->pipe_sel);
1477                         if (SEL_WAITING(&wpipe->pipe_sel))
1478                                 wpipe->pipe_state |= PIPE_SEL;
1479                 }
1480         }
1481 #ifdef MAC
1482 locked_error:
1483 #endif
1484         PIPE_UNLOCK(rpipe);
1485
1486         return (revents);
1487 }
1488
1489 /*
1490  * We shouldn't need locks here as we're doing a read and this should
1491  * be a natural race.
1492  */
1493 static int
1494 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1495     struct thread *td)
1496 {
1497         struct pipe *pipe;
1498 #ifdef MAC
1499         int error;
1500 #endif
1501
1502         pipe = fp->f_data;
1503 #ifdef MAC
1504         if (mac_pipe_check_stat_enabled()) {
1505                 PIPE_LOCK(pipe);
1506                 error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1507                 PIPE_UNLOCK(pipe);
1508                 if (error) {
1509                         return (error);
1510                 }
1511         }
1512 #endif
1513
1514         /* For named pipes ask the underlying filesystem. */
1515         if (pipe->pipe_type & PIPE_TYPE_NAMED) {
1516                 return (vnops.fo_stat(fp, ub, active_cred, td));
1517         }
1518
1519         bzero(ub, sizeof(*ub));
1520         ub->st_mode = S_IFIFO;
1521         ub->st_blksize = PAGE_SIZE;
1522         if (pipe->pipe_pages.cnt != 0)
1523                 ub->st_size = pipe->pipe_pages.cnt;
1524         else
1525                 ub->st_size = pipe->pipe_buffer.cnt;
1526         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1527         ub->st_atim = pipe->pipe_atime;
1528         ub->st_mtim = pipe->pipe_mtime;
1529         ub->st_ctim = pipe->pipe_ctime;
1530         ub->st_uid = fp->f_cred->cr_uid;
1531         ub->st_gid = fp->f_cred->cr_gid;
1532         ub->st_dev = pipedev_ino;
1533         ub->st_ino = pipe->pipe_ino;
1534         /*
1535          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1536          */
1537         return (0);
1538 }
1539
1540 /* ARGSUSED */
1541 static int
1542 pipe_close(struct file *fp, struct thread *td)
1543 {
1544
1545         if (fp->f_vnode != NULL)
1546                 return vnops.fo_close(fp, td);
1547         fp->f_ops = &badfileops;
1548         pipe_dtor(fp->f_data);
1549         fp->f_data = NULL;
1550         return (0);
1551 }
1552
1553 static int
1554 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1555 {
1556         struct pipe *cpipe;
1557         int error;
1558
1559         cpipe = fp->f_data;
1560         if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1561                 error = vn_chmod(fp, mode, active_cred, td);
1562         else
1563                 error = invfo_chmod(fp, mode, active_cred, td);
1564         return (error);
1565 }
1566
1567 static int
1568 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1569     struct thread *td)
1570 {
1571         struct pipe *cpipe;
1572         int error;
1573
1574         cpipe = fp->f_data;
1575         if (cpipe->pipe_type & PIPE_TYPE_NAMED)
1576                 error = vn_chown(fp, uid, gid, active_cred, td);
1577         else
1578                 error = invfo_chown(fp, uid, gid, active_cred, td);
1579         return (error);
1580 }
1581
1582 static int
1583 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1584 {
1585         struct pipe *pi;
1586
1587         if (fp->f_type == DTYPE_FIFO)
1588                 return (vn_fill_kinfo(fp, kif, fdp));
1589         kif->kf_type = KF_TYPE_PIPE;
1590         pi = fp->f_data;
1591         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1592         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1593         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1594         return (0);
1595 }
1596
1597 static void
1598 pipe_free_kmem(struct pipe *cpipe)
1599 {
1600
1601         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1602             ("pipe_free_kmem: pipe mutex locked"));
1603
1604         if (cpipe->pipe_buffer.buffer != NULL) {
1605                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1606                 vm_map_remove(pipe_map,
1607                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1608                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1609                 cpipe->pipe_buffer.buffer = NULL;
1610         }
1611 #ifndef PIPE_NODIRECT
1612         {
1613                 cpipe->pipe_pages.cnt = 0;
1614                 cpipe->pipe_pages.pos = 0;
1615                 cpipe->pipe_pages.npages = 0;
1616         }
1617 #endif
1618 }
1619
1620 /*
1621  * shutdown the pipe
1622  */
1623 static void
1624 pipeclose(struct pipe *cpipe)
1625 {
1626         struct pipepair *pp;
1627         struct pipe *ppipe;
1628
1629         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1630
1631         PIPE_LOCK(cpipe);
1632         pipelock(cpipe, 0);
1633         pp = cpipe->pipe_pair;
1634
1635         /*
1636          * If the other side is blocked, wake it up saying that
1637          * we want to close it down.
1638          */
1639         cpipe->pipe_state |= PIPE_EOF;
1640         while (cpipe->pipe_busy) {
1641                 wakeup(cpipe);
1642                 cpipe->pipe_state |= PIPE_WANT;
1643                 pipeunlock(cpipe);
1644                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1645                 pipelock(cpipe, 0);
1646         }
1647
1648         pipeselwakeup(cpipe);
1649
1650         /*
1651          * Disconnect from peer, if any.
1652          */
1653         ppipe = cpipe->pipe_peer;
1654         if (ppipe->pipe_present == PIPE_ACTIVE) {
1655                 ppipe->pipe_state |= PIPE_EOF;
1656                 wakeup(ppipe);
1657                 pipeselwakeup(ppipe);
1658         }
1659
1660         /*
1661          * Mark this endpoint as free.  Release kmem resources.  We
1662          * don't mark this endpoint as unused until we've finished
1663          * doing that, or the pipe might disappear out from under
1664          * us.
1665          */
1666         PIPE_UNLOCK(cpipe);
1667         pipe_free_kmem(cpipe);
1668         PIPE_LOCK(cpipe);
1669         cpipe->pipe_present = PIPE_CLOSING;
1670         pipeunlock(cpipe);
1671
1672         /*
1673          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1674          * PIPE_FINALIZED, that allows other end to free the
1675          * pipe_pair, only after the knotes are completely dismantled.
1676          */
1677         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1678         cpipe->pipe_present = PIPE_FINALIZED;
1679         seldrain(&cpipe->pipe_sel);
1680         knlist_destroy(&cpipe->pipe_sel.si_note);
1681
1682         /*
1683          * If both endpoints are now closed, release the memory for the
1684          * pipe pair.  If not, unlock.
1685          */
1686         if (ppipe->pipe_present == PIPE_FINALIZED) {
1687                 PIPE_UNLOCK(cpipe);
1688 #ifdef MAC
1689                 mac_pipe_destroy(pp);
1690 #endif
1691                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1692         } else
1693                 PIPE_UNLOCK(cpipe);
1694 }
1695
1696 /*ARGSUSED*/
1697 static int
1698 pipe_kqfilter(struct file *fp, struct knote *kn)
1699 {
1700         struct pipe *cpipe;
1701
1702         /*
1703          * If a filter is requested that is not supported by this file
1704          * descriptor, don't return an error, but also don't ever generate an
1705          * event.
1706          */
1707         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1708                 kn->kn_fop = &pipe_nfiltops;
1709                 return (0);
1710         }
1711         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1712                 kn->kn_fop = &pipe_nfiltops;
1713                 return (0);
1714         }
1715         cpipe = fp->f_data;
1716         PIPE_LOCK(cpipe);
1717         switch (kn->kn_filter) {
1718         case EVFILT_READ:
1719                 kn->kn_fop = &pipe_rfiltops;
1720                 break;
1721         case EVFILT_WRITE:
1722                 kn->kn_fop = &pipe_wfiltops;
1723                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1724                         /* other end of pipe has been closed */
1725                         PIPE_UNLOCK(cpipe);
1726                         return (EPIPE);
1727                 }
1728                 cpipe = PIPE_PEER(cpipe);
1729                 break;
1730         default:
1731                 PIPE_UNLOCK(cpipe);
1732                 return (EINVAL);
1733         }
1734
1735         kn->kn_hook = cpipe;
1736         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1737         PIPE_UNLOCK(cpipe);
1738         return (0);
1739 }
1740
1741 static void
1742 filt_pipedetach(struct knote *kn)
1743 {
1744         struct pipe *cpipe = kn->kn_hook;
1745
1746         PIPE_LOCK(cpipe);
1747         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1748         PIPE_UNLOCK(cpipe);
1749 }
1750
1751 /*ARGSUSED*/
1752 static int
1753 filt_piperead(struct knote *kn, long hint)
1754 {
1755         struct file *fp = kn->kn_fp;
1756         struct pipe *rpipe = kn->kn_hook;
1757
1758         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1759         kn->kn_data = rpipe->pipe_buffer.cnt;
1760         if (kn->kn_data == 0)
1761                 kn->kn_data = rpipe->pipe_pages.cnt;
1762
1763         if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1764             ((rpipe->pipe_type & PIPE_TYPE_NAMED) == 0 ||
1765             fp->f_pipegen != rpipe->pipe_wgen)) {
1766                 kn->kn_flags |= EV_EOF;
1767                 return (1);
1768         }
1769         kn->kn_flags &= ~EV_EOF;
1770         return (kn->kn_data > 0);
1771 }
1772
1773 /*ARGSUSED*/
1774 static int
1775 filt_pipewrite(struct knote *kn, long hint)
1776 {
1777         struct pipe *wpipe = kn->kn_hook;
1778
1779         /*
1780          * If this end of the pipe is closed, the knote was removed from the
1781          * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1782          */
1783         if (wpipe->pipe_present == PIPE_ACTIVE ||
1784             (wpipe->pipe_type & PIPE_TYPE_NAMED) != 0) {
1785                 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1786
1787                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1788                         kn->kn_data = 0;
1789                 } else if (wpipe->pipe_buffer.size > 0) {
1790                         kn->kn_data = wpipe->pipe_buffer.size -
1791                             wpipe->pipe_buffer.cnt;
1792                 } else {
1793                         kn->kn_data = PIPE_BUF;
1794                 }
1795         }
1796
1797         if (wpipe->pipe_present != PIPE_ACTIVE ||
1798             (wpipe->pipe_state & PIPE_EOF)) {
1799                 kn->kn_flags |= EV_EOF;
1800                 return (1);
1801         }
1802         kn->kn_flags &= ~EV_EOF;
1803         return (kn->kn_data >= PIPE_BUF);
1804 }
1805
1806 static void
1807 filt_pipedetach_notsup(struct knote *kn)
1808 {
1809
1810 }
1811
1812 static int
1813 filt_pipenotsup(struct knote *kn, long hint)
1814 {
1815
1816         return (0);
1817 }