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 1.49 2008/06/05 18:06:32 swildner 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_bigcount;
 158 static int pipe_nbig;
 159 static int pipe_bcache_alloc;
 160 static int pipe_bkmem_alloc;
 161 /*
 162  * There's a bug in the sfbuf-based direct write code, not yet located.
 163  * Disable it for now.
 164  */
 165 static int pipe_dwrite_enable = 0;      /* 0:copy, 1:kmem/sfbuf 2:force */
 166 static int pipe_dwrite_sfbuf = 1;       /* 0:kmem_map 1:sfbufs 2:sfbufs_dmap */
 167                                         /* 3:sfbuf_dmap w/ forced invlpg */
 168
 169 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 170 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 171         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 172 SYSCTL_INT(_kern_pipe, OID_AUTO, bigcount,
 173         CTLFLAG_RW, &pipe_bigcount, 0, "number of times pipe expanded");
 174 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 175         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 176 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 177         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 178 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_enable,
 179         CTLFLAG_RW, &pipe_dwrite_enable, 0, "1:enable/2:force direct writes");
 180 SYSCTL_INT(_kern_pipe, OID_AUTO, dwrite_sfbuf,
 181         CTLFLAG_RW, &pipe_dwrite_sfbuf, 0,
 182         "(if dwrite_enable) 0:kmem 1:sfbuf 2:sfbuf_dmap 3:sfbuf_dmap_forceinvlpg");
 183 #if !defined(NO_PIPE_SYSCTL_STATS)
 184 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 185         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 186 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 187         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 188 #endif
 189
 190 static void pipeclose (struct pipe *cpipe);
 191 static void pipe_free_kmem (struct pipe *cpipe);
 192 static int pipe_create (struct pipe **cpipep);
 193 static __inline int pipelock (struct pipe *cpipe, int catch);
 194 static __inline void pipeunlock (struct pipe *cpipe);
 195 static __inline void pipeselwakeup (struct pipe *cpipe);
 196 #ifndef PIPE_NODIRECT
 197 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 198 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 199 static void pipe_clone_write_buffer (struct pipe *wpipe);
 200 #endif
 201 static int pipespace (struct pipe *cpipe, int size);
 202
 203 /*
 204  * The pipe system call for the DTYPE_PIPE type of pipes
 205  *
 206  * pipe_ARgs(int dummy)
 207  */
 208
 209 /* ARGSUSED */
 210 int
 211 sys_pipe(struct pipe_args *uap)
 212 {
 213         struct thread *td = curthread;
 214         struct proc *p = td->td_proc;
 215         struct file *rf, *wf;
 216         struct pipe *rpipe, *wpipe;
 217         int fd1, fd2, error;
 218
 219         KKASSERT(p);
 220
 221         rpipe = wpipe = NULL;
 222         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 223                 pipeclose(rpipe);
 224                 pipeclose(wpipe);
 225                 return (ENFILE);
 226         }
 227
 228         rpipe->pipe_state |= PIPE_DIRECTOK;
 229         wpipe->pipe_state |= PIPE_DIRECTOK;
 230
 231         /*
 232          * Select the direct-map features to use for this pipe.  Since the
 233          * sysctl's can change on the fly we record the settings when the
 234          * pipe is created.
 235          *
 236          * Generally speaking the system will default to what we consider
 237          * to be the best-balanced and most stable option.  Right now this
 238          * is SFBUF1.  Modes 2 and 3 are considered experiemental at the
 239          * moment.
 240          */
 241         wpipe->pipe_feature = PIPE_COPY;
 242         if (pipe_dwrite_enable) {
 243                 switch(pipe_dwrite_sfbuf) {
 244                 case 0:
 245                         wpipe->pipe_feature = PIPE_KMEM;
 246                         break;
 247                 case 1:
 248                         wpipe->pipe_feature = PIPE_SFBUF1;
 249                         break;
 250                 case 2:
 251                 case 3:
 252                         wpipe->pipe_feature = PIPE_SFBUF2;
 253                         break;
 254                 }
 255         }
 256         rpipe->pipe_feature = wpipe->pipe_feature;
 257
 258         error = falloc(p, &rf, &fd1);
 259         if (error) {
 260                 pipeclose(rpipe);
 261                 pipeclose(wpipe);
 262                 return (error);
 263         }
 264         uap->sysmsg_fds[0] = fd1;
 265
 266         /*
 267          * Warning: once we've gotten past allocation of the fd for the
 268          * read-side, we can only drop the read side via fdrop() in order
 269          * to avoid races against processes which manage to dup() the read
 270          * side while we are blocked trying to allocate the write side.
 271          */
 272         rf->f_type = DTYPE_PIPE;
 273         rf->f_flag = FREAD | FWRITE;
 274         rf->f_ops = &pipeops;
 275         rf->f_data = rpipe;
 276         error = falloc(p, &wf, &fd2);
 277         if (error) {
 278                 fsetfd(p, NULL, fd1);
 279                 fdrop(rf);
 280                 /* rpipe has been closed by fdrop(). */
 281                 pipeclose(wpipe);
 282                 return (error);
 283         }
 284         wf->f_type = DTYPE_PIPE;
 285         wf->f_flag = FREAD | FWRITE;
 286         wf->f_ops = &pipeops;
 287         wf->f_data = wpipe;
 288         uap->sysmsg_fds[1] = fd2;
 289
 290         rpipe->pipe_peer = wpipe;
 291         wpipe->pipe_peer = rpipe;
 292
 293         fsetfd(p, rf, fd1);
 294         fsetfd(p, wf, fd2);
 295         fdrop(rf);
 296         fdrop(wf);
 297
 298         return (0);
 299 }
 300
 301 /*
 302  * Allocate kva for pipe circular buffer, the space is pageable
 303  * This routine will 'realloc' the size of a pipe safely, if it fails
 304  * it will retain the old buffer.
 305  * If it fails it will return ENOMEM.
 306  */
 307 static int
 308 pipespace(struct pipe *cpipe, int size)
 309 {
 310         struct vm_object *object;
 311         caddr_t buffer;
 312         int npages, error;
 313
 314         npages = round_page(size) / PAGE_SIZE;
 315         object = cpipe->pipe_buffer.object;
 316
 317         /*
 318          * [re]create the object if necessary and reserve space for it
 319          * in the kernel_map.  The object and memory are pageable.  On
 320          * success, free the old resources before assigning the new
 321          * ones.
 322          */
 323         if (object == NULL || object->size != npages) {
 324                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 325                 buffer = (caddr_t)vm_map_min(&kernel_map);
 326
 327                 error = vm_map_find(&kernel_map, object, 0,
 328                                     (vm_offset_t *)&buffer, size,
 329                                     1,
 330                                     VM_MAPTYPE_NORMAL,
 331                                     VM_PROT_ALL, VM_PROT_ALL,
 332                                     0);
 333
 334                 if (error != KERN_SUCCESS) {
 335                         vm_object_deallocate(object);
 336                         return (ENOMEM);
 337                 }
 338                 pipe_free_kmem(cpipe);
 339                 cpipe->pipe_buffer.object = object;
 340                 cpipe->pipe_buffer.buffer = buffer;
 341                 cpipe->pipe_buffer.size = size;
 342                 ++pipe_bkmem_alloc;
 343         } else {
 344                 ++pipe_bcache_alloc;
 345         }
 346         cpipe->pipe_buffer.in = 0;
 347         cpipe->pipe_buffer.out = 0;
 348         cpipe->pipe_buffer.cnt = 0;
 349         return (0);
 350 }
 351
 352 /*
 353  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 354  * our per-cpu cache if possible.  For now make sure it is sized for the
 355  * smaller PIPE_SIZE default.
 356  */
 357 static int
 358 pipe_create(struct pipe **cpipep)
 359 {
 360         globaldata_t gd = mycpu;
 361         struct pipe *cpipe;
 362         int error;
 363
 364         if ((cpipe = gd->gd_pipeq) != NULL) {
 365                 gd->gd_pipeq = cpipe->pipe_peer;
 366                 --gd->gd_pipeqcount;
 367                 cpipe->pipe_peer = NULL;
 368         } else {
 369                 cpipe = kmalloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 370         }
 371         *cpipep = cpipe;
 372         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 373                 return (error);
 374         vfs_timestamp(&cpipe->pipe_ctime);
 375         cpipe->pipe_atime = cpipe->pipe_ctime;
 376         cpipe->pipe_mtime = cpipe->pipe_ctime;
 377         return (0);
 378 }
 379
 380
 381 /*
 382  * lock a pipe for I/O, blocking other access
 383  */
 384 static __inline int
 385 pipelock(struct pipe *cpipe, int catch)
 386 {
 387         int error;
 388
 389         while (cpipe->pipe_state & PIPE_LOCK) {
 390                 cpipe->pipe_state |= PIPE_LWANT;
 391                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 392                 if (error != 0)
 393                         return (error);
 394         }
 395         cpipe->pipe_state |= PIPE_LOCK;
 396         return (0);
 397 }
 398
 399 /*
 400  * unlock a pipe I/O lock
 401  */
 402 static __inline void
 403 pipeunlock(struct pipe *cpipe)
 404 {
 405
 406         cpipe->pipe_state &= ~PIPE_LOCK;
 407         if (cpipe->pipe_state & PIPE_LWANT) {
 408                 cpipe->pipe_state &= ~PIPE_LWANT;
 409                 wakeup(cpipe);
 410         }
 411 }
 412
 413 static __inline void
 414 pipeselwakeup(struct pipe *cpipe)
 415 {
 416
 417         if (cpipe->pipe_state & PIPE_SEL) {
 418                 cpipe->pipe_state &= ~PIPE_SEL;
 419                 selwakeup(&cpipe->pipe_sel);
 420         }
 421         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 422                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 423         KNOTE(&cpipe->pipe_sel.si_note, 0);
 424 }
 425
 426 /*
 427  * MPALMOSTSAFE (acquires mplock)
 428  */
 429 static int
 430 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 431 {
 432         struct pipe *rpipe;
 433         int error;
 434         int nread = 0;
 435         int nbio;
 436         u_int size;
 437
 438         get_mplock();
 439         rpipe = (struct pipe *) fp->f_data;
 440         ++rpipe->pipe_busy;
 441         error = pipelock(rpipe, 1);
 442         if (error)
 443                 goto unlocked_error;
 444
 445         if (fflags & O_FBLOCKING)
 446                 nbio = 0;
 447         else if (fflags & O_FNONBLOCKING)
 448                 nbio = 1;
 449         else if (fp->f_flag & O_NONBLOCK)
 450                 nbio = 1;
 451         else
 452                 nbio = 0;
 453
 454         while (uio->uio_resid) {
 455                 caddr_t va;
 456
 457                 if (rpipe->pipe_buffer.cnt > 0) {
 458                         /*
 459                          * normal pipe buffer receive
 460                          */
 461                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 462                         if (size > rpipe->pipe_buffer.cnt)
 463                                 size = rpipe->pipe_buffer.cnt;
 464                         if (size > (u_int) uio->uio_resid)
 465                                 size = (u_int) uio->uio_resid;
 466
 467                         error = uiomove(&rpipe->pipe_buffer.buffer
 468                                           [rpipe->pipe_buffer.out],
 469                                         size, uio);
 470                         if (error)
 471                                 break;
 472
 473                         rpipe->pipe_buffer.out += size;
 474                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 475                                 rpipe->pipe_buffer.out = 0;
 476
 477                         rpipe->pipe_buffer.cnt -= size;
 478
 479                         /*
 480                          * If there is no more to read in the pipe, reset
 481                          * its pointers to the beginning.  This improves
 482                          * cache hit stats.
 483                          */
 484                         if (rpipe->pipe_buffer.cnt == 0) {
 485                                 rpipe->pipe_buffer.in = 0;
 486                                 rpipe->pipe_buffer.out = 0;
 487                         }
 488                         nread += size;
 489 #ifndef PIPE_NODIRECT
 490                 } else if (rpipe->pipe_kva &&
 491                            rpipe->pipe_feature == PIPE_KMEM &&
 492                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 493                                == PIPE_DIRECTW
 494                 ) {
 495                         /*
 496                          * Direct copy using source-side kva mapping
 497                          */
 498                         size = rpipe->pipe_map.xio_bytes -
 499                                 rpipe->pipe_buffer.out;
 500                         if (size > (u_int)uio->uio_resid)
 501                                 size = (u_int)uio->uio_resid;
 502                         va = (caddr_t)rpipe->pipe_kva +
 503                                 xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 504                         error = uiomove(va, size, uio);
 505                         if (error)
 506                                 break;
 507                         nread += size;
 508                         rpipe->pipe_buffer.out += size;
 509                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 510                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 511                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 512                                 /* reset out index for copy mode */
 513                                 rpipe->pipe_buffer.out = 0;
 514                                 wakeup(rpipe);
 515                         }
 516                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 517                            rpipe->pipe_kva &&
 518                            rpipe->pipe_feature == PIPE_SFBUF2 &&
 519                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 520                                == PIPE_DIRECTW
 521                 ) {
 522                         /*
 523                          * Direct copy, bypassing a kernel buffer.  We cannot
 524                          * mess with the direct-write buffer until
 525                          * PIPE_DIRECTIP is cleared.  In order to prevent
 526                          * the pipe_write code from racing itself in
 527                          * direct_write, we set DIRECTIP when we clear
 528                          * DIRECTW after we have exhausted the buffer.
 529                          */
 530                         if (pipe_dwrite_sfbuf == 3)
 531                                 rpipe->pipe_kvamask = 0;
 532                         pmap_qenter2(rpipe->pipe_kva, rpipe->pipe_map.xio_pages,
 533                                     rpipe->pipe_map.xio_npages,
 534                                     &rpipe->pipe_kvamask);
 535                         size = rpipe->pipe_map.xio_bytes -
 536                                 rpipe->pipe_buffer.out;
 537                         if (size > (u_int)uio->uio_resid)
 538                                 size = (u_int)uio->uio_resid;
 539                         va = (caddr_t)rpipe->pipe_kva + xio_kvaoffset(&rpipe->pipe_map, rpipe->pipe_buffer.out);
 540                         error = uiomove(va, size, uio);
 541                         if (error)
 542                                 break;
 543                         nread += size;
 544                         rpipe->pipe_buffer.out += size;
 545                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 546                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 547                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 548                                 /* reset out index for copy mode */
 549                                 rpipe->pipe_buffer.out = 0;
 550                                 wakeup(rpipe);
 551                         }
 552                 } else if (rpipe->pipe_buffer.out != rpipe->pipe_map.xio_bytes &&
 553                            rpipe->pipe_feature == PIPE_SFBUF1 &&
 554                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP))
 555                                 == PIPE_DIRECTW
 556                 ) {
 557                         /*
 558                          * Direct copy, bypassing a kernel buffer.  We cannot
 559                          * mess with the direct-write buffer until
 560                          * PIPE_DIRECTIP is cleared.  In order to prevent
 561                          * the pipe_write code from racing itself in
 562                          * direct_write, we set DIRECTIP when we clear
 563                          * DIRECTW after we have exhausted the buffer.
 564                          */
 565                         error = xio_uio_copy(&rpipe->pipe_map, rpipe->pipe_buffer.out, uio, &size);
 566                         if (error)
 567                                 break;
 568                         nread += size;
 569                         rpipe->pipe_buffer.out += size;
 570                         if (rpipe->pipe_buffer.out == rpipe->pipe_map.xio_bytes) {
 571                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 572                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 573                                 /* reset out index for copy mode */
 574                                 rpipe->pipe_buffer.out = 0;
 575                                 wakeup(rpipe);
 576                         }
 577 #endif
 578                 } else {
 579                         /*
 580                          * detect EOF condition
 581                          * read returns 0 on EOF, no need to set error
 582                          */
 583                         if (rpipe->pipe_state & PIPE_EOF)
 584                                 break;
 585
 586                         /*
 587                          * If the "write-side" has been blocked, wake it up now.
 588                          */
 589                         if (rpipe->pipe_state & PIPE_WANTW) {
 590                                 rpipe->pipe_state &= ~PIPE_WANTW;
 591                                 wakeup(rpipe);
 592                         }
 593
 594                         /*
 595                          * Break if some data was read.
 596                          */
 597                         if (nread > 0)
 598                                 break;
 599
 600                         /*
 601                          * Unlock the pipe buffer for our remaining
 602                          * processing.  We will either break out with an
 603                          * error or we will sleep and relock to loop.
 604                          */
 605                         pipeunlock(rpipe);
 606
 607                         /*
 608                          * Handle non-blocking mode operation or
 609                          * wait for more data.
 610                          */
 611                         if (nbio) {
 612                                 error = EAGAIN;
 613                         } else {
 614                                 rpipe->pipe_state |= PIPE_WANTR;
 615                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 616                                     "piperd", 0)) == 0) {
 617                                         error = pipelock(rpipe, 1);
 618                                 }
 619                         }
 620                         if (error)
 621                                 goto unlocked_error;
 622                 }
 623         }
 624         pipeunlock(rpipe);
 625
 626         if (error == 0)
 627                 vfs_timestamp(&rpipe->pipe_atime);
 628 unlocked_error:
 629         --rpipe->pipe_busy;
 630
 631         /*
 632          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 633          */
 634         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 635                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 636                 wakeup(rpipe);
 637         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 638                 /*
 639                  * Handle write blocking hysteresis.
 640                  */
 641                 if (rpipe->pipe_state & PIPE_WANTW) {
 642                         rpipe->pipe_state &= ~PIPE_WANTW;
 643                         wakeup(rpipe);
 644                 }
 645         }
 646
 647         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 648                 pipeselwakeup(rpipe);
 649         rel_mplock();
 650         return (error);
 651 }
 652
 653 #ifndef PIPE_NODIRECT
 654 /*
 655  * Map the sending processes' buffer into kernel space and wire it.
 656  * This is similar to a physical write operation.
 657  */
 658 static int
 659 pipe_build_write_buffer(struct pipe *wpipe, struct uio *uio)
 660 {
 661         int error;
 662         u_int size;
 663
 664         size = (u_int) uio->uio_iov->iov_len;
 665         if (size > wpipe->pipe_buffer.size)
 666                 size = wpipe->pipe_buffer.size;
 667
 668         if (uio->uio_segflg == UIO_SYSSPACE) {
 669                 error = xio_init_kbuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 670                                         size);
 671         } else {
 672                 error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 673                                         size, XIOF_READ);
 674         }
 675         wpipe->pipe_buffer.out = 0;
 676         if (error)
 677                 return(error);
 678
 679         /*
 680          * Create a kernel map for KMEM and SFBUF2 copy modes.  SFBUF2 will
 681          * map the pages on the target while KMEM maps the pages now.
 682          */
 683         switch(wpipe->pipe_feature) {
 684         case PIPE_KMEM:
 685         case PIPE_SFBUF2:
 686                 if (wpipe->pipe_kva == 0) {
 687                         wpipe->pipe_kva =
 688                             kmem_alloc_nofault(&kernel_map, XIO_INTERNAL_SIZE);
 689                         wpipe->pipe_kvamask = 0;
 690                 }
 691                 if (wpipe->pipe_feature == PIPE_KMEM) {
 692                         pmap_qenter(wpipe->pipe_kva, wpipe->pipe_map.xio_pages,
 693                                     wpipe->pipe_map.xio_npages);
 694                 }
 695                 break;
 696         default:
 697                 break;
 698         }
 699
 700         /*
 701          * And update the uio data.  The XIO might have loaded fewer bytes
 702          * then requested so reload 'size'.
 703          */
 704         size = wpipe->pipe_map.xio_bytes;
 705         uio->uio_iov->iov_len -= size;
 706         uio->uio_iov->iov_base += size;
 707         if (uio->uio_iov->iov_len == 0)
 708                 uio->uio_iov++;
 709         uio->uio_resid -= size;
 710         uio->uio_offset += size;
 711         return (0);
 712 }
 713
 714 /*
 715  * In the case of a signal, the writing process might go away.  This
 716  * code copies the data into the circular buffer so that the source
 717  * pages can be freed without loss of data.
 718  *
 719  * Note that in direct mode pipe_buffer.out is used to track the
 720  * XIO offset.  We are converting the direct mode into buffered mode
 721  * which changes the meaning of pipe_buffer.out.
 722  */
 723 static void
 724 pipe_clone_write_buffer(struct pipe *wpipe)
 725 {
 726         int size;
 727         int offset;
 728
 729         offset = wpipe->pipe_buffer.out;
 730         size = wpipe->pipe_map.xio_bytes - offset;
 731
 732         KKASSERT(size <= wpipe->pipe_buffer.size);
 733
 734         wpipe->pipe_buffer.in = size;
 735         wpipe->pipe_buffer.out = 0;
 736         wpipe->pipe_buffer.cnt = size;
 737         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 738
 739         xio_copy_xtok(&wpipe->pipe_map, offset, wpipe->pipe_buffer.buffer, size);
 740         xio_release(&wpipe->pipe_map);
 741         if (wpipe->pipe_kva) {
 742                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 743                 kmem_free(&kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 744                 wpipe->pipe_kva = 0;
 745         }
 746 }
 747
 748 /*
 749  * This implements the pipe buffer write mechanism.  Note that only
 750  * a direct write OR a normal pipe write can be pending at any given time.
 751  * If there are any characters in the pipe buffer, the direct write will
 752  * be deferred until the receiving process grabs all of the bytes from
 753  * the pipe buffer.  Then the direct mapping write is set-up.
 754  */
 755 static int
 756 pipe_direct_write(struct pipe *wpipe, struct uio *uio)
 757 {
 758         int error;
 759
 760 retry:
 761         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 762                 if (wpipe->pipe_state & PIPE_WANTR) {
 763                         wpipe->pipe_state &= ~PIPE_WANTR;
 764                         wakeup(wpipe);
 765                 }
 766                 wpipe->pipe_state |= PIPE_WANTW;
 767                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 768                 if (error)
 769                         goto error2;
 770                 if (wpipe->pipe_state & PIPE_EOF) {
 771                         error = EPIPE;
 772                         goto error2;
 773                 }
 774         }
 775         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 776         if (wpipe->pipe_buffer.cnt > 0) {
 777                 if (wpipe->pipe_state & PIPE_WANTR) {
 778                         wpipe->pipe_state &= ~PIPE_WANTR;
 779                         wakeup(wpipe);
 780                 }
 781
 782                 wpipe->pipe_state |= PIPE_WANTW;
 783                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 784                 if (error)
 785                         goto error2;
 786                 if (wpipe->pipe_state & PIPE_EOF) {
 787                         error = EPIPE;
 788                         goto error2;
 789                 }
 790                 goto retry;
 791         }
 792
 793         /*
 794          * Build our direct-write buffer
 795          */
 796         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 797         error = pipe_build_write_buffer(wpipe, uio);
 798         if (error)
 799                 goto error1;
 800         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 801
 802         /*
 803          * Wait until the receiver has snarfed the data.  Since we are likely
 804          * going to sleep we optimize the case and yield synchronously,
 805          * possibly avoiding the tsleep().
 806          */
 807         error = 0;
 808         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 809                 if (wpipe->pipe_state & PIPE_EOF) {
 810                         pipelock(wpipe, 0);
 811                         xio_release(&wpipe->pipe_map);
 812                         if (wpipe->pipe_kva) {
 813                                 pmap_qremove(wpipe->pipe_kva, XIO_INTERNAL_PAGES);
 814                                 kmem_free(&kernel_map, wpipe->pipe_kva, XIO_INTERNAL_SIZE);
 815                                 wpipe->pipe_kva = 0;
 816                         }
 817                         pipeunlock(wpipe);
 818                         pipeselwakeup(wpipe);
 819                         error = EPIPE;
 820                         goto error1;
 821                 }
 822                 if (wpipe->pipe_state & PIPE_WANTR) {
 823                         wpipe->pipe_state &= ~PIPE_WANTR;
 824                         wakeup(wpipe);
 825                 }
 826                 pipeselwakeup(wpipe);
 827                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 828         }
 829         pipelock(wpipe,0);
 830         if (wpipe->pipe_state & PIPE_DIRECTW) {
 831                 /*
 832                  * this bit of trickery substitutes a kernel buffer for
 833                  * the process that might be going away.
 834                  */
 835                 pipe_clone_write_buffer(wpipe);
 836                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 837         } else {
 838                 /*
 839                  * note: The pipe_kva mapping is not qremove'd here.  For
 840                  * legacy PIPE_KMEM mode this constitutes an improvement
 841                  * over the original FreeBSD-4 algorithm.  For PIPE_SFBUF2
 842                  * mode the kva mapping must not be removed to get the
 843                  * caching benefit.
 844                  *
 845                  * For testing purposes we will give the original algorithm
 846                  * the benefit of the doubt 'what it could have been', and
 847                  * keep the optimization.
 848                  */
 849                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 850                 xio_release(&wpipe->pipe_map);
 851                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 852         }
 853         pipeunlock(wpipe);
 854         return (error);
 855
 856         /*
 857          * Direct-write error, clear the direct write flags.
 858          */
 859 error1:
 860         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 861         /* fallthrough */
 862
 863         /*
 864          * General error, wakeup the other side if it happens to be sleeping.
 865          */
 866 error2:
 867         wakeup(wpipe);
 868         return (error);
 869 }
 870 #endif
 871
 872 /*
 873  * MPALMOSTSAFE - acquires mplock
 874  */
 875 static int
 876 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred, int fflags)
 877 {
 878         int error = 0;
 879         int orig_resid;
 880         int nbio;
 881         struct pipe *wpipe, *rpipe;
 882
 883         get_mplock();
 884         rpipe = (struct pipe *) fp->f_data;
 885         wpipe = rpipe->pipe_peer;
 886
 887         /*
 888          * detect loss of pipe read side, issue SIGPIPE if lost.
 889          */
 890         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 891                 rel_mplock();
 892                 return (EPIPE);
 893         }
 894         ++wpipe->pipe_busy;
 895
 896         if (fflags & O_FBLOCKING)
 897                 nbio = 0;
 898         else if (fflags & O_FNONBLOCKING)
 899                 nbio = 1;
 900         else if (fp->f_flag & O_NONBLOCK)
 901                 nbio = 1;
 902         else
 903                 nbio = 0;
 904
 905         /*
 906          * If it is advantageous to resize the pipe buffer, do
 907          * so.
 908          */
 909         if ((uio->uio_resid > PIPE_SIZE) &&
 910             (pipe_nbig < pipe_maxbig) &&
 911             (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 912             (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 913             (wpipe->pipe_buffer.cnt == 0) &&
 914             (error = pipelock(wpipe, 1)) == 0) {
 915                 /*
 916                  * Recheck after lock.
 917                  */
 918                 if ((pipe_nbig < pipe_maxbig) &&
 919                     (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 920                     (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 921                     (wpipe->pipe_buffer.cnt == 0)) {
 922                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0) {
 923                                 ++pipe_bigcount;
 924                                 pipe_nbig++;
 925                         }
 926                 }
 927                 pipeunlock(wpipe);
 928         }
 929
 930         /*
 931          * If an early error occured unbusy and return, waking up any pending
 932          * readers.
 933          */
 934         if (error) {
 935                 --wpipe->pipe_busy;
 936                 if ((wpipe->pipe_busy == 0) &&
 937                     (wpipe->pipe_state & PIPE_WANT)) {
 938                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 939                         wakeup(wpipe);
 940                 }
 941                 rel_mplock();
 942                 return(error);
 943         }
 944
 945         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 946
 947         orig_resid = uio->uio_resid;
 948
 949         while (uio->uio_resid) {
 950                 int space;
 951
 952 #ifndef PIPE_NODIRECT
 953                 /*
 954                  * If direct process-to-process writes are enabled and
 955                  * the buffer is large enough, and this is blocking IO,
 956                  * then use the direct write feature.
 957                  *
 958                  * If the pipe was opened at a time when the direct write
 959                  * feature was not enabled pipe_feature will be set to
 960                  * PIPE_COPY and we do not use the feature.
 961                  *
 962                  * The direct write mechanism will detect the reader going
 963                  * away on us.
 964                  */
 965                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT ||
 966                      pipe_dwrite_enable > 1) &&
 967                     nbio == 0 && pipe_dwrite_enable &&
 968                     wpipe->pipe_feature != PIPE_COPY
 969                 ) {
 970                         error = pipe_direct_write( wpipe, uio);
 971                         if (error)
 972                                 break;
 973                         continue;
 974                 }
 975 #endif
 976
 977                 /*
 978                  * Pipe buffered writes cannot be coincidental with
 979                  * direct writes.  We wait until the currently executing
 980                  * direct write is completed before we start filling the
 981                  * pipe buffer.  We break out if a signal occurs or the
 982                  * reader goes away.
 983                  */
 984         retrywrite:
 985                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 986                         if (wpipe->pipe_state & PIPE_WANTR) {
 987                                 wpipe->pipe_state &= ~PIPE_WANTR;
 988                                 wakeup(wpipe);
 989                         }
 990                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 991                         if (wpipe->pipe_state & PIPE_EOF)
 992                                 break;
 993                         if (error)
 994                                 break;
 995                 }
 996                 if (wpipe->pipe_state & PIPE_EOF) {
 997                         error = EPIPE;
 998                         break;
 999                 }
1000
1001                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1002
1003                 /* Writes of size <= PIPE_BUF must be atomic. */
1004                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
1005                         space = 0;
1006
1007                 /*
1008                  * Write to fill, read size handles write hysteresis.  Also
1009                  * additional restrictions can cause select-based non-blocking
1010                  * writes to spin.
1011                  */
1012                 if (space > 0) {
1013                         if ((error = pipelock(wpipe,1)) == 0) {
1014                                 int size;       /* Transfer size */
1015                                 int segsize;    /* first segment to transfer */
1016
1017                                 /*
1018                                  * It is possible for a direct write to
1019                                  * slip in on us... handle it here...
1020                                  */
1021                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
1022                                         pipeunlock(wpipe);
1023                                         goto retrywrite;
1024                                 }
1025                                 /*
1026                                  * If a process blocked in uiomove, our
1027                                  * value for space might be bad.
1028                                  *
1029                                  * XXX will we be ok if the reader has gone
1030                                  * away here?
1031                                  */
1032                                 if (space > wpipe->pipe_buffer.size -
1033                                     wpipe->pipe_buffer.cnt) {
1034                                         pipeunlock(wpipe);
1035                                         goto retrywrite;
1036                                 }
1037
1038                                 /*
1039                                  * Transfer size is minimum of uio transfer
1040                                  * and free space in pipe buffer.
1041                                  */
1042                                 if (space > uio->uio_resid)
1043                                         size = uio->uio_resid;
1044                                 else
1045                                         size = space;
1046                                 /*
1047                                  * First segment to transfer is minimum of
1048                                  * transfer size and contiguous space in
1049                                  * pipe buffer.  If first segment to transfer
1050                                  * is less than the transfer size, we've got
1051                                  * a wraparound in the buffer.
1052                                  */
1053                                 segsize = wpipe->pipe_buffer.size -
1054                                         wpipe->pipe_buffer.in;
1055                                 if (segsize > size)
1056                                         segsize = size;
1057
1058                                 /* Transfer first segment */
1059
1060                                 error = uiomove(&wpipe->pipe_buffer.buffer
1061                                                   [wpipe->pipe_buffer.in],
1062                                                 segsize, uio);
1063
1064                                 if (error == 0 && segsize < size) {
1065                                         /*
1066                                          * Transfer remaining part now, to
1067                                          * support atomic writes.  Wraparound
1068                                          * happened.
1069                                          */
1070                                         if (wpipe->pipe_buffer.in + segsize !=
1071                                             wpipe->pipe_buffer.size)
1072                                                 panic("Expected pipe buffer wraparound disappeared");
1073
1074                                         error = uiomove(&wpipe->pipe_buffer.
1075                                                           buffer[0],
1076                                                         size - segsize, uio);
1077                                 }
1078                                 if (error == 0) {
1079                                         wpipe->pipe_buffer.in += size;
1080                                         if (wpipe->pipe_buffer.in >=
1081                                             wpipe->pipe_buffer.size) {
1082                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
1083                                                         panic("Expected wraparound bad");
1084                                                 wpipe->pipe_buffer.in = size - segsize;
1085                                         }
1086
1087                                         wpipe->pipe_buffer.cnt += size;
1088                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
1089                                                 panic("Pipe buffer overflow");
1090
1091                                 }
1092                                 pipeunlock(wpipe);
1093                         }
1094                         if (error)
1095                                 break;
1096
1097                 } else {
1098                         /*
1099                          * If the "read-side" has been blocked, wake it up now
1100                          * and yield to let it drain synchronously rather
1101                          * then block.
1102                          */
1103                         if (wpipe->pipe_state & PIPE_WANTR) {
1104                                 wpipe->pipe_state &= ~PIPE_WANTR;
1105                                 wakeup(wpipe);
1106                         }
1107
1108                         /*
1109                          * don't block on non-blocking I/O
1110                          */
1111                         if (nbio) {
1112                                 error = EAGAIN;
1113                                 break;
1114                         }
1115
1116                         /*
1117                          * We have no more space and have something to offer,
1118                          * wake up select/poll.
1119                          */
1120                         pipeselwakeup(wpipe);
1121
1122                         wpipe->pipe_state |= PIPE_WANTW;
1123                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
1124                         if (error != 0)
1125                                 break;
1126                         /*
1127                          * If read side wants to go away, we just issue a signal
1128                          * to ourselves.
1129                          */
1130                         if (wpipe->pipe_state & PIPE_EOF) {
1131                                 error = EPIPE;
1132                                 break;
1133                         }
1134                 }
1135         }
1136
1137         --wpipe->pipe_busy;
1138
1139         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
1140                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
1141                 wakeup(wpipe);
1142         } else if (wpipe->pipe_buffer.cnt > 0) {
1143                 /*
1144                  * If we have put any characters in the buffer, we wake up
1145                  * the reader.
1146                  */
1147                 if (wpipe->pipe_state & PIPE_WANTR) {
1148                         wpipe->pipe_state &= ~PIPE_WANTR;
1149                         wakeup(wpipe);
1150                 }
1151         }
1152
1153         /*
1154          * Don't return EPIPE if I/O was successful
1155          */
1156         if ((wpipe->pipe_buffer.cnt == 0) &&
1157             (uio->uio_resid == 0) &&
1158             (error == EPIPE)) {
1159                 error = 0;
1160         }
1161
1162         if (error == 0)
1163                 vfs_timestamp(&wpipe->pipe_mtime);
1164
1165         /*
1166          * We have something to offer,
1167          * wake up select/poll.
1168          */
1169         if (wpipe->pipe_buffer.cnt)
1170                 pipeselwakeup(wpipe);
1171         rel_mplock();
1172         return (error);
1173 }
1174
1175 /*
1176  * MPALMOSTSAFE - acquires mplock
1177  *
1178  * we implement a very minimal set of ioctls for compatibility with sockets.
1179  */
1180 int
1181 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct ucred *cred)
1182 {
1183         struct pipe *mpipe;
1184         int error;
1185
1186         get_mplock();
1187         mpipe = (struct pipe *)fp->f_data;
1188
1189         switch (cmd) {
1190         case FIOASYNC:
1191                 if (*(int *)data) {
1192                         mpipe->pipe_state |= PIPE_ASYNC;
1193                 } else {
1194                         mpipe->pipe_state &= ~PIPE_ASYNC;
1195                 }
1196                 error = 0;
1197                 break;
1198         case FIONREAD:
1199                 if (mpipe->pipe_state & PIPE_DIRECTW) {
1200                         *(int *)data = mpipe->pipe_map.xio_bytes -
1201                                         mpipe->pipe_buffer.out;
1202                 } else {
1203                         *(int *)data = mpipe->pipe_buffer.cnt;
1204                 }
1205                 error = 0;
1206                 break;
1207         case FIOSETOWN:
1208                 error = fsetown(*(int *)data, &mpipe->pipe_sigio);
1209                 break;
1210         case FIOGETOWN:
1211                 *(int *)data = fgetown(mpipe->pipe_sigio);
1212                 error = 0;
1213                 break;
1214         case TIOCSPGRP:
1215                 /* This is deprecated, FIOSETOWN should be used instead. */
1216                 error = fsetown(-(*(int *)data), &mpipe->pipe_sigio);
1217                 break;
1218
1219         case TIOCGPGRP:
1220                 /* This is deprecated, FIOGETOWN should be used instead. */
1221                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1222                 error = 0;
1223                 break;
1224         default:
1225                 error = ENOTTY;
1226                 break;
1227         }
1228         rel_mplock();
1229         return (error);
1230 }
1231
1232 /*
1233  * MPALMOSTSAFE - acquires mplock
1234  */
1235 int
1236 pipe_poll(struct file *fp, int events, struct ucred *cred)
1237 {
1238         struct pipe *rpipe;
1239         struct pipe *wpipe;
1240         int revents = 0;
1241
1242         get_mplock();
1243         rpipe = (struct pipe *)fp->f_data;
1244         wpipe = rpipe->pipe_peer;
1245         if (events & (POLLIN | POLLRDNORM))
1246                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1247                     (rpipe->pipe_buffer.cnt > 0) ||
1248                     (rpipe->pipe_state & PIPE_EOF))
1249                         revents |= events & (POLLIN | POLLRDNORM);
1250
1251         if (events & (POLLOUT | POLLWRNORM))
1252                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1253                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1254                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1255                         revents |= events & (POLLOUT | POLLWRNORM);
1256
1257         if ((rpipe->pipe_state & PIPE_EOF) ||
1258             (wpipe == NULL) ||
1259             (wpipe->pipe_state & PIPE_EOF))
1260                 revents |= POLLHUP;
1261
1262         if (revents == 0) {
1263                 if (events & (POLLIN | POLLRDNORM)) {
1264                         selrecord(curthread, &rpipe->pipe_sel);
1265                         rpipe->pipe_state |= PIPE_SEL;
1266                 }
1267
1268                 if (events & (POLLOUT | POLLWRNORM)) {
1269                         selrecord(curthread, &wpipe->pipe_sel);
1270                         wpipe->pipe_state |= PIPE_SEL;
1271                 }
1272         }
1273         rel_mplock();
1274         return (revents);
1275 }
1276
1277 /*
1278  * MPALMOSTSAFE - acquires mplock
1279  */
1280 static int
1281 pipe_stat(struct file *fp, struct stat *ub, struct ucred *cred)
1282 {
1283         struct pipe *pipe;
1284
1285         get_mplock();
1286         pipe = (struct pipe *)fp->f_data;
1287
1288         bzero((caddr_t)ub, sizeof(*ub));
1289         ub->st_mode = S_IFIFO;
1290         ub->st_blksize = pipe->pipe_buffer.size;
1291         ub->st_size = pipe->pipe_buffer.cnt;
1292         if (ub->st_size == 0 && (pipe->pipe_state & PIPE_DIRECTW)) {
1293                 ub->st_size = pipe->pipe_map.xio_bytes -
1294                                 pipe->pipe_buffer.out;
1295         }
1296         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1297         ub->st_atimespec = pipe->pipe_atime;
1298         ub->st_mtimespec = pipe->pipe_mtime;
1299         ub->st_ctimespec = pipe->pipe_ctime;
1300         /*
1301          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1302          * st_flags, st_gen.
1303          * XXX (st_dev, st_ino) should be unique.
1304          */
1305         rel_mplock();
1306         return (0);
1307 }
1308
1309 /*
1310  * MPALMOSTSAFE - acquires mplock
1311  */
1312 static int
1313 pipe_close(struct file *fp)
1314 {
1315         struct pipe *cpipe;
1316
1317         get_mplock();
1318         cpipe = (struct pipe *)fp->f_data;
1319         fp->f_ops = &badfileops;
1320         fp->f_data = NULL;
1321         funsetown(cpipe->pipe_sigio);
1322         pipeclose(cpipe);
1323         rel_mplock();
1324         return (0);
1325 }
1326
1327 /*
1328  * Shutdown one or both directions of a full-duplex pipe.
1329  *
1330  * MPALMOSTSAFE - acquires mplock
1331  */
1332 static int
1333 pipe_shutdown(struct file *fp, int how)
1334 {
1335         struct pipe *rpipe;
1336         struct pipe *wpipe;
1337         int error = EPIPE;
1338
1339         get_mplock();
1340         rpipe = (struct pipe *)fp->f_data;
1341
1342         switch(how) {
1343         case SHUT_RDWR:
1344         case SHUT_RD:
1345                 if (rpipe) {
1346                         rpipe->pipe_state |= PIPE_EOF;
1347                         pipeselwakeup(rpipe);
1348                         if (rpipe->pipe_busy)
1349                                 wakeup(rpipe);
1350                         error = 0;
1351                 }
1352                 if (how == SHUT_RD)
1353                         break;
1354                 /* fall through */
1355         case SHUT_WR:
1356                 if (rpipe && (wpipe = rpipe->pipe_peer) != NULL) {
1357                         wpipe->pipe_state |= PIPE_EOF;
1358                         pipeselwakeup(wpipe);
1359                         if (wpipe->pipe_busy)
1360                                 wakeup(wpipe);
1361                         error = 0;
1362                 }
1363         }
1364         rel_mplock();
1365         return (error);
1366 }
1367
1368 static void
1369 pipe_free_kmem(struct pipe *cpipe)
1370 {
1371         if (cpipe->pipe_buffer.buffer != NULL) {
1372                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1373                         --pipe_nbig;
1374                 kmem_free(&kernel_map,
1375                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1376                         cpipe->pipe_buffer.size);
1377                 cpipe->pipe_buffer.buffer = NULL;
1378                 cpipe->pipe_buffer.object = NULL;
1379         }
1380 #ifndef PIPE_NODIRECT
1381         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1382                 cpipe->pipe_map.xio_offset == 0 &&
1383                 cpipe->pipe_map.xio_npages == 0);
1384 #endif
1385 }
1386
1387 /*
1388  * shutdown the pipe
1389  */
1390 static void
1391 pipeclose(struct pipe *cpipe)
1392 {
1393         globaldata_t gd;
1394         struct pipe *ppipe;
1395
1396         if (cpipe == NULL)
1397                 return;
1398
1399         pipeselwakeup(cpipe);
1400
1401         /*
1402          * If the other side is blocked, wake it up saying that
1403          * we want to close it down.
1404          */
1405         while (cpipe->pipe_busy) {
1406                 wakeup(cpipe);
1407                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1408                 tsleep(cpipe, 0, "pipecl", 0);
1409         }
1410
1411         /*
1412          * Disconnect from peer
1413          */
1414         if ((ppipe = cpipe->pipe_peer) != NULL) {
1415                 pipeselwakeup(ppipe);
1416
1417                 ppipe->pipe_state |= PIPE_EOF;
1418                 wakeup(ppipe);
1419                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1420                 ppipe->pipe_peer = NULL;
1421         }
1422
1423         if (cpipe->pipe_kva) {
1424                 pmap_qremove(cpipe->pipe_kva, XIO_INTERNAL_PAGES);
1425                 kmem_free(&kernel_map, cpipe->pipe_kva, XIO_INTERNAL_SIZE);
1426                 cpipe->pipe_kva = 0;
1427         }
1428
1429         /*
1430          * free or cache resources
1431          */
1432         gd = mycpu;
1433         if (gd->gd_pipeqcount >= pipe_maxcache ||
1434             cpipe->pipe_buffer.size != PIPE_SIZE
1435         ) {
1436                 pipe_free_kmem(cpipe);
1437                 kfree(cpipe, M_PIPE);
1438         } else {
1439                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1440                         cpipe->pipe_map.xio_bytes == 0 &&
1441                         cpipe->pipe_map.xio_offset == 0);
1442                 cpipe->pipe_state = 0;
1443                 cpipe->pipe_busy = 0;
1444                 cpipe->pipe_peer = gd->gd_pipeq;
1445                 gd->gd_pipeq = cpipe;
1446                 ++gd->gd_pipeqcount;
1447         }
1448 }
1449
1450 /*
1451  * MPALMOSTSAFE - acquires mplock
1452  */
1453 static int
1454 pipe_kqfilter(struct file *fp, struct knote *kn)
1455 {
1456         struct pipe *cpipe;
1457
1458         get_mplock();
1459         cpipe = (struct pipe *)kn->kn_fp->f_data;
1460
1461         switch (kn->kn_filter) {
1462         case EVFILT_READ:
1463                 kn->kn_fop = &pipe_rfiltops;
1464                 break;
1465         case EVFILT_WRITE:
1466                 kn->kn_fop = &pipe_wfiltops;
1467                 cpipe = cpipe->pipe_peer;
1468                 if (cpipe == NULL) {
1469                         /* other end of pipe has been closed */
1470                         rel_mplock();
1471                         return (EPIPE);
1472                 }
1473                 break;
1474         default:
1475                 return (1);
1476         }
1477         kn->kn_hook = (caddr_t)cpipe;
1478
1479         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1480         rel_mplock();
1481         return (0);
1482 }
1483
1484 static void
1485 filt_pipedetach(struct knote *kn)
1486 {
1487         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1488
1489         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1490 }
1491
1492 /*ARGSUSED*/
1493 static int
1494 filt_piperead(struct knote *kn, long hint)
1495 {
1496         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1497         struct pipe *wpipe = rpipe->pipe_peer;
1498
1499         kn->kn_data = rpipe->pipe_buffer.cnt;
1500         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW)) {
1501                 kn->kn_data = rpipe->pipe_map.xio_bytes -
1502                                 rpipe->pipe_buffer.out;
1503         }
1504
1505         if ((rpipe->pipe_state & PIPE_EOF) ||
1506             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1507                 kn->kn_flags |= EV_EOF;
1508                 return (1);
1509         }
1510         return (kn->kn_data > 0);
1511 }
1512
1513 /*ARGSUSED*/
1514 static int
1515 filt_pipewrite(struct knote *kn, long hint)
1516 {
1517         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1518         struct pipe *wpipe = rpipe->pipe_peer;
1519
1520         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1521                 kn->kn_data = 0;
1522                 kn->kn_flags |= EV_EOF;
1523                 return (1);
1524         }
1525         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1526         if (wpipe->pipe_state & PIPE_DIRECTW)
1527                 kn->kn_data = 0;
1528
1529         return (kn->kn_data >= PIPE_BUF);
1530 }