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 1.17 2004/04/01 17:58:02 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
  76 #include <vm/vm.h>
  77 #include <vm/vm_param.h>
  78 #include <sys/lock.h>
  79 #include <vm/vm_object.h>
  80 #include <vm/vm_kern.h>
  81 #include <vm/vm_extern.h>
  82 #include <vm/pmap.h>
  83 #include <vm/vm_map.h>
  84 #include <vm/vm_page.h>
  85 #include <vm/vm_zone.h>
  86
  87 #include <sys/file2.h>
  88
  89 /*
  90  * Use this define if you want to disable *fancy* VM things.  Expect an
  91  * approx 30% decrease in transfer rate.  This could be useful for
  92  * NetBSD or OpenBSD.
  93  */
  94 /* #define PIPE_NODIRECT */
  95
  96 /*
  97  * interfaces to the outside world
  98  */
  99 static int pipe_read (struct file *fp, struct uio *uio,
 100                 struct ucred *cred, int flags, struct thread *td);
 101 static int pipe_write (struct file *fp, struct uio *uio,
 102                 struct ucred *cred, int flags, struct thread *td);
 103 static int pipe_close (struct file *fp, struct thread *td);
 104 static int pipe_poll (struct file *fp, int events, struct ucred *cred,
 105                 struct thread *td);
 106 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 107 static int pipe_stat (struct file *fp, struct stat *sb, struct thread *td);
 108 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct thread *td);
 109
 110 static struct fileops pipeops = {
 111         NULL,   /* port */
 112         0,      /* autoq */
 113         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 114         pipe_stat, pipe_close
 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
 161 SYSCTL_NODE(_kern, OID_AUTO, pipe, CTLFLAG_RW, 0, "Pipe operation");
 162 SYSCTL_INT(_kern_pipe, OID_AUTO, nbig,
 163         CTLFLAG_RD, &pipe_nbig, 0, "numer of big pipes allocated");
 164 SYSCTL_INT(_kern_pipe, OID_AUTO, maxcache,
 165         CTLFLAG_RW, &pipe_maxcache, 0, "max pipes cached per-cpu");
 166 SYSCTL_INT(_kern_pipe, OID_AUTO, maxbig,
 167         CTLFLAG_RW, &pipe_maxbig, 0, "max number of big pipes");
 168 #if !defined(NO_PIPE_SYSCTL_STATS)
 169 SYSCTL_INT(_kern_pipe, OID_AUTO, bcache_alloc,
 170         CTLFLAG_RW, &pipe_bcache_alloc, 0, "pipe buffer from pcpu cache");
 171 SYSCTL_INT(_kern_pipe, OID_AUTO, bkmem_alloc,
 172         CTLFLAG_RW, &pipe_bkmem_alloc, 0, "pipe buffer from kmem");
 173 #endif
 174
 175 static void pipeclose (struct pipe *cpipe);
 176 static void pipe_free_kmem (struct pipe *cpipe);
 177 static int pipe_create (struct pipe **cpipep);
 178 static __inline int pipelock (struct pipe *cpipe, int catch);
 179 static __inline void pipeunlock (struct pipe *cpipe);
 180 static __inline void pipeselwakeup (struct pipe *cpipe);
 181 #ifndef PIPE_NODIRECT
 182 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 183 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 184 static void pipe_clone_write_buffer (struct pipe *wpipe);
 185 #endif
 186 static int pipespace (struct pipe *cpipe, int size);
 187
 188 /*
 189  * The pipe system call for the DTYPE_PIPE type of pipes
 190  *
 191  * pipe_ARgs(int dummy)
 192  */
 193
 194 /* ARGSUSED */
 195 int
 196 pipe(struct pipe_args *uap)
 197 {
 198         struct thread *td = curthread;
 199         struct proc *p = td->td_proc;
 200         struct filedesc *fdp;
 201         struct file *rf, *wf;
 202         struct pipe *rpipe, *wpipe;
 203         int fd1, fd2, error;
 204
 205         KKASSERT(p);
 206         fdp = p->p_fd;
 207
 208         rpipe = wpipe = NULL;
 209         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 210                 pipeclose(rpipe);
 211                 pipeclose(wpipe);
 212                 return (ENFILE);
 213         }
 214
 215         rpipe->pipe_state |= PIPE_DIRECTOK;
 216         wpipe->pipe_state |= PIPE_DIRECTOK;
 217
 218         error = falloc(p, &rf, &fd1);
 219         if (error) {
 220                 pipeclose(rpipe);
 221                 pipeclose(wpipe);
 222                 return (error);
 223         }
 224         fhold(rf);
 225         uap->sysmsg_fds[0] = fd1;
 226
 227         /*
 228          * Warning: once we've gotten past allocation of the fd for the
 229          * read-side, we can only drop the read side via fdrop() in order
 230          * to avoid races against processes which manage to dup() the read
 231          * side while we are blocked trying to allocate the write side.
 232          */
 233         rf->f_flag = FREAD | FWRITE;
 234         rf->f_type = DTYPE_PIPE;
 235         rf->f_data = (caddr_t)rpipe;
 236         rf->f_ops = &pipeops;
 237         error = falloc(p, &wf, &fd2);
 238         if (error) {
 239                 if (fdp->fd_ofiles[fd1] == rf) {
 240                         fdp->fd_ofiles[fd1] = NULL;
 241                         fdrop(rf, td);
 242                 }
 243                 fdrop(rf, td);
 244                 /* rpipe has been closed by fdrop(). */
 245                 pipeclose(wpipe);
 246                 return (error);
 247         }
 248         wf->f_flag = FREAD | FWRITE;
 249         wf->f_type = DTYPE_PIPE;
 250         wf->f_data = (caddr_t)wpipe;
 251         wf->f_ops = &pipeops;
 252         uap->sysmsg_fds[1] = fd2;
 253
 254         rpipe->pipe_peer = wpipe;
 255         wpipe->pipe_peer = rpipe;
 256         fdrop(rf, td);
 257
 258         return (0);
 259 }
 260
 261 /*
 262  * Allocate kva for pipe circular buffer, the space is pageable
 263  * This routine will 'realloc' the size of a pipe safely, if it fails
 264  * it will retain the old buffer.
 265  * If it fails it will return ENOMEM.
 266  */
 267 static int
 268 pipespace(struct pipe *cpipe, int size)
 269 {
 270         struct vm_object *object;
 271         caddr_t buffer;
 272         int npages, error;
 273
 274         npages = round_page(size) / PAGE_SIZE;
 275         object = cpipe->pipe_buffer.object;
 276
 277         /*
 278          * [re]create the object if necessary and reserve space for it
 279          * in the kernel_map.  The object and memory are pageable.  On
 280          * success, free the old resources before assigning the new
 281          * ones.
 282          */
 283         if (object == NULL || object->size != npages) {
 284                 object = vm_object_allocate(OBJT_DEFAULT, npages);
 285                 buffer = (caddr_t) vm_map_min(kernel_map);
 286
 287                 error = vm_map_find(kernel_map, object, 0,
 288                         (vm_offset_t *) &buffer, size, 1,
 289                         VM_PROT_ALL, VM_PROT_ALL, 0);
 290
 291                 if (error != KERN_SUCCESS) {
 292                         vm_object_deallocate(object);
 293                         return (ENOMEM);
 294                 }
 295                 pipe_free_kmem(cpipe);
 296                 cpipe->pipe_buffer.object = object;
 297                 cpipe->pipe_buffer.buffer = buffer;
 298                 cpipe->pipe_buffer.size = size;
 299                 ++pipe_bkmem_alloc;
 300         } else {
 301                 ++pipe_bcache_alloc;
 302         }
 303         cpipe->pipe_buffer.in = 0;
 304         cpipe->pipe_buffer.out = 0;
 305         cpipe->pipe_buffer.cnt = 0;
 306         return (0);
 307 }
 308
 309 /*
 310  * Initialize and allocate VM and memory for pipe, pulling the pipe from
 311  * our per-cpu cache if possible.  For now make sure it is sized for the
 312  * smaller PIPE_SIZE default.
 313  */
 314 static int
 315 pipe_create(cpipep)
 316         struct pipe **cpipep;
 317 {
 318         globaldata_t gd = mycpu;
 319         struct pipe *cpipe;
 320         int error;
 321
 322         if ((cpipe = gd->gd_pipeq) != NULL) {
 323                 gd->gd_pipeq = cpipe->pipe_peer;
 324                 --gd->gd_pipeqcount;
 325                 cpipe->pipe_peer = NULL;
 326         } else {
 327                 cpipe = malloc(sizeof(struct pipe), M_PIPE, M_WAITOK|M_ZERO);
 328         }
 329         *cpipep = cpipe;
 330         if ((error = pipespace(cpipe, PIPE_SIZE)) != 0)
 331                 return (error);
 332         vfs_timestamp(&cpipe->pipe_ctime);
 333         cpipe->pipe_atime = cpipe->pipe_ctime;
 334         cpipe->pipe_mtime = cpipe->pipe_ctime;
 335         return (0);
 336 }
 337
 338
 339 /*
 340  * lock a pipe for I/O, blocking other access
 341  */
 342 static __inline int
 343 pipelock(cpipe, catch)
 344         struct pipe *cpipe;
 345         int catch;
 346 {
 347         int error;
 348
 349         while (cpipe->pipe_state & PIPE_LOCK) {
 350                 cpipe->pipe_state |= PIPE_LWANT;
 351                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 352                 if (error != 0)
 353                         return (error);
 354         }
 355         cpipe->pipe_state |= PIPE_LOCK;
 356         return (0);
 357 }
 358
 359 /*
 360  * unlock a pipe I/O lock
 361  */
 362 static __inline void
 363 pipeunlock(cpipe)
 364         struct pipe *cpipe;
 365 {
 366
 367         cpipe->pipe_state &= ~PIPE_LOCK;
 368         if (cpipe->pipe_state & PIPE_LWANT) {
 369                 cpipe->pipe_state &= ~PIPE_LWANT;
 370                 wakeup(cpipe);
 371         }
 372 }
 373
 374 static __inline void
 375 pipeselwakeup(cpipe)
 376         struct pipe *cpipe;
 377 {
 378
 379         if (cpipe->pipe_state & PIPE_SEL) {
 380                 cpipe->pipe_state &= ~PIPE_SEL;
 381                 selwakeup(&cpipe->pipe_sel);
 382         }
 383         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 384                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 385         KNOTE(&cpipe->pipe_sel.si_note, 0);
 386 }
 387
 388 /* ARGSUSED */
 389 static int
 390 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
 391         int flags, struct thread *td)
 392 {
 393         struct pipe *rpipe = (struct pipe *) fp->f_data;
 394         int error;
 395         int nread = 0;
 396         u_int size;
 397
 398         ++rpipe->pipe_busy;
 399         error = pipelock(rpipe, 1);
 400         if (error)
 401                 goto unlocked_error;
 402
 403         while (uio->uio_resid) {
 404                 /*
 405                  * normal pipe buffer receive
 406                  */
 407                 if (rpipe->pipe_buffer.cnt > 0) {
 408                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 409                         if (size > rpipe->pipe_buffer.cnt)
 410                                 size = rpipe->pipe_buffer.cnt;
 411                         if (size > (u_int) uio->uio_resid)
 412                                 size = (u_int) uio->uio_resid;
 413
 414                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 415                                         size, uio);
 416                         if (error)
 417                                 break;
 418
 419                         rpipe->pipe_buffer.out += size;
 420                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 421                                 rpipe->pipe_buffer.out = 0;
 422
 423                         rpipe->pipe_buffer.cnt -= size;
 424
 425                         /*
 426                          * If there is no more to read in the pipe, reset
 427                          * its pointers to the beginning.  This improves
 428                          * cache hit stats.
 429                          */
 430                         if (rpipe->pipe_buffer.cnt == 0) {
 431                                 rpipe->pipe_buffer.in = 0;
 432                                 rpipe->pipe_buffer.out = 0;
 433                         }
 434                         nread += size;
 435 #ifndef PIPE_NODIRECT
 436                 /*
 437                  * Direct copy, bypassing a kernel buffer.  We cannot mess
 438                  * with the direct-write buffer until PIPE_DIRECTIP is
 439                  * cleared.  In order to prevent the pipe_write code from
 440                  * racing itself in direct_write, we set DIRECTIP when we
 441                  * clear DIRECTW after we have exhausted the buffer.
 442                  */
 443                 } else if (rpipe->pipe_map.xio_bytes &&
 444                            (rpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == PIPE_DIRECTW) {
 445                         error = xio_uio_copy(&rpipe->pipe_map, uio, &size);
 446                         if (error)
 447                                 break;
 448                         nread += size;
 449                         if (rpipe->pipe_map.xio_bytes == 0) {
 450                                 rpipe->pipe_state |= PIPE_DIRECTIP;
 451                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 452                                 wakeup(rpipe);
 453                         }
 454 #endif
 455                 } else {
 456                         /*
 457                          * detect EOF condition
 458                          * read returns 0 on EOF, no need to set error
 459                          */
 460                         if (rpipe->pipe_state & PIPE_EOF)
 461                                 break;
 462
 463                         /*
 464                          * If the "write-side" has been blocked, wake it up now.
 465                          */
 466                         if (rpipe->pipe_state & PIPE_WANTW) {
 467                                 rpipe->pipe_state &= ~PIPE_WANTW;
 468                                 wakeup(rpipe);
 469                         }
 470
 471                         /*
 472                          * Break if some data was read.
 473                          */
 474                         if (nread > 0)
 475                                 break;
 476
 477                         /*
 478                          * Unlock the pipe buffer for our remaining processing.  We
 479                          * will either break out with an error or we will sleep and
 480                          * relock to loop.
 481                          */
 482                         pipeunlock(rpipe);
 483
 484                         /*
 485                          * Handle non-blocking mode operation or
 486                          * wait for more data.
 487                          */
 488                         if (fp->f_flag & FNONBLOCK) {
 489                                 error = EAGAIN;
 490                         } else {
 491                                 rpipe->pipe_state |= PIPE_WANTR;
 492                                 if ((error = tsleep(rpipe, PCATCH|PNORESCHED,
 493                                     "piperd", 0)) == 0) {
 494                                         error = pipelock(rpipe, 1);
 495                                 }
 496                         }
 497                         if (error)
 498                                 goto unlocked_error;
 499                 }
 500         }
 501         pipeunlock(rpipe);
 502
 503         if (error == 0)
 504                 vfs_timestamp(&rpipe->pipe_atime);
 505 unlocked_error:
 506         --rpipe->pipe_busy;
 507
 508         /*
 509          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 510          */
 511         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 512                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 513                 wakeup(rpipe);
 514         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 515                 /*
 516                  * Handle write blocking hysteresis.
 517                  */
 518                 if (rpipe->pipe_state & PIPE_WANTW) {
 519                         rpipe->pipe_state &= ~PIPE_WANTW;
 520                         wakeup(rpipe);
 521                 }
 522         }
 523
 524         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 525                 pipeselwakeup(rpipe);
 526
 527         return (error);
 528 }
 529
 530 #ifndef PIPE_NODIRECT
 531 /*
 532  * Map the sending processes' buffer into kernel space and wire it.
 533  * This is similar to a physical write operation.
 534  */
 535 static int
 536 pipe_build_write_buffer(wpipe, uio)
 537         struct pipe *wpipe;
 538         struct uio *uio;
 539 {
 540         int error;
 541         u_int size;
 542
 543         size = (u_int) uio->uio_iov->iov_len;
 544         if (size > wpipe->pipe_buffer.size)
 545                 size = wpipe->pipe_buffer.size;
 546         if (size > XIO_INTERNAL_SIZE)
 547                 size = XIO_INTERNAL_SIZE;
 548
 549         error = xio_init_ubuf(&wpipe->pipe_map, uio->uio_iov->iov_base,
 550                                 size, XIOF_READ);
 551         if (error)
 552                 return(error);
 553
 554         /*
 555          * and update the uio data
 556          */
 557         uio->uio_iov->iov_len -= size;
 558         uio->uio_iov->iov_base += size;
 559         if (uio->uio_iov->iov_len == 0)
 560                 uio->uio_iov++;
 561         uio->uio_resid -= size;
 562         uio->uio_offset += size;
 563         return (0);
 564 }
 565
 566 /*
 567  * In the case of a signal, the writing process might go away.  This
 568  * code copies the data into the circular buffer so that the source
 569  * pages can be freed without loss of data.
 570  */
 571 static void
 572 pipe_clone_write_buffer(wpipe)
 573         struct pipe *wpipe;
 574 {
 575         int size;
 576
 577         size = wpipe->pipe_map.xio_bytes;
 578
 579         wpipe->pipe_buffer.in = size;
 580         wpipe->pipe_buffer.out = 0;
 581         wpipe->pipe_buffer.cnt = size;
 582         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 583
 584         xio_copy_xtok(&wpipe->pipe_map, wpipe->pipe_buffer.buffer, size);
 585         xio_release(&wpipe->pipe_map);
 586 }
 587
 588 /*
 589  * This implements the pipe buffer write mechanism.  Note that only
 590  * a direct write OR a normal pipe write can be pending at any given time.
 591  * If there are any characters in the pipe buffer, the direct write will
 592  * be deferred until the receiving process grabs all of the bytes from
 593  * the pipe buffer.  Then the direct mapping write is set-up.
 594  */
 595 static int
 596 pipe_direct_write(wpipe, uio)
 597         struct pipe *wpipe;
 598         struct uio *uio;
 599 {
 600         int error;
 601
 602 retry:
 603         while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 604                 if (wpipe->pipe_state & PIPE_WANTR) {
 605                         wpipe->pipe_state &= ~PIPE_WANTR;
 606                         wakeup(wpipe);
 607                 }
 608                 wpipe->pipe_state |= PIPE_WANTW;
 609                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 610                 if (error)
 611                         goto error2;
 612                 if (wpipe->pipe_state & PIPE_EOF) {
 613                         error = EPIPE;
 614                         goto error2;
 615                 }
 616         }
 617         KKASSERT(wpipe->pipe_map.xio_bytes == 0);
 618         if (wpipe->pipe_buffer.cnt > 0) {
 619                 if (wpipe->pipe_state & PIPE_WANTR) {
 620                         wpipe->pipe_state &= ~PIPE_WANTR;
 621                         wakeup(wpipe);
 622                 }
 623
 624                 wpipe->pipe_state |= PIPE_WANTW;
 625                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 626                 if (error)
 627                         goto error2;
 628                 if (wpipe->pipe_state & PIPE_EOF) {
 629                         error = EPIPE;
 630                         goto error2;
 631                 }
 632                 goto retry;
 633         }
 634
 635         /*
 636          * Build our direct-write buffer
 637          */
 638         wpipe->pipe_state |= PIPE_DIRECTW | PIPE_DIRECTIP;
 639         error = pipe_build_write_buffer(wpipe, uio);
 640         if (error)
 641                 goto error1;
 642         wpipe->pipe_state &= ~PIPE_DIRECTIP;
 643
 644         /*
 645          * Wait until the receiver has snarfed the data.  Since we are likely
 646          * going to sleep we optimize the case and yield synchronously,
 647          * possibly avoiding the tsleep().
 648          */
 649         error = 0;
 650         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 651                 if (wpipe->pipe_state & PIPE_EOF) {
 652                         pipelock(wpipe, 0);
 653                         xio_release(&wpipe->pipe_map);
 654                         pipeunlock(wpipe);
 655                         pipeselwakeup(wpipe);
 656                         error = EPIPE;
 657                         goto error1;
 658                 }
 659                 if (wpipe->pipe_state & PIPE_WANTR) {
 660                         wpipe->pipe_state &= ~PIPE_WANTR;
 661                         wakeup(wpipe);
 662                 }
 663                 pipeselwakeup(wpipe);
 664                 error = tsleep(wpipe, PCATCH|PNORESCHED, "pipdwt", 0);
 665         }
 666         pipelock(wpipe,0);
 667         if (wpipe->pipe_state & PIPE_DIRECTW) {
 668                 /*
 669                  * this bit of trickery substitutes a kernel buffer for
 670                  * the process that might be going away.
 671                  */
 672                 pipe_clone_write_buffer(wpipe);
 673                 KKASSERT((wpipe->pipe_state & PIPE_DIRECTIP) == 0);
 674         } else {
 675                 KKASSERT(wpipe->pipe_state & PIPE_DIRECTIP);
 676                 xio_release(&wpipe->pipe_map);
 677                 wpipe->pipe_state &= ~PIPE_DIRECTIP;
 678         }
 679         pipeunlock(wpipe);
 680         return (error);
 681
 682         /*
 683          * Direct-write error, clear the direct write flags.
 684          */
 685 error1:
 686         wpipe->pipe_state &= ~(PIPE_DIRECTW | PIPE_DIRECTIP);
 687         /* fallthrough */
 688
 689         /*
 690          * General error, wakeup the other side if it happens to be sleeping.
 691          */
 692 error2:
 693         wakeup(wpipe);
 694         return (error);
 695 }
 696 #endif
 697
 698 static int
 699 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
 700         int flags, struct thread *td)
 701 {
 702         int error = 0;
 703         int orig_resid;
 704         struct pipe *wpipe, *rpipe;
 705
 706         rpipe = (struct pipe *) fp->f_data;
 707         wpipe = rpipe->pipe_peer;
 708
 709         /*
 710          * detect loss of pipe read side, issue SIGPIPE if lost.
 711          */
 712         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 713                 return (EPIPE);
 714         }
 715         ++wpipe->pipe_busy;
 716
 717         /*
 718          * If it is advantageous to resize the pipe buffer, do
 719          * so.
 720          */
 721         if ((uio->uio_resid > PIPE_SIZE) &&
 722                 (pipe_nbig < pipe_maxbig) &&
 723                 (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) == 0 &&
 724                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 725                 (wpipe->pipe_buffer.cnt == 0)) {
 726
 727                 if ((error = pipelock(wpipe,1)) == 0) {
 728                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 729                                 pipe_nbig++;
 730                         pipeunlock(wpipe);
 731                 }
 732         }
 733
 734         /*
 735          * If an early error occured unbusy and return, waking up any pending
 736          * readers.
 737          */
 738         if (error) {
 739                 --wpipe->pipe_busy;
 740                 if ((wpipe->pipe_busy == 0) &&
 741                     (wpipe->pipe_state & PIPE_WANT)) {
 742                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 743                         wakeup(wpipe);
 744                 }
 745                 return(error);
 746         }
 747
 748         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 749
 750         orig_resid = uio->uio_resid;
 751
 752         while (uio->uio_resid) {
 753                 int space;
 754
 755 #ifndef PIPE_NODIRECT
 756                 /*
 757                  * If the transfer is large, we can gain performance if
 758                  * we do process-to-process copies directly.
 759                  * If the write is non-blocking, we don't use the
 760                  * direct write mechanism.
 761                  *
 762                  * The direct write mechanism will detect the reader going
 763                  * away on us.
 764                  */
 765                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
 766                     (fp->f_flag & FNONBLOCK) == 0) {
 767                         error = pipe_direct_write( wpipe, uio);
 768                         if (error)
 769                                 break;
 770                         continue;
 771                 }
 772 #endif
 773
 774                 /*
 775                  * Pipe buffered writes cannot be coincidental with
 776                  * direct writes.  We wait until the currently executing
 777                  * direct write is completed before we start filling the
 778                  * pipe buffer.  We break out if a signal occurs or the
 779                  * reader goes away.
 780                  */
 781         retrywrite:
 782                 while (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 783                         if (wpipe->pipe_state & PIPE_WANTR) {
 784                                 wpipe->pipe_state &= ~PIPE_WANTR;
 785                                 wakeup(wpipe);
 786                         }
 787                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 788                         if (wpipe->pipe_state & PIPE_EOF)
 789                                 break;
 790                         if (error)
 791                                 break;
 792                 }
 793                 if (wpipe->pipe_state & PIPE_EOF) {
 794                         error = EPIPE;
 795                         break;
 796                 }
 797
 798                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 799
 800                 /* Writes of size <= PIPE_BUF must be atomic. */
 801                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 802                         space = 0;
 803
 804                 /*
 805                  * Write to fill, read size handles write hysteresis.  Also
 806                  * additional restrictions can cause select-based non-blocking
 807                  * writes to spin.
 808                  */
 809                 if (space > 0) {
 810                         if ((error = pipelock(wpipe,1)) == 0) {
 811                                 int size;       /* Transfer size */
 812                                 int segsize;    /* first segment to transfer */
 813
 814                                 /*
 815                                  * It is possible for a direct write to
 816                                  * slip in on us... handle it here...
 817                                  */
 818                                 if (wpipe->pipe_state & (PIPE_DIRECTW|PIPE_DIRECTIP)) {
 819                                         pipeunlock(wpipe);
 820                                         goto retrywrite;
 821                                 }
 822                                 /*
 823                                  * If a process blocked in uiomove, our
 824                                  * value for space might be bad.
 825                                  *
 826                                  * XXX will we be ok if the reader has gone
 827                                  * away here?
 828                                  */
 829                                 if (space > wpipe->pipe_buffer.size -
 830                                     wpipe->pipe_buffer.cnt) {
 831                                         pipeunlock(wpipe);
 832                                         goto retrywrite;
 833                                 }
 834
 835                                 /*
 836                                  * Transfer size is minimum of uio transfer
 837                                  * and free space in pipe buffer.
 838                                  */
 839                                 if (space > uio->uio_resid)
 840                                         size = uio->uio_resid;
 841                                 else
 842                                         size = space;
 843                                 /*
 844                                  * First segment to transfer is minimum of
 845                                  * transfer size and contiguous space in
 846                                  * pipe buffer.  If first segment to transfer
 847                                  * is less than the transfer size, we've got
 848                                  * a wraparound in the buffer.
 849                                  */
 850                                 segsize = wpipe->pipe_buffer.size -
 851                                         wpipe->pipe_buffer.in;
 852                                 if (segsize > size)
 853                                         segsize = size;
 854
 855                                 /* Transfer first segment */
 856
 857                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 858                                                 segsize, uio);
 859
 860                                 if (error == 0 && segsize < size) {
 861                                         /*
 862                                          * Transfer remaining part now, to
 863                                          * support atomic writes.  Wraparound
 864                                          * happened.
 865                                          */
 866                                         if (wpipe->pipe_buffer.in + segsize !=
 867                                             wpipe->pipe_buffer.size)
 868                                                 panic("Expected pipe buffer wraparound disappeared");
 869
 870                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
 871                                                         size - segsize, uio);
 872                                 }
 873                                 if (error == 0) {
 874                                         wpipe->pipe_buffer.in += size;
 875                                         if (wpipe->pipe_buffer.in >=
 876                                             wpipe->pipe_buffer.size) {
 877                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
 878                                                         panic("Expected wraparound bad");
 879                                                 wpipe->pipe_buffer.in = size - segsize;
 880                                         }
 881
 882                                         wpipe->pipe_buffer.cnt += size;
 883                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
 884                                                 panic("Pipe buffer overflow");
 885
 886                                 }
 887                                 pipeunlock(wpipe);
 888                         }
 889                         if (error)
 890                                 break;
 891
 892                 } else {
 893                         /*
 894                          * If the "read-side" has been blocked, wake it up now
 895                          * and yield to let it drain synchronously rather
 896                          * then block.
 897                          */
 898                         if (wpipe->pipe_state & PIPE_WANTR) {
 899                                 wpipe->pipe_state &= ~PIPE_WANTR;
 900                                 wakeup(wpipe);
 901                         }
 902
 903                         /*
 904                          * don't block on non-blocking I/O
 905                          */
 906                         if (fp->f_flag & FNONBLOCK) {
 907                                 error = EAGAIN;
 908                                 break;
 909                         }
 910
 911                         /*
 912                          * We have no more space and have something to offer,
 913                          * wake up select/poll.
 914                          */
 915                         pipeselwakeup(wpipe);
 916
 917                         wpipe->pipe_state |= PIPE_WANTW;
 918                         error = tsleep(wpipe, PCATCH|PNORESCHED, "pipewr", 0);
 919                         if (error != 0)
 920                                 break;
 921                         /*
 922                          * If read side wants to go away, we just issue a signal
 923                          * to ourselves.
 924                          */
 925                         if (wpipe->pipe_state & PIPE_EOF) {
 926                                 error = EPIPE;
 927                                 break;
 928                         }
 929                 }
 930         }
 931
 932         --wpipe->pipe_busy;
 933
 934         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
 935                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 936                 wakeup(wpipe);
 937         } else if (wpipe->pipe_buffer.cnt > 0) {
 938                 /*
 939                  * If we have put any characters in the buffer, we wake up
 940                  * the reader.
 941                  */
 942                 if (wpipe->pipe_state & PIPE_WANTR) {
 943                         wpipe->pipe_state &= ~PIPE_WANTR;
 944                         wakeup(wpipe);
 945                 }
 946         }
 947
 948         /*
 949          * Don't return EPIPE if I/O was successful
 950          */
 951         if ((wpipe->pipe_buffer.cnt == 0) &&
 952             (uio->uio_resid == 0) &&
 953             (error == EPIPE)) {
 954                 error = 0;
 955         }
 956
 957         if (error == 0)
 958                 vfs_timestamp(&wpipe->pipe_mtime);
 959
 960         /*
 961          * We have something to offer,
 962          * wake up select/poll.
 963          */
 964         if (wpipe->pipe_buffer.cnt)
 965                 pipeselwakeup(wpipe);
 966
 967         return (error);
 968 }
 969
 970 /*
 971  * we implement a very minimal set of ioctls for compatibility with sockets.
 972  */
 973 int
 974 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
 975 {
 976         struct pipe *mpipe = (struct pipe *)fp->f_data;
 977
 978         switch (cmd) {
 979
 980         case FIONBIO:
 981                 return (0);
 982
 983         case FIOASYNC:
 984                 if (*(int *)data) {
 985                         mpipe->pipe_state |= PIPE_ASYNC;
 986                 } else {
 987                         mpipe->pipe_state &= ~PIPE_ASYNC;
 988                 }
 989                 return (0);
 990
 991         case FIONREAD:
 992                 if (mpipe->pipe_state & PIPE_DIRECTW) {
 993                         *(int *)data = mpipe->pipe_map.xio_bytes;
 994                 } else {
 995                         *(int *)data = mpipe->pipe_buffer.cnt;
 996                 }
 997                 return (0);
 998
 999         case FIOSETOWN:
1000                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1001
1002         case FIOGETOWN:
1003                 *(int *)data = fgetown(mpipe->pipe_sigio);
1004                 return (0);
1005
1006         /* This is deprecated, FIOSETOWN should be used instead. */
1007         case TIOCSPGRP:
1008                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1009
1010         /* This is deprecated, FIOGETOWN should be used instead. */
1011         case TIOCGPGRP:
1012                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1013                 return (0);
1014
1015         }
1016         return (ENOTTY);
1017 }
1018
1019 int
1020 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1021 {
1022         struct pipe *rpipe = (struct pipe *)fp->f_data;
1023         struct pipe *wpipe;
1024         int revents = 0;
1025
1026         wpipe = rpipe->pipe_peer;
1027         if (events & (POLLIN | POLLRDNORM))
1028                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1029                     (rpipe->pipe_buffer.cnt > 0) ||
1030                     (rpipe->pipe_state & PIPE_EOF))
1031                         revents |= events & (POLLIN | POLLRDNORM);
1032
1033         if (events & (POLLOUT | POLLWRNORM))
1034                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1035                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1036                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1037                         revents |= events & (POLLOUT | POLLWRNORM);
1038
1039         if ((rpipe->pipe_state & PIPE_EOF) ||
1040             (wpipe == NULL) ||
1041             (wpipe->pipe_state & PIPE_EOF))
1042                 revents |= POLLHUP;
1043
1044         if (revents == 0) {
1045                 if (events & (POLLIN | POLLRDNORM)) {
1046                         selrecord(td, &rpipe->pipe_sel);
1047                         rpipe->pipe_state |= PIPE_SEL;
1048                 }
1049
1050                 if (events & (POLLOUT | POLLWRNORM)) {
1051                         selrecord(td, &wpipe->pipe_sel);
1052                         wpipe->pipe_state |= PIPE_SEL;
1053                 }
1054         }
1055
1056         return (revents);
1057 }
1058
1059 static int
1060 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1061 {
1062         struct pipe *pipe = (struct pipe *)fp->f_data;
1063
1064         bzero((caddr_t)ub, sizeof(*ub));
1065         ub->st_mode = S_IFIFO;
1066         ub->st_blksize = pipe->pipe_buffer.size;
1067         ub->st_size = pipe->pipe_buffer.cnt;
1068         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1069         ub->st_atimespec = pipe->pipe_atime;
1070         ub->st_mtimespec = pipe->pipe_mtime;
1071         ub->st_ctimespec = pipe->pipe_ctime;
1072         /*
1073          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1074          * st_flags, st_gen.
1075          * XXX (st_dev, st_ino) should be unique.
1076          */
1077         return (0);
1078 }
1079
1080 /* ARGSUSED */
1081 static int
1082 pipe_close(struct file *fp, struct thread *td)
1083 {
1084         struct pipe *cpipe = (struct pipe *)fp->f_data;
1085
1086         fp->f_ops = &badfileops;
1087         fp->f_data = NULL;
1088         funsetown(cpipe->pipe_sigio);
1089         pipeclose(cpipe);
1090         return (0);
1091 }
1092
1093 static void
1094 pipe_free_kmem(struct pipe *cpipe)
1095 {
1096         if (cpipe->pipe_buffer.buffer != NULL) {
1097                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1098                         --pipe_nbig;
1099                 kmem_free(kernel_map,
1100                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1101                         cpipe->pipe_buffer.size);
1102                 cpipe->pipe_buffer.buffer = NULL;
1103                 cpipe->pipe_buffer.object = NULL;
1104         }
1105 #ifndef PIPE_NODIRECT
1106         KKASSERT(cpipe->pipe_map.xio_bytes == 0 &&
1107                 cpipe->pipe_map.xio_offset == 0 &&
1108                 cpipe->pipe_map.xio_npages == 0);
1109 #endif
1110 }
1111
1112 /*
1113  * shutdown the pipe
1114  */
1115 static void
1116 pipeclose(struct pipe *cpipe)
1117 {
1118         globaldata_t gd;
1119         struct pipe *ppipe;
1120
1121         if (cpipe == NULL)
1122                 return;
1123
1124         pipeselwakeup(cpipe);
1125
1126         /*
1127          * If the other side is blocked, wake it up saying that
1128          * we want to close it down.
1129          */
1130         while (cpipe->pipe_busy) {
1131                 wakeup(cpipe);
1132                 cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1133                 tsleep(cpipe, 0, "pipecl", 0);
1134         }
1135
1136         /*
1137          * Disconnect from peer
1138          */
1139         if ((ppipe = cpipe->pipe_peer) != NULL) {
1140                 pipeselwakeup(ppipe);
1141
1142                 ppipe->pipe_state |= PIPE_EOF;
1143                 wakeup(ppipe);
1144                 KNOTE(&ppipe->pipe_sel.si_note, 0);
1145                 ppipe->pipe_peer = NULL;
1146         }
1147
1148         /*
1149          * free or cache resources
1150          */
1151         gd = mycpu;
1152         if (gd->gd_pipeqcount >= pipe_maxcache ||
1153             cpipe->pipe_buffer.size != PIPE_SIZE
1154         ) {
1155                 pipe_free_kmem(cpipe);
1156                 free(cpipe, M_PIPE);
1157         } else {
1158                 KKASSERT(cpipe->pipe_map.xio_npages == 0 &&
1159                         cpipe->pipe_map.xio_bytes == 0 &&
1160                         cpipe->pipe_map.xio_offset == 0);
1161                 cpipe->pipe_state = 0;
1162                 cpipe->pipe_busy = 0;
1163                 cpipe->pipe_peer = gd->gd_pipeq;
1164                 gd->gd_pipeq = cpipe;
1165                 ++gd->gd_pipeqcount;
1166         }
1167 }
1168
1169 /*ARGSUSED*/
1170 static int
1171 pipe_kqfilter(struct file *fp, struct knote *kn)
1172 {
1173         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1174
1175         switch (kn->kn_filter) {
1176         case EVFILT_READ:
1177                 kn->kn_fop = &pipe_rfiltops;
1178                 break;
1179         case EVFILT_WRITE:
1180                 kn->kn_fop = &pipe_wfiltops;
1181                 cpipe = cpipe->pipe_peer;
1182                 if (cpipe == NULL)
1183                         /* other end of pipe has been closed */
1184                         return (EPIPE);
1185                 break;
1186         default:
1187                 return (1);
1188         }
1189         kn->kn_hook = (caddr_t)cpipe;
1190
1191         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1192         return (0);
1193 }
1194
1195 static void
1196 filt_pipedetach(struct knote *kn)
1197 {
1198         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1199
1200         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1201 }
1202
1203 /*ARGSUSED*/
1204 static int
1205 filt_piperead(struct knote *kn, long hint)
1206 {
1207         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1208         struct pipe *wpipe = rpipe->pipe_peer;
1209
1210         kn->kn_data = rpipe->pipe_buffer.cnt;
1211         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1212                 kn->kn_data = rpipe->pipe_map.xio_bytes;
1213
1214         if ((rpipe->pipe_state & PIPE_EOF) ||
1215             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1216                 kn->kn_flags |= EV_EOF;
1217                 return (1);
1218         }
1219         return (kn->kn_data > 0);
1220 }
1221
1222 /*ARGSUSED*/
1223 static int
1224 filt_pipewrite(struct knote *kn, long hint)
1225 {
1226         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1227         struct pipe *wpipe = rpipe->pipe_peer;
1228
1229         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1230                 kn->kn_data = 0;
1231                 kn->kn_flags |= EV_EOF;
1232                 return (1);
1233         }
1234         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1235         if (wpipe->pipe_state & PIPE_DIRECTW)
1236                 kn->kn_data = 0;
1237
1238         return (kn->kn_data >= PIPE_BUF);
1239 }