sys/kern/sys_pipe.c

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