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 1.32 2005/09/02 07:16:58 hsu 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, struct thread *td);
  97 static int pipe_write (struct file *fp, struct uio *uio,
  98                 struct ucred *cred, int flags, struct thread *td);
  99 static int pipe_close (struct file *fp, struct thread *td);
 100 static int pipe_shutdown (struct file *fp, int how, struct thread *td);
 101 static int pipe_poll (struct file *fp, int events, struct ucred *cred,
 102                 struct thread *td);
 103 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 104 static int pipe_stat (struct file *fp, struct stat *sb, struct thread *td);
 105 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct thread *td);
 106
 107 static struct fileops pipeops = {
 108         NULL,   /* port */
 109         NULL,   /* clone */
 110         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 111         pipe_stat, pipe_close, pipe_shutdown
 112 };
 113
 114 static void     filt_pipedetach(struct knote *kn);
 115 static int      filt_piperead(struct knote *kn, long hint);
 116 static int      filt_pipewrite(struct knote *kn, long hint);
 117
 118 static struct filterops pipe_rfiltops =
 119         { 1, NULL, filt_pipedetach, filt_piperead };
 120 static struct filterops pipe_wfiltops =
 121         { 1, NULL, filt_pipedetach, filt_pipewrite };
 122
 123 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 124
 125 /*
 126  * Default pipe buffer size(s), this can be kind-of large now because pipe
 127  * space is pageable.  The pipe code will try to maintain locality of
 128  * reference for performance reasons, so small amounts of outstanding I/O
 129  * will not wipe the cache.
 130  */
 131 #define MINPIPESIZE (PIPE_SIZE/3)
 132 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 133
 134 /*
 135  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 136  * is there so that on large systems, we don't exhaust it.
 137  */
 138 #define MAXPIPEKVA (8*1024*1024)
 139
 140 /*
 141  * Limit for direct transfers, we cannot, of course limit
 142  * the amount of kva for pipes in general though.
 143  */
 144 #define LIMITPIPEKVA (16*1024*1024)
 145
 146 /*
 147  * Limit the number of "big" pipes
 148  */
 149 #define LIMITBIGPIPES   32
 150 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 151
 152 static int pipe_maxbig = LIMITBIGPIPES;
 153 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 154 static int pipe_nbig;
 155 static int pipe_bcache_alloc;
 156 static int pipe_bkmem_alloc;
 157 static int pipe_dwrite_enable = 1;      /* 0:copy, 1:kmem/sfbuf 2:force */
 158 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 159                                         /* 3:sfbuf_dmap w/ forced invlpg */
 160
 161 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 162 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 163         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 164 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 165         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 166 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 167         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 168 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 169         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 170 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 171         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
 172         "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 173 #if !defined(NO_PIPE_SYSCTL_STATS)
 174 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 175         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 176 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 177         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 178 #endif
 179
 180 static void pipeclose (struct pipe *cpipe);
 181 static void pipe_free_kmem (struct pipe *cpipe);
 182 static int pipe_create (struct pipe **cpipep);
 183 static __inline int pipelock (struct pipe *cpipe, int catch);
 184 static __inline void pipeunlock (struct pipe *cpipe);
 185 static __inline void pipeselwakeup (struct pipe *cpipe);
 186 #ifndef PIPE_NODIRECT
 187 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 188 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 189 static void pipe_clone_write_buffer (struct pipe *wpipe);
 190 #endif
 191 static int pipespace (struct pipe *cpipe, int size);
 192
 193 /*
 194  * The pipe system call for the DTYPE_PIPE type of pipes
 195  *
 196  * pipe_ARgs(int dummy)
 197  */
 198
 199 /* ARGSUSED */
 200 int
 201 pipe(struct pipe_args *uap)
 202 {
 203         struct thread *td = curthread;
 204         struct proc *p = td->td_proc;
 205         struct filedesc *fdp;
 206         struct file *rf, *wf;
 207         struct pipe *rpipe, *wpipe;
 208         int fd1, fd2, error;
 209
 210         KKASSERT(p);
 211         fdp = p->p_fd;
 212
 213         rpipe = wpipe = NULL;
 214         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 215                 pipeclose(rpipe);
 216                 pipeclose(wpipe);
 217                 return (ENFILE);
 218         }
 219
 220         rpipe->pipe_state |= PIPE_DIRECTOK;
 221         wpipe->pipe_state |= PIPE_DIRECTOK;
 222
 223         /*
 224          * Select the direct-map features to use for this pipe.  Since the
 225          * sysctl's can change on the fly we record the settings when the
 226          * pipe is created.
 227          *
 228          * Generally speaking the system will default to what we consider
 229          * to be the best-balanced and most stable option.  Right now this
 230          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 231          * moment.
 232          */
 233         wpipe->pipe_feature = PIPE_COPY;
 234         if (pipe_dwrite_enable) {
 235                 switch(pipe_dwrite_sfbuf) {
 236                 case 0:
 237                         wpipe->pipe_feature = PIPE_KMEM;
 238                         break;
 239                 case 1:
 240                         wpipe->pipe_feature = PIPE_SFBUF1;
 241                         break;
 242                 case 2:
 243                 case 3:
 244                         wpipe->pipe_feature = PIPE_SFBUF2;
 245                         break;
 246                 }
 247         }
 248         rpipe->pipe_feature = wpipe->pipe_feature;
 249
 250         error = falloc(p, &rf, &fd1);
 251         if (error) {
 252                 pipeclose(rpipe);
 253                 pipeclose(wpipe);
 254                 return (error);
 255         }
 256         uap->sysmsg_fds[0] = fd1;
 257
 258         /*
 259          * Warning: once we've gotten past allocation of the fd for the
 260          * read-side, we can only drop the read side via fdrop() in order
 261          * to avoid races against processes which manage to dup() the read
 262          * side while we are blocked trying to allocate the write side.
 263          */
 264         rf->f_type = DTYPE_PIPE;
 265         rf->f_flag = FREAD | FWRITE;
 266         rf->f_ops = &pipeops;
 267         rf->f_data = rpipe;
 268         error = falloc(p, &wf, &fd2);
 269         if (error) {
 270                 if (fdp->fd_files[fd1].fp == rf) {
 271                         funsetfd(fdp, fd1);
 272                         fdrop(rf, td);
 273                 }
 274                 fdrop(rf, td);
 275                 /* rpipe has been closed by fdrop(). */
 276                 pipeclose(wpipe);
 277                 return (error);
 278         }
 279         wf->f_type = DTYPE_PIPE;
 280         wf->f_flag = FREAD | FWRITE;
 281         wf->f_ops = &pipeops;
 282         wf->f_data = wpipe;
 283         uap->sysmsg_fds[1] = fd2;
 284
 285         rpipe->pipe_peer = wpipe;
 286         wpipe->pipe_peer = rpipe;
 287         fdrop(rf, td);
 288         fdrop(wf, td);
 289
 290         return (0);
 291 }
 292
 293 /*
 294  * Allocate kva for pipe circular buffer, the space is pageable
 295  * This routine will 'realloc' the size of a pipe safely, if it fails
 296  * it will retain the old buffer.
 297  * If it fails it will return ENOMEM.
 298  */
 299 static int
 300 pipespace(struct pipe *cpipe, int size)
 301 {
 302         struct vm_object *object;
 303         caddr_t buffer;
 304         int npages, error;
 305
 306         npages = round_page(size) / PAGE_SIZE;
 307         object = cpipe->pipe_buffer.object;
 308
 309         /*
 310          * [re]create the object if necessary and reserve space for it
 311          * in the kernel_map.  The object and memory are pageable.  On
 312          * success, free the old resources before assigning the new
 313          * ones.
 314          */
 315         if (object == NULL || object->size != npages) {
 316                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 317                 buffer = (caddr_t) vm_map_min(kernel_map);
 318
 319                 error = vm_map_find(kernel_map, object, 0,
 320                         (vm_offset_t *) &buffer, size, 1,
 321                         VM_PROT_ALL, VM_PROT_ALL, 0);
 322
 323                 if (error != KERN_SUCCESS) {
 324                         vm_object_deallocate(object);
 325                         return (ENOMEM);
 326                 }
 327                 pipe_free_kmem(cpipe);
 328                 cpipe->pipe_buffer.object = object;
 329                 cpipe->pipe_buffer.buffer = buffer;
 330                 cpipe->pipe_buffer.size = size;
 331                 ++pipe_bkmem_alloc;
 332         } else {
 333                 ++pipe_bcache_alloc;
 334         }
 335         cpipe->pipe_buffer.in = 0;
 336         cpipe->pipe_buffer.out = 0;
 337         cpipe->pipe_buffer.cnt = 0;
 338         return (0);
 339 }
 340
 341 /*
 342  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 343  * our per-cpu cache if possible.  For now make sure it is sized for the
 344  * smaller PIPE_SIZE default.
 345  */
 346 static int
 347 pipe_create(cpipep)
 348         struct pipe **cpipep;
 349 {
 350         globaldata_t gd = mycpu;
 351         struct pipe *cpipe;
 352         int error;
 353
 354         if ((cpipe = gd->gd_pipeq) != NULL) {
 355                 gd->gd_pipeq = cpipe->pipe_peer;
 356                 --gd->gd_pipeqcount;
 357                 cpipe->pipe_peer = NULL;
 358         } else {
 359                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 360         }
 361         *cpipep = cpipe;
 362         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 363                 return (error);
 364         vfs_timestamp(&cpipe->pipe_ctime);
 365         cpipe->pipe_atime = cpipe->pipe_ctime;
 366         cpipe->pipe_mtime = cpipe->pipe_ctime;
 367         return (0);
 368 }
 369
 370
 371 /*
 372  * lock a pipe for I/O, blocking other access
 373  */
 374 static __inline int
 375 pipelock(cpipe, catch)
 376         struct pipe *cpipe;
 377         int catch;
 378 {
 379         int error;
 380
 381         while (cpipe->pipe_state & PIPE_LOCK) {
 382                 cpipe->pipe_state |= PIPE_LWANT;
 383                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 384                 if (error != 0)
 385                         return (error);
 386         }
 387         cpipe->pipe_state |= PIPE_LOCK;
 388         return (0);
 389 }
 390
 391 /*
 392  * unlock a pipe I/O lock
 393  */
 394 static __inline void
 395 pipeunlock(cpipe)
 396         struct pipe *cpipe;
 397 {
 398
 399         cpipe->pipe_state &= ~PIPE_LOCK;
 400         if (cpipe->pipe_state & PIPE_LWANT) {
 401                 cpipe->pipe_state &= ~PIPE_LWANT;
 402                 wakeup(cpipe);
 403         }
 404 }
 405
 406 static __inline void
 407 pipeselwakeup(cpipe)
 408         struct pipe *cpipe;
 409 {
 410
 411         if (cpipe->pipe_state & PIPE_SEL) {
 412                 cpipe->pipe_state &= ~PIPE_SEL;
 413                 selwakeup(&cpipe->pipe_sel);
 414         }
 415         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 416                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 417         KNOTE(&cpipe->pipe_sel.si_note, 0);
 418 }
 419
 420 /* ARGSUSED */
 421 static int
 422 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
 423         int flags, struct thread *td)
 424 {
 425         struct pipe *rpipe = (struct pipe *) fp->f_data;
 426         int error;
 427         int nread = 0;
 428         u_int size;
 429
 430         ++rpipe->pipe_busy;
 431         error = pipelock(rpipe, 1);
 432         if (error)
 433                 goto unlocked_error;
 434
 435         while (uio->uio_resid) {
 436                 caddr_t va;
 437
 438                 if (rpipe->pipe_buffer.cnt > 0) {
 439                         /*
 440                          * normal pipe buffer receive
 441                          */
 442                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 443                         if (size > rpipe->pipe_buffer.cnt)
 444                                 size = rpipe->pipe_buffer.cnt;
 445                         if (size > (u_int) uio->uio_resid)
 446                                 size = (u_int) uio->uio_resid;
 447
 448                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 449                                         size, uio);
 450                         if (error)
 451                                 break;
 452
 453                         rpipe->pipe_buffer.out += size;
 454                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 455                                 rpipe->pipe_buffer.out = 0;
 456
 457                         rpipe->pipe_buffer.cnt -= size;
 458
 459                         /*
 460                          * If there is no more to read in the pipe, reset
 461                          * its pointers to the beginning.  This improves
 462                          * cache hit stats.
 463                          */
 464                         if (rpipe->pipe_buffer.cnt == 0) {
 465                                 rpipe->pipe_buffer.in = 0;
 466                                 rpipe->pipe_buffer.out = 0;
 467                         }
 468                         nread += size;
 469 #ifndef PIPE_NODIRECT
 470                 } else if (rpipe->pipe_kva &&
 471                            rpipe->pipe_feature == PIPE_KMEM &&
 472                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 473                                == PIPE_DIRECTW
 474                 ) {
 475                         /*
 476                          * Direct copy using source-side kva mapping
 477                          */
 478                         size = rpipe->pipe_map.xio_bytes -
 479                                 rpipe->pipe_buffer.out;
 480                         if (size > (u_int)uio->uio_resid)
 481                                 size = (u_int)uio->uio_resid;
 482                         va = (caddr_t)rpipe->pipe_kva +
 483                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 484                         error = uiomove(va, size, uio);
 485                         if (error)
 486                                 break;
 487                         nread += size;
 488                         rpipe->pipe_buffer.out += size;
 489                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 490                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 491                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 492                                 /* reset out index for copy mode */
 493                                 rpipe->pipe_buffer.out = 0;
 494                                 wakeup(rpipe);
 495                         }
 496                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 497                            rpipe->pipe_kva &&
 498                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 499                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 500                                == PIPE_DIRECTW
 501                 ) {
 502                         /*
 503                          * Direct copy, bypassing a kernel buffer.  We cannot
 504                          * mess with the direct-write buffer until
 505                          * PIPE_DIRECTIP is cleared.  In order to prevent
 506                          * the pipe_write code from racing itself in
 507                          * direct_write, we set DIRECTIP when we clear
 508                          * DIRECTW after we have exhausted the buffer.
 509                          */
 510                         if (pipe_dwrite_sfbuf == 3)
 511                                 rpipe->pipe_kvamask = 0;
 512                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 513                                     rpipe->pipe_map.xio_npages,
 514                                     &rpipe->pipe_kvamask);
 515                         size = rpipe->pipe_map.xio_bytes -
 516                                 rpipe->pipe_buffer.out;
 517                         if (size > (u_int)uio->uio_resid)
 518                                 size = (u_int)uio->uio_resid;
 519                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 520                         error = uiomove(va, size, uio);
 521                         if (error)
 522                                 break;
 523                         nread += size;
 524                         rpipe->pipe_buffer.out += size;
 525                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 526                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 527                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 528                                 /* reset out index for copy mode */
 529                                 rpipe->pipe_buffer.out = 0;
 530                                 wakeup(rpipe);
 531                         }
 532                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 533                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 534                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 535                                 == PIPE_DIRECTW
 536                 ) {
 537                         /*
 538                          * Direct copy, bypassing a kernel buffer.  We cannot
 539                          * mess with the direct-write buffer until
 540                          * PIPE_DIRECTIP is cleared.  In order to prevent
 541                          * the pipe_write code from racing itself in
 542                          * direct_write, we set DIRECTIP when we clear
 543                          * DIRECTW after we have exhausted the buffer.
 544                          */
 545                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 546                         if (error)
 547                                 break;
 548                         nread += size;
 549                         rpipe->pipe_buffer.out += size;
 550                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 551                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 552                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 553                                 /* reset out index for copy mode */
 554                                 rpipe->pipe_buffer.out = 0;
 555                                 wakeup(rpipe);
 556                         }
 557 #endif
 558                 } else {
 559                         /*
 560                          * detect EOF condition
 561                          * read returns 0 on EOF, no need to set error
 562                          */
 563                         if (rpipe->pipe_state & PIPE_EOF)
 564                                 break;
 565
 566                         /*
 567                          * If the "write-side" has been blocked, wake it up now.
 568                          */
 569                         if (rpipe->pipe_state & PIPE_WANTW) {
 570                                 rpipe->pipe_state &= ~PIPE_WANTW;
 571                                 wakeup(rpipe);
 572                         }
 573
 574                         /*
 575                          * Break if some data was read.
 576                          */
 577                         if (nread > 0)
 578                                 break;
 579
 580                         /*
 581                          * Unlock the pipe buffer for our remaining
 582                          * processing.  We will either break out with an
 583                          * error or we will sleep and relock to loop.
 584                          */
 585                         pipeunlock(rpipe);
 586
 587                         /*
 588                          * Handle non-blocking mode operation or
 589                          * wait for more data.
 590                          */
 591                         if (fp->f_flag & FNONBLOCK) {
 592                                 error = EAGAIN;
 593                         } else {
 594                                 rpipe->pipe_state |= PIPE_WANTR;
 595                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 596                                     "piperd", 0)) == 0) {
 597                                         error = pipelock(rpipe, 1);
 598                                 }
 599                         }
 600                         if (error)
 601                                 goto unlocked_error;
 602                 }
 603         }
 604         pipeunlock(rpipe);
 605
 606         if (error == 0)
 607                 vfs_timestamp(&rpipe->pipe_atime);
 608 unlocked_error:
 609         --rpipe->pipe_busy;
 610
 611         /*
 612          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 613          */
 614         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 615                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 616                 wakeup(rpipe);
 617         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 618                 /*
 619                  * Handle write blocking hysteresis.
 620                  */
 621                 if (rpipe->pipe_state & PIPE_WANTW) {
 622                         rpipe->pipe_state &= ~PIPE_WANTW;
 623                         wakeup(rpipe);
 624                 }
 625         }
 626
 627         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 628                 pipeselwakeup(rpipe);
 629         return (error);
 630 }
 631
 632 #ifndef PIPE_NODIRECT
 633 /*
 634  * Map the sending processes' buffer into kernel space and wire it.
 635  * This is similar to a physical write operation.
 636  */
 637 static int
 638 pipe_build_write_buffer(wpipe, uio)
 639         struct pipe *wpipe;
 640         struct uio *uio;
 641 {
 642         int error;
 643         u_int size;
 644
 645         size = (u_int) uio->uio_iov->iov_len;
 646         if (size > wpipe->pipe_buffer.size)
 647                 size = wpipe->pipe_buffer.size;
 648
 649         if (uio->uio_segflg == UIO_SYSSPACE) {
 650                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 651                                         size);
 652         } else {
 653                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 654                                         size, XIOF_READ);
 655         }
 656         wpipe->pipe_buffer.out = 0;
 657         if (error)
 658                 return(error);
 659
 660         /*
 661          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 662          * map the pages on the target while KMEM maps the pages now.
 663          */
 664         switch(wpipe->pipe_feature) {
 665         case PIPE_KMEM:
 666         case PIPE_SFBUF2:
 667                 if (wpipe->pipe_kva == NULL) {
 668                         wpipe->pipe_kva =
 669                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
 670                         wpipe->pipe_kvamask = 0;
 671                 }
 672                 if (wpipe->pipe_feature == PIPE_KMEM) {
 673                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 674                                     wpipe->pipe_map.xio_npages);
 675                 }
 676                 break;
 677         default:
 678                 break;
 679         }
 680
 681         /*
 682          * And update the uio data.  The XIO might have loaded fewer bytes
 683          * then requested so reload 'size'.
 684          */
 685         size = wpipe->pipe_map.xio_bytes;
 686         uio->uio_iov->iov_len -= size;
 687         uio->uio_iov->iov_base += size;
 688         if (uio->uio_iov->iov_len == 0)
 689                 uio->uio_iov++;
 690         uio->uio_resid -= size;
 691         uio->uio_offset += size;
 692         return (0);
 693 }
 694
 695 /*
 696  * In the case of a signal, the writing process might go away.  This
 697  * code copies the data into the circular buffer so that the source
 698  * pages can be freed without loss of data.
 699  *
 700  * Note that in direct mode pipe_buffer.out is used to track the
 701  * XIO offset.  We are converting the direct mode into buffered mode
 702  * which changes the meaning of pipe_buffer.out.
 703  */
 704 static void
 705 pipe_clone_write_buffer(wpipe)
 706         struct pipe *wpipe;
 707 {
 708         int size;
 709         int offset;
 710
 711         offset = wpipe->pipe_buffer.out;
 712         size = wpipe->pipe_map.xio_bytes - offset;
 713
 714         KKASSERT(size <= wpipe->pipe_buffer.size);
 715
 716         wpipe->pipe_buffer.in = size;
 717         wpipe->pipe_buffer.out = 0;
 718         wpipe->pipe_buffer.cnt = size;
 719         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 720
 721         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 722         xio_release(&wpipe->pipe_map);
 723         if (wpipe->pipe_kva) {
 724                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 725                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 726                 wpipe->pipe_kva = NULL;
 727         }
 728 }
 729
 730 /*
 731  * This implements the pipe buffer write mechanism.  Note that only
 732  * a direct write OR a normal pipe write can be pending at any given time.
 733  * If there are any characters in the pipe buffer, the direct write will
 734  * be deferred until the receiving process grabs all of the bytes from
 735  * the pipe buffer.  Then the direct mapping write is set-up.
 736  */
 737 static int
 738 pipe_direct_write(wpipe, uio)
 739         struct pipe *wpipe;
 740         struct uio *uio;
 741 {
 742         int error;
 743
 744 retry:
 745         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 746                 if (wpipe->pipe_state & PIPE_WANTR) {
 747                         wpipe->pipe_state &= ~PIPE_WANTR;
 748                         wakeup(wpipe);
 749                 }
 750                 wpipe->pipe_state |= PIPE_WANTW;
 751                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 752                 if (error)
 753                         goto error2;
 754                 if (wpipe->pipe_state & PIPE_EOF) {
 755                         error = EPIPE;
 756                         goto error2;
 757                 }
 758         }
 759         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 760         if (wpipe->pipe_buffer.cnt > 0) {
 761                 if (wpipe->pipe_state & PIPE_WANTR) {
 762                         wpipe->pipe_state &= ~PIPE_WANTR;
 763                         wakeup(wpipe);
 764                 }
 765
 766                 wpipe->pipe_state |= PIPE_WANTW;
 767                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 768                 if (error)
 769                         goto error2;
 770                 if (wpipe->pipe_state & PIPE_EOF) {
 771                         error = EPIPE;
 772                         goto error2;
 773                 }
 774                 goto retry;
 775         }
 776
 777         /*
 778          * Build our direct-write buffer
 779          */
 780         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 781         error = pipe_build_write_buffer(wpipe, uio);
 782         if (error)
 783                 goto error1;
 784         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 785
 786         /*
 787          * Wait until the receiver has snarfed the data.  Since we are likely
 788          * going to sleep we optimize the case and yield synchronously,
 789          * possibly avoiding the tsleep().
 790          */
 791         error = 0;
 792         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 793                 if (wpipe->pipe_state & PIPE_EOF) {
 794                         pipelock(wpipe, 0);
 795                         xio_release(&wpipe->pipe_map);
 796                         if (wpipe->pipe_kva) {
 797                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 798                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 799                                 wpipe->pipe_kva = NULL;
 800                         }
 801                         pipeunlock(wpipe);
 802                         pipeselwakeup(wpipe);
 803                         error = EPIPE;
 804                         goto error1;
 805                 }
 806                 if (wpipe->pipe_state & PIPE_WANTR) {
 807                         wpipe->pipe_state &= ~PIPE_WANTR;
 808                         wakeup(wpipe);
 809                 }
 810                 pipeselwakeup(wpipe);
 811                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 812         }
 813         pipelock(wpipe,0);
 814         if (wpipe->pipe_state & PIPE_DIRECTW) {
 815                 /*
 816                  * this bit of trickery substitutes a kernel buffer for
 817                  * the process that might be going away.
 818                  */
 819                 pipe_clone_write_buffer(wpipe);
 820                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 821         } else {
 822                 /*
 823                  * note: The pipe_kva mapping is not qremove'd here.  For
 824                  * legacy PIPE_KMEM mode this constitutes an improvement
 825                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 826                  * mode the kva mapping must not be removed to get the
 827                  * caching benefit.
 828                  *
 829                  * For testing purposes we will give the original algorithm
 830                  * the benefit of the doubt 'what it could have been', and
 831                  * keep the optimization.
 832                  */
 833                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 834                 xio_release(&wpipe->pipe_map);
 835                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 836         }
 837         pipeunlock(wpipe);
 838         return (error);
 839
 840         /*
 841          * Direct-write error, clear the direct write flags.
 842          */
 843 error1:
 844         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 845         /* fallthrough */
 846
 847         /*
 848          * General error, wakeup the other side if it happens to be sleeping.
 849          */
 850 error2:
 851         wakeup(wpipe);
 852         return (error);
 853 }
 854 #endif
 855
 856 static int
 857 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
 858         int flags, struct thread *td)
 859 {
 860         int error = 0;
 861         int orig_resid;
 862         struct pipe *wpipe, *rpipe;
 863
 864         rpipe = (struct pipe *) fp->f_data;
 865         wpipe = rpipe->pipe_peer;
 866
 867         /*
 868          * detect loss of pipe read side, issue SIGPIPE if lost.
 869          */
 870         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 871                 return (EPIPE);
 872         }
 873         ++wpipe->pipe_busy;
 874
 875         /*
 876          * If it is advantageous to resize the pipe buffer, do
 877          * so.
 878          */
 879         if ((uio->uio_resid > PIPE_SIZE) &&
 880                 (pipe_nbig < pipe_maxbig) &&
 881                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 882                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 883                 (wpipe->pipe_buffer.cnt == 0)) {
 884
 885                 if ((error = pipelock(wpipe,1)) == 0) {
 886                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 887                                 pipe_nbig++;
 888                         pipeunlock(wpipe);
 889                 }
 890         }
 891
 892         /*
 893          * If an early error occured unbusy and return, waking up any pending
 894          * readers.
 895          */
 896         if (error) {
 897                 --wpipe->pipe_busy;
 898                 if ((wpipe->pipe_busy == 0) &&
 899                     (wpipe->pipe_state & PIPE_WANT)) {
 900                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 901                         wakeup(wpipe);
 902                 }
 903                 return(error);
 904         }
 905
 906         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 907
 908         orig_resid = uio->uio_resid;
 909
 910         while (uio->uio_resid) {
 911                 int space;
 912
 913 #ifndef PIPE_NODIRECT
 914                 /*
 915                  * If the transfer is large, we can gain performance if
 916                  * we do process-to-process copies directly.
 917                  * If the write is non-blocking, we don't use the
 918                  * direct write mechanism.
 919                  *
 920                  * The direct write mechanism will detect the reader going
 921                  * away on us.
 922                  */
 923                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 924                     pipe_dwrite_enable > 1) &&
 925                     (fp->f_flag & FNONBLOCK) == 0 &&
 926                     pipe_dwrite_enable) {
 927                         error = pipe_direct_write( wpipe, uio);
 928                         if (error)
 929                                 break;
 930                         continue;
 931                 }
 932 #endif
 933
 934                 /*
 935                  * Pipe buffered writes cannot be coincidental with
 936                  * direct writes.  We wait until the currently executing
 937                  * direct write is completed before we start filling the
 938                  * pipe buffer.  We break out if a signal occurs or the
 939                  * reader goes away.
 940                  */
 941         retrywrite:
 942                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 943                         if (wpipe->pipe_state & PIPE_WANTR) {
 944                                 wpipe->pipe_state &= ~PIPE_WANTR;
 945                                 wakeup(wpipe);
 946                         }
 947                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 948                         if (wpipe->pipe_state & PIPE_EOF)
 949                                 break;
 950                         if (error)
 951                                 break;
 952                 }
 953                 if (wpipe->pipe_state & PIPE_EOF) {
 954                         error = EPIPE;
 955                         break;
 956                 }
 957
 958                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 959
 960                 /* Writes of size <= PIPE_BUF must be atomic. */
 961                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 962                         space = 0;
 963
 964                 /*
 965                  * Write to fill, read size handles write hysteresis.  Also
 966                  * additional restrictions can cause select-based non-blocking
 967                  * writes to spin.
 968                  */
 969                 if (space > 0) {
 970                         if ((error = pipelock(wpipe,1)) == 0) {
 971                                 int size;       /* Transfer size */
 972                                 int segsize;    /* first segment to transfer */
 973
 974                                 /*
 975                                  * It is possible for a direct write to
 976                                  * slip in on us... handle it here...
 977                                  */
 978                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 979                                         pipeunlock(wpipe);
 980                                         goto retrywrite;
 981                                 }
 982                                 /*
 983                                  * If a process blocked in uiomove, our
 984                                  * value for space might be bad.
 985                                  *
 986                                  * XXX will we be ok if the reader has gone
 987                                  * away here?
 988                                  */
 989                                 if (space > wpipe->pipe_buffer.size -
 990                                     wpipe->pipe_buffer.cnt) {
 991                                         pipeunlock(wpipe);
 992                                         goto retrywrite;
 993                                 }
 994
 995                                 /*
 996                                  * Transfer size is minimum of uio transfer
 997                                  * and free space in pipe buffer.
 998                                  */
 999                                 if (space > uio->uio_resid)
1000                                         size = uio->uio_resid;
1001                                 else
1002                                         size = space;
1003                                 /*
1004                                  * First segment to transfer is minimum of
1005                                  * transfer size and contiguous space in
1006                                  * pipe buffer.  If first segment to transfer
1007                                  * is less than the transfer size, we've got
1008                                  * a wraparound in the buffer.
1009                                  */
1010                                 segsize = wpipe->pipe_buffer.size -
1011                                         wpipe->pipe_buffer.in;
1012                                 if (segsize > size)
1013                                         segsize = size;
1014
1015                                 /* Transfer first segment */
1016
1017                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1018                                                 segsize, uio);
1019
1020                                 if (error == 0 && segsize < size) {
1021                                         /*
1022                                          * Transfer remaining part now, to
1023                                          * support atomic writes.  Wraparound
1024                                          * happened.
1025                                          */
1026                                         if (wpipe->pipe_buffer.in + segsize !=
1027                                             wpipe->pipe_buffer.size)
1028                                                 panic("Expected pipe buffer wraparound disappeared");
1029
1030                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1031                                                         size - segsize, uio);
1032                                 }
1033                                 if (error == 0) {
1034                                         wpipe->pipe_buffer.in += size;
1035                                         if (wpipe->pipe_buffer.in >=
1036                                             wpipe->pipe_buffer.size) {
1037                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1038                                                         panic("Expected wraparound bad");
1039                                                 wpipe->pipe_buffer.in = size - segsize;
1040                                         }
1041
1042                                         wpipe->pipe_buffer.cnt += size;
1043                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1044                                                 panic("Pipe buffer overflow");
1045
1046                                 }
1047                                 pipeunlock(wpipe);
1048                         }
1049                         if (error)
1050                                 break;
1051
1052                 } else {
1053                         /*
1054                          * If the "read-side" has been blocked, wake it up now
1055                          * and yield to let it drain synchronously rather
1056                          * then block.
1057                          */
1058                         if (wpipe->pipe_state & PIPE_WANTR) {
1059                                 wpipe->pipe_state &= ~PIPE_WANTR;
1060                                 wakeup(wpipe);
1061                         }
1062
1063                         /*
1064                          * don't block on non-blocking I/O
1065                          */
1066                         if (fp->f_flag & FNONBLOCK) {
1067                                 error = EAGAIN;
1068                                 break;
1069                         }
1070
1071                         /*
1072                          * We have no more space and have something to offer,
1073                          * wake up select/poll.
1074                          */
1075                         pipeselwakeup(wpipe);
1076
1077                         wpipe->pipe_state |= PIPE_WANTW;
1078                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1079                         if (error != 0)
1080                                 break;
1081                         /*
1082                          * If read side wants to go away, we just issue a signal
1083                          * to ourselves.
1084                          */
1085                         if (wpipe->pipe_state & PIPE_EOF) {
1086                                 error = EPIPE;
1087                                 break;
1088                         }
1089                 }
1090         }
1091
1092         --wpipe->pipe_busy;
1093
1094         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1095                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1096                 wakeup(wpipe);
1097         } else if (wpipe->pipe_buffer.cnt > 0) {
1098                 /*
1099                  * If we have put any characters in the buffer, we wake up
1100                  * the reader.
1101                  */
1102                 if (wpipe->pipe_state & PIPE_WANTR) {
1103                         wpipe->pipe_state &= ~PIPE_WANTR;
1104                         wakeup(wpipe);
1105                 }
1106         }
1107
1108         /*
1109          * Don't return EPIPE if I/O was successful
1110          */
1111         if ((wpipe->pipe_buffer.cnt == 0) &&
1112             (uio->uio_resid == 0) &&
1113             (error == EPIPE)) {
1114                 error = 0;
1115         }
1116
1117         if (error == 0)
1118                 vfs_timestamp(&wpipe->pipe_mtime);
1119
1120         /*
1121          * We have something to offer,
1122          * wake up select/poll.
1123          */
1124         if (wpipe->pipe_buffer.cnt)
1125                 pipeselwakeup(wpipe);
1126
1127         return (error);
1128 }
1129
1130 /*
1131  * we implement a very minimal set of ioctls for compatibility with sockets.
1132  */
1133 int
1134 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1135 {
1136         struct pipe *mpipe = (struct pipe *)fp->f_data;
1137
1138         switch (cmd) {
1139
1140         case FIONBIO:
1141                 return (0);
1142
1143         case FIOASYNC:
1144                 if (*(int *)data) {
1145                         mpipe->pipe_state |= PIPE_ASYNC;
1146                 } else {
1147                         mpipe->pipe_state &= ~PIPE_ASYNC;
1148                 }
1149                 return (0);
1150
1151         case FIONREAD:
1152                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1153                         *(int *)data = mpipe->pipe_map.xio_bytes -
1154                                         mpipe->pipe_buffer.out;
1155                 } else {
1156                         *(int *)data = mpipe->pipe_buffer.cnt;
1157                 }
1158                 return (0);
1159
1160         case FIOSETOWN:
1161                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1162
1163         case FIOGETOWN:
1164                 *(int *)data = fgetown(mpipe->pipe_sigio);
1165                 return (0);
1166
1167         /* This is deprecated, FIOSETOWN should be used instead. */
1168         case TIOCSPGRP:
1169                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1170
1171         /* This is deprecated, FIOGETOWN should be used instead. */
1172         case TIOCGPGRP:
1173                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1174                 return (0);
1175
1176         }
1177         return (ENOTTY);
1178 }
1179
1180 int
1181 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1182 {
1183         struct pipe *rpipe = (struct pipe *)fp->f_data;
1184         struct pipe *wpipe;
1185         int revents = 0;
1186
1187         wpipe = rpipe->pipe_peer;
1188         if (events & (POLLIN | POLLRDNORM))
1189                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1190                     (rpipe->pipe_buffer.cnt > 0) ||
1191                     (rpipe->pipe_state & PIPE_EOF))
1192                         revents |= events & (POLLIN | POLLRDNORM);
1193
1194         if (events & (POLLOUT | POLLWRNORM))
1195                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1196                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1197                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1198                         revents |= events & (POLLOUT | POLLWRNORM);
1199
1200         if ((rpipe->pipe_state & PIPE_EOF) ||
1201             (wpipe == NULL) ||
1202             (wpipe->pipe_state & PIPE_EOF))
1203                 revents |= POLLHUP;
1204
1205         if (revents == 0) {
1206                 if (events & (POLLIN | POLLRDNORM)) {
1207                         selrecord(td, &rpipe->pipe_sel);
1208                         rpipe->pipe_state |= PIPE_SEL;
1209                 }
1210
1211                 if (events & (POLLOUT | POLLWRNORM)) {
1212                         selrecord(td, &wpipe->pipe_sel);
1213                         wpipe->pipe_state |= PIPE_SEL;
1214                 }
1215         }
1216
1217         return (revents);
1218 }
1219
1220 static int
1221 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1222 {
1223         struct pipe *pipe = (struct pipe *)fp->f_data;
1224
1225         bzero((caddr_t)ub, sizeof(*ub));
1226         ub->st_mode = S_IFIFO;
1227         ub->st_blksize = pipe->pipe_buffer.size;
1228         ub->st_size = pipe->pipe_buffer.cnt;
1229         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1230                 ub->st_size = pipe->pipe_map.xio_bytes -
1231                                 pipe->pipe_buffer.out;
1232         }
1233         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1234         ub->st_atimespec = pipe->pipe_atime;
1235         ub->st_mtimespec = pipe->pipe_mtime;
1236         ub->st_ctimespec = pipe->pipe_ctime;
1237         /*
1238          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1239          * st_flags, st_gen.
1240          * XXX (st_dev, st_ino) should be unique.
1241          */
1242         return (0);
1243 }
1244
1245 /* ARGSUSED */
1246 static int
1247 pipe_close(struct file *fp, struct thread *td)
1248 {
1249         struct pipe *cpipe = (struct pipe *)fp->f_data;
1250
1251         fp->f_ops = &badfileops;
1252         fp->f_data = NULL;
1253         funsetown(cpipe->pipe_sigio);
1254         pipeclose(cpipe);
1255         return (0);
1256 }
1257
1258 /*
1259  * Shutdown one or both directions of a full-duplex pipe.
1260  */
1261 /* ARGSUSED */
1262 static int
1263 pipe_shutdown(struct file *fp, int how, struct thread *td)
1264 {
1265         struct pipe *rpipe = (struct pipe *)fp->f_data;
1266         struct pipe *wpipe;
1267         int error = EPIPE;
1268
1269         switch(how) {
1270         case SHUT_RDWR:
1271         case SHUT_RD:
1272                 if (rpipe) {
1273                         rpipe->pipe_state |= PIPE_EOF;
1274                         pipeselwakeup(rpipe);
1275                         if (rpipe->pipe_busy)
1276                                 wakeup(rpipe);
1277                         error = 0;
1278                 }
1279                 if (how == SHUT_RD)
1280                         break;
1281                 /* fall through */
1282         case SHUT_WR:
1283                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1284                         wpipe->pipe_state |= PIPE_EOF;
1285                         pipeselwakeup(wpipe);
1286                         if (wpipe->pipe_busy)
1287                                 wakeup(wpipe);
1288                         error = 0;
1289                 }
1290         }
1291         return (error);
1292 }
1293
1294 static void
1295 pipe_free_kmem(struct pipe *cpipe)
1296 {
1297         if (cpipe->pipe_buffer.buffer != NULL) {
1298                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1299                         --pipe_nbig;
1300                 kmem_free(kernel_map,
1301                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1302                         cpipe->pipe_buffer.size);
1303                 cpipe->pipe_buffer.buffer = NULL;
1304                 cpipe->pipe_buffer.object = NULL;
1305         }
1306 #ifndef PIPE_NODIRECT
1307         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1308                 cpipe->pipe_map.xio_offset == 0 &&
1309                 cpipe->pipe_map.xio_npages == 0);
1310 #endif
1311 }
1312
1313 /*
1314  * shutdown the pipe
1315  */
1316 static void
1317 pipeclose(struct pipe *cpipe)
1318 {
1319         globaldata_t gd;
1320         struct pipe *ppipe;
1321
1322         if (cpipe == NULL)
1323                 return;
1324
1325         pipeselwakeup(cpipe);
1326
1327         /*
1328          * If the other side is blocked, wake it up saying that
1329          * we want to close it down.
1330          */
1331         while (cpipe->pipe_busy) {
1332                 wakeup(cpipe);
1333                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1334                 tsleep(cpipe, 0, "pipecl", 0);
1335         }
1336
1337         /*
1338          * Disconnect from peer
1339          */
1340         if ((ppipe = cpipe->pipe_peer) != NULL) {
1341                 pipeselwakeup(ppipe);
1342
1343                 ppipe->pipe_state |= PIPE_EOF;
1344                 wakeup(ppipe);
1345                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1346                 ppipe->pipe_peer = NULL;
1347         }
1348
1349         if (cpipe->pipe_kva) {
1350                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1351                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1352                 cpipe->pipe_kva = NULL;
1353         }
1354
1355         /*
1356          * free or cache resources
1357          */
1358         gd = mycpu;
1359         if (gd->gd_pipeqcount >= pipe_maxcache ||
1360             cpipe->pipe_buffer.size != PIPE_SIZE
1361         ) {
1362                 pipe_free_kmem(cpipe);
1363                 free(cpipe, M_PIPE);
1364         } else {
1365                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1366                         cpipe->pipe_map.xio_bytes == 0 &&
1367                         cpipe->pipe_map.xio_offset == 0);
1368                 cpipe->pipe_state = 0;
1369                 cpipe->pipe_busy = 0;
1370                 cpipe->pipe_peer = gd->gd_pipeq;
1371                 gd->gd_pipeq = cpipe;
1372                 ++gd->gd_pipeqcount;
1373         }
1374 }
1375
1376 /*ARGSUSED*/
1377 static int
1378 pipe_kqfilter(struct file *fp, struct knote *kn)
1379 {
1380         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1381
1382         switch (kn->kn_filter) {
1383         case EVFILT_READ:
1384                 kn->kn_fop = &pipe_rfiltops;
1385                 break;
1386         case EVFILT_WRITE:
1387                 kn->kn_fop = &pipe_wfiltops;
1388                 cpipe = cpipe->pipe_peer;
1389                 if (cpipe == NULL)
1390                         /* other end of pipe has been closed */
1391                         return (EPIPE);
1392                 break;
1393         default:
1394                 return (1);
1395         }
1396         kn->kn_hook = (caddr_t)cpipe;
1397
1398         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1399         return (0);
1400 }
1401
1402 static void
1403 filt_pipedetach(struct knote *kn)
1404 {
1405         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1406
1407         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1408 }
1409
1410 /*ARGSUSED*/
1411 static int
1412 filt_piperead(struct knote *kn, long hint)
1413 {
1414         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1415         struct pipe *wpipe = rpipe->pipe_peer;
1416
1417         kn->kn_data = rpipe->pipe_buffer.cnt;
1418         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1419                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1420                                 rpipe->pipe_buffer.out;
1421         }
1422
1423         if ((rpipe->pipe_state & PIPE_EOF) ||
1424             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1425                 kn->kn_flags |= EV_EOF;
1426                 return (1);
1427         }
1428         return (kn->kn_data > 0);
1429 }
1430
1431 /*ARGSUSED*/
1432 static int
1433 filt_pipewrite(struct knote *kn, long hint)
1434 {
1435         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1436         struct pipe *wpipe = rpipe->pipe_peer;
1437
1438         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1439                 kn->kn_data = 0;
1440                 kn->kn_flags |= EV_EOF;
1441                 return (1);
1442         }
1443         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1444         if (wpipe->pipe_state & PIPE_DIRECTW)
1445                 kn->kn_data = 0;
1446
1447         return (kn->kn_data >= PIPE_BUF);
1448 }