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, the sending process pins the underlying pages in
  33  * memory, and the receiving process copies directly from these pinned pages
  34  * in the sending process.
  35  *
  36  * If the sending process receives a signal, it is possible that it will
  37  * go away, and certainly its address space can change, because control
  38  * is returned back to the user-mode side.  In that case, the pipe code
  39  * arranges to copy the buffer supplied by the user process, to a pageable
  40  * kernel buffer, and the receiving process will grab the data from the
  41  * pageable kernel buffer.  Since signals don't happen all that often,
  42  * the copy operation is normally eliminated.
  43  *
  44  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  45  * happen for small transfers so that the system will not spend all of
  46  * its time context switching.
  47  *
  48  * In order to limit the resource use of pipes, two sysctls exist:
  49  *
  50  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  51  * address space available to us in pipe_map. This value is normally
  52  * autotuned, but may also be loader tuned.
  53  *
  54  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  55  * memory in use by pipes.
  56  *
  57  * Based on how large pipekva is relative to maxpipekva, the following
  58  * will happen:
  59  *
  60  * 0% - 50%:
  61  *     New pipes are given 16K of memory backing, pipes may dynamically
  62  *     grow to as large as 64K where needed.
  63  * 50% - 75%:
  64  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  65  *     existing pipes may NOT grow.
  66  * 75% - 100%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes will be shrunk down to 4K whenever possible.
  69  *
  70  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  71  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  72  * resize which MUST occur for reverse-direction pipes when they are
  73  * first used.
  74  *
  75  * Additional information about the current state of pipes may be obtained
  76  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  77  * and kern.ipc.piperesizefail.
  78  *
  79  * Locking rules:  There are two locks present here:  A mutex, used via
  80  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  81  * the flag, as mutexes can not persist over uiomove.  The mutex
  82  * exists only to guard access to the flag, and is not in itself a
  83  * locking mechanism.  Also note that there is only a single mutex for
  84  * both directions of a pipe.
  85  *
  86  * As pipelock() may have to sleep before it can acquire the flag, it
  87  * is important to reread all data after a call to pipelock(); everything
  88  * in the structure may have changed.
  89  */
  90
  91 #include <sys/cdefs.h>
  92 __FBSDID("$FreeBSD$");
  93
  94 #include <sys/param.h>
  95 #include <sys/systm.h>
  96 #include <sys/conf.h>
  97 #include <sys/fcntl.h>
  98 #include <sys/file.h>
  99 #include <sys/filedesc.h>
 100 #include <sys/filio.h>
 101 #include <sys/kernel.h>
 102 #include <sys/lock.h>
 103 #include <sys/mutex.h>
 104 #include <sys/ttycom.h>
 105 #include <sys/stat.h>
 106 #include <sys/malloc.h>
 107 #include <sys/poll.h>
 108 #include <sys/selinfo.h>
 109 #include <sys/signalvar.h>
 110 #include <sys/syscallsubr.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_chmod = invfo_chmod,
 160         .fo_chown = invfo_chown,
 161         .fo_flags = DFLAG_PASSABLE
 162 };
 163
 164 static void     filt_pipedetach(struct knote *kn);
 165 static int      filt_piperead(struct knote *kn, long hint);
 166 static int      filt_pipewrite(struct knote *kn, long hint);
 167
 168 static struct filterops pipe_rfiltops = {
 169         .f_isfd = 1,
 170         .f_detach = filt_pipedetach,
 171         .f_event = filt_piperead
 172 };
 173 static struct filterops pipe_wfiltops = {
 174         .f_isfd = 1,
 175         .f_detach = filt_pipedetach,
 176         .f_event = filt_pipewrite
 177 };
 178
 179 /*
 180  * Default pipe buffer size(s), this can be kind-of large now because pipe
 181  * space is pageable.  The pipe code will try to maintain locality of
 182  * reference for performance reasons, so small amounts of outstanding I/O
 183  * will not wipe the cache.
 184  */
 185 #define MINPIPESIZE (PIPE_SIZE/3)
 186 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 187
 188 static long amountpipekva;
 189 static int pipefragretry;
 190 static int pipeallocfail;
 191 static int piperesizefail;
 192 static int piperesizeallowed = 1;
 193
 194 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 195            &maxpipekva, 0, "Pipe KVA limit");
 196 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 197            &amountpipekva, 0, "Pipe KVA usage");
 198 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 199           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 200 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 201           &pipeallocfail, 0, "Pipe allocation failures");
 202 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 203           &piperesizefail, 0, "Pipe resize failures");
 204 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 205           &piperesizeallowed, 0, "Pipe resizing allowed");
 206
 207 static void pipeinit(void *dummy __unused);
 208 static void pipeclose(struct pipe *cpipe);
 209 static void pipe_free_kmem(struct pipe *cpipe);
 210 static int pipe_create(struct pipe *pipe, int backing);
 211 static __inline int pipelock(struct pipe *cpipe, int catch);
 212 static __inline void pipeunlock(struct pipe *cpipe);
 213 static __inline void pipeselwakeup(struct pipe *cpipe);
 214 #ifndef PIPE_NODIRECT
 215 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 216 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 217 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 218 static void pipe_clone_write_buffer(struct pipe *wpipe);
 219 #endif
 220 static int pipespace(struct pipe *cpipe, int size);
 221 static int pipespace_new(struct pipe *cpipe, int size);
 222
 223 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 224 static int      pipe_zone_init(void *mem, int size, int flags);
 225 static void     pipe_zone_fini(void *mem, int size);
 226
 227 static uma_zone_t pipe_zone;
 228 static struct unrhdr *pipeino_unr;
 229 static dev_t pipedev_ino;
 230
 231 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 232
 233 static void
 234 pipeinit(void *dummy __unused)
 235 {
 236
 237         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 238             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 239             UMA_ALIGN_PTR, 0);
 240         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 241         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 242         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 243         pipedev_ino = devfs_alloc_cdp_inode();
 244         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 245 }
 246
 247 static int
 248 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 249 {
 250         struct pipepair *pp;
 251         struct pipe *rpipe, *wpipe;
 252
 253         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 254
 255         pp = (struct pipepair *)mem;
 256
 257         /*
 258          * We zero both pipe endpoints to make sure all the kmem pointers
 259          * are NULL, flag fields are zero'd, etc.  We timestamp both
 260          * endpoints with the same time.
 261          */
 262         rpipe = &pp->pp_rpipe;
 263         bzero(rpipe, sizeof(*rpipe));
 264         vfs_timestamp(&rpipe->pipe_ctime);
 265         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 266
 267         wpipe = &pp->pp_wpipe;
 268         bzero(wpipe, sizeof(*wpipe));
 269         wpipe->pipe_ctime = rpipe->pipe_ctime;
 270         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 271
 272         rpipe->pipe_peer = wpipe;
 273         rpipe->pipe_pair = pp;
 274         wpipe->pipe_peer = rpipe;
 275         wpipe->pipe_pair = pp;
 276
 277         /*
 278          * Mark both endpoints as present; they will later get free'd
 279          * one at a time.  When both are free'd, then the whole pair
 280          * is released.
 281          */
 282         rpipe->pipe_present = PIPE_ACTIVE;
 283         wpipe->pipe_present = PIPE_ACTIVE;
 284
 285         /*
 286          * Eventually, the MAC Framework may initialize the label
 287          * in ctor or init, but for now we do it elswhere to avoid
 288          * blocking in ctor or init.
 289          */
 290         pp->pp_label = NULL;
 291
 292         return (0);
 293 }
 294
 295 static int
 296 pipe_zone_init(void *mem, int size, int flags)
 297 {
 298         struct pipepair *pp;
 299
 300         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 301
 302         pp = (struct pipepair *)mem;
 303
 304         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 305         return (0);
 306 }
 307
 308 static void
 309 pipe_zone_fini(void *mem, int size)
 310 {
 311         struct pipepair *pp;
 312
 313         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 314
 315         pp = (struct pipepair *)mem;
 316
 317         mtx_destroy(&pp->pp_mtx);
 318 }
 319
 320 /*
 321  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 322  * the zone pick up the pieces via pipeclose().
 323  */
 324 int
 325 kern_pipe(struct thread *td, int fildes[2])
 326 {
 327         struct filedesc *fdp = td->td_proc->p_fd;
 328         struct file *rf, *wf;
 329         struct pipepair *pp;
 330         struct pipe *rpipe, *wpipe;
 331         int fd, error;
 332
 333         pp = uma_zalloc(pipe_zone, M_WAITOK);
 334 #ifdef MAC
 335         /*
 336          * The MAC label is shared between the connected endpoints.  As a
 337          * result mac_pipe_init() and mac_pipe_create() are called once
 338          * for the pair, and not on the endpoints.
 339          */
 340         mac_pipe_init(pp);
 341         mac_pipe_create(td->td_ucred, pp);
 342 #endif
 343         rpipe = &pp->pp_rpipe;
 344         wpipe = &pp->pp_wpipe;
 345
 346         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 347         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 348
 349         /* Only the forward direction pipe is backed by default */
 350         if ((error = pipe_create(rpipe, 1)) != 0 ||
 351             (error = pipe_create(wpipe, 0)) != 0) {
 352                 pipeclose(rpipe);
 353                 pipeclose(wpipe);
 354                 return (error);
 355         }
 356
 357         rpipe->pipe_state |= PIPE_DIRECTOK;
 358         wpipe->pipe_state |= PIPE_DIRECTOK;
 359
 360         error = falloc(td, &rf, &fd, 0);
 361         if (error) {
 362                 pipeclose(rpipe);
 363                 pipeclose(wpipe);
 364                 return (error);
 365         }
 366         /* An extra reference on `rf' has been held for us by falloc(). */
 367         fildes[0] = fd;
 368
 369         /*
 370          * Warning: once we've gotten past allocation of the fd for the
 371          * read-side, we can only drop the read side via fdrop() in order
 372          * to avoid races against processes which manage to dup() the read
 373          * side while we are blocked trying to allocate the write side.
 374          */
 375         finit(rf, FREAD | FWRITE, DTYPE_PIPE, rpipe, &pipeops);
 376         error = falloc(td, &wf, &fd, 0);
 377         if (error) {
 378                 fdclose(fdp, rf, fildes[0], td);
 379                 fdrop(rf, td);
 380                 /* rpipe has been closed by fdrop(). */
 381                 pipeclose(wpipe);
 382                 return (error);
 383         }
 384         /* An extra reference on `wf' has been held for us by falloc(). */
 385         finit(wf, FREAD | FWRITE, DTYPE_PIPE, wpipe, &pipeops);
 386         fdrop(wf, td);
 387         fildes[1] = fd;
 388         fdrop(rf, td);
 389
 390         return (0);
 391 }
 392
 393 /* ARGSUSED */
 394 int
 395 sys_pipe(struct thread *td, struct pipe_args *uap)
 396 {
 397         int error;
 398         int fildes[2];
 399
 400         error = kern_pipe(td, fildes);
 401         if (error)
 402                 return (error);
 403
 404         td->td_retval[0] = fildes[0];
 405         td->td_retval[1] = fildes[1];
 406
 407         return (0);
 408 }
 409
 410 /*
 411  * Allocate kva for pipe circular buffer, the space is pageable
 412  * This routine will 'realloc' the size of a pipe safely, if it fails
 413  * it will retain the old buffer.
 414  * If it fails it will return ENOMEM.
 415  */
 416 static int
 417 pipespace_new(cpipe, size)
 418         struct pipe *cpipe;
 419         int size;
 420 {
 421         caddr_t buffer;
 422         int error, cnt, firstseg;
 423         static int curfail = 0;
 424         static struct timeval lastfail;
 425
 426         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 427         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 428                 ("pipespace: resize of direct writes not allowed"));
 429 retry:
 430         cnt = cpipe->pipe_buffer.cnt;
 431         if (cnt > size)
 432                 size = cnt;
 433
 434         size = round_page(size);
 435         buffer = (caddr_t) vm_map_min(pipe_map);
 436
 437         error = vm_map_find(pipe_map, NULL, 0,
 438                 (vm_offset_t *) &buffer, size, 1,
 439                 VM_PROT_ALL, VM_PROT_ALL, 0);
 440         if (error != KERN_SUCCESS) {
 441                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 442                         (size > SMALL_PIPE_SIZE)) {
 443                         size = SMALL_PIPE_SIZE;
 444                         pipefragretry++;
 445                         goto retry;
 446                 }
 447                 if (cpipe->pipe_buffer.buffer == NULL) {
 448                         pipeallocfail++;
 449                         if (ppsratecheck(&lastfail, &curfail, 1))
 450                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 451                 } else {
 452                         piperesizefail++;
 453                 }
 454                 return (ENOMEM);
 455         }
 456
 457         /* copy data, then free old resources if we're resizing */
 458         if (cnt > 0) {
 459                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 460                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 461                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 462                                 buffer, firstseg);
 463                         if ((cnt - firstseg) > 0)
 464                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 465                                         cpipe->pipe_buffer.in);
 466                 } else {
 467                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 468                                 buffer, cnt);
 469                 }
 470         }
 471         pipe_free_kmem(cpipe);
 472         cpipe->pipe_buffer.buffer = buffer;
 473         cpipe->pipe_buffer.size = size;
 474         cpipe->pipe_buffer.in = cnt;
 475         cpipe->pipe_buffer.out = 0;
 476         cpipe->pipe_buffer.cnt = cnt;
 477         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 478         return (0);
 479 }
 480
 481 /*
 482  * Wrapper for pipespace_new() that performs locking assertions.
 483  */
 484 static int
 485 pipespace(cpipe, size)
 486         struct pipe *cpipe;
 487         int size;
 488 {
 489
 490         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 491                 ("Unlocked pipe passed to pipespace"));
 492         return (pipespace_new(cpipe, size));
 493 }
 494
 495 /*
 496  * lock a pipe for I/O, blocking other access
 497  */
 498 static __inline int
 499 pipelock(cpipe, catch)
 500         struct pipe *cpipe;
 501         int catch;
 502 {
 503         int error;
 504
 505         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 506         while (cpipe->pipe_state & PIPE_LOCKFL) {
 507                 cpipe->pipe_state |= PIPE_LWANT;
 508                 error = msleep(cpipe, PIPE_MTX(cpipe),
 509                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 510                     "pipelk", 0);
 511                 if (error != 0)
 512                         return (error);
 513         }
 514         cpipe->pipe_state |= PIPE_LOCKFL;
 515         return (0);
 516 }
 517
 518 /*
 519  * unlock a pipe I/O lock
 520  */
 521 static __inline void
 522 pipeunlock(cpipe)
 523         struct pipe *cpipe;
 524 {
 525
 526         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 527         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 528                 ("Unlocked pipe passed to pipeunlock"));
 529         cpipe->pipe_state &= ~PIPE_LOCKFL;
 530         if (cpipe->pipe_state & PIPE_LWANT) {
 531                 cpipe->pipe_state &= ~PIPE_LWANT;
 532                 wakeup(cpipe);
 533         }
 534 }
 535
 536 static __inline void
 537 pipeselwakeup(cpipe)
 538         struct pipe *cpipe;
 539 {
 540
 541         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 542         if (cpipe->pipe_state & PIPE_SEL) {
 543                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 544                 if (!SEL_WAITING(&cpipe->pipe_sel))
 545                         cpipe->pipe_state &= ~PIPE_SEL;
 546         }
 547         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 548                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 549         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 550 }
 551
 552 /*
 553  * Initialize and allocate VM and memory for pipe.  The structure
 554  * will start out zero'd from the ctor, so we just manage the kmem.
 555  */
 556 static int
 557 pipe_create(pipe, backing)
 558         struct pipe *pipe;
 559         int backing;
 560 {
 561         int error;
 562
 563         if (backing) {
 564                 if (amountpipekva > maxpipekva / 2)
 565                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 566                 else
 567                         error = pipespace_new(pipe, PIPE_SIZE);
 568         } else {
 569                 /* If we're not backing this pipe, no need to do anything. */
 570                 error = 0;
 571         }
 572         pipe->pipe_ino = -1;
 573         return (error);
 574 }
 575
 576 /* ARGSUSED */
 577 static int
 578 pipe_read(fp, uio, active_cred, flags, td)
 579         struct file *fp;
 580         struct uio *uio;
 581         struct ucred *active_cred;
 582         struct thread *td;
 583         int flags;
 584 {
 585         struct pipe *rpipe = fp->f_data;
 586         int error;
 587         int nread = 0;
 588         int size;
 589
 590         PIPE_LOCK(rpipe);
 591         ++rpipe->pipe_busy;
 592         error = pipelock(rpipe, 1);
 593         if (error)
 594                 goto unlocked_error;
 595
 596 #ifdef MAC
 597         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 598         if (error)
 599                 goto locked_error;
 600 #endif
 601         if (amountpipekva > (3 * maxpipekva) / 4) {
 602                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 603                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 604                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 605                         (piperesizeallowed == 1)) {
 606                         PIPE_UNLOCK(rpipe);
 607                         pipespace(rpipe, SMALL_PIPE_SIZE);
 608                         PIPE_LOCK(rpipe);
 609                 }
 610         }
 611
 612         while (uio->uio_resid) {
 613                 /*
 614                  * normal pipe buffer receive
 615                  */
 616                 if (rpipe->pipe_buffer.cnt > 0) {
 617                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 618                         if (size > rpipe->pipe_buffer.cnt)
 619                                 size = rpipe->pipe_buffer.cnt;
 620                         if (size > uio->uio_resid)
 621                                 size = uio->uio_resid;
 622
 623                         PIPE_UNLOCK(rpipe);
 624                         error = uiomove(
 625                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 626                             size, uio);
 627                         PIPE_LOCK(rpipe);
 628                         if (error)
 629                                 break;
 630
 631                         rpipe->pipe_buffer.out += size;
 632                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 633                                 rpipe->pipe_buffer.out = 0;
 634
 635                         rpipe->pipe_buffer.cnt -= size;
 636
 637                         /*
 638                          * If there is no more to read in the pipe, reset
 639                          * its pointers to the beginning.  This improves
 640                          * cache hit stats.
 641                          */
 642                         if (rpipe->pipe_buffer.cnt == 0) {
 643                                 rpipe->pipe_buffer.in = 0;
 644                                 rpipe->pipe_buffer.out = 0;
 645                         }
 646                         nread += size;
 647 #ifndef PIPE_NODIRECT
 648                 /*
 649                  * Direct copy, bypassing a kernel buffer.
 650                  */
 651                 } else if ((size = rpipe->pipe_map.cnt) &&
 652                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 653                         if (size > uio->uio_resid)
 654                                 size = (u_int) uio->uio_resid;
 655
 656                         PIPE_UNLOCK(rpipe);
 657                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 658                             rpipe->pipe_map.pos, size, uio);
 659                         PIPE_LOCK(rpipe);
 660                         if (error)
 661                                 break;
 662                         nread += size;
 663                         rpipe->pipe_map.pos += size;
 664                         rpipe->pipe_map.cnt -= size;
 665                         if (rpipe->pipe_map.cnt == 0) {
 666                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 667                                 wakeup(rpipe);
 668                         }
 669 #endif
 670                 } else {
 671                         /*
 672                          * detect EOF condition
 673                          * read returns 0 on EOF, no need to set error
 674                          */
 675                         if (rpipe->pipe_state & PIPE_EOF)
 676                                 break;
 677
 678                         /*
 679                          * If the "write-side" has been blocked, wake it up now.
 680                          */
 681                         if (rpipe->pipe_state & PIPE_WANTW) {
 682                                 rpipe->pipe_state &= ~PIPE_WANTW;
 683                                 wakeup(rpipe);
 684                         }
 685
 686                         /*
 687                          * Break if some data was read.
 688                          */
 689                         if (nread > 0)
 690                                 break;
 691
 692                         /*
 693                          * Unlock the pipe buffer for our remaining processing.
 694                          * We will either break out with an error or we will
 695                          * sleep and relock to loop.
 696                          */
 697                         pipeunlock(rpipe);
 698
 699                         /*
 700                          * Handle non-blocking mode operation or
 701                          * wait for more data.
 702                          */
 703                         if (fp->f_flag & FNONBLOCK) {
 704                                 error = EAGAIN;
 705                         } else {
 706                                 rpipe->pipe_state |= PIPE_WANTR;
 707                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 708                                     PRIBIO | PCATCH,
 709                                     "piperd", 0)) == 0)
 710                                         error = pipelock(rpipe, 1);
 711                         }
 712                         if (error)
 713                                 goto unlocked_error;
 714                 }
 715         }
 716 #ifdef MAC
 717 locked_error:
 718 #endif
 719         pipeunlock(rpipe);
 720
 721         /* XXX: should probably do this before getting any locks. */
 722         if (error == 0)
 723                 vfs_timestamp(&rpipe->pipe_atime);
 724 unlocked_error:
 725         --rpipe->pipe_busy;
 726
 727         /*
 728          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 729          */
 730         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 731                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 732                 wakeup(rpipe);
 733         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 734                 /*
 735                  * Handle write blocking hysteresis.
 736                  */
 737                 if (rpipe->pipe_state & PIPE_WANTW) {
 738                         rpipe->pipe_state &= ~PIPE_WANTW;
 739                         wakeup(rpipe);
 740                 }
 741         }
 742
 743         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 744                 pipeselwakeup(rpipe);
 745
 746         PIPE_UNLOCK(rpipe);
 747         return (error);
 748 }
 749
 750 #ifndef PIPE_NODIRECT
 751 /*
 752  * Map the sending processes' buffer into kernel space and wire it.
 753  * This is similar to a physical write operation.
 754  */
 755 static int
 756 pipe_build_write_buffer(wpipe, uio)
 757         struct pipe *wpipe;
 758         struct uio *uio;
 759 {
 760         u_int size;
 761         int i;
 762
 763         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 764         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 765                 ("Clone attempt on non-direct write pipe!"));
 766
 767         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 768                 size = wpipe->pipe_buffer.size;
 769         else
 770                 size = uio->uio_iov->iov_len;
 771
 772         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 773             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 774             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 775                 return (EFAULT);
 776
 777 /*
 778  * set up the control block
 779  */
 780         wpipe->pipe_map.npages = i;
 781         wpipe->pipe_map.pos =
 782             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 783         wpipe->pipe_map.cnt = size;
 784
 785 /*
 786  * and update the uio data
 787  */
 788
 789         uio->uio_iov->iov_len -= size;
 790         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 791         if (uio->uio_iov->iov_len == 0)
 792                 uio->uio_iov++;
 793         uio->uio_resid -= size;
 794         uio->uio_offset += size;
 795         return (0);
 796 }
 797
 798 /*
 799  * unmap and unwire the process buffer
 800  */
 801 static void
 802 pipe_destroy_write_buffer(wpipe)
 803         struct pipe *wpipe;
 804 {
 805
 806         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 807         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 808         wpipe->pipe_map.npages = 0;
 809 }
 810
 811 /*
 812  * In the case of a signal, the writing process might go away.  This
 813  * code copies the data into the circular buffer so that the source
 814  * pages can be freed without loss of data.
 815  */
 816 static void
 817 pipe_clone_write_buffer(wpipe)
 818         struct pipe *wpipe;
 819 {
 820         struct uio uio;
 821         struct iovec iov;
 822         int size;
 823         int pos;
 824
 825         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 826         size = wpipe->pipe_map.cnt;
 827         pos = wpipe->pipe_map.pos;
 828
 829         wpipe->pipe_buffer.in = size;
 830         wpipe->pipe_buffer.out = 0;
 831         wpipe->pipe_buffer.cnt = size;
 832         wpipe->pipe_state &= ~PIPE_DIRECTW;
 833
 834         PIPE_UNLOCK(wpipe);
 835         iov.iov_base = wpipe->pipe_buffer.buffer;
 836         iov.iov_len = size;
 837         uio.uio_iov = &iov;
 838         uio.uio_iovcnt = 1;
 839         uio.uio_offset = 0;
 840         uio.uio_resid = size;
 841         uio.uio_segflg = UIO_SYSSPACE;
 842         uio.uio_rw = UIO_READ;
 843         uio.uio_td = curthread;
 844         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 845         PIPE_LOCK(wpipe);
 846         pipe_destroy_write_buffer(wpipe);
 847 }
 848
 849 /*
 850  * This implements the pipe buffer write mechanism.  Note that only
 851  * a direct write OR a normal pipe write can be pending at any given time.
 852  * If there are any characters in the pipe buffer, the direct write will
 853  * be deferred until the receiving process grabs all of the bytes from
 854  * the pipe buffer.  Then the direct mapping write is set-up.
 855  */
 856 static int
 857 pipe_direct_write(wpipe, uio)
 858         struct pipe *wpipe;
 859         struct uio *uio;
 860 {
 861         int error;
 862
 863 retry:
 864         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 865         error = pipelock(wpipe, 1);
 866         if (wpipe->pipe_state & PIPE_EOF)
 867                 error = EPIPE;
 868         if (error) {
 869                 pipeunlock(wpipe);
 870                 goto error1;
 871         }
 872         while (wpipe->pipe_state & PIPE_DIRECTW) {
 873                 if (wpipe->pipe_state & PIPE_WANTR) {
 874                         wpipe->pipe_state &= ~PIPE_WANTR;
 875                         wakeup(wpipe);
 876                 }
 877                 pipeselwakeup(wpipe);
 878                 wpipe->pipe_state |= PIPE_WANTW;
 879                 pipeunlock(wpipe);
 880                 error = msleep(wpipe, PIPE_MTX(wpipe),
 881                     PRIBIO | PCATCH, "pipdww", 0);
 882                 if (error)
 883                         goto error1;
 884                 else
 885                         goto retry;
 886         }
 887         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 888         if (wpipe->pipe_buffer.cnt > 0) {
 889                 if (wpipe->pipe_state & PIPE_WANTR) {
 890                         wpipe->pipe_state &= ~PIPE_WANTR;
 891                         wakeup(wpipe);
 892                 }
 893                 pipeselwakeup(wpipe);
 894                 wpipe->pipe_state |= PIPE_WANTW;
 895                 pipeunlock(wpipe);
 896                 error = msleep(wpipe, PIPE_MTX(wpipe),
 897                     PRIBIO | PCATCH, "pipdwc", 0);
 898                 if (error)
 899                         goto error1;
 900                 else
 901                         goto retry;
 902         }
 903
 904         wpipe->pipe_state |= PIPE_DIRECTW;
 905
 906         PIPE_UNLOCK(wpipe);
 907         error = pipe_build_write_buffer(wpipe, uio);
 908         PIPE_LOCK(wpipe);
 909         if (error) {
 910                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 911                 pipeunlock(wpipe);
 912                 goto error1;
 913         }
 914
 915         error = 0;
 916         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 917                 if (wpipe->pipe_state & PIPE_EOF) {
 918                         pipe_destroy_write_buffer(wpipe);
 919                         pipeselwakeup(wpipe);
 920                         pipeunlock(wpipe);
 921                         error = EPIPE;
 922                         goto error1;
 923                 }
 924                 if (wpipe->pipe_state & PIPE_WANTR) {
 925                         wpipe->pipe_state &= ~PIPE_WANTR;
 926                         wakeup(wpipe);
 927                 }
 928                 pipeselwakeup(wpipe);
 929                 pipeunlock(wpipe);
 930                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 931                     "pipdwt", 0);
 932                 pipelock(wpipe, 0);
 933         }
 934
 935         if (wpipe->pipe_state & PIPE_EOF)
 936                 error = EPIPE;
 937         if (wpipe->pipe_state & PIPE_DIRECTW) {
 938                 /*
 939                  * this bit of trickery substitutes a kernel buffer for
 940                  * the process that might be going away.
 941                  */
 942                 pipe_clone_write_buffer(wpipe);
 943         } else {
 944                 pipe_destroy_write_buffer(wpipe);
 945         }
 946         pipeunlock(wpipe);
 947         return (error);
 948
 949 error1:
 950         wakeup(wpipe);
 951         return (error);
 952 }
 953 #endif
 954
 955 static int
 956 pipe_write(fp, uio, active_cred, flags, td)
 957         struct file *fp;
 958         struct uio *uio;
 959         struct ucred *active_cred;
 960         struct thread *td;
 961         int flags;
 962 {
 963         int error = 0;
 964         int desiredsize;
 965         ssize_t orig_resid;
 966         struct pipe *wpipe, *rpipe;
 967
 968         rpipe = fp->f_data;
 969         wpipe = rpipe->pipe_peer;
 970
 971         PIPE_LOCK(rpipe);
 972         error = pipelock(wpipe, 1);
 973         if (error) {
 974                 PIPE_UNLOCK(rpipe);
 975                 return (error);
 976         }
 977         /*
 978          * detect loss of pipe read side, issue SIGPIPE if lost.
 979          */
 980         if (wpipe->pipe_present != PIPE_ACTIVE ||
 981             (wpipe->pipe_state & PIPE_EOF)) {
 982                 pipeunlock(wpipe);
 983                 PIPE_UNLOCK(rpipe);
 984                 return (EPIPE);
 985         }
 986 #ifdef MAC
 987         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
 988         if (error) {
 989                 pipeunlock(wpipe);
 990                 PIPE_UNLOCK(rpipe);
 991                 return (error);
 992         }
 993 #endif
 994         ++wpipe->pipe_busy;
 995
 996         /* Choose a larger size if it's advantageous */
 997         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
 998         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
 999                 if (piperesizeallowed != 1)
1000                         break;
1001                 if (amountpipekva > maxpipekva / 2)
1002                         break;
1003                 if (desiredsize == BIG_PIPE_SIZE)
1004                         break;
1005                 desiredsize = desiredsize * 2;
1006         }
1007
1008         /* Choose a smaller size if we're in a OOM situation */
1009         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1010                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1011                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1012                 (piperesizeallowed == 1))
1013                 desiredsize = SMALL_PIPE_SIZE;
1014
1015         /* Resize if the above determined that a new size was necessary */
1016         if ((desiredsize != wpipe->pipe_buffer.size) &&
1017                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1018                 PIPE_UNLOCK(wpipe);
1019                 pipespace(wpipe, desiredsize);
1020                 PIPE_LOCK(wpipe);
1021         }
1022         if (wpipe->pipe_buffer.size == 0) {
1023                 /*
1024                  * This can only happen for reverse direction use of pipes
1025                  * in a complete OOM situation.
1026                  */
1027                 error = ENOMEM;
1028                 --wpipe->pipe_busy;
1029                 pipeunlock(wpipe);
1030                 PIPE_UNLOCK(wpipe);
1031                 return (error);
1032         }
1033
1034         pipeunlock(wpipe);
1035
1036         orig_resid = uio->uio_resid;
1037
1038         while (uio->uio_resid) {
1039                 int space;
1040
1041                 pipelock(wpipe, 0);
1042                 if (wpipe->pipe_state & PIPE_EOF) {
1043                         pipeunlock(wpipe);
1044                         error = EPIPE;
1045                         break;
1046                 }
1047 #ifndef PIPE_NODIRECT
1048                 /*
1049                  * If the transfer is large, we can gain performance if
1050                  * we do process-to-process copies directly.
1051                  * If the write is non-blocking, we don't use the
1052                  * direct write mechanism.
1053                  *
1054                  * The direct write mechanism will detect the reader going
1055                  * away on us.
1056                  */
1057                 if (uio->uio_segflg == UIO_USERSPACE &&
1058                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1059                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1060                     (fp->f_flag & FNONBLOCK) == 0) {
1061                         pipeunlock(wpipe);
1062                         error = pipe_direct_write(wpipe, uio);
1063                         if (error)
1064                                 break;
1065                         continue;
1066                 }
1067 #endif
1068
1069                 /*
1070                  * Pipe buffered writes cannot be coincidental with
1071                  * direct writes.  We wait until the currently executing
1072                  * direct write is completed before we start filling the
1073                  * pipe buffer.  We break out if a signal occurs or the
1074                  * reader goes away.
1075                  */
1076                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1077                         if (wpipe->pipe_state & PIPE_WANTR) {
1078                                 wpipe->pipe_state &= ~PIPE_WANTR;
1079                                 wakeup(wpipe);
1080                         }
1081                         pipeselwakeup(wpipe);
1082                         wpipe->pipe_state |= PIPE_WANTW;
1083                         pipeunlock(wpipe);
1084                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1085                             "pipbww", 0);
1086                         if (error)
1087                                 break;
1088                         else
1089                                 continue;
1090                 }
1091
1092                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1093
1094                 /* Writes of size <= PIPE_BUF must be atomic. */
1095                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1096                         space = 0;
1097
1098                 if (space > 0) {
1099                         int size;       /* Transfer size */
1100                         int segsize;    /* first segment to transfer */
1101
1102                         /*
1103                          * Transfer size is minimum of uio transfer
1104                          * and free space in pipe buffer.
1105                          */
1106                         if (space > uio->uio_resid)
1107                                 size = uio->uio_resid;
1108                         else
1109                                 size = space;
1110                         /*
1111                          * First segment to transfer is minimum of
1112                          * transfer size and contiguous space in
1113                          * pipe buffer.  If first segment to transfer
1114                          * is less than the transfer size, we've got
1115                          * a wraparound in the buffer.
1116                          */
1117                         segsize = wpipe->pipe_buffer.size -
1118                                 wpipe->pipe_buffer.in;
1119                         if (segsize > size)
1120                                 segsize = size;
1121
1122                         /* Transfer first segment */
1123
1124                         PIPE_UNLOCK(rpipe);
1125                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1126                                         segsize, uio);
1127                         PIPE_LOCK(rpipe);
1128
1129                         if (error == 0 && segsize < size) {
1130                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1131                                         wpipe->pipe_buffer.size,
1132                                         ("Pipe buffer wraparound disappeared"));
1133                                 /*
1134                                  * Transfer remaining part now, to
1135                                  * support atomic writes.  Wraparound
1136                                  * happened.
1137                                  */
1138
1139                                 PIPE_UNLOCK(rpipe);
1140                                 error = uiomove(
1141                                     &wpipe->pipe_buffer.buffer[0],
1142                                     size - segsize, uio);
1143                                 PIPE_LOCK(rpipe);
1144                         }
1145                         if (error == 0) {
1146                                 wpipe->pipe_buffer.in += size;
1147                                 if (wpipe->pipe_buffer.in >=
1148                                     wpipe->pipe_buffer.size) {
1149                                         KASSERT(wpipe->pipe_buffer.in ==
1150                                                 size - segsize +
1151                                                 wpipe->pipe_buffer.size,
1152                                                 ("Expected wraparound bad"));
1153                                         wpipe->pipe_buffer.in = size - segsize;
1154                                 }
1155
1156                                 wpipe->pipe_buffer.cnt += size;
1157                                 KASSERT(wpipe->pipe_buffer.cnt <=
1158                                         wpipe->pipe_buffer.size,
1159                                         ("Pipe buffer overflow"));
1160                         }
1161                         pipeunlock(wpipe);
1162                         if (error != 0)
1163                                 break;
1164                 } else {
1165                         /*
1166                          * If the "read-side" has been blocked, wake it up now.
1167                          */
1168                         if (wpipe->pipe_state & PIPE_WANTR) {
1169                                 wpipe->pipe_state &= ~PIPE_WANTR;
1170                                 wakeup(wpipe);
1171                         }
1172
1173                         /*
1174                          * don't block on non-blocking I/O
1175                          */
1176                         if (fp->f_flag & FNONBLOCK) {
1177                                 error = EAGAIN;
1178                                 pipeunlock(wpipe);
1179                                 break;
1180                         }
1181
1182                         /*
1183                          * We have no more space and have something to offer,
1184                          * wake up select/poll.
1185                          */
1186                         pipeselwakeup(wpipe);
1187
1188                         wpipe->pipe_state |= PIPE_WANTW;
1189                         pipeunlock(wpipe);
1190                         error = msleep(wpipe, PIPE_MTX(rpipe),
1191                             PRIBIO | PCATCH, "pipewr", 0);
1192                         if (error != 0)
1193                                 break;
1194                 }
1195         }
1196
1197         pipelock(wpipe, 0);
1198         --wpipe->pipe_busy;
1199
1200         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1201                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1202                 wakeup(wpipe);
1203         } else if (wpipe->pipe_buffer.cnt > 0) {
1204                 /*
1205                  * If we have put any characters in the buffer, we wake up
1206                  * the reader.
1207                  */
1208                 if (wpipe->pipe_state & PIPE_WANTR) {
1209                         wpipe->pipe_state &= ~PIPE_WANTR;
1210                         wakeup(wpipe);
1211                 }
1212         }
1213
1214         /*
1215          * Don't return EPIPE if I/O was successful
1216          */
1217         if ((wpipe->pipe_buffer.cnt == 0) &&
1218             (uio->uio_resid == 0) &&
1219             (error == EPIPE)) {
1220                 error = 0;
1221         }
1222
1223         if (error == 0)
1224                 vfs_timestamp(&wpipe->pipe_mtime);
1225
1226         /*
1227          * We have something to offer,
1228          * wake up select/poll.
1229          */
1230         if (wpipe->pipe_buffer.cnt)
1231                 pipeselwakeup(wpipe);
1232
1233         pipeunlock(wpipe);
1234         PIPE_UNLOCK(rpipe);
1235         return (error);
1236 }
1237
1238 /* ARGSUSED */
1239 static int
1240 pipe_truncate(fp, length, active_cred, td)
1241         struct file *fp;
1242         off_t length;
1243         struct ucred *active_cred;
1244         struct thread *td;
1245 {
1246
1247         return (EINVAL);
1248 }
1249
1250 /*
1251  * we implement a very minimal set of ioctls for compatibility with sockets.
1252  */
1253 static int
1254 pipe_ioctl(fp, cmd, data, active_cred, td)
1255         struct file *fp;
1256         u_long cmd;
1257         void *data;
1258         struct ucred *active_cred;
1259         struct thread *td;
1260 {
1261         struct pipe *mpipe = fp->f_data;
1262         int error;
1263
1264         PIPE_LOCK(mpipe);
1265
1266 #ifdef MAC
1267         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1268         if (error) {
1269                 PIPE_UNLOCK(mpipe);
1270                 return (error);
1271         }
1272 #endif
1273
1274         error = 0;
1275         switch (cmd) {
1276
1277         case FIONBIO:
1278                 break;
1279
1280         case FIOASYNC:
1281                 if (*(int *)data) {
1282                         mpipe->pipe_state |= PIPE_ASYNC;
1283                 } else {
1284                         mpipe->pipe_state &= ~PIPE_ASYNC;
1285                 }
1286                 break;
1287
1288         case FIONREAD:
1289                 if (mpipe->pipe_state & PIPE_DIRECTW)
1290                         *(int *)data = mpipe->pipe_map.cnt;
1291                 else
1292                         *(int *)data = mpipe->pipe_buffer.cnt;
1293                 break;
1294
1295         case FIOSETOWN:
1296                 PIPE_UNLOCK(mpipe);
1297                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1298                 goto out_unlocked;
1299
1300         case FIOGETOWN:
1301                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1302                 break;
1303
1304         /* This is deprecated, FIOSETOWN should be used instead. */
1305         case TIOCSPGRP:
1306                 PIPE_UNLOCK(mpipe);
1307                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1308                 goto out_unlocked;
1309
1310         /* This is deprecated, FIOGETOWN should be used instead. */
1311         case TIOCGPGRP:
1312                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1313                 break;
1314
1315         default:
1316                 error = ENOTTY;
1317                 break;
1318         }
1319         PIPE_UNLOCK(mpipe);
1320 out_unlocked:
1321         return (error);
1322 }
1323
1324 static int
1325 pipe_poll(fp, events, active_cred, td)
1326         struct file *fp;
1327         int events;
1328         struct ucred *active_cred;
1329         struct thread *td;
1330 {
1331         struct pipe *rpipe = fp->f_data;
1332         struct pipe *wpipe;
1333         int revents = 0;
1334 #ifdef MAC
1335         int error;
1336 #endif
1337
1338         wpipe = rpipe->pipe_peer;
1339         PIPE_LOCK(rpipe);
1340 #ifdef MAC
1341         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1342         if (error)
1343                 goto locked_error;
1344 #endif
1345         if (events & (POLLIN | POLLRDNORM))
1346                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1347                     (rpipe->pipe_buffer.cnt > 0))
1348                         revents |= events & (POLLIN | POLLRDNORM);
1349
1350         if (events & (POLLOUT | POLLWRNORM))
1351                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1352                     (wpipe->pipe_state & PIPE_EOF) ||
1353                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1354                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1355                          wpipe->pipe_buffer.size == 0)))
1356                         revents |= events & (POLLOUT | POLLWRNORM);
1357
1358         if ((events & POLLINIGNEOF) == 0) {
1359                 if (rpipe->pipe_state & PIPE_EOF) {
1360                         revents |= (events & (POLLIN | POLLRDNORM));
1361                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1362                             (wpipe->pipe_state & PIPE_EOF))
1363                                 revents |= POLLHUP;
1364                 }
1365         }
1366
1367         if (revents == 0) {
1368                 if (events & (POLLIN | POLLRDNORM)) {
1369                         selrecord(td, &rpipe->pipe_sel);
1370                         if (SEL_WAITING(&rpipe->pipe_sel))
1371                                 rpipe->pipe_state |= PIPE_SEL;
1372                 }
1373
1374                 if (events & (POLLOUT | POLLWRNORM)) {
1375                         selrecord(td, &wpipe->pipe_sel);
1376                         if (SEL_WAITING(&wpipe->pipe_sel))
1377                                 wpipe->pipe_state |= PIPE_SEL;
1378                 }
1379         }
1380 #ifdef MAC
1381 locked_error:
1382 #endif
1383         PIPE_UNLOCK(rpipe);
1384
1385         return (revents);
1386 }
1387
1388 /*
1389  * We shouldn't need locks here as we're doing a read and this should
1390  * be a natural race.
1391  */
1392 static int
1393 pipe_stat(fp, ub, active_cred, td)
1394         struct file *fp;
1395         struct stat *ub;
1396         struct ucred *active_cred;
1397         struct thread *td;
1398 {
1399         struct pipe *pipe;
1400         int new_unr;
1401 #ifdef MAC
1402         int error;
1403 #endif
1404
1405         pipe = fp->f_data;
1406         PIPE_LOCK(pipe);
1407 #ifdef MAC
1408         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1409         if (error) {
1410                 PIPE_UNLOCK(pipe);
1411                 return (error);
1412         }
1413 #endif
1414         /*
1415          * Lazily allocate an inode number for the pipe.  Most pipe
1416          * users do not call fstat(2) on the pipe, which means that
1417          * postponing the inode allocation until it is must be
1418          * returned to userland is useful.  If alloc_unr failed,
1419          * assign st_ino zero instead of returning an error.
1420          * Special pipe_ino values:
1421          *  -1 - not yet initialized;
1422          *  0  - alloc_unr failed, return 0 as st_ino forever.
1423          */
1424         if (pipe->pipe_ino == (ino_t)-1) {
1425                 new_unr = alloc_unr(pipeino_unr);
1426                 if (new_unr != -1)
1427                         pipe->pipe_ino = new_unr;
1428                 else
1429                         pipe->pipe_ino = 0;
1430         }
1431         PIPE_UNLOCK(pipe);
1432
1433         bzero(ub, sizeof(*ub));
1434         ub->st_mode = S_IFIFO;
1435         ub->st_blksize = PAGE_SIZE;
1436         if (pipe->pipe_state & PIPE_DIRECTW)
1437                 ub->st_size = pipe->pipe_map.cnt;
1438         else
1439                 ub->st_size = pipe->pipe_buffer.cnt;
1440         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1441         ub->st_atim = pipe->pipe_atime;
1442         ub->st_mtim = pipe->pipe_mtime;
1443         ub->st_ctim = pipe->pipe_ctime;
1444         ub->st_uid = fp->f_cred->cr_uid;
1445         ub->st_gid = fp->f_cred->cr_gid;
1446         ub->st_dev = pipedev_ino;
1447         ub->st_ino = pipe->pipe_ino;
1448         /*
1449          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1450          */
1451         return (0);
1452 }
1453
1454 /* ARGSUSED */
1455 static int
1456 pipe_close(fp, td)
1457         struct file *fp;
1458         struct thread *td;
1459 {
1460         struct pipe *cpipe = fp->f_data;
1461
1462         fp->f_ops = &badfileops;
1463         fp->f_data = NULL;
1464         funsetown(&cpipe->pipe_sigio);
1465         pipeclose(cpipe);
1466         return (0);
1467 }
1468
1469 static void
1470 pipe_free_kmem(cpipe)
1471         struct pipe *cpipe;
1472 {
1473
1474         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1475             ("pipe_free_kmem: pipe mutex locked"));
1476
1477         if (cpipe->pipe_buffer.buffer != NULL) {
1478                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1479                 vm_map_remove(pipe_map,
1480                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1481                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1482                 cpipe->pipe_buffer.buffer = NULL;
1483         }
1484 #ifndef PIPE_NODIRECT
1485         {
1486                 cpipe->pipe_map.cnt = 0;
1487                 cpipe->pipe_map.pos = 0;
1488                 cpipe->pipe_map.npages = 0;
1489         }
1490 #endif
1491 }
1492
1493 /*
1494  * shutdown the pipe
1495  */
1496 static void
1497 pipeclose(cpipe)
1498         struct pipe *cpipe;
1499 {
1500         struct pipepair *pp;
1501         struct pipe *ppipe;
1502         ino_t ino;
1503
1504         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1505
1506         PIPE_LOCK(cpipe);
1507         pipelock(cpipe, 0);
1508         pp = cpipe->pipe_pair;
1509
1510         pipeselwakeup(cpipe);
1511
1512         /*
1513          * If the other side is blocked, wake it up saying that
1514          * we want to close it down.
1515          */
1516         cpipe->pipe_state |= PIPE_EOF;
1517         while (cpipe->pipe_busy) {
1518                 wakeup(cpipe);
1519                 cpipe->pipe_state |= PIPE_WANT;
1520                 pipeunlock(cpipe);
1521                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1522                 pipelock(cpipe, 0);
1523         }
1524
1525
1526         /*
1527          * Disconnect from peer, if any.
1528          */
1529         ppipe = cpipe->pipe_peer;
1530         if (ppipe->pipe_present == PIPE_ACTIVE) {
1531                 pipeselwakeup(ppipe);
1532
1533                 ppipe->pipe_state |= PIPE_EOF;
1534                 wakeup(ppipe);
1535                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1536         }
1537
1538         /*
1539          * Mark this endpoint as free.  Release kmem resources.  We
1540          * don't mark this endpoint as unused until we've finished
1541          * doing that, or the pipe might disappear out from under
1542          * us.
1543          */
1544         PIPE_UNLOCK(cpipe);
1545         pipe_free_kmem(cpipe);
1546         PIPE_LOCK(cpipe);
1547         cpipe->pipe_present = PIPE_CLOSING;
1548         pipeunlock(cpipe);
1549
1550         /*
1551          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1552          * PIPE_FINALIZED, that allows other end to free the
1553          * pipe_pair, only after the knotes are completely dismantled.
1554          */
1555         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1556         cpipe->pipe_present = PIPE_FINALIZED;
1557         seldrain(&cpipe->pipe_sel);
1558         knlist_destroy(&cpipe->pipe_sel.si_note);
1559
1560         /*
1561          * Postpone the destroy of the fake inode number allocated for
1562          * our end, until pipe mtx is unlocked.
1563          */
1564         ino = cpipe->pipe_ino;
1565
1566         /*
1567          * If both endpoints are now closed, release the memory for the
1568          * pipe pair.  If not, unlock.
1569          */
1570         if (ppipe->pipe_present == PIPE_FINALIZED) {
1571                 PIPE_UNLOCK(cpipe);
1572 #ifdef MAC
1573                 mac_pipe_destroy(pp);
1574 #endif
1575                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1576         } else
1577                 PIPE_UNLOCK(cpipe);
1578
1579         if (ino != 0 && ino != (ino_t)-1)
1580                 free_unr(pipeino_unr, ino);
1581 }
1582
1583 /*ARGSUSED*/
1584 static int
1585 pipe_kqfilter(struct file *fp, struct knote *kn)
1586 {
1587         struct pipe *cpipe;
1588
1589         cpipe = kn->kn_fp->f_data;
1590         PIPE_LOCK(cpipe);
1591         switch (kn->kn_filter) {
1592         case EVFILT_READ:
1593                 kn->kn_fop = &pipe_rfiltops;
1594                 break;
1595         case EVFILT_WRITE:
1596                 kn->kn_fop = &pipe_wfiltops;
1597                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1598                         /* other end of pipe has been closed */
1599                         PIPE_UNLOCK(cpipe);
1600                         return (EPIPE);
1601                 }
1602                 cpipe = cpipe->pipe_peer;
1603                 break;
1604         default:
1605                 PIPE_UNLOCK(cpipe);
1606                 return (EINVAL);
1607         }
1608
1609         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1610         PIPE_UNLOCK(cpipe);
1611         return (0);
1612 }
1613
1614 static void
1615 filt_pipedetach(struct knote *kn)
1616 {
1617         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1618
1619         PIPE_LOCK(cpipe);
1620         if (kn->kn_filter == EVFILT_WRITE)
1621                 cpipe = cpipe->pipe_peer;
1622         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1623         PIPE_UNLOCK(cpipe);
1624 }
1625
1626 /*ARGSUSED*/
1627 static int
1628 filt_piperead(struct knote *kn, long hint)
1629 {
1630         struct pipe *rpipe = kn->kn_fp->f_data;
1631         struct pipe *wpipe = rpipe->pipe_peer;
1632         int ret;
1633
1634         PIPE_LOCK(rpipe);
1635         kn->kn_data = rpipe->pipe_buffer.cnt;
1636         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1637                 kn->kn_data = rpipe->pipe_map.cnt;
1638
1639         if ((rpipe->pipe_state & PIPE_EOF) ||
1640             wpipe->pipe_present != PIPE_ACTIVE ||
1641             (wpipe->pipe_state & PIPE_EOF)) {
1642                 kn->kn_flags |= EV_EOF;
1643                 PIPE_UNLOCK(rpipe);
1644                 return (1);
1645         }
1646         ret = kn->kn_data > 0;
1647         PIPE_UNLOCK(rpipe);
1648         return ret;
1649 }
1650
1651 /*ARGSUSED*/
1652 static int
1653 filt_pipewrite(struct knote *kn, long hint)
1654 {
1655         struct pipe *rpipe = kn->kn_fp->f_data;
1656         struct pipe *wpipe = rpipe->pipe_peer;
1657
1658         PIPE_LOCK(rpipe);
1659         if (wpipe->pipe_present != PIPE_ACTIVE ||
1660             (wpipe->pipe_state & PIPE_EOF)) {
1661                 kn->kn_data = 0;
1662                 kn->kn_flags |= EV_EOF;
1663                 PIPE_UNLOCK(rpipe);
1664                 return (1);
1665         }
1666         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1667             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1668         if (wpipe->pipe_state & PIPE_DIRECTW)
1669                 kn->kn_data = 0;
1670
1671         PIPE_UNLOCK(rpipe);
1672         return (kn->kn_data >= PIPE_BUF);
1673 }