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 1.39 2006/06/13 08:12:03 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 sys_pipe(struct pipe_args *uap)
 201 {
 202         struct thread *td = curthread;
 203         struct proc *p = td->td_proc;
 204         struct file *rf, *wf;
 205         struct pipe *rpipe, *wpipe;
 206         int fd1, fd2, error;
 207
 208         KKASSERT(p);
 209
 210         rpipe = wpipe = NULL;
 211         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 212                 pipeclose(rpipe);
 213                 pipeclose(wpipe);
 214                 return (ENFILE);
 215         }
 216
 217         rpipe->pipe_state |= PIPE_DIRECTOK;
 218         wpipe->pipe_state |= PIPE_DIRECTOK;
 219
 220         /*
 221          * Select the direct-map features to use for this pipe.  Since the
 222          * sysctl's can change on the fly we record the settings when the
 223          * pipe is created.
 224          *
 225          * Generally speaking the system will default to what we consider
 226          * to be the best-balanced and most stable option.  Right now this
 227          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 228          * moment.
 229          */
 230         wpipe->pipe_feature = PIPE_COPY;
 231         if (pipe_dwrite_enable) {
 232                 switch(pipe_dwrite_sfbuf) {
 233                 case 0:
 234                         wpipe->pipe_feature = PIPE_KMEM;
 235                         break;
 236                 case 1:
 237                         wpipe->pipe_feature = PIPE_SFBUF1;
 238                         break;
 239                 case 2:
 240                 case 3:
 241                         wpipe->pipe_feature = PIPE_SFBUF2;
 242                         break;
 243                 }
 244         }
 245         rpipe->pipe_feature = wpipe->pipe_feature;
 246
 247         error = falloc(p, &rf, &fd1);
 248         if (error) {
 249                 pipeclose(rpipe);
 250                 pipeclose(wpipe);
 251                 return (error);
 252         }
 253         uap->sysmsg_fds[0] = fd1;
 254
 255         /*
 256          * Warning: once we've gotten past allocation of the fd for the
 257          * read-side, we can only drop the read side via fdrop() in order
 258          * to avoid races against processes which manage to dup() the read
 259          * side while we are blocked trying to allocate the write side.
 260          */
 261         rf->f_type = DTYPE_PIPE;
 262         rf->f_flag = FREAD | FWRITE;
 263         rf->f_ops = &pipeops;
 264         rf->f_data = rpipe;
 265         error = falloc(p, &wf, &fd2);
 266         if (error) {
 267                 fsetfd(p, NULL, fd1);
 268                 fdrop(rf);
 269                 /* rpipe has been closed by fdrop(). */
 270                 pipeclose(wpipe);
 271                 return (error);
 272         }
 273         wf->f_type = DTYPE_PIPE;
 274         wf->f_flag = FREAD | FWRITE;
 275         wf->f_ops = &pipeops;
 276         wf->f_data = wpipe;
 277         uap->sysmsg_fds[1] = fd2;
 278
 279         rpipe->pipe_peer = wpipe;
 280         wpipe->pipe_peer = rpipe;
 281
 282         fsetfd(p, rf, fd1);
 283         fsetfd(p, wf, fd2);
 284         fdrop(rf);
 285         fdrop(wf);
 286
 287         return (0);
 288 }
 289
 290 /*
 291  * Allocate kva for pipe circular buffer, the space is pageable
 292  * This routine will 'realloc' the size of a pipe safely, if it fails
 293  * it will retain the old buffer.
 294  * If it fails it will return ENOMEM.
 295  */
 296 static int
 297 pipespace(struct pipe *cpipe, int size)
 298 {
 299         struct vm_object *object;
 300         caddr_t buffer;
 301         int npages, error;
 302
 303         npages = round_page(size) / PAGE_SIZE;
 304         object = cpipe->pipe_buffer.object;
 305
 306         /*
 307          * [re]create the object if necessary and reserve space for it
 308          * in the kernel_map.  The object and memory are pageable.  On
 309          * success, free the old resources before assigning the new
 310          * ones.
 311          */
 312         if (object == NULL || object->size != npages) {
 313                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 314                 buffer = (caddr_t) vm_map_min(kernel_map);
 315
 316                 error = vm_map_find(kernel_map, object, 0,
 317                         (vm_offset_t *) &buffer, size, 1,
 318                         VM_PROT_ALL, VM_PROT_ALL, 0);
 319
 320                 if (error != KERN_SUCCESS) {
 321                         vm_object_deallocate(object);
 322                         return (ENOMEM);
 323                 }
 324                 pipe_free_kmem(cpipe);
 325                 cpipe->pipe_buffer.object = object;
 326                 cpipe->pipe_buffer.buffer = buffer;
 327                 cpipe->pipe_buffer.size = size;
 328                 ++pipe_bkmem_alloc;
 329         } else {
 330                 ++pipe_bcache_alloc;
 331         }
 332         cpipe->pipe_buffer.in = 0;
 333         cpipe->pipe_buffer.out = 0;
 334         cpipe->pipe_buffer.cnt = 0;
 335         return (0);
 336 }
 337
 338 /*
 339  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 340  * our per-cpu cache if possible.  For now make sure it is sized for the
 341  * smaller PIPE_SIZE default.
 342  */
 343 static int
 344 pipe_create(cpipep)
 345         struct pipe **cpipep;
 346 {
 347         globaldata_t gd = mycpu;
 348         struct pipe *cpipe;
 349         int error;
 350
 351         if ((cpipe = gd->gd_pipeq) != NULL) {
 352                 gd->gd_pipeq = cpipe->pipe_peer;
 353                 --gd->gd_pipeqcount;
 354                 cpipe->pipe_peer = NULL;
 355         } else {
 356                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 357         }
 358         *cpipep = cpipe;
 359         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 360                 return (error);
 361         vfs_timestamp(&cpipe->pipe_ctime);
 362         cpipe->pipe_atime = cpipe->pipe_ctime;
 363         cpipe->pipe_mtime = cpipe->pipe_ctime;
 364         return (0);
 365 }
 366
 367
 368 /*
 369  * lock a pipe for I/O, blocking other access
 370  */
 371 static __inline int
 372 pipelock(cpipe, catch)
 373         struct pipe *cpipe;
 374         int catch;
 375 {
 376         int error;
 377
 378         while (cpipe->pipe_state & PIPE_LOCK) {
 379                 cpipe->pipe_state |= PIPE_LWANT;
 380                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 381                 if (error != 0)
 382                         return (error);
 383         }
 384         cpipe->pipe_state |= PIPE_LOCK;
 385         return (0);
 386 }
 387
 388 /*
 389  * unlock a pipe I/O lock
 390  */
 391 static __inline void
 392 pipeunlock(cpipe)
 393         struct pipe *cpipe;
 394 {
 395
 396         cpipe->pipe_state &= ~PIPE_LOCK;
 397         if (cpipe->pipe_state & PIPE_LWANT) {
 398                 cpipe->pipe_state &= ~PIPE_LWANT;
 399                 wakeup(cpipe);
 400         }
 401 }
 402
 403 static __inline void
 404 pipeselwakeup(cpipe)
 405         struct pipe *cpipe;
 406 {
 407
 408         if (cpipe->pipe_state & PIPE_SEL) {
 409                 cpipe->pipe_state &= ~PIPE_SEL;
 410                 selwakeup(&cpipe->pipe_sel);
 411         }
 412         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 413                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 414         KNOTE(&cpipe->pipe_sel.si_note, 0);
 415 }
 416
 417 /*
 418  * MPALMOSTSAFE (acquires mplock)
 419  */
 420 static int
 421 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 422 {
 423         struct pipe *rpipe;
 424         int error;
 425         int nread = 0;
 426         int nbio;
 427         u_int size;
 428
 429         get_mplock();
 430         rpipe = (struct pipe *) fp->f_data;
 431         ++rpipe->pipe_busy;
 432         error = pipelock(rpipe, 1);
 433         if (error)
 434                 goto unlocked_error;
 435
 436         if (fflags & O_FBLOCKING)
 437                 nbio = 0;
 438         else if (fflags & O_FNONBLOCKING)
 439                 nbio = 1;
 440         else if (fp->f_flag & O_NONBLOCK)
 441                 nbio = 1;
 442         else
 443                 nbio = 0;
 444
 445         while (uio->uio_resid) {
 446                 caddr_t va;
 447
 448                 if (rpipe->pipe_buffer.cnt > 0) {
 449                         /*
 450                          * normal pipe buffer receive
 451                          */
 452                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 453                         if (size > rpipe->pipe_buffer.cnt)
 454                                 size = rpipe->pipe_buffer.cnt;
 455                         if (size > (u_int) uio->uio_resid)
 456                                 size = (u_int) uio->uio_resid;
 457
 458                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 459                                         size, uio);
 460                         if (error)
 461                                 break;
 462
 463                         rpipe->pipe_buffer.out += size;
 464                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 465                                 rpipe->pipe_buffer.out = 0;
 466
 467                         rpipe->pipe_buffer.cnt -= size;
 468
 469                         /*
 470                          * If there is no more to read in the pipe, reset
 471                          * its pointers to the beginning.  This improves
 472                          * cache hit stats.
 473                          */
 474                         if (rpipe->pipe_buffer.cnt == 0) {
 475                                 rpipe->pipe_buffer.in = 0;
 476                                 rpipe->pipe_buffer.out = 0;
 477                         }
 478                         nread += size;
 479 #ifndef PIPE_NODIRECT
 480                 } else if (rpipe->pipe_kva &&
 481                            rpipe->pipe_feature == PIPE_KMEM &&
 482                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 483                                == PIPE_DIRECTW
 484                 ) {
 485                         /*
 486                          * Direct copy using source-side kva mapping
 487                          */
 488                         size = rpipe->pipe_map.xio_bytes -
 489                                 rpipe->pipe_buffer.out;
 490                         if (size > (u_int)uio->uio_resid)
 491                                 size = (u_int)uio->uio_resid;
 492                         va = (caddr_t)rpipe->pipe_kva +
 493                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 494                         error = uiomove(va, size, uio);
 495                         if (error)
 496                                 break;
 497                         nread += size;
 498                         rpipe->pipe_buffer.out += size;
 499                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 500                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 501                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 502                                 /* reset out index for copy mode */
 503                                 rpipe->pipe_buffer.out = 0;
 504                                 wakeup(rpipe);
 505                         }
 506                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 507                            rpipe->pipe_kva &&
 508                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 509                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 510                                == PIPE_DIRECTW
 511                 ) {
 512                         /*
 513                          * Direct copy, bypassing a kernel buffer.  We cannot
 514                          * mess with the direct-write buffer until
 515                          * PIPE_DIRECTIP is cleared.  In order to prevent
 516                          * the pipe_write code from racing itself in
 517                          * direct_write, we set DIRECTIP when we clear
 518                          * DIRECTW after we have exhausted the buffer.
 519                          */
 520                         if (pipe_dwrite_sfbuf == 3)
 521                                 rpipe->pipe_kvamask = 0;
 522                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 523                                     rpipe->pipe_map.xio_npages,
 524                                     &rpipe->pipe_kvamask);
 525                         size = rpipe->pipe_map.xio_bytes -
 526                                 rpipe->pipe_buffer.out;
 527                         if (size > (u_int)uio->uio_resid)
 528                                 size = (u_int)uio->uio_resid;
 529                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 530                         error = uiomove(va, size, uio);
 531                         if (error)
 532                                 break;
 533                         nread += size;
 534                         rpipe->pipe_buffer.out += size;
 535                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 536                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 537                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 538                                 /* reset out index for copy mode */
 539                                 rpipe->pipe_buffer.out = 0;
 540                                 wakeup(rpipe);
 541                         }
 542                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 543                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 544                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 545                                 == PIPE_DIRECTW
 546                 ) {
 547                         /*
 548                          * Direct copy, bypassing a kernel buffer.  We cannot
 549                          * mess with the direct-write buffer until
 550                          * PIPE_DIRECTIP is cleared.  In order to prevent
 551                          * the pipe_write code from racing itself in
 552                          * direct_write, we set DIRECTIP when we clear
 553                          * DIRECTW after we have exhausted the buffer.
 554                          */
 555                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 556                         if (error)
 557                                 break;
 558                         nread += size;
 559                         rpipe->pipe_buffer.out += size;
 560                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 561                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 562                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 563                                 /* reset out index for copy mode */
 564                                 rpipe->pipe_buffer.out = 0;
 565                                 wakeup(rpipe);
 566                         }
 567 #endif
 568                 } else {
 569                         /*
 570                          * detect EOF condition
 571                          * read returns 0 on EOF, no need to set error
 572                          */
 573                         if (rpipe->pipe_state & PIPE_EOF)
 574                                 break;
 575
 576                         /*
 577                          * If the "write-side" has been blocked, wake it up now.
 578                          */
 579                         if (rpipe->pipe_state & PIPE_WANTW) {
 580                                 rpipe->pipe_state &= ~PIPE_WANTW;
 581                                 wakeup(rpipe);
 582                         }
 583
 584                         /*
 585                          * Break if some data was read.
 586                          */
 587                         if (nread > 0)
 588                                 break;
 589
 590                         /*
 591                          * Unlock the pipe buffer for our remaining
 592                          * processing.  We will either break out with an
 593                          * error or we will sleep and relock to loop.
 594                          */
 595                         pipeunlock(rpipe);
 596
 597                         /*
 598                          * Handle non-blocking mode operation or
 599                          * wait for more data.
 600                          */
 601                         if (nbio) {
 602                                 error = EAGAIN;
 603                         } else {
 604                                 rpipe->pipe_state |= PIPE_WANTR;
 605                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 606                                     "piperd", 0)) == 0) {
 607                                         error = pipelock(rpipe, 1);
 608                                 }
 609                         }
 610                         if (error)
 611                                 goto unlocked_error;
 612                 }
 613         }
 614         pipeunlock(rpipe);
 615
 616         if (error == 0)
 617                 vfs_timestamp(&rpipe->pipe_atime);
 618 unlocked_error:
 619         --rpipe->pipe_busy;
 620
 621         /*
 622          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 623          */
 624         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 625                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 626                 wakeup(rpipe);
 627         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 628                 /*
 629                  * Handle write blocking hysteresis.
 630                  */
 631                 if (rpipe->pipe_state & PIPE_WANTW) {
 632                         rpipe->pipe_state &= ~PIPE_WANTW;
 633                         wakeup(rpipe);
 634                 }
 635         }
 636
 637         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 638                 pipeselwakeup(rpipe);
 639         rel_mplock();
 640         return (error);
 641 }
 642
 643 #ifndef PIPE_NODIRECT
 644 /*
 645  * Map the sending processes' buffer into kernel space and wire it.
 646  * This is similar to a physical write operation.
 647  */
 648 static int
 649 pipe_build_write_buffer(wpipe, uio)
 650         struct pipe *wpipe;
 651         struct uio *uio;
 652 {
 653         int error;
 654         u_int size;
 655
 656         size = (u_int) uio->uio_iov->iov_len;
 657         if (size > wpipe->pipe_buffer.size)
 658                 size = wpipe->pipe_buffer.size;
 659
 660         if (uio->uio_segflg == UIO_SYSSPACE) {
 661                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 662                                         size);
 663         } else {
 664                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 665                                         size, XIOF_READ);
 666         }
 667         wpipe->pipe_buffer.out = 0;
 668         if (error)
 669                 return(error);
 670
 671         /*
 672          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 673          * map the pages on the target while KMEM maps the pages now.
 674          */
 675         switch(wpipe->pipe_feature) {
 676         case PIPE_KMEM:
 677         case PIPE_SFBUF2:
 678                 if (wpipe->pipe_kva == NULL) {
 679                         wpipe->pipe_kva =
 680                             kmem_alloc_nofault(kernel_map, XIO_INTERNAL_SIZE);
 681                         wpipe->pipe_kvamask = 0;
 682                 }
 683                 if (wpipe->pipe_feature == PIPE_KMEM) {
 684                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 685                                     wpipe->pipe_map.xio_npages);
 686                 }
 687                 break;
 688         default:
 689                 break;
 690         }
 691
 692         /*
 693          * And update the uio data.  The XIO might have loaded fewer bytes
 694          * then requested so reload 'size'.
 695          */
 696         size = wpipe->pipe_map.xio_bytes;
 697         uio->uio_iov->iov_len -= size;
 698         uio->uio_iov->iov_base += size;
 699         if (uio->uio_iov->iov_len == 0)
 700                 uio->uio_iov++;
 701         uio->uio_resid -= size;
 702         uio->uio_offset += size;
 703         return (0);
 704 }
 705
 706 /*
 707  * In the case of a signal, the writing process might go away.  This
 708  * code copies the data into the circular buffer so that the source
 709  * pages can be freed without loss of data.
 710  *
 711  * Note that in direct mode pipe_buffer.out is used to track the
 712  * XIO offset.  We are converting the direct mode into buffered mode
 713  * which changes the meaning of pipe_buffer.out.
 714  */
 715 static void
 716 pipe_clone_write_buffer(wpipe)
 717         struct pipe *wpipe;
 718 {
 719         int size;
 720         int offset;
 721
 722         offset = wpipe->pipe_buffer.out;
 723         size = wpipe->pipe_map.xio_bytes - offset;
 724
 725         KKASSERT(size <= wpipe->pipe_buffer.size);
 726
 727         wpipe->pipe_buffer.in = size;
 728         wpipe->pipe_buffer.out = 0;
 729         wpipe->pipe_buffer.cnt = size;
 730         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 731
 732         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 733         xio_release(&wpipe->pipe_map);
 734         if (wpipe->pipe_kva) {
 735                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 736                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 737                 wpipe->pipe_kva = NULL;
 738         }
 739 }
 740
 741 /*
 742  * This implements the pipe buffer write mechanism.  Note that only
 743  * a direct write OR a normal pipe write can be pending at any given time.
 744  * If there are any characters in the pipe buffer, the direct write will
 745  * be deferred until the receiving process grabs all of the bytes from
 746  * the pipe buffer.  Then the direct mapping write is set-up.
 747  */
 748 static int
 749 pipe_direct_write(wpipe, uio)
 750         struct pipe *wpipe;
 751         struct uio *uio;
 752 {
 753         int error;
 754
 755 retry:
 756         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 757                 if (wpipe->pipe_state & PIPE_WANTR) {
 758                         wpipe->pipe_state &= ~PIPE_WANTR;
 759                         wakeup(wpipe);
 760                 }
 761                 wpipe->pipe_state |= PIPE_WANTW;
 762                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 763                 if (error)
 764                         goto error2;
 765                 if (wpipe->pipe_state & PIPE_EOF) {
 766                         error = EPIPE;
 767                         goto error2;
 768                 }
 769         }
 770         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 771         if (wpipe->pipe_buffer.cnt > 0) {
 772                 if (wpipe->pipe_state & PIPE_WANTR) {
 773                         wpipe->pipe_state &= ~PIPE_WANTR;
 774                         wakeup(wpipe);
 775                 }
 776
 777                 wpipe->pipe_state |= PIPE_WANTW;
 778                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 779                 if (error)
 780                         goto error2;
 781                 if (wpipe->pipe_state & PIPE_EOF) {
 782                         error = EPIPE;
 783                         goto error2;
 784                 }
 785                 goto retry;
 786         }
 787
 788         /*
 789          * Build our direct-write buffer
 790          */
 791         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 792         error = pipe_build_write_buffer(wpipe, uio);
 793         if (error)
 794                 goto error1;
 795         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 796
 797         /*
 798          * Wait until the receiver has snarfed the data.  Since we are likely
 799          * going to sleep we optimize the case and yield synchronously,
 800          * possibly avoiding the tsleep().
 801          */
 802         error = 0;
 803         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 804                 if (wpipe->pipe_state & PIPE_EOF) {
 805                         pipelock(wpipe, 0);
 806                         xio_release(&wpipe->pipe_map);
 807                         if (wpipe->pipe_kva) {
 808                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 809                                 kmem_free(kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 810                                 wpipe->pipe_kva = NULL;
 811                         }
 812                         pipeunlock(wpipe);
 813                         pipeselwakeup(wpipe);
 814                         error = EPIPE;
 815                         goto error1;
 816                 }
 817                 if (wpipe->pipe_state & PIPE_WANTR) {
 818                         wpipe->pipe_state &= ~PIPE_WANTR;
 819                         wakeup(wpipe);
 820                 }
 821                 pipeselwakeup(wpipe);
 822                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 823         }
 824         pipelock(wpipe,0);
 825         if (wpipe->pipe_state & PIPE_DIRECTW) {
 826                 /*
 827                  * this bit of trickery substitutes a kernel buffer for
 828                  * the process that might be going away.
 829                  */
 830                 pipe_clone_write_buffer(wpipe);
 831                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 832         } else {
 833                 /*
 834                  * note: The pipe_kva mapping is not qremove'd here.  For
 835                  * legacy PIPE_KMEM mode this constitutes an improvement
 836                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 837                  * mode the kva mapping must not be removed to get the
 838                  * caching benefit.
 839                  *
 840                  * For testing purposes we will give the original algorithm
 841                  * the benefit of the doubt 'what it could have been', and
 842                  * keep the optimization.
 843                  */
 844                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 845                 xio_release(&wpipe->pipe_map);
 846                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 847         }
 848         pipeunlock(wpipe);
 849         return (error);
 850
 851         /*
 852          * Direct-write error, clear the direct write flags.
 853          */
 854 error1:
 855         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 856         /* fallthrough */
 857
 858         /*
 859          * General error, wakeup the other side if it happens to be sleeping.
 860          */
 861 error2:
 862         wakeup(wpipe);
 863         return (error);
 864 }
 865 #endif
 866
 867 /*
 868  * MPALMOSTSAFE - acquires mplock
 869  */
 870 static int
 871 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 872 {
 873         int error = 0;
 874         int orig_resid;
 875         int nbio;
 876         struct pipe *wpipe, *rpipe;
 877
 878         get_mplock();
 879         rpipe = (struct pipe *) fp->f_data;
 880         wpipe = rpipe->pipe_peer;
 881
 882         /*
 883          * detect loss of pipe read side, issue SIGPIPE if lost.
 884          */
 885         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 886                 rel_mplock();
 887                 return (EPIPE);
 888         }
 889         ++wpipe->pipe_busy;
 890
 891         if (fflags & O_FBLOCKING)
 892                 nbio = 0;
 893         else if (fflags & O_FNONBLOCKING)
 894                 nbio = 1;
 895         else if (fp->f_flag & O_NONBLOCK)
 896                 nbio = 1;
 897         else
 898                 nbio = 0;
 899
 900         /*
 901          * If it is advantageous to resize the pipe buffer, do
 902          * so.
 903          */
 904         if ((uio->uio_resid > PIPE_SIZE) &&
 905                 (pipe_nbig < pipe_maxbig) &&
 906                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 907                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 908                 (wpipe->pipe_buffer.cnt == 0)) {
 909
 910                 if ((error = pipelock(wpipe,1)) == 0) {
 911                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 912                                 pipe_nbig++;
 913                         pipeunlock(wpipe);
 914                 }
 915         }
 916
 917         /*
 918          * If an early error occured unbusy and return, waking up any pending
 919          * readers.
 920          */
 921         if (error) {
 922                 --wpipe->pipe_busy;
 923                 if ((wpipe->pipe_busy == 0) &&
 924                     (wpipe->pipe_state & PIPE_WANT)) {
 925                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 926                         wakeup(wpipe);
 927                 }
 928                 rel_mplock();
 929                 return(error);
 930         }
 931
 932         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 933
 934         orig_resid = uio->uio_resid;
 935
 936         while (uio->uio_resid) {
 937                 int space;
 938
 939 #ifndef PIPE_NODIRECT
 940                 /*
 941                  * If the transfer is large, we can gain performance if
 942                  * we do process-to-process copies directly.
 943                  * If the write is non-blocking, we don't use the
 944                  * direct write mechanism.
 945                  *
 946                  * The direct write mechanism will detect the reader going
 947                  * away on us.
 948                  */
 949                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 950                     pipe_dwrite_enable > 1) &&
 951                     nbio == 0 &&
 952                     pipe_dwrite_enable) {
 953                         error = pipe_direct_write( wpipe, uio);
 954                         if (error)
 955                                 break;
 956                         continue;
 957                 }
 958 #endif
 959
 960                 /*
 961                  * Pipe buffered writes cannot be coincidental with
 962                  * direct writes.  We wait until the currently executing
 963                  * direct write is completed before we start filling the
 964                  * pipe buffer.  We break out if a signal occurs or the
 965                  * reader goes away.
 966                  */
 967         retrywrite:
 968                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 969                         if (wpipe->pipe_state & PIPE_WANTR) {
 970                                 wpipe->pipe_state &= ~PIPE_WANTR;
 971                                 wakeup(wpipe);
 972                         }
 973                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 974                         if (wpipe->pipe_state & PIPE_EOF)
 975                                 break;
 976                         if (error)
 977                                 break;
 978                 }
 979                 if (wpipe->pipe_state & PIPE_EOF) {
 980                         error = EPIPE;
 981                         break;
 982                 }
 983
 984                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 985
 986                 /* Writes of size <= PIPE_BUF must be atomic. */
 987                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 988                         space = 0;
 989
 990                 /*
 991                  * Write to fill, read size handles write hysteresis.  Also
 992                  * additional restrictions can cause select-based non-blocking
 993                  * writes to spin.
 994                  */
 995                 if (space > 0) {
 996                         if ((error = pipelock(wpipe,1)) == 0) {
 997                                 int size;       /* Transfer size */
 998                                 int segsize;    /* first segment to transfer */
 999
1000                                 /*
1001                                  * It is possible for a direct write to
1002                                  * slip in on us... handle it here...
1003                                  */
1004                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
1005                                         pipeunlock(wpipe);
1006                                         goto retrywrite;
1007                                 }
1008                                 /*
1009                                  * If a process blocked in uiomove, our
1010                                  * value for space might be bad.
1011                                  *
1012                                  * XXX will we be ok if the reader has gone
1013                                  * away here?
1014                                  */
1015                                 if (space > wpipe->pipe_buffer.size -
1016                                     wpipe->pipe_buffer.cnt) {
1017                                         pipeunlock(wpipe);
1018                                         goto retrywrite;
1019                                 }
1020
1021                                 /*
1022                                  * Transfer size is minimum of uio transfer
1023                                  * and free space in pipe buffer.
1024                                  */
1025                                 if (space > uio->uio_resid)
1026                                         size = uio->uio_resid;
1027                                 else
1028                                         size = space;
1029                                 /*
1030                                  * First segment to transfer is minimum of
1031                                  * transfer size and contiguous space in
1032                                  * pipe buffer.  If first segment to transfer
1033                                  * is less than the transfer size, we've got
1034                                  * a wraparound in the buffer.
1035                                  */
1036                                 segsize = wpipe->pipe_buffer.size -
1037                                         wpipe->pipe_buffer.in;
1038                                 if (segsize > size)
1039                                         segsize = size;
1040
1041                                 /* Transfer first segment */
1042
1043                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1044                                                 segsize, uio);
1045
1046                                 if (error == 0 && segsize < size) {
1047                                         /*
1048                                          * Transfer remaining part now, to
1049                                          * support atomic writes.  Wraparound
1050                                          * happened.
1051                                          */
1052                                         if (wpipe->pipe_buffer.in + segsize !=
1053                                             wpipe->pipe_buffer.size)
1054                                                 panic("Expected pipe buffer wraparound disappeared");
1055
1056                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1057                                                         size - segsize, uio);
1058                                 }
1059                                 if (error == 0) {
1060                                         wpipe->pipe_buffer.in += size;
1061                                         if (wpipe->pipe_buffer.in >=
1062                                             wpipe->pipe_buffer.size) {
1063                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1064                                                         panic("Expected wraparound bad");
1065                                                 wpipe->pipe_buffer.in = size - segsize;
1066                                         }
1067
1068                                         wpipe->pipe_buffer.cnt += size;
1069                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1070                                                 panic("Pipe buffer overflow");
1071
1072                                 }
1073                                 pipeunlock(wpipe);
1074                         }
1075                         if (error)
1076                                 break;
1077
1078                 } else {
1079                         /*
1080                          * If the "read-side" has been blocked, wake it up now
1081                          * and yield to let it drain synchronously rather
1082                          * then block.
1083                          */
1084                         if (wpipe->pipe_state & PIPE_WANTR) {
1085                                 wpipe->pipe_state &= ~PIPE_WANTR;
1086                                 wakeup(wpipe);
1087                         }
1088
1089                         /*
1090                          * don't block on non-blocking I/O
1091                          */
1092                         if (nbio) {
1093                                 error = EAGAIN;
1094                                 break;
1095                         }
1096
1097                         /*
1098                          * We have no more space and have something to offer,
1099                          * wake up select/poll.
1100                          */
1101                         pipeselwakeup(wpipe);
1102
1103                         wpipe->pipe_state |= PIPE_WANTW;
1104                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1105                         if (error != 0)
1106                                 break;
1107                         /*
1108                          * If read side wants to go away, we just issue a signal
1109                          * to ourselves.
1110                          */
1111                         if (wpipe->pipe_state & PIPE_EOF) {
1112                                 error = EPIPE;
1113                                 break;
1114                         }
1115                 }
1116         }
1117
1118         --wpipe->pipe_busy;
1119
1120         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1121                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1122                 wakeup(wpipe);
1123         } else if (wpipe->pipe_buffer.cnt > 0) {
1124                 /*
1125                  * If we have put any characters in the buffer, we wake up
1126                  * the reader.
1127                  */
1128                 if (wpipe->pipe_state & PIPE_WANTR) {
1129                         wpipe->pipe_state &= ~PIPE_WANTR;
1130                         wakeup(wpipe);
1131                 }
1132         }
1133
1134         /*
1135          * Don't return EPIPE if I/O was successful
1136          */
1137         if ((wpipe->pipe_buffer.cnt == 0) &&
1138             (uio->uio_resid == 0) &&
1139             (error == EPIPE)) {
1140                 error = 0;
1141         }
1142
1143         if (error == 0)
1144                 vfs_timestamp(&wpipe->pipe_mtime);
1145
1146         /*
1147          * We have something to offer,
1148          * wake up select/poll.
1149          */
1150         if (wpipe->pipe_buffer.cnt)
1151                 pipeselwakeup(wpipe);
1152         rel_mplock();
1153         return (error);
1154 }
1155
1156 /*
1157  * MPALMOSTSAFE - acquires mplock
1158  *
1159  * we implement a very minimal set of ioctls for compatibility with sockets.
1160  */
1161 int
1162 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1163 {
1164         struct pipe *mpipe;
1165         int error;
1166
1167         get_mplock();
1168         mpipe = (struct pipe *)fp->f_data;
1169
1170         switch (cmd) {
1171         case FIOASYNC:
1172                 if (*(int *)data) {
1173                         mpipe->pipe_state |= PIPE_ASYNC;
1174                 } else {
1175                         mpipe->pipe_state &= ~PIPE_ASYNC;
1176                 }
1177                 error = 0;
1178                 break;
1179         case FIONREAD:
1180                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1181                         *(int *)data = mpipe->pipe_map.xio_bytes -
1182                                         mpipe->pipe_buffer.out;
1183                 } else {
1184                         *(int *)data = mpipe->pipe_buffer.cnt;
1185                 }
1186                 error = 0;
1187                 break;
1188         case FIOSETOWN:
1189                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1190                 break;
1191         case FIOGETOWN:
1192                 *(int *)data = fgetown(mpipe->pipe_sigio);
1193                 error = 0;
1194                 break;
1195         case TIOCSPGRP:
1196                 /* This is deprecated, FIOSETOWN should be used instead. */
1197                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1198                 break;
1199
1200         case TIOCGPGRP:
1201                 /* This is deprecated, FIOGETOWN should be used instead. */
1202                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1203                 error = 0;
1204                 break;
1205         default:
1206                 error = ENOTTY;
1207                 break;
1208         }
1209         rel_mplock();
1210         return (error);
1211 }
1212
1213 /*
1214  * MPALMOSTSAFE - acquires mplock
1215  */
1216 int
1217 pipe_poll(struct file *fp, int events, struct ucred *cred)
1218 {
1219         struct pipe *rpipe;
1220         struct pipe *wpipe;
1221         int revents = 0;
1222
1223         get_mplock();
1224         rpipe = (struct pipe *)fp->f_data;
1225         wpipe = rpipe->pipe_peer;
1226         if (events & (POLLIN | POLLRDNORM))
1227                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1228                     (rpipe->pipe_buffer.cnt > 0) ||
1229                     (rpipe->pipe_state & PIPE_EOF))
1230                         revents |= events & (POLLIN | POLLRDNORM);
1231
1232         if (events & (POLLOUT | POLLWRNORM))
1233                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1234                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1235                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1236                         revents |= events & (POLLOUT | POLLWRNORM);
1237
1238         if ((rpipe->pipe_state & PIPE_EOF) ||
1239             (wpipe == NULL) ||
1240             (wpipe->pipe_state & PIPE_EOF))
1241                 revents |= POLLHUP;
1242
1243         if (revents == 0) {
1244                 if (events & (POLLIN | POLLRDNORM)) {
1245                         selrecord(curthread, &rpipe->pipe_sel);
1246                         rpipe->pipe_state |= PIPE_SEL;
1247                 }
1248
1249                 if (events & (POLLOUT | POLLWRNORM)) {
1250                         selrecord(curthread, &wpipe->pipe_sel);
1251                         wpipe->pipe_state |= PIPE_SEL;
1252                 }
1253         }
1254         rel_mplock();
1255         return (revents);
1256 }
1257
1258 /*
1259  * MPALMOSTSAFE - acquires mplock
1260  */
1261 static int
1262 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1263 {
1264         struct pipe *pipe;
1265
1266         get_mplock();
1267         pipe = (struct pipe *)fp->f_data;
1268
1269         bzero((caddr_t)ub, sizeof(*ub));
1270         ub->st_mode = S_IFIFO;
1271         ub->st_blksize = pipe->pipe_buffer.size;
1272         ub->st_size = pipe->pipe_buffer.cnt;
1273         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1274                 ub->st_size = pipe->pipe_map.xio_bytes -
1275                                 pipe->pipe_buffer.out;
1276         }
1277         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1278         ub->st_atimespec = pipe->pipe_atime;
1279         ub->st_mtimespec = pipe->pipe_mtime;
1280         ub->st_ctimespec = pipe->pipe_ctime;
1281         /*
1282          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1283          * st_flags, st_gen.
1284          * XXX (st_dev, st_ino) should be unique.
1285          */
1286         rel_mplock();
1287         return (0);
1288 }
1289
1290 /*
1291  * MPALMOSTSAFE - acquires mplock
1292  */
1293 static int
1294 pipe_close(struct file *fp)
1295 {
1296         struct pipe *cpipe = (struct pipe *)fp->f_data;
1297
1298         get_mplock();
1299         fp->f_ops = &badfileops;
1300         fp->f_data = NULL;
1301         funsetown(cpipe->pipe_sigio);
1302         pipeclose(cpipe);
1303         rel_mplock();
1304         return (0);
1305 }
1306
1307 /*
1308  * Shutdown one or both directions of a full-duplex pipe.
1309  *
1310  * MPALMOSTSAFE - acquires mplock
1311  */
1312 static int
1313 pipe_shutdown(struct file *fp, int how)
1314 {
1315         struct pipe *rpipe;
1316         struct pipe *wpipe;
1317         int error = EPIPE;
1318
1319         get_mplock();
1320         rpipe = (struct pipe *)fp->f_data;
1321
1322         switch(how) {
1323         case SHUT_RDWR:
1324         case SHUT_RD:
1325                 if (rpipe) {
1326                         rpipe->pipe_state |= PIPE_EOF;
1327                         pipeselwakeup(rpipe);
1328                         if (rpipe->pipe_busy)
1329                                 wakeup(rpipe);
1330                         error = 0;
1331                 }
1332                 if (how == SHUT_RD)
1333                         break;
1334                 /* fall through */
1335         case SHUT_WR:
1336                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1337                         wpipe->pipe_state |= PIPE_EOF;
1338                         pipeselwakeup(wpipe);
1339                         if (wpipe->pipe_busy)
1340                                 wakeup(wpipe);
1341                         error = 0;
1342                 }
1343         }
1344         rel_mplock();
1345         return (error);
1346 }
1347
1348 static void
1349 pipe_free_kmem(struct pipe *cpipe)
1350 {
1351         if (cpipe->pipe_buffer.buffer != NULL) {
1352                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1353                         --pipe_nbig;
1354                 kmem_free(kernel_map,
1355                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1356                         cpipe->pipe_buffer.size);
1357                 cpipe->pipe_buffer.buffer = NULL;
1358                 cpipe->pipe_buffer.object = NULL;
1359         }
1360 #ifndef PIPE_NODIRECT
1361         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1362                 cpipe->pipe_map.xio_offset == 0 &&
1363                 cpipe->pipe_map.xio_npages == 0);
1364 #endif
1365 }
1366
1367 /*
1368  * shutdown the pipe
1369  */
1370 static void
1371 pipeclose(struct pipe *cpipe)
1372 {
1373         globaldata_t gd;
1374         struct pipe *ppipe;
1375
1376         if (cpipe == NULL)
1377                 return;
1378
1379         pipeselwakeup(cpipe);
1380
1381         /*
1382          * If the other side is blocked, wake it up saying that
1383          * we want to close it down.
1384          */
1385         while (cpipe->pipe_busy) {
1386                 wakeup(cpipe);
1387                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1388                 tsleep(cpipe, 0, "pipecl", 0);
1389         }
1390
1391         /*
1392          * Disconnect from peer
1393          */
1394         if ((ppipe = cpipe->pipe_peer) != NULL) {
1395                 pipeselwakeup(ppipe);
1396
1397                 ppipe->pipe_state |= PIPE_EOF;
1398                 wakeup(ppipe);
1399                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1400                 ppipe->pipe_peer = NULL;
1401         }
1402
1403         if (cpipe->pipe_kva) {
1404                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1405                 kmem_free(kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1406                 cpipe->pipe_kva = NULL;
1407         }
1408
1409         /*
1410          * free or cache resources
1411          */
1412         gd = mycpu;
1413         if (gd->gd_pipeqcount >= pipe_maxcache ||
1414             cpipe->pipe_buffer.size != PIPE_SIZE
1415         ) {
1416                 pipe_free_kmem(cpipe);
1417                 free(cpipe, M_PIPE);
1418         } else {
1419                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1420                         cpipe->pipe_map.xio_bytes == 0 &&
1421                         cpipe->pipe_map.xio_offset == 0);
1422                 cpipe->pipe_state = 0;
1423                 cpipe->pipe_busy = 0;
1424                 cpipe->pipe_peer = gd->gd_pipeq;
1425                 gd->gd_pipeq = cpipe;
1426                 ++gd->gd_pipeqcount;
1427         }
1428 }
1429
1430 /*
1431  * MPALMOSTSAFE - acquires mplock
1432  */
1433 static int
1434 pipe_kqfilter(struct file *fp, struct knote *kn)
1435 {
1436         struct pipe *cpipe;
1437
1438         get_mplock();
1439         cpipe = (struct pipe *)kn->kn_fp->f_data;
1440
1441         switch (kn->kn_filter) {
1442         case EVFILT_READ:
1443                 kn->kn_fop = &pipe_rfiltops;
1444                 break;
1445         case EVFILT_WRITE:
1446                 kn->kn_fop = &pipe_wfiltops;
1447                 cpipe = cpipe->pipe_peer;
1448                 if (cpipe == NULL) {
1449                         /* other end of pipe has been closed */
1450                         rel_mplock();
1451                         return (EPIPE);
1452                 }
1453                 break;
1454         default:
1455                 return (1);
1456         }
1457         kn->kn_hook = (caddr_t)cpipe;
1458
1459         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1460         rel_mplock();
1461         return (0);
1462 }
1463
1464 static void
1465 filt_pipedetach(struct knote *kn)
1466 {
1467         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1468
1469         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1470 }
1471
1472 /*ARGSUSED*/
1473 static int
1474 filt_piperead(struct knote *kn, long hint)
1475 {
1476         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1477         struct pipe *wpipe = rpipe->pipe_peer;
1478
1479         kn->kn_data = rpipe->pipe_buffer.cnt;
1480         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1481                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1482                                 rpipe->pipe_buffer.out;
1483         }
1484
1485         if ((rpipe->pipe_state & PIPE_EOF) ||
1486             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1487                 kn->kn_flags |= EV_EOF;
1488                 return (1);
1489         }
1490         return (kn->kn_data > 0);
1491 }
1492
1493 /*ARGSUSED*/
1494 static int
1495 filt_pipewrite(struct knote *kn, long hint)
1496 {
1497         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1498         struct pipe *wpipe = rpipe->pipe_peer;
1499
1500         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1501                 kn->kn_data = 0;
1502                 kn->kn_flags |= EV_EOF;
1503                 return (1);
1504         }
1505         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1506         if (wpipe->pipe_state & PIPE_DIRECTW)
1507                 kn->kn_data = 0;
1508
1509         return (kn->kn_data >= PIPE_BUF);
1510 }