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 /* XXX */
 132 int     do_pipe(struct thread *td, int fildes[2], int flags);
 133
 134 /*
 135  * Use this define if you want to disable *fancy* VM things.  Expect an
 136  * approx 30% decrease in transfer rate.  This could be useful for
 137  * NetBSD or OpenBSD.
 138  */
 139 /* #define PIPE_NODIRECT */
 140
 141 /*
 142  * interfaces to the outside world
 143  */
 144 static fo_rdwr_t        pipe_read;
 145 static fo_rdwr_t        pipe_write;
 146 static fo_truncate_t    pipe_truncate;
 147 static fo_ioctl_t       pipe_ioctl;
 148 static fo_poll_t        pipe_poll;
 149 static fo_kqfilter_t    pipe_kqfilter;
 150 static fo_stat_t        pipe_stat;
 151 static fo_close_t       pipe_close;
 152
 153 static struct fileops pipeops = {
 154         .fo_read = pipe_read,
 155         .fo_write = pipe_write,
 156         .fo_truncate = pipe_truncate,
 157         .fo_ioctl = pipe_ioctl,
 158         .fo_poll = pipe_poll,
 159         .fo_kqfilter = pipe_kqfilter,
 160         .fo_stat = pipe_stat,
 161         .fo_close = pipe_close,
 162         .fo_chmod = invfo_chmod,
 163         .fo_chown = invfo_chown,
 164         .fo_flags = DFLAG_PASSABLE
 165 };
 166
 167 static void     filt_pipedetach(struct knote *kn);
 168 static int      filt_piperead(struct knote *kn, long hint);
 169 static int      filt_pipewrite(struct knote *kn, long hint);
 170
 171 static struct filterops pipe_rfiltops = {
 172         .f_isfd = 1,
 173         .f_detach = filt_pipedetach,
 174         .f_event = filt_piperead
 175 };
 176 static struct filterops pipe_wfiltops = {
 177         .f_isfd = 1,
 178         .f_detach = filt_pipedetach,
 179         .f_event = filt_pipewrite
 180 };
 181
 182 /*
 183  * Default pipe buffer size(s), this can be kind-of large now because pipe
 184  * space is pageable.  The pipe code will try to maintain locality of
 185  * reference for performance reasons, so small amounts of outstanding I/O
 186  * will not wipe the cache.
 187  */
 188 #define MINPIPESIZE (PIPE_SIZE/3)
 189 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 190
 191 static long amountpipekva;
 192 static int pipefragretry;
 193 static int pipeallocfail;
 194 static int piperesizefail;
 195 static int piperesizeallowed = 1;
 196
 197 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 198            &maxpipekva, 0, "Pipe KVA limit");
 199 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 200            &amountpipekva, 0, "Pipe KVA usage");
 201 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 202           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 203 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 204           &pipeallocfail, 0, "Pipe allocation failures");
 205 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 206           &piperesizefail, 0, "Pipe resize failures");
 207 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 208           &piperesizeallowed, 0, "Pipe resizing allowed");
 209
 210 static void pipeinit(void *dummy __unused);
 211 static void pipeclose(struct pipe *cpipe);
 212 static void pipe_free_kmem(struct pipe *cpipe);
 213 static int pipe_create(struct pipe *pipe, int backing);
 214 static __inline int pipelock(struct pipe *cpipe, int catch);
 215 static __inline void pipeunlock(struct pipe *cpipe);
 216 static __inline void pipeselwakeup(struct pipe *cpipe);
 217 #ifndef PIPE_NODIRECT
 218 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 219 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 220 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 221 static void pipe_clone_write_buffer(struct pipe *wpipe);
 222 #endif
 223 static int pipespace(struct pipe *cpipe, int size);
 224 static int pipespace_new(struct pipe *cpipe, int size);
 225
 226 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 227 static int      pipe_zone_init(void *mem, int size, int flags);
 228 static void     pipe_zone_fini(void *mem, int size);
 229
 230 static uma_zone_t pipe_zone;
 231 static struct unrhdr *pipeino_unr;
 232 static dev_t pipedev_ino;
 233
 234 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 235
 236 static void
 237 pipeinit(void *dummy __unused)
 238 {
 239
 240         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 241             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 242             UMA_ALIGN_PTR, 0);
 243         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 244         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 245         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 246         pipedev_ino = devfs_alloc_cdp_inode();
 247         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 248 }
 249
 250 static int
 251 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 252 {
 253         struct pipepair *pp;
 254         struct pipe *rpipe, *wpipe;
 255
 256         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 257
 258         pp = (struct pipepair *)mem;
 259
 260         /*
 261          * We zero both pipe endpoints to make sure all the kmem pointers
 262          * are NULL, flag fields are zero'd, etc.  We timestamp both
 263          * endpoints with the same time.
 264          */
 265         rpipe = &pp->pp_rpipe;
 266         bzero(rpipe, sizeof(*rpipe));
 267         vfs_timestamp(&rpipe->pipe_ctime);
 268         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 269
 270         wpipe = &pp->pp_wpipe;
 271         bzero(wpipe, sizeof(*wpipe));
 272         wpipe->pipe_ctime = rpipe->pipe_ctime;
 273         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 274
 275         rpipe->pipe_peer = wpipe;
 276         rpipe->pipe_pair = pp;
 277         wpipe->pipe_peer = rpipe;
 278         wpipe->pipe_pair = pp;
 279
 280         /*
 281          * Mark both endpoints as present; they will later get free'd
 282          * one at a time.  When both are free'd, then the whole pair
 283          * is released.
 284          */
 285         rpipe->pipe_present = PIPE_ACTIVE;
 286         wpipe->pipe_present = PIPE_ACTIVE;
 287
 288         /*
 289          * Eventually, the MAC Framework may initialize the label
 290          * in ctor or init, but for now we do it elswhere to avoid
 291          * blocking in ctor or init.
 292          */
 293         pp->pp_label = NULL;
 294
 295         return (0);
 296 }
 297
 298 static int
 299 pipe_zone_init(void *mem, int size, int flags)
 300 {
 301         struct pipepair *pp;
 302
 303         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 304
 305         pp = (struct pipepair *)mem;
 306
 307         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_RECURSE);
 308         return (0);
 309 }
 310
 311 static void
 312 pipe_zone_fini(void *mem, int size)
 313 {
 314         struct pipepair *pp;
 315
 316         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 317
 318         pp = (struct pipepair *)mem;
 319
 320         mtx_destroy(&pp->pp_mtx);
 321 }
 322
 323 /*
 324  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 325  * the zone pick up the pieces via pipeclose().
 326  */
 327 int
 328 kern_pipe(struct thread *td, int fildes[2])
 329 {
 330
 331         return (do_pipe(td, fildes, 0));
 332 }
 333
 334 int
 335 do_pipe(struct thread *td, int fildes[2], int flags)
 336 {
 337         struct filedesc *fdp = td->td_proc->p_fd;
 338         struct file *rf, *wf;
 339         struct pipepair *pp;
 340         struct pipe *rpipe, *wpipe;
 341         int fd, fflags, error;
 342
 343         pp = uma_zalloc(pipe_zone, M_WAITOK);
 344 #ifdef MAC
 345         /*
 346          * The MAC label is shared between the connected endpoints.  As a
 347          * result mac_pipe_init() and mac_pipe_create() are called once
 348          * for the pair, and not on the endpoints.
 349          */
 350         mac_pipe_init(pp);
 351         mac_pipe_create(td->td_ucred, pp);
 352 #endif
 353         rpipe = &pp->pp_rpipe;
 354         wpipe = &pp->pp_wpipe;
 355
 356         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 357         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 358
 359         /* Only the forward direction pipe is backed by default */
 360         if ((error = pipe_create(rpipe, 1)) != 0 ||
 361             (error = pipe_create(wpipe, 0)) != 0) {
 362                 pipeclose(rpipe);
 363                 pipeclose(wpipe);
 364                 return (error);
 365         }
 366
 367         rpipe->pipe_state |= PIPE_DIRECTOK;
 368         wpipe->pipe_state |= PIPE_DIRECTOK;
 369
 370         error = falloc(td, &rf, &fd, flags);
 371         if (error) {
 372                 pipeclose(rpipe);
 373                 pipeclose(wpipe);
 374                 return (error);
 375         }
 376         /* An extra reference on `rf' has been held for us by falloc(). */
 377         fildes[0] = fd;
 378
 379         fflags = FREAD | FWRITE;
 380         if ((flags & O_NONBLOCK) != 0)
 381                 fflags |= FNONBLOCK;
 382
 383         /*
 384          * Warning: once we've gotten past allocation of the fd for the
 385          * read-side, we can only drop the read side via fdrop() in order
 386          * to avoid races against processes which manage to dup() the read
 387          * side while we are blocked trying to allocate the write side.
 388          */
 389         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 390         error = falloc(td, &wf, &fd, flags);
 391         if (error) {
 392                 fdclose(fdp, rf, fildes[0], td);
 393                 fdrop(rf, td);
 394                 /* rpipe has been closed by fdrop(). */
 395                 pipeclose(wpipe);
 396                 return (error);
 397         }
 398         /* An extra reference on `wf' has been held for us by falloc(). */
 399         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 400         fdrop(wf, td);
 401         fildes[1] = fd;
 402         fdrop(rf, td);
 403
 404         return (0);
 405 }
 406
 407 /* ARGSUSED */
 408 int
 409 sys_pipe(struct thread *td, struct pipe_args *uap)
 410 {
 411         int error;
 412         int fildes[2];
 413
 414         error = kern_pipe(td, fildes);
 415         if (error)
 416                 return (error);
 417
 418         td->td_retval[0] = fildes[0];
 419         td->td_retval[1] = fildes[1];
 420
 421         return (0);
 422 }
 423
 424 /*
 425  * Allocate kva for pipe circular buffer, the space is pageable
 426  * This routine will 'realloc' the size of a pipe safely, if it fails
 427  * it will retain the old buffer.
 428  * If it fails it will return ENOMEM.
 429  */
 430 static int
 431 pipespace_new(cpipe, size)
 432         struct pipe *cpipe;
 433         int size;
 434 {
 435         caddr_t buffer;
 436         int error, cnt, firstseg;
 437         static int curfail = 0;
 438         static struct timeval lastfail;
 439
 440         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 441         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 442                 ("pipespace: resize of direct writes not allowed"));
 443 retry:
 444         cnt = cpipe->pipe_buffer.cnt;
 445         if (cnt > size)
 446                 size = cnt;
 447
 448         size = round_page(size);
 449         buffer = (caddr_t) vm_map_min(pipe_map);
 450
 451         error = vm_map_find(pipe_map, NULL, 0,
 452                 (vm_offset_t *) &buffer, size, 1,
 453                 VM_PROT_ALL, VM_PROT_ALL, 0);
 454         if (error != KERN_SUCCESS) {
 455                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 456                         (size > SMALL_PIPE_SIZE)) {
 457                         size = SMALL_PIPE_SIZE;
 458                         pipefragretry++;
 459                         goto retry;
 460                 }
 461                 if (cpipe->pipe_buffer.buffer == NULL) {
 462                         pipeallocfail++;
 463                         if (ppsratecheck(&lastfail, &curfail, 1))
 464                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 465                 } else {
 466                         piperesizefail++;
 467                 }
 468                 return (ENOMEM);
 469         }
 470
 471         /* copy data, then free old resources if we're resizing */
 472         if (cnt > 0) {
 473                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 474                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 475                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 476                                 buffer, firstseg);
 477                         if ((cnt - firstseg) > 0)
 478                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 479                                         cpipe->pipe_buffer.in);
 480                 } else {
 481                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 482                                 buffer, cnt);
 483                 }
 484         }
 485         pipe_free_kmem(cpipe);
 486         cpipe->pipe_buffer.buffer = buffer;
 487         cpipe->pipe_buffer.size = size;
 488         cpipe->pipe_buffer.in = cnt;
 489         cpipe->pipe_buffer.out = 0;
 490         cpipe->pipe_buffer.cnt = cnt;
 491         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 492         return (0);
 493 }
 494
 495 /*
 496  * Wrapper for pipespace_new() that performs locking assertions.
 497  */
 498 static int
 499 pipespace(cpipe, size)
 500         struct pipe *cpipe;
 501         int size;
 502 {
 503
 504         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 505                 ("Unlocked pipe passed to pipespace"));
 506         return (pipespace_new(cpipe, size));
 507 }
 508
 509 /*
 510  * lock a pipe for I/O, blocking other access
 511  */
 512 static __inline int
 513 pipelock(cpipe, catch)
 514         struct pipe *cpipe;
 515         int catch;
 516 {
 517         int error;
 518
 519         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 520         while (cpipe->pipe_state & PIPE_LOCKFL) {
 521                 cpipe->pipe_state |= PIPE_LWANT;
 522                 error = msleep(cpipe, PIPE_MTX(cpipe),
 523                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 524                     "pipelk", 0);
 525                 if (error != 0)
 526                         return (error);
 527         }
 528         cpipe->pipe_state |= PIPE_LOCKFL;
 529         return (0);
 530 }
 531
 532 /*
 533  * unlock a pipe I/O lock
 534  */
 535 static __inline void
 536 pipeunlock(cpipe)
 537         struct pipe *cpipe;
 538 {
 539
 540         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 541         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 542                 ("Unlocked pipe passed to pipeunlock"));
 543         cpipe->pipe_state &= ~PIPE_LOCKFL;
 544         if (cpipe->pipe_state & PIPE_LWANT) {
 545                 cpipe->pipe_state &= ~PIPE_LWANT;
 546                 wakeup(cpipe);
 547         }
 548 }
 549
 550 static __inline void
 551 pipeselwakeup(cpipe)
 552         struct pipe *cpipe;
 553 {
 554
 555         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 556         if (cpipe->pipe_state & PIPE_SEL) {
 557                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 558                 if (!SEL_WAITING(&cpipe->pipe_sel))
 559                         cpipe->pipe_state &= ~PIPE_SEL;
 560         }
 561         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 562                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 563         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 564 }
 565
 566 /*
 567  * Initialize and allocate VM and memory for pipe.  The structure
 568  * will start out zero'd from the ctor, so we just manage the kmem.
 569  */
 570 static int
 571 pipe_create(pipe, backing)
 572         struct pipe *pipe;
 573         int backing;
 574 {
 575         int error;
 576
 577         if (backing) {
 578                 if (amountpipekva > maxpipekva / 2)
 579                         error = pipespace_new(pipe, SMALL_PIPE_SIZE);
 580                 else
 581                         error = pipespace_new(pipe, PIPE_SIZE);
 582         } else {
 583                 /* If we're not backing this pipe, no need to do anything. */
 584                 error = 0;
 585         }
 586         pipe->pipe_ino = -1;
 587         return (error);
 588 }
 589
 590 /* ARGSUSED */
 591 static int
 592 pipe_read(fp, uio, active_cred, flags, td)
 593         struct file *fp;
 594         struct uio *uio;
 595         struct ucred *active_cred;
 596         struct thread *td;
 597         int flags;
 598 {
 599         struct pipe *rpipe = fp->f_data;
 600         int error;
 601         int nread = 0;
 602         int size;
 603
 604         PIPE_LOCK(rpipe);
 605         ++rpipe->pipe_busy;
 606         error = pipelock(rpipe, 1);
 607         if (error)
 608                 goto unlocked_error;
 609
 610 #ifdef MAC
 611         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 612         if (error)
 613                 goto locked_error;
 614 #endif
 615         if (amountpipekva > (3 * maxpipekva) / 4) {
 616                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 617                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 618                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 619                         (piperesizeallowed == 1)) {
 620                         PIPE_UNLOCK(rpipe);
 621                         pipespace(rpipe, SMALL_PIPE_SIZE);
 622                         PIPE_LOCK(rpipe);
 623                 }
 624         }
 625
 626         while (uio->uio_resid) {
 627                 /*
 628                  * normal pipe buffer receive
 629                  */
 630                 if (rpipe->pipe_buffer.cnt > 0) {
 631                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 632                         if (size > rpipe->pipe_buffer.cnt)
 633                                 size = rpipe->pipe_buffer.cnt;
 634                         if (size > uio->uio_resid)
 635                                 size = uio->uio_resid;
 636
 637                         PIPE_UNLOCK(rpipe);
 638                         error = uiomove(
 639                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 640                             size, uio);
 641                         PIPE_LOCK(rpipe);
 642                         if (error)
 643                                 break;
 644
 645                         rpipe->pipe_buffer.out += size;
 646                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 647                                 rpipe->pipe_buffer.out = 0;
 648
 649                         rpipe->pipe_buffer.cnt -= size;
 650
 651                         /*
 652                          * If there is no more to read in the pipe, reset
 653                          * its pointers to the beginning.  This improves
 654                          * cache hit stats.
 655                          */
 656                         if (rpipe->pipe_buffer.cnt == 0) {
 657                                 rpipe->pipe_buffer.in = 0;
 658                                 rpipe->pipe_buffer.out = 0;
 659                         }
 660                         nread += size;
 661 #ifndef PIPE_NODIRECT
 662                 /*
 663                  * Direct copy, bypassing a kernel buffer.
 664                  */
 665                 } else if ((size = rpipe->pipe_map.cnt) &&
 666                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 667                         if (size > uio->uio_resid)
 668                                 size = (u_int) uio->uio_resid;
 669
 670                         PIPE_UNLOCK(rpipe);
 671                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 672                             rpipe->pipe_map.pos, size, uio);
 673                         PIPE_LOCK(rpipe);
 674                         if (error)
 675                                 break;
 676                         nread += size;
 677                         rpipe->pipe_map.pos += size;
 678                         rpipe->pipe_map.cnt -= size;
 679                         if (rpipe->pipe_map.cnt == 0) {
 680                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 681                                 wakeup(rpipe);
 682                         }
 683 #endif
 684                 } else {
 685                         /*
 686                          * detect EOF condition
 687                          * read returns 0 on EOF, no need to set error
 688                          */
 689                         if (rpipe->pipe_state & PIPE_EOF)
 690                                 break;
 691
 692                         /*
 693                          * If the "write-side" has been blocked, wake it up now.
 694                          */
 695                         if (rpipe->pipe_state & PIPE_WANTW) {
 696                                 rpipe->pipe_state &= ~PIPE_WANTW;
 697                                 wakeup(rpipe);
 698                         }
 699
 700                         /*
 701                          * Break if some data was read.
 702                          */
 703                         if (nread > 0)
 704                                 break;
 705
 706                         /*
 707                          * Unlock the pipe buffer for our remaining processing.
 708                          * We will either break out with an error or we will
 709                          * sleep and relock to loop.
 710                          */
 711                         pipeunlock(rpipe);
 712
 713                         /*
 714                          * Handle non-blocking mode operation or
 715                          * wait for more data.
 716                          */
 717                         if (fp->f_flag & FNONBLOCK) {
 718                                 error = EAGAIN;
 719                         } else {
 720                                 rpipe->pipe_state |= PIPE_WANTR;
 721                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 722                                     PRIBIO | PCATCH,
 723                                     "piperd", 0)) == 0)
 724                                         error = pipelock(rpipe, 1);
 725                         }
 726                         if (error)
 727                                 goto unlocked_error;
 728                 }
 729         }
 730 #ifdef MAC
 731 locked_error:
 732 #endif
 733         pipeunlock(rpipe);
 734
 735         /* XXX: should probably do this before getting any locks. */
 736         if (error == 0)
 737                 vfs_timestamp(&rpipe->pipe_atime);
 738 unlocked_error:
 739         --rpipe->pipe_busy;
 740
 741         /*
 742          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 743          */
 744         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 745                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 746                 wakeup(rpipe);
 747         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 748                 /*
 749                  * Handle write blocking hysteresis.
 750                  */
 751                 if (rpipe->pipe_state & PIPE_WANTW) {
 752                         rpipe->pipe_state &= ~PIPE_WANTW;
 753                         wakeup(rpipe);
 754                 }
 755         }
 756
 757         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 758                 pipeselwakeup(rpipe);
 759
 760         PIPE_UNLOCK(rpipe);
 761         return (error);
 762 }
 763
 764 #ifndef PIPE_NODIRECT
 765 /*
 766  * Map the sending processes' buffer into kernel space and wire it.
 767  * This is similar to a physical write operation.
 768  */
 769 static int
 770 pipe_build_write_buffer(wpipe, uio)
 771         struct pipe *wpipe;
 772         struct uio *uio;
 773 {
 774         u_int size;
 775         int i;
 776
 777         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 778         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 779                 ("Clone attempt on non-direct write pipe!"));
 780
 781         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 782                 size = wpipe->pipe_buffer.size;
 783         else
 784                 size = uio->uio_iov->iov_len;
 785
 786         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 787             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 788             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 789                 return (EFAULT);
 790
 791 /*
 792  * set up the control block
 793  */
 794         wpipe->pipe_map.npages = i;
 795         wpipe->pipe_map.pos =
 796             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 797         wpipe->pipe_map.cnt = size;
 798
 799 /*
 800  * and update the uio data
 801  */
 802
 803         uio->uio_iov->iov_len -= size;
 804         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 805         if (uio->uio_iov->iov_len == 0)
 806                 uio->uio_iov++;
 807         uio->uio_resid -= size;
 808         uio->uio_offset += size;
 809         return (0);
 810 }
 811
 812 /*
 813  * unmap and unwire the process buffer
 814  */
 815 static void
 816 pipe_destroy_write_buffer(wpipe)
 817         struct pipe *wpipe;
 818 {
 819
 820         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 821         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 822         wpipe->pipe_map.npages = 0;
 823 }
 824
 825 /*
 826  * In the case of a signal, the writing process might go away.  This
 827  * code copies the data into the circular buffer so that the source
 828  * pages can be freed without loss of data.
 829  */
 830 static void
 831 pipe_clone_write_buffer(wpipe)
 832         struct pipe *wpipe;
 833 {
 834         struct uio uio;
 835         struct iovec iov;
 836         int size;
 837         int pos;
 838
 839         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 840         size = wpipe->pipe_map.cnt;
 841         pos = wpipe->pipe_map.pos;
 842
 843         wpipe->pipe_buffer.in = size;
 844         wpipe->pipe_buffer.out = 0;
 845         wpipe->pipe_buffer.cnt = size;
 846         wpipe->pipe_state &= ~PIPE_DIRECTW;
 847
 848         PIPE_UNLOCK(wpipe);
 849         iov.iov_base = wpipe->pipe_buffer.buffer;
 850         iov.iov_len = size;
 851         uio.uio_iov = &iov;
 852         uio.uio_iovcnt = 1;
 853         uio.uio_offset = 0;
 854         uio.uio_resid = size;
 855         uio.uio_segflg = UIO_SYSSPACE;
 856         uio.uio_rw = UIO_READ;
 857         uio.uio_td = curthread;
 858         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 859         PIPE_LOCK(wpipe);
 860         pipe_destroy_write_buffer(wpipe);
 861 }
 862
 863 /*
 864  * This implements the pipe buffer write mechanism.  Note that only
 865  * a direct write OR a normal pipe write can be pending at any given time.
 866  * If there are any characters in the pipe buffer, the direct write will
 867  * be deferred until the receiving process grabs all of the bytes from
 868  * the pipe buffer.  Then the direct mapping write is set-up.
 869  */
 870 static int
 871 pipe_direct_write(wpipe, uio)
 872         struct pipe *wpipe;
 873         struct uio *uio;
 874 {
 875         int error;
 876
 877 retry:
 878         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 879         error = pipelock(wpipe, 1);
 880         if (wpipe->pipe_state & PIPE_EOF)
 881                 error = EPIPE;
 882         if (error) {
 883                 pipeunlock(wpipe);
 884                 goto error1;
 885         }
 886         while (wpipe->pipe_state & PIPE_DIRECTW) {
 887                 if (wpipe->pipe_state & PIPE_WANTR) {
 888                         wpipe->pipe_state &= ~PIPE_WANTR;
 889                         wakeup(wpipe);
 890                 }
 891                 pipeselwakeup(wpipe);
 892                 wpipe->pipe_state |= PIPE_WANTW;
 893                 pipeunlock(wpipe);
 894                 error = msleep(wpipe, PIPE_MTX(wpipe),
 895                     PRIBIO | PCATCH, "pipdww", 0);
 896                 if (error)
 897                         goto error1;
 898                 else
 899                         goto retry;
 900         }
 901         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 902         if (wpipe->pipe_buffer.cnt > 0) {
 903                 if (wpipe->pipe_state & PIPE_WANTR) {
 904                         wpipe->pipe_state &= ~PIPE_WANTR;
 905                         wakeup(wpipe);
 906                 }
 907                 pipeselwakeup(wpipe);
 908                 wpipe->pipe_state |= PIPE_WANTW;
 909                 pipeunlock(wpipe);
 910                 error = msleep(wpipe, PIPE_MTX(wpipe),
 911                     PRIBIO | PCATCH, "pipdwc", 0);
 912                 if (error)
 913                         goto error1;
 914                 else
 915                         goto retry;
 916         }
 917
 918         wpipe->pipe_state |= PIPE_DIRECTW;
 919
 920         PIPE_UNLOCK(wpipe);
 921         error = pipe_build_write_buffer(wpipe, uio);
 922         PIPE_LOCK(wpipe);
 923         if (error) {
 924                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 925                 pipeunlock(wpipe);
 926                 goto error1;
 927         }
 928
 929         error = 0;
 930         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 931                 if (wpipe->pipe_state & PIPE_EOF) {
 932                         pipe_destroy_write_buffer(wpipe);
 933                         pipeselwakeup(wpipe);
 934                         pipeunlock(wpipe);
 935                         error = EPIPE;
 936                         goto error1;
 937                 }
 938                 if (wpipe->pipe_state & PIPE_WANTR) {
 939                         wpipe->pipe_state &= ~PIPE_WANTR;
 940                         wakeup(wpipe);
 941                 }
 942                 pipeselwakeup(wpipe);
 943                 pipeunlock(wpipe);
 944                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
 945                     "pipdwt", 0);
 946                 pipelock(wpipe, 0);
 947         }
 948
 949         if (wpipe->pipe_state & PIPE_EOF)
 950                 error = EPIPE;
 951         if (wpipe->pipe_state & PIPE_DIRECTW) {
 952                 /*
 953                  * this bit of trickery substitutes a kernel buffer for
 954                  * the process that might be going away.
 955                  */
 956                 pipe_clone_write_buffer(wpipe);
 957         } else {
 958                 pipe_destroy_write_buffer(wpipe);
 959         }
 960         pipeunlock(wpipe);
 961         return (error);
 962
 963 error1:
 964         wakeup(wpipe);
 965         return (error);
 966 }
 967 #endif
 968
 969 static int
 970 pipe_write(fp, uio, active_cred, flags, td)
 971         struct file *fp;
 972         struct uio *uio;
 973         struct ucred *active_cred;
 974         struct thread *td;
 975         int flags;
 976 {
 977         int error = 0;
 978         int desiredsize;
 979         ssize_t orig_resid;
 980         struct pipe *wpipe, *rpipe;
 981
 982         rpipe = fp->f_data;
 983         wpipe = rpipe->pipe_peer;
 984
 985         PIPE_LOCK(rpipe);
 986         error = pipelock(wpipe, 1);
 987         if (error) {
 988                 PIPE_UNLOCK(rpipe);
 989                 return (error);
 990         }
 991         /*
 992          * detect loss of pipe read side, issue SIGPIPE if lost.
 993          */
 994         if (wpipe->pipe_present != PIPE_ACTIVE ||
 995             (wpipe->pipe_state & PIPE_EOF)) {
 996                 pipeunlock(wpipe);
 997                 PIPE_UNLOCK(rpipe);
 998                 return (EPIPE);
 999         }
1000 #ifdef MAC
1001         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1002         if (error) {
1003                 pipeunlock(wpipe);
1004                 PIPE_UNLOCK(rpipe);
1005                 return (error);
1006         }
1007 #endif
1008         ++wpipe->pipe_busy;
1009
1010         /* Choose a larger size if it's advantageous */
1011         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1012         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1013                 if (piperesizeallowed != 1)
1014                         break;
1015                 if (amountpipekva > maxpipekva / 2)
1016                         break;
1017                 if (desiredsize == BIG_PIPE_SIZE)
1018                         break;
1019                 desiredsize = desiredsize * 2;
1020         }
1021
1022         /* Choose a smaller size if we're in a OOM situation */
1023         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1024                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1025                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1026                 (piperesizeallowed == 1))
1027                 desiredsize = SMALL_PIPE_SIZE;
1028
1029         /* Resize if the above determined that a new size was necessary */
1030         if ((desiredsize != wpipe->pipe_buffer.size) &&
1031                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1032                 PIPE_UNLOCK(wpipe);
1033                 pipespace(wpipe, desiredsize);
1034                 PIPE_LOCK(wpipe);
1035         }
1036         if (wpipe->pipe_buffer.size == 0) {
1037                 /*
1038                  * This can only happen for reverse direction use of pipes
1039                  * in a complete OOM situation.
1040                  */
1041                 error = ENOMEM;
1042                 --wpipe->pipe_busy;
1043                 pipeunlock(wpipe);
1044                 PIPE_UNLOCK(wpipe);
1045                 return (error);
1046         }
1047
1048         pipeunlock(wpipe);
1049
1050         orig_resid = uio->uio_resid;
1051
1052         while (uio->uio_resid) {
1053                 int space;
1054
1055                 pipelock(wpipe, 0);
1056                 if (wpipe->pipe_state & PIPE_EOF) {
1057                         pipeunlock(wpipe);
1058                         error = EPIPE;
1059                         break;
1060                 }
1061 #ifndef PIPE_NODIRECT
1062                 /*
1063                  * If the transfer is large, we can gain performance if
1064                  * we do process-to-process copies directly.
1065                  * If the write is non-blocking, we don't use the
1066                  * direct write mechanism.
1067                  *
1068                  * The direct write mechanism will detect the reader going
1069                  * away on us.
1070                  */
1071                 if (uio->uio_segflg == UIO_USERSPACE &&
1072                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1073                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1074                     (fp->f_flag & FNONBLOCK) == 0) {
1075                         pipeunlock(wpipe);
1076                         error = pipe_direct_write(wpipe, uio);
1077                         if (error)
1078                                 break;
1079                         continue;
1080                 }
1081 #endif
1082
1083                 /*
1084                  * Pipe buffered writes cannot be coincidental with
1085                  * direct writes.  We wait until the currently executing
1086                  * direct write is completed before we start filling the
1087                  * pipe buffer.  We break out if a signal occurs or the
1088                  * reader goes away.
1089                  */
1090                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1091                         if (wpipe->pipe_state & PIPE_WANTR) {
1092                                 wpipe->pipe_state &= ~PIPE_WANTR;
1093                                 wakeup(wpipe);
1094                         }
1095                         pipeselwakeup(wpipe);
1096                         wpipe->pipe_state |= PIPE_WANTW;
1097                         pipeunlock(wpipe);
1098                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1099                             "pipbww", 0);
1100                         if (error)
1101                                 break;
1102                         else
1103                                 continue;
1104                 }
1105
1106                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1107
1108                 /* Writes of size <= PIPE_BUF must be atomic. */
1109                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1110                         space = 0;
1111
1112                 if (space > 0) {
1113                         int size;       /* Transfer size */
1114                         int segsize;    /* first segment to transfer */
1115
1116                         /*
1117                          * Transfer size is minimum of uio transfer
1118                          * and free space in pipe buffer.
1119                          */
1120                         if (space > uio->uio_resid)
1121                                 size = uio->uio_resid;
1122                         else
1123                                 size = space;
1124                         /*
1125                          * First segment to transfer is minimum of
1126                          * transfer size and contiguous space in
1127                          * pipe buffer.  If first segment to transfer
1128                          * is less than the transfer size, we've got
1129                          * a wraparound in the buffer.
1130                          */
1131                         segsize = wpipe->pipe_buffer.size -
1132                                 wpipe->pipe_buffer.in;
1133                         if (segsize > size)
1134                                 segsize = size;
1135
1136                         /* Transfer first segment */
1137
1138                         PIPE_UNLOCK(rpipe);
1139                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1140                                         segsize, uio);
1141                         PIPE_LOCK(rpipe);
1142
1143                         if (error == 0 && segsize < size) {
1144                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1145                                         wpipe->pipe_buffer.size,
1146                                         ("Pipe buffer wraparound disappeared"));
1147                                 /*
1148                                  * Transfer remaining part now, to
1149                                  * support atomic writes.  Wraparound
1150                                  * happened.
1151                                  */
1152
1153                                 PIPE_UNLOCK(rpipe);
1154                                 error = uiomove(
1155                                     &wpipe->pipe_buffer.buffer[0],
1156                                     size - segsize, uio);
1157                                 PIPE_LOCK(rpipe);
1158                         }
1159                         if (error == 0) {
1160                                 wpipe->pipe_buffer.in += size;
1161                                 if (wpipe->pipe_buffer.in >=
1162                                     wpipe->pipe_buffer.size) {
1163                                         KASSERT(wpipe->pipe_buffer.in ==
1164                                                 size - segsize +
1165                                                 wpipe->pipe_buffer.size,
1166                                                 ("Expected wraparound bad"));
1167                                         wpipe->pipe_buffer.in = size - segsize;
1168                                 }
1169
1170                                 wpipe->pipe_buffer.cnt += size;
1171                                 KASSERT(wpipe->pipe_buffer.cnt <=
1172                                         wpipe->pipe_buffer.size,
1173                                         ("Pipe buffer overflow"));
1174                         }
1175                         pipeunlock(wpipe);
1176                         if (error != 0)
1177                                 break;
1178                 } else {
1179                         /*
1180                          * If the "read-side" has been blocked, wake it up now.
1181                          */
1182                         if (wpipe->pipe_state & PIPE_WANTR) {
1183                                 wpipe->pipe_state &= ~PIPE_WANTR;
1184                                 wakeup(wpipe);
1185                         }
1186
1187                         /*
1188                          * don't block on non-blocking I/O
1189                          */
1190                         if (fp->f_flag & FNONBLOCK) {
1191                                 error = EAGAIN;
1192                                 pipeunlock(wpipe);
1193                                 break;
1194                         }
1195
1196                         /*
1197                          * We have no more space and have something to offer,
1198                          * wake up select/poll.
1199                          */
1200                         pipeselwakeup(wpipe);
1201
1202                         wpipe->pipe_state |= PIPE_WANTW;
1203                         pipeunlock(wpipe);
1204                         error = msleep(wpipe, PIPE_MTX(rpipe),
1205                             PRIBIO | PCATCH, "pipewr", 0);
1206                         if (error != 0)
1207                                 break;
1208                 }
1209         }
1210
1211         pipelock(wpipe, 0);
1212         --wpipe->pipe_busy;
1213
1214         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1215                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1216                 wakeup(wpipe);
1217         } else if (wpipe->pipe_buffer.cnt > 0) {
1218                 /*
1219                  * If we have put any characters in the buffer, we wake up
1220                  * the reader.
1221                  */
1222                 if (wpipe->pipe_state & PIPE_WANTR) {
1223                         wpipe->pipe_state &= ~PIPE_WANTR;
1224                         wakeup(wpipe);
1225                 }
1226         }
1227
1228         /*
1229          * Don't return EPIPE if I/O was successful
1230          */
1231         if ((wpipe->pipe_buffer.cnt == 0) &&
1232             (uio->uio_resid == 0) &&
1233             (error == EPIPE)) {
1234                 error = 0;
1235         }
1236
1237         if (error == 0)
1238                 vfs_timestamp(&wpipe->pipe_mtime);
1239
1240         /*
1241          * We have something to offer,
1242          * wake up select/poll.
1243          */
1244         if (wpipe->pipe_buffer.cnt)
1245                 pipeselwakeup(wpipe);
1246
1247         pipeunlock(wpipe);
1248         PIPE_UNLOCK(rpipe);
1249         return (error);
1250 }
1251
1252 /* ARGSUSED */
1253 static int
1254 pipe_truncate(fp, length, active_cred, td)
1255         struct file *fp;
1256         off_t length;
1257         struct ucred *active_cred;
1258         struct thread *td;
1259 {
1260
1261         return (EINVAL);
1262 }
1263
1264 /*
1265  * we implement a very minimal set of ioctls for compatibility with sockets.
1266  */
1267 static int
1268 pipe_ioctl(fp, cmd, data, active_cred, td)
1269         struct file *fp;
1270         u_long cmd;
1271         void *data;
1272         struct ucred *active_cred;
1273         struct thread *td;
1274 {
1275         struct pipe *mpipe = fp->f_data;
1276         int error;
1277
1278         PIPE_LOCK(mpipe);
1279
1280 #ifdef MAC
1281         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1282         if (error) {
1283                 PIPE_UNLOCK(mpipe);
1284                 return (error);
1285         }
1286 #endif
1287
1288         error = 0;
1289         switch (cmd) {
1290
1291         case FIONBIO:
1292                 break;
1293
1294         case FIOASYNC:
1295                 if (*(int *)data) {
1296                         mpipe->pipe_state |= PIPE_ASYNC;
1297                 } else {
1298                         mpipe->pipe_state &= ~PIPE_ASYNC;
1299                 }
1300                 break;
1301
1302         case FIONREAD:
1303                 if (mpipe->pipe_state & PIPE_DIRECTW)
1304                         *(int *)data = mpipe->pipe_map.cnt;
1305                 else
1306                         *(int *)data = mpipe->pipe_buffer.cnt;
1307                 break;
1308
1309         case FIOSETOWN:
1310                 PIPE_UNLOCK(mpipe);
1311                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1312                 goto out_unlocked;
1313
1314         case FIOGETOWN:
1315                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1316                 break;
1317
1318         /* This is deprecated, FIOSETOWN should be used instead. */
1319         case TIOCSPGRP:
1320                 PIPE_UNLOCK(mpipe);
1321                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1322                 goto out_unlocked;
1323
1324         /* This is deprecated, FIOGETOWN should be used instead. */
1325         case TIOCGPGRP:
1326                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1327                 break;
1328
1329         default:
1330                 error = ENOTTY;
1331                 break;
1332         }
1333         PIPE_UNLOCK(mpipe);
1334 out_unlocked:
1335         return (error);
1336 }
1337
1338 static int
1339 pipe_poll(fp, events, active_cred, td)
1340         struct file *fp;
1341         int events;
1342         struct ucred *active_cred;
1343         struct thread *td;
1344 {
1345         struct pipe *rpipe = fp->f_data;
1346         struct pipe *wpipe;
1347         int revents = 0;
1348 #ifdef MAC
1349         int error;
1350 #endif
1351
1352         wpipe = rpipe->pipe_peer;
1353         PIPE_LOCK(rpipe);
1354 #ifdef MAC
1355         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1356         if (error)
1357                 goto locked_error;
1358 #endif
1359         if (events & (POLLIN | POLLRDNORM))
1360                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1361                     (rpipe->pipe_buffer.cnt > 0))
1362                         revents |= events & (POLLIN | POLLRDNORM);
1363
1364         if (events & (POLLOUT | POLLWRNORM))
1365                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1366                     (wpipe->pipe_state & PIPE_EOF) ||
1367                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1368                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1369                          wpipe->pipe_buffer.size == 0)))
1370                         revents |= events & (POLLOUT | POLLWRNORM);
1371
1372         if ((events & POLLINIGNEOF) == 0) {
1373                 if (rpipe->pipe_state & PIPE_EOF) {
1374                         revents |= (events & (POLLIN | POLLRDNORM));
1375                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1376                             (wpipe->pipe_state & PIPE_EOF))
1377                                 revents |= POLLHUP;
1378                 }
1379         }
1380
1381         if (revents == 0) {
1382                 if (events & (POLLIN | POLLRDNORM)) {
1383                         selrecord(td, &rpipe->pipe_sel);
1384                         if (SEL_WAITING(&rpipe->pipe_sel))
1385                                 rpipe->pipe_state |= PIPE_SEL;
1386                 }
1387
1388                 if (events & (POLLOUT | POLLWRNORM)) {
1389                         selrecord(td, &wpipe->pipe_sel);
1390                         if (SEL_WAITING(&wpipe->pipe_sel))
1391                                 wpipe->pipe_state |= PIPE_SEL;
1392                 }
1393         }
1394 #ifdef MAC
1395 locked_error:
1396 #endif
1397         PIPE_UNLOCK(rpipe);
1398
1399         return (revents);
1400 }
1401
1402 /*
1403  * We shouldn't need locks here as we're doing a read and this should
1404  * be a natural race.
1405  */
1406 static int
1407 pipe_stat(fp, ub, active_cred, td)
1408         struct file *fp;
1409         struct stat *ub;
1410         struct ucred *active_cred;
1411         struct thread *td;
1412 {
1413         struct pipe *pipe;
1414         int new_unr;
1415 #ifdef MAC
1416         int error;
1417 #endif
1418
1419         pipe = fp->f_data;
1420         PIPE_LOCK(pipe);
1421 #ifdef MAC
1422         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1423         if (error) {
1424                 PIPE_UNLOCK(pipe);
1425                 return (error);
1426         }
1427 #endif
1428         /*
1429          * Lazily allocate an inode number for the pipe.  Most pipe
1430          * users do not call fstat(2) on the pipe, which means that
1431          * postponing the inode allocation until it is must be
1432          * returned to userland is useful.  If alloc_unr failed,
1433          * assign st_ino zero instead of returning an error.
1434          * Special pipe_ino values:
1435          *  -1 - not yet initialized;
1436          *  0  - alloc_unr failed, return 0 as st_ino forever.
1437          */
1438         if (pipe->pipe_ino == (ino_t)-1) {
1439                 new_unr = alloc_unr(pipeino_unr);
1440                 if (new_unr != -1)
1441                         pipe->pipe_ino = new_unr;
1442                 else
1443                         pipe->pipe_ino = 0;
1444         }
1445         PIPE_UNLOCK(pipe);
1446
1447         bzero(ub, sizeof(*ub));
1448         ub->st_mode = S_IFIFO;
1449         ub->st_blksize = PAGE_SIZE;
1450         if (pipe->pipe_state & PIPE_DIRECTW)
1451                 ub->st_size = pipe->pipe_map.cnt;
1452         else
1453                 ub->st_size = pipe->pipe_buffer.cnt;
1454         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1455         ub->st_atim = pipe->pipe_atime;
1456         ub->st_mtim = pipe->pipe_mtime;
1457         ub->st_ctim = pipe->pipe_ctime;
1458         ub->st_uid = fp->f_cred->cr_uid;
1459         ub->st_gid = fp->f_cred->cr_gid;
1460         ub->st_dev = pipedev_ino;
1461         ub->st_ino = pipe->pipe_ino;
1462         /*
1463          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1464          */
1465         return (0);
1466 }
1467
1468 /* ARGSUSED */
1469 static int
1470 pipe_close(fp, td)
1471         struct file *fp;
1472         struct thread *td;
1473 {
1474         struct pipe *cpipe = fp->f_data;
1475
1476         fp->f_ops = &badfileops;
1477         fp->f_data = NULL;
1478         funsetown(&cpipe->pipe_sigio);
1479         pipeclose(cpipe);
1480         return (0);
1481 }
1482
1483 static void
1484 pipe_free_kmem(cpipe)
1485         struct pipe *cpipe;
1486 {
1487
1488         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1489             ("pipe_free_kmem: pipe mutex locked"));
1490
1491         if (cpipe->pipe_buffer.buffer != NULL) {
1492                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1493                 vm_map_remove(pipe_map,
1494                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1495                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1496                 cpipe->pipe_buffer.buffer = NULL;
1497         }
1498 #ifndef PIPE_NODIRECT
1499         {
1500                 cpipe->pipe_map.cnt = 0;
1501                 cpipe->pipe_map.pos = 0;
1502                 cpipe->pipe_map.npages = 0;
1503         }
1504 #endif
1505 }
1506
1507 /*
1508  * shutdown the pipe
1509  */
1510 static void
1511 pipeclose(cpipe)
1512         struct pipe *cpipe;
1513 {
1514         struct pipepair *pp;
1515         struct pipe *ppipe;
1516         ino_t ino;
1517
1518         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1519
1520         PIPE_LOCK(cpipe);
1521         pipelock(cpipe, 0);
1522         pp = cpipe->pipe_pair;
1523
1524         pipeselwakeup(cpipe);
1525
1526         /*
1527          * If the other side is blocked, wake it up saying that
1528          * we want to close it down.
1529          */
1530         cpipe->pipe_state |= PIPE_EOF;
1531         while (cpipe->pipe_busy) {
1532                 wakeup(cpipe);
1533                 cpipe->pipe_state |= PIPE_WANT;
1534                 pipeunlock(cpipe);
1535                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1536                 pipelock(cpipe, 0);
1537         }
1538
1539
1540         /*
1541          * Disconnect from peer, if any.
1542          */
1543         ppipe = cpipe->pipe_peer;
1544         if (ppipe->pipe_present == PIPE_ACTIVE) {
1545                 pipeselwakeup(ppipe);
1546
1547                 ppipe->pipe_state |= PIPE_EOF;
1548                 wakeup(ppipe);
1549                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1550         }
1551
1552         /*
1553          * Mark this endpoint as free.  Release kmem resources.  We
1554          * don't mark this endpoint as unused until we've finished
1555          * doing that, or the pipe might disappear out from under
1556          * us.
1557          */
1558         PIPE_UNLOCK(cpipe);
1559         pipe_free_kmem(cpipe);
1560         PIPE_LOCK(cpipe);
1561         cpipe->pipe_present = PIPE_CLOSING;
1562         pipeunlock(cpipe);
1563
1564         /*
1565          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1566          * PIPE_FINALIZED, that allows other end to free the
1567          * pipe_pair, only after the knotes are completely dismantled.
1568          */
1569         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1570         cpipe->pipe_present = PIPE_FINALIZED;
1571         seldrain(&cpipe->pipe_sel);
1572         knlist_destroy(&cpipe->pipe_sel.si_note);
1573
1574         /*
1575          * Postpone the destroy of the fake inode number allocated for
1576          * our end, until pipe mtx is unlocked.
1577          */
1578         ino = cpipe->pipe_ino;
1579
1580         /*
1581          * If both endpoints are now closed, release the memory for the
1582          * pipe pair.  If not, unlock.
1583          */
1584         if (ppipe->pipe_present == PIPE_FINALIZED) {
1585                 PIPE_UNLOCK(cpipe);
1586 #ifdef MAC
1587                 mac_pipe_destroy(pp);
1588 #endif
1589                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1590         } else
1591                 PIPE_UNLOCK(cpipe);
1592
1593         if (ino != 0 && ino != (ino_t)-1)
1594                 free_unr(pipeino_unr, ino);
1595 }
1596
1597 /*ARGSUSED*/
1598 static int
1599 pipe_kqfilter(struct file *fp, struct knote *kn)
1600 {
1601         struct pipe *cpipe;
1602
1603         cpipe = kn->kn_fp->f_data;
1604         PIPE_LOCK(cpipe);
1605         switch (kn->kn_filter) {
1606         case EVFILT_READ:
1607                 kn->kn_fop = &pipe_rfiltops;
1608                 break;
1609         case EVFILT_WRITE:
1610                 kn->kn_fop = &pipe_wfiltops;
1611                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1612                         /* other end of pipe has been closed */
1613                         PIPE_UNLOCK(cpipe);
1614                         return (EPIPE);
1615                 }
1616                 cpipe = cpipe->pipe_peer;
1617                 break;
1618         default:
1619                 PIPE_UNLOCK(cpipe);
1620                 return (EINVAL);
1621         }
1622
1623         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1624         PIPE_UNLOCK(cpipe);
1625         return (0);
1626 }
1627
1628 static void
1629 filt_pipedetach(struct knote *kn)
1630 {
1631         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1632
1633         PIPE_LOCK(cpipe);
1634         if (kn->kn_filter == EVFILT_WRITE)
1635                 cpipe = cpipe->pipe_peer;
1636         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1637         PIPE_UNLOCK(cpipe);
1638 }
1639
1640 /*ARGSUSED*/
1641 static int
1642 filt_piperead(struct knote *kn, long hint)
1643 {
1644         struct pipe *rpipe = kn->kn_fp->f_data;
1645         struct pipe *wpipe = rpipe->pipe_peer;
1646         int ret;
1647
1648         PIPE_LOCK(rpipe);
1649         kn->kn_data = rpipe->pipe_buffer.cnt;
1650         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1651                 kn->kn_data = rpipe->pipe_map.cnt;
1652
1653         if ((rpipe->pipe_state & PIPE_EOF) ||
1654             wpipe->pipe_present != PIPE_ACTIVE ||
1655             (wpipe->pipe_state & PIPE_EOF)) {
1656                 kn->kn_flags |= EV_EOF;
1657                 PIPE_UNLOCK(rpipe);
1658                 return (1);
1659         }
1660         ret = kn->kn_data > 0;
1661         PIPE_UNLOCK(rpipe);
1662         return ret;
1663 }
1664
1665 /*ARGSUSED*/
1666 static int
1667 filt_pipewrite(struct knote *kn, long hint)
1668 {
1669         struct pipe *rpipe = kn->kn_fp->f_data;
1670         struct pipe *wpipe = rpipe->pipe_peer;
1671
1672         PIPE_LOCK(rpipe);
1673         if (wpipe->pipe_present != PIPE_ACTIVE ||
1674             (wpipe->pipe_state & PIPE_EOF)) {
1675                 kn->kn_data = 0;
1676                 kn->kn_flags |= EV_EOF;
1677                 PIPE_UNLOCK(rpipe);
1678                 return (1);
1679         }
1680         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1681             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1682         if (wpipe->pipe_state & PIPE_DIRECTW)
1683                 kn->kn_data = 0;
1684
1685         PIPE_UNLOCK(rpipe);
1686         return (kn->kn_data >= PIPE_BUF);
1687 }