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