sys/kern/sys_pipe.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-4-Clause
   3  *
   4  * Copyright (c) 1996 John S. Dyson
   5  * Copyright (c) 2012 Giovanni Trematerra
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice immediately at the beginning of the file, without modification,
  13  *    this list of conditions, and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Absolutely no warranty of function or purpose is made by the author
  18  *    John S. Dyson.
  19  * 4. Modifications may be freely made to this file if the above conditions
  20  *    are met.
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, the sending process pins the underlying pages in
  36  * memory, and the receiving process copies directly from these pinned pages
  37  * in the sending process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.
  50  *
  51  * In order to limit the resource use of pipes, two sysctls exist:
  52  *
  53  * kern.ipc.maxpipekva - This is a hard limit on the amount of pageable
  54  * address space available to us in pipe_map. This value is normally
  55  * autotuned, but may also be loader tuned.
  56  *
  57  * kern.ipc.pipekva - This read-only sysctl tracks the current amount of
  58  * memory in use by pipes.
  59  *
  60  * Based on how large pipekva is relative to maxpipekva, the following
  61  * will happen:
  62  *
  63  * 0% - 50%:
  64  *     New pipes are given 16K of memory backing, pipes may dynamically
  65  *     grow to as large as 64K where needed.
  66  * 50% - 75%:
  67  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  68  *     existing pipes may NOT grow.
  69  * 75% - 100%:
  70  *     New pipes are given 4K (or PAGE_SIZE) of memory backing,
  71  *     existing pipes will be shrunk down to 4K whenever possible.
  72  *
  73  * Resizing may be disabled by setting kern.ipc.piperesizeallowed=0.  If
  74  * that is set,  the only resize that will occur is the 0 -> SMALL_PIPE_SIZE
  75  * resize which MUST occur for reverse-direction pipes when they are
  76  * first used.
  77  *
  78  * Additional information about the current state of pipes may be obtained
  79  * from kern.ipc.pipes, kern.ipc.pipefragretry, kern.ipc.pipeallocfail,
  80  * and kern.ipc.piperesizefail.
  81  *
  82  * Locking rules:  There are two locks present here:  A mutex, used via
  83  * PIPE_LOCK, and a flag, used via pipelock().  All locking is done via
  84  * the flag, as mutexes can not persist over uiomove.  The mutex
  85  * exists only to guard access to the flag, and is not in itself a
  86  * locking mechanism.  Also note that there is only a single mutex for
  87  * both directions of a pipe.
  88  *
  89  * As pipelock() may have to sleep before it can acquire the flag, it
  90  * is important to reread all data after a call to pipelock(); everything
  91  * in the structure may have changed.
  92  */
  93
  94 #include <sys/cdefs.h>
  95 __FBSDID("$FreeBSD$");
  96
  97 #include <sys/param.h>
  98 #include <sys/systm.h>
  99 #include <sys/conf.h>
 100 #include <sys/fcntl.h>
 101 #include <sys/file.h>
 102 #include <sys/filedesc.h>
 103 #include <sys/filio.h>
 104 #include <sys/kernel.h>
 105 #include <sys/lock.h>
 106 #include <sys/mutex.h>
 107 #include <sys/ttycom.h>
 108 #include <sys/stat.h>
 109 #include <sys/malloc.h>
 110 #include <sys/poll.h>
 111 #include <sys/selinfo.h>
 112 #include <sys/signalvar.h>
 113 #include <sys/syscallsubr.h>
 114 #include <sys/sysctl.h>
 115 #include <sys/sysproto.h>
 116 #include <sys/pipe.h>
 117 #include <sys/proc.h>
 118 #include <sys/vnode.h>
 119 #include <sys/uio.h>
 120 #include <sys/user.h>
 121 #include <sys/event.h>
 122
 123 #include <security/mac/mac_framework.h>
 124
 125 #include <vm/vm.h>
 126 #include <vm/vm_param.h>
 127 #include <vm/vm_object.h>
 128 #include <vm/vm_kern.h>
 129 #include <vm/vm_extern.h>
 130 #include <vm/pmap.h>
 131 #include <vm/vm_map.h>
 132 #include <vm/vm_page.h>
 133 #include <vm/uma.h>
 134
 135 /*
 136  * Use this define if you want to disable *fancy* VM things.  Expect an
 137  * approx 30% decrease in transfer rate.  This could be useful for
 138  * NetBSD or OpenBSD.
 139  */
 140 /* #define PIPE_NODIRECT */
 141
 142 #define PIPE_PEER(pipe) \
 143         (((pipe)->pipe_state & PIPE_NAMED) ? (pipe) : ((pipe)->pipe_peer))
 144
 145 /*
 146  * interfaces to the outside world
 147  */
 148 static fo_rdwr_t        pipe_read;
 149 static fo_rdwr_t        pipe_write;
 150 static fo_truncate_t    pipe_truncate;
 151 static fo_ioctl_t       pipe_ioctl;
 152 static fo_poll_t        pipe_poll;
 153 static fo_kqfilter_t    pipe_kqfilter;
 154 static fo_stat_t        pipe_stat;
 155 static fo_close_t       pipe_close;
 156 static fo_chmod_t       pipe_chmod;
 157 static fo_chown_t       pipe_chown;
 158 static fo_fill_kinfo_t  pipe_fill_kinfo;
 159
 160 struct fileops pipeops = {
 161         .fo_read = pipe_read,
 162         .fo_write = pipe_write,
 163         .fo_truncate = pipe_truncate,
 164         .fo_ioctl = pipe_ioctl,
 165         .fo_poll = pipe_poll,
 166         .fo_kqfilter = pipe_kqfilter,
 167         .fo_stat = pipe_stat,
 168         .fo_close = pipe_close,
 169         .fo_chmod = pipe_chmod,
 170         .fo_chown = pipe_chown,
 171         .fo_sendfile = invfo_sendfile,
 172         .fo_fill_kinfo = pipe_fill_kinfo,
 173         .fo_flags = DFLAG_PASSABLE
 174 };
 175
 176 static void     filt_pipedetach(struct knote *kn);
 177 static void     filt_pipedetach_notsup(struct knote *kn);
 178 static int      filt_pipenotsup(struct knote *kn, long hint);
 179 static int      filt_piperead(struct knote *kn, long hint);
 180 static int      filt_pipewrite(struct knote *kn, long hint);
 181
 182 static struct filterops pipe_nfiltops = {
 183         .f_isfd = 1,
 184         .f_detach = filt_pipedetach_notsup,
 185         .f_event = filt_pipenotsup
 186 };
 187 static struct filterops pipe_rfiltops = {
 188         .f_isfd = 1,
 189         .f_detach = filt_pipedetach,
 190         .f_event = filt_piperead
 191 };
 192 static struct filterops pipe_wfiltops = {
 193         .f_isfd = 1,
 194         .f_detach = filt_pipedetach,
 195         .f_event = filt_pipewrite
 196 };
 197
 198 /*
 199  * Default pipe buffer size(s), this can be kind-of large now because pipe
 200  * space is pageable.  The pipe code will try to maintain locality of
 201  * reference for performance reasons, so small amounts of outstanding I/O
 202  * will not wipe the cache.
 203  */
 204 #define MINPIPESIZE (PIPE_SIZE/3)
 205 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 206
 207 static long amountpipekva;
 208 static int pipefragretry;
 209 static int pipeallocfail;
 210 static int piperesizefail;
 211 static int piperesizeallowed = 1;
 212
 213 SYSCTL_LONG(_kern_ipc, OID_AUTO, maxpipekva, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
 214            &maxpipekva, 0, "Pipe KVA limit");
 215 SYSCTL_LONG(_kern_ipc, OID_AUTO, pipekva, CTLFLAG_RD,
 216            &amountpipekva, 0, "Pipe KVA usage");
 217 SYSCTL_INT(_kern_ipc, OID_AUTO, pipefragretry, CTLFLAG_RD,
 218           &pipefragretry, 0, "Pipe allocation retries due to fragmentation");
 219 SYSCTL_INT(_kern_ipc, OID_AUTO, pipeallocfail, CTLFLAG_RD,
 220           &pipeallocfail, 0, "Pipe allocation failures");
 221 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizefail, CTLFLAG_RD,
 222           &piperesizefail, 0, "Pipe resize failures");
 223 SYSCTL_INT(_kern_ipc, OID_AUTO, piperesizeallowed, CTLFLAG_RW,
 224           &piperesizeallowed, 0, "Pipe resizing allowed");
 225
 226 static void pipeinit(void *dummy __unused);
 227 static void pipeclose(struct pipe *cpipe);
 228 static void pipe_free_kmem(struct pipe *cpipe);
 229 static int pipe_create(struct pipe *pipe, bool backing);
 230 static int pipe_paircreate(struct thread *td, struct pipepair **p_pp);
 231 static __inline int pipelock(struct pipe *cpipe, int catch);
 232 static __inline void pipeunlock(struct pipe *cpipe);
 233 #ifndef PIPE_NODIRECT
 234 static int pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio);
 235 static void pipe_destroy_write_buffer(struct pipe *wpipe);
 236 static int pipe_direct_write(struct pipe *wpipe, struct uio *uio);
 237 static void pipe_clone_write_buffer(struct pipe *wpipe);
 238 #endif
 239 static int pipespace(struct pipe *cpipe, int size);
 240 static int pipespace_new(struct pipe *cpipe, int size);
 241
 242 static int      pipe_zone_ctor(void *mem, int size, void *arg, int flags);
 243 static int      pipe_zone_init(void *mem, int size, int flags);
 244 static void     pipe_zone_fini(void *mem, int size);
 245
 246 static uma_zone_t pipe_zone;
 247 static struct unrhdr64 pipeino_unr;
 248 static dev_t pipedev_ino;
 249
 250 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, pipeinit, NULL);
 251
 252 static void
 253 pipeinit(void *dummy __unused)
 254 {
 255
 256         pipe_zone = uma_zcreate("pipe", sizeof(struct pipepair),
 257             pipe_zone_ctor, NULL, pipe_zone_init, pipe_zone_fini,
 258             UMA_ALIGN_PTR, 0);
 259         KASSERT(pipe_zone != NULL, ("pipe_zone not initialized"));
 260         new_unrhdr64(&pipeino_unr, 1);
 261         pipedev_ino = devfs_alloc_cdp_inode();
 262         KASSERT(pipedev_ino > 0, ("pipe dev inode not initialized"));
 263 }
 264
 265 static int
 266 pipe_zone_ctor(void *mem, int size, void *arg, int flags)
 267 {
 268         struct pipepair *pp;
 269         struct pipe *rpipe, *wpipe;
 270
 271         KASSERT(size == sizeof(*pp), ("pipe_zone_ctor: wrong size"));
 272
 273         pp = (struct pipepair *)mem;
 274
 275         /*
 276          * We zero both pipe endpoints to make sure all the kmem pointers
 277          * are NULL, flag fields are zero'd, etc.  We timestamp both
 278          * endpoints with the same time.
 279          */
 280         rpipe = &pp->pp_rpipe;
 281         bzero(rpipe, sizeof(*rpipe));
 282         vfs_timestamp(&rpipe->pipe_ctime);
 283         rpipe->pipe_atime = rpipe->pipe_mtime = rpipe->pipe_ctime;
 284
 285         wpipe = &pp->pp_wpipe;
 286         bzero(wpipe, sizeof(*wpipe));
 287         wpipe->pipe_ctime = rpipe->pipe_ctime;
 288         wpipe->pipe_atime = wpipe->pipe_mtime = rpipe->pipe_ctime;
 289
 290         rpipe->pipe_peer = wpipe;
 291         rpipe->pipe_pair = pp;
 292         wpipe->pipe_peer = rpipe;
 293         wpipe->pipe_pair = pp;
 294
 295         /*
 296          * Mark both endpoints as present; they will later get free'd
 297          * one at a time.  When both are free'd, then the whole pair
 298          * is released.
 299          */
 300         rpipe->pipe_present = PIPE_ACTIVE;
 301         wpipe->pipe_present = PIPE_ACTIVE;
 302
 303         /*
 304          * Eventually, the MAC Framework may initialize the label
 305          * in ctor or init, but for now we do it elswhere to avoid
 306          * blocking in ctor or init.
 307          */
 308         pp->pp_label = NULL;
 309
 310         return (0);
 311 }
 312
 313 static int
 314 pipe_zone_init(void *mem, int size, int flags)
 315 {
 316         struct pipepair *pp;
 317
 318         KASSERT(size == sizeof(*pp), ("pipe_zone_init: wrong size"));
 319
 320         pp = (struct pipepair *)mem;
 321
 322         mtx_init(&pp->pp_mtx, "pipe mutex", NULL, MTX_DEF | MTX_NEW);
 323         return (0);
 324 }
 325
 326 static void
 327 pipe_zone_fini(void *mem, int size)
 328 {
 329         struct pipepair *pp;
 330
 331         KASSERT(size == sizeof(*pp), ("pipe_zone_fini: wrong size"));
 332
 333         pp = (struct pipepair *)mem;
 334
 335         mtx_destroy(&pp->pp_mtx);
 336 }
 337
 338 static int
 339 pipe_paircreate(struct thread *td, struct pipepair **p_pp)
 340 {
 341         struct pipepair *pp;
 342         struct pipe *rpipe, *wpipe;
 343         int error;
 344
 345         *p_pp = pp = uma_zalloc(pipe_zone, M_WAITOK);
 346 #ifdef MAC
 347         /*
 348          * The MAC label is shared between the connected endpoints.  As a
 349          * result mac_pipe_init() and mac_pipe_create() are called once
 350          * for the pair, and not on the endpoints.
 351          */
 352         mac_pipe_init(pp);
 353         mac_pipe_create(td->td_ucred, pp);
 354 #endif
 355         rpipe = &pp->pp_rpipe;
 356         wpipe = &pp->pp_wpipe;
 357
 358         knlist_init_mtx(&rpipe->pipe_sel.si_note, PIPE_MTX(rpipe));
 359         knlist_init_mtx(&wpipe->pipe_sel.si_note, PIPE_MTX(wpipe));
 360
 361         /*
 362          * Only the forward direction pipe is backed by big buffer by
 363          * default.
 364          */
 365         error = pipe_create(rpipe, true);
 366         if (error != 0)
 367                 goto fail;
 368         error = pipe_create(wpipe, false);
 369         if (error != 0) {
 370                 /*
 371                  * This cleanup leaves the pipe inode number for rpipe
 372                  * still allocated, but never used.  We do not free
 373                  * inode numbers for opened pipes, which is required
 374                  * for correctness because numbers must be unique.
 375                  * But also it avoids any memory use by the unr
 376                  * allocator, so stashing away the transient inode
 377                  * number is reasonable.
 378                  */
 379                 pipe_free_kmem(rpipe);
 380                 goto fail;
 381         }
 382
 383         rpipe->pipe_state |= PIPE_DIRECTOK;
 384         wpipe->pipe_state |= PIPE_DIRECTOK;
 385         return (0);
 386
 387 fail:
 388         knlist_destroy(&rpipe->pipe_sel.si_note);
 389         knlist_destroy(&wpipe->pipe_sel.si_note);
 390 #ifdef MAC
 391         mac_pipe_destroy(pp);
 392 #endif
 393         return (error);
 394 }
 395
 396 int
 397 pipe_named_ctor(struct pipe **ppipe, struct thread *td)
 398 {
 399         struct pipepair *pp;
 400         int error;
 401
 402         error = pipe_paircreate(td, &pp);
 403         if (error != 0)
 404                 return (error);
 405         pp->pp_rpipe.pipe_state |= PIPE_NAMED;
 406         *ppipe = &pp->pp_rpipe;
 407         return (0);
 408 }
 409
 410 void
 411 pipe_dtor(struct pipe *dpipe)
 412 {
 413         struct pipe *peer;
 414         ino_t ino;
 415
 416         ino = dpipe->pipe_ino;
 417         peer = (dpipe->pipe_state & PIPE_NAMED) != 0 ? dpipe->pipe_peer : NULL;
 418         funsetown(&dpipe->pipe_sigio);
 419         pipeclose(dpipe);
 420         if (peer != NULL) {
 421                 funsetown(&peer->pipe_sigio);
 422                 pipeclose(peer);
 423         }
 424 }
 425
 426 /*
 427  * The pipe system call for the DTYPE_PIPE type of pipes.  If we fail, let
 428  * the zone pick up the pieces via pipeclose().
 429  */
 430 int
 431 kern_pipe(struct thread *td, int fildes[2], int flags, struct filecaps *fcaps1,
 432     struct filecaps *fcaps2)
 433 {
 434         struct file *rf, *wf;
 435         struct pipe *rpipe, *wpipe;
 436         struct pipepair *pp;
 437         int fd, fflags, error;
 438
 439         error = pipe_paircreate(td, &pp);
 440         if (error != 0)
 441                 return (error);
 442         rpipe = &pp->pp_rpipe;
 443         wpipe = &pp->pp_wpipe;
 444         error = falloc_caps(td, &rf, &fd, flags, fcaps1);
 445         if (error) {
 446                 pipeclose(rpipe);
 447                 pipeclose(wpipe);
 448                 return (error);
 449         }
 450         /* An extra reference on `rf' has been held for us by falloc_caps(). */
 451         fildes[0] = fd;
 452
 453         fflags = FREAD | FWRITE;
 454         if ((flags & O_NONBLOCK) != 0)
 455                 fflags |= FNONBLOCK;
 456
 457         /*
 458          * Warning: once we've gotten past allocation of the fd for the
 459          * read-side, we can only drop the read side via fdrop() in order
 460          * to avoid races against processes which manage to dup() the read
 461          * side while we are blocked trying to allocate the write side.
 462          */
 463         finit(rf, fflags, DTYPE_PIPE, rpipe, &pipeops);
 464         error = falloc_caps(td, &wf, &fd, flags, fcaps2);
 465         if (error) {
 466                 fdclose(td, rf, fildes[0]);
 467                 fdrop(rf, td);
 468                 /* rpipe has been closed by fdrop(). */
 469                 pipeclose(wpipe);
 470                 return (error);
 471         }
 472         /* An extra reference on `wf' has been held for us by falloc_caps(). */
 473         finit(wf, fflags, DTYPE_PIPE, wpipe, &pipeops);
 474         fdrop(wf, td);
 475         fildes[1] = fd;
 476         fdrop(rf, td);
 477
 478         return (0);
 479 }
 480
 481 #ifdef COMPAT_FREEBSD10
 482 /* ARGSUSED */
 483 int
 484 freebsd10_pipe(struct thread *td, struct freebsd10_pipe_args *uap __unused)
 485 {
 486         int error;
 487         int fildes[2];
 488
 489         error = kern_pipe(td, fildes, 0, NULL, NULL);
 490         if (error)
 491                 return (error);
 492
 493         td->td_retval[0] = fildes[0];
 494         td->td_retval[1] = fildes[1];
 495
 496         return (0);
 497 }
 498 #endif
 499
 500 int
 501 sys_pipe2(struct thread *td, struct pipe2_args *uap)
 502 {
 503         int error, fildes[2];
 504
 505         if (uap->flags & ~(O_CLOEXEC | O_NONBLOCK))
 506                 return (EINVAL);
 507         error = kern_pipe(td, fildes, uap->flags, NULL, NULL);
 508         if (error)
 509                 return (error);
 510         error = copyout(fildes, uap->fildes, 2 * sizeof(int));
 511         if (error) {
 512                 (void)kern_close(td, fildes[0]);
 513                 (void)kern_close(td, fildes[1]);
 514         }
 515         return (error);
 516 }
 517
 518 /*
 519  * Allocate kva for pipe circular buffer, the space is pageable
 520  * This routine will 'realloc' the size of a pipe safely, if it fails
 521  * it will retain the old buffer.
 522  * If it fails it will return ENOMEM.
 523  */
 524 static int
 525 pipespace_new(struct pipe *cpipe, int size)
 526 {
 527         caddr_t buffer;
 528         int error, cnt, firstseg;
 529         static int curfail = 0;
 530         static struct timeval lastfail;
 531
 532         KASSERT(!mtx_owned(PIPE_MTX(cpipe)), ("pipespace: pipe mutex locked"));
 533         KASSERT(!(cpipe->pipe_state & PIPE_DIRECTW),
 534                 ("pipespace: resize of direct writes not allowed"));
 535 retry:
 536         cnt = cpipe->pipe_buffer.cnt;
 537         if (cnt > size)
 538                 size = cnt;
 539
 540         size = round_page(size);
 541         buffer = (caddr_t) vm_map_min(pipe_map);
 542
 543         error = vm_map_find(pipe_map, NULL, 0, (vm_offset_t *)&buffer, size, 0,
 544             VMFS_ANY_SPACE, VM_PROT_RW, VM_PROT_RW, 0);
 545         if (error != KERN_SUCCESS) {
 546                 if (cpipe->pipe_buffer.buffer == NULL &&
 547                     size > SMALL_PIPE_SIZE) {
 548                         size = SMALL_PIPE_SIZE;
 549                         pipefragretry++;
 550                         goto retry;
 551                 }
 552                 if (cpipe->pipe_buffer.buffer == NULL) {
 553                         pipeallocfail++;
 554                         if (ppsratecheck(&lastfail, &curfail, 1))
 555                                 printf("kern.ipc.maxpipekva exceeded; see tuning(7)\n");
 556                 } else {
 557                         piperesizefail++;
 558                 }
 559                 return (ENOMEM);
 560         }
 561
 562         /* copy data, then free old resources if we're resizing */
 563         if (cnt > 0) {
 564                 if (cpipe->pipe_buffer.in <= cpipe->pipe_buffer.out) {
 565                         firstseg = cpipe->pipe_buffer.size - cpipe->pipe_buffer.out;
 566                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 567                                 buffer, firstseg);
 568                         if ((cnt - firstseg) > 0)
 569                                 bcopy(cpipe->pipe_buffer.buffer, &buffer[firstseg],
 570                                         cpipe->pipe_buffer.in);
 571                 } else {
 572                         bcopy(&cpipe->pipe_buffer.buffer[cpipe->pipe_buffer.out],
 573                                 buffer, cnt);
 574                 }
 575         }
 576         pipe_free_kmem(cpipe);
 577         cpipe->pipe_buffer.buffer = buffer;
 578         cpipe->pipe_buffer.size = size;
 579         cpipe->pipe_buffer.in = cnt;
 580         cpipe->pipe_buffer.out = 0;
 581         cpipe->pipe_buffer.cnt = cnt;
 582         atomic_add_long(&amountpipekva, cpipe->pipe_buffer.size);
 583         return (0);
 584 }
 585
 586 /*
 587  * Wrapper for pipespace_new() that performs locking assertions.
 588  */
 589 static int
 590 pipespace(struct pipe *cpipe, int size)
 591 {
 592
 593         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 594             ("Unlocked pipe passed to pipespace"));
 595         return (pipespace_new(cpipe, size));
 596 }
 597
 598 /*
 599  * lock a pipe for I/O, blocking other access
 600  */
 601 static __inline int
 602 pipelock(struct pipe *cpipe, int catch)
 603 {
 604         int error;
 605
 606         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 607         while (cpipe->pipe_state & PIPE_LOCKFL) {
 608                 cpipe->pipe_state |= PIPE_LWANT;
 609                 error = msleep(cpipe, PIPE_MTX(cpipe),
 610                     catch ? (PRIBIO | PCATCH) : PRIBIO,
 611                     "pipelk", 0);
 612                 if (error != 0)
 613                         return (error);
 614         }
 615         cpipe->pipe_state |= PIPE_LOCKFL;
 616         return (0);
 617 }
 618
 619 /*
 620  * unlock a pipe I/O lock
 621  */
 622 static __inline void
 623 pipeunlock(struct pipe *cpipe)
 624 {
 625
 626         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 627         KASSERT(cpipe->pipe_state & PIPE_LOCKFL,
 628                 ("Unlocked pipe passed to pipeunlock"));
 629         cpipe->pipe_state &= ~PIPE_LOCKFL;
 630         if (cpipe->pipe_state & PIPE_LWANT) {
 631                 cpipe->pipe_state &= ~PIPE_LWANT;
 632                 wakeup(cpipe);
 633         }
 634 }
 635
 636 void
 637 pipeselwakeup(struct pipe *cpipe)
 638 {
 639
 640         PIPE_LOCK_ASSERT(cpipe, MA_OWNED);
 641         if (cpipe->pipe_state & PIPE_SEL) {
 642                 selwakeuppri(&cpipe->pipe_sel, PSOCK);
 643                 if (!SEL_WAITING(&cpipe->pipe_sel))
 644                         cpipe->pipe_state &= ~PIPE_SEL;
 645         }
 646         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 647                 pgsigio(&cpipe->pipe_sigio, SIGIO, 0);
 648         KNOTE_LOCKED(&cpipe->pipe_sel.si_note, 0);
 649 }
 650
 651 /*
 652  * Initialize and allocate VM and memory for pipe.  The structure
 653  * will start out zero'd from the ctor, so we just manage the kmem.
 654  */
 655 static int
 656 pipe_create(struct pipe *pipe, bool large_backing)
 657 {
 658         int error;
 659
 660         error = pipespace_new(pipe, !large_backing || amountpipekva >
 661             maxpipekva / 2 ? SMALL_PIPE_SIZE : PIPE_SIZE);
 662         if (error == 0)
 663                 pipe->pipe_ino = alloc_unr64(&pipeino_unr);
 664         return (error);
 665 }
 666
 667 /* ARGSUSED */
 668 static int
 669 pipe_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
 670     int flags, struct thread *td)
 671 {
 672         struct pipe *rpipe;
 673         int error;
 674         int nread = 0;
 675         int size;
 676
 677         rpipe = fp->f_data;
 678         PIPE_LOCK(rpipe);
 679         ++rpipe->pipe_busy;
 680         error = pipelock(rpipe, 1);
 681         if (error)
 682                 goto unlocked_error;
 683
 684 #ifdef MAC
 685         error = mac_pipe_check_read(active_cred, rpipe->pipe_pair);
 686         if (error)
 687                 goto locked_error;
 688 #endif
 689         if (amountpipekva > (3 * maxpipekva) / 4) {
 690                 if ((rpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 691                     rpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
 692                     rpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
 693                     piperesizeallowed == 1) {
 694                         PIPE_UNLOCK(rpipe);
 695                         pipespace(rpipe, SMALL_PIPE_SIZE);
 696                         PIPE_LOCK(rpipe);
 697                 }
 698         }
 699
 700         while (uio->uio_resid) {
 701                 /*
 702                  * normal pipe buffer receive
 703                  */
 704                 if (rpipe->pipe_buffer.cnt > 0) {
 705                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 706                         if (size > rpipe->pipe_buffer.cnt)
 707                                 size = rpipe->pipe_buffer.cnt;
 708                         if (size > uio->uio_resid)
 709                                 size = uio->uio_resid;
 710
 711                         PIPE_UNLOCK(rpipe);
 712                         error = uiomove(
 713                             &rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 714                             size, uio);
 715                         PIPE_LOCK(rpipe);
 716                         if (error)
 717                                 break;
 718
 719                         rpipe->pipe_buffer.out += size;
 720                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 721                                 rpipe->pipe_buffer.out = 0;
 722
 723                         rpipe->pipe_buffer.cnt -= size;
 724
 725                         /*
 726                          * If there is no more to read in the pipe, reset
 727                          * its pointers to the beginning.  This improves
 728                          * cache hit stats.
 729                          */
 730                         if (rpipe->pipe_buffer.cnt == 0) {
 731                                 rpipe->pipe_buffer.in = 0;
 732                                 rpipe->pipe_buffer.out = 0;
 733                         }
 734                         nread += size;
 735 #ifndef PIPE_NODIRECT
 736                 /*
 737                  * Direct copy, bypassing a kernel buffer.
 738                  */
 739                 } else if ((size = rpipe->pipe_pages.cnt) != 0) {
 740                         if (size > uio->uio_resid)
 741                                 size = (u_int) uio->uio_resid;
 742                         PIPE_UNLOCK(rpipe);
 743                         error = uiomove_fromphys(rpipe->pipe_pages.ms,
 744                             rpipe->pipe_pages.pos, size, uio);
 745                         PIPE_LOCK(rpipe);
 746                         if (error)
 747                                 break;
 748                         nread += size;
 749                         rpipe->pipe_pages.pos += size;
 750                         rpipe->pipe_pages.cnt -= size;
 751                         if (rpipe->pipe_pages.cnt == 0) {
 752                                 rpipe->pipe_state &= ~PIPE_WANTW;
 753                                 wakeup(rpipe);
 754                         }
 755 #endif
 756                 } else {
 757                         /*
 758                          * detect EOF condition
 759                          * read returns 0 on EOF, no need to set error
 760                          */
 761                         if (rpipe->pipe_state & PIPE_EOF)
 762                                 break;
 763
 764                         /*
 765                          * If the "write-side" has been blocked, wake it up now.
 766                          */
 767                         if (rpipe->pipe_state & PIPE_WANTW) {
 768                                 rpipe->pipe_state &= ~PIPE_WANTW;
 769                                 wakeup(rpipe);
 770                         }
 771
 772                         /*
 773                          * Break if some data was read.
 774                          */
 775                         if (nread > 0)
 776                                 break;
 777
 778                         /*
 779                          * Unlock the pipe buffer for our remaining processing.
 780                          * We will either break out with an error or we will
 781                          * sleep and relock to loop.
 782                          */
 783                         pipeunlock(rpipe);
 784
 785                         /*
 786                          * Handle non-blocking mode operation or
 787                          * wait for more data.
 788                          */
 789                         if (fp->f_flag & FNONBLOCK) {
 790                                 error = EAGAIN;
 791                         } else {
 792                                 rpipe->pipe_state |= PIPE_WANTR;
 793                                 if ((error = msleep(rpipe, PIPE_MTX(rpipe),
 794                                     PRIBIO | PCATCH,
 795                                     "piperd", 0)) == 0)
 796                                         error = pipelock(rpipe, 1);
 797                         }
 798                         if (error)
 799                                 goto unlocked_error;
 800                 }
 801         }
 802 #ifdef MAC
 803 locked_error:
 804 #endif
 805         pipeunlock(rpipe);
 806
 807         /* XXX: should probably do this before getting any locks. */
 808         if (error == 0)
 809                 vfs_timestamp(&rpipe->pipe_atime);
 810 unlocked_error:
 811         --rpipe->pipe_busy;
 812
 813         /*
 814          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 815          */
 816         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 817                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 818                 wakeup(rpipe);
 819         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 820                 /*
 821                  * Handle write blocking hysteresis.
 822                  */
 823                 if (rpipe->pipe_state & PIPE_WANTW) {
 824                         rpipe->pipe_state &= ~PIPE_WANTW;
 825                         wakeup(rpipe);
 826                 }
 827         }
 828
 829         /*
 830          * Only wake up writers if there was actually something read.
 831          * Otherwise, when calling read(2) at EOF, a spurious wakeup occurs.
 832          */
 833         if (nread > 0 &&
 834             rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt >= PIPE_BUF)
 835                 pipeselwakeup(rpipe);
 836
 837         PIPE_UNLOCK(rpipe);
 838         return (error);
 839 }
 840
 841 #ifndef PIPE_NODIRECT
 842 /*
 843  * Map the sending processes' buffer into kernel space and wire it.
 844  * This is similar to a physical write operation.
 845  */
 846 static int
 847 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 848 {
 849         u_int size;
 850         int i;
 851
 852         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 853         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
 854             ("%s: PIPE_DIRECTW set on %p", __func__, wpipe));
 855         KASSERT(wpipe->pipe_pages.cnt == 0,
 856             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 857
 858         if (uio->uio_iov->iov_len > wpipe->pipe_buffer.size)
 859                 size = wpipe->pipe_buffer.size;
 860         else
 861                 size = uio->uio_iov->iov_len;
 862
 863         wpipe->pipe_state |= PIPE_DIRECTW;
 864         PIPE_UNLOCK(wpipe);
 865         i = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
 866             (vm_offset_t)uio->uio_iov->iov_base, size, VM_PROT_READ,
 867             wpipe->pipe_pages.ms, PIPENPAGES);
 868         PIPE_LOCK(wpipe);
 869         if (i < 0) {
 870                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 871                 return (EFAULT);
 872         }
 873
 874         wpipe->pipe_pages.npages = i;
 875         wpipe->pipe_pages.pos =
 876             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 877         wpipe->pipe_pages.cnt = size;
 878
 879         uio->uio_iov->iov_len -= size;
 880         uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + size;
 881         if (uio->uio_iov->iov_len == 0)
 882                 uio->uio_iov++;
 883         uio->uio_resid -= size;
 884         uio->uio_offset += size;
 885         return (0);
 886 }
 887
 888 /*
 889  * Unwire the process buffer.
 890  */
 891 static void
 892 pipe_destroy_write_buffer(struct pipe *wpipe)
 893 {
 894
 895         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 896         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 897             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 898         KASSERT(wpipe->pipe_pages.cnt == 0,
 899             ("%s: pipe map for %p contains residual data", __func__, wpipe));
 900
 901         wpipe->pipe_state &= ~PIPE_DIRECTW;
 902         vm_page_unhold_pages(wpipe->pipe_pages.ms, wpipe->pipe_pages.npages);
 903         wpipe->pipe_pages.npages = 0;
 904 }
 905
 906 /*
 907  * In the case of a signal, the writing process might go away.  This
 908  * code copies the data into the circular buffer so that the source
 909  * pages can be freed without loss of data.
 910  */
 911 static void
 912 pipe_clone_write_buffer(struct pipe *wpipe)
 913 {
 914         struct uio uio;
 915         struct iovec iov;
 916         int size;
 917         int pos;
 918
 919         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 920         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) != 0,
 921             ("%s: PIPE_DIRECTW not set on %p", __func__, wpipe));
 922
 923         size = wpipe->pipe_pages.cnt;
 924         pos = wpipe->pipe_pages.pos;
 925         wpipe->pipe_pages.cnt = 0;
 926
 927         wpipe->pipe_buffer.in = size;
 928         wpipe->pipe_buffer.out = 0;
 929         wpipe->pipe_buffer.cnt = size;
 930
 931         PIPE_UNLOCK(wpipe);
 932         iov.iov_base = wpipe->pipe_buffer.buffer;
 933         iov.iov_len = size;
 934         uio.uio_iov = &iov;
 935         uio.uio_iovcnt = 1;
 936         uio.uio_offset = 0;
 937         uio.uio_resid = size;
 938         uio.uio_segflg = UIO_SYSSPACE;
 939         uio.uio_rw = UIO_READ;
 940         uio.uio_td = curthread;
 941         uiomove_fromphys(wpipe->pipe_pages.ms, pos, size, &uio);
 942         PIPE_LOCK(wpipe);
 943         pipe_destroy_write_buffer(wpipe);
 944 }
 945
 946 /*
 947  * This implements the pipe buffer write mechanism.  Note that only
 948  * a direct write OR a normal pipe write can be pending at any given time.
 949  * If there are any characters in the pipe buffer, the direct write will
 950  * be deferred until the receiving process grabs all of the bytes from
 951  * the pipe buffer.  Then the direct mapping write is set-up.
 952  */
 953 static int
 954 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 955 {
 956         int error;
 957
 958 retry:
 959         PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
 960         error = pipelock(wpipe, 1);
 961         if (error != 0)
 962                 goto error1;
 963         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
 964                 error = EPIPE;
 965                 pipeunlock(wpipe);
 966                 goto error1;
 967         }
 968         if (wpipe->pipe_state & PIPE_DIRECTW) {
 969                 if (wpipe->pipe_state & PIPE_WANTR) {
 970                         wpipe->pipe_state &= ~PIPE_WANTR;
 971                         wakeup(wpipe);
 972                 }
 973                 pipeselwakeup(wpipe);
 974                 wpipe->pipe_state |= PIPE_WANTW;
 975                 pipeunlock(wpipe);
 976                 error = msleep(wpipe, PIPE_MTX(wpipe),
 977                     PRIBIO | PCATCH, "pipdww", 0);
 978                 if (error)
 979                         goto error1;
 980                 else
 981                         goto retry;
 982         }
 983         if (wpipe->pipe_buffer.cnt > 0) {
 984                 if (wpipe->pipe_state & PIPE_WANTR) {
 985                         wpipe->pipe_state &= ~PIPE_WANTR;
 986                         wakeup(wpipe);
 987                 }
 988                 pipeselwakeup(wpipe);
 989                 wpipe->pipe_state |= PIPE_WANTW;
 990                 pipeunlock(wpipe);
 991                 error = msleep(wpipe, PIPE_MTX(wpipe),
 992                     PRIBIO | PCATCH, "pipdwc", 0);
 993                 if (error)
 994                         goto error1;
 995                 else
 996                         goto retry;
 997         }
 998
 999         error = pipe_build_write_buffer(wpipe, uio);
1000         if (error) {
1001                 pipeunlock(wpipe);
1002                 goto error1;
1003         }
1004
1005         while (wpipe->pipe_pages.cnt != 0 &&
1006             (wpipe->pipe_state & PIPE_EOF) == 0) {
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                 if (error != 0)
1018                         break;
1019         }
1020
1021         if ((wpipe->pipe_state & PIPE_EOF) != 0) {
1022                 wpipe->pipe_pages.cnt = 0;
1023                 pipe_destroy_write_buffer(wpipe);
1024                 pipeselwakeup(wpipe);
1025                 error = EPIPE;
1026         } else if (error == EINTR || error == ERESTART) {
1027                 pipe_clone_write_buffer(wpipe);
1028         } else {
1029                 pipe_destroy_write_buffer(wpipe);
1030         }
1031         pipeunlock(wpipe);
1032         KASSERT((wpipe->pipe_state & PIPE_DIRECTW) == 0,
1033             ("pipe %p leaked PIPE_DIRECTW", wpipe));
1034         return (error);
1035
1036 error1:
1037         wakeup(wpipe);
1038         return (error);
1039 }
1040 #endif
1041
1042 static int
1043 pipe_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
1044     int flags, struct thread *td)
1045 {
1046         struct pipe *wpipe, *rpipe;
1047         ssize_t orig_resid;
1048         int desiredsize, error;
1049
1050         rpipe = fp->f_data;
1051         wpipe = PIPE_PEER(rpipe);
1052         PIPE_LOCK(rpipe);
1053         error = pipelock(wpipe, 1);
1054         if (error) {
1055                 PIPE_UNLOCK(rpipe);
1056                 return (error);
1057         }
1058         /*
1059          * detect loss of pipe read side, issue SIGPIPE if lost.
1060          */
1061         if (wpipe->pipe_present != PIPE_ACTIVE ||
1062             (wpipe->pipe_state & PIPE_EOF)) {
1063                 pipeunlock(wpipe);
1064                 PIPE_UNLOCK(rpipe);
1065                 return (EPIPE);
1066         }
1067 #ifdef MAC
1068         error = mac_pipe_check_write(active_cred, wpipe->pipe_pair);
1069         if (error) {
1070                 pipeunlock(wpipe);
1071                 PIPE_UNLOCK(rpipe);
1072                 return (error);
1073         }
1074 #endif
1075         ++wpipe->pipe_busy;
1076
1077         /* Choose a larger size if it's advantageous */
1078         desiredsize = max(SMALL_PIPE_SIZE, wpipe->pipe_buffer.size);
1079         while (desiredsize < wpipe->pipe_buffer.cnt + uio->uio_resid) {
1080                 if (piperesizeallowed != 1)
1081                         break;
1082                 if (amountpipekva > maxpipekva / 2)
1083                         break;
1084                 if (desiredsize == BIG_PIPE_SIZE)
1085                         break;
1086                 desiredsize = desiredsize * 2;
1087         }
1088
1089         /* Choose a smaller size if we're in a OOM situation */
1090         if (amountpipekva > (3 * maxpipekva) / 4 &&
1091             wpipe->pipe_buffer.size > SMALL_PIPE_SIZE &&
1092             wpipe->pipe_buffer.cnt <= SMALL_PIPE_SIZE &&
1093             piperesizeallowed == 1)
1094                 desiredsize = SMALL_PIPE_SIZE;
1095
1096         /* Resize if the above determined that a new size was necessary */
1097         if (desiredsize != wpipe->pipe_buffer.size &&
1098             (wpipe->pipe_state & PIPE_DIRECTW) == 0) {
1099                 PIPE_UNLOCK(wpipe);
1100                 pipespace(wpipe, desiredsize);
1101                 PIPE_LOCK(wpipe);
1102         }
1103         MPASS(wpipe->pipe_buffer.size != 0);
1104
1105         pipeunlock(wpipe);
1106
1107         orig_resid = uio->uio_resid;
1108
1109         while (uio->uio_resid) {
1110                 int space;
1111
1112                 pipelock(wpipe, 0);
1113                 if (wpipe->pipe_state & PIPE_EOF) {
1114                         pipeunlock(wpipe);
1115                         error = EPIPE;
1116                         break;
1117                 }
1118 #ifndef PIPE_NODIRECT
1119                 /*
1120                  * If the transfer is large, we can gain performance if
1121                  * we do process-to-process copies directly.
1122                  * If the write is non-blocking, we don't use the
1123                  * direct write mechanism.
1124                  *
1125                  * The direct write mechanism will detect the reader going
1126                  * away on us.
1127                  */
1128                 if (uio->uio_segflg == UIO_USERSPACE &&
1129                     uio->uio_iov->iov_len >= PIPE_MINDIRECT &&
1130                     wpipe->pipe_buffer.size >= PIPE_MINDIRECT &&
1131                     (fp->f_flag & FNONBLOCK) == 0) {
1132                         pipeunlock(wpipe);
1133                         error = pipe_direct_write(wpipe, uio);
1134                         if (error)
1135                                 break;
1136                         continue;
1137                 }
1138 #endif
1139
1140                 /*
1141                  * Pipe buffered writes cannot be coincidental with
1142                  * direct writes.  We wait until the currently executing
1143                  * direct write is completed before we start filling the
1144                  * pipe buffer.  We break out if a signal occurs or the
1145                  * reader goes away.
1146                  */
1147                 if (wpipe->pipe_pages.cnt != 0) {
1148                         if (wpipe->pipe_state & PIPE_WANTR) {
1149                                 wpipe->pipe_state &= ~PIPE_WANTR;
1150                                 wakeup(wpipe);
1151                         }
1152                         pipeselwakeup(wpipe);
1153                         wpipe->pipe_state |= PIPE_WANTW;
1154                         pipeunlock(wpipe);
1155                         error = msleep(wpipe, PIPE_MTX(rpipe), PRIBIO | PCATCH,
1156                             "pipbww", 0);
1157                         if (error)
1158                                 break;
1159                         else
1160                                 continue;
1161                 }
1162
1163                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1164
1165                 /* Writes of size <= PIPE_BUF must be atomic. */
1166                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1167                         space = 0;
1168
1169                 if (space > 0) {
1170                         int size;       /* Transfer size */
1171                         int segsize;    /* first segment to transfer */
1172
1173                         /*
1174                          * Transfer size is minimum of uio transfer
1175                          * and free space in pipe buffer.
1176                          */
1177                         if (space > uio->uio_resid)
1178                                 size = uio->uio_resid;
1179                         else
1180                                 size = space;
1181                         /*
1182                          * First segment to transfer is minimum of
1183                          * transfer size and contiguous space in
1184                          * pipe buffer.  If first segment to transfer
1185                          * is less than the transfer size, we've got
1186                          * a wraparound in the buffer.
1187                          */
1188                         segsize = wpipe->pipe_buffer.size -
1189                                 wpipe->pipe_buffer.in;
1190                         if (segsize > size)
1191                                 segsize = size;
1192
1193                         /* Transfer first segment */
1194
1195                         PIPE_UNLOCK(rpipe);
1196                         error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1197                                         segsize, uio);
1198                         PIPE_LOCK(rpipe);
1199
1200                         if (error == 0 && segsize < size) {
1201                                 KASSERT(wpipe->pipe_buffer.in + segsize ==
1202                                         wpipe->pipe_buffer.size,
1203                                         ("Pipe buffer wraparound disappeared"));
1204                                 /*
1205                                  * Transfer remaining part now, to
1206                                  * support atomic writes.  Wraparound
1207                                  * happened.
1208                                  */
1209
1210                                 PIPE_UNLOCK(rpipe);
1211                                 error = uiomove(
1212                                     &wpipe->pipe_buffer.buffer[0],
1213                                     size - segsize, uio);
1214                                 PIPE_LOCK(rpipe);
1215                         }
1216                         if (error == 0) {
1217                                 wpipe->pipe_buffer.in += size;
1218                                 if (wpipe->pipe_buffer.in >=
1219                                     wpipe->pipe_buffer.size) {
1220                                         KASSERT(wpipe->pipe_buffer.in ==
1221                                                 size - segsize +
1222                                                 wpipe->pipe_buffer.size,
1223                                                 ("Expected wraparound bad"));
1224                                         wpipe->pipe_buffer.in = size - segsize;
1225                                 }
1226
1227                                 wpipe->pipe_buffer.cnt += size;
1228                                 KASSERT(wpipe->pipe_buffer.cnt <=
1229                                         wpipe->pipe_buffer.size,
1230                                         ("Pipe buffer overflow"));
1231                         }
1232                         pipeunlock(wpipe);
1233                         if (error != 0)
1234                                 break;
1235                 } else {
1236                         /*
1237                          * If the "read-side" has been blocked, wake it up now.
1238                          */
1239                         if (wpipe->pipe_state & PIPE_WANTR) {
1240                                 wpipe->pipe_state &= ~PIPE_WANTR;
1241                                 wakeup(wpipe);
1242                         }
1243
1244                         /*
1245                          * don't block on non-blocking I/O
1246                          */
1247                         if (fp->f_flag & FNONBLOCK) {
1248                                 error = EAGAIN;
1249                                 pipeunlock(wpipe);
1250                                 break;
1251                         }
1252
1253                         /*
1254                          * We have no more space and have something to offer,
1255                          * wake up select/poll.
1256                          */
1257                         pipeselwakeup(wpipe);
1258
1259                         wpipe->pipe_state |= PIPE_WANTW;
1260                         pipeunlock(wpipe);
1261                         error = msleep(wpipe, PIPE_MTX(rpipe),
1262                             PRIBIO | PCATCH, "pipewr", 0);
1263                         if (error != 0)
1264                                 break;
1265                 }
1266         }
1267
1268         pipelock(wpipe, 0);
1269         --wpipe->pipe_busy;
1270
1271         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1272                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1273                 wakeup(wpipe);
1274         } else if (wpipe->pipe_buffer.cnt > 0) {
1275                 /*
1276                  * If we have put any characters in the buffer, we wake up
1277                  * the reader.
1278                  */
1279                 if (wpipe->pipe_state & PIPE_WANTR) {
1280                         wpipe->pipe_state &= ~PIPE_WANTR;
1281                         wakeup(wpipe);
1282                 }
1283         }
1284
1285         /*
1286          * Don't return EPIPE if any byte was written.
1287          * EINTR and other interrupts are handled by generic I/O layer.
1288          * Do not pretend that I/O succeeded for obvious user error
1289          * like EFAULT.
1290          */
1291         if (uio->uio_resid != orig_resid && error == EPIPE)
1292                 error = 0;
1293
1294         if (error == 0)
1295                 vfs_timestamp(&wpipe->pipe_mtime);
1296
1297         /*
1298          * We have something to offer,
1299          * wake up select/poll.
1300          */
1301         if (wpipe->pipe_buffer.cnt)
1302                 pipeselwakeup(wpipe);
1303
1304         pipeunlock(wpipe);
1305         PIPE_UNLOCK(rpipe);
1306         return (error);
1307 }
1308
1309 /* ARGSUSED */
1310 static int
1311 pipe_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1312     struct thread *td)
1313 {
1314         struct pipe *cpipe;
1315         int error;
1316
1317         cpipe = fp->f_data;
1318         if (cpipe->pipe_state & PIPE_NAMED)
1319                 error = vnops.fo_truncate(fp, length, active_cred, td);
1320         else
1321                 error = invfo_truncate(fp, length, active_cred, td);
1322         return (error);
1323 }
1324
1325 /*
1326  * we implement a very minimal set of ioctls for compatibility with sockets.
1327  */
1328 static int
1329 pipe_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
1330     struct thread *td)
1331 {
1332         struct pipe *mpipe = fp->f_data;
1333         int error;
1334
1335         PIPE_LOCK(mpipe);
1336
1337 #ifdef MAC
1338         error = mac_pipe_check_ioctl(active_cred, mpipe->pipe_pair, cmd, data);
1339         if (error) {
1340                 PIPE_UNLOCK(mpipe);
1341                 return (error);
1342         }
1343 #endif
1344
1345         error = 0;
1346         switch (cmd) {
1347
1348         case FIONBIO:
1349                 break;
1350
1351         case FIOASYNC:
1352                 if (*(int *)data) {
1353                         mpipe->pipe_state |= PIPE_ASYNC;
1354                 } else {
1355                         mpipe->pipe_state &= ~PIPE_ASYNC;
1356                 }
1357                 break;
1358
1359         case FIONREAD:
1360                 if (!(fp->f_flag & FREAD)) {
1361                         *(int *)data = 0;
1362                         PIPE_UNLOCK(mpipe);
1363                         return (0);
1364                 }
1365                 if (mpipe->pipe_pages.cnt != 0)
1366                         *(int *)data = mpipe->pipe_pages.cnt;
1367                 else
1368                         *(int *)data = mpipe->pipe_buffer.cnt;
1369                 break;
1370
1371         case FIOSETOWN:
1372                 PIPE_UNLOCK(mpipe);
1373                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1374                 goto out_unlocked;
1375
1376         case FIOGETOWN:
1377                 *(int *)data = fgetown(&mpipe->pipe_sigio);
1378                 break;
1379
1380         /* This is deprecated, FIOSETOWN should be used instead. */
1381         case TIOCSPGRP:
1382                 PIPE_UNLOCK(mpipe);
1383                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1384                 goto out_unlocked;
1385
1386         /* This is deprecated, FIOGETOWN should be used instead. */
1387         case TIOCGPGRP:
1388                 *(int *)data = -fgetown(&mpipe->pipe_sigio);
1389                 break;
1390
1391         default:
1392                 error = ENOTTY;
1393                 break;
1394         }
1395         PIPE_UNLOCK(mpipe);
1396 out_unlocked:
1397         return (error);
1398 }
1399
1400 static int
1401 pipe_poll(struct file *fp, int events, struct ucred *active_cred,
1402     struct thread *td)
1403 {
1404         struct pipe *rpipe;
1405         struct pipe *wpipe;
1406         int levents, revents;
1407 #ifdef MAC
1408         int error;
1409 #endif
1410
1411         revents = 0;
1412         rpipe = fp->f_data;
1413         wpipe = PIPE_PEER(rpipe);
1414         PIPE_LOCK(rpipe);
1415 #ifdef MAC
1416         error = mac_pipe_check_poll(active_cred, rpipe->pipe_pair);
1417         if (error)
1418                 goto locked_error;
1419 #endif
1420         if (fp->f_flag & FREAD && events & (POLLIN | POLLRDNORM))
1421                 if (rpipe->pipe_pages.cnt > 0 || rpipe->pipe_buffer.cnt > 0)
1422                         revents |= events & (POLLIN | POLLRDNORM);
1423
1424         if (fp->f_flag & FWRITE && events & (POLLOUT | POLLWRNORM))
1425                 if (wpipe->pipe_present != PIPE_ACTIVE ||
1426                     (wpipe->pipe_state & PIPE_EOF) ||
1427                     ((wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
1428                      ((wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF ||
1429                          wpipe->pipe_buffer.size == 0)))
1430                         revents |= events & (POLLOUT | POLLWRNORM);
1431
1432         levents = events &
1433             (POLLIN | POLLINIGNEOF | POLLPRI | POLLRDNORM | POLLRDBAND);
1434         if (rpipe->pipe_state & PIPE_NAMED && fp->f_flag & FREAD && levents &&
1435             fp->f_pipegen == rpipe->pipe_wgen)
1436                 events |= POLLINIGNEOF;
1437
1438         if ((events & POLLINIGNEOF) == 0) {
1439                 if (rpipe->pipe_state & PIPE_EOF) {
1440                         if (fp->f_flag & FREAD)
1441                                 revents |= (events & (POLLIN | POLLRDNORM));
1442                         if (wpipe->pipe_present != PIPE_ACTIVE ||
1443                             (wpipe->pipe_state & PIPE_EOF))
1444                                 revents |= POLLHUP;
1445                 }
1446         }
1447
1448         if (revents == 0) {
1449                 /*
1450                  * Add ourselves regardless of eventmask as we have to return
1451                  * POLLHUP even if it was not asked for.
1452                  */
1453                 if ((fp->f_flag & FREAD) != 0) {
1454                         selrecord(td, &rpipe->pipe_sel);
1455                         if (SEL_WAITING(&rpipe->pipe_sel))
1456                                 rpipe->pipe_state |= PIPE_SEL;
1457                 }
1458
1459                 if ((fp->f_flag & FWRITE) != 0) {
1460                         selrecord(td, &wpipe->pipe_sel);
1461                         if (SEL_WAITING(&wpipe->pipe_sel))
1462                                 wpipe->pipe_state |= PIPE_SEL;
1463                 }
1464         }
1465 #ifdef MAC
1466 locked_error:
1467 #endif
1468         PIPE_UNLOCK(rpipe);
1469
1470         return (revents);
1471 }
1472
1473 /*
1474  * We shouldn't need locks here as we're doing a read and this should
1475  * be a natural race.
1476  */
1477 static int
1478 pipe_stat(struct file *fp, struct stat *ub, struct ucred *active_cred,
1479     struct thread *td)
1480 {
1481         struct pipe *pipe;
1482 #ifdef MAC
1483         int error;
1484 #endif
1485
1486         pipe = fp->f_data;
1487         PIPE_LOCK(pipe);
1488 #ifdef MAC
1489         error = mac_pipe_check_stat(active_cred, pipe->pipe_pair);
1490         if (error) {
1491                 PIPE_UNLOCK(pipe);
1492                 return (error);
1493         }
1494 #endif
1495
1496         /* For named pipes ask the underlying filesystem. */
1497         if (pipe->pipe_state & PIPE_NAMED) {
1498                 PIPE_UNLOCK(pipe);
1499                 return (vnops.fo_stat(fp, ub, active_cred, td));
1500         }
1501
1502         PIPE_UNLOCK(pipe);
1503
1504         bzero(ub, sizeof(*ub));
1505         ub->st_mode = S_IFIFO;
1506         ub->st_blksize = PAGE_SIZE;
1507         if (pipe->pipe_pages.cnt != 0)
1508                 ub->st_size = pipe->pipe_pages.cnt;
1509         else
1510                 ub->st_size = pipe->pipe_buffer.cnt;
1511         ub->st_blocks = howmany(ub->st_size, ub->st_blksize);
1512         ub->st_atim = pipe->pipe_atime;
1513         ub->st_mtim = pipe->pipe_mtime;
1514         ub->st_ctim = pipe->pipe_ctime;
1515         ub->st_uid = fp->f_cred->cr_uid;
1516         ub->st_gid = fp->f_cred->cr_gid;
1517         ub->st_dev = pipedev_ino;
1518         ub->st_ino = pipe->pipe_ino;
1519         /*
1520          * Left as 0: st_nlink, st_rdev, st_flags, st_gen.
1521          */
1522         return (0);
1523 }
1524
1525 /* ARGSUSED */
1526 static int
1527 pipe_close(struct file *fp, struct thread *td)
1528 {
1529
1530         if (fp->f_vnode != NULL)
1531                 return vnops.fo_close(fp, td);
1532         fp->f_ops = &badfileops;
1533         pipe_dtor(fp->f_data);
1534         fp->f_data = NULL;
1535         return (0);
1536 }
1537
1538 static int
1539 pipe_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td)
1540 {
1541         struct pipe *cpipe;
1542         int error;
1543
1544         cpipe = fp->f_data;
1545         if (cpipe->pipe_state & PIPE_NAMED)
1546                 error = vn_chmod(fp, mode, active_cred, td);
1547         else
1548                 error = invfo_chmod(fp, mode, active_cred, td);
1549         return (error);
1550 }
1551
1552 static int
1553 pipe_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1554     struct thread *td)
1555 {
1556         struct pipe *cpipe;
1557         int error;
1558
1559         cpipe = fp->f_data;
1560         if (cpipe->pipe_state & PIPE_NAMED)
1561                 error = vn_chown(fp, uid, gid, active_cred, td);
1562         else
1563                 error = invfo_chown(fp, uid, gid, active_cred, td);
1564         return (error);
1565 }
1566
1567 static int
1568 pipe_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1569 {
1570         struct pipe *pi;
1571
1572         if (fp->f_type == DTYPE_FIFO)
1573                 return (vn_fill_kinfo(fp, kif, fdp));
1574         kif->kf_type = KF_TYPE_PIPE;
1575         pi = fp->f_data;
1576         kif->kf_un.kf_pipe.kf_pipe_addr = (uintptr_t)pi;
1577         kif->kf_un.kf_pipe.kf_pipe_peer = (uintptr_t)pi->pipe_peer;
1578         kif->kf_un.kf_pipe.kf_pipe_buffer_cnt = pi->pipe_buffer.cnt;
1579         return (0);
1580 }
1581
1582 static void
1583 pipe_free_kmem(struct pipe *cpipe)
1584 {
1585
1586         KASSERT(!mtx_owned(PIPE_MTX(cpipe)),
1587             ("pipe_free_kmem: pipe mutex locked"));
1588
1589         if (cpipe->pipe_buffer.buffer != NULL) {
1590                 atomic_subtract_long(&amountpipekva, cpipe->pipe_buffer.size);
1591                 vm_map_remove(pipe_map,
1592                     (vm_offset_t)cpipe->pipe_buffer.buffer,
1593                     (vm_offset_t)cpipe->pipe_buffer.buffer + cpipe->pipe_buffer.size);
1594                 cpipe->pipe_buffer.buffer = NULL;
1595         }
1596 #ifndef PIPE_NODIRECT
1597         {
1598                 cpipe->pipe_pages.cnt = 0;
1599                 cpipe->pipe_pages.pos = 0;
1600                 cpipe->pipe_pages.npages = 0;
1601         }
1602 #endif
1603 }
1604
1605 /*
1606  * shutdown the pipe
1607  */
1608 static void
1609 pipeclose(struct pipe *cpipe)
1610 {
1611         struct pipepair *pp;
1612         struct pipe *ppipe;
1613
1614         KASSERT(cpipe != NULL, ("pipeclose: cpipe == NULL"));
1615
1616         PIPE_LOCK(cpipe);
1617         pipelock(cpipe, 0);
1618         pp = cpipe->pipe_pair;
1619
1620         /*
1621          * If the other side is blocked, wake it up saying that
1622          * we want to close it down.
1623          */
1624         cpipe->pipe_state |= PIPE_EOF;
1625         while (cpipe->pipe_busy) {
1626                 wakeup(cpipe);
1627                 cpipe->pipe_state |= PIPE_WANT;
1628                 pipeunlock(cpipe);
1629                 msleep(cpipe, PIPE_MTX(cpipe), PRIBIO, "pipecl", 0);
1630                 pipelock(cpipe, 0);
1631         }
1632
1633         pipeselwakeup(cpipe);
1634
1635         /*
1636          * Disconnect from peer, if any.
1637          */
1638         ppipe = cpipe->pipe_peer;
1639         if (ppipe->pipe_present == PIPE_ACTIVE) {
1640                 ppipe->pipe_state |= PIPE_EOF;
1641                 wakeup(ppipe);
1642                 pipeselwakeup(ppipe);
1643         }
1644
1645         /*
1646          * Mark this endpoint as free.  Release kmem resources.  We
1647          * don't mark this endpoint as unused until we've finished
1648          * doing that, or the pipe might disappear out from under
1649          * us.
1650          */
1651         PIPE_UNLOCK(cpipe);
1652         pipe_free_kmem(cpipe);
1653         PIPE_LOCK(cpipe);
1654         cpipe->pipe_present = PIPE_CLOSING;
1655         pipeunlock(cpipe);
1656
1657         /*
1658          * knlist_clear() may sleep dropping the PIPE_MTX. Set the
1659          * PIPE_FINALIZED, that allows other end to free the
1660          * pipe_pair, only after the knotes are completely dismantled.
1661          */
1662         knlist_clear(&cpipe->pipe_sel.si_note, 1);
1663         cpipe->pipe_present = PIPE_FINALIZED;
1664         seldrain(&cpipe->pipe_sel);
1665         knlist_destroy(&cpipe->pipe_sel.si_note);
1666
1667         /*
1668          * If both endpoints are now closed, release the memory for the
1669          * pipe pair.  If not, unlock.
1670          */
1671         if (ppipe->pipe_present == PIPE_FINALIZED) {
1672                 PIPE_UNLOCK(cpipe);
1673 #ifdef MAC
1674                 mac_pipe_destroy(pp);
1675 #endif
1676                 uma_zfree(pipe_zone, cpipe->pipe_pair);
1677         } else
1678                 PIPE_UNLOCK(cpipe);
1679 }
1680
1681 /*ARGSUSED*/
1682 static int
1683 pipe_kqfilter(struct file *fp, struct knote *kn)
1684 {
1685         struct pipe *cpipe;
1686
1687         /*
1688          * If a filter is requested that is not supported by this file
1689          * descriptor, don't return an error, but also don't ever generate an
1690          * event.
1691          */
1692         if ((kn->kn_filter == EVFILT_READ) && !(fp->f_flag & FREAD)) {
1693                 kn->kn_fop = &pipe_nfiltops;
1694                 return (0);
1695         }
1696         if ((kn->kn_filter == EVFILT_WRITE) && !(fp->f_flag & FWRITE)) {
1697                 kn->kn_fop = &pipe_nfiltops;
1698                 return (0);
1699         }
1700         cpipe = fp->f_data;
1701         PIPE_LOCK(cpipe);
1702         switch (kn->kn_filter) {
1703         case EVFILT_READ:
1704                 kn->kn_fop = &pipe_rfiltops;
1705                 break;
1706         case EVFILT_WRITE:
1707                 kn->kn_fop = &pipe_wfiltops;
1708                 if (cpipe->pipe_peer->pipe_present != PIPE_ACTIVE) {
1709                         /* other end of pipe has been closed */
1710                         PIPE_UNLOCK(cpipe);
1711                         return (EPIPE);
1712                 }
1713                 cpipe = PIPE_PEER(cpipe);
1714                 break;
1715         default:
1716                 PIPE_UNLOCK(cpipe);
1717                 return (EINVAL);
1718         }
1719
1720         kn->kn_hook = cpipe;
1721         knlist_add(&cpipe->pipe_sel.si_note, kn, 1);
1722         PIPE_UNLOCK(cpipe);
1723         return (0);
1724 }
1725
1726 static void
1727 filt_pipedetach(struct knote *kn)
1728 {
1729         struct pipe *cpipe = kn->kn_hook;
1730
1731         PIPE_LOCK(cpipe);
1732         knlist_remove(&cpipe->pipe_sel.si_note, kn, 1);
1733         PIPE_UNLOCK(cpipe);
1734 }
1735
1736 /*ARGSUSED*/
1737 static int
1738 filt_piperead(struct knote *kn, long hint)
1739 {
1740         struct file *fp = kn->kn_fp;
1741         struct pipe *rpipe = kn->kn_hook;
1742
1743         PIPE_LOCK_ASSERT(rpipe, MA_OWNED);
1744         kn->kn_data = rpipe->pipe_buffer.cnt;
1745         if (kn->kn_data == 0)
1746                 kn->kn_data = rpipe->pipe_pages.cnt;
1747
1748         if ((rpipe->pipe_state & PIPE_EOF) != 0 &&
1749             ((rpipe->pipe_state & PIPE_NAMED) == 0 ||
1750             fp->f_pipegen != rpipe->pipe_wgen)) {
1751                 kn->kn_flags |= EV_EOF;
1752                 return (1);
1753         }
1754         kn->kn_flags &= ~EV_EOF;
1755         return (kn->kn_data > 0);
1756 }
1757
1758 /*ARGSUSED*/
1759 static int
1760 filt_pipewrite(struct knote *kn, long hint)
1761 {
1762         struct pipe *wpipe = kn->kn_hook;
1763
1764         /*
1765          * If this end of the pipe is closed, the knote was removed from the
1766          * knlist and the list lock (i.e., the pipe lock) is therefore not held.
1767          */
1768         if (wpipe->pipe_present == PIPE_ACTIVE ||
1769             (wpipe->pipe_state & PIPE_NAMED) != 0) {
1770                 PIPE_LOCK_ASSERT(wpipe, MA_OWNED);
1771
1772                 if (wpipe->pipe_state & PIPE_DIRECTW) {
1773                         kn->kn_data = 0;
1774                 } else if (wpipe->pipe_buffer.size > 0) {
1775                         kn->kn_data = wpipe->pipe_buffer.size -
1776                             wpipe->pipe_buffer.cnt;
1777                 } else {
1778                         kn->kn_data = PIPE_BUF;
1779                 }
1780         }
1781
1782         if (wpipe->pipe_present != PIPE_ACTIVE ||
1783             (wpipe->pipe_state & PIPE_EOF)) {
1784                 kn->kn_flags |= EV_EOF;
1785                 return (1);
1786         }
1787         kn->kn_flags &= ~EV_EOF;
1788         return (kn->kn_data >= PIPE_BUF);
1789 }
1790
1791 static void
1792 filt_pipedetach_notsup(struct knote *kn)
1793 {
1794
1795 }
1796
1797 static int
1798 filt_pipenotsup(struct knote *kn, long hint)
1799 {
1800
1801         return (0);
1802 }