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 1.13 2003/11/03 17:11:21 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 #include <sys/file2.h>
  83
  84 /*
  85  * Use this define if you want to disable *fancy* VM things.  Expect an
  86  * approx 30% decrease in transfer rate.  This could be useful for
  87  * NetBSD or OpenBSD.
  88  */
  89 /* #define PIPE_NODIRECT */
  90
  91 /*
  92  * interfaces to the outside world
  93  */
  94 static int pipe_read (struct file *fp, struct uio *uio,
  95                 struct ucred *cred, int flags, struct thread *td);
  96 static int pipe_write (struct file *fp, struct uio *uio,
  97                 struct ucred *cred, int flags, struct thread *td);
  98 static int pipe_close (struct file *fp, struct thread *td);
  99 static int pipe_poll (struct file *fp, int events, struct ucred *cred,
 100                 struct thread *td);
 101 static int pipe_kqfilter (struct file *fp, struct knote *kn);
 102 static int pipe_stat (struct file *fp, struct stat *sb, struct thread *td);
 103 static int pipe_ioctl (struct file *fp, u_long cmd, caddr_t data, struct thread *td);
 104
 105 static struct fileops pipeops = {
 106         NULL,   /* port */
 107         0,      /* autoq */
 108         pipe_read, pipe_write, pipe_ioctl, pipe_poll, pipe_kqfilter,
 109         pipe_stat, pipe_close
 110 };
 111
 112 static void     filt_pipedetach(struct knote *kn);
 113 static int      filt_piperead(struct knote *kn, long hint);
 114 static int      filt_pipewrite(struct knote *kn, long hint);
 115
 116 static struct filterops pipe_rfiltops =
 117         { 1, NULL, filt_pipedetach, filt_piperead };
 118 static struct filterops pipe_wfiltops =
 119         { 1, NULL, filt_pipedetach, filt_pipewrite };
 120
 121
 122 /*
 123  * Default pipe buffer size(s), this can be kind-of large now because pipe
 124  * space is pageable.  The pipe code will try to maintain locality of
 125  * reference for performance reasons, so small amounts of outstanding I/O
 126  * will not wipe the cache.
 127  */
 128 #define MINPIPESIZE (PIPE_SIZE/3)
 129 #define MAXPIPESIZE (2*PIPE_SIZE/3)
 130
 131 /*
 132  * Maximum amount of kva for pipes -- this is kind-of a soft limit, but
 133  * is there so that on large systems, we don't exhaust it.
 134  */
 135 #define MAXPIPEKVA (8*1024*1024)
 136
 137 /*
 138  * Limit for direct transfers, we cannot, of course limit
 139  * the amount of kva for pipes in general though.
 140  */
 141 #define LIMITPIPEKVA (16*1024*1024)
 142
 143 /*
 144  * Limit the number of "big" pipes
 145  */
 146 #define LIMITBIGPIPES   32
 147 static int nbigpipe;
 148
 149 static int amountpipekva;
 150
 151 static void pipeclose (struct pipe *cpipe);
 152 static void pipe_free_kmem (struct pipe *cpipe);
 153 static int pipe_create (struct pipe **cpipep);
 154 static __inline int pipelock (struct pipe *cpipe, int catch);
 155 static __inline void pipeunlock (struct pipe *cpipe);
 156 static __inline void pipeselwakeup (struct pipe *cpipe);
 157 #ifndef PIPE_NODIRECT
 158 static int pipe_build_write_buffer (struct pipe *wpipe, struct uio *uio);
 159 static void pipe_destroy_write_buffer (struct pipe *wpipe);
 160 static int pipe_direct_write (struct pipe *wpipe, struct uio *uio);
 161 static void pipe_clone_write_buffer (struct pipe *wpipe);
 162 #endif
 163 static int pipespace (struct pipe *cpipe, int size);
 164
 165 static vm_zone_t pipe_zone;
 166
 167 /*
 168  * The pipe system call for the DTYPE_PIPE type of pipes
 169  *
 170  * pipe_ARgs(int dummy)
 171  */
 172
 173 /* ARGSUSED */
 174 int
 175 pipe(struct pipe_args *uap)
 176 {
 177         struct thread *td = curthread;
 178         struct proc *p = td->td_proc;
 179         struct filedesc *fdp;
 180         struct file *rf, *wf;
 181         struct pipe *rpipe, *wpipe;
 182         int fd1, fd2, error;
 183
 184         KKASSERT(p);
 185         fdp = p->p_fd;
 186
 187         if (pipe_zone == NULL)
 188                 pipe_zone = zinit("PIPE", sizeof(struct pipe), 0, 0, 4);
 189
 190         rpipe = wpipe = NULL;
 191         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 192                 pipeclose(rpipe);
 193                 pipeclose(wpipe);
 194                 return (ENFILE);
 195         }
 196
 197         rpipe->pipe_state |= PIPE_DIRECTOK;
 198         wpipe->pipe_state |= PIPE_DIRECTOK;
 199
 200         error = falloc(p, &rf, &fd1);
 201         if (error) {
 202                 pipeclose(rpipe);
 203                 pipeclose(wpipe);
 204                 return (error);
 205         }
 206         fhold(rf);
 207         uap->sysmsg_fds[0] = fd1;
 208
 209         /*
 210          * Warning: once we've gotten past allocation of the fd for the
 211          * read-side, we can only drop the read side via fdrop() in order
 212          * to avoid races against processes which manage to dup() the read
 213          * side while we are blocked trying to allocate the write side.
 214          */
 215         rf->f_flag = FREAD | FWRITE;
 216         rf->f_type = DTYPE_PIPE;
 217         rf->f_data = (caddr_t)rpipe;
 218         rf->f_ops = &pipeops;
 219         error = falloc(p, &wf, &fd2);
 220         if (error) {
 221                 if (fdp->fd_ofiles[fd1] == rf) {
 222                         fdp->fd_ofiles[fd1] = NULL;
 223                         fdrop(rf, td);
 224                 }
 225                 fdrop(rf, td);
 226                 /* rpipe has been closed by fdrop(). */
 227                 pipeclose(wpipe);
 228                 return (error);
 229         }
 230         wf->f_flag = FREAD | FWRITE;
 231         wf->f_type = DTYPE_PIPE;
 232         wf->f_data = (caddr_t)wpipe;
 233         wf->f_ops = &pipeops;
 234         uap->sysmsg_fds[1] = fd2;
 235
 236         rpipe->pipe_peer = wpipe;
 237         wpipe->pipe_peer = rpipe;
 238         fdrop(rf, td);
 239
 240         return (0);
 241 }
 242
 243 /*
 244  * Allocate kva for pipe circular buffer, the space is pageable
 245  * This routine will 'realloc' the size of a pipe safely, if it fails
 246  * it will retain the old buffer.
 247  * If it fails it will return ENOMEM.
 248  */
 249 static int
 250 pipespace(cpipe, size)
 251         struct pipe *cpipe;
 252         int size;
 253 {
 254         struct vm_object *object;
 255         caddr_t buffer;
 256         int npages, error;
 257
 258         npages = round_page(size)/PAGE_SIZE;
 259         /*
 260          * Create an object, I don't like the idea of paging to/from
 261          * kernel_object.
 262          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 263          */
 264         object = vm_object_allocate(OBJT_DEFAULT, npages);
 265         buffer = (caddr_t) vm_map_min(kernel_map);
 266
 267         /*
 268          * Insert the object into the kernel map, and allocate kva for it.
 269          * The map entry is, by default, pageable.
 270          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 271          */
 272         error = vm_map_find(kernel_map, object, 0,
 273                 (vm_offset_t *) &buffer, size, 1,
 274                 VM_PROT_ALL, VM_PROT_ALL, 0);
 275
 276         if (error != KERN_SUCCESS) {
 277                 vm_object_deallocate(object);
 278                 return (ENOMEM);
 279         }
 280
 281         /* free old resources if we're resizing */
 282         pipe_free_kmem(cpipe);
 283         cpipe->pipe_buffer.object = object;
 284         cpipe->pipe_buffer.buffer = buffer;
 285         cpipe->pipe_buffer.size = size;
 286         cpipe->pipe_buffer.in = 0;
 287         cpipe->pipe_buffer.out = 0;
 288         cpipe->pipe_buffer.cnt = 0;
 289         amountpipekva += cpipe->pipe_buffer.size;
 290         return (0);
 291 }
 292
 293 /*
 294  * initialize and allocate VM and memory for pipe
 295  */
 296 static int
 297 pipe_create(cpipep)
 298         struct pipe **cpipep;
 299 {
 300         struct pipe *cpipe;
 301         int error;
 302
 303         *cpipep = zalloc(pipe_zone);
 304         if (*cpipep == NULL)
 305                 return (ENOMEM);
 306
 307         cpipe = *cpipep;
 308
 309         /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */
 310         cpipe->pipe_buffer.object = NULL;
 311 #ifndef PIPE_NODIRECT
 312         cpipe->pipe_map.kva = NULL;
 313 #endif
 314         /*
 315          * protect so pipeclose() doesn't follow a junk pointer
 316          * if pipespace() fails.
 317          */
 318         bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel));
 319         cpipe->pipe_state = 0;
 320         cpipe->pipe_peer = NULL;
 321         cpipe->pipe_busy = 0;
 322
 323 #ifndef PIPE_NODIRECT
 324         /*
 325          * pipe data structure initializations to support direct pipe I/O
 326          */
 327         cpipe->pipe_map.cnt = 0;
 328         cpipe->pipe_map.kva = 0;
 329         cpipe->pipe_map.pos = 0;
 330         cpipe->pipe_map.npages = 0;
 331         /* cpipe->pipe_map.ms[] = invalid */
 332 #endif
 333
 334         error = pipespace(cpipe, PIPE_SIZE);
 335         if (error)
 336                 return (error);
 337
 338         vfs_timestamp(&cpipe->pipe_ctime);
 339         cpipe->pipe_atime = cpipe->pipe_ctime;
 340         cpipe->pipe_mtime = cpipe->pipe_ctime;
 341
 342         return (0);
 343 }
 344
 345
 346 /*
 347  * lock a pipe for I/O, blocking other access
 348  */
 349 static __inline int
 350 pipelock(cpipe, catch)
 351         struct pipe *cpipe;
 352         int catch;
 353 {
 354         int error;
 355
 356         while (cpipe->pipe_state & PIPE_LOCK) {
 357                 cpipe->pipe_state |= PIPE_LWANT;
 358                 error = tsleep(cpipe, (catch ? PCATCH : 0), "pipelk", 0);
 359                 if (error != 0)
 360                         return (error);
 361         }
 362         cpipe->pipe_state |= PIPE_LOCK;
 363         return (0);
 364 }
 365
 366 /*
 367  * unlock a pipe I/O lock
 368  */
 369 static __inline void
 370 pipeunlock(cpipe)
 371         struct pipe *cpipe;
 372 {
 373
 374         cpipe->pipe_state &= ~PIPE_LOCK;
 375         if (cpipe->pipe_state & PIPE_LWANT) {
 376                 cpipe->pipe_state &= ~PIPE_LWANT;
 377                 wakeup(cpipe);
 378         }
 379 }
 380
 381 static __inline void
 382 pipeselwakeup(cpipe)
 383         struct pipe *cpipe;
 384 {
 385
 386         if (cpipe->pipe_state & PIPE_SEL) {
 387                 cpipe->pipe_state &= ~PIPE_SEL;
 388                 selwakeup(&cpipe->pipe_sel);
 389         }
 390         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 391                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 392         KNOTE(&cpipe->pipe_sel.si_note, 0);
 393 }
 394
 395 /* ARGSUSED */
 396 static int
 397 pipe_read(struct file *fp, struct uio *uio, struct ucred *cred,
 398         int flags, struct thread *td)
 399 {
 400         struct pipe *rpipe = (struct pipe *) fp->f_data;
 401         int error;
 402         int nread = 0;
 403         u_int size;
 404
 405         ++rpipe->pipe_busy;
 406         error = pipelock(rpipe, 1);
 407         if (error)
 408                 goto unlocked_error;
 409
 410         while (uio->uio_resid) {
 411                 /*
 412                  * normal pipe buffer receive
 413                  */
 414                 if (rpipe->pipe_buffer.cnt > 0) {
 415                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 416                         if (size > rpipe->pipe_buffer.cnt)
 417                                 size = rpipe->pipe_buffer.cnt;
 418                         if (size > (u_int) uio->uio_resid)
 419                                 size = (u_int) uio->uio_resid;
 420
 421                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 422                                         size, uio);
 423                         if (error)
 424                                 break;
 425
 426                         rpipe->pipe_buffer.out += size;
 427                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 428                                 rpipe->pipe_buffer.out = 0;
 429
 430                         rpipe->pipe_buffer.cnt -= size;
 431
 432                         /*
 433                          * If there is no more to read in the pipe, reset
 434                          * its pointers to the beginning.  This improves
 435                          * cache hit stats.
 436                          */
 437                         if (rpipe->pipe_buffer.cnt == 0) {
 438                                 rpipe->pipe_buffer.in = 0;
 439                                 rpipe->pipe_buffer.out = 0;
 440                         }
 441                         nread += size;
 442 #ifndef PIPE_NODIRECT
 443                 /*
 444                  * Direct copy, bypassing a kernel buffer.
 445                  */
 446                 } else if ((size = rpipe->pipe_map.cnt) &&
 447                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 448                         caddr_t va;
 449                         if (size > (u_int) uio->uio_resid)
 450                                 size = (u_int) uio->uio_resid;
 451
 452                         va = (caddr_t) rpipe->pipe_map.kva +
 453                             rpipe->pipe_map.pos;
 454                         error = uiomove(va, size, uio);
 455                         if (error)
 456                                 break;
 457                         nread += size;
 458                         rpipe->pipe_map.pos += size;
 459                         rpipe->pipe_map.cnt -= size;
 460                         if (rpipe->pipe_map.cnt == 0) {
 461                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 462                                 wakeup(rpipe);
 463                         }
 464 #endif
 465                 } else {
 466                         /*
 467                          * detect EOF condition
 468                          * read returns 0 on EOF, no need to set error
 469                          */
 470                         if (rpipe->pipe_state & PIPE_EOF)
 471                                 break;
 472
 473                         /*
 474                          * If the "write-side" has been blocked, wake it up now.
 475                          */
 476                         if (rpipe->pipe_state & PIPE_WANTW) {
 477                                 rpipe->pipe_state &= ~PIPE_WANTW;
 478                                 wakeup(rpipe);
 479                         }
 480
 481                         /*
 482                          * Break if some data was read.
 483                          */
 484                         if (nread > 0)
 485                                 break;
 486
 487                         /*
 488                          * Unlock the pipe buffer for our remaining processing.  We
 489                          * will either break out with an error or we will sleep and
 490                          * relock to loop.
 491                          */
 492                         pipeunlock(rpipe);
 493
 494                         /*
 495                          * Handle non-blocking mode operation or
 496                          * wait for more data.
 497                          */
 498                         if (fp->f_flag & FNONBLOCK) {
 499                                 error = EAGAIN;
 500                         } else {
 501                                 rpipe->pipe_state |= PIPE_WANTR;
 502                                 if ((error = tsleep(rpipe, PCATCH,
 503                                     "piperd", 0)) == 0) {
 504                                         error = pipelock(rpipe, 1);
 505                                 }
 506                         }
 507                         if (error)
 508                                 goto unlocked_error;
 509                 }
 510         }
 511         pipeunlock(rpipe);
 512
 513         if (error == 0)
 514                 vfs_timestamp(&rpipe->pipe_atime);
 515 unlocked_error:
 516         --rpipe->pipe_busy;
 517
 518         /*
 519          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 520          */
 521         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 522                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 523                 wakeup(rpipe);
 524         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 525                 /*
 526                  * Handle write blocking hysteresis.
 527                  */
 528                 if (rpipe->pipe_state & PIPE_WANTW) {
 529                         rpipe->pipe_state &= ~PIPE_WANTW;
 530                         wakeup(rpipe);
 531                 }
 532         }
 533
 534         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 535                 pipeselwakeup(rpipe);
 536
 537         return (error);
 538 }
 539
 540 #ifndef PIPE_NODIRECT
 541 /*
 542  * Map the sending processes' buffer into kernel space and wire it.
 543  * This is similar to a physical write operation.
 544  */
 545 static int
 546 pipe_build_write_buffer(wpipe, uio)
 547         struct pipe *wpipe;
 548         struct uio *uio;
 549 {
 550         u_int size;
 551         int i;
 552         vm_offset_t addr, endaddr;
 553         vm_paddr_t paddr;
 554
 555         size = (u_int) uio->uio_iov->iov_len;
 556         if (size > wpipe->pipe_buffer.size)
 557                 size = wpipe->pipe_buffer.size;
 558
 559         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 560         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 561         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 562                 vm_page_t m;
 563
 564                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
 565                     (paddr = pmap_kextract(addr)) == 0) {
 566                         int j;
 567
 568                         for (j = 0; j < i; j++)
 569                                 vm_page_unhold(wpipe->pipe_map.ms[j]);
 570                         return (EFAULT);
 571                 }
 572
 573                 m = PHYS_TO_VM_PAGE(paddr);
 574                 vm_page_hold(m);
 575                 wpipe->pipe_map.ms[i] = m;
 576         }
 577
 578 /*
 579  * set up the control block
 580  */
 581         wpipe->pipe_map.npages = i;
 582         wpipe->pipe_map.pos =
 583             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 584         wpipe->pipe_map.cnt = size;
 585
 586 /*
 587  * and map the buffer
 588  */
 589         if (wpipe->pipe_map.kva == 0) {
 590                 /*
 591                  * We need to allocate space for an extra page because the
 592                  * address range might (will) span pages at times.
 593                  */
 594                 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
 595                         wpipe->pipe_buffer.size + PAGE_SIZE);
 596                 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE;
 597         }
 598         pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
 599                 wpipe->pipe_map.npages);
 600
 601 /*
 602  * and update the uio data
 603  */
 604
 605         uio->uio_iov->iov_len -= size;
 606         uio->uio_iov->iov_base += size;
 607         if (uio->uio_iov->iov_len == 0)
 608                 uio->uio_iov++;
 609         uio->uio_resid -= size;
 610         uio->uio_offset += size;
 611         return (0);
 612 }
 613
 614 /*
 615  * unmap and unwire the process buffer
 616  */
 617 static void
 618 pipe_destroy_write_buffer(wpipe)
 619         struct pipe *wpipe;
 620 {
 621         int i;
 622
 623         if (wpipe->pipe_map.kva) {
 624                 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
 625
 626                 if (amountpipekva > MAXPIPEKVA) {
 627                         vm_offset_t kva = wpipe->pipe_map.kva;
 628                         wpipe->pipe_map.kva = 0;
 629                         kmem_free(kernel_map, kva,
 630                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 631                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 632                 }
 633         }
 634         for (i = 0; i < wpipe->pipe_map.npages; i++)
 635                 vm_page_unhold(wpipe->pipe_map.ms[i]);
 636         wpipe->pipe_map.npages = 0;
 637 }
 638
 639 /*
 640  * In the case of a signal, the writing process might go away.  This
 641  * code copies the data into the circular buffer so that the source
 642  * pages can be freed without loss of data.
 643  */
 644 static void
 645 pipe_clone_write_buffer(wpipe)
 646         struct pipe *wpipe;
 647 {
 648         int size;
 649         int pos;
 650
 651         size = wpipe->pipe_map.cnt;
 652         pos = wpipe->pipe_map.pos;
 653         bcopy((caddr_t) wpipe->pipe_map.kva + pos,
 654             (caddr_t) wpipe->pipe_buffer.buffer, size);
 655
 656         wpipe->pipe_buffer.in = size;
 657         wpipe->pipe_buffer.out = 0;
 658         wpipe->pipe_buffer.cnt = size;
 659         wpipe->pipe_state &= ~PIPE_DIRECTW;
 660
 661         pipe_destroy_write_buffer(wpipe);
 662 }
 663
 664 /*
 665  * This implements the pipe buffer write mechanism.  Note that only
 666  * a direct write OR a normal pipe write can be pending at any given time.
 667  * If there are any characters in the pipe buffer, the direct write will
 668  * be deferred until the receiving process grabs all of the bytes from
 669  * the pipe buffer.  Then the direct mapping write is set-up.
 670  */
 671 static int
 672 pipe_direct_write(wpipe, uio)
 673         struct pipe *wpipe;
 674         struct uio *uio;
 675 {
 676         int error;
 677
 678 retry:
 679         while (wpipe->pipe_state & PIPE_DIRECTW) {
 680                 if (wpipe->pipe_state & PIPE_WANTR) {
 681                         wpipe->pipe_state &= ~PIPE_WANTR;
 682                         wakeup(wpipe);
 683                 }
 684                 wpipe->pipe_state |= PIPE_WANTW;
 685                 error = tsleep(wpipe, PCATCH, "pipdww", 0);
 686                 if (error)
 687                         goto error1;
 688                 if (wpipe->pipe_state & PIPE_EOF) {
 689                         error = EPIPE;
 690                         goto error1;
 691                 }
 692         }
 693         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 694         if (wpipe->pipe_buffer.cnt > 0) {
 695                 if (wpipe->pipe_state & PIPE_WANTR) {
 696                         wpipe->pipe_state &= ~PIPE_WANTR;
 697                         wakeup(wpipe);
 698                 }
 699
 700                 wpipe->pipe_state |= PIPE_WANTW;
 701                 error = tsleep(wpipe, PCATCH, "pipdwc", 0);
 702                 if (error)
 703                         goto error1;
 704                 if (wpipe->pipe_state & PIPE_EOF) {
 705                         error = EPIPE;
 706                         goto error1;
 707                 }
 708                 goto retry;
 709         }
 710
 711         wpipe->pipe_state |= PIPE_DIRECTW;
 712
 713         error = pipe_build_write_buffer(wpipe, uio);
 714         if (error) {
 715                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 716                 goto error1;
 717         }
 718
 719         error = 0;
 720         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 721                 if (wpipe->pipe_state & PIPE_EOF) {
 722                         pipelock(wpipe, 0);
 723                         pipe_destroy_write_buffer(wpipe);
 724                         pipeunlock(wpipe);
 725                         pipeselwakeup(wpipe);
 726                         error = EPIPE;
 727                         goto error1;
 728                 }
 729                 if (wpipe->pipe_state & PIPE_WANTR) {
 730                         wpipe->pipe_state &= ~PIPE_WANTR;
 731                         wakeup(wpipe);
 732                 }
 733                 pipeselwakeup(wpipe);
 734                 error = tsleep(wpipe, PCATCH, "pipdwt", 0);
 735         }
 736
 737         pipelock(wpipe,0);
 738         if (wpipe->pipe_state & PIPE_DIRECTW) {
 739                 /*
 740                  * this bit of trickery substitutes a kernel buffer for
 741                  * the process that might be going away.
 742                  */
 743                 pipe_clone_write_buffer(wpipe);
 744         } else {
 745                 pipe_destroy_write_buffer(wpipe);
 746         }
 747         pipeunlock(wpipe);
 748         return (error);
 749
 750 error1:
 751         wakeup(wpipe);
 752         return (error);
 753 }
 754 #endif
 755
 756 static int
 757 pipe_write(struct file *fp, struct uio *uio, struct ucred *cred,
 758         int flags, struct thread *td)
 759 {
 760         int error = 0;
 761         int orig_resid;
 762         struct pipe *wpipe, *rpipe;
 763
 764         rpipe = (struct pipe *) fp->f_data;
 765         wpipe = rpipe->pipe_peer;
 766
 767         /*
 768          * detect loss of pipe read side, issue SIGPIPE if lost.
 769          */
 770         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
 771                 return (EPIPE);
 772         }
 773         ++wpipe->pipe_busy;
 774
 775         /*
 776          * If it is advantageous to resize the pipe buffer, do
 777          * so.
 778          */
 779         if ((uio->uio_resid > PIPE_SIZE) &&
 780                 (nbigpipe < LIMITBIGPIPES) &&
 781                 (wpipe->pipe_state & PIPE_DIRECTW) == 0 &&
 782                 (wpipe->pipe_buffer.size <= PIPE_SIZE) &&
 783                 (wpipe->pipe_buffer.cnt == 0)) {
 784
 785                 if ((error = pipelock(wpipe,1)) == 0) {
 786                         if (pipespace(wpipe, BIG_PIPE_SIZE) == 0)
 787                                 nbigpipe++;
 788                         pipeunlock(wpipe);
 789                 }
 790         }
 791
 792         /*
 793          * If an early error occured unbusy and return, waking up any pending
 794          * readers.
 795          */
 796         if (error) {
 797                 --wpipe->pipe_busy;
 798                 if ((wpipe->pipe_busy == 0) &&
 799                     (wpipe->pipe_state & PIPE_WANT)) {
 800                         wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 801                         wakeup(wpipe);
 802                 }
 803                 return(error);
 804         }
 805
 806         KASSERT(wpipe->pipe_buffer.buffer != NULL, ("pipe buffer gone"));
 807
 808         orig_resid = uio->uio_resid;
 809
 810         while (uio->uio_resid) {
 811                 int space;
 812
 813 #ifndef PIPE_NODIRECT
 814                 /*
 815                  * If the transfer is large, we can gain performance if
 816                  * we do process-to-process copies directly.
 817                  * If the write is non-blocking, we don't use the
 818                  * direct write mechanism.
 819                  *
 820                  * The direct write mechanism will detect the reader going
 821                  * away on us.
 822                  */
 823                 if ((uio->uio_iov->iov_len >= PIPE_MINDIRECT) &&
 824                     (fp->f_flag & FNONBLOCK) == 0 &&
 825                         (wpipe->pipe_map.kva || (amountpipekva < LIMITPIPEKVA)) &&
 826                         (uio->uio_iov->iov_len >= PIPE_MINDIRECT)) {
 827                         error = pipe_direct_write( wpipe, uio);
 828                         if (error)
 829                                 break;
 830                         continue;
 831                 }
 832 #endif
 833
 834                 /*
 835                  * Pipe buffered writes cannot be coincidental with
 836                  * direct writes.  We wait until the currently executing
 837                  * direct write is completed before we start filling the
 838                  * pipe buffer.  We break out if a signal occurs or the
 839                  * reader goes away.
 840                  */
 841         retrywrite:
 842                 while (wpipe->pipe_state & PIPE_DIRECTW) {
 843                         if (wpipe->pipe_state & PIPE_WANTR) {
 844                                 wpipe->pipe_state &= ~PIPE_WANTR;
 845                                 wakeup(wpipe);
 846                         }
 847                         error = tsleep(wpipe, PCATCH, "pipbww", 0);
 848                         if (wpipe->pipe_state & PIPE_EOF)
 849                                 break;
 850                         if (error)
 851                                 break;
 852                 }
 853                 if (wpipe->pipe_state & PIPE_EOF) {
 854                         error = EPIPE;
 855                         break;
 856                 }
 857
 858                 space = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
 859
 860                 /* Writes of size <= PIPE_BUF must be atomic. */
 861                 if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 862                         space = 0;
 863
 864                 if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
 865                         if ((error = pipelock(wpipe,1)) == 0) {
 866                                 int size;       /* Transfer size */
 867                                 int segsize;    /* first segment to transfer */
 868
 869                                 /*
 870                                  * It is possible for a direct write to
 871                                  * slip in on us... handle it here...
 872                                  */
 873                                 if (wpipe->pipe_state & PIPE_DIRECTW) {
 874                                         pipeunlock(wpipe);
 875                                         goto retrywrite;
 876                                 }
 877                                 /*
 878                                  * If a process blocked in uiomove, our
 879                                  * value for space might be bad.
 880                                  *
 881                                  * XXX will we be ok if the reader has gone
 882                                  * away here?
 883                                  */
 884                                 if (space > wpipe->pipe_buffer.size -
 885                                     wpipe->pipe_buffer.cnt) {
 886                                         pipeunlock(wpipe);
 887                                         goto retrywrite;
 888                                 }
 889
 890                                 /*
 891                                  * Transfer size is minimum of uio transfer
 892                                  * and free space in pipe buffer.
 893                                  */
 894                                 if (space > uio->uio_resid)
 895                                         size = uio->uio_resid;
 896                                 else
 897                                         size = space;
 898                                 /*
 899                                  * First segment to transfer is minimum of
 900                                  * transfer size and contiguous space in
 901                                  * pipe buffer.  If first segment to transfer
 902                                  * is less than the transfer size, we've got
 903                                  * a wraparound in the buffer.
 904                                  */
 905                                 segsize = wpipe->pipe_buffer.size -
 906                                         wpipe->pipe_buffer.in;
 907                                 if (segsize > size)
 908                                         segsize = size;
 909
 910                                 /* Transfer first segment */
 911
 912                                 error = uiomove(&wpipe->pipe_buffer.buffer[wpipe->pipe_buffer.in],
 913                                                 segsize, uio);
 914
 915                                 if (error == 0 && segsize < size) {
 916                                         /*
 917                                          * Transfer remaining part now, to
 918                                          * support atomic writes.  Wraparound
 919                                          * happened.
 920                                          */
 921                                         if (wpipe->pipe_buffer.in + segsize !=
 922                                             wpipe->pipe_buffer.size)
 923                                                 panic("Expected pipe buffer wraparound disappeared");
 924
 925                                         error = uiomove(&wpipe->pipe_buffer.buffer[0],
 926                                                         size - segsize, uio);
 927                                 }
 928                                 if (error == 0) {
 929                                         wpipe->pipe_buffer.in += size;
 930                                         if (wpipe->pipe_buffer.in >=
 931                                             wpipe->pipe_buffer.size) {
 932                                                 if (wpipe->pipe_buffer.in != size - segsize + wpipe->pipe_buffer.size)
 933                                                         panic("Expected wraparound bad");
 934                                                 wpipe->pipe_buffer.in = size - segsize;
 935                                         }
 936
 937                                         wpipe->pipe_buffer.cnt += size;
 938                                         if (wpipe->pipe_buffer.cnt > wpipe->pipe_buffer.size)
 939                                                 panic("Pipe buffer overflow");
 940
 941                                 }
 942                                 pipeunlock(wpipe);
 943                         }
 944                         if (error)
 945                                 break;
 946
 947                 } else {
 948                         /*
 949                          * If the "read-side" has been blocked, wake it up now.
 950                          */
 951                         if (wpipe->pipe_state & PIPE_WANTR) {
 952                                 wpipe->pipe_state &= ~PIPE_WANTR;
 953                                 wakeup(wpipe);
 954                         }
 955
 956                         /*
 957                          * don't block on non-blocking I/O
 958                          */
 959                         if (fp->f_flag & FNONBLOCK) {
 960                                 error = EAGAIN;
 961                                 break;
 962                         }
 963
 964                         /*
 965                          * We have no more space and have something to offer,
 966                          * wake up select/poll.
 967                          */
 968                         pipeselwakeup(wpipe);
 969
 970                         wpipe->pipe_state |= PIPE_WANTW;
 971                         error = tsleep(wpipe, PCATCH, "pipewr", 0);
 972                         if (error != 0)
 973                                 break;
 974                         /*
 975                          * If read side wants to go away, we just issue a signal
 976                          * to ourselves.
 977                          */
 978                         if (wpipe->pipe_state & PIPE_EOF) {
 979                                 error = EPIPE;
 980                                 break;
 981                         }
 982                 }
 983         }
 984
 985         --wpipe->pipe_busy;
 986
 987         if ((wpipe->pipe_busy == 0) && (wpipe->pipe_state & PIPE_WANT)) {
 988                 wpipe->pipe_state &= ~(PIPE_WANT | PIPE_WANTR);
 989                 wakeup(wpipe);
 990         } else if (wpipe->pipe_buffer.cnt > 0) {
 991                 /*
 992                  * If we have put any characters in the buffer, we wake up
 993                  * the reader.
 994                  */
 995                 if (wpipe->pipe_state & PIPE_WANTR) {
 996                         wpipe->pipe_state &= ~PIPE_WANTR;
 997                         wakeup(wpipe);
 998                 }
 999         }
1000
1001         /*
1002          * Don't return EPIPE if I/O was successful
1003          */
1004         if ((wpipe->pipe_buffer.cnt == 0) &&
1005             (uio->uio_resid == 0) &&
1006             (error == EPIPE)) {
1007                 error = 0;
1008         }
1009
1010         if (error == 0)
1011                 vfs_timestamp(&wpipe->pipe_mtime);
1012
1013         /*
1014          * We have something to offer,
1015          * wake up select/poll.
1016          */
1017         if (wpipe->pipe_buffer.cnt)
1018                 pipeselwakeup(wpipe);
1019
1020         return (error);
1021 }
1022
1023 /*
1024  * we implement a very minimal set of ioctls for compatibility with sockets.
1025  */
1026 int
1027 pipe_ioctl(struct file *fp, u_long cmd, caddr_t data, struct thread *td)
1028 {
1029         struct pipe *mpipe = (struct pipe *)fp->f_data;
1030
1031         switch (cmd) {
1032
1033         case FIONBIO:
1034                 return (0);
1035
1036         case FIOASYNC:
1037                 if (*(int *)data) {
1038                         mpipe->pipe_state |= PIPE_ASYNC;
1039                 } else {
1040                         mpipe->pipe_state &= ~PIPE_ASYNC;
1041                 }
1042                 return (0);
1043
1044         case FIONREAD:
1045                 if (mpipe->pipe_state & PIPE_DIRECTW)
1046                         *(int *)data = mpipe->pipe_map.cnt;
1047                 else
1048                         *(int *)data = mpipe->pipe_buffer.cnt;
1049                 return (0);
1050
1051         case FIOSETOWN:
1052                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1053
1054         case FIOGETOWN:
1055                 *(int *)data = fgetown(mpipe->pipe_sigio);
1056                 return (0);
1057
1058         /* This is deprecated, FIOSETOWN should be used instead. */
1059         case TIOCSPGRP:
1060                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1061
1062         /* This is deprecated, FIOGETOWN should be used instead. */
1063         case TIOCGPGRP:
1064                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1065                 return (0);
1066
1067         }
1068         return (ENOTTY);
1069 }
1070
1071 int
1072 pipe_poll(struct file *fp, int events, struct ucred *cred, struct thread *td)
1073 {
1074         struct pipe *rpipe = (struct pipe *)fp->f_data;
1075         struct pipe *wpipe;
1076         int revents = 0;
1077
1078         wpipe = rpipe->pipe_peer;
1079         if (events & (POLLIN | POLLRDNORM))
1080                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1081                     (rpipe->pipe_buffer.cnt > 0) ||
1082                     (rpipe->pipe_state & PIPE_EOF))
1083                         revents |= events & (POLLIN | POLLRDNORM);
1084
1085         if (events & (POLLOUT | POLLWRNORM))
1086                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1087                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1088                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1089                         revents |= events & (POLLOUT | POLLWRNORM);
1090
1091         if ((rpipe->pipe_state & PIPE_EOF) ||
1092             (wpipe == NULL) ||
1093             (wpipe->pipe_state & PIPE_EOF))
1094                 revents |= POLLHUP;
1095
1096         if (revents == 0) {
1097                 if (events & (POLLIN | POLLRDNORM)) {
1098                         selrecord(td, &rpipe->pipe_sel);
1099                         rpipe->pipe_state |= PIPE_SEL;
1100                 }
1101
1102                 if (events & (POLLOUT | POLLWRNORM)) {
1103                         selrecord(td, &wpipe->pipe_sel);
1104                         wpipe->pipe_state |= PIPE_SEL;
1105                 }
1106         }
1107
1108         return (revents);
1109 }
1110
1111 static int
1112 pipe_stat(struct file *fp, struct stat *ub, struct thread *td)
1113 {
1114         struct pipe *pipe = (struct pipe *)fp->f_data;
1115
1116         bzero((caddr_t)ub, sizeof(*ub));
1117         ub->st_mode = S_IFIFO;
1118         ub->st_blksize = pipe->pipe_buffer.size;
1119         ub->st_size = pipe->pipe_buffer.cnt;
1120         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1121         ub->st_atimespec = pipe->pipe_atime;
1122         ub->st_mtimespec = pipe->pipe_mtime;
1123         ub->st_ctimespec = pipe->pipe_ctime;
1124         /*
1125          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1126          * st_flags, st_gen.
1127          * XXX (st_dev, st_ino) should be unique.
1128          */
1129         return (0);
1130 }
1131
1132 /* ARGSUSED */
1133 static int
1134 pipe_close(struct file *fp, struct thread *td)
1135 {
1136         struct pipe *cpipe = (struct pipe *)fp->f_data;
1137
1138         fp->f_ops = &badfileops;
1139         fp->f_data = NULL;
1140         funsetown(cpipe->pipe_sigio);
1141         pipeclose(cpipe);
1142         return (0);
1143 }
1144
1145 static void
1146 pipe_free_kmem(struct pipe *cpipe)
1147 {
1148
1149         if (cpipe->pipe_buffer.buffer != NULL) {
1150                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1151                         --nbigpipe;
1152                 amountpipekva -= cpipe->pipe_buffer.size;
1153                 kmem_free(kernel_map,
1154                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1155                         cpipe->pipe_buffer.size);
1156                 cpipe->pipe_buffer.buffer = NULL;
1157         }
1158 #ifndef PIPE_NODIRECT
1159         if (cpipe->pipe_map.kva != NULL) {
1160                 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE;
1161                 kmem_free(kernel_map,
1162                         cpipe->pipe_map.kva,
1163                         cpipe->pipe_buffer.size + PAGE_SIZE);
1164                 cpipe->pipe_map.cnt = 0;
1165                 cpipe->pipe_map.kva = 0;
1166                 cpipe->pipe_map.pos = 0;
1167                 cpipe->pipe_map.npages = 0;
1168         }
1169 #endif
1170 }
1171
1172 /*
1173  * shutdown the pipe
1174  */
1175 static void
1176 pipeclose(struct pipe *cpipe)
1177 {
1178         struct pipe *ppipe;
1179
1180         if (cpipe) {
1181
1182                 pipeselwakeup(cpipe);
1183
1184                 /*
1185                  * If the other side is blocked, wake it up saying that
1186                  * we want to close it down.
1187                  */
1188                 while (cpipe->pipe_busy) {
1189                         wakeup(cpipe);
1190                         cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1191                         tsleep(cpipe, 0, "pipecl", 0);
1192                 }
1193
1194                 /*
1195                  * Disconnect from peer
1196                  */
1197                 if ((ppipe = cpipe->pipe_peer) != NULL) {
1198                         pipeselwakeup(ppipe);
1199
1200                         ppipe->pipe_state |= PIPE_EOF;
1201                         wakeup(ppipe);
1202                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1203                         ppipe->pipe_peer = NULL;
1204                 }
1205                 /*
1206                  * free resources
1207                  */
1208                 pipe_free_kmem(cpipe);
1209                 zfree(pipe_zone, cpipe);
1210         }
1211 }
1212
1213 /*ARGSUSED*/
1214 static int
1215 pipe_kqfilter(struct file *fp, struct knote *kn)
1216 {
1217         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1218
1219         switch (kn->kn_filter) {
1220         case EVFILT_READ:
1221                 kn->kn_fop = &pipe_rfiltops;
1222                 break;
1223         case EVFILT_WRITE:
1224                 kn->kn_fop = &pipe_wfiltops;
1225                 cpipe = cpipe->pipe_peer;
1226                 if (cpipe == NULL)
1227                         /* other end of pipe has been closed */
1228                         return (EPIPE);
1229                 break;
1230         default:
1231                 return (1);
1232         }
1233         kn->kn_hook = (caddr_t)cpipe;
1234
1235         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1236         return (0);
1237 }
1238
1239 static void
1240 filt_pipedetach(struct knote *kn)
1241 {
1242         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1243
1244         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1245 }
1246
1247 /*ARGSUSED*/
1248 static int
1249 filt_piperead(struct knote *kn, long hint)
1250 {
1251         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1252         struct pipe *wpipe = rpipe->pipe_peer;
1253
1254         kn->kn_data = rpipe->pipe_buffer.cnt;
1255         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1256                 kn->kn_data = rpipe->pipe_map.cnt;
1257
1258         if ((rpipe->pipe_state & PIPE_EOF) ||
1259             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1260                 kn->kn_flags |= EV_EOF;
1261                 return (1);
1262         }
1263         return (kn->kn_data > 0);
1264 }
1265
1266 /*ARGSUSED*/
1267 static int
1268 filt_pipewrite(struct knote *kn, long hint)
1269 {
1270         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1271         struct pipe *wpipe = rpipe->pipe_peer;
1272
1273         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1274                 kn->kn_data = 0;
1275                 kn->kn_flags |= EV_EOF;
1276                 return (1);
1277         }
1278         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1279         if (wpipe->pipe_state & PIPE_DIRECTW)
1280                 kn->kn_data = 0;
1281
1282         return (kn->kn_data >= PIPE_BUF);
1283 }