sys/kern/sys_pipe.c

   1 /*-
   2  * Copyright (c) 1996 John S. Dyson
   3  * Copyright (c) 2012 Giovanni Trematerra
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice immediately at the beginning of the file, without modification,
  11  *    this list of conditions, and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Absolutely no warranty of function or purpose is made by the author
  16  *    John S. Dyson.
  17  * 4. Modifications may be freely made to this file if the above conditions
  18  *    are met.
  19  */
  20
  21 /*
  22  * This file contains a high-performance replacement for the socket-based
  23  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  24  * all features of sockets, but does do everything that pipes normally
  25  * do.
  26  */
  27
  28 /*
  29  * This code has two modes of operation, a small write mode and a large
  30  * write mode.  The small write mode acts like conventional pipes with
  31  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  32  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  33  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  34  * memory, and the receiving process copies directly from these pinned pages
  35  * in the sending process.
  36  *
  37  * If the sending process receives a signal, it is possible that it will
  38  * go away, and certainly its address space can change, because control
  39  * is returned back to the user-mode side.  In that case, the pipe code
  40  * arranges to copy the buffer supplied by the user process, to a pageable
  41  * kernel buffer, and the receiving process will grab the data from the
  42  * pageable kernel buffer.  Since signals don't happen all that often,
  43  * the copy operation is normally eliminated.
  44  *
  45  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  46  * happen for small transfers so that the system will not spend all of
  47  * its time context switching.
  48  *
  49  * In order to limit the resource use of pipes, two sysctls exist:
  50  *
  51  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  52  * address space available to us in pipe_map. This value is normally
  53  * autotuned, but may also be loader tuned.
  54  *
  55  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  56  * memory in use by pipes.
  57  *
  58  * Based on how large pipekva is relative to maxpipekva, the following
  59  * will happen:
  60  *
  61  * 0% - 50%:
  62  *     New pipes are given 16K of memory backing, pipes may dynamically
  63  *     grow to as large as 64K where needed.
  64  * 50% - 75%:
  65  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  66  *     existing pipes may NOT grow.
  67  * 75% - 100%:
  68  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  69  *     existing pipes will be shrunk down to 4K whenever possible.
  70  *
  71  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  72  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  73  * resize which MUST occur for reverse-direction pipes when they are
  74  * first used.
  75  *
  76  * Additional information about the current state of pipes may be obtained
  77  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  78  * and kern.ipc.piperesizefail.
  79  *
  80  * Locking rules:  There are two locks present here:  A mutex, used via
  81  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  82  * the flag, as mutexes can not persist over uiomove.  The mutex
  83  * exists only to guard access to the flag, and is not in itself a
  84  * locking mechanism.  Also note that there is only a single mutex for
  85  * both directions of a pipe.
  86  *
  87  * As pipelock() may have to sleep before it can acquire the flag, it
  88  * is important to reread all data after a call to pipelock(); everything
  89  * in the structure may have changed.
  90  */
  91
  92 #include <sys/cdefs.h>
  93 __FBSDID("$FreeBSD$");
  94
  95 #include <sys/param.h>
  96 #include <sys/systm.h>
  97 #include <sys/conf.h>
  98 #include <sys/fcntl.h>
  99 #include <sys/file.h>
 100 #include <sys/filedesc.h>
 101 #include <sys/filio.h>
 102 #include <sys/kernel.h>
 103 #include <sys/lock.h>
 104 #include <sys/mutex.h>
 105 #include <sys/ttycom.h>
 106 #include <sys/stat.h>
 107 #include <sys/malloc.h>
 108 #include <sys/poll.h>
 109 #include <sys/selinfo.h>
 110 #include <sys/signalvar.h>
 111 #include <sys/syscallsubr.h>
 112 #include <sys/sysctl.h>
 113 #include <sys/sysproto.h>
 114 #include <sys/pipe.h>
 115 #include <sys/proc.h>
 116 #include <sys/vnode.h>
 117 #include <sys/uio.h>
 118 #include <sys/event.h>
 119
 120 #include <security/mac/mac_framework.h>
 121
 122 #include <vm/vm.h>
 123 #include <vm/vm_param.h>
 124 #include <vm/vm_object.h>
 125 #include <vm/vm_kern.h>
 126 #include <vm/vm_extern.h>
 127 #include <vm/pmap.h>
 128 #include <vm/vm_map.h>
 129 #include <vm/vm_page.h>
 130 #include <vm/uma.h>
 131
 132 /*
 133  * Use this define if you want to disable *fancy* VM things.  Expect an
 134  * approx 30% decrease in transfer rate.  This could be useful for
 135  * NetBSD or OpenBSD.
 136  */
 137 /* #define PIPE_NODIRECT */
 138
 139 #define PIPE_PEER(pipe) \
 140         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 141
 142 /*
 143  * interfaces to the outside world
 144  */
 145 static fo_rdwr_t        pipe_read;
 146 static fo_rdwr_t        pipe_write;
 147 static fo_truncate_t    pipe_truncate;
 148 static fo_ioctl_t       pipe_ioctl;
 149 static fo_poll_t        pipe_poll;
 150 static fo_kqfilter_t    pipe_kqfilter;
 151 static fo_stat_t        pipe_stat;
 152 static fo_close_t       pipe_close;
 153 static fo_chmod_t       pipe_chmod;
 154 static fo_chown_t       pipe_chown;
 155
 156 struct fileops pipeops = {
 157         .fo_read = pipe_read,
 158         .fo_write = pipe_write,
 159         .fo_truncate = pipe_truncate,
 160         .fo_ioctl = pipe_ioctl,
 161         .fo_poll = pipe_poll,
 162         .fo_kqfilter = pipe_kqfilter,
 163         .fo_stat = pipe_stat,
 164         .fo_close = pipe_close,
 165         .fo_chmod = pipe_chmod,
 166         .fo_chown = pipe_chown,
 167         .fo_sendfile = invfo_sendfile,
 168         .fo_flags = DFLAG_PASSABLE
 169 };
 170
 171 static void     filt_pipedetach(struct knote *kn);
 172 static void     filt_pipedetach_notsup(struct knote *kn);
 173 static int      filt_pipenotsup(struct knote *kn, long hint);
 174 static int      filt_piperead(struct knote *kn, long hint);
 175 static int      filt_pipewrite(struct knote *kn, long hint);
 176
 177 static struct filterops pipe_nfiltops = {
 178         .f_isfd = 1,
 179         .f_detach = filt_pipedetach_notsup,
 180         .f_event = filt_pipenotsup
 181 };
 182 static struct filterops pipe_rfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach,
 185         .f_event = filt_piperead
 186 };
 187 static struct filterops pipe_wfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_pipewrite
 191 };
 192
 193 /*
 194  * Default pipe buffer size(s), this can be kind-of large now because pipe
 195  * space is pageable.  The pipe code will try to maintain locality of
 196  * reference for performance reasons, so small amounts of outstanding I/O
 197  * will not wipe the cache.
 198  */
 199 #define MINPIPESIZE (PIPE_SIZE/3)
 200 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 201
 202 static long amountpipekva;
 203 static int pipefragretry;
 204 static int pipeallocfail;
 205 static int piperesizefail;
 206 static int piperesizeallowed = 1;
 207
 208 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN,
 209            &maxpipekva, 0, "Pipe KVA limit");
 210 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 211            &amountpipekva, 0, "Pipe KVA usage");
 212 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 213           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 214 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 215           &pipeallocfail, 0, "Pipe allocation failures");
 216 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 217           &piperesizefail, 0, "Pipe resize failures");
 218 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 219           &piperesizeallowed, 0, "Pipe resizing allowed");
 220
 221 static void pipeinit(void *dummy __unused);
 222 static void pipeclose(struct pipe *cpipe);
 223 static void pipe_free_kmem(struct pipe *cpipe);
 224 static void pipe_create(struct pipe *pipe, int backing);
 225 static void pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 226 static __inline int pipelock(struct pipe *cpipe, int catch);
 227 static __inline void pipeunlock(struct pipe *cpipe);
 228 #ifndef PIPE_NODIRECT
 229 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 230 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 231 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 232 static void pipe_clone_write_buffer(struct pipe *wpipe);
 233 #endif
 234 static int pipespace(struct pipe *cpipe, int size);
 235 static int pipespace_new(struct pipe *cpipe, int size);
 236
 237 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 238 static int      pipe_zone_init(void *mem, int size, int flags);
 239 static void     pipe_zone_fini(void *mem, int size);
 240
 241 static uma_zone_t pipe_zone;
 242 static struct unrhdr *pipeino_unr;
 243 static dev_t pipedev_ino;
 244
 245 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 246
 247 static void
 248 pipeinit(void *dummy __unused)
 249 {
 250
 251         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 252             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 253             UMA_ALIGN_PTR, 0);
 254         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 255         pipeino_unr = new_unrhdr(1, INT32_MAX, NULL);
 256         KASSERT(pipeino_unr != NULL, ("pipe fake inodes not initialized"));
 257         pipedev_ino = devfs_alloc_cdp_inode();
 258         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 259 }
 260
 261 static int
 262 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 263 {
 264         struct pipepair *pp;
 265         struct pipe *rpipe, *wpipe;
 266
 267         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 268
 269         pp = (struct pipepair *)mem;
 270
 271         /*
 272          * We zero both pipe endpoints to make sure all the kmem pointers
 273          * are NULL, flag fields are zero'd, etc.  We timestamp both
 274          * endpoints with the same time.
 275          */
 276         rpipe = &pp->pp_rpipe;
 277         bzero(rpipe, sizeof(*rpipe));
 278         vfs_timestamp(&rpipe->pipe_ctime);
 279         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 280
 281         wpipe = &pp->pp_wpipe;
 282         bzero(wpipe, sizeof(*wpipe));
 283         wpipe->pipe_ctime = rpipe->pipe_ctime;
 284         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 285
 286         rpipe->pipe_peer = wpipe;
 287         rpipe->pipe_pair = pp;
 288         wpipe->pipe_peer = rpipe;
 289         wpipe->pipe_pair = pp;
 290
 291         /*
 292          * Mark both endpoints as present; they will later get free'd
 293          * one at a time.  When both are free'd, then the whole pair
 294          * is released.
 295          */
 296         rpipe->pipe_present = PIPE_ACTIVE;
 297         wpipe->pipe_present = PIPE_ACTIVE;
 298
 299         /*
 300          * Eventually, the MAC Framework may initialize the label
 301          * in ctor or init, but for now we do it elswhere to avoid
 302          * blocking in ctor or init.
 303          */
 304         pp->pp_label = NULL;
 305
 306         return (0);
 307 }
 308
 309 static int
 310 pipe_zone_init(void *mem, int size, int flags)
 311 {
 312         struct pipepair *pp;
 313
 314         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 315
 316         pp = (struct pipepair *)mem;
 317
 318         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF);
 319         return (0);
 320 }
 321
 322 static void
 323 pipe_zone_fini(void *mem, int size)
 324 {
 325         struct pipepair *pp;
 326
 327         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 328
 329         pp = (struct pipepair *)mem;
 330
 331         mtx_destroy(&pp->pp_mtx);
 332 }
 333
 334 static void
 335 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 336 {
 337         struct pipepair *pp;
 338         struct pipe *rpipe, *wpipe;
 339
 340         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 341 #ifdef MAC
 342         /*
 343          * The MAC label is shared between the connected endpoints.  As a
 344          * result mac_pipe_init() and mac_pipe_create() are called once
 345          * for the pair, and not on the endpoints.
 346          */
 347         mac_pipe_init(pp);
 348         mac_pipe_create(td->td_ucred, pp);
 349 #endif
 350         rpipe = &pp->pp_rpipe;
 351         wpipe = &pp->pp_wpipe;
 352
 353         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 354         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 355
 356         /* Only the forward direction pipe is backed by default */
 357         pipe_create(rpipe, 1);
 358         pipe_create(wpipe, 0);
 359
 360         rpipe->pipe_state |= PIPE_DIRECTOK;
 361         wpipe->pipe_state |= PIPE_DIRECTOK;
 362 }
 363
 364 void
 365 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 366 {
 367         struct pipepair *pp;
 368
 369         pipe_paircreate(td, &pp);
 370         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 371         *ppipe = &pp->pp_rpipe;
 372 }
 373
 374 void
 375 pipe_dtor(struct pipe *dpipe)
 376 {
 377         struct pipe *peer;
 378         ino_t ino;
 379
 380         ino = dpipe->pipe_ino;
 381         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 382         funsetown(&dpipe->pipe_sigio);
 383         pipeclose(dpipe);
 384         if (peer != NULL) {
 385                 funsetown(&peer->pipe_sigio);
 386                 pipeclose(peer);
 387         }
 388         if (ino != 0 && ino != (ino_t)-1)
 389                 free_unr(pipeino_unr, ino);
 390 }
 391
 392 /*
 393  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 394  * the zone pick up the pieces via pipeclose().
 395  */
 396 int
 397 kern_pipe(struct thread *td, int fildes[2])
 398 {
 399
 400         return (kern_pipe2(td, fildes, 0));
 401 }
 402
 403 int
 404 kern_pipe2(struct thread *td, int fildes[2], int flags)
 405 {
 406         struct filedesc *fdp;
 407         struct file *rf, *wf;
 408         struct pipe *rpipe, *wpipe;
 409         struct pipepair *pp;
 410         int fd, fflags, error;
 411
 412         fdp = td->td_proc->p_fd;
 413         pipe_paircreate(td, &pp);
 414         rpipe = &pp->pp_rpipe;
 415         wpipe = &pp->pp_wpipe;
 416         error = falloc(td, &rf, &fd, flags);
 417         if (error) {
 418                 pipeclose(rpipe);
 419                 pipeclose(wpipe);
 420                 return (error);
 421         }
 422         /* An extra reference on `rf' has been held for us by falloc(). */
 423         fildes[0] = fd;
 424
 425         fflags = FREAD | FWRITE;
 426         if ((flags & O_NONBLOCK) != 0)
 427                 fflags |= FNONBLOCK;
 428
 429         /*
 430          * Warning: once we've gotten past allocation of the fd for the
 431          * read-side, we can only drop the read side via fdrop() in order
 432          * to avoid races against processes which manage to dup() the read
 433          * side while we are blocked trying to allocate the write side.
 434          */
 435         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 436         error = falloc(td, &wf, &fd, flags);
 437         if (error) {
 438                 fdclose(fdp, rf, fildes[0], td);
 439                 fdrop(rf, td);
 440                 /* rpipe has been closed by fdrop(). */
 441                 pipeclose(wpipe);
 442                 return (error);
 443         }
 444         /* An extra reference on `wf' has been held for us by falloc(). */
 445         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 446         fdrop(wf, td);
 447         fildes[1] = fd;
 448         fdrop(rf, td);
 449
 450         return (0);
 451 }
 452
 453 /* ARGSUSED */
 454 int
 455 sys_pipe(struct thread *td, struct pipe_args *uap)
 456 {
 457         int error;
 458         int fildes[2];
 459
 460         error = kern_pipe(td, fildes);
 461         if (error)
 462                 return (error);
 463
 464         td->td_retval[0] = fildes[0];
 465         td->td_retval[1] = fildes[1];
 466
 467         return (0);
 468 }
 469
 470 int
 471 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 472 {
 473         int error, fildes[2];
 474
 475         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 476                 return (EINVAL);
 477         error = kern_pipe2(td, fildes, uap->flags);
 478         if (error)
 479                 return (error);
 480         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 481         if (error) {
 482                 (void)kern_close(td, fildes[0]);
 483                 (void)kern_close(td, fildes[1]);
 484         }
 485         return (error);
 486 }
 487
 488 /*
 489  * Allocate kva for pipe circular buffer, the space is pageable
 490  * This routine will 'realloc' the size of a pipe safely, if it fails
 491  * it will retain the old buffer.
 492  * If it fails it will return ENOMEM.
 493  */
 494 static int
 495 pipespace_new(cpipe, size)
 496         struct pipe *cpipe;
 497         int size;
 498 {
 499         caddr_t buffer;
 500         int error, cnt, firstseg;
 501         static int curfail = 0;
 502         static struct timeval lastfail;
 503
 504         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 505         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 506                 ("pipespace: resize of direct writes not allowed"));
 507 retry:
 508         cnt = cpipe->pipe_buffer.cnt;
 509         if (cnt > size)
 510                 size = cnt;
 511
 512         size = round_page(size);
 513         buffer = (caddr_t) vm_map_min(pipe_map);
 514
 515         error = vm_map_find(pipe_map, NULL, 0,
 516                 (vm_offset_t *) &buffer, size, 0, VMFS_ANY_SPACE,
 517                 VM_PROT_ALL, VM_PROT_ALL, 0);
 518         if (error != KERN_SUCCESS) {
 519                 if ((cpipe->pipe_buffer.buffer == NULL) &&
 520                         (size > SMALL_PIPE_SIZE)) {
 521                         size = SMALL_PIPE_SIZE;
 522                         pipefragretry++;
 523                         goto retry;
 524                 }
 525                 if (cpipe->pipe_buffer.buffer == NULL) {
 526                         pipeallocfail++;
 527                         if (ppsratecheck(&lastfail, &curfail, 1))
 528                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 529                 } else {
 530                         piperesizefail++;
 531                 }
 532                 return (ENOMEM);
 533         }
 534
 535         /* copy data, then free old resources if we're resizing */
 536         if (cnt > 0) {
 537                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 538                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 539                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 540                                 buffer, firstseg);
 541                         if ((cnt - firstseg) > 0)
 542                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 543                                         cpipe->pipe_buffer.in);
 544                 } else {
 545                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 546                                 buffer, cnt);
 547                 }
 548         }
 549         pipe_free_kmem(cpipe);
 550         cpipe->pipe_buffer.buffer = buffer;
 551         cpipe->pipe_buffer.size = size;
 552         cpipe->pipe_buffer.in = cnt;
 553         cpipe->pipe_buffer.out = 0;
 554         cpipe->pipe_buffer.cnt = cnt;
 555         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 556         return (0);
 557 }
 558
 559 /*
 560  * Wrapper for pipespace_new() that performs locking assertions.
 561  */
 562 static int
 563 pipespace(cpipe, size)
 564         struct pipe *cpipe;
 565         int size;
 566 {
 567
 568         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 569                 ("Unlocked pipe passed to pipespace"));
 570         return (pipespace_new(cpipe, size));
 571 }
 572
 573 /*
 574  * lock a pipe for I/O, blocking other access
 575  */
 576 static __inline int
 577 pipelock(cpipe, catch)
 578         struct pipe *cpipe;
 579         int catch;
 580 {
 581         int error;
 582
 583         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 584         while (cpipe->pipe_state & PIPE_LOCKFL) {
 585                 cpipe->pipe_state |= PIPE_LWANT;
 586                 error = msleep(cpipe, PIPE_MTX(cpipe),
 587                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 588                     "pipelk", 0);
 589                 if (error != 0)
 590                         return (error);
 591         }
 592         cpipe->pipe_state |= PIPE_LOCKFL;
 593         return (0);
 594 }
 595
 596 /*
 597  * unlock a pipe I/O lock
 598  */
 599 static __inline void
 600 pipeunlock(cpipe)
 601         struct pipe *cpipe;
 602 {
 603
 604         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 605         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 606                 ("Unlocked pipe passed to pipeunlock"));
 607         cpipe->pipe_state &= ~PIPE_LOCKFL;
 608         if (cpipe->pipe_state & PIPE_LWANT) {
 609                 cpipe->pipe_state &= ~PIPE_LWANT;
 610                 wakeup(cpipe);
 611         }
 612 }
 613
 614 void
 615 pipeselwakeup(cpipe)
 616         struct pipe *cpipe;
 617 {
 618
 619         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 620         if (cpipe->pipe_state & PIPE_SEL) {
 621                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 622                 if (!SEL_WAITING(&cpipe->pipe_sel))
 623                         cpipe->pipe_state &= ~PIPE_SEL;
 624         }
 625         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 626                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 627         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 628 }
 629
 630 /*
 631  * Initialize and allocate VM and memory for pipe.  The structure
 632  * will start out zero'd from the ctor, so we just manage the kmem.
 633  */
 634 static void
 635 pipe_create(pipe, backing)
 636         struct pipe *pipe;
 637         int backing;
 638 {
 639
 640         if (backing) {
 641                 /*
 642                  * Note that these functions can fail if pipe map is exhausted
 643                  * (as a result of too many pipes created), but we ignore the
 644                  * error as it is not fatal and could be provoked by
 645                  * unprivileged users. The only consequence is worse performance
 646                  * with given pipe.
 647                  */
 648                 if (amountpipekva > maxpipekva / 2)
 649                         (void)pipespace_new(pipe, SMALL_PIPE_SIZE);
 650                 else
 651                         (void)pipespace_new(pipe, PIPE_SIZE);
 652         }
 653
 654         pipe->pipe_ino = -1;
 655 }
 656
 657 /* ARGSUSED */
 658 static int
 659 pipe_read(fp, uio, active_cred, flags, td)
 660         struct file *fp;
 661         struct uio *uio;
 662         struct ucred *active_cred;
 663         struct thread *td;
 664         int flags;
 665 {
 666         struct pipe *rpipe;
 667         int error;
 668         int nread = 0;
 669         int size;
 670
 671         rpipe = fp->f_data;
 672         PIPE_LOCK(rpipe);
 673         ++rpipe->pipe_busy;
 674         error = pipelock(rpipe, 1);
 675         if (error)
 676                 goto unlocked_error;
 677
 678 #ifdef MAC
 679         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 680         if (error)
 681                 goto locked_error;
 682 #endif
 683         if (amountpipekva > (3 * maxpipekva) / 4) {
 684                 if (!(rpipe->pipe_state & PIPE_DIRECTW) &&
 685                         (rpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
 686                         (rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
 687                         (piperesizeallowed == 1)) {
 688                         PIPE_UNLOCK(rpipe);
 689                         pipespace(rpipe, SMALL_PIPE_SIZE);
 690                         PIPE_LOCK(rpipe);
 691                 }
 692         }
 693
 694         while (uio->uio_resid) {
 695                 /*
 696                  * normal pipe buffer receive
 697                  */
 698                 if (rpipe->pipe_buffer.cnt > 0) {
 699                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 700                         if (size > rpipe->pipe_buffer.cnt)
 701                                 size = rpipe->pipe_buffer.cnt;
 702                         if (size > uio->uio_resid)
 703                                 size = uio->uio_resid;
 704
 705                         PIPE_UNLOCK(rpipe);
 706                         error = uiomove(
 707                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 708                             size, uio);
 709                         PIPE_LOCK(rpipe);
 710                         if (error)
 711                                 break;
 712
 713                         rpipe->pipe_buffer.out += size;
 714                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 715                                 rpipe->pipe_buffer.out = 0;
 716
 717                         rpipe->pipe_buffer.cnt -= size;
 718
 719                         /*
 720                          * If there is no more to read in the pipe, reset
 721                          * its pointers to the beginning.  This improves
 722                          * cache hit stats.
 723                          */
 724                         if (rpipe->pipe_buffer.cnt == 0) {
 725                                 rpipe->pipe_buffer.in = 0;
 726                                 rpipe->pipe_buffer.out = 0;
 727                         }
 728                         nread += size;
 729 #ifndef PIPE_NODIRECT
 730                 /*
 731                  * Direct copy, bypassing a kernel buffer.
 732                  */
 733                 } else if ((size = rpipe->pipe_map.cnt) &&
 734                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 735                         if (size > uio->uio_resid)
 736                                 size = (u_int) uio->uio_resid;
 737
 738                         PIPE_UNLOCK(rpipe);
 739                         error = uiomove_fromphys(rpipe->pipe_map.ms,
 740                             rpipe->pipe_map.pos, size, uio);
 741                         PIPE_LOCK(rpipe);
 742                         if (error)
 743                                 break;
 744                         nread += size;
 745                         rpipe->pipe_map.pos += size;
 746                         rpipe->pipe_map.cnt -= size;
 747                         if (rpipe->pipe_map.cnt == 0) {
 748                                 rpipe->pipe_state &= ~(PIPE_DIRECTW|PIPE_WANTW);
 749                                 wakeup(rpipe);
 750                         }
 751 #endif
 752                 } else {
 753                         /*
 754                          * detect EOF condition
 755                          * read returns 0 on EOF, no need to set error
 756                          */
 757                         if (rpipe->pipe_state & PIPE_EOF)
 758                                 break;
 759
 760                         /*
 761                          * If the "write-side" has been blocked, wake it up now.
 762                          */
 763                         if (rpipe->pipe_state & PIPE_WANTW) {
 764                                 rpipe->pipe_state &= ~PIPE_WANTW;
 765                                 wakeup(rpipe);
 766                         }
 767
 768                         /*
 769                          * Break if some data was read.
 770                          */
 771                         if (nread > 0)
 772                                 break;
 773
 774                         /*
 775                          * Unlock the pipe buffer for our remaining processing.
 776                          * We will either break out with an error or we will
 777                          * sleep and relock to loop.
 778                          */
 779                         pipeunlock(rpipe);
 780
 781                         /*
 782                          * Handle non-blocking mode operation or
 783                          * wait for more data.
 784                          */
 785                         if (fp->f_flag & FNONBLOCK) {
 786                                 error = EAGAIN;
 787                         } else {
 788                                 rpipe->pipe_state |= PIPE_WANTR;
 789                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 790                                     PRIBIO | PCATCH,
 791                                     "piperd", 0)) == 0)
 792                                         error = pipelock(rpipe, 1);
 793                         }
 794                         if (error)
 795                                 goto unlocked_error;
 796                 }
 797         }
 798 #ifdef MAC
 799 locked_error:
 800 #endif
 801         pipeunlock(rpipe);
 802
 803         /* XXX: should probably do this before getting any locks. */
 804         if (error == 0)
 805                 vfs_timestamp(&rpipe->pipe_atime);
 806 unlocked_error:
 807         --rpipe->pipe_busy;
 808
 809         /*
 810          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 811          */
 812         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 813                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 814                 wakeup(rpipe);
 815         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 816                 /*
 817                  * Handle write blocking hysteresis.
 818                  */
 819                 if (rpipe->pipe_state & PIPE_WANTW) {
 820                         rpipe->pipe_state &= ~PIPE_WANTW;
 821                         wakeup(rpipe);
 822                 }
 823         }
 824
 825         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 826                 pipeselwakeup(rpipe);
 827
 828         PIPE_UNLOCK(rpipe);
 829         return (error);
 830 }
 831
 832 #ifndef PIPE_NODIRECT
 833 /*
 834  * Map the sending processes' buffer into kernel space and wire it.
 835  * This is similar to a physical write operation.
 836  */
 837 static int
 838 pipe_build_write_buffer(wpipe, uio)
 839         struct pipe *wpipe;
 840         struct uio *uio;
 841 {
 842         u_int size;
 843         int i;
 844
 845         PIPE_LOCK_ASSERT(wpipe, MA_NOTOWNED);
 846         KASSERT(wpipe->pipe_state & PIPE_DIRECTW,
 847                 ("Clone attempt on non-direct write pipe!"));
 848
 849         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 850                 size = wpipe->pipe_buffer.size;
 851         else
 852                 size = uio->uio_iov->iov_len;
 853
 854         if ((i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 855             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 856             wpipe->pipe_map.ms, PIPENPAGES)) < 0)
 857                 return (EFAULT);
 858
 859 /*
 860  * set up the control block
 861  */
 862         wpipe->pipe_map.npages = i;
 863         wpipe->pipe_map.pos =
 864             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 865         wpipe->pipe_map.cnt = size;
 866
 867 /*
 868  * and update the uio data
 869  */
 870
 871         uio->uio_iov->iov_len -= size;
 872         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 873         if (uio->uio_iov->iov_len == 0)
 874                 uio->uio_iov++;
 875         uio->uio_resid -= size;
 876         uio->uio_offset += size;
 877         return (0);
 878 }
 879
 880 /*
 881  * unmap and unwire the process buffer
 882  */
 883 static void
 884 pipe_destroy_write_buffer(wpipe)
 885         struct pipe *wpipe;
 886 {
 887
 888         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 889         vm_page_unhold_pages(wpipe->pipe_map.ms, wpipe->pipe_map.npages);
 890         wpipe->pipe_map.npages = 0;
 891 }
 892
 893 /*
 894  * In the case of a signal, the writing process might go away.  This
 895  * code copies the data into the circular buffer so that the source
 896  * pages can be freed without loss of data.
 897  */
 898 static void
 899 pipe_clone_write_buffer(wpipe)
 900         struct pipe *wpipe;
 901 {
 902         struct uio uio;
 903         struct iovec iov;
 904         int size;
 905         int pos;
 906
 907         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 908         size = wpipe->pipe_map.cnt;
 909         pos = wpipe->pipe_map.pos;
 910
 911         wpipe->pipe_buffer.in = size;
 912         wpipe->pipe_buffer.out = 0;
 913         wpipe->pipe_buffer.cnt = size;
 914         wpipe->pipe_state &= ~PIPE_DIRECTW;
 915
 916         PIPE_UNLOCK(wpipe);
 917         iov.iov_base = wpipe->pipe_buffer.buffer;
 918         iov.iov_len = size;
 919         uio.uio_iov = &iov;
 920         uio.uio_iovcnt = 1;
 921         uio.uio_offset = 0;
 922         uio.uio_resid = size;
 923         uio.uio_segflg = UIO_SYSSPACE;
 924         uio.uio_rw = UIO_READ;
 925         uio.uio_td = curthread;
 926         uiomove_fromphys(wpipe->pipe_map.ms, pos, size, &uio);
 927         PIPE_LOCK(wpipe);
 928         pipe_destroy_write_buffer(wpipe);
 929 }
 930
 931 /*
 932  * This implements the pipe buffer write mechanism.  Note that only
 933  * a direct write OR a normal pipe write can be pending at any given time.
 934  * If there are any characters in the pipe buffer, the direct write will
 935  * be deferred until the receiving process grabs all of the bytes from
 936  * the pipe buffer.  Then the direct mapping write is set-up.
 937  */
 938 static int
 939 pipe_direct_write(wpipe, uio)
 940         struct pipe *wpipe;
 941         struct uio *uio;
 942 {
 943         int error;
 944
 945 retry:
 946         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 947         error = pipelock(wpipe, 1);
 948         if (error != 0)
 949                 goto error1;
 950         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 951                 error = EPIPE;
 952                 pipeunlock(wpipe);
 953                 goto error1;
 954         }
 955         while (wpipe->pipe_state & PIPE_DIRECTW) {
 956                 if (wpipe->pipe_state & PIPE_WANTR) {
 957                         wpipe->pipe_state &= ~PIPE_WANTR;
 958                         wakeup(wpipe);
 959                 }
 960                 pipeselwakeup(wpipe);
 961                 wpipe->pipe_state |= PIPE_WANTW;
 962                 pipeunlock(wpipe);
 963                 error = msleep(wpipe, PIPE_MTX(wpipe),
 964                     PRIBIO | PCATCH, "pipdww", 0);
 965                 if (error)
 966                         goto error1;
 967                 else
 968                         goto retry;
 969         }
 970         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 971         if (wpipe->pipe_buffer.cnt > 0) {
 972                 if (wpipe->pipe_state & PIPE_WANTR) {
 973                         wpipe->pipe_state &= ~PIPE_WANTR;
 974                         wakeup(wpipe);
 975                 }
 976                 pipeselwakeup(wpipe);
 977                 wpipe->pipe_state |= PIPE_WANTW;
 978                 pipeunlock(wpipe);
 979                 error = msleep(wpipe, PIPE_MTX(wpipe),
 980                     PRIBIO | PCATCH, "pipdwc", 0);
 981                 if (error)
 982                         goto error1;
 983                 else
 984                         goto retry;
 985         }
 986
 987         wpipe->pipe_state |= PIPE_DIRECTW;
 988
 989         PIPE_UNLOCK(wpipe);
 990         error = pipe_build_write_buffer(wpipe, uio);
 991         PIPE_LOCK(wpipe);
 992         if (error) {
 993                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 994                 pipeunlock(wpipe);
 995                 goto error1;
 996         }
 997
 998         error = 0;
 999         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
1000                 if (wpipe->pipe_state & PIPE_EOF) {
1001                         pipe_destroy_write_buffer(wpipe);
1002                         pipeselwakeup(wpipe);
1003                         pipeunlock(wpipe);
1004                         error = EPIPE;
1005                         goto error1;
1006                 }
1007                 if (wpipe->pipe_state & PIPE_WANTR) {
1008                         wpipe->pipe_state &= ~PIPE_WANTR;
1009                         wakeup(wpipe);
1010                 }
1011                 pipeselwakeup(wpipe);
1012                 wpipe->pipe_state |= PIPE_WANTW;
1013                 pipeunlock(wpipe);
1014                 error = msleep(wpipe, PIPE_MTX(wpipe), PRIBIO | PCATCH,
1015                     "pipdwt", 0);
1016                 pipelock(wpipe, 0);
1017         }
1018
1019         if (wpipe->pipe_state & PIPE_EOF)
1020                 error = EPIPE;
1021         if (wpipe->pipe_state & PIPE_DIRECTW) {
1022                 /*
1023                  * this bit of trickery substitutes a kernel buffer for
1024                  * the process that might be going away.
1025                  */
1026                 pipe_clone_write_buffer(wpipe);
1027         } else {
1028                 pipe_destroy_write_buffer(wpipe);
1029         }
1030         pipeunlock(wpipe);
1031         return (error);
1032
1033 error1:
1034         wakeup(wpipe);
1035         return (error);
1036 }
1037 #endif
1038
1039 static int
1040 pipe_write(fp, uio, active_cred, flags, td)
1041         struct file *fp;
1042         struct uio *uio;
1043         struct ucred *active_cred;
1044         struct thread *td;
1045         int flags;
1046 {
1047         int error = 0;
1048         int desiredsize;
1049         ssize_t orig_resid;
1050         struct pipe *wpipe, *rpipe;
1051
1052         rpipe = fp->f_data;
1053         wpipe = PIPE_PEER(rpipe);
1054         PIPE_LOCK(rpipe);
1055         error = pipelock(wpipe, 1);
1056         if (error) {
1057                 PIPE_UNLOCK(rpipe);
1058                 return (error);
1059         }
1060         /*
1061          * detect loss of pipe read side, issue SIGPIPE if lost.
1062          */
1063         if (wpipe->pipe_present != PIPE_ACTIVE ||
1064             (wpipe->pipe_state & PIPE_EOF)) {
1065                 pipeunlock(wpipe);
1066                 PIPE_UNLOCK(rpipe);
1067                 return (EPIPE);
1068         }
1069 #ifdef MAC
1070         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1071         if (error) {
1072                 pipeunlock(wpipe);
1073                 PIPE_UNLOCK(rpipe);
1074                 return (error);
1075         }
1076 #endif
1077         ++wpipe->pipe_busy;
1078
1079         /* Choose a larger size if it's advantageous */
1080         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1081         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1082                 if (piperesizeallowed != 1)
1083                         break;
1084                 if (amountpipekva > maxpipekva / 2)
1085                         break;
1086                 if (desiredsize == BIG_PIPE_SIZE)
1087                         break;
1088                 desiredsize = desiredsize * 2;
1089         }
1090
1091         /* Choose a smaller size if we're in a OOM situation */
1092         if ((amountpipekva > (3 * maxpipekva) / 4) &&
1093                 (wpipe->pipe_buffer.size > SMALL_PIPE_SIZE) &&
1094                 (wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE) &&
1095                 (piperesizeallowed == 1))
1096                 desiredsize = SMALL_PIPE_SIZE;
1097
1098         /* Resize if the above determined that a new size was necessary */
1099         if ((desiredsize != wpipe->pipe_buffer.size) &&
1100                 ((wpipe->pipe_state & PIPE_DIRECTW) == 0)) {
1101                 PIPE_UNLOCK(wpipe);
1102                 pipespace(wpipe, desiredsize);
1103                 PIPE_LOCK(wpipe);
1104         }
1105         if (wpipe->pipe_buffer.size == 0) {
1106                 /*
1107                  * This can only happen for reverse direction use of pipes
1108                  * in a complete OOM situation.
1109                  */
1110                 error = ENOMEM;
1111                 --wpipe->pipe_busy;
1112                 pipeunlock(wpipe);
1113                 PIPE_UNLOCK(wpipe);
1114                 return (error);
1115         }
1116
1117         pipeunlock(wpipe);
1118
1119         orig_resid = uio->uio_resid;
1120
1121         while (uio->uio_resid) {
1122                 int space;
1123
1124                 pipelock(wpipe, 0);
1125                 if (wpipe->pipe_state & PIPE_EOF) {
1126                         pipeunlock(wpipe);
1127                         error = EPIPE;
1128                         break;
1129                 }
1130 #ifndef PIPE_NODIRECT
1131                 /*
1132                  * If the transfer is large, we can gain performance if
1133                  * we do process-to-process copies directly.
1134                  * If the write is non-blocking, we don't use the
1135                  * direct write mechanism.
1136                  *
1137                  * The direct write mechanism will detect the reader going
1138                  * away on us.
1139                  */
1140                 if (uio->uio_segflg == UIO_USERSPACE &&
1141                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1142                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1143                     (fp->f_flag & FNONBLOCK) == 0) {
1144                         pipeunlock(wpipe);
1145                         error = pipe_direct_write(wpipe, uio);
1146                         if (error)
1147                                 break;
1148                         continue;
1149                 }
1150 #endif
1151
1152                 /*
1153                  * Pipe buffered writes cannot be coincidental with
1154                  * direct writes.  We wait until the currently executing
1155                  * direct write is completed before we start filling the
1156                  * pipe buffer.  We break out if a signal occurs or the
1157                  * reader goes away.
1158                  */
1159                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1160                         if (wpipe->pipe_state & PIPE_WANTR) {
1161                                 wpipe->pipe_state &= ~PIPE_WANTR;
1162                                 wakeup(wpipe);
1163                         }
1164                         pipeselwakeup(wpipe);
1165                         wpipe->pipe_state |= PIPE_WANTW;
1166                         pipeunlock(wpipe);
1167                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1168                             "pipbww", 0);
1169                         if (error)
1170                                 break;
1171                         else
1172                                 continue;
1173                 }
1174
1175                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1176
1177                 /* Writes of size <= PIPE_BUF must be atomic. */
1178                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1179                         space = 0;
1180
1181                 if (space > 0) {
1182                         int size;       /* Transfer size */
1183                         int segsize;    /* first segment to transfer */
1184
1185                         /*
1186                          * Transfer size is minimum of uio transfer
1187                          * and free space in pipe buffer.
1188                          */
1189                         if (space > uio->uio_resid)
1190                                 size = uio->uio_resid;
1191                         else
1192                                 size = space;
1193                         /*
1194                          * First segment to transfer is minimum of
1195                          * transfer size and contiguous space in
1196                          * pipe buffer.  If first segment to transfer
1197                          * is less than the transfer size, we've got
1198                          * a wraparound in the buffer.
1199                          */
1200                         segsize = wpipe->pipe_buffer.size -
1201                                 wpipe->pipe_buffer.in;
1202                         if (segsize > size)
1203                                 segsize = size;
1204
1205                         /* Transfer first segment */
1206
1207                         PIPE_UNLOCK(rpipe);
1208                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1209                                         segsize, uio);
1210                         PIPE_LOCK(rpipe);
1211
1212                         if (error == 0 && segsize < size) {
1213                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1214                                         wpipe->pipe_buffer.size,
1215                                         ("Pipe buffer wraparound disappeared"));
1216                                 /*
1217                                  * Transfer remaining part now, to
1218                                  * support atomic writes.  Wraparound
1219                                  * happened.
1220                                  */
1221
1222                                 PIPE_UNLOCK(rpipe);
1223                                 error = uiomove(
1224                                     &wpipe->pipe_buffer.buffer[0],
1225                                     size - segsize, uio);
1226                                 PIPE_LOCK(rpipe);
1227                         }
1228                         if (error == 0) {
1229                                 wpipe->pipe_buffer.in += size;
1230                                 if (wpipe->pipe_buffer.in >=
1231                                     wpipe->pipe_buffer.size) {
1232                                         KASSERT(wpipe->pipe_buffer.in ==
1233                                                 size - segsize +
1234                                                 wpipe->pipe_buffer.size,
1235                                                 ("Expected wraparound bad"));
1236                                         wpipe->pipe_buffer.in = size - segsize;
1237                                 }
1238
1239                                 wpipe->pipe_buffer.cnt += size;
1240                                 KASSERT(wpipe->pipe_buffer.cnt <=
1241                                         wpipe->pipe_buffer.size,
1242                                         ("Pipe buffer overflow"));
1243                         }
1244                         pipeunlock(wpipe);
1245                         if (error != 0)
1246                                 break;
1247                 } else {
1248                         /*
1249                          * If the "read-side" has been blocked, wake it up now.
1250                          */
1251                         if (wpipe->pipe_state & PIPE_WANTR) {
1252                                 wpipe->pipe_state &= ~PIPE_WANTR;
1253                                 wakeup(wpipe);
1254                         }
1255
1256                         /*
1257                          * don't block on non-blocking I/O
1258                          */
1259                         if (fp->f_flag & FNONBLOCK) {
1260                                 error = EAGAIN;
1261                                 pipeunlock(wpipe);
1262                                 break;
1263                         }
1264
1265                         /*
1266                          * We have no more space and have something to offer,
1267                          * wake up select/poll.
1268                          */
1269                         pipeselwakeup(wpipe);
1270
1271                         wpipe->pipe_state |= PIPE_WANTW;
1272                         pipeunlock(wpipe);
1273                         error = msleep(wpipe, PIPE_MTX(rpipe),
1274                             PRIBIO | PCATCH, "pipewr", 0);
1275                         if (error != 0)
1276                                 break;
1277                 }
1278         }
1279
1280         pipelock(wpipe, 0);
1281         --wpipe->pipe_busy;
1282
1283         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1284                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1285                 wakeup(wpipe);
1286         } else if (wpipe->pipe_buffer.cnt > 0) {
1287                 /*
1288                  * If we have put any characters in the buffer, we wake up
1289                  * the reader.
1290                  */
1291                 if (wpipe->pipe_state & PIPE_WANTR) {
1292                         wpipe->pipe_state &= ~PIPE_WANTR;
1293                         wakeup(wpipe);
1294                 }
1295         }
1296
1297         /*
1298          * Don't return EPIPE if any byte was written.
1299          * EINTR and other interrupts are handled by generic I/O layer.
1300          * Do not pretend that I/O succeeded for obvious user error
1301          * like EFAULT.
1302          */
1303         if (uio->uio_resid != orig_resid && error == EPIPE)
1304                 error = 0;
1305
1306         if (error == 0)
1307                 vfs_timestamp(&wpipe->pipe_mtime);
1308
1309         /*
1310          * We have something to offer,
1311          * wake up select/poll.
1312          */
1313         if (wpipe->pipe_buffer.cnt)
1314                 pipeselwakeup(wpipe);
1315
1316         pipeunlock(wpipe);
1317         PIPE_UNLOCK(rpipe);
1318         return (error);
1319 }
1320
1321 /* ARGSUSED */
1322 static int
1323 pipe_truncate(fp, length, active_cred, td)
1324         struct file *fp;
1325         off_t length;
1326         struct ucred *active_cred;
1327         struct thread *td;
1328 {
1329
1330         /* For named pipes call the vnode operation. */
1331         if (fp->f_vnode != NULL)
1332                 return (vnops.fo_truncate(fp, length, active_cred, td));
1333         return (EINVAL);
1334 }
1335
1336 /*
1337  * we implement a very minimal set of ioctls for compatibility with sockets.
1338  */
1339 static int
1340 pipe_ioctl(fp, cmd, data, active_cred, td)
1341         struct file *fp;
1342         u_long cmd;
1343         void *data;
1344         struct ucred *active_cred;
1345         struct thread *td;
1346 {
1347         struct pipe *mpipe = fp->f_data;
1348         int error;
1349
1350         PIPE_LOCK(mpipe);
1351
1352 #ifdef MAC
1353         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1354         if (error) {
1355                 PIPE_UNLOCK(mpipe);
1356                 return (error);
1357         }
1358 #endif
1359
1360         error = 0;
1361         switch (cmd) {
1362
1363         case FIONBIO:
1364                 break;
1365
1366         case FIOASYNC:
1367                 if (*(int *)data) {
1368                         mpipe->pipe_state |= PIPE_ASYNC;
1369                 } else {
1370                         mpipe->pipe_state &= ~PIPE_ASYNC;
1371                 }
1372                 break;
1373
1374         case FIONREAD:
1375                 if (!(fp->f_flag & FREAD)) {
1376                         *(int *)data = 0;
1377                         PIPE_UNLOCK(mpipe);
1378                         return (0);
1379                 }
1380                 if (mpipe->pipe_state & PIPE_DIRECTW)
1381                         *(int *)data = mpipe->pipe_map.cnt;
1382                 else
1383                         *(int *)data = mpipe->pipe_buffer.cnt;
1384                 break;
1385
1386         case FIOSETOWN:
1387                 PIPE_UNLOCK(mpipe);
1388                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1389                 goto out_unlocked;
1390
1391         case FIOGETOWN:
1392                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1393                 break;
1394
1395         /* This is deprecated, FIOSETOWN should be used instead. */
1396         case TIOCSPGRP:
1397                 PIPE_UNLOCK(mpipe);
1398                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1399                 goto out_unlocked;
1400
1401         /* This is deprecated, FIOGETOWN should be used instead. */
1402         case TIOCGPGRP:
1403                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1404                 break;
1405
1406         default:
1407                 error = ENOTTY;
1408                 break;
1409         }
1410         PIPE_UNLOCK(mpipe);
1411 out_unlocked:
1412         return (error);
1413 }
1414
1415 static int
1416 pipe_poll(fp, events, active_cred, td)
1417         struct file *fp;
1418         int events;
1419         struct ucred *active_cred;
1420         struct thread *td;
1421 {
1422         struct pipe *rpipe;
1423         struct pipe *wpipe;
1424         int levents, revents;
1425 #ifdef MAC
1426         int error;
1427 #endif
1428
1429         revents = 0;
1430         rpipe = fp->f_data;
1431         wpipe = PIPE_PEER(rpipe);
1432         PIPE_LOCK(rpipe);
1433 #ifdef MAC
1434         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1435         if (error)
1436                 goto locked_error;
1437 #endif
1438         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1439                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1440                     (rpipe->pipe_buffer.cnt > 0))
1441                         revents |= events & (POLLIN | POLLRDNORM);
1442
1443         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1444                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1445                     (wpipe->pipe_state & PIPE_EOF) ||
1446                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1447                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1448                          wpipe->pipe_buffer.size == 0)))
1449                         revents |= events & (POLLOUT | POLLWRNORM);
1450
1451         levents = events &
1452             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1453         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1454             fp->f_seqcount == rpipe->pipe_wgen)
1455                 events |= POLLINIGNEOF;
1456
1457         if ((events & POLLINIGNEOF) == 0) {
1458                 if (rpipe->pipe_state & PIPE_EOF) {
1459                         revents |= (events & (POLLIN | POLLRDNORM));
1460                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1461                             (wpipe->pipe_state & PIPE_EOF))
1462                                 revents |= POLLHUP;
1463                 }
1464         }
1465
1466         if (revents == 0) {
1467                 if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM)) {
1468                         selrecord(td, &rpipe->pipe_sel);
1469                         if (SEL_WAITING(&rpipe->pipe_sel))
1470                                 rpipe->pipe_state |= PIPE_SEL;
1471                 }
1472
1473                 if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM)) {
1474                         selrecord(td, &wpipe->pipe_sel);
1475                         if (SEL_WAITING(&wpipe->pipe_sel))
1476                                 wpipe->pipe_state |= PIPE_SEL;
1477                 }
1478         }
1479 #ifdef MAC
1480 locked_error:
1481 #endif
1482         PIPE_UNLOCK(rpipe);
1483
1484         return (revents);
1485 }
1486
1487 /*
1488  * We shouldn't need locks here as we're doing a read and this should
1489  * be a natural race.
1490  */
1491 static int
1492 pipe_stat(fp, ub, active_cred, td)
1493         struct file *fp;
1494         struct stat *ub;
1495         struct ucred *active_cred;
1496         struct thread *td;
1497 {
1498         struct pipe *pipe;
1499         int new_unr;
1500 #ifdef MAC
1501         int error;
1502 #endif
1503
1504         pipe = fp->f_data;
1505         PIPE_LOCK(pipe);
1506 #ifdef MAC
1507         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1508         if (error) {
1509                 PIPE_UNLOCK(pipe);
1510                 return (error);
1511         }
1512 #endif
1513
1514         /* For named pipes ask the underlying filesystem. */
1515         if (pipe->pipe_state & PIPE_NAMED) {
1516                 PIPE_UNLOCK(pipe);
1517                 return (vnops.fo_stat(fp, ub, active_cred, td));
1518         }
1519
1520         /*
1521          * Lazily allocate an inode number for the pipe.  Most pipe
1522          * users do not call fstat(2) on the pipe, which means that
1523          * postponing the inode allocation until it is must be
1524          * returned to userland is useful.  If alloc_unr failed,
1525          * assign st_ino zero instead of returning an error.
1526          * Special pipe_ino values:
1527          *  -1 - not yet initialized;
1528          *  0  - alloc_unr failed, return 0 as st_ino forever.
1529          */
1530         if (pipe->pipe_ino == (ino_t)-1) {
1531                 new_unr = alloc_unr(pipeino_unr);
1532                 if (new_unr != -1)
1533                         pipe->pipe_ino = new_unr;
1534                 else
1535                         pipe->pipe_ino = 0;
1536         }
1537         PIPE_UNLOCK(pipe);
1538
1539         bzero(ub, sizeof(*ub));
1540         ub->st_mode = S_IFIFO;
1541         ub->st_blksize = PAGE_SIZE;
1542         if (pipe->pipe_state & PIPE_DIRECTW)
1543                 ub->st_size = pipe->pipe_map.cnt;
1544         else
1545                 ub->st_size = pipe->pipe_buffer.cnt;
1546         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1547         ub->st_atim = pipe->pipe_atime;
1548         ub->st_mtim = pipe->pipe_mtime;
1549         ub->st_ctim = pipe->pipe_ctime;
1550         ub->st_uid = fp->f_cred->cr_uid;
1551         ub->st_gid = fp->f_cred->cr_gid;
1552         ub->st_dev = pipedev_ino;
1553         ub->st_ino = pipe->pipe_ino;
1554         /*
1555          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1556          */
1557         return (0);
1558 }
1559
1560 /* ARGSUSED */
1561 static int
1562 pipe_close(fp, td)
1563         struct file *fp;
1564         struct thread *td;
1565 {
1566
1567         if (fp->f_vnode != NULL)
1568                 return vnops.fo_close(fp, td);
1569         fp->f_ops = &badfileops;
1570         pipe_dtor(fp->f_data);
1571         fp->f_data = NULL;
1572         return (0);
1573 }
1574
1575 static int
1576 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1577 {
1578         struct pipe *cpipe;
1579         int error;
1580
1581         cpipe = fp->f_data;
1582         if (cpipe->pipe_state & PIPE_NAMED)
1583                 error = vn_chmod(fp, mode, active_cred, td);
1584         else
1585                 error = invfo_chmod(fp, mode, active_cred, td);
1586         return (error);
1587 }
1588
1589 static int
1590 pipe_chown(fp, uid, gid, active_cred, td)
1591         struct file *fp;
1592         uid_t uid;
1593         gid_t gid;
1594         struct ucred *active_cred;
1595         struct thread *td;
1596 {
1597         struct pipe *cpipe;
1598         int error;
1599
1600         cpipe = fp->f_data;
1601         if (cpipe->pipe_state & PIPE_NAMED)
1602                 error = vn_chown(fp, uid, gid, active_cred, td);
1603         else
1604                 error = invfo_chown(fp, uid, gid, active_cred, td);
1605         return (error);
1606 }
1607
1608 static void
1609 pipe_free_kmem(cpipe)
1610         struct pipe *cpipe;
1611 {
1612
1613         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1614             ("pipe_free_kmem: pipe mutex locked"));
1615
1616         if (cpipe->pipe_buffer.buffer != NULL) {
1617                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1618                 vm_map_remove(pipe_map,
1619                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1620                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1621                 cpipe->pipe_buffer.buffer = NULL;
1622         }
1623 #ifndef PIPE_NODIRECT
1624         {
1625                 cpipe->pipe_map.cnt = 0;
1626                 cpipe->pipe_map.pos = 0;
1627                 cpipe->pipe_map.npages = 0;
1628         }
1629 #endif
1630 }
1631
1632 /*
1633  * shutdown the pipe
1634  */
1635 static void
1636 pipeclose(cpipe)
1637         struct pipe *cpipe;
1638 {
1639         struct pipepair *pp;
1640         struct pipe *ppipe;
1641
1642         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1643
1644         PIPE_LOCK(cpipe);
1645         pipelock(cpipe, 0);
1646         pp = cpipe->pipe_pair;
1647
1648         pipeselwakeup(cpipe);
1649
1650         /*
1651          * If the other side is blocked, wake it up saying that
1652          * we want to close it down.
1653          */
1654         cpipe->pipe_state |= PIPE_EOF;
1655         while (cpipe->pipe_busy) {
1656                 wakeup(cpipe);
1657                 cpipe->pipe_state |= PIPE_WANT;
1658                 pipeunlock(cpipe);
1659                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1660                 pipelock(cpipe, 0);
1661         }
1662
1663
1664         /*
1665          * Disconnect from peer, if any.
1666          */
1667         ppipe = cpipe->pipe_peer;
1668         if (ppipe->pipe_present == PIPE_ACTIVE) {
1669                 pipeselwakeup(ppipe);
1670
1671                 ppipe->pipe_state |= PIPE_EOF;
1672                 wakeup(ppipe);
1673                 KNOTE_LOCKED(&ppipe->pipe_sel.si_note, 0);
1674         }
1675
1676         /*
1677          * Mark this endpoint as free.  Release kmem resources.  We
1678          * don't mark this endpoint as unused until we've finished
1679          * doing that, or the pipe might disappear out from under
1680          * us.
1681          */
1682         PIPE_UNLOCK(cpipe);
1683         pipe_free_kmem(cpipe);
1684         PIPE_LOCK(cpipe);
1685         cpipe->pipe_present = PIPE_CLOSING;
1686         pipeunlock(cpipe);
1687
1688         /*
1689          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1690          * PIPE_FINALIZED, that allows other end to free the
1691          * pipe_pair, only after the knotes are completely dismantled.
1692          */
1693         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1694         cpipe->pipe_present = PIPE_FINALIZED;
1695         seldrain(&cpipe->pipe_sel);
1696         knlist_destroy(&cpipe->pipe_sel.si_note);
1697
1698         /*
1699          * If both endpoints are now closed, release the memory for the
1700          * pipe pair.  If not, unlock.
1701          */
1702         if (ppipe->pipe_present == PIPE_FINALIZED) {
1703                 PIPE_UNLOCK(cpipe);
1704 #ifdef MAC
1705                 mac_pipe_destroy(pp);
1706 #endif
1707                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1708         } else
1709                 PIPE_UNLOCK(cpipe);
1710 }
1711
1712 /*ARGSUSED*/
1713 static int
1714 pipe_kqfilter(struct file *fp, struct knote *kn)
1715 {
1716         struct pipe *cpipe;
1717
1718         /*
1719          * If a filter is requested that is not supported by this file
1720          * descriptor, don't return an error, but also don't ever generate an
1721          * event.
1722          */
1723         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1724                 kn->kn_fop = &pipe_nfiltops;
1725                 return (0);
1726         }
1727         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1728                 kn->kn_fop = &pipe_nfiltops;
1729                 return (0);
1730         }
1731         cpipe = fp->f_data;
1732         PIPE_LOCK(cpipe);
1733         switch (kn->kn_filter) {
1734         case EVFILT_READ:
1735                 kn->kn_fop = &pipe_rfiltops;
1736                 break;
1737         case EVFILT_WRITE:
1738                 kn->kn_fop = &pipe_wfiltops;
1739                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1740                         /* other end of pipe has been closed */
1741                         PIPE_UNLOCK(cpipe);
1742                         return (EPIPE);
1743                 }
1744                 cpipe = PIPE_PEER(cpipe);
1745                 break;
1746         default:
1747                 PIPE_UNLOCK(cpipe);
1748                 return (EINVAL);
1749         }
1750
1751         kn->kn_hook = cpipe;
1752         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1753         PIPE_UNLOCK(cpipe);
1754         return (0);
1755 }
1756
1757 static void
1758 filt_pipedetach(struct knote *kn)
1759 {
1760         struct pipe *cpipe = kn->kn_hook;
1761
1762         PIPE_LOCK(cpipe);
1763         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1764         PIPE_UNLOCK(cpipe);
1765 }
1766
1767 /*ARGSUSED*/
1768 static int
1769 filt_piperead(struct knote *kn, long hint)
1770 {
1771         struct pipe *rpipe = kn->kn_hook;
1772         struct pipe *wpipe = rpipe->pipe_peer;
1773         int ret;
1774
1775         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1776         kn->kn_data = rpipe->pipe_buffer.cnt;
1777         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1778                 kn->kn_data = rpipe->pipe_map.cnt;
1779
1780         if ((rpipe->pipe_state & PIPE_EOF) ||
1781             wpipe->pipe_present != PIPE_ACTIVE ||
1782             (wpipe->pipe_state & PIPE_EOF)) {
1783                 kn->kn_flags |= EV_EOF;
1784                 return (1);
1785         }
1786         ret = kn->kn_data > 0;
1787         return ret;
1788 }
1789
1790 /*ARGSUSED*/
1791 static int
1792 filt_pipewrite(struct knote *kn, long hint)
1793 {
1794         struct pipe *wpipe;
1795
1796         wpipe = kn->kn_hook;
1797         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1798         if (wpipe->pipe_present != PIPE_ACTIVE ||
1799             (wpipe->pipe_state & PIPE_EOF)) {
1800                 kn->kn_data = 0;
1801                 kn->kn_flags |= EV_EOF;
1802                 return (1);
1803         }
1804         kn->kn_data = (wpipe->pipe_buffer.size > 0) ?
1805             (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) : PIPE_BUF;
1806         if (wpipe->pipe_state & PIPE_DIRECTW)
1807                 kn->kn_data = 0;
1808
1809         return (kn->kn_data >= PIPE_BUF);
1810 }
1811
1812 static void
1813 filt_pipedetach_notsup(struct knote *kn)
1814 {
1815
1816 }
1817
1818 static int
1819 filt_pipenotsup(struct knote *kn, long hint)
1820 {
1821
1822         return (0);
1823 }