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