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