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 1.2 2003/06/17 04:28:41 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
 167 /* ARGSUSED */
 168 int
 169 pipe(p, uap)
 170         struct proc *p;
 171         struct pipe_args /* {
 172                 int     dummy;
 173         } */ *uap;
 174 {
 175         struct filedesc *fdp = p->p_fd;
 176         struct file *rf, *wf;
 177         struct pipe *rpipe, *wpipe;
 178         int fd, error;
 179
 180         if (pipe_zone == NULL)
 181                 pipe_zone = zinit("PIPE", sizeof(struct pipe), 0, 0, 4);
 182
 183         rpipe = wpipe = NULL;
 184         if (pipe_create(&rpipe) || pipe_create(&wpipe)) {
 185                 pipeclose(rpipe);
 186                 pipeclose(wpipe);
 187                 return (ENFILE);
 188         }
 189
 190         rpipe->pipe_state |= PIPE_DIRECTOK;
 191         wpipe->pipe_state |= PIPE_DIRECTOK;
 192
 193         error = falloc(p, &rf, &fd);
 194         if (error) {
 195                 pipeclose(rpipe);
 196                 pipeclose(wpipe);
 197                 return (error);
 198         }
 199         fhold(rf);
 200         p->p_retval[0] = fd;
 201
 202         /*
 203          * Warning: once we've gotten past allocation of the fd for the
 204          * read-side, we can only drop the read side via fdrop() in order
 205          * to avoid races against processes which manage to dup() the read
 206          * side while we are blocked trying to allocate the write side.
 207          */
 208         rf->f_flag = FREAD | FWRITE;
 209         rf->f_type = DTYPE_PIPE;
 210         rf->f_data = (caddr_t)rpipe;
 211         rf->f_ops = &pipeops;
 212         error = falloc(p, &wf, &fd);
 213         if (error) {
 214                 if (fdp->fd_ofiles[p->p_retval[0]] == rf) {
 215                         fdp->fd_ofiles[p->p_retval[0]] = NULL;
 216                         fdrop(rf, p);
 217                 }
 218                 fdrop(rf, p);
 219                 /* rpipe has been closed by fdrop(). */
 220                 pipeclose(wpipe);
 221                 return (error);
 222         }
 223         wf->f_flag = FREAD | FWRITE;
 224         wf->f_type = DTYPE_PIPE;
 225         wf->f_data = (caddr_t)wpipe;
 226         wf->f_ops = &pipeops;
 227         p->p_retval[1] = fd;
 228
 229         rpipe->pipe_peer = wpipe;
 230         wpipe->pipe_peer = rpipe;
 231         fdrop(rf, p);
 232
 233         return (0);
 234 }
 235
 236 /*
 237  * Allocate kva for pipe circular buffer, the space is pageable
 238  * This routine will 'realloc' the size of a pipe safely, if it fails
 239  * it will retain the old buffer.
 240  * If it fails it will return ENOMEM.
 241  */
 242 static int
 243 pipespace(cpipe, size)
 244         struct pipe *cpipe;
 245         int size;
 246 {
 247         struct vm_object *object;
 248         caddr_t buffer;
 249         int npages, error;
 250
 251         npages = round_page(size)/PAGE_SIZE;
 252         /*
 253          * Create an object, I don't like the idea of paging to/from
 254          * kernel_object.
 255          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 256          */
 257         object = vm_object_allocate(OBJT_DEFAULT, npages);
 258         buffer = (caddr_t) vm_map_min(kernel_map);
 259
 260         /*
 261          * Insert the object into the kernel map, and allocate kva for it.
 262          * The map entry is, by default, pageable.
 263          * XXX -- minor change needed here for NetBSD/OpenBSD VM systems.
 264          */
 265         error = vm_map_find(kernel_map, object, 0,
 266                 (vm_offset_t *) &buffer, size, 1,
 267                 VM_PROT_ALL, VM_PROT_ALL, 0);
 268
 269         if (error != KERN_SUCCESS) {
 270                 vm_object_deallocate(object);
 271                 return (ENOMEM);
 272         }
 273
 274         /* free old resources if we're resizing */
 275         pipe_free_kmem(cpipe);
 276         cpipe->pipe_buffer.object = object;
 277         cpipe->pipe_buffer.buffer = buffer;
 278         cpipe->pipe_buffer.size = size;
 279         cpipe->pipe_buffer.in = 0;
 280         cpipe->pipe_buffer.out = 0;
 281         cpipe->pipe_buffer.cnt = 0;
 282         amountpipekva += cpipe->pipe_buffer.size;
 283         return (0);
 284 }
 285
 286 /*
 287  * initialize and allocate VM and memory for pipe
 288  */
 289 static int
 290 pipe_create(cpipep)
 291         struct pipe **cpipep;
 292 {
 293         struct pipe *cpipe;
 294         int error;
 295
 296         *cpipep = zalloc(pipe_zone);
 297         if (*cpipep == NULL)
 298                 return (ENOMEM);
 299
 300         cpipe = *cpipep;
 301
 302         /* so pipespace()->pipe_free_kmem() doesn't follow junk pointer */
 303         cpipe->pipe_buffer.object = NULL;
 304 #ifndef PIPE_NODIRECT
 305         cpipe->pipe_map.kva = NULL;
 306 #endif
 307         /*
 308          * protect so pipeclose() doesn't follow a junk pointer
 309          * if pipespace() fails.
 310          */
 311         bzero(&cpipe->pipe_sel, sizeof(cpipe->pipe_sel));
 312         cpipe->pipe_state = 0;
 313         cpipe->pipe_peer = NULL;
 314         cpipe->pipe_busy = 0;
 315
 316 #ifndef PIPE_NODIRECT
 317         /*
 318          * pipe data structure initializations to support direct pipe I/O
 319          */
 320         cpipe->pipe_map.cnt = 0;
 321         cpipe->pipe_map.kva = 0;
 322         cpipe->pipe_map.pos = 0;
 323         cpipe->pipe_map.npages = 0;
 324         /* cpipe->pipe_map.ms[] = invalid */
 325 #endif
 326
 327         error = pipespace(cpipe, PIPE_SIZE);
 328         if (error)
 329                 return (error);
 330
 331         vfs_timestamp(&cpipe->pipe_ctime);
 332         cpipe->pipe_atime = cpipe->pipe_ctime;
 333         cpipe->pipe_mtime = cpipe->pipe_ctime;
 334
 335         return (0);
 336 }
 337
 338
 339 /*
 340  * lock a pipe for I/O, blocking other access
 341  */
 342 static __inline int
 343 pipelock(cpipe, catch)
 344         struct pipe *cpipe;
 345         int catch;
 346 {
 347         int error;
 348
 349         while (cpipe->pipe_state & PIPE_LOCK) {
 350                 cpipe->pipe_state |= PIPE_LWANT;
 351                 error = tsleep(cpipe, catch ? (PRIBIO | PCATCH) : PRIBIO,
 352                     "pipelk", 0);
 353                 if (error != 0)
 354                         return (error);
 355         }
 356         cpipe->pipe_state |= PIPE_LOCK;
 357         return (0);
 358 }
 359
 360 /*
 361  * unlock a pipe I/O lock
 362  */
 363 static __inline void
 364 pipeunlock(cpipe)
 365         struct pipe *cpipe;
 366 {
 367
 368         cpipe->pipe_state &= ~PIPE_LOCK;
 369         if (cpipe->pipe_state & PIPE_LWANT) {
 370                 cpipe->pipe_state &= ~PIPE_LWANT;
 371                 wakeup(cpipe);
 372         }
 373 }
 374
 375 static __inline void
 376 pipeselwakeup(cpipe)
 377         struct pipe *cpipe;
 378 {
 379
 380         if (cpipe->pipe_state & PIPE_SEL) {
 381                 cpipe->pipe_state &= ~PIPE_SEL;
 382                 selwakeup(&cpipe->pipe_sel);
 383         }
 384         if ((cpipe->pipe_state & PIPE_ASYNC) && cpipe->pipe_sigio)
 385                 pgsigio(cpipe->pipe_sigio, SIGIO, 0);
 386         KNOTE(&cpipe->pipe_sel.si_note, 0);
 387 }
 388
 389 /* ARGSUSED */
 390 static int
 391 pipe_read(fp, uio, cred, flags, p)
 392         struct file *fp;
 393         struct uio *uio;
 394         struct ucred *cred;
 395         struct proc *p;
 396         int flags;
 397 {
 398         struct pipe *rpipe = (struct pipe *) fp->f_data;
 399         int error;
 400         int nread = 0;
 401         u_int size;
 402
 403         ++rpipe->pipe_busy;
 404         error = pipelock(rpipe, 1);
 405         if (error)
 406                 goto unlocked_error;
 407
 408         while (uio->uio_resid) {
 409                 /*
 410                  * normal pipe buffer receive
 411                  */
 412                 if (rpipe->pipe_buffer.cnt > 0) {
 413                         size = rpipe->pipe_buffer.size - rpipe->pipe_buffer.out;
 414                         if (size > rpipe->pipe_buffer.cnt)
 415                                 size = rpipe->pipe_buffer.cnt;
 416                         if (size > (u_int) uio->uio_resid)
 417                                 size = (u_int) uio->uio_resid;
 418
 419                         error = uiomove(&rpipe->pipe_buffer.buffer[rpipe->pipe_buffer.out],
 420                                         size, uio);
 421                         if (error)
 422                                 break;
 423
 424                         rpipe->pipe_buffer.out += size;
 425                         if (rpipe->pipe_buffer.out >= rpipe->pipe_buffer.size)
 426                                 rpipe->pipe_buffer.out = 0;
 427
 428                         rpipe->pipe_buffer.cnt -= size;
 429
 430                         /*
 431                          * If there is no more to read in the pipe, reset
 432                          * its pointers to the beginning.  This improves
 433                          * cache hit stats.
 434                          */
 435                         if (rpipe->pipe_buffer.cnt == 0) {
 436                                 rpipe->pipe_buffer.in = 0;
 437                                 rpipe->pipe_buffer.out = 0;
 438                         }
 439                         nread += size;
 440 #ifndef PIPE_NODIRECT
 441                 /*
 442                  * Direct copy, bypassing a kernel buffer.
 443                  */
 444                 } else if ((size = rpipe->pipe_map.cnt) &&
 445                            (rpipe->pipe_state & PIPE_DIRECTW)) {
 446                         caddr_t va;
 447                         if (size > (u_int) uio->uio_resid)
 448                                 size = (u_int) uio->uio_resid;
 449
 450                         va = (caddr_t) rpipe->pipe_map.kva +
 451                             rpipe->pipe_map.pos;
 452                         error = uiomove(va, size, uio);
 453                         if (error)
 454                                 break;
 455                         nread += size;
 456                         rpipe->pipe_map.pos += size;
 457                         rpipe->pipe_map.cnt -= size;
 458                         if (rpipe->pipe_map.cnt == 0) {
 459                                 rpipe->pipe_state &= ~PIPE_DIRECTW;
 460                                 wakeup(rpipe);
 461                         }
 462 #endif
 463                 } else {
 464                         /*
 465                          * detect EOF condition
 466                          * read returns 0 on EOF, no need to set error
 467                          */
 468                         if (rpipe->pipe_state & PIPE_EOF)
 469                                 break;
 470
 471                         /*
 472                          * If the "write-side" has been blocked, wake it up now.
 473                          */
 474                         if (rpipe->pipe_state & PIPE_WANTW) {
 475                                 rpipe->pipe_state &= ~PIPE_WANTW;
 476                                 wakeup(rpipe);
 477                         }
 478
 479                         /*
 480                          * Break if some data was read.
 481                          */
 482                         if (nread > 0)
 483                                 break;
 484
 485                         /*
 486                          * Unlock the pipe buffer for our remaining processing.  We
 487                          * will either break out with an error or we will sleep and
 488                          * relock to loop.
 489                          */
 490                         pipeunlock(rpipe);
 491
 492                         /*
 493                          * Handle non-blocking mode operation or
 494                          * wait for more data.
 495                          */
 496                         if (fp->f_flag & FNONBLOCK) {
 497                                 error = EAGAIN;
 498                         } else {
 499                                 rpipe->pipe_state |= PIPE_WANTR;
 500                                 if ((error = tsleep(rpipe, PRIBIO | PCATCH,
 501                                     "piperd", 0)) == 0)
 502                                         error = pipelock(rpipe, 1);
 503                         }
 504                         if (error)
 505                                 goto unlocked_error;
 506                 }
 507         }
 508         pipeunlock(rpipe);
 509
 510         if (error == 0)
 511                 vfs_timestamp(&rpipe->pipe_atime);
 512 unlocked_error:
 513         --rpipe->pipe_busy;
 514
 515         /*
 516          * PIPE_WANT processing only makes sense if pipe_busy is 0.
 517          */
 518         if ((rpipe->pipe_busy == 0) && (rpipe->pipe_state & PIPE_WANT)) {
 519                 rpipe->pipe_state &= ~(PIPE_WANT|PIPE_WANTW);
 520                 wakeup(rpipe);
 521         } else if (rpipe->pipe_buffer.cnt < MINPIPESIZE) {
 522                 /*
 523                  * Handle write blocking hysteresis.
 524                  */
 525                 if (rpipe->pipe_state & PIPE_WANTW) {
 526                         rpipe->pipe_state &= ~PIPE_WANTW;
 527                         wakeup(rpipe);
 528                 }
 529         }
 530
 531         if ((rpipe->pipe_buffer.size - rpipe->pipe_buffer.cnt) >= PIPE_BUF)
 532                 pipeselwakeup(rpipe);
 533
 534         return (error);
 535 }
 536
 537 #ifndef PIPE_NODIRECT
 538 /*
 539  * Map the sending processes' buffer into kernel space and wire it.
 540  * This is similar to a physical write operation.
 541  */
 542 static int
 543 pipe_build_write_buffer(wpipe, uio)
 544         struct pipe *wpipe;
 545         struct uio *uio;
 546 {
 547         u_int size;
 548         int i;
 549         vm_offset_t addr, endaddr, paddr;
 550
 551         size = (u_int) uio->uio_iov->iov_len;
 552         if (size > wpipe->pipe_buffer.size)
 553                 size = wpipe->pipe_buffer.size;
 554
 555         endaddr = round_page((vm_offset_t)uio->uio_iov->iov_base + size);
 556         addr = trunc_page((vm_offset_t)uio->uio_iov->iov_base);
 557         for (i = 0; addr < endaddr; addr += PAGE_SIZE, i++) {
 558                 vm_page_t m;
 559
 560                 if (vm_fault_quick((caddr_t)addr, VM_PROT_READ) < 0 ||
 561                     (paddr = pmap_kextract(addr)) == 0) {
 562                         int j;
 563
 564                         for (j = 0; j < i; j++)
 565                                 vm_page_unwire(wpipe->pipe_map.ms[j], 1);
 566                         return (EFAULT);
 567                 }
 568
 569                 m = PHYS_TO_VM_PAGE(paddr);
 570                 vm_page_wire(m);
 571                 wpipe->pipe_map.ms[i] = m;
 572         }
 573
 574 /*
 575  * set up the control block
 576  */
 577         wpipe->pipe_map.npages = i;
 578         wpipe->pipe_map.pos =
 579             ((vm_offset_t) uio->uio_iov->iov_base) & PAGE_MASK;
 580         wpipe->pipe_map.cnt = size;
 581
 582 /*
 583  * and map the buffer
 584  */
 585         if (wpipe->pipe_map.kva == 0) {
 586                 /*
 587                  * We need to allocate space for an extra page because the
 588                  * address range might (will) span pages at times.
 589                  */
 590                 wpipe->pipe_map.kva = kmem_alloc_pageable(kernel_map,
 591                         wpipe->pipe_buffer.size + PAGE_SIZE);
 592                 amountpipekva += wpipe->pipe_buffer.size + PAGE_SIZE;
 593         }
 594         pmap_qenter(wpipe->pipe_map.kva, wpipe->pipe_map.ms,
 595                 wpipe->pipe_map.npages);
 596
 597 /*
 598  * and update the uio data
 599  */
 600
 601         uio->uio_iov->iov_len -= size;
 602         uio->uio_iov->iov_base += size;
 603         if (uio->uio_iov->iov_len == 0)
 604                 uio->uio_iov++;
 605         uio->uio_resid -= size;
 606         uio->uio_offset += size;
 607         return (0);
 608 }
 609
 610 /*
 611  * unmap and unwire the process buffer
 612  */
 613 static void
 614 pipe_destroy_write_buffer(wpipe)
 615         struct pipe *wpipe;
 616 {
 617         int i;
 618
 619         if (wpipe->pipe_map.kva) {
 620                 pmap_qremove(wpipe->pipe_map.kva, wpipe->pipe_map.npages);
 621
 622                 if (amountpipekva > MAXPIPEKVA) {
 623                         vm_offset_t kva = wpipe->pipe_map.kva;
 624                         wpipe->pipe_map.kva = 0;
 625                         kmem_free(kernel_map, kva,
 626                                 wpipe->pipe_buffer.size + PAGE_SIZE);
 627                         amountpipekva -= wpipe->pipe_buffer.size + PAGE_SIZE;
 628                 }
 629         }
 630         for (i = 0; i < wpipe->pipe_map.npages; i++)
 631                 vm_page_unwire(wpipe->pipe_map.ms[i], 1);
 632         wpipe->pipe_map.npages = 0;
 633 }
 634
 635 /*
 636  * In the case of a signal, the writing process might go away.  This
 637  * code copies the data into the circular buffer so that the source
 638  * pages can be freed without loss of data.
 639  */
 640 static void
 641 pipe_clone_write_buffer(wpipe)
 642         struct pipe *wpipe;
 643 {
 644         int size;
 645         int pos;
 646
 647         size = wpipe->pipe_map.cnt;
 648         pos = wpipe->pipe_map.pos;
 649         bcopy((caddr_t) wpipe->pipe_map.kva + pos,
 650             (caddr_t) wpipe->pipe_buffer.buffer, size);
 651
 652         wpipe->pipe_buffer.in = size;
 653         wpipe->pipe_buffer.out = 0;
 654         wpipe->pipe_buffer.cnt = size;
 655         wpipe->pipe_state &= ~PIPE_DIRECTW;
 656
 657         pipe_destroy_write_buffer(wpipe);
 658 }
 659
 660 /*
 661  * This implements the pipe buffer write mechanism.  Note that only
 662  * a direct write OR a normal pipe write can be pending at any given time.
 663  * If there are any characters in the pipe buffer, the direct write will
 664  * be deferred until the receiving process grabs all of the bytes from
 665  * the pipe buffer.  Then the direct mapping write is set-up.
 666  */
 667 static int
 668 pipe_direct_write(wpipe, uio)
 669         struct pipe *wpipe;
 670         struct uio *uio;
 671 {
 672         int error;
 673
 674 retry:
 675         while (wpipe->pipe_state & PIPE_DIRECTW) {
 676                 if (wpipe->pipe_state & PIPE_WANTR) {
 677                         wpipe->pipe_state &= ~PIPE_WANTR;
 678                         wakeup(wpipe);
 679                 }
 680                 wpipe->pipe_state |= PIPE_WANTW;
 681                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdww", 0);
 682                 if (error)
 683                         goto error1;
 684                 if (wpipe->pipe_state & PIPE_EOF) {
 685                         error = EPIPE;
 686                         goto error1;
 687                 }
 688         }
 689         wpipe->pipe_map.cnt = 0;        /* transfer not ready yet */
 690         if (wpipe->pipe_buffer.cnt > 0) {
 691                 if (wpipe->pipe_state & PIPE_WANTR) {
 692                         wpipe->pipe_state &= ~PIPE_WANTR;
 693                         wakeup(wpipe);
 694                 }
 695
 696                 wpipe->pipe_state |= PIPE_WANTW;
 697                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwc", 0);
 698                 if (error)
 699                         goto error1;
 700                 if (wpipe->pipe_state & PIPE_EOF) {
 701                         error = EPIPE;
 702                         goto error1;
 703                 }
 704                 goto retry;
 705         }
 706
 707         wpipe->pipe_state |= PIPE_DIRECTW;
 708
 709         error = pipe_build_write_buffer(wpipe, uio);
 710         if (error) {
 711                 wpipe->pipe_state &= ~PIPE_DIRECTW;
 712                 goto error1;
 713         }
 714
 715         error = 0;
 716         while (!error && (wpipe->pipe_state & PIPE_DIRECTW)) {
 717                 if (wpipe->pipe_state & PIPE_EOF) {
 718                         pipelock(wpipe, 0);
 719                         pipe_destroy_write_buffer(wpipe);
 720                         pipeunlock(wpipe);
 721                         pipeselwakeup(wpipe);
 722                         error = EPIPE;
 723                         goto error1;
 724                 }
 725                 if (wpipe->pipe_state & PIPE_WANTR) {
 726                         wpipe->pipe_state &= ~PIPE_WANTR;
 727                         wakeup(wpipe);
 728                 }
 729                 pipeselwakeup(wpipe);
 730                 error = tsleep(wpipe, PRIBIO | PCATCH, "pipdwt", 0);
 731         }
 732
 733         pipelock(wpipe,0);
 734         if (wpipe->pipe_state & PIPE_DIRECTW) {
 735                 /*
 736                  * this bit of trickery substitutes a kernel buffer for
 737                  * the process that might be going away.
 738                  */
 739                 pipe_clone_write_buffer(wpipe);
 740         } else {
 741                 pipe_destroy_write_buffer(wpipe);
 742         }
 743         pipeunlock(wpipe);
 744         return (error);
 745
 746 error1:
 747         wakeup(wpipe);
 748         return (error);
 749 }
 750 #endif
 751
 752 static int
 753 pipe_write(fp, uio, cred, flags, p)
 754         struct file *fp;
 755         struct uio *uio;
 756         struct ucred *cred;
 757         struct proc *p;
 758         int flags;
 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, PRIBIO | 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, PRIBIO | 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(fp, cmd, data, p)
1028         struct file *fp;
1029         u_long cmd;
1030         caddr_t data;
1031         struct proc *p;
1032 {
1033         struct pipe *mpipe = (struct pipe *)fp->f_data;
1034
1035         switch (cmd) {
1036
1037         case FIONBIO:
1038                 return (0);
1039
1040         case FIOASYNC:
1041                 if (*(int *)data) {
1042                         mpipe->pipe_state |= PIPE_ASYNC;
1043                 } else {
1044                         mpipe->pipe_state &= ~PIPE_ASYNC;
1045                 }
1046                 return (0);
1047
1048         case FIONREAD:
1049                 if (mpipe->pipe_state & PIPE_DIRECTW)
1050                         *(int *)data = mpipe->pipe_map.cnt;
1051                 else
1052                         *(int *)data = mpipe->pipe_buffer.cnt;
1053                 return (0);
1054
1055         case FIOSETOWN:
1056                 return (fsetown(*(int *)data, &mpipe->pipe_sigio));
1057
1058         case FIOGETOWN:
1059                 *(int *)data = fgetown(mpipe->pipe_sigio);
1060                 return (0);
1061
1062         /* This is deprecated, FIOSETOWN should be used instead. */
1063         case TIOCSPGRP:
1064                 return (fsetown(-(*(int *)data), &mpipe->pipe_sigio));
1065
1066         /* This is deprecated, FIOGETOWN should be used instead. */
1067         case TIOCGPGRP:
1068                 *(int *)data = -fgetown(mpipe->pipe_sigio);
1069                 return (0);
1070
1071         }
1072         return (ENOTTY);
1073 }
1074
1075 int
1076 pipe_poll(fp, events, cred, p)
1077         struct file *fp;
1078         int events;
1079         struct ucred *cred;
1080         struct proc *p;
1081 {
1082         struct pipe *rpipe = (struct pipe *)fp->f_data;
1083         struct pipe *wpipe;
1084         int revents = 0;
1085
1086         wpipe = rpipe->pipe_peer;
1087         if (events & (POLLIN | POLLRDNORM))
1088                 if ((rpipe->pipe_state & PIPE_DIRECTW) ||
1089                     (rpipe->pipe_buffer.cnt > 0) ||
1090                     (rpipe->pipe_state & PIPE_EOF))
1091                         revents |= events & (POLLIN | POLLRDNORM);
1092
1093         if (events & (POLLOUT | POLLWRNORM))
1094                 if (wpipe == NULL || (wpipe->pipe_state & PIPE_EOF) ||
1095                     (((wpipe->pipe_state & PIPE_DIRECTW) == 0) &&
1096                      (wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt) >= PIPE_BUF))
1097                         revents |= events & (POLLOUT | POLLWRNORM);
1098
1099         if ((rpipe->pipe_state & PIPE_EOF) ||
1100             (wpipe == NULL) ||
1101             (wpipe->pipe_state & PIPE_EOF))
1102                 revents |= POLLHUP;
1103
1104         if (revents == 0) {
1105                 if (events & (POLLIN | POLLRDNORM)) {
1106                         selrecord(p, &rpipe->pipe_sel);
1107                         rpipe->pipe_state |= PIPE_SEL;
1108                 }
1109
1110                 if (events & (POLLOUT | POLLWRNORM)) {
1111                         selrecord(p, &wpipe->pipe_sel);
1112                         wpipe->pipe_state |= PIPE_SEL;
1113                 }
1114         }
1115
1116         return (revents);
1117 }
1118
1119 static int
1120 pipe_stat(fp, ub, p)
1121         struct file *fp;
1122         struct stat *ub;
1123         struct proc *p;
1124 {
1125         struct pipe *pipe = (struct pipe *)fp->f_data;
1126
1127         bzero((caddr_t)ub, sizeof(*ub));
1128         ub->st_mode = S_IFIFO;
1129         ub->st_blksize = pipe->pipe_buffer.size;
1130         ub->st_size = pipe->pipe_buffer.cnt;
1131         ub->st_blocks = (ub->st_size + ub->st_blksize - 1) / ub->st_blksize;
1132         ub->st_atimespec = pipe->pipe_atime;
1133         ub->st_mtimespec = pipe->pipe_mtime;
1134         ub->st_ctimespec = pipe->pipe_ctime;
1135         /*
1136          * Left as 0: st_dev, st_ino, st_nlink, st_uid, st_gid, st_rdev,
1137          * st_flags, st_gen.
1138          * XXX (st_dev, st_ino) should be unique.
1139          */
1140         return (0);
1141 }
1142
1143 /* ARGSUSED */
1144 static int
1145 pipe_close(fp, p)
1146         struct file *fp;
1147         struct proc *p;
1148 {
1149         struct pipe *cpipe = (struct pipe *)fp->f_data;
1150
1151         fp->f_ops = &badfileops;
1152         fp->f_data = NULL;
1153         funsetown(cpipe->pipe_sigio);
1154         pipeclose(cpipe);
1155         return (0);
1156 }
1157
1158 static void
1159 pipe_free_kmem(cpipe)
1160         struct pipe *cpipe;
1161 {
1162
1163         if (cpipe->pipe_buffer.buffer != NULL) {
1164                 if (cpipe->pipe_buffer.size > PIPE_SIZE)
1165                         --nbigpipe;
1166                 amountpipekva -= cpipe->pipe_buffer.size;
1167                 kmem_free(kernel_map,
1168                         (vm_offset_t)cpipe->pipe_buffer.buffer,
1169                         cpipe->pipe_buffer.size);
1170                 cpipe->pipe_buffer.buffer = NULL;
1171         }
1172 #ifndef PIPE_NODIRECT
1173         if (cpipe->pipe_map.kva != NULL) {
1174                 amountpipekva -= cpipe->pipe_buffer.size + PAGE_SIZE;
1175                 kmem_free(kernel_map,
1176                         cpipe->pipe_map.kva,
1177                         cpipe->pipe_buffer.size + PAGE_SIZE);
1178                 cpipe->pipe_map.cnt = 0;
1179                 cpipe->pipe_map.kva = 0;
1180                 cpipe->pipe_map.pos = 0;
1181                 cpipe->pipe_map.npages = 0;
1182         }
1183 #endif
1184 }
1185
1186 /*
1187  * shutdown the pipe
1188  */
1189 static void
1190 pipeclose(cpipe)
1191         struct pipe *cpipe;
1192 {
1193         struct pipe *ppipe;
1194
1195         if (cpipe) {
1196
1197                 pipeselwakeup(cpipe);
1198
1199                 /*
1200                  * If the other side is blocked, wake it up saying that
1201                  * we want to close it down.
1202                  */
1203                 while (cpipe->pipe_busy) {
1204                         wakeup(cpipe);
1205                         cpipe->pipe_state |= PIPE_WANT | PIPE_EOF;
1206                         tsleep(cpipe, PRIBIO, "pipecl", 0);
1207                 }
1208
1209                 /*
1210                  * Disconnect from peer
1211                  */
1212                 if ((ppipe = cpipe->pipe_peer) != NULL) {
1213                         pipeselwakeup(ppipe);
1214
1215                         ppipe->pipe_state |= PIPE_EOF;
1216                         wakeup(ppipe);
1217                         KNOTE(&ppipe->pipe_sel.si_note, 0);
1218                         ppipe->pipe_peer = NULL;
1219                 }
1220                 /*
1221                  * free resources
1222                  */
1223                 pipe_free_kmem(cpipe);
1224                 zfree(pipe_zone, cpipe);
1225         }
1226 }
1227
1228 /*ARGSUSED*/
1229 static int
1230 pipe_kqfilter(struct file *fp, struct knote *kn)
1231 {
1232         struct pipe *cpipe = (struct pipe *)kn->kn_fp->f_data;
1233
1234         switch (kn->kn_filter) {
1235         case EVFILT_READ:
1236                 kn->kn_fop = &pipe_rfiltops;
1237                 break;
1238         case EVFILT_WRITE:
1239                 kn->kn_fop = &pipe_wfiltops;
1240                 cpipe = cpipe->pipe_peer;
1241                 if (cpipe == NULL)
1242                         /* other end of pipe has been closed */
1243                         return (EBADF);
1244                 break;
1245         default:
1246                 return (1);
1247         }
1248         kn->kn_hook = (caddr_t)cpipe;
1249
1250         SLIST_INSERT_HEAD(&cpipe->pipe_sel.si_note, kn, kn_selnext);
1251         return (0);
1252 }
1253
1254 static void
1255 filt_pipedetach(struct knote *kn)
1256 {
1257         struct pipe *cpipe = (struct pipe *)kn->kn_hook;
1258
1259         SLIST_REMOVE(&cpipe->pipe_sel.si_note, kn, knote, kn_selnext);
1260 }
1261
1262 /*ARGSUSED*/
1263 static int
1264 filt_piperead(struct knote *kn, long hint)
1265 {
1266         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1267         struct pipe *wpipe = rpipe->pipe_peer;
1268
1269         kn->kn_data = rpipe->pipe_buffer.cnt;
1270         if ((kn->kn_data == 0) && (rpipe->pipe_state & PIPE_DIRECTW))
1271                 kn->kn_data = rpipe->pipe_map.cnt;
1272
1273         if ((rpipe->pipe_state & PIPE_EOF) ||
1274             (wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1275                 kn->kn_flags |= EV_EOF;
1276                 return (1);
1277         }
1278         return (kn->kn_data > 0);
1279 }
1280
1281 /*ARGSUSED*/
1282 static int
1283 filt_pipewrite(struct knote *kn, long hint)
1284 {
1285         struct pipe *rpipe = (struct pipe *)kn->kn_fp->f_data;
1286         struct pipe *wpipe = rpipe->pipe_peer;
1287
1288         if ((wpipe == NULL) || (wpipe->pipe_state & PIPE_EOF)) {
1289                 kn->kn_data = 0;
1290                 kn->kn_flags |= EV_EOF;
1291                 return (1);
1292         }
1293         kn->kn_data = wpipe->pipe_buffer.size - wpipe->pipe_buffer.cnt;
1294         if (wpipe->pipe_state & PIPE_DIRECTW)
1295                 kn->kn_data = 0;
1296
1297         return (kn->kn_data >= PIPE_BUF);
1298 }