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 1.44 2006/12/28 21:24:01 dillon Exp $
  21  */
  22
  23 /*
  24  * This file contains a high-performance replacement for the socket-based
  25  * pipes scheme originally used in FreeBSD/4.4Lite.  It does not support
  26  * all features of sockets, but does do everything that pipes normally
  27  * do.
  28  */
  29
  30 /*
  31  * This code has two modes of operation, a small write mode and a large
  32  * write mode.  The small write mode acts like conventional pipes with
  33  * a kernel buffer.  If the buffer is less than PIPE_MINDIRECT, then the
  34  * "normal" pipe buffering is done.  If the buffer is between PIPE_MINDIRECT
  35  * and PIPE_SIZE in size, it is fully mapped and wired into the kernel, and
  36  * the receiving process can copy it directly from the pages in the sending
  37  * process.
  38  *
  39  * If the sending process receives a signal, it is possible that it will
  40  * go away, and certainly its address space can change, because control
  41  * is returned back to the user-mode side.  In that case, the pipe code
  42  * arranges to copy the buffer supplied by the user process, to a pageable
  43  * kernel buffer, and the receiving process will grab the data from the
  44  * pageable kernel buffer.  Since signals don't happen all that often,
  45  * the copy operation is normally eliminated.
  46  *
  47  * The constant PIPE_MINDIRECT is chosen to make sure that buffering will
  48  * happen for small transfers so that the system will not spend all of
  49  * its time context switching.  PIPE_SIZE is constrained by the
  50  * amount of kernel virtual memory.
  51  */
  52
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/file.h>
  59 #include <sys/filedesc.h>
  60 #include <sys/filio.h>
  61 #include <sys/ttycom.h>
  62 #include <sys/stat.h>
  63 #include <sys/poll.h>
  64 #include <sys/select.h>
  65 #include <sys/signalvar.h>
  66 #include <sys/sysproto.h>
  67 #include <sys/pipe.h>
  68 #include <sys/vnode.h>
  69 #include <sys/uio.h>
  70 #include <sys/event.h>
  71 #include <sys/globaldata.h>
  72 #include <sys/module.h>
  73 #include <sys/malloc.h>
  74 #include <sys/sysctl.h>
  75 #include <sys/socket.h>
  76
  77 #include <vm/vm.h>
  78 #include <vm/vm_param.h>
  79 #include <sys/lock.h>
  80 #include <vm/vm_object.h>
  81 #include <vm/vm_kern.h>
  82 #include <vm/vm_extern.h>
  83 #include <vm/pmap.h>
  84 #include <vm/vm_map.h>
  85 #include <vm/vm_page.h>
  86 #include <vm/vm_zone.h>
  87
  88 #include <sys/file2.h>
  89
  90 #include <machine/cpufunc.h>
  91
  92 /*
  93  * interfaces to the outside world
  94  */
  95 static int pipe_read (struct file *fp, struct uio *uio,
  96                 struct ucred *cred, int flags);
  97 static int pipe_write (struct file *fp, struct uio *uio,
  98                 struct ucred *cred, int flags);
  99 static int pipe_close (struct file *fp);
 100 static int pipe_shutdown (struct file *fp, int how);
 101 static int pipe_poll (struct file *fp, int events, struct ucred *cred);
 102 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 103 static int pipe_stat (struct file *fp, struct stat *sb, struct ucred *cred);
 104 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct ucred *cred);
 105
 106 static struct fileops pipeops = {
 107         .fo_read = pipe_read,
 108         .fo_write = pipe_write,
 109         .fo_ioctl = pipe_ioctl,
 110         .fo_poll = pipe_poll,
 111         .fo_kqfilter = pipe_kqfilter,
 112         .fo_stat = pipe_stat,
 113         .fo_close = pipe_close,
 114         .fo_shutdown = pipe_shutdown
 115 };
 116
 117 static void     filt_pipedetach(struct knote *kn);
 118 static int      filt_piperead(struct knote *kn, long hint);
 119 static int      filt_pipewrite(struct knote *kn, long hint);
 120
 121 static struct filterops pipe_rfiltops =
 122         { 1, NULL, filt_pipedetach, filt_piperead };
 123 static struct filterops pipe_wfiltops =
 124         { 1, NULL, filt_pipedetach, filt_pipewrite };
 125
 126 MALLOC_DEFINE(M_PIPE, "pipe", "pipe structures");
 127
 128 /*
 129  * Default pipe buffer size(s), this can be kind-of large now because pipe
 130  * space is pageable.  The pipe code will try to maintain locality of
 131  * reference for performance reasons, so small amounts of outstanding I/O
 132  * will not wipe the cache.
 133  */
 134 #define MINPIPESIZE (PIPE_SIZE/3)
 135 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 136
 137 /*
 138  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 139  * is there so that on large systems, we don't exhaust it.
 140  */
 141 #define MAXPIPEKVA (8*1024*1024)
 142
 143 /*
 144  * Limit for direct transfers, we cannot, of course limit
 145  * the amount of kva for pipes in general though.
 146  */
 147 #define LIMITPIPEKVA (16*1024*1024)
 148
 149 /*
 150  * Limit the number of "big" pipes
 151  */
 152 #define LIMITBIGPIPES   32
 153 #define PIPEQ_MAX_CACHE 16      /* per-cpu pipe structure cache */
 154
 155 static int pipe_maxbig = LIMITBIGPIPES;
 156 static int pipe_maxcache = PIPEQ_MAX_CACHE;
 157 static int pipe_nbig;
 158 static int pipe_bcache_alloc;
 159 static int pipe_bkmem_alloc;
 160 static int pipe_dwrite_enable = 1;      /* 0:copy, 1:kmem/sfbuf 2:force */
 161 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 162                                         /* 3:sfbuf_dmap w/ forced invlpg */
 163
 164 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 165 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 166         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 167 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 168         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 169 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 170         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 171 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 172         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 173 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 174         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
 175         "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 176 #if !defined(NO_PIPE_SYSCTL_STATS)
 177 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 178         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 179 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 180         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 181 #endif
 182
 183 static void pipeclose (struct pipe *cpipe);
 184 static void pipe_free_kmem (struct pipe *cpipe);
 185 static int pipe_create (struct pipe **cpipep);
 186 static __inline int pipelock (struct pipe *cpipe, int catch);
 187 static __inline void pipeunlock (struct pipe *cpipe);
 188 static __inline void pipeselwakeup (struct pipe *cpipe);
 189 #ifndef PIPE_NODIRECT
 190 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 191 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 192 static void pipe_clone_write_buffer (struct pipe *wpipe);
 193 #endif
 194 static int pipespace (struct pipe *cpipe, int size);
 195
 196 /*
 197  * The pipe system call for the DTYPE_PIPE type of pipes
 198  *
 199  * pipe_ARgs(int dummy)
 200  */
 201
 202 /* ARGSUSED */
 203 int
 204 sys_pipe(struct pipe_args *uap)
 205 {
 206         struct thread *td = curthread;
 207         struct proc *p = td->td_proc;
 208         struct file *rf, *wf;
 209         struct pipe *rpipe, *wpipe;
 210         int fd1, fd2, error;
 211
 212         KKASSERT(p);
 213
 214         rpipe = wpipe = NULL;
 215         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 216                 pipeclose(rpipe);
 217                 pipeclose(wpipe);
 218                 return (ENFILE);
 219         }
 220
 221         rpipe->pipe_state |= PIPE_DIRECTOK;
 222         wpipe->pipe_state |= PIPE_DIRECTOK;
 223
 224         /*
 225          * Select the direct-map features to use for this pipe.  Since the
 226          * sysctl's can change on the fly we record the settings when the
 227          * pipe is created.
 228          *
 229          * Generally speaking the system will default to what we consider
 230          * to be the best-balanced and most stable option.  Right now this
 231          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 232          * moment.
 233          */
 234         wpipe->pipe_feature = PIPE_COPY;
 235         if (pipe_dwrite_enable) {
 236                 switch(pipe_dwrite_sfbuf) {
 237                 case 0:
 238                         wpipe->pipe_feature = PIPE_KMEM;
 239                         break;
 240                 case 1:
 241                         wpipe->pipe_feature = PIPE_SFBUF1;
 242                         break;
 243                 case 2:
 244                 case 3:
 245                         wpipe->pipe_feature = PIPE_SFBUF2;
 246                         break;
 247                 }
 248         }
 249         rpipe->pipe_feature = wpipe->pipe_feature;
 250
 251         error = falloc(p, &rf, &fd1);
 252         if (error) {
 253                 pipeclose(rpipe);
 254                 pipeclose(wpipe);
 255                 return (error);
 256         }
 257         uap->sysmsg_fds[0] = fd1;
 258
 259         /*
 260          * Warning: once we've gotten past allocation of the fd for the
 261          * read-side, we can only drop the read side via fdrop() in order
 262          * to avoid races against processes which manage to dup() the read
 263          * side while we are blocked trying to allocate the write side.
 264          */
 265         rf->f_type = DTYPE_PIPE;
 266         rf->f_flag = FREAD | FWRITE;
 267         rf->f_ops = &pipeops;
 268         rf->f_data = rpipe;
 269         error = falloc(p, &wf, &fd2);
 270         if (error) {
 271                 fsetfd(p, NULL, fd1);
 272                 fdrop(rf);
 273                 /* rpipe has been closed by fdrop(). */
 274                 pipeclose(wpipe);
 275                 return (error);
 276         }
 277         wf->f_type = DTYPE_PIPE;
 278         wf->f_flag = FREAD | FWRITE;
 279         wf->f_ops = &pipeops;
 280         wf->f_data = wpipe;
 281         uap->sysmsg_fds[1] = fd2;
 282
 283         rpipe->pipe_peer = wpipe;
 284         wpipe->pipe_peer = rpipe;
 285
 286         fsetfd(p, rf, fd1);
 287         fsetfd(p, wf, fd2);
 288         fdrop(rf);
 289         fdrop(wf);
 290
 291         return (0);
 292 }
 293
 294 /*
 295  * Allocate kva for pipe circular buffer, the space is pageable
 296  * This routine will 'realloc' the size of a pipe safely, if it fails
 297  * it will retain the old buffer.
 298  * If it fails it will return ENOMEM.
 299  */
 300 static int
 301 pipespace(struct pipe *cpipe, int size)
 302 {
 303         struct vm_object *object;
 304         caddr_t buffer;
 305         int npages, error;
 306
 307         npages = round_page(size) / PAGE_SIZE;
 308         object = cpipe->pipe_buffer.object;
 309
 310         /*
 311          * [re]create the object if necessary and reserve space for it
 312          * in the kernel_map.  The object and memory are pageable.  On
 313          * success, free the old resources before assigning the new
 314          * ones.
 315          */
 316         if (object == NULL || object->size != npages) {
 317                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 318                 buffer = (caddr_t)vm_map_min(&kernel_map);
 319
 320                 error = vm_map_find(&kernel_map, object, 0,
 321                                     (vm_offset_t *)&buffer, size,
 322                                     1,
 323                                     VM_MAPTYPE_NORMAL,
 324                                     VM_PROT_ALL, VM_PROT_ALL,
 325                                     0);
 326
 327                 if (error != KERN_SUCCESS) {
 328                         vm_object_deallocate(object);
 329                         return (ENOMEM);
 330                 }
 331                 pipe_free_kmem(cpipe);
 332                 cpipe->pipe_buffer.object = object;
 333                 cpipe->pipe_buffer.buffer = buffer;
 334                 cpipe->pipe_buffer.size = size;
 335                 ++pipe_bkmem_alloc;
 336         } else {
 337                 ++pipe_bcache_alloc;
 338         }
 339         cpipe->pipe_buffer.in = 0;
 340         cpipe->pipe_buffer.out = 0;
 341         cpipe->pipe_buffer.cnt = 0;
 342         return (0);
 343 }
 344
 345 /*
 346  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 347  * our per-cpu cache if possible.  For now make sure it is sized for the
 348  * smaller PIPE_SIZE default.
 349  */
 350 static int
 351 pipe_create(struct pipe **cpipep)
 352 {
 353         globaldata_t gd = mycpu;
 354         struct pipe *cpipe;
 355         int error;
 356
 357         if ((cpipe = gd->gd_pipeq) != NULL) {
 358                 gd->gd_pipeq = cpipe->pipe_peer;
 359                 --gd->gd_pipeqcount;
 360                 cpipe->pipe_peer = NULL;
 361         } else {
 362                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 363         }
 364         *cpipep = cpipe;
 365         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 366                 return (error);
 367         vfs_timestamp(&cpipe->pipe_ctime);
 368         cpipe->pipe_atime = cpipe->pipe_ctime;
 369         cpipe->pipe_mtime = cpipe->pipe_ctime;
 370         return (0);
 371 }
 372
 373
 374 /*
 375  * lock a pipe for I/O, blocking other access
 376  */
 377 static __inline int
 378 pipelock(struct pipe *cpipe, int catch)
 379 {
 380         int error;
 381
 382         while (cpipe->pipe_state & PIPE_LOCK) {
 383                 cpipe->pipe_state |= PIPE_LWANT;
 384                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 385                 if (error != 0)
 386                         return (error);
 387         }
 388         cpipe->pipe_state |= PIPE_LOCK;
 389         return (0);
 390 }
 391
 392 /*
 393  * unlock a pipe I/O lock
 394  */
 395 static __inline void
 396 pipeunlock(struct pipe *cpipe)
 397 {
 398
 399         cpipe->pipe_state &= ~PIPE_LOCK;
 400         if (cpipe->pipe_state & PIPE_LWANT) {
 401                 cpipe->pipe_state &= ~PIPE_LWANT;
 402                 wakeup(cpipe);
 403         }
 404 }
 405
 406 static __inline void
 407 pipeselwakeup(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 /*
 420  * MPALMOSTSAFE (acquires mplock)
 421  */
 422 static int
 423 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 424 {
 425         struct pipe *rpipe;
 426         int error;
 427         int nread = 0;
 428         int nbio;
 429         u_int size;
 430
 431         get_mplock();
 432         rpipe = (struct pipe *) fp->f_data;
 433         ++rpipe->pipe_busy;
 434         error = pipelock(rpipe, 1);
 435         if (error)
 436                 goto unlocked_error;
 437
 438         if (fflags & O_FBLOCKING)
 439                 nbio = 0;
 440         else if (fflags & O_FNONBLOCKING)
 441                 nbio = 1;
 442         else if (fp->f_flag & O_NONBLOCK)
 443                 nbio = 1;
 444         else
 445                 nbio = 0;
 446
 447         while (uio->uio_resid) {
 448                 caddr_t va;
 449
 450                 if (rpipe->pipe_buffer.cnt > 0) {
 451                         /*
 452                          * normal pipe buffer receive
 453                          */
 454                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 455                         if (size > rpipe->pipe_buffer.cnt)
 456                                 size = rpipe->pipe_buffer.cnt;
 457                         if (size > (u_int) uio->uio_resid)
 458                                 size = (u_int) uio->uio_resid;
 459
 460                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 461                                         size, uio);
 462                         if (error)
 463                                 break;
 464
 465                         rpipe->pipe_buffer.out += size;
 466                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 467                                 rpipe->pipe_buffer.out = 0;
 468
 469                         rpipe->pipe_buffer.cnt -= size;
 470
 471                         /*
 472                          * If there is no more to read in the pipe, reset
 473                          * its pointers to the beginning.  This improves
 474                          * cache hit stats.
 475                          */
 476                         if (rpipe->pipe_buffer.cnt == 0) {
 477                                 rpipe->pipe_buffer.in = 0;
 478                                 rpipe->pipe_buffer.out = 0;
 479                         }
 480                         nread += size;
 481 #ifndef PIPE_NODIRECT
 482                 } else if (rpipe->pipe_kva &&
 483                            rpipe->pipe_feature == PIPE_KMEM &&
 484                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 485                                == PIPE_DIRECTW
 486                 ) {
 487                         /*
 488                          * Direct copy using source-side kva mapping
 489                          */
 490                         size = rpipe->pipe_map.xio_bytes -
 491                                 rpipe->pipe_buffer.out;
 492                         if (size > (u_int)uio->uio_resid)
 493                                 size = (u_int)uio->uio_resid;
 494                         va = (caddr_t)rpipe->pipe_kva +
 495                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 496                         error = uiomove(va, size, uio);
 497                         if (error)
 498                                 break;
 499                         nread += size;
 500                         rpipe->pipe_buffer.out += size;
 501                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 502                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 503                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 504                                 /* reset out index for copy mode */
 505                                 rpipe->pipe_buffer.out = 0;
 506                                 wakeup(rpipe);
 507                         }
 508                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 509                            rpipe->pipe_kva &&
 510                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 511                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 512                                == PIPE_DIRECTW
 513                 ) {
 514                         /*
 515                          * Direct copy, bypassing a kernel buffer.  We cannot
 516                          * mess with the direct-write buffer until
 517                          * PIPE_DIRECTIP is cleared.  In order to prevent
 518                          * the pipe_write code from racing itself in
 519                          * direct_write, we set DIRECTIP when we clear
 520                          * DIRECTW after we have exhausted the buffer.
 521                          */
 522                         if (pipe_dwrite_sfbuf == 3)
 523                                 rpipe->pipe_kvamask = 0;
 524                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 525                                     rpipe->pipe_map.xio_npages,
 526                                     &rpipe->pipe_kvamask);
 527                         size = rpipe->pipe_map.xio_bytes -
 528                                 rpipe->pipe_buffer.out;
 529                         if (size > (u_int)uio->uio_resid)
 530                                 size = (u_int)uio->uio_resid;
 531                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 532                         error = uiomove(va, size, uio);
 533                         if (error)
 534                                 break;
 535                         nread += size;
 536                         rpipe->pipe_buffer.out += size;
 537                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 538                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 539                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 540                                 /* reset out index for copy mode */
 541                                 rpipe->pipe_buffer.out = 0;
 542                                 wakeup(rpipe);
 543                         }
 544                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 545                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 546                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 547                                 == PIPE_DIRECTW
 548                 ) {
 549                         /*
 550                          * Direct copy, bypassing a kernel buffer.  We cannot
 551                          * mess with the direct-write buffer until
 552                          * PIPE_DIRECTIP is cleared.  In order to prevent
 553                          * the pipe_write code from racing itself in
 554                          * direct_write, we set DIRECTIP when we clear
 555                          * DIRECTW after we have exhausted the buffer.
 556                          */
 557                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 558                         if (error)
 559                                 break;
 560                         nread += size;
 561                         rpipe->pipe_buffer.out += size;
 562                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 563                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 564                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 565                                 /* reset out index for copy mode */
 566                                 rpipe->pipe_buffer.out = 0;
 567                                 wakeup(rpipe);
 568                         }
 569 #endif
 570                 } else {
 571                         /*
 572                          * detect EOF condition
 573                          * read returns 0 on EOF, no need to set error
 574                          */
 575                         if (rpipe->pipe_state & PIPE_EOF)
 576                                 break;
 577
 578                         /*
 579                          * If the "write-side" has been blocked, wake it up now.
 580                          */
 581                         if (rpipe->pipe_state & PIPE_WANTW) {
 582                                 rpipe->pipe_state &= ~PIPE_WANTW;
 583                                 wakeup(rpipe);
 584                         }
 585
 586                         /*
 587                          * Break if some data was read.
 588                          */
 589                         if (nread > 0)
 590                                 break;
 591
 592                         /*
 593                          * Unlock the pipe buffer for our remaining
 594                          * processing.  We will either break out with an
 595                          * error or we will sleep and relock to loop.
 596                          */
 597                         pipeunlock(rpipe);
 598
 599                         /*
 600                          * Handle non-blocking mode operation or
 601                          * wait for more data.
 602                          */
 603                         if (nbio) {
 604                                 error = EAGAIN;
 605                         } else {
 606                                 rpipe->pipe_state |= PIPE_WANTR;
 607                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 608                                     "piperd", 0)) == 0) {
 609                                         error = pipelock(rpipe, 1);
 610                                 }
 611                         }
 612                         if (error)
 613                                 goto unlocked_error;
 614                 }
 615         }
 616         pipeunlock(rpipe);
 617
 618         if (error == 0)
 619                 vfs_timestamp(&rpipe->pipe_atime);
 620 unlocked_error:
 621         --rpipe->pipe_busy;
 622
 623         /*
 624          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 625          */
 626         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 627                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 628                 wakeup(rpipe);
 629         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 630                 /*
 631                  * Handle write blocking hysteresis.
 632                  */
 633                 if (rpipe->pipe_state & PIPE_WANTW) {
 634                         rpipe->pipe_state &= ~PIPE_WANTW;
 635                         wakeup(rpipe);
 636                 }
 637         }
 638
 639         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 640                 pipeselwakeup(rpipe);
 641         rel_mplock();
 642         return (error);
 643 }
 644
 645 #ifndef PIPE_NODIRECT
 646 /*
 647  * Map the sending processes' buffer into kernel space and wire it.
 648  * This is similar to a physical write operation.
 649  */
 650 static int
 651 pipe_build_write_buffer(struct pipe *wpipe, 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(struct pipe *wpipe)
 717 {
 718         int size;
 719         int offset;
 720
 721         offset = wpipe->pipe_buffer.out;
 722         size = wpipe->pipe_map.xio_bytes - offset;
 723
 724         KKASSERT(size <= wpipe->pipe_buffer.size);
 725
 726         wpipe->pipe_buffer.in = size;
 727         wpipe->pipe_buffer.out = 0;
 728         wpipe->pipe_buffer.cnt = size;
 729         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 730
 731         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 732         xio_release(&wpipe->pipe_map);
 733         if (wpipe->pipe_kva) {
 734                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 735                 kmem_free(&kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 736                 wpipe->pipe_kva = NULL;
 737         }
 738 }
 739
 740 /*
 741  * This implements the pipe buffer write mechanism.  Note that only
 742  * a direct write OR a normal pipe write can be pending at any given time.
 743  * If there are any characters in the pipe buffer, the direct write will
 744  * be deferred until the receiving process grabs all of the bytes from
 745  * the pipe buffer.  Then the direct mapping write is set-up.
 746  */
 747 static int
 748 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 749 {
 750         int error;
 751
 752 retry:
 753         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 754                 if (wpipe->pipe_state & PIPE_WANTR) {
 755                         wpipe->pipe_state &= ~PIPE_WANTR;
 756                         wakeup(wpipe);
 757                 }
 758                 wpipe->pipe_state |= PIPE_WANTW;
 759                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 760                 if (error)
 761                         goto error2;
 762                 if (wpipe->pipe_state & PIPE_EOF) {
 763                         error = EPIPE;
 764                         goto error2;
 765                 }
 766         }
 767         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 768         if (wpipe->pipe_buffer.cnt > 0) {
 769                 if (wpipe->pipe_state & PIPE_WANTR) {
 770                         wpipe->pipe_state &= ~PIPE_WANTR;
 771                         wakeup(wpipe);
 772                 }
 773
 774                 wpipe->pipe_state |= PIPE_WANTW;
 775                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 776                 if (error)
 777                         goto error2;
 778                 if (wpipe->pipe_state & PIPE_EOF) {
 779                         error = EPIPE;
 780                         goto error2;
 781                 }
 782                 goto retry;
 783         }
 784
 785         /*
 786          * Build our direct-write buffer
 787          */
 788         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 789         error = pipe_build_write_buffer(wpipe, uio);
 790         if (error)
 791                 goto error1;
 792         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 793
 794         /*
 795          * Wait until the receiver has snarfed the data.  Since we are likely
 796          * going to sleep we optimize the case and yield synchronously,
 797          * possibly avoiding the tsleep().
 798          */
 799         error = 0;
 800         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 801                 if (wpipe->pipe_state & PIPE_EOF) {
 802                         pipelock(wpipe, 0);
 803                         xio_release(&wpipe->pipe_map);
 804                         if (wpipe->pipe_kva) {
 805                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 806                                 kmem_free(&kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 807                                 wpipe->pipe_kva = NULL;
 808                         }
 809                         pipeunlock(wpipe);
 810                         pipeselwakeup(wpipe);
 811                         error = EPIPE;
 812                         goto error1;
 813                 }
 814                 if (wpipe->pipe_state & PIPE_WANTR) {
 815                         wpipe->pipe_state &= ~PIPE_WANTR;
 816                         wakeup(wpipe);
 817                 }
 818                 pipeselwakeup(wpipe);
 819                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 820         }
 821         pipelock(wpipe,0);
 822         if (wpipe->pipe_state & PIPE_DIRECTW) {
 823                 /*
 824                  * this bit of trickery substitutes a kernel buffer for
 825                  * the process that might be going away.
 826                  */
 827                 pipe_clone_write_buffer(wpipe);
 828                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 829         } else {
 830                 /*
 831                  * note: The pipe_kva mapping is not qremove'd here.  For
 832                  * legacy PIPE_KMEM mode this constitutes an improvement
 833                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 834                  * mode the kva mapping must not be removed to get the
 835                  * caching benefit.
 836                  *
 837                  * For testing purposes we will give the original algorithm
 838                  * the benefit of the doubt 'what it could have been', and
 839                  * keep the optimization.
 840                  */
 841                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 842                 xio_release(&wpipe->pipe_map);
 843                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 844         }
 845         pipeunlock(wpipe);
 846         return (error);
 847
 848         /*
 849          * Direct-write error, clear the direct write flags.
 850          */
 851 error1:
 852         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 853         /* fallthrough */
 854
 855         /*
 856          * General error, wakeup the other side if it happens to be sleeping.
 857          */
 858 error2:
 859         wakeup(wpipe);
 860         return (error);
 861 }
 862 #endif
 863
 864 /*
 865  * MPALMOSTSAFE - acquires mplock
 866  */
 867 static int
 868 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 869 {
 870         int error = 0;
 871         int orig_resid;
 872         int nbio;
 873         struct pipe *wpipe, *rpipe;
 874
 875         get_mplock();
 876         rpipe = (struct pipe *) fp->f_data;
 877         wpipe = rpipe->pipe_peer;
 878
 879         /*
 880          * detect loss of pipe read side, issue SIGPIPE if lost.
 881          */
 882         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 883                 rel_mplock();
 884                 return (EPIPE);
 885         }
 886         ++wpipe->pipe_busy;
 887
 888         if (fflags & O_FBLOCKING)
 889                 nbio = 0;
 890         else if (fflags & O_FNONBLOCKING)
 891                 nbio = 1;
 892         else if (fp->f_flag & O_NONBLOCK)
 893                 nbio = 1;
 894         else
 895                 nbio = 0;
 896
 897         /*
 898          * If it is advantageous to resize the pipe buffer, do
 899          * so.
 900          */
 901         if ((uio->uio_resid > PIPE_SIZE) &&
 902                 (pipe_nbig < pipe_maxbig) &&
 903                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 904                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 905                 (wpipe->pipe_buffer.cnt == 0)) {
 906
 907                 if ((error = pipelock(wpipe,1)) == 0) {
 908                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 909                                 pipe_nbig++;
 910                         pipeunlock(wpipe);
 911                 }
 912         }
 913
 914         /*
 915          * If an early error occured unbusy and return, waking up any pending
 916          * readers.
 917          */
 918         if (error) {
 919                 --wpipe->pipe_busy;
 920                 if ((wpipe->pipe_busy == 0) &&
 921                     (wpipe->pipe_state & PIPE_WANT)) {
 922                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 923                         wakeup(wpipe);
 924                 }
 925                 rel_mplock();
 926                 return(error);
 927         }
 928
 929         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 930
 931         orig_resid = uio->uio_resid;
 932
 933         while (uio->uio_resid) {
 934                 int space;
 935
 936 #ifndef PIPE_NODIRECT
 937                 /*
 938                  * If the transfer is large, we can gain performance if
 939                  * we do process-to-process copies directly.
 940                  * If the write is non-blocking, we don't use the
 941                  * direct write mechanism.
 942                  *
 943                  * The direct write mechanism will detect the reader going
 944                  * away on us.
 945                  */
 946                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 947                     pipe_dwrite_enable > 1) &&
 948                     nbio == 0 &&
 949                     pipe_dwrite_enable) {
 950                         error = pipe_direct_write( wpipe, uio);
 951                         if (error)
 952                                 break;
 953                         continue;
 954                 }
 955 #endif
 956
 957                 /*
 958                  * Pipe buffered writes cannot be coincidental with
 959                  * direct writes.  We wait until the currently executing
 960                  * direct write is completed before we start filling the
 961                  * pipe buffer.  We break out if a signal occurs or the
 962                  * reader goes away.
 963                  */
 964         retrywrite:
 965                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 966                         if (wpipe->pipe_state & PIPE_WANTR) {
 967                                 wpipe->pipe_state &= ~PIPE_WANTR;
 968                                 wakeup(wpipe);
 969                         }
 970                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 971                         if (wpipe->pipe_state & PIPE_EOF)
 972                                 break;
 973                         if (error)
 974                                 break;
 975                 }
 976                 if (wpipe->pipe_state & PIPE_EOF) {
 977                         error = EPIPE;
 978                         break;
 979                 }
 980
 981                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 982
 983                 /* Writes of size <= PIPE_BUF must be atomic. */
 984                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 985                         space = 0;
 986
 987                 /*
 988                  * Write to fill, read size handles write hysteresis.  Also
 989                  * additional restrictions can cause select-based non-blocking
 990                  * writes to spin.
 991                  */
 992                 if (space > 0) {
 993                         if ((error = pipelock(wpipe,1)) == 0) {
 994                                 int size;       /* Transfer size */
 995                                 int segsize;    /* first segment to transfer */
 996
 997                                 /*
 998                                  * It is possible for a direct write to
 999                                  * slip in on us... handle it here...
1000                                  */
1001                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
1002                                         pipeunlock(wpipe);
1003                                         goto retrywrite;
1004                                 }
1005                                 /*
1006                                  * If a process blocked in uiomove, our
1007                                  * value for space might be bad.
1008                                  *
1009                                  * XXX will we be ok if the reader has gone
1010                                  * away here?
1011                                  */
1012                                 if (space > wpipe->pipe_buffer.size -
1013                                     wpipe->pipe_buffer.cnt) {
1014                                         pipeunlock(wpipe);
1015                                         goto retrywrite;
1016                                 }
1017
1018                                 /*
1019                                  * Transfer size is minimum of uio transfer
1020                                  * and free space in pipe buffer.
1021                                  */
1022                                 if (space > uio->uio_resid)
1023                                         size = uio->uio_resid;
1024                                 else
1025                                         size = space;
1026                                 /*
1027                                  * First segment to transfer is minimum of
1028                                  * transfer size and contiguous space in
1029                                  * pipe buffer.  If first segment to transfer
1030                                  * is less than the transfer size, we've got
1031                                  * a wraparound in the buffer.
1032                                  */
1033                                 segsize = wpipe->pipe_buffer.size -
1034                                         wpipe->pipe_buffer.in;
1035                                 if (segsize > size)
1036                                         segsize = size;
1037
1038                                 /* Transfer first segment */
1039
1040                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
1041                                                 segsize, uio);
1042
1043                                 if (error == 0 && segsize < size) {
1044                                         /*
1045                                          * Transfer remaining part now, to
1046                                          * support atomic writes.  Wraparound
1047                                          * happened.
1048                                          */
1049                                         if (wpipe->pipe_buffer.in + segsize !=
1050                                             wpipe->pipe_buffer.size)
1051                                                 panic("Expected pipe buffer wraparound disappeared");
1052
1053                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
1054                                                         size - segsize, uio);
1055                                 }
1056                                 if (error == 0) {
1057                                         wpipe->pipe_buffer.in += size;
1058                                         if (wpipe->pipe_buffer.in >=
1059                                             wpipe->pipe_buffer.size) {
1060                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1061                                                         panic("Expected wraparound bad");
1062                                                 wpipe->pipe_buffer.in = size - segsize;
1063                                         }
1064
1065                                         wpipe->pipe_buffer.cnt += size;
1066                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1067                                                 panic("Pipe buffer overflow");
1068
1069                                 }
1070                                 pipeunlock(wpipe);
1071                         }
1072                         if (error)
1073                                 break;
1074
1075                 } else {
1076                         /*
1077                          * If the "read-side" has been blocked, wake it up now
1078                          * and yield to let it drain synchronously rather
1079                          * then block.
1080                          */
1081                         if (wpipe->pipe_state & PIPE_WANTR) {
1082                                 wpipe->pipe_state &= ~PIPE_WANTR;
1083                                 wakeup(wpipe);
1084                         }
1085
1086                         /*
1087                          * don't block on non-blocking I/O
1088                          */
1089                         if (nbio) {
1090                                 error = EAGAIN;
1091                                 break;
1092                         }
1093
1094                         /*
1095                          * We have no more space and have something to offer,
1096                          * wake up select/poll.
1097                          */
1098                         pipeselwakeup(wpipe);
1099
1100                         wpipe->pipe_state |= PIPE_WANTW;
1101                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1102                         if (error != 0)
1103                                 break;
1104                         /*
1105                          * If read side wants to go away, we just issue a signal
1106                          * to ourselves.
1107                          */
1108                         if (wpipe->pipe_state & PIPE_EOF) {
1109                                 error = EPIPE;
1110                                 break;
1111                         }
1112                 }
1113         }
1114
1115         --wpipe->pipe_busy;
1116
1117         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1118                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1119                 wakeup(wpipe);
1120         } else if (wpipe->pipe_buffer.cnt > 0) {
1121                 /*
1122                  * If we have put any characters in the buffer, we wake up
1123                  * the reader.
1124                  */
1125                 if (wpipe->pipe_state & PIPE_WANTR) {
1126                         wpipe->pipe_state &= ~PIPE_WANTR;
1127                         wakeup(wpipe);
1128                 }
1129         }
1130
1131         /*
1132          * Don't return EPIPE if I/O was successful
1133          */
1134         if ((wpipe->pipe_buffer.cnt == 0) &&
1135             (uio->uio_resid == 0) &&
1136             (error == EPIPE)) {
1137                 error = 0;
1138         }
1139
1140         if (error == 0)
1141                 vfs_timestamp(&wpipe->pipe_mtime);
1142
1143         /*
1144          * We have something to offer,
1145          * wake up select/poll.
1146          */
1147         if (wpipe->pipe_buffer.cnt)
1148                 pipeselwakeup(wpipe);
1149         rel_mplock();
1150         return (error);
1151 }
1152
1153 /*
1154  * MPALMOSTSAFE - acquires mplock
1155  *
1156  * we implement a very minimal set of ioctls for compatibility with sockets.
1157  */
1158 int
1159 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1160 {
1161         struct pipe *mpipe;
1162         int error;
1163
1164         get_mplock();
1165         mpipe = (struct pipe *)fp->f_data;
1166
1167         switch (cmd) {
1168         case FIOASYNC:
1169                 if (*(int *)data) {
1170                         mpipe->pipe_state |= PIPE_ASYNC;
1171                 } else {
1172                         mpipe->pipe_state &= ~PIPE_ASYNC;
1173                 }
1174                 error = 0;
1175                 break;
1176         case FIONREAD:
1177                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1178                         *(int *)data = mpipe->pipe_map.xio_bytes -
1179                                         mpipe->pipe_buffer.out;
1180                 } else {
1181                         *(int *)data = mpipe->pipe_buffer.cnt;
1182                 }
1183                 error = 0;
1184                 break;
1185         case FIOSETOWN:
1186                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1187                 break;
1188         case FIOGETOWN:
1189                 *(int *)data = fgetown(mpipe->pipe_sigio);
1190                 error = 0;
1191                 break;
1192         case TIOCSPGRP:
1193                 /* This is deprecated, FIOSETOWN should be used instead. */
1194                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1195                 break;
1196
1197         case TIOCGPGRP:
1198                 /* This is deprecated, FIOGETOWN should be used instead. */
1199                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1200                 error = 0;
1201                 break;
1202         default:
1203                 error = ENOTTY;
1204                 break;
1205         }
1206         rel_mplock();
1207         return (error);
1208 }
1209
1210 /*
1211  * MPALMOSTSAFE - acquires mplock
1212  */
1213 int
1214 pipe_poll(struct file *fp, int events, struct ucred *cred)
1215 {
1216         struct pipe *rpipe;
1217         struct pipe *wpipe;
1218         int revents = 0;
1219
1220         get_mplock();
1221         rpipe = (struct pipe *)fp->f_data;
1222         wpipe = rpipe->pipe_peer;
1223         if (events & (POLLIN | POLLRDNORM))
1224                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1225                     (rpipe->pipe_buffer.cnt > 0) ||
1226                     (rpipe->pipe_state & PIPE_EOF))
1227                         revents |= events & (POLLIN | POLLRDNORM);
1228
1229         if (events & (POLLOUT | POLLWRNORM))
1230                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1231                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1232                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1233                         revents |= events & (POLLOUT | POLLWRNORM);
1234
1235         if ((rpipe->pipe_state & PIPE_EOF) ||
1236             (wpipe == NULL) ||
1237             (wpipe->pipe_state & PIPE_EOF))
1238                 revents |= POLLHUP;
1239
1240         if (revents == 0) {
1241                 if (events & (POLLIN | POLLRDNORM)) {
1242                         selrecord(curthread, &rpipe->pipe_sel);
1243                         rpipe->pipe_state |= PIPE_SEL;
1244                 }
1245
1246                 if (events & (POLLOUT | POLLWRNORM)) {
1247                         selrecord(curthread, &wpipe->pipe_sel);
1248                         wpipe->pipe_state |= PIPE_SEL;
1249                 }
1250         }
1251         rel_mplock();
1252         return (revents);
1253 }
1254
1255 /*
1256  * MPALMOSTSAFE - acquires mplock
1257  */
1258 static int
1259 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1260 {
1261         struct pipe *pipe;
1262
1263         get_mplock();
1264         pipe = (struct pipe *)fp->f_data;
1265
1266         bzero((caddr_t)ub, sizeof(*ub));
1267         ub->st_mode = S_IFIFO;
1268         ub->st_blksize = pipe->pipe_buffer.size;
1269         ub->st_size = pipe->pipe_buffer.cnt;
1270         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1271                 ub->st_size = pipe->pipe_map.xio_bytes -
1272                                 pipe->pipe_buffer.out;
1273         }
1274         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1275         ub->st_atimespec = pipe->pipe_atime;
1276         ub->st_mtimespec = pipe->pipe_mtime;
1277         ub->st_ctimespec = pipe->pipe_ctime;
1278         /*
1279          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1280          * st_flags, st_gen.
1281          * XXX (st_dev, st_ino) should be unique.
1282          */
1283         rel_mplock();
1284         return (0);
1285 }
1286
1287 /*
1288  * MPALMOSTSAFE - acquires mplock
1289  */
1290 static int
1291 pipe_close(struct file *fp)
1292 {
1293         struct pipe *cpipe = (struct pipe *)fp->f_data;
1294
1295         get_mplock();
1296         fp->f_ops = &badfileops;
1297         fp->f_data = NULL;
1298         funsetown(cpipe->pipe_sigio);
1299         pipeclose(cpipe);
1300         rel_mplock();
1301         return (0);
1302 }
1303
1304 /*
1305  * Shutdown one or both directions of a full-duplex pipe.
1306  *
1307  * MPALMOSTSAFE - acquires mplock
1308  */
1309 static int
1310 pipe_shutdown(struct file *fp, int how)
1311 {
1312         struct pipe *rpipe;
1313         struct pipe *wpipe;
1314         int error = EPIPE;
1315
1316         get_mplock();
1317         rpipe = (struct pipe *)fp->f_data;
1318
1319         switch(how) {
1320         case SHUT_RDWR:
1321         case SHUT_RD:
1322                 if (rpipe) {
1323                         rpipe->pipe_state |= PIPE_EOF;
1324                         pipeselwakeup(rpipe);
1325                         if (rpipe->pipe_busy)
1326                                 wakeup(rpipe);
1327                         error = 0;
1328                 }
1329                 if (how == SHUT_RD)
1330                         break;
1331                 /* fall through */
1332         case SHUT_WR:
1333                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1334                         wpipe->pipe_state |= PIPE_EOF;
1335                         pipeselwakeup(wpipe);
1336                         if (wpipe->pipe_busy)
1337                                 wakeup(wpipe);
1338                         error = 0;
1339                 }
1340         }
1341         rel_mplock();
1342         return (error);
1343 }
1344
1345 static void
1346 pipe_free_kmem(struct pipe *cpipe)
1347 {
1348         if (cpipe->pipe_buffer.buffer != NULL) {
1349                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1350                         --pipe_nbig;
1351                 kmem_free(&kernel_map,
1352                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1353                         cpipe->pipe_buffer.size);
1354                 cpipe->pipe_buffer.buffer = NULL;
1355                 cpipe->pipe_buffer.object = NULL;
1356         }
1357 #ifndef PIPE_NODIRECT
1358         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1359                 cpipe->pipe_map.xio_offset == 0 &&
1360                 cpipe->pipe_map.xio_npages == 0);
1361 #endif
1362 }
1363
1364 /*
1365  * shutdown the pipe
1366  */
1367 static void
1368 pipeclose(struct pipe *cpipe)
1369 {
1370         globaldata_t gd;
1371         struct pipe *ppipe;
1372
1373         if (cpipe == NULL)
1374                 return;
1375
1376         pipeselwakeup(cpipe);
1377
1378         /*
1379          * If the other side is blocked, wake it up saying that
1380          * we want to close it down.
1381          */
1382         while (cpipe->pipe_busy) {
1383                 wakeup(cpipe);
1384                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1385                 tsleep(cpipe, 0, "pipecl", 0);
1386         }
1387
1388         /*
1389          * Disconnect from peer
1390          */
1391         if ((ppipe = cpipe->pipe_peer) != NULL) {
1392                 pipeselwakeup(ppipe);
1393
1394                 ppipe->pipe_state |= PIPE_EOF;
1395                 wakeup(ppipe);
1396                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1397                 ppipe->pipe_peer = NULL;
1398         }
1399
1400         if (cpipe->pipe_kva) {
1401                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1402                 kmem_free(&kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1403                 cpipe->pipe_kva = NULL;
1404         }
1405
1406         /*
1407          * free or cache resources
1408          */
1409         gd = mycpu;
1410         if (gd->gd_pipeqcount >= pipe_maxcache ||
1411             cpipe->pipe_buffer.size != PIPE_SIZE
1412         ) {
1413                 pipe_free_kmem(cpipe);
1414                 kfree(cpipe, M_PIPE);
1415         } else {
1416                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1417                         cpipe->pipe_map.xio_bytes == 0 &&
1418                         cpipe->pipe_map.xio_offset == 0);
1419                 cpipe->pipe_state = 0;
1420                 cpipe->pipe_busy = 0;
1421                 cpipe->pipe_peer = gd->gd_pipeq;
1422                 gd->gd_pipeq = cpipe;
1423                 ++gd->gd_pipeqcount;
1424         }
1425 }
1426
1427 /*
1428  * MPALMOSTSAFE - acquires mplock
1429  */
1430 static int
1431 pipe_kqfilter(struct file *fp, struct knote *kn)
1432 {
1433         struct pipe *cpipe;
1434
1435         get_mplock();
1436         cpipe = (struct pipe *)kn->kn_fp->f_data;
1437
1438         switch (kn->kn_filter) {
1439         case EVFILT_READ:
1440                 kn->kn_fop = &pipe_rfiltops;
1441                 break;
1442         case EVFILT_WRITE:
1443                 kn->kn_fop = &pipe_wfiltops;
1444                 cpipe = cpipe->pipe_peer;
1445                 if (cpipe == NULL) {
1446                         /* other end of pipe has been closed */
1447                         rel_mplock();
1448                         return (EPIPE);
1449                 }
1450                 break;
1451         default:
1452                 return (1);
1453         }
1454         kn->kn_hook = (caddr_t)cpipe;
1455
1456         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1457         rel_mplock();
1458         return (0);
1459 }
1460
1461 static void
1462 filt_pipedetach(struct knote *kn)
1463 {
1464         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1465
1466         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1467 }
1468
1469 /*ARGSUSED*/
1470 static int
1471 filt_piperead(struct knote *kn, long hint)
1472 {
1473         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1474         struct pipe *wpipe = rpipe->pipe_peer;
1475
1476         kn->kn_data = rpipe->pipe_buffer.cnt;
1477         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1478                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1479                                 rpipe->pipe_buffer.out;
1480         }
1481
1482         if ((rpipe->pipe_state & PIPE_EOF) ||
1483             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1484                 kn->kn_flags |= EV_EOF;
1485                 return (1);
1486         }
1487         return (kn->kn_data > 0);
1488 }
1489
1490 /*ARGSUSED*/
1491 static int
1492 filt_pipewrite(struct knote *kn, long hint)
1493 {
1494         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1495         struct pipe *wpipe = rpipe->pipe_peer;
1496
1497         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1498                 kn->kn_data = 0;
1499                 kn->kn_flags |= EV_EOF;
1500                 return (1);
1501         }
1502         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1503         if (wpipe->pipe_state & PIPE_DIRECTW)
1504                 kn->kn_data = 0;
1505
1506         return (kn->kn_data >= PIPE_BUF);
1507 }