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 1.3 2003/06/23 17:55:41 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/proc.h>
  56 #include <sys/fcntl.h>
  57 #include <sys/file.h>
  58 #include <sys/filedesc.h>
  59 #include <sys/filio.h>
  60 #include <sys/ttycom.h>
  61 #include <sys/stat.h>
  62 #include <sys/poll.h>
  63 #include <sys/select.h>
  64 #include <sys/signalvar.h>
  65 #include <sys/sysproto.h>
  66 #include <sys/pipe.h>
  67 #include <sys/vnode.h>
  68 #include <sys/uio.h>
  69 #include <sys/event.h>
  70
  71 #include <vm/vm.h>
  72 #include <vm/vm_param.h>
  73 #include <sys/lock.h>
  74 #include <vm/vm_object.h>
  75 #include <vm/vm_kern.h>
  76 #include <vm/vm_extern.h>
  77 #include <vm/pmap.h>
  78 #include <vm/vm_map.h>
  79 #include <vm/vm_page.h>
  80 #include <vm/vm_zone.h>
  81
  82 /*
  83  * Use this define if you want to disable *fancy* VM things.  Expect an
  84  * approx 30% decrease in transfer rate.  This could be useful for
  85  * NetBSD or OpenBSD.
  86  */
  87 /* #define PIPE_NODIRECT */
  88
  89 /*
  90  * interfaces to the outside world
  91  */
  92 static int pipe_read __P((struct file *fp, struct uio *uio,
  93                 struct ucred *cred, int flags, struct proc *p));
  94 static int pipe_write __P((struct file *fp, struct uio *uio,
  95                 struct ucred *cred, int flags, struct proc *p));
  96 static int pipe_close __P((struct file *fp, struct proc *p));
  97 static int pipe_poll __P((struct file *fp, int events, struct ucred *cred,
  98                 struct proc *p));
  99 static int pipe_kqfilter __P((struct file *fp, struct knote *kn));
 100 static int pipe_stat __P((struct file *fp, struct stat *sb, struct proc *p));
 101 static int pipe_ioctl __P((struct file *fp, u_long cmd, caddr_t data, struct proc *p));
 102
 103 static struct fileops pipeops = {
 104         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 105         pipe_stat, pipe_close
 106 };
 107
 108 static void     filt_pipedetach(struct knote *kn);
 109 static int      filt_piperead(struct knote *kn, long hint);
 110 static int      filt_pipewrite(struct knote *kn, long hint);
 111
 112 static struct filterops pipe_rfiltops =
 113         { 1, NULL, filt_pipedetach, filt_piperead };
 114 static struct filterops pipe_wfiltops =
 115         { 1, NULL, filt_pipedetach, filt_pipewrite };
 116
 117
 118 /*
 119  * Default pipe buffer size(s), this can be kind-of large now because pipe
 120  * space is pageable.  The pipe code will try to maintain locality of
 121  * reference for performance reasons, so small amounts of outstanding I/O
 122  * will not wipe the cache.
 123  */
 124 #define MINPIPESIZE (PIPE_SIZE/3)
 125 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 126
 127 /*
 128  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 129  * is there so that on large systems, we don't exhaust it.
 130  */
 131 #define MAXPIPEKVA (8*1024*1024)
 132
 133 /*
 134  * Limit for direct transfers, we cannot, of course limit
 135  * the amount of kva for pipes in general though.
 136  */
 137 #define LIMITPIPEKVA (16*1024*1024)
 138
 139 /*
 140  * Limit the number of "big" pipes
 141  */
 142 #define LIMITBIGPIPES   32
 143 static int nbigpipe;
 144
 145 static int amountpipekva;
 146
 147 static void pipeclose __P((struct pipe *cpipe));
 148 static void pipe_free_kmem __P((struct pipe *cpipe));
 149 static int pipe_create __P((struct pipe **cpipep));
 150 static __inline int pipelock __P((struct pipe *cpipe, int catch));
 151 static __inline void pipeunlock __P((struct pipe *cpipe));
 152 static __inline void pipeselwakeup __P((struct pipe *cpipe));
 153 #ifndef PIPE_NODIRECT
 154 static int pipe_build_write_buffer __P((struct pipe *wpipe, struct uio *uio));
 155 static void pipe_destroy_write_buffer __P((struct pipe *wpipe));
 156 static int pipe_direct_write __P((struct pipe *wpipe, struct uio *uio));
 157 static void pipe_clone_write_buffer __P((struct pipe *wpipe));
 158 #endif
 159 static int pipespace __P((struct pipe *cpipe, int size));
 160
 161 static vm_zone_t pipe_zone;
 162
 163 /*
 164  * The pipe system call for the DTYPE_PIPE type of pipes
 165  *
 166  * pipe_ARgs(int dummy)
 167  */
 168
 169 /* ARGSUSED */
 170 int
 171 pipe(struct pipe_args *uap)
 172 {
 173         struct proc *p = curproc;
 174         struct filedesc *fdp = p->p_fd;
 175         struct file *rf, *wf;
 176         struct pipe *rpipe, *wpipe;
 177         int fd, error;
 178
 179         if (pipe_zone == NULL)
 180                 pipe_zone = zinit("PIPE", sizeof(struct pipe), 0, 0, 4);
 181
 182         rpipe = wpipe = NULL;
 183         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 184                 pipeclose(rpipe);
 185                 pipeclose(wpipe);
 186                 return (ENFILE);
 187         }
 188
 189         rpipe->pipe_state |= PIPE_DIRECTOK;
 190         wpipe->pipe_state |= PIPE_DIRECTOK;
 191
 192         error = falloc(p, &rf, &fd);
 193         if (error) {
 194                 pipeclose(rpipe);
 195                 pipeclose(wpipe);
 196                 return (error);
 197         }
 198         fhold(rf);
 199         p->p_retval[0] = fd;
 200
 201         /*
 202          * Warning: once we've gotten past allocation of the fd for the
 203          * read-side, we can only drop the read side via fdrop() in order
 204          * to avoid races against processes which manage to dup() the read
 205          * side while we are blocked trying to allocate the write side.
 206          */
 207         rf->f_flag = FREAD | FWRITE;
 208         rf->f_type = DTYPE_PIPE;
 209         rf->f_data = (caddr_t)rpipe;
 210         rf->f_ops = &pipeops;
 211         error = falloc(p, &wf, &fd);
 212         if (error) {
 213                 if (fdp->fd_ofiles[p->p_retval[0]] == rf) {
 214                         fdp->fd_ofiles[p->p_retval[0]] = NULL;
 215                         fdrop(rf, p);
 216                 }
 217                 fdrop(rf, p);
 218                 /* rpipe has been closed by fdrop(). */
 219                 pipeclose(wpipe);
 220                 return (error);
 221         }
 222         wf->f_flag = FREAD | FWRITE;
 223         wf->f_type = DTYPE_PIPE;
 224         wf->f_data = (caddr_t)wpipe;
 225         wf->f_ops = &pipeops;
 226         p->p_retval[1] = fd;
 227
 228         rpipe->pipe_peer = wpipe;
 229         wpipe->pipe_peer = rpipe;
 230         fdrop(rf, p);
 231
 232         return (0);
 233 }
 234
 235 /*
 236  * Allocate kva for pipe circular buffer, the space is pageable
 237  * This routine will 'realloc' the size of a pipe safely, if it fails
 238  * it will retain the old buffer.
 239  * If it fails it will return ENOMEM.
 240  */
 241 static int
 242 pipespace(cpipe, size)
 243         struct pipe *cpipe;
 244         int size;
 245 {
 246         struct vm_object *object;
 247         caddr_t buffer;
 248         int npages, error;
 249
 250         npages = round_page(size)/PAGE_SIZE;
 251         /*
 252          * Create an object, I don't like the idea of paging to/from
 253          * kernel_object.
 254          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 255          */
 256         object = vm_object_allocate(OBJT_DEFAULT, npages);
 257         buffer = (caddr_t) vm_map_min(kernel_map);
 258
 259         /*
 260          * Insert the object into the kernel map, and allocate kva for it.
 261          * The map entry is, by default, pageable.
 262          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 263          */
 264         error = vm_map_find(kernel_map, object, 0,
 265                 (vm_offset_t *) &buffer, size, 1,
 266                 VM_PROT_ALL, VM_PROT_ALL, 0);
 267
 268         if (error != KERN_SUCCESS) {
 269                 vm_object_deallocate(object);
 270                 return (ENOMEM);
 271         }
 272
 273         /* free old resources if we're resizing */
 274         pipe_free_kmem(cpipe);
 275         cpipe->pipe_buffer.object = object;
 276         cpipe->pipe_buffer.buffer = buffer;
 277         cpipe->pipe_buffer.size = size;
 278         cpipe->pipe_buffer.in = 0;
 279         cpipe->pipe_buffer.out = 0;
 280         cpipe->pipe_buffer.cnt = 0;
 281         amountpipekva += cpipe->pipe_buffer.size;
 282         return (0);
 283 }
 284
 285 /*
 286  * initialize and allocate VM and memory for pipe
 287  */
 288 static int
 289 pipe_create(cpipep)
 290         struct pipe **cpipep;
 291 {
 292         struct pipe *cpipe;
 293         int error;
 294
 295         *cpipep = zalloc(pipe_zone);
 296         if (*cpipep == NULL)
 297                 return (ENOMEM);
 298
 299         cpipe = *cpipep;
 300
 301         /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */
 302         cpipe->pipe_buffer.object = NULL;
 303 #ifndef PIPE_NODIRECT
 304         cpipe->pipe_map.kva = NULL;
 305 #endif
 306         /*
 307          * protect so pipeclose() doesn't follow a junk pointer
 308          * if pipespace() fails.
 309          */
 310         bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel));
 311         cpipe->pipe_state = 0;
 312         cpipe->pipe_peer = NULL;
 313         cpipe->pipe_busy = 0;
 314
 315 #ifndef PIPE_NODIRECT
 316         /*
 317          * pipe data structure initializations to support direct pipe I/O
 318          */
 319         cpipe->pipe_map.cnt = 0;
 320         cpipe->pipe_map.kva = 0;
 321         cpipe->pipe_map.pos = 0;
 322         cpipe->pipe_map.npages = 0;
 323         /* cpipe->pipe_map.ms[] = invalid */
 324 #endif
 325
 326         error = pipespace(cpipe, PIPE_SIZE);
 327         if (error)
 328                 return (error);
 329
 330         vfs_timestamp(&cpipe->pipe_ctime);
 331         cpipe->pipe_atime = cpipe->pipe_ctime;
 332         cpipe->pipe_mtime = cpipe->pipe_ctime;
 333
 334         return (0);
 335 }
 336
 337
 338 /*
 339  * lock a pipe for I/O, blocking other access
 340  */
 341 static __inline int
 342 pipelock(cpipe, catch)
 343         struct pipe *cpipe;
 344         int catch;
 345 {
 346         int error;
 347
 348         while (cpipe->pipe_state & PIPE_LOCK) {
 349                 cpipe->pipe_state |= PIPE_LWANT;
 350                 error = tsleep(cpipe, catch ? (PRIBIO | PCATCH) : PRIBIO,
 351                     "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(fp, uio, cred, flags, p)
 391         struct file *fp;
 392         struct uio *uio;
 393         struct ucred *cred;
 394         struct proc *p;
 395         int flags;
 396 {
 397         struct pipe *rpipe = (struct pipe *) fp->f_data;
 398         int error;
 399         int nread = 0;
 400         u_int size;
 401
 402         ++rpipe->pipe_busy;
 403         error = pipelock(rpipe, 1);
 404         if (error)
 405                 goto unlocked_error;
 406
 407         while (uio->uio_resid) {
 408                 /*
 409                  * normal pipe buffer receive
 410                  */
 411                 if (rpipe->pipe_buffer.cnt > 0) {
 412                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 413                         if (size > rpipe->pipe_buffer.cnt)
 414                                 size = rpipe->pipe_buffer.cnt;
 415                         if (size > (u_int) uio->uio_resid)
 416                                 size = (u_int) uio->uio_resid;
 417
 418                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 419                                         size, uio);
 420                         if (error)
 421                                 break;
 422
 423                         rpipe->pipe_buffer.out += size;
 424                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 425                                 rpipe->pipe_buffer.out = 0;
 426
 427                         rpipe->pipe_buffer.cnt -= size;
 428
 429                         /*
 430                          * If there is no more to read in the pipe, reset
 431                          * its pointers to the beginning.  This improves
 432                          * cache hit stats.
 433                          */
 434                         if (rpipe->pipe_buffer.cnt == 0) {
 435                                 rpipe->pipe_buffer.in = 0;
 436                                 rpipe->pipe_buffer.out = 0;
 437                         }
 438                         nread += size;
 439 #ifndef PIPE_NODIRECT
 440                 /*
 441                  * Direct copy, bypassing a kernel buffer.
 442                  */
 443                 } else if ((size = rpipe->pipe_map.cnt) &&
 444                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 445                         caddr_t va;
 446                         if (size > (u_int) uio->uio_resid)
 447                                 size = (u_int) uio->uio_resid;
 448
 449                         va = (caddr_t) rpipe->pipe_map.kva +
 450                             rpipe->pipe_map.pos;
 451                         error = uiomove(va, size, uio);
 452                         if (error)
 453                                 break;
 454                         nread += size;
 455                         rpipe->pipe_map.pos += size;
 456                         rpipe->pipe_map.cnt -= size;
 457                         if (rpipe->pipe_map.cnt == 0) {
 458                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 459                                 wakeup(rpipe);
 460                         }
 461 #endif
 462                 } else {
 463                         /*
 464                          * detect EOF condition
 465                          * read returns 0 on EOF, no need to set error
 466                          */
 467                         if (rpipe->pipe_state & PIPE_EOF)
 468                                 break;
 469
 470                         /*
 471                          * If the "write-side" has been blocked, wake it up now.
 472                          */
 473                         if (rpipe->pipe_state & PIPE_WANTW) {
 474                                 rpipe->pipe_state &= ~PIPE_WANTW;
 475                                 wakeup(rpipe);
 476                         }
 477
 478                         /*
 479                          * Break if some data was read.
 480                          */
 481                         if (nread > 0)
 482                                 break;
 483
 484                         /*
 485                          * Unlock the pipe buffer for our remaining processing.  We
 486                          * will either break out with an error or we will sleep and
 487                          * relock to loop.
 488                          */
 489                         pipeunlock(rpipe);
 490
 491                         /*
 492                          * Handle non-blocking mode operation or
 493                          * wait for more data.
 494                          */
 495                         if (fp->f_flag & FNONBLOCK) {
 496                                 error = EAGAIN;
 497                         } else {
 498                                 rpipe->pipe_state |= PIPE_WANTR;
 499                                 if ((error = tsleep(rpipe, PRIBIO | PCATCH,
 500                                     "piperd", 0)) == 0)
 501                                         error = pipelock(rpipe, 1);
 502                         }
 503                         if (error)
 504                                 goto unlocked_error;
 505                 }
 506         }
 507         pipeunlock(rpipe);
 508
 509         if (error == 0)
 510                 vfs_timestamp(&rpipe->pipe_atime);
 511 unlocked_error:
 512         --rpipe->pipe_busy;
 513
 514         /*
 515          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 516          */
 517         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 518                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 519                 wakeup(rpipe);
 520         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 521                 /*
 522                  * Handle write blocking hysteresis.
 523                  */
 524                 if (rpipe->pipe_state & PIPE_WANTW) {
 525                         rpipe->pipe_state &= ~PIPE_WANTW;
 526                         wakeup(rpipe);
 527                 }
 528         }
 529
 530         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 531                 pipeselwakeup(rpipe);
 532
 533         return (error);
 534 }
 535
 536 #ifndef PIPE_NODIRECT
 537 /*
 538  * Map the sending processes' buffer into kernel space and wire it.
 539  * This is similar to a physical write operation.
 540  */
 541 static int
 542 pipe_build_write_buffer(wpipe, uio)
 543         struct pipe *wpipe;
 544         struct uio *uio;
 545 {
 546         u_int size;
 547         int i;
 548         vm_offset_t addr, endaddr, paddr;
 549
 550         size = (u_int) uio->uio_iov->iov_len;
 551         if (size > wpipe->pipe_buffer.size)
 552                 size = wpipe->pipe_buffer.size;
 553
 554         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 555         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 556         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 557                 vm_page_t m;
 558
 559                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
 560                     (paddr = pmap_kextract(addr)) == 0) {
 561                         int j;
 562
 563                         for (j = 0; j < i; j++)
 564                                 vm_page_unwire(wpipe->pipe_map.ms[j], 1);
 565                         return (EFAULT);
 566                 }
 567
 568                 m = PHYS_TO_VM_PAGE(paddr);
 569                 vm_page_wire(m);
 570                 wpipe->pipe_map.ms[i] = m;
 571         }
 572
 573 /*
 574  * set up the control block
 575  */
 576         wpipe->pipe_map.npages = i;
 577         wpipe->pipe_map.pos =
 578             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 579         wpipe->pipe_map.cnt = size;
 580
 581 /*
 582  * and map the buffer
 583  */
 584         if (wpipe->pipe_map.kva == 0) {
 585                 /*
 586                  * We need to allocate space for an extra page because the
 587                  * address range might (will) span pages at times.
 588                  */
 589                 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
 590                         wpipe->pipe_buffer.size + PAGE_SIZE);
 591                 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE;
 592         }
 593         pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
 594                 wpipe->pipe_map.npages);
 595
 596 /*
 597  * and update the uio data
 598  */
 599
 600         uio->uio_iov->iov_len -= size;
 601         uio->uio_iov->iov_base += size;
 602         if (uio->uio_iov->iov_len == 0)
 603                 uio->uio_iov++;
 604         uio->uio_resid -= size;
 605         uio->uio_offset += size;
 606         return (0);
 607 }
 608
 609 /*
 610  * unmap and unwire the process buffer
 611  */
 612 static void
 613 pipe_destroy_write_buffer(wpipe)
 614         struct pipe *wpipe;
 615 {
 616         int i;
 617
 618         if (wpipe->pipe_map.kva) {
 619                 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
 620
 621                 if (amountpipekva > MAXPIPEKVA) {
 622                         vm_offset_t kva = wpipe->pipe_map.kva;
 623                         wpipe->pipe_map.kva = 0;
 624                         kmem_free(kernel_map, kva,
 625                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 626                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 627                 }
 628         }
 629         for (i = 0; i < wpipe->pipe_map.npages; i++)
 630                 vm_page_unwire(wpipe->pipe_map.ms[i], 1);
 631         wpipe->pipe_map.npages = 0;
 632 }
 633
 634 /*
 635  * In the case of a signal, the writing process might go away.  This
 636  * code copies the data into the circular buffer so that the source
 637  * pages can be freed without loss of data.
 638  */
 639 static void
 640 pipe_clone_write_buffer(wpipe)
 641         struct pipe *wpipe;
 642 {
 643         int size;
 644         int pos;
 645
 646         size = wpipe->pipe_map.cnt;
 647         pos = wpipe->pipe_map.pos;
 648         bcopy((caddr_t) wpipe->pipe_map.kva + pos,
 649             (caddr_t) wpipe->pipe_buffer.buffer, size);
 650
 651         wpipe->pipe_buffer.in = size;
 652         wpipe->pipe_buffer.out = 0;
 653         wpipe->pipe_buffer.cnt = size;
 654         wpipe->pipe_state &= ~PIPE_DIRECTW;
 655
 656         pipe_destroy_write_buffer(wpipe);
 657 }
 658
 659 /*
 660  * This implements the pipe buffer write mechanism.  Note that only
 661  * a direct write OR a normal pipe write can be pending at any given time.
 662  * If there are any characters in the pipe buffer, the direct write will
 663  * be deferred until the receiving process grabs all of the bytes from
 664  * the pipe buffer.  Then the direct mapping write is set-up.
 665  */
 666 static int
 667 pipe_direct_write(wpipe, uio)
 668         struct pipe *wpipe;
 669         struct uio *uio;
 670 {
 671         int error;
 672
 673 retry:
 674         while (wpipe->pipe_state & PIPE_DIRECTW) {
 675                 if (wpipe->pipe_state & PIPE_WANTR) {
 676                         wpipe->pipe_state &= ~PIPE_WANTR;
 677                         wakeup(wpipe);
 678                 }
 679                 wpipe->pipe_state |= PIPE_WANTW;
 680                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdww", 0);
 681                 if (error)
 682                         goto error1;
 683                 if (wpipe->pipe_state & PIPE_EOF) {
 684                         error = EPIPE;
 685                         goto error1;
 686                 }
 687         }
 688         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 689         if (wpipe->pipe_buffer.cnt > 0) {
 690                 if (wpipe->pipe_state & PIPE_WANTR) {
 691                         wpipe->pipe_state &= ~PIPE_WANTR;
 692                         wakeup(wpipe);
 693                 }
 694
 695                 wpipe->pipe_state |= PIPE_WANTW;
 696                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwc", 0);
 697                 if (error)
 698                         goto error1;
 699                 if (wpipe->pipe_state & PIPE_EOF) {
 700                         error = EPIPE;
 701                         goto error1;
 702                 }
 703                 goto retry;
 704         }
 705
 706         wpipe->pipe_state |= PIPE_DIRECTW;
 707
 708         error = pipe_build_write_buffer(wpipe, uio);
 709         if (error) {
 710                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 711                 goto error1;
 712         }
 713
 714         error = 0;
 715         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 716                 if (wpipe->pipe_state & PIPE_EOF) {
 717                         pipelock(wpipe, 0);
 718                         pipe_destroy_write_buffer(wpipe);
 719                         pipeunlock(wpipe);
 720                         pipeselwakeup(wpipe);
 721                         error = EPIPE;
 722                         goto error1;
 723                 }
 724                 if (wpipe->pipe_state & PIPE_WANTR) {
 725                         wpipe->pipe_state &= ~PIPE_WANTR;
 726                         wakeup(wpipe);
 727                 }
 728                 pipeselwakeup(wpipe);
 729                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwt", 0);
 730         }
 731
 732         pipelock(wpipe,0);
 733         if (wpipe->pipe_state & PIPE_DIRECTW) {
 734                 /*
 735                  * this bit of trickery substitutes a kernel buffer for
 736                  * the process that might be going away.
 737                  */
 738                 pipe_clone_write_buffer(wpipe);
 739         } else {
 740                 pipe_destroy_write_buffer(wpipe);
 741         }
 742         pipeunlock(wpipe);
 743         return (error);
 744
 745 error1:
 746         wakeup(wpipe);
 747         return (error);
 748 }
 749 #endif
 750
 751 static int
 752 pipe_write(fp, uio, cred, flags, p)
 753         struct file *fp;
 754         struct uio *uio;
 755         struct ucred *cred;
 756         struct proc *p;
 757         int flags;
 758 {
 759         int error = 0;
 760         int orig_resid;
 761         struct pipe *wpipe, *rpipe;
 762
 763         rpipe = (struct pipe *) fp->f_data;
 764         wpipe = rpipe->pipe_peer;
 765
 766         /*
 767          * detect loss of pipe read side, issue SIGPIPE if lost.
 768          */
 769         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 770                 return (EPIPE);
 771         }
 772         ++wpipe->pipe_busy;
 773
 774         /*
 775          * If it is advantageous to resize the pipe buffer, do
 776          * so.
 777          */
 778         if ((uio->uio_resid > PIPE_SIZE) &&
 779                 (nbigpipe < LIMITBIGPIPES) &&
 780                 (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 781                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 782                 (wpipe->pipe_buffer.cnt == 0)) {
 783
 784                 if ((error = pipelock(wpipe,1)) == 0) {
 785                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 786                                 nbigpipe++;
 787                         pipeunlock(wpipe);
 788                 }
 789         }
 790
 791         /*
 792          * If an early error occured unbusy and return, waking up any pending
 793          * readers.
 794          */
 795         if (error) {
 796                 --wpipe->pipe_busy;
 797                 if ((wpipe->pipe_busy == 0) &&
 798                     (wpipe->pipe_state & PIPE_WANT)) {
 799                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 800                         wakeup(wpipe);
 801                 }
 802                 return(error);
 803         }
 804
 805         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 806
 807         orig_resid = uio->uio_resid;
 808
 809         while (uio->uio_resid) {
 810                 int space;
 811
 812 #ifndef PIPE_NODIRECT
 813                 /*
 814                  * If the transfer is large, we can gain performance if
 815                  * we do process-to-process copies directly.
 816                  * If the write is non-blocking, we don't use the
 817                  * direct write mechanism.
 818                  *
 819                  * The direct write mechanism will detect the reader going
 820                  * away on us.
 821                  */
 822                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
 823                     (fp->f_flag & FNONBLOCK) == 0 &&
 824                         (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) &&
 825                         (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) {
 826                         error = pipe_direct_write( wpipe, uio);
 827                         if (error)
 828                                 break;
 829                         continue;
 830                 }
 831 #endif
 832
 833                 /*
 834                  * Pipe buffered writes cannot be coincidental with
 835                  * direct writes.  We wait until the currently executing
 836                  * direct write is completed before we start filling the
 837                  * pipe buffer.  We break out if a signal occurs or the
 838                  * reader goes away.
 839                  */
 840         retrywrite:
 841                 while (wpipe->pipe_state & PIPE_DIRECTW) {
 842                         if (wpipe->pipe_state & PIPE_WANTR) {
 843                                 wpipe->pipe_state &= ~PIPE_WANTR;
 844                                 wakeup(wpipe);
 845                         }
 846                         error = tsleep(wpipe, PRIBIO | PCATCH, "pipbww", 0);
 847                         if (wpipe->pipe_state & PIPE_EOF)
 848                                 break;
 849                         if (error)
 850                                 break;
 851                 }
 852                 if (wpipe->pipe_state & PIPE_EOF) {
 853                         error = EPIPE;
 854                         break;
 855                 }
 856
 857                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 858
 859                 /* Writes of size <= PIPE_BUF must be atomic. */
 860                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 861                         space = 0;
 862
 863                 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
 864                         if ((error = pipelock(wpipe,1)) == 0) {
 865                                 int size;       /* Transfer size */
 866                                 int segsize;    /* first segment to transfer */
 867
 868                                 /*
 869                                  * It is possible for a direct write to
 870                                  * slip in on us... handle it here...
 871                                  */
 872                                 if (wpipe->pipe_state & PIPE_DIRECTW) {
 873                                         pipeunlock(wpipe);
 874                                         goto retrywrite;
 875                                 }
 876                                 /*
 877                                  * If a process blocked in uiomove, our
 878                                  * value for space might be bad.
 879                                  *
 880                                  * XXX will we be ok if the reader has gone
 881                                  * away here?
 882                                  */
 883                                 if (space > wpipe->pipe_buffer.size -
 884                                     wpipe->pipe_buffer.cnt) {
 885                                         pipeunlock(wpipe);
 886                                         goto retrywrite;
 887                                 }
 888
 889                                 /*
 890                                  * Transfer size is minimum of uio transfer
 891                                  * and free space in pipe buffer.
 892                                  */
 893                                 if (space > uio->uio_resid)
 894                                         size = uio->uio_resid;
 895                                 else
 896                                         size = space;
 897                                 /*
 898                                  * First segment to transfer is minimum of
 899                                  * transfer size and contiguous space in
 900                                  * pipe buffer.  If first segment to transfer
 901                                  * is less than the transfer size, we've got
 902                                  * a wraparound in the buffer.
 903                                  */
 904                                 segsize = wpipe->pipe_buffer.size -
 905                                         wpipe->pipe_buffer.in;
 906                                 if (segsize > size)
 907                                         segsize = size;
 908
 909                                 /* Transfer first segment */
 910
 911                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 912                                                 segsize, uio);
 913
 914                                 if (error == 0 && segsize < size) {
 915                                         /*
 916                                          * Transfer remaining part now, to
 917                                          * support atomic writes.  Wraparound
 918                                          * happened.
 919                                          */
 920                                         if (wpipe->pipe_buffer.in + segsize !=
 921                                             wpipe->pipe_buffer.size)
 922                                                 panic("Expected pipe buffer wraparound disappeared");
 923
 924                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
 925                                                         size - segsize, uio);
 926                                 }
 927                                 if (error == 0) {
 928                                         wpipe->pipe_buffer.in += size;
 929                                         if (wpipe->pipe_buffer.in >=
 930                                             wpipe->pipe_buffer.size) {
 931                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
 932                                                         panic("Expected wraparound bad");
 933                                                 wpipe->pipe_buffer.in = size - segsize;
 934                                         }
 935
 936                                         wpipe->pipe_buffer.cnt += size;
 937                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
 938                                                 panic("Pipe buffer overflow");
 939
 940                                 }
 941                                 pipeunlock(wpipe);
 942                         }
 943                         if (error)
 944                                 break;
 945
 946                 } else {
 947                         /*
 948                          * If the "read-side" has been blocked, wake it up now.
 949                          */
 950                         if (wpipe->pipe_state & PIPE_WANTR) {
 951                                 wpipe->pipe_state &= ~PIPE_WANTR;
 952                                 wakeup(wpipe);
 953                         }
 954
 955                         /*
 956                          * don't block on non-blocking I/O
 957                          */
 958                         if (fp->f_flag & FNONBLOCK) {
 959                                 error = EAGAIN;
 960                                 break;
 961                         }
 962
 963                         /*
 964                          * We have no more space and have something to offer,
 965                          * wake up select/poll.
 966                          */
 967                         pipeselwakeup(wpipe);
 968
 969                         wpipe->pipe_state |= PIPE_WANTW;
 970                         error = tsleep(wpipe, PRIBIO | PCATCH, "pipewr", 0);
 971                         if (error != 0)
 972                                 break;
 973                         /*
 974                          * If read side wants to go away, we just issue a signal
 975                          * to ourselves.
 976                          */
 977                         if (wpipe->pipe_state & PIPE_EOF) {
 978                                 error = EPIPE;
 979                                 break;
 980                         }
 981                 }
 982         }
 983
 984         --wpipe->pipe_busy;
 985
 986         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
 987                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 988                 wakeup(wpipe);
 989         } else if (wpipe->pipe_buffer.cnt > 0) {
 990                 /*
 991                  * If we have put any characters in the buffer, we wake up
 992                  * the reader.
 993                  */
 994                 if (wpipe->pipe_state & PIPE_WANTR) {
 995                         wpipe->pipe_state &= ~PIPE_WANTR;
 996                         wakeup(wpipe);
 997                 }
 998         }
 999
1000         /*
1001          * Don't return EPIPE if I/O was successful
1002          */
1003         if ((wpipe->pipe_buffer.cnt == 0) &&
1004             (uio->uio_resid == 0) &&
1005             (error == EPIPE)) {
1006                 error = 0;
1007         }
1008
1009         if (error == 0)
1010                 vfs_timestamp(&wpipe->pipe_mtime);
1011
1012         /*
1013          * We have something to offer,
1014          * wake up select/poll.
1015          */
1016         if (wpipe->pipe_buffer.cnt)
1017                 pipeselwakeup(wpipe);
1018
1019         return (error);
1020 }
1021
1022 /*
1023  * we implement a very minimal set of ioctls for compatibility with sockets.
1024  */
1025 int
1026 pipe_ioctl(fp, cmd, data, p)
1027         struct file *fp;
1028         u_long cmd;
1029         caddr_t data;
1030         struct proc *p;
1031 {
1032         struct pipe *mpipe = (struct pipe *)fp->f_data;
1033
1034         switch (cmd) {
1035
1036         case FIONBIO:
1037                 return (0);
1038
1039         case FIOASYNC:
1040                 if (*(int *)data) {
1041                         mpipe->pipe_state |= PIPE_ASYNC;
1042                 } else {
1043                         mpipe->pipe_state &= ~PIPE_ASYNC;
1044                 }
1045                 return (0);
1046
1047         case FIONREAD:
1048                 if (mpipe->pipe_state & PIPE_DIRECTW)
1049                         *(int *)data = mpipe->pipe_map.cnt;
1050                 else
1051                         *(int *)data = mpipe->pipe_buffer.cnt;
1052                 return (0);
1053
1054         case FIOSETOWN:
1055                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1056
1057         case FIOGETOWN:
1058                 *(int *)data = fgetown(mpipe->pipe_sigio);
1059                 return (0);
1060
1061         /* This is deprecated, FIOSETOWN should be used instead. */
1062         case TIOCSPGRP:
1063                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1064
1065         /* This is deprecated, FIOGETOWN should be used instead. */
1066         case TIOCGPGRP:
1067                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1068                 return (0);
1069
1070         }
1071         return (ENOTTY);
1072 }
1073
1074 int
1075 pipe_poll(fp, events, cred, p)
1076         struct file *fp;
1077         int events;
1078         struct ucred *cred;
1079         struct proc *p;
1080 {
1081         struct pipe *rpipe = (struct pipe *)fp->f_data;
1082         struct pipe *wpipe;
1083         int revents = 0;
1084
1085         wpipe = rpipe->pipe_peer;
1086         if (events & (POLLIN | POLLRDNORM))
1087                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1088                     (rpipe->pipe_buffer.cnt > 0) ||
1089                     (rpipe->pipe_state & PIPE_EOF))
1090                         revents |= events & (POLLIN | POLLRDNORM);
1091
1092         if (events & (POLLOUT | POLLWRNORM))
1093                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1094                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1095                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1096                         revents |= events & (POLLOUT | POLLWRNORM);
1097
1098         if ((rpipe->pipe_state & PIPE_EOF) ||
1099             (wpipe == NULL) ||
1100             (wpipe->pipe_state & PIPE_EOF))
1101                 revents |= POLLHUP;
1102
1103         if (revents == 0) {
1104                 if (events & (POLLIN | POLLRDNORM)) {
1105                         selrecord(p->p_thread, &rpipe->pipe_sel);
1106                         rpipe->pipe_state |= PIPE_SEL;
1107                 }
1108
1109                 if (events & (POLLOUT | POLLWRNORM)) {
1110                         selrecord(p->p_thread, &wpipe->pipe_sel);
1111                         wpipe->pipe_state |= PIPE_SEL;
1112                 }
1113         }
1114
1115         return (revents);
1116 }
1117
1118 static int
1119 pipe_stat(fp, ub, p)
1120         struct file *fp;
1121         struct stat *ub;
1122         struct proc *p;
1123 {
1124         struct pipe *pipe = (struct pipe *)fp->f_data;
1125
1126         bzero((caddr_t)ub, sizeof(*ub));
1127         ub->st_mode = S_IFIFO;
1128         ub->st_blksize = pipe->pipe_buffer.size;
1129         ub->st_size = pipe->pipe_buffer.cnt;
1130         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1131         ub->st_atimespec = pipe->pipe_atime;
1132         ub->st_mtimespec = pipe->pipe_mtime;
1133         ub->st_ctimespec = pipe->pipe_ctime;
1134         /*
1135          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1136          * st_flags, st_gen.
1137          * XXX (st_dev, st_ino) should be unique.
1138          */
1139         return (0);
1140 }
1141
1142 /* ARGSUSED */
1143 static int
1144 pipe_close(fp, p)
1145         struct file *fp;
1146         struct proc *p;
1147 {
1148         struct pipe *cpipe = (struct pipe *)fp->f_data;
1149
1150         fp->f_ops = &badfileops;
1151         fp->f_data = NULL;
1152         funsetown(cpipe->pipe_sigio);
1153         pipeclose(cpipe);
1154         return (0);
1155 }
1156
1157 static void
1158 pipe_free_kmem(cpipe)
1159         struct pipe *cpipe;
1160 {
1161
1162         if (cpipe->pipe_buffer.buffer != NULL) {
1163                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1164                         --nbigpipe;
1165                 amountpipekva -= cpipe->pipe_buffer.size;
1166                 kmem_free(kernel_map,
1167                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1168                         cpipe->pipe_buffer.size);
1169                 cpipe->pipe_buffer.buffer = NULL;
1170         }
1171 #ifndef PIPE_NODIRECT
1172         if (cpipe->pipe_map.kva != NULL) {
1173                 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE;
1174                 kmem_free(kernel_map,
1175                         cpipe->pipe_map.kva,
1176                         cpipe->pipe_buffer.size + PAGE_SIZE);
1177                 cpipe->pipe_map.cnt = 0;
1178                 cpipe->pipe_map.kva = 0;
1179                 cpipe->pipe_map.pos = 0;
1180                 cpipe->pipe_map.npages = 0;
1181         }
1182 #endif
1183 }
1184
1185 /*
1186  * shutdown the pipe
1187  */
1188 static void
1189 pipeclose(cpipe)
1190         struct pipe *cpipe;
1191 {
1192         struct pipe *ppipe;
1193
1194         if (cpipe) {
1195
1196                 pipeselwakeup(cpipe);
1197
1198                 /*
1199                  * If the other side is blocked, wake it up saying that
1200                  * we want to close it down.
1201                  */
1202                 while (cpipe->pipe_busy) {
1203                         wakeup(cpipe);
1204                         cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1205                         tsleep(cpipe, PRIBIO, "pipecl", 0);
1206                 }
1207
1208                 /*
1209                  * Disconnect from peer
1210                  */
1211                 if ((ppipe = cpipe->pipe_peer) != NULL) {
1212                         pipeselwakeup(ppipe);
1213
1214                         ppipe->pipe_state |= PIPE_EOF;
1215                         wakeup(ppipe);
1216                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1217                         ppipe->pipe_peer = NULL;
1218                 }
1219                 /*
1220                  * free resources
1221                  */
1222                 pipe_free_kmem(cpipe);
1223                 zfree(pipe_zone, cpipe);
1224         }
1225 }
1226
1227 /*ARGSUSED*/
1228 static int
1229 pipe_kqfilter(struct file *fp, struct knote *kn)
1230 {
1231         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1232
1233         switch (kn->kn_filter) {
1234         case EVFILT_READ:
1235                 kn->kn_fop = &pipe_rfiltops;
1236                 break;
1237         case EVFILT_WRITE:
1238                 kn->kn_fop = &pipe_wfiltops;
1239                 cpipe = cpipe->pipe_peer;
1240                 if (cpipe == NULL)
1241                         /* other end of pipe has been closed */
1242                         return (EBADF);
1243                 break;
1244         default:
1245                 return (1);
1246         }
1247         kn->kn_hook = (caddr_t)cpipe;
1248
1249         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1250         return (0);
1251 }
1252
1253 static void
1254 filt_pipedetach(struct knote *kn)
1255 {
1256         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1257
1258         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1259 }
1260
1261 /*ARGSUSED*/
1262 static int
1263 filt_piperead(struct knote *kn, long hint)
1264 {
1265         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1266         struct pipe *wpipe = rpipe->pipe_peer;
1267
1268         kn->kn_data = rpipe->pipe_buffer.cnt;
1269         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1270                 kn->kn_data = rpipe->pipe_map.cnt;
1271
1272         if ((rpipe->pipe_state & PIPE_EOF) ||
1273             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1274                 kn->kn_flags |= EV_EOF;
1275                 return (1);
1276         }
1277         return (kn->kn_data > 0);
1278 }
1279
1280 /*ARGSUSED*/
1281 static int
1282 filt_pipewrite(struct knote *kn, long hint)
1283 {
1284         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1285         struct pipe *wpipe = rpipe->pipe_peer;
1286
1287         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1288                 kn->kn_data = 0;
1289                 kn->kn_flags |= EV_EOF;
1290                 return (1);
1291         }
1292         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1293         if (wpipe->pipe_state & PIPE_DIRECTW)
1294                 kn->kn_data = 0;
1295
1296         return (kn->kn_data >= PIPE_BUF);
1297 }