sys/kern/sys_pipe.c

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