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 1.42 2006/09/11 20:25:01 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,
 322                                     1,
 323                                     VM_MAPTYPE_NORMAL,
 324                                     VM_PROT_ALL, VM_PROT_ALL,
 325                                     0);
 326
 327                 if (error != KERN_SUCCESS) {
 328                         vm_object_deallocate(object);
 329                         return (ENOMEM);
 330                 }
 331                 pipe_free_kmem(cpipe);
 332                 cpipe->pipe_buffer.object = object;
 333                 cpipe->pipe_buffer.buffer = buffer;
 334                 cpipe->pipe_buffer.size = size;
 335                 ++pipe_bkmem_alloc;
 336         } else {
 337                 ++pipe_bcache_alloc;
 338         }
 339         cpipe->pipe_buffer.in = 0;
 340         cpipe->pipe_buffer.out = 0;
 341         cpipe->pipe_buffer.cnt = 0;
 342         return (0);
 343 }
 344
 345 /*
 346  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 347  * our per-cpu cache if possible.  For now make sure it is sized for the
 348  * smaller PIPE_SIZE default.
 349  */
 350 static int
 351 pipe_create(cpipep)
 352         struct pipe **cpipep;
 353 {
 354         globaldata_t gd = mycpu;
 355         struct pipe *cpipe;
 356         int error;
 357
 358         if ((cpipe = gd->gd_pipeq) != NULL) {
 359                 gd->gd_pipeq = cpipe->pipe_peer;
 360                 --gd->gd_pipeqcount;
 361                 cpipe->pipe_peer = NULL;
 362         } else {
 363                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 364         }
 365         *cpipep = cpipe;
 366         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 367                 return (error);
 368         vfs_timestamp(&cpipe->pipe_ctime);
 369         cpipe->pipe_atime = cpipe->pipe_ctime;
 370         cpipe->pipe_mtime = cpipe->pipe_ctime;
 371         return (0);
 372 }
 373
 374
 375 /*
 376  * lock a pipe for I/O, blocking other access
 377  */
 378 static __inline int
 379 pipelock(cpipe, catch)
 380         struct pipe *cpipe;
 381         int catch;
 382 {
 383         int error;
 384
 385         while (cpipe->pipe_state & PIPE_LOCK) {
 386                 cpipe->pipe_state |= PIPE_LWANT;
 387                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 388                 if (error != 0)
 389                         return (error);
 390         }
 391         cpipe->pipe_state |= PIPE_LOCK;
 392         return (0);
 393 }
 394
 395 /*
 396  * unlock a pipe I/O lock
 397  */
 398 static __inline void
 399 pipeunlock(cpipe)
 400         struct pipe *cpipe;
 401 {
 402
 403         cpipe->pipe_state &= ~PIPE_LOCK;
 404         if (cpipe->pipe_state & PIPE_LWANT) {
 405                 cpipe->pipe_state &= ~PIPE_LWANT;
 406                 wakeup(cpipe);
 407         }
 408 }
 409
 410 static __inline void
 411 pipeselwakeup(cpipe)
 412         struct pipe *cpipe;
 413 {
 414
 415         if (cpipe->pipe_state & PIPE_SEL) {
 416                 cpipe->pipe_state &= ~PIPE_SEL;
 417                 selwakeup(&cpipe->pipe_sel);
 418         }
 419         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 420                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 421         KNOTE(&cpipe->pipe_sel.si_note, 0);
 422 }
 423
 424 /*
 425  * MPALMOSTSAFE (acquires mplock)
 426  */
 427 static int
 428 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 429 {
 430         struct pipe *rpipe;
 431         int error;
 432         int nread = 0;
 433         int nbio;
 434         u_int size;
 435
 436         get_mplock();
 437         rpipe = (struct pipe *) fp->f_data;
 438         ++rpipe->pipe_busy;
 439         error = pipelock(rpipe, 1);
 440         if (error)
 441                 goto unlocked_error;
 442
 443         if (fflags & O_FBLOCKING)
 444                 nbio = 0;
 445         else if (fflags & O_FNONBLOCKING)
 446                 nbio = 1;
 447         else if (fp->f_flag & O_NONBLOCK)
 448                 nbio = 1;
 449         else
 450                 nbio = 0;
 451
 452         while (uio->uio_resid) {
 453                 caddr_t va;
 454
 455                 if (rpipe->pipe_buffer.cnt > 0) {
 456                         /*
 457                          * normal pipe buffer receive
 458                          */
 459                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 460                         if (size > rpipe->pipe_buffer.cnt)
 461                                 size = rpipe->pipe_buffer.cnt;
 462                         if (size > (u_int) uio->uio_resid)
 463                                 size = (u_int) uio->uio_resid;
 464
 465                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 466                                         size, uio);
 467                         if (error)
 468                                 break;
 469
 470                         rpipe->pipe_buffer.out += size;
 471                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 472                                 rpipe->pipe_buffer.out = 0;
 473
 474                         rpipe->pipe_buffer.cnt -= size;
 475
 476                         /*
 477                          * If there is no more to read in the pipe, reset
 478                          * its pointers to the beginning.  This improves
 479                          * cache hit stats.
 480                          */
 481                         if (rpipe->pipe_buffer.cnt == 0) {
 482                                 rpipe->pipe_buffer.in = 0;
 483                                 rpipe->pipe_buffer.out = 0;
 484                         }
 485                         nread += size;
 486 #ifndef PIPE_NODIRECT
 487                 } else if (rpipe->pipe_kva &&
 488                            rpipe->pipe_feature == PIPE_KMEM &&
 489                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 490                                == PIPE_DIRECTW
 491                 ) {
 492                         /*
 493                          * Direct copy using source-side kva mapping
 494                          */
 495                         size = rpipe->pipe_map.xio_bytes -
 496                                 rpipe->pipe_buffer.out;
 497                         if (size > (u_int)uio->uio_resid)
 498                                 size = (u_int)uio->uio_resid;
 499                         va = (caddr_t)rpipe->pipe_kva +
 500                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 501                         error = uiomove(va, size, uio);
 502                         if (error)
 503                                 break;
 504                         nread += size;
 505                         rpipe->pipe_buffer.out += size;
 506                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 507                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 508                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 509                                 /* reset out index for copy mode */
 510                                 rpipe->pipe_buffer.out = 0;
 511                                 wakeup(rpipe);
 512                         }
 513                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 514                            rpipe->pipe_kva &&
 515                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 516                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 517                                == PIPE_DIRECTW
 518                 ) {
 519                         /*
 520                          * Direct copy, bypassing a kernel buffer.  We cannot
 521                          * mess with the direct-write buffer until
 522                          * PIPE_DIRECTIP is cleared.  In order to prevent
 523                          * the pipe_write code from racing itself in
 524                          * direct_write, we set DIRECTIP when we clear
 525                          * DIRECTW after we have exhausted the buffer.
 526                          */
 527                         if (pipe_dwrite_sfbuf == 3)
 528                                 rpipe->pipe_kvamask = 0;
 529                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 530                                     rpipe->pipe_map.xio_npages,
 531                                     &rpipe->pipe_kvamask);
 532                         size = rpipe->pipe_map.xio_bytes -
 533                                 rpipe->pipe_buffer.out;
 534                         if (size > (u_int)uio->uio_resid)
 535                                 size = (u_int)uio->uio_resid;
 536                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 537                         error = uiomove(va, size, uio);
 538                         if (error)
 539                                 break;
 540                         nread += size;
 541                         rpipe->pipe_buffer.out += size;
 542                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 543                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 544                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 545                                 /* reset out index for copy mode */
 546                                 rpipe->pipe_buffer.out = 0;
 547                                 wakeup(rpipe);
 548                         }
 549                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 550                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 551                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 552                                 == PIPE_DIRECTW
 553                 ) {
 554                         /*
 555                          * Direct copy, bypassing a kernel buffer.  We cannot
 556                          * mess with the direct-write buffer until
 557                          * PIPE_DIRECTIP is cleared.  In order to prevent
 558                          * the pipe_write code from racing itself in
 559                          * direct_write, we set DIRECTIP when we clear
 560                          * DIRECTW after we have exhausted the buffer.
 561                          */
 562                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 563                         if (error)
 564                                 break;
 565                         nread += size;
 566                         rpipe->pipe_buffer.out += size;
 567                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 568                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 569                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 570                                 /* reset out index for copy mode */
 571                                 rpipe->pipe_buffer.out = 0;
 572                                 wakeup(rpipe);
 573                         }
 574 #endif
 575                 } else {
 576                         /*
 577                          * detect EOF condition
 578                          * read returns 0 on EOF, no need to set error
 579                          */
 580                         if (rpipe->pipe_state & PIPE_EOF)
 581                                 break;
 582
 583                         /*
 584                          * If the "write-side" has been blocked, wake it up now.
 585                          */
 586                         if (rpipe->pipe_state & PIPE_WANTW) {
 587                                 rpipe->pipe_state &= ~PIPE_WANTW;
 588                                 wakeup(rpipe);
 589                         }
 590
 591                         /*
 592                          * Break if some data was read.
 593                          */
 594                         if (nread > 0)
 595                                 break;
 596
 597                         /*
 598                          * Unlock the pipe buffer for our remaining
 599                          * processing.  We will either break out with an
 600                          * error or we will sleep and relock to loop.
 601                          */
 602                         pipeunlock(rpipe);
 603
 604                         /*
 605                          * Handle non-blocking mode operation or
 606                          * wait for more data.
 607                          */
 608                         if (nbio) {
 609                                 error = EAGAIN;
 610                         } else {
 611                                 rpipe->pipe_state |= PIPE_WANTR;
 612                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 613                                     "piperd", 0)) == 0) {
 614                                         error = pipelock(rpipe, 1);
 615                                 }
 616                         }
 617                         if (error)
 618                                 goto unlocked_error;
 619                 }
 620         }
 621         pipeunlock(rpipe);
 622
 623         if (error == 0)
 624                 vfs_timestamp(&rpipe->pipe_atime);
 625 unlocked_error:
 626         --rpipe->pipe_busy;
 627
 628         /*
 629          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 630          */
 631         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 632                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 633                 wakeup(rpipe);
 634         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 635                 /*
 636                  * Handle write blocking hysteresis.
 637                  */
 638                 if (rpipe->pipe_state & PIPE_WANTW) {
 639                         rpipe->pipe_state &= ~PIPE_WANTW;
 640                         wakeup(rpipe);
 641                 }
 642         }
 643
 644         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 645                 pipeselwakeup(rpipe);
 646         rel_mplock();
 647         return (error);
 648 }
 649
 650 #ifndef PIPE_NODIRECT
 651 /*
 652  * Map the sending processes' buffer into kernel space and wire it.
 653  * This is similar to a physical write operation.
 654  */
 655 static int
 656 pipe_build_write_buffer(wpipe, uio)
 657         struct pipe *wpipe;
 658         struct uio *uio;
 659 {
 660         int error;
 661         u_int size;
 662
 663         size = (u_int) uio->uio_iov->iov_len;
 664         if (size > wpipe->pipe_buffer.size)
 665                 size = wpipe->pipe_buffer.size;
 666
 667         if (uio->uio_segflg == UIO_SYSSPACE) {
 668                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 669                                         size);
 670         } else {
 671                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 672                                         size, XIOF_READ);
 673         }
 674         wpipe->pipe_buffer.out = 0;
 675         if (error)
 676                 return(error);
 677
 678         /*
 679          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 680          * map the pages on the target while KMEM maps the pages now.
 681          */
 682         switch(wpipe->pipe_feature) {
 683         case PIPE_KMEM:
 684         case PIPE_SFBUF2:
 685                 if (wpipe->pipe_kva == NULL) {
 686                         wpipe->pipe_kva =
 687                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
 688                         wpipe->pipe_kvamask = 0;
 689                 }
 690                 if (wpipe->pipe_feature == PIPE_KMEM) {
 691                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 692                                     wpipe->pipe_map.xio_npages);
 693                 }
 694                 break;
 695         default:
 696                 break;
 697         }
 698
 699         /*
 700          * And update the uio data.  The XIO might have loaded fewer bytes
 701          * then requested so reload 'size'.
 702          */
 703         size = wpipe->pipe_map.xio_bytes;
 704         uio->uio_iov->iov_len -= size;
 705         uio->uio_iov->iov_base += size;
 706         if (uio->uio_iov->iov_len == 0)
 707                 uio->uio_iov++;
 708         uio->uio_resid -= size;
 709         uio->uio_offset += size;
 710         return (0);
 711 }
 712
 713 /*
 714  * In the case of a signal, the writing process might go away.  This
 715  * code copies the data into the circular buffer so that the source
 716  * pages can be freed without loss of data.
 717  *
 718  * Note that in direct mode pipe_buffer.out is used to track the
 719  * XIO offset.  We are converting the direct mode into buffered mode
 720  * which changes the meaning of pipe_buffer.out.
 721  */
 722 static void
 723 pipe_clone_write_buffer(wpipe)
 724         struct pipe *wpipe;
 725 {
 726         int size;
 727         int offset;
 728
 729         offset = wpipe->pipe_buffer.out;
 730         size = wpipe->pipe_map.xio_bytes - offset;
 731
 732         KKASSERT(size <= wpipe->pipe_buffer.size);
 733
 734         wpipe->pipe_buffer.in = size;
 735         wpipe->pipe_buffer.out = 0;
 736         wpipe->pipe_buffer.cnt = size;
 737         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 738
 739         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 740         xio_release(&wpipe->pipe_map);
 741         if (wpipe->pipe_kva) {
 742                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 743                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 744                 wpipe->pipe_kva = NULL;
 745         }
 746 }
 747
 748 /*
 749  * This implements the pipe buffer write mechanism.  Note that only
 750  * a direct write OR a normal pipe write can be pending at any given time.
 751  * If there are any characters in the pipe buffer, the direct write will
 752  * be deferred until the receiving process grabs all of the bytes from
 753  * the pipe buffer.  Then the direct mapping write is set-up.
 754  */
 755 static int
 756 pipe_direct_write(wpipe, uio)
 757         struct pipe *wpipe;
 758         struct uio *uio;
 759 {
 760         int error;
 761
 762 retry:
 763         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 764                 if (wpipe->pipe_state & PIPE_WANTR) {
 765                         wpipe->pipe_state &= ~PIPE_WANTR;
 766                         wakeup(wpipe);
 767                 }
 768                 wpipe->pipe_state |= PIPE_WANTW;
 769                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 770                 if (error)
 771                         goto error2;
 772                 if (wpipe->pipe_state & PIPE_EOF) {
 773                         error = EPIPE;
 774                         goto error2;
 775                 }
 776         }
 777         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 778         if (wpipe->pipe_buffer.cnt > 0) {
 779                 if (wpipe->pipe_state & PIPE_WANTR) {
 780                         wpipe->pipe_state &= ~PIPE_WANTR;
 781                         wakeup(wpipe);
 782                 }
 783
 784                 wpipe->pipe_state |= PIPE_WANTW;
 785                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 786                 if (error)
 787                         goto error2;
 788                 if (wpipe->pipe_state & PIPE_EOF) {
 789                         error = EPIPE;
 790                         goto error2;
 791                 }
 792                 goto retry;
 793         }
 794
 795         /*
 796          * Build our direct-write buffer
 797          */
 798         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 799         error = pipe_build_write_buffer(wpipe, uio);
 800         if (error)
 801                 goto error1;
 802         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 803
 804         /*
 805          * Wait until the receiver has snarfed the data.  Since we are likely
 806          * going to sleep we optimize the case and yield synchronously,
 807          * possibly avoiding the tsleep().
 808          */
 809         error = 0;
 810         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 811                 if (wpipe->pipe_state & PIPE_EOF) {
 812                         pipelock(wpipe, 0);
 813                         xio_release(&wpipe->pipe_map);
 814                         if (wpipe->pipe_kva) {
 815                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 816                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 817                                 wpipe->pipe_kva = NULL;
 818                         }
 819                         pipeunlock(wpipe);
 820                         pipeselwakeup(wpipe);
 821                         error = EPIPE;
 822                         goto error1;
 823                 }
 824                 if (wpipe->pipe_state & PIPE_WANTR) {
 825                         wpipe->pipe_state &= ~PIPE_WANTR;
 826                         wakeup(wpipe);
 827                 }
 828                 pipeselwakeup(wpipe);
 829                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 830         }
 831         pipelock(wpipe,0);
 832         if (wpipe->pipe_state & PIPE_DIRECTW) {
 833                 /*
 834                  * this bit of trickery substitutes a kernel buffer for
 835                  * the process that might be going away.
 836                  */
 837                 pipe_clone_write_buffer(wpipe);
 838                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 839         } else {
 840                 /*
 841                  * note: The pipe_kva mapping is not qremove'd here.  For
 842                  * legacy PIPE_KMEM mode this constitutes an improvement
 843                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 844                  * mode the kva mapping must not be removed to get the
 845                  * caching benefit.
 846                  *
 847                  * For testing purposes we will give the original algorithm
 848                  * the benefit of the doubt 'what it could have been', and
 849                  * keep the optimization.
 850                  */
 851                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 852                 xio_release(&wpipe->pipe_map);
 853                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 854         }
 855         pipeunlock(wpipe);
 856         return (error);
 857
 858         /*
 859          * Direct-write error, clear the direct write flags.
 860          */
 861 error1:
 862         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 863         /* fallthrough */
 864
 865         /*
 866          * General error, wakeup the other side if it happens to be sleeping.
 867          */
 868 error2:
 869         wakeup(wpipe);
 870         return (error);
 871 }
 872 #endif
 873
 874 /*
 875  * MPALMOSTSAFE - acquires mplock
 876  */
 877 static int
 878 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 879 {
 880         int error = 0;
 881         int orig_resid;
 882         int nbio;
 883         struct pipe *wpipe, *rpipe;
 884
 885         get_mplock();
 886         rpipe = (struct pipe *) fp->f_data;
 887         wpipe = rpipe->pipe_peer;
 888
 889         /*
 890          * detect loss of pipe read side, issue SIGPIPE if lost.
 891          */
 892         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 893                 rel_mplock();
 894                 return (EPIPE);
 895         }
 896         ++wpipe->pipe_busy;
 897
 898         if (fflags & O_FBLOCKING)
 899                 nbio = 0;
 900         else if (fflags & O_FNONBLOCKING)
 901                 nbio = 1;
 902         else if (fp->f_flag & O_NONBLOCK)
 903                 nbio = 1;
 904         else
 905                 nbio = 0;
 906
 907         /*
 908          * If it is advantageous to resize the pipe buffer, do
 909          * so.
 910          */
 911         if ((uio->uio_resid > PIPE_SIZE) &&
 912                 (pipe_nbig < pipe_maxbig) &&
 913                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 914                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 915                 (wpipe->pipe_buffer.cnt == 0)) {
 916
 917                 if ((error = pipelock(wpipe,1)) == 0) {
 918                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 919                                 pipe_nbig++;
 920                         pipeunlock(wpipe);
 921                 }
 922         }
 923
 924         /*
 925          * If an early error occured unbusy and return, waking up any pending
 926          * readers.
 927          */
 928         if (error) {
 929                 --wpipe->pipe_busy;
 930                 if ((wpipe->pipe_busy == 0) &&
 931                     (wpipe->pipe_state & PIPE_WANT)) {
 932                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 933                         wakeup(wpipe);
 934                 }
 935                 rel_mplock();
 936                 return(error);
 937         }
 938
 939         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 940
 941         orig_resid = uio->uio_resid;
 942
 943         while (uio->uio_resid) {
 944                 int space;
 945
 946 #ifndef PIPE_NODIRECT
 947                 /*
 948                  * If the transfer is large, we can gain performance if
 949                  * we do process-to-process copies directly.
 950                  * If the write is non-blocking, we don't use the
 951                  * direct write mechanism.
 952                  *
 953                  * The direct write mechanism will detect the reader going
 954                  * away on us.
 955                  */
 956                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 957                     pipe_dwrite_enable > 1) &&
 958                     nbio == 0 &&
 959                     pipe_dwrite_enable) {
 960                         error = pipe_direct_write( wpipe, uio);
 961                         if (error)
 962                                 break;
 963                         continue;
 964                 }
 965 #endif
 966
 967                 /*
 968                  * Pipe buffered writes cannot be coincidental with
 969                  * direct writes.  We wait until the currently executing
 970                  * direct write is completed before we start filling the
 971                  * pipe buffer.  We break out if a signal occurs or the
 972                  * reader goes away.
 973                  */
 974         retrywrite:
 975                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 976                         if (wpipe->pipe_state & PIPE_WANTR) {
 977                                 wpipe->pipe_state &= ~PIPE_WANTR;
 978                                 wakeup(wpipe);
 979                         }
 980                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 981                         if (wpipe->pipe_state & PIPE_EOF)
 982                                 break;
 983                         if (error)
 984                                 break;
 985                 }
 986                 if (wpipe->pipe_state & PIPE_EOF) {
 987                         error = EPIPE;
 988                         break;
 989                 }
 990
 991                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 992
 993                 /* Writes of size <= PIPE_BUF must be atomic. */
 994                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 995                         space = 0;
 996
 997                 /*
 998                  * Write to fill, read size handles write hysteresis.  Also
 999                  * additional restrictions can cause select-based non-blocking
1000                  * writes to spin.
1001                  */
1002                 if (space > 0) {
1003                         if ((error = pipelock(wpipe,1)) == 0) {
1004                                 int size;       /* Transfer size */
1005                                 int segsize;    /* first segment to transfer */
1006
1007                                 /*
1008                                  * It is possible for a direct write to
1009                                  * slip in on us... handle it here...
1010                                  */
1011                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
1012                                         pipeunlock(wpipe);
1013                                         goto retrywrite;
1014                                 }
1015                                 /*
1016                                  * If a process blocked in uiomove, our
1017                                  * value for space might be bad.
1018                                  *
1019                                  * XXX will we be ok if the reader has gone
1020                                  * away here?
1021                                  */
1022                                 if (space > wpipe->pipe_buffer.size -
1023                                     wpipe->pipe_buffer.cnt) {
1024                                         pipeunlock(wpipe);
1025                                         goto retrywrite;
1026                                 }
1027
1028                                 /*
1029                                  * Transfer size is minimum of uio transfer
1030                                  * and free space in pipe buffer.
1031                                  */
1032                                 if (space > uio->uio_resid)
1033                                         size = uio->uio_resid;
1034                                 else
1035                                         size = space;
1036                                 /*
1037                                  * First segment to transfer is minimum of
1038                                  * transfer size and contiguous space in
1039                                  * pipe buffer.  If first segment to transfer
1040                                  * is less than the transfer size, we've got
1041                                  * a wraparound in the buffer.
1042                                  */
1043                                 segsize = wpipe->pipe_buffer.size -
1044                                         wpipe->pipe_buffer.in;
1045                                 if (segsize > size)
1046                                         segsize = size;
1047
1048                                 /* Transfer first segment */
1049
1050                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1051                                                 segsize, uio);
1052
1053                                 if (error == 0 && segsize < size) {
1054                                         /*
1055                                          * Transfer remaining part now, to
1056                                          * support atomic writes.  Wraparound
1057                                          * happened.
1058                                          */
1059                                         if (wpipe->pipe_buffer.in + segsize !=
1060                                             wpipe->pipe_buffer.size)
1061                                                 panic("Expected pipe buffer wraparound disappeared");
1062
1063                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1064                                                         size - segsize, uio);
1065                                 }
1066                                 if (error == 0) {
1067                                         wpipe->pipe_buffer.in += size;
1068                                         if (wpipe->pipe_buffer.in >=
1069                                             wpipe->pipe_buffer.size) {
1070                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1071                                                         panic("Expected wraparound bad");
1072                                                 wpipe->pipe_buffer.in = size - segsize;
1073                                         }
1074
1075                                         wpipe->pipe_buffer.cnt += size;
1076                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1077                                                 panic("Pipe buffer overflow");
1078
1079                                 }
1080                                 pipeunlock(wpipe);
1081                         }
1082                         if (error)
1083                                 break;
1084
1085                 } else {
1086                         /*
1087                          * If the "read-side" has been blocked, wake it up now
1088                          * and yield to let it drain synchronously rather
1089                          * then block.
1090                          */
1091                         if (wpipe->pipe_state & PIPE_WANTR) {
1092                                 wpipe->pipe_state &= ~PIPE_WANTR;
1093                                 wakeup(wpipe);
1094                         }
1095
1096                         /*
1097                          * don't block on non-blocking I/O
1098                          */
1099                         if (nbio) {
1100                                 error = EAGAIN;
1101                                 break;
1102                         }
1103
1104                         /*
1105                          * We have no more space and have something to offer,
1106                          * wake up select/poll.
1107                          */
1108                         pipeselwakeup(wpipe);
1109
1110                         wpipe->pipe_state |= PIPE_WANTW;
1111                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1112                         if (error != 0)
1113                                 break;
1114                         /*
1115                          * If read side wants to go away, we just issue a signal
1116                          * to ourselves.
1117                          */
1118                         if (wpipe->pipe_state & PIPE_EOF) {
1119                                 error = EPIPE;
1120                                 break;
1121                         }
1122                 }
1123         }
1124
1125         --wpipe->pipe_busy;
1126
1127         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1128                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1129                 wakeup(wpipe);
1130         } else if (wpipe->pipe_buffer.cnt > 0) {
1131                 /*
1132                  * If we have put any characters in the buffer, we wake up
1133                  * the reader.
1134                  */
1135                 if (wpipe->pipe_state & PIPE_WANTR) {
1136                         wpipe->pipe_state &= ~PIPE_WANTR;
1137                         wakeup(wpipe);
1138                 }
1139         }
1140
1141         /*
1142          * Don't return EPIPE if I/O was successful
1143          */
1144         if ((wpipe->pipe_buffer.cnt == 0) &&
1145             (uio->uio_resid == 0) &&
1146             (error == EPIPE)) {
1147                 error = 0;
1148         }
1149
1150         if (error == 0)
1151                 vfs_timestamp(&wpipe->pipe_mtime);
1152
1153         /*
1154          * We have something to offer,
1155          * wake up select/poll.
1156          */
1157         if (wpipe->pipe_buffer.cnt)
1158                 pipeselwakeup(wpipe);
1159         rel_mplock();
1160         return (error);
1161 }
1162
1163 /*
1164  * MPALMOSTSAFE - acquires mplock
1165  *
1166  * we implement a very minimal set of ioctls for compatibility with sockets.
1167  */
1168 int
1169 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1170 {
1171         struct pipe *mpipe;
1172         int error;
1173
1174         get_mplock();
1175         mpipe = (struct pipe *)fp->f_data;
1176
1177         switch (cmd) {
1178         case FIOASYNC:
1179                 if (*(int *)data) {
1180                         mpipe->pipe_state |= PIPE_ASYNC;
1181                 } else {
1182                         mpipe->pipe_state &= ~PIPE_ASYNC;
1183                 }
1184                 error = 0;
1185                 break;
1186         case FIONREAD:
1187                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1188                         *(int *)data = mpipe->pipe_map.xio_bytes -
1189                                         mpipe->pipe_buffer.out;
1190                 } else {
1191                         *(int *)data = mpipe->pipe_buffer.cnt;
1192                 }
1193                 error = 0;
1194                 break;
1195         case FIOSETOWN:
1196                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1197                 break;
1198         case FIOGETOWN:
1199                 *(int *)data = fgetown(mpipe->pipe_sigio);
1200                 error = 0;
1201                 break;
1202         case TIOCSPGRP:
1203                 /* This is deprecated, FIOSETOWN should be used instead. */
1204                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1205                 break;
1206
1207         case TIOCGPGRP:
1208                 /* This is deprecated, FIOGETOWN should be used instead. */
1209                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1210                 error = 0;
1211                 break;
1212         default:
1213                 error = ENOTTY;
1214                 break;
1215         }
1216         rel_mplock();
1217         return (error);
1218 }
1219
1220 /*
1221  * MPALMOSTSAFE - acquires mplock
1222  */
1223 int
1224 pipe_poll(struct file *fp, int events, struct ucred *cred)
1225 {
1226         struct pipe *rpipe;
1227         struct pipe *wpipe;
1228         int revents = 0;
1229
1230         get_mplock();
1231         rpipe = (struct pipe *)fp->f_data;
1232         wpipe = rpipe->pipe_peer;
1233         if (events & (POLLIN | POLLRDNORM))
1234                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1235                     (rpipe->pipe_buffer.cnt > 0) ||
1236                     (rpipe->pipe_state & PIPE_EOF))
1237                         revents |= events & (POLLIN | POLLRDNORM);
1238
1239         if (events & (POLLOUT | POLLWRNORM))
1240                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1241                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1242                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1243                         revents |= events & (POLLOUT | POLLWRNORM);
1244
1245         if ((rpipe->pipe_state & PIPE_EOF) ||
1246             (wpipe == NULL) ||
1247             (wpipe->pipe_state & PIPE_EOF))
1248                 revents |= POLLHUP;
1249
1250         if (revents == 0) {
1251                 if (events & (POLLIN | POLLRDNORM)) {
1252                         selrecord(curthread, &rpipe->pipe_sel);
1253                         rpipe->pipe_state |= PIPE_SEL;
1254                 }
1255
1256                 if (events & (POLLOUT | POLLWRNORM)) {
1257                         selrecord(curthread, &wpipe->pipe_sel);
1258                         wpipe->pipe_state |= PIPE_SEL;
1259                 }
1260         }
1261         rel_mplock();
1262         return (revents);
1263 }
1264
1265 /*
1266  * MPALMOSTSAFE - acquires mplock
1267  */
1268 static int
1269 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1270 {
1271         struct pipe *pipe;
1272
1273         get_mplock();
1274         pipe = (struct pipe *)fp->f_data;
1275
1276         bzero((caddr_t)ub, sizeof(*ub));
1277         ub->st_mode = S_IFIFO;
1278         ub->st_blksize = pipe->pipe_buffer.size;
1279         ub->st_size = pipe->pipe_buffer.cnt;
1280         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1281                 ub->st_size = pipe->pipe_map.xio_bytes -
1282                                 pipe->pipe_buffer.out;
1283         }
1284         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1285         ub->st_atimespec = pipe->pipe_atime;
1286         ub->st_mtimespec = pipe->pipe_mtime;
1287         ub->st_ctimespec = pipe->pipe_ctime;
1288         /*
1289          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1290          * st_flags, st_gen.
1291          * XXX (st_dev, st_ino) should be unique.
1292          */
1293         rel_mplock();
1294         return (0);
1295 }
1296
1297 /*
1298  * MPALMOSTSAFE - acquires mplock
1299  */
1300 static int
1301 pipe_close(struct file *fp)
1302 {
1303         struct pipe *cpipe = (struct pipe *)fp->f_data;
1304
1305         get_mplock();
1306         fp->f_ops = &badfileops;
1307         fp->f_data = NULL;
1308         funsetown(cpipe->pipe_sigio);
1309         pipeclose(cpipe);
1310         rel_mplock();
1311         return (0);
1312 }
1313
1314 /*
1315  * Shutdown one or both directions of a full-duplex pipe.
1316  *
1317  * MPALMOSTSAFE - acquires mplock
1318  */
1319 static int
1320 pipe_shutdown(struct file *fp, int how)
1321 {
1322         struct pipe *rpipe;
1323         struct pipe *wpipe;
1324         int error = EPIPE;
1325
1326         get_mplock();
1327         rpipe = (struct pipe *)fp->f_data;
1328
1329         switch(how) {
1330         case SHUT_RDWR:
1331         case SHUT_RD:
1332                 if (rpipe) {
1333                         rpipe->pipe_state |= PIPE_EOF;
1334                         pipeselwakeup(rpipe);
1335                         if (rpipe->pipe_busy)
1336                                 wakeup(rpipe);
1337                         error = 0;
1338                 }
1339                 if (how == SHUT_RD)
1340                         break;
1341                 /* fall through */
1342         case SHUT_WR:
1343                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1344                         wpipe->pipe_state |= PIPE_EOF;
1345                         pipeselwakeup(wpipe);
1346                         if (wpipe->pipe_busy)
1347                                 wakeup(wpipe);
1348                         error = 0;
1349                 }
1350         }
1351         rel_mplock();
1352         return (error);
1353 }
1354
1355 static void
1356 pipe_free_kmem(struct pipe *cpipe)
1357 {
1358         if (cpipe->pipe_buffer.buffer != NULL) {
1359                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1360                         --pipe_nbig;
1361                 kmem_free(kernel_map,
1362                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1363                         cpipe->pipe_buffer.size);
1364                 cpipe->pipe_buffer.buffer = NULL;
1365                 cpipe->pipe_buffer.object = NULL;
1366         }
1367 #ifndef PIPE_NODIRECT
1368         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1369                 cpipe->pipe_map.xio_offset == 0 &&
1370                 cpipe->pipe_map.xio_npages == 0);
1371 #endif
1372 }
1373
1374 /*
1375  * shutdown the pipe
1376  */
1377 static void
1378 pipeclose(struct pipe *cpipe)
1379 {
1380         globaldata_t gd;
1381         struct pipe *ppipe;
1382
1383         if (cpipe == NULL)
1384                 return;
1385
1386         pipeselwakeup(cpipe);
1387
1388         /*
1389          * If the other side is blocked, wake it up saying that
1390          * we want to close it down.
1391          */
1392         while (cpipe->pipe_busy) {
1393                 wakeup(cpipe);
1394                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1395                 tsleep(cpipe, 0, "pipecl", 0);
1396         }
1397
1398         /*
1399          * Disconnect from peer
1400          */
1401         if ((ppipe = cpipe->pipe_peer) != NULL) {
1402                 pipeselwakeup(ppipe);
1403
1404                 ppipe->pipe_state |= PIPE_EOF;
1405                 wakeup(ppipe);
1406                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1407                 ppipe->pipe_peer = NULL;
1408         }
1409
1410         if (cpipe->pipe_kva) {
1411                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1412                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1413                 cpipe->pipe_kva = NULL;
1414         }
1415
1416         /*
1417          * free or cache resources
1418          */
1419         gd = mycpu;
1420         if (gd->gd_pipeqcount >= pipe_maxcache ||
1421             cpipe->pipe_buffer.size != PIPE_SIZE
1422         ) {
1423                 pipe_free_kmem(cpipe);
1424                 kfree(cpipe, M_PIPE);
1425         } else {
1426                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1427                         cpipe->pipe_map.xio_bytes == 0 &&
1428                         cpipe->pipe_map.xio_offset == 0);
1429                 cpipe->pipe_state = 0;
1430                 cpipe->pipe_busy = 0;
1431                 cpipe->pipe_peer = gd->gd_pipeq;
1432                 gd->gd_pipeq = cpipe;
1433                 ++gd->gd_pipeqcount;
1434         }
1435 }
1436
1437 /*
1438  * MPALMOSTSAFE - acquires mplock
1439  */
1440 static int
1441 pipe_kqfilter(struct file *fp, struct knote *kn)
1442 {
1443         struct pipe *cpipe;
1444
1445         get_mplock();
1446         cpipe = (struct pipe *)kn->kn_fp->f_data;
1447
1448         switch (kn->kn_filter) {
1449         case EVFILT_READ:
1450                 kn->kn_fop = &pipe_rfiltops;
1451                 break;
1452         case EVFILT_WRITE:
1453                 kn->kn_fop = &pipe_wfiltops;
1454                 cpipe = cpipe->pipe_peer;
1455                 if (cpipe == NULL) {
1456                         /* other end of pipe has been closed */
1457                         rel_mplock();
1458                         return (EPIPE);
1459                 }
1460                 break;
1461         default:
1462                 return (1);
1463         }
1464         kn->kn_hook = (caddr_t)cpipe;
1465
1466         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1467         rel_mplock();
1468         return (0);
1469 }
1470
1471 static void
1472 filt_pipedetach(struct knote *kn)
1473 {
1474         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1475
1476         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1477 }
1478
1479 /*ARGSUSED*/
1480 static int
1481 filt_piperead(struct knote *kn, long hint)
1482 {
1483         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1484         struct pipe *wpipe = rpipe->pipe_peer;
1485
1486         kn->kn_data = rpipe->pipe_buffer.cnt;
1487         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1488                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1489                                 rpipe->pipe_buffer.out;
1490         }
1491
1492         if ((rpipe->pipe_state & PIPE_EOF) ||
1493             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1494                 kn->kn_flags |= EV_EOF;
1495                 return (1);
1496         }
1497         return (kn->kn_data > 0);
1498 }
1499
1500 /*ARGSUSED*/
1501 static int
1502 filt_pipewrite(struct knote *kn, long hint)
1503 {
1504         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1505         struct pipe *wpipe = rpipe->pipe_peer;
1506
1507         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1508                 kn->kn_data = 0;
1509                 kn->kn_flags |= EV_EOF;
1510                 return (1);
1511         }
1512         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1513         if (wpipe->pipe_state & PIPE_DIRECTW)
1514                 kn->kn_data = 0;
1515
1516         return (kn->kn_data >= PIPE_BUF);
1517 }