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 1.19 2004/05/02 07:57:45 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
  76 #include <vm/vm.h>
  77 #include <vm/vm_param.h>
  78 #include <sys/lock.h>
  79 #include <vm/vm_object.h>
  80 #include <vm/vm_kern.h>
  81 #include <vm/vm_extern.h>
  82 #include <vm/pmap.h>
  83 #include <vm/vm_map.h>
  84 #include <vm/vm_page.h>
  85 #include <vm/vm_zone.h>
  86
  87 #include <sys/file2.h>
  88
  89 #include <machine/cpufunc.h>
  90
  91 /*
  92  * interfaces to the outside world
  93  */
  94 static int pipe_read (struct file *fp, struct uio *uio,
  95                 struct ucred *cred, int flags, struct thread *td);
  96 static int pipe_write (struct file *fp, struct uio *uio,
  97                 struct ucred *cred, int flags, struct thread *td);
  98 static int pipe_close (struct file *fp, struct thread *td);
  99 static int pipe_poll (struct file *fp, int events, struct ucred *cred,
 100                 struct thread *td);
 101 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 102 static int pipe_stat (struct file *fp, struct stat *sb, struct thread *td);
 103 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct thread *td);
 104
 105 static struct fileops pipeops = {
 106         NULL,   /* port */
 107         0,      /* autoq */
 108         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 109         pipe_stat, pipe_close
 110 };
 111
 112 static void     filt_pipedetach(struct knote *kn);
 113 static int      filt_piperead(struct knote *kn, long hint);
 114 static int      filt_pipewrite(struct knote *kn, long hint);
 115
 116 static struct filterops pipe_rfiltops =
 117         { 1, NULL, filt_pipedetach, filt_piperead };
 118 static struct filterops pipe_wfiltops =
 119         { 1, NULL, filt_pipedetach, filt_pipewrite };
 120
 121 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 122
 123 /*
 124  * Default pipe buffer size(s), this can be kind-of large now because pipe
 125  * space is pageable.  The pipe code will try to maintain locality of
 126  * reference for performance reasons, so small amounts of outstanding I/O
 127  * will not wipe the cache.
 128  */
 129 #define MINPIPESIZE (PIPE_SIZE/3)
 130 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 131
 132 /*
 133  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 134  * is there so that on large systems, we don't exhaust it.
 135  */
 136 #define MAXPIPEKVA (8*1024*1024)
 137
 138 /*
 139  * Limit for direct transfers, we cannot, of course limit
 140  * the amount of kva for pipes in general though.
 141  */
 142 #define LIMITPIPEKVA (16*1024*1024)
 143
 144 /*
 145  * Limit the number of "big" pipes
 146  */
 147 #define LIMITBIGPIPES   32
 148 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 149
 150 static int pipe_maxbig = LIMITBIGPIPES;
 151 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 152 static int pipe_nbig;
 153 static int pipe_bcache_alloc;
 154 static int pipe_bkmem_alloc;
 155 static int pipe_dwrite_enable = 1;      /* 0:copy, 1:kmem/sfbuf 2:force */
 156 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 157                                         /* 3:sfbuf_dmap w/ forced invlpg */
 158
 159 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 160 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 161         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 162 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 163         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 164 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 165         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 166 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 167         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 168 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 169         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0, "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 170 #if !defined(NO_PIPE_SYSCTL_STATS)
 171 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 172         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 173 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 174         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 175 #endif
 176
 177 static void pipeclose (struct pipe *cpipe);
 178 static void pipe_free_kmem (struct pipe *cpipe);
 179 static int pipe_create (struct pipe **cpipep);
 180 static __inline int pipelock (struct pipe *cpipe, int catch);
 181 static __inline void pipeunlock (struct pipe *cpipe);
 182 static __inline void pipeselwakeup (struct pipe *cpipe);
 183 #ifndef PIPE_NODIRECT
 184 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 185 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 186 static void pipe_clone_write_buffer (struct pipe *wpipe);
 187 #endif
 188 static int pipespace (struct pipe *cpipe, int size);
 189
 190 /*
 191  * The pipe system call for the DTYPE_PIPE type of pipes
 192  *
 193  * pipe_ARgs(int dummy)
 194  */
 195
 196 /* ARGSUSED */
 197 int
 198 pipe(struct pipe_args *uap)
 199 {
 200         struct thread *td = curthread;
 201         struct proc *p = td->td_proc;
 202         struct filedesc *fdp;
 203         struct file *rf, *wf;
 204         struct pipe *rpipe, *wpipe;
 205         int fd1, fd2, error;
 206
 207         KKASSERT(p);
 208         fdp = p->p_fd;
 209
 210         rpipe = wpipe = NULL;
 211         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 212                 pipeclose(rpipe);
 213                 pipeclose(wpipe);
 214                 return (ENFILE);
 215         }
 216
 217         rpipe->pipe_state |= PIPE_DIRECTOK;
 218         wpipe->pipe_state |= PIPE_DIRECTOK;
 219
 220         /*
 221          * Select the direct-map features to use for this pipe.  Since the
 222          * sysctl's can change on the fly we record the settings when the
 223          * pipe is created.
 224          *
 225          * Generally speaking the system will default to what we consider
 226          * to be the best-balanced and most stable option.  Right now this
 227          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 228          * moment.
 229          */
 230         wpipe->pipe_feature = PIPE_COPY;
 231         if (pipe_dwrite_enable) {
 232                 switch(pipe_dwrite_sfbuf) {
 233                 case 0:
 234                         wpipe->pipe_feature = PIPE_KMEM;
 235                         break;
 236                 case 1:
 237                         wpipe->pipe_feature = PIPE_SFBUF1;
 238                         break;
 239                 case 2:
 240                 case 3:
 241                         wpipe->pipe_feature = PIPE_SFBUF2;
 242                         break;
 243                 }
 244         }
 245         rpipe->pipe_feature = wpipe->pipe_feature;
 246
 247         error = falloc(p, &rf, &fd1);
 248         if (error) {
 249                 pipeclose(rpipe);
 250                 pipeclose(wpipe);
 251                 return (error);
 252         }
 253         fhold(rf);
 254         uap->sysmsg_fds[0] = fd1;
 255
 256         /*
 257          * Warning: once we've gotten past allocation of the fd for the
 258          * read-side, we can only drop the read side via fdrop() in order
 259          * to avoid races against processes which manage to dup() the read
 260          * side while we are blocked trying to allocate the write side.
 261          */
 262         rf->f_flag = FREAD | FWRITE;
 263         rf->f_type = DTYPE_PIPE;
 264         rf->f_data = (caddr_t)rpipe;
 265         rf->f_ops = &pipeops;
 266         error = falloc(p, &wf, &fd2);
 267         if (error) {
 268                 if (fdp->fd_ofiles[fd1] == rf) {
 269                         fdp->fd_ofiles[fd1] = NULL;
 270                         fdrop(rf, td);
 271                 }
 272                 fdrop(rf, td);
 273                 /* rpipe has been closed by fdrop(). */
 274                 pipeclose(wpipe);
 275                 return (error);
 276         }
 277         wf->f_flag = FREAD | FWRITE;
 278         wf->f_type = DTYPE_PIPE;
 279         wf->f_data = (caddr_t)wpipe;
 280         wf->f_ops = &pipeops;
 281         uap->sysmsg_fds[1] = fd2;
 282
 283         rpipe->pipe_peer = wpipe;
 284         wpipe->pipe_peer = rpipe;
 285         fdrop(rf, td);
 286
 287         return (0);
 288 }
 289
 290 /*
 291  * Allocate kva for pipe circular buffer, the space is pageable
 292  * This routine will 'realloc' the size of a pipe safely, if it fails
 293  * it will retain the old buffer.
 294  * If it fails it will return ENOMEM.
 295  */
 296 static int
 297 pipespace(struct pipe *cpipe, int size)
 298 {
 299         struct vm_object *object;
 300         caddr_t buffer;
 301         int npages, error;
 302
 303         npages = round_page(size) / PAGE_SIZE;
 304         object = cpipe->pipe_buffer.object;
 305
 306         /*
 307          * [re]create the object if necessary and reserve space for it
 308          * in the kernel_map.  The object and memory are pageable.  On
 309          * success, free the old resources before assigning the new
 310          * ones.
 311          */
 312         if (object == NULL || object->size != npages) {
 313                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 314                 buffer = (caddr_t) vm_map_min(kernel_map);
 315
 316                 error = vm_map_find(kernel_map, object, 0,
 317                         (vm_offset_t *) &buffer, size, 1,
 318                         VM_PROT_ALL, VM_PROT_ALL, 0);
 319
 320                 if (error != KERN_SUCCESS) {
 321                         vm_object_deallocate(object);
 322                         return (ENOMEM);
 323                 }
 324                 pipe_free_kmem(cpipe);
 325                 cpipe->pipe_buffer.object = object;
 326                 cpipe->pipe_buffer.buffer = buffer;
 327                 cpipe->pipe_buffer.size = size;
 328                 ++pipe_bkmem_alloc;
 329         } else {
 330                 ++pipe_bcache_alloc;
 331         }
 332         cpipe->pipe_buffer.in = 0;
 333         cpipe->pipe_buffer.out = 0;
 334         cpipe->pipe_buffer.cnt = 0;
 335         return (0);
 336 }
 337
 338 /*
 339  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 340  * our per-cpu cache if possible.  For now make sure it is sized for the
 341  * smaller PIPE_SIZE default.
 342  */
 343 static int
 344 pipe_create(cpipep)
 345         struct pipe **cpipep;
 346 {
 347         globaldata_t gd = mycpu;
 348         struct pipe *cpipe;
 349         int error;
 350
 351         if ((cpipe = gd->gd_pipeq) != NULL) {
 352                 gd->gd_pipeq = cpipe->pipe_peer;
 353                 --gd->gd_pipeqcount;
 354                 cpipe->pipe_peer = NULL;
 355         } else {
 356                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 357         }
 358         *cpipep = cpipe;
 359         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 360                 return (error);
 361         vfs_timestamp(&cpipe->pipe_ctime);
 362         cpipe->pipe_atime = cpipe->pipe_ctime;
 363         cpipe->pipe_mtime = cpipe->pipe_ctime;
 364         return (0);
 365 }
 366
 367
 368 /*
 369  * lock a pipe for I/O, blocking other access
 370  */
 371 static __inline int
 372 pipelock(cpipe, catch)
 373         struct pipe *cpipe;
 374         int catch;
 375 {
 376         int error;
 377
 378         while (cpipe->pipe_state & PIPE_LOCK) {
 379                 cpipe->pipe_state |= PIPE_LWANT;
 380                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 381                 if (error != 0)
 382                         return (error);
 383         }
 384         cpipe->pipe_state |= PIPE_LOCK;
 385         return (0);
 386 }
 387
 388 /*
 389  * unlock a pipe I/O lock
 390  */
 391 static __inline void
 392 pipeunlock(cpipe)
 393         struct pipe *cpipe;
 394 {
 395
 396         cpipe->pipe_state &= ~PIPE_LOCK;
 397         if (cpipe->pipe_state & PIPE_LWANT) {
 398                 cpipe->pipe_state &= ~PIPE_LWANT;
 399                 wakeup(cpipe);
 400         }
 401 }
 402
 403 static __inline void
 404 pipeselwakeup(cpipe)
 405         struct pipe *cpipe;
 406 {
 407
 408         if (cpipe->pipe_state & PIPE_SEL) {
 409                 cpipe->pipe_state &= ~PIPE_SEL;
 410                 selwakeup(&cpipe->pipe_sel);
 411         }
 412         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 413                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 414         KNOTE(&cpipe->pipe_sel.si_note, 0);
 415 }
 416
 417 /* ARGSUSED */
 418 static int
 419 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
 420         int flags, struct thread *td)
 421 {
 422         struct pipe *rpipe = (struct pipe *) fp->f_data;
 423         int error;
 424         int nread = 0;
 425         u_int size;
 426
 427         ++rpipe->pipe_busy;
 428         error = pipelock(rpipe, 1);
 429         if (error)
 430                 goto unlocked_error;
 431
 432         while (uio->uio_resid) {
 433                 caddr_t va;
 434
 435                 if (rpipe->pipe_buffer.cnt > 0) {
 436                         /*
 437                          * normal pipe buffer receive
 438                          */
 439                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 440                         if (size > rpipe->pipe_buffer.cnt)
 441                                 size = rpipe->pipe_buffer.cnt;
 442                         if (size > (u_int) uio->uio_resid)
 443                                 size = (u_int) uio->uio_resid;
 444
 445                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 446                                         size, uio);
 447                         if (error)
 448                                 break;
 449
 450                         rpipe->pipe_buffer.out += size;
 451                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 452                                 rpipe->pipe_buffer.out = 0;
 453
 454                         rpipe->pipe_buffer.cnt -= size;
 455
 456                         /*
 457                          * If there is no more to read in the pipe, reset
 458                          * its pointers to the beginning.  This improves
 459                          * cache hit stats.
 460                          */
 461                         if (rpipe->pipe_buffer.cnt == 0) {
 462                                 rpipe->pipe_buffer.in = 0;
 463                                 rpipe->pipe_buffer.out = 0;
 464                         }
 465                         nread += size;
 466 #ifndef PIPE_NODIRECT
 467                 } else if (rpipe->pipe_kva &&
 468                            rpipe->pipe_feature == PIPE_KMEM &&
 469                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 470                                == PIPE_DIRECTW
 471                 ) {
 472                         /*
 473                          * Direct copy using source-side kva mapping
 474                          */
 475                         size = rpipe->pipe_map.xio_bytes;
 476                         if (size > (u_int)uio->uio_resid)
 477                                 size = (u_int)uio->uio_resid;
 478                         va = (caddr_t)rpipe->pipe_kva + rpipe->pipe_map.xio_offset;
 479                         error = uiomove(va, size, uio);
 480                         if (error)
 481                                 break;
 482                         nread += size;
 483                         rpipe->pipe_map.xio_offset += size;
 484                         rpipe->pipe_map.xio_bytes -= size;
 485                         if (rpipe->pipe_map.xio_bytes == 0) {
 486                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 487                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 488                                 wakeup(rpipe);
 489                         }
 490                 } else if (rpipe->pipe_map.xio_bytes &&
 491                            rpipe->pipe_kva &&
 492                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 493                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 494                                == PIPE_DIRECTW
 495                 ) {
 496                         /*
 497                          * Direct copy, bypassing a kernel buffer.  We cannot
 498                          * mess with the direct-write buffer until
 499                          * PIPE_DIRECTIP is cleared.  In order to prevent
 500                          * the pipe_write code from racing itself in
 501                          * direct_write, we set DIRECTIP when we clear
 502                          * DIRECTW after we have exhausted the buffer.
 503                          */
 504                         if (pipe_dwrite_sfbuf == 3)
 505                                 rpipe->pipe_kvamask = 0;
 506                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 507                                     rpipe->pipe_map.xio_npages,
 508                                     &rpipe->pipe_kvamask);
 509                         size = rpipe->pipe_map.xio_bytes;
 510                         if (size > (u_int)uio->uio_resid)
 511                                 size = (u_int)uio->uio_resid;
 512                         va = (caddr_t)rpipe->pipe_kva +
 513                                 rpipe->pipe_map.xio_offset;
 514                         error = uiomove(va, size, uio);
 515                         if (error)
 516                                 break;
 517                         nread += size;
 518                         rpipe->pipe_map.xio_offset += size;
 519                         rpipe->pipe_map.xio_bytes -= size;
 520                         if (rpipe->pipe_map.xio_bytes == 0) {
 521                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 522                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 523                                 wakeup(rpipe);
 524                         }
 525                 } else if (rpipe->pipe_map.xio_bytes &&
 526                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 527                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 528                                 == PIPE_DIRECTW
 529                 ) {
 530                         /*
 531                          * Direct copy, bypassing a kernel buffer.  We cannot
 532                          * mess with the direct-write buffer until
 533                          * PIPE_DIRECTIP is cleared.  In order to prevent
 534                          * the pipe_write code from racing itself in
 535                          * direct_write, we set DIRECTIP when we clear
 536                          * DIRECTW after we have exhausted the buffer.
 537                          */
 538                         error = xio_uio_copy(&rpipe->pipe_map, uio, &size);
 539                         if (error)
 540                                 break;
 541                         nread += size;
 542                         if (rpipe->pipe_map.xio_bytes == 0) {
 543                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 544                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 545                                 wakeup(rpipe);
 546                         }
 547 #endif
 548                 } else {
 549                         /*
 550                          * detect EOF condition
 551                          * read returns 0 on EOF, no need to set error
 552                          */
 553                         if (rpipe->pipe_state & PIPE_EOF)
 554                                 break;
 555
 556                         /*
 557                          * If the "write-side" has been blocked, wake it up now.
 558                          */
 559                         if (rpipe->pipe_state & PIPE_WANTW) {
 560                                 rpipe->pipe_state &= ~PIPE_WANTW;
 561                                 wakeup(rpipe);
 562                         }
 563
 564                         /*
 565                          * Break if some data was read.
 566                          */
 567                         if (nread > 0)
 568                                 break;
 569
 570                         /*
 571                          * Unlock the pipe buffer for our remaining
 572                          * processing.  We will either break out with an
 573                          * error or we will sleep and relock to loop.
 574                          */
 575                         pipeunlock(rpipe);
 576
 577                         /*
 578                          * Handle non-blocking mode operation or
 579                          * wait for more data.
 580                          */
 581                         if (fp->f_flag & FNONBLOCK) {
 582                                 error = EAGAIN;
 583                         } else {
 584                                 rpipe->pipe_state |= PIPE_WANTR;
 585                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 586                                     "piperd", 0)) == 0) {
 587                                         error = pipelock(rpipe, 1);
 588                                 }
 589                         }
 590                         if (error)
 591                                 goto unlocked_error;
 592                 }
 593         }
 594         pipeunlock(rpipe);
 595
 596         if (error == 0)
 597                 vfs_timestamp(&rpipe->pipe_atime);
 598 unlocked_error:
 599         --rpipe->pipe_busy;
 600
 601         /*
 602          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 603          */
 604         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 605                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 606                 wakeup(rpipe);
 607         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 608                 /*
 609                  * Handle write blocking hysteresis.
 610                  */
 611                 if (rpipe->pipe_state & PIPE_WANTW) {
 612                         rpipe->pipe_state &= ~PIPE_WANTW;
 613                         wakeup(rpipe);
 614                 }
 615         }
 616
 617         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 618                 pipeselwakeup(rpipe);
 619         return (error);
 620 }
 621
 622 #ifndef PIPE_NODIRECT
 623 /*
 624  * Map the sending processes' buffer into kernel space and wire it.
 625  * This is similar to a physical write operation.
 626  */
 627 static int
 628 pipe_build_write_buffer(wpipe, uio)
 629         struct pipe *wpipe;
 630         struct uio *uio;
 631 {
 632         int error;
 633         u_int size;
 634
 635         size = (u_int) uio->uio_iov->iov_len;
 636         if (size > wpipe->pipe_buffer.size)
 637                 size = wpipe->pipe_buffer.size;
 638         if (size > XIO_INTERNAL_SIZE)
 639                 size = XIO_INTERNAL_SIZE;
 640
 641         error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 642                                 size, XIOF_READ);
 643         if (error)
 644                 return(error);
 645
 646         /*
 647          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 648          * map the pages on the target while KMEM maps the pages now.
 649          */
 650         switch(wpipe->pipe_feature) {
 651         case PIPE_KMEM:
 652         case PIPE_SFBUF2:
 653                 if (wpipe->pipe_kva == NULL) {
 654                         wpipe->pipe_kva =
 655                             kmem_alloc_pageable(kernel_map, XIO_INTERNAL_SIZE);
 656                         wpipe->pipe_kvamask = 0;
 657                 }
 658                 if (wpipe->pipe_feature == PIPE_KMEM) {
 659                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 660                                     wpipe->pipe_map.xio_npages);
 661                 }
 662                 break;
 663         default:
 664                 break;
 665         }
 666
 667         /*
 668          * and update the uio data
 669          */
 670         uio->uio_iov->iov_len -= size;
 671         uio->uio_iov->iov_base += size;
 672         if (uio->uio_iov->iov_len == 0)
 673                 uio->uio_iov++;
 674         uio->uio_resid -= size;
 675         uio->uio_offset += size;
 676         return (0);
 677 }
 678
 679 /*
 680  * In the case of a signal, the writing process might go away.  This
 681  * code copies the data into the circular buffer so that the source
 682  * pages can be freed without loss of data.
 683  */
 684 static void
 685 pipe_clone_write_buffer(wpipe)
 686         struct pipe *wpipe;
 687 {
 688         int size;
 689
 690         size = wpipe->pipe_map.xio_bytes;
 691
 692         wpipe->pipe_buffer.in = size;
 693         wpipe->pipe_buffer.out = 0;
 694         wpipe->pipe_buffer.cnt = size;
 695         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 696
 697         xio_copy_xtok(&wpipe->pipe_map, wpipe->pipe_buffer.buffer, size);
 698         xio_release(&wpipe->pipe_map);
 699         if (wpipe->pipe_kva) {
 700                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 701                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 702                 wpipe->pipe_kva = NULL;
 703         }
 704 }
 705
 706 /*
 707  * This implements the pipe buffer write mechanism.  Note that only
 708  * a direct write OR a normal pipe write can be pending at any given time.
 709  * If there are any characters in the pipe buffer, the direct write will
 710  * be deferred until the receiving process grabs all of the bytes from
 711  * the pipe buffer.  Then the direct mapping write is set-up.
 712  */
 713 static int
 714 pipe_direct_write(wpipe, uio)
 715         struct pipe *wpipe;
 716         struct uio *uio;
 717 {
 718         int error;
 719
 720 retry:
 721         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 722                 if (wpipe->pipe_state & PIPE_WANTR) {
 723                         wpipe->pipe_state &= ~PIPE_WANTR;
 724                         wakeup(wpipe);
 725                 }
 726                 wpipe->pipe_state |= PIPE_WANTW;
 727                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 728                 if (error)
 729                         goto error2;
 730                 if (wpipe->pipe_state & PIPE_EOF) {
 731                         error = EPIPE;
 732                         goto error2;
 733                 }
 734         }
 735         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 736         if (wpipe->pipe_buffer.cnt > 0) {
 737                 if (wpipe->pipe_state & PIPE_WANTR) {
 738                         wpipe->pipe_state &= ~PIPE_WANTR;
 739                         wakeup(wpipe);
 740                 }
 741
 742                 wpipe->pipe_state |= PIPE_WANTW;
 743                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 744                 if (error)
 745                         goto error2;
 746                 if (wpipe->pipe_state & PIPE_EOF) {
 747                         error = EPIPE;
 748                         goto error2;
 749                 }
 750                 goto retry;
 751         }
 752
 753         /*
 754          * Build our direct-write buffer
 755          */
 756         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 757         error = pipe_build_write_buffer(wpipe, uio);
 758         if (error)
 759                 goto error1;
 760         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 761
 762         /*
 763          * Wait until the receiver has snarfed the data.  Since we are likely
 764          * going to sleep we optimize the case and yield synchronously,
 765          * possibly avoiding the tsleep().
 766          */
 767         error = 0;
 768         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 769                 if (wpipe->pipe_state & PIPE_EOF) {
 770                         pipelock(wpipe, 0);
 771                         xio_release(&wpipe->pipe_map);
 772                         if (wpipe->pipe_kva) {
 773                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 774                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 775                                 wpipe->pipe_kva = NULL;
 776                         }
 777                         pipeunlock(wpipe);
 778                         pipeselwakeup(wpipe);
 779                         error = EPIPE;
 780                         goto error1;
 781                 }
 782                 if (wpipe->pipe_state & PIPE_WANTR) {
 783                         wpipe->pipe_state &= ~PIPE_WANTR;
 784                         wakeup(wpipe);
 785                 }
 786                 pipeselwakeup(wpipe);
 787                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 788         }
 789         pipelock(wpipe,0);
 790         if (wpipe->pipe_state & PIPE_DIRECTW) {
 791                 /*
 792                  * this bit of trickery substitutes a kernel buffer for
 793                  * the process that might be going away.
 794                  */
 795                 pipe_clone_write_buffer(wpipe);
 796                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 797         } else {
 798                 /*
 799                  * note: The pipe_kva mapping is not qremove'd here.  For
 800                  * legacy PIPE_KMEM mode this constitutes an improvement
 801                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 802                  * mode the kva mapping must not be removed to get the
 803                  * caching benefit.
 804                  *
 805                  * For testing purposes we will give the original algorithm
 806                  * the benefit of the doubt 'what it could have been', and
 807                  * keep the optimization.
 808                  */
 809                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 810                 xio_release(&wpipe->pipe_map);
 811                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 812         }
 813         pipeunlock(wpipe);
 814         return (error);
 815
 816         /*
 817          * Direct-write error, clear the direct write flags.
 818          */
 819 error1:
 820         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 821         /* fallthrough */
 822
 823         /*
 824          * General error, wakeup the other side if it happens to be sleeping.
 825          */
 826 error2:
 827         wakeup(wpipe);
 828         return (error);
 829 }
 830 #endif
 831
 832 static int
 833 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
 834         int flags, struct thread *td)
 835 {
 836         int error = 0;
 837         int orig_resid;
 838         struct pipe *wpipe, *rpipe;
 839
 840         rpipe = (struct pipe *) fp->f_data;
 841         wpipe = rpipe->pipe_peer;
 842
 843         /*
 844          * detect loss of pipe read side, issue SIGPIPE if lost.
 845          */
 846         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 847                 return (EPIPE);
 848         }
 849         ++wpipe->pipe_busy;
 850
 851         /*
 852          * If it is advantageous to resize the pipe buffer, do
 853          * so.
 854          */
 855         if ((uio->uio_resid > PIPE_SIZE) &&
 856                 (pipe_nbig < pipe_maxbig) &&
 857                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 858                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 859                 (wpipe->pipe_buffer.cnt == 0)) {
 860
 861                 if ((error = pipelock(wpipe,1)) == 0) {
 862                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 863                                 pipe_nbig++;
 864                         pipeunlock(wpipe);
 865                 }
 866         }
 867
 868         /*
 869          * If an early error occured unbusy and return, waking up any pending
 870          * readers.
 871          */
 872         if (error) {
 873                 --wpipe->pipe_busy;
 874                 if ((wpipe->pipe_busy == 0) &&
 875                     (wpipe->pipe_state & PIPE_WANT)) {
 876                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 877                         wakeup(wpipe);
 878                 }
 879                 return(error);
 880         }
 881
 882         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 883
 884         orig_resid = uio->uio_resid;
 885
 886         while (uio->uio_resid) {
 887                 int space;
 888
 889 #ifndef PIPE_NODIRECT
 890                 /*
 891                  * If the transfer is large, we can gain performance if
 892                  * we do process-to-process copies directly.
 893                  * If the write is non-blocking, we don't use the
 894                  * direct write mechanism.
 895                  *
 896                  * The direct write mechanism will detect the reader going
 897                  * away on us.
 898                  */
 899                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 900                     pipe_dwrite_enable > 1) &&
 901                     (fp->f_flag & FNONBLOCK) == 0 &&
 902                     pipe_dwrite_enable) {
 903                         error = pipe_direct_write( wpipe, uio);
 904                         if (error)
 905                                 break;
 906                         continue;
 907                 }
 908 #endif
 909
 910                 /*
 911                  * Pipe buffered writes cannot be coincidental with
 912                  * direct writes.  We wait until the currently executing
 913                  * direct write is completed before we start filling the
 914                  * pipe buffer.  We break out if a signal occurs or the
 915                  * reader goes away.
 916                  */
 917         retrywrite:
 918                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 919                         if (wpipe->pipe_state & PIPE_WANTR) {
 920                                 wpipe->pipe_state &= ~PIPE_WANTR;
 921                                 wakeup(wpipe);
 922                         }
 923                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 924                         if (wpipe->pipe_state & PIPE_EOF)
 925                                 break;
 926                         if (error)
 927                                 break;
 928                 }
 929                 if (wpipe->pipe_state & PIPE_EOF) {
 930                         error = EPIPE;
 931                         break;
 932                 }
 933
 934                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 935
 936                 /* Writes of size <= PIPE_BUF must be atomic. */
 937                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 938                         space = 0;
 939
 940                 /*
 941                  * Write to fill, read size handles write hysteresis.  Also
 942                  * additional restrictions can cause select-based non-blocking
 943                  * writes to spin.
 944                  */
 945                 if (space > 0) {
 946                         if ((error = pipelock(wpipe,1)) == 0) {
 947                                 int size;       /* Transfer size */
 948                                 int segsize;    /* first segment to transfer */
 949
 950                                 /*
 951                                  * It is possible for a direct write to
 952                                  * slip in on us... handle it here...
 953                                  */
 954                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 955                                         pipeunlock(wpipe);
 956                                         goto retrywrite;
 957                                 }
 958                                 /*
 959                                  * If a process blocked in uiomove, our
 960                                  * value for space might be bad.
 961                                  *
 962                                  * XXX will we be ok if the reader has gone
 963                                  * away here?
 964                                  */
 965                                 if (space > wpipe->pipe_buffer.size -
 966                                     wpipe->pipe_buffer.cnt) {
 967                                         pipeunlock(wpipe);
 968                                         goto retrywrite;
 969                                 }
 970
 971                                 /*
 972                                  * Transfer size is minimum of uio transfer
 973                                  * and free space in pipe buffer.
 974                                  */
 975                                 if (space > uio->uio_resid)
 976                                         size = uio->uio_resid;
 977                                 else
 978                                         size = space;
 979                                 /*
 980                                  * First segment to transfer is minimum of
 981                                  * transfer size and contiguous space in
 982                                  * pipe buffer.  If first segment to transfer
 983                                  * is less than the transfer size, we've got
 984                                  * a wraparound in the buffer.
 985                                  */
 986                                 segsize = wpipe->pipe_buffer.size -
 987                                         wpipe->pipe_buffer.in;
 988                                 if (segsize > size)
 989                                         segsize = size;
 990
 991                                 /* Transfer first segment */
 992
 993                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 994                                                 segsize, uio);
 995
 996                                 if (error == 0 && segsize < size) {
 997                                         /*
 998                                          * Transfer remaining part now, to
 999                                          * support atomic writes.  Wraparound
1000                                          * happened.
1001                                          */
1002                                         if (wpipe->pipe_buffer.in + segsize !=
1003                                             wpipe->pipe_buffer.size)
1004                                                 panic("Expected pipe buffer wraparound disappeared");
1005
1006                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1007                                                         size - segsize, uio);
1008                                 }
1009                                 if (error == 0) {
1010                                         wpipe->pipe_buffer.in += size;
1011                                         if (wpipe->pipe_buffer.in >=
1012                                             wpipe->pipe_buffer.size) {
1013                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1014                                                         panic("Expected wraparound bad");
1015                                                 wpipe->pipe_buffer.in = size - segsize;
1016                                         }
1017
1018                                         wpipe->pipe_buffer.cnt += size;
1019                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1020                                                 panic("Pipe buffer overflow");
1021
1022                                 }
1023                                 pipeunlock(wpipe);
1024                         }
1025                         if (error)
1026                                 break;
1027
1028                 } else {
1029                         /*
1030                          * If the "read-side" has been blocked, wake it up now
1031                          * and yield to let it drain synchronously rather
1032                          * then block.
1033                          */
1034                         if (wpipe->pipe_state & PIPE_WANTR) {
1035                                 wpipe->pipe_state &= ~PIPE_WANTR;
1036                                 wakeup(wpipe);
1037                         }
1038
1039                         /*
1040                          * don't block on non-blocking I/O
1041                          */
1042                         if (fp->f_flag & FNONBLOCK) {
1043                                 error = EAGAIN;
1044                                 break;
1045                         }
1046
1047                         /*
1048                          * We have no more space and have something to offer,
1049                          * wake up select/poll.
1050                          */
1051                         pipeselwakeup(wpipe);
1052
1053                         wpipe->pipe_state |= PIPE_WANTW;
1054                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1055                         if (error != 0)
1056                                 break;
1057                         /*
1058                          * If read side wants to go away, we just issue a signal
1059                          * to ourselves.
1060                          */
1061                         if (wpipe->pipe_state & PIPE_EOF) {
1062                                 error = EPIPE;
1063                                 break;
1064                         }
1065                 }
1066         }
1067
1068         --wpipe->pipe_busy;
1069
1070         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1071                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1072                 wakeup(wpipe);
1073         } else if (wpipe->pipe_buffer.cnt > 0) {
1074                 /*
1075                  * If we have put any characters in the buffer, we wake up
1076                  * the reader.
1077                  */
1078                 if (wpipe->pipe_state & PIPE_WANTR) {
1079                         wpipe->pipe_state &= ~PIPE_WANTR;
1080                         wakeup(wpipe);
1081                 }
1082         }
1083
1084         /*
1085          * Don't return EPIPE if I/O was successful
1086          */
1087         if ((wpipe->pipe_buffer.cnt == 0) &&
1088             (uio->uio_resid == 0) &&
1089             (error == EPIPE)) {
1090                 error = 0;
1091         }
1092
1093         if (error == 0)
1094                 vfs_timestamp(&wpipe->pipe_mtime);
1095
1096         /*
1097          * We have something to offer,
1098          * wake up select/poll.
1099          */
1100         if (wpipe->pipe_buffer.cnt)
1101                 pipeselwakeup(wpipe);
1102
1103         return (error);
1104 }
1105
1106 /*
1107  * we implement a very minimal set of ioctls for compatibility with sockets.
1108  */
1109 int
1110 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1111 {
1112         struct pipe *mpipe = (struct pipe *)fp->f_data;
1113
1114         switch (cmd) {
1115
1116         case FIONBIO:
1117                 return (0);
1118
1119         case FIOASYNC:
1120                 if (*(int *)data) {
1121                         mpipe->pipe_state |= PIPE_ASYNC;
1122                 } else {
1123                         mpipe->pipe_state &= ~PIPE_ASYNC;
1124                 }
1125                 return (0);
1126
1127         case FIONREAD:
1128                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1129                         *(int *)data = mpipe->pipe_map.xio_bytes;
1130                 } else {
1131                         *(int *)data = mpipe->pipe_buffer.cnt;
1132                 }
1133                 return (0);
1134
1135         case FIOSETOWN:
1136                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1137
1138         case FIOGETOWN:
1139                 *(int *)data = fgetown(mpipe->pipe_sigio);
1140                 return (0);
1141
1142         /* This is deprecated, FIOSETOWN should be used instead. */
1143         case TIOCSPGRP:
1144                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1145
1146         /* This is deprecated, FIOGETOWN should be used instead. */
1147         case TIOCGPGRP:
1148                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1149                 return (0);
1150
1151         }
1152         return (ENOTTY);
1153 }
1154
1155 int
1156 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1157 {
1158         struct pipe *rpipe = (struct pipe *)fp->f_data;
1159         struct pipe *wpipe;
1160         int revents = 0;
1161
1162         wpipe = rpipe->pipe_peer;
1163         if (events & (POLLIN | POLLRDNORM))
1164                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1165                     (rpipe->pipe_buffer.cnt > 0) ||
1166                     (rpipe->pipe_state & PIPE_EOF))
1167                         revents |= events & (POLLIN | POLLRDNORM);
1168
1169         if (events & (POLLOUT | POLLWRNORM))
1170                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1171                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1172                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1173                         revents |= events & (POLLOUT | POLLWRNORM);
1174
1175         if ((rpipe->pipe_state & PIPE_EOF) ||
1176             (wpipe == NULL) ||
1177             (wpipe->pipe_state & PIPE_EOF))
1178                 revents |= POLLHUP;
1179
1180         if (revents == 0) {
1181                 if (events & (POLLIN | POLLRDNORM)) {
1182                         selrecord(td, &rpipe->pipe_sel);
1183                         rpipe->pipe_state |= PIPE_SEL;
1184                 }
1185
1186                 if (events & (POLLOUT | POLLWRNORM)) {
1187                         selrecord(td, &wpipe->pipe_sel);
1188                         wpipe->pipe_state |= PIPE_SEL;
1189                 }
1190         }
1191
1192         return (revents);
1193 }
1194
1195 static int
1196 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1197 {
1198         struct pipe *pipe = (struct pipe *)fp->f_data;
1199
1200         bzero((caddr_t)ub, sizeof(*ub));
1201         ub->st_mode = S_IFIFO;
1202         ub->st_blksize = pipe->pipe_buffer.size;
1203         ub->st_size = pipe->pipe_buffer.cnt;
1204         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1205         ub->st_atimespec = pipe->pipe_atime;
1206         ub->st_mtimespec = pipe->pipe_mtime;
1207         ub->st_ctimespec = pipe->pipe_ctime;
1208         /*
1209          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1210          * st_flags, st_gen.
1211          * XXX (st_dev, st_ino) should be unique.
1212          */
1213         return (0);
1214 }
1215
1216 /* ARGSUSED */
1217 static int
1218 pipe_close(struct file *fp, struct thread *td)
1219 {
1220         struct pipe *cpipe = (struct pipe *)fp->f_data;
1221
1222         fp->f_ops = &badfileops;
1223         fp->f_data = NULL;
1224         funsetown(cpipe->pipe_sigio);
1225         pipeclose(cpipe);
1226         return (0);
1227 }
1228
1229 static void
1230 pipe_free_kmem(struct pipe *cpipe)
1231 {
1232         if (cpipe->pipe_buffer.buffer != NULL) {
1233                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1234                         --pipe_nbig;
1235                 kmem_free(kernel_map,
1236                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1237                         cpipe->pipe_buffer.size);
1238                 cpipe->pipe_buffer.buffer = NULL;
1239                 cpipe->pipe_buffer.object = NULL;
1240         }
1241 #ifndef PIPE_NODIRECT
1242         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1243                 cpipe->pipe_map.xio_offset == 0 &&
1244                 cpipe->pipe_map.xio_npages == 0);
1245 #endif
1246 }
1247
1248 /*
1249  * shutdown the pipe
1250  */
1251 static void
1252 pipeclose(struct pipe *cpipe)
1253 {
1254         globaldata_t gd;
1255         struct pipe *ppipe;
1256
1257         if (cpipe == NULL)
1258                 return;
1259
1260         pipeselwakeup(cpipe);
1261
1262         /*
1263          * If the other side is blocked, wake it up saying that
1264          * we want to close it down.
1265          */
1266         while (cpipe->pipe_busy) {
1267                 wakeup(cpipe);
1268                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1269                 tsleep(cpipe, 0, "pipecl", 0);
1270         }
1271
1272         /*
1273          * Disconnect from peer
1274          */
1275         if ((ppipe = cpipe->pipe_peer) != NULL) {
1276                 pipeselwakeup(ppipe);
1277
1278                 ppipe->pipe_state |= PIPE_EOF;
1279                 wakeup(ppipe);
1280                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1281                 ppipe->pipe_peer = NULL;
1282         }
1283
1284         if (cpipe->pipe_kva) {
1285                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1286                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1287                 cpipe->pipe_kva = NULL;
1288         }
1289
1290         /*
1291          * free or cache resources
1292          */
1293         gd = mycpu;
1294         if (gd->gd_pipeqcount >= pipe_maxcache ||
1295             cpipe->pipe_buffer.size != PIPE_SIZE
1296         ) {
1297                 pipe_free_kmem(cpipe);
1298                 free(cpipe, M_PIPE);
1299         } else {
1300                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1301                         cpipe->pipe_map.xio_bytes == 0 &&
1302                         cpipe->pipe_map.xio_offset == 0);
1303                 cpipe->pipe_state = 0;
1304                 cpipe->pipe_busy = 0;
1305                 cpipe->pipe_peer = gd->gd_pipeq;
1306                 gd->gd_pipeq = cpipe;
1307                 ++gd->gd_pipeqcount;
1308         }
1309 }
1310
1311 /*ARGSUSED*/
1312 static int
1313 pipe_kqfilter(struct file *fp, struct knote *kn)
1314 {
1315         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1316
1317         switch (kn->kn_filter) {
1318         case EVFILT_READ:
1319                 kn->kn_fop = &pipe_rfiltops;
1320                 break;
1321         case EVFILT_WRITE:
1322                 kn->kn_fop = &pipe_wfiltops;
1323                 cpipe = cpipe->pipe_peer;
1324                 if (cpipe == NULL)
1325                         /* other end of pipe has been closed */
1326                         return (EPIPE);
1327                 break;
1328         default:
1329                 return (1);
1330         }
1331         kn->kn_hook = (caddr_t)cpipe;
1332
1333         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1334         return (0);
1335 }
1336
1337 static void
1338 filt_pipedetach(struct knote *kn)
1339 {
1340         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1341
1342         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1343 }
1344
1345 /*ARGSUSED*/
1346 static int
1347 filt_piperead(struct knote *kn, long hint)
1348 {
1349         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1350         struct pipe *wpipe = rpipe->pipe_peer;
1351
1352         kn->kn_data = rpipe->pipe_buffer.cnt;
1353         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1354                 kn->kn_data = rpipe->pipe_map.xio_bytes;
1355
1356         if ((rpipe->pipe_state & PIPE_EOF) ||
1357             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1358                 kn->kn_flags |= EV_EOF;
1359                 return (1);
1360         }
1361         return (kn->kn_data > 0);
1362 }
1363
1364 /*ARGSUSED*/
1365 static int
1366 filt_pipewrite(struct knote *kn, long hint)
1367 {
1368         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1369         struct pipe *wpipe = rpipe->pipe_peer;
1370
1371         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1372                 kn->kn_data = 0;
1373                 kn->kn_flags |= EV_EOF;
1374                 return (1);
1375         }
1376         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1377         if (wpipe->pipe_state & PIPE_DIRECTW)
1378                 kn->kn_data = 0;
1379
1380         return (kn->kn_data >= PIPE_BUF);
1381 }