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 1.34 2006/05/06 06:38:38 dillon Exp $
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
  36  * the receiving process can copy it directly from the pages in the sending
  37  * process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.  PIPE_SIZE is constrained by the
  50  * amount of kernel virtual memory.
  51  */
  52
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/file.h>
  59 #include <sys/filedesc.h>
  60 #include <sys/filio.h>
  61 #include <sys/ttycom.h>
  62 #include <sys/stat.h>
  63 #include <sys/poll.h>
  64 #include <sys/select.h>
  65 #include <sys/signalvar.h>
  66 #include <sys/sysproto.h>
  67 #include <sys/pipe.h>
  68 #include <sys/vnode.h>
  69 #include <sys/uio.h>
  70 #include <sys/event.h>
  71 #include <sys/globaldata.h>
  72 #include <sys/module.h>
  73 #include <sys/malloc.h>
  74 #include <sys/sysctl.h>
  75 #include <sys/socket.h>
  76
  77 #include <vm/vm.h>
  78 #include <vm/vm_param.h>
  79 #include <sys/lock.h>
  80 #include <vm/vm_object.h>
  81 #include <vm/vm_kern.h>
  82 #include <vm/vm_extern.h>
  83 #include <vm/pmap.h>
  84 #include <vm/vm_map.h>
  85 #include <vm/vm_page.h>
  86 #include <vm/vm_zone.h>
  87
  88 #include <sys/file2.h>
  89
  90 #include <machine/cpufunc.h>
  91
  92 /*
  93  * interfaces to the outside world
  94  */
  95 static int pipe_read (struct file *fp, struct uio *uio,
  96                 struct ucred *cred, int flags);
  97 static int pipe_write (struct file *fp, struct uio *uio,
  98                 struct ucred *cred, int flags);
  99 static int pipe_close (struct file *fp);
 100 static int pipe_shutdown (struct file *fp, int how);
 101 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
 102 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 103 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
 104 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred);
 105
 106 static struct fileops pipeops = {
 107         NULL,   /* port */
 108         NULL,   /* clone */
 109         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 110         pipe_stat, pipe_close, pipe_shutdown
 111 };
 112
 113 static void     filt_pipedetach(struct knote *kn);
 114 static int      filt_piperead(struct knote *kn, long hint);
 115 static int      filt_pipewrite(struct knote *kn, long hint);
 116
 117 static struct filterops pipe_rfiltops =
 118         { 1, NULL, filt_pipedetach, filt_piperead };
 119 static struct filterops pipe_wfiltops =
 120         { 1, NULL, filt_pipedetach, filt_pipewrite };
 121
 122 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 123
 124 /*
 125  * Default pipe buffer size(s), this can be kind-of large now because pipe
 126  * space is pageable.  The pipe code will try to maintain locality of
 127  * reference for performance reasons, so small amounts of outstanding I/O
 128  * will not wipe the cache.
 129  */
 130 #define MINPIPESIZE (PIPE_SIZE/3)
 131 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 132
 133 /*
 134  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 135  * is there so that on large systems, we don't exhaust it.
 136  */
 137 #define MAXPIPEKVA (8*1024*1024)
 138
 139 /*
 140  * Limit for direct transfers, we cannot, of course limit
 141  * the amount of kva for pipes in general though.
 142  */
 143 #define LIMITPIPEKVA (16*1024*1024)
 144
 145 /*
 146  * Limit the number of "big" pipes
 147  */
 148 #define LIMITBIGPIPES   32
 149 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 150
 151 static int pipe_maxbig = LIMITBIGPIPES;
 152 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 153 static int pipe_nbig;
 154 static int pipe_bcache_alloc;
 155 static int pipe_bkmem_alloc;
 156 static int pipe_dwrite_enable = 1;      /* 0:copy, 1:kmem/sfbuf 2:force */
 157 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 158                                         /* 3:sfbuf_dmap w/ forced invlpg */
 159
 160 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 161 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 162         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 163 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 164         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 165 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 166         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 167 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 168         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 169 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 170         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
 171         "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 172 #if !defined(NO_PIPE_SYSCTL_STATS)
 173 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 174         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 175 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 176         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 177 #endif
 178
 179 static void pipeclose (struct pipe *cpipe);
 180 static void pipe_free_kmem (struct pipe *cpipe);
 181 static int pipe_create (struct pipe **cpipep);
 182 static __inline int pipelock (struct pipe *cpipe, int catch);
 183 static __inline void pipeunlock (struct pipe *cpipe);
 184 static __inline void pipeselwakeup (struct pipe *cpipe);
 185 #ifndef PIPE_NODIRECT
 186 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 187 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 188 static void pipe_clone_write_buffer (struct pipe *wpipe);
 189 #endif
 190 static int pipespace (struct pipe *cpipe, int size);
 191
 192 /*
 193  * The pipe system call for the DTYPE_PIPE type of pipes
 194  *
 195  * pipe_ARgs(int dummy)
 196  */
 197
 198 /* ARGSUSED */
 199 int
 200 pipe(struct pipe_args *uap)
 201 {
 202         struct thread *td = curthread;
 203         struct proc *p = td->td_proc;
 204         struct filedesc *fdp;
 205         struct file *rf, *wf;
 206         struct pipe *rpipe, *wpipe;
 207         int fd1, fd2, error;
 208
 209         KKASSERT(p);
 210         fdp = p->p_fd;
 211
 212         rpipe = wpipe = NULL;
 213         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 214                 pipeclose(rpipe);
 215                 pipeclose(wpipe);
 216                 return (ENFILE);
 217         }
 218
 219         rpipe->pipe_state |= PIPE_DIRECTOK;
 220         wpipe->pipe_state |= PIPE_DIRECTOK;
 221
 222         /*
 223          * Select the direct-map features to use for this pipe.  Since the
 224          * sysctl's can change on the fly we record the settings when the
 225          * pipe is created.
 226          *
 227          * Generally speaking the system will default to what we consider
 228          * to be the best-balanced and most stable option.  Right now this
 229          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 230          * moment.
 231          */
 232         wpipe->pipe_feature = PIPE_COPY;
 233         if (pipe_dwrite_enable) {
 234                 switch(pipe_dwrite_sfbuf) {
 235                 case 0:
 236                         wpipe->pipe_feature = PIPE_KMEM;
 237                         break;
 238                 case 1:
 239                         wpipe->pipe_feature = PIPE_SFBUF1;
 240                         break;
 241                 case 2:
 242                 case 3:
 243                         wpipe->pipe_feature = PIPE_SFBUF2;
 244                         break;
 245                 }
 246         }
 247         rpipe->pipe_feature = wpipe->pipe_feature;
 248
 249         error = falloc(p, &rf, &fd1);
 250         if (error) {
 251                 pipeclose(rpipe);
 252                 pipeclose(wpipe);
 253                 return (error);
 254         }
 255         uap->sysmsg_fds[0] = fd1;
 256
 257         /*
 258          * Warning: once we've gotten past allocation of the fd for the
 259          * read-side, we can only drop the read side via fdrop() in order
 260          * to avoid races against processes which manage to dup() the read
 261          * side while we are blocked trying to allocate the write side.
 262          */
 263         rf->f_type = DTYPE_PIPE;
 264         rf->f_flag = FREAD | FWRITE;
 265         rf->f_ops = &pipeops;
 266         rf->f_data = rpipe;
 267         error = falloc(p, &wf, &fd2);
 268         if (error) {
 269                 if (fdp->fd_files[fd1].fp == rf) {
 270                         funsetfd(fdp, fd1);
 271                         fdrop(rf);
 272                 }
 273                 fdrop(rf);
 274                 /* rpipe has been closed by fdrop(). */
 275                 pipeclose(wpipe);
 276                 return (error);
 277         }
 278         wf->f_type = DTYPE_PIPE;
 279         wf->f_flag = FREAD | FWRITE;
 280         wf->f_ops = &pipeops;
 281         wf->f_data = wpipe;
 282         uap->sysmsg_fds[1] = fd2;
 283
 284         rpipe->pipe_peer = wpipe;
 285         wpipe->pipe_peer = rpipe;
 286         fdrop(rf);
 287         fdrop(wf);
 288
 289         return (0);
 290 }
 291
 292 /*
 293  * Allocate kva for pipe circular buffer, the space is pageable
 294  * This routine will 'realloc' the size of a pipe safely, if it fails
 295  * it will retain the old buffer.
 296  * If it fails it will return ENOMEM.
 297  */
 298 static int
 299 pipespace(struct pipe *cpipe, int size)
 300 {
 301         struct vm_object *object;
 302         caddr_t buffer;
 303         int npages, error;
 304
 305         npages = round_page(size) / PAGE_SIZE;
 306         object = cpipe->pipe_buffer.object;
 307
 308         /*
 309          * [re]create the object if necessary and reserve space for it
 310          * in the kernel_map.  The object and memory are pageable.  On
 311          * success, free the old resources before assigning the new
 312          * ones.
 313          */
 314         if (object == NULL || object->size != npages) {
 315                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 316                 buffer = (caddr_t) vm_map_min(kernel_map);
 317
 318                 error = vm_map_find(kernel_map, object, 0,
 319                         (vm_offset_t *) &buffer, size, 1,
 320                         VM_PROT_ALL, VM_PROT_ALL, 0);
 321
 322                 if (error != KERN_SUCCESS) {
 323                         vm_object_deallocate(object);
 324                         return (ENOMEM);
 325                 }
 326                 pipe_free_kmem(cpipe);
 327                 cpipe->pipe_buffer.object = object;
 328                 cpipe->pipe_buffer.buffer = buffer;
 329                 cpipe->pipe_buffer.size = size;
 330                 ++pipe_bkmem_alloc;
 331         } else {
 332                 ++pipe_bcache_alloc;
 333         }
 334         cpipe->pipe_buffer.in = 0;
 335         cpipe->pipe_buffer.out = 0;
 336         cpipe->pipe_buffer.cnt = 0;
 337         return (0);
 338 }
 339
 340 /*
 341  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 342  * our per-cpu cache if possible.  For now make sure it is sized for the
 343  * smaller PIPE_SIZE default.
 344  */
 345 static int
 346 pipe_create(cpipep)
 347         struct pipe **cpipep;
 348 {
 349         globaldata_t gd = mycpu;
 350         struct pipe *cpipe;
 351         int error;
 352
 353         if ((cpipe = gd->gd_pipeq) != NULL) {
 354                 gd->gd_pipeq = cpipe->pipe_peer;
 355                 --gd->gd_pipeqcount;
 356                 cpipe->pipe_peer = NULL;
 357         } else {
 358                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 359         }
 360         *cpipep = cpipe;
 361         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 362                 return (error);
 363         vfs_timestamp(&cpipe->pipe_ctime);
 364         cpipe->pipe_atime = cpipe->pipe_ctime;
 365         cpipe->pipe_mtime = cpipe->pipe_ctime;
 366         return (0);
 367 }
 368
 369
 370 /*
 371  * lock a pipe for I/O, blocking other access
 372  */
 373 static __inline int
 374 pipelock(cpipe, catch)
 375         struct pipe *cpipe;
 376         int catch;
 377 {
 378         int error;
 379
 380         while (cpipe->pipe_state & PIPE_LOCK) {
 381                 cpipe->pipe_state |= PIPE_LWANT;
 382                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 383                 if (error != 0)
 384                         return (error);
 385         }
 386         cpipe->pipe_state |= PIPE_LOCK;
 387         return (0);
 388 }
 389
 390 /*
 391  * unlock a pipe I/O lock
 392  */
 393 static __inline void
 394 pipeunlock(cpipe)
 395         struct pipe *cpipe;
 396 {
 397
 398         cpipe->pipe_state &= ~PIPE_LOCK;
 399         if (cpipe->pipe_state & PIPE_LWANT) {
 400                 cpipe->pipe_state &= ~PIPE_LWANT;
 401                 wakeup(cpipe);
 402         }
 403 }
 404
 405 static __inline void
 406 pipeselwakeup(cpipe)
 407         struct pipe *cpipe;
 408 {
 409
 410         if (cpipe->pipe_state & PIPE_SEL) {
 411                 cpipe->pipe_state &= ~PIPE_SEL;
 412                 selwakeup(&cpipe->pipe_sel);
 413         }
 414         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 415                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 416         KNOTE(&cpipe->pipe_sel.si_note, 0);
 417 }
 418
 419 /* ARGSUSED */
 420 static int
 421 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
 422 {
 423         struct pipe *rpipe = (struct pipe *) fp->f_data;
 424         int error;
 425         int nread = 0;
 426         u_int size;
 427
 428         ++rpipe->pipe_busy;
 429         error = pipelock(rpipe, 1);
 430         if (error)
 431                 goto unlocked_error;
 432
 433         while (uio->uio_resid) {
 434                 caddr_t va;
 435
 436                 if (rpipe->pipe_buffer.cnt > 0) {
 437                         /*
 438                          * normal pipe buffer receive
 439                          */
 440                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 441                         if (size > rpipe->pipe_buffer.cnt)
 442                                 size = rpipe->pipe_buffer.cnt;
 443                         if (size > (u_int) uio->uio_resid)
 444                                 size = (u_int) uio->uio_resid;
 445
 446                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 447                                         size, uio);
 448                         if (error)
 449                                 break;
 450
 451                         rpipe->pipe_buffer.out += size;
 452                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 453                                 rpipe->pipe_buffer.out = 0;
 454
 455                         rpipe->pipe_buffer.cnt -= size;
 456
 457                         /*
 458                          * If there is no more to read in the pipe, reset
 459                          * its pointers to the beginning.  This improves
 460                          * cache hit stats.
 461                          */
 462                         if (rpipe->pipe_buffer.cnt == 0) {
 463                                 rpipe->pipe_buffer.in = 0;
 464                                 rpipe->pipe_buffer.out = 0;
 465                         }
 466                         nread += size;
 467 #ifndef PIPE_NODIRECT
 468                 } else if (rpipe->pipe_kva &&
 469                            rpipe->pipe_feature == PIPE_KMEM &&
 470                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 471                                == PIPE_DIRECTW
 472                 ) {
 473                         /*
 474                          * Direct copy using source-side kva mapping
 475                          */
 476                         size = rpipe->pipe_map.xio_bytes -
 477                                 rpipe->pipe_buffer.out;
 478                         if (size > (u_int)uio->uio_resid)
 479                                 size = (u_int)uio->uio_resid;
 480                         va = (caddr_t)rpipe->pipe_kva +
 481                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 482                         error = uiomove(va, size, uio);
 483                         if (error)
 484                                 break;
 485                         nread += size;
 486                         rpipe->pipe_buffer.out += size;
 487                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 488                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 489                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 490                                 /* reset out index for copy mode */
 491                                 rpipe->pipe_buffer.out = 0;
 492                                 wakeup(rpipe);
 493                         }
 494                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 495                            rpipe->pipe_kva &&
 496                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 497                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 498                                == PIPE_DIRECTW
 499                 ) {
 500                         /*
 501                          * Direct copy, bypassing a kernel buffer.  We cannot
 502                          * mess with the direct-write buffer until
 503                          * PIPE_DIRECTIP is cleared.  In order to prevent
 504                          * the pipe_write code from racing itself in
 505                          * direct_write, we set DIRECTIP when we clear
 506                          * DIRECTW after we have exhausted the buffer.
 507                          */
 508                         if (pipe_dwrite_sfbuf == 3)
 509                                 rpipe->pipe_kvamask = 0;
 510                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 511                                     rpipe->pipe_map.xio_npages,
 512                                     &rpipe->pipe_kvamask);
 513                         size = rpipe->pipe_map.xio_bytes -
 514                                 rpipe->pipe_buffer.out;
 515                         if (size > (u_int)uio->uio_resid)
 516                                 size = (u_int)uio->uio_resid;
 517                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 518                         error = uiomove(va, size, uio);
 519                         if (error)
 520                                 break;
 521                         nread += size;
 522                         rpipe->pipe_buffer.out += size;
 523                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 524                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 525                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 526                                 /* reset out index for copy mode */
 527                                 rpipe->pipe_buffer.out = 0;
 528                                 wakeup(rpipe);
 529                         }
 530                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 531                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 532                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 533                                 == PIPE_DIRECTW
 534                 ) {
 535                         /*
 536                          * Direct copy, bypassing a kernel buffer.  We cannot
 537                          * mess with the direct-write buffer until
 538                          * PIPE_DIRECTIP is cleared.  In order to prevent
 539                          * the pipe_write code from racing itself in
 540                          * direct_write, we set DIRECTIP when we clear
 541                          * DIRECTW after we have exhausted the buffer.
 542                          */
 543                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 544                         if (error)
 545                                 break;
 546                         nread += size;
 547                         rpipe->pipe_buffer.out += size;
 548                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 549                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 550                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 551                                 /* reset out index for copy mode */
 552                                 rpipe->pipe_buffer.out = 0;
 553                                 wakeup(rpipe);
 554                         }
 555 #endif
 556                 } else {
 557                         /*
 558                          * detect EOF condition
 559                          * read returns 0 on EOF, no need to set error
 560                          */
 561                         if (rpipe->pipe_state & PIPE_EOF)
 562                                 break;
 563
 564                         /*
 565                          * If the "write-side" has been blocked, wake it up now.
 566                          */
 567                         if (rpipe->pipe_state & PIPE_WANTW) {
 568                                 rpipe->pipe_state &= ~PIPE_WANTW;
 569                                 wakeup(rpipe);
 570                         }
 571
 572                         /*
 573                          * Break if some data was read.
 574                          */
 575                         if (nread > 0)
 576                                 break;
 577
 578                         /*
 579                          * Unlock the pipe buffer for our remaining
 580                          * processing.  We will either break out with an
 581                          * error or we will sleep and relock to loop.
 582                          */
 583                         pipeunlock(rpipe);
 584
 585                         /*
 586                          * Handle non-blocking mode operation or
 587                          * wait for more data.
 588                          */
 589                         if (fp->f_flag & FNONBLOCK) {
 590                                 error = EAGAIN;
 591                         } else {
 592                                 rpipe->pipe_state |= PIPE_WANTR;
 593                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 594                                     "piperd", 0)) == 0) {
 595                                         error = pipelock(rpipe, 1);
 596                                 }
 597                         }
 598                         if (error)
 599                                 goto unlocked_error;
 600                 }
 601         }
 602         pipeunlock(rpipe);
 603
 604         if (error == 0)
 605                 vfs_timestamp(&rpipe->pipe_atime);
 606 unlocked_error:
 607         --rpipe->pipe_busy;
 608
 609         /*
 610          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 611          */
 612         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 613                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 614                 wakeup(rpipe);
 615         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 616                 /*
 617                  * Handle write blocking hysteresis.
 618                  */
 619                 if (rpipe->pipe_state & PIPE_WANTW) {
 620                         rpipe->pipe_state &= ~PIPE_WANTW;
 621                         wakeup(rpipe);
 622                 }
 623         }
 624
 625         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 626                 pipeselwakeup(rpipe);
 627         return (error);
 628 }
 629
 630 #ifndef PIPE_NODIRECT
 631 /*
 632  * Map the sending processes' buffer into kernel space and wire it.
 633  * This is similar to a physical write operation.
 634  */
 635 static int
 636 pipe_build_write_buffer(wpipe, uio)
 637         struct pipe *wpipe;
 638         struct uio *uio;
 639 {
 640         int error;
 641         u_int size;
 642
 643         size = (u_int) uio->uio_iov->iov_len;
 644         if (size > wpipe->pipe_buffer.size)
 645                 size = wpipe->pipe_buffer.size;
 646
 647         if (uio->uio_segflg == UIO_SYSSPACE) {
 648                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 649                                         size);
 650         } else {
 651                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 652                                         size, XIOF_READ);
 653         }
 654         wpipe->pipe_buffer.out = 0;
 655         if (error)
 656                 return(error);
 657
 658         /*
 659          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 660          * map the pages on the target while KMEM maps the pages now.
 661          */
 662         switch(wpipe->pipe_feature) {
 663         case PIPE_KMEM:
 664         case PIPE_SFBUF2:
 665                 if (wpipe->pipe_kva == NULL) {
 666                         wpipe->pipe_kva =
 667                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
 668                         wpipe->pipe_kvamask = 0;
 669                 }
 670                 if (wpipe->pipe_feature == PIPE_KMEM) {
 671                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 672                                     wpipe->pipe_map.xio_npages);
 673                 }
 674                 break;
 675         default:
 676                 break;
 677         }
 678
 679         /*
 680          * And update the uio data.  The XIO might have loaded fewer bytes
 681          * then requested so reload 'size'.
 682          */
 683         size = wpipe->pipe_map.xio_bytes;
 684         uio->uio_iov->iov_len -= size;
 685         uio->uio_iov->iov_base += size;
 686         if (uio->uio_iov->iov_len == 0)
 687                 uio->uio_iov++;
 688         uio->uio_resid -= size;
 689         uio->uio_offset += size;
 690         return (0);
 691 }
 692
 693 /*
 694  * In the case of a signal, the writing process might go away.  This
 695  * code copies the data into the circular buffer so that the source
 696  * pages can be freed without loss of data.
 697  *
 698  * Note that in direct mode pipe_buffer.out is used to track the
 699  * XIO offset.  We are converting the direct mode into buffered mode
 700  * which changes the meaning of pipe_buffer.out.
 701  */
 702 static void
 703 pipe_clone_write_buffer(wpipe)
 704         struct pipe *wpipe;
 705 {
 706         int size;
 707         int offset;
 708
 709         offset = wpipe->pipe_buffer.out;
 710         size = wpipe->pipe_map.xio_bytes - offset;
 711
 712         KKASSERT(size <= wpipe->pipe_buffer.size);
 713
 714         wpipe->pipe_buffer.in = size;
 715         wpipe->pipe_buffer.out = 0;
 716         wpipe->pipe_buffer.cnt = size;
 717         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 718
 719         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 720         xio_release(&wpipe->pipe_map);
 721         if (wpipe->pipe_kva) {
 722                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 723                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 724                 wpipe->pipe_kva = NULL;
 725         }
 726 }
 727
 728 /*
 729  * This implements the pipe buffer write mechanism.  Note that only
 730  * a direct write OR a normal pipe write can be pending at any given time.
 731  * If there are any characters in the pipe buffer, the direct write will
 732  * be deferred until the receiving process grabs all of the bytes from
 733  * the pipe buffer.  Then the direct mapping write is set-up.
 734  */
 735 static int
 736 pipe_direct_write(wpipe, uio)
 737         struct pipe *wpipe;
 738         struct uio *uio;
 739 {
 740         int error;
 741
 742 retry:
 743         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 744                 if (wpipe->pipe_state & PIPE_WANTR) {
 745                         wpipe->pipe_state &= ~PIPE_WANTR;
 746                         wakeup(wpipe);
 747                 }
 748                 wpipe->pipe_state |= PIPE_WANTW;
 749                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 750                 if (error)
 751                         goto error2;
 752                 if (wpipe->pipe_state & PIPE_EOF) {
 753                         error = EPIPE;
 754                         goto error2;
 755                 }
 756         }
 757         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 758         if (wpipe->pipe_buffer.cnt > 0) {
 759                 if (wpipe->pipe_state & PIPE_WANTR) {
 760                         wpipe->pipe_state &= ~PIPE_WANTR;
 761                         wakeup(wpipe);
 762                 }
 763
 764                 wpipe->pipe_state |= PIPE_WANTW;
 765                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 766                 if (error)
 767                         goto error2;
 768                 if (wpipe->pipe_state & PIPE_EOF) {
 769                         error = EPIPE;
 770                         goto error2;
 771                 }
 772                 goto retry;
 773         }
 774
 775         /*
 776          * Build our direct-write buffer
 777          */
 778         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 779         error = pipe_build_write_buffer(wpipe, uio);
 780         if (error)
 781                 goto error1;
 782         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 783
 784         /*
 785          * Wait until the receiver has snarfed the data.  Since we are likely
 786          * going to sleep we optimize the case and yield synchronously,
 787          * possibly avoiding the tsleep().
 788          */
 789         error = 0;
 790         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 791                 if (wpipe->pipe_state & PIPE_EOF) {
 792                         pipelock(wpipe, 0);
 793                         xio_release(&wpipe->pipe_map);
 794                         if (wpipe->pipe_kva) {
 795                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 796                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 797                                 wpipe->pipe_kva = NULL;
 798                         }
 799                         pipeunlock(wpipe);
 800                         pipeselwakeup(wpipe);
 801                         error = EPIPE;
 802                         goto error1;
 803                 }
 804                 if (wpipe->pipe_state & PIPE_WANTR) {
 805                         wpipe->pipe_state &= ~PIPE_WANTR;
 806                         wakeup(wpipe);
 807                 }
 808                 pipeselwakeup(wpipe);
 809                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 810         }
 811         pipelock(wpipe,0);
 812         if (wpipe->pipe_state & PIPE_DIRECTW) {
 813                 /*
 814                  * this bit of trickery substitutes a kernel buffer for
 815                  * the process that might be going away.
 816                  */
 817                 pipe_clone_write_buffer(wpipe);
 818                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 819         } else {
 820                 /*
 821                  * note: The pipe_kva mapping is not qremove'd here.  For
 822                  * legacy PIPE_KMEM mode this constitutes an improvement
 823                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 824                  * mode the kva mapping must not be removed to get the
 825                  * caching benefit.
 826                  *
 827                  * For testing purposes we will give the original algorithm
 828                  * the benefit of the doubt 'what it could have been', and
 829                  * keep the optimization.
 830                  */
 831                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 832                 xio_release(&wpipe->pipe_map);
 833                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 834         }
 835         pipeunlock(wpipe);
 836         return (error);
 837
 838         /*
 839          * Direct-write error, clear the direct write flags.
 840          */
 841 error1:
 842         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 843         /* fallthrough */
 844
 845         /*
 846          * General error, wakeup the other side if it happens to be sleeping.
 847          */
 848 error2:
 849         wakeup(wpipe);
 850         return (error);
 851 }
 852 #endif
 853
 854 static int
 855 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
 856 {
 857         int error = 0;
 858         int orig_resid;
 859         struct pipe *wpipe, *rpipe;
 860
 861         rpipe = (struct pipe *) fp->f_data;
 862         wpipe = rpipe->pipe_peer;
 863
 864         /*
 865          * detect loss of pipe read side, issue SIGPIPE if lost.
 866          */
 867         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 868                 return (EPIPE);
 869         }
 870         ++wpipe->pipe_busy;
 871
 872         /*
 873          * If it is advantageous to resize the pipe buffer, do
 874          * so.
 875          */
 876         if ((uio->uio_resid > PIPE_SIZE) &&
 877                 (pipe_nbig < pipe_maxbig) &&
 878                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 879                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 880                 (wpipe->pipe_buffer.cnt == 0)) {
 881
 882                 if ((error = pipelock(wpipe,1)) == 0) {
 883                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 884                                 pipe_nbig++;
 885                         pipeunlock(wpipe);
 886                 }
 887         }
 888
 889         /*
 890          * If an early error occured unbusy and return, waking up any pending
 891          * readers.
 892          */
 893         if (error) {
 894                 --wpipe->pipe_busy;
 895                 if ((wpipe->pipe_busy == 0) &&
 896                     (wpipe->pipe_state & PIPE_WANT)) {
 897                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 898                         wakeup(wpipe);
 899                 }
 900                 return(error);
 901         }
 902
 903         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 904
 905         orig_resid = uio->uio_resid;
 906
 907         while (uio->uio_resid) {
 908                 int space;
 909
 910 #ifndef PIPE_NODIRECT
 911                 /*
 912                  * If the transfer is large, we can gain performance if
 913                  * we do process-to-process copies directly.
 914                  * If the write is non-blocking, we don't use the
 915                  * direct write mechanism.
 916                  *
 917                  * The direct write mechanism will detect the reader going
 918                  * away on us.
 919                  */
 920                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 921                     pipe_dwrite_enable > 1) &&
 922                     (fp->f_flag & FNONBLOCK) == 0 &&
 923                     pipe_dwrite_enable) {
 924                         error = pipe_direct_write( wpipe, uio);
 925                         if (error)
 926                                 break;
 927                         continue;
 928                 }
 929 #endif
 930
 931                 /*
 932                  * Pipe buffered writes cannot be coincidental with
 933                  * direct writes.  We wait until the currently executing
 934                  * direct write is completed before we start filling the
 935                  * pipe buffer.  We break out if a signal occurs or the
 936                  * reader goes away.
 937                  */
 938         retrywrite:
 939                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 940                         if (wpipe->pipe_state & PIPE_WANTR) {
 941                                 wpipe->pipe_state &= ~PIPE_WANTR;
 942                                 wakeup(wpipe);
 943                         }
 944                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 945                         if (wpipe->pipe_state & PIPE_EOF)
 946                                 break;
 947                         if (error)
 948                                 break;
 949                 }
 950                 if (wpipe->pipe_state & PIPE_EOF) {
 951                         error = EPIPE;
 952                         break;
 953                 }
 954
 955                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 956
 957                 /* Writes of size <= PIPE_BUF must be atomic. */
 958                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 959                         space = 0;
 960
 961                 /*
 962                  * Write to fill, read size handles write hysteresis.  Also
 963                  * additional restrictions can cause select-based non-blocking
 964                  * writes to spin.
 965                  */
 966                 if (space > 0) {
 967                         if ((error = pipelock(wpipe,1)) == 0) {
 968                                 int size;       /* Transfer size */
 969                                 int segsize;    /* first segment to transfer */
 970
 971                                 /*
 972                                  * It is possible for a direct write to
 973                                  * slip in on us... handle it here...
 974                                  */
 975                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 976                                         pipeunlock(wpipe);
 977                                         goto retrywrite;
 978                                 }
 979                                 /*
 980                                  * If a process blocked in uiomove, our
 981                                  * value for space might be bad.
 982                                  *
 983                                  * XXX will we be ok if the reader has gone
 984                                  * away here?
 985                                  */
 986                                 if (space > wpipe->pipe_buffer.size -
 987                                     wpipe->pipe_buffer.cnt) {
 988                                         pipeunlock(wpipe);
 989                                         goto retrywrite;
 990                                 }
 991
 992                                 /*
 993                                  * Transfer size is minimum of uio transfer
 994                                  * and free space in pipe buffer.
 995                                  */
 996                                 if (space > uio->uio_resid)
 997                                         size = uio->uio_resid;
 998                                 else
 999                                         size = space;
1000                                 /*
1001                                  * First segment to transfer is minimum of
1002                                  * transfer size and contiguous space in
1003                                  * pipe buffer.  If first segment to transfer
1004                                  * is less than the transfer size, we've got
1005                                  * a wraparound in the buffer.
1006                                  */
1007                                 segsize = wpipe->pipe_buffer.size -
1008                                         wpipe->pipe_buffer.in;
1009                                 if (segsize > size)
1010                                         segsize = size;
1011
1012                                 /* Transfer first segment */
1013
1014                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1015                                                 segsize, uio);
1016
1017                                 if (error == 0 && segsize < size) {
1018                                         /*
1019                                          * Transfer remaining part now, to
1020                                          * support atomic writes.  Wraparound
1021                                          * happened.
1022                                          */
1023                                         if (wpipe->pipe_buffer.in + segsize !=
1024                                             wpipe->pipe_buffer.size)
1025                                                 panic("Expected pipe buffer wraparound disappeared");
1026
1027                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1028                                                         size - segsize, uio);
1029                                 }
1030                                 if (error == 0) {
1031                                         wpipe->pipe_buffer.in += size;
1032                                         if (wpipe->pipe_buffer.in >=
1033                                             wpipe->pipe_buffer.size) {
1034                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1035                                                         panic("Expected wraparound bad");
1036                                                 wpipe->pipe_buffer.in = size - segsize;
1037                                         }
1038
1039                                         wpipe->pipe_buffer.cnt += size;
1040                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1041                                                 panic("Pipe buffer overflow");
1042
1043                                 }
1044                                 pipeunlock(wpipe);
1045                         }
1046                         if (error)
1047                                 break;
1048
1049                 } else {
1050                         /*
1051                          * If the "read-side" has been blocked, wake it up now
1052                          * and yield to let it drain synchronously rather
1053                          * then block.
1054                          */
1055                         if (wpipe->pipe_state & PIPE_WANTR) {
1056                                 wpipe->pipe_state &= ~PIPE_WANTR;
1057                                 wakeup(wpipe);
1058                         }
1059
1060                         /*
1061                          * don't block on non-blocking I/O
1062                          */
1063                         if (fp->f_flag & FNONBLOCK) {
1064                                 error = EAGAIN;
1065                                 break;
1066                         }
1067
1068                         /*
1069                          * We have no more space and have something to offer,
1070                          * wake up select/poll.
1071                          */
1072                         pipeselwakeup(wpipe);
1073
1074                         wpipe->pipe_state |= PIPE_WANTW;
1075                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1076                         if (error != 0)
1077                                 break;
1078                         /*
1079                          * If read side wants to go away, we just issue a signal
1080                          * to ourselves.
1081                          */
1082                         if (wpipe->pipe_state & PIPE_EOF) {
1083                                 error = EPIPE;
1084                                 break;
1085                         }
1086                 }
1087         }
1088
1089         --wpipe->pipe_busy;
1090
1091         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1092                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1093                 wakeup(wpipe);
1094         } else if (wpipe->pipe_buffer.cnt > 0) {
1095                 /*
1096                  * If we have put any characters in the buffer, we wake up
1097                  * the reader.
1098                  */
1099                 if (wpipe->pipe_state & PIPE_WANTR) {
1100                         wpipe->pipe_state &= ~PIPE_WANTR;
1101                         wakeup(wpipe);
1102                 }
1103         }
1104
1105         /*
1106          * Don't return EPIPE if I/O was successful
1107          */
1108         if ((wpipe->pipe_buffer.cnt == 0) &&
1109             (uio->uio_resid == 0) &&
1110             (error == EPIPE)) {
1111                 error = 0;
1112         }
1113
1114         if (error == 0)
1115                 vfs_timestamp(&wpipe->pipe_mtime);
1116
1117         /*
1118          * We have something to offer,
1119          * wake up select/poll.
1120          */
1121         if (wpipe->pipe_buffer.cnt)
1122                 pipeselwakeup(wpipe);
1123
1124         return (error);
1125 }
1126
1127 /*
1128  * we implement a very minimal set of ioctls for compatibility with sockets.
1129  */
1130 int
1131 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1132 {
1133         struct pipe *mpipe = (struct pipe *)fp->f_data;
1134
1135         switch (cmd) {
1136
1137         case FIONBIO:
1138                 return (0);
1139
1140         case FIOASYNC:
1141                 if (*(int *)data) {
1142                         mpipe->pipe_state |= PIPE_ASYNC;
1143                 } else {
1144                         mpipe->pipe_state &= ~PIPE_ASYNC;
1145                 }
1146                 return (0);
1147
1148         case FIONREAD:
1149                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1150                         *(int *)data = mpipe->pipe_map.xio_bytes -
1151                                         mpipe->pipe_buffer.out;
1152                 } else {
1153                         *(int *)data = mpipe->pipe_buffer.cnt;
1154                 }
1155                 return (0);
1156
1157         case FIOSETOWN:
1158                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1159
1160         case FIOGETOWN:
1161                 *(int *)data = fgetown(mpipe->pipe_sigio);
1162                 return (0);
1163
1164         /* This is deprecated, FIOSETOWN should be used instead. */
1165         case TIOCSPGRP:
1166                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1167
1168         /* This is deprecated, FIOGETOWN should be used instead. */
1169         case TIOCGPGRP:
1170                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1171                 return (0);
1172
1173         }
1174         return (ENOTTY);
1175 }
1176
1177 int
1178 pipe_poll(struct file *fp, int events, struct ucred *cred)
1179 {
1180         struct pipe *rpipe = (struct pipe *)fp->f_data;
1181         struct pipe *wpipe;
1182         int revents = 0;
1183
1184         wpipe = rpipe->pipe_peer;
1185         if (events & (POLLIN | POLLRDNORM))
1186                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1187                     (rpipe->pipe_buffer.cnt > 0) ||
1188                     (rpipe->pipe_state & PIPE_EOF))
1189                         revents |= events & (POLLIN | POLLRDNORM);
1190
1191         if (events & (POLLOUT | POLLWRNORM))
1192                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1193                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1194                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1195                         revents |= events & (POLLOUT | POLLWRNORM);
1196
1197         if ((rpipe->pipe_state & PIPE_EOF) ||
1198             (wpipe == NULL) ||
1199             (wpipe->pipe_state & PIPE_EOF))
1200                 revents |= POLLHUP;
1201
1202         if (revents == 0) {
1203                 if (events & (POLLIN | POLLRDNORM)) {
1204                         selrecord(curthread, &rpipe->pipe_sel);
1205                         rpipe->pipe_state |= PIPE_SEL;
1206                 }
1207
1208                 if (events & (POLLOUT | POLLWRNORM)) {
1209                         selrecord(curthread, &wpipe->pipe_sel);
1210                         wpipe->pipe_state |= PIPE_SEL;
1211                 }
1212         }
1213
1214         return (revents);
1215 }
1216
1217 static int
1218 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1219 {
1220         struct pipe *pipe = (struct pipe *)fp->f_data;
1221
1222         bzero((caddr_t)ub, sizeof(*ub));
1223         ub->st_mode = S_IFIFO;
1224         ub->st_blksize = pipe->pipe_buffer.size;
1225         ub->st_size = pipe->pipe_buffer.cnt;
1226         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1227                 ub->st_size = pipe->pipe_map.xio_bytes -
1228                                 pipe->pipe_buffer.out;
1229         }
1230         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1231         ub->st_atimespec = pipe->pipe_atime;
1232         ub->st_mtimespec = pipe->pipe_mtime;
1233         ub->st_ctimespec = pipe->pipe_ctime;
1234         /*
1235          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1236          * st_flags, st_gen.
1237          * XXX (st_dev, st_ino) should be unique.
1238          */
1239         return (0);
1240 }
1241
1242 /* ARGSUSED */
1243 static int
1244 pipe_close(struct file *fp)
1245 {
1246         struct pipe *cpipe = (struct pipe *)fp->f_data;
1247
1248         fp->f_ops = &badfileops;
1249         fp->f_data = NULL;
1250         funsetown(cpipe->pipe_sigio);
1251         pipeclose(cpipe);
1252         return (0);
1253 }
1254
1255 /*
1256  * Shutdown one or both directions of a full-duplex pipe.
1257  */
1258 /* ARGSUSED */
1259 static int
1260 pipe_shutdown(struct file *fp, int how)
1261 {
1262         struct pipe *rpipe = (struct pipe *)fp->f_data;
1263         struct pipe *wpipe;
1264         int error = EPIPE;
1265
1266         switch(how) {
1267         case SHUT_RDWR:
1268         case SHUT_RD:
1269                 if (rpipe) {
1270                         rpipe->pipe_state |= PIPE_EOF;
1271                         pipeselwakeup(rpipe);
1272                         if (rpipe->pipe_busy)
1273                                 wakeup(rpipe);
1274                         error = 0;
1275                 }
1276                 if (how == SHUT_RD)
1277                         break;
1278                 /* fall through */
1279         case SHUT_WR:
1280                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1281                         wpipe->pipe_state |= PIPE_EOF;
1282                         pipeselwakeup(wpipe);
1283                         if (wpipe->pipe_busy)
1284                                 wakeup(wpipe);
1285                         error = 0;
1286                 }
1287         }
1288         return (error);
1289 }
1290
1291 static void
1292 pipe_free_kmem(struct pipe *cpipe)
1293 {
1294         if (cpipe->pipe_buffer.buffer != NULL) {
1295                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1296                         --pipe_nbig;
1297                 kmem_free(kernel_map,
1298                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1299                         cpipe->pipe_buffer.size);
1300                 cpipe->pipe_buffer.buffer = NULL;
1301                 cpipe->pipe_buffer.object = NULL;
1302         }
1303 #ifndef PIPE_NODIRECT
1304         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1305                 cpipe->pipe_map.xio_offset == 0 &&
1306                 cpipe->pipe_map.xio_npages == 0);
1307 #endif
1308 }
1309
1310 /*
1311  * shutdown the pipe
1312  */
1313 static void
1314 pipeclose(struct pipe *cpipe)
1315 {
1316         globaldata_t gd;
1317         struct pipe *ppipe;
1318
1319         if (cpipe == NULL)
1320                 return;
1321
1322         pipeselwakeup(cpipe);
1323
1324         /*
1325          * If the other side is blocked, wake it up saying that
1326          * we want to close it down.
1327          */
1328         while (cpipe->pipe_busy) {
1329                 wakeup(cpipe);
1330                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1331                 tsleep(cpipe, 0, "pipecl", 0);
1332         }
1333
1334         /*
1335          * Disconnect from peer
1336          */
1337         if ((ppipe = cpipe->pipe_peer) != NULL) {
1338                 pipeselwakeup(ppipe);
1339
1340                 ppipe->pipe_state |= PIPE_EOF;
1341                 wakeup(ppipe);
1342                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1343                 ppipe->pipe_peer = NULL;
1344         }
1345
1346         if (cpipe->pipe_kva) {
1347                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1348                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1349                 cpipe->pipe_kva = NULL;
1350         }
1351
1352         /*
1353          * free or cache resources
1354          */
1355         gd = mycpu;
1356         if (gd->gd_pipeqcount >= pipe_maxcache ||
1357             cpipe->pipe_buffer.size != PIPE_SIZE
1358         ) {
1359                 pipe_free_kmem(cpipe);
1360                 free(cpipe, M_PIPE);
1361         } else {
1362                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1363                         cpipe->pipe_map.xio_bytes == 0 &&
1364                         cpipe->pipe_map.xio_offset == 0);
1365                 cpipe->pipe_state = 0;
1366                 cpipe->pipe_busy = 0;
1367                 cpipe->pipe_peer = gd->gd_pipeq;
1368                 gd->gd_pipeq = cpipe;
1369                 ++gd->gd_pipeqcount;
1370         }
1371 }
1372
1373 /*ARGSUSED*/
1374 static int
1375 pipe_kqfilter(struct file *fp, struct knote *kn)
1376 {
1377         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1378
1379         switch (kn->kn_filter) {
1380         case EVFILT_READ:
1381                 kn->kn_fop = &pipe_rfiltops;
1382                 break;
1383         case EVFILT_WRITE:
1384                 kn->kn_fop = &pipe_wfiltops;
1385                 cpipe = cpipe->pipe_peer;
1386                 if (cpipe == NULL)
1387                         /* other end of pipe has been closed */
1388                         return (EPIPE);
1389                 break;
1390         default:
1391                 return (1);
1392         }
1393         kn->kn_hook = (caddr_t)cpipe;
1394
1395         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1396         return (0);
1397 }
1398
1399 static void
1400 filt_pipedetach(struct knote *kn)
1401 {
1402         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1403
1404         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1405 }
1406
1407 /*ARGSUSED*/
1408 static int
1409 filt_piperead(struct knote *kn, long hint)
1410 {
1411         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1412         struct pipe *wpipe = rpipe->pipe_peer;
1413
1414         kn->kn_data = rpipe->pipe_buffer.cnt;
1415         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1416                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1417                                 rpipe->pipe_buffer.out;
1418         }
1419
1420         if ((rpipe->pipe_state & PIPE_EOF) ||
1421             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1422                 kn->kn_flags |= EV_EOF;
1423                 return (1);
1424         }
1425         return (kn->kn_data > 0);
1426 }
1427
1428 /*ARGSUSED*/
1429 static int
1430 filt_pipewrite(struct knote *kn, long hint)
1431 {
1432         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1433         struct pipe *wpipe = rpipe->pipe_peer;
1434
1435         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1436                 kn->kn_data = 0;
1437                 kn->kn_flags |= EV_EOF;
1438                 return (1);
1439         }
1440         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1441         if (wpipe->pipe_state & PIPE_DIRECTW)
1442                 kn->kn_data = 0;
1443
1444         return (kn->kn_data >= PIPE_BUF);
1445 }