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  * $FreeBSD: src/sys/kern/sys_pipe.c,v 1.60.2.13 2002/08/05 15:05:15 des Exp $
  20  * $DragonFly: src/sys/kern/sys_pipe.c,v 1.40 2006/08/02 01:25:25 dillon Exp $
  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, it is fully mapped and wired into the kernel, and
  36  * the receiving process can copy it directly from the pages in the sending
  37  * 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.  PIPE_SIZE is constrained by the
  50  * amount of kernel virtual memory.
  51  */
  52
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/file.h>
  59 #include <sys/filedesc.h>
  60 #include <sys/filio.h>
  61 #include <sys/ttycom.h>
  62 #include <sys/stat.h>
  63 #include <sys/poll.h>
  64 #include <sys/select.h>
  65 #include <sys/signalvar.h>
  66 #include <sys/sysproto.h>
  67 #include <sys/pipe.h>
  68 #include <sys/vnode.h>
  69 #include <sys/uio.h>
  70 #include <sys/event.h>
  71 #include <sys/globaldata.h>
  72 #include <sys/module.h>
  73 #include <sys/malloc.h>
  74 #include <sys/sysctl.h>
  75 #include <sys/socket.h>
  76
  77 #include <vm/vm.h>
  78 #include <vm/vm_param.h>
  79 #include <sys/lock.h>
  80 #include <vm/vm_object.h>
  81 #include <vm/vm_kern.h>
  82 #include <vm/vm_extern.h>
  83 #include <vm/pmap.h>
  84 #include <vm/vm_map.h>
  85 #include <vm/vm_page.h>
  86 #include <vm/vm_zone.h>
  87
  88 #include <sys/file2.h>
  89
  90 #include <machine/cpufunc.h>
  91
  92 /*
  93  * interfaces to the outside world
  94  */
  95 static int pipe_read (struct file *fp, struct uio *uio,
  96                 struct ucred *cred, int flags);
  97 static int pipe_write (struct file *fp, struct uio *uio,
  98                 struct ucred *cred, int flags);
  99 static int pipe_close (struct file *fp);
 100 static int pipe_shutdown (struct file *fp, int how);
 101 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
 102 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 103 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
 104 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred);
 105
 106 static struct fileops pipeops = {
 107         .fo_read = pipe_read,
 108         .fo_write = pipe_write,
 109         .fo_ioctl = pipe_ioctl,
 110         .fo_poll = pipe_poll,
 111         .fo_kqfilter = pipe_kqfilter,
 112         .fo_stat = pipe_stat,
 113         .fo_close = pipe_close,
 114         .fo_shutdown = pipe_shutdown
 115 };
 116
 117 static void     filt_pipedetach(struct knote *kn);
 118 static int      filt_piperead(struct knote *kn, long hint);
 119 static int      filt_pipewrite(struct knote *kn, long hint);
 120
 121 static struct filterops pipe_rfiltops =
 122         { 1, NULL, filt_pipedetach, filt_piperead };
 123 static struct filterops pipe_wfiltops =
 124         { 1, NULL, filt_pipedetach, filt_pipewrite };
 125
 126 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 127
 128 /*
 129  * Default pipe buffer size(s), this can be kind-of large now because pipe
 130  * space is pageable.  The pipe code will try to maintain locality of
 131  * reference for performance reasons, so small amounts of outstanding I/O
 132  * will not wipe the cache.
 133  */
 134 #define MINPIPESIZE (PIPE_SIZE/3)
 135 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 136
 137 /*
 138  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 139  * is there so that on large systems, we don't exhaust it.
 140  */
 141 #define MAXPIPEKVA (8*1024*1024)
 142
 143 /*
 144  * Limit for direct transfers, we cannot, of course limit
 145  * the amount of kva for pipes in general though.
 146  */
 147 #define LIMITPIPEKVA (16*1024*1024)
 148
 149 /*
 150  * Limit the number of "big" pipes
 151  */
 152 #define LIMITBIGPIPES   32
 153 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 154
 155 static int pipe_maxbig = LIMITBIGPIPES;
 156 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 157 static int pipe_nbig;
 158 static int pipe_bcache_alloc;
 159 static int pipe_bkmem_alloc;
 160 static int pipe_dwrite_enable = 1;      /* 0:copy, 1:kmem/sfbuf 2:force */
 161 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 162                                         /* 3:sfbuf_dmap w/ forced invlpg */
 163
 164 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 165 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 166         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 167 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 168         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 169 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 170         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 171 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 172         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 173 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 174         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
 175         "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 176 #if !defined(NO_PIPE_SYSCTL_STATS)
 177 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 178         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 179 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 180         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 181 #endif
 182
 183 static void pipeclose (struct pipe *cpipe);
 184 static void pipe_free_kmem (struct pipe *cpipe);
 185 static int pipe_create (struct pipe **cpipep);
 186 static __inline int pipelock (struct pipe *cpipe, int catch);
 187 static __inline void pipeunlock (struct pipe *cpipe);
 188 static __inline void pipeselwakeup (struct pipe *cpipe);
 189 #ifndef PIPE_NODIRECT
 190 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 191 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 192 static void pipe_clone_write_buffer (struct pipe *wpipe);
 193 #endif
 194 static int pipespace (struct pipe *cpipe, int size);
 195
 196 /*
 197  * The pipe system call for the DTYPE_PIPE type of pipes
 198  *
 199  * pipe_ARgs(int dummy)
 200  */
 201
 202 /* ARGSUSED */
 203 int
 204 sys_pipe(struct pipe_args *uap)
 205 {
 206         struct thread *td = curthread;
 207         struct proc *p = td->td_proc;
 208         struct file *rf, *wf;
 209         struct pipe *rpipe, *wpipe;
 210         int fd1, fd2, error;
 211
 212         KKASSERT(p);
 213
 214         rpipe = wpipe = NULL;
 215         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 216                 pipeclose(rpipe);
 217                 pipeclose(wpipe);
 218                 return (ENFILE);
 219         }
 220
 221         rpipe->pipe_state |= PIPE_DIRECTOK;
 222         wpipe->pipe_state |= PIPE_DIRECTOK;
 223
 224         /*
 225          * Select the direct-map features to use for this pipe.  Since the
 226          * sysctl's can change on the fly we record the settings when the
 227          * pipe is created.
 228          *
 229          * Generally speaking the system will default to what we consider
 230          * to be the best-balanced and most stable option.  Right now this
 231          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 232          * moment.
 233          */
 234         wpipe->pipe_feature = PIPE_COPY;
 235         if (pipe_dwrite_enable) {
 236                 switch(pipe_dwrite_sfbuf) {
 237                 case 0:
 238                         wpipe->pipe_feature = PIPE_KMEM;
 239                         break;
 240                 case 1:
 241                         wpipe->pipe_feature = PIPE_SFBUF1;
 242                         break;
 243                 case 2:
 244                 case 3:
 245                         wpipe->pipe_feature = PIPE_SFBUF2;
 246                         break;
 247                 }
 248         }
 249         rpipe->pipe_feature = wpipe->pipe_feature;
 250
 251         error = falloc(p, &rf, &fd1);
 252         if (error) {
 253                 pipeclose(rpipe);
 254                 pipeclose(wpipe);
 255                 return (error);
 256         }
 257         uap->sysmsg_fds[0] = fd1;
 258
 259         /*
 260          * Warning: once we've gotten past allocation of the fd for the
 261          * read-side, we can only drop the read side via fdrop() in order
 262          * to avoid races against processes which manage to dup() the read
 263          * side while we are blocked trying to allocate the write side.
 264          */
 265         rf->f_type = DTYPE_PIPE;
 266         rf->f_flag = FREAD | FWRITE;
 267         rf->f_ops = &pipeops;
 268         rf->f_data = rpipe;
 269         error = falloc(p, &wf, &fd2);
 270         if (error) {
 271                 fsetfd(p, NULL, fd1);
 272                 fdrop(rf);
 273                 /* rpipe has been closed by fdrop(). */
 274                 pipeclose(wpipe);
 275                 return (error);
 276         }
 277         wf->f_type = DTYPE_PIPE;
 278         wf->f_flag = FREAD | FWRITE;
 279         wf->f_ops = &pipeops;
 280         wf->f_data = wpipe;
 281         uap->sysmsg_fds[1] = fd2;
 282
 283         rpipe->pipe_peer = wpipe;
 284         wpipe->pipe_peer = rpipe;
 285
 286         fsetfd(p, rf, fd1);
 287         fsetfd(p, wf, fd2);
 288         fdrop(rf);
 289         fdrop(wf);
 290
 291         return (0);
 292 }
 293
 294 /*
 295  * Allocate kva for pipe circular buffer, the space is pageable
 296  * This routine will 'realloc' the size of a pipe safely, if it fails
 297  * it will retain the old buffer.
 298  * If it fails it will return ENOMEM.
 299  */
 300 static int
 301 pipespace(struct pipe *cpipe, int size)
 302 {
 303         struct vm_object *object;
 304         caddr_t buffer;
 305         int npages, error;
 306
 307         npages = round_page(size) / PAGE_SIZE;
 308         object = cpipe->pipe_buffer.object;
 309
 310         /*
 311          * [re]create the object if necessary and reserve space for it
 312          * in the kernel_map.  The object and memory are pageable.  On
 313          * success, free the old resources before assigning the new
 314          * ones.
 315          */
 316         if (object == NULL || object->size != npages) {
 317                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 318                 buffer = (caddr_t) vm_map_min(kernel_map);
 319
 320                 error = vm_map_find(kernel_map, object, 0,
 321                         (vm_offset_t *) &buffer, size, 1,
 322                         VM_PROT_ALL, VM_PROT_ALL, 0);
 323
 324                 if (error != KERN_SUCCESS) {
 325                         vm_object_deallocate(object);
 326                         return (ENOMEM);
 327                 }
 328                 pipe_free_kmem(cpipe);
 329                 cpipe->pipe_buffer.object = object;
 330                 cpipe->pipe_buffer.buffer = buffer;
 331                 cpipe->pipe_buffer.size = size;
 332                 ++pipe_bkmem_alloc;
 333         } else {
 334                 ++pipe_bcache_alloc;
 335         }
 336         cpipe->pipe_buffer.in = 0;
 337         cpipe->pipe_buffer.out = 0;
 338         cpipe->pipe_buffer.cnt = 0;
 339         return (0);
 340 }
 341
 342 /*
 343  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 344  * our per-cpu cache if possible.  For now make sure it is sized for the
 345  * smaller PIPE_SIZE default.
 346  */
 347 static int
 348 pipe_create(cpipep)
 349         struct pipe **cpipep;
 350 {
 351         globaldata_t gd = mycpu;
 352         struct pipe *cpipe;
 353         int error;
 354
 355         if ((cpipe = gd->gd_pipeq) != NULL) {
 356                 gd->gd_pipeq = cpipe->pipe_peer;
 357                 --gd->gd_pipeqcount;
 358                 cpipe->pipe_peer = NULL;
 359         } else {
 360                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 361         }
 362         *cpipep = cpipe;
 363         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 364                 return (error);
 365         vfs_timestamp(&cpipe->pipe_ctime);
 366         cpipe->pipe_atime = cpipe->pipe_ctime;
 367         cpipe->pipe_mtime = cpipe->pipe_ctime;
 368         return (0);
 369 }
 370
 371
 372 /*
 373  * lock a pipe for I/O, blocking other access
 374  */
 375 static __inline int
 376 pipelock(cpipe, catch)
 377         struct pipe *cpipe;
 378         int catch;
 379 {
 380         int error;
 381
 382         while (cpipe->pipe_state & PIPE_LOCK) {
 383                 cpipe->pipe_state |= PIPE_LWANT;
 384                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 385                 if (error != 0)
 386                         return (error);
 387         }
 388         cpipe->pipe_state |= PIPE_LOCK;
 389         return (0);
 390 }
 391
 392 /*
 393  * unlock a pipe I/O lock
 394  */
 395 static __inline void
 396 pipeunlock(cpipe)
 397         struct pipe *cpipe;
 398 {
 399
 400         cpipe->pipe_state &= ~PIPE_LOCK;
 401         if (cpipe->pipe_state & PIPE_LWANT) {
 402                 cpipe->pipe_state &= ~PIPE_LWANT;
 403                 wakeup(cpipe);
 404         }
 405 }
 406
 407 static __inline void
 408 pipeselwakeup(cpipe)
 409         struct pipe *cpipe;
 410 {
 411
 412         if (cpipe->pipe_state & PIPE_SEL) {
 413                 cpipe->pipe_state &= ~PIPE_SEL;
 414                 selwakeup(&cpipe->pipe_sel);
 415         }
 416         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 417                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 418         KNOTE(&cpipe->pipe_sel.si_note, 0);
 419 }
 420
 421 /*
 422  * MPALMOSTSAFE (acquires mplock)
 423  */
 424 static int
 425 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 426 {
 427         struct pipe *rpipe;
 428         int error;
 429         int nread = 0;
 430         int nbio;
 431         u_int size;
 432
 433         get_mplock();
 434         rpipe = (struct pipe *) fp->f_data;
 435         ++rpipe->pipe_busy;
 436         error = pipelock(rpipe, 1);
 437         if (error)
 438                 goto unlocked_error;
 439
 440         if (fflags & O_FBLOCKING)
 441                 nbio = 0;
 442         else if (fflags & O_FNONBLOCKING)
 443                 nbio = 1;
 444         else if (fp->f_flag & O_NONBLOCK)
 445                 nbio = 1;
 446         else
 447                 nbio = 0;
 448
 449         while (uio->uio_resid) {
 450                 caddr_t va;
 451
 452                 if (rpipe->pipe_buffer.cnt > 0) {
 453                         /*
 454                          * normal pipe buffer receive
 455                          */
 456                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 457                         if (size > rpipe->pipe_buffer.cnt)
 458                                 size = rpipe->pipe_buffer.cnt;
 459                         if (size > (u_int) uio->uio_resid)
 460                                 size = (u_int) uio->uio_resid;
 461
 462                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 463                                         size, uio);
 464                         if (error)
 465                                 break;
 466
 467                         rpipe->pipe_buffer.out += size;
 468                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 469                                 rpipe->pipe_buffer.out = 0;
 470
 471                         rpipe->pipe_buffer.cnt -= size;
 472
 473                         /*
 474                          * If there is no more to read in the pipe, reset
 475                          * its pointers to the beginning.  This improves
 476                          * cache hit stats.
 477                          */
 478                         if (rpipe->pipe_buffer.cnt == 0) {
 479                                 rpipe->pipe_buffer.in = 0;
 480                                 rpipe->pipe_buffer.out = 0;
 481                         }
 482                         nread += size;
 483 #ifndef PIPE_NODIRECT
 484                 } else if (rpipe->pipe_kva &&
 485                            rpipe->pipe_feature == PIPE_KMEM &&
 486                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 487                                == PIPE_DIRECTW
 488                 ) {
 489                         /*
 490                          * Direct copy using source-side kva mapping
 491                          */
 492                         size = rpipe->pipe_map.xio_bytes -
 493                                 rpipe->pipe_buffer.out;
 494                         if (size > (u_int)uio->uio_resid)
 495                                 size = (u_int)uio->uio_resid;
 496                         va = (caddr_t)rpipe->pipe_kva +
 497                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 498                         error = uiomove(va, size, uio);
 499                         if (error)
 500                                 break;
 501                         nread += size;
 502                         rpipe->pipe_buffer.out += size;
 503                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 504                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 505                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 506                                 /* reset out index for copy mode */
 507                                 rpipe->pipe_buffer.out = 0;
 508                                 wakeup(rpipe);
 509                         }
 510                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 511                            rpipe->pipe_kva &&
 512                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 513                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 514                                == PIPE_DIRECTW
 515                 ) {
 516                         /*
 517                          * Direct copy, bypassing a kernel buffer.  We cannot
 518                          * mess with the direct-write buffer until
 519                          * PIPE_DIRECTIP is cleared.  In order to prevent
 520                          * the pipe_write code from racing itself in
 521                          * direct_write, we set DIRECTIP when we clear
 522                          * DIRECTW after we have exhausted the buffer.
 523                          */
 524                         if (pipe_dwrite_sfbuf == 3)
 525                                 rpipe->pipe_kvamask = 0;
 526                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 527                                     rpipe->pipe_map.xio_npages,
 528                                     &rpipe->pipe_kvamask);
 529                         size = rpipe->pipe_map.xio_bytes -
 530                                 rpipe->pipe_buffer.out;
 531                         if (size > (u_int)uio->uio_resid)
 532                                 size = (u_int)uio->uio_resid;
 533                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 534                         error = uiomove(va, size, uio);
 535                         if (error)
 536                                 break;
 537                         nread += size;
 538                         rpipe->pipe_buffer.out += size;
 539                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 540                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 541                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 542                                 /* reset out index for copy mode */
 543                                 rpipe->pipe_buffer.out = 0;
 544                                 wakeup(rpipe);
 545                         }
 546                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 547                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 548                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 549                                 == PIPE_DIRECTW
 550                 ) {
 551                         /*
 552                          * Direct copy, bypassing a kernel buffer.  We cannot
 553                          * mess with the direct-write buffer until
 554                          * PIPE_DIRECTIP is cleared.  In order to prevent
 555                          * the pipe_write code from racing itself in
 556                          * direct_write, we set DIRECTIP when we clear
 557                          * DIRECTW after we have exhausted the buffer.
 558                          */
 559                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 560                         if (error)
 561                                 break;
 562                         nread += size;
 563                         rpipe->pipe_buffer.out += size;
 564                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 565                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 566                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 567                                 /* reset out index for copy mode */
 568                                 rpipe->pipe_buffer.out = 0;
 569                                 wakeup(rpipe);
 570                         }
 571 #endif
 572                 } else {
 573                         /*
 574                          * detect EOF condition
 575                          * read returns 0 on EOF, no need to set error
 576                          */
 577                         if (rpipe->pipe_state & PIPE_EOF)
 578                                 break;
 579
 580                         /*
 581                          * If the "write-side" has been blocked, wake it up now.
 582                          */
 583                         if (rpipe->pipe_state & PIPE_WANTW) {
 584                                 rpipe->pipe_state &= ~PIPE_WANTW;
 585                                 wakeup(rpipe);
 586                         }
 587
 588                         /*
 589                          * Break if some data was read.
 590                          */
 591                         if (nread > 0)
 592                                 break;
 593
 594                         /*
 595                          * Unlock the pipe buffer for our remaining
 596                          * processing.  We will either break out with an
 597                          * error or we will sleep and relock to loop.
 598                          */
 599                         pipeunlock(rpipe);
 600
 601                         /*
 602                          * Handle non-blocking mode operation or
 603                          * wait for more data.
 604                          */
 605                         if (nbio) {
 606                                 error = EAGAIN;
 607                         } else {
 608                                 rpipe->pipe_state |= PIPE_WANTR;
 609                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 610                                     "piperd", 0)) == 0) {
 611                                         error = pipelock(rpipe, 1);
 612                                 }
 613                         }
 614                         if (error)
 615                                 goto unlocked_error;
 616                 }
 617         }
 618         pipeunlock(rpipe);
 619
 620         if (error == 0)
 621                 vfs_timestamp(&rpipe->pipe_atime);
 622 unlocked_error:
 623         --rpipe->pipe_busy;
 624
 625         /*
 626          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 627          */
 628         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 629                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 630                 wakeup(rpipe);
 631         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 632                 /*
 633                  * Handle write blocking hysteresis.
 634                  */
 635                 if (rpipe->pipe_state & PIPE_WANTW) {
 636                         rpipe->pipe_state &= ~PIPE_WANTW;
 637                         wakeup(rpipe);
 638                 }
 639         }
 640
 641         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 642                 pipeselwakeup(rpipe);
 643         rel_mplock();
 644         return (error);
 645 }
 646
 647 #ifndef PIPE_NODIRECT
 648 /*
 649  * Map the sending processes' buffer into kernel space and wire it.
 650  * This is similar to a physical write operation.
 651  */
 652 static int
 653 pipe_build_write_buffer(wpipe, uio)
 654         struct pipe *wpipe;
 655         struct uio *uio;
 656 {
 657         int error;
 658         u_int size;
 659
 660         size = (u_int) uio->uio_iov->iov_len;
 661         if (size > wpipe->pipe_buffer.size)
 662                 size = wpipe->pipe_buffer.size;
 663
 664         if (uio->uio_segflg == UIO_SYSSPACE) {
 665                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 666                                         size);
 667         } else {
 668                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 669                                         size, XIOF_READ);
 670         }
 671         wpipe->pipe_buffer.out = 0;
 672         if (error)
 673                 return(error);
 674
 675         /*
 676          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 677          * map the pages on the target while KMEM maps the pages now.
 678          */
 679         switch(wpipe->pipe_feature) {
 680         case PIPE_KMEM:
 681         case PIPE_SFBUF2:
 682                 if (wpipe->pipe_kva == NULL) {
 683                         wpipe->pipe_kva =
 684                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
 685                         wpipe->pipe_kvamask = 0;
 686                 }
 687                 if (wpipe->pipe_feature == PIPE_KMEM) {
 688                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 689                                     wpipe->pipe_map.xio_npages);
 690                 }
 691                 break;
 692         default:
 693                 break;
 694         }
 695
 696         /*
 697          * And update the uio data.  The XIO might have loaded fewer bytes
 698          * then requested so reload 'size'.
 699          */
 700         size = wpipe->pipe_map.xio_bytes;
 701         uio->uio_iov->iov_len -= size;
 702         uio->uio_iov->iov_base += size;
 703         if (uio->uio_iov->iov_len == 0)
 704                 uio->uio_iov++;
 705         uio->uio_resid -= size;
 706         uio->uio_offset += size;
 707         return (0);
 708 }
 709
 710 /*
 711  * In the case of a signal, the writing process might go away.  This
 712  * code copies the data into the circular buffer so that the source
 713  * pages can be freed without loss of data.
 714  *
 715  * Note that in direct mode pipe_buffer.out is used to track the
 716  * XIO offset.  We are converting the direct mode into buffered mode
 717  * which changes the meaning of pipe_buffer.out.
 718  */
 719 static void
 720 pipe_clone_write_buffer(wpipe)
 721         struct pipe *wpipe;
 722 {
 723         int size;
 724         int offset;
 725
 726         offset = wpipe->pipe_buffer.out;
 727         size = wpipe->pipe_map.xio_bytes - offset;
 728
 729         KKASSERT(size <= wpipe->pipe_buffer.size);
 730
 731         wpipe->pipe_buffer.in = size;
 732         wpipe->pipe_buffer.out = 0;
 733         wpipe->pipe_buffer.cnt = size;
 734         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 735
 736         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 737         xio_release(&wpipe->pipe_map);
 738         if (wpipe->pipe_kva) {
 739                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 740                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 741                 wpipe->pipe_kva = NULL;
 742         }
 743 }
 744
 745 /*
 746  * This implements the pipe buffer write mechanism.  Note that only
 747  * a direct write OR a normal pipe write can be pending at any given time.
 748  * If there are any characters in the pipe buffer, the direct write will
 749  * be deferred until the receiving process grabs all of the bytes from
 750  * the pipe buffer.  Then the direct mapping write is set-up.
 751  */
 752 static int
 753 pipe_direct_write(wpipe, uio)
 754         struct pipe *wpipe;
 755         struct uio *uio;
 756 {
 757         int error;
 758
 759 retry:
 760         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 761                 if (wpipe->pipe_state & PIPE_WANTR) {
 762                         wpipe->pipe_state &= ~PIPE_WANTR;
 763                         wakeup(wpipe);
 764                 }
 765                 wpipe->pipe_state |= PIPE_WANTW;
 766                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 767                 if (error)
 768                         goto error2;
 769                 if (wpipe->pipe_state & PIPE_EOF) {
 770                         error = EPIPE;
 771                         goto error2;
 772                 }
 773         }
 774         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 775         if (wpipe->pipe_buffer.cnt > 0) {
 776                 if (wpipe->pipe_state & PIPE_WANTR) {
 777                         wpipe->pipe_state &= ~PIPE_WANTR;
 778                         wakeup(wpipe);
 779                 }
 780
 781                 wpipe->pipe_state |= PIPE_WANTW;
 782                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 783                 if (error)
 784                         goto error2;
 785                 if (wpipe->pipe_state & PIPE_EOF) {
 786                         error = EPIPE;
 787                         goto error2;
 788                 }
 789                 goto retry;
 790         }
 791
 792         /*
 793          * Build our direct-write buffer
 794          */
 795         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 796         error = pipe_build_write_buffer(wpipe, uio);
 797         if (error)
 798                 goto error1;
 799         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 800
 801         /*
 802          * Wait until the receiver has snarfed the data.  Since we are likely
 803          * going to sleep we optimize the case and yield synchronously,
 804          * possibly avoiding the tsleep().
 805          */
 806         error = 0;
 807         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 808                 if (wpipe->pipe_state & PIPE_EOF) {
 809                         pipelock(wpipe, 0);
 810                         xio_release(&wpipe->pipe_map);
 811                         if (wpipe->pipe_kva) {
 812                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 813                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 814                                 wpipe->pipe_kva = NULL;
 815                         }
 816                         pipeunlock(wpipe);
 817                         pipeselwakeup(wpipe);
 818                         error = EPIPE;
 819                         goto error1;
 820                 }
 821                 if (wpipe->pipe_state & PIPE_WANTR) {
 822                         wpipe->pipe_state &= ~PIPE_WANTR;
 823                         wakeup(wpipe);
 824                 }
 825                 pipeselwakeup(wpipe);
 826                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 827         }
 828         pipelock(wpipe,0);
 829         if (wpipe->pipe_state & PIPE_DIRECTW) {
 830                 /*
 831                  * this bit of trickery substitutes a kernel buffer for
 832                  * the process that might be going away.
 833                  */
 834                 pipe_clone_write_buffer(wpipe);
 835                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 836         } else {
 837                 /*
 838                  * note: The pipe_kva mapping is not qremove'd here.  For
 839                  * legacy PIPE_KMEM mode this constitutes an improvement
 840                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 841                  * mode the kva mapping must not be removed to get the
 842                  * caching benefit.
 843                  *
 844                  * For testing purposes we will give the original algorithm
 845                  * the benefit of the doubt 'what it could have been', and
 846                  * keep the optimization.
 847                  */
 848                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 849                 xio_release(&wpipe->pipe_map);
 850                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 851         }
 852         pipeunlock(wpipe);
 853         return (error);
 854
 855         /*
 856          * Direct-write error, clear the direct write flags.
 857          */
 858 error1:
 859         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 860         /* fallthrough */
 861
 862         /*
 863          * General error, wakeup the other side if it happens to be sleeping.
 864          */
 865 error2:
 866         wakeup(wpipe);
 867         return (error);
 868 }
 869 #endif
 870
 871 /*
 872  * MPALMOSTSAFE - acquires mplock
 873  */
 874 static int
 875 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 876 {
 877         int error = 0;
 878         int orig_resid;
 879         int nbio;
 880         struct pipe *wpipe, *rpipe;
 881
 882         get_mplock();
 883         rpipe = (struct pipe *) fp->f_data;
 884         wpipe = rpipe->pipe_peer;
 885
 886         /*
 887          * detect loss of pipe read side, issue SIGPIPE if lost.
 888          */
 889         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 890                 rel_mplock();
 891                 return (EPIPE);
 892         }
 893         ++wpipe->pipe_busy;
 894
 895         if (fflags & O_FBLOCKING)
 896                 nbio = 0;
 897         else if (fflags & O_FNONBLOCKING)
 898                 nbio = 1;
 899         else if (fp->f_flag & O_NONBLOCK)
 900                 nbio = 1;
 901         else
 902                 nbio = 0;
 903
 904         /*
 905          * If it is advantageous to resize the pipe buffer, do
 906          * so.
 907          */
 908         if ((uio->uio_resid > PIPE_SIZE) &&
 909                 (pipe_nbig < pipe_maxbig) &&
 910                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 911                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 912                 (wpipe->pipe_buffer.cnt == 0)) {
 913
 914                 if ((error = pipelock(wpipe,1)) == 0) {
 915                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 916                                 pipe_nbig++;
 917                         pipeunlock(wpipe);
 918                 }
 919         }
 920
 921         /*
 922          * If an early error occured unbusy and return, waking up any pending
 923          * readers.
 924          */
 925         if (error) {
 926                 --wpipe->pipe_busy;
 927                 if ((wpipe->pipe_busy == 0) &&
 928                     (wpipe->pipe_state & PIPE_WANT)) {
 929                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 930                         wakeup(wpipe);
 931                 }
 932                 rel_mplock();
 933                 return(error);
 934         }
 935
 936         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 937
 938         orig_resid = uio->uio_resid;
 939
 940         while (uio->uio_resid) {
 941                 int space;
 942
 943 #ifndef PIPE_NODIRECT
 944                 /*
 945                  * If the transfer is large, we can gain performance if
 946                  * we do process-to-process copies directly.
 947                  * If the write is non-blocking, we don't use the
 948                  * direct write mechanism.
 949                  *
 950                  * The direct write mechanism will detect the reader going
 951                  * away on us.
 952                  */
 953                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 954                     pipe_dwrite_enable > 1) &&
 955                     nbio == 0 &&
 956                     pipe_dwrite_enable) {
 957                         error = pipe_direct_write( wpipe, uio);
 958                         if (error)
 959                                 break;
 960                         continue;
 961                 }
 962 #endif
 963
 964                 /*
 965                  * Pipe buffered writes cannot be coincidental with
 966                  * direct writes.  We wait until the currently executing
 967                  * direct write is completed before we start filling the
 968                  * pipe buffer.  We break out if a signal occurs or the
 969                  * reader goes away.
 970                  */
 971         retrywrite:
 972                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 973                         if (wpipe->pipe_state & PIPE_WANTR) {
 974                                 wpipe->pipe_state &= ~PIPE_WANTR;
 975                                 wakeup(wpipe);
 976                         }
 977                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 978                         if (wpipe->pipe_state & PIPE_EOF)
 979                                 break;
 980                         if (error)
 981                                 break;
 982                 }
 983                 if (wpipe->pipe_state & PIPE_EOF) {
 984                         error = EPIPE;
 985                         break;
 986                 }
 987
 988                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 989
 990                 /* Writes of size <= PIPE_BUF must be atomic. */
 991                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 992                         space = 0;
 993
 994                 /*
 995                  * Write to fill, read size handles write hysteresis.  Also
 996                  * additional restrictions can cause select-based non-blocking
 997                  * writes to spin.
 998                  */
 999                 if (space > 0) {
1000                         if ((error = pipelock(wpipe,1)) == 0) {
1001                                 int size;       /* Transfer size */
1002                                 int segsize;    /* first segment to transfer */
1003
1004                                 /*
1005                                  * It is possible for a direct write to
1006                                  * slip in on us... handle it here...
1007                                  */
1008                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
1009                                         pipeunlock(wpipe);
1010                                         goto retrywrite;
1011                                 }
1012                                 /*
1013                                  * If a process blocked in uiomove, our
1014                                  * value for space might be bad.
1015                                  *
1016                                  * XXX will we be ok if the reader has gone
1017                                  * away here?
1018                                  */
1019                                 if (space > wpipe->pipe_buffer.size -
1020                                     wpipe->pipe_buffer.cnt) {
1021                                         pipeunlock(wpipe);
1022                                         goto retrywrite;
1023                                 }
1024
1025                                 /*
1026                                  * Transfer size is minimum of uio transfer
1027                                  * and free space in pipe buffer.
1028                                  */
1029                                 if (space > uio->uio_resid)
1030                                         size = uio->uio_resid;
1031                                 else
1032                                         size = space;
1033                                 /*
1034                                  * First segment to transfer is minimum of
1035                                  * transfer size and contiguous space in
1036                                  * pipe buffer.  If first segment to transfer
1037                                  * is less than the transfer size, we've got
1038                                  * a wraparound in the buffer.
1039                                  */
1040                                 segsize = wpipe->pipe_buffer.size -
1041                                         wpipe->pipe_buffer.in;
1042                                 if (segsize > size)
1043                                         segsize = size;
1044
1045                                 /* Transfer first segment */
1046
1047                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1048                                                 segsize, uio);
1049
1050                                 if (error == 0 && segsize < size) {
1051                                         /*
1052                                          * Transfer remaining part now, to
1053                                          * support atomic writes.  Wraparound
1054                                          * happened.
1055                                          */
1056                                         if (wpipe->pipe_buffer.in + segsize !=
1057                                             wpipe->pipe_buffer.size)
1058                                                 panic("Expected pipe buffer wraparound disappeared");
1059
1060                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1061                                                         size - segsize, uio);
1062                                 }
1063                                 if (error == 0) {
1064                                         wpipe->pipe_buffer.in += size;
1065                                         if (wpipe->pipe_buffer.in >=
1066                                             wpipe->pipe_buffer.size) {
1067                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1068                                                         panic("Expected wraparound bad");
1069                                                 wpipe->pipe_buffer.in = size - segsize;
1070                                         }
1071
1072                                         wpipe->pipe_buffer.cnt += size;
1073                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1074                                                 panic("Pipe buffer overflow");
1075
1076                                 }
1077                                 pipeunlock(wpipe);
1078                         }
1079                         if (error)
1080                                 break;
1081
1082                 } else {
1083                         /*
1084                          * If the "read-side" has been blocked, wake it up now
1085                          * and yield to let it drain synchronously rather
1086                          * then block.
1087                          */
1088                         if (wpipe->pipe_state & PIPE_WANTR) {
1089                                 wpipe->pipe_state &= ~PIPE_WANTR;
1090                                 wakeup(wpipe);
1091                         }
1092
1093                         /*
1094                          * don't block on non-blocking I/O
1095                          */
1096                         if (nbio) {
1097                                 error = EAGAIN;
1098                                 break;
1099                         }
1100
1101                         /*
1102                          * We have no more space and have something to offer,
1103                          * wake up select/poll.
1104                          */
1105                         pipeselwakeup(wpipe);
1106
1107                         wpipe->pipe_state |= PIPE_WANTW;
1108                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1109                         if (error != 0)
1110                                 break;
1111                         /*
1112                          * If read side wants to go away, we just issue a signal
1113                          * to ourselves.
1114                          */
1115                         if (wpipe->pipe_state & PIPE_EOF) {
1116                                 error = EPIPE;
1117                                 break;
1118                         }
1119                 }
1120         }
1121
1122         --wpipe->pipe_busy;
1123
1124         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1125                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1126                 wakeup(wpipe);
1127         } else if (wpipe->pipe_buffer.cnt > 0) {
1128                 /*
1129                  * If we have put any characters in the buffer, we wake up
1130                  * the reader.
1131                  */
1132                 if (wpipe->pipe_state & PIPE_WANTR) {
1133                         wpipe->pipe_state &= ~PIPE_WANTR;
1134                         wakeup(wpipe);
1135                 }
1136         }
1137
1138         /*
1139          * Don't return EPIPE if I/O was successful
1140          */
1141         if ((wpipe->pipe_buffer.cnt == 0) &&
1142             (uio->uio_resid == 0) &&
1143             (error == EPIPE)) {
1144                 error = 0;
1145         }
1146
1147         if (error == 0)
1148                 vfs_timestamp(&wpipe->pipe_mtime);
1149
1150         /*
1151          * We have something to offer,
1152          * wake up select/poll.
1153          */
1154         if (wpipe->pipe_buffer.cnt)
1155                 pipeselwakeup(wpipe);
1156         rel_mplock();
1157         return (error);
1158 }
1159
1160 /*
1161  * MPALMOSTSAFE - acquires mplock
1162  *
1163  * we implement a very minimal set of ioctls for compatibility with sockets.
1164  */
1165 int
1166 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1167 {
1168         struct pipe *mpipe;
1169         int error;
1170
1171         get_mplock();
1172         mpipe = (struct pipe *)fp->f_data;
1173
1174         switch (cmd) {
1175         case FIOASYNC:
1176                 if (*(int *)data) {
1177                         mpipe->pipe_state |= PIPE_ASYNC;
1178                 } else {
1179                         mpipe->pipe_state &= ~PIPE_ASYNC;
1180                 }
1181                 error = 0;
1182                 break;
1183         case FIONREAD:
1184                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1185                         *(int *)data = mpipe->pipe_map.xio_bytes -
1186                                         mpipe->pipe_buffer.out;
1187                 } else {
1188                         *(int *)data = mpipe->pipe_buffer.cnt;
1189                 }
1190                 error = 0;
1191                 break;
1192         case FIOSETOWN:
1193                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1194                 break;
1195         case FIOGETOWN:
1196                 *(int *)data = fgetown(mpipe->pipe_sigio);
1197                 error = 0;
1198                 break;
1199         case TIOCSPGRP:
1200                 /* This is deprecated, FIOSETOWN should be used instead. */
1201                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1202                 break;
1203
1204         case TIOCGPGRP:
1205                 /* This is deprecated, FIOGETOWN should be used instead. */
1206                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1207                 error = 0;
1208                 break;
1209         default:
1210                 error = ENOTTY;
1211                 break;
1212         }
1213         rel_mplock();
1214         return (error);
1215 }
1216
1217 /*
1218  * MPALMOSTSAFE - acquires mplock
1219  */
1220 int
1221 pipe_poll(struct file *fp, int events, struct ucred *cred)
1222 {
1223         struct pipe *rpipe;
1224         struct pipe *wpipe;
1225         int revents = 0;
1226
1227         get_mplock();
1228         rpipe = (struct pipe *)fp->f_data;
1229         wpipe = rpipe->pipe_peer;
1230         if (events & (POLLIN | POLLRDNORM))
1231                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1232                     (rpipe->pipe_buffer.cnt > 0) ||
1233                     (rpipe->pipe_state & PIPE_EOF))
1234                         revents |= events & (POLLIN | POLLRDNORM);
1235
1236         if (events & (POLLOUT | POLLWRNORM))
1237                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1238                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1239                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1240                         revents |= events & (POLLOUT | POLLWRNORM);
1241
1242         if ((rpipe->pipe_state & PIPE_EOF) ||
1243             (wpipe == NULL) ||
1244             (wpipe->pipe_state & PIPE_EOF))
1245                 revents |= POLLHUP;
1246
1247         if (revents == 0) {
1248                 if (events & (POLLIN | POLLRDNORM)) {
1249                         selrecord(curthread, &rpipe->pipe_sel);
1250                         rpipe->pipe_state |= PIPE_SEL;
1251                 }
1252
1253                 if (events & (POLLOUT | POLLWRNORM)) {
1254                         selrecord(curthread, &wpipe->pipe_sel);
1255                         wpipe->pipe_state |= PIPE_SEL;
1256                 }
1257         }
1258         rel_mplock();
1259         return (revents);
1260 }
1261
1262 /*
1263  * MPALMOSTSAFE - acquires mplock
1264  */
1265 static int
1266 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1267 {
1268         struct pipe *pipe;
1269
1270         get_mplock();
1271         pipe = (struct pipe *)fp->f_data;
1272
1273         bzero((caddr_t)ub, sizeof(*ub));
1274         ub->st_mode = S_IFIFO;
1275         ub->st_blksize = pipe->pipe_buffer.size;
1276         ub->st_size = pipe->pipe_buffer.cnt;
1277         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1278                 ub->st_size = pipe->pipe_map.xio_bytes -
1279                                 pipe->pipe_buffer.out;
1280         }
1281         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1282         ub->st_atimespec = pipe->pipe_atime;
1283         ub->st_mtimespec = pipe->pipe_mtime;
1284         ub->st_ctimespec = pipe->pipe_ctime;
1285         /*
1286          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1287          * st_flags, st_gen.
1288          * XXX (st_dev, st_ino) should be unique.
1289          */
1290         rel_mplock();
1291         return (0);
1292 }
1293
1294 /*
1295  * MPALMOSTSAFE - acquires mplock
1296  */
1297 static int
1298 pipe_close(struct file *fp)
1299 {
1300         struct pipe *cpipe = (struct pipe *)fp->f_data;
1301
1302         get_mplock();
1303         fp->f_ops = &badfileops;
1304         fp->f_data = NULL;
1305         funsetown(cpipe->pipe_sigio);
1306         pipeclose(cpipe);
1307         rel_mplock();
1308         return (0);
1309 }
1310
1311 /*
1312  * Shutdown one or both directions of a full-duplex pipe.
1313  *
1314  * MPALMOSTSAFE - acquires mplock
1315  */
1316 static int
1317 pipe_shutdown(struct file *fp, int how)
1318 {
1319         struct pipe *rpipe;
1320         struct pipe *wpipe;
1321         int error = EPIPE;
1322
1323         get_mplock();
1324         rpipe = (struct pipe *)fp->f_data;
1325
1326         switch(how) {
1327         case SHUT_RDWR:
1328         case SHUT_RD:
1329                 if (rpipe) {
1330                         rpipe->pipe_state |= PIPE_EOF;
1331                         pipeselwakeup(rpipe);
1332                         if (rpipe->pipe_busy)
1333                                 wakeup(rpipe);
1334                         error = 0;
1335                 }
1336                 if (how == SHUT_RD)
1337                         break;
1338                 /* fall through */
1339         case SHUT_WR:
1340                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1341                         wpipe->pipe_state |= PIPE_EOF;
1342                         pipeselwakeup(wpipe);
1343                         if (wpipe->pipe_busy)
1344                                 wakeup(wpipe);
1345                         error = 0;
1346                 }
1347         }
1348         rel_mplock();
1349         return (error);
1350 }
1351
1352 static void
1353 pipe_free_kmem(struct pipe *cpipe)
1354 {
1355         if (cpipe->pipe_buffer.buffer != NULL) {
1356                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1357                         --pipe_nbig;
1358                 kmem_free(kernel_map,
1359                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1360                         cpipe->pipe_buffer.size);
1361                 cpipe->pipe_buffer.buffer = NULL;
1362                 cpipe->pipe_buffer.object = NULL;
1363         }
1364 #ifndef PIPE_NODIRECT
1365         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1366                 cpipe->pipe_map.xio_offset == 0 &&
1367                 cpipe->pipe_map.xio_npages == 0);
1368 #endif
1369 }
1370
1371 /*
1372  * shutdown the pipe
1373  */
1374 static void
1375 pipeclose(struct pipe *cpipe)
1376 {
1377         globaldata_t gd;
1378         struct pipe *ppipe;
1379
1380         if (cpipe == NULL)
1381                 return;
1382
1383         pipeselwakeup(cpipe);
1384
1385         /*
1386          * If the other side is blocked, wake it up saying that
1387          * we want to close it down.
1388          */
1389         while (cpipe->pipe_busy) {
1390                 wakeup(cpipe);
1391                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1392                 tsleep(cpipe, 0, "pipecl", 0);
1393         }
1394
1395         /*
1396          * Disconnect from peer
1397          */
1398         if ((ppipe = cpipe->pipe_peer) != NULL) {
1399                 pipeselwakeup(ppipe);
1400
1401                 ppipe->pipe_state |= PIPE_EOF;
1402                 wakeup(ppipe);
1403                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1404                 ppipe->pipe_peer = NULL;
1405         }
1406
1407         if (cpipe->pipe_kva) {
1408                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1409                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1410                 cpipe->pipe_kva = NULL;
1411         }
1412
1413         /*
1414          * free or cache resources
1415          */
1416         gd = mycpu;
1417         if (gd->gd_pipeqcount >= pipe_maxcache ||
1418             cpipe->pipe_buffer.size != PIPE_SIZE
1419         ) {
1420                 pipe_free_kmem(cpipe);
1421                 free(cpipe, M_PIPE);
1422         } else {
1423                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1424                         cpipe->pipe_map.xio_bytes == 0 &&
1425                         cpipe->pipe_map.xio_offset == 0);
1426                 cpipe->pipe_state = 0;
1427                 cpipe->pipe_busy = 0;
1428                 cpipe->pipe_peer = gd->gd_pipeq;
1429                 gd->gd_pipeq = cpipe;
1430                 ++gd->gd_pipeqcount;
1431         }
1432 }
1433
1434 /*
1435  * MPALMOSTSAFE - acquires mplock
1436  */
1437 static int
1438 pipe_kqfilter(struct file *fp, struct knote *kn)
1439 {
1440         struct pipe *cpipe;
1441
1442         get_mplock();
1443         cpipe = (struct pipe *)kn->kn_fp->f_data;
1444
1445         switch (kn->kn_filter) {
1446         case EVFILT_READ:
1447                 kn->kn_fop = &pipe_rfiltops;
1448                 break;
1449         case EVFILT_WRITE:
1450                 kn->kn_fop = &pipe_wfiltops;
1451                 cpipe = cpipe->pipe_peer;
1452                 if (cpipe == NULL) {
1453                         /* other end of pipe has been closed */
1454                         rel_mplock();
1455                         return (EPIPE);
1456                 }
1457                 break;
1458         default:
1459                 return (1);
1460         }
1461         kn->kn_hook = (caddr_t)cpipe;
1462
1463         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1464         rel_mplock();
1465         return (0);
1466 }
1467
1468 static void
1469 filt_pipedetach(struct knote *kn)
1470 {
1471         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1472
1473         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1474 }
1475
1476 /*ARGSUSED*/
1477 static int
1478 filt_piperead(struct knote *kn, long hint)
1479 {
1480         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1481         struct pipe *wpipe = rpipe->pipe_peer;
1482
1483         kn->kn_data = rpipe->pipe_buffer.cnt;
1484         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1485                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1486                                 rpipe->pipe_buffer.out;
1487         }
1488
1489         if ((rpipe->pipe_state & PIPE_EOF) ||
1490             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1491                 kn->kn_flags |= EV_EOF;
1492                 return (1);
1493         }
1494         return (kn->kn_data > 0);
1495 }
1496
1497 /*ARGSUSED*/
1498 static int
1499 filt_pipewrite(struct knote *kn, long hint)
1500 {
1501         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1502         struct pipe *wpipe = rpipe->pipe_peer;
1503
1504         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1505                 kn->kn_data = 0;
1506                 kn->kn_flags |= EV_EOF;
1507                 return (1);
1508         }
1509         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1510         if (wpipe->pipe_state & PIPE_DIRECTW)
1511                 kn->kn_data = 0;
1512
1513         return (kn->kn_data >= PIPE_BUF);
1514 }