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